content
stringlengths 1
15.9M
|
|---|
\section{Introduction}\label{sec:introduction}
It is known that a speaker's voice contains {{personal traits}} of the speaker, given the unique pronunciation organs and speaking manner of the speaker, e.g. the unique vocal tract shape, larynx size, accent, and rhythm \cite{kinnunen2010overview}. Therefore, it is possible to identify a speaker from his/her voice automatically via a computer. This technology is termed as \textit{automatic speaker recognition}, which is the core topic of this paper. We do not discuss speaker recognition by humans. Speaker recognition is a fundamental task of speech processing, and finds its wide applications in real-world scenarios. For example, it is used for the voice-based authentication of personal smart devices, such as cellular phones, vehicles, and laptops. It guarantees the transaction security of bank trading and remote payment. It has been widely applied to forensics for investigating a suspect to be guilty or non-guilty \cite{kinnunen2010overview,campbell2009forensic,champod2000inference}, or surveillance and automatic identity tagging \cite{togneri2011overview}. It is important in audio-based information retrieval for broadcast news, meeting recordings and telephone calls. It can also serve as a frontend of automatic speech recognition (ASR) for improving the transcription performance of multi-speaker conversations.
\begin{figure*}[t]
\centering
\includegraphics[width=18cm]{veri-iden-diar.pdf}
\caption{Flowcharts of speaker verification, speaker identification, and speaker diarization. Fig. A describes speaker verification, which is a task of verifying whether a test utterance and an enrollment utterance are uttered by the same speaker via comparing the similarity score of the utterances with a pre-defined threshold. Fig. B describes speaker identification, which is a task of determining the speaker identity of a test utterance from a set of speakers. If the utterance must be produced from the set of the speakers, then it is a closed set identification problem; otherwise, it is an open set problem. Fig. C describes speaker diarization, which addresses the problem of ``who spoke when'', i.e., partitioning a conversation recording into several speech recordings, each of which belongs to a single speaker.}
\label{fig:veri-iden-diar}
\end{figure*}
The research on speaker recognition can be dated back to at least 1960s \cite{pruzansky1964talker}. In the following forty years, many advanced technologies promoted the development of speaker recognition.
For example, a number of acoustic features (e.g. the linear predictive cepstral coefficients, the perceptual linear prediction coefficient, and the mel-frequency cepstral coefficients) and template models (e.g. vector quantization, and dynamic time warping) have been applied, see \cite{kinnunen2010overview} for the details. Later on, \cite{reynolds2000speaker} proposed the Gaussian mixture model based universal background model (GMM-UBM), which has been the foundation of speaker recognition for more than ten years since year 2000.
Several representative models based on GMM-UBM have been developed, including the applications of support vector machines \cite{campbell2006support} and joint factor analysis \cite{kenny2007joint}. Among the models, the GMM-UBM/i-vector frontend \cite{dehak2010front} with probabilistic linear discriminant analysis (PLDA) backend \cite{kenny2010bayesian,garcia2011analysis} provided the state-of-the-art performance for several years, until the new era of deep learning based speaker recognition.
Recently, motivated by the powerful feature extraction capability of deep neural networks (DNNs), a lot of deep learning based speaker recognition methods were proposed \cite{lei2014novel,variani2014deep,snyder2018x} right after the great success of deep learning based speech recognition, which significantly boosts the performance of speaker recognition to a new level, even in wild environments \cite{nagrani2017voxceleb,mclaren2016speakers}.
In this survey article, we give a comprehensive overview to the deep learning based speaker recognition methods in terms of the vital subtasks and research topics, including speaker verification, identification, diarization, and robust speaker recognition. By doing the survey, we hope to provide a useful resource for the speaker recognition community. The main contributions of this paper are summarized as follows:
\begin{itemize}
\item We summarize deep learning based speaker feature extraction techniques for speaker verification and identification, from the aspects of inputs, network structures, temporal pooling strategies, and objective functions which are also the fundamental components of many other speaker recognition subtasks beyond speaker verification and identification.
\item We make an overview to the deep learning based speaker diarization, with an emphasis of recent supervised, end-to-end, and online diarization.
\item We survey robust speaker recognition from the perspectives of domain adaptation and speech enhancement, which are two major approaches to deal with domain mismatch and noise problems.
\end{itemize}
In the last two decades, many excellent overviews on speaker recognition have been published. This paper is fundamentally different from previous overviews. First, this paper focuses on the recently development of deep learning based speaker recognition techniques, while most previous overviews are based on traditional speaker recognition methods \cite{kinnunen2010overview,hansen2015speaker,togneri2011overview,reynolds2002overview,fazel2011overview,wu2015spoofing,anguera2012speaker}.
Although \cite{das2020attacker,irum2019speaker} summarized deep learning based speaker recognition methods in certain aspects, our paper summarizes different subtasks and topics from new perspectives. Specifically, \cite{das2020attacker} presents an overview to the potential threats of adversarial attacks to speaker verification as well as the spoofing countermeasures, which is not the focus of this overview. We provide a broad and comprehensive overview to a wide aspects of speaker verification, speaker diarization, domain adaptation, {{most}} of which have not been mentioned in \cite{irum2019speaker}.
This article is targeted at three categories of readers: The beginners who wish to study speaker recognition, the researchers who want to learn the whole picture of speaker recognition based on deep learning, and the engineers who need to understand or implement specific algorithms for their speaker recognition related products. In addition, we assume that the readers have basic knowledge of speech signal processing, machine leaning and pattern recognition.
The rest of the survey is organized as follows. In Section \ref{sec:overview}, we give a general overview and define some notations. In Sections \ref{sec:DNN/i-vector} to \ref{sec:robust}, we survey the deep learning based speaker recognition methods in various aspects. In Section \ref{sec:data_tool}, we summarize some speaker recognition challenges and publicly available data. Finally, we conclude this article in Section \ref{sec:discussion_conclusion}.
\section{Overview and scope}
\label{sec:overview}
This overview summarizes four major research branches of speaker recognition, which are speaker verification, identification, diarization, and robust speaker recognition respectively. The flowcharts of the first three branches are described in Fig. \ref{fig:veri-iden-diar}, while robust speaker recognition deals with the challenges of noise and domain mismatch problems. The contents of the overview are organized in Fig. \ref{fig:overview}, which are described briefly as follows.
\begin{figure*}[t]
\centering
\includegraphics[width=6.5in]{overview-eps-converted-to.pdf}
\caption{ Overview of deep learning based speaker recognition.}
\label{fig:overview}
\end{figure*}
\textit{Speaker verification} aims at verifying whether an utterance is pronounced by a hypothesized speaker based on his/her pre-recorded utterances. Speaker verification algorithms can be categorized into \textit{stage-wise} and \textit{end-to-end} ones. A stage-wise speaker verification system usually consists of a front-end for the extraction of speaker features and a back-end for the similarity calculation of speaker features.
The front-end transforms an utterance in time domain or time-frequency domain into a high-dimensional feature vector. It accounts for the recent advantage of the deep learning based speaker recognition.
We survey the research on the front-end comprehensively in Sections \ref{sec:DNN/i-vector} to \ref{sec:embedding_objectives}.
The back-end first calculates a similarity score between enrollment and test speaker features and then compare the score with a threshold:
\begin{equation}\label{eq:verifi_def}
f(\mathbf{x}^e, \mathbf{x}^t;\mathbf{w}) \mathop{\gtrless}\limits_{H_1}^{H_0} \xi
\end{equation}
where $f(\cdot)$ denotes a function for calculating the similarity, $\mathbf{w}$ stands for the parameters of the back-end, $\mathbf{x}^e$ and $\mathbf{x}^t$ are the enrollment and test speaker features respectively, $\xi$ is the threshold, $H_0$ represents the hypothesis of $\mathbf{x}^e$ and $\mathbf{x}^t$ belonging to the same speaker, and $H_1$ is the opposite hypothesis of $H_0$. One of the major responsibilities of the back-end is to compensate the channel variability and reduce interferences, e.g. language mismatch.
Because most back-ends aim at alleviating the interferences, which belongs to the problem of robust speaker recognition, we put the overview of the back-ends in Section \ref{sec:robust}.
In contrast to the stage-wise techniques, {{end-to-end}} speaker verification takes a pair of speech utterances as the input, and produces their similarity score directly.
Because a fundamental difference between the end-to-end speaker verification and the deep embedding techniques in the stage-wise speaker verification is the loss function, we mainly summarize the loss functions of the end-to-end speaker verification in Section \ref{sec:end2end_loss}.
\textit{Speaker identification} aims at detecting the speaker identity of a test utterance $\mathbf{x}^t$ from an enrollment database $\{\mathbf{x}_k^e|k=1,2,\cdots,K\}$ by\footnote{Although some work used all speakers in a given database for both training and test which is essentially regarded as a close-set speaker classification problem \cite{nagrani2017voxceleb}, most real world speaker recognition systems must be able to ``enroll'' and ``test'' new speakers dynamically.}:
\begin{equation}
\label{eq:ident_def}
k^{*} = \arg \max_{k} \{f(\mathbf{x}^e_1, \mathbf{x}^t; \mathbf{w}), f(\mathbf{x}^e_2, \mathbf{x}^t; \mathbf{w}), \cdots, f(\mathbf{x}^e_K, \mathbf{x}^t; \mathbf{w})\}
\end{equation}
where $K>1$ denotes the number of the enrollment speakers. If $\mathbf{x}^t$ can never be out of the $K$ registered speakers, then the speaker identification problem is a {{closed set problem}}; otherwise, it is an open set problem. Comparing \eqref{eq:verifi_def} with \eqref{eq:ident_def}, we see that speaker verification is a special case of the open set speaker identification problem with $K=1$, therefore, it is possible that the fundamental techniques of speaker identification and verification are similar, as what we have observed in \cite{yadav2018learning,wang2019centroid,ji2018end,9053640,9053152}. Taking this point into consideration, we make a joint overview to speaker verification and identification with an emphasis on the former.
\textit{Speaker diarization} addresses the problem of ``who spoke when'',
which is a process of partitioning a conversation recording into several speech recordings, each of which belongs to a single speaker.
As shown in Fig. \ref{fig:overview}, a conventional framework of speaker diarization is stage-wise with multiple modules.
Although the stage-wise speaker verification and diarization share some common modules, e.g. voice activity detection and speaker feature extraction, they have many differences. First, speaker verification assumes that each utterance belongs to a single speaker, while the number of speakers of a conversation in speaker diarization changes case by case. Moreover, speaker verification has an explicit registration/enrollment procedure, while speaker diarization intends to detect speakers on-the-fly without an enrollment procedure. At last, overlapped speech is one of the biggest challenges of speaker diarization, while speaker verification usually assumes that the enrollment or test utterance contains a single speaker only. Therefore, we focus on reviewing the work on the above distinguished properties of the stage-wise speaker diarization in Section \ref{sec:diarization}.
Recently, end-to-end speaker diarization, which outputs the diarization result directly, attracted much attention. Online speaker diarization, which meets the requirement of real-world applications, is also an emerging direction. Furthermore, multimodal speaker diarization, which integrates speech with video or text signals, was also studied extensively.
We review the aforementioned end-to-end, online, and multimodal speaker diarization techniques in Section \ref{sec:diarization}.
Besides, speech is easily contaminated by additive noise, reverberation, channel distortions. {{Therefore,}} robust speaker recognition is also one of the main topics. It mainly includes speech enhancement and domain adaptation techniques, which will be summarized in detail in Section \ref{sec:robust}. At last, we survey benchmark corpora in Section \ref{sec:data_tool}.
To summarize, the aforementioned contents will be organized as listed in Table \ref{tab:contens}. The notations are summarized in Table \ref{tab:notations}.
\renewcommand\arraystretch{1.5}
\begin{table}[t]
\footnotesize
\centering
\caption{Organization of the contents of this paper.}\label{tab:contens}
\begin{tabular}{m{1.5cm}<{\centering} m{5.5cm}}
\hline
\hline
\textbf{ Sections} & \textbf{Contents} \\
\hline
\ref{sec:introduction}, \ref{sec:overview}& Introduction and brief overview. \\
\ref{sec:DNN/i-vector}, \ref{sec:embedding}, \ref{sec:embedding_input_network}, \ref{sec:embedding_pooling}, \ref{sec:embedding_objectives} & Speaker feature extraction. \\
\ref{sec:end2end_loss} & The loss functions of the end-to-end speaker verification. \\
\ref{sec:diarization} & Speaker diarization. \\
\ref{sec:robust} & Robust speaker recognition. \\
\ref{sec:data_tool} & Benchmark corpora. \\
\ref{sec:discussion_conclusion} & Conclusions and discussions.\\
\hline
\hline
\end{tabular}
\end{table}
\renewcommand\arraystretch{1.5}
\begin{table}[t]
\footnotesize
\centering
\caption{Summary of the notations in this paper.}
\label{tab:notations}
\scalebox{1}{
\begin{tabular}{p{1.5cm} p{6.2cm}}
\hline
\hline
\textbf{Notation} & \textbf{Description}\\
\hline
$\mathbb{R}$ & Set of real numbers \\
$\mathbb{R}^d$ & Set of $d$-dimensional real-valued vectors\\
$\mathbb{R}^{d_1 \times d_2}$ & Set of $d_1 \times d_2$ real-valued matrices\\
\hline
$\mathcal{Y} $ & Set of acoustic features \\
$\mathcal{H} $ & Set of the last frame-level hidden layer's outputs \\
$\mathcal{E}$ & Set of embeddings \\
$\mathcal{X} $ & Set of the inputs to loss functions \\
$\mathcal{L} $ & The symbol of loss functions \\
\hline
$\mathbf{y}$ & A frame acoustic feature \\
$\mathbf{h}$ & A hidden feature of the last frame-level layer's output \\
$\mathbf{e}$ & An embedding feature of the embedding layer's output\\
$\mathbf{x}$ & An input feature to loss functions \\
$\mathbf{u}$ & An output of the temporal pooling layer \\
\hline
$t,\, T$ & Index and total number of frames in an utterance \\
$i,\, I$ & Index and total number of the utterance \\
$j,\, J$ & Index and total number of speakers in the training set\\
\hline
$\|\cdot\|$ & The $\ell_2$ norm \\
$\odot$ & The Hadamard product \\
$\delta(\cdot)$ & Indicator function \\
$(\cdot)^T$ & The transform of matrix or vector\\
\hline
\hline
\end{tabular}
}
\end{table}
\section{Speaker feature extraction with DNN/i-vector}\label{sec:DNN/i-vector}
In this section, we first introduce two main streams of the deep learning based improvement to the i-vector framework in Section \ref{sec:GMM-UBM-ivector}, and then comprehensively review the two streams in Sections \ref{sec:DNN-UBM} and \ref{sec:DNN-BNF} respectively. Finally, we make some discussions to the DNN/i-vector in Section \ref{sec:discuss_DNN/i-vector}.
\begin{figure*}[!ht]
\centering
\includegraphics[width=7in]{IvectorFramework-eps-converted-to.pdf}
\put(-318,-5){\small\bfseries\color{black}{$\uparrow$}}
\put(-335,-15){\small\bfseries\color{black}{$\{\omega_c, \bm{\mu}_c, \bm{\Sigma}_c\}_{c=1}^{C}$}}
\put(-85,-5){\small\bfseries\color{black}{$\uparrow$}}
\put(-88,-15){\small\bfseries\color{black}{$\{\mathbf{T}_c\}_{c=1}^{C}$}}
\put(-373,-4){\small\bfseries\color{black}{$\uparrow$}}
\put(-380,-15){\small\bfseries\color{black}{$\{\mathbf{y}_t^{(i)}\}_{t=1}^T$}}
\put(-23,-4){\small\bfseries\color{black}{$\uparrow$}}
\put(-25,-15){\small\bfseries\color{black}{$\bm{\upsilon}^{(i)}$}}
\put(-142,-4){\small\bfseries\color{black}{$\uparrow$}}
\put(-163,-15){\small\bfseries\color{black}{$\{N_c^{(i)},\mathbf{f}_c^{(i)}, \mathbf{\widetilde{f}}_c^{(i)} \}_{c=1}^C$}}
\caption{The traditional GMM/i-vector framework. The term MFCC denotes Mel-frequency cepstral coefficent.}
\label{fig:IvectorFramework}
\end{figure*}
\subsection{From GMM/i-vector to DNN/i-vector}
\label{sec:GMM-UBM-ivector}
The performance of the conventional GMM-UBM based speaker recognition is largely affected by the speaker and channel variations of utterances. To address this issue, \cite{dehak2010front} proposed to reduce the high-dimensional GMM-UBM supervectors into low-dimensional vectors, named \textit{i-vectors} by factor analysis. The GMM/i-vector system eliminates the within-speaker and channel variabilities effectively, which leads to significant performance improvement.
The GMM/i-vector system is shown in Fig. \ref{fig:IvectorFramework}. We assume that $\mathcal{Y}=\{\mathbf{y}_t^{(i)} \in \mathbb{R}^{d_1}|t=1,2,\cdots,T\}$ represents the $i$th ($i=1,2,\cdots,I$ ) utterance of $T$ successive Mel-frequency cepstral coefficent (MFCC) frames, and $\Omega =\{\omega_c \in \mathbb{R}, \bm{\mu}_c \in \mathbb{R}^{d_1},\bm{\Sigma}_c \in \mathbb{R}^{d_1 \times d_1}| c=1,2,\cdots, C\}$ ($\omega_c \geq 0$ for all $c$, and $\sum_{c=1}^C \omega_c = 1$) denotes a GMM-UBM model where $C$ is the total number of components and $\omega_c$, $\bm{\mu}_c$ and $\bm{\Sigma}_c$ are the weight, mean, and
covariance matrix of the $c$th component respectively. Then, $\mathbf{y}_t^{(i)}$ is assumed to be generated by the following distribution \cite{lei2014novel, snyder2020x}:
\begin{equation}
\mathbf{y}_t^{(i)} \sim \sum_{c=1}^{C} \omega_c N(\bm{\mu}_c+\mathbf{T}_c \bm{\upsilon}^{(i)}, \bm{\Sigma}_c )
\end{equation}
where $\{\mathbf{T}_c\}_{c=1}^{C}$ is a so called total variability subspace, $\bm{\upsilon}^{(i)}$ is a segment-specific standard normal-distributed latent vector. The i-vector used to represent the speech signal is the maximum a posterior (MAP) point estimate of the latent vector $\bm{\upsilon}^{(i)}$, and it can be regarded as a kind of ``\textit{speaker embedding}''\footnote{In this paper, the 'embedding' denotes the problem of learning a vector space where speakers are ``embedded''. The i-vectors, d-vectors (introduced in Section \ref{sec:d-vector}), and x-vectors (introduced in Section \ref{sec:x-vector}) are different embedding models for learning the vector spaces.}.
Given a speech segment, the following sufficient statistics can be accumulated from the GMM-UBM:
\begin{equation}
\label{eq:BW0}
N_c^{(i)}= \sum_{t=1}^{T} p(c|\mathbf{y}_t^{(i)})
\end{equation}
\begin{equation}
\label{eq:BW1}
\mathbf{f}_c^{(i)}= \sum_{t=1}^{T} p(c|\mathbf{y}_t^{(i)})\mathbf{y}_t^{(i)}
\end{equation}
\begin{equation}
\label{eq:BW2}
\mathbf{S}_c^{(i)}= \sum_{t=1}^{T} p(c|\mathbf{y}_t^{(i)})\mathbf{y}_t^{(i)}(\mathbf{y}_t^{(i)})^T
\end{equation}
where $ p(c|\mathbf{y}_t^{(i)})= \frac{\omega_c N(\mathbf{y}_t^{(i)};\bm{\mu}_c,\bm{\Sigma}_c)}{\sum_{c^\prime=1}^{C}\omega_c N(\mathbf{y}_t^{(i)};\bm{\mu}_{c^\prime},\bm{\Sigma}_{c^\prime})} $ denotes the posterior probability of $\mathbf{y}_t^{(i)}$ against the $c$th Gaussian component. These sufficient statistics are all that are needed to train the subspace $\{\mathbf{T}_c\}_{c=1}^{C}$ and extract the i-vector $\bm{\upsilon}^{(i)}$ \cite{lei2014novel}. See \cite{dehak2010front,kenny2008study} for the details of training $\mathbf{T}_c$ and estimating the i-vectors.
Motivated by the success of deep learning for speech recognition, many efforts have been made to replace the GMM-UBM module of the GMM/i-vector system by DNN, which can be categorized to two main streams---DNN-UBM/i-vector and DNN based bottleneck feature (DNN-BNF)/i-vector. The two main streams will be presented in detail in the following two subsections, with selected references summarized in Table \ref{tab:DNN_ivector}.
\begin{table}[t]
\centering
\footnotesize
\caption{Two main streams of the DNN/i-vector techniques. }
\label{tab:DNN_ivector}
\begin{tabular}{m{2.3cm}m{5.3cm}}
\hline
\hline
\textbf{Approaches} & \textbf{References}\\
\hline
DNN-UBM/i-vector &\cite{lei2014novel} \cite{dey2016deep} \cite{zeinali2016deep} \cite{lei2014deep} \cite{dey2017exploiting} \cite{zeinali2017text} \cite{richardson2015deep} \cite{kenny2014deep} \cite{snyder2015time} \cite{mclaren2015advances} \cite{chen2015phone} \cite{richardson2015unified} \cite{sadjadi2016ibm} \cite{mclaren2014application} \cite{garcia2015insights} \cite{zheng2015exploring} \\
DNN-BNF/i-vector & \cite{ghalehjegh2015deep} \cite{zeinali2016deep} \cite{richardson2015deep} \cite{mclaren2015advances} \cite{richardson2015unified} \cite{mclaren2016exploring} \cite{lozano2016analysis} \cite{do2013augmenting} \cite{sarkar2014combination} \\
\hline
\hline
\end{tabular}
\end{table}
\begin{figure}[t]
\centering
\includegraphics[width=2.5in]{DNN-UBM-ivector.pdf}
\caption{DNN-UBM/i-vector (from \cite{lei2014novel}). The posteriors are produced from the DNN acoustic model of an automatic speech recognition (ASR) system that is trained with, e.g. Logmel filterbank features. On the contrary, the sufficient statistics are computed from a speaker verification (SV) system that is trained with, e.g. MFCC which is not necessarily the same as the features for ASR. That is to say, one does not have to find a feature that works well for both ASR and SV in this framework.}
\label{fig:DNN-ASR/i-vector}
\end{figure}
\subsection{DNN-UBM/i-vector}
\label{sec:DNN-UBM}
From \eqref{eq:BW0}, \eqref{eq:BW1}, and \eqref{eq:BW2}, one can see that only the posteriors of speech frames are needed to collect sufficient statistics for producing the i-vectors. Thus, we can use any probabilistic models beyond GMM-UBM to produce the posteriors theoretically \cite{lei2014novel}. Motivated by this insight, \cite{lei2014novel} proposed the DNN-UBM/i-vector framework (Fig.\ref{fig:DNN-ASR/i-vector}) which takes a DNN acoustic model trained for ASR, denoted as DNN-UBM, to generate the posterior probabilities instead of GMM-UBM.
Specifically, DNN-UBM uses a set of senones $\mathcal{Q}=\{Q_c|c=1,2,\cdots,C\}$, e.g., the tied-triphone states, to mimic the mixture components of the GMM-UBM. It first trains a DNN-based ASR acoustic model to align each training frame with a senone, and then generates the posterior probabilities of each frame over the senones from the softmax output layer of the DNN acoustic model. The posteriors can be directly applied to \eqref{eq:BW0}, \eqref{eq:BW1} and \eqref{eq:BW2} to extract the DNN-UBM based i-vector. Due to the strong representation ability of DNN over GMM, DNN-UBM/i-vector yields 30\% relative equal error rate (EER) reduction over GMM/i-vector on the telephone condition of the 2012 NIST speaker recognition evaluation (SRE) \cite{lei2014novel}. Later on, the authors in \cite{lei2014deep,mclaren2015advances,mclaren2014application} further analyzed the performance of the DNN-UBM/i-vector in microphone and noisy conditions.
A lot of further studies bloomed the DNN-UBM/i-vector related techniques. For example, \cite{richardson2015deep,richardson2015unified} proposed to use a single ASR-DNN for both the speaker and language recognition tasks simultaneously. Additionally, \cite{snyder2015time} employed a time delay deep neural network (TDNN), which was originally applied to speech recognition, to compute the posteriors. It achieved the state-of-the-art performance on the NIST SRE10 corpus at the time. As a third instance, \cite{zheng2015exploring} replaced the feedforward DNN by a long short-term memory (LSTM) recurrent neural network (RNN). The last but not all, \cite{garcia2015insights} studied a number of open issues relating to performance, computational complexity, and applicability of different types of DNNs.
The advantage of the DNN acoustic model may be brought by its strong ability in modeling content-related phonetic states explicitly, which not only generates highly compact representation of data but also provides precise frame alignment. This advantage is particularly apparent in text-dependent speaker verification \cite{chen2015phone,dey2016deep,dey2017exploiting,zeinali2016deep,zeinali2017text}.
However, this comes at the cost of greatly increased computational complexity over the traditional GMM-UBM/i-vector systems \cite{snyder2015time, snyder2017deep}, since that a DNN usually has more parameters than GMM. In addition, the training of the DNN based acoustic model requires a large number of labeled training data.
To overcome the computational complexity, a supervised GMM-UBM was also investigated based on the DNN acoustic model \cite{snyder2015time}. In specific, a GMM is obtained by:
\begin{equation}\label{eq:dnn-asr-poster}
\begin{split}
\gamma_{ct}^{(i)} = & p(c|\mathbf{\overline{y}}_t^{(i)}) \\
\omega_{c} = & \sum_{i,t}\gamma_{ct}^{(i)} \\
\bm{\mu}_c = & \frac{\sum_{i,t}\gamma_{ct}^{(i)}\mathbf{y}_t^{(i)}}{\sum_{i,t}\gamma_{ct}^{(i)}} \\
\bm{\Sigma}_c = & \frac{\sum_{i,t}\gamma_{ct}^{(i)}\mathbf{y}_t^{(i)}(\mathbf{y}_t^{(i)})^T}{\sum_{i,t}\gamma_{ct}^{(i)}}-\bm{\mu}_c\bm{\mu}_c^T\\
\end{split}
\end{equation}
where $\mathbf{\overline{y}}_t^{(i)}$ and $\mathbf{y}_t^{(i)}$ denote the acoustic features for ASR and speaker recognition respectively, and $p(c|\mathbf{\overline{y}}_t^{(i)})$ is the posterior probability corresponding to the $c$th senones. By this way, the
supervised-GMM maintains the training computational complexity of the traditional unsupervised-GMM, with a 20\% relative EER reduction on the NIST
SRE10 corpus \cite{snyder2015time}. Similar idea was also studied in \cite{lei2014novel}, though no performance improvement over the baseline is observed. Although the supervised-GMM reduces the training computational complexity, training the DNN acoustic model still needs a large amount of labeled training data.
\subsection{DNN-BNF/i-vector}
\label{sec:DNN-BNF}
The fundamental idea of DNN-BNF/i-vector is to extract a compact feature from the bottleneck layer of a DNN as the input of the factor analysis, where the bottleneck layer is a special hidden layer of the DNN that has much less hidden units than the other hidden layers. In practice, DNN-BNF/i-vector has many variants, as we have summarized in Fig. \ref{fig:DNN-BNF/i-vector}. Like DNN-UBM/i-vector, the deep model in DNN-BNF/i-vector is mainly trained to discriminate senones \cite{zeinali2016deep,mclaren2015advances,mclaren2016exploring,lozano2016analysis,richardson2015deep,richardson2015unified} or phonemes \cite{do2013augmenting,sarkar2014combination}.
The input of the factor analysis can be either the bottleneck feature (BNF) produced from the bottleneck layer, a concatenation of BNF with other acoustic feature \cite{do2013augmenting,sarkar2014combination}, or a post-processed feature by principal components analysis (PCA) or linear discriminant analysis (LDA) \cite{do2013augmenting,sarkar2014combination}. One can find that no matter whether we apply BNF alone \cite{richardson2015unified} or concatenate it with other acoustic features \cite{mclaren2015advances}, DNN-BNF/i-vector can significantly outperform the conventional GMM/i-vector, which indicates the effectiveness of the framework \cite{richardson2015deep}.
\begin{figure}[t]
\centering
\includegraphics[width=3.2in]{DNN-BNF-ivector-eps-converted-to.pdf}
\caption{The DNN-BNF/i-vector framework. The framework is a summarization of related work. The term ``PCA'', ``LDA'', and ``BNF'' are short for principal components analysis, linear discriminant analysis, and bottleneck feature respectively.}
\label{fig:DNN-BNF/i-vector}
\end{figure}
However, it is unclear why a deep model trained to discriminate phonemes or senones can produce speaker-sensitive BNF. To address this issue, {{the authors of}} \cite{mclaren2016exploring} assumed that speaker information is traded for dense phonetic information when the bottleneck layer moves toward the DNN output layer. Under this hypothesis, they experimentally analyzed the role of BNF by placing the bottleneck layer at different depths of the DNN. They found that, if the training and test conditions match, the closer the bottleneck layer is to the output layer, the better the performance is; otherwise, the bottleneck layer should be placed around the middle of the DNN.
{{The authors of}} \cite{lozano2016analysis} explored whether weakening the accuracy of the acoustic model on speech recognition yields better BNF for speaker recognition. They analyzed the speaker recognition performance in different respects of the acoustic model, including
under-trained DNN, different inputs, and different feature normalization strategies. Results indicate that high speech recognition performance in terms of phonetic accuracy does not necessarily imply {{increased speaker recognition accuracy}}.
In addition, {{\cite{ghalehjegh2015deep}}} proposed to take speaker identity as the training target, under the conjecture that this training target should be able to improve the robustness of the phonetic variability of BNF.
\subsection{Discussion to the DNN/i-vector}
\label{sec:discuss_DNN/i-vector}
\renewcommand\arraystretch{1.5}
\begin{table}[t]
\centering
\footnotesize
\caption{Comparison results between DNN/i-vector and GMM/i-vector. Each row denotes a comparison. The last three columns list the EER of the main models, the EER of the baselines, and the relative EER reductions, respectively. \textbf{The results across rows are not comparable}, since they are collected from different references, and their comparisons are not apple-to-apple comparisons.}
\label{tab:experiments_DNN/i-vector}
\scalebox{0.75}{
\begin{tabular}{m{2.45cm} m{1.45cm} m{2.5cm}<{\centering} m{0.8cm}<{\centering} m{0.8cm}<{\centering} m{1.2cm}<{\centering}}
\hline
\hline
\multicolumn{2}{c}{Comparisons} & \multirow{2}*{Test dataset [condition]} &\multicolumn{3}{c}{EER}\\
\cline{1-2} \cline{4-6}
Main models & Baselines & & Main & Baseline & Relative reduction \\
\hline
DNN-UBM \cite{lei2014novel} & GMM-UBM & NIST SRE12 C2 & 1.39\% & 1.81\% & \textbf{23\%}\\
DNN-UBM \cite{lei2014novel} & GMM-UBM & NIST SRE12 C5 & 1.92\% & 2.55\% & \textbf{25\%}\\
\hline
TDNN-UBM\cite{snyder2015time}& GMM-UBM & NIST SRE10 C5 & 1.20\% & 2.42\% & \textbf{50\%}\\
Sup-GMM-UBM\cite{snyder2015time}& GMM-UBM & NIST SRE10 C5 & 1.94\% & 2.42\% & \textbf{20\%}\\
\hline
BNF \cite{richardson2015deep}& MFCC &In-domain DAC13 & 2.00\% &2.71\% & \textbf{26\%}\\
BNF \cite{richardson2015deep}& MFCC &Out-domain DAC13 & 2.79\% &6.18\% & \textbf{55\%}\\
\hline
\hline
\end{tabular}}
\end{table}
It is known that a major difference between DNN-UBM and GMM-UBM is that DNN-UBM is a discriminant model, while GMM-UBM is a generative one. {{DNN-UBM is more powerful than GMM-UBM in modeling a complicated data distribution \cite{lei2014novel, snyder2015time}}}. Moreover, the DNN acoustic model is trained to align each speech frame to its corresponding senone in a supervised fashion. Its output nodes have a clear physical explanation. It mines the pronunciation characteristics of speakers. On the contrary, GMM-UBM is trained by the expectation-maximum algorithm in an unsupervised manner. Its mixtures have no inherent meaning.
Although DNN-UBM/i-vector needs labeled training data and heavier computation power than GMM-UBM, it does yield excellent performance. In addition, many corpora are also developed for the demand of training strong DNN, which will be reviewed in Section \ref{sec:data_tool}.
{{To demonstrate general performance differences of DNN/i-vector and GMM/i-vector,}} some carefully selected experimental results from literatures are listed in Table \ref{tab:experiments_DNN/i-vector}. Compared to the GMM-UBM/i-vector baseline, one can find that DNN-UBM/i-vector achieves more than 20\% relative EER reduction over GMM-UBM/i-vector. In addition, the supervised GMM-UBM in \eqref{eq:dnn-asr-poster} can also get 20\% relative improvement according to the fourth row. Finally, from the last two rows, one can see that, when taking DNN-BNF and MFCC as the input features of the GMM-UBM/i-vector respectively, the former achieves better performance than the latter.
It should be note that, as far as we know, different test conditions may yield slightly different conclusions from those in Table \ref{tab:experiments_DNN/i-vector}. However, to our knowledge, the results in the table can be a representative of the research trend.
\section{Speaker feature extraction with deep embedding}
\label{sec:embedding}
In this section, we first introduce two representative deep embeddings---d-vector and x-vector in Section \ref{subsec:twoseminal} with some discussions in Section \ref{subsec:discussion}, and then identify their key components in Section \ref{subsec:four}, which provide a taxonomy to existing algorithms.
\subsection{Two seminal work of deep embeddings}\label{subsec:twoseminal}
\subsubsection{Frame-level embedding --- d-vector}
\label{sec:d-vector}
D-vector is one of the earliest DNN-based embeddings \cite{variani2014deep}. The core idea of d-vector is to assign the ground-truth speaker identity of a training utterance as the labels of the training frames belonging to the utterance in the training stage, which transforms the model training as a classification problem. As shown in Fig. \ref{fig:d-vector}, d-vector expands each training frame with its context, and employs a \textit{maxout} DNN to classify the frames of a training utterance to the speaker identity of the utterance, where the DNN takes softmax as the output layer to minimize the cross-entropy loss between the ground-truth labels of the frames and the network output.
In the test stage, d-vector takes the output activation of each frame from the last hidden layer of the DNN as the deep embedding feature of the frame, and averages the deep embedding features of all frames of an utterance as a new compact representation of the utterance, named \textit{d-vector}.
An underlying hypothesis of d-vector is that the compact representation space produced from a development set may generalize well to unseen speakers in the test stage.
\begin{figure}[t]
\centering
\includegraphics[width=2.75in]{d-vector-eps-converted-to.pdf}
\caption{Diagram of the d-vector framework (from \cite{variani2014deep}). }
\label{fig:d-vector}
\end{figure}
\subsubsection{Segment-level embedding --- x-vector}
\label{sec:x-vector}
\begin{figure}[t]
\centering
\includegraphics[width=2.85in]{x-vector-eps-converted-to.pdf}
\caption{Diagram of the DNN model for extracting x-vectors (from \cite{snyder2017deep}). Note that segment-level embeddings (e.g., $\mathbf{a}$ or $\mathbf{b}$) can be extracted from any layer of the network after the statistics pooling layer \cite{snyder2017deep}. \cite{snyder2018x} where the name ``x-vector'' comes from uses the embedding \textbf{a} as the speaker feature. }
\label{fig:x-vector}
\end{figure}
X-vector \cite{snyder2017deep,snyder2018x} is an important evolution of d-vector that evolves speaker recognition from frame-by-frame speaker labels to utterance-level speaker labels with an aggregation process. The network structure of x-vector is shown in Fig. \ref{fig:x-vector}.
It first extracts frame-level embeddings of speech frames by time-delay layers, then concatenates the mean and standard deviation of the frame-level embeddings of an utterance as a segment-level (a.k.a., utterance-level) feature by a statistical pooling layer, and finally classifies the segment-level feature to its speaker by a standard feedforward network. The time-delay layers, statistical pooling layer, and feedforward network are jointly trained.
\textit{X-vector} is defined as the segment-level speaker embedding produced from the second to last hidden layer of the feedforward network, i.e. the variable $\mathbf{a}$ in Fig. \ref{fig:x-vector}.
{{The authors in \cite{snyder2018x}}} found that data augmentation is important in improving the performance of x-vector. We will introduce the data augmentation techniques in Section \ref{sec:data_augmentation}.
\subsection{Discussion to the speaker embedding}\label{subsec:discussion}
\renewcommand\arraystretch{1.5}
\begin{table*}[t]
\footnotesize
\centering
\caption{Characteristics of different speaker embeddings and their favorite back-ends. }
\label{tab:comparsion_speaker_embedding}
\begin{tabular}{m{2.5cm}| m{3.2cm}<{\centering} m{3.2cm}<{\centering} m{2.1cm}<{\centering} m{1.2cm}<{\centering} m{2.5cm}<{\centering}}
\hline \hline
Model name & Type & Training strategy & Label for model training & Back-end name & Label for back-end training \\
\hline
GMM-UBM/i-vector & Generative/Generative & Unsupervised/Unsupervised & \ding{55} / \ding{55} & PLDA &Speaker identity\\
DNN-UBM/i-vector & Discriminative/Generative & Supervised /Unsupervised & Phonetic labels/\ding{55} & PLDA &Speaker identity\\
DNN-BNF/i-vector & Discriminative/Generative & Supervised/Unsupervised & Phonetic labels /\ding{55} & PLDA &Speaker identity \\
D-vector & Discriminative & Supervised & Speaker identity & Cosine& \ding{55} \\
X-vector & Discriminative & Supervised & Speaker identity & PLDA &Speaker identity \\
\hline
\hline
\end{tabular}
\end{table*}
Similar to the i-vector, the d-vector and x-vector are also a kind of speaker embedding, which discriminatively embeds speakers into a vector space by using DNNs. We call this type of speaker embedding as {{\textit{deep speaker embedding}, or \textit{deep embedding} for short}}. The main characteristics between different speaker embeddings are summarized in Table \ref{tab:comparsion_speaker_embedding}. Compared to the traditional GMM-UBM/i-vector, the deep embedding is a discriminant model and trained in a supervised fashion. Compared to DNN-UBM, its training data does not need phonetic-level labels. Therefore, the training of the deep embedding is much simpler than that of DNN-UBM and DNN-BNF. In addition, the deep embedding is a new framework, while DNN-UBM/i-vector and DNN-BNF/i-vector are hybrid ones.
\renewcommand\arraystretch{1.5}
\begin{table*}[t]
\centering
\footnotesize
\caption{Selected results on deep embedding in literature. Each row represents a comparison. The results across rows are not comparable.}
\label{tab:experiments_deep_embedding}
\scalebox{0.9}{
\begin{tabular}{m{3.5cm}<{\centering} m{2.45cm}<{\centering} m{5cm}<{\centering} m{2cm}<{\centering} m{2cm}<{\centering} m{1.8cm}<{\centering}}
\hline
\hline
\multicolumn{2}{c}{Comparison methods} & \multirow{2}*{Test dataset [condition]} &\multicolumn{3}{c}{EER}\\
\cline{1-2} \cline{4-6}
Deep embedding & Baseline & & Deep embedding & Baseline & Relative reduction \\
\hline
d-vector \cite{variani2014deep} & GMM-UBM/i-vector & Google data & 4.54\% & 2.83\% & \textbf{-37\%} \\
d-vector+i-vector \cite{variani2014deep} & GMM-UBM/i-vector & Google data [clean , noisy] & ----- & ----- & \textbf{[14\% , 25\%]} \\
\hline
embedding \textbf{a+b} (in Fig.\ref{fig:x-vector}) \cite{snyder2017deep} & GMM-UBM/i-vector & NIST SRE10 [10s-10s , 60s ] & [7.9\% , 2.9\%] &[11.0\% , 2.3\% ] & \textbf{[28\% , -21\% ]} \\
embedding \textbf{a+b} (in Fig.\ref{fig:x-vector}) \cite{snyder2017deep} & GMM-UBM/i-vector & NIST SRE16 [Cantonese , Tagalog ] & [6.5\% , 16.3\% ] &[8.3\% , 17.6\% ] & \textbf{[22\% , 7\%]} \\
\hline
x-vector (embedding \textbf{a}) \cite{snyder2018x} & GMM-UBM/i-vector & SITW Core [ PLDA and extractor aug. , Incl. VoxCeleb] & [ 6.00\% , 4.16\% ] & [ 8.04\% , 7.45\%] & \textbf{[25\%, 44\%]} \\
x-vector (embedding \textbf{a}) \cite{snyder2018x} & GMM-UBM/i-vector & SRE16 Cantonese [ PLDA and extractor aug. , Incl. VoxCeleb] & [ 5.86\%, 5.71\% ] & [ 8.95\% , 9.23\%] & \textbf{[ 34\%, 38\%]} \\
\hline
\hline
\end{tabular}}
\end{table*}
Some experimental results on deep embedding are listed in Table \ref{tab:experiments_deep_embedding}. From the table, one can find that, the d-vector alone yields higher EER than the i-vector. When fusing the d-vector and i-vector, the combined system achieves 14\% and 25\% relative EER reduction in clean and noisy test conditions respectively over the i-vector. The ``embedding a+b'' model, which is the predecessor of the x-vector, achieves lower EER than the GMM-UBM/i-vector baseline on the 10-second short utterances of NIST SRE10, and higher EER than the latter on the 60-second long utterances of NIST SRE10. With enlarged training data and data augmentation, the x-vector achieves significant performance improvement over the GMM-UBM/i-vector.
Fig. \ref{fig:Embedding_time_line} shows the number of the related papers. {{We observe the following phenomena.}} First, the d-vector and DNN-UBM/i-vector was proposed both in 2014, where the former achieved better performance at the time. Second, the research on DNN/i-vector was mainly conducted in the first few years after its appearance, and then became less studied. Third, the research on deep embedding becomes bloom along with its performance improvement after that the x-vector achieved the state-of-the-art performance. At present, the deep embedding is the trend of speaker recognition, which has been developed in several aspects as summarized in the following subsection.
\begin{figure}[t]
\centering
\includegraphics[width=3.2in]{Embedding_time_line.pdf}
\put(-170,52){\small\bfseries\color{black}{\begin{turn}{90} \scriptsize d-vector \cite{variani2014deep} \end{turn}}}
\put(-169,42){\small\bfseries\color{black}{$\uparrow$}}
\put(-178,60){\small\bfseries\color{black}{\begin{turn}{90}\scriptsize DNN-UBM/i-vector \cite{lei2014novel}\end{turn}}}
\put(-177,53){\small\bfseries\color{black}{$\uparrow$}}
\put(-80,105){\small\bfseries\color{black}{\scriptsize x-vector \cite{snyder2018x} }}
\put(-114,65){\small\bfseries\color{black}{\scriptsize Embedding \cite{snyder2017deep}}}
\put(-91,54){\small\bfseries\color{black}{$\uparrow$}}
\put(-65,95){\small\bfseries\color{black}{$\uparrow$}}
\caption{Statistics of the published papers on DNN/i-vector and deep embedding cited by this article.}
\label{fig:Embedding_time_line}
\end{figure}
\subsection{Four key components of deep embedding}\label{subsec:four}
Motivated by the seminal work d-vector and x-vector, many deep embedding techniques were proposed, most of which are composed of four key components---network input, network structure, temporal pooling, and training objective. These components include but not limited to the following contents:
\begin{itemize}
\item \textit{Network inputs and structures}:
The network input can be categorized into two classes---raw wave signals in time domain and acoustic features in time-frequency domain, including spectrogram, Mel-filterbanks (f-bank), and MFCC. The network structure is diverse, which is rooted essentially at DNN, RNN/LSTM, and CNN. Because the network input and structure were jointly designed case by case in practice, we will jointly summarize them in Section \ref{sec:embedding_input_network}.
\item \textit{Temporal pooling}: Temporal pooling represents the transition layer of a neural network that transforms frame-level embedding features to utterance-level embedding features. The temporal pooling strategies consist of two classes---statistical pooling and learning based pooling. We will introduce them in Section \ref{sec:embedding_pooling}.
\item \textit{Objective functions}: Objective functions affect the effectiveness of speaker recognition much. Both d-vector and x-vector adopt softmax as the output layer and take the cross-entropy minimization as the objective function, which may not be optimal. Recently, many objectives were designed to further improve the performance. We will survey the objective functions in Section \ref{sec:embedding_objectives}.
\end{itemize}
\section{Deep embedding: network structures and inputs}
\label{sec:embedding_input_network}
\begin{table*}[t]
\centering
\footnotesize
\caption{A brief summary of the inputs and neural network structures in deep speaker feature extraction.}
\label{tab:networks_inputs_summary}
\scalebox{1}{
\begin{tabular}{m{2cm} <{\centering}| m{6cm} | m{2cm} <{\centering} | m{5.5cm} }
\hline
\hline
\textbf{Inputs}&\makecell[c]{ \textbf{CNN } }& \makecell[c]{\textbf{LSTM } } & \makecell[c]{\textbf{Hybrid structures}} \\
\hline
\textbf{Wave} &Others \cite{muckenhirn2018towards, ravanelli2018speaker}. &------ &CNN-LSTM \cite{jung2018avoiding, jung2018complete}; CNN-GRU \cite{jung2019rawnet, jung2019short}.\\
\textbf{Spectrogram}& ResNet \cite{yu2019ensemble, chung2018voxceleb2, xie2019utterance, 9054440}; VGGNet \cite{nagrani2017voxceleb,
yadav2018learning}; Inception-resnet-v1 \cite{zhang2017end, zhang2018text}. &------ & CNN-GRU \cite{zhang2019seq2seq} \\
\textbf{F-bank} & TDNN \cite{snyder2018x,9053198,Zhu2020Orthogonality,9053209}; ResNet \cite{9053217,li2017deep,kim2019deep,Garcia-Romero2020MagNetO}; VGGNet \cite{bhattacharya2017deep}; Inception-resnet-v1 \cite{zhang2018text,li2018deep,li2019boundary}; Others \cite{torfi2018text,li2017deepspeaker}. & \cite{rahman2018attention,wan2018generalized,heigold2016end}. &BLSTM-ResNet \cite{9053767}, TDNN-LSTM \cite{tang2019deep} \\
\textbf{MFCC} &TDNN \cite{snyder2019speaker,snyder2017deep,liu2018speaker,zhu2018self,okabe2018attentive,bai2020partial,li2019gaussian,
xiang2019margin,9053209,villalba2019state,9054350,Li2020Bayesian}; ResNet \cite{zhou2019deep}; Others \cite{gao2018improved,jiang2019effective}. & ------ & TDNN-LSTM \cite{chen2019speaker} \\
\hline
\hline
\end{tabular}
}
\end{table*}
Although deep neural networks can be divided roughly into DNN, CNN, and RNN/LSTM structures, the network structure and input for speaker recognition are quite flexible. Each component of a network has many candidates. For example, the hidden layer of a neural network may be a standard convolutional layer \cite{bhattacharya2017deep}, a dilated convolution layer \cite{gao2018improved}, a LSTM layer \cite{jung2018avoiding}, a gated recurrent unit (GRU) layer \cite{jung2019rawnet}, a multi-head attention layer, a fully-connected layer, and even a combination of these different layers \cite{jung2018avoiding,jung2019rawnet}, etc. The activation functions can be Sigmoid, Rectified Linear Unit (ReLU), Leaky ReLU, or Parametric Rectified Linear Unit (PReLU) etc. Besides, the topology of a network and connection mode between layers are all variables. Even the number of layers and number of hidden units at a layer can also affect the performance. To prevent enumerating the networks case by case, here we first review some commonly used networks for the speaker feature extraction, and then briefly review their inputs.
\textit{Time delay neural network} (TDNN) \cite{snyder2017deep}: TDNN takes a one-dimensional convolution structure along the time axis as a feature extractor \cite{peddinti2015time}. It is adopted by the well known x-vector, as shown in Fig. \ref{fig:x-vector}. Due to the success of the x-vector \cite{snyder2018x,snyder2019speaker}, TDNN becomes one of the most popular structures for speaker recognition.
For example, \cite{liu2018speaker} introduced phonetic information to the TDNN architecture based embedding extractor.
\cite{stafylakis2019self} trained a TDNN embedding extractor without speaker labels via self-supervised training. \cite{9053198,Zhu2020Orthogonality} explored the effectiveness of the orthogonality regularization by TDNN. Generally, TDNN has been frequently used as a framework to study other key components of the deep embedding models, such as the temporal pooling layers \cite{zhu2018self, okabe2018attentive} and objective functions \cite{bai2020partial,li2019gaussian,xiang2019margin}.
The TDNN structure has also been intensively improved. For instance, an extended TDNN architecture (E-TDNN) was introduced in \cite{snyder2019speaker}, which greatly outperforms the x-vector baseline \cite{snyder2018x}. It adopts a slightly wider temporal context than TDNN, and interleaves affine layers in between the convolutional layers \cite{garcia2020jhu}. \cite{povey2018semi} developed a factorized TDNN (F-TDNN) to reduce the number of parameters. It factorizes the weight matrix of each TDNN layer into the product of two low-rank matrices. It further constrains the first low-rank matrix to be semi-orthogonal under the assumption that the semi-orthogonal constraint prevents information loss. The application of F-TDNN to deep embedding was also investigated \cite{villalba2019state, garcia2020jhu, snyder2019jhu}. Some other parameter reduction works can be found in \cite{Yu2020,Georges2020Compact}. Recently, \cite{9054350} integrated TDNN with statistics pooling at each layer for compensating the variation of temporal context in the frame-level transforms. Similarly, \cite{tang2019deep,chen2019speaker} inserted LSTM layers into TDNN to capture the temporal information for remedying the weakness of TDNN whose time delay layers focus on local patterns only. \cite{Li2020Bayesian} alleviated the mismatch problem between training and evaluation by incorporating Bayesian neural networks into TDNN.
\textit{Residual networks} (ResNet) \cite{he2016deep}: it is another popular structure in speaker embedding. Its trunk architecture is a 2-dimensional CNN with convolutions in both the time and frequency domains. Some work directly used the standard ResNet as their speaker feature extractors \cite{9053217,yu2019ensemble,chung2018voxceleb2,li2017deep}. Some other work employed ResNet as a backbone and modified it for specific purposes or applications \cite{xie2019utterance,zhou2019deep,9053767,9054440,kim2019deep,Garcia-Romero2020MagNetO}.
For example, to reduce the number of parameters, \cite{xie2019utterance} modified the standard ResNet-34 to a \textit{thin} ResNet by cutting down the number of channels in each residual block. {{The authors in \cite{9053767}}} combined bi-directional LSTM (BLSTM) and ResNet into a unified architecture, where the BLSTM is used to model long temporal contexts. {{The authors in \cite{zhou2019deep}}} incorporated a so-called ``squeeze-and-excitation'' block into ResNet.
\textit{Raw wave neural networks} \cite{muckenhirn2018towards, ravanelli2018speaker, jung2018avoiding, jung2018complete,jung2019rawnet, jung2019short, Jung2020Improved, LinWeiwei2020}: some work takes raw waves in the time domain as the input, which aims to extract learnable acoustic features instead of handcrafted features. For example, \cite{muckenhirn2018towards} applied CNN to capture raw speech signal. The experimental results indicate that the filters of the first convolution layer give emphasis to speaker information in low frequency regions. {{The authors in \cite{ravanelli2018speaker}}} believed that the first layer is critical for the waveform-based CNNs, since it not only deals with high-dimensional inputs, but suffers more from the gradient vanishing problem than the other layers. Therefore, they proposed a SincNet architecture based on parametrized sinc functions, where only low and high cutoff frequencies of band-pass filters are learned from data \cite{ravanelli2018speaker}.
In \cite{jung2018complete}, {{the authors}} thought that the difficulty of processing raw audio signals by DNN is mainly caused by the fluctuating scales of the signals. To stabilize the scales, they employed a convolutional layer, named pre-emphasis layer, to mimic the well-known signal pre-emphasis technique $p(t)=s(t)-\alpha s(t-1)$. They also made several improvements to the original raw wave network \cite{jung2019rawnet, jung2019short, Jung2020Improved} which results in excellent performance. \cite{LinWeiwei2020} designed a \textit{Wav2Spk} architecture to learn speaker embeddings from waveforms, where the traditional MFCC extraction, voice activity detection, and cepstral mean and variance normalization are replaced by a feature encoder, a temporal gating unit and an instance normalization
scheme respectively. Wav2Spk performs better than the convention x-vector network.
\textit{Other neural networks}: in addition to TDNN and ResNet, many other well-known neural network architectures have also been applied to speaker recognition, including VGGNet \cite{bhattacharya2017deep,yadav2018learning,nagrani2017voxceleb}, Inception-resnet-v1 \cite{zhang2017end, zhang2018text, li2018deep, li2019boundary}, BERT \cite{Ling2020BERTphone}, and Transformer \cite{Safari2020}.
Besides, recurrent neural networks, such as LSTM and gated recurrent units, are often used for text-dependent speaker verification \cite{rahman2018attention,wan2018generalized,heigold2016end}. The CNN models can also be improved by inserting LSTM or gated recurrent units into the backbone networks \cite{jung2018avoiding,jung2018complete,jung2019rawnet,jung2019short,zhang2019seq2seq,9053767,tang2019deep,chen2019speaker}.
Finally, apart from the above handcrafted neural architectures, neural architecture search was also recently applied to speaker recognition \cite{Ding2020AutoSpeech,Qu2020Evolutionary}.
\textit{Neural network inputs}: Table \ref{tab:networks_inputs_summary} provides a summary to the common inputs and neural networks for the deep embedding based speaker feature extraction. From the table, one can see that CNN-based neural networks and f-bank/MFCC acoustic features are becoming popular, while some 2-dimensional convolution structures, e.g. ResNet, use spectrogram as the input feature.
In addition to the above common inputs, such as MFCC, spectrum and mel-filterbanks, \cite{Liu2020Comparative} recently presented an extensive re-assessment of 14 acoustic feature extractors. They found that the acoustic features equipped with the techniques of spectral centroids, group delay function, and integrated noise suppression provide promising alternatives to MFCC.
\subsection{Discussion to the networks}
\renewcommand\arraystretch{1.5}
\begin{table}[t]
\centering
\footnotesize
\caption{Selected examples of the effect of neural network structures on performance (from \cite{villalba2020state}).}
\label{tab:experiments_network}
\scalebox{0.8}{
\begin{tabular}{m{1.8cm} m{1.35cm} m{2.5cm}<{\centering} m{0.8cm}<{\centering} m{0.8cm}<{\centering} m{1.2cm}<{\centering}}
\hline
\hline
\multicolumn{2}{c}{Comparison methods} & \multirow{2}*{Test dataset [condition]} &\multicolumn{3}{c}{EER}\\
\cline{1-2} \cline{4-6}
Main models & Baselines & & Main & Baseline & Relative reduction \\
\hline
E-TDNN(10M) & TDNN(8.5M) & SITW EVAL CORE (\textit{16 kHz systems}) & 2.74\% & 3.40\% & \textbf{19\%}\\
F-TDNN(9M) & TDNN(8.5M) & SITW EVAL CORE (\textit{16 kHz systems}) & 2.39\% & 3.40\% & \textbf{30\%}\\
F-TDNN(17M) & TDNN(8.5M) & SITW EVAL CORE (\textit{16 kHz systems}) & 1.89\% & 3.40\% & \textbf{44\%}\\
ResNet(8M) & TDNN(8.5M) & SITW EVAL CORE (\textit{16 kHz systems}) & 3.01\% & 3.40\% & \textbf{11\%}\\
\hline
\hline
\end{tabular}}
\end{table}
The network structure plays a key role on performance. For example, as shown in Table \ref{tab:experiments_network}, E-TDNN and F-TDNN significantly reduced EER on the SITW dataset \cite{mclaren2016speakers}, where F-TDNN achieves more than 40\% relative EER reduction over the original TDNN. Although this promotion is not consistent across all datasets \cite{villalba2020state}, it demonstrates the importance of the network structure on performance.
For the acoustic features, the delta and double-delta features are helpful in statistical model based speaker recognition, e.g. the GMM-UBM/i-vector. However, they are not very effective in convolution and time-delay neural networks. This may be caused by that, the statistical model needs the delta and double-delta operations to capture the time dependency between frames, while the neural networks are able to achieve this goal intrinsically.
Although the deep embedding networks have achieved a great success, {{in our view}}, the following aspects can be further studied. First, the raw wave networks did not attract much attention. The mainstream of speaker recognition still adopts handcrafted features, which may lose useful information, e.g. the phase information, and finally may result in suboptimal performance as what we have observed in speech separation.
Second, the model size and inference efficiency, which is important for the devices with limited computation source, e.g. edge or mobile devices, have not been fully studied. The topic was just recently investigated in \cite{Safari2020, Georges2020Compact, nunes2020mobilenet1d}.
\section{Deep embedding: Temporal pooling layers}
\label{sec:embedding_pooling}
As shown in Fig. \ref{fig:x-vector}, the temporal pooling layer is a bridge between the frame-level and utterance-level hidden layers. Given a speech segment, we assume that the input and output of the temporal pooling layer are $\mathcal{H}=\{\mathbf{h}_t \in \mathbb{R}^{d_2}| t=1,2,\cdots, T\}$ and $\mathbf{u}$, respectively, where $\mathbf{h}_t$ denotes the $t$th frame-level speaker feature produced from the frame-level hidden layers. In this section, we introduce a number of temporal pooling functions.
\subsection{Average pooling}
\label{sec:average_pooling}
Average pooling \cite{zhang2017end,yadav2018learning,li2017deep,li2018deep} is the most common pooling function:
\begin{equation}\label{eq:avg_pooling}
\mathbf{u}=\frac{1}{T}\sum_{t=1}^{T} \mathbf{h}_t
\end{equation}
\subsection{Statistics pooling}
Statistics pooling \cite{snyder2017deep,snyder2018x} calculates both the statistic mean $\mathbf{m}$ and standard deviation $\mathbf{d}$ of $\mathcal{H}$:
\begin{equation}\label{eq:sta_pooling_avg}
\mathbf{m} = \frac{1}{T}\sum_{t=1}^{T} \mathbf{h}_t
\end{equation}
\begin{equation}\label{eq:sta_pooling_dev}
\mathbf{d} = \sqrt{\frac{1}{T}\sum_{t=1}^{T}\mathbf{h}_t \odot \mathbf{h}_t - \mathbf{m} \odot \mathbf{m} } \\
\end{equation}
where $\odot$ denotes the Hadamard product. The output of the statistics pooling layer is a concatenation of $\mathbf{m}$ and $\mathbf{d}$, i.e. $\mathbf{u} = [\mathbf{m}^T, \mathbf{d}^T]^T$.
\subsection{Self-attention-based pooling}
Obviously, \eqref{eq:avg_pooling}, \eqref{eq:sta_pooling_avg}, and \eqref{eq:sta_pooling_dev} assume that all elements of $\mathcal{H}$ contribute equally to $\mathbf{u}$. However, the assumption may not be true, since that the frames may not provide equal speaker-discriminative information. To address this issue, many works applied self attention mechanisms for weighted statistics pooling layers.
Specifically, the attention can be broadly interpreted as a vector of importance weights\footnote{https://lilianweng.github.io/lil-log/2018/06/24/attention-attention.html}, which allows a neural network to focus on a specific portion of its input. Further more, the self attention computes attentive weights within a single sequence.
In the following two subsections, we first present a general self attention framework which produces weighted means and standard deviations of the input from a self-attentive scoring function in Section \ref{subsubsec:g_att}, and then list a number of specific self-attention-based pooling methods under the framework in Section \ref{subsubsec:att_pool}.
\subsubsection{A self attention pooling framework}\label{subsubsec:g_att}
Without loss of generality, self-attentive scoring is defined as:
\begin{equation}
\big\{f_{\rm Att}^{(k)}(\cdot)|k=1,2,\cdots, K \big\}
\end{equation}
where $f_{\rm Att}^{(k)}(\cdot)$ is usually referred as one-head, and $K$ is the total number of heads. If $K \ge 2$,
the self attention mechanism is usually called multi-head self attention which allows the model to jointly attend to information from different representation subspaces \cite{vaswani2017attention}; otherwise, it degenerates into a single-head one.
Although $f_{\rm Att}^{(k)}(\cdot)$ has many different implementations, many of the implementations share similar forms with the structured self-attentive function \cite{lin2017structured} which obtains the importance weights by:
\begin{equation}\label{eq:multi_head_atten}
f_{\rm Att}^{(k)}(\mathbf{h}_t) = \mathbf{v}^{(k)^T} {\rm tanh}(\mathbf{W}^{(k)}\mathbf{h}_t+\mathbf{g}^{(k)})+b^{(k)},\quad k=1,2, \cdots, K
\end{equation}
where $\mathbf{W}^{(k)} \in \mathbb{R}^{d_3 \times d_2}$, $\mathbf{g}^{(k)} \in \mathbb{R}^{d_3}$, $\mathbf{v}^{(k)} \in \mathbb{R}^{d_3}$ and $b^{(k)} \in \mathbb{R}$ are learnable parameters of the $k$th scoring function. Suppose
$ s_t^{(k)} = f_{\rm Att}^{(k)}(\mathbf{h}_t)$, $\quad k=1,2, \cdots, K$, then the importance weights for the frame-level feature $\mathbf{h}_t$ are obtained by normalizing $s_t^{(k)}$ with a softmax function:
\begin{equation}\label{eq:multi_head_weights}
\alpha_t^{(k)} = \frac{{\rm exp}\big(s_t^{(k)}\big)}{\sum_{t^\prime}^{T} {\rm exp}\big(s_{t^\prime}^{(k)} \big)}, \quad k=1,2, \cdots, K
\end{equation}
where the normalization guarantees that the weights satisfy $0 \leq \alpha_t^{(k)} \leq 1$ and $\sum_{t=1}^{T}\alpha_t^{(k)} =1$.
Finally, the weighted mean and standard deviation produced from the $k$th self-attentive scoring function can be derived as follows:
\begin{equation}\label{eq:atten_pooling_avg}
\widetilde{\mathbf{m}}^{(k)} = \sum_{t=1}^{T} \alpha_t^{(k)} \mathbf{h}_t, \quad k=1,2, \cdots, K
\end{equation}
\begin{equation}\label{eq:atten_pooling_dev}
\widetilde{\mathbf{d}}^{(k)}= \sqrt{\sum_{t=1}^{T} \alpha_t^{(k)} \mathbf{h}_t \odot \mathbf{h}_t - \mathbf{\widetilde{m}}^{(k)} \odot \mathbf{\widetilde{m}}^{(k)}}, \quad k=1,2, \cdots, K
\end{equation}
Finally, $\widetilde{\mathbf{m}}^{(k)}$ and $\widetilde{\mathbf{d}}^{(k)}$ are used to calculate an utterance-level representation as described in the following subsection.
\subsubsection{Attention pooling methods}\label{subsubsec:att_pool}
Under the above attention framework, this subsection categorizes existing self-attention based pooling layers into the following six classes, where all methods take \eqref{eq:multi_head_atten} as the self-attentive scoring function and take \eqref{eq:multi_head_weights} as the normalization function, unless otherwise stated.
\begin{itemize}
\item \textbf{Single-head attentive average pooling} \cite{bhattacharya2017deep, bhattacharya2018deeply, rahman2018attention}: \cite{bhattacharya2017deep} takes a fully-connected layer as $f_{\rm Att}(\cdot)$. \cite{bhattacharya2018deeply} adopts the cosine function to compute attention scores:
\begin{equation}\label{eq:saap}
s_t = f_{\rm Att}(\mathbf{h}_t,\mathbf{r}) = \frac{\mathbf{h}_t^T\mathbf{r}}{\|\mathbf{h}_t\|_2\|\mathbf{r}\|_2}
\end{equation}
where $\mathbf{r}$ is a nonlinearly transformed i-vector from the same utterance as $\mathbf{h}$. Obviously, the attention weights in \eqref{eq:saap} are determined by both the frame-level $\mathbf{h}_t$ and the utterance-level information $\mathbf{r}$. In \cite{rahman2018attention}, several attentive functions similar to \eqref{eq:multi_head_atten} are investigated.
The output of the single-head attentive average pooling is set to the weighted mean:
\begin{equation}
\mathbf{u}= \mathbf{\widetilde{m}}^{(1)}.
\end{equation}
\item \textbf{Single-head attentive statistics pooling} \cite{okabe2018attentive}: It uses a single-head attention function, i.e. $K=1$. Its output is a concatenation of both the weighted mean and weighted standard deviation:
\begin{equation}\label{eq:Single-head attentive statistics pooling}
\mathbf{u} = [\widetilde{\mathbf{m}}^{(1)^T}, \widetilde{\mathbf{d}}^{(1)^T}]^T.
\end{equation}
\item \textbf{Single-head Baum-Welch statistics attention mechanism based statistics pooling} \cite{9054151}:
To overcome the weakness of \eqref{eq:multi_head_atten} which cannot fully mine the inner relationship between an utterance and its frames, \cite{9054151} integrated the Baum-Welch statistics into the attention mechanism:
\begin{equation}
\label{eq:multi_head_score}
s_t = \mathbf{v}^T {\rm tanh}(\mathbf{K}\mathbf{q}_t+\mathbf{g})
\end{equation}
where $\mathbf{K}$ is named the key matrix and $\mathbf{q}_t$ is a query vector calculated by:
\begin{equation}
\mathbf{q}_t=f(\mathbf{h}_t^{(-1)})
\end{equation}
where $f(\cdot)$ is a nonlinear function, and $\mathbf{h}_t^{(-1)}$ denotes the output of a penultimate frame-level hidden layer.
The key matrix $ \mathbf{K}$ is calculated from the Baum-Welch statistics. Specifically, \cite{9054151} first calculates the normalized first order statistics $\mathbf{f}_c$ from the $c$th component of a GMM-UBM model $\Omega$ (see \eqref{eq:BW1}), and then conducts the following nonlinear transform:
\begin{equation}
\mathbf{f}_c^\prime=\mathbf{V}_2 {\rm tanh}(\mathbf{V}_1\mathbf{f}_c+\mathbf{g}),\quad \forall c = 1,\ldots,C
\end{equation}
where $\mathbf{V}_1$, $\mathbf{V}_2$ and $\mathbf{g}$ are the parameters of DNN.
Finally, it concatenates $\mathbf{F}^\prime=[\mathbf{f}_1^{\prime},\mathbf{f}_2^{\prime}, \cdots, \mathbf{f}_C^{\prime}]$ and the trainable matrix $\mathbf{W}$ as the key matrix:
\begin{equation}
\mathbf{K}=[\mathbf{F}^\prime, \mathbf{W}]^T
\end{equation}
After obtaining $s_t$, $\mathbf{u}$ is obtained in the same way as \eqref{eq:Single-head attentive statistics pooling}.
\item \textbf{Global multi-head attentive average pooling}:
It first applies a $K$-head ($K \geq 2$) attention function to $\mathcal{H}$ by \eqref{eq:multi_head_atten}. Then, the attentive weights and weighted means are calculated by \eqref{eq:multi_head_weights} and \eqref{eq:atten_pooling_avg} respectively \cite{9053217}.
Finally, the output of the pooling layer $\mathbf{u}$ is the concatenation of the weighted means:
\begin{equation}\label{eq:Global_multi_head}
\mathbf{u} = [\widetilde{\mathbf{m}}^{(1)^T}, \widetilde{\mathbf{m}}^{(2)^T}, \cdots, \widetilde{\mathbf{m}}^{(K)^T}]^T
\end{equation}
It can be seen that $\mathbf{u} \in \mathbb{R}^{Kd_2}$. Similar ideas can also be found in \cite{zhu2018self,zhou2019cnn}.
\cite{zhu2018self} also added an additional penalty term into the objective function to enlarge the diversity between the heads.
\item \textbf{Sub-vectors based multi-head attentive average pooling} \cite{safari2019self}:
It first splits $\mathbf{h}_t$ into $K$ ($K \geq 2$) non-overlapping homogeneous sub-vectors $\mathbf{h}_t= [\mathbf{h}_t^{(1)^T},\mathbf{h}_t^{(2)^T},\cdots,\mathbf{h}_t^{(K)^T} ]^T$, where $\mathbf{h}_t^{(k)} \in \mathbb{R}^{d_2/K}$. Then, it applies single-head attention to each of the sub-vectors
$\mathcal{H}^{(k)}=\{\mathbf{h}_t^{(k)}\in \mathbb{R}^{d_2/K}| t=1,2, \cdots, T\}$. Finally, it obtains the sub-pooling outputs by:
\begin{equation}
\mathbf{u}^{(k)}=\sum_{t=1}^{T} \alpha_t^{(k)} \mathbf{h}_t^{(k)}, \quad k=1,2, \cdots, K
\end{equation}
It can be seen that $\mathbf{u}^{(k)} \in \mathbb{R}^{d_2/K}$.
The output of the pooling layer is a concatenation of the sub-pooling outputs:
\begin{equation}
\mathbf{u} = [\mathbf{u}^{(1)^T},\mathbf{u}^{(2)^T},\cdots,\mathbf{u}^{(K)^T}]^T.
\end{equation}
\item \textbf{Multi-resolution multi-head attentive average pooling} \cite{9053217}: Because the speaker characteristics are obtained through the aggregation of the attentive weights reweighted frame-level features, \cite{9053217} proposed to control the resolution of the attentive weights with a temperature parameter. They modify the softmax function as:
\begin{equation}
\label{eq:softmax_temperature}
\alpha_t=\frac{{\rm exp}(s_t/E)}{\sum_{t^\prime=1}^{T}{\rm exp}(s_{t^\prime}/E)}
\end{equation}
where $E$ is the temperature parameter. It is obvious that increasing $E$ makes the distribution of $ \alpha_t$ less sharp, i.e. lower resolution. By incorporating the above intuition, the weighting equation \eqref{eq:multi_head_weights} is changed to:
\begin{equation}\label{eq:softmax_temperature_k}
\alpha_{t}^{(k)} = \frac{{\rm exp}\big(s_t^{(k)}/E_k\big)}{\sum_{t^\prime}^{T} {\rm exp}\big(s_{t^\prime}^{(k)}/E_k\big)}
\end{equation}
where $E_k\geq 1$ is a temperature hyperparameter of the $k$th head. Finally, the output $\mathbf{u}$ is calculated in a similar way with that of the global multi-head attentive average pooling except that $\alpha_t^{(k)}$ is replaced by \eqref{eq:softmax_temperature_k}.
\end{itemize}
It is clear that the above attentive pooling methods all employ scalar attention weights for each frame-level vector. \cite{Wu2020} further proposed a vector-based attentive pooling method, which adopts vectorial attention weights for each frame-level vector.
\subsection{NetVLAD $\&$ GhostVLAD pooling}
In \cite{xie2019utterance}, the authors applied a dictionary-based NetVLAD layer to aggregate features across time, which can be intuitively regarded as trainable discriminative clustering: every frame-level descriptor will be softly assigned to different clusters, making the residuals encoded as the output feature \cite{xie2019utterance}.
\begin{figure}[t]
\centering
\includegraphics[width=3.3in]{NetVLAD_pooling-eps-converted-to.pdf}
\caption{ Diagram of the NetVLAD pooling layer (from \cite{arandjelovic2016netvlad}).}
\label{fig:NetVLAD_pooling}
\end{figure}
Specifically, as shown in Fig. \ref{fig:NetVLAD_pooling}, suppose that the input of the NetVLAD layer is a three-dimensional tensor $\mathbf{H}^{(W \times H) \times D}$, where $W$, $H$ and $D$ depend on the speech length, the dimensions of the spectrum frequency bins, and the number of convolution kernels respectively. By only retaining the third dimension, $\mathbf{H}$ can be converted to $N$ one-dimensional tensors, i.e. $\overline{\mathcal{H}}=\{\overline{\mathbf{h}}_{n} \in \mathbb{R}^D|n=1,2,\cdots,N \}$ where $N=W\times H$.
As shown in Fig. \ref{fig:NetVLAD_pooling}, the NetVLAD pooling layer consists of the following four steps \cite{arandjelovic2016netvlad}:
\begin{enumerate}[1)]
\item Calculate a matrix $\mathbf{U} \in \mathbb{R}^{D \times M}$ from $\overline{\mathcal{H}}$ by:
\begin{equation}\label{eq:NetVLAD}
\mathbf{U}(:,m)=\sum_{n=1}^{N} \beta_{m}(\mathbf{\overline{h}}_n)(\mathbf{\overline{h}}_n-\mathbf{o}_m)
\end{equation}
where $M$ is the number of the chosen clusters $\mathcal{O}=\{\mathbf{o}_m \in \mathbb{R}^D|m=1,2,\cdots,M\}$, and $\beta_{m}(\mathbf{\overline{h}}_n)$ is an assignment weight calculated by:
\begin{equation}\label{eq:NetVLAD1}
\beta_{m}(\mathbf{\overline{h}}_n)=\frac{{\rm
exp}\big(\mathbf{w}_m^T \mathbf{\overline{h}}_n+b_m \big)}{\sum_{m^\prime=1}^{M}{\rm exp}\big(\mathbf{w}_{m^\prime}^T \mathbf{\overline{h}}_n+b_{m^\prime}\big)}
\end{equation}
with $\{\mathbf{w}_m \}$, $\{b_m\}$ and $\{\mathbf{o}_m\}$ as the parameters of the network.
\item Normalize $\mathbf{U}$ by $\ell_2$-norm column-wisely. This step is termed as the intra-normalization.
\item Convert the normalized $\mathbf{U}$ into a vector:
\begin{equation}
\mathbf{u}=[\mathbf{U}(:,1)^T,\mathbf{U}(:,2)^T,\cdots,\mathbf{U}(:,M)^T]^T
\end{equation}
\item Normalize $\mathbf{u}$ by $\ell_2$-norm to generate an $M \times D$ dimensional output vector. This step is termed as the $\ell_2$-normalization.
\end{enumerate}
In addition, \cite{xie2019utterance} also applied a variant of NetVLAD, named GhostVLAD. The main difference between them is that some of the clusters in the GhostVLAD layer, named \textit{``ghost clusters''}, are not included in the final concatenation, and hence do not contribute to the final representation. When aggregating the frame-level features, the contribution of the noisy and undesirable sections of a speech segment to the normal VLAD clusters will be effectively down-weighted, since that larger weights are assigned to the ``ghost cluster''. See \cite{zhong2018ghostvlad} for the details.
\subsection{Learnable dictionary encoding pooling}
Motivated by GMM-UBM, {{\cite{cai2018exploring}}} proposed a learnable dictionary encoding (LDE) pooling layer which models the distribution of the frame-level features $\mathcal{H}$ by a dictionary. The dictionary learns a set of dictionary component centers $\mathcal{\overline{O}}=\big \{\mathbf{\overline{o}}_m \in \mathbb{R}^{d_2}|m = 1,2,\cdots, M \big\} $, and assigns weights to the frame-level features by:
\begin{equation}
\overline{\beta}_{tm}=\frac{{\rm exp}(-\tau_m||\mathbf{h}_t-\mathbf{\overline{o}}_m||^2)}{\sum_{m^\prime=1}^{M}{\rm exp}(-\tau_{m^\prime}||\mathbf{h}_t-\mathbf{\overline{o}}_{m^\prime}||^2)}
\end{equation}
where the smoothing factor $\tau_m$ for each dictionary center $\mathbf{\overline{o}}_m$ is learnable. The aggregated output of the pooling layer with respect to the center $\mathbf{\overline{o}}_m$ is:
\begin{equation}
\label{eq:LDEP}
\mathbf{u}_m=\frac{\sum_{t=1}^{T}\overline{\beta}_{tm}(\mathbf{h}_t-\overline{\mathbf{o}}_m)}{\sum_{t=1}^{T}\overline{\beta}_{tm}}
\end{equation}
In order to facilitate the derivation, \eqref{eq:LDEP} is simplified to:
\begin{equation}
\mathbf{u}_m=\frac{\sum_{t=1}^{T}\overline{\beta}_{tm}(\mathbf{h}_t-\overline{\mathbf{o}}_m)}{T}
\end{equation}
Finally, the output of the pooling layer is $\mathbf{u}=[\mathbf{u}_1^T,\mathbf{u}_2^T,\cdots,\mathbf{u}_M^T]^T$.
\subsection{Spatial pyramid pooling}
\label{sec:spatial_pyramid_pooling}
In order to handle variable-length utterances, \cite{zhang2018text} incorporated a Spatial Pyramid Pooling operation into a CNN-based network, which can directly produce fixed-length feature vectors from variable-length utterances.
As shown in Fig. \ref{fig:pyramid_pooling}, the spatial pyramid pooling layer outputs a fixed length vector by first dividing the input feature maps into $1 \times 1$, $ 2 \times 2$, and $3 \times 3$ small patches and then performing average pooling over these patches.
An exceptional advantage of the spatial pyramid pooling layer is that it maintains spatial information of the last frame-level feature maps by making average pooling in each local small patches.
\cite{jung2019spatial} further extracted embeddings from the divided small patches via a parameter-sharing LDE layer instead of applying the averaging pooling on them.
\begin{figure}[t]
\centering
\includegraphics[width=2.5in]{pyramid_pooling-eps-converted-to.pdf}
\caption{ Spatial pyramid pooling (from \cite{zhang2018text}).}
\label{fig:pyramid_pooling}
\end{figure}
\subsection{Other temporal pooling functions}
There are many other successful pooling methods. For example, \cite{gao2018improved} proposed a cross-convolutional-layer pooling method to capture the first-order statistics for modelling long-term speaker characteristics. \cite{travadi2019total} reported a total variability model based pooling layer. \cite{heigold2016end} connected the last output of LSTM to the loss function for an utterance-level speaker representation.
Apart from the single-scale aggregation methods in Sections \ref{sec:average_pooling} to \ref{sec:spatial_pyramid_pooling} which generate the pooling output from the last frame-level layer, multiscale aggregation methods have also been proposed \cite{tang2019deep, gao2019improving, seo2019shortcut, hajavi2019deep, Jung2020Improving} which utilize multiscale features from different frame-level layers to generate the pooling output.
\subsection{Discussion to the temporal pooling layers}
Because temporal pooling layers behave fundamentally different in different datasets, network structures, or loss functions, it is difficult to conclude which one is the best. To our knowledge, temporal pooling functions with learnable parameters achieved better (at least competitive) results than the simple pooling layers such as the average pooling and statistical pooling in most cases, with a weakness of higher computational complexity than the latter. Some examples are listed in Table \ref{tab:experiments_pooling}.
\renewcommand\arraystretch{1.5}
\begin{table}[t]
\centering
\footnotesize
\caption{Experimental results of different temporal pooling functions selected from literature. Each row represents a comparison. The results across rows are not comparable.}
\label{tab:experiments_pooling}
\scalebox{0.8}{
\begin{tabular}{m{2cm} m{1.1cm} m{2.5cm}<{\centering} m{0.8cm}<{\centering} m{0.8cm}<{\centering} m{1.2cm}<{\centering}}
\hline
\hline
\multicolumn{2}{c}{Comparison methods} & \multirow{2}*{Test dataset [condition]} &\multicolumn{3}{c}{EER}\\
\cline{1-2} \cline{4-6}
Main models & Baselines & & Main & Baseline & Relative reduction \\
\hline
Attention\cite{zhu2018self} & Average &SRE16 Cantonese & 5.81\% & 7.33\% & \textbf{21\%}\\
NetVLAD \cite{xie2019utterance} & Average &VoxCeleb1 test set & 3.57\% & 10.48\% & \textbf{66\%}\\
GhostVLAD \cite{xie2019utterance}& Average & VoxCeleb1 test set& 3.22\% & 10.48\% & \textbf{69\%}\\
LDE \cite{villalba2020state} & Statistics & SITW & 2.50\% & 3.01\% & \textbf{17\%}\\
\hline
\hline
\end{tabular}}
\end{table}
\section{Deep embedding: Classification-based objective functions}
\label{sec:embedding_objectives}
The objective function largely determines the performance of a neural network. Deep-embedding-based speaker recognition systems usually adopt classification-based objective functions. Before reviewing the objective functions, we first summarize the deep-embedding-based speaker recognition as the following multi-class classification problem.
Let $\mathcal{X}=\{(\mathbf{x}_n, l_n)|n=1,2, \cdots, N\}$ denote the training samples in a mini-batch\footnote{DNN is often trained using a mini-batch data in an iteration}, where $\mathbf{x}_n \in \mathbb{R}^{d_4}$ represents the input of the last fully connected layer, $ l_n \in \{1,2,\cdots,J\}$ is the class label of $\mathbf{x}_n$ with $J$ as the number of speakers in the training set, and $N$ is the batch size.
In addition, $\mathbf{W}=[\mathbf{w}_1,\mathbf{w}_2,\cdots,\mathbf{w}_J]$ and $\mathbf{b}=[b_1,b_2,\cdots,b_J]$ denote the weight matrix and bias vector of the last fully connected layer respectively.
In this section, we comprehensively summarize the classification based objective functions. Without loss of generality, the ``loss function'',``cost function'' and ``objective function'' are equivalent in this article.
\subsection{The variants of softmax loss }\label{sec:softmax_variants}
As shown in Section \ref{sec:embedding}, both the d-vector and x-vector extractors take the minimum cross entropy as the objective function, and take softmax as the output layer. For short, we denote the objective function as the \textit{Softmax loss}\footnote{Following \cite{liu2017sphereface}, we define the softmax loss as a combination of the last fully connected layer, softmax function, and cross-entropy loss function.}.
{{For a multiclass classification problem, the cross-entropy error function over $\mathcal{X}$ can be calculated as:
\begin{equation}
\label{eq:CE}
L = -\frac{1}{N} \sum_{n=1}^{N} {\sum_{j=1}^{J} t_{nj} {\rm log} p_{nj} }
\end{equation}
where $[t_{n1}, t_{n2}, \cdots, t_{nJ}]$ is a one-hot vector encoded from the label $l_n$, in other words, $t_{nj}$ equals to 1 if and only if sample $\mathbf{x}_n$ belongs to class $j$. $p_{nj}$ is the posterior probability of $\mathbf{x}_n$ belonging to class $j$. It is produced from neural networks with the following Softmax function:
\begin{equation}
\label{eq:softmax_function}
p_{nj} = \frac{{\rm exp}{(\mathbf{w}_{j}^T \mathbf{x}_n+b_{j}})}{\sum_{j=1}^{J}{\rm exp}{(\mathbf{w}_{j}^T \mathbf{x}_n+b_{j}})}
\end{equation}
Combining \eqref{eq:CE} and \eqref{eq:softmax_function} derives an equivalent form of the Softmax loss:
\begin{equation}
\label{eq:softmax}
\mathcal{L}_{\rm S}= -\frac{1}{N}\sum_{n=1}^{N} {\rm log} \frac{{\rm exp}{(\mathbf{w}_{l_n}^T \mathbf{x}_n+b_{l_n}})}{\sum_{j=1}^{J}{\rm exp}{(\mathbf{w}_{j}^T \mathbf{x}_n+b_{j}})}
\end{equation}
}}
Softmax loss is the most common objective function for deep embedding. However, from \eqref{eq:softmax}, one can see that Softmax loss is only good at maximizing the between-class distance, but does not have an explicit constraint on minimizing the within-class variance. Therefore, the performance of deep embedding has much room of improvement. Here we present some representative variants of Softmax loss as follows.
\begin{itemize}
\item \textbf{Angular softmax (ASoftmax) loss} \cite{huang2018angular,novoselov2018deep,cai2018exploring}: Because the inner product between $ \mathbf{w}_j$ and $\mathbf{x}_n$ in \eqref{eq:softmax} can be rewritten as:
\begin{equation}
\mathbf{w}_j^T\mathbf{x}_n=\|\mathbf{w}_j\|\,\|\mathbf{x}_n\|{\rm cos}(\theta_{j,n})
\end{equation}
where $\theta_{j,n} (0 \leq \theta_{j,n} \leq \pi)$ denotes the angle between $ \mathbf{w}_j$ and $\mathbf{x}_n$, Softmax loss can be rewritten as:
\begin{equation}
\label{eq:softmax_variant}
\mathcal{L}_{\rm S}= -\frac{1}{N}\sum_{n=1}^{N} {\rm log} \frac{{\rm exp}{\big(\|\mathbf{w}_{l_n}\|\,\| \mathbf{x}_n\| {\rm cos}(\theta_{l_n,n})+b_{l_n}}\big)}{\sum_{j=1}^{J}{\rm exp}{\big(\|\mathbf{w}_{j}\|\, \|\mathbf{x}_n\|{\rm cos}(\theta_{j,n})+b_{j}}\big)}
\end{equation}
If we further set the bias terms to zero, normalize the weights at the forward propagation stage, and add a margin to the angle:
\begin{equation}
\label{eq:ASoftmax_margin}
b_j=0, \quad \|\mathbf{w}_j\|=1, \quad \psi(\theta_{l_n,n})= (-1)^{a}{\rm cos}(m_1 \theta_{l_n,n})-2a
\end{equation}
then, we explicitly constrain the learned features to have a small intra-speaker variation, where $m_1 \geq 1$ is an integer margin hyperparameter, $ \theta_{l_n,n} \in [\frac{a \pi}{m_1}, \frac{(a+1)\pi}{m_1} ]$, and $a \in \{0,1,\cdots,m_1-1\}$. The intuition behind the angle function $\psi(\theta_{l_n,n})$ is illustrated in Fig. \ref{fig:ASoftmax_margin}.
Then, we obtain ASoftmax loss as follows:
\begin{equation}
\label{eq:A-Softmax}
\begin{split}
&\mathcal{L}_{\rm AS}= \\
&-\frac{1}{N}\sum_{n=1}^{N} {\rm log} \frac{{\rm exp}{\big(\| \mathbf{x}_n\| \psi(\theta_{l_n,n})})}{{\rm exp}{\big(\| \mathbf{x}_n\| \psi(\theta_{l_n,n})})+\sum_{j=1,j\neq l_n}^{J}{\rm exp}{\big(\|\mathbf{x}_n\|{\rm cos}(\theta_{j,n})}\big)}
\end{split}
\end{equation}
Note that, because $m_1$ is limited to a positive integer instead of a real number, the margin is not flexible enough.
\begin{figure}[t]
\centering
\includegraphics[width=2.5in]{ASoftmax_margin-eps-converted-to.pdf}
\put(-162,40){\small\bfseries{------ }}
\put(-140,40){\small\bfseries\color{black}{$\psi_1(\theta_{l_n,n})= {\rm cos}(m_1 \theta_{l_n,n})$}}
\put(-162,22){\small\bfseries\color{red}{------}}
\put(-140,22){\small\bfseries\color{black}{$\psi_2(\theta_{l_n,n})= (-1)^{a}{\rm cos}(m_1 \theta_{l_n,n})-2a$}}
\put(-150,-12){\small\bfseries\color{black}{$m_1 \theta_{l_n,n}$ \textit{with} $m_1=4$ \textit{and} $\theta_{l_n,n} \in [0, \pi]$}}
\put(5,2){\small\bfseries\color{black}{$ \leftarrow m_1 \theta_{l_n,n}$ }}
\caption{Illustration of the angle function of the ASoftmax loss. Obviously, the angle $\theta_{l_n,n}$ in the training stage is in $[0,\pi]$. If we simply multiply an integer margin $m_1$ to $\theta_{l_n,n}$, then the angle function $\psi_1(\cdot)$ is monotonic when $\theta_{l_n,n} \in [0,\frac{\pi}{m_1}]$ only. Therefore, in practice, $\psi_1(\cdot)$ is generalized to $\psi_2(\cdot)$ to ensure that the angle function is monotonically decreasing when $\theta \in [0,\pi]$.}
\label{fig:ASoftmax_margin}
\end{figure}
\item
\textbf{Additive margin softmax (AMSoftmax) loss} \cite{xie2019utterance,hajibabaei2018unified,yu2019ensemble}: It is a revision of ASoftmax loss by replacing $ \psi(\theta_{l_n,n})$ in \eqref{eq:A-Softmax} with $ ({\rm cos}(\theta_{l_n,n})-m_2)$, and normalizing $\|\mathbf{x}_n\|=1$:
\begin{small}
\begin{equation}\label{eq:AMSoftmax}
\begin{split}
&\mathcal{L}_{\rm AMS} = \\
&-\frac{1}{N}\sum_{n=1}^{N} {\rm log} \frac{{\rm exp}{\big(\tau ({\rm cos}(\theta_{l_n,n})-m_2)}\big)}{{\rm exp}{\big(\tau ({\rm cos}( \theta_{l_n,n})-m_2)\big)}+\sum_{j=1,j\neq l_n}^{J}{\rm exp}{\big(\tau ({\rm cos}(\theta_{j,n})})\big)}
\end{split}
\end{equation}
\end{small}
where $\tau$ is a scaling factor for preventing gradients too small during the training process \cite{xiang2019margin}. In addition, \cite{Zhou2020} also proposed a dynamic-additive margin softmax, where $m_2$ is replaced by a dynamic margin for each training sample.
\item
\textbf{Additive angular margin softmax (AAMSoftmax)} \cite{xiang2019margin,liu2019large}: It replaces $({\rm cos}(\theta_{l_n,n})-m_2)$ in \eqref{eq:AMSoftmax} by ${\rm cos}(\theta_{l_n,n}+m_3)$:
\begin{small}
\begin{equation}\label{eq:AAMSoftmax}
\begin{split}
&\mathcal{L}_{\rm AAMS} = \\
&-\frac{1}{N}\sum_{n=1}^{N}{\rm log} \frac{{\rm exp}{\big(\tau ({\rm cos}(\theta_{l_n,n}+m_3))}\big)}{{\rm exp}{\big(\tau ({\rm cos}( \theta_{l_n,n}+m_3))\big)}+\sum_{j=1,j\neq l_n}^{J}{\rm exp}{\big(\tau ({\rm cos}(\theta_{j,n})})\big)}
\end{split}
\end{equation}
\end{small}
\end{itemize}
To further improve the convergence speed and accuracy, \cite{Rybicka2020} recently proposed a parameter adaptation method which adapts the scaling factor $\tau$ and margin $m_3$ at each iteration.
{
Compared to Softmax, both ASoftmax, AMSoftmax, and AAMSoftmax benefit from the following two aspects: first,
the learned features are angularly distributed, which matches with the cosine similarity scoring back-end; second, they introduce an angle, i.e. a cosine margin, to quantitatively control the decision boundary between training speakers for minimizing the within-class variance.
More information can be found in \cite{liu2017sphereface,wang2018additive,wang2018cosface,deng2019arcface}.
\subsection{Regularization for Softmax loss and its variants}
\label{sec:class_regular}
As illustrated in Section \ref{sec:softmax_variants}, the learned feature by the Softmax loss is not discriminative enough. To address this issue, an alterative way is to combine the Softmax loss with some regularizers \cite{liu2017sphereface}:
\begin{equation}
\mathcal{L} = \mathcal{L}_{\rm S} + \lambda \mathcal{L}_{\rm Regular}
\end{equation}
where $\lambda$ is a hyperparameter for balancing the Softmax loss $\mathcal{L}_{\rm S}$ and the regularizer $\mathcal{L}_{\rm Regular}$. Besides, the regularizer is also applicable to other Softmax variants.
Because the embedding layer that produces the embedding speaker features is not always the last hidden layer, e.g. the x-vector in Fig.\ref{fig:x-vector}, the regularizer was sometimes added to the embedding layer. For clarity, we define the output of the embedding layer as $\mathcal{E}=\{(\mathbf{e}_n,l_n)|n=1,2,\cdots,N\}$ , where $\mathbf{e}_n \in \mathbb{R}^{d_5}$. Here we introduce some regularizers as follows:
\begin{itemize}
\item
\textbf{Center loss} \cite{cai2018exploring,li2018deep,wang2019discriminative}:
It is a typical regularizer for Softmax loss. It explicitly minimizes the within-class variance by:
\begin{equation}\label{eq:center_loss}
\mathcal{L}_{\rm C}=\frac{1}{2}\sum_{n=1}^{N}\|\mathbf{e}_n-\mathbf{c}_{l_{n}}\|^2
\end{equation}
where $\mathbf{c}_{l_n} \in \mathbb{R}^{d_5}$ denotes the $l_n$th class center of the elements in $\mathcal{E}$. At each training iteration, the centers are updated as follows \cite{li2018deep}:
\begin{equation}\label{eq:center1}
\mathbf{c}_j^{t+1}=\mathbf{c}_j^t-\epsilon \Delta\mathbf{c}_j^t
\end{equation}
\begin{equation}\label{eq:center2}
\Delta\mathbf{c}_j=\frac{\sum_{n=1}^{N}\delta(l_n=j)\cdot(\mathbf{c}_j-\mathbf{e}_n)}{1+\sum_{n=1}^{N}\delta(l_n=j)}
\end{equation}
where $\epsilon \in [0,1]$ controls the learning rate of the centers, the superscript ``$t$'' represents the number of iterations, and $\delta(\cdot)$ is an indicator function. If the condition of the indicator function, i.e. $l_n=j$, is satisfied, then $\delta=1$; otherwise, $\delta=0$.
The center loss is usually combined with Softmax loss:
\begin{equation}
\mathcal{L} = \mathcal{L}_{\rm S}+\lambda\mathcal{L}_{\rm C}
\end{equation}
See \cite{wen2016discriminative} for more information about the center loss.
\item
\textbf{Ring loss} \cite{liu2019large}: It restricts $\|\mathbf{e}_n\|$ to be close to a target value $R$ for AMSoftmax loss:
\begin{equation}\label{eq:Ring loss}
\mathcal{L}=\mathcal{L}_{\rm AMS}+\lambda \times \frac{1}{N}\sum_{n=1}^{N}(\|\mathbf{e}_n\|-R)^2
\end{equation}
where the target norm $R$ is optimized during the network training. Equation \eqref{eq:Ring loss} essentially applies a normalization constraint to the features.
\item
\textbf{Minimum hyperspherical energy criterion} \cite{liu2019large}:
It enforces the weights of the output layer to distribute evenly on a hypersphere:
\begin{equation}\label{eq:MHE}
\mathcal{L}=\mathcal{L}_{\rm AMS}+\frac{\lambda}{N(J-1)}\sum_{n=1}^{N}\sum_{j=1,j \neq l_n}^{J} h(\|\hat{\mathbf{w}}_{l_n}-\hat{\mathbf{w}}_j\|)
\end{equation}
where $ \hat{\mathbf{w}}_{l_n}$ and $\hat{\mathbf{w}}_j$ are the
$\ell_2$-normalized $\mathbf{w}_{l_n}$ and $\mathbf{w}_j$ respectively, and $h(z)=\frac{1}{z^2}$ is a decreasing function. Intuitively, the minimum hyperspherical energy based regularizer enlarges the inter-class separability.
\item
\textbf{Gaussian prior} \cite{li2019gaussian}:
To reduce information leak, \cite{li2019gaussian} introduced a Gaussian prior to the output of the embedding layer, which results in the following objective function:
\begin{equation}
\label{eq:reglar_Gaussian}
\mathcal{L}=\mathcal{L}_{\rm S}+\lambda \sum_{j}\sum_{\mathbf{e}_n \in\varepsilon(j)}\|\mathbf{e}_n-\mathbf{w}_j\|
\end{equation}
where $\varepsilon(j)$ is the set of the utterances belonging to the $j$th speaker, $\mathbf{e}_n$ represents the x-vector, $\mathbf{w}_j$ represents the parameters of the last layer corresponding to the output unit of speaker $j$.
\item
\textbf{Triplet loss} \cite{jati2019multi}: Because Softmax loss does not explicitly reduce intra-class variance, triplet loss was introduced to directly bring samples from the same class closer than the samples from different classes. Formally, the triplet loss weighted Softmax loss is written as:
\begin{equation}
\label{eq:cross_trip}
\mathcal{L}=\lambda \mathcal{L}_{\rm S}+(1-\lambda)\mathcal{L}_{\rm Triplet}
\end{equation}
where $\mathcal{L}_{\rm Triplet}$ denotes the triplet loss, which will be introduced in Section \ref{sec:end2end_loss}.
\end{itemize}
There are also many other regularization approaches. For example, \cite{yu2019ensemble} added a Hilbert-Schmidt independence criterion based constraint to the embedding layer for regularizing AMSoftmax loss $\mathcal{L}_{\rm AMS}$. See \cite{yu2019ensemble} for the details.
\subsection{Multi-task learning for deep embedding}
\label{sec:multi_task}
Phonetic information is important in improving the performance of speaker recognition. As illustrated in Section \ref{sec:DNN/i-vector}, one way to incorporate phonetic information into the i-vector-based systems is to employ an ASR acoustic model, e.g. the DNN-UBM/i-vector or the DNN-BNF/i-vector. As for the deep-embedding-based speaker recognition, the phonetic information was usually incorporated by multi-task learning.
For example, \cite{chen2015multi,liu2015deep} trained
a deep embedding network to discriminate the speaker identity and text phrases simultaneously. The training objective is to minimize:
\begin{equation}\label{eq:multi_task}
{\rm CE}([\bm{l}_1,\bm{l}_2],[\bm{l}_1^\prime,\bm{l}_2^\prime]) = {\rm CE}_1(\bm{l}_1,\bm{l}_1^\prime) +
{\rm CE}_2(\bm{l}_2,\bm{l}_2^\prime)
\end{equation}
where ${\rm CE}_1$ and ${\rm CE}_2$ are two cross-entropy criteria for speaker and text phrase respectively. $\bm{l}_1$ and $\bm{l}_2$ indicate the true labels for speakers and text individually, and $\bm{l}_1^\prime$ and $\bm{l}_2^\prime$ are the two outputs of the network respectively. Some similar ideas can also be found in \cite{dey2018dnn}.
Although the text content may be a harmful source to text-independent speaker recognition, some positive results were observed with the multi-task learning. {{The authors in \cite{liu2018speaker}}} added phonetic information to the frame-level layers of the x-vector extractor with an auxiliary ASR acoustic model by multi-task learning. {{The authors in \cite{wang2019usage}}} conjectured that the phonetic information is helpful for frame-level feature learning, however, it is useless in utterance-level speaker embeddings. They experimentally verified their assumptions by multitask learning and adversarial training, where the phonetic information was used as positive and negative effects respectively.
Besides the phonetic information mining, some multi-task learning approaches intend to improve the performance of the auxiliary and main tasks together. For example, \cite{tang2016collaborative} proposed a collaborative learning approach based on multi-task recurrent neural model to improve the performance of both speech and speaker recognition. \cite{yao2018snr} proposed a multitask DNN structure to denoise i-vectors and classify speakers simultaneously. Considering that the acoustic and speaker domains are complementary, \cite{Jung2020MultTask} recently proposed a multi-task network that performs keyword spotting and speaker verification simultaneously to fully utilize the interrelated domain information.
\subsection{Discussion to the classification-based loss functions}
Because speaker verification is an open set recognition task, the deep embedding space produced from a training dataset with a limited number of speakers is required to generalize well to unseen test speakers. Therefore, it is the speaker discriminative ability of the embedding rather than the classification accuracy that is important, which accounts for the motivation why many classification-based loss functions are designed to minimize the within-class variance by adding a margin or a regularizer into the Softmax loss.
From the experimental results in literature, one can concluded that the design of loss functions is very important to performance. At present, nearly all stat-of-the-art deep embedding systems replaced the traditional Softmax by its variants, especially AMSoftmax and AAMSoftmax. In addition, the Softmax loss, its variants and regularizers are not mutually exclusive. For instance, the regularization terms \eqref{eq:Ring loss} and \eqref{eq:MHE} were originally added to AMSoftmax \cite{liu2019large}.
\section{End-to-end speaker verification: Verification-based objective functions}
\label{sec:end2end_loss}
An emerging direction of speaker recognition is end-to-end speaker verification. It is able to produce the similarity score of a pair of utterances in a test trial directly. The main difference between deep embedding and end-to-end speaker verification is the objective function. Therefore, in this section, we mainly review the verification-based loss functions, and skip the other components that are similar to deep embedding, e.g. the network structures or temporal pooling layers.
Here we emphasize that the borderline between the classification-based deep embedding and verification-based end-to-end speaker verification is unclear in literature. Some work also called the end-to-end speaker verification systems as deep embedding extractors. The main reason for this confusion is that, although the end-to-end speaker verification systems have different objective functions and training strategies from the deep embedding extractors, they need to extract utterance-level speaker embeddings from the hidden layers as the input of some independent back-ends, e.g. PLDA, in the test stage, so as to achieve the state-of-the-art performance. Despite the confusion usage of the terms in literature, here we clearly regard the speaker verification systems whose loss functions yield similarity scores from training trials as end-to-end speaker verification.
In this section, we focus on summarizing verification-based objective functions, each of which needs to address the following three core issues:
\begin{itemize}
\item How to design a \textbf{training loss} that pushes DNN towards our desired direction: As shown in Fig. \ref{fig:veri-iden-diar}, speaker verification can be viewed as a binary classification problem of whether a pair of utterances are from the same speaker. A natural solution to this problem is to train a binary classifier in an end-to-end fashion from a large number of manually constructed pairs of training utterances, i.e. training trials. The training loss of the binary classifier largely determines the effectiveness of the classifier.
\item How to define a \textbf{similarity metric} between a pair of utterances: The similarity of a pair of utterances is calculated from the embeddings of the utterances at the output layer where a proper similarity metric for evaluating the similarity between the embeddings boosts the performance.
\item How to \textbf{select and construct training trials} from an exponentially large number of training trials: Because the number of all possible training trials is at least the square of the number of training utterances, and also because many of the training trials are less informative, we need to select or even construct some informative training trials instead of using all training trials.
\end{itemize}
\subsection{Pairwise loss}
\label{sec:pairwise_loss}
Pairwise loss is a kind of training loss of the end-to-end speaker verification where each training trial contributes to the accumulation of the training objective value independently.
Suppose there is a set of pairwise training trials as $\mathcal{X}_{\rm pair}=\{(\mathbf{x}_n^e,\mathbf{x}_n^t; l_n)|n=1,2,\cdots,N\}$ where $\mathbf{x}_n^e$ and $\mathbf{x}_n^t$ denote a pair of speaker embedding features at the output layer, and $l_n \in \{0, 1 \}$ is the ground-truth label. If $\mathbf{x}_n^e$ and $\mathbf{x}_n^t$ belong to the same speaker, then $l_n=1$; otherwise, $l_n=0$.
\textit{Binary cross-entropy loss} is the most common pairwise loss \cite{heigold2016end,rahman2018attention,zhang2016end,snyder2016deep,zhang2019seq2seq,rohdin2018end}:
\begin{equation}
\label{eq:binary_cross_entropy}
\mathcal{L}_{\rm BCE}=- \sum_{n=1}^{N} {\bigg [}l_n {\rm ln}\Big(p(\mathbf{x}_{n}^e,\mathbf{x}_{n}^t)\Big)+\eta (1-l_n) {\rm ln}\Big(1-p(\mathbf{x}_{n}^e,\mathbf{x}_{n}^t)\Big){\bigg ]}
\end{equation}
where $\eta$ is a balance factor between positive ($l_n=1$) and negative ($l_n=0$) trials, and $p(\mathbf{x}^e_n,\mathbf{x}^t_n)$ denotes the acceptance probability, i.e. the probability of $\mathbf{x}^e_n$ and $\mathbf{x}^t_n$ belonging to the same speaker. The reason why there needs a balance factor is that the number of the negative trials is usually much larger than that of positive trials. The difference between the variants of the binary cross-entropy loss is on the calculation method of $p(\mathbf{x}^e_n,\mathbf{x}^t_n)$ which is summarized as follows:
\begin{itemize}
\item In \cite{heigold2016end,rahman2018attention}, {{the authors}} applied sigmoid function to cosine similarity:
\begin{equation}
\label{eq:end-to-end-sigmod-cosine}
\begin{split}
& p(\mathbf{x}^e_n,\mathbf{x}^t_n) =\frac{1}{1+{\rm exp}{\big(-wS(\mathbf{x}^e_n, \mathbf{x}^t_n)-b}\big)} \\
& S(\mathbf{x}^e_n,\mathbf{x}^t_n) =\frac{\mathbf{x}^{e^T}_n \mathbf{x}^t_n}{\|\mathbf{x}^e_n\| \, \|\mathbf{x}^t_n\|}
\end{split}
\end{equation}
where $w$ and $b$ are two learnable parameters, and $-b/w$ corresponds to the verification threshold. Some similar idea can also be found in \cite{zhang2016end}.
\item In \cite{snyder2016deep}, {{the authors}} further introduced a PLDA-like similarity metric:
\begin{equation}
\label{eq:end-to-end-sigmod-PLDA}
\begin{split}
&p(\mathbf{x}^e_n,\mathbf{x}^t_n)= \frac{1}{1+{\rm exp}{\big(-S(\mathbf{x}^e_n, \mathbf{x}^t_n))}} \\
& S(\mathbf{x}^e_n, \mathbf{x}^t_n) = {(\mathbf{x}^e_n)}^T\mathbf{x}^t_n-{(\mathbf{x}^e_n)}^T\mathbf{S}\mathbf{x}^e_n-{(\mathbf{x}^t_n)}^T\mathbf{S}\mathbf{x}^{t}_n+b
\end{split}
\end{equation}
where $\mathbf{S}$ and $b$ are learnable parameters. Another similar PLDA-based similarity metric was proposed in \cite{rohdin2018end}.
\item The third calculation method is to learn a score from a joint vector ${\mathbf{x}^{e,t}_n}$ by \cite{zhang2019seq2seq}:
\begin{equation}
\label{eq:end-to-end-sigmod-PLDA}
\begin{split}
&p(\mathbf{x}^e_n,\mathbf{x}^t_n)= \frac{1}{1+{\rm exp}{\big(-s^{e,t}_n)}} \\
& s^{e,t}_n = S(\mathbf{x}^{e,t}_n)
\end{split}
\end{equation}
where $s^{e,t}_n$ is a scalar produced from a fully-connected feedforward neural network $S(\cdot)$, and ${\mathbf{x}^{e,t}_n}$ is a joint vector of ${\mathbf{x}^e_n}$ and ${\mathbf{x}^t_n}$ produced by a sequence-to-sequence attention mechanism \cite{zhang2019seq2seq}. Similar ideas can also be found in \cite{heo2017joint}.
\end{itemize}
The training trials of the aforementioned end-to-end speaker verification are constructed from two utterances.
To reduce the variability of the training trials, some work \cite{heigold2016end,rahman2018attention,zhang2016end} obtains the embedding of the enrollment speech $\mathbf{x}_n^e$ from an average of a small amount of utterances.
\textit{Contrastive loss} \cite{chung2018voxceleb2,yu2019ensemble} is another commonly used pairwise loss:
\begin{equation}
\mathcal{L}_{\rm C} = \frac{1}{2N} \sum_{n=1}^{N}\Big(l_n d_n^2+(1-l_n){\rm max}(\rho-d_n,0)^2 \Big)
\end{equation}
where $d_n$ denotes the Euclidean distance between $\mathbf{x}_n^e$ and $\mathbf{x}_n^t$, and $\rho$ is a manually-defined margin.
Unfortunately, training an end-to-end network with the contrastive loss is notoriously difficult. In order to avoid bad local minima in the early training stage, \cite{chung2018voxceleb2} proposed to first pre-train a speaker embedding system using Softmax loss, and then fine-tune the system with the contrastive loss.
\cite{wan2018generalized} proposed a generalization of the contrastive loss as follows:
\begin{equation}
\label{eq:GE2E-contrast}
L_{\rm GC}=\sum_{n=1}^{N}\Big( l_n \big(1-p(\mathbf{x}^e_n,\mathbf{x}^t_n)\big)+ (1-l_n)\mathop{\mathop{\rm max}_{\mathbf{x}^{e}_{n} \in \{\mathbf{c}_{j^\prime}\}_{j^\prime=1}^{J^\prime}}}p(\mathbf{x}^e_n,\mathbf{x}^t_n) {\Big )}
\end{equation}
where $p(\mathbf{x}^e_n,\mathbf{x}^t_n)$ is the same as \eqref{eq:end-to-end-sigmod-cosine}, and $\mathbf{c}_{j^\prime}$ with $j^\prime=1,2,\ldots,J^\prime$ is the speaker centroid of the $j$th speaker in a mini-batch which is obtained by averaging the utterances that belong to the $j^\prime$th speaker.
Besides the above two common training losses, some other loss functions are as follows.
In \cite{gao2019improving}, the authors proposed a \textit{discriminant analysis loss} $\mathcal{L}_{\rm DALoss}$ to learn discriminative embeddings:
\begin{equation}\label{eq:DALoss}
\mathcal{L}_{\rm DALoss} = \eta_1 \mathcal{L}_{\rm intra} + \eta_2 \mathcal{L}_{\rm inter}
\end{equation}
where $\eta_1$ and $\eta_2$ are the weights of the two loss items, and $\mathcal{L}_{\rm intra}$ and $\mathcal{L}_{\rm inter}$ are described as follows. $\mathcal{L}_{\rm intra}$ represents the intra-speaker variabilities which is defined as:
\begin{equation}
\mathcal{L}_{\rm intra}=\sum_{j^\prime=1}^{J^\prime}\frac{C^{j^\prime}}{\sum_{k=1}^{C^{j^\prime}}\frac{1}{d_k(\mathbf{x}_{n_1}^{j^\prime},\mathbf{x}_{n_2}^{j^\prime})}}
\end{equation}
where $j^\prime=1,2,\cdots,J^\prime$ denotes the index of the training speaker in each mini-batch, and $d_k(\mathbf{x}_{n_1}^{j^\prime},\mathbf{x}_{n_2}^{j^\prime})$ denotes the $k$th largest squared Euclidean distance between the embeddings of the $j^\prime$th speaker. The overall cost is the mean of the first $C^{j^\prime}$th largest distances within each speaker. $\mathcal{L}_{\rm inter}$ represents the inter-speaker variabilities:
\begin{equation}
\mathcal{L}_{\rm inter} = {\rm max}(0,\zeta-{\rm min}(d(\widetilde{\mathbf{x}}^{j^\prime_1},\widetilde{\mathbf{x}}^{j^\prime_2})))
\end{equation}
where $\mathbf{\widetilde{x}}^{j_1^\prime}$ and $\mathbf{\widetilde{x}}^{j_2^\prime}$ denote the centers of the feature vectors of the $j_1^\prime$th and $j_2^\prime$th speakers respectively with $j_1^\prime \neq j_2^\prime \in \{1,2,\cdots,J^\prime \}$, $d(\cdot)$ denotes the distance (e.g. the squared Euclidean distance), and $\zeta$ denotes a margin. Thus, minimizing $ \mathcal{L}_{\rm inter} $ is equivalent to maximizing the distances between the centers to be larger than the minimum margin $\zeta$.
In \cite{mingote2019optimization}, the authors proposed to minimize both the empirical false alarm rate $P_{\rm fa}$ and miss detection rate $P_{\rm miss}$:
\begin{equation}
\mathcal{L}=\eta_1 \cdot {P}_{\rm fa}(\xi) + \eta_2 \cdot {P}_{\rm miss}(\xi)
\end{equation}
\begin{equation}
\label{eq:p_fa}
P_{\rm miss}(\xi) = \frac{ \sum_{n=1}^{N} l_n \ \delta \Big (S(\mathbf{x}^e_n,\mathbf{x}^t_n) < \xi \Big)}{\sum_{n=1}^{N} \delta(l_n=1)}
\end{equation}
\begin{equation}
\label{eq:p_miss}
P_{\rm fa}(\xi) = \frac{\sum_{n=1}^{N} (1-l_n) \ \delta \Big (S(\mathbf{x}^e_n,\mathbf{x}^t_n) > \xi \Big)}{\sum_{n=1}^{N} \delta(l_n=0)}
\end{equation}
where $\delta(\cdot)$ denotes an indicator function, $\xi$ denotes a decision threshold which is optimized with the neural network, and $\eta_1$ and $\eta_2$ are two tunable hyperparameters. The score $S(\mathbf{x}^e_n,\mathbf{x}^t_n)$ is obtained from the output linear layer of the neural network, where the number of units of the output layer equals to the number of the speakers in the training data. Specifically, it uses a batch of input vectors $\{\mathbf{x}_i\}_{i=1}^{I}$ and the parameters $\{\mathbf{w}_j,b_j\}_{j=1}^{J}$ of the output linear layer to construct training trials in a mini-batch:
\begin{equation}
S(\mathbf{x}^e_n,\mathbf{x}^t_n)={(\mathbf{x}^{e}_{n})}^T \mathbf{x}^t_n +b_n^e,
\end{equation}
where $(\mathbf{x}^{e}_{n},b_n^e) \in \{\mathbf{w}_j,b_j\}_{j=1}^{J}$ and $\mathbf{x}^{t}_{n} \in \{\mathbf{x}_i\}_{i=1}^{I}$. To make \eqref{eq:p_fa} and \eqref{eq:p_miss} differentiable, the indicator function $\delta(z>0)$ is relaxed to a sigmoid function $\sigma(z)=1 / (1+{\rm exp}{(-z)})$.
\subsection{Triplet loss}
\label{sec:triplet}
Triplet loss is a kind of training loss that each training sample that contributes to the accumulation of the training objective value independently is constructed from three utterances.
A triplet training sample consists of three utterances, including an anchor utterance, a positive utterance that is produced from the same speaker as the anchor utterance, and a negative utterance from a different speaker. Suppose the speaker features of a training sample produced from the top hidden layer are $\mathbf{x}^a$ (anchor), $\mathbf{x}^p$ (positive), and $\mathbf{x}^n$ (negative), respectively. We denote the training set as $\mathcal{X}_{\rm trip} = \{ (\mathbf{x}_n^a,\mathbf{x}_n^p,\mathbf{x}_n^n)|n=1,2,\cdots,N \}$.
Triplet loss designs a margin-based loss to push the positive utterance $\mathbf{x}_n^p$ closer to the anchor $\mathbf{x}_n^a$ than the negative utterance $\mathbf{x}_n^n$ in a trial as shown in Fig. \ref{fig:Triplet_Loss}.
For any training sample in $\mathcal{X}_{\rm trip}$, we require:
\begin{equation}
\label{eq:triplet}
s_n^{an} - s_n^{ap}+\zeta \le 0
\end{equation}
where, without loss of generality, $s_n^{an}$ denotes the cosine similarity between $\mathbf{x}^a$ and $\mathbf{x}^n$, $s_n^{ap}$ denotes the cosine similarity between $\mathbf{x}^a$ and $\mathbf{x}^p$, and $\zeta \in \mathbb{R}^+$ is a manually-defined \textit{safety} margin between positive and negative pairs. Note that $s_n^{an}$ and $s_n^{ap}$ could be the scores of any similarity measurement instead of merely the cosine similarity. Given $\eqref{eq:triplet}$, the triplet loss is defined as:
\begin{equation}
\label{eq:triplet_loss}
\mathcal{L}_{\rm trip}=\sum_{n=1}^{N} {\rm max}(0, s_n^{an} - s_n^{ap}+\zeta)
\end{equation}
Cosine similarity \cite{li2017deep} and squared Euclidean distance \cite{zhang2017end,huang2018joint,bredin2017tristounet} are the most common similarity metric for the triplet loss. Before calculating the similarities, each speaker embedding in the training samples needs to be length-normalized. It is easy to prove that the two similarity metrics are equivalent \cite{zhang2018text,bai2020speaker} after the length normalization.
Besides the two similarity metrics, {{the authors in}} \cite{dey2018end} proposed several distance functions to explore phonetic information for text-dependent speaker verification. They first compute the Euclidean distance between any pair of frame-level hidden representations of the two input utterances $\mathcal{H}_1=\{\mathbf{h}_{1,t_1}|t_1=1,2,\cdots,T_1\}$ and $\mathcal{H}_2=\{\mathbf{h}_{2,t_2}|t_2=1,2,\cdots,T_2 \}$ via $\mathbf{D}=\{d(\mathbf{h}_{1,t_1},\mathbf{h}_{2,t_2})|t_1=1,2,\cdots,T_1, t_2=1,2,\cdots,T_2\} \in \mathbb{R}^{T_1,T_2}$. Then, they integrate the $T_1 \times T_2$ frame-level Euclidean distances into an utterance-level similarity score of $\mathcal{H}_1$ and $\mathcal{H}_2$ by, e.g. the attention mechanism.
\begin{figure}[t]
\centering
\includegraphics[width=3in]{Triplet_Loss-eps-converted-to.pdf}
\caption{Triplet loss based on cosine similarity (from \cite{li2017deep}).}
\label{fig:Triplet_Loss}
\end{figure}
Given a training set, we can see that the number of all possible triplet training samples is cubically larger than the number of training utterances. It is neither efficient nor effective to enumerate all possible triplets \cite{bredin2017tristounet}, and only those that violate the constraint of $s_n^{an} - s_n^{ap}+\zeta \le 0$ contributes to the training process. Therefore, how to select informative triplet training samples is fundamental to the effectiveness of the model training.
In practice, the \textit{``hard negative''} sampling strategy is popular \cite{bredin2017tristounet}. It consists of the following two steps at each epoch:
\begin{enumerate}[1)]
\item Randomly sample $m$ utterances from each of the $M$ speakers of the training set, which constructs $Mm(m-1)/2$ anchor-positive pairs.
\item For each of the anchor-positive pairs, randomly choose one negative utterance that satisfies $s_n^{an} - s_n^{ap} + \zeta > 0$ from the $(M-1)m$ negative candidates.
\end{enumerate}
Several variants of the \textit{``hard negative''} sampling were proposed as well. For example, \cite{zhang2017end} changed the first step by randomly selecting a small number of speakers from the speaker pool instead of from all speakers. In \cite{huang2018joint}, {{the authors}} divided training speakers into different groups and constructed each triplet training sample from a single group. Besides the \textit{``hard negative''} sampling, the ``semi-hard'' negative sample selection \cite{schroff2015facenet,jati2019multi}
and softmax pre-training \cite{li2017deep} are all used to stabilize the training process of the triplet loss.
\subsection{Quadruplet loss}
Quadruplet loss is a kind of training loss for end-to-end speaker verification where each training sample that contributes to the accumulation of the training objective value independently is constructed from four utterances. Suppose there are
a positive pairwise training set $\mathcal{X}_{\rm same} = \{ (\mathbf{x}_{n_1}^e,\mathbf{x}_{n_1}^t)|n_1=1,2,\cdots,N_1 \}$ and a negative pairwise training set $\mathcal{X}_{\rm diff} = \{ (\mathbf{x}_{n_2}^e,\mathbf{x}_{n_2}^t)|n_2=1,2,\cdots, N_2\}$ respectively, where $\mathbf{x}_{n_1}^e$ and $\mathbf{x}_{n_1}^t$ are from the same speaker while $\mathbf{x}_{n_2}^e$ and $\mathbf{x}_{n_2}^t$ are from different speakers. We have $\mathcal{X}_{\rm same} \cup \mathcal{X}_{\rm diff} = \mathcal{X}_{\rm pair}$. Currently, the quadruplet loss is {{formulated}} as the maximization of the partial interested area under the ROC curve (pAUC) \cite{bai2020partial}.
\begin{figure}[t]
\centering
\includegraphics[width=2in]{pAUC_diagram-eps-converted-to.pdf}
\caption{Illustration of the ROC curve, AUC, and pAUC (from \cite{bai2020partial}).}
\label{fig:pAUC_Loss}
\end{figure}
The maximization of pAUC is to maximize an interested gray area of Fig. \ref{fig:pAUC_Loss} that is defined by two hyperparameters $\alpha$ and $\beta$. It has the following three steps:
\begin{enumerate}[1)~]
\item Rank the similarity scores of all pairwise trials in $\mathcal{X}_{\rm diff}$ in descending order, and selecting those
elements that rank between the $(\lceil N_2 \times \alpha \rceil+1)$-th to $\lfloor N_2 \times \beta \rfloor$-th positions to construct $\mathcal{X}_{\rm diff}^\prime = \{(\mathbf{x}_{n_3}^e,\mathbf{x}_{n_3}^t)|n_3= 1,2,\cdots,N_3\} $ where $N_3 = \lfloor N_2 \times \beta \rfloor - (\lceil N_2 \times \alpha \rceil+1)$.
\item Calculate pAUC on $\mathcal{X}_{\rm same}$ and $\mathcal{X}_{\rm diff}^\prime$:
\begin{equation}
\label{eq:pAUC}
{\rm pAUC} =1-\frac{1}{N_1N_3} \sum_{n_1=1}^{N_1}\sum_{n_3=1}^{N_3}\Big (\delta(s_{n_1}<s_{n_3})+\frac{1}{2}\delta(s_{n_1}=s_{n_3}) \Big)
\end{equation}
where $\delta(\cdot)$ denotes the indicator function, and $s_{n_1}$ and $s_{n_3}$ denote the cosine similarity of the pairwise trials in $\mathcal{X}_{\rm same}$ and $\mathcal{X}_{\rm diff}^\prime$ respectively.
\item Relax the indicator function by the hinge loss which reformulates \eqref{eq:pAUC} to:
\begin{equation}
\label{eq:pAUC_relax}
\mathcal{L}_{\rm pAUC}=\frac{1}{N_1N_3} \sum_{n_1=1}^{N_1}\sum_{n_3=1}^{N_3} {\rm max}\Big (0, \zeta-\big(
s_{n_1}-s_{n_3})\big)\Big)^2
\end{equation}
where $\zeta$ is a margin hyperparameter.
\end{enumerate}
It can be seen clearly that \eqref{eq:pAUC_relax} is a quadruplet loss, since that $(s_{n_1}-s_{n_3})$ is calculated from four utterances.
\cite{bai2020partial} proposed two training sample construction methods. The first one is named random sampling. For a mini-batch, it first randomly chooses a mini-batch number of speakers, then randomly selects two utterances for each of the selected speaker, and finally generates the training trials of the batch by pairing all the selected utterances. The second one is named class-center learning. Before training, it first assigns a class center to each speaker in the training set. Then, for each training iteration, it generates the training trials of a mini-batch by pairing each of the class centers with each of the utterances in the batch, where the class centers are updated together with the DNN parameters.
The pAUC based loss has several advantages: (i) it directly optimizes the detection error tradeoff (DET) curve which is the major evaluation metric of speaker verification \cite{bai2020speaker}; (ii) it naturally overcomes the class-imbalanced problem; (iii) it is able to select difficult quadruplet training samples by setting $\alpha=0$ and $\beta$ to a small value, e.g. 0.01; (iv) triplet training samples is a subset of quadruplet training samples when given the same training utterances \cite{bai2020speaker}.
{{ Actually the pAUC should be calculated on the entire dataset, however, due to the limited computation resource, \eqref{eq:pAUC} is an empirical approximation to it within a mini-batch. Therefore, a large batch size is usually set to reduce the approximation error as much as possible. Fortunately, experimental results demonstrate that a good approximation can be obtained with a batch size of no larger than 512.
}}
\subsection{ Prototypical network loss}
In \cite{wang2019centroid}, the \textit{prototypical network loss} \cite{snell2017prototypical}, which was originally proposed for few-shot learning, was applied to speaker embedding models. Suppose that a mini-batch contains a support set of $N$ labeled samples $\mathcal{S}=\{(\mathbf{x}_n,l_n)|n=1, 2,\cdots, N\}$ where $ l_n \in \{1,2,\cdots,J\}$ is the label of the sample $\mathbf{x}_n$, and $\mathcal{S}_j$ denotes the set of all samples of class $j$. Then, the \textit{prototype} of each class is the mean vector of the support points belonging to the class:
\begin{equation}
\label{eq:PNL-centroid}
\mathbf{c}_j = \frac{1}{|\mathcal{S}_j|} \sum_{(\mathbf{x}_n,l_n)\in \mathcal{S}_j}^{} \mathbf{x}_n, \quad j=1,2,\cdots,J
\end{equation}
Given a query set $\mathcal{Q}=\{(\mathbf{x}_q,l_q)|q=1, 2,\cdots, Q\}$ with $l_q \in \{1,2,\cdots,J\}$, the prototypical network loss classifies each query point $\mathbf{x}_q$ against $J$ prototypes $\{\mathbf{c}_j|j=1,2,\cdots,J\}$ via a softmax function:
\begin{equation}
\mathcal{L}_{\rm PNL} = - \sum_{(\mathbf{x}_q,l_q) \in Q }^{}{\rm log}\frac{{\rm exp}\Big(-d\big(\mathbf{x}_q,\mathbf{c}_{l_q} \big) \Big)}{\sum_{j\prime=1}^{J}{\rm exp}\Big(-d\big(\mathbf{x}_q, \mathbf{c}_{j\prime} \big) \Big)}
\end{equation}
where $d(\cdot)$ denotes the squared Euclidean distance.
For each mini-batch, \cite{wang2019centroid} first randomly selects a number of speakers from the training speaker pool, and then randomly chooses a support set and a query set for each of the selected speakers, where the samples of the support set and query set do not overlap.
Similar works were also conducted in \cite{anand2019few,Chung2020In, Kye2020}.
Before \cite{wang2019centroid}, {{\cite{wan2018generalized}}} proposed a generalized end-to-end loss based on the softmax function, which shares a similar idea with the prototypical network loss except that it uses a single set as both the support and query sets. In addition, \cite{Wei2020} recently proposed an AM-Centroid loss which replaced the weights of the AAMSoftmax loss function with speaker centroids proposed in \cite{wan2018generalized}. This loss function aims to overcome the weakness of the AAMSoftmax loss based deep networks whose number of parameters at the output layer grows linearly with the number of training speakers.
\subsection{Other end-to-end loss functions}
Some loss functions cannot be categorized to the above categories. For example, given learnable speaker bases $\{\mathbf{w}_j\}_j^{J}$ and a mini-batch of utterances $\{(\mathbf{x}_n,l_n)\}_n^{N}$ where $l_n \in \{1,2,\cdots,J\}$, \cite{Heo2019} proposed a between-class variation based loss $\mathcal{L}_{\rm BC}$,
\begin{equation}
\mathcal{L}_{\rm BC}= \sum_{j_2=1}^{J}\sum_{j_1=1, j_1 \neq j_2}^{J} \frac{\mathbf{w}_{j_1}^T\mathbf{w}_{j_2}}{\|\mathbf{w}_{j_1}\| \ \|\mathbf{w}_{j_2}\|}
\end{equation}
and a hard negative mining loss $\mathcal{L}_{\rm H}$,
\begin{equation}
\mathcal{L}_{\rm H} = \sum_{n=1}^{N} \sum_{\mathbf{w}_h \in {\rm Hard}_n}^{} \!\! {\rm log}\Big(1+{\rm exp}\big(S(\mathbf{w}_h,\mathbf{x}_n)-S(\mathbf{w}_{l_n},\mathbf{x}_n)\big)\Big)
\end{equation}
where $\mathbf{x}_n$ denotes the \textit{n}th utterance, $\mathbf{w}_{l_n}$ denotes the basis that $\mathbf{x}_n$ belongs to, $S(\cdot)$ is the cosine similarity, and ${\rm Hard}_n$ is a set of so-called hard negative speaker bases of $\mathbf{x}_n$ which correspond to the top $H$ largest values in $\{S(\mathbf{w}_{j},\mathbf{x}_n) | {j\neq l_n}, j=1,2,\cdots,J\}$.
\subsection{Discussion to the verification-based loss functions}
The verification-based loss functions are fundamentally different from the classification-based loss functions in at least the following aspects.
First, speaker verification is essentially an open-set metric learning problem instead of a {{closed set}} classification problem. The verification-based losses are consistent with the test pipeline, which directly outputs verification scores.
Second, the output layers of the verification-based losses are very small and irrelevant to the number of training speakers, which is an important advantage of the verification-based methods over classification-based methods. Specifically, the number of parameters of a classification-based network at the output layer grows linearly with the increase of the number of training speakers, which make the network large-scale and easily overfit to the training data. For example, if a training set consists of 50000 speakers and if the top hidden layer of a classification network has 512 hidden units, then the output layer of the network contains 25.6 million parameters. On the contrary, the verification-based systems do not suffer the aforementioned weakness. Aware of this issue, \cite{Wei2020} tried to solve the parameter explosion problem by drawing lessons from the prototypical network loss.
The main weakness of the verification-based systems is that they are harder to train than the classification-based systems, since that they need to construct a large number of training trials and then select those that contributes significantly to the effectiveness of the systems, while the classification-based systems just classify each training utterance to its corresponding speaker. To overcome this weakness, many sample selection strategies for selecting highly-informative trials have been developed, such as the hard negative sampling in the triplet loss and the pAUC optimization in the quadruplet loss. Because the highly-informative trials are dynamically changing during the training process, the optimization process is not very stable and consistent. Some unstable examples include the triplet loss or pAUC maximization with the random sampling strategy. Fortunately, this weakness can be alleviated by constructing trials with speaker centroids via, e.g. the class-center learning \cite{bai2020partial}.
In respect of the performance, although the classification-based systems outperformed the verification-based systems once, recently results shown that the latter can achieve competitive performance with the former \cite{bai2020partial,Kye2020}.
A remark: the term ``end-to-end'' in this section intends to make a difference from the embedding systems in Section \ref{sec:embedding_objectives}. However, the term in many other speech processing tasks, which takes raw wave signals as the input and directly output decisions, is broader than the concept here. From the broader concept of ``end-to-end'', an end-to-end speaker verification system needs to further integrate additional procedures, including the voice activity detection, cepstral mean and variance normalization, into the network. It also has to prevent using additional back-ends, such as PLDA \cite{LinWeiwei2020}.
\section{Speaker diarization}
\label{sec:diarization}
In this section, we overview four kinds of speaker diarization technologies---stage-wise diarization, end-to-end diarization, online diarization, and multimodal diarization, where the stage-wise diarization has been studied for a long time, while the last three are emerging directions.
\subsection{Stage-wise speaker diarization}
\label{sec:stagewise}
\begin{figure*}[t]
\centering
\includegraphics[width=5in]{SpeakerDiarization_Diagram.jpg}
\caption{Diagram of the stage-wise speaker diarization, where speech re-segmentation is an optional module.}
\label{fig:SpeakerDiarization_Diagram}
\end{figure*}
Stage-wise speaker diarization is composed of multiple independent modules. As shown in Fig. \ref{fig:SpeakerDiarization_Diagram}, most stage-wise speaker diarization systems consist of four modules---voice activity detection, speech segmentation, speaker feature extraction, and speaker clustering. Some systems also have an optional re-segmentation module. This subsection briefly reviews deep learning based methods for each module. Voice activity detection detects speech in an audio recording and removes non-speech regions. Although it is an important module, it is usually studied independently. Therefore, we focus on reviewing the other modules.
\subsubsection{Speech segmentation}
Speech segmentation splits speech into multiple speaker-homogeneous segments where each segment belongs to a single speaker. It usually can be categorized to two classes---\textit{uniform segmentation} and \textit{speaker change detection} (SCD). Uniform segmentation divides a long audio stream into short segments evenly by a sliding window \cite{Lin2020DIHARD, garcia2017speaker,sell2018diarization,9053952,wang2018speaker}. For example, a 1.5 seconds sliding window with 0.75 second overlap is a common setting of uniform segmentation.
SCD partitions an audio recording according to the detected speaker change points, which results in non-uniform segments.
Generally, it first partitions an audio recording into small segments, then computes the similarity between the two adjacent speech segments in terms of the distance between their representations, and finally decides whether the two adjacent segments are produced from the same speaker by thresholding the distance or finding a local extremum in the consecutive distance stream \cite{bredin2017tristounet}.
A common method for segmenting the audio recording into short segments is to use a sliding window \cite{bredin2017tristounet,wang2017speaker}. Recently, an ASR based segmentation \cite{9053280,sari2019pre} is also employed.
Conventional SCD algorithms usually adopt common handcrafted features, e.g. MFCC, as the acoustic representation \cite{chen1998speaker,siegler1997automatic}. An important advantage of the conventional methods is that only the step of tuning the threshold needs some experience, while the other parts do not need training \cite{yin2017speaker}.
Recently, the deep speaker embedding features are used as the representation of speech segments instead of conventional handcrafted features \cite{wang2017speaker,jati2018unsupervised,bredin2017tristounet,9053280,sari2019pre}.
To calculate the similarity between two adjacent segments, Euclidean distance \cite{bredin2017tristounet} and cosine similarity \cite{wang2017speaker} are two common similarity measurements. Some methods also feed two \cite{sari2019pre, jati2018unsupervised} or more \cite{9053280} consecutive embeddings together into a pre-trained DNN to predict the similarity between two adjacent segments. Recently, some work formulates SCD as a sequence labeling task, which directly predicts if there is a change point in a speech segment \cite{yin2017speaker, zajic2017speaker, hruz2017convolutional}.
To summarize, on one side, the uniform segmentation is simple and works fine in many cases, which is the choice of many real-world diarization systems \cite{Lin2020DIHARD,sell2014speaker,9053280}; on the other side, the research on SCD is important not only for speaker diarization but also for many other applications, such as the closed captioning of broadcast television for hearing-impaired people \cite{9053280}.
\subsubsection{Speaker feature extraction}
The speaker feature extraction module in diarization shares similar technologies with speaker verification. Both of them map speech segments into speaker embeddings by i-vector \cite{sell2014speaker}, DNN-UBM/i-vector \cite{sell2015speaker}, x-vector \cite{sell2018diarization,diez2019bayesian,9054251,9053982}, or some other deep embedding extractors \cite{yella2015comparison, wang2018speaker, sun2019speaker}. See Sections \ref{sec:DNN/i-vector} to \ref{sec:end2end_loss} for the details. Here we only review some embedding methods that utilize additional information for speaker diarization. {{The authors in \cite{9054273}}} incorporated acoustic conditions, such as the distances between speakers and microphones in a meeting or the channel conditions of different speakers in a telephone conversation, into the speaker embeddings, given the fact that the acoustic conditions provide discriminative information for diarization. {{\cite{9054176}}} utilized a graph neural network to refine speaker embeddings, where the local structural information between speech segments is utilized as additional information.
\subsubsection{Speaker clustering}
Given the segment-level embedding features of an audio recording, speaker clustering aims to partition the speech segments into several groups, each of which belongs to a single speaker. It first defines a similarity measurement for evaluating the similarity of two segments, and then conducts clustering according to the similarity scores.
Popular similarity measurements include cosine similarity \cite{wang2018speaker} and PLDA-bases similarity \cite{sell2018diarization,sell2014speaker}. Recently, some deep-learning-based similarity measurements were also introduced, such as the LSTM-based scoring \cite{lin2019lstm}, self-attentive similarity measurement strategies \cite{Lin2020SelfAttentive}, and joint training of speaker embedding and PLDA scoring \cite{garcia2017speaker}.
Common clustering algorithms include k-means \cite{wang2018speaker}, agglomerative hierarchical clustering \cite{sell2018diarization}, spectral clustering \cite{lin2019lstm,wang2018speaker}, Bayesian Hidden Markov Model based clustering \cite{diez2019bayesian,9054251,9053982} etc. Recently, Zhang \cite{zhang2018multilayer} proposed a non-neural-network deep model, named multilayer bootstrap networks, and applied it to speaker clustering \cite{zhang2016universal,li2018investigation}, which demonstrates competitive performance to the common speaker clustering algorithms.
However, these clustering algorithms are unsupervised, which is difficult to utilize manually-labeled training data, e.g. the time-stamped speaker ground truth \cite{zhang2019fully}.
To address the problem, some work formulated speaker clustering as a semi-supervised learning problem \cite{milner2016dnn,yu2017active}.
Specifically, \cite{milner2016dnn} built a new DNN iteratively from a pre-trained speaker separation DNN for the speaker clustering of each audio file. In \cite{yu2017active}, {{the authors}} proposed an active-learning-based speaker clustering algorithm, which needs some involvement of human labor during the clustering process.
Some work formulated speaker clustering as a supervised learning problem \cite{zhang2019fully,li2019discriminative,9053477}. Specifically, \cite{zhang2019fully} and \cite{li2019discriminative} defined the speaker labels of training data according to their first appearance in the training data.
As shown in Fig. \ref{fig:Diarization_labeling}, given a sequence of speech segments $\mathcal{E}=\{(\mathbf{e}_n,l_n)|n=1,2,\cdots,N\}$ where $l_n \in \{\rm A,B,C,D,E,\cdots\}$ is the ground truth label of the $n$th segment with each capital letter representing a speaker, the training label of the sequence $\hat{l}_n \in \{1,2,3,4,5,\cdots\}$ is tagged with positive integers in the order of the speaker appearance in the sequence. For example, the training labels of the two sequences $[\rm E,A,C,A,E,E,C]$ and $[\rm A,C,A,B,B,C,D,B,D]$ are tagged as $[\rm 1,2,3,2,1,1,3]$ and $[\rm 1,2,1,3,3,2,4,3,4]$ respectively.
{{ Under this labeling manner,}} \cite{zhang2019fully} trained a parameter-sharing RNN clustering model in a supervised way by multiple-instance-learning. They further integrated the RNN with a distance-dependent Chinese restaurant process to address the difficult problem of the unknown number of speakers. An improvement to \cite{zhang2019fully} was further presented in \cite{9053477}. Additionally, the clustering procedure was also modeled by a discriminative sequence-to-sequence neural network \cite{li2019discriminative}.
\begin{figure}[t]
\centering
\includegraphics[width=3.3in]{Diarization_labeling-eps-converted-to.pdf}
\caption{Two examples of the training label sequence definition of the supervised clustering in \cite{li2019discriminative} (from \cite{li2019discriminative}).}
\label{fig:Diarization_labeling}
\end{figure}
\subsubsection{Speech re-segmentation}
Re-segmentation is an optional step after the speaker clustering. It refines speech boundaries between the speech segments. Variational-Bayesian refinement \cite{sell2015diarization,sell2018diarization,diez2018speaker} is the most famous conventional method. Recently, deep learning based re-segmentation methods were also developed. For example, following the successful application of the RNN to voice activity detection and SCD, \cite{yin2018neural} proposed to address re-segmentation with LSTM.
\subsubsection{Speech overlap detection}
Most stage-wise speaker diarization systems simply assume that only one person is speaking at any time. In other words, they do not consider the speech overlap problem. However, speech overlap is one of the most important factors that hinder the diarization performance \cite{sell2018diarization,ryant2018first,ryant2019second,Lin2020DIHARD}, since it happens frequently in practice, e.g. a fast multi-speaker conversation. There have been several traditional studies on overlap detection for speaker diarization \cite{otterson2007efficient,boakye2008overlapped,huijbregts2009speech,yella2014overlapping}.
Recently, some deep learning based speech overlap detection methods were also proposed \cite{9053096,9053760}. Specifically, \cite{9053096} first addressed the overlapped speech detection as a sequence labeling problem by an LSTM-based architecture, and then assigned the detected overlap regions to two speakers by a simple yet effective overlap-aware re-segmentation module. {{In \cite{9053760}, }} a region proposal network {{was first used}} to detect overlapped speech, and then removed the highly overlapped segments in the post-processing stage. In addition to the above methods, end-to-end speaker diarization, which will be reviewed in the next section, is another way to deal with the speech overlap problem.
\subsection{End-to-end speaker diarization}
Because conventional clustering algorithms are unsupervised, it cannot minimize the diarization error rate directly \cite{fujita2020end} and is difficult to deal with the speech overlap problem \cite{fujita2020end}. Moreover, because each module of the stage-wise speaker diarization in Fig. \ref{fig:SpeakerDiarization_Diagram} is optimized independently, the performance is difficult to be boosted.
Although several semi-supervised and supervised speaker clustering methods have recently been proposed, the potential of deep neural networks, which are mainly used to extract speaker embeddings in the stage-wise speaker diarization, has not been fully explored yet.
To address these problems, \cite{fujita2019end2,fujita2019end,fujita2020end} proposed an end-to-end diarization method by formulating speaker diarization as a multi-label classification problem.
\begin{figure}[t]
\centering
\includegraphics[width=3in]{E2E_diarization_PIT-eps-converted-to.pdf}
\caption{End-to-end neural diarization with the permutation-free loss for two-speaker diarization (from \cite{fujita2020end}).}
\label{Fig:E2E_diarization_PIT}
\end{figure}
As shown in Fig. \ref{Fig:E2E_diarization_PIT}, given an acoustic feature sequence of an audio recording $\mathcal{Y}=\{\mathbf{y}_t \in \mathbb{R}^{d_1}|t=1,2,\cdots,T\}$, speaker diarization estimates the speaker label sequence $L=\{\bm{l}_t|t=1,2,\cdots,T\}$, where $\bm{l}_t=\{l_{t,j}\in \{0,1\}|j=1,2,\cdots,J\}$ denotes a joint activity of a total number of $J$ speakers at time $t$. This multi-label formulation can represent speaker overlap regions properly. For example, $l_{t,j}=l_{t,j^\prime}=1$ with $j\neq j^\prime$ represents that the speech of the speakers $j$ and $j^\prime$ overlaps at time $t$.
With the assumption that each speaker is present independently, the frame-wise posteriors are estimated by a neural network as follows \cite{fujita2020end}:
\begin{equation}\label{eq:EEND_MultiLabel}
\!\! \mathbf{z}_t=[P(l_{t,1}|\mathcal{Y}),P(l_{t,2}|\mathcal{Y}),\cdots,P(l_{t,J}|\mathcal{Y})]={\rm NN}_t(\mathcal{Y})\in(0,1)^J
\end{equation}
where ${\rm NN}_t(\cdot)$ denotes a neural network, which was implemented by a BLSTM \cite{fujita2019end}, or a self-attention based neural network \cite{fujita2019end2}.
However, a difficult problem for the end-to-end speaker diarization is the speaker-label permutation ambiguity problem when aligning the ground truth label with the speech recording in preparing the training data.
For example, given an audio recording with three speakers A, B and C. If the ground-truth label of the recording is ``AAABBC'' , then the encoded labels ``111223'' and ``222113'' are equally correct, making the neural network hard to define a unique training label sequence \cite{lin2020optimal}.
To cope with this problem, as shown in Fig. \ref{Fig:E2E_diarization_PIT}, a neural network is trained to minimize the permutation-invariant training (PIT) loss between the output $\mathbf{z}_t$ and the reference speaker label $\mathbf{1}_t \in \{0,1 \}^J$ \cite{fujita2019end2,fujita2019end,fujita2020end}:
\begin{equation}
\mathcal{L}^{PF}=\frac{1}{TJ} \mathop{{\rm min}}_{\phi \in {\rm perm}(J)}\sum_{t}^{}{\rm BCE}(\mathbf{1}_t^\phi,\mathbf{z}_t)
\end{equation}
where ${\rm perm}(J)$ is the set of all possible permutations of the speaker identifiers $\{1,2\cdots, J\}$, and $\mathbf{1}_t^\phi$ is the $\phi$-th permutation of the ground-truth label sequence, and ${\rm BCE}(\cdot, \cdot)$ is the binary cross entropy between the label and the network output.
Under the multi-label classification framework \eqref{eq:EEND_MultiLabel}, the end-to-end diarization system is unable to deal with the test scenario where the number of speakers is larger than the maximum number of speakers in any of the training conversations. Here we bravely call this problem the \textit{fixed speaker capacity} issue for short.
To this end, the end-to-end diarization framework is less flexible than the stage-wise methods which can handle any number of speakers in a test conversation \cite{horiguchi2020end}. To enlarge the speaker capacity, the number of the output nodes of the neural network \eqref{eq:EEND_MultiLabel} tends to be set large.
Unfortunately, the computational resource for training with the PIT loss will be exponentially increased along with the increase of the number of the output nodes. To deal with this contradiction, \cite{lin2020optimal} proposed an optimal mapping loss, which directly computes the best match between the output speaker sequence and the ground-truth speaker sequence through a so-called Hungarian algorithm. It reduces the computational complexity to polynomial time, and meanwhile yields similar performance as the PIT loss. It should be noted that, the fixed {speaker capacity} issue remains unsolved in the optimal mapping loss.
\begin{figure}[t]
\centering
\includegraphics[width=3in]{SC-EEND-eps-converted-to.pdf}
\caption{Diagrams of the conventional end-to-end neural diarization (EEND) method with fixed speaker numbers in \cite{fujita2020end}, and the speaker-wise conditional end-to-end neural diarization (SC-EEND) method (from \cite{fujita2020neural}).}
\label{Fig:SC-EEND}
\end{figure}
To address the fixed speaker capacity issue, \cite{fujita2020neural} proposed a speaker-wise conditional end-to-end (SC-EEND) speaker diarization method.
As shown in Fig. \ref{Fig:SC-EEND}, it uses an encoder-decoder architecture to decode each speaker's speech activity iteratively conditioned on the estimated historical speech activities. Although the SC-EEND method achieves {{increased}} speaker counting accuracy, it is difficult to handle more than four speakers \cite{fujita2020neural}, since it still has to deal with the PIT problem during decoding. Besides, they also proposed an encoder-decoder based attractor calculation method, where a flexible number of speaker attractors are calculated from a speech embedding sequence \cite{horiguchi2020end}. However, the speaker capacity of the method remains limited, due to the PIT loss which determines the assignment of the attractors to the training speakers.
\subsection{Online speaker diarization}
\label{sec:online_diarization}
Most state-of-the-art speaker diarization systems work in an offline manner. Online speaker diarization, which outputs the diarization result right after the audio segment arrives, is not an easy task, since that future information is unavailable when analyzing the current segment \cite{xue2020online}. In history, a number of online speaker diarization and speaker tracking solutions have been reported \cite{patino2018low,dimitriadis2017developing}. Here we focus on the deep learning based ones, which can be categorized to stage-wise online diarization and end-to-end online diarization methods.
In respect of the stage-wise online diarization, \cite{zhang2019fully} replaced the commonly used clustering module with a trainable unbounded interleaved-state RNN to make prediction in an online fashion{{, and it was further improved by introducing a \textit {sample mean loss} \cite{9053477}}}.
Additionally, a transformer-based discriminative neural clustering model proposed in \cite{li2019discriminative}
can also perform online diarization. However, although it is possible to transfer the online clustering to end-to-end speaker diarization, these methods still suffer from the assumption that only one speaker is present in a single segment \cite{xue2020online,lin2020optimal}.
In respect of the end-to-end online diarization, \cite{xue2020online} extended the self-attention-based end-to-end speaker diarization method in Fig. \ref{Fig:E2E_diarization_PIT} to an online version, with a speaker-tracing mechanism that is important for the success of the online diarization. Specifically, a straightforward online extension to the framework in Fig. \ref{Fig:E2E_diarization_PIT} is to perform diarization independently for each chunked recording. However, it is observed that the extension degrades the diarization error rate, due to the speaker permutation inconsistency across the chunk, especially for the short-duration chunks. To overcome this weakness, they applied a speaker-tracing buffer to record the speaker permutation information determined in previous chunks for a correct speaker-tracing in the following chunk. Because this method is limited to two-speaker diarization, more flexible speaker online end-to-end diarization methods need to be further explored \cite{xue2020online}. In addition, \cite{von2019all} presented an all-neural approach to simultaneously conduct speaker counting, diarization, and source separation in a block-online manner, where a speaker-tracing mechanism is also employed to avoid the permutation inconsistency problem across time chunks.
\subsection{Multimodal speaker diarization}
Although speaker diarization is conventionally an audio-only task, the linguistic content carried by speech signals \cite{Flemotomos2020Linguistically} and the speaker behaviors, e.g., the movement of lips, recorded by videos \cite{9054376} offer valuable supplementary cues to the detection of active speakers. To incorporate the aforementioned knowledge, \textit{multimodal speaker diarization} is emerging. Here we summarize some work as follows.
The first class is \textit{audio-linguistic} speaker diarization, \cite{park2018multimodal} integrated lexical cues and acoustic cues together by a gated recurrent unit-based sequence-to-sequence model, which improves the diarization performance by exploring linguistic variability deeply. The effectiveness of using both the linguistic and acoustic cues for diarization has been manifested further in structured scenarios \cite{el2019joint,Flemotomos2020Linguistically} where the speakers are assumed to produce distinguishable linguistic patterns. For instance, a teacher is likely to speak in a more didactic style while a student tends to be more inquisitive; a doctor is likely to inquire on symptoms and prescribe while a patient describe symptoms, etc. Another emerging direction is \textit{audio-visual} speaker diarization. In \cite{9054376}, the authors proposed a self-supervised audio-video synchronization learning method for the scenario where there lacks massive labeled data. The authors in \cite{chung2019said} proposed an iterative audio-visual approach which first enrolls speaker models using audio-visual correspondence, and then combines the enrolled models together with the visual information to determine the active speaker. In addition, microphone arrays provide important spatial information as well. For example, \cite{9053122} recently combined d-vectors with the spatial information produced from beamforming for the multimodal speaker diarization.
\section{Robust speaker recognition}
\label{sec:robust}
\renewcommand\arraystretch{1.5}
\begin{table*}[t]
\centering
\footnotesize
\caption{Summary of adversarial-training-based domain adaptation methods.}
\label{tab:adversarial_adaptation}
\scalebox{1}{
\begin{tabular}{m{0.2 cm} m{ 4.5cm} m{6.5cm} m{3.5cm} }
\hline
\hline
&\textbf{Variable factors} & \textbf{Instantiation} & \textbf{References}\\
\hline
1. & Source and target mappings $M_s$ and $M_t$& Shared parameters $M_s=M_t=M$, or unshared parameters $M_s \neq M_t$. & \cite{wang2018unsupervised,9053323,9053905,fang2019channel}; \cite{xia2019cross} \\
2. & Domain discriminator loss $\mathcal{L}_{{\rm adv}D}$ & Binary cross-entropy, multi-class cross-entropy, or Wasserstein distance.& \cite{wang2018unsupervised}; \cite{zhou2019training}; \cite{rohdin2019speaker}\\
3. & Adversarial loss $\mathcal{L}_{{\rm adv}M}$ & The gradient reversal layer, or the GAN loss function. & \cite{wang2018unsupervised,9053323, bhattacharya2019adapting}; \cite{bhattacharya2019generative,xia2019cross} \\
4. & Input feature $\mathbf{X}_s$ and $\mathbf{X}_t$ & Utterance-level speaker features (e.g. i-vector and x-vector), or frame-level acoustic features (e.g. MFCC and F-bank). & \cite{wang2018unsupervised,tu2019variational} ; \cite{9053323,bhattacharya2019adapting} \\
5. & Adaptation target & Channel invariant, language invariant, phoneme invariant, noise robust, or short utterance compensation. & \cite{wang2018unsupervised};\cite{bhattacharya2019adapting};\cite{9053871};\cite{zhou2019training};\cite{zhang2018vector} \\
\hline
\hline
\end{tabular}}
\end{table*}
Along with the rapid progress of speaker recognition, the frontier turns to ``\textit{recognition in the wild}'' \cite{nagrani2017voxceleb,mclaren2016speakers,ryant2018first}, where lots of domain mismatch and noisy problems arise. To overcome these difficulties, many domain adaptation and noise reduction methods were proposed. In this section, we comprehensively review these robust speaker recognition methods, including domain adaptation in Section \ref{subsec:da}, speech enhancement preprocessing in Section \ref{subsec:se}, and data augmentation techniques in Section \ref{sec:data_augmentation}.
\subsection{Domain adaptation}\label{subsec:da}
Over the past few years, speaker recognition has achieved great success due to the application of deep learning and large amount of labeled speech data. However, because collecting and annotating data for every new application is extremely expensive and time-consuming, sufficient training data may not be always available. For example, although large-scale labeled English databases are publicly available, Cantonese databases may be scarce \cite{sadjadi20172016}. Hence, it is needed to improve the performance of low-resource speaker recognition by using the large amount of auxiliary data. However, there are many distribution mismatch or domain shift problems between the low-resource data and auxiliary data, including different languages, phonemes, recording equipments, etc., which hinder the effectiveness of the auxiliary data. Fortunately, the mismatch problem can be alleviated by domain adaptation techniques.
Without loss of generality, the domain of interest is called the \textit{target} domain, and the domain with sufficient labeled training data is called the \textit{source} domain. The data distributions of the target domain and source domain are denoted as $p_t(\mathbf{x}, y)$ and $p_s(\mathbf{x}, y)$ respectively, where $p_s(\mathbf{x}, y) \neq p_t(\mathbf{x}, y)$.
Domain adaptation uses large amount of labeled data in the source domain to solve the problems in the target domain. If the training data in the target domain is manually labeled, then the domain adaptation is supervised; otherwise, it is unsupervised.
This paper focuses on the unsupervised domain adaptation, since it is more common and technically more challenging than the supervised domain adaptation in speaker recognition.
Domain adaptation has long been a significant topic in speaker recognition. It received much attention after the domain adaptation challenge in 2013\footnote{https://www.clsp.jhu.edu/workshops/13-workshop/speaker-and-language-recognition/}. Over the past years, lots of supervised \cite{garcia2014supervised,wang2016domain,wang2020generalized} and unsupervised \cite{kanagasundaram2015improving,misra2018maximum,alam2018speaker,glembek2014domain,garcia2014unsupervised,lee2019coral,villalba2014unsupervised,wang2020generalized,shum2014unsupervised,bousquet2019robustness} shallow domain adaptation methods based on the well known i-vector/PLDA pipline have been developed.
Among them, the adaptation is usually accomplished at the back-end, including the methods of compensating the domain mismatch in the i-vector space by an independent module before LDA and PLDA \cite{aronowitz2014inter, kanagasundaram2015improving,misra2018maximum,alam2018speaker}, and conducting domain adaptation at LDA \cite{glembek2014domain,mclaren2011source} or PLDA \cite{garcia2014supervised,garcia2014unsupervised,wang2016domain,lee2019coral,villalba2014unsupervised,wang2020generalized,shum2014unsupervised}. Although the methods are quite effective, here we do not discuss their details, as this article focuses on deep learning based ones.
Recently, many deep learning based domain adaptation methods were proposed. Following the categorization of the domain adaptation techniques \cite{wang2018deep,tzeng2017adversarial}, this paper categorize the deep-learning-based domain adaptation in speaker recognition into the following three classes:
\begin{itemize}
\item \textbf{Adversarial-training-based domain adaptation:} It seeks to minimize an approximate domain discrepancy distance through an adversarial objective with a domain discriminator.
\item \textbf{Reconstruction-based domain adaptation:} It assumes that the data reconstruction of the source or target samples can be helpful for improving the performance of domain adaptation.
\item \textbf{Discrepancy-based domain adaptation:} It aligns the statistical distribution shift between the source and target domains using some mechanisms.
\end{itemize}
\subsubsection{Adversarial-training-based domain adaptation} \label{sec:adversarial-training-based}
\begin{figure}[t]
\centering
\includegraphics[width=3.3in]{Adversarial-training.pdf}
\put(-166,97){\small\bfseries\color{black}{$M_s$}}
\put(-166,27){\small\bfseries\color{black}{$M_t$}}
\put(-77,27){\small\bfseries\color{black}{$D$}}
\put(-77,97){\small\bfseries\color{black}{$C$}}
\put(-220,102){\small\bfseries\color{black}{$\mathbf{X}_s, Y_s$}}
\put(-215,35){\small\bfseries\color{black}{$\mathbf{X}_t$}}
\put(-242,62){\small\bfseries\color{black}{$p_s(\mathbf{x}, y) \!\! \neq \!\! p_t(\mathbf{x}, y)$}}
\put(-112,62){\small\bfseries\color{black}{$ p(M_s(\mathbf{X}_s)) \!\! \approx \!\! p(M_t(\mathbf{X}_t))$}}
\put(-35,102){\small\bfseries\color{purple}{$\mathcal{L}_{\rm cls}$}}
\put(-133,102){\small\bfseries\color{black}{$M_s(\mathbf{X}_s)$}}
\put(-133,35){\small\bfseries\color{black}{$M_t(\mathbf{X}_t)$}}
\put(-35,35){\small\bfseries\color{purple}{$\mathcal{L}_{{\rm adv}_D}$}}
\put(-35,22){\small\bfseries\color{purple}{$\mathcal{L}_{{\rm adv}_M}$}}
\put(-125,-10){\small\bfseries\color{purple}{Gradient descent}}
\put(-125,140){\small\bfseries\color{purple}{Gradient descent}}
\put(-205,2){\small\bfseries\color{red}{ $M_s \neq M_t$ or $M_s = M_t$}}
\caption{A unified framework of adversarial-training-based domain adaptation.}
\label{fig:Adversarial-training}
\end{figure}
\begin{figure*}[t]
\centering
\includegraphics[width=5in]{Y-GAN-robust.jpg}
\put(-158,40){\small\bfseries\color{black}{$D$}}
\put(-158,87){\small\bfseries\color{black}{$C$}}
\put(-250,64){\small\bfseries\color{black}{$M$}}
\caption{Domain-adversarial neural network (DANN) architecture with $M_s = M_t=M$.}
\label{fig:Y-GAN-robust}
\end{figure*}
Before presenting the literature, we first build a unified framework of adversarial-training-based domain adaptation for speaker recognition {{by drawing lessons from \cite{tzeng2017adversarial}}}. As shown in Fig. \ref{fig:Adversarial-training}, the framework consists of:
\begin{itemize}
\item $\mathbf{X}_s$ and $Y_s$ drawn from $p_s(\mathbf{x}, y)$ are the source features and speaker labels, and $\mathbf{X}_t$ drawn from $p_t(\mathbf{x}, y)$ are the target features without labels;
\item $M_s$ and $M_t$ are the feature mappings for the source and target domains respectively;
\item $C$ is a classifier for discriminating speakers in the source domain.
\item $D$ is a domain discriminator for discriminating the source domain and target domain.
\end{itemize}
The adversarial adaptation methods aim at learning $M_s$ and $M_t$ for minimizing the distance between the empirical source and target data distributions in the feature space, i.e. making $ p(M_s(\mathbf{X}_s))\approx p(M_t(\mathbf{X}_t))$. After the adaptation, the recognition models trained on $M_s(\mathbf{X}_s)$ can be directly applied to the target domain.
In the training stage, $M_s$ and $C$ are jointly trained using the standard supervised loss:
\begin{equation}
\label{eq:adv_class}
\begin{split}
(\hat{M_s} , \hat{C})&= {\rm arg} \mathop{{\rm min}}_{M_s,C} \mathcal{L}_{\rm cls}(\mathbf{X}_s,Y_s;M_s,C) \\
&= -\mathbb{E}_{(\mathbf{x}_s,y_s) \sim (\mathbf{X}_s,Y_s)} \sum_{j=1}^{J}\mathbb{I}_{[j=y_s]}{\rm log}[C(M_s(\mathbf{x}_s))]
\end{split}
\end{equation}
where $J$ is the number of speakers, and $\mathcal{L}_{\rm cls}$ is either the standard softmax or its variants described in Section \ref{sec:embedding_objectives}, which ensures that the outputs of $M_s$ are speaker-discriminative.
To minimize the difference between the source and target representations, the adversarial adaptation methods conduct the following two steps alternatively for $D$ and $M_t$:
\begin{equation}
\label{eq:adv_D}
\hat{D} ={\rm arg} \mathop{{\rm min}}_{D} \mathcal{L}_{{\rm adv}_D}(\mathbf{X}_s,\mathbf{X}_t; \hat{M_s},\hat{M_t},D)
\end{equation}
\begin{equation}
\label{eq:adv_Map}
\hat{M_t} ={\rm arg} \mathop{{\rm min}}_{M_t} \mathcal{L}_{{\rm adv}_M}(\mathbf{X}_s,\mathbf{X}_t; \hat{M_s},M_t,\hat{D})
\end{equation}
where the symbols with hat, such as $\hat{M_s}$, denote that they are fixed during the alternative optimization.
By alternatively minimizing \eqref{eq:adv_D} and \eqref{eq:adv_Map}, $D$ and $M_t$ play an adversarial game: $D$ is optimized to predict the domain labels of $M_s(\mathbf{X}_s)$ and $M_t(\mathbf{X}_t)$, while $M_t$ is trained to make the prediction as incorrect as possible. When the training process converges, we have $ p(M_s(\mathbf{X}_s))\approx p(M_t(\mathbf{X}_t))$ since that $D$ is unable to discriminate $M_s(\mathbf{X}_s)$ and $M_t(\mathbf{X}_t)$. In addition, because the output of $M_s$ is speaker-discriminative which is ensured by the speaker classifier $C$, the output of $M_t$ is supposed to be speaker-discriminative too. For clarity, the manual label requirement in \eqref{eq:adv_class}, \eqref{eq:adv_D}, and \eqref{eq:adv_Map} are summarized in Table \ref{tab:adv_labels}.
\renewcommand\arraystretch{1.2}
\begin{table}[t]
\footnotesize
\centering
\caption{The manual label requirement in the loss functions \eqref{eq:adv_class}, \eqref{eq:adv_D}, and \eqref{eq:adv_Map}. The domain labels can be any factors that are needed to be mitigated, including the types of channels, languages, phonemes, noise, etc.}
\label{tab:adv_labels}
\begin{tabular}{m{1cm}<{\centering}| m{1.8cm}<{\centering} m{1.8cm}<{\centering} m{1.8cm}<{\centering}}
\hline
\hline
Losses & \textbf{Target} domain speaker labels & \textbf{Source} domain speaker labels & \textbf{Domain} labels\\
\hline
$\mathcal{L}_{\rm cls}$ &\checkmark & \ding{55} &\ding{55} \\
$\mathcal{L}_{{\rm adv}_D}$ &\ding{55} & \ding{55} &\checkmark \\
$\mathcal{L}_{{\rm adv}_M}$ &\ding{55} & \ding{55} &\checkmark \\
\hline
\hline
\end{tabular}
\end{table}
Different implementations of the framework in Fig. \ref{fig:Adversarial-training} are summarized in Table \ref{tab:adversarial_adaptation} which will be reviewed in detail as follows.
One of the most popular domain-adversarial neural network (DANN) architectures in literature is shown in Fig. \ref{fig:Y-GAN-robust} \cite{ganin2015unsupervised}. It consists of an encoder $M$, a speaker classifier $C$, and a domain discriminator $D$.
It follows the framework in Fig. \ref{fig:Adversarial-training} with a constraint $M_s=M_t=M$. Its training alternates the following steps:
\begin{enumerate}[1)~]
\item The encoder is optimized by merging \eqref{eq:adv_class} and \eqref{eq:adv_Map}:
$\hat{M} = {\rm arg} \mathop{{\rm min}}_{M} {\large[}\mathcal{L}_{\rm cls}(\mathbf{X}_s,Y_s;M,\hat{C}) + \lambda \mathcal{L}_{{\rm adv}_M}(\mathbf{X}_s,\mathbf{X}_t; M,\hat{D}){\large]}$, where $\lambda$ is a balance factor.
\item The speaker classifier is obtained by minimizing \eqref{eq:adv_class}: $\hat{C} = {\rm arg} \mathop{{\rm min}}_{C} \mathcal{L}_{\rm cls}(\mathbf{X}_s,Y_s;\hat{M},C) $.
\item The domain discriminator is obtained by minimizing \eqref{eq:adv_D}: $\hat{D} ={\rm arg} \mathop{{\rm min}}_{D} \mathcal{L}_{{\rm adv}_D}(\mathbf{X}_s,\mathbf{X}_t; \hat{M},D) $.
\end{enumerate}
After training, the encoder is responsible for extracting domain-invariant and speaker-discriminative speaker features, while the speaker classifier and domain discriminator will be discarded.
The domain-adversarial architecture has been used to explore channel-invariant \cite{wang2018unsupervised,9053323,9053905,fang2019channel}, language-invariant \cite{bhattacharya2019adapting,bhattacharya2019generative,rohdin2019speaker,tu2019variational,9053735}, phoneme-invariant \cite{9053871,wang2019usage}, noise-robust \cite{zhou2019training,meng2019adversarial,9054601} speaker features, etc. Here we present them briefly as follows.
\begin{figure*}[t]
\centering
\includegraphics[width=5in]{Channel-invariant-eps-converted-to.pdf}
\caption{ Channel-invariant domain adaptation (from \cite{9053323}). The classifier is trained in the same way as the ordinary x-vector. The discriminator is trained on concatenated pairs of within-speaker pairs. The blue arrows represent the forward propagation. The red arrows represent the backward propagation of gradients.}
\label{fig:Channel-invariant}
\end{figure*}
\cite{wang2018unsupervised} applied DANN to learn a channel invariant feature extractor from the i-vector subspace.
It takes the binary cross-entropy loss as the loss function of $\mathcal{L}_{{\rm adv}_D}$ to train the discriminator $D$. It adds a gradient reversal layer between the encoder and the discriminator to realize a minimax game $\mathcal{L}_{{\rm adv}_M}= - \mathcal{L}_{{\rm adv}_D}$.
\cite{9053323} developed a similar framework with Fig. \ref{fig:Y-GAN-robust} on an x-vector extractor. The framework produces speaker features that are invariant to the granularity of the recording channels. Instead of predicting the concrete domain labels, its domain discriminator predicts whether a pair of speaker embeddings that comes from the same speaker belong to the same recording in a Siamese fashion, see Fig. \ref{fig:Channel-invariant} for the above process. There is also a gradient reversal layer between then x-vector extractor and the domain discriminator. \cite{9053905} also proposed a similar work to suppress the channel variability.
In addition to the channel mismatch problem, language mismatch is another challenge in speaker recognition. In \cite{bhattacharya2019adapting}, {{the authors}} applied the domain-adversarial architecture directly to acoustic features for a language invariant feature extractor, where the binary cross-entropy loss and a gradient reversal layer are applied to \eqref{eq:adv_D} and \eqref{eq:adv_Map} respectively. In \cite{bhattacharya2019generative}, they further
replaced the gradient reversal layer with a generative adversarial network (GAN) loss, a.k.a. inverted-label loss. Specifically,
$\mathcal{L}_{{\rm adv}_D}$ still adopts the binary cross-entropy loss, while $\mathcal{L}_{{\rm adv}_M}$ fools the domain discriminator by inverting the domain labels instead of using a gradient reversal layer, i.e.:
\begin{equation}
\begin{split}
\mathcal{L}&_{{\rm adv}_D}(\mathbf{X}_s,\mathbf{X}_t;\hat{M},D) = \\
& -\mathbb{E}_{\mathbf{x}_s \sim \mathbf{X}_s}[{\rm log}(D(\hat{M}(\mathbf{x}_s)))]-\mathbb{E}_{\mathbf{x}_t \sim \mathbf{X}_t}[{\rm log}(1-D(\hat{M}(\mathbf{x}_t)))]
\end{split}
\end{equation}
\begin{equation}
\label{eq:GAN_loss}
\begin{split}
\mathcal{L}&_{{\rm adv}_M}(\mathbf{X}_s,\mathbf{X}_t;M,\hat{D}) = \\
& -\mathbb{E}_{\mathbf{x}_s \sim \mathbf{X}_s}[{\rm log}(1-\hat{D}(M(\mathbf{x}_s)))]-\mathbb{E}_{\mathbf{x}_t \sim \mathbf{X}_t}[{\rm log}(\hat{D}(M(\mathbf{x}_t)))]
\end{split}
\end{equation}
where $\mathcal{L}_{{\rm adv}_D}$ tags the data from the source domain as ``1'' and the data from the target domain as ``0'', and $\mathcal{L}_{{\rm adv}_M}$ takes the opposite domain labels. This objective has the same fixed-point properties as the gradient reversal layer but provides stronger gradients to the encoder than the latter \cite{bhattacharya2019generative, tzeng2017adversarial}.
\begin{figure}[t]
\centering
\includegraphics[width=3in]{DAT_E2E_Wass_distance_adaptation-eps-converted-to.pdf}
\caption{End-to-end adversarial language adaptation (from \cite{rohdin2019speaker}).}
\label{fig:DAT_E2E_Wass_distance_adaptation}
\end{figure}
{{The authors in \cite{rohdin2019speaker}}} trained a language-invariant embedding extractor in an end-to-end fashion, where the embedding extractor is a standard TDNN based x-vector extractor (Fig. \ref{fig:DAT_E2E_Wass_distance_adaptation}). They utilized the discriminator to estimate the empirical Wasserstein distance between the source and target samples, and optimized the feature extractor network to minimize the distance in an adversarial manner.
\begin{figure}[t]
\centering
\includegraphics[width=2.5in]{VAE_regularizer_adaptation-eps-converted-to.pdf}
\caption{Variational domain adversarial neural network (from \cite{tu2019variational}).}
\label{fig:VAE_regularizer_adaptation}
\end{figure}
\cite{tu2019variational} added an additional variational autoencoder (VAE) branch to the standard DANN structure of Fig. \ref{fig:Y-GAN-robust} for a language-invariant feature extractor as shown in Fig. \ref{fig:VAE_regularizer_adaptation}, where the state-of-the-art x-vector was used as the input. The VAE branch performs like a variational regularization which constrains the learned features to be Gaussian. As we know, Gaussian distribution is essential for the effectiveness of the standard PLDA backend. In \cite{9053735}, they further replaced the VAE with an information-maximized VAE, which not only retains the variational regularization but also inclines to preserve more speaker discriminative information than the VAE.
For text-independent speaker recognition, phonetic information is sometimes harmful, given the fact that it is difficult to ensure shared phonetic coverage across short enrollment and test utterances. However, for long recordings, phonetic information and even more higher level cues, such as word usages, actually differentiates speakers.
Quest for such high-level speaker features was an active area of research in about two decades ago. Recently, some studies also investigated its effect on text-independent speaker recognition in the era of deep learning. For example, \cite{wang2019usage} applied multi-task learning on frame-level layers to enhance the phonetic information in the frame-level features, and used the adversarial training on segment-level layers to learn phoneme-independent representations. Finally, both operations result in improved performance. \cite{9053871} concluded that phonetic information should be suppressed in text-independent speaker recognition working with frame-wise or extremely short utterances.
Additive noise is one of the most serious interferences of speaker recognition. To explore a noise-robust feature extractor, many works resorted to DANN.
For instance, \cite{zhou2019training} used a multi-class cross entropy loss as the loss function of a noise discriminator to train a noise-condition-invariant feature extractor. \cite{meng2019adversarial} applied the adversarial training to learn a feature extractor that are invariant to two kinds of conditions---different environments and different SNRs, where the different environments are represented as a categorical variable and the range of SNRs is formulated as a continuous variable.
\cite{9054601} applied an unsupervised adversarial invariance (UAI) architecture to disentangle speaker-discriminative information.
As shown in Fig. \ref{fig:UAI-domain-adaptation}, the encoder generates two latent representations $\mathbf{h}_1$ and $\mathbf{h}_2$ from the
x-vector $\mathbf{x}$, where $\mathbf{h}_1$ only contains the speaker-discriminative information, and $\mathbf{h}_2$ contains all other information of $\mathbf{x}$. This was implemented by optimizing the encoder, predictor and decoder together, where the predictor aims to predict speaker labels $\hat{\mathbf{y}}$ from $\mathbf{h}_1$, and the decoder aims to recover $\hat{\mathbf{x}}$ from a concatenation of $\mathbf{h}_2$ and a noise corrupted version of $\mathbf{h}_1$, denoted as $\mathbf{h}_1^\prime$. In order to further encourage the ``disentanglement'', a minimax game between the disentanglers and the encoder was conducted, where the disentanglers tries to reconstructs the two latent representations from each other, while the encoder is optimized against the disentanglers. An important merit of UAI over DANN is that the adversarial game does not need domain labels.
\begin{figure}[t]
\centering
\includegraphics[width=3.2in]{UAI_domain_adaptation.pdf}
\put(-222,106){\small\bfseries\color{white}{$\mathbf{x}$}}
\put(-122,133){\small\bfseries\color{black}{$\mathbf{h}_1$}}
\put(-122,80){\small\bfseries\color{black}{$\mathbf{h}_2$}}
\put(-102,88){\small\bfseries\color{black}{$\mathbf{h}_1^\prime$}}
\put(-14,80){\small\bfseries\color{black}{$\hat{\mathbf{x}}$}}
\put(-14,133){\small\bfseries\color{black}{$\hat{\mathbf{y}}$}}
\put(-14,43){\small\bfseries\color{black}{$\hat{\mathbf{h}}_1$}}
\put(-14,13){\small\bfseries\color{black}{$\hat{\mathbf{h}}_2$}}
\caption{A brief diagram of unsupervised adversarial invariance.}
\label{fig:UAI-domain-adaptation}
\end{figure}
In addition to DANN where $M_s=M_t=M$, another kind of adversarial-training-based domain adaptation methods trains an encoder for each domain, i.e. $M_s \neq M_t$.
Adversarial discriminative domain adaptation (ADDA) shown in Fig. \ref{fig:GAN-adaptation} is such a representative approach \cite{tzeng2017adversarial}.
In \cite{xia2019cross}, {{the authors}} applied ADDA to learn an asymmetric mapping that adapts the target domain encoder to the source domain encoder, where the two domains are in different languages.
\begin{figure}[t]
\centering
\includegraphics[width=3.5in]{GAN-adaptation-eps-converted-to.pdf}
\put(-165,87){\small\bfseries\color{red}{$C$}}
\put(-165,87){\small\bfseries\color{red}{$C$}}
\put(-200,80){\small\bfseries\color{red}{$M_s$}}
\put(-200,80){\small\bfseries\color{red}{$M_s$}}
\put(-150,65){\small\bfseries\color{red}{$\mathcal{L}_{\rm cls}$}}
\put(-150,65){\small\bfseries\color{red}{$\mathcal{L}_{\rm cls}$}}
\put(-47,98){\small\bfseries\color{red}{$D$}}
\put(-47,98){\small\bfseries\color{red}{$D$}}
\put(-80,90){\small\bfseries\color{red}{$M_s$}}
\put(-80,90){\small\bfseries\color{red}{$M_s$}}
\put(-80,45){\small\bfseries\color{red}{$M_t$}}
\put(-80,45){\small\bfseries\color{red}{$M_t$}}
\put(-32,65){\small\bfseries\color{red}{$\mathcal{L}_{{\rm adv}_D}$}}
\put(-32,65){\small\bfseries\color{red}{$\mathcal{L}_{{\rm adv}_D}$}}
\put(-32,45){\small\bfseries\color{red}{$\mathcal{L}_{{\rm adv}_M}$}}
\put(-32,45){\small\bfseries\color{red}{$\mathcal{L}_{{\rm adv}_M}$}}
\caption{Adversarial discriminative domain adaptation (ADDA) approach (from \cite{xia2019cross}). Note that the source encoder is fixed during the adversarial adaptation stage.}
\label{fig:GAN-adaptation}
\end{figure}
From the view of the framework in Fig. \ref{fig:Adversarial-training}, ADDA is trained by the following two successive steps:
\begin{enumerate}[1)~]
\item Pre-train a source domain encoder $M_s$ and a speaker classifier $C$ with the labeled source data by \eqref{eq:adv_class}.
\item Fix the source domain encoder $M_s$, and perform adversarial training on the target encoder $M_t$ and domain discriminator $D$ by alternatively minimizing $\mathcal{L}_{{\rm adv}_D}$ and $\mathcal{L}_{{\rm adv}_M}$ via \eqref{eq:adv_D} and \eqref{eq:adv_Map} respectively. The domain discriminator $D$ minimizes the binary cross-entropy loss, i.e. $ \mathcal{L}_{{\rm adv}_D}= -\mathbb{E}_{\mathbf{x}_s \sim \mathbf{X}_s}[{\rm log}(D(\hat{M}_s(\mathbf{x}_s)))]-\mathbb{E}_{\mathbf{x}_t \sim \mathbf{X}_t}[{\rm log}(1-D(\hat{M}_t(\mathbf{x}_t)))] $. The target encoder $M_t$ minimizes an inverted label loss, i.e. $\mathcal{L}_{{\rm adv}_M}=-\mathbb{E}_{\mathbf{x}_t \sim \mathbf{X}_t}[{\rm log}(\hat{D}(M_t(\mathbf{x}_t)))]$\footnote{Different from \eqref{eq:GAN_loss}, the constant part $-\mathbb{E}_{\mathbf{x}_s \sim \mathbf{X}_s}[{\rm log}(1-\hat{D}(\hat{M}_s(\mathbf{x}_s)))]$ was removed from $\mathcal{L}_{{\rm adv}_M}$, given that $M_s$ is fixed.}.
\end{enumerate}
In the test stage, the data from the target domain is first mapped to the shared feature space by the target domain encoder, and then classified by a back-end classifier trained in the source domain, e.g. PLDA.
In order to compensate the unreliability of short utterances, \cite{zhang2018vector} proposed to compensate short utterances by long utterances using GAN. Specifically, it uses a generator to generate compensated i-vectors from short-utterance i-vectors, and uses a discriminator to determine whether an i-vector is generated by the generator or extracted from a long utterance. Similarly, \cite{9054036} addressed the problem by adversely learning a mapping function that maps short embedding features to enhanced embedding features.
Despite of the success of the adversarial-training-based domain adaptation, its training is not easy in practice. For example, DANN learns a common encoder for both domains, which may make the optimization poorly conditioned, since a single encoder has to handle features from two separate domains \cite{tzeng2017adversarial}. Although ADDA is able to learn domain specific encoders, its target domain has no labels. As a result, without shared weights between the encoders, it may quickly fall into a degenerate solution if not properly initialized \cite{tzeng2017adversarial}. To remedy this weakness, a pre-trained source encoder is used to initialize the target encoder, leaving the source encoder fixed during the adversarial training \cite{tzeng2017adversarial,xia2019cross}. Besides, many training difficulties observed in related areas, such as image processing, have been encountered in speaker recognition as well, though seldom discussed in depth.
\subsubsection{Reconstruction-based domain adaptation}
In the machine learning community, CycleGAN, which was originally proposed for image-to-image translation \cite{zhu2017unpaired}, is one of the most common reconstruction-based domain adaptation methods \cite{wang2018deep}. Recently, it was introduced to speaker recognition \cite{nidadavolu2019cycle, nidadavolu2019low, 9053823}.
\begin{figure}[t]
\centering
\includegraphics[width=3.5in]{Cycle-GAN-eps-converted-to.pdf}
\caption{ The cycle GAN architecture (from \cite{zhu2017unpaired}).}
\label{fig:Cycle-GAN}
\end{figure}
As shown in Fig. \ref{fig:Cycle-GAN}, CycleGAN comprises two generators and two discriminators. The generator $G$ transforms the feature $X$ in a domain to the feature $Y$ in another domain, producing an approximation of $Y$, i.e. $\hat{Y}=G(X)$. The discriminator $D_Y$, which aligns with $G$, discriminates between $\hat{Y}$ and $Y$. The other generator-discriminator pair, i.e. $F$ and $D_X$, is intended to transfer features from $Y$ to $X$. The generators and discriminators are trained using a cycle consistency loss and a combination of two adversarial losses, where the cycle consistency loss measures how well the original input is reconstructed after a sequence of two generators, i.e. $F(G(X)) \approx X$ or $G(F(Y)) \approx Y $. Because of the adversarial and cycle reconstruction mechanisms of CycleGAN, it has an outstanding advantage that its training needs neither speaker labels nor paired data between the source and target domains.
{{At the acoustic feature level, \cite{nidadavolu2019cycle} explored a domain adaptation approach }} by learning feature mappings between a microphone domain and a telephone domain using CycleGAN. It maps the acoustic features from the target domain (microphone) back to the source domain (telephone), and conducts speaker recognition using the system trained in the source domain.
In \cite{nidadavolu2019low}, they further investigated the effectiveness of CycleGAN in low resource scenarios where the target domain only has limited amount of data. They found that, the adaptation system trained on limited amount of target domain data performs slightly better than the adaptation system trained on a larger amount of target domain data, when some noise was added to the data.
In \cite{9053823}, they developed a CycleGAN-based feature enhancement approach in the log-filter bank space to improve the performance of speaker verification in noisy and reverberant environments.
Besides CycleGAN, the encoder-decoder structure is another popular reconstruction-based domain adaptation method in machine learning \cite{wang2018deep}. As for speaker recognition, \cite{shon2017autoencoder} combined an autoencoder with a denoising autoencoder to adapt resource-rich source domain data to the target domain.
\subsubsection{Discrepancy-based domain adaptation}
Discrepancy-based domain adaptation aligns the statistical distribution shift between the source and target domains by using some statistic criteria, including maximum mean discrepancy (MMD), correlation alignment (CORAL), Kullback-Leibler divergence, etc. {{For speaker verification, those criteria has been widely studied in shallow domain adaptation models \cite{alam2018speaker,lee2019coral}. Recently, some of them were also introduced to the deep learning based adaptation approaches. Specifically, \cite{lin2018reducing}}} added a MMD based loss to the reconstruction loss of an autoencoder to train a domain-invariant encoder for multi-source adaptation of i-vectors.
In \cite{lin2018multisource,lin2019semi}, they further proposed a nuisance-attribute autoencoder based on MMD. In \cite{9054134}, they proposed a multi-level deep neural network adaptation method using MMD and consistency regularization.
\subsection{Speech enhancement and de-reverberation preprocessing}\label{subsec:se}
Speech is always distorted by noise and reverberation in real-world scenarios. A natural choice of coping with these distortions for speaker recognition is to add a speech enhancement or de-reverberation preprocessing module. Recently, deep learning based speech enhancement and de-reverberation techniques \cite{wang2018supervised} have been applied to speaker recognition, which can be roughly divided into three categories, i.e. masking-based \cite{zhao2014robust,kolboek2016speech,chang2017robust,zhao2019robust,shon2019voiceid}, mapping-based \cite{plchot2016audio,novotny2018on,novotny2019analysis,oo2016dnn,sun2018speaker}, and GAN-based \cite{michelsanti2017conditional} techniques.
{{It should be note that this section only reviews some general concepts and provides useful cues without considering the details of the speech enhancement techniques, since it is out the scope of this paper. More details can be found in speech enhancement related papers \cite{wang2018supervised}.}}
Masking-based speech enhancement has received a lot of attention and shown impressive performance in speech quality and speech intelligibility. It uses a DNN to estimate a time-frequency mask of noisy speech, and then uses the mask to recover the corresponding clean speech. In \cite{zhao2014robust,kolboek2016speech,chang2017robust}, the authors applied masking-based speech enhancement techniques {{as an independently noise reduction module}} for speaker recognition. In \cite{zhao2019robust}, {{the authors}} jointly optimized speech separation and speaker verification networks together. {{Besides,}} \cite{shon2019voiceid} designed a \textit{VoiceID loss} which jointly optimize a pre-trained speaker embedding system and a speech enhancement network where the speaker embedding system is fixed during the joint optimization, as shown in Fig. \ref{fig:VoiceID_Loss}.
\begin{figure*}[t]
\centering
\includegraphics[width=5 in]{VoiceID_Loss-eps-converted-to.pdf}
\caption{Flow chart of the VoiceID loss (from \cite{shon2019voiceid}).}
\label{fig:VoiceID_Loss}
\end{figure*}
DNN-based autoencoder, which maps noisy speech directly to its clean counterpart, is another speech enhancement method for speaker recognition. {{\cite{plchot2016audio}}} trained an autoencoder to map the log magnitude spectrum of noisy speech to its clean counterpart by minimizing the mean squared error, which demonstrated its effectiveness on the text-dependent GMM-MAP and text-independent i-vector systems.
The authors of \cite{novotny2018on} explored a similar noise reduction method with \cite{plchot2016audio}, and applied it to an x-vector system \cite{novotny2019analysis}. In \cite{oo2016dnn}, the authors enhanced both the amplitude feature and the phase feature, where they used MFCC as the amplitude feature, and modified group delay cepstral coefficients as the phase feature. They concluded that simultaneous enhancing the amplitude and phase features is more effective than enhancing their individual components alone. Apart from the simplest feedforward DNN, more complicated LSTM based speech enhancement methods were also explored \cite{sun2018speaker}.
Besides the masking and mapping based speech enhancement, GAN-based speech enhancement methods for speaker recognition were also developed. In specific, {{\cite{michelsanti2017conditional}}} used conditional GANs to learn a mapping from a noisy spectrum to its enhanced counterpart. The conditional GAN consists of a generator and a discriminator which are trained in an adversarial manner. The generator enhances the noisy spectrum; the discriminator aims to distinguish the enhanced spectrum from their clean counterpart using the noisy spectrum as a condition. In addition, {{\cite{yu2017adversarial}}} proposed a noise robust bottleneck feature extraction method based on adversarial training.
In addition to the above speech enhancement methods for speaker recognition, speech de-reverberation has also been studied in speaker recognition \cite{guzewich2017improving,movsner2018dereverberation,Mošner2018}. Recently, far-field and multi-channel speaker recognition also attracted much attention \cite{qin2019far,taherian2019deep,cai2019multi,taherian2020robust}.
\subsection{Data augmentation for robust speaker recognition}
\label{sec:data_augmentation}
Large-scale multi-condition training is an effective way to improve the generalization of speaker recognition in noisy environments. Particularly, we have observed that the performance of deep speaker embedding systems appear to be highly dependent on the amount of training data. One way to prepare large-scale noisy training data is \textit{data augmentation}. In \cite{snyder2018x}, {{the authors}} employed additive noises and reverberation to the original training data for the data augmentation of x-vectors, which has shown to be very effective. {{\cite{zhu2019mixup}}} applied a mixup learning strategy to improve the generalization of x-vector extractors.
To improve the performance of speaker verification for children with limited data, \cite{9053891} made speed and pitch perturbation as well as voice conversion to increase the amount of training data. {{Besides, }} \cite{9053481} validated the effectiveness of \textit{spectral augmentation}, which was originally proposed for speech recognition, for deep speaker embeddings.
\subsection{Other robust methods}
Apart from the aforementioned methods, some other robust speaker recognition methods are as follows. For example, \cite{9053110} optimized a deep feature loss for feature-domain enhancement of x-vector extractors. \cite{9053407} designed a new loss function for noise-robust speaker recognition. \cite{kim2019deep} proposed an orthogonal vector pooling strategy to remove unwanted factors. There are also many robust back-ends for speaker verification \cite{ghahabi2017deep,bhattacharya2016modelling,yang2017applying,mahto2017vector,guo2018deep}.
\section{Datasets}
\label{sec:data_tool}
\renewcommand\arraystretch{1.2}
\begin{table*}[t]
\centering
{{
\caption{Popular databases and challenges for speaker recognition. The term ``wild'' denotes that the audio is acquired across unconstrained conditions. The term ``quite'' denotes that the data were recorded indoor under a typical office environment. The term ``SV'',``SI'' and ``SR'' are the abbreviations of ``speaker verification'', ``speaker identification'' and ``speaker recognition'' respectively. } \label{tab:challenge-dataset}
\scalebox{0.655}{
\begin{tabular}{p{3.4cm} p{1cm} p{3cm} p{2.2cm} p{2cm} p{3.5cm} p{3cm} p{3cm} p{2.8cm}}
\hline
\hline
Dataset & Year& Condition & Language & Sample rate& Application & Speakers & Data amount & Annotation method \\
\hline
NIST SRE \cite{greenberg2020two} & 1996 $\sim$ 2020 & Clean, noisy & Multilingual &--- & Text-independent SR & --- & --- & Hand annotated \\
VoxCeleb1 \cite{nagrani2017voxceleb} & 2017& Multi-media (wild) &Mostly English &16kHZ &\makecell[l]{Text-independent SV \\ and SI} & 1,251 (690 males) &\makecell[l]{153,516 utterances, \\ 352 hours} & Automated pipeline \\
VoxCeleb2 \cite{chung2018voxceleb2} & 2018 & Multi-media (wild) & \makecell[l]{Multilingual \\ Mostly English}&--- & \makecell[l]{Text-independent SV \\ and SI}& 6,112 (3761 males) &\makecell[l]{1,128,246 utterances, \\ 2,442 hours} & Automated pipeline \\
SITW \cite{mclaren2016speakers} & 2016 & Multi-media (wild) & --- & 16kHZ & \makecell[l]{Text-independent single \\ and multi-speaker SV} & 299 (203 males) & 2,800 utterances & Hand annotated \\
RSR2015 \cite{larcher2014text} & 2015& \makecell[l]{Smart-phones and \\tablets (quite)} & English &16 kHZ&Text-dependent SV& 300 (157 males) &\makecell[l]{196,844 files, \\ 151 hours}& Hand annotated \\
RedDots \cite{lee2015reddots} & 2015 & \makecell[l]{Mobile devices \\(through internet)} & English &---& \makecell[l]{Fixed phrase, free speech \\ and text-prompted SV}& 45 & --- & Manual, automatic, or semi-automatic \\
VOICES \cite{richey2018voices,nandwana2019voices} & 2018 & \makecell[l]{Far-field microphones \\(noisy room)} & English &48 kHZ& \makecell[l]{Text-independent SV \\and SI}& 300 & \makecell[l]{374,688 files, \\ 1440 hours} & Hand annotated \\
Librispeech\cite{panayotov2015librispeech}& 2015 & ---& English & 16 kHz & ASR and SR &Over 9,000 & 1000 hours & Manually annotated \\
CN-CELEB\cite{9054017} & 2019 & Multi-media (wild)& Chinese & --- &Text-independent SV &1000 & \makecell[l]{130,109 utterances, \\ 274 hours} & Automated pipeline with human check \\
BookTubeSpeech \cite{9053258}& 2020 & Multi-media & --- & --- &Text-independent SV& 8450 & --- & Automatic pipeline \\
Hi-MIA \cite{9054423} & 2020 & \makecell[l]{Microphone arrays \\ (rooms, far-field)}& \makecell[l]{Chinese,\\ English} &16kHZ, 44.1kHz & Text-dependent SV & 340 (175 male) & \makecell[l]{More than 3,936,003 \\utterance, 1,561 hours} & Hand annotated \\
FFSVC 2020 \cite{qin2020ffsvc} & 2020 & Close-talk cellphone, far-field microphone arrays (far-field)& Mandarin & 16kHZ, 48kHZ & \makecell[l]{Text-dependent and \\ text-independent SV} & --- & --- & Hand annotated \\
DIHARD1 \cite{ryant2018first} & 2018 & \makecell[l]{Single channel (wild)}& English (most), Mandarin & 16kHZ& Speaker diarization& --- & 40 hours & Hand annotated \\
DIHARD2 \cite{ryant2019secondDIHARD}& 2019 & \makecell[l]{Single channel and \\ multichannel (wild)} & English (most), Mandarin &16kHZ& Speaker diarization & --- & \makecell[l]{503 files, \\ 339.95 hours} & Hand annotated \\
AMI \cite{carletta2005ami} & 2005 & Multi-modal& English & --- & Speaker diarization & --- & 100 hours & Hand annotated \\
\hline
\hline
\end{tabular}
}
}}
\end{table*}
In this section, we make an overview to existing challenges, and datasets for speaker recognition.
Table \ref{tab:challenge-dataset} summarizes the brief information of most common and some recently developed datasets. The recording methods of the databases are briefly introduced as follows.
\begin{itemize}
\item NIST SRE: The National Institute of Standards and Technology (NIST) of America has successfully conducted 15 Speaker Recognition Evaluations (SREs) in the past 20 years. It is the largest and most popular challenge in speaker recognition. More information can be found in \cite{greenberg2020two,gonzalez2014evaluating}.
\item VoxCeleb1,2: The VoxCeleb dataset was collected by a fully automated pipeline based on computer vision techniques from open-source media. The pipeline obtains videos from YouTube, performs active speaker verification using a two-stream synchronization CNN, and confirms the identity of the speaker using CNN based facial recognition \cite{nagrani2017voxceleb,chung2018voxceleb2}.
\item SITW: The speakers in the wild (SITW) database contains hand-annotated speech samples from open-source media for the purpose of benchmarking text-independent speaker recognition technology on single and multi-speaker audio acquired across unconstrained conditions \cite{mclaren2016speakers}.
\item RSR2015: The RSR2015 database were recorded indoor under a typical office environment with six mobile devices (five smart-phones and one tablet) \cite{larcher2014text}.
\item RedDots: The RedDots project used a mobile app as the recording front-end. Speakers recorded their voices offline and later on uploaded the recordings to an Apache web server when Internet connection was available \cite{lee2015reddots}.
\item VOICES: The voices obscured in complex environmental settings (VOICES) corpus was recorded in furnished rooms with background noise played in conjunction with foreground speech selected from the LibriSpeech corpus. The displayed noises includes television, music, or overlapping speech from multiple speakers (referred to as babble) \cite{richey2018voices,nandwana2019voices}.
\item LibriSpeech: While intended created for speech recognition rather than verification or diarization, LibriSpeech does include labels of speaker identities and is thus useful for speaker recognition \cite{panayotov2015librispeech,9053258}.
\item CN-CELEB: CN-CELEB was collected following a two-stage strategy: firstly {{the authors}} used an automated pipeline to extract potential segments of the Person of Interest from ``bilibili.com'', and then they applied a human check to remove incorrect segments \cite{9054017}.
\item BookTubeSpeech: The BookTubeSpeech was collected by an automatic pipeline from BookTube videos \cite{9053258}.
\item Hi-MIA: Hi-MIA database was designed for far-field scenarios. Recordings were captured by multiple microphone arrays located in different directions and distance to the speaker and a high-fidelity close-talking microphone \cite{9054423}.
\item FFSVC 2020: The far-field speaker verification challenge 2020 (FFSVC20) is designed to boost the speaker verification research with special focus on far-field distributed microphone arrays under noisy conditions in real scenarios \cite{qin2020interspeech,qin2020ffsvc}.
\item DIHARD1,2: The DIHARD challenge intended to improve the robustness of diarization systems to variation in recording equipment, noise conditions, and conversational domain \cite{ryant2018first,ryant2019secondDIHARD}.
\item AMI: The AMI meeting corpus is a multi-modal data set consisting of 100 hours of meeting recordings, and it was recorded using a wide range of devices including close-talking and far-field microphones, individual and room-view video cameras, projection, a whiteboard, and individual pens, all of which produce output signals that are synchronized with each other \cite{carletta2005ami}.
\end{itemize}
\section{Conclusions and discussions}
\label{sec:discussion_conclusion}
This paper has provided a comprehensive overview of the deep learning based speaker recognition. We have analysed the relationship between different subtasks, including speaker verification, identification, and diarization, and summarized some common difficulties. Based on the analysis, we summarized the subtasks from three widely studied core issues---speaker feature extraction, speaker diarization, and robust speaker recognition. For speaker feature extraction, we reviewed two kinds of hybrid structures, which are DNN-UBM/i-vector and DNN-BNF/i-vector. In addition, the overview of the state-of-the-art deep speaker embedding were made in respect of four key components, which are the inputs, network structures, temporal pooling strategies, and loss functions respectively. Particularly, we reviewed the loss functions of the end-to-end speaker verification for feature learning from the perspective of different training sample construction methods.
For speaker diarization, we reviewed stage-wise diarization, supervised end-to-end diarization, online diarization, and multimodal diarization. For robust speaker recognition, we surveyed three kinds of deep learning based domain adaptation methods as well as several speech preprocessing methods, which deal with the domain mismatch and back-ground noise respectively. Some popular and recently developed datasets were summarized as well.
To conclude, deep learning has boosted the performance of speaker recognition to a new high level. We make our best to summarize the recent rapid progress of the deep learning based speaker recognition, hopefully this provides a knowledge resource and further blooms the research community.
Although the deep learning based speaker recognition has achieved a great success, many issues remain to be addressed. Here we list some open problems from the perspectives of network training, loss functions, real-world diarization, and domain adaptation respectively. For the network training, most speaker feature extraction methods need handcraft acoustic features as the input, which may not be optimal. The state-of-the-art deep models have a large number of parameters, which are difficult to be applied to portable devices. The network training also needs large amounts of labeled training data and heavy computation resources.
For the loss functions, although so many loss functions have been proposed, there is lack of strong theoretical base for the success of the loss functions, nor theoretical guidance that could lead to better loss functions. Although the verification losses for the end-to-end speaker verification meet the verification process tightly, their potentials have not been fully developed yet.
For the real-world diarization, from the recent DIHARD challenges, one can see that speaker diarization is still a hard problem. In a real-world conversation, a speech recording may be contaminated by serious speech overlap and strong background noise. Some technically difficult problems, such as the unknown number of speakers and rapid speaker changes, also hinder the performance of speaker diarization in real-world applications severely.
Finally, although many domain adaptation algorithms have been proposed, especially those based on adversarial learning, they have not made landmark progress compared to traditional shallow adaptation methods, e.g. the PLDA based adaptation, which needs further efforts.
\section{Acknowledgement}
The authors are grateful to Prof. DeLiang Wang, the Co-Editor-in-Chief, and the anonymous reviewers for their valuable comments,
which helped greatly improve the quality of the paper. The authors would also like to thank our colleagues Meng-Zhen Li and Rui Wang for their helpful discussions.
\bibliographystyle{elsarticle-num}
|
\section{Introduction}
\vspace{-4pt}
The visual/LiDAR odometry is one of the key technologies in autonomous driving. This task uses two consecutive images or point clouds to obtain the relative pose transformation between two frames, and acts as the base of the subsequential planning and decision making of mobile robots \cite{robot_}. Recently, learning-based odometry methods have shown impressive accuracy on datasets compared with conventional methods based on hand-crafted features. It is found that learning-based methods can deal with sparse features and dynamic environments \cite{dasgil,lpdnet}, which are usually difficult for conventional methods. To our knowledge, most learning-based methods are on the 2D visual odometry \cite{wang2017deepvo,zhou2017unsupervised,min2020voldor,yang2020d3vo,Unsupervised_depth,chizhang} or utilize 2D projection of LiDAR \cite{nicolai2016deep,velas2018cnn,wang2019deeppco,li2019net,li2020dmlo} and the LiDAR odometry in 3D point clouds is underexplored. This paper aims to estimate the LiDAR odometry directly through raw 3D point clouds.
\begin{figure}[t]
\centering
\resizebox{0.90\columnwidth}{!}
{
\includegraphics[scale=1.0]{fignetwork_introduction_1-eps-converted-to.pdf}}
\vspace{0mm}
\caption{The Point Feature Pyramid, Pose Warping, and Attentive Cost Volume (PWC) structure in our proposed PWCLO-Net. The pose is refined layer by layer through iterative pose warp-refinement. The whole process is realized end-to-end by making all modules differentiable. In the LiDAR point clouds, the small red points are the whole point cloud. The big black points are the sampled points in $PC_1$. Distinctive colors of big points in embedding mask measure the contribution of sampled points to the pose estimation.}
\label{fig:network_little}
\vspace{-12pt}
\end{figure}
For the LiDAR odometry in 3D point clouds, there are three challenges: 1) As the discrete LiDAR point data are obtained in two consecutive frames separately, it is hard to find an precisely corresponding point pair between two frames; 2) Some points belonging to an object in a frame may not be seen in the other view if they are occluded by other objects or are not captured because of the limitation of LiDAR resolution; 3) Some points belonging to dynamic objects are not suitable to be used for the pose estimation because these points have uncertainty motions of the dynamic objects.
For the first challenge, Zheng et al. \cite{zheng2020lodonet} use matched keypoint pairs judged in 2D depth images. However, the correspondence is rough because of the discrete perception of LiDAR. The cost volume for 3D point cloud \cite{wu2019pointpwc,wang2020hierarchical} is adopted in this paper to obtain a weighted soft correspondence between two consecutive frames. For the second and third challenges, the mismatched points or dynamic objects, which do not conform to the overall pose, need to be filtered. LO-Net~\cite{li2019net} trains an extra mask estimation network~\cite{zhou2017unsupervised, yang2018unsupervised} by weighting the consistency error of the normal of 3D points. In our network, an internal trainable embedding mask is proposed to weigh local motion patterns from the cost volume to regress the overall pose. In this way, the mask can be optimized for more accurate pose estimation rather than depending on geometry correspondence. In addition, the PWC structure is built to capture the large motion in the layer of sparse points and refine the estimated pose in dense layers. As shown in Fig.~\ref{fig:network_little}, the embedding mask is also optimized hierarchically to obtain more accurate filtering information to refine pose estimation.
Overall, our contributions are as follows:
\vspace{-0.1cm}
\begin{itemize}
\vspace{-0.2cm}
\item The Point Feature Pyramid, Pose Warping, and Cost Volume (PWC) structure for the 3D LiDAR odometry task is built to capture the large motion between two frames and accomplish the trainable iterative 3D feature matching and pose regression.
\vspace{-0.3cm}
\item In this structure, the hierarchical embedding mask is proposed to filter mismatched points and convert the cost volume embedded in points to the overall ego-motion in each refinement level. Meanwhile, the embedding mask is optimized and refined hierarchically to obtain more accurate filtering information for pose refinement following the density of points.
\vspace{-0.3cm}
\item Based on the characteristic of the pose transformation,
the pose warping and pose refinement are proposed to refine the estimated pose layer by layer iteratively. A totally end-to-end framework, named PWCLO-Net, is established, in which all modules are fully differentiable so that each process is no longer independent and combinedly optimized.
\vspace{-0.3cm}
\item Finally, our method is demonstrated on KITTI odometry dataset \cite{geiger2012we,geiger2013vision}. The evaluation experiments and ablation studies show the superiority of the proposed method and the effectiveness of each design. To the best of our knowledge, our method outperforms all recent learning-based LiDAR odometry and even outperforms the geometry-based LOAM with mapping optimization \cite{zhang2017low} on most sequences.
\vspace{-0.2cm}
\end{itemize}
\section{Related Work}
\vspace{-4pt}
\subsection{Deep LiDAR Odometry}
\vspace{-4pt}
Deep learning has gained impressive progress in visual odometry \cite{min2020voldor,yang2020d3vo}. However, the 3D LiDAR odometry with deep learning is still a challenging problem. In the beginning, Nicolai et al. \cite{nicolai2016deep} project two consecutive LiDAR point clouds to the 2D plane to obtain two 2D depth images, and then use the 2D convolution and fully connected (FC) layers to realize learning-based LiDAR odometry. Their work verified that the learning-based method is serviceable for the LiDAR odometry although their experiment results are not superior. Velas et al. \cite{velas2018cnn} also project LiDAR points to the 2D plane but use three channels to encode the information, including height, range, and intensity. Then the convolution and FC layers are used for the pose regression. The performance is superb when only estimating the translation but is poor when estimating the 6-DOF pose. Wang et al. \cite{wang2019deeppco} project point clouds to panoramic depth images and stack two frames together as input. Then the translation sub-network and FlowNet \cite{dosovitskiy2015flownet} orientation sub-network are used to estimate the translation and orientation respectively. \cite{li2019net} also preprocess 3D LiDAR points to 2D information but use the cylindrical projection \cite{chen2017multi}. Then, the normal of each 3D point is estimated to build consistency constraint between frames, and an uncertainty mask is estimated to mask dynamic regions. Zheng et al. \cite{zheng2020lodonet} extract matched keypoint pairs by classic detecting and matching algorithm in 2D spherical depth images projected from 3D LiDAR point clouds. Then the PointNet \cite{qi2017pointnet} based structure is used to regress the pose from the matched keypoint pairs. \cite{li2020dmlo} proposes a learning-based network to generate matching pairs with high confidence, then applies Singular Value Decomposition (SVD) to get the 6-DoF pose. \cite{cho2020} introduces an unsupervised learning method on LiDAR odometry.
\begin{figure*}[t]
\centering
\resizebox{1.00\textwidth}{!}
{
\includegraphics[scale=1.00]{fignetwork-eps-converted-to.pdf}}
\vspace{-7mm}
\caption{The details of proposed PWCLO-Net architecture. The network is composed of four set conv layers in point feature pyramid, one attentive cost volume, one initial embedding mask and pose generation module, and three pose warp-refinement modules. The network outputs the predicted poses from four levels for supervised training. }
\label{fig:network}
\vspace{-12pt}
\end{figure*}
\vspace{-0.1cm}
\subsection{Deep Point Correlation}
\vspace{-4pt}
The above studies all use the 2D projection information for the LiDAR odometry learning, which converts the LiDAR odometry to the 2D learning problem. Wang et al. \cite{wang2019deeppco} compared the 3D point input and 2D projection information input based on the same 2D convolution model. It is found that the 3D input based method has a poor performance. As the development of 3D deep learning \cite{qi2017pointnet,qi2017pointnet++}, FlowNet3D \cite{liu2019flownet3d} proposes an embedding layer to learn the correlation between the points in two consecutive frames. After that, Wu et al. \cite{wu2019pointpwc} propose the cost volume method on point clouds, and Wang et al. \cite{wang2020hierarchical} develop the attentive cost volume method. The point cost volume involves the motion patterns of each point. It becomes a new direction and challenge to regress pose from the cost volume, and meanwhile, not all point motions are for the overall pose motion. We exploit to estimate the pose directly from raw 3D point cloud data and deal with the new challenges encountered.
What is more, we are inspired by the Pyramid, Warping, and Cost volume (PWC) structure in the flow network proposed by Su et al. \cite{sun2018pwc}. The work uses the three modules (Pyramid, Warping, and Cost volume) to refine the optical flow through a coarse-to-fine approach. The works \cite{wu2019pointpwc, wang2020hierarchical} on 3D scene flow also use the PWC structure to refine the estimated 3D scene flow in point clouds. In this paper, the idea is applied to the pose estimation refinement, and a PWC structure for LiDAR odometry is built for the first time.
\vspace{-0.2cm}
\section{PWCLO-Net}
\vspace{-4pt}
Our method learns the LiDAR odometry from raw 3D point clouds in an end-to-end approach with no need to pre-project the point cloud into 2D data, which is significantly different from the deep LiDAR odometry methods introduced in the related work section.
The Fig.~\ref{fig:network} illustrates the overall structure of our PWCLO-Net. The inputs to the network are two point clouds $PC_1 = \{ {x_i}|{x_i} \in {\mathbb{R}^3}\} _{i = 1}^N$ and $PC_2 = \{ {y_j}|{y_j} \in {\mathbb{R}^3}\} _{j = 1}^N$ respectively sampled from two adjacent frames. These two point clouds are firstly encoded by the siamese feature pyramid consisting of several set conv layers as introduced in Sec.~\ref{sec:pyramid}. Then the attentive cost volume is used to generate embedding features, which will be described in Sec.~\ref{sec:cost}. To regress pose transformation from the embedding features, hierarchical embedding mask optimization is proposed in Sec.~\ref{sec:mask}. Next, the pose warp-refinement method is proposed in Sec.~\ref{sec:refine} to refine the pose estimation in a coarse-to-fine approach. Finally, the network outputs the quaternion $q \in {\mathbb{R}^4}$ and translation vector $t \in {\mathbb{R}^3}$.
\subsection{Siamese Point Feature Pyramid}\label{sec:pyramid}
\vspace{-4pt}
The input point clouds are usually disorganized and sparse in a large 3D space. A siamese feature pyramid consisting of several set conv layers is built to encode and extract the hierarchical features of each point cloud. Farthest Point Sampling (FPS)~\cite{qi2017pointnet++} and shared Multi-Layer Perceptron (MLP) are used. The formula of set conv is:
{\setlength\abovedisplayskip{4pt}\setlength\belowdisplayskip{4pt}\begin{equation} {f_i} = \mathop {MAX}\limits_{k = 1,2,...K} (MLP((x_i^k - x_i) \oplus f_i^k \oplus f_i^c)),\end{equation}}where $x_i$ is the obtained $i$-th sampled point by FPS. And $K$ points $x_i^k$ $(k=1,2,...,K)$ are selected by $K$ Nearest Neighbors (KNN) around $x_i$. $f_i^c$ and $f_i^k$ are the local features of $x_i$ and $x_i^k$ (they are null for the first layer in the pyramid). ${f_i}$ is the output feature located at the central point $x_i$. $\oplus$ denotes the concatenation of two vectors, and $\mathop {MAX}\limits_{k = 1,2,...K}() $ indicates the max pooling operation. The hierarchical feature pyramid is constructed as shown in Fig.~\ref{fig:network}. The siamese pyramid \cite{chopra2005learning} means that the learning parameters of the built pyramid are shared for these two point clouds.
\subsection{Attentive Cost Volume}\label{sec:cost}
\vspace{-4pt}
Next, the point cost volume with attention in \cite{wang2020hierarchical} is adopted here to associate two point clouds. The cost volume generates point embedding features by associating two point clouds after the feature pyramid.
The embedding features contain the point correlation information between two point clouds. As shown in Fig.~\ref{fig:cost}, $F_1 = \{ {f_i}|{f_i} \in {\mathbb{R}^c}\} _{i = 1}^n$ is the features of point cloud $PC_1 = \{ {x_i}|{x_i} \in {\mathbb{R}^3}\} _{i = 1}^n$ and $F_2 = \{ {g_j}|{g_j} \in {\mathbb{R}^c}\} _{j = 1}^n$ is the features of point cloud $PC_2 = \{ {y_j}|{y_i} \in {\mathbb{R}^3}\} _{i = 1}^n$. The embedding features between two point clouds are calculated as follows:
\begin{equation}w_{1, i}^k = softmax(u(x_i,y_j^k,f_i,g_j^k))_{k = 1}^{K_1},\end{equation}\vspace{-0.5cm}
\begin{equation}pe_{i} = \sum\limits_{k = 1}^{k_1} {{w_{1,i}^k} \odot v(x_i,y_j^k,f_i,g_j^k)},\end{equation}\vspace{-0.2cm}
\begin{equation}w_{2,i}^k = softmax(u(x_i,x_i^k,pe_{i},pe_{i}^k))_{k = 1}^{K_2},\end{equation}\vspace{-0.2cm}
\begin{equation}e_{i} = \sum\limits_{k = 1}^{k_2} {w_{2,i}^k \odot v(x_i,x_i^k,pe_{i},pe_{i}^k)},\end{equation}
where $y_j^k$ and $g_j^k$ represent the coordinates and local features of the selected $K_1$ points in $PC_2$ respectively. $\odot$ represents dot product. $u(\cdot)$ and $v(\cdot )$ represent the attention encoding and feature encoding functions referring to \cite{wang2020hierarchical}. $u(\cdot)$ encodes the 3D Euclidean space information and point features to generate attention weights, and $v(\cdot)$ represents the further feature encoding on spacial information and features of two frames of point cloud. The output $E = \{ {e_{i}}|{e_{i}} \in {\mathbb{R}^c}\} _{i = 1}^n$ is the embeding features in $PC_1$.
\begin{figure}[t]
\centering
\resizebox{1.0\columnwidth}{!}
{
\includegraphics[scale=1.00]{figcost-half.pdf}}
\vspace{-5mm}
\caption{Attention Cost-volume. This module takes two frames of point clouds with their local features as input and associates the two point clouds. Finally, the module outputs the embedding features located in $PC_1$.}
\label{fig:cost}
\vspace{-8pt}
\end{figure}
\subsection{Hierarchical Embedding Mask Optimization}\label{sec:mask}
\vspace{-4pt}
It is a new problem to convert the embedding features $E$ to a global consistent pose transformation between two frames. In this subsection, a novel embedding mask is proposed to generate pose transformation from embedding features.
It should be noted that some points may belong to dynamic objects or be occluded in the other frame. It is necessary to filter out these points and keep the points that are of value to the LiDAR odometry task. To deal with this, the embedding features $E = \{ {e_{i}}|{e_{i}} \in {\mathbb{R}^c}\} _{i = 1}^n$ and the features $F_1$ of $PC_1$ are input to a shared MLP followed by the softmax operation along the point dimension to obtain the embedding mask (as the initial embedding mask in Fig.~\ref{fig:network}):
\vspace{-0.2cm}
\begin{equation}
{M} = softmax(sharedMLP({E} \oplus F_1)),\vspace{-0.2cm}\end{equation}
where $M = \{ {m_{i}}|{m_{i}} \in {\mathbb{R}^c}\} _{i = 1}^n$ represents trainable masks for prioritizing embedding features of $n$ points in $PC_1$. Each point has a characteristic weight ranging from 0 to 1. The smaller the weight of a point is, the more likely the point needs to be filtered out, and vice versa. Then, the quaternion $q \in {\mathbb{R}^4}$ and translation vector $t \in {\mathbb{R}^3}$ can be generated by weighting embedding features and FC layers separately, and $q$ is normalized to accord with the characteristic of rotation.\vspace{-0.3cm}
\begin{equation}
q = \frac{{FC(\sum\limits_{i = 1}^n {{e_i} \odot {m_i}} )}}{{\left| {FC(\sum\limits_{i = 1}^n {{e_i} \odot {m_i}} )} \right|}}\label{eqn:q},\vspace{-0.2cm}\end{equation}
\vspace{-0.1cm}
\begin{equation}
t = FC(\sum\limits_{i = 1}^n {{e_{i}} \odot {m_{i}}} )\label{eqn:t}.\vspace{-0.2cm}\end{equation}
The trainable mask $M$ is also part of the hierarchical refinement. As shown in Fig.~\ref{fig:network}, the embedding mask is propagated to denser layers of the point cloud just like embedding features $E$ and pose. The embedding mask is also optimized in a coarse-to-fine approach during the warp-refinement process, making the final mask estimation and the calculation of pose transformation accurate and reliable. We call this process the hierarchical embedding mask optimization.
\subsection{Pose Warp-Refinement Module}\label{sec:refine}
\vspace{-4pt}
To achieve the coarse-to-fine refinement process in an end-to-end fashion, we propose the differentiable warp-refinement module based on pose transformation as shown in Fig.~\ref{fig:warping}. This module contains several key parts: set upconv layer, pose warping, embedding feature and embedding mask refinement, and pose refinement.
\vspace{5pt}
\noindent{}{\bf Set Upconv Layer:}
To refine the pose estimation in a coarse-to-fine approach, the set upconv layer \cite{liu2019flownet3d} is adopted here to enable features of point cloud to propagate from sparse level to dense level. Embedding features $E^{l+1}$ and embedding masks $M^{l+1}$ of $l+1$ layer are propagated through set upconv layer to obtain coarse embedding features $CE^{l}=\{ {ce_{i}^l}|{ce_{i}^l} \in {\mathbb{R}^{c^l}}\} _{i = 1}^{n^l}$ and coarse embedding masks $CM^{l}= \{ {cm_{i}^l}|{cm_{i}^l} \in {\mathbb{R}^{c^l}}\} _{i = 1}^{n^l}$ that need to be optimized at the $l$-th level.
\vspace{5pt}
\noindent{}{\bf Pose Warping:}
The process of pose warping means that the quaternion $q^{l+1}$ and translation vector $t^{l+1}$ from the ${(l+1)}$-th level are applied to warp $PC_1^l=\{ {x_{i}^l}|{x_{i}^l} \in {\mathbb{R}^{c^l}}\} _{i = 1}^{n^l}$ to generate $PC_{1,warped}^l=\{ {x_{i,warped}^l}|{x_{i,warped}^l} \in {\mathbb{R}^{c^l}}\} _{i = 1}^{n^l}$. The warped $PC_{1,warped}^l$ is closer to $PC_2^l$ than original $PC_{1}^l$, which makes the residual motion estimation easier at the $l$-th level. The equation of warping transformation is as follows:\vspace{-0.2cm}
\begin{equation}
[0,x_{i,warped}^l] = {q^{l+1}}[0,x_{i}^l](q^{l+1})^{-1}+[0, t^{l+1}].\vspace{-0.2cm}\end{equation}
Then, the attentive cost volume between $PC{_{1,warped}^l}$ and $PC_2^l$ is re-calculated to estimate the residual motion. Following the approach introduced in Sec.~\ref{sec:cost}, re-embedding features between $PC{_{1,warped}^l}$ and $PC_2^l$ are calculated and denoted as $RE^l = \{ {re_{i}^l}|{re_{i}^l} \in {\mathbb{R}^{c^l}}\} _{i = 1}^{n^l}$.
\begin{figure}[t]
\centering
\vspace{-0mm}
\resizebox{1.00\columnwidth}{!}
{
\includegraphics[scale=1.00]{fignetwork_warp.pdf}}
\vspace{-5mm}
\caption{The details of the proposed Pose Warp-Refinement module at the $l$-th level.}
\label{fig:warping}
\vspace{-10pt}
\end{figure}
\vspace{5pt}
\noindent{}{\bf Embedding Feature and Embedding Mask Refinement:}
The generated coarse embedding feature $ce_i^{l}$, the re-embedding feature $re_{i}^l$, and the features $f_i^l$ of $PC_1^l$ are concatenated and input to a shared MLP to obtain embedding features $E^{l}=\{ {e_{i}^l}|{e_{i}^l} \in {\mathbb{R}^{c^l}}\} _{i = 1}^{n^l}$ at the $l$-th level:
\vspace{-0.2cm}
\begin{equation}
\vspace{-0.1cm}
e_{i}^l = MLP(ce_{i}^l \oplus re_{i}^{l} \oplus f_i^l).
\vspace{-0.2cm}\end{equation}The output of this MLP is the optimized embedding features of $l$-th level, which will not only participate in the following pose generation operation but also be output as the input to the next level warp-refinement module.
Like the refinement of the embedding feature, the newly generated embedding feature $e_{i}^l$, the generated coarse embedding mask $cm_i^{l}$, and the local feature $f_i^l$ of $PC_1^l$ are concatenated and input to a shared MLP and softmax operation along the point dimension to obtain the embedding mask $M^{l}= \{ {m_{i}^l}|{m_{i}^l} \in {\mathbb{R}^{c^l}}\} _{i = 1}^{n^l}$ at the $l$-th level
\vspace{-0.1cm}
\begin{equation}
M^l = softmax(sharedMLP(E^l \oplus CM^{l} \oplus F^l)).
\vspace{-0.2cm}
\end{equation}
\vspace{0pt}
\noindent{}{\bf Pose Refinement:}
The residual $\Delta q^l$ and $\Delta t^l$ can be obtained from the refined embedding features and mask following the formulas~(\ref{eqn:q}) and (\ref{eqn:t}) in Sec.~\ref{sec:mask}. Lastly, the refined quaternion $q^l$ and translation vector $t^l$ of the $l$-th level can be calculated by:\vspace{-0.2cm}\begin{equation}{q^l} = {\Delta q^{l}}{q^{l + 1}},\end{equation}\vspace{-0.5cm}\begin{equation}[0,t^l] = {\Delta q^{l}}[0,t^{l+1}](\Delta q^{l})^{-1}+[0, \Delta t^{l}].\end{equation}
\subsection{Training Loss}
\vspace{-4pt}
The network outputs quaternion $q^l$ and translation vector $t^l$ from four different levels of point clouds. The outputs of each level will enter into a designed loss function and be used to calculate the supervised loss $\ell^l$. Due to the difference in scale and units between translation vector $t$ and quaternion $q$, two learnable parameters $s_x$ and $s_q$ are introduced like previous deep odometry work \cite{li2019net}. The training loss function at the $l$-th level is: \begin{equation}\begin{gathered}
{\ell ^l} = {\left\| {{t_{gt}} - {t^l}} \right\|}exp( - {s_x}) + {s_x} \hfill \\
{\text{ }} + {\left\| {{q_{gt}} - \frac{{{q^l}}}{{\left\| {{q^l}} \right\|}}} \right\|_2}exp( - {s_q}) + {s_q},
\end{gathered} \label{eqn:sxsq} \end{equation}
where $\left\| \cdot \right\|$ and ${\left\| \cdot \right\|_2}$ denotes the ${\ell _1}$-norm and the ${\ell _2}$-norm respectively. ${{t_{gt}}}$ and ${{q_{gt}}}$ are the ground-truth translation vector and quaternion respectively generated by the ground-truth pose transformation matrix. Then, a multi-scale supervised approach is adopted. The total training loss $\ell$ is:\vspace{-0.2cm}
\begin{equation}\ell = \sum\nolimits_{l = 1}^L {{\alpha ^l}} {\ell ^l}\label{eqn:l},\end{equation}
where $L$ is the total number of warp-refinement levels and ${{\alpha ^l}}$ denotes the weight of the $l$-th level.
\setlength{\tabcolsep}{0.9mm}
\begin{table*}[t]
\centering
\footnotesize
\begin{center}
\resizebox{1.0\textwidth}{!}
{
\begin{tabular}{l||cc|cc|cc|cc|cc|cc|cc|cc|cc|cc|cc||cc}
\toprule
& \multicolumn{2}{c|}{00$^*$} &\multicolumn{2}{c|}{01$^*$} & \multicolumn{2}{c|}{02$^*$} & \multicolumn{2}{c|}{03$^*$} & \multicolumn{2}{c|}{04$^*$} & \multicolumn{2}{c|}{05$^*$} & \multicolumn{2}{c|}{06$^*$} & \multicolumn{2}{c|}{07} & \multicolumn{2}{c|}{08} & \multicolumn{2}{c|}{09} &\multicolumn{2}{c||}{10} &\multicolumn{2}{c}{Mean on 07-10} \\
\cline{2-25}\noalign{\smallskip}
\multirow{-2}{*}{\begin{tabular}[c]{@{}c@{}}Method \end{tabular}}
& $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ \\
\hline\hline
\noalign{\smallskip}
Full LOAM \cite{zhang2017low}
&1.10 & 0.53
&2.79 & 0.55
&1.54 & 0.55
&1.13 & 0.65
&1.45 & 0.50
&0.75 & 0.38
&0.72 & 0.39
&0.69 & 0.50
&1.18 & 0.44
&1.20 & 0.48
&1.51 & \bf0.57
& 1.145 & 0.498
\\
\hline
ICP-po2po
&6.88& 2.99
&11.21& 2.58
& 8.21 &3.39
&11.07& 5.05
& 6.64 &4.02
& 3.97& 1.93
&1.95 &1.59
&5.17 &3.35
&10.04 &4.93
&6.93 &2.89
&8.91 &4.74
&7.763 &3.978
\\
ICP-po2pl
&3.80 & 1.73
&13.53 & 2.58
&9.00 & 2.74
&2.72 & 1.63
&2.96 & 2.58
&2.29 & 1.08
&1.77 & 1.00
&1.55 & 1.42
&4.42 & 2.14
&3.95 & 1.71
&6.13 & 2.60
&4.013 & 1.968
\\
GICP \cite{segal2009generalized}
&1.29 & 0.64
&4.39 & 0.91
&2.53 & 0.77
&1.68 & 1.08
&3.76 & 1.07
&1.02 & 0.54
&0.92 & 0.46
&0.64 & 0.45
&1.58 & 0.75
&1.97 & 0.77
&1.31 &0.62
&1.375 & 0.648
\\
CLS \cite{velas2016collar}
&2.11 & 0.95
&4.22 & 1.05
&2.29 & 0.86
&1.63 & 1.09
&1.59 & 0.71
&1.98 & 0.92
&0.92 & 0.46
&1.04 & 0.73
&2.14 & 1.05
&1.95 & 0.92
&3.46 & 1.28
&2.148 & 0.995
\\
Velas et al. \cite{velas2018cnn}
&3.02 & NA
&4.44 & NA
&3.42 & NA
&4.94 & NA
&1.77 & NA
&2.35 & NA
&1.88 & NA
&1.77 & NA
&2.89 & NA
&4.94 & NA
&3.27 & NA
&3.218 & NA
\\
LO-Net \cite{li2019net}
&1.47 & 0.72
&1.36 & 0.47
&1.52 & 0.71
&1.03 & 0.66
&0.51 & 0.65
&1.04 & 0.69
&0.71 & 0.50
&1.70 & 0.89
&2.12 & 0.77
&1.37 & 0.58
&1.80 & 0.93
&1.748 & 0.793
\\
DMLO \cite{li2020dmlo}
& NA & NA
& NA & NA
& NA & NA
& NA & NA
& NA & NA
& NA & NA
& NA & NA
&0.73 & 0.48
&\bf 1.08 & \bf 0.42
&1.10 & 0.61
&\bf 1.12 & 0.64
&\bf 1.008 & 0.538
\\
LOAM w/o mapping
&15.99 &6.25
& 3.43 &0.93
& 9.40 & 3.68
& 18.18 &9.91
& 9.59 & 4.57
& 9.16 & 4.10
& 8.91 & 4.63
& 10.87 & 6.76
& 12.72 & 5.77
& 8.10 & 4.30
& 12.67 & 8.79
& 11.090 & 6.405
\\
Ours
&\bf0.78 & \bf0.42
&\bf0.67 & \bf0.23
&\bf0.86 & \bf0.41
&\bf0.76 & \bf0.44
&\bf0.37 & \bf0.40
&\bf0.45 & \bf0.27
&\bf0.27 & \bf0.22
&\bf0.60 & \bf0.44
&1.26 & 0.55
&\bf0.79 & \bf0.35
&1.69 & 0.62
&1.085 &\bf 0.490
\\ \bottomrule
\end{tabular}
}
\end{center}
\vspace{-8pt}
\caption{The LiDAR odometry experiment results on KITTI odometry dataset \cite{geiger2013vision}. $t_{rel}$ and $r_{rel}$ mean the average translational RMSE (\%) and rotational RMSE ($^{\circ}$/100m) respectively on all possible subsequences in the length of $100,200,...,800m$. `$^*$' means the training sequence. LOAM is a complete SLAM system, including back-end optimization and others only include odometry. The data other than the last three lines are from \cite{li2019net}. The results of LOAM w/o mapping are obtained by running their published codes. The best results are bold.}
\vspace{-8pt}
\label{table:lidar}
\end{table*}
\setlength{\tabcolsep}{0.9mm}
\begin{table*}[t]
\centering
\footnotesize
\begin{center}
\resizebox{0.9\textwidth}{!}
{
\begin{tabular}{l||cc|cc|cc|cc|cc|cc|cc|cc|cc|cc|cc||cc}
\toprule
& \multicolumn{2}{c|}{00$^*$} &\multicolumn{2}{c|}{01$^*$} & \multicolumn{2}{c|}{02$^*$} & \multicolumn{2}{c|}{03$^*$} & \multicolumn{2}{c|}{04$^*$} & \multicolumn{2}{c|}{05$^*$} & \multicolumn{2}{c|}{06$^*$} & \multicolumn{2}{c|}{07} & \multicolumn{2}{c|}{08} & \multicolumn{2}{c|}{09$^*$} &\multicolumn{2}{c||}{10$^*$} &\multicolumn{2}{c}{Mean on 07-08} \\
\cline{2-25}\noalign{\smallskip}
\multirow{-2}{*}{\begin{tabular}[c]{@{}c@{}}Method \end{tabular}}
& $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ \\
\hline\hline
\noalign{\smallskip}
LodoNet \cite{zheng2020lodonet}
&1.43 &0.69
&0.96 &0.28
&1.46 &0.57
&2.12 &0.98
&0.65 &0.45
&1.07 &0.59
&0.62 &0.34
&1.86 &1.64
&2.04 &0.97
&0.63 &0.35
&1.18 &0.45
&1.950 &1.305
\\
Ours
&\bf0.75 & \bf0.36
&\bf0.57 & \bf0.21
&\bf0.83 & \bf0.37
&\bf0.90 & \bf0.38
&\bf0.45 & \bf0.39
&\bf0.53 & \bf0.31
&\bf0.37 & \bf0.22
&\bf0.61 & \bf0.43
&\bf1.29 & \bf0.57
&\bf0.55 & \bf0.24
&\bf0.61 &\bf0.39
&\bf0.950 &\bf0.500
\\ \bottomrule
\end{tabular}
}
\end{center}
\vspace{-8pt}
\caption{The LiDAR odometry experiment results on KITTI odometry dataset \cite{geiger2013vision} compared with \cite{zheng2020lodonet}. As \cite{zheng2020lodonet} is trained on sequences 00-06, 09-10 and tested on sequences 07-08, we train and test our model like this to make comparisons with \cite{zheng2020lodonet}.}
\vspace{-8pt}
\label{table:lidar_7_8}
\end{table*}
\setlength{\tabcolsep}{1.3mm}
\begin{table}[t]
\centering
\footnotesize
\begin{center}
\resizebox{0.8\columnwidth}{!}
{
\begin{tabular}{l||cc|cc||cc}
\toprule
& \multicolumn{2}{c|}{04} &\multicolumn{2}{c||}{10} &\multicolumn{2}{c}{Mean} \\
\cline{2-7}\noalign{\smallskip}
\multirow{-2}{*}{\begin{tabular}[c]{@{}c@{}}Method \end{tabular}}
& $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ \\
\hline\hline
\noalign{\smallskip}
DeepPCO \cite{wang2019deeppco}
&2.63 & 3.05
&2.47 & 6.59
&2.550 & 4.820
\\
Ours
&\bf0.73 & \bf0.43
&\bf1.57 & \bf0.57
&\bf1.150 & \bf 0.500
\\ \bottomrule
\end{tabular}
}
\end{center}
\vspace{-8pt}
\caption{The LiDAR odometry results on sequences 04 and 10 of KITTI odometry dataset \cite{geiger2013vision}. As \cite{wang2019deeppco} is trained on sequences 00-03, 05-09 and only reports testing results on sequences 04 and 10, we train and test our model like this to make comparisons with it.}
\vspace{-8pt}
\label{table:lidar4_10}
\vspace{0.2cm}
\end{table}
\vspace{-0.2cm}
\setlength{\tabcolsep}{1.3mm}
\begin{table}[t]
\centering
\footnotesize
\begin{center}
\resizebox{0.8\columnwidth}{!}
{
\begin{tabular}{l||cc|cc||cc}
\toprule
& \multicolumn{2}{c|}{09} &\multicolumn{2}{c||}{10} &\multicolumn{2}{c}{Mean} \\
\cline{2-7}\noalign{\smallskip}
\multirow{-2}{*}{\begin{tabular}[c]{@{}c@{}}Method \end{tabular}}
& $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ \\
\hline\hline
\noalign{\smallskip}
Cho et al. \cite{cho2020}
&4.87 & 1.95
&5.02 & 1.83
&4.945 & 1.890
\\
Ours
&\bf0.73 & \bf0.40
&\bf1.16 & \bf0.78
&\bf0.945 & \bf 0.295
\\ \bottomrule
\end{tabular}
}
\end{center}
\vspace{-8pt}
\caption{The LiDAR odometry results on sequences 09 and 10 of KITTI odometry dataset \cite{geiger2013vision}. As \cite{cho2020} applies unsupervised training on sequences 00-08, and reports testing results on sequences 09 and 10, we train and test our model like this to make fair comparisons.}
\vspace{-6pt}
\label{table:lidar09_10}
\vspace{-0.2cm}
\end{table}
\section{Implementation}\label{sec:implementation}
\vspace{-4pt}
\subsection{KITTI Odometry Dataset}
\vspace{-4pt}
KITTI odometry dataset \cite{geiger2012we,geiger2013vision} is composed of 22 independent sequences. The Velodyne LiDAR point clouds in the dataset are used in our experiments. All scans of point clouds have XYZ coordinates and the reflectance information. Sequences 00-10 (23201 scans) contain ground truth pose (trajectory), while there is no ground truth publicly available for the remaining sequences 11-21 (20351 scans). By driving under different road environments, such as highways, residential roads, and campus roads, the sampling vehicle captures point clouds for the LiDAR odometry task from different environments.
\vspace{5pt}
\noindent{}{\bf Data Preprocessing:}
Only coordinates of LiDAR points are used in our method. Because the ground truth poses are denoted in the left camera coordinate system, all the training and evaluation processes of this network are carried out in the left camera coordinate system. Thus, the captured point clouds from the Velodyne LiDAR are firstly transformed to the left camera coordinate system by:
\vspace{-0.2cm}
\begin{equation}
{P_{cam}} = {T_r}P_{vel},
\vspace{-0.2cm}\end{equation}
where ${P_{cam}}$ and ${P_{vel}}$ are the point cloud coordinates in the left camera coordinate system and LiDAR coordinate system respectively, and ${T_r}$ is the calibration matrix of each sequence.
Moreover, the point cloud collected by the LiDAR sensor often contains outliers at the edge of the point cloud in each frame. This is often because objects are far away from the LiDAR sensor, thus forming incomplete point clouds in the edge. To filter out these outlier points, the points out of the $30 \times 30m^2$ square area around the vehicle are filtered out for each point cloud $P_{cam}$. To speed up the data reading and training, the ground less than $0.55m$ in height is removed. For our model, the performances of removing and reserving the ground are similar. The detailed comparison is shown in the supplementary material.
\vspace{4pt}
\noindent{}{\bf Data Augmentation:}
We augment the training dataset by the augmentation matrix ${T_{aug}}$, generated by the rotation matrix ${R_{aug}}$ and translation vector ${t_{aug}}$.
Varied values of yaw-pitch-roll Euler angles are generated by Gaussian distribution around $0^{\circ}$. Then the ${R_{aug}}$ can be obtained from these random Euler angles. Similarly, the ${t_{aug}}$ is generated by the same process.
The composed ${T_{aug}}$ is then used to augment the $PC_1$ to obtain new point clouds $PC_{1,aug}$ by: \vspace{-0.2cm}
\begin{equation} \vspace{-0.2cm}
{PC_{1,aug}} = {T_{aug}}{PC_1}.\end{equation}
Correspondingly, the ground truth transformation matrix is also modified as:\vspace{-0.2cm}
\begin{equation}\vspace{-0.2cm}
{T_{trans}} = {T_{aug}}{T_p},\end{equation}
where ${T_p}$ denotes the original ground truth pose transformation matrix from $PC_1$ to $PC_2$.
${T_{trans}}$ is then used to generate ${q_{gt}}$ and ${t_{gt}}$ to supervise the training of the network.
\subsection{Network Details}
\vspace{-4pt}
In the training and evaluation process, the input $N$ points are randomly sampled from the point clouds of two frames separately. It is unnecessary that the original input two point clouds have the same number of points. And $N$ is set to be 8192 in the proposed network. Each layer in MLP contains the ReLU activation function, except for the FC layer. For shared MLP, $1 \times 1$ convolution with $1$ stride is the implement manner. The detailed layer parameters including each linear layer width in MLP are described in the supplementary material. All training and evaluation experiments are conducted on a single NVIDIA RTX 2080Ti GPU with TensorFlow 1.9.0. The Adam optimizer is adopted with ${\beta _1} = 0.9$, ${\beta _2} = 0.999$. The initial learning rate is 0.001 and exponentially decays every 200000 steps until 0.00001. The initial values of the trainable parameters $s_x$ and $s_q$ are set as 0.0 and -2.5 respectively in formula~(\ref{eqn:sxsq}). For formula~(\ref{eqn:l}), ${\alpha _1}$ = 1.6, ${\alpha _2}$ = 0.8, ${\alpha _3}$ = 0.4, and $L=4$. The batch size is 8.
\vspace{-0.1cm}
\section{Experimental Results}
\vspace{-4pt}
In this section, the quantitative and qualitative results of the network performance on the LiDAR odometry task are demonstrated and compared with state-of-the-art methods. Then, the ablation studies are presented. Finally, the embedding mask is visualized and discussed.
\subsection{Performance Evaluation}
\vspace{-4pt}
As there are several modes to divide the training/testing sets for published papers \cite{li2019net,zheng2020lodonet,wang2019deeppco,zheng2020lodonet,cho2020}, in order to fairly compare with all current methods as we know, we train/test our model four times.
\noindent{}{\bf Using sequences 00-06/07-10 as training/testing sets:} Quantitative results are listed in Table~\ref{table:lidar}. ICP-point2point (ICP-po2po), ICP-point2plane (ICP-po2pl), GICP \cite{segal2009generalized} , CLS \cite{velas2016collar} are several classic LiDAR odometry estimation methods based on Iterative Closest Point (ICP) \cite{besl1992method}. LOAM \cite{zhang2017low} is based on the matching of extracted line and plane features, which has a similar idea with our method, and it is a hand-crafted method. It achieves the best results among LiDAR-based methods in the KITTI odometry evaluation benchmark \cite{geiger2013vision}. Velas et al. \cite{velas2018cnn} is a learning-based method. It has a good performance when only the translation is estimated, but the performance will decrease when estimating the 6-DOF pose. LO-Net \cite{li2019net} and DMLO \cite{li2020dmlo} are learning-based LiDAR odometry methods that have comparable results, but they have no codes publicly available, so we adopt the same training and testing sequences with \cite{li2019net,li2020dmlo}. Compared with \cite{li2020dmlo}, our method utilizes soft correspondence rather than exact matching pairs so as to relize end-to-end pose estimation, and we achieved similar results on $t_{rel}$ and better results on $r_{rel}$. Compared with \cite{li2019net}, our method does not need an extra mask network and can filter the outliers with the proposed hierarchical embedding mask. Moreover, our method does not need to project 3D point clouds to 2D \cite{li2019net,li2020dmlo} and obtains the LiDAR odometry directly from the raw 3D point clouds. We achieved better results than LO-Net~\cite{li2019net}, even than LOAM~\cite{zhang2017low} on most sequences. We believe the pose refinement target makes our internal trainable mask effective for various outliers other than only dynamic regions~\cite{li2019net} as shown in Fig.~\ref{fig:visual}. Moreover, the trainable iterative refinement makes the estimated pose refined multiple times in one network inference.
\vspace{0.1cm}
\setlength{\tabcolsep}{0.9mm}
\begin{table*}[t]
\footnotesize
\begin{center}
\resizebox{1.0\textwidth}{!}
{
\begin{tabular}{l|l||cc|cc|cc|cc|cc|cc|cc|cc|cc|cc|cc||cc}
\toprule
& & \multicolumn{2}{c|}{00$^*$} &\multicolumn{2}{c|}{01$^*$} & \multicolumn{2}{c|}{02$^*$} & \multicolumn{2}{c|}{03$^*$} & \multicolumn{2}{c|}{04$^*$} & \multicolumn{2}{c|}{05$^*$} & \multicolumn{2}{c|}{06$^*$} & \multicolumn{2}{c|}{07} & \multicolumn{2}{c|}{08} & \multicolumn{2}{c|}{09} &\multicolumn{2}{c||}{10} &\multicolumn{2}{c}{Mean on 07-10} \\
\cline{3-26}\noalign{\smallskip}
& \multirow{-2}{*}{\begin{tabular}[c]{@{}c@{}}Method \end{tabular}}
& $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ \\
\hline\hline
\noalign{\smallskip}
(a)
&Ours w/o mask
&0.99 & 0.58
&0.61 & 0.25
& 1.07 & 0.55
& 0.84 & 0.66
& 0.42 & 0.58
& 0.68 & 0.41
& 0.33 & 0.25
& 0.76 & 0.58
& 1.54 & 0.66
& 1.30 & 0.61
& 2.36 & 0.86
& 1.490 & 0.678
\\
& Ours w/o mask optimazation
& 0.91& 0.54
&\bf0.59 & 0.26
& 1.10 & 0.49
& 0.81 & \bf0.43
& 0.41 & 1.05
&0.57 & 0.39
& 0.34 & \bf0.22
& 0.67 & 0.54
& 1.41 & 0.59
& \bf0.78 & 0.39
& \bf1.32 & 0.66
& \bf1.045 & 0.545
\\
&Ours (full, with mask and mask optimazation)
&\bf0.78 & \bf0.42
&0.67 & \bf0.23
&\bf0.86 & \bf0.41
&\bf0.76 & 0.44
&\bf0.37 & \bf0.40
&\bf0.45 & \bf0.27
&\bf0.27 & \bf0.22
&\bf0.60 & \bf0.44
&\bf1.26 & \bf0.55
&0.79 & \bf0.35
&1.69 & \bf0.62
&1.085 &\bf 0.490
\\
\cline{1-26}\noalign{\smallskip}
(b) &Ours (w/o cost volume)
& 42.52 & 19.44
& 22.73 & 4.44
& 37.53 & 12.55
& 22.21 & 13.18
& 15.38 & 3.59
& 31.95 & 12.88
& 12.82 & 3.83
& 30.78 & 18.95
& 67.08 & 25.72
& 38.63 & 16.38
& 40.55 & 19.80
& 44.260 & 20.213
\\
&Ours (with the cost volume in \cite{liu2019flownet3d})
&1.11 & 0.51
&0.61 & \bf0.22
&0.97 & 0.47
& \bf0.68 & 0.61
& 0.50 & 0.67
& 0.51 & 0.30
& 0.38 & \bf 0.17
& 1.18 & 0.73
& 1.40 & 0.59
& 1.15 & 0.46
&\bf1.50 & 0.88
& 1.308 & 0.665
\\
&Ours (with the cost volume in \cite{wu2019pointpwc})
&\bf0.69 &\bf 0.35
&\bf0.59 & 0.23
& 0.87 & 0.42
& 0.79 & 0.55
&\bf 0.34 & \bf 0.34
& 0.49 & 0.28
& 0.28 & \bf0.17
& 0.62 & 0.47
& 1.47 & 0.62
& 0.97 & 0.48
&1.54 & \bf0.59
& 1.150 & 0.540
\\
&Ours (full, with the cost volume in \cite{wang2020hierarchical})
&0.78 & 0.42
&0.67 & 0.23
&\bf0.86 & \bf0.41
&0.76 & \bf0.44
&0.37 & 0.40
&\bf0.45 & \bf0.27
&\bf0.27 & 0.22
&\bf0.60 & \bf0.44
&\bf1.26 & \bf 0.55
&\bf0.79 & \bf0.35
&1.69 & 0.62
&\bf1.085 &\bf 0.490
\\
\cline{1-26}\noalign{\smallskip}
(c)
&Ours w/o Pose Warp-Refinement
&15.71 & 8.26
&13.92 & 5.62
& 13.38 & 7.07
& 15.74 & 16.62
& 5.00 & 6.89
& 14.19 & 7.87
& 6.38 & 5.33
& 12.52 & 11.73
& 15.64 & 9.11
& 10.17 & 7.60
& 16.09 & 11.29
& 13.605 & 9.933
\\
&Ours w/o Pose Warp
&3.07 & 1.40
&4.04 & 1.56
&3.27 & 1.27
&2.56 & 2.01
&0.91 & 1.13
&2.28 & 1.15
&2.08 & 1.18
&4.29 & 2.75
&5.00 & 2.14
&4.48 & 2.19
&5.39 & 3.36
&4.790 & 2.610
\\
&Ours (full, with Pose Warp-Refinement)
&\bf0.78 & \bf0.42
&\bf0.67 & \bf0.23
&\bf0.86 & \bf0.41
&\bf0.76 & \bf0.44
&\bf0.37 & \bf0.40
&\bf0.45 & \bf0.27
&\bf0.27 & \bf0.22
&\bf0.60 & \bf0.44
&\bf1.26 & \bf 0.55
&\bf0.79 & \bf0.35
&\bf1.69 & \bf0.62
&\bf1.085 &\bf 0.490
\\
\cline{1-26}\noalign{\smallskip}
(d)
&Ours (first embedding on the last level)
& 0.95 &0.46
&\bf0.67 & 0.29
& 1.03 & 0.44
& 0.96 & 0.77
& 0.48 & \bf0.40
& 0.74 & 0.40
& 0.41 & 0.26
& 0.94 & 0.49
& 1.41 & 0.56
& 1.01 & 0.47
&\bf1.67 & 0.82
& 1.258 & 0.585
\\
&Ours (full, first embedding on the penultimate level)
&\bf0.78 & \bf0.42
&\bf0.67 & \bf0.23
&\bf0.86 & \bf0.41
&\bf0.76 & \bf0.44
&\bf0.37 & \bf0.40
&\bf0.45 & \bf0.27
&\bf0.27 & \bf0.22
&\bf0.60 & \bf0.44
&\bf1.26 & \bf0.55
&\bf0.79 & \bf0.35
&1.69 &\bf 0.62
&\bf1.085 &\bf 0.490
\\
\bottomrule
\end{tabular}
}
\end{center}
\vspace{-8pt}
\caption{The ablation study results of LiDAR odometry on KITTI odometry dataset \cite{geiger2013vision}.}
\vspace{-9pt}
\label{table:ablation}
\end{table*}
\begin{figure}[t]
\begin{center}
\resizebox{0.95\columnwidth}{!}
{
\includegraphics[scale=1.00]{0106path2d-eps-converted-to.pdf}}
\end{center}
\vspace{-8pt}
\caption{Trajectory results of LOAM and ours on KITTI training sequences with ground truth. Ours is much better than the LOAM without mapping. And ours also has better performance on the two sequences than full LOAM, though ours is for odometry and LOAM has the mapping optimization.}
\label{fig:odometry_path2d}
\vspace{-8pt}
\end{figure}
\begin{figure}[t]
\begin{center}
\resizebox{0.95\columnwidth}{!}
{
\includegraphics[scale=1.00]{07path2d3d-eps-converted-to.pdf}}
\end{center}
\vspace{-8pt}
\caption{3D and 2D trajectory results on KITTI validation sequence 7 with ground truth. Our method obtained the most accurate trajectory.}
\label{fig:odometry_path2d3d}
\vspace{-9pt}
\end{figure}
\begin{figure}[t]
\begin{center}
\resizebox{0.95\columnwidth}{!}
{
\includegraphics[scale=1.00]{error_line-eps-converted-to.pdf}}
\end{center}
\vspace{-8pt}
\caption{Average translational and rotational error on KITTI sequences 00-10 on all possible subsequences in the length of $100,200,...,800m$. Our method has the best performance.}
\label{fig:odometry_error_line}
\vspace{-8pt}
\end{figure}
\vspace{-0.05cm}
\noindent{}{\bf Using other sequences as training/testing sets: }
\noindent{}{\bf 00-06, 09-10/07-08:} Quantitative results are listed in Table~\ref{table:lidar_7_8} to compare with a recent method \cite{zheng2020lodonet}.
\noindent{}{\bf 00-03, 05-09/04, 10:} As shown in Table~\ref{table:lidar4_10}, we adopted the same training/testing sets of the KITTI odometry dataset \cite{geiger2013vision} and compared with \cite{wang2019deeppco}.
\noindent{}{\bf 00-08/09-10:} \cite{cho2020} proposes an unsupervised method on LiDAR odometry. Table~\ref{table:lidar09_10} shows the quantitative comparison of this method with ours.
The results in Table~\ref{table:lidar_7_8}, Table~\ref{table:lidar4_10}, and Table~\ref{table:lidar09_10} show that our method outperforms them. \cite{wang2019deeppco} is based on 2D convolutional network, which loses the raw 3D information. \cite{zheng2020lodonet} finds matched keypoint pairs in 2D depth images to regress pose. \cite{cho2020} is a method of unsupervised training. The results demonstrate the superiority of our full 3D learning based LiDAR odometry.
The qualitative results are shown in Figs.~\ref{fig:odometry_path2d},~\ref{fig:odometry_path2d3d}, and~\ref{fig:odometry_error_line}. We compared our method with LOAM \cite{zhang2017low} and LOAM without mapping since \cite{li2019net,zheng2020lodonet,wang2019deeppco,li2020dmlo} do not release their codes. Full LOAM has a superb performance but degrades significantly without mapping. Ours is only for odometry and is better than the LOAM without mapping. At the same time, ours is even better than LOAM \cite{zhang2017low} on average evaluation.
\vspace{-0.15cm}
\subsection{Ablation Study}
\vspace{-4pt}
In order to analyze the effectiveness of each module, we remove or change components of our model to do the ablation studies on KITTI odometry dataset. The training/testing details are the same as described in Sec.~\ref{sec:implementation}.
\vspace{5pt}
\noindent{}{\bf Benefits of Embedding Mask:}
We first remove the optimization of the mask, which means that the embedding masks are independently estimated at each level. Then we entirely remove the mask and apply the average pooling to embedding features to calculate the pose. The results in Table~\ref{table:ablation}(a) show that the proposed embedding mask and its hierarchical optimization both contribute to better results.
The mismatched objects are different in various scenes. In some scenes, the car is dynamic, while the car is static in another scene (Fig.~\ref{fig:visual}). Through the learnable embedding mask, the network learns to filter outlier points from the overall motion pattern. The network achieves the effect of RANSAC~\cite{fischler1981random} by only once network inference.
\vspace{3pt}
\noindent{}{\bf Different Cost Volume:}
As there are different point feature embedding methods in point clouds, we compare three recent methods in our odometry task, including the flow embedding layer in FlowNet3D \cite{liu2019flownet3d}, point cost volume~\cite{wu2019pointpwc} and the attentive cost volume~\cite{wang2020hierarchical}. The results in Table~\ref{table:ablation}(b) show that the model with double attentive cost volume has the best results in the three. Therefore, robust point association also contributes to the LiDAR odometry task.
\vspace{5pt}
\noindent{}{\bf Effect of Pose Warp-Refinement:}
We first remove the pose warping and reserve the hierarchical refinement of the embedding feature and the mask. The results show a decline as in Table~\ref{table:ablation}(c). Then we remove the full pose warp-refinement module, which means that the embedding features, mask, and pose are only calculated once. As a result, the performance degrades a lot, which demonstrates the importance of the coarse-to-fine refinement.
\vspace{5pt}
\noindent{}{\bf Suitable Level for the First Embedding:}
As fewer points have less structural information, it is needed to decide which level is used to firstly correlate the two point clouds and generate the embedding features. We test the different levels of the first feature embedding. The results in Table~\ref{table:ablation}(d) demonstrate that the most suitable level for the first embedding is the penultimate level in the point feature pyramid, which shows that the 3D structure is also important for the coarse pose regression.
\begin{figure}[t]
\centering
\resizebox{0.95\columnwidth}{!}
{
\includegraphics[scale=1.00]{fig_mask_visual-eps-converted-to.pdf}}
\vspace{-0.1mm}
\caption{The Visualization of embedding mask. In the LiDAR point clouds, the small red points are the whole point cloud of the first frame, and the big points with distinctive colors are the sampled 2048 points with contributions to the pose transformation. In the example on the left, the buildings and steel bars around the highway have high weights, while the bushes and the moving car have low weights. In the example on the right, the buildings, the tree trunks, and the static car have high weights, while the weeds and the cyclist have low weights.}
\label{fig:visual}
\vspace{-0.4cm}
\end{figure}
\vspace{-0.2cm}
\subsection{Visualization of Embedding Mask}
\vspace{-4pt}
The proposed network uses embedding features of 2048 points to calculate the pose transformation in the final pose output layer. We visualize the embedding mask in the 2048 points to show the contribution of each point to the final pose transformation. As illustrated in Fig.~\ref{fig:visual}, the points that are sampled from the static and structured rigid objects such as buildings, steel bars, and the static car have higher weights. On the contrary, the points from the dynamic and irregular objects such as the moving car, the cyclist, bushes, and weeds have lower weights. In LO-Net~\cite{li2019net}, only the dynamic regions are masked, while our method also gives small weight to unreliable weeds and bushes. Therefore, the embedding mask can effectively reduce the influence of dynamic objects and other outliers on the estimation of pose transformation from adjacent frames. \vspace{-0.2cm}
\section{Conclusion}
\vspace{-4pt}
Different from the 2D projection-based methods \cite{nicolai2016deep,velas2018cnn,wang2019deeppco,li2019net}, a full 3D learning method based on PWC structure for LiDAR odometry is proposed in this paper. The pose warp-refinement module is proposed to realize the hierarchical refinement of 3D LiDAR odometry in a coarse-to-fine approach. The hierarchical embedding mask optimization is proposed to deal with various outlier points. And our method achieved a new state-of-the-art end-to-end LiDAR odometry method with no need for an extra mask network \cite{li2019net}. Since our mask can filter a variety of outlier points that are not suitable for calculating overall motion, it is a direction worth exploring to combine our new mask with mapping optimization in the future.
\vspace{4pt}
{\bf \small Acknowledgement.} {\small This work was supported in part by the Natural Science Foundation of China under Grant U1613218, U1913204, 62073222, in part by "Shu Guang" project supported by Shanghai Municipal Education Commission and Shanghai Education Development Foundation under Grant 19SG08, in part by grants from NVIDIA Corporation.}
{\small
\bibliographystyle{ieee_fullname}
\section{Introduction}
\vspace{-4pt}
The visual/LiDAR odometry is one of the key technologies in autonomous driving. This task uses two consecutive images or point clouds to obtain the relative pose transformation between two frames, and acts as the base of the subsequential planning and decision making of mobile robots \cite{robot_}. Recently, learning-based odometry methods have shown impressive accuracy on datasets compared with conventional methods based on hand-crafted features. It is found that learning-based methods can deal with sparse features and dynamic environments \cite{dasgil,lpdnet}, which are usually difficult for conventional methods. To our knowledge, most learning-based methods are on the 2D visual odometry \cite{wang2017deepvo,zhou2017unsupervised,min2020voldor,yang2020d3vo,Unsupervised_depth,chizhang} or utilize 2D projection of LiDAR \cite{nicolai2016deep,velas2018cnn,wang2019deeppco,li2019net,li2020dmlo} and the LiDAR odometry in 3D point clouds is underexplored. This paper aims to estimate the LiDAR odometry directly through raw 3D point clouds.
\begin{figure}[t]
\centering
\resizebox{0.90\columnwidth}{!}
{
\includegraphics[scale=1.0]{fignetwork_introduction_1-eps-converted-to.pdf}}
\vspace{0mm}
\caption{The Point Feature Pyramid, Pose Warping, and Attentive Cost Volume (PWC) structure in our proposed PWCLO-Net. The pose is refined layer by layer through iterative pose warp-refinement. The whole process is realized end-to-end by making all modules differentiable. In the LiDAR point clouds, the small red points are the whole point cloud. The big black points are the sampled points in $PC_1$. Distinctive colors of big points in embedding mask measure the contribution of sampled points to the pose estimation.}
\label{fig:network_little}
\vspace{-12pt}
\end{figure}
For the LiDAR odometry in 3D point clouds, there are three challenges: 1) As the discrete LiDAR point data are obtained in two consecutive frames separately, it is hard to find an precisely corresponding point pair between two frames; 2) Some points belonging to an object in a frame may not be seen in the other view if they are occluded by other objects or are not captured because of the limitation of LiDAR resolution; 3) Some points belonging to dynamic objects are not suitable to be used for the pose estimation because these points have uncertainty motions of the dynamic objects.
For the first challenge, Zheng et al. \cite{zheng2020lodonet} use matched keypoint pairs judged in 2D depth images. However, the correspondence is rough because of the discrete perception of LiDAR. The cost volume for 3D point cloud \cite{wu2019pointpwc,wang2020hierarchical} is adopted in this paper to obtain a weighted soft correspondence between two consecutive frames. For the second and third challenges, the mismatched points or dynamic objects, which do not conform to the overall pose, need to be filtered. LO-Net~\cite{li2019net} trains an extra mask estimation network~\cite{zhou2017unsupervised, yang2018unsupervised} by weighting the consistency error of the normal of 3D points. In our network, an internal trainable embedding mask is proposed to weigh local motion patterns from the cost volume to regress the overall pose. In this way, the mask can be optimized for more accurate pose estimation rather than depending on geometry correspondence. In addition, the PWC structure is built to capture the large motion in the layer of sparse points and refine the estimated pose in dense layers. As shown in Fig.~\ref{fig:network_little}, the embedding mask is also optimized hierarchically to obtain more accurate filtering information to refine pose estimation.
Overall, our contributions are as follows:
\vspace{-0.1cm}
\begin{itemize}
\vspace{-0.2cm}
\item The Point Feature Pyramid, Pose Warping, and Cost Volume (PWC) structure for the 3D LiDAR odometry task is built to capture the large motion between two frames and accomplish the trainable iterative 3D feature matching and pose regression.
\vspace{-0.3cm}
\item In this structure, the hierarchical embedding mask is proposed to filter mismatched points and convert the cost volume embedded in points to the overall ego-motion in each refinement level. Meanwhile, the embedding mask is optimized and refined hierarchically to obtain more accurate filtering information for pose refinement following the density of points.
\vspace{-0.3cm}
\item Based on the characteristic of the pose transformation,
the pose warping and pose refinement are proposed to refine the estimated pose layer by layer iteratively. A totally end-to-end framework, named PWCLO-Net, is established, in which all modules are fully differentiable so that each process is no longer independent and combinedly optimized.
\vspace{-0.3cm}
\item Finally, our method is demonstrated on KITTI odometry dataset \cite{geiger2012we,geiger2013vision}. The evaluation experiments and ablation studies show the superiority of the proposed method and the effectiveness of each design. To the best of our knowledge, our method outperforms all recent learning-based LiDAR odometry and even outperforms the geometry-based LOAM with mapping optimization \cite{zhang2017low} on most sequences.
\vspace{-0.2cm}
\end{itemize}
\section{Related Work}
\vspace{-4pt}
\subsection{Deep LiDAR Odometry}
\vspace{-4pt}
Deep learning has gained impressive progress in visual odometry \cite{min2020voldor,yang2020d3vo}. However, the 3D LiDAR odometry with deep learning is still a challenging problem. In the beginning, Nicolai et al. \cite{nicolai2016deep} project two consecutive LiDAR point clouds to the 2D plane to obtain two 2D depth images, and then use the 2D convolution and fully connected (FC) layers to realize learning-based LiDAR odometry. Their work verified that the learning-based method is serviceable for the LiDAR odometry although their experiment results are not superior. Velas et al. \cite{velas2018cnn} also project LiDAR points to the 2D plane but use three channels to encode the information, including height, range, and intensity. Then the convolution and FC layers are used for the pose regression. The performance is superb when only estimating the translation but is poor when estimating the 6-DOF pose. Wang et al. \cite{wang2019deeppco} project point clouds to panoramic depth images and stack two frames together as input. Then the translation sub-network and FlowNet \cite{dosovitskiy2015flownet} orientation sub-network are used to estimate the translation and orientation respectively. \cite{li2019net} also preprocess 3D LiDAR points to 2D information but use the cylindrical projection \cite{chen2017multi}. Then, the normal of each 3D point is estimated to build consistency constraint between frames, and an uncertainty mask is estimated to mask dynamic regions. Zheng et al. \cite{zheng2020lodonet} extract matched keypoint pairs by classic detecting and matching algorithm in 2D spherical depth images projected from 3D LiDAR point clouds. Then the PointNet \cite{qi2017pointnet} based structure is used to regress the pose from the matched keypoint pairs. \cite{li2020dmlo} proposes a learning-based network to generate matching pairs with high confidence, then applies Singular Value Decomposition (SVD) to get the 6-DoF pose. \cite{cho2020} introduces an unsupervised learning method on LiDAR odometry.
\begin{figure*}[t]
\centering
\resizebox{1.00\textwidth}{!}
{
\includegraphics[scale=1.00]{fignetwork-eps-converted-to.pdf}}
\vspace{-7mm}
\caption{The details of proposed PWCLO-Net architecture. The network is composed of four set conv layers in point feature pyramid, one attentive cost volume, one initial embedding mask and pose generation module, and three pose warp-refinement modules. The network outputs the predicted poses from four levels for supervised training. }
\label{fig:network}
\vspace{-12pt}
\end{figure*}
\vspace{-0.1cm}
\subsection{Deep Point Correlation}
\vspace{-4pt}
The above studies all use the 2D projection information for the LiDAR odometry learning, which converts the LiDAR odometry to the 2D learning problem. Wang et al. \cite{wang2019deeppco} compared the 3D point input and 2D projection information input based on the same 2D convolution model. It is found that the 3D input based method has a poor performance. As the development of 3D deep learning \cite{qi2017pointnet,qi2017pointnet++}, FlowNet3D \cite{liu2019flownet3d} proposes an embedding layer to learn the correlation between the points in two consecutive frames. After that, Wu et al. \cite{wu2019pointpwc} propose the cost volume method on point clouds, and Wang et al. \cite{wang2020hierarchical} develop the attentive cost volume method. The point cost volume involves the motion patterns of each point. It becomes a new direction and challenge to regress pose from the cost volume, and meanwhile, not all point motions are for the overall pose motion. We exploit to estimate the pose directly from raw 3D point cloud data and deal with the new challenges encountered.
What is more, we are inspired by the Pyramid, Warping, and Cost volume (PWC) structure in the flow network proposed by Su et al. \cite{sun2018pwc}. The work uses the three modules (Pyramid, Warping, and Cost volume) to refine the optical flow through a coarse-to-fine approach. The works \cite{wu2019pointpwc, wang2020hierarchical} on 3D scene flow also use the PWC structure to refine the estimated 3D scene flow in point clouds. In this paper, the idea is applied to the pose estimation refinement, and a PWC structure for LiDAR odometry is built for the first time.
\vspace{-0.2cm}
\section{PWCLO-Net}
\vspace{-4pt}
Our method learns the LiDAR odometry from raw 3D point clouds in an end-to-end approach with no need to pre-project the point cloud into 2D data, which is significantly different from the deep LiDAR odometry methods introduced in the related work section.
The Fig.~\ref{fig:network} illustrates the overall structure of our PWCLO-Net. The inputs to the network are two point clouds $PC_1 = \{ {x_i}|{x_i} \in {\mathbb{R}^3}\} _{i = 1}^N$ and $PC_2 = \{ {y_j}|{y_j} \in {\mathbb{R}^3}\} _{j = 1}^N$ respectively sampled from two adjacent frames. These two point clouds are firstly encoded by the siamese feature pyramid consisting of several set conv layers as introduced in Sec.~\ref{sec:pyramid}. Then the attentive cost volume is used to generate embedding features, which will be described in Sec.~\ref{sec:cost}. To regress pose transformation from the embedding features, hierarchical embedding mask optimization is proposed in Sec.~\ref{sec:mask}. Next, the pose warp-refinement method is proposed in Sec.~\ref{sec:refine} to refine the pose estimation in a coarse-to-fine approach. Finally, the network outputs the quaternion $q \in {\mathbb{R}^4}$ and translation vector $t \in {\mathbb{R}^3}$.
\subsection{Siamese Point Feature Pyramid}\label{sec:pyramid}
\vspace{-4pt}
The input point clouds are usually disorganized and sparse in a large 3D space. A siamese feature pyramid consisting of several set conv layers is built to encode and extract the hierarchical features of each point cloud. Farthest Point Sampling (FPS)~\cite{qi2017pointnet++} and shared Multi-Layer Perceptron (MLP) are used. The formula of set conv is:
{\setlength\abovedisplayskip{4pt}\setlength\belowdisplayskip{4pt}\begin{equation} {f_i} = \mathop {MAX}\limits_{k = 1,2,...K} (MLP((x_i^k - x_i) \oplus f_i^k \oplus f_i^c)),\end{equation}}where $x_i$ is the obtained $i$-th sampled point by FPS. And $K$ points $x_i^k$ $(k=1,2,...,K)$ are selected by $K$ Nearest Neighbors (KNN) around $x_i$. $f_i^c$ and $f_i^k$ are the local features of $x_i$ and $x_i^k$ (they are null for the first layer in the pyramid). ${f_i}$ is the output feature located at the central point $x_i$. $\oplus$ denotes the concatenation of two vectors, and $\mathop {MAX}\limits_{k = 1,2,...K}() $ indicates the max pooling operation. The hierarchical feature pyramid is constructed as shown in Fig.~\ref{fig:network}. The siamese pyramid \cite{chopra2005learning} means that the learning parameters of the built pyramid are shared for these two point clouds.
\subsection{Attentive Cost Volume}\label{sec:cost}
\vspace{-4pt}
Next, the point cost volume with attention in \cite{wang2020hierarchical} is adopted here to associate two point clouds. The cost volume generates point embedding features by associating two point clouds after the feature pyramid.
The embedding features contain the point correlation information between two point clouds. As shown in Fig.~\ref{fig:cost}, $F_1 = \{ {f_i}|{f_i} \in {\mathbb{R}^c}\} _{i = 1}^n$ is the features of point cloud $PC_1 = \{ {x_i}|{x_i} \in {\mathbb{R}^3}\} _{i = 1}^n$ and $F_2 = \{ {g_j}|{g_j} \in {\mathbb{R}^c}\} _{j = 1}^n$ is the features of point cloud $PC_2 = \{ {y_j}|{y_i} \in {\mathbb{R}^3}\} _{i = 1}^n$. The embedding features between two point clouds are calculated as follows:
\begin{equation}w_{1, i}^k = softmax(u(x_i,y_j^k,f_i,g_j^k))_{k = 1}^{K_1},\end{equation}\vspace{-0.5cm}
\begin{equation}pe_{i} = \sum\limits_{k = 1}^{k_1} {{w_{1,i}^k} \odot v(x_i,y_j^k,f_i,g_j^k)},\end{equation}\vspace{-0.2cm}
\begin{equation}w_{2,i}^k = softmax(u(x_i,x_i^k,pe_{i},pe_{i}^k))_{k = 1}^{K_2},\end{equation}\vspace{-0.2cm}
\begin{equation}e_{i} = \sum\limits_{k = 1}^{k_2} {w_{2,i}^k \odot v(x_i,x_i^k,pe_{i},pe_{i}^k)},\end{equation}
where $y_j^k$ and $g_j^k$ represent the coordinates and local features of the selected $K_1$ points in $PC_2$ respectively. $\odot$ represents dot product. $u(\cdot)$ and $v(\cdot )$ represent the attention encoding and feature encoding functions referring to \cite{wang2020hierarchical}. $u(\cdot)$ encodes the 3D Euclidean space information and point features to generate attention weights, and $v(\cdot)$ represents the further feature encoding on spacial information and features of two frames of point cloud. The output $E = \{ {e_{i}}|{e_{i}} \in {\mathbb{R}^c}\} _{i = 1}^n$ is the embeding features in $PC_1$.
\begin{figure}[t]
\centering
\resizebox{1.0\columnwidth}{!}
{
\includegraphics[scale=1.00]{figcost-half.pdf}}
\vspace{-5mm}
\caption{Attention Cost-volume. This module takes two frames of point clouds with their local features as input and associates the two point clouds. Finally, the module outputs the embedding features located in $PC_1$.}
\label{fig:cost}
\vspace{-8pt}
\end{figure}
\subsection{Hierarchical Embedding Mask Optimization}\label{sec:mask}
\vspace{-4pt}
It is a new problem to convert the embedding features $E$ to a global consistent pose transformation between two frames. In this subsection, a novel embedding mask is proposed to generate pose transformation from embedding features.
It should be noted that some points may belong to dynamic objects or be occluded in the other frame. It is necessary to filter out these points and keep the points that are of value to the LiDAR odometry task. To deal with this, the embedding features $E = \{ {e_{i}}|{e_{i}} \in {\mathbb{R}^c}\} _{i = 1}^n$ and the features $F_1$ of $PC_1$ are input to a shared MLP followed by the softmax operation along the point dimension to obtain the embedding mask (as the initial embedding mask in Fig.~\ref{fig:network}):
\vspace{-0.2cm}
\begin{equation}
{M} = softmax(sharedMLP({E} \oplus F_1)),\vspace{-0.2cm}\end{equation}
where $M = \{ {m_{i}}|{m_{i}} \in {\mathbb{R}^c}\} _{i = 1}^n$ represents trainable masks for prioritizing embedding features of $n$ points in $PC_1$. Each point has a characteristic weight ranging from 0 to 1. The smaller the weight of a point is, the more likely the point needs to be filtered out, and vice versa. Then, the quaternion $q \in {\mathbb{R}^4}$ and translation vector $t \in {\mathbb{R}^3}$ can be generated by weighting embedding features and FC layers separately, and $q$ is normalized to accord with the characteristic of rotation.\vspace{-0.3cm}
\begin{equation}
q = \frac{{FC(\sum\limits_{i = 1}^n {{e_i} \odot {m_i}} )}}{{\left| {FC(\sum\limits_{i = 1}^n {{e_i} \odot {m_i}} )} \right|}}\label{eqn:q},\vspace{-0.2cm}\end{equation}
\vspace{-0.1cm}
\begin{equation}
t = FC(\sum\limits_{i = 1}^n {{e_{i}} \odot {m_{i}}} )\label{eqn:t}.\vspace{-0.2cm}\end{equation}
The trainable mask $M$ is also part of the hierarchical refinement. As shown in Fig.~\ref{fig:network}, the embedding mask is propagated to denser layers of the point cloud just like embedding features $E$ and pose. The embedding mask is also optimized in a coarse-to-fine approach during the warp-refinement process, making the final mask estimation and the calculation of pose transformation accurate and reliable. We call this process the hierarchical embedding mask optimization.
\subsection{Pose Warp-Refinement Module}\label{sec:refine}
\vspace{-4pt}
To achieve the coarse-to-fine refinement process in an end-to-end fashion, we propose the differentiable warp-refinement module based on pose transformation as shown in Fig.~\ref{fig:warping}. This module contains several key parts: set upconv layer, pose warping, embedding feature and embedding mask refinement, and pose refinement.
\vspace{5pt}
\noindent{}{\bf Set Upconv Layer:}
To refine the pose estimation in a coarse-to-fine approach, the set upconv layer \cite{liu2019flownet3d} is adopted here to enable features of point cloud to propagate from sparse level to dense level. Embedding features $E^{l+1}$ and embedding masks $M^{l+1}$ of $l+1$ layer are propagated through set upconv layer to obtain coarse embedding features $CE^{l}=\{ {ce_{i}^l}|{ce_{i}^l} \in {\mathbb{R}^{c^l}}\} _{i = 1}^{n^l}$ and coarse embedding masks $CM^{l}= \{ {cm_{i}^l}|{cm_{i}^l} \in {\mathbb{R}^{c^l}}\} _{i = 1}^{n^l}$ that need to be optimized at the $l$-th level.
\vspace{5pt}
\noindent{}{\bf Pose Warping:}
The process of pose warping means that the quaternion $q^{l+1}$ and translation vector $t^{l+1}$ from the ${(l+1)}$-th level are applied to warp $PC_1^l=\{ {x_{i}^l}|{x_{i}^l} \in {\mathbb{R}^{c^l}}\} _{i = 1}^{n^l}$ to generate $PC_{1,warped}^l=\{ {x_{i,warped}^l}|{x_{i,warped}^l} \in {\mathbb{R}^{c^l}}\} _{i = 1}^{n^l}$. The warped $PC_{1,warped}^l$ is closer to $PC_2^l$ than original $PC_{1}^l$, which makes the residual motion estimation easier at the $l$-th level. The equation of warping transformation is as follows:\vspace{-0.2cm}
\begin{equation}
[0,x_{i,warped}^l] = {q^{l+1}}[0,x_{i}^l](q^{l+1})^{-1}+[0, t^{l+1}].\vspace{-0.2cm}\end{equation}
Then, the attentive cost volume between $PC{_{1,warped}^l}$ and $PC_2^l$ is re-calculated to estimate the residual motion. Following the approach introduced in Sec.~\ref{sec:cost}, re-embedding features between $PC{_{1,warped}^l}$ and $PC_2^l$ are calculated and denoted as $RE^l = \{ {re_{i}^l}|{re_{i}^l} \in {\mathbb{R}^{c^l}}\} _{i = 1}^{n^l}$.
\begin{figure}[t]
\centering
\vspace{-0mm}
\resizebox{1.00\columnwidth}{!}
{
\includegraphics[scale=1.00]{fignetwork_warp.pdf}}
\vspace{-5mm}
\caption{The details of the proposed Pose Warp-Refinement module at the $l$-th level.}
\label{fig:warping}
\vspace{-10pt}
\end{figure}
\vspace{5pt}
\noindent{}{\bf Embedding Feature and Embedding Mask Refinement:}
The generated coarse embedding feature $ce_i^{l}$, the re-embedding feature $re_{i}^l$, and the features $f_i^l$ of $PC_1^l$ are concatenated and input to a shared MLP to obtain embedding features $E^{l}=\{ {e_{i}^l}|{e_{i}^l} \in {\mathbb{R}^{c^l}}\} _{i = 1}^{n^l}$ at the $l$-th level:
\vspace{-0.2cm}
\begin{equation}
\vspace{-0.1cm}
e_{i}^l = MLP(ce_{i}^l \oplus re_{i}^{l} \oplus f_i^l).
\vspace{-0.2cm}\end{equation}The output of this MLP is the optimized embedding features of $l$-th level, which will not only participate in the following pose generation operation but also be output as the input to the next level warp-refinement module.
Like the refinement of the embedding feature, the newly generated embedding feature $e_{i}^l$, the generated coarse embedding mask $cm_i^{l}$, and the local feature $f_i^l$ of $PC_1^l$ are concatenated and input to a shared MLP and softmax operation along the point dimension to obtain the embedding mask $M^{l}= \{ {m_{i}^l}|{m_{i}^l} \in {\mathbb{R}^{c^l}}\} _{i = 1}^{n^l}$ at the $l$-th level
\vspace{-0.1cm}
\begin{equation}
M^l = softmax(sharedMLP(E^l \oplus CM^{l} \oplus F^l)).
\vspace{-0.2cm}
\end{equation}
\vspace{0pt}
\noindent{}{\bf Pose Refinement:}
The residual $\Delta q^l$ and $\Delta t^l$ can be obtained from the refined embedding features and mask following the formulas~(\ref{eqn:q}) and (\ref{eqn:t}) in Sec.~\ref{sec:mask}. Lastly, the refined quaternion $q^l$ and translation vector $t^l$ of the $l$-th level can be calculated by:\vspace{-0.2cm}\begin{equation}{q^l} = {\Delta q^{l}}{q^{l + 1}},\end{equation}\vspace{-0.5cm}\begin{equation}[0,t^l] = {\Delta q^{l}}[0,t^{l+1}](\Delta q^{l})^{-1}+[0, \Delta t^{l}].\end{equation}
\subsection{Training Loss}
\vspace{-4pt}
The network outputs quaternion $q^l$ and translation vector $t^l$ from four different levels of point clouds. The outputs of each level will enter into a designed loss function and be used to calculate the supervised loss $\ell^l$. Due to the difference in scale and units between translation vector $t$ and quaternion $q$, two learnable parameters $s_x$ and $s_q$ are introduced like previous deep odometry work \cite{li2019net}. The training loss function at the $l$-th level is: \begin{equation}\begin{gathered}
{\ell ^l} = {\left\| {{t_{gt}} - {t^l}} \right\|}exp( - {s_x}) + {s_x} \hfill \\
{\text{ }} + {\left\| {{q_{gt}} - \frac{{{q^l}}}{{\left\| {{q^l}} \right\|}}} \right\|_2}exp( - {s_q}) + {s_q},
\end{gathered} \label{eqn:sxsq} \end{equation}
where $\left\| \cdot \right\|$ and ${\left\| \cdot \right\|_2}$ denotes the ${\ell _1}$-norm and the ${\ell _2}$-norm respectively. ${{t_{gt}}}$ and ${{q_{gt}}}$ are the ground-truth translation vector and quaternion respectively generated by the ground-truth pose transformation matrix. Then, a multi-scale supervised approach is adopted. The total training loss $\ell$ is:\vspace{-0.2cm}
\begin{equation}\ell = \sum\nolimits_{l = 1}^L {{\alpha ^l}} {\ell ^l}\label{eqn:l},\end{equation}
where $L$ is the total number of warp-refinement levels and ${{\alpha ^l}}$ denotes the weight of the $l$-th level.
\setlength{\tabcolsep}{0.9mm}
\begin{table*}[t]
\centering
\footnotesize
\begin{center}
\resizebox{1.0\textwidth}{!}
{
\begin{tabular}{l||cc|cc|cc|cc|cc|cc|cc|cc|cc|cc|cc||cc}
\toprule
& \multicolumn{2}{c|}{00$^*$} &\multicolumn{2}{c|}{01$^*$} & \multicolumn{2}{c|}{02$^*$} & \multicolumn{2}{c|}{03$^*$} & \multicolumn{2}{c|}{04$^*$} & \multicolumn{2}{c|}{05$^*$} & \multicolumn{2}{c|}{06$^*$} & \multicolumn{2}{c|}{07} & \multicolumn{2}{c|}{08} & \multicolumn{2}{c|}{09} &\multicolumn{2}{c||}{10} &\multicolumn{2}{c}{Mean on 07-10} \\
\cline{2-25}\noalign{\smallskip}
\multirow{-2}{*}{\begin{tabular}[c]{@{}c@{}}Method \end{tabular}}
& $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ \\
\hline\hline
\noalign{\smallskip}
Full LOAM \cite{zhang2017low}
&1.10 & 0.53
&2.79 & 0.55
&1.54 & 0.55
&1.13 & 0.65
&1.45 & 0.50
&0.75 & 0.38
&0.72 & 0.39
&0.69 & 0.50
&1.18 & 0.44
&1.20 & 0.48
&1.51 & \bf0.57
& 1.145 & 0.498
\\
\hline
ICP-po2po
&6.88& 2.99
&11.21& 2.58
& 8.21 &3.39
&11.07& 5.05
& 6.64 &4.02
& 3.97& 1.93
&1.95 &1.59
&5.17 &3.35
&10.04 &4.93
&6.93 &2.89
&8.91 &4.74
&7.763 &3.978
\\
ICP-po2pl
&3.80 & 1.73
&13.53 & 2.58
&9.00 & 2.74
&2.72 & 1.63
&2.96 & 2.58
&2.29 & 1.08
&1.77 & 1.00
&1.55 & 1.42
&4.42 & 2.14
&3.95 & 1.71
&6.13 & 2.60
&4.013 & 1.968
\\
GICP \cite{segal2009generalized}
&1.29 & 0.64
&4.39 & 0.91
&2.53 & 0.77
&1.68 & 1.08
&3.76 & 1.07
&1.02 & 0.54
&0.92 & 0.46
&0.64 & 0.45
&1.58 & 0.75
&1.97 & 0.77
&1.31 &0.62
&1.375 & 0.648
\\
CLS \cite{velas2016collar}
&2.11 & 0.95
&4.22 & 1.05
&2.29 & 0.86
&1.63 & 1.09
&1.59 & 0.71
&1.98 & 0.92
&0.92 & 0.46
&1.04 & 0.73
&2.14 & 1.05
&1.95 & 0.92
&3.46 & 1.28
&2.148 & 0.995
\\
Velas et al. \cite{velas2018cnn}
&3.02 & NA
&4.44 & NA
&3.42 & NA
&4.94 & NA
&1.77 & NA
&2.35 & NA
&1.88 & NA
&1.77 & NA
&2.89 & NA
&4.94 & NA
&3.27 & NA
&3.218 & NA
\\
LO-Net \cite{li2019net}
&1.47 & 0.72
&1.36 & 0.47
&1.52 & 0.71
&1.03 & 0.66
&0.51 & 0.65
&1.04 & 0.69
&0.71 & 0.50
&1.70 & 0.89
&2.12 & 0.77
&1.37 & 0.58
&1.80 & 0.93
&1.748 & 0.793
\\
DMLO \cite{li2020dmlo}
& NA & NA
& NA & NA
& NA & NA
& NA & NA
& NA & NA
& NA & NA
& NA & NA
&0.73 & 0.48
&\bf 1.08 & \bf 0.42
&1.10 & 0.61
&\bf 1.12 & 0.64
&\bf 1.008 & 0.538
\\
LOAM w/o mapping
&15.99 &6.25
& 3.43 &0.93
& 9.40 & 3.68
& 18.18 &9.91
& 9.59 & 4.57
& 9.16 & 4.10
& 8.91 & 4.63
& 10.87 & 6.76
& 12.72 & 5.77
& 8.10 & 4.30
& 12.67 & 8.79
& 11.090 & 6.405
\\
Ours
&\bf0.78 & \bf0.42
&\bf0.67 & \bf0.23
&\bf0.86 & \bf0.41
&\bf0.76 & \bf0.44
&\bf0.37 & \bf0.40
&\bf0.45 & \bf0.27
&\bf0.27 & \bf0.22
&\bf0.60 & \bf0.44
&1.26 & 0.55
&\bf0.79 & \bf0.35
&1.69 & 0.62
&1.085 &\bf 0.490
\\ \bottomrule
\end{tabular}
}
\end{center}
\vspace{-8pt}
\caption{The LiDAR odometry experiment results on KITTI odometry dataset \cite{geiger2013vision}. $t_{rel}$ and $r_{rel}$ mean the average translational RMSE (\%) and rotational RMSE ($^{\circ}$/100m) respectively on all possible subsequences in the length of $100,200,...,800m$. `$^*$' means the training sequence. LOAM is a complete SLAM system, including back-end optimization and others only include odometry. The data other than the last three lines are from \cite{li2019net}. The results of LOAM w/o mapping are obtained by running their published codes. The best results are bold.}
\vspace{-8pt}
\label{table:lidar}
\end{table*}
\setlength{\tabcolsep}{0.9mm}
\begin{table*}[t]
\centering
\footnotesize
\begin{center}
\resizebox{0.9\textwidth}{!}
{
\begin{tabular}{l||cc|cc|cc|cc|cc|cc|cc|cc|cc|cc|cc||cc}
\toprule
& \multicolumn{2}{c|}{00$^*$} &\multicolumn{2}{c|}{01$^*$} & \multicolumn{2}{c|}{02$^*$} & \multicolumn{2}{c|}{03$^*$} & \multicolumn{2}{c|}{04$^*$} & \multicolumn{2}{c|}{05$^*$} & \multicolumn{2}{c|}{06$^*$} & \multicolumn{2}{c|}{07} & \multicolumn{2}{c|}{08} & \multicolumn{2}{c|}{09$^*$} &\multicolumn{2}{c||}{10$^*$} &\multicolumn{2}{c}{Mean on 07-08} \\
\cline{2-25}\noalign{\smallskip}
\multirow{-2}{*}{\begin{tabular}[c]{@{}c@{}}Method \end{tabular}}
& $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ \\
\hline\hline
\noalign{\smallskip}
LodoNet \cite{zheng2020lodonet}
&1.43 &0.69
&0.96 &0.28
&1.46 &0.57
&2.12 &0.98
&0.65 &0.45
&1.07 &0.59
&0.62 &0.34
&1.86 &1.64
&2.04 &0.97
&0.63 &0.35
&1.18 &0.45
&1.950 &1.305
\\
Ours
&\bf0.75 & \bf0.36
&\bf0.57 & \bf0.21
&\bf0.83 & \bf0.37
&\bf0.90 & \bf0.38
&\bf0.45 & \bf0.39
&\bf0.53 & \bf0.31
&\bf0.37 & \bf0.22
&\bf0.61 & \bf0.43
&\bf1.29 & \bf0.57
&\bf0.55 & \bf0.24
&\bf0.61 &\bf0.39
&\bf0.950 &\bf0.500
\\ \bottomrule
\end{tabular}
}
\end{center}
\vspace{-8pt}
\caption{The LiDAR odometry experiment results on KITTI odometry dataset \cite{geiger2013vision} compared with \cite{zheng2020lodonet}. As \cite{zheng2020lodonet} is trained on sequences 00-06, 09-10 and tested on sequences 07-08, we train and test our model like this to make comparisons with \cite{zheng2020lodonet}.}
\vspace{-8pt}
\label{table:lidar_7_8}
\end{table*}
\setlength{\tabcolsep}{1.3mm}
\begin{table}[t]
\centering
\footnotesize
\begin{center}
\resizebox{0.8\columnwidth}{!}
{
\begin{tabular}{l||cc|cc||cc}
\toprule
& \multicolumn{2}{c|}{04} &\multicolumn{2}{c||}{10} &\multicolumn{2}{c}{Mean} \\
\cline{2-7}\noalign{\smallskip}
\multirow{-2}{*}{\begin{tabular}[c]{@{}c@{}}Method \end{tabular}}
& $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ \\
\hline\hline
\noalign{\smallskip}
DeepPCO \cite{wang2019deeppco}
&2.63 & 3.05
&2.47 & 6.59
&2.550 & 4.820
\\
Ours
&\bf0.73 & \bf0.43
&\bf1.57 & \bf0.57
&\bf1.150 & \bf 0.500
\\ \bottomrule
\end{tabular}
}
\end{center}
\vspace{-8pt}
\caption{The LiDAR odometry results on sequences 04 and 10 of KITTI odometry dataset \cite{geiger2013vision}. As \cite{wang2019deeppco} is trained on sequences 00-03, 05-09 and only reports testing results on sequences 04 and 10, we train and test our model like this to make comparisons with it.}
\vspace{-8pt}
\label{table:lidar4_10}
\vspace{0.2cm}
\end{table}
\vspace{-0.2cm}
\setlength{\tabcolsep}{1.3mm}
\begin{table}[t]
\centering
\footnotesize
\begin{center}
\resizebox{0.8\columnwidth}{!}
{
\begin{tabular}{l||cc|cc||cc}
\toprule
& \multicolumn{2}{c|}{09} &\multicolumn{2}{c||}{10} &\multicolumn{2}{c}{Mean} \\
\cline{2-7}\noalign{\smallskip}
\multirow{-2}{*}{\begin{tabular}[c]{@{}c@{}}Method \end{tabular}}
& $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ \\
\hline\hline
\noalign{\smallskip}
Cho et al. \cite{cho2020}
&4.87 & 1.95
&5.02 & 1.83
&4.945 & 1.890
\\
Ours
&\bf0.73 & \bf0.40
&\bf1.16 & \bf0.78
&\bf0.945 & \bf 0.295
\\ \bottomrule
\end{tabular}
}
\end{center}
\vspace{-8pt}
\caption{The LiDAR odometry results on sequences 09 and 10 of KITTI odometry dataset \cite{geiger2013vision}. As \cite{cho2020} applies unsupervised training on sequences 00-08, and reports testing results on sequences 09 and 10, we train and test our model like this to make fair comparisons.}
\vspace{-6pt}
\label{table:lidar09_10}
\vspace{-0.2cm}
\end{table}
\section{Implementation}\label{sec:implementation}
\vspace{-4pt}
\subsection{KITTI Odometry Dataset}
\vspace{-4pt}
KITTI odometry dataset \cite{geiger2012we,geiger2013vision} is composed of 22 independent sequences. The Velodyne LiDAR point clouds in the dataset are used in our experiments. All scans of point clouds have XYZ coordinates and the reflectance information. Sequences 00-10 (23201 scans) contain ground truth pose (trajectory), while there is no ground truth publicly available for the remaining sequences 11-21 (20351 scans). By driving under different road environments, such as highways, residential roads, and campus roads, the sampling vehicle captures point clouds for the LiDAR odometry task from different environments.
\vspace{5pt}
\noindent{}{\bf Data Preprocessing:}
Only coordinates of LiDAR points are used in our method. Because the ground truth poses are denoted in the left camera coordinate system, all the training and evaluation processes of this network are carried out in the left camera coordinate system. Thus, the captured point clouds from the Velodyne LiDAR are firstly transformed to the left camera coordinate system by:
\vspace{-0.2cm}
\begin{equation}
{P_{cam}} = {T_r}P_{vel},
\vspace{-0.2cm}\end{equation}
where ${P_{cam}}$ and ${P_{vel}}$ are the point cloud coordinates in the left camera coordinate system and LiDAR coordinate system respectively, and ${T_r}$ is the calibration matrix of each sequence.
Moreover, the point cloud collected by the LiDAR sensor often contains outliers at the edge of the point cloud in each frame. This is often because objects are far away from the LiDAR sensor, thus forming incomplete point clouds in the edge. To filter out these outlier points, the points out of the $30 \times 30m^2$ square area around the vehicle are filtered out for each point cloud $P_{cam}$. To speed up the data reading and training, the ground less than $0.55m$ in height is removed. For our model, the performances of removing and reserving the ground are similar. The detailed comparison is shown in the supplementary material.
\vspace{4pt}
\noindent{}{\bf Data Augmentation:}
We augment the training dataset by the augmentation matrix ${T_{aug}}$, generated by the rotation matrix ${R_{aug}}$ and translation vector ${t_{aug}}$.
Varied values of yaw-pitch-roll Euler angles are generated by Gaussian distribution around $0^{\circ}$. Then the ${R_{aug}}$ can be obtained from these random Euler angles. Similarly, the ${t_{aug}}$ is generated by the same process.
The composed ${T_{aug}}$ is then used to augment the $PC_1$ to obtain new point clouds $PC_{1,aug}$ by: \vspace{-0.2cm}
\begin{equation} \vspace{-0.2cm}
{PC_{1,aug}} = {T_{aug}}{PC_1}.\end{equation}
Correspondingly, the ground truth transformation matrix is also modified as:\vspace{-0.2cm}
\begin{equation}\vspace{-0.2cm}
{T_{trans}} = {T_{aug}}{T_p},\end{equation}
where ${T_p}$ denotes the original ground truth pose transformation matrix from $PC_1$ to $PC_2$.
${T_{trans}}$ is then used to generate ${q_{gt}}$ and ${t_{gt}}$ to supervise the training of the network.
\subsection{Network Details}
\vspace{-4pt}
In the training and evaluation process, the input $N$ points are randomly sampled from the point clouds of two frames separately. It is unnecessary that the original input two point clouds have the same number of points. And $N$ is set to be 8192 in the proposed network. Each layer in MLP contains the ReLU activation function, except for the FC layer. For shared MLP, $1 \times 1$ convolution with $1$ stride is the implement manner. The detailed layer parameters including each linear layer width in MLP are described in the supplementary material. All training and evaluation experiments are conducted on a single NVIDIA RTX 2080Ti GPU with TensorFlow 1.9.0. The Adam optimizer is adopted with ${\beta _1} = 0.9$, ${\beta _2} = 0.999$. The initial learning rate is 0.001 and exponentially decays every 200000 steps until 0.00001. The initial values of the trainable parameters $s_x$ and $s_q$ are set as 0.0 and -2.5 respectively in formula~(\ref{eqn:sxsq}). For formula~(\ref{eqn:l}), ${\alpha _1}$ = 1.6, ${\alpha _2}$ = 0.8, ${\alpha _3}$ = 0.4, and $L=4$. The batch size is 8.
\vspace{-0.1cm}
\section{Experimental Results}
\vspace{-4pt}
In this section, the quantitative and qualitative results of the network performance on the LiDAR odometry task are demonstrated and compared with state-of-the-art methods. Then, the ablation studies are presented. Finally, the embedding mask is visualized and discussed.
\subsection{Performance Evaluation}
\vspace{-4pt}
As there are several modes to divide the training/testing sets for published papers \cite{li2019net,zheng2020lodonet,wang2019deeppco,zheng2020lodonet,cho2020}, in order to fairly compare with all current methods as we know, we train/test our model four times.
\noindent{}{\bf Using sequences 00-06/07-10 as training/testing sets:} Quantitative results are listed in Table~\ref{table:lidar}. ICP-point2point (ICP-po2po), ICP-point2plane (ICP-po2pl), GICP \cite{segal2009generalized} , CLS \cite{velas2016collar} are several classic LiDAR odometry estimation methods based on Iterative Closest Point (ICP) \cite{besl1992method}. LOAM \cite{zhang2017low} is based on the matching of extracted line and plane features, which has a similar idea with our method, and it is a hand-crafted method. It achieves the best results among LiDAR-based methods in the KITTI odometry evaluation benchmark \cite{geiger2013vision}. Velas et al. \cite{velas2018cnn} is a learning-based method. It has a good performance when only the translation is estimated, but the performance will decrease when estimating the 6-DOF pose. LO-Net \cite{li2019net} and DMLO \cite{li2020dmlo} are learning-based LiDAR odometry methods that have comparable results, but they have no codes publicly available, so we adopt the same training and testing sequences with \cite{li2019net,li2020dmlo}. Compared with \cite{li2020dmlo}, our method utilizes soft correspondence rather than exact matching pairs so as to relize end-to-end pose estimation, and we achieved similar results on $t_{rel}$ and better results on $r_{rel}$. Compared with \cite{li2019net}, our method does not need an extra mask network and can filter the outliers with the proposed hierarchical embedding mask. Moreover, our method does not need to project 3D point clouds to 2D \cite{li2019net,li2020dmlo} and obtains the LiDAR odometry directly from the raw 3D point clouds. We achieved better results than LO-Net~\cite{li2019net}, even than LOAM~\cite{zhang2017low} on most sequences. We believe the pose refinement target makes our internal trainable mask effective for various outliers other than only dynamic regions~\cite{li2019net} as shown in Fig.~\ref{fig:visual}. Moreover, the trainable iterative refinement makes the estimated pose refined multiple times in one network inference.
\vspace{0.1cm}
\setlength{\tabcolsep}{0.9mm}
\begin{table*}[t]
\footnotesize
\begin{center}
\resizebox{1.0\textwidth}{!}
{
\begin{tabular}{l|l||cc|cc|cc|cc|cc|cc|cc|cc|cc|cc|cc||cc}
\toprule
& & \multicolumn{2}{c|}{00$^*$} &\multicolumn{2}{c|}{01$^*$} & \multicolumn{2}{c|}{02$^*$} & \multicolumn{2}{c|}{03$^*$} & \multicolumn{2}{c|}{04$^*$} & \multicolumn{2}{c|}{05$^*$} & \multicolumn{2}{c|}{06$^*$} & \multicolumn{2}{c|}{07} & \multicolumn{2}{c|}{08} & \multicolumn{2}{c|}{09} &\multicolumn{2}{c||}{10} &\multicolumn{2}{c}{Mean on 07-10} \\
\cline{3-26}\noalign{\smallskip}
& \multirow{-2}{*}{\begin{tabular}[c]{@{}c@{}}Method \end{tabular}}
& $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ & $t_{rel}$ & $r_{rel}$ \\
\hline\hline
\noalign{\smallskip}
(a)
&Ours w/o mask
&0.99 & 0.58
&0.61 & 0.25
& 1.07 & 0.55
& 0.84 & 0.66
& 0.42 & 0.58
& 0.68 & 0.41
& 0.33 & 0.25
& 0.76 & 0.58
& 1.54 & 0.66
& 1.30 & 0.61
& 2.36 & 0.86
& 1.490 & 0.678
\\
& Ours w/o mask optimazation
& 0.91& 0.54
&\bf0.59 & 0.26
& 1.10 & 0.49
& 0.81 & \bf0.43
& 0.41 & 1.05
&0.57 & 0.39
& 0.34 & \bf0.22
& 0.67 & 0.54
& 1.41 & 0.59
& \bf0.78 & 0.39
& \bf1.32 & 0.66
& \bf1.045 & 0.545
\\
&Ours (full, with mask and mask optimazation)
&\bf0.78 & \bf0.42
&0.67 & \bf0.23
&\bf0.86 & \bf0.41
&\bf0.76 & 0.44
&\bf0.37 & \bf0.40
&\bf0.45 & \bf0.27
&\bf0.27 & \bf0.22
&\bf0.60 & \bf0.44
&\bf1.26 & \bf0.55
&0.79 & \bf0.35
&1.69 & \bf0.62
&1.085 &\bf 0.490
\\
\cline{1-26}\noalign{\smallskip}
(b) &Ours (w/o cost volume)
& 42.52 & 19.44
& 22.73 & 4.44
& 37.53 & 12.55
& 22.21 & 13.18
& 15.38 & 3.59
& 31.95 & 12.88
& 12.82 & 3.83
& 30.78 & 18.95
& 67.08 & 25.72
& 38.63 & 16.38
& 40.55 & 19.80
& 44.260 & 20.213
\\
&Ours (with the cost volume in \cite{liu2019flownet3d})
&1.11 & 0.51
&0.61 & \bf0.22
&0.97 & 0.47
& \bf0.68 & 0.61
& 0.50 & 0.67
& 0.51 & 0.30
& 0.38 & \bf 0.17
& 1.18 & 0.73
& 1.40 & 0.59
& 1.15 & 0.46
&\bf1.50 & 0.88
& 1.308 & 0.665
\\
&Ours (with the cost volume in \cite{wu2019pointpwc})
&\bf0.69 &\bf 0.35
&\bf0.59 & 0.23
& 0.87 & 0.42
& 0.79 & 0.55
&\bf 0.34 & \bf 0.34
& 0.49 & 0.28
& 0.28 & \bf0.17
& 0.62 & 0.47
& 1.47 & 0.62
& 0.97 & 0.48
&1.54 & \bf0.59
& 1.150 & 0.540
\\
&Ours (full, with the cost volume in \cite{wang2020hierarchical})
&0.78 & 0.42
&0.67 & 0.23
&\bf0.86 & \bf0.41
&0.76 & \bf0.44
&0.37 & 0.40
&\bf0.45 & \bf0.27
&\bf0.27 & 0.22
&\bf0.60 & \bf0.44
&\bf1.26 & \bf 0.55
&\bf0.79 & \bf0.35
&1.69 & 0.62
&\bf1.085 &\bf 0.490
\\
\cline{1-26}\noalign{\smallskip}
(c)
&Ours w/o Pose Warp-Refinement
&15.71 & 8.26
&13.92 & 5.62
& 13.38 & 7.07
& 15.74 & 16.62
& 5.00 & 6.89
& 14.19 & 7.87
& 6.38 & 5.33
& 12.52 & 11.73
& 15.64 & 9.11
& 10.17 & 7.60
& 16.09 & 11.29
& 13.605 & 9.933
\\
&Ours w/o Pose Warp
&3.07 & 1.40
&4.04 & 1.56
&3.27 & 1.27
&2.56 & 2.01
&0.91 & 1.13
&2.28 & 1.15
&2.08 & 1.18
&4.29 & 2.75
&5.00 & 2.14
&4.48 & 2.19
&5.39 & 3.36
&4.790 & 2.610
\\
&Ours (full, with Pose Warp-Refinement)
&\bf0.78 & \bf0.42
&\bf0.67 & \bf0.23
&\bf0.86 & \bf0.41
&\bf0.76 & \bf0.44
&\bf0.37 & \bf0.40
&\bf0.45 & \bf0.27
&\bf0.27 & \bf0.22
&\bf0.60 & \bf0.44
&\bf1.26 & \bf 0.55
&\bf0.79 & \bf0.35
&\bf1.69 & \bf0.62
&\bf1.085 &\bf 0.490
\\
\cline{1-26}\noalign{\smallskip}
(d)
&Ours (first embedding on the last level)
& 0.95 &0.46
&\bf0.67 & 0.29
& 1.03 & 0.44
& 0.96 & 0.77
& 0.48 & \bf0.40
& 0.74 & 0.40
& 0.41 & 0.26
& 0.94 & 0.49
& 1.41 & 0.56
& 1.01 & 0.47
&\bf1.67 & 0.82
& 1.258 & 0.585
\\
&Ours (full, first embedding on the penultimate level)
&\bf0.78 & \bf0.42
&\bf0.67 & \bf0.23
&\bf0.86 & \bf0.41
&\bf0.76 & \bf0.44
&\bf0.37 & \bf0.40
&\bf0.45 & \bf0.27
&\bf0.27 & \bf0.22
&\bf0.60 & \bf0.44
&\bf1.26 & \bf0.55
&\bf0.79 & \bf0.35
&1.69 &\bf 0.62
&\bf1.085 &\bf 0.490
\\
\bottomrule
\end{tabular}
}
\end{center}
\vspace{-8pt}
\caption{The ablation study results of LiDAR odometry on KITTI odometry dataset \cite{geiger2013vision}.}
\vspace{-9pt}
\label{table:ablation}
\end{table*}
\begin{figure}[t]
\begin{center}
\resizebox{0.95\columnwidth}{!}
{
\includegraphics[scale=1.00]{0106path2d-eps-converted-to.pdf}}
\end{center}
\vspace{-8pt}
\caption{Trajectory results of LOAM and ours on KITTI training sequences with ground truth. Ours is much better than the LOAM without mapping. And ours also has better performance on the two sequences than full LOAM, though ours is for odometry and LOAM has the mapping optimization.}
\label{fig:odometry_path2d}
\vspace{-8pt}
\end{figure}
\begin{figure}[t]
\begin{center}
\resizebox{0.95\columnwidth}{!}
{
\includegraphics[scale=1.00]{07path2d3d-eps-converted-to.pdf}}
\end{center}
\vspace{-8pt}
\caption{3D and 2D trajectory results on KITTI validation sequence 7 with ground truth. Our method obtained the most accurate trajectory.}
\label{fig:odometry_path2d3d}
\vspace{-9pt}
\end{figure}
\begin{figure}[t]
\begin{center}
\resizebox{0.95\columnwidth}{!}
{
\includegraphics[scale=1.00]{error_line-eps-converted-to.pdf}}
\end{center}
\vspace{-8pt}
\caption{Average translational and rotational error on KITTI sequences 00-10 on all possible subsequences in the length of $100,200,...,800m$. Our method has the best performance.}
\label{fig:odometry_error_line}
\vspace{-8pt}
\end{figure}
\vspace{-0.05cm}
\noindent{}{\bf Using other sequences as training/testing sets: }
\noindent{}{\bf 00-06, 09-10/07-08:} Quantitative results are listed in Table~\ref{table:lidar_7_8} to compare with a recent method \cite{zheng2020lodonet}.
\noindent{}{\bf 00-03, 05-09/04, 10:} As shown in Table~\ref{table:lidar4_10}, we adopted the same training/testing sets of the KITTI odometry dataset \cite{geiger2013vision} and compared with \cite{wang2019deeppco}.
\noindent{}{\bf 00-08/09-10:} \cite{cho2020} proposes an unsupervised method on LiDAR odometry. Table~\ref{table:lidar09_10} shows the quantitative comparison of this method with ours.
The results in Table~\ref{table:lidar_7_8}, Table~\ref{table:lidar4_10}, and Table~\ref{table:lidar09_10} show that our method outperforms them. \cite{wang2019deeppco} is based on 2D convolutional network, which loses the raw 3D information. \cite{zheng2020lodonet} finds matched keypoint pairs in 2D depth images to regress pose. \cite{cho2020} is a method of unsupervised training. The results demonstrate the superiority of our full 3D learning based LiDAR odometry.
The qualitative results are shown in Figs.~\ref{fig:odometry_path2d},~\ref{fig:odometry_path2d3d}, and~\ref{fig:odometry_error_line}. We compared our method with LOAM \cite{zhang2017low} and LOAM without mapping since \cite{li2019net,zheng2020lodonet,wang2019deeppco,li2020dmlo} do not release their codes. Full LOAM has a superb performance but degrades significantly without mapping. Ours is only for odometry and is better than the LOAM without mapping. At the same time, ours is even better than LOAM \cite{zhang2017low} on average evaluation.
\vspace{-0.15cm}
\subsection{Ablation Study}
\vspace{-4pt}
In order to analyze the effectiveness of each module, we remove or change components of our model to do the ablation studies on KITTI odometry dataset. The training/testing details are the same as described in Sec.~\ref{sec:implementation}.
\vspace{5pt}
\noindent{}{\bf Benefits of Embedding Mask:}
We first remove the optimization of the mask, which means that the embedding masks are independently estimated at each level. Then we entirely remove the mask and apply the average pooling to embedding features to calculate the pose. The results in Table~\ref{table:ablation}(a) show that the proposed embedding mask and its hierarchical optimization both contribute to better results.
The mismatched objects are different in various scenes. In some scenes, the car is dynamic, while the car is static in another scene (Fig.~\ref{fig:visual}). Through the learnable embedding mask, the network learns to filter outlier points from the overall motion pattern. The network achieves the effect of RANSAC~\cite{fischler1981random} by only once network inference.
\vspace{3pt}
\noindent{}{\bf Different Cost Volume:}
As there are different point feature embedding methods in point clouds, we compare three recent methods in our odometry task, including the flow embedding layer in FlowNet3D \cite{liu2019flownet3d}, point cost volume~\cite{wu2019pointpwc} and the attentive cost volume~\cite{wang2020hierarchical}. The results in Table~\ref{table:ablation}(b) show that the model with double attentive cost volume has the best results in the three. Therefore, robust point association also contributes to the LiDAR odometry task.
\vspace{5pt}
\noindent{}{\bf Effect of Pose Warp-Refinement:}
We first remove the pose warping and reserve the hierarchical refinement of the embedding feature and the mask. The results show a decline as in Table~\ref{table:ablation}(c). Then we remove the full pose warp-refinement module, which means that the embedding features, mask, and pose are only calculated once. As a result, the performance degrades a lot, which demonstrates the importance of the coarse-to-fine refinement.
\vspace{5pt}
\noindent{}{\bf Suitable Level for the First Embedding:}
As fewer points have less structural information, it is needed to decide which level is used to firstly correlate the two point clouds and generate the embedding features. We test the different levels of the first feature embedding. The results in Table~\ref{table:ablation}(d) demonstrate that the most suitable level for the first embedding is the penultimate level in the point feature pyramid, which shows that the 3D structure is also important for the coarse pose regression.
\begin{figure}[t]
\centering
\resizebox{0.95\columnwidth}{!}
{
\includegraphics[scale=1.00]{fig_mask_visual-eps-converted-to.pdf}}
\vspace{-0.1mm}
\caption{The Visualization of embedding mask. In the LiDAR point clouds, the small red points are the whole point cloud of the first frame, and the big points with distinctive colors are the sampled 2048 points with contributions to the pose transformation. In the example on the left, the buildings and steel bars around the highway have high weights, while the bushes and the moving car have low weights. In the example on the right, the buildings, the tree trunks, and the static car have high weights, while the weeds and the cyclist have low weights.}
\label{fig:visual}
\vspace{-0.4cm}
\end{figure}
\vspace{-0.2cm}
\subsection{Visualization of Embedding Mask}
\vspace{-4pt}
The proposed network uses embedding features of 2048 points to calculate the pose transformation in the final pose output layer. We visualize the embedding mask in the 2048 points to show the contribution of each point to the final pose transformation. As illustrated in Fig.~\ref{fig:visual}, the points that are sampled from the static and structured rigid objects such as buildings, steel bars, and the static car have higher weights. On the contrary, the points from the dynamic and irregular objects such as the moving car, the cyclist, bushes, and weeds have lower weights. In LO-Net~\cite{li2019net}, only the dynamic regions are masked, while our method also gives small weight to unreliable weeds and bushes. Therefore, the embedding mask can effectively reduce the influence of dynamic objects and other outliers on the estimation of pose transformation from adjacent frames. \vspace{-0.2cm}
\section{Conclusion}
\vspace{-4pt}
Different from the 2D projection-based methods \cite{nicolai2016deep,velas2018cnn,wang2019deeppco,li2019net}, a full 3D learning method based on PWC structure for LiDAR odometry is proposed in this paper. The pose warp-refinement module is proposed to realize the hierarchical refinement of 3D LiDAR odometry in a coarse-to-fine approach. The hierarchical embedding mask optimization is proposed to deal with various outlier points. And our method achieved a new state-of-the-art end-to-end LiDAR odometry method with no need for an extra mask network \cite{li2019net}. Since our mask can filter a variety of outlier points that are not suitable for calculating overall motion, it is a direction worth exploring to combine our new mask with mapping optimization in the future.
\vspace{4pt}
{\bf \small Acknowledgement.} {\small This work was supported in part by the Natural Science Foundation of China under Grant U1613218, U1913204, 62073222, in part by "Shu Guang" project supported by Shanghai Municipal Education Commission and Shanghai Education Development Foundation under Grant 19SG08, in part by grants from NVIDIA Corporation.}
{\small
\bibliographystyle{ieee_fullname}
|
\section{Introduction}
\IEEEPARstart{W}{ith} the rapid development of the Internet of Things (IoT) and edge computing technologies, IoT users can be distributed in order to provide wide coverage services in remote areas, e.g., environment monitoring, ocean transportation, smart grid, etc., \cite{7289337}. Considering that IoT users have low latency requirements, limited computing capabilities and battery power, computation tasks from IoT users can be offloaded to nearby edge servers for further performing, where edge servers are usually deployed at base stations (BSs) \cite{8030322}. However, terrestrial networks have not been established in some remote areas of deserts, oceans, and mountains, due to high network construction costs and specific geographical conditions \cite{9048610}. Therefore, it is hard to offer data collection and computation offloading for IoT users only by terrestrial networks in these remote areas. As a supplement to terrestrial networks, low earth orbit (LEO) satellite networks, which have global seamless coverage and low transmission delay time, play an important role in satellite-based IoT and edge computing \cite{8681409,8610431,8002583}.\par
For some remote areas without the coverage of terrestrial networks, LEO satellite networks can assist in gathering data from remote IoT users and transmitting them to cloud data centers on the ground for further processing \cite{8681409}. Due to the nature of LEO satellite networks, the transmission delay between remote IoT users and cloud data centers will be difficult to meet the real-time requirements of IoT users. Besides, the available network bandwidths will decrease to result in the network congestion as the number of IoT users increases. Considering the limited network bandwidths and real-time requirements, we can deploy edge servers on LEO satellites and provide edge computing services for remote IoT users to reduce their end-to-end delay \cite{8610431}, such as ocean transportation and smart grid \cite{7289337}. However, LEO satellites have limited resource capacities of computing, storage, bandwidth, and energy \cite{8896440}. In order to improve the operational efficiency of LEO satellite networks, multiple LEO satellites can provide computing services by the network function virtualization (NFV) technology \cite{7534741} for a IoT user in a cooperative manner.\par
\begin{figure}[tbp]
\centering
\includegraphics[width = 0.9\columnwidth]{Fig/1.eps}
\caption{Satellite edge computing framework with NFV.}
\label{Satellite edge computing framework with NFV}
\end{figure}
As a new paradigm in allocating network resources on-demand, NFV can support the decoupling of software and hardware equipments and enable service functions to run on commodity servers \cite{7534741}. By introducing NFV to satellite edge computing, we can abstract the available resources of LEO satellite networks into a resource pool and provide agile service provisioning for IoT users on-demand \cite{9062531}. Fig.~\ref{Satellite edge computing framework with NFV} shows the satellite edge computing framework with NFV, which consists of physical layer, virtual layer, and service layer. Physical layer consists of remote IoT sensors, actuators, ground networks, etc., and can provide sensing data collection and actuator interaction. For virtual layer, the available resources of LEO satellite networks, e.g., computing, storage, bandwidth, etc., can be abstracted into a resource pool by NFV for allocating available resources to IoT users in a flexible and scalable way. Service layer is responsible for managing LEO satellite network resources and orchestrating virtual network functions (VNFs) for IoT users.\par
For computation offloading in satellite edge computing, most of the existing work focuses on addressing the problem of resource allocation without considering the cooperative operation of multiple LEO satellites, where a computation task from a IoT user is considered as a whole for allocating network resources \cite{wang2019satellite,8672604}. However, these resource allocation strategies can not fully utilize the limited resources of LEO satellite networks. According to service function chaining (SFC), we can decompose a computation task from a IoT user into multiple VNFs in order and deploy these VNFs to multiple LEO satellites correspondingly to improve the network operational efficiency \cite{8690972,8406159}.\par
In this paper, we investigate the VNF placement problem in satellite edge computing. When a IoT user needs to offload its computation task to LEO satellites for obtaining computing services, a user request from the IoT user will be first sent to LEO satellite networks in order to obtain an access permission. The user request is considered as an SFC, consists of multiple VNFs in order, and carries the information concerning service type and resource requirements. Our aim is to maximize the number of allocated IoT users with minimum overall deployment cost, which is composed of energy consumption, bandwidth, and service delay costs. Considering that the user payoff is inversely proportional to the deployment cost, we establish the VNF placement problem with maximum network payoff, which is the sum of all allocated user payoffs.\par
To address the optimization problem of VNF placement, we formulate the problem as a potential game \cite{1996Potential}, which can be performed for making decisions in distributed computing as a non-cooperative game theory and widely used for handling the resource allocation problem in decentralized optimization algorithms \cite{7102682,8823046}. In potential game, a user request from a IoT user is considered as a player for finding a VNF placement strategy with maximum user payoff in a self-interested way and these players have potential conflicts in maximizing their payoffs \cite{8823046}. The payoff for each player can be improved by competing available resources with other players and then a Nash equilibrium can be acquired in a gradual iteration \cite{1996Potential}. Therefore, we implement a decentralized resource allocation algorithm based on a potential game, called as PGRA, to optimize the VNF placement solution. In each iteration, we traverse all the available paths for a user request to find a feasible strategy with maximum user payoff, where the Viterbi algorithm \cite{7469866} is used to address the VNF placement problem for each path. We assume that the resource allocation strategies for these players can be shared by a message synchronization mechanism. Our main contributions of this paper are summarized as follows:
\begin{itemize}
\item In the perspective of LEO satellite networks, we build the VNF placement problem with maximum network payoff, which is an integer non-linear programming problem. Our aim is to maximize the number of allocated IoT users while minimizing the overall deployment cost including energy consumption, bandwidth, and service delay costs.
\item To address the VNF placement problem, we formulate the problem as a potential game and analyze the problem by a game-theoretical approach. We implement a decentralized resource allocation algorithm based on a potential game to obtain an approximate strategy profile by finding a Nash equilibrium, where the Viterbi algorithm is used to deploy the VNFs for each user request.
\item We conduct the experiments to simulate and evaluate the performance of the proposed PGRA algorithm in LEO satellite networks. The simulation results show that the proposed PGRA algorithm outperforms two existing baseline algorithms of Greedy and Viterbi.
\end{itemize}
The remainder of this paper is organized as follows. Section \ref{Related Work} briefly reviews related work about resource allocation in satellite edge computing and decentralized algorithms. Section \ref{System Model} introduces the system model of VNF placement in satellite edge computing. In Section \ref{Problem Formulation}, we model the problem of VNF placement with maximum network payoff and prove it to be NP-hard. The VNF placement problem is formulated as a potential game and a decentralized resource allocation algorithm is implemented for tackling the problem in Section \ref{VNF Placement Game and Proposed Algorithm}. Section \ref{Performance Evaluation} discusses the performance of the proposed PGRA algorithm in LEO satellite networks. Finally, we provide the conclusion of this paper in Section \ref{Conclusion}.\par
\section{Related Work}\label{Related Work}
In satellite edge computing, most of the existing work focuses on offloading computation tasks from IoT devices to satellite edge nodes \cite{wang2019satellite,denby2019orbital,9013467}. In \cite{wang2019satellite}, considering traditional satellites as space edge computing nodes, the authors presented an approach of satellite edge computing to share on-board resources for IoT devices and provided computing services combined with the cloud. The orbital edge computing for nano-satellite constellations was discussed by formulation flying in \cite{denby2019orbital}. A fine-grained resource management in satellite edge computing was presented by the advanced K-means algorithm in \cite{9013467}. This existing literature considers a computation task as a whole to allocate the network resources. In this paper, we assume that each user request is viewed as an SFC and consists of multiple VNFs in order. We need to deploy these VNFs to LEO satellites and route traffic flows between two adjacent VNFs by inter-satellite links.\par
In order to effectively utilize the limited network resources, the VNF orchestration in software defined satellite networks has been investigated in \cite{8690972,8406159,9062531}. The authors in \cite{9062531} formulated the SFC planning problem as an integer non-linear programming problem and proposed a greedy algorithm to address it. The authors in \cite{8690972} discussed the problem of VNF placement to minimize the cost in software defined satellite networks and proposed a time-slot decoupled heuristic algorithm to solve this problem. In \cite{8406159}, an approach of deploying SFCs in satellite networks was presented to minimize the end-to-end service delay. In this paper, we jointly consider three deployment costs of network energy, network bandwidth, and user service delay to formulate the VNF placement problem with maximum network payoff.\par
As the number of remote IoT devices increases, decentralized mechanisms of network resource management have been a research topic in distributed networks \cite{8519737,8360511,8327930}, where potential game is widely used to address the problem of resource allocation in distributed computing \cite{7102682,8823046,8945402,8039433}. The authors in \cite{7102682} proposed a game-theoretical algorithm to minimize the number and power of anchor nodes in a wireless sensor network. In \cite{8823046}, a cost-effective edge user allocation (EUA) problem in edge computing was presented to maximize the number of served users with minimum system cost and the authors designed a decentralized algorithm by a potential game to address the EUA problem. Furthermore, computation offloading in satellite edge computing was discussed by a game-theoretical approach in \cite{8945402}. However, in this paper, we formulate the VNF placement problem with maximum network payoff as a potential game and use a decentralized resource allocation algorithm to optimize the problem.\par
\section{System Model}\label{System Model}
In this section, we discuss the system model of VNF placement in satellite edge computing \cite{wei2019satellite,9048610}, including physical network, user requests, and the VNF placement problem.\par
\subsection{Satellite Network}
We denote a satellite network as a directed graph $G(V,E)$. The parameter $V=\left \{v_n|n=1,2,\cdots,N \right\}$ indicates the set of satellite nodes, where the number of satellite nodes is $N$. We assume that the set of resource types supported by satellite node $v_n$ is denoted by $R_n$ and each satellite node has limited resource capacities, where two resource types of central processing unit (CPU) and storage are considered in this paper. Let us denote the $r$-th resource capacity of satellite node $v_n$ by $C_{n}^{r}$, $r \in R_n$. The parameter $E$ indicates the set of links between satellite nodes. We assume that each satellite has four inter-satellite links (ISLs) with neighbouring satellites. For link $e$, $e \in E$, we denote the bandwidth capacity by $B_{e}$ and the transmission delay time by $t_{e}$. The parameter $P_{n_{1}}^{n_{2}}$ indicates the set of the $d$ shortest paths between satellites $v_{n_1}$ and $v_{n_2}$.\par
Due to the limited power of satellites, one of our aims is to minimize the overall energy cost of a satellite network. We introduce a power consumption model for edge servers on satellite nodes \cite{7279063}. An edge server can be considered in four states of \emph{on}, \emph{idle}, \emph{unavailable off}, and \emph{available off}. In an \emph{on} state, an edge server can provide computing services for IoT users and will consume energy. We denote $P_{n}^{on}$ as the average power consumption of an edge server on satellite node $v_n$ in an \emph{on} state. If an edge server on satellite node $v_n$ does not provide computing services for any IoT users the edge server is considered into an \emph{idle} state and we use $P_{n}^{idle}$ to indicate the average idle power consumption. When the idle time for an edge server on satellite node $v_n$ is over the maximum idle threshold $t_{n}^{idle}$ the edge server will be into an \emph{unavailable off} state. If an edge server is in an \emph{unavailable off} state, it can not provide computing services for IoT users in the next time slot. When the off time for an edge server on satellite node $v_n$ is greater than the minimum off threshold $t_{n}^{off}$ the edge server can be in an \emph{available off} state, that is, the edge server can provide computing services for IoT users in the next time slot. If an edge server in an \emph{available off} state needs to provide computing services for IoT users there will exist a setup procedure for the edge server, where the state of the edge server will be converted from \emph{off} to \emph{on}. We assume that the period of the setup procedure is $1$ time slot and the average setup power consumption is considered as the maximum power $P_{n}^{max}$ of an edge server on satellite node $v_n$. For \emph{available} and \emph{unavailable} states, the power consumption of an edge server is zero. Note that a satellite node can be considered as a router for routing traffic flows when its edge server is in an \emph{off} state. Based on the above discussion, the state transition diagram of an edge server is shown in Fig.~\ref{State transition diagram of an edge server}.\par
\begin{figure}[tbp]
\centering
\includegraphics[width = 0.9\columnwidth]{Fig/2.eps}
\caption{State transition diagram of an edge server.}
\label{State transition diagram of an edge server}
\end{figure}
\subsection{User Requests}
When a IoT user needs to offload its computation task to satellites, a user request will be first sent to the LEO satellite network for obtaining an access permission, where the user request consists of multiple VNFs in specific order and can be considered as an SFC. We denote a set of user requests by $U=\left \{u_m|m=1,2,\cdots,M \right\}$ with $M$ user requests. The user request $u_{m}$, $u_{m} \in U$ is defined as a directed graph $G(F_{m},H_{m})$. The set $F_{m}=\left \{f_{m,1}=s_m,f_{m,2},\cdots,f_{m,\left |F_{m} \right |}=d_m \right\}$ indicates the set of VNFs for user request $u_m$, where $f_{m,i}$ indicates the $i$-th VNF of user request $u_m$, while the parameters $s_m$ and $d_m$ indicate the source and the destination of user request $u_m$, respectively. In satellite edge computing, we assume that the results of computation tasks processed by satellite nodes can be sent back to IoT users or transmitted to cloud data centers on ground. In these scenarios, the source and the destination of a user request can either be the same node or not. The resource requirements of each VNF include computing, storage, and execution time. We use $c_{m,i}^{r}$ to indicate the $r$-th resource requirement of $f_{m,i}$ and $t_{m,i}$ to represent the execution time of $f_{m,i}$. The set $H_{m}$ describes the set of edges for user request $u_m$. An edge between $f_{m,i_1}$ and $f_{m,i_2}$ is denoted by $h_{m}^{i_{1},i_{2}}$ and the bandwidth requirement of edge $h_{m}^{i_{1},i_{2}}$ is denoted by $b_{m}^{i_{1},i_{2}}$ accordingly. We define the maximum acceptable delay time for user request $u_m$ as $t_{m}^{delay}$.\par
\subsection{VNF Placement Problem}
\begin{figure}[tbp]
\centering
\includegraphics[width = 0.9\columnwidth]{Fig/3.eps}
\caption{Example of placing VNFs for three IoT users.}
\label{Example of placing VNFs for three IoT users}
\end{figure}
In this paper, we discuss the problem of dynamically deploying VNFs over varying time slots. A batch processing mode is simply used for allocating available resources of a LEO satellite network to user requests. We assume that a batch of user requests arrive concurrently at the beginning of each time slot and they can make decisions for deploying the VNFs to satellite nodes by their optimization strategies, where user requests have different resource requirements and maximum acceptable delay time, however, the source and the destination for each user request can be known. In addition, we assume that the satellite network topology is constant during the computing service period for each user request. Thus, the variability of a satellite network does not need to be considered when we deploy VNFs for user requests to satellite nodes. For deploying VNFs in dynamic cloud environment, we first abstract all available resources of a satellite network into a resource pool. Before performing resource allocation algorithms for the VNF placement, we need to free the resources that are used by the completed user requests in the previous time slot into the resource pool as available resources for new user requests in the current time slot. Then under remaining available resource and service requirement constraints, we can run resource allocation algorithms to orchestrate VNFs for new user requests with maximum network payoff. Our aim is to provide computing services for as many user requests as possible with minimum overall deployment cost, which consists of energy, bandwidth, and service delay costs.\par
For a user request, the computation task can be offloaded to satellites by different resource allocation strategies. However, the VNF placement strategy has an impact on the network payoff. Fig.~\ref{Example of placing VNFs for three IoT users} shows an example of placing VNFs for three IoT users. There are 9 satellite nodes, which are represented by $\left \{Sat1,Sat2,\cdots,Sat9\right\}$, and each satellite node deploys an edge server. The number of inter-satellite links for a satellite is 4. The user requests are composed of multiple specific VNFs in sequence. The three user requests can be described by $SFC1=\left \{s_1,f_{1,1},f_{1,2},d_1\right\}$, $SFC2=\left \{s_2,f_{2,1},f_{2,2},f_{2,3},d_2\right\}$, and $SFC3=\left \{s_3,f_{3,1},f_{3,2},d_3\right\}$, respectively. We assume that source $s_1$ and destination $d_1$ for $SFC1$ are on the same satellite node $Sat9$, source $s_2$ and destination $d_2$ for $SFC2$ are also on the same satellite node $Sat8$. In addition, source $s_3$ and destination $d_3$ for $SFC3$ are on satellite nodes $Sat4$ and $Sat7$, respectively. We deploy VNFs $f_{1,1}$ and $f_{1,2}$ for $SFC1$ to satellite node $Sat9$, VNFs $f_{2,1},f_{2,2}$, and $f_{2,3}$ for $SFC2$ to satellite node $Sat5$. For $SFC3$, we deploy VNF $f_{3,1}$ to satellite node $Sat2$ and VNF $f_{3,2}$ to satellite node $Sat1$, respectively. For $SFC1$, all VNFs are deployed on satellite node $Sat9$. The routing path for $SFC2$ is $\left \{Sat8,Sat5,Sat8\right\}$ and for $SFC3$ is $\left \{Sat5,Sat2,Sat1,Sat4,Sat7\right\}$. The strategies for the three IoT users are shown in Table \ref{VNF Placement Solutions for Three IoT Users}.\par
\begin{table}[tbp]
\centering
\caption{VNF Placement Solutions for Three IoT Users.}
\label{VNF Placement Solutions for Three IoT Users
\resizebox{\columnwidth}{!}{
\begin{tabular}{cc}
\hline
User & Strategy \\
\hline
SFC1 & $s_1$(Sat9)$\rightarrow$ $f_{1,1}$(Sat9) $\rightarrow$ $f_{1,2}$(Sat9)$\rightarrow$ $d_1$(Sat9) \\
\hline
SFC2 & $s_2$(Sat8) $\rightarrow$ $f_{2,1}$(Sat5) $\rightarrow$ $f_{2,2}$(Sat5) $\rightarrow$ $f_{2,3}$(Sat5) $\rightarrow$ $d_2$(Sat8) \\
\hline
SFC3 & $s_3$(Sat5) $\rightarrow$ $f_{3,1}$(Sat2) $\rightarrow$ $f_{3,2}$(Sat1) $\rightarrow$ Sat4 $\rightarrow$ $d_3$(Sat7)\\
\hline
\end{tabular
}
\end{table
\section{Problem Formulation}\label{Problem Formulation}
In this section, the problem of VNF placement is proposed by a mathematical method in satellite edge computing and proved to be NP-hard.
\subsection{Problem Description}
\begin{table}[tbp]
\caption{List of Symbols.}
\label{List of Symbols}
\centering
\resizebox{\columnwidth}{!}{
\begin{tabular}{cl
\hline
\multicolumn{2}{c}{\bfseries Satellite Network}\\
\hline
$V$ & Set of satellites with the number of $N$.\\
$v_n$ & The $n$-th satellite.\\
$R_n$ & Set of resources offered by satellite $v_n$.\\
$C_{n}^{r}$ & The $r$-th resource capacity for satellite $v_n$.\\
$P_{n}^{idle}$ & Idle power of an edge server on satellite $v_n$.\\
$P_{n}^{on}$ & Active power of an edge server on satellite $v_n$.\\
$P_{n}^{max}$ & Maximum power of an edge server on satellite $v_n$.\\
$t_{n}^{idle}$ & Maximum idle time of an edge server on satellite $v_n$.\\
$t_{n}^{off}$ & Minimum off time of an edge server on satellite $v_n$.\\
$E$ & Set of links between satellites.\\
$e$ & The $e$-th link.\\
$B_{e}$ & Bandwidth capacity for link $e$.\\
$t_{e}$ & Transmission delay for link $e$.\\
$P_{n_{1}}^{n_{2}}$ & Set of the $d$ shortest paths between $v_{n_{1}}$ and $v_{n_{2}}$.\\
\hline
\multicolumn{2}{c}{\bfseries User Requests}\\
\hline
$U$ & Set of $M$ user requests.\\
$u_{m}$ & The $m$-th user request.\\
$F_{m}$ & Set of VNFs for user request $u_m$.\\
$f_{m,i}$ & The $i$-th VNF for $u_m$.\\
$s_m,d_m$ & Source and destination of user request $u_m$.\\
$c_{m,i}^{r}$ & The $r$-th resource requirement for VNF $f_{m,i}$.\\
$t_{m,i}$ & Execute time for VNF $f_{m,i}$.\\
$H_{m}$ & Set of edges for user request $u_m$.\\
$h_{m}^{i_{1},i_{2}}$ & Edge between VNFs $f_{m,i_{1}}$ and $f_{m,i_{2}}$.\\
$b_{m}^{i_{1},i_{2}}$ & Bandwidth resource requirement for $h_{m}^{i_{1},i_{2}}$. \\
$t_{m}^{delay}$ & Maximum acceptable delay time for user request $u_m$. \\
\hline
\multicolumn{2}{c}{\bfseries Binary Decision Variables}\\
\hline
$x_{m,i}^{n}$ & $x_{m,i}^{n} = 1$ if $f_{m,i}$ is placed on satellite $v_n$ or $x_{m,i}^{n} = 0$.\\
$y_{m,p}^{i_{1},i_{2}}$ & $y_{m,p}^{i_{1},i_{2}}=1$ if path $p$ is used by $h_{m}^{i_{1},i_{2}}$ or $y_{m,p}^{i_{1},i_{2}}=0$.\\
\hline
\multicolumn{2}{c}{\bfseries Variables}\\
\hline
$q_{e}^{p}$ & $q_{e}^{p}=1$ if link $e$ is used by path $p$, otherwise $q_{e}^{p}=0$.\\
$z_m$ & $z_m=1$ if $u_m$ is allocated, otherwise $z_m=0$.\\
$\varphi_m^{bw}$ & Bandwidth cost for user request $u_m$.\\
$\varphi_m^{power}$ & Energy cost for user request $u_m$.\\
$\varphi_m^{delay}$ & Delay cost for user request $u_m$.\\
$\varphi_m$ & Payoff function for user request $u_m$.\\
$\Phi$ & Total payoff function. \\
$\alpha$ & Weight value. \\
\hline
\end{tabular}
}
\end{table}
In the perspective of a satellite network, we formulate the VNF placement problem as a constrained optimization problem with maximum network payoff in satellite edge computing, where the VNF placement problem is viewed as an integer non-linear programming problem. We assume that when a satellite network provides computing services for user requests, it can acquire uncertain user payoffs, which are affected by the deployment costs of user requests. Considering the limited physical resources of satellites and the real-time service requirement of IoT users, the goal of VNF placement in satellite edge computing is to reduce the energy consumption and bandwidth usage for a satellite network while minimizing the end-to-end user delay. Therefore, we assume that the deployment cost consists of three parts as: (1) energy cost, (2) bandwidth cost, and (3) service delay cost. The user payoff for a user request is non-negative and inversely proportional to the deployment cost. Therefore, the lower the deployment cost for a user request is, the higher the user payoff will be. When a user request is not deployed to satellite nodes, we denote the user payoff as zero. The overall network payoff is the sum of all user payoffs and our optimization aim is to maximize the network payoff within the network physical resource and service requirement constraints. That is, we make the number of allocated user requests maximum while minimizing the overall deployment cost. To better discuss the problem of VNF placement, we list the main symbols for our problem description in Table \ref{List of Symbols}.\par
Let us denote a binary decision variable $x_{m,i}^{n}=\left \{0,1\right \}$ to indicate whether VNF $f_{m,i}$ is placed on satellite node $v_n$, where $x_{m,i}^{n}=1$ if VNF $f_{m,i}$ is placed on satellite node $v_n$, otherwise $x_{m,i}^{n}=0$.\par
Another binary decision variable $y_{m,p}^{i_{1},i_{2}}=\left \{0,1\right \}$ is defined to describe which path is used by edge $h_{m}^{i_1,i_2}$. If path $p$ is used by $h_{m}^{i_1,i_2}$, then $y_{m,p}^{i_{1},i_{2}}=1$, otherwise $y_{m,p}^{i_{1},i_{2}}=0$.\par
We also denote a binary variable $q_{e}^{p}$ to indicate whether link $e$ is used by path $p$, as:
\begin{equation}\label{equation0}
q_{e}^{p}=
\begin{cases}
1& \text{if link $e$ is used by path $p$}, \\
0& \text{otherwise}.
\end{cases}
\end{equation}
A binary variable $z_m=\left \{0,1\right \}$ is used to indicate whether user request $u_m$ is deployed to satellite nodes, as:
\begin{equation}\label{equation1}
z_m=
\begin{cases}
1& \text{if user request $u_m$ is deployed to satellites}, \\
0& \text{otherwise}.
\end{cases}
\end{equation}
When we deploy the VNFs for a user request to satellites, the deployment cost can be composed of energy consumption, bandwidth, and service delay costs, where the three costs are normalized values.\par
\noindent The bandwidth resources $u_m^{bw}$ used by user request $u_m$ can be described as:
\begin{equation}\label{equation2}
u_m^{bw} =\!\! \sum\limits_{h_m^{{i_1},{i_2}} \in {H_m}}\!\! {\sum\limits_{{v_{{n_1}}},{v_{{n_2}}}}\! {\sum\limits_{p \in P_{{n_1}}^{{n_2}}} {\sum\limits_{e \in p} {x_{m,{i_1}}^{{n_1}} \!\cdot\! x_{m,{i_2}}^{{n_2}} \!\cdot\! y_{m,p}^{{i_1},{i_2}} \!\cdot\! q_e^p \!\cdot\! b_m^{{i_1},{i_2}}} } } }.
\end{equation}
The total bandwidth resource capacities in satellite network $G(V,E)$ are $\sum\limits_{e \in E}B_e$. Therefore, the normalized bandwidth cost for user request $u_m$ can be described as:
\begin{equation}\label{equation3}
\varphi_m^{bw}=\frac{u_m^{bw}}{\sum\limits_{e \in E}B_e}.
\end{equation}
According to the power consumption model \cite{7279063}, we assume that the energy consumption for an edge server on a satellite is mainly produced by running CPU. The energy consumption for an active edge server on satellite node $v_n$ can be expressed as:
\begin{equation}\label{equation4}
P_n^{on}= P_n^{idle} + \frac{\sum\limits_{u_m \in U}\sum\limits_{f_{m,i} \in F_m}x_{m,i}^{n}\cdot c_{m,i}^{cpu}}{C_{n}^{cpu}}\cdot (P_n^{max} -P_n^{idle}).
\end{equation}
Considering that there are different energy consumptions for placing VNFs on an edge server in various running states, we divide the VNF placement into four solutions based on energy consumption costs as follows:
\begin{itemize}
\item \emph{Case 1:} An edge server on satellite node $v_n$ is \emph{off} in the current time slot and will provide computing services for VNFs in the next time slot. Then if we place VNF $f_{m,i}$ on satellite node $v_n$, the power produced by VNF $f_{m,i}$ in the current time slot can be denoted as $u_{m,i}^{power}=0$.
\item \emph{Case 2:} An edge server on satellite node $v_n$ is \emph{off} in the current time slot and will not provide computing services for any VNFs in the next time slot. Then if we place VNF $f_{m,i}$ on satellite node $v_n$, the power produced by VNF $f_{m,i}$ in the current time slot can be denoted as $u_{m,i}^{power}=P_{n}^{max}$.
\item \emph{Case 3:} An edge server on satellite node $v_n$ is \emph{idle} in the current time slot and will not provide computing services for any VNFs in the next time slot. Then if we place VNF $f_{m,i}$ on satellite node $v_n$, the power produced by VNF $f_{m,i}$ in the current time slot could be denoted as $u_{m,i}^{power}=P_n^{idle} + \frac{x_{m,i}^{n}\cdot c_{m,i}^{cpu}}{C_{n}^{cpu}} \cdot (P_n^{max} -P_n^{idle}).$
\item \emph{Case 4:} An edge server on satellite node $v_n$ is \emph{idle} or \emph{on} in the current time slot and will provide computing services for VNFs in the next time slot. Then if we place VNF $f_{m,i}$ on satellite node $v_n$, the power produced by VNF $f_{m,i}$ in the current time slot could be seen as $u_{m,i}^{power}=\frac{x_{m,i}^{n}\cdot c_{m,i}^{cpu}}{C_{n}^{cpu}}\cdot (P_n^{max} -P_n^{idle})$.
\end{itemize}
The total energy consumption in satellite network $G(V,E)$ is represented by $\sum\limits_{v_n \in V}P_n^{max}$. The normalized energy cost for user request $u_m$ can be expressed as:
\begin{equation}\label{equation5}
\varphi_m^{power}=\frac{1}{\sum\limits_{v_n \in V}P_n^{max}} \sum\limits_{f_{m,i}\in F_m}\sum\limits_{v_n \in V} x_{m,i}^{n}\cdot u_{m,i}^{power}.
\end{equation}
For a user request, the source-to-destination delay time consists of two parts: computing delay for VNFs and transmission delay. The computing delay time for user request $u_m$ can be denoted by:
\begin{equation}\label{equation6}
t_m^{exec}=\sum_{f_{m,i}\in F_m}\sum\limits_{v_n \in V} x_{m,i}^{n} \cdot t_{m,i}.
\end{equation}
The transmission delay time for user request $u_m$ can be indicated as:
\begin{equation}\label{equation7}
t_m^{trans} =\!\! \sum\limits_{h_m^{{i_1},{i_2}} \in {H_m}}\!\! {\sum\limits_{{v_{{n_1}}},{v_{{n_2}}}}\! {\sum\limits_{p \in P_{{n_1}}^{{n_2}}} {\sum\limits_{e \in p} {x_{m,{i_1}}^{{n_1}} \!\cdot\! x_{m,{i_2}}^{{n_2}} \!\cdot\! y_{m,p}^{{i_1},{i_2}} \!\cdot\! q_e^p \!\cdot\! t_e} } } }.
\end{equation}
The maximum acceptable delay time for user request $u_m$ is $t_{m}^{delay}$, then we can denote the normalized service delay time cost for user request $u_m$ by:
\begin{equation}\label{equation8}
\varphi_m^{delay}=\frac{1}{t_{m}^{delay}}\left ( t_m^{exec}+ t_m^{trans}\right ).
\end{equation}
For user request $u_m$, the deployment cost is the weighted sum of energy consumption cost $\varphi_m^{power}$, bandwidth cost $\varphi_m^{bw}$, and service delay cost $\varphi_m^{delay}$, then the user payoff $\varphi_m$ can be indicated by:
\begin{equation}\label{equation9}
\varphi_m= (1 - \alpha_{1} \cdot \varphi_m^{bw} - \alpha_{2} \cdot \varphi_m^{power} - \alpha_{3} \cdot \varphi_m^{delay} ) \cdot z_m,
\end{equation}
where $\alpha_{1}$, $\alpha_{2}$, and $\alpha_{3}$, $\alpha_{1}+\alpha_{2}+\alpha_{3}=1$, are the weighted factors and can adjust the preferences of the three costs.\par
\noindent The overall network payoff $\Phi$ is the sum of all user payoffs and can be represented by:
\begin{equation}\label{equation10}
\Phi= \sum\limits_{u_m \in U} \varphi_m.
\end{equation}
In order to address the problem of VNF placement to maximize the overall network payoff, the following physical constraints need to be considered.\par
\noindent When the VNFs for user request $u_m$ are deployed to satellite nodes, each VNF $f_{m,i} \in F_m$ can be placed on one and only one satellite node. We describe the VNF deployment constraint as follows:
\begin{equation}\label{equation11}
\sum\limits_{v_n \in V} x_{m,i}^{n} = 1,\forall f_{m,i}\in F_m,\forall u_m \in U.
\end{equation}
For user request $u_m$, if two adjacent VNFs $f_{m,i_1}$ and $f_{m,i_2}$ are deployed on satellite nodes $v_{n_1}$ and $v_{n_2}$, respectively, we need to guarantee that there exists one path between $v_{n_1}$ and $v_{n_2}$ that can be used to route traffic flows from $f_{m,i_1}$ to $f_{m,i_2}$. The path selection constraint can be expressed by:
\begin{equation}\label{equation12}
\sum\limits_{p \in P_{n_1}^{n_2}}\!\!\! y_{m,p}^{i_1,i_2}\! = \!x_{m,i_1}^{n_1}\!\!\cdot \!\!x_{m,i_2}^{n_2}, f_{m,i_1},f_{m,i_2}\!\in \!F_m, v_{n_1},v_{n_2}\! \in \!V.
\end{equation}
When we place VNFs for user requests to satellite nodes, the resource requirements for each satellite node can not exceed its resource capacities. The satellite resource constraint can be described as:
\begin{equation}\label{equation13}
\sum\limits_{u_m \in U}\sum\limits_{f_{m,i}\in F_m} x_{m,i}^{n}\cdot c_{m,i}^{r}\leq C_{n}^{r},\forall v_{n}\in V,\forall r \in R_n.
\end{equation}
When we choose a path between two satellite nodes to route traffic flows, we also guarantee that the bandwidth requirements for each link can not be greater than the bandwidth capacity. The bandwidth resource constraint for $\forall e \in E$ can be indicated by:
\begin{equation}\label{equation14}
\sum\limits_{u_m \in U}\! {\sum\limits_{h_m^{{i_1},{i_2}} \in {H_m}}\!\! {\sum\limits_{{v_{{n_1}}},{v_{{n_2}}}}\! {\sum\limits_{p \in P_{{n_1}}^{{n_2}}} {\!\!\!\!{x_{m,{i_1}}^{{n_1}} \!\!\cdot\! x_{m,{i_2}}^{{n_2}} \!\cdot\! y_{m,p}^{{i_1},{i_2}} \!\cdot\! q_e^p \!\!\cdot\! b_m^{{i_1},{i_2}}} } } }}\!\!\leq \!\!B_{e}.
\end{equation}
For user request $u_m$, the source-to-destination delay time is less than the maximum acceptable delay time. The service delay time constraint can be described as:
\begin{equation}\label{equation15}
t_m^{exec}+ t_m^{trans}\leq t_{m}^{delay},\forall u_m \in U,
\end{equation}
where the maximum acceptable delay time for a user request is the sum of the executed time for all VNFs and the acceptable path transmission time. In this paper, we assume that the acceptable path transmission time is equal to the average transmission time of all source-to-destination paths $P_{s_m,all}^{d_m}$ in satellite network $G(V,E)$. Thus, we can denote the maximum acceptable delay time for user request $u_m$ by:
\begin{equation}\label{equation16}
t_{m}^{delay} = \sum_{f_{m,i}\in F_m}t_{m,i} + \frac{1}{\left |P_{s_m,all}^{d_m}\right |}\sum\limits_{p\in P_{s_m,all}^{d_m}}\sum\limits_{e\in p} t_{e}.
\end{equation}
In addition, we need to ensure that the idle time $t_{n,gap}^{idle}$ for an edge server on satellite $v_n$ can not exceed the maximum idle time threshold $t_{n}^{idle}$. For an edge server on satellite $v_n$, we denote the earliest idle time in the current \emph{idle} state by $t_{n,idle}^{old}$ and the current idle time by $t_{n,idle}^{new}$. The idle time constraint for an edge server on satellite $v_n$ can be indicated by:
\begin{equation}\label{equation17}
t_{n,gap}^{idle}=t_{n,idle}^{new} - t_{n,idle}^{old}\leq t_{n}^{idle},\forall v_n \in V.
\end{equation}
We also guarantee that the off time $t_{n,gap}^{off}$ for an edge server on satellite $v_n$ should be greater than the minimum off time threshold $t_{n}^{off}$. For an edge server on satellite $v_n$, we use $t_{n,off}^{old}$ to indicate the earliest off time in the current \emph{off} state and $t_{n,off}^{new}$ to indicate the current off time. The off time constraint for an edge server on satellite $v_n$ can be indicated by:
\begin{equation}\label{equation18}
t_{n,gap}^{off}=t_{n,off}^{new} - t_{n,off}^{old} \geq t_{n}^{off},\forall v_n \in V.
\end{equation}
Under the physical network resource and service requirement constraints in equations \eqref{equation11}-\eqref{equation18}, we formulate the VNF placement problem in satellite edge computing using \eqref{equation10} as an optimization problem with maximum network payoff. The optimization problem in this paper can be indicated by:
\begin{equation}\label{equation19}
\begin{aligned}
\text{max}\quad & \Phi \\
s.t.\quad & \eqref{equation11}-\eqref{equation18}.
\end{aligned}
\end{equation}
\subsection{Problem Analysis}
In this section, we reduce the capacitated plant location problem with single source constraints (CPLPSS) \cite{1995The} to the above VNF placement problem and prove that the formulated VNF placement problem is NP-hard \cite{7469866}.
In CPLPSS, a set of potential locations is given for deploying plants with fixed capacities and costs, and goods from the plants need to be provided to a set of customers with fixed demands. There are transportation costs for supplying goods from the plants to the customers. In addition, the goods demanded by a customer are provided only by a single plant. The problem aims to find an optimal solution of placing plants within the capacity and demand constraints to minimize the total operational and transportation costs.
To reduce a CPLPSS problem to our proposed VNF placement, we re-construct the problem description. Satellite nodes indicate plants and satellite resources represent plant capacities. User requests can be viewed as customers and resource demands for their computation tasks are described as customer required goods. We assume that all demanded resources from a user request are offered only by a satellite node. The operational cost of a plant is denoted by energy consumption and the transportation cost can be indicated by bandwidth and source-to-destination delay costs. For the proposed VNF placement problem, maximum overall network payoff is equal to minimizing the sum cost of energy consumption, bandwidth, and service delay costs within the constant number of user requests. Thus, a CPLPSS problem can be reduced to the proposed optimization problem. As a CPLPSS problem is NP-hard, the VNF placement problem is also NP-hard.
\section{VNF Placement Game and Proposed Algorithm}\label{VNF Placement Game and Proposed Algorithm}
In this section, we formulate the VNF placement problem as a potential game and analyze its property by a game-theoretical approach. Then a decentralized resource allocation algorithm based on a potential game is proposed for addressing the VNF placement problem.
\subsection{VNF Placement Game}
Game theory is a mathematical tool for analyzing interactive decision-making processes. A non-cooperative game can be denoted by $\Gamma =\left \{U,A,\varphi\right \}$. The term $U$ indicates the set of players. The strategy of the $m$-th player is denoted by $a_m \in A_m$, where $A_m$ represents the strategy set of the $m$-th player. $A=\prod\limits_{m = 1}^M {{A_m}}$ denotes the strategy combination of all players and $a=\left \{a_1,a_2,\cdots,a_M\right \}$ indicates the strategies of all players. We denote the strategies of all players except the $m$-th player as $a_{-m}=\left \{a_1,\cdots,a_{m-1},a_{m+1},\cdots,a_M\right \}$ and the strategy combination of all players except the $m$-th player as $A_{-m}=\prod\limits_{j \neq m}^M {{A_j}}$. The term $\varphi_{m}(a_m) \in \varphi$ indicates the payoff of the $m$-th player with the strategy $a_m$ and the term $\varphi(a)=\left \{\varphi_{1}(a_1),\varphi_{2}(a_2),\cdots,\varphi_{M}(a_M)\right \}$ is the payoff set of all players.
For a non-cooperative game, the players can make strategy decisions in a self-interested way for maximizing their payoffs until a Nash equilibrium is generated, where each player can not unilaterally deviate its strategy for improving the payoff.
\begin{definition}\label{definition1}
(\textbf{Nash equilibrium}) A strategy decision profile $a^*=\left \{a_1^*,a_2^*,\cdots,a_M^*\right \}$ of all user requests is a Nash equilibrium if no player has an incentive for unilaterally deviating the strategy, such that,
\begin{equation}\label{equation20}
\varphi_{m}(a_m^*,a_{-m}^*)\ge \varphi_{m}(a_m,a_{-m}^*), \forall u_m \in U, \forall a_m \in A_m.
\end{equation}
\end{definition}
Before finding a Nash equilibrium of the game, we should ensure whether the game admits at least a Nash equilibrium. As a special instance of non-cooperative games, potential game possesses a pure strategy Nash equilibrium \cite{1996Potential}, where a global potential function is used to map the payoffs of all players. The game can be considered as an exact potential game if an exact potential function is admitted.
\begin{definition}\label{definition2}
(\textbf{Exact Potential Game}) A game is an exact potential game if, for an exact potential function $\Phi(a), \forall u_m \in U, a_m, a_m' \in A_m$ and $a_{-m} \in A_{-m}$, there is,
\begin{equation}\label{equation21}
\Phi(a_m',a_{-m})\! - \!\Phi(a_m,a_{-m})\! = \!\varphi_{m}(a_m',a_{-m})\! - \!\varphi_{m}(a_m,a_{-m}).
\end{equation}
\end{definition}
The key for the existence of a Nash equilibrium is to prove the VNF placement game is a potential game. For the VNF placement in satellite edge computing, user requests have potential conflicts in maximizing their payoffs and the strategy of a user request has an effect on that of other user requests. We formulate the VNF placement problem as a non-cooperative game, where $M$ user requests are $M$ players, the VNF placement solution for user request $u_m$ represents the decision strategy $a_m$ and the payoff for user request $u_m$ indicates the payoff of the $m$-th player. The term $A_m$ indicates the strategy set of user request $u_m$ and $A$ is the strategy combination of all user requests. The exact potential function is denoted by $\Phi(a)$, which is the sum of all user request payoffs. Then we prove that the VNF placement game is a potential game.
\begin{proposition}\label{proposition1}
The VNF placement game is an exact potential game, where the payoff of the $m$-th player is indicated by $\varphi_{m}(a_m)$ and the exact potential function is the network payoff $\Phi(a)$.
\end{proposition}
\begin{proof}\label{proof1}
For $a_m',a_m \in A_m$, $a_{-m} \in A_{-m}$, according to equation \eqref{equation9}, there is,
\begin{equation}\label{equation22}
\Delta\varphi_{m} = \varphi_{m}(a_m',a_{-m}) - \varphi_{m}(a_m,a_{-m}).
\end{equation}
Similarly, according to equation \eqref{equation10}, we can obtain the potential function difference as:
\begin{equation}\label{equation23}
\begin{aligned}
\Delta\Phi\! & = \!\Phi(a_m',a_{-m}) - \Phi(a_m,a_{-m}) \\
& \! = \!\!\left [\!\varphi_{m}(a_m',a_{-m}) \!+ \!\!\!\!\!\sum\limits_{u_l\neq u_m}^U \!\!\!\!\varphi_l(a_l)\! \right ] \!\!\!-\!\!\! \left [ \!\varphi_{m}(a_m,a_{-m}) \!+ \!\!\!\!\!\sum\limits_{u_l\neq u_m}^U \!\!\!\!\varphi_l(a_l)\!\right ]\\
& \! = \!\varphi_{m}(a_m',a_{-m}) - \varphi_{m}(a_m,a_{-m}).
\end{aligned}
\end{equation}
There is $\Delta\Phi \equiv \Delta\varphi_{m}$ \cite{2015ChoiA,7102682}. Thus, we prove that the VNF placement game is an exact potential game and $\Phi(a)$ is an exact potential function.
\end{proof}
Considering that the VNF placement game is a potential game and has the finite improvement property \cite{2018A}, the VNF placement problem can be addressed by finding a Nash equilibrium in a finite gradual iteration. A decentralized resource allocation algorithm based on a potential game, which is discussed in the following subsection, is implemented to tackle the VNF placement problem.
\subsection{Decentralized Resource Allocation Algorithm}
\begin{algorithm}[tbp]
\caption{Resource Allocation Based on a Potential Game.}
\label{Resource Allocation Based on a Potential Game}
\hspace*{0.02in} {\bf Input:} User requests $U$;\\
\hspace*{0.02in} {\bf Output:} $a^{*}=\left[a_{1}^{*},a_{2}^{*},\cdots, a_{M}^{*}\right]$;
\begin{algorithmic}[1]
\STATE \textbf{Initialize:} $k=0,\forall u_m \in U,a_m(k)=\varnothing$;
\WHILE {$k<K_{max}$}
\FOR {$\forall u_m \in U$}
\FOR {$\forall p \in P_{s_m}^{d_m}$}
\STATE Search an optimal strategy $a'_m(k,p)$ for path $p$ by the Viterbi algorithm and compute the user payoff $\varphi (a'_m(k,p),a_{-m}(k))$;
\ENDFOR
\STATE Find the strategy $a'_m(k)$ with maximum user payoff according to\\ $a'_m(k) = \arg\max\limits_{a'_m(k,p)} \varphi (a'_m(k,p),a_{-m}(k))$;
\IF {$a'_m(k) \neq a_m(k)$}
\STATE Share decision $a'_m(k)$ with other user requests;
\ENDIF
\ENDFOR
\STATE Run a competitive mechanism for all user requests, and update strategy $a'(k)$ with the winning decision $a'_m(k)$;
\IF {$\left | \Phi(a'(k))- \Phi(a(k))\right |< \epsilon $}
\STATE $a^{*}=a'(k)$ and break;
\ELSE
\STATE $k \leftarrow k+1$;
\ENDIF
\ENDWHILE
\RETURN $a^{*}$;
\end{algorithmic}
\end{algorithm}
As the VNF placement problem is NP-hard \cite{7469866}, we implement a decentralized resource allocation algorithm by a potential game (PGRA) to find an approximate solution. The proposed PGRA algorithm can perform on the satellite network to improve the real-time decision-making capacity, where multiple satellites share the network resource states and the VNF placement strategies by interacting with each other, but satellites does not need to exchange the information with IoT users during the running time of the proposed PGRA algorithm. The procedure of the proposed PGRA algorithm is shown in Algorithm \ref{Resource Allocation Based on a Potential Game}. The maximum number of iterations is $K_{max}$. At the beginning, the initial iteration time is $k=0$, for $\forall u_m \in U$, we initialize the strategy as $a_m(k)=\varnothing$ and the user payoff as $\varphi_{m}(a_m(k),a_{-m}(k))$ accordingly. In iteration time $k$, each user request $u_m \in U$ needs to find an optimal strategy $a'_m(k)$ with maximum user payoff $\varphi_{m}(a'_m(k),a_{-m}(k))$ while the strategies of other user requests remain unchanged. For obtaining the optimal strategy of user request $u_m$, we search the $d$ shortest paths between source $s_m$ and destination $d_m$ successively and the local optimal strategy $a'_m(k,p)$ for each path $p$ is calculated by the Viterbi algorithm \cite{8933111}, which will be discussed later. Thus we can acquire an optimal strategy of user request $u_m$ in the current iteration by:
\begin{equation}\label{equation24}
a'_m(k) = \arg\max\limits_{a'_m(k,p)} \varphi (a'_m(k,p),a_{-m}(k)).
\end{equation}
If $a'_m(k) \neq a_m(k)$, the strategy $a'_m(k)$ will be shared by a message synchronization mechanism to other user requests for competing available resources of the satellite network. In each iteration, only a user request, which makes the overall network payoff maximum, can win the opportunity for updating its decision-making strategy and other user requests need to keep their old decision-making strategies. Note that the strategy decision-making processes for all user requests are performed simultaneously in parallel. The iteration process will terminate when no user request has an incentive to deviate its strategy unilaterally or the number of iterations exceeds the maximum iteration threshold. The final strategy profile $a^{*}=\left[a_{1}^{*},a_{2}^{*},\cdots, a_{M}^{*}\right]$ for all user requests is a Nash equilibrium of the VNF placement game and represents an approximate solution of VNF placement.
\begin{table*}[tbp]
\centering
\caption{Simulation Parameters.}
\label{Simulation Parameters}
\resizebox{\textwidth}{!}{
\begin{tabular}{cccccccc}
\hline
\multicolumn{8}{c}{\bfseries Satellite Network} \\
\hline
Name & \multicolumn{3}{c}{Total Number of Satellites} & \multicolumn{2}{c}{Number of Planes} & \multicolumn{2}{c}{Number of Satellites per Plane} \\
Value & \multicolumn{3}{c}{6,9,12,15} & \multicolumn{2}{c}{3} & \multicolumn{2}{c}{2,3,4,5} \\
\hline
\multicolumn{8}{c}{\bfseries Inter-Satellite Links} \\
\hline
Name & \multicolumn{3}{c}{Distance for a Plane} & \multicolumn{2}{c}{Distance for different Planes} & \multicolumn{2}{c}{Bandwidth} \\
Value & \multicolumn{3}{c}{600 km} & \multicolumn{2}{c}{400 km} & \multicolumn{2}{c}{100 Mbps} \\
\hline
\multicolumn{8}{c}{\bfseries Edge Servers on Satellites} \\
\hline
Name & vCPUs & Memory & Idle Power & Maximum Active Power & Setup Power & Maximum Idle Time & Minimum Off Time\\
Value & 112 & 192 GB & 49.9 W & 415 W & 415 W & 3 slots & 1 slot \\
\hline
\multicolumn{8}{c}{\bfseries User Requests} \\
\hline
Name & VNFs & vCPUs & Memory & Execution Time for VNFs & Bandwidth & \multicolumn{2}{c}{Running Time} \\
Lower & 5 & 4 & 4 GB & 10 ms & 10 Mbps & \multicolumn{2}{c}{1 slot} \\
Upper & 10 & 8 & 16 GB & 30 ms & 30 Mbps & \multicolumn{2}{c}{4 slots} \\
\hline
\end{tabular
}
\end{table*
\begin{figure}[tbp]
\centering
\includegraphics[width = \columnwidth]{Fig/4.eps}
\caption{Example of placing VNFs for a user request by the Viterbi algorithm.}
\label{Example of placing VNFs for a user request by the Viterbi algorithm}
\end{figure}
In this paper, we search the $d$ shortest source-to-destination paths for user request $\forall u_m \in U$ and use the Viterbi algorithm to place the VNFs for each path. the Viterbi algorithm can be viewed as a multi-stage graph for the states and their relationships and its aim is to find the most likely sequence of the states. The number of stages is equal to the length of an observed event sequence. Each node in a stage represents a possible state with a fixed cost for the current observed event and each stage in a graph consists of all possible states for the current observed event. The weighted edge between two states from adjacent stages represents a transmission cost. We can compute the cumulative cost of each state by the Viterbi algorithm in an increasing stage order, where the cumulative cost is composed of fixed costs and transmission costs. In the final stage, there are multiple possible states and each possible state corresponds to a path with a cumulative cost. We select a state path with minimum cumulative cost as the most likely path and obtain the best sequence by tracing the path. As the number of observed events and states increases the search space becomes large. Therefore, we can cut the search tree width to reduce the computational complexity.
\begin{algorithm}[tbp]
\caption{Viterbi Algorithm.}
\label{Viterbi Algorithm}
\hspace*{0.02in} {\bf Input:} $u_m,p$;\\
\hspace*{0.02in} {\bf Output:} $a'_m(k,p)$;
\begin{algorithmic}[1]
\STATE \textbf{Initialize:} $a'_m(k,p)=\varnothing,A_{layer}=\varnothing$;
\STATE Obtain topological sort sequence $\Phi_{m}$ for the VNFs;
\FOR {each $f_{m,i} \in \Phi_{m}$}
\STATE ${A}'_{layer}=\varnothing$;
\IF {$f_{m,i}=s_m$}
\STATE $A_{layer}\leftarrow$ configure information about $s_m$;
\STATE continue;
\ENDIF
\FOR {$\forall a_{m,i}^{pre}(k) \in A_{layer}$}
\STATE Update network resource conditions under $a_{m,i}^{pre}(k)$;
\STATE Obtain the set $\Omega_m$ of available satellites for path $p$;
\FOR {each $u \in \Omega_m$}
\STATE Deploy $f_{m,i}$ to satellite $u$ by strategy $a_{m,i}^{curr}(k)$;
\IF {the constraints in \eqref{equation11}-\eqref{equation18} are satisfied}
\STATE ${A}'_{layer}\leftarrow \left [a_{m,i}^{pre}(k), a_{m,i}^{curr}(k)\right ]$;
\ENDIF
\ENDFOR
\ENDFOR
\STATE List ${A}'_{layer}$ by user payoffs in descending order.
\STATE ${A_{layer}\leftarrow A}'_{layer}[:B]$;
\ENDFOR
\STATE Obtain strategy $a'_m(k,p)$ with maximum user payoff;
\RETURN $a'_m(k,p)$;
\end{algorithmic}
\end{algorithm}
For finding an approximate solution of the VNF placement by the Viterbi algorithm, we construct the VNF placement states and their relationships as a multi-stage graph. For a user request, the SFC is considered as an observed event sequence, each VNF indicates an observed event and the number of the VNFs represents the number of stages. The strategy of deploying a VNF indicates a possible state in each stage and there has the VNF deployment cost accordingly. The path between two adjacent VNFs indicates the edge between two states from two adjacent stages and the path cost is equal to the edge cost. Thus, we can perform the Viterbi algorithm to obtain an approximate strategy of VNF placement for a user request. Fig.~\ref{Example of placing VNFs for a user request by the Viterbi algorithm} illustrates an example of placing VNFs for a user request by the Viterbi algorithm, where the approximate VNF placement strategy is indicated with a red line.
The Viterbi algorithm for the VNF placement is shown in Algorithm \ref{Viterbi Algorithm}. The input parameters are user request $u_m$ and path $p$. The output parameter is an approximate strategy $a'_m(k,p)$ for path $p$. Initially, we denote a state set of the first stage by $A_{layer}=\varnothing$ and set $a'_m(k,p)=\varnothing$. Then we can obtain the ordered VNF sequence $\Gamma_m$ by a topological sorting method. For each stage of VNF $f_{m,i}\in \Gamma_m$, we search all possible VNF placement states and calculate their cumulative costs. respectively. If $f_{m,i}=s_m$, we can directly update the configure information concerning $s_m$ to $A_{layer}$. If $f_{m,i} \neq s_m$, for $\forall a_{m,i}^{pre}(k) \in A_{layer}$, we first update network resource conditions by $a_{m,i}^{pre}(k)$ and obtain a set $\Omega_{m,i}$ of current available satellites for path $p$. Within the network resource and service requirement constraints, we deploy VNF $f_{m,i}$ to satellite $u \in \Omega_{m,i}$ by strategy $a_{m,i}^{curr}(k)$. When the constraints in equations \eqref{equation11}-\eqref{equation18} are satisfied, $a_{m,i}^{curr}(k)$ will be put into a new strategy set $A'_{layer}$, which is initialized as $A'_{layer}=\varnothing$ at the beginning of each stage. When each stage is over, we will sort $A'_{layer}$ by user payoffs in descending order and use the first $B$ paths from $A'_{layer}$ to update $A_{layer}$. If all VNFs are deployed we can obtain an approximate strategy $a'_m(k,p)$ with maximum user payoff.
For Algorithm \ref{Resource Allocation Based on a Potential Game}, the computation complexity can be indicated as $O(K_{max}Md)$, where $K_{max}$ is the maximum number of iterations, $M$ is the number of players for the current time slot, and $d$ represents the size of a candidate path set. When we perform the Viterbi algorithm to deploy the VNFs to satellite nodes, the search tree width is $B$, the number of satellite nodes for an available path is less than $N$, and the maximum number of VNFs for a user request is $F$. The computation complexity of Algorithm \ref{Viterbi Algorithm} can be described as $O(FBN)$.
\section{Performance Evaluation}\label{Performance Evaluation}
In this section, we make the experiments to evaluate the performance of the proposed PGRA algorithm for addressing the VNF placement problem in satellite edge computing. We setup the system parameters for performance evaluation. In order to investigate the effects of system parameters on the performance of the proposed PGRA algorithm, we design the experiments by the Taguchi method with two factors, which are the shortest paths $d$ between the source and the destination for a user request and the width $B$ of the Viterbi search tree. Furthermore, we compare the proposed PGRA algorithm with two existing centralized baseline algorithms of Viterbi \cite{7469866} and Greedy \cite{7332796} in terms of network payoff and percentage of allocated users. Finally, we discuss the performance of the proposed PGRA algorithm in on-line strategy.
\subsection{Simulation Setup}
\begin{figure*}[tbp]
\centering
\subfigure[Main effects for $M=10$]{\includegraphics[width=0.3\textwidth]{Fig/5_a.eps}
\label{Main effects for 30 user requests}}
\subfigure[Main effects for $M=20$]{\includegraphics[width=0.3\textwidth]{Fig/5_b.eps}
\label{Main effects for 30 user requests}}
\subfigure[Main effects for $M=30$]{\includegraphics[width=0.3\textwidth]{Fig/5_c.eps}
\label{Main effects for 30 user requests}}
\caption{Main effects of two factors for different number $M=10,20,$ and $30$ of user requests.}
\label{Main effects of two factors for different number of user requests}
\end{figure*}
In the simulation experiments, we set the number of LEO satellite nodes by 6, 9, 12, and 15, respectively, where there are 3 orbital planes and each plane consists of 2, 3, 4, and 5 satellite nodes accordingly. There is an edge server on each satellite node. For each edge server, the resource capacities are configured as 112 vCPUs and 192 GB Memory, we consider the idle power as 49.9 W and the maximum active power as 415 W \cite{2019SPECpower}. We assume that the maximum idle time interval is 3 time slots and the minimum off time interval is 1 time slot. In addition, we assume that the inter-satellite link distance for the same orbital plane is 600 km and for the different orbital planes is 400 km, respectively. The bandwidth capacity for each inter-satellite link is 100 Mbps.
\begin{table}[tbp]
\centering
\caption{Parameters for Taguchi Method.}
\label{Parameters for Taguchi Method
\resizebox{\columnwidth}{!}{
\begin{tabular}{p{0.08\textwidth}<{\centering}p{0.07\textwidth}<{\centering}p{0.07\textwidth}<{\centering}p{0.07\textwidth}<{\centering}p{0.07\textwidth}<{\centering}}
\hline
\multirow{2}{*}{Factor} & \multicolumn{4}{c}{Level} \\
\cline{2-5} & 1 & 2 & 3 & 4 \\
\hline
d & 1 & 2 & 4 & 8 \\
B & 1 & 2 & 4 & 8 \\
\hline
\end{tabular
}
\end{table
\begin{table}[tbp]
\centering
\caption{Orthogonal Table $L_{16}(4^2)$ and Network Payoff Results.}
\label{Orthogonal Table and Network Payoff Results
\resizebox{\columnwidth}{!}{
\begin{tabular}{p{0.05\textwidth}<{\centering}p{0.04\textwidth}<{\centering}p{0.04\textwidth}<{\centering}p{0.07\textwidth}<{\centering}p{0.07\textwidth}<{\centering}p{0.07\textwidth}<{\centering}}
\hline
\multirow{2}{*}{Number} & \multicolumn{2}{c}{Factor} & \multicolumn{3}{c}{Network Payoff} \\
\cline{2-6} & d & B & \multicolumn{1}{c}{M=10} & \multicolumn{1}{c}{M=20} & \multicolumn{1}{c}{M=30} \\
\hline
0 & 1 & 1 & 9.1585 & 15.1844 & 16.1857 \\
1 & 1 & 2 & 9.3734 & 15.1888 & 16.2896 \\
2 & 1 & 4 & 9.3748 & 15.2886 & 16.1899 \\
3 & 1 & 8 & 9.3751 & 15.2891 & 16.1899 \\
4 & 2 & 1 & 9.2622 & 15.3828 & 16.3832 \\
5 & 2 & 2 & 9.3822 & 15.5843 & 16.3890 \\
6 & 2 & 4 & 9.3832 & 15.5844 & 16.4879 \\
7 & 2 & 8 & 9.3846 & 15.5844 & 16.4879 \\
8 & 4 & 1 & 9.2680 & 15.4805 & 16.2845 \\
9 & 4 & 2 & 9.3916 & 15.6820 & 16.3895 \\
10 & 4 & 4 & 9.3825 & 15.6821 & 16.4884 \\
11 & 4 & 8 & 9.3838 & 15.6814 & 16.4884 \\
12 & 8 & 1 & 9.2680 & 15.5785 & 16.2846 \\
13 & 8 & 2 & 9.3919 & 15.8785 & 16.4872 \\
14 & 8 & 4 & 9.3828 & 15.6819 & 16.5862 \\
15 & 8 & 8 & 9.3840 & 15.6812 & 16.5862 \\
\hline
\end{tabular
}
\end{table
In order not to lose generality, we randomly generate the SFC and resource requirements for each user request. The number of VNFs is from 5 to 10. Each VNF, except source and destination, has a predecessor and a successor, and requires $\left[4,8\right]$ vCPUs and $\left[4 GB,16 GB \right]$ Memory, respectively. The execution time for each VNF is $\left[10,30\right]$ ms. The bandwidth demand between two adjacent VNFs for routing traffic flows is $\left[10,30\right]$ Mbps. The source and the destination are randomly generated from the set of satellite nodes and can be known in advance. In addition, we define the weighted values in equation \eqref{equation9} as $\alpha_{1}=\alpha_{2}=\alpha_{3}=\frac{1}{3}$. Table \ref{Simulation Parameters} summarizes the main simulation parameters in our evaluation. We use a commodity server to be the simulation platform, where the configuration information is i7-4790K CPU, 16 GB Memory, and Windows 10. The programming language is PYTHON.
\subsection{System Parameters Evaluation}
For the proposed PGRA algorithm, the performance results of addressing the VNF placement problem can be influenced by two important parameters, which are the shortest paths $d$ between the source and the destination for a user request and the width $B$ of the Viterbi search tree, respectively. Consequently, in a satellite network with 6 satellite nodes, we use the Taguchi method of design-of-experiment (DOE) \cite{2013An} to investigate the simulation results under different values of $d$ and $B$ \cite{2015A}. The two factors are denoted by $d$ and $B$, respectively, where each factor includes 4 levels, e.g., 1, 2, 4, and 8. The parameters for the Taguchi method \cite{2013An} are shown in Table \ref{Parameters for Taguchi Method}. The orthogonal table $L_{16}(4^2)$ consists of 16 instances and we run each instance, for $M=10$, 20, and 30, 10 times to obtain the average network payoffs, respectively. Table \ref{Orthogonal Table and Network Payoff Results} describes the orthogonal table and network payoff results in our simulation. The main effects of the two factors for different user requests are illustrated in Fig. \ref{Main effects of two factors for different number of user requests}. We can find from Fig. \ref{Main effects of two factors for different number of user requests} that the network payoff results for $M=10$, 20, and 30 are better as parameters $d$ and $B$ increase. Large $d$ can improve the exploration ability of the proposed PGRA algorithm by searching for a larger solution space. For the Viterbi algorithm, large $B$ can keep more possible states of the current stage to the next search stage and also promote the network performance. However, large $d$ and $B$ can lead to a high computational complexity of the proposed PGRA algorithm. When $d$ and $B$ are small, the performance of the proposed PGAR algorithm will degrade. Therefore, their values should be considered in a tradeoff way. According to the simulation results of the Taguchi method, we can find that the ideal values of $d$ and $B$ are 8 and 4, respectively.
\begin{figure*}[tbp]
\centering
\subfigure[Average energy cost]{\includegraphics[width=0.3\textwidth]{Fig/6_a.eps}
\label{Average energy cost}}
\subfigure[Average bandwidth cost]{\includegraphics[width=0.3\textwidth]{Fig/6_b.eps}
\label{Average bandwidth cost}}
\subfigure[Average service delay cost]{\includegraphics[width=0.3\textwidth]{Fig/6_c.eps}
\label{Average service delay cost}}
\subfigure[Average network payoff]{\includegraphics[width=0.3\textwidth]{Fig/6_d.eps}
\label{Average network payoff}}
\subfigure[Average percentage of allocated users]{\includegraphics[width=0.3\textwidth]{Fig/6_e.eps}
\label{Average percentage of allocated users}}
\caption{Performance comparison with PGRA, Viterbi, and Greedy in a satellite network with 6 satellite nodes.}
\label{Performance comparison with PGRA, Viterbi, and Greedy in a satellite network with 6 satellite nodes}
\end{figure*}
\subsection{Performance Comparison with Baseline Algorithms}
To further discuss the effectiveness of the proposed PGRA algorithm in terms of energy consumption, bandwidth, and service delay, we compare the proposed PGRA algorithm with two existing baseline algorithms of Viterbi \cite{7469866} and Greedy \cite{7332796}. Based on the parameter analysis of the Taguchi method, the values of $d$ and $B$ are set as 8 and 4, respectively. A satellite network with 6 satellite nodes is used to run the experiment $L=\left \{L_1,L_2,\cdots,L_{10}\right \}$, which is composed of 10 group experiments and the corresponding number of user requests is denoted by $\left \{5,10,\cdots,5*i,\cdots,50\right \}$. Each experiment is run 10 times and the average results are obtained.
The experiment results for the performance comparison with three optimization algorithms are shown in Fig.~\ref{Performance comparison with PGRA, Viterbi, and Greedy in a satellite network with 6 satellite nodes}. Fig.~\ref{Average energy cost} illustrates the average energy costs for different number of user requests. We can find from Fig.~\ref{Average energy cost} that the energy costs obtained by the proposed PGRA algorithm are better than that of Viterbi and Greedy when there is a small number of user requests, e.g., $M=5$, 10, and 15. For instance $L_2$ with 10 user requests, the average energy costs for PGRA, Viterbi, and Greedy are 0.6833, 0.7666, and 0.7833, respectively. The average energy cost obtained by the proposed PGRA algorithm reduces by $10.87\%$ for Viterbi and $12.77\%$ for Greedy. However, as the number of user requests increases, the energy costs are increasing until they reach the maximum values. Then new user requests can not be deployed to satellite nodes due to the limitation of network resource capacities.
The average bandwidth costs for deploying different user requests to satellite nodes are described in Fig.~\ref{Average bandwidth cost}. We can observe from Fig.~\ref{Average bandwidth cost} that the proposed PGRA algorithm performs better than the two baseline algorithms of Viterbi and Greedy. For the small number of user requests, such as $M=5$, 10, and 15, the performance of the proposed PGRA algorithm is slightly better than that of Viterbi and Greedy. For an example of $M=10$, the average bandwidth costs are 0.1560, 0.1711, and 0.1915 for PGRA, Viterbi, and Greedy, respectively. The performance improvement of the proposed PGRA algorithm is $8.82\%$ for Viterbi and $18.53\%$ for Greedy. As the number of user requests increases, we can observe that the performance differences between the proposed PGRA algorithm and the two baseline algorithms are also increasing. In the case of $M=35$, the average bandwidth costs for PGRA, Viterbi, and Greedy are 0.1756, 0.3581, and 0.3923, respectively. We can observe that the performance of the proposed PGRA algorithm improves by $50.96\%$ for Viterbi and $55.23\%$ for Greedy. That is due to the fact that the players in a potential game want to make decisions by competing with each others for optimizing their objectives in a self-interested behavior. Thus, under the limitation of network resource capacities, the players, which have better strategies of placing the VNFs, can win the opportunities of updating their strategy information. In our experiments, when there are enough user requests to require available network resources, these resource requirements are greater than the network resource capacities and the energy costs produced by the user requests tend to constant values. Therefore, the used bandwidth resources can be considered as an important optimization objective. For the 10 group experiments, the average performance improvement of the proposed PGRA algorithm is $40.05\%$ for Viterbi and $47.93\%$ for Greedy. In Fig.~\ref{Average service delay cost}, Average service delay costs for different user requests are indicated. Similar to the average bandwidth costs in Fig.~\ref{Average bandwidth cost}, we can observe that the average service delay costs obtained by the proposed PGRA algorithm are better than that of the two baseline algorithms. On average, the service delay costs for PGRA, Viterbi, and Greedy are 0.9734, 0.9811, and 0.9810, respectively. The performance of the proposed PGRA algorithm improves by $0.78\%$ over both Viterbi and Greedy.
\begin{figure*}[tbp]
\centering
\subfigure[Satellite network with $N=9$]{\includegraphics[width=0.3\textwidth]{Fig/7_a.eps}
\label{Satellite network with N=9}}
\subfigure[Satellite network with $N=12$]{\includegraphics[width=0.3\textwidth]{Fig/7_b.eps}
\label{Satellite network with N=12}}
\subfigure[Satellite network with $N=15$]{\includegraphics[width=0.3\textwidth]{Fig/7_c.eps}
\label{Satellite network with N=15}}
\caption{Network payoffs for PGRA, Viterbi, and Greedy in satellite networks with $N=9$, 12, and 15.}
\label{Network payoffs for PGRA, Viterbi, and Greedy in satellite networks}
\end{figure*}
Fig.~\ref{Average network payoff} describes the average network payoffs for different number of user requests. It can be found from Fig.~\ref{Average network payoff} that the proposed PGRA algorithm outperforms the two baseline algorithms of Viterbi and Greedy in all the experiments. For an example of $M=15$, the network payoffs for PGRA, Viterbi, and Greedy are 13.7703, 13.0588, and 12.6343, respectively, and the result obtained by the proposed PGRA algorithm is over $5.44\%$ for Viterbi and $8.99\%$ for Greedy. Overall, the average performance improvement of the proposed PGRA algorithm is $5.16\%$ for Viterbi and $6.15\%$ for Greedy, respectively. In Fig.~\ref{Average percentage of allocated users}, we illustrate the average percentages of allocated user requests. We can find from Fig.~\ref{Average percentage of allocated users} that all user requests can be deployed to satellite nodes when the number of user requests is small, e.g., $M=10$. As $M$ increases, the percentage of allocated user requests begins to decrease due to the resource limitation of a satellite network. In these cases, the proposed PGRA algorithm also outperforms Viterbi and Greedy in terms of the percentage of allocated user requests. On average, the performance of the proposed PGRA algorithm is better $3.18\%$ than Viterbi and $4.60\%$ than Greedy. From Fig.~\ref{Performance comparison with PGRA, Viterbi, and Greedy in a satellite network with 6 satellite nodes}, we can demonstrate the effectiveness of the proposed PGRA algorithm compared with two baseline algorithms of Viterbi and Greedy, meanwhile, it is shown that the proposed PGRA algorithm outperforms Viterbi and Greedy for deploying user requests to satellite nodes.
In order to further evaluate the performance of the proposed PGRA algorithm in different scale satellite networks, three satellite networks, which consist of $N=9$, 12, and 15 satellite nodes, are used for conducting the experiments with user requests $\left \{5,10,15,20,25,30,35,40\right \}$, respectively. Each experiment is run 10 times and we obtain the average results. Fig.~\ref{Network payoffs for PGRA, Viterbi, and Greedy in satellite networks} shows the network payoffs for different user requests in three satellite networks with $N=9$, 12, and 15, respectively. In Fig.~\ref{Satellite network with N=9}, we describe the network payoffs for PGRA, Viterbi, and Greedy in the satellite network with $N=9$. Overall, we can find that the proposed PGRA algorithm performs better than Viterbi and Greedy. When the number of user requests is small, the performance of the proposed PGRA algorithm is close to that of Viterbi and over that of Greedy. As the number of user requests increases, the network payoff achieved by the proposed PGRA algorithm is more than that of Viterbi and Greedy. For example, when $M=10$, the network payoffs for PGRA, Viterbi, and Greedy are 9.4731, 9.4428, and 9.1489, respectively. The performance improvement of the proposed PGRA algorithm is $0.32\%$ for Viterbi and $3.54\%$ for Greedy. When $M=25$, the network payoffs for PGRA, Viterbi, and Greedy are 23.0444, 22.2306, and 21.9208, respectively. The proposed PGRA algorithm performs better $3.66\%$ than Viterbi and $5.12\%$ than Greedy. We can also find that the network payoffs begin to be stable when the number of user requests is greater than $25$. That is due to the fact that more user requests can not be deployed to satellite nodes as the resource limitation of a satellite network. On average, the performance of the proposed PGRA algorithm improves by $2.78\%$ for Viterbi and $3.11\%$ for Greedy. Fig.~\ref{Satellite network with N=12} and Fig.~\ref{Satellite network with N=15} illustrate the network payoffs for PGRA, Viterbi, and Greedy in satellite networks with $N=12$ and 15, respectively. Similar to the results in Fig.~\ref{Satellite network with N=9}, the proposed PGRA algorithm outperforms Viterbi and Greedy in satellite networks with $N=12$ and 15. For $N=12$, the network payoff obtained by the proposed PGRA algorithm increases by $0.96\%$ for Viterbi and $2.08\%$ for Greedy. For $N=15$, the network payoff obtained by the proposed PGRA algorithm increases by $0.65\%$ for Viterbi and $2.64\%$ for Greedy. In addition, as the increasing number in satellite nodes, a satellite network can provide more available network resources for user requests. Thus, more user requests can be placed to satellite nodes and that can lead to an increase in network payoff. For an instance of $M=35$, the network payoffs obtained by the proposed PGRA algorithm are 24.8568, 32.4339, and 34.2962 for satellite networks with $N=9$, 12, and 15, respectively. The network payoff obtained by the proposed PGRA algorithm for the satellite network with $N=15$ is over $37.97\%$ for the satellite network with $N=9$ and $5.74\%$ for the satellite network with $N=12$.
\begin{figure*}[tbp]
\centering
\subfigure[Satellite network with $N=9$]{\includegraphics[width=0.3\textwidth]{Fig/8_a.eps}
\label{Percentages of allocated user requests in a satellite network with N=9}}
\subfigure[Satellite network with $N=12$]{\includegraphics[width=0.3\textwidth]{Fig/8_b.eps}
\label{Percentages of allocated user requests in a satellite network with N=12}}
\subfigure[Satellite network with $N=15$]{\includegraphics[width=0.3\textwidth]{Fig/8_c.eps}
\label{Percentages of allocated user requests in a satellite network with N=15}}
\caption{Percentages of allocated user requests for PGRA, Viterbi, and Greedy in satellite networks with $N=9$, 12, and 15.}
\label{Percentages of allocated user requests for PGRA, Viterbi, and Greedy in satellite networks}
\end{figure*}
We also provide the percentages of allocated user requests for PGRA, Viterbi, and Greedy in three satellite networks with $N=9$, 12, and 15, as shown in Fig.~\ref{Percentages of allocated user requests for PGRA, Viterbi, and Greedy in satellite networks}. Fig.~\ref{Percentages of allocated user requests in a satellite network with N=9} describes the percentages of allocated user requests for different number of user requests in the satellite network with $N=9$. We can find from Fig.~\ref{Percentages of allocated user requests in a satellite network with N=9} that the percentages of allocated user requests are close to 1 for small number of user requests, e.g., $M=5$, 10, and 15. For small $M$, the resource requirements of user requests are less than the network resource capacities and all of them can be deployed to satellite nodes. As the increasing number in user requests, the available network resources are less and less. When the resource requirements of user requests exceed the network resource capacities, the satellite network can not provide any available resources for new user requests and that can result in a decrease in the percentages of allocated user requests, as shown in Fig.~\ref{Percentages of allocated user requests in a satellite network with N=9}. However, for almost all experiments, the proposed PGRA algorithm performs better than Viterbi and Greedy. On average, the percentage of allocated user requests obtained by the proposed PGRA algorithm improves by $1.52\%$ for Viterbi and $2.56\%$ for Greedy. Fig.~\ref{Percentages of allocated user requests in a satellite network with N=12} and Fig.~\ref{Percentages of allocated user requests in a satellite network with N=15} illustrate the percentages of allocated user requests for different user requests in satellite networks with $N=12$ and 15, respectively, where similar results in Fig.~\ref{Percentages of allocated user requests in a satellite network with N=9} can be obtained. When the number of user requests is small, the satellite network can provide enough available resources for deploying user requests. All of user requests can be deployed to satellite nodes and the percentages of allocated user requests are close to $1$. As $M$ increases, the available resources of satellite nodes are gradually decreasing and the resource requirements of more user requests will be not satisfied. Therefore, the percentage of allocated user requests will reduce as $M$ increases. We can also find from Fig.~\ref{Percentages of allocated user requests in a satellite network with N=12} and Fig.~\ref{Percentages of allocated user requests in a satellite network with N=15} that the performance of the proposed PGRA algorithm is better than that of Viterbi and Greedy. For $N=12$, the performance improvement of the proposed PGRA algorithm is $0.43\%$ for Viterbi and $1.46\%$ for Greedy. For $N=15$, the proposed PGRA algorithm performs better $0.28\%$ than Viterbi and $1.69\%$ than Greedy. Besides, as $N$ increases, more resources of the satellite network are available for deploying user requests and thus the percentages of allocated use requests will increase. For $M=35$, the percentages of allocated user requests obtained by the proposed PGRA algorithm are 0.7314, 0.9485, and 1.0 for satellite networks with $N=9$, 12, and 15, respectively.\par
\subsection{Performance Analysis in On-line Strategy}
\begin{figure*}[tbp]
\centering
\subfigure[Average energy cost]{\includegraphics[width=0.3\textwidth]{Fig/9_a.eps}
\label{Average energy cost in on-line strategy}}
\subfigure[Average bandwidth cost]{\includegraphics[width=0.3\textwidth]{Fig/9_b.eps}
\label{Average bandwidth cost in on-line strategy}}
\subfigure[Average service delay cost]{\includegraphics[width=0.3\textwidth]{Fig/9_c.eps}
\label{Average service delay cost in on-line strategy}}
\subfigure[Average network payoff]{\includegraphics[width=0.3\textwidth]{Fig/9_d.eps}
\label{Average network payoff in on-line strategy}}
\subfigure[Average percentage of allocated users]{\includegraphics[width=0.3\textwidth]{Fig/9_e.eps}
\label{Average percentage of allocated users in on-line strategy}}
\caption{Performance comparison with PGRA, Viterbi, and Greedy in satellite networks with 6, 9, 12, and 15 satellite nodes, respectively.}
\label{Performance comparison with PGRA, Viterbi, and Greedy in satellite networks with different satellite nodes}
\end{figure*}
To evaluate the effectiveness of the proposed PGRA algorithm for addressing the VNF placement problem in dynamic environment, we make the following experiments in satellite networks with 6, 9, 12, and 15 satellite nodes, respectively. The total number of time slots for each instance is 50 and we randomly generate the number of user requests from 5 to 10 in each time slot. The running periods for user requests can be randomly selected from 1 to 4 time slots. When the running time for a user request is over, the resources used by the user request can free and be available for deploying new user requests in the next time slot. All experiments are performed for 10 times and we obtain the average results.
In the four satellite networks, the average experiment results for deploying user requests in on-line strategy are shown in Fig.~\ref{Performance comparison with PGRA, Viterbi, and Greedy in satellite networks with different satellite nodes}. Fig.~\ref{Average energy cost in on-line strategy} describes the average energy costs for different satellite networks in on-line strategy. We can find from Fig.~\ref{Average energy cost in on-line strategy} that overall the energy costs obtained by the proposed PGRA algorithm for satellite networks with $N=6$, 9, 12, and 15 are less than that of Viterbi and Greedy. For an example of $N=12$, the energy costs for PGRA, Viterbi, and Greedy are 0.6145, 0.6360, and 0.6335, respectively, the energy cost of the proposed PGRA algorithm decreases by $3.38\%$ for Viterbi and $2.99\%$ for Greedy. The average bandwidth costs for different satellite networks in on-line strategy are illustrated in Fig.~\ref{Average bandwidth cost in on-line strategy}. We can find that the bandwidth costs obtained by the proposed PGRA algorithm for $N=6$ and 9 are less than that of Viterbi and Greedy, however, for $N=12$ and 15, are more than that of Viterbi and less than that of Greedy. That is due to the fact that the number of satellite nodes used by user requests can have an impact on the strategy of deploying VNFs. When the number of used satellite nodes is small, adjacent VNFs for a user request may not be deployed to the same satellite node due to the available resource limitation. If adjacent VNFs are deployed two different satellite nodes there will lead to the bandwidth costs as a result of routing traffic flows. In Fig.~\ref{Average service delay cost in on-line strategy}, we describe the service delay costs for different satellite networks. Similar to the results in Fig.~\ref{Average bandwidth cost in on-line strategy}. For $N=12$ and 15, the proposed PGRA algorithm performs better than Viterbi and Greedy in terms of energy consumption, however, the delay costs obtained by the proposed PGRA algorithm are more than that of Viterbi and Greedy, as the result of the tradeoff between bandwidth usage and energy consumption. In addition, we can find from Fig.~\ref{Average service delay cost in on-line strategy} that the average delay cost can decrease as the number of satellite nodes increases. The delay cost obtained by the proposed PGRA algorithm is 0.9793 for $N=6$ and 0.9449 for $N=12$. As the increase in the number of satellite nodes, there are more available resources for allocating user requests. Therefore, the paths with lower delay costs can be chosen to route traffic flows for user requests.
Note that the proposed PGRA outperforms both Viterbi and Greedy for network payoffs and percentages of allocated users, where the average network payoffs and percentages of allocated user requests are shown in Fig.~\ref{Average network payoff in on-line strategy} and Fig.~\ref{Average percentage of allocated users in on-line strategy}, respectively. From Fig.~\ref{Average network payoff in on-line strategy}, we can find that the network payoffs obtained by the proposed PGRA algorithm for satellite networks with $N=6$, 9, 12, and 15 are better than that of Viterbi and Greedy. For these four satellite networks, the network payoff improvement of the proposed PGRA algorithm is $1.31\%$, $0.21\%$, $0.31\%$, and $0.42\%$ for Viterbi, and $2.92\%$, $1.45\%$, $0.73\%$, and $1.18\%$ for Greedy, respectively. Furthermore, we can also find that the network payoff can increase as the number of satellite nodes increases. That is the fact that as available resources in satellite networks increase more user requests can be deployed to satellite nodes. For example, the network payoffs obtained by the proposed PGRA algorithm, in satellite networks with $N=6$, 9, 12, and 15, are 11.5633, 14.3481, 19.8609, and 19.8865, respectively. The average percentages of allocated user requests are shown in Fig.~\ref{Average percentage of allocated users in on-line strategy}. The proposed PGRA algorithm performs better than Viterbi and Greedy. For satellite networks with $N=6$, 9, 12, and 15, the percentages of allocated user requests obtained by the proposed PGRA algorithm increase by $0.47\%$, $0.04\%$, $0.09\%$, and $0.18\%$ for Viterbi, and $0.93\%$, $0.34\%$, $0.16\%$, and $0.42\%$ for Greedy, respectively. As the number of satellite nodes increases, we can also find that the percentage of allocated user requests increases. For the proposed PGRA algorithm, the percentages of allocated user requests in satellite networks with $N=6$, 9, 12, and 15 are $96.01\%$, $99.50\%$, $99.96\%$, and $99.92\%$, respectively. From Fig.~\ref{Performance comparison with PGRA, Viterbi, and Greedy in satellite networks with different satellite nodes}, we can demonstrate the effectiveness of the proposed PGRA algorithm compared with Viterbi and Greedy in dynamic environment.
\section{Conclusion}\label{Conclusion}
In this paper, we study the VNF placement problem by a potential game in satellite edge computing. Our aim is to maximize the number of allocated user requests with minimum overall deployment cost, which include energy, bandwidth, and service delay costs. We prove the VNF placement problem to be NP-hard and formulate the problem as a potential game with maximum network payoff. Each user request can make the deployment decision in a self-interested way and all user requests can optimize their strategies by competing with each other in a distributed manner. Considering that a potential game admits at least a Nash equilibrium we implement a decentralized resource allocation algorithm based on a potential game to find an approximate solution.
To evaluate the effectiveness of the proposed PGRA algorithm, we first discuss the influence of system parameters on the proposed PGRA algorithm performance by the Taguchi method. Then we make the experiments for different number of user requests in satellite networks with 6, 9, 12, and 15 satellite nodes and compare the simulation results with two baseline algorithms of Viterbi and Greedy. For example, in the case of $N=12$, the average network payoff obtained by the proposed PGRA algorithm increases by $0.96\%$ for Viterbi and $2.08\%$ for Greedy, the average percentage of allocated users obtained by the proposed PGAR algorithm improves by $0.43\%$ for Viterbi and $1.46\%$ for Greedy. In addition, we investigate the performance of the proposed PGRA algorithm in dynamic environment. In dynamic environment, for $N=12$, the proposed PGRA algorithm improves by $0.31\%$ for Viterbi and $0.73\%$ for Greedy in terms of network payoffs, the percentage of allocated users obtained by the proposed PGAR algorithm increases by $0.09\%$ for Viterbi and $0.16\%$ for Greedy. All the simulation results show that the proposed PGRA algorithm is an effective approach for addressing the VNF placement in satellite edge computing and performs better than Viterbi and Greedy.
\ifCLASSOPTIONcaptionsoff
\newpage
\fi
\bibliographystyle{IEEEtran}
|
\section{Introduction}
Quantum tunneling, allowing an object to pass a potential energy barrier, even when
classically it does not have enough energy, is one of the most striking concepts of quantum physics.
In Nature, it produces energy in stars via nuclear fusion, it heats the Earth's interior by $\alpha-$decay,
it produces new elements in nucleosynthesis, and it is believed to cause DNA
mutations, ageing and cancer. Quantum tunneling also underpins many technological
applications such as the tunneling electron microscope, FLASH memory, tunnel junctions
in solar cells, and tunnel diodes in high speed devices.
Although well-understood for simple systems such as
electrons, the description of quantum tunneling for most systems (e.g., molecules, atomic
nuclei, Bose-Einstein condensates) remains challenging due to their composite
nature and the interaction between their constituents.
Several methods have been developed in recent years to describe tunneling of interacting particles.
Due to the complexity of the problem, these methods are either limited to few-body systems \cite{ahsan2010,hunn2013,rontani2013,lundmark2015,dobrzyniecki18} or they require approximations
to the quantum many-body dynamics \cite{umar2009,Hagino2012,simenel2013,fasshauer2016,wen2017,simenel2017,godbey2019b,lode2020}.
As a result, mean-field driven non-exponential decay \cite{zhao2017,alcala2017,potnis2017} as well as
substantial deviation from mean-field dynamics \cite{mclain18} were found in cold atoms systems.
Pairing effects have also been demonstrated in systems of two and three atoms \cite{zurn2012,zurn2013,gharashi2015,lundmark2015,dobrzyniecki18,dobrzyniecki19}.
Indeed, two atoms can tunnel as a correlated pair when the interaction is strong and attractive \cite{lundmark2015}.
Similar cluster tunneling is found in nuclear systems, e.g., in $\alpha-$decay.
Despite being composed of four strongly interacting nucleons, $\alpha-$clusters can be approximated
as inert particles due to the large difference between their ground-state and first excited state,
thus reducing theoretical description of $\alpha-$decay to a single-particle tunneling problem.
However, the situation is much more complicated with heavier systems which can encounter
non-trivial many-body dynamics while tunneling, as, e.g., in spontaneous fission and in low-energy heavy-ion fusion reactions.
Unlike cold atoms which tunnel sequentially or as small clusters, all nucleons are usually tunneling together as self-bound systems in such reactions.
However, unlike in $\alpha-$decay, individual nucleonic degrees of freedom need to be accounted for.
Indeed, microscopic nuclear dynamics calculations (see \cite{simenel2018} for a review) show that fusion
is affected by nucleon transfer \cite{vo-phuoc2016,godbey2017}
leading to dissipation \cite{Williams2018,simenel2020} and potentially decoherence \cite{dasgupta2007} in many-body tunneling.
At present, there is no model of quantum many-body tunneling for strongly interacting
systems such as nuclei that explicitly accounts for effects of
dissipation or decoherence that are induced by nucleonic degrees of
freedom.
Nevertheless, early works based on instantons and path integral description of quantum mechanics
opened the possibility for mean-field description of tunneling based on imaginary-time techniques \cite{Levit1980a,Reinhardt1981,pudu1987,Arve1987,Negele1989,Skalski2008}.
Direct implementation of this method, however, is challenging due to the difficulty to find quantum many-particle closed trajectories in imaginary
time.
To overcome this difficulty, we propose an initial value approach to describe tunneling through
an imaginary time mean-field
evolution akin to the way standard real-time mean-field evolution
techniques such as time-dependent Hartree-Fock (TDHF) can be used to study vibrations (without
requantization) (see, e.g., \cite{simenel2009,avez2013}).
This approach is tested in a simple toy-model where
an exact solution exists, and where the evolution in classically forbidden region can be easily visualised.
In particular, we show how tunneling probabilities can be extracted at the mean-field level.
These predictions are compared with the exact solution for various strength of the interaction.
The toy model has been chosen so that generalisations to more realistic systems are in principle feasible.
The two-well toy model is described in section~\ref{sec:toy}, where both exact and real-time mean-field dynamics are studied.
The path integral approach and its application within the toy model are discussed in section~\ref{sec:path}.
The method to compute tunneling probabilities is introduced in section~\ref{sec:proba}.
Potential extensions and applications to more realistic systems are discussed in section~\ref{sec:realistic}, before we conclude in section~\ref{sec:conclusions}.
\section{Two-well model\label{sec:toy}}
We consider a simple toy model with two interacting distinguishable particles evolving according to the Hamiltonian
\begin{equation}
\hat{H}(1,2)=\hat{h}_0(1)+\hat{h}_0(2)+\hat{v}(1,2).
\end{equation}
The single-particle Hamiltonian is written
$$\hat{h}_0=\alpha \ket{-}\bra{-},$$ where $\ket{-}$ is the excited state. Its ground-state $\ket{+}$ has energy 0.
This corresponds to a two-well potential with two possible positions, $\ket{L}$ and $\ket{R}$, for a particle in the left and right well, respectively. These states are related to the eigenstates of $\hat{h}_0$ by $\ket{\pm}=\frac{1}{\sqrt{2}}(\ket{L}\pm\ket{R})$.
In this model, a particle initially in one well can tunnel to the other well through a potential energy barrier decreasing with $\alpha$.
The interaction is assumed to occur when both particles are in the same well with $\hat{v}(1,2)=\mu\,(\ket{LL}\bra{LL}+\ket{RR}\bra{RR})$, where $\mu$ is a parameter controlling the interaction strength.
The initial state is chosen as $\ket{LL}$.
It is possible to extend this toy model by considering identical particles (bosons or fermions), adding more particles and modes, and using a more realistic interaction.
Nevertheless, despite its simplicity this model accounts for the essential aspects of tunneling of interacting particles.
Moreover, such simplifications allow for easier visualisation of the configuration space available to the system, and thus it gives us a better understanding of its dynamics.
Most importantly, this model is exactly solvable analytically, thus providing a benchmark to test various approximations.
\subsection{Exact dynamics}
The exact evolution is determined from the time-evolution operator $\hat{U}=e^{-i\hat{H}t}$ (we set $\hbar=1$).
The Hamiltonian of the two-well model is
\begin{eqnarray}
\hat{H}&=&\alpha\left(\ket{-}_1\bra{-}_1\mathbb{1}_2+\mathbb{1}_1\ket{-}_2\bra{-}_2\right)\nonumber\\
&&+\mu\left(\ket{LL}\bra{LL}+\ket{RR}\bra{RR}\right)\nonumber
\end{eqnarray}
with $\ket{-}=\frac{1}{\sqrt{2}}(\ket{L}-\ket{R})$. The indices refer to particles 1 and 2.
In the $\{L,R\}$ basis it is expressed as
\begin{equation}\hat{H}=\,\,\begin{blockarray}{cccc}
\ket{LL} & \ket{LR} & \ket{RL} & \ket{RR} \\
\begin{block}{(cccc)}
\alpha +\mu& -\alpha/2 & -\alpha/2 & 0 \\
-\alpha/2 & \alpha & 0 & -\alpha/2 \\
-\alpha/2 & 0 & \alpha & -\alpha/2 \\
0 & -\alpha/2 & -\alpha/2 & \alpha+\mu \\
\end{block}
\end{blockarray}\,\,\,\,\,.\nonumber\end{equation}
In this basis, the operator counting the particles in the left well is
\begin{equation}
\hat{N}_L= \begin{pmatrix}
2&0&0&0\\
0&1&0&0\\
0&0&1&0\\
0&0&0&0
\end{pmatrix}.\nonumber
\end{equation}
With the condition that the particles are initially in $\ket{LL}$, the state of the system at time $t$ is
$$\ket{\Psi}(t)=\exp(-i\hat{H}t)\ket{LL}.$$
The expectation value of $\hat{N}_L$ in this state can be expressed as
\begin{equation}\label{eq:Nlcount}
\langle\hat{N}_L\rangle=1+\frac{\beta-\mu}{2\beta}\cos(\frac{\beta+\mu}{2}t)+\frac{\beta+\mu}{2\beta}\cos(\frac{\beta-\mu}{2}t),
\end{equation}
where $\beta=\sqrt{4\alpha^2+\mu^2}$.
Note that $\langle\hat{N}_L\rangle$ does not depend on the sign of $\mu$, i.e., if the interaction is attractive ($\mu<0$) or repulsive ($\mu>0$).
From now on, we set $\alpha=1$, so that the only parameter is the interaction strength $\mu$.
The exact evolution $\langle\hat{N}_L\rangle(t)$ is represented in Fig. \ref{fig:2part} (blue solid line) for two values of $\mu$.
The oscillatory behaviour shows that the particles can tunnel from the left well ($\langle\hat{N}_L\rangle/2=1$) to the right well ($\langle\hat{N}_L\rangle/2=0$). The effect of increasing the interaction strength $\mu$ is to slow down tunneling as the particles are found in the right well at a later time.
\begin{figure}[t]
\includegraphics[width=8.5cm]{mfu}
\caption
Exact solutions (blue solid lines) and real-time mean field predictions (orange solid lines) of $\langle\hat{N}_L\rangle(t)/2$ for (a) a weak interaction with $|\mu|=1$ and (b) a strong interaction with $|\mu|=4$. Time unit is fixed by our choice of $\alpha=1$.}
\label{fig:2part}
\end{figure}
\subsection{Real time mean-field evolution}
Approximations to describe the dynamics of quantum many-body systems are often based on the self-consistent mean-field theory, or time-dependent Hartree theory in the case of distinguishable particles. The latter assumes that the particles remain independent at all times, i.e., with a state $\ket{\Psi}=\ket{\psi_1}\otimes\ket{\psi_2}$.
The Hamiltonian is then approximated by $\hat{H}_H(1,2)=\hat{h}_H(1)+\hat{h}_H(2)$ with time-dependent single-particle Hartree Hamiltonian
$$\hat{h}_H(i)=\hat{h}_0(i)+\bra{\psi_j}\hat{v}(1,2)\ket{\psi_j}$$
where $i,j=1,2$ and $j\ne i$.
Note that in this form there is no spurious self-interaction.
As both particles are initially in the same state, they encounter the same mean-field evolution.
As a result, they remain in identical states $\ket{\psi(t)}=L(t)\ket{L}+R(t)\ket{R}$ obeying the time-dependent Hartree equation
\begin{equation}
i\frac{d}{dt}\ket{\psi(t)}=\hat{h}_H(t) \ket{\psi(t)}.\label{eq:RTDH}
\end{equation}
The time-dependence of $\hat{h}_H(t)$ is due to its self-consistency.
The equations of motion for $L(t)$ and $R(t)$ then become
\begin{subequations}
\begin{align}
i\frac{d}{dt}L&=\frac{\alpha}{2}(L-R)+\mu L|L|^2 \mbox{ and}\label{eq:dL}\\
i\frac{d}{dt}R&=\frac{\alpha}{2}(R-L)+\mu R|R|^2.\label{eq:dR}
\end{align}
\end{subequations}
Let us define new coordinates $\theta\in[-\pi/2,\pi/2]$ and $\phi\in[-\pi,\pi]$:
\begin{equation}
\theta=\arcsin(|L|^2-|R|^2) \mbox{ and }\phi=\arg(R/L).
\label{eq:coordinates}
\end{equation}
This choice preserves the normalisation $|L|^2+|R|^2=1$.
Both sets of coordinates are equivalent up to a global phase.
From Eqs.~(\ref{eq:coordinates}) and the normalisation condition we can write
$$L=\sqrt{\frac{1+\sin\theta}{2}}e^{i\phi_L} \mbox{ and }
R=\sqrt{\frac{1-\sin\theta}{2}}e^{i\phi_R}$$
where we have introduced the phases $\phi_{L,R}$.
According to Eqs.~(\ref{eq:coordinates}), only their difference $\phi=\phi_R-\phi_L$ is relevant.
Inserting into Eqs.~(\ref{eq:dL}) and~(\ref{eq:dR}) gives
\begin{subequations}
\begin{align}
i\dot\theta\cos\theta-2(1+\sin\theta))\dot\phi_L=&1+\sin\theta-e^{i\phi}\cos\theta \nonumber\\
&+\mu(1+\sin\theta)^2\label{eq:d1}\\
-i\dot\theta\cos\theta-2(1-\sin\theta))\dot\phi_R=&1-\sin\theta-e^{-i\phi}\cos\theta \nonumber\\
&+\mu(1-\sin\theta)^2,\label{eq:d2}
\end{align}
\end{subequations}
where we have set $\alpha=1$.
Taking the imaginary part gives the first differential equation
\begin{equation}
\dot{\theta}=-\sin\phi\label{eq:thetadot}.
\end{equation}
Taking the real part of Eqs.~(\ref{eq:d1}) and~(\ref{eq:d2}) and rearranging gives the second differential equation
\begin{equation}
\dot{\phi}=\tan\theta\cos\phi+\mu\sin\theta.\label{eq:phidot}
\end{equation}
Equations (\ref{eq:thetadot}) and (\ref{eq:phidot}) provide a closed set of equations for the real-time mean-field dynamics of the system.
Solving these equations numerically with initial condition $(\theta,\phi)=(\frac{\pi}{2},0)$ corresponding to both particles in the left well,
we get the real-time mean-field prediction for $\langle N_L \rangle(t) = 1+\sin\theta(t)$ plotted in Fig.~\ref{fig:2part} (orange solid line).
As in the exact case, the latter does not depend on the sign of $\mu$.
Apart for the non-interacting case $\mu=0$ (for which mean-field dynamics is obviously exact), we see that mean-field predictions rapidly deviate from the exact solution.
Although for $|\mu|\leq2$, which we loosely refer to as the ``weakly'' interacting regime, tunneling is observed in the mean-field solution, a transition appears at $|\mu|=2$, above which (``strongly'' interacting regime) the particles are ``trapped" in one well, unable to tunnel completely to the other well. This spurious phenomenon, called ``discrete self-trapping" \cite{Eilbeck1985,Smerzi1997,Milburn1997a}, illustrates the inability of real-time mean-field theory to describe tunneling dynamics in strongly interacting systems.
\begin{figure*}
\centering
\includegraphics[width=18cm]{energy_contours}
\caption
Mean-field energy contours are shown for (a) free particles, (b) a weak attraction, and (c) a strong attraction. Energy increases from blue to yellow.
Solid red lines show mean-field trajectories starting from the left well. The dashed lines represent constant values of $\phi$.}
\label{fig:enconts}
\end{figure*}
\subsection{Hartree energy}
In order to investigate the origin of self-trapping, let us first determine the energy of the system in the mean-field approximation.
Without interaction, the total energy is given by
\begin{eqnarray}
K&=&\sum_{i=1}^{2}\bra{\psi_i}\hat{h}_0\ket{\psi_i}=
\begin{pmatrix}L^*&R^*\end{pmatrix}
\begin{pmatrix}
1&-1\\
-1&1
\end{pmatrix}
\begin{pmatrix}L\\R\end{pmatrix}\nonumber\\
&=&1-2\text{Re}[L^*R]=1-\cos\theta\cos\phi,\nonumber
\end{eqnarray}
where we have used $\ket{\psi_{1,2}}=L\ket{L}+R\ket{R}$, Eqs.~(\ref{eq:coordinates}) and
\begin{equation}\hat{h}_0=\frac{\alpha}{2}
\begin{pmatrix}
1 & -1 \\
-1 & 1 \\
\end{pmatrix}\nonumber
\end{equation}
in the $\{\ket{L},\ket{R}\}$ basis, with our choice of $\alpha=1$.
With an interaction treated at the mean-field level, an additional term $U=\frac{1}{2}\sum_{i,j\ne i}\bra{\psi_i\psi_j}\hat{v}(1,2)\ket{\psi_i\psi_j}$ contributes to the total energy $E=K+U$.
Using $\hat{v}(1,2)=\mu\,(\ket{LL}\bra{LL}+\ket{RR}\bra{RR})$, we get
\begin{eqnarray}
U&=&\mu\begin{pmatrix}L^*&R^*\end{pmatrix}\begin{pmatrix}
|L|^2&0\\
0&|R|^2
\end{pmatrix}\begin{pmatrix}L\\R\end{pmatrix}\nonumber\\
&=&\mu(|L|^4+|R|^4)
=\frac{\mu}{2}\left[1+\sin^2\theta\right].\nonumber
\end{eqnarray}
As a result, the total (Hartree) energy is expressed as
\begin{equation}
E=1+\frac{\mu}{2}(1+\sin^2\theta)-\cos\theta\cos\phi\label{eq:totalen}
\end{equation}
which is conserved under Eqs. (\ref{eq:thetadot}) and (\ref{eq:phidot}).
\subsection{Self-trapping}
The self-trapping phenomenon can be understood by examining the mean-field dynamics of the system in the configuration space.
The latter is entirely defined by the coordinates $\theta$ and $\phi$, allowing for a simple two-dimensional representation as in Figure~\ref{fig:enconts}(a) showing the position of the $\ket{L}$, $\ket{R}$ and $\ket{\pm}$ states.
Figuress~\ref{fig:enconts}(a-c) show contour plots of the Hartree energy in configuration space
for various interaction strengths.
As in the exact case, mean-field dynamics conserve total energy. The system is thus bound to follow iso-energy contour lines.
We see that with no (Fig.~\ref{fig:enconts}(a)) or ``weak'' attraction (Fig.~\ref{fig:enconts}(b)), the system is able to go from one well to the other following a classically allowed path (thick solid red lines).
However, in the case of ``strong'' attraction (Fig.~\ref{fig:enconts}(c)), there is no iso-energy contour line connecting both wells. The transition is classically forbidden, preventing tunneling and leading to self-trapping.
Self-trapping occurs when the states at $\theta=\pm\pi/2$ are not connected by any energy contour line in the $(\theta,\phi)$ plane.
According to Eq.~(\ref{eq:totalen}), the energy at $\theta=\pm\pi/2$ is
\begin{equation}
E_1=E\big|_{\theta=\pm\pi/2}=1+\mu,\nonumber
\end{equation}
while at $\theta=0$ it is
\begin{equation}
E_2(\phi)=E\big|_{\theta=0}=1+\frac{\mu}{2}-\cos\phi.\nonumber
\end{equation}
A condition for the system to tunnel from one well to another in the realtime mean-field dynamics is that there exist a $\phi$ for which $E_2(\phi)=E_1$
(otherwise the system is unable to cross the $\theta=0$ line).
This is only possible for $\mu/2=-\cos\phi$, leading to the condition $|\mu|\leq 2$.
Self-trapping then occurs when $|\mu|\geq 2$.
This condition does not depend on the sign of $\mu$. Thus, despite the fact that the contour lines are different for attractive or repulsive interactions, self-trapping occurs in both cases for the same magnitude of the interaction strength.
\subsection{``Weak'' and ``strong'' interactions}
Some comments are in order regarding the distinction between ``weakly''
and ``strongly'' interacting regimes.
The situation in terms of realistic applications will depend on the system.
For cold atoms, the interaction can be tuned experimentally and the
``transition'' region could then be explored.
For nuclear systems, in particular in the case of fusion, fission and cluster
decay, the systems are expected to be well in the ``strong'' interaction
regime.
In our toy model, the energy splitting $\Delta E$
between the ground and first excited state of the exact Hamiltonian is the quantity that drives the tunneling rate.
Arve {\it et al.} used
a two-well potential with parameters adjusted to describe a typical fission
problem, with energy splittings between the quasi-degenerate eigenstates
of the order of $10^{-13}$~MeV for the ground-state, up to $10^{-2}$~MeV near the
barrier \cite{Arve1987}. In the two-well model we use, the energy splitting $\Delta E$
is $\sim\alpha^2/|\mu|$ for $|\mu|\gg \alpha$. The interaction $\mu$ is of the order
of the binding energy per nucleon ($\sim8$~MeV). To get similar splitting as
Arve et al, we would then set $\alpha\sim10^{-6}$~MeV for the ground state,
up to $\sim0.3$~MeV near the barrier. In any case, $\alpha$ remains
smaller than $2|\mu|$ (recall that the transition appears at $|\mu|=2\alpha$),
indicating a system clearly in the ``strong'' interaction regime.
\section{Path integral approach\label{sec:path}}
Our goal is now to search for an initial-value mean-field based description of the system which would account for tunneling in the strongly interacting regime.
Following Feynman's many-path approach to quantum mechanics~\cite{Feynman1948}, the amplitude of probability for the system to go from the state $\ket{i}$ at time $t_i$ to the state $\ket{f}$ at time $t_f$ is written as a path integral
\begin{equation}\bra{f}\hat{U}(t_f,t_i)\ket{i}=\int D[\sigma]\exp(iS[\sigma])
\nonumber\end{equation}
where $\hat{U}(t_f,t_i)$ is the evolution operator associated with the hamiltonian $\hat{H}$ of the system and $S[\sigma]$ is the action for the path $\sigma(t)$ in configuration space between $t_i$ and~$t_f$.
\subsection{Imaginary time-dependent mean-field equation}
Though elegant, this path integral approach is often too complicated in practice and requires approximations such as the stationary phase approximation (SPA).
For a single-particle following a classical path $\sigma(t)\equiv x(t)$,
the SPA leads to the stationary action principle $\delta{S}=0$ of classical mechanics in which quantum tunneling is forbidden.
Nevertheless, the latter can be recovered approximatively through a Wick rotation changing real time to imaginary time $t\rightarrow -i\tau$ \cite{Zichichi1977}.
Its effect is indeed to change the sign of the potential, thus allowing the system to explore classically forbidden regions.
This approach is formally equivalent to the WKB semiclassical approximation \cite{Holstein1982,Holstein1982a}.
For a many-particle system, mean-field equations can be recovered from the stationary action principle with the Dirac action
\begin{equation}
S=\int_{t_i}^{t_f}dt \bra{\Psi}i\partial_t-\hat{H}\ket{\Psi}
\label{eq:Dirac}
\end{equation}
while restricting the variational space to independent particle states.
However, as illustrated by our toy model, this theory does not account for tunneling in the strong interaction regime.
Nevertheless, applying a Wick rotation should produce imaginary-time mean-field equations accounting for tunneling.
Replacing $t\rightarrow-i\tau$ in Eq.~(\ref{eq:RTDH})
leads to an imaginary-time dependent Hartree equation
\begin{equation}
\frac{d}{d\tau}\ket{\psi(\tau)}=-\hat{h}_H(\tau) \ket{\psi(\tau)}
\end{equation}
As in real time evolution, computing observables $\langle\hat{Q}\rangle(\tau)$ also requires a conjugate state which we now define.
\subsection{Wave-function of conjugate state}
In real time, the wave-function of a single-particle conjugate state is given by $\tilde{\psi}(x,t)=\bra{\psi(t)}\ket{x}=\psi^*(x,t)$.
It is convenient to write the wave-function as
\begin{equation}
\psi(x,t)=\sqrt{\rho(x,t)}e^{i\phi(x,t)}, \label{eq:psit}
\end{equation}
with
\begin{subequations}
\begin{align}
\rho(x,t)&=\tilde{\psi}(x,t)\psi(x,t) \,\,\mbox{ and}\label{eq:rhot}\\
\phi(x,t)&=\frac{1}{2i}\ln\left(\frac{\psi(x,t)}{\tilde{\psi}(x,t)}\right), \label{eq:phit}
\end{align}
\end{subequations}
leading to $\tilde{\psi}(x,t)=\sqrt{\rho(x,t)}e^{-i\phi(x,t)}$.
Imaginary-time evolution is obtained from a Wick rotation $t\rightarrow-i\tau$.
This has a consequence for how the conjugate of a single-particle wave-function is defined.
The conjugate is used to compute expectation values
\begin{equation}
\langle\hat{Q}\rangle(\tau)=\int dx \,\,\,\tilde{\psi}(x,\tau) \,Q(x,y)\, \psi(x,\tau)\nonumber
\end{equation}
which are transformed under the Wick rotation as\footnote{\label{foot1}This expression is only correct for a time-independent Hamiltonian for which the imaginary-time evolution operator is given by $\exp(-\hat{H}\tau)$. For a time-dependent Hamiltonian, such as in the self-consistent mean-field approximation, it should be replaced by $T\exp[-\int_0^\tau\hat{H}(\tau)d\tau]$, where $T$ denotes time ordering.}
$$\langle\Psi(0)|e^{i\hat{H}t}\hat{Q}e^{-i\hat{H}t}|\Psi(0)\rangle\longrightarrow \langle\Psi(0)|e^{\hat{H}\tau}\hat{Q}e^{-\hat{H}\tau}|\Psi(0)\rangle.$$
The expectation value is then given by
\begin{equation}
\langle\hat{Q}\rangle(\tau)=\langle\Psi(-\tau)|\hat{Q}|\Psi(\tau)\rangle
\end{equation}
implying
\begin{equation}\tilde{\psi}(x,\tau)=\psi^*(x,-\tau).\label{eq:psitau}\end{equation}
In imaginary time, $\rho$ and $\phi$ become complex.
It is easy to show from Eqs.~(\ref{eq:psit}),~(\ref{eq:rhot}) and~(\ref{eq:phit}) that $\rho(\tau)=\rho^*(-\tau)$ and $\phi(\tau)=\phi^*(-\tau)$.
As a result, Eq.~(\ref{eq:psitau}) becomes
\begin{equation}
\tilde{\psi}(x,\tau)=\sqrt{\rho(x,\tau)}\exp(-i\phi(x,\tau)).
\end{equation}
As a result, the conjugate in imaginary time has the same structure as in real time.
\subsection{Classically forbidden region and final condition}
Computing expectation values of observables in imaginary time thus requires both forward and backward evolutions.
These are nevertheless initial value equations as only $|\Psi(0)\rangle$ is required to compute both evolutions in the classically forbidden region.
Criteria must be defined for where to ``stop'' the calculation.
One (or several) observable $O_f$ can be used to define such criteria.
The system needs then to be evolved in imaginary time until the condition (see footnote \ref{foot1})
$$O(\tau) = \langle\psi_0| e^{\hat{H}\tau} \hat{O} e^{-\hat{H}\tau} |\psi_0\rangle = O_f$$
is reached.
This defines the mean-field path from which the action and then the
probability (see Sec.~\ref{sec:proba}) to reach $\langle\hat{O}\rangle=O_f$ can be computed.
For instance, $\hat{O}$ could be
the quadrupole operator in fusion/fission problems, or the center of
mass in cluster decay (see section~\ref{sec:applications}).
Note that the final state is not necessarily in
the classically allowed region (in that case, however, a connection to
real time dynamics cannot be performed).
This is a major difference with earlier implementation of the imaginary-time mean-field approximation~\cite{Levit1980a}
which required bounce solutions with the condition
$$\ket{\Psi(\tau)}=\ket{\Psi(-\tau)}.$$
\subsection{Application to two-well model}
\begin{figure}[h!]
\includegraphics[width=7cm]{imag_path}
\caption{The real and imaginary-time mean-field energy is computed with $\mu=-3$ and is increasing from blue to yellow. The $(\theta_R,\phi_R)$ horizontal plane shows the energy for $\phi_I=\theta_I=0$.
The $(\theta_R,\phi_I)$ vertical plane shows the energy for $\phi_R=\theta_I=0$.
The red (respectively green) solid line shows an iso-energy contour connecting the left and right wells (resp. the left and right mean-field ground-states).}\label{fig:3contours}
\end{figure}
The equations of motion in imaginary time are obtained from a Witck rotation of Eqs.~(\ref{eq:thetadot}) and~(\ref{eq:phidot}):
\begin{equation}
\frac{d\theta}{d\tau}=i\sin\phi\,\,\,\mbox{ and }\,\,\,\frac{d\phi}{d\tau}=-i\tan\theta\cos\phi-i\mu\sin\theta.\label{eq:ITEOM}
\end{equation}
The coordinates $\theta=\theta_R+i\theta_I$ and $\phi=\phi_R+i\phi_I$ are now complex.
The total energy of the system in the imaginary-time-dependent Hartree theory is still conserved and given by Eq.~(\ref{eq:totalen}).
As a result, as long as the initial condition is in the classically allowed region, i.e., with $\theta_I(t_i)=\phi_I(t_i)=0$, this energy remains real.
This condition, together with the constant norm $\langle\psi(-\tau)|\psi(\tau)\rangle=1$ also impose relationships between $\theta$ and $\phi$.
Figure~\ref{fig:3contours} shows this energy for a strongly attractive system.
The horizontal plane gives the energy for real time evolution, as in Fig.~\ref{fig:enconts}.
The vertical plane represents the energy for imaginary time evolution with $\phi_R=0$.
It is now possible for the system to go from the left well to the right one following a combination of real and imaginary time evolutions (solid red line).
This demonstrates the ability of imaginary time mean-field evolution to explore classically forbidden region through quantum tunneling.
\section{Tunneling probability\label{sec:proba}}
\begin{figure}[htbp]
\includegraphics[width=\columnwidth]{proba}
\caption
Exact (solid line) and imaginary-time-dependent mean-field (symbols) tunneling probabilities per unit of time (in units of $\alpha/\hbar$) are plotted as a function of the attraction strength $-\mu$ (in units of $\alpha$).\label{fig:exactandapprox} Error bars correspond to numerical uncertainty of 1\% in the action.
}
\end{figure}
Now that we found mean-field tunneling paths, our next task is to calculate their associated tunneling probability per unit of time
(tunneling rate) and compare it with the exact case.
Consider a mean-field evolution from $\ket{i}$ to $\ket{f}$ over a time $T$.
In real-time, the probability to end up in $\ket{f}$ is $|\bra{f}\hat{U}(T)\ket{i}|^2=|e^{iS}|^2=1$ as the Dirac action
in Eq.~(\ref{eq:Dirac}) for this path is real.
The energy being constant, the global phase $ET$ is irrelevant.
In imaginary-time, this probability is now given by $|e^{iW}|^2$ with
\begin{eqnarray}
W(T)&=&-\int_0^Td\tau \bra{\Psi}\partial_\tau\ket{\Psi}\nonumber\\
&=&-\sum_{i=1}^N\int_{0}^T d\tau \bra{\psi_i(\tau)}\pdv{\tau}\ket{\psi_i(\tau)}\nonumber
\end{eqnarray}
where $N$ is the number of particles.
Using results from the previous section, we find
\begin{subequations}
\begin{align}
\bra{\psi}\pdv{\tau}\ket{\psi
&=\int dx \,\left(\frac{1}{2}\pdv{\rho(x,\tau)}{\tau}+i\rho(x,\tau)\pdv{\phi(x,\tau)}{\tau}\right)\nonumber\\%\label{eq:dtau}\\
&=\int dx\, i\rho(x,\tau)\pdv{\phi(x,\tau)}{\tau},\nonumber
\end{align}
\end{subequations}
where we used the fact that $\int dx\, \rho(x,\tau)$ is constant.
In the toy model, the space integral $\int dx$ is simply replaced by a discrete sum $\sum_{L,R}$ over the left and right states, giving
\begin{equation}
W[T]=-i\int_0^Td\tau\,\,\left(\rho_L\pdv{\phi_L}{\tau}+\rho_R\pdv{\phi_R}{\tau}\right)\nonumber
\end{equation}
where
$$\rho_{L}=\frac{1+\sin\theta}{2} \mbox{ and } \rho_{R}=\frac{1-\sin\theta}{2}.$$
Once again, only the relative phase $\phi=\phi_R-\phi_L$ matters, thus we set $\phi_{R}=\phi/2$ and $\phi_{L}=-\phi/2$, giving
\begin{equation}
W[T]=\frac{i}{2}\int_0^Td\tau\,\,\sin\theta\partial_\tau\phi.\nonumber
\end{equation}
Using the second equation of motion~(\ref{eq:ITEOM}), we finally get
\begin{equation}
W[T]=\frac{1}{2}\int_0^Td\tau\,\,(\tan\theta\cos\phi+\mu\sin\theta)\sin\theta. \nonumbe
\end{equation}
Up to an irrelevant global phase,
the tunneling probability amplitude for an imaginary-time-dependent mean-field path $\Psi(\tau)$ connecting the two wells is then given by
\begin{equation}
e^{iS[\Psi]}\equiv e^{-i\int d\tau\bra{\Psi}\partial_\tau\ket{\Psi}}= e^{\frac{i}{2}\int_0^Td\tau\,(\tan\theta\cos\phi+\mu\sin\theta)\sin\theta}.\nonumber
\end{equation}
As $\theta$ and $\phi$ are now complex quantities, the tunneling probability $|e^{iS[\Psi]}|^2$ associated with this path can be less than one.
By analogy with the standard semi-classical treatment of $\alpha-$decay \cite{Gamow1928}, the tunneling rate is given by the tunneling probability
multiplied by the frequency at which the system ``hits'' the potential barrier,
i.e., the frequency of the oscillation observed in mean-field trapping [see orange line in Fig.~\ref{fig:2part}(b)].
As the system may tunnel from different configurations along a real-time iso-energy contour [see solid red line in Fig.~\ref{fig:enconts}(c)],
the tunneling rate is in principle obtained by averaging over the associated imaginary-time paths.
To a good approximation, this corresponds to the tunneling rate for the path connecting the left and right degenerate mean-field ground-states, indicated by the solid green line in Fig.~\ref{fig:3contours}.
In the exact case, the tunneling rate is simply given by twice the frequency at which the system oscillates between left and right wells.
As shown in Eq.~(\ref{eq:Nlcount}), this oscillation has two modes at $\omega_\pm=\frac{\sqrt{4\alpha^2+\mu^2}\pm|\mu|}{2\pi}$.
Only the lowest frequency is associated with tunneling, giving an exact tunneling rate $2\omega_-=\frac{2}{\pi}\Delta E$
where $\Delta E=\frac{\sqrt{4\alpha^2+\mu^2}-|\mu|}{2}\xrightarrow[|\mu|\gg\alpha]{}\frac{\alpha^2}{|\mu|}$ is the energy difference between the ground and first excited states.
The mean-field and exact tunneling rates are compared in Fig.~\ref{fig:exactandapprox}.
Although the imaginary-time mean-field predictions are wrong for weakly interacting ($|\mu|<2$) systems -- in which case the real-time mean-field prediction can be used anyway -- it is in excellent agreement with the exact case in the strongly interacting regime,
reproducing well the slowing down of tunneling with increased interactions.
\section{Towards realistic applications\label{sec:realistic}}
The purpose of the two-well model is to illustrate the imaginary time mean-field
method and compare with an exact solution (which would be hard to obtain
with more realistic models). This is of course a first step and for the method
to be useful, its applicability to more realistic systems needs to be demonstrated.
\subsubsection{Cartesian grids}
The toy model has only two states per particle, while the single-particle Hilbert
space for a one-dimensional discretised cartesian grid has as many states
as the number of points in the grid -- typically $\sim100$.
Naturally, numerical simulations with non-unitary evolution operators
such as $\exp(\pm\hat{H}\tau)$
present additional technical challenges in terms of stability and convergence.
The generalisation from one to three dimensions will be another computational
challenge, though it does not bring additional formal difficulty in terms of the
algorithm itself.
\subsubsection{Spin and exchange terms}
Other extensions include the inclusion of spin and exchange terms.
Spin $1/2$ degrees of freedom can be accounted for in the same way as in
real-time calculations where each single-particle is treated as a 2-spinor
$(\varphi_\uparrow(\mathbf{r}), \varphi_\downarrow(\mathbf{r}))$.
An exchange (Fock) term also appears in the case of identical
particles. In general, it is non-local and often requires a major extra
computation cost. However, in the case of contact interactions (often used in cold atoms
systems as well as in nuclear physics, e.g., with the Skyrme effective interaction \cite{skyrme1956}),
the exchange terms are easily accounted for.
For Coulomb interactions, the exchange term can also be included via the Slater approximation.
\subsubsection{Potential applications\label{sec:applications}}
Several applications could be considered:
\begin{itemize}
\item {\it Interacting particles in an external potential}\\
Cluster dynamics such as $\alpha-$decay or emission of atom clusters
can be studied with an external potential.
In this case, the external potential simulates the mean-field of the
particles which do not belong to the cluster, while each particle of the cluster is
treated explicitly.
One could study the effect of the internal degrees of freedom of the cluster
while it tunnels as a whole.
In this case the classically forbidden region could be defined as the
turning points of the external potential in the usual way.
\item {\it Merging of two self-bound systems}\\
A typical example is the fusion of two atomic nuclei. Note that the nuclei
are self-bound and thus there is no external potential in this problem, i.e.,
$h_0$ only contains the kinetic energy of the nucleons. The nuclei are
bound thanks to their strong nuclear interaction. The Coulomb barrier
preventing fusion in the classical case is produced by the competition of
long-range Coulomb repulsion between protons and the short range
nuclear attraction of all nucleons, both terms being part of the interaction $v(1,2)$. In this
case, the real-time mean-field dynamics is only able to reach fusion if
the kinetic energy at large distance exceeds the Coulomb barrier height
(for a central collision). At lower energy, tunneling will be obtained
through the imaginary time mean-field method.
As illustrated by the red
line in Fig.~\ref{fig:3contours}, the system may explore different configurations through
real-time dynamics, with each of these configurations potentially serving
as initial condition for the imaginary time evolution. In principle, a weighting
of each possibility should be determined. In practice, however, the
transmission through the barrier is expected to be dominated by the
trajectory starting from the distance of closest approach.
\item {\it Scission of a self-bound metastable system}\\
Self-bound systems can be in a local minimum of their potential energy
surface, with more stable configurations corresponding to disconnected
fragments. This is the case of fission in heavy nuclei. Here, again, the
parent nucleus is self-bound and no external potential is required ($h_0$
only contains a kinetic energy term). In the case of spontaneous fission
in particular, all directions in the multidimensional potential energy surface
are classically forbidden, thus the initial condition for the imaginary time
evolution is well defined\footnote{\label{foot2}In practice, a small deviation from the
mean-field ground state would be needed to initiate the evolution in the
right ``direction'' (e.g., a small increase of the quadrupole moment). }.
\end{itemize}
\section{Conclusions\label{sec:conclusions}}
Theoretical description of tunneling in strongly interacting systems such as atomic nuclei remains a challenging problem.
Standard real-time mean-field approaches are unable to account for many-body tunneling due to spurious ``self-trapping''.
Using a simple model with two particles in a two-well potential, we demonstrated the possibility to overcome this limitation by allowing imaginary-time mean-field evolution.
Tunneling probabilities are in excellent agreement with the exact solution in the strongly interacting regime.
These results are promising and encourage applications to more realistic systems.
A first natural extension is to increase the number of modes, e.g., using cartesian grids with one or more dimensions.
Computational effort only increases linearly with the number of particles at the mean-field level, so simulating tunneling dynamics of larger systems (out of reach to exact and few-body techniques) should not be an issue.
The imaginary-time mean-field equations could also be extended to include pairing correlations \cite{Levit2020}.\\
\begin{acknowledgments}
We are grateful to R. Bernard for useful discussions.
This work has been supported by the Australian Research Council Discovery Project (project number DP190100256) funding scheme.
\end{acknowledgments}
\twocolumngrid
|
\section{Introduction}
The success of deep networks largely relies on large-scale datasets with high-quality label annotations \cite{han2018masking,wang2018iterative,jiang2020beyond}. However, it is quite costly, time-consuming, or even infeasible to collect such data. Instead, in practice, many large-scale datasets are collected by cheap ways, e.g., from search engines or web crawlers. The obtained data contains label noise \cite{yao2020searching,wang2020fair}. The presence of label noise may adversely affect the model prediction and generalization performance \cite{zhang2017understanding}. It is therefore of great importance to train deep networks robustly against label noise.
\begin{figure}[t]
\centering
\includegraphics[width=0.40\textwidth]{main.png}
\caption{Illustration of the cluster-dependent extended transition matrix. The face images are collected from MS1MV0 \cite{guo2016ms}. The proposed method learns the transition matrix for different clusters and extends the traditional transition matrix to be able to model mixed label noise.}
\label{fig:main}
\end{figure}
The types of label noise studied so far can be fell into two categories: \textit{closed-set} and \textit{open-set} label noise. The closed-set label noise occurs when instances with noisy labels have true class labels within the noisy label set \cite{wang2018iterative}. Oppositely, the open-set label noise occurs when instances with noisy labels have some true class labels outside the noisy label set \cite{wang2018iterative}.
The closed-set label noise has been extensively studied, e.g., \cite{liu2016classification,han2018masking,zheng2020error,yao2020dual,menon2020can,li2020dividemix,nguyen2020self,liu2019peer,liu2020early,han2020sigua,wei2020optimizing,cheng2020learning}. There are also some pioneer work studying the open-set label noise, e.g., \cite{wang2018iterative,zhang2018generalized,yu2019does}. All these methods are designed for handling the closed-set and open-set label noise independently and cannot handle the mixed closed-set and open-set label noise well. However, it is more practical that the two types of label noise co-occur in real-world applications \cite{sachdeva2020evidentialmix}. For example, many large-scale face recognition datasets are automatically collected via image search engines. The face images in the datasets contain both two types of label noise \cite{wang2019co}.
One promising strategy for combating label noise is to model label noise. Compared with model-free methods which empirically work well but does not model the label noise explicitly, the reliability of model-based methods is better guaranteed \cite{patrini2017making,xia2020parts,berthon2020idn}. By utilizing the \textit{transition matrix} which denotes the probabilities that clean labels flip into noisy ones, model-based methods have been verified to be able to deal with closed-set label noise, mainly with the kind of class-dependent (instance-independent) label noise \cite{cheng2017learning,natarajan2013learning,shu2020meta}. However, the traditional model-based methods have the following limitations. Firstly, they cannot model open-set label noise and will provide \textit{biased} solutions when there exists mixed closed-set and open-set label noise. Secondly, the instance-dependent label noise is common in real-world applications as difficult instances are prone to have inaccurate labels. It is \textit{ill-posed} to
learn the instance-dependent transition matrix by only exploiting the noisy training data as discussed in \cite{xia2020parts}. Class-dependent transition matrix has been exploited to approximate the instance-dependent transition matrix. However, the approximation error is large when the label noise rate is high.
In this paper, we present a novel method for learning under the mixed closed-set and open-set label noise. The proposed method extends the traditional transition matrix to be able to model the mixed label noise and better approximate the instance-dependent label noise. Specifically, we integrate all open-set classes as a \textit{meta class}, which is paratactic with the other true classes in the closed set. We then identify \textit{anchor points} belonging to the meta class of the open-set and the true classes of the closed set. The extended transition matrix involving the meta class can be unbiasedly estimated by exploiting anchor points. To further handle the instance-dependent label noise in reality, we exploit cluster-dependent transition matrices to better approximate the instance-dependent transition matrix. Specifically, we divide all training examples into several clusters (with the constraint that cluster contains anchor points for the meta class of the open set and true classes of the closed set). The cluster-dependent transition matrix can then be unbiasedly estimated for each cluster. The cluster-dependent transition matrices capture the geometric information of instances and thus can better approximate the instance-dependent transition matrix than the class-dependent transition matrix. The illustration of the proposed method is provided in Fig.~{\ref{fig:main}}.
The main contributions of this paper are three-fold:
\begin{itemize}
\item We focus on learning under the mixed closed-set and open-set label noise, and extend the traditional transition matrix to be able to model the mixed label noise.
\item We propose the cluster-dependent extended transition matrices to handle instance-dependent label noise in real-world applications, which produces a more reliable solution.
\item We conduct comprehensive experiments on synthetic and real-world label-noise datasets to demonstrate that the proposed method achieves superior robustness over the baselines under the mixed label noise.
\end{itemize}
The rest of the paper is organized as follows. In Section \ref{sec2} , we review related works on learning with label noise. In section \ref{sec3}, we introduce some notations and background knowledge. In Section \ref{sec4}, we introduce the proposed method in details. Experimental results and analyses are provided in Section \ref{sec5}. Finally, we conclude the paper in Section \ref{sec6}.
\section{Related Work}\label{sec2}
\textbf{Learning with the noise transition matrix.} The label noise transition matrix plays an essential role in modeling the label noise, which reflects the probabilities that true labels flip into other noisy ones \cite{natarajan2013learning,scott2015rate,hendrycks2018using,wu2020class2simi}. For learning under closed-set label noise, where true classes are known classes in the training data, the noise transition matrix can effectively infer the \textit{clean class posterior probability} with the \textit{noisy class posterior probability} \cite{liu2016classification}. Thus, we can assign clean labels for given instances, even though only noisy training data are available \cite{patrini2017making}. Many advanced methods borrow this idea and learn the noise transition matrix to handle label noise \cite{liu2016classification,shu2020meta}. For example, for estimating the class-dependent (closed-set label noise) transition matrix more accurately, a slack variable can be introduced to revise the initialized transition matrix \cite{xia2019revision}. An intermediate class can be used to avoid directly estimating the noisy class posterior \cite{yao2020dual}. For modeling the instance-dependent closed-set label noise more accurately, the label noise can be assumed to depend only on the parts of instances \cite{xia2020parts}. For learning under open-set label noise, true classes of noisy training data are \textit{outside} the set of known classes. Recall the definition of the (traditional) transition matrix, the flip probabilities indicate the rates of the true classes flipped to the noisy ones. If we use the method of modeling closed-set label noise to model open-set label noise, we will mistakenly treat some unknown incorrect classes as true classes, which leads to poor classification performance. To the best of our knowledge, how to effectively model the mixed closed-set and open-set label noise is a new challenge. \\
\textbf{Other methods of learning under label noise.} We first briefly introduce other algorithms for dealing with closed-set and open-set label noise separately without modeling the noise explicitly. For example, some approaches to handle label noise are to exploit the memorization effects of deep models and extract reliable examples with small loss from the training data \cite{jiang2018mentornet,han2018co,yu2019does}. Some approaches to handle label noise are to reweight examples \cite{ren2018learning,wang2018iterative,shu2019meta}. Some works aim to design robust loss functions to cope with label noise \cite{wang2019symmetric,menon2020can,liu2019peer}. Some other works (implicitly) add regularization \cite{guo2018curriculumnet,veit2017learning,vahdat2017toward,jiang2020beyond,lukasik2020does}. Then, we introduce the pioneer work for dealing with mixed closed-set and open-set label noise without modeling the noise explicitly. EvidentialMix \cite{sachdeva2020evidentialmix} focuses on synthetic mixed label noise and achieves promising performance by combining DivideMix \cite{li2020dividemix} and SL loss \cite{sensoy2018evidential}. Note that our work is fundamentally different from EvidentialMix. The main reasons are as follows. (1) EvidentialMix combines several of advanced approaches, but our work focuses on one, i.e., learning with the noise transition matrix. (2) EvidentialMix works well, but does not model the mixed label noise. Our work models the mixed label noise and improves the reliability of the method. We suggest that the readers can refer \cite{han2020survey} for more methods of learning under label noise. \\
\textbf{Deep clustering.} As a typical unsupervised learning method, clustering has been widely used in various tasks. It aims to keep similar data points in the same cluster while dissimilar ones in different clusters. Clustering is proven to be able to find representative data points among all data points \cite{wang2008clustering,teichgraeber2019clustering}. Benefit from the power of deep learning, lots of approaches boost traditional clustering techniques, e.g., $k$-means \cite{jain2010data} and spectral clustering \cite{ng2002spectral}, by using deep models \cite{li2018discriminatively,yang2016joint}. They cluster on deep representations instead of original features as deep representations have a higher degree of discrimination \cite{yang2019deep}.
\section{Preliminaries}\label{sec3}
In this section, we introduce the notation, problem setting and background knowledge for the label noise transition matrix.
In this paper, we consider the $c$-class classification problem. Let $\mathcal{X}$ and $\mathcal{Y}$ be the input and output domains, where $\mathcal{Y}=\{1,\ldots,c\}$ is the label set. Let $D$ be the clean joint distribution of a pair of random variables $(X,Y)\in\mathcal{X}\times\mathcal{Y}$, and $\{(x_i,y_i)\}_{i=1}^n$ be a training sample drawn from $D$.
Let $\Tilde{D}$ be the noisy joint distribution of the noisy random variables $(X,\Tilde{Y})\in\mathcal{X}\times\mathcal{Y}$.
In learning under the mixed label noise, what we can access is a noisy training sample $\{(x_i,\Tilde{y}_i)\}_{i=1}^n$ drawn from a noisy distribution $\Tilde{D}$, where $\Tilde{y}$ is the possibly corrupted label of the underlying clean label $y$. Our aim is to learn a robust classifier from the noisy sample that can assign clean labels for test instances.
We exploit a deep neural network to train our classifier. The network comprises a transformation $h:\mathcal{X}\rightarrow\mathbb{R}^c$, where $h=h^{(d)}\circ h^{(d-1)}\circ\cdots\circ h^{(1)}$ is the composition of a series of transformation layers $h^{(i)}$. Specifically,
\begin{align*}
&h^{(i)}(z)=\sigma(W^{(i)}z+b^{(i)}), i=1,\ldots,d-1,\\
&h^{(d)}(z)=W^{(d)}z,
\end{align*}
where $W^{(i)}$ and $b^{(i)}$ are parameters to be learned, and $\sigma$ is the activation function, e.g., Sigmoid \cite{yin2003flexible} and ReLU \cite{glorot2011deep}. Note that the outputs after the layer $h^{(d-1)}$ can be regarded as the \textit{deep representations} of the instances. For the classification task, the output layer $f$ is always set as a \textit{softmax} layer to imitate classificaition probabilities, i.e., $f_i(x)=\text{exp}(h_i(x))/\sum_{k=1}^c\text{exp}(h_k(x))$. Namely, $f(X)=\hat{P}(Y|X)$. For traditional supervised learning (the classification task), the optimal classifier is defined by that one that minimizes the expected risk:
\begin{align}\label{eq:expect_risk}
R_{\ell,D}(f)=\mathbb{E}_{(X,Y)\sim D}[\ell(f(X),Y)],
\end{align}
where $\mathbb{E}_{(X,Y)\sim D}$ denotes the expectation over the distribution $D$, and $\ell:\mathbb{R}^c\times\mathcal{Y}\rightarrow\mathbb{R}_{+}$ is a surrogate loss function for $c$-class classification, e.g., the \textit{cross-entropy loss}. As $D$ is usually unknown, the empirical risk has been exploited as follows to approximate the expected risk:
\begin{align}\label{eq:empirical_risk_clean}
R_{\ell,n}(f)=\frac{1}{n}\sum_{i=1}^n\ell(f(x_i),y_i).
\end{align}
However, in learning under label noise, the clean training sample is not available.
We only have the access to the noisy training sample. By exploiting the label noise transition matrix, the following modified empirical risk has been proposed to well approximate $R_{\ell,D}(f)$:
\begin{align}\label{eq:empirical_risk_noisy}
\Tilde{R}_{\tilde{\ell},n}(f)=\frac{1}{n}\sum_{i=1}^n\tilde{\ell}(f(x_i),\Tilde{y}_i),
\end{align}
where $\tilde{\ell}$ represents the modified loss function by exploiting the transition matrix. It has been proven that the empirical risk $\Tilde{R}_{\tilde{\ell},n}(f)$ will converge to the expected risk $R_{\ell,D}(f)$ if the transition matrix can be unbiasedly estimated \cite{liu2016classification,patrini2017making,yao2020dual}. Then the corresponding learned classifier will be statistically consistent because it will converge to the optimal one defined by the clean data.
In this work, we will study how to use the transition matrix to model the mixed closed-set and open-set label noise, based on which, robust classifiers can be designed. We first formally introduce the traditional transition matrix, i.e., $T\in[0,1]^{c\times c}$, which is only capable of modeling the closed-set label noise. The transition matrix generally depends on the instances and the true labels, i.e., $T_{ij}(x)=P(\Tilde{Y}=j|Y=i,X=x)$ \cite{cheng2017learning}. Unfortunately, the instance-dependent transition matrix is unidentifiable without any assumption \cite{xia2019revision,xia2020parts}. The existing methods approximate the instance-dependent transition matrix by assuming that the noise transition matrix is \textit{class-dependent} and \textit{instance-independent}, i.e., $T(x)=P(\Tilde{Y}=j|Y=i,X=x)=P(\Tilde{Y}=j|Y=i)$. When there is no confusion, we will short-hand $T(x)$ as $T$ for the class-dependent transition matrix. Note that the noisy class posterior $P(\Tilde{Y}|X)$ can be estimated by using the noisy training data. With the transition matrix, we can bridge the noisy class posterior $P(\Tilde{Y}|X)$ and the clean class posterior $P(Y|X)$ as follows:
\begin{align}\label{eq:transition}
P(\Tilde{Y}=j|X=x) = \sum_{i=1}^cT_{ij}P(Y=i|X=x).
\end{align}
The matrix form of Eq.~(\ref{eq:transition}) can be written as $P(\Tilde{Y}|X)=T^\top P(Y|X)$. Namely, when learning under label noise, if we have the access to the ground-truth transition matrix, we can recover $P(Y|X)$ with $P(\Tilde{Y}|X)$. The transition matrix can be therefore used to build \textit{classifier-consistent} or \textit{risk-consistent} algorithms \cite{liu2016classification,xia2019revision}.
However, the traditional transition matrix fails to handle the mixed closed-set and open-set label noise as it encodes no open-set class information. We will discuss how to extend it to better handle the instance-dependent mixed closed-set and open-set label noise in the next section.
\section{Methodology}\label{sec4}
In this section, we first show how to model the mixed label noise by extending the traditional label noise transition matrix.
Then we present how to learn the cluster-dependent extended transition matrices for better approximating the instance-dependent mixed label noise. Finally, we show how to exploit the extended transition matrix for training a robust classifier.
\subsection{Class-dependent extended T}\label{sec:4.1}
As discussed in Section \ref{sec3}, the traditional label noise transition matrix is a $c\times c$ matrix linking the noisy class information to the closed-set clean class information without considering the open-set clean class information. It is therefore limited to handle the open-set label noise problems. Taking the traditional class-dependent transition matrix as an example, we discuss how to extend it to handle the open-set label noise. Note that in the next subsection, we will discuss how to handle the instance-dependent mixed label noise.
To model the open-set label noise, we introduce a \textit{meta class} which is an integration of all the possible open-set classes. As shown in Figure \ref{fig:main}, we extend the traditional transition matrix to $(c+1)\times c$ dimensional, where the additional $1\times c$ vector denoted by $T^{\circ}$ represents how the meta class (or the open-set class) flips into the closed-set classes, i.e., $P(\tilde{Y}=j|Y=m,X=x)$, where $j=1,...,c$ and $m$ represents the meta class label. The extended transition matrix encodes the open-set class information and thus can be exploited to better reduce the side-effect of the open-set label noise.
We then discuss how to estimate the extended transition matrix by exploiting \textit{anchor points}.
Anchor points are defined in the clean data domain \cite{liu2016classification}. Formally, an instance $x$ is an anchor point for the class $i$ if $P(Y=i|X=x)$ is equal to one or approaches one. Given an anchor point $x$, if $P(Y=i|X=x)=1$, we have that for $k\neq i$, $P(Y=k|X=x)=0$. Then, we have,
\begin{equation}\label{eq:anchor}
P(\Tilde{Y}=j|X=x)=\sum_{k=1}^cT_{kj}P(Y=k|X=x)=T_{ij}.
\end{equation}
The equation holds because we assume that the transition matrix is class-dependent and instance-independent, e.g., $T_{ij}(x)=T_{ij}$.
Therefore, the transition matrix $T$ can be unbiasedly estimated via estimating the noisy class posteriors for the anchor point of each class (including the meta class). Note that the anchor point assumption is a widely adopted in the literature of learning with noisy labels \cite{liu2016classification,yao2020dual,yu2018learning}. We could follow them and assume the availability of anchor points.
However, assuming that anchor points are given could be strong for many real-world applications. We could release the assumption by just assuming that the anchor points exist in the training data and then design algorithms to locate them. By the definition of anchor points, they are the representatives of each class.
The $k$-means cluster technique \cite{jain2010data} therefore could be utilized to detect anchor points. Note that to fulfil the power of deep learning, we could make use of the deep features to find anchor points. Note that even if the training labels are noisy, the deep features obtained by exploiting a deep neural network and the noisy training sample still capture the class-discriminative information and could be exploited for finding anchor points. (Note that the obtained deep features are not sufficient for learning the class posteriors.)
Specifically, for the training example $(x_1,\ldots,x_n)$, the number of clusters is set to $k=c+1$. The aim is to detect the anchor points of the $c$ classes in the closed set and the meta class in the open set.
For $k$-means, we formulate the loss function as:
\begin{equation}
\ell_k=\sum_{i=1}^n\sum_{k=1}^{c+1}M_{ik}\|x_i-\mu_k\|_2^2,
\end{equation}
where $M$ is the cluster matrix, $M_{ik}=1$ if $x_i$ belongs to the $k$-th cluster, otherwise $M_{ik}=0$. The symbol $\mu_k$ represents the $k$-th cluster centroid \cite{jain2010data}. By assuming that the open-set examples are of minority, the top $c$ largest clusters are selected for the closed-set classes and the remaining cluster is deemed as the meta-class cluster. To assign class labels to the closed-set clusters, we first count the noisy class labels in each cluster. Then we regard the cluster with the most examples of a certain noisy class $i$ as the $i$-class cluster. When detecting the anchor points belonging to the closed-set classes, we follow \cite{patrini2017making,xia2019revision}. Then, we use these anchor points to estimate the transition matrix $T$ for modeling the closed-set label noise with Eq.~(\ref{eq:anchor}).
For detecting the anchor points belonging to the meta class, we determine that the anchor points are the data points which are close to the centroid of the meta class cluster. This way of detection is reasonable. Recall the definition of anchor points (i.e., data points that belong to a specific class almost surely), it reveals that anchor points are the most representative data points. Clustering has been verified to be able to detect the the representative data points. The cluster centroid is the representative of the cluster. Thus for the meta class, we intuitively set the data point close to the centroid as the anchor point. Then with Eq.~(\ref{eq:anchor}), we use the noisy class posterior probabilities of the anchor point to estimate the transition matrix for modeling the open-set label noise. We denote the transition matrix for open-set label noise as $T^{\circ}\in[0,1]^{1\times c}$. When finish the estimation of the transition matrix for both types of label noise, we combine $T$ and $T^{\circ}$ to obtain the extended transition matrix $T^{\star}\in[0,1]^{(c+1)\times c}$ to model the mixed label noise.
\subsection{Cluster-dependent extended T}\label{sec:4.2}
We have presented how to model the mixed label noise by using class-dependent extended transition matrix and how to estimate the extended transition matrix. However, many real-world label noise is instance-dependent. To handle this problem, we propose to use cluster-dependent extended transition matrices to better model the instance-dependent label noise, which is based on the intuition that the instances which have similar features are more prone to have the similar label flip process \cite{cheng2017learning}. We thus can employ the same extended transition matrix to model the mixed label noise for the instances which have similar features. We term such extended transition matrices as cluster-dependent extended transition matrices.
We show how to learn the cluster-dependent transition matrices as follows. Consider the training examples $(x_1,\ldots,x_n)$, we cluster on their deep representations again to obtain clusters. The total number of the clusters is set to a small number, i.e., ${k}$. Note that when we set the value of ${k}$ for clustering, we have to ensure that there are anchor points for each classes in each cluster. The overall procedure to learn the cluster-dependent extended transition matrices is summarized in Algorithm \ref{alg:main}.
\begin{algorithm}[h]
{\bfseries Input}: Noisy training sample $\mathcal{D}_t$, noisy validation sample $\mathcal{D}_v$, the number of cluster-dependent transtion matrices $k$;\\
1: Train a deep model by using the noisy data $\mathcal{D}_t$ and $\mathcal{D}_v$;\\
2: Get the deep representations of the examples by employing the trained deep network;\\
3: Detect the anchor points used for estimation with clustering; \\
4: Cluster on the deep representations of the examples to obtain $k$ clusters;\\
\For{$i=1,\dots,k$}{
5: Estimate the transition matrix $T_i$ for the closed-set label nosie;\\
6: Estimate the transition matrix $T_i^\circ$ for the open-set label noise;\\
7: Obtain the cluster-dependent transition matrix $T_i^{\star}$;\\
}
8: {\bfseries Output}: $T_1^{\star},\ldots,T_{k}^{\star}$.
\caption{Cluster-dependent Transition Matrices Learning Algorithm}
\label{alg:main}
\end{algorithm}
\subsection{Learning with importance reweighting}\label{sec:4.3}
In previous subsection, we have presented the methods about how to learn the cluster-dependent extended transition matrices to model the mixed instance-dependent label noise. With the learned extended transition matrix $T_i^{\star} (i=1,\ldots,{k})$, we employ the \textit{importance reweighting} technique \cite{gretton2009covariate,liu2016classification,fang2020rethinking} to train a robust classifier against the label noise. Importance reweighting can be exploited to rewrite the expected risk w.r.t. clean data (Eq.~(\ref{eq:expect_risk})) as follows:
\begin{eqnarray*}
&R_{\ell,D}(f)&=\mathbb{E}_{(X,Y)\sim D}\left[\ell(f(X),Y)\right]\\
&&=\mathbb{E}_{(X,\Tilde{Y})\sim \Tilde{D}}\left[\frac{P_{D}(X,Y)}{P_{\Tilde{D}}(X,\Tilde{Y})}\ell(f(X),\Tilde{Y})\right]\\
&&=\mathbb{E}_{(X,\Tilde{Y})\sim \Tilde{D}}\left[\beta(X,\Tilde{Y})\ell(f(X),\Tilde{Y})\right]\\
&&=\mathbb{E}_{(X,\Tilde{Y})\sim \Tilde{D}}\left[\bar{\ell}(f(X),\Tilde{Y})\right].
\end{eqnarray*}
If $\Tilde{Y}=i$, $\bar{\ell}(f(X),\Tilde{Y})=\bar{\ell}(f(X),i)=\frac{P_D(Y=i|X=x)}{P_{\Tilde{D}}(\Tilde{Y}=i|X=x)}\ell(f(X),i)$. As $P_{\Tilde{D}}(\Tilde{Y}=i|X=x)\neq 0$ during training, the importance reweighting method is stable without truncating the importance ratios \cite{liu2016classification}.
For the $c$-class classification problem under the mixed label noise, by exploiting the cluster-dependent extended transition matrices $T^{\star}$ (we hide the index for simplify), the empirical risk can be formulated as:
\begin{align}
\label{eq:em_importance}%
{
\tilde{R}_{\tilde{\ell},n}=\frac{1}{n}\sum_{i=1}^{n}\frac{{g}_{\Tilde{y}_i}(x_i)}{(T^{\star\top}{g})_{\Tilde{y}_i}(x_i)}\ell(f(x_i),\Tilde{y}_i),}
\end{align}
where $f(x)=(T^{\star\top}{g})(x)$ and $g(x)$ is the output of the softmax layer. We use $\arg\max_{j\in\{1,\ldots,c\}}{g}_j(x)$ to assign labels for the test data. Note that during training, the $T^{\star}$ is determined according to the cluster to which the example $x_i$ belongs. As we detect the anchor points from the noisy training data, as did in \cite{xia2019revision,xia2020parts,patrini2017making}, data points that are similar to anchor points will be detected if there are no anchor points available. Then, the cluster-dependent extended transition matrices will be poorly estimated. To handle this problem, we follow \cite{xia2019revision} to revise the cluster-dependent extended transition matrices, which helps lead to a better classifier. We term the systemic proposed method for training a robust classifier against mixed label noise as \textit{Extended T}. In more detail, Extended T-$i$ means that the number of the cluster-dependent transition matrices is $i$.
\section{Experiments}\label{sec5}
In this section, we first introduce the methods for comparison in the experiments (Section \ref{sec:5.1}). We then introduce the details of the experiments on synthetic datasets (Section \ref{sec:5.2}). The experiments on real-world datasets are finally presented (Section \ref{sec:5.3}).
\subsection{Comparison methods}\label{sec:5.1}
We compare the proposed method with the following methods: (1) CE, which trains the deep models with the standard cross entropy loss on noisy datasets. (2) GCE \cite{zhang2018generalized}, which handles label noise by exploiting the negative Box-Cox transformation. (3) APL \cite{ma2020normalized}, which combines two robust loss functions which boost each other. (4) DMI \cite{xu2019l_dmi}, which handles label noise from the perspective of the information theory. (5) NLNL \cite{kim2019nlnl}, which proposes a novel learning method called Negative Learning (NL) to reduce the side effect of label noise. (6) Co-teaching \cite{han2018co}, which trains two networks simultaneously and exchanges the selected examples for network updating. (7) Co-teaching+ \cite{yu2019does}, which trains two networks simultaneously and finds confident examples among the prediction disagreement data. (8) JoCor \cite{wei2020combating}, which reduces the diversity of networks to improve the robustness. (9) S2E \cite{yao2020searching}, which utilizes AutoML to handle label noise. (10) Forward \cite{patrini2017making}, which estimates the class-dependent transition matrix to correct the training loss. (11) T-Revision \cite{xia2019revision}, which introduces a slack variable to revise the estimated transition matrix and leads to a better classifier. Note that we
do not compare with some state-of-the-art methods like SELF \cite{nguyen2020self}, DivideMix \cite{li2020dividemix}, and EvidentialMix \cite{sachdeva2020evidentialmix}. It is because that their proposed methods are aggregations of multiple advanced approachs while this work only focuses on one, therefore the comparison is not fair. To measure the performance, we use the classification accuracy on the test set. Intuitively, higher accuracy means that the method is more robust to the mixed label noise. Note that for a fair comparison, we implement all methods with default parameters by PyTorch, and conduct all the experiments on NVIDIA Tesla V100 GPUs.
\subsection{Experiments on synthetic datasets}\label{sec:5.2}
\subsubsection{Datasets and Implementation details}
\textbf{Datasets.} We evaluate the robustness of the proposed method to label noise on synthetic CIFAR-10 \cite{krizhevsky2009learning} with comprehensive experiments. Original CIFAR-10 consists of 50,000 training images and 10,000 test images with 10 classes. The size of images is 32$\times$32$\times$3. Note that the original CIFAR-10 contains clean training data. We thus corrupt the training data manually to generate label noise. Specifically, for the open-set label noise, we follow \cite{wang2018iterative} and borrow the images from SVHN \cite{netzer2011svhn}, CIFAR-100 \cite{krizhevsky2009learning}, and ImageNet32 (32$\times$32$\times$3 ImageNet images) \cite{chrabaszcz2017downsampled} to act as outside images. Note that only images whose labels exclude 10 classes in CIFAR-10 are considered. We then use the outside images to replace some training images in CIFAR-10. For closed-set label noise, we consider the symmetric noise, which is more challenging than asymmetric noise \cite{patrini2017making}. Note that we consider controlled
class-dependent label noise. It is because we can show the estimation results more clearly. The overall label noise rate is denoted by $\tau\in\{0.2,0.4,0.6,0.8\}$. The proportion of open-set noise in the label noise is denoted by $\rho\in\{0.25,0.5,0.75\}$, and the proportion of closed-set noise is $1-\rho$. We leave out 10\% of the training data as a validation set, which is used for model selection. The experimental results are reported over five trials. We conduct significance tests to show whether the experimental results are statistically significant. The details for significance tests can be found in Appendix A.\\
\textbf{Implementation.} We employ a PreAct ResNet-18 network \cite{he2016identity}. For learning the transition matrix, we follow the optimization method in \cite{patrini2017making,xia2019revision}. We next use SGD optimizer with momentum 0.9, batch size 128, and weight decay 5$\times10^{-4}$ to initialize the network. The initial learning rate is set to $10^{-2}$, and divided by 10 after the 40th epochs and 80th epochs. 100 epochs are set totally. Following \cite{xia2019revision}, we then exploit Adam optimizer with a learning rate 5$\times10^{-7}$ to revise the transition matrix. Typical data augmentations including random crop and horizontal flip are applied.
\begin{table*}[!t]
\centering
\tiny
\begin{tabular}{lp{10mm}p{0.001mm}|ccc|ccc|ccc|ccc}
\toprule
&\multirow{2}{*}{Method}&$\tau$& \multicolumn{3}{c|}{0.2}&\multicolumn{3}{c|}{0.4}&\multicolumn{3}{c|}{0.6}&\multicolumn{3}{c}{0.8}\\
\cmidrule{4-15}
& &$\rho$ & 0.25 & 0.5 & 0.75 & 0.25 & 0.5 & 0.75 & 0.25 & 0.5 & 0.75 & 0.25 & 0.5 & 0.75 \\
\midrule
\parbox[t]{0.1mm}{\multirow{12}{*}{\rotatebox{90}{CIFAR-10+SVHN}}}& CE & & 90.01$\pm$0.12 & 89.82$\pm$0.35&90.18$\pm$0.16&87.82$\pm$0.35&88.20$\pm$0.16&88.27$\pm$0.17&83.18$\pm$0.65&84.91$\pm$0.11&86.44$\pm$0.67&75.07$\pm$0.85&78.44$\pm$0.92&81.72$\pm$0.34\\
& GCE & &90.71$\pm$0.20&90.56$\pm$0.22&90.88$\pm$0.20&90.04$\pm$0.21&90.06$\pm$0.25&89.54$\pm$0.14& 85.82$\pm$0.22&86.21$\pm$0.26&86.26$\pm$0.16&82.57$\pm$0.76&83.80$\pm$0.28&84.12$\pm$0.29 \\
& APL & &90.23$\pm$0.17&90.53$\pm$0.14&90.01$\pm$0.38&90.21$\pm$0.25&88.25$\pm$0.29&88.21$\pm$0.43&84.27$\pm$0.39 &84.51$\pm$0.37&84.52$\pm$0.29&82.95$\pm$0.18&82.02$\pm$0.50&82.11$\pm$0.36\\
& DMI & &90.71$\pm$0.10&90.04$\pm$0.33&90.58$\pm$0.17&89.87$\pm$0.72&89.94$\pm$0.27&89.02$\pm$0.47&85.96$\pm$0.85&85.21$\pm$0.64&85.76$\pm$0.95&81.95$\pm$1.07&82.01$\pm$0.78&80.27$\pm$0.60\\
& NLNL &&89.85$\pm$0.19&89.72$\pm$0.32&89.98$\pm$0.27&87.39$\pm$0.28&87.80$\pm$0.50&88.78$\pm$0.29&82.90$\pm$0.55 &84.58$\pm$0.41&84.37$\pm$0.72&76.39$\pm$0.59&79.65$\pm$0.54&78.39$\pm$1.08\\
& Co-teaching& &90.50$\pm$0.07&90.62$\pm$0.15&90.87$\pm$0.13&86.64$\pm$0.25&87.48$\pm$0.40&87.31$\pm$0.28&75.10$\pm$0.48&76.25$\pm$1.08&78.78$\pm$2.48&46.36$\pm$2.22&49.20$\pm$2.89&53.44$\pm$2.34\\
& Co-teaching+&&88.46$\pm$0.54&89.51$\pm$0.38&88.32$\pm$0.40&86.39$\pm$0.51&86.71$\pm$0.21&84.92$\pm$0.56&63.18$\pm$4.87&65.29$\pm$9.84&56.41$\pm$8.83&10.07$\pm$1.07&17.06$\pm$8.07&15.38$\pm$2.93\\
& JoCor&&88.25$\pm$0.06&89.15$\pm$0.21&89.15$\pm$0.45&84.16$\pm$1.07&82.12$\pm$1.03&82.02$\pm$0.74&67.29$\pm$1.23&69.02$\pm$1.72&71.70$\pm$1.73&43.93$\pm$0.32&42.82$\pm$1.31&40.12$\pm$3.44 \\
& S2E&&89.42$\pm$1.35&89.68$\pm$1.13&89.87$\pm$1.80&88.24$\pm$2.48&88.99$\pm$1.94&88.78$\pm$1.57&81.16$\pm$2.49&85.44$\pm$1.62&85.48$\pm$2.32&57.45$\pm$4.17&74.16$\pm$4.52&78.39$\pm$3.46 \\
& Forward & &89.37$\pm$0.14&89.27$\pm$0.86&89.64$\pm$0.80&87.54$\pm$0.24&87.76$\pm$0.54& 87.01$\pm$0.39&80.19$\pm$2.72&83.21$\pm$0.89&83.92$\pm$1.98&78.05$\pm$2.02&80.32$\pm$1.84&78.66$\pm$1.72\\
& T-Revision&&90.23$\pm$0.14&89.97$\pm$0.23&90.02$\pm$0.14&88.68$\pm$0.24 &88.79$\pm$0.29&89.02$\pm$0.47&85.07$\pm$1.03&85.37$\pm$1.09&85.42$\pm$0.83&81.04$\pm$2.04&81.36$\pm$0.97&82.98$\pm$1.17\\
\midrule
& Extended T& &90.86$\pm$0.13&\textbf{90.89$\pm$0.28}&90.78$\pm$0.16&\textbf{90.94$\pm$0.22}&\textbf{90.72$\pm$0.37}&90.68$\pm$0.38&87.34$\pm$0.38& 86.92$\pm$0.93&87.18$\pm$0.75&83.68$\pm$0.47&83.94$\pm$1.02&84.83$\pm$1.42 \\
& Extended T-2& &\textbf{90.92$\pm$0.08}&90.58$\pm$0.54&\textbf{91.03$\pm$0.22}&90.73$\pm$0.28&90.54$\pm$0.30&90.42$\pm$0.51&87.32$\pm$0.77&87.03$\pm$1.07&87.08$\pm$0.99&84.02$\pm$0.76&84.02$\pm$1.08&84.77$\pm$1.53 \\
& Extended T-3& &90.89$\pm$0.17&90.77$\pm$0.18&91.02$\pm$0.17&90.67$\pm$0.42&90.67$\pm$0.41&\textbf{90.72$\pm$0.69}&\textbf{87.48$\pm$0.63}&\textbf{87.19$\pm$0.92}&\textbf{87.29$\pm$0.73}&\textbf{84.42$\pm$0.93}&\textbf{84.31$\pm$0.94}&\textbf{84.88$\pm$1.07} \\
\midrule
\midrule
\parbox[t]{0.1mm}{\multirow{12}{*}{\rotatebox{90}{CIFAR-10+CIFAR-100}}}& CE & &87.90$\pm$0.38&87.87$\pm$0.33&87.81$\pm$0.21&86.77$\pm$0.35&86.32$\pm$0.29&86.35$\pm$0.49&82.56$\pm$0.83 &83.81$\pm$0.32&83.87$\pm$0.32&75.55$\pm$0.87&77.85$\pm$0.58&78.40$\pm$0.42\\
& GCE & &89.65$\pm$0.29&89.33$\pm$0.30&89.68$\pm$0.32&\textbf{89.61$\pm$0.27}&88.26$\pm$0.21&88.87$\pm$0.24& 86.16$\pm$0.72&85.80$\pm$0.18&85.64$\pm$0.38&82.75$\pm$0.44&81.38$\pm$0.78&82.47$\pm$0.51 \\
& APL & &89.01$\pm$0.12&89.01$\pm$0.25&88.42$\pm$0.53&89.23$\pm$0.37&88.14$\pm$0.33&88.23$\pm$0.19&86.75$\pm$0.26 &86.89$\pm$0.20&86.01$\pm$0.36&83.17$\pm$0.40&80.85$\pm$0.38&83.11$\pm$0.30\\
& DMI & &89.09$\pm$0.17&88.47$\pm$0.86&88.97$\pm$0.45&87.62$\pm$0.64&87.98$\pm$0.39&87.66$\pm$0.48&83.92$\pm$0.79&84.02$\pm$0.96&84.57$\pm$0.76&80.27$\pm$0.62&80.16$\pm$0.99&79.81$\pm$1.38\\
& NLNL &&89.27$\pm$0.24&89.88$\pm$0.59&90.06$\pm$0.09&87.06$\pm$0.20&87.55$\pm$0.28&87.77$\pm$0.25&83.72$\pm$0.27&84.00$\pm$0.19&84.25$\pm$0.27&78.02$\pm$0.43&79.10$\pm$0.67&81.04$\pm$1.16\\
& Co-teaching& &90.06$\pm$0.16&90.01$\pm$0.33&90.01$\pm$0.12&86.47$\pm$0.15&87.27$\pm$0.33&87.39$\pm$0.30&75.34$\pm$1.60&76.52$\pm$0.87&76.08$\pm$2.73&47.56$\pm$2.39&44.91$\pm$2.81&42.60$\pm$2.48\\
& Co-teaching+&&88.28$\pm$0.47&88.48$\pm$0.50&87.52$\pm$0.16&86.34$\pm$0.67&86.18$\pm$0.19&84.25$\pm$0.23&66.59$\pm$6.22&70.97$\pm$9.27&59.27$\pm$9.06&16.77$\pm$5.29&10.87$\pm$1.63&14.26$\pm$3.01\\
& JoCor&&88.13$\pm$0.20&88.77$\pm$0.21&89.25$\pm$0.14&79.71$\pm$0.73&81.56$\pm$0.95&82.14$\pm$0.15&65.88$\pm$1.19&67.50$\pm$0.90&68.35$\pm$0.62&39.90$\pm$3.62&37.67$\pm$3.24&40.07$\pm$3.23 \\
& S2E&&89.32$\pm$1.42&89.21$\pm$1.09&89.36$\pm$1.45&88.18$\pm$2.31&87.51$\pm$2.31&86.07$\pm$3.17&79.21$\pm$4.81&81.52$\pm$3.62&80.15$\pm$3.61&54.51$\pm$4.65&66.79$\pm$4.76&66.73$\pm$4.76 \\
& Forward & &88.25$\pm$0.14&88.03$\pm$0.92&87.10$\pm$0.47&85.21$\pm$0.93&85.22$\pm$1.93&86.23$\pm$0.80&81.30$\pm$0.54&83.98$\pm$0.75&81.77$\pm$1.06&80.02$\pm$2.42&79.68$\pm$2.87&78.05$\pm$3.08\\
& T-Revision&&88.81$\pm$0.09&89.05$\pm$0.19&89.57$\pm$0.32&86.63$\pm$0.24&87.33$\pm$0.28&87.26$\pm$0.48&85.78$\pm$0.93 &85.63$\pm$1.25&85.36$\pm$1.28&82.49$\pm$0.97&82.20$\pm$1.05&82.27$\pm$1.73\\
\midrule
& Extended T& &90.27$\pm$0.20&90.34$\pm$0.37&90.03$\pm$0.80&89.52$\pm$0.71&89.32$\pm$0.32&89.42$\pm$0.36&87.33$\pm$0.67&87.26$\pm$0.43&\textbf{87.39$\pm$0.90}&84.06$\pm$1.35&84.07$\pm$1.42&84.62$\pm$1.40 \\
& Extended T-2& &90.26$\pm$0.18&90.38$\pm$0.40&\textbf{90.17$\pm$0.58}&89.54$\pm$0.68&\textbf{89.73$\pm$0.65}&89.60$\pm$0.74&87.02$\pm$0.50&87.13$\pm$0.32&86.74$\pm$1.56&84.03$\pm$0.72&\textbf{84.10$\pm$0.87}&\textbf{84.74$\pm$0.92} \\
& Extended T-3& &\textbf{90.33$\pm$0.25}&\textbf{90.41$\pm$0.39}&90.14$\pm$0.39&89.47$\pm$0.53&89.37$\pm$0.65&\textbf{89.64$\pm$0.45}&\textbf{87.51$\pm$0.72}&\textbf{87.33$\pm$0.70}&87.30$\pm$1.08&\textbf{84.38$\pm$0.92}&83.92$\pm$1.17&84.59$\pm$1.14 \\
\midrule
\midrule
\parbox[t]{0.1mm}{\multirow{12}{*}{\rotatebox{90}{CIFAR-10+ImageNet32}}}& CE & &90.18$\pm$0.12&90.59$\pm$0.32&90.54$\pm$0.52&87.69$\pm$0.64&88.97$\pm$0.34&89.02$\pm$0.52&82.39$\pm$0.57&84.64$\pm$1.16&86.03$\pm$0.88&75.94$\pm$0.52&79.71$\pm$1.22&81.05$\pm$2.57\\
& GCE & &90.59$\pm$0.15& 90.63$\pm$0.17&\textbf{90.73$\pm$0.12}&90.01$\pm$0.34&89.98$\pm$0.37&89.20$\pm$0.68&86.07$\pm$0.51&86.25$\pm$0.63&86.37$\pm$0.42&82.92$\pm$1.07&82.03$\pm$1.89&82.36$\pm$0.82 \\
& APL & &89.02$\pm$0.48&89.37$\pm$0.23&88.65$\pm$1.07&88.92$\pm$0.64&89.07$\pm$0.23&88.65$\pm$0.29&84.64$\pm$0.50 &83.88$\pm$0.79&84.56$\pm$1.01&78.62$\pm$3.46&79.39$\pm$3.87&80.67$\pm$0.86\\
& DMI & &90.48$\pm$0.09&\textbf{90.77$\pm$0.23}&90.71$\pm$0.17&89.90$\pm$0.19&89.16$\pm$0.32&89.27$\pm$0.37&85.62$\pm$0.83&85.02$\pm$0.77&85.32$\pm$0.80&77.53$\pm$2.08&80.45$\pm$1.39&80.19$\pm$2.68\\
& NLNL &&89.92$\pm$0.28&89.93$\pm$0.18&89.97$\pm$0.11&87.06$\pm$0.22&87.33$\pm$0.24&87.50$\pm$0.48&83.30$\pm$0.49&84.70$\pm$0.53&85.02$\pm$0.50&76.57$\pm$0.61&79.66$\pm$0.82&80.05$\pm$1.02\\
& Co-teaching& &90.31$\pm$0.06&90.20$\pm$0.07&90.37$\pm$0.20&86.48$\pm$0.58&87.66$\pm$0.31&88.34$\pm$0.25&75.60$\pm$1.04&77.33$\pm$1.50&78.53$\pm$3.61&45.89$\pm$2.15&48.09$\pm$3.09&46.67$\pm$5.90\\
& Co-teaching+&&89.38$\pm$0.23&89.36$\pm$0.40&89.01$\pm$0.65&87.90$\pm$0.54&87.82$\pm$0.23& 86.51$\pm$0.38&76.58$\pm$3.33&79.76$\pm$4.21&75.20$\pm$3.00&10.34$\pm$0.47&10.73$\pm$0.84&20.07$\pm$9.62\\
& JoCor&&88.18$\pm$0.18&88.90$\pm$0.16&89.41$\pm$0.12&80.89$\pm$0.51&82.62$\pm$0.71&83.29$\pm$0.46&67.16$\pm$1.25&68.69$\pm$1.02&71.00$\pm$1.21&43.63$\pm$2.87&40.70$\pm$4.62&38.88$\pm$4.26 \\
& S2E&&89.43$\pm$0.75&89.63$\pm$1.21&89.53$\pm$0.99&88.72$\pm$2.38&88.72$\pm$1.03&87.70$\pm$2.34&82.24$\pm$5.18&82.71$\pm$2.82&83.10$\pm$1.64&60.08$\pm$7.93&75.43$\pm$4.44&78.84$\pm$3.28\\
& Forward & &89.34$\pm$0.96&89.33$\pm$0.53&89.14$\pm$0.73&88.92$\pm$0.73&88.53$\pm$0.62&88.32$\pm$0.98&84.58$\pm$0.71&83.49$\pm$0.74&83.77$\pm$0.97&76.02$\pm$3.09&76.82$\pm$3.46&78.89$\pm$4.52\\
& T-Revision&&89.72$\pm$0.28&90.01$\pm$0.24&89.95$\pm$0.37&89.45$\pm$0.65&89.09$\pm$0.78&89.02$\pm$0.96&85.72$\pm$0.93&85.25$\pm$1.01&85.03$\pm$1.78&81.17$\pm$2.77&81.34$\pm$2.53&81.37$\pm$2.86\\
\midrule
& Extended T& &90.62$\pm$0.13&90.47$\pm$0.28&90.63$\pm$0.19&90.17$\pm$0.32&90.10$\pm$0.29&90.18$\pm$0.50&87.07$\pm$0.37&87.53$\pm$0.92&87.47$\pm$0.93&84.03$\pm$0.94&84.51$\pm$1.07&84.57$\pm$1.96 \\
& Extended T-2& &\textbf{90.73$\pm$0.30}&90.60$\pm$0.48&90.68$\pm$0.29& \textbf{90.26$\pm$0.75}&90.29$\pm$0.47&90.23$\pm$1.02&\textbf{87.62$\pm$0.77}&87.58$\pm$0.76&87.42$\pm$1.17&\textbf{84.32$\pm$0.93}&84.42$\pm$1.25&84.60$\pm$1.20 \\
& Extended T-3& &90.69$\pm$0.26&90.58$\pm$0.18&\textbf{90.73$\pm$0.21}&\textbf{90.26$\pm$0.68}&\textbf{90.34$\pm$0.52}&\textbf{90.31$\pm$1.06}&87.49$\pm$0.82&\textbf{87.62$\pm$1.04}&\textbf{87.57$\pm$1.34}&84.19$\pm$1.18&\textbf{84.52$\pm$0.96}&\textbf{84.64$\pm$1.48} \\
\bottomrule
\end{tabular}
\caption
{
Mean and standard deviations of test accuracies (\%) on synthetic noisy datasets with different noise settings. The best experimental results are in bolded.
}
\label{tab:cifar}
\end{table*}
\subsubsection{Classification accuracy}
We report comprehensive experimental results on the synthetic datasets in Table \ref{tab:cifar}. For CIFAR-10+SVHN, we can clearly see that the proposed method consistently outperform the the prior state-of-the-art approaches for learning with mixed label noise. Specially, in the cases of high label noise rates, e.g., $\tau=0.6$ and $\tau=0.8$, our method Extended T-3 achieves the best classification performance. This means that the proposed cluster-dependent transiton matrix can reduce the estimation error brought by randomness and uncertainty in the estimation process, which helps lead to better classifiers. For CIFAR-10+CIFAR-100, our method achieves the superior performance again, and get ahead of the other methods in almost all cases. Note that the baseline GCE achieves the best performance when $\tau=0.4$ and $\rho=0.25$. The reasons may be that the open-set label noise rate and overall label noise rate are both low, and GCE can cope with the symmetric closed-set label noise well in this case. However, when in the face of more challenging cases, our method is better for handling the mixed label noise. For instance, when $\tau=0.8$ and $\rho=0.75$, the open-set label noise rate and overall label noise rate are both high, the proposed Extended T achieves a more than 2\% lead over GCE. Compared with the model-based baselines, e.g., Forward and T-Revision, the proposed method always works better, which means that we can effectively extend to model the mixed label noise and thus obtain better robustness. For CIFAR-10+ImageNet32, in the easier cases, e.g., $\tau=2$, we can see that DMI and GCE perform better than our method in some cases. In more challenging cases, the proposed method again surpasses all baselines. To sum up, the synthetic experiments reveal that our method is powerful in handling mixed label noise, particularly in the case of high noise rates.
\subsubsection{Estimation results}
We report the estimation error of the transition matrix $T^{\circ}$ for open-set label noise, and of the transition matrix $T^{\star}$ for mixed label noise. Note that only our method extends the traditional transition matrix to model the open-set label noise and further model the mixed label noise. We thus do not report the estimation results of other model-based methods. The experiments are conducted on CIFAR10+SVHN by exploiting the proposed Extended T. The experimental settings are same as before. The estimation errors are calculated with $\ell_1$ norm. The results are presented in Fig.~\ref{fig2}.
\subsection{Experiments on real-world face datasets}\label{sec:5.3}
\subsubsection{Datasets}
\textbf{Training data.} The training datasets include VggFace-2 \cite{cao2018vggface2} and MS1MV0 \cite{guo2016ms}. VggFace-2 is a noisy dataset collected from Google image search. MS1MV0 is a raw dataset with a large amount of noisy labels \cite{wang2018devil,deng2020subcenter}. \\
\textbf{Test data.} We use four popular benchmarks as test datasets, including CFP \cite{sengupta2016frontal}, AgeDB \cite{moschoglou2017agedb}, CALFW \cite{zheng2017cross}, and CPLFW \cite{zheng2018cross}. CFP consist of face images of celebrities in frontal and profile views. AgeDB contains images annotated with accurate to the year, noise-free labels. CALFW considers a more general cross-age situation and provides a face image set with large intra-class variations. CPLFW is similar to CALFW, but considers a cross-pose case. We summarize the important statistics of the used datasets in Appendix B.
\subsubsection{Experimental settings}
\textbf{Data processing.} We follow ArcFace \cite{deng2019arcface} to generate the normalised face crops by exploiting five facial points (two eyes, nose tip, and two mouth corners) predicted by RetinaFace \cite{deng2020retinaface}. The size of the face crops is 112$\times$112. \\
\textbf{Network structure and activation function.} To be fair, in this paper, we employ the same architecture and the activation function for testing different baselines. We use MobileFaceNet \cite{chen2018mobilefacenets} and Arc-Softmax \cite{deng2019arcface}, which are popular in the face recognition task. The dimension of the face embedding feature is 512. For the angular margin $m$ and feature scale $s$, we set 0.5 and 32, respectively. \\
\textbf{Training and testing.} At the train stage, we train the deep models with SGD with momentum 0.9, with total batch size 512 on 4 GPUs parallelly and weight decay $5\times10^{-4}$. For learning an initial classifer, the learning rate is initially 0.1 and divided by 10 at the 5th, 10th, and 15th epochs. We set 20 epochs in total. For learning the classifier and slack variable, Adam is used and the learning rate is changed to $5\times10^{-7}$. At the test stage, we use MobileFaceNet to extract the 512-$D$ feature embeddings of test face images. We follow the unrestricted with labelled outside data protocol \cite{huang2008labeled} to report the verification performance on test face datasets.
\begin{figure}[t]
\centering
\subfigure[]{\label{fig2:a}
\begin{minipage}[t]{0.5\linewidth}
\centering
\includegraphics[width=2.20in]{estimation1.pdf}
\label{error}
\end{minipage}
}%
\subfigure[]{\label{fig2:b}
\begin{minipage}[t]{0.5\linewidth}
\centering
\includegraphics[width=2.20in]{estimation2.pdf}
\end{minipage}
}
\centering
\caption{Illustration of the transition matrix estimation error. Figure (a) illustrates the estimation error for modeling open-set label noise. Figure (b) illustrates the estimation error for modeling mixed label noise. The error bar for standard deviation in each figure has been shaded.}
\label{fig2}
\end{figure}
\subsubsection{Experimental results}
\begin{table}[h]
\centering
\begin{tabular}{c|cccc|c}
\toprule
Method&CFP&AgeDB&CALFW&CPLFW&Ave.\\
\midrule
CE&95.30&92.69&89.94&85.97&90.98\\
GCE&94.26&91.06&89.98&85.28&90.15\\
APL&94.16&90.29&88.32&85.09&89.47\\
DMI&94.39&92.83&90.06&85.86&90.79\\
NLNL&86.74&87.63&84.79&80.06&84.81\\
Co-teaching&95.47&92.53&89.58&85.32&90.73\\
Co-teaching+&95.26&92.01&85.12&85.19&89.40\\
JoCor&92.32&90.97&84.58&82.11&87.50\\
S2E&92.09&91.64&89.70&84.63&89.52\\
Forward&95.07&92.40&89.10&85.79&90.59\\
T-Revision&95.38&92.79&89.86&85.94&90.99\\
\midrule
Extended T&95.57&93.06&90.33&86.24&91.30\\
Extended T-2&95.59&\textbf{93.15}&90.42&86.36&91.38\\
Extended T-3&\textbf{95.73}&93.14&\textbf{90.67}&\textbf{86.52}&\textbf{91.51}\\
\bottomrule
\end{tabular}
\caption
{
Test accuracies (\%) of different methods training on VggFace-2. The best results are in bolded.
}
\label{tab:vggface2}
\end{table}
\begin{table}[h]
\centering
\begin{tabular}{c|cccc|c}
\toprule
Method&CFP&AgeDB&CALFW&CPLFW&Ave.\\
\midrule
CE&88.79&92.35&91.36&82.92&88.86\\
GCE&89.02&91.87&91.32&82.77&88.75\\
APL&87.06&91.55&91.29&82.02&87.98\\
DMI&89.02&92.18&90.94&83.01&88.79\\
NLNL&83.06&85.47&84.21&74.02&81.69\\
Co-teaching&91.25&93.05&91.24&84.02&89.89\\
Co-teaching+&91.36&91.87&90.93&84.53&89.67\\
JoCor&86.16&89.30&88.09&79.84&85.85\\
S2E&91.52&91.61&91.03&81.47&88.91\\
Forward&90.02&92.32&91.28&82.95&89.14\\
T-Revision&90.29&92.49&91.60&83.82&89.55\\
\midrule
Extended T&90.33&92.59&91.65&83.91&89.62\\
Extended T-2&91.65&93.05&92.30&84.67&90.42\\
Extended T-3&\textbf{92.08}&\textbf{93.71}&\textbf{92.61}&\textbf{85.31}&\textbf{90.93}\\
\bottomrule
\end{tabular}
\caption
{
Test accuracies (\%) of different methods training on MS1MV0. The best results are in bolded.
}
\label{tab:ms1mv0ab}
\end{table}
We use two training datasets, i.e., VggFace-2 and MS1MV0, to separately train the deep networks. In Table \ref{tab:vggface2} and \ref{tab:ms1mv0ab}, we show the results of the proposed method and baselines on CFP \cite{sengupta2016frontal}, AgeDB \cite{moschoglou2017agedb}, CALFW \cite{zheng2017cross}, and CPLFW \cite{zheng2018cross}, respectively. We can observe that the most of the results obtained by training on VggFace-2 are higher than the results on MS1MV0. It is because that MS1MV0 contains more noisy labels, and therefore is more challenging \cite{deng2020subcenter,wang2019co}. To our method, we can effectively model the mixed label noise, and thus can achieve higher performance than the baselines. Specifically, when training on VggFace-2, Extended T consistently outperforms the baselines. Note that NLNL achieve poor performance in most cases, it is because NLNL exploits negative learning and choose the complementary labels to avoid wrong information brought by noisy labels. Real-world datasets are more complex, though NLNL aims to avoid wrong information, it is hard to select the possible clean labels included in complementary labels. The improvement of classification performance brought by changing the number of cluster is not too obvious. It is because that the label noise rate of VggFace-2 is low \cite{wang2019co}. When training on MS1MV0, we can see that the proposed method surpass the baselines again. Specifically, compared with Forward and T-Revision, the proposed method leads them clearly. Compared with the methods which empirically work well, the proposed method still outperforms them. It is worth noting that the cluster-dependent transition matrices brings a significant performance improvement. The improvement shows that the cluster-dependent transition matrices can model the complex label noise more accurately in this realistic scenario.
\section{Conclusion}\label{sec6}
In this paper, we investigate into learning with mixed closed-set and open-set noisy labels, which is more practical but lacks systematic study in current works. We extend the traditional transition matrix to be able to model mixed label noise, and exploit the cluster-dependent extended transition matrices to better approximate the instance-dependent label noise. Empirical evaluations on synthetic and real-world datasets show the effectiveness of the proposed method for modeling label noise and leading to better classifiers. We believe that our work will urge the research community to explore the robustness of algorithms in this realistic noisy label scenario.
\newpage
|
\section{Introduction}
In this paper we prove $L^{2}$ instability of the soliton for the focusing, mass-critical, generalized KdV equation
\begin{equation}\label{1.1}
u_{t} = -(u_{xx} + u^{5})_{x}, \qquad u(0,x) = u_{0} \in L^{2}(\mathbb{R}).
\end{equation}
This equation is called mass-critical because the scaling leaving \eqref{1.1} invariant, i.e.
$$
u(t, x) \mapsto \lambda^{\frac{1}{2}} u\left(\lambda^{3} t, \lambda x\right)
$$
leaves the $L^{2}$ norm, or mass, invariant. The mass of a solution, defined by
$$
M(u(t)):=\int_{\mathbb{R}}|u(t, x)|^{2} d x
$$
is conserved.
Recently, \cite{dodson2017global} proved that the defocusing, mass-critical generalized KdV equation
\begin{equation}\label{1.2}
u_{t} = -(u_{xx} - u^{5})_{x}, \qquad u(0,x) = u_{0} \in L^{2}(\mathbb{R}),
\end{equation}
is globally well-posed and scattering for any $u_{0} \in L^{2}(\mathbb{R})$. The proof of the defocusing result used the concentration compactness method. Namely, a result of \cite{killip2009mass} combined with a scattering result of \cite{defocusing2016global} for the defocusing nonlinear Schr{\"o}dinger equation,
\begin{equation}\label{1.3}
i u_{t} + u_{xx} = |u|^{4} u, \qquad u(0,x) = u_{0} \in L^{2}(\mathbb{R}),
\end{equation}
implies that for scattering to fail for $(\ref{1.2})$, there must exist a nonzero, almost periodic solution to $(\ref{1.2})$.
\begin{definition}[Almost periodic solution]\label{d1.1}
Suppose $u$ is a strong solution to $(\ref{1.1})$ on the maximal interval of existence $I$. Such a solution $u$ is said to be almost periodic (modulo symmetries) if there exist continuous functions $ N(t) : I \rightarrow (0, \infty)$ and $x(t) : I \rightarrow \mathbb{R}$, such that
\begin{equation}\label{1.4}
\{ v(t,x) = N(t)^{-1/2} u(t, N(t)^{-1} x + x(t)) : t \in I\}
\end{equation}
is contained in a compact subset of $L^{2}(\mathbb{R})$. See also section \ref{sec:almost:per} for an equivalent condition.
\end{definition}
Then \cite{defocusing2016global} proved that in the defocusing case, there does not exist a nonzero, almost periodic solution to $(\ref{1.2})$, which implies scattering for the defocusing equation $(\ref{1.2})$. The proof used an interaction Morawetz estimate based upon the argument in \cite{Tao2006TwoRO}, which proved there does not exist a soliton for the defocusing, generalized KdV equation. \medskip
For the focusing generalized KdV equation, there exists the soliton $u(t,x) = Q(x - t)$, where
\begin{equation}\label{1.6}
Q(x) = \frac{3^{1/4}}{\cosh^{1/2}(2x)} > 0.
\end{equation}
The function $Q(x)$ solves the elliptic equation
\begin{equation}\label{1.7}
Q_{xx} + Q^{5} = Q,
\end{equation}
so therefore, $Q(x - t)$ solves $(\ref{1.1})$. Note that $Q(x - t)$ is an almost periodic solution to $(\ref{1.1})$. Meanwhile, for the focusing, mass-critical nonlinear Schr{\"o}dinger equation,
\begin{equation}\label{1.5}
i u_{t} + u_{xx} = -|u|^{4} u, \qquad u(0,x) = u_{0} \in L^{2}(\mathbb{R}),
\end{equation}
$u(t,x) = e^{it} Q(x)$ gives a soliton solution.\medskip
The paper \cite{dodson2016global} proved that the focusing nonlinear Schr{\"o}dinger equation $(\ref{1.5})$ is scattering for initial data below the ground state, $\| u_{0} \|_{L^{2}} < \| Q \|_{L^{2}}$. It is conjectured that the same is also true for the focusing, generalized KdV equation.
\begin{conjecture}\label{c1.2}
If $\| u_{0} \|_{L^{2}} < \| Q \|_{L^{2}}$, then the solution to $(\ref{1.1})$ is globally well-posed and scattering.
\end{conjecture}
It can be verified that if Conjecture $\ref{c1.2}$ is true, then this implies that there does not exist an almost periodic solution to $(\ref{1.1})$ below the ground state.
\begin{conjecture}\label{c1.3}
There does not exist a nonzero, almost periodic solution $u$ to $(\ref{1.1})$ satisfying $0 < \| u \|_{L^{2}} < \| Q \|_{L^{2}}$.
\end{conjecture}
However, unlike in the defocusing case, Conjecture $\ref{c1.3}$ does not imply Conjecture $\ref{c1.2}$. This is because \cite{killip2009mass} states that if $(\ref{1.5})$ is globally well-posed and scattering when $\| u \|_{L^{2}} < \| Q \|_{L^{2}}$, Conjecture $\ref{c1.3}$ implies Conjecture $\ref{c1.2}$ when $0 < \| u \|_{L^{2}} < \sqrt{\frac{5}{6}} \| Q \|_{L^{2}}$. In the defocusing case, the presence of the constant $\sqrt{\frac{5}{6}}$ is unimportant, because scattering for the defocusing nonlinear Schr{\"o}dinger equation holds for any finite mass. However, in the focusing case, the constant $\sqrt{\frac{5}{6}}$ becomes quite important, since it is conjectured that $(\ref{1.1})$ scatters for any $\| u_{0} \|_{L^{2}} < \| Q \|_{L^{2}}$.\medskip
Conjecture $\ref{c1.2}$ would also imply instability of the soliton in an $L^{2}$-sense. For any initial data $u_{0} \in L^{2}$, $\| u_{0} \|_{L^{2}} < \| Q \|_{L^{2}}$, the solution to $(\ref{1.1})$ would scatter to a free solution, and thus the solution would approach distance
$$(\| Q \|_{L^{2}}^{2} + \| u_{0} \|_{L^{2}}^{2})^{1/2}$$ from any translation or rescaling of the soliton as $t \rightarrow \pm \infty$.\medskip
In a remarkable series of works, \cite{merle2001existence}, \cite{martel2000liouville}, \cite{martel2001instability}, \cite{martel2002blow}, \cite{merle1998blow}, \cite{martel2002stability} proved, among many nice results, the instability of the soliton in an $H^{1}$ sense, for initial data with mass greater than or equal to the soliton. In fact, they proved something more, that there initial data arbitrarily close to the soliton in $H^{1}$-norm, which eventually move away from the soliton in an $L^{2}$-sense. See \cite{martel2014blow} and \cite{MR3372073} for results in a weighted $L^{2}$ space.\medskip
In this paper we show that there are no almost periodic solutions to $(\ref{1.1})$ which are uniformly close to $Q(x)$ in $ L^2_x $ modulo symmetries.
\begin{definition}
If a maximal-lifespan strong solution $ u $ to \eqref{1.1} on $ I $ satisfies
\begin{equation}\label{1.8}
\sup_{t \in I} \inf_{\lambda_{0}, x_{0}} \| u(t,x) - \frac{1}{\lambda_{0}^{1/2}} Q(\frac{x - x_{0}}{\lambda_{0}}) \|_{L^{2}(\mathbb{R})} \leq \delta
\end{equation}
then we say $ u $ is $ \delta $-close to $ Q $. It is readily seen that the infimum is attained and the values $ \lambda_{0}(t), x_{0}(t)$ which attain the minimum can be chosen to be continuous.
\end{definition}
The main result is
\begin{theorem}\label{main.thm}
There exists $\delta > 0$ sufficiently small such that there does not exist a maximal-lifespan solution to $(\ref{1.1})$ with $\| u_{0} \|_{L^{2}} < \| Q \|_{L^{2}}$ satisfying \eqref{1.8}.
\end{theorem}
In other words, Theorem \ref{main.thm} states that there no solutions $ \delta $-close to $ Q $. A consequence of this fact is that for any initial data satisfying $\| u_{0} \|_{L^{2}} < \| Q \|_{L^{2}}$, the solution to $(\ref{1.1})$ with such initial data must eventually move a distance $\delta > 0$ away from the soliton, modulo translations and rescalings, where $\delta > 0$ is a small, fixed constant.\medskip
We split Theorem \ref{main.thm} into two statements. The first part reduces the study to the existence of almost-periodic solutions.
\begin{theorem} \label{almost_periodic_prop}
Suppose $ u:I \times \mathbb{R} \to \mathbb{R} $ is a maximal-lifespan strong solution with $ \vn{u_0}_{L^2} < \vn{Q}_{L^2} $ to the mass-critical focusing gKdV equation \eqref{1.1} which is $ \delta $-close to $ Q $. Then, if $ \delta $ is small enough, there exists an almost periodic modulo symmetries maximal-lifespan (strong) solution $ v $ which is $ \delta $-close to $ Q $ with mass less than $ Q $.
\end{theorem}
The proof is given in Section \ref{Sec.reduction.to.almost.periodic} and it relies essentially on a Palais-Smale result based on the Airy linear profile decomposition, decoupling and an approximation of gKdV solutions by NLS solutions, which are tools developed in \cite{shao2009}, \cite{Tao2006TwoRO}, \cite{killip2009mass} and reviewed in Section \ref{Sec.prelim}. See \cite{10.1093/imrn/rny164} for a similar argument in the case of the mass-critical nonlinear Schr{\"o}dinger equation.\medskip
Once we have this reduction, we prove that such solutions cannot exist.
\begin{theorem}\label{t1.1}
There are no almost periodic solutions to \eqref{1.1} with mass less than $ Q $ which are $ \delta $-close to $ Q $, if $ \delta $ is small enough.
\end{theorem}
The proof of Theorem $\ref{t1.1}$ combines the ideas of \cite{dodson2017global} and in \cite{merle2001existence}, \cite{martel2000liouville}, \cite{martel2001instability}, \cite{martel2002blow}, \cite{merle1998blow}, \cite{martel2002stability}, \cite{martel2014blow}, and \cite{MR3372073}. The proof of scattering in \cite{dodson2017global} reduced an almost periodic solution to three scenarios: a self-similar solution, a double rapid cascade solution, and a quasisoliton solution. The arguments used in excluding the self-similar and double rapid cascade solutions can also be used to exclude an almost periodic solution to $(\ref{1.1})$ with mass less than the soliton, regardless of whether it is close to the soliton or not. However, in the defocusing case, the interaction Morawetz estimate developed in \cite{dodson2017global} used \cite{Tao2006TwoRO}, and there is no analog to \cite{Tao2006TwoRO}, even for a solution with mass below the mass of the soliton. Instead, we rely on the Morawetz arguments in \cite{merle2001existence}, \cite{martel2000liouville}, \cite{martel2001instability}, \cite{martel2002blow}, \cite{merle1998blow}, \cite{martel2002stability}, \cite{martel2014blow}, and \cite{MR3372073}. These Morawetz estimates depend very much on the fact that the solution is close to the soliton in an $L^{2}$-sense, and can be used to show that a solution cannot stay close to the soliton for the entire time of its existence.\medskip
\textbf{Acknowledgements:} The authors are grateful to Jonas L{\"u}hrmann, Yvann Martel, and Daniel Tataru for several helpful conversations concerning this problem. The first author also acknowleges the support of NSF grant DMS-$1764358$.
\section{Preliminaries} \label{Sec.prelim}
\subsection{Notation and linear estimates}
We will write $ x \lesssim y $ to denote $ x \leq C y$ for a uniform constant $ C >0 $. We denote $ \jb{x} = (1+x^2)^{1/2}$.
The one-dimensional Fourier transform is defined by
$$
\hat{f}(\xi):=\frac{1}{(2 \pi)^{1 / 2}} \int_{\mathbb{R}} e^{-i x \xi} f(x) \,\mathrm{d} x, \qquad \xi \in \mathbb{R}
$$
which is used to define the linear propagator and fractional differentiation operators by
$$
\widehat{e^{-t \partial_{x}^{3}} f} (\xi)= e^{-it \xi^3} \hat{f} (\xi), \qquad \widehat{\left|\partial_{x}\right|^{s} f}(\xi):=|\xi|^{s} \hat{f}(\xi).
$$
For an interval $ I $ one considers the mixed norms on $ I \times \mathbb{R} $
\begin{align*}
\vn{F}_{L_{t}^{p} L_{x}^{q}(I \times \mathbb{R})} = \Big(\int_{I}\big(\int_{\mathbb{R}}|F(t, x)|^{q} \,\mathrm{d} x \big)^{p / q} \,\mathrm{d} t \Big)^{1 / p}, \\
\vn{F}_{L_{x}^{p} L_{t}^{q}(I \times \mathbb{R}) } = \Big(\int_{\mathbb{R}}\big(\int_{I}|F(t,x )|^{q} \,\mathrm{d} t\big)^{p / q} \,\mathrm{d} x\Big)^{1 / p},
\end{align*}
with the standard modification when $ p = \infty $ or $ q = \infty $.
We recall the dispersive estimate
$$
\left\|e^{-t \partial_{x}^{3}} u_{0}\right\|_{L_{x}^{p}(\mathbb{R})} \lesssim t^{-\frac{2}{3}\left(\frac{1}{2}-\frac{1}{p}\right)}\left\|u_{0}\right\|_{L_{x}^{p^{\prime}}(\mathbb{R})}, \qquad 2 \leq p \leq \infty.
$$
We will consider weakly convergent sequences in $ L_{x}^{2}(\mathbb{R}) $, i.e. $ f_n \rightharpoonup f $ if
$$ \langle f_n, g \rangle = \int_{\mathbb{R}} f_n(x) \bar{g}(x) \,\mathrm{d} x \to \int_{\mathbb{R}} f(x) \bar{g}(x) \,\mathrm{d} x \qquad \forall \ g \in L_{x}^{2}(\mathbb{R}).
$$
By approximation arguments, it sufficies to check this condition for all $ g \in C_c(\mathbb{R}) $.
A basic fact which we will be using tacitly is that if $ f_n \rightharpoonup f $ then $$ \vn{f}_{L_{x}^{2}(\mathbb{R})} \leq
\liminf_{n \to \infty} \vn{f_n}_{L_{x}^{2}(\mathbb{R})}. $$
\subsection{Solutions to gKdV}
Throughout this paper we will consider strong solutions, defined as follows.
\begin{definition} \label{deff} \
\begin{enumerate}
\item $A$ function $u: I \times \mathbb{R} \rightarrow \mathbb{R}$ on a non-empty interval $0 \in I \subset \mathbb{R}$ is $a$ (strong) solution to \eqref{1.1} if it lies in the class $C_{t}^{0} L_{x}^{2}(J \times \mathbb{R}) \cap L_{x}^{5} L_{t}^{10}(J \times \mathbb{R})$ for any compact $J \subset I$ and obeys the Duhamel formula
$$
u(t)=e^{-t \partial_{x}^{3}} u_{0}-\int_{0}^{t} e^{-(t-\tau) \partial_{x}^{3}} \partial_{x}\left(u^{5}(\tau)\right) d \tau.
$$
We say that $u$ is a maximal-lifespan solution if the solution cannot be extended to any strictly larger interval. We say that u is a global solution if $I=\mathbb{R}$.
\item The scattering size is defined to be
$$S_{I}(u)=\int_{\mathbb{R}}\left(\int_{I}|u(t, x)|^{10} d t\right)^{1 / 2} d x=\|u\|_{L_{x}^{5} L_{t}^{10}(I \times \mathbb{R})}^{5}. $$
\item We say that a solution $u$ to \eqref{1.1} blows up forward in time if there exists $t_{1} \in I$ such that $
S_{\left[t_{1}, \sup (I)\right)}(u)=\infty
$
and that $u$ blows up backward in time if there exists a time $t_{1} \in I$ such that
$
S_{\left(\inf (I), t_{1}\right]}(u)=\infty.
$
\item We say that $u$ scatters forward/backwards in time if there exists a unique $u_{\pm} \in L_{x}^{2}(\mathbb{R})$ such that
\begin{equation} \label{scattering}
\lim _{t \rightarrow \pm \infty}\left\|u(t)-e^{-t \partial_{x}^{3}} u_{\pm}\right\|_{L_{x}^{2}(\mathbb{R})}=0.
\end{equation}
\item The symmetry group $ G $ is defined as the set of unitary transformations
$$
G=\{ g_{x_{0}, \lambda}: L_{x}^{2}(\mathbb{R}) \rightarrow L_{x}^{2}(\mathbb{R}) | \ (x_0,\lambda) \in \mathbb{R} \times (0,\infty), \ g_{x_{0}, \lambda} f (x):=\lambda^{-\frac{1}{2}} f\left(\lambda^{-1}\left(x-x_{0}\right)\right)
\}.
$$
For $u: I \times \mathbb{R} \rightarrow \mathbb{R}$, one defines $T_{g_{x_{0}, \lambda}} u: \lambda^{3} I \times \mathbb{R} \rightarrow \mathbb{R}$ by
$$
T_{g_{x_{0}, \lambda}} u (t, x):=\lambda^{-\frac{1}{2}} u\left(\lambda^{-3} t, \lambda^{-1}\left(x-x_{0}\right)\right).
$$
\end{enumerate}
\end{definition}
$T_{g} u$ solves \eqref{1.1} with initial data $g u_{0}$ if $ u $ is a solution. Moreover, scattering sizes are invariant
$$ S_{\lambda^3 I}( T_{g} u ) = S_{I}(u) , \qquad g \in G. $$
We note that $ G $ is a Lie group and the map $ g \mapsto T_g $ is a homomorphism. Giving the operators in $ G $ the strong operator topology, then the identification $ (x_0,\lambda) \mapsto g_{x_{0}, \lambda} $ is a homeomorphism between $ \mathbb{R} \times (0,\infty) $ and $ G $.
Thus we say $ g_{x_n, \lambda_n} \to \infty $ if $ \vm{x_n} + \lambda_n + \lambda_n^{-1} \to \infty $. Moreover, in that case $ g_{x_n, \lambda_n}$ converges to $ 0 $ in the weak operator topology.
\
The $ L^2 $ local well-posedness theory of \eqref{1.1} was established by Kenig, Ponce, Vega in \cite{KenigPonceVega}.
\begin{theorem}[Local well-posedness \cite{KenigPonceVega}] For any $u_{0} \in L_{x}^{2}(\mathbb{R})$ and $t_{0} \in \mathbb{R},$ there exists a unique solution $ u $ to \eqref{1.1} with $u\left(t_{0}\right)=u_{0}$ which has maximal lifespan. Let $ I $ denote the lifespan of $ u $. Then:
\begin{enumerate}
\item I is an open neighborhood of $t_{0}$.
\item If $\sup (I) / \inf (I)$ is finite then $u$ blows up forward / backward in time.
\item If sup $(I)=+\infty$ and $u$ does not blow up forward in time, then u scatters forward in time. Conversely, given $u_{+} \in L_{x}^{2}(\mathbb{R})$ there is a unique solution to \eqref{1.1} in a neighborhood of $\infty$ so that \eqref{scattering} holds. One can define scattering backward in time in a completely analogous manner.
\item If $M\left(u_{0}\right)$ is sufficiently small then $u$ is a global solution which does not blow up either forward or backward in time and
$
S_{\mathbb{R}}(u) \lesssim M(u)^{5 / 2}.
$
\item Uniformly continuous dependence on initial data holds, see Corollary \ref{sol.conv}.
\end{enumerate}
\end{theorem}
\subsection{Stability and corollaries}
The stability theory of the generalized KdV equation \eqref{1.1} is discussed in detail in \cite{killip2009mass}.
\begin{lemma}[Short time stability \cite{killip2009mass}(Lemma 3.3)] \label{short.time.stability}
Let $ I $ be an interval with $ 0 \in I $. Suppose $ \tilde{u} : I \times \mathbb{R} \to \mathbb{R} $ is a solution to
\begin{align} \label{approx.kdv}
(\partial_t + \partial_x^3) \tilde{u} + \partial_x(\tilde{u}^5) &=e, \\
\tilde{u}(0,x) &=\tilde{u}_0(x), \nonumber
\end{align}
for some function $ e $ such that
$$
\vn{\tilde{u}}_{L^{\infty}_t L^2_x(I \times \mathbb{R})} \leq M,
$$
for some $ M > 0 $. Let $ u_0 $ be such that
$$
\vn{u_0 - \tilde{u}_0}_{L^2_x} \leq M',
$$
for some $ M' \geq 0 $. Assume the smallness conditions
\begin{align}
\vn{\tilde{u}}_{L^5_x L^{10}_t(I \times \mathbb{R}) } & \leq \varepsilon_0, \\
\vn{ e^{-t \partial_x^3} (u_0 - \tilde{u}_0) }_{L^5_x L^{10}_t(I \times \mathbb{R}) } & \leq \varepsilon, \\
\vn{\vm{\partial_x}^{-1} e }_{L^1_x L^{2}_t(I \times \mathbb{R}) } & \leq \varepsilon,
\end{align}
for some small $ 0< \varepsilon< \varepsilon_0=\varepsilon_0(M,M') $. Then there exists a solution $ u:I \times \mathbb{R} \to \mathbb{R} $ to \eqref{1.1} with initial data $ u(0)=u_0 $ satisfying
\begin{align}
\vn{u-\tilde{u}}_{ L^5_x L^{10}_t(I \times \mathbb{R}) } + \vn{u^5-\tilde{u}^5}_{L^1_x L^{2}_t(I \times \mathbb{R})} & \lesssim \varepsilon ,\\
\vn{u-\tilde{u}}_{L^{\infty}_t L^2_x(I \times \mathbb{R})} +\vn{\vm{\partial_x}^{1/6} ( u-\tilde{u})}_{ L^6_{t,x} (I \times \mathbb{R})} & \lesssim M'+\varepsilon.
\end{align}
\end{lemma}
Iterating this lemma over small intervals also a long-time stability result can be obtained, see \cite[Theorem 3.1]{killip2009mass}. To keep track of the number of small intervals one uses the following bound.
\begin{lemma} \label{num.intervals}
Let $ v \in L^5_x L^{10}_t(J \times \mathbb{R}) $ for an interval $ J $. Divide $ J $ into $ N $ intervals $ [t_k,t_{k+1}] $ such that $ \vn{v}_{ L^5_x L^{10}_t([t_k,t_{k+1}] \times \mathbb{R})} \simeq \varepsilon_0 $ for every $ 1 \leq k \leq N-1 $ and $ \vn{v}_{ L^5_x L^{10}_t([t_N,t_{N+1}] \times \mathbb{R})} \lesssim \varepsilon_0 $, for a fixed $ \varepsilon_0 > 0$. Then the number of intervals $ N $ is finite and $ N \lesssim (1+ \vn{v}_{ L^5_x L^{10}_t ( J \times \mathbb{R})} / \varepsilon_0 )^{10} $.
\end{lemma}
\begin{proof}
See \cite[Thm 3.1 -first part]{killip2009mass}
\end{proof}
As a consequence, one has
\begin{corollary}[Uniformly continuous dependence on initial data] \label{sol.conv}
Consider solutions $ v \in L^5_x L^{10}_t(J \times \mathbb{R}) $ to \eqref{1.1}. For every $ \varepsilon > 0 $ there exists $ \delta = \delta(\varepsilon, \vn{v(0)}_{L^2_x} , \vn{v}_{ L^5_x L^{10}_t ( J \times \mathbb{R})} )$ such that if $ \vn{u_0-v(0)}_{L^2_x} \leq \delta $, then there exists a solution to \eqref{1.1} defined on $ J $, with initial data $ u(0)=u_0 $ such that
$$
\vn{u-v}_{ L^{\infty}_t L^2_x \cap L^5_x L^{10}_t(J \times \mathbb{R}) } \leq \varepsilon.
$$
\end{corollary}
Finally, we can use stability to prove a compactness property for the the transformations associated to solutions that are $ \delta$-close to $ Q $.
\begin{lemma} \label{group.compactness}
There exists $ \delta>0 $ such that the following statement holds. Let $ u:I \times \mathbb{R} \to \mathbb{R} $ be a strong solution to \eqref{1.1} such that
\begin{equation} \label{epp.close}
\vn{ g(t) u(t)- Q}_{L^2_x} \leq \delta, \qquad \forall \ t \in I,
\end{equation}
with $ g(0)=g_{0,1} $ (the identity), $ 0 \in I $, $ g(t) \in G $. Then for any $ t \in I $ there exists a compact set $ K_t $ depending only on $ \vm{t}, \ \vn{u}_{ L^5_x L^{10}_t([0,t] \times \mathbb{R})} $ and $ M(u) $ such that $ g(t) \in K_t $.
\end{lemma}
\begin{proof} Without loss of generality suppose $ t>0 $ is fixed. We split $ [0,t] $ into $ N $ intervals as in Lemma \ref{num.intervals} where $ \varepsilon_0 $ is the constant in Lemma \ref{short.time.stability}. Then $ N $ depends on $ M(u) $ and $ \vn{u}_{ L^5_x L^{10}_t([0,t] \times \mathbb{R})}. $
We do an induction argument. We prove that if the statement holds for $ t_k $ then it also holds for $ s \in [t_k, t_{k+1} ] $. At $ t=0 $ we have $ K_0 = \{ g_{0,1} \}. $
From $ \vn{ u(t_k)- g(t_k)^{-1} Q}_{L^2_x} \leq \delta $ using Lemma \ref{short.time.stability} we deduce
$$
\vn{u(s) - g(t_k)^{-1} Q(\cdot - \frac{s-t_k}{\lambda_k^3} ) }_{L^2} \lesssim \delta,
$$
From this and \eqref{epp.close} at $ s $ we find
$$
\vn{Q - g(s) g(t_k)^{-1} g_{(s-t_k)/\lambda_k^3,1} Q}_{L^2} \lesssim \delta.
$$
For $ \delta $ small enough, $ g(s) g(t_k)^{-1} g_{(s-t_k)/\lambda_k^3,1} $ will lie in a small compact neighborhood of the identity parametrized by $ (x,\lambda) \in [-\eta,\eta] \times [r,R] $ for some $ \eta > 0 $ and $ 0<r < 1 <R $. Therefore $ g(s) $ has to be in a compact set. Moreover, denoting $ g(t_k) = g_{x_k, \lambda_k} $, one checks inductively that
$$
\lambda_k \in [r^{k-1},R^{k-1} ], \qquad \vm{x_k} \leq (1+R)^{k-2} \eta + \frac{R^{k-1}}{r^{3 (k-2)}} t_k.
$$
which implies the stated dependence.
\end{proof}
\subsection{Almost periodicity} \label{sec:almost:per}
As a consequence of the Arzela-Ascoli theorem we know that precompactness of a family of functions in $L_{x}^{2}(\mathbb{R})$ is equivalent to it being bounded in $L_{x}^{2}(\mathbb{R})$ and the existence of a function $C(\eta) $ so that
$$
\int_{|x| \geq C(\eta)}|f(x)|^{2} \,\mathrm{d} x+\int_{|\xi| \geq C(\eta)}|\hat{f}(\xi)|^{2} \,\mathrm{d} \xi \leq \eta \qquad \forall \ \eta >0,
$$
holds for all the functions. Therefore, the almost periodicity condition \eqref{1.4} is equivalent to
$$
\int_{\vm{x-x(t)} \geq C(\eta)/N(t)} \vm{u(t,x)}^2 \,\mathrm{d} x + \int_{|\xi| \geq C(\eta) N(t)}|\hat{u}(t,\xi)|^{2} \,\mathrm{d} \xi \leq \eta \qquad \forall \ \eta >0.
$$
\subsection{The embedding of NLS into gKdV}
We now review the approximation of solutions to gKdV by certain modulated, rescaled versions of solutions to NLS discussed in \cite{Tao2006TwoRO}, \cite{Christ2003AsymptoticsFM}, \cite{killip2009mass}.
We cite the following theorem from \cite[Thm. 4.1]{killip2009mass}, which was initially conditional on the global well-posedness and scattering of the focusing NLS below the ground state, which was subsequently proved in \cite{DODSON20151589}. We will only need this theorem for small data (in which case the existence part is automatic), and specifically we will use the approximations \eqref{smoothapprox}, \eqref{unifapproxu}. Here
\begin{equation} \label{unt}
\tilde{u}_{n}^T(t, x) \vcentcolon= \left\{\begin{array}{ll}
\operatorname{Re}\left[ e^{i x \xi_{n} \lambda_n +i t (\xi_{n} \lambda_n)^{3}} V_{n}\left(3 \xi_{n} \lambda_n t, x+3 (\xi_{n} \lambda_n)^{2} t\right)\right], & \text { when }|t| \leq \frac{T}{3 \xi_{n} \lambda_n} \\
\exp \left\{-\left(t- \frac{T}{3 \xi_{n} \lambda_n}\right) \partial_{x}^{3}\right\} \tilde{u}_{n}\left(\frac{T}{3 \xi_{n} \lambda_n}\right), & \text { when } t>\frac{T}{3 \xi_{n} \lambda_n} \\
\exp \left\{-\left(t+ \frac{T}{3 \xi_{n} \lambda_n}\right) \partial_{x}^{3}\right\} \tilde{u}_{n}\left(- \frac{T}{3 \xi_{n} \lambda_n}\right), & \text { when } t<-\frac{T}{3 \xi_{n} \lambda_n}
\end{array}\right.
\end{equation}
is defined in terms of certain frequency-localized solutions $ V_n $ and $ V $ to NLS such that
\begin{equation} \label{VnV}
\vn{V_n- V}_{L_{t}^{\infty} L_{x}^{2}(\mathbb{R} \times \mathbb{R})} \to 0.
\end{equation}
\begin{theorem}[Oscillatory profiles \cite{killip2009mass}] \label{oscilatory.profiles}
Let $\phi \in L_{x}^{2}$ with $M(\phi)<2 \sqrt{\frac{6}{5}} M(Q)$. Let $ (\lambda_{n})_{n \geq 1}, (\xi_{n})_{n \geq 1} \subset(0, \infty)$, with $\xi_{n} \lambda_n \rightarrow \infty$ and let $ (t_{n} )_{n \geq 1} \subset \mathbb{R}$ such that $3 \xi_{n} \lambda_n t_{n}$ converges to some
$T_{0} \in[-\infty, \infty] .$ Then, for $n$ sufficiently large there exists a global solution $ \tilde{v}_{n}$ to \eqref{1.1} with initial data at time $t=t_{n}$ given by
$$
\tilde{v}_{n} \left(t_{n}, x\right)=e^{-t_{n} \partial_{x}^{3}} \operatorname{Re}[ e^{i x \xi_{n} \lambda_n } \phi(x)]
$$
The solution obeys the global spacetime bounds
$$
\vn{ \left|\partial_{x}\right|^{1 / 6} \tilde{v}_{n} }_{L_{t, x}^{6}(\mathbb{R} \times \mathbb{R})}+\left\|\tilde{v}_{n} \right\|_{L_{x}^{5} L_{t}^{10}(\mathbb{R} \times \mathbb{R})} \lesssim_{\phi} 1
$$
and for every $\varepsilon>0$ there exist $n_{\varepsilon} \in \mathbb{N}$ and $\psi_{\varepsilon} \in C_{c}^{\infty}(\mathbb{R} \times \mathbb{R})$ so that, for all $n \geq n_{\varepsilon}$ one has
\begin{equation} \label{smoothapprox}
\vn{ \tilde{v}_{n} (t, x)-\operatorname{Re} [e^{i x \xi_{n} \lambda_n +i t (\xi_{n} \lambda_n)^{3}} \psi_{\varepsilon} \left(3 \xi_{n} \lambda_n t, x+3 (\xi_{n} \lambda_n)^{2} t\right) ] }_{L_{x}^{5} L_{t}^{10}(\mathbb{R} \times \mathbb{R})} \leq \varepsilon.
\end{equation}
Moreover, defining $ \tilde{u}_{n}^T $ by \eqref{unt}, one has the approximation
\begin{equation} \label{unifapproxu}
\lim_{T \to \infty} \lim _{n \rightarrow \infty} \left\| \tilde{v}_{n}-\tilde{u}_{n}^T \right\|_{L_{t}^{\infty} L_{x}^{2}(\mathbb{R} \times \mathbb{R})} = 0
\end{equation}
\end{theorem}
We note that \eqref{unifapproxu} is obtained in the proof of \cite[Thm. 4.1]{killip2009mass}.
\subsection{The Airy profile decomposition and decoupling}
\begin{definition} \label{asym.orth}
\begin{enumerate}
\item We say that two sequences $ (\Gamma_n^1)_{n \geq 1} =(\lambda_n^1, \xi_n^1, x_n^1, t_n^1 )_{n \geq 1} $ and $ (\Gamma_n^2)_{n \geq 1} =(\lambda_n^2, \xi_n^2, x_n^2, t_n^2 )_{n \geq 1} $ in $ (0,\infty) \times \mathbb{R}^3 $ are
asymptotically orthogonal if
\begin{multline*}
\frac{\lambda_n^1}{\lambda_n^2} + \frac{\lambda_n^2}{\lambda_n^1} + \sqrt{ \lambda_n^1 \lambda_n^2} \vm{\xi_n^1 - \xi_n^2} + \jb{ \lambda_n^1 \xi_n^1 \lambda_n^2 \xi_n^2}^{\frac{1}{2}}
\vm{\frac{ (\lambda_n^1)^3 t^1_n- (\lambda_n^2)^3 t^2_n }{ (\lambda_n^1 \lambda_n^2 )^{3/2} } } \\
+ (\lambda_n^1 \lambda_n^2 )^{-\frac{1}{2}} \vm{x_n^1-x_n^2 +\frac{3}{2} [(\lambda_n^1)^3 t^1_n- (\lambda_n^2)^3 t^2_n ] [ (\xi^1_n)^2 +(\xi^2_n)^2 ] }
\overset{n}{\longrightarrow} \infty.
\end{multline*}
\item We say that
$$ (\Gamma_n)_{n \geq 1} =(\lambda_n, \xi_n, x_n, t_n )_{n \geq 1} \overset{n}{\longrightarrow} \infty, $$
if
$ (\Gamma_n)_{n \geq 1} $ and $ (1,0,0,0)_{n \geq 1} $ are asymptotically orthogonal , i.e.
$$
\lambda_n + \frac{1}{ \lambda_n} + \vm{\xi_n} + \vm{t_n} + \vm{x_n} \overset{n}{\longrightarrow} \infty.
$$
\end{enumerate}
\end{definition}
Thus one can think of $ \infty $ as an element in the one-point compactification of $ (0,\infty)\times \mathbb{R}^3 $.
\
If $ \Gamma_n^1 =(\lambda_n^1, \xi_n^1, x_n^1, t_n^1 ) $ and $ \Gamma_n^2 =(\lambda_n^2, \xi_n^2, x_n^2, t_n^2 ) $ are asymptotically orthogonal then
\begin{equation} \label{inner.orthog}
\lim_{n \to \infty} \langle g_{x_n^1, \lambda_n^1 } e^{-t^1_n \partial^3_x} [e^{ix \xi_n^1 \lambda_n^1 } \phi ] , \ g_{x_n^2, \lambda_n^2 } e^{-t^2_n \partial^3_x} [e^{ix \xi_n^2 \lambda_n^2 } \varphi ]
\rangle = 0, \qquad \phi, \varphi \in L^2.
\end{equation}
where either $ \xi_n^j = 0 $ for all $ n \geq 1 $ or $ | \lambda_n^j \xi_n^j | \to \infty $.
See \cite[Lemma 5.2, 5.1 Cor. 3.7]{shao2009}.
This implies, in particular, the following statement.
\begin{lemma} \label{weakgconv}
Let $ \Gamma_n = (\lambda_n, \xi_n, z_n, s_n ) \to \infty $ and $ \theta_n \in \mathbb{R} $. Then, weakly in $ L^2 $ one has
\begin{equation} \label{gamma.weak.conv}
e^{i \theta_n} g_{z_n, \lambda_n} e^{-s_n \partial^3_x} [e^{\pm ix \xi_n \lambda_n } h ] \rightharpoonup 0,
\end{equation}
for any $ h \in L^2 $.
\end{lemma}
\
We are ready to state the profile decomposition for the Airy propagator obtained by Shao in \cite{shao2009}.
\begin{lemma} \label{profile.dec}
(Airy linear profile decomposition \cite{shao2009})
Let $ v_n : \mathbb{R} \to \mathbb{R} $ be a sequence of functions bounded in $ L^2_x(\mathbb{R}) $. Then, after passing to a subsequence, there exist functions $ \phi^j: \mathbb{R} \to \mathbb{C} $ in $ L^2_x(\mathbb{R}) $, group elements $ g_n^j := g_{x_n^j, \lambda_n^j } \in G $, frequency parameters $ \xi_n^j \in [0,\infty) $ and times $ t_n^j \in \mathbb{R} $ such that for all $ J \geq 1 $ one can write
\begin{equation}
v_n = \sum_{1 \leq j \leq J} g^j_n e^{-t^j_n \partial^3_x} \text{Re}[e^{ix \xi_n^j \lambda_n^j } \phi^j ] + w_n^J,
\end{equation}
for some real-valued sequence $ w_n^J $ in $ L^2_x(\mathbb{R}) $ with
\begin{equation}
\lim_{J \to \infty} \limsup_{n \to \infty} \vn{| \partial_x |^{1/6} e^{-t \partial^3_x} w_n^J }_{L^6_{t,x}(\mathbb{R} \times \mathbb{R} )} = \lim_{J \to \infty} \limsup_{n \to \infty} \vn{e^{-t \partial^3_x} w_n^J }_{L^5_x L^{10}_t (\mathbb{R} \times \mathbb{R} )}=0.
\end{equation}
For each $ 1 \leq j \leq J $, the frequency parameters $ \xi_n^j $ satisfy: either $ \xi_n^j =0 $ for all $ n \geq 1 $ or $ \xi_n^j \lambda_n^j \to \infty $ as $ n \to \infty $ (If $ \xi_n^j =0 $ we assume $ \phi^j $ is real). For any $ J \geq 1 $ one has
\begin{equation} \label{L2.orthog}
\vn{v_n}_{L^2}^2- \sum_{1 \leq j \leq J} \vn{\text{Re}[e^{ix \xi_n^j \lambda_n^j } \phi^j ] }_{L^2}^2 - \vn{w_n^J}_{L^2}^2 \overset{n}{\longrightarrow} 0.
\end{equation}
The family of sequences $ \Gamma_n^j =(\lambda_n^j, \xi_n^j, x_n^j, t_n^j ) \in (0,\infty) \times \mathbb{R}^3 $ are pair-wise asymptotically orthogonal in the sense of Definition \ref{asym.orth} and
for any $ 1 \leq j \leq J $
\begin{equation} \label{orthog2}
\lim_{n \to \infty} \langle g^j_n e^{-t^j_n \partial^3_x} \text{Re}[e^{ix \xi_n^j \lambda_n^j } \phi^j ] , w_n^J \rangle = 0.
\end{equation}
\end{lemma}
For more discussion of the properties stated above we refer to Lemma 2.4, Remark 2.5. in \cite{killip2009mass} and Corollary 3.7, Lemma 5.2 in \cite{shao2009}.
\begin{corollary} \label{cor.profile.dec}
Under the assumptions and notations of Lemma \ref{profile.dec}, if $ v_n \rightharpoonup 0 $ weakly in $ L^2 $, then also $ w_n^J \rightharpoonup 0 $ weakly in $ L^2 $ for all $ J \geq 1 $ after passing to a subsequence. For any $ 1 \leq j \leq J $ one has $ \phi^j=0 $ or $\Gamma_n^j =(\lambda_n^j, \xi_n^j, x_n^j, t_n^j ) \to \infty $ in the sense of Definition \ref{asym.orth} and therefore
\begin{equation}
g^j_n e^{-t^j_n \partial^3_x} \text{Re}[e^{ix \xi_n^j \lambda_n^j } \phi^j ] \rightharpoonup 0.
\end{equation}
\end{corollary}
\begin{proof}
After passing to a subsequence, we can arrange so that for each $ j \in \overline{1,J} $, either $ \Gamma_n^j $ converges to a finite $ \Gamma_0^j $ in $ (0,\infty) \times \mathbb{R}^3 $ or $ \Gamma_n^j \to \infty $. By pair-wise asymptotic orthogonality we have \eqref{inner.orthog} and therefore at most one of the sequences $ \{ ( \Gamma_n^j)_{n \geq 1} \ | \ 1 \leq j \leq J, \ \phi^j \neq 0 \} $ can converge to a finite value. Assume this happens for $ j=1 $ and then $ \xi^1_n =0 $ for all $ n \geq 1 $ and $ \phi^1 $ is assumed real. Since $ v_n \rightharpoonup 0 $ we obtain
$$ g^1_0 e^{-t^1_0 \partial^3_x} \phi^1 + w_n^J \rightharpoonup 0.
$$
Taking inner product with $ g^1_n e^{-t^1_n \partial^3_x} \phi^1 $ and using \eqref{orthog2} we obtain $ \vn{g^1_0 e^{-t^1_0 \partial^3_x} \phi^1}_{L^2}^2=0 $ and then $ \phi^1=0 $, which is a contradiction. \end{proof}
\
Finally, we recall the decoupling property of nonlinear profiles proved in \cite[Lemma 2.6]{killip2009mass}. When $ \xi_n \lambda_n \to \infty $ the decoupling will follow from this lemma together with the approximation \eqref{smoothapprox} from Theorem \ref{oscilatory.profiles}
\begin{lemma}[\cite{killip2009mass}] \label{decoupling}
Let $ \psi^1, \psi^2 \in C_c^{\infty} (\mathbb{R} \times \mathbb{R}) $ and sequences
$$ (\Gamma_n^1)_{n \geq 1} =(\lambda_n^1, \xi_n^1, x_n^1, t_n^1 )_{n \geq 1}, \quad (\Gamma_n^2)_{n \geq 1} =(\lambda_n^2, \xi_n^2, x_n^2, t_n^2 )_{n \geq 1}, $$ in $ (0,\infty) \times \mathbb{R}^3 $ assumed
asymptotically orthogonal in the sense of Definition \ref{asym.orth}. Then one has:
$$
\lim_{n \to \infty} \vn{ T_{g_{x^1_n, \lambda^2_n}} \psi^1(t+t_n^1) \ T_{g_{x^2_n, \lambda^2_n}} \psi^2(t+t_n^2)}_{L^{\frac{5}{2}}_x L^5_t } =0,
$$
in the case $ \xi^1_n \equiv \xi^2_n \equiv 0 $, and
$$
\lim_{n \to \infty} \vn{ T_{g_{x^1_n, \lambda^1_n}} \left[ \psi^1( 3 \lambda^1_n \xi^1_n (t+t_n^1), x+ 3 (\lambda^1_n \xi^1_n)^2 (t+t_n^1) ) \right] T_{g_{x^2_n, \lambda^2_n}} \psi^2(t+t_n^2)}_{L^{\frac{5}{2}}_x L^5_t } =0,
$$
when $ \xi^1_n \lambda^1_n \to \infty $ and $ \xi^2_n \equiv 0 $, while
\begin{multline*}
\lim_{n \to \infty} \vn{
T_{g_{x^1_n, \lambda^2_n}} \left[ \psi^1( 3 \lambda^1_n \xi^1_n (t+t_n^1), x+ 3 (\lambda^1_n \xi^1_n)^2 (t+t_n^1) ) \right] \\
T_{g_{x^2_n, \lambda^2_n}} \left[ \psi^2( 3 \lambda^2_n \xi^2_n (t+t_n^2), x+ 3 (\lambda^2_n \xi^2_n)^2 (t+t_n^2) ) \right]
}_{L^{\frac{5}{2}}_x L^5_t } =0,
\end{multline*}
when $ \xi^1_n \lambda^1_n \to \infty $ and $ \xi^2_n \lambda^2_n \to \infty $.
\end{lemma}
\section{Reduction to an almost periodic solution - Proof of Theorem \ref{almost_periodic_prop}} \label{Sec.reduction.to.almost.periodic}
This section is devoted to the proof of Theorem \ref{almost_periodic_prop}. Therefore we will assume at least one $ \delta $-close solution exists. Then we define the set
$$
S(\delta) \vcentcolon= \{ u \ | \ u = \text{solution} \ \delta-close \ \text{to} \ Q \ \text{with} \ M(u) < M_Q \}
$$
and the minimal mass:
$$
m_0(\delta) \vcentcolon= \inf \{ M(u) \ | u \in S(\delta) \}.
$$
By the triangle inequality, if $ u \in S(\delta) \neq \emptyset $ and $ t_0 \in I $ we have the basic bounds
\begin{equation} \label{mass.closeness}
M_Q^{\frac{1}{2}} -\delta \leq \vn{ u(t_0)}_{L^2_x} < M_Q^{\frac{1}{2}}, \quad \text{and} \quad M_Q^{\frac{1}{2}} -\delta \leq m_0^{\frac{1}{2}} \leq M_Q^{\frac{1}{2}}.
\end{equation}
\
The crux of the proof is the following Palais-Smale -type proposition which is used
to extract subsequences convergent in $ L^2 $.
\begin{proposition} \label{L2convergence_prop}
There exists an $ \delta>0 $ small enough such that the following holds.
Let $ u_n:I_n \times \mathbb{R} \to \mathbb{R} $ be maximal-lifespan (strong) solutions to the mass-critical focusing gKdV equation \eqref{1.1} which are $ \delta $-close to $ Q $, i.e. for some continuous $ g_n : I_n \to G $ one has
\begin{equation} \label{epcloseseq}
\vn{ g_n(t) u_n(t) - Q}_{L^2} \leq \delta \qquad \forall \ t \in I_n, \ n \geq 1.
\end{equation}
Suppose $ M(u_n) \searrow m_0=m_0(\delta) $ and let $ t_n \in I_n $ be a sequence of times. Then the sequence $ g_n(t_n) u_n(t_n) $ has a subsequence which converges in $ L^2 $ to a function $ \phi $ with $ M(\phi) =m_0 $.
\end{proposition}
Assuming Proposition \ref{L2convergence_prop} we can now construct almost periodic solutions.
\begin{proof}[{\bf Proof of Theorem \ref{almost_periodic_prop}}]
We first show that if $ M(u)>m_0 $ then there exists a maximal-lifespan solution $ v : J \times \mathbb{R} \to \mathbb{R} $ with minimal mass $ M(v)=m_0 $ which is $ \delta $-close to $ Q $. In that case there exists a sequence of maximal-lifespan solutions $ u_n:I_n \times \mathbb{R} \to \mathbb{R} $ with $ M(u_n) \searrow m_0 $ such that \eqref{epcloseseq} holds for some continuous $ g_n : I_n \to G $. Then we apply Prop. \ref{L2convergence_prop} with some $ t_n \in I_n $ and obtain a $ \phi \in L^2 $ with $ \vn{\phi}_{L^2}=m_0^{1/2} $. By translating time we may assume all $ t_n=0 $ and by applying transformations $ T_{g_n(0)^{-1}} $ we may assume without loss of generality that all $ g_n(0) $ are the identity.
Let $ v $ be the strong solution to \eqref{1.1} with initial data $ v(0) =\phi $, defined on a maximal interval $ J $, which then satisfies
$$ \vn{ u_n(0) - v(0)}_{L^2_x} \to 0. $$
Then for any $ t \in J $, by continuous dependence on initial data, see Corollary \ref{sol.conv} applied on $ [0,t] $, one has $ t \in I_n $ for $ n $ large enough and
\begin{equation} \label{un.v.conv}
\vn{ u_n(t) - v(t)}_{L^2_x} + \vn{u_n-v}_{ L^5_x L^{10}_t([0,t] \times \mathbb{R}) } \to 0.
\end{equation}
By Lemma \ref{group.compactness} we have $ g_n(t) \in K_t $ for a compact set $ K_t $. Then we can extract a subsequence such that $ g_n(t) $ converges to some $ g(t) \in G $ in the strong operator topology. Therefore \eqref{un.v.conv} and \eqref{epcloseseq} imply
$$ \vn{ g(t) v(t) - Q}_{L^2} \leq \delta \qquad \forall \ t \in J, $$
which gives the desired $ \delta $-closeness to $ Q $. Note that $ g(t) $ is continuous.
\
We now show that $ v $ is almost periodic modulo symmetries. This follows by considering a new arbitrary sequence of times $ t_n \in J $ and applying Prop. \ref{L2convergence_prop} with $ g_n =g $, $ u_n =v $ and $ t_n \in I_n=J $ to conclude that $ g(t_n) v(t_n) $ has a limit point in $ L^2 $.
\end{proof}
\
It remains to prove the key convergence result.
\begin{proof}[{\bf Proof of Proposition \ref{L2convergence_prop}}] By translating time we may assume all $ t_n=0 $ and by applying transformations $ T_{g_n(0)^{-1}} $ we may assume without loss of generality that all $ g_n(0) $ are the identity.
We divide the proof into several steps and for the first steps we largely follow the outline of \cite[Prop. 5.1 -Case II]{killip2009mass}, with the mention that here one needs to insure that the bulk of $ m_0 $, except for $ O(\delta) $ mass, has to fall onto the first profile.
{\bf Step 1.} (Decomposing the sequence)
By passing to a subsequence, using the Banach-Alaoglu theorem, we obtain a function $ \phi^1 \in L^2 $ such that $ u_n(0) \rightharpoonup \phi^1 $ weakly in $ L^2 $. Note that $ \vn{\phi^1}_{L^2}^2 \leq m_0 $ and since $ u_n(0) - Q \rightharpoonup \phi^1 -Q $ we obtain
\begin{equation}
\vn{\phi^1 -Q}_{L^2} \leq \delta.
\end{equation}
Moreover,
\begin{equation} \label{unphi.conv}
\vn{ u_n(0) - \phi^1}_{L^2}^2=\vn{u_n(0) }_{L^2}^2+ \vn{\phi^1}_{L^2}^2- 2 \langle u_n(0), \phi^1 \rangle \overset{n}{\longrightarrow} m_0-\vn{\phi^1}_{L^2}^2
\end{equation}
If $ \vn{\phi^1}_{L^2}^2 = m_0 $ this implies the desired convergence. Now assume $ \vn{\phi^1}_{L^2}^2 < m_0 $ and we will obtain a contradiction. We use the profile decomposition in Lemma \ref{profile.dec} and its Corollary \ref{cor.profile.dec} applied to $ v_n=u_n(0) - \phi^1 \rightharpoonup 0 $ to write for any $ J \geq 2 $
$$
u_n(0) - \phi^1 = \sum_{2 \leq j \leq J} g^j_n e^{-t^j_n \partial^3_x} \text{Re}[e^{ix \xi_n^j \lambda_n^j } \phi^j ] + w_n^J.
$$
By \eqref{unphi.conv}, the limit \eqref{L2.orthog} becomes
\begin{equation} \label{L2.orthog2}
m_0-\vn{\phi^1}_{L^2}^2- \sum_{2 \leq j \leq J} \vn{\text{Re}[e^{ix \xi_n^j \lambda_n^j } \phi^j ] }_{L^2}^2 - \vn{w_n^J}_{L^2}^2 \overset{n}{\longrightarrow} 0.
\end{equation}
By re-denoting some indices, we may assume that all the $ \phi^j $'s are nonzero. Defining $\Gamma_n^1 =(1, 0, 0, 0 ) $ corresponding to $ \phi^1 $, from Corollary \ref{cor.profile.dec} we obtain that $ \Gamma_n^j =(\lambda_n^j, \xi_n^j, x_n^j, t_n^j ) \to \infty $ for $ j \geq 2 $, and thus all $ (\Gamma_n^j)_{j \geq 1} $ are pair-wise asymptotically orthogonal and
\begin{equation} \label{wnJ.weaklimit}
w_n^J \rightharpoonup 0.
\end{equation}
From \eqref{mass.closeness} and $ \vn{\phi^1}_{L^2}^2 \leq m_0 $ we obtain the smallness condition
\begin{equation} \label{smallness}
\sum_{2 \leq j \leq J} \vn{\text{Re}[e^{ix \xi_n^j \lambda_n^j } \phi^j ] }_{L^2}^2 + \vn{w_n^J}_{L^2}^2 < 2 \delta M_Q^{\frac{1}{2}}, \qquad \forall \ n \gg_J 1.
\end{equation}
\
{\bf Step 2.} (Construct nonlinear profiles)
\
Let $ v^1 :I \times \mathbb{R} \to \mathbb{R} $ be the maximal-lifespan solution to \eqref{1.1} with initial data $ v^1(0)=\phi $.
We continue with defining solutions associated to the profiles for $ j \geq 2 $. For each $ J \geq 2 $ we reorder the indices such that:
A) For $ j \in \overline{2,J_0} $ one has $ \xi_n^j \equiv 0 $. Then one can refine the sequence for each $ j $ and by a diagonal argument one can assume that each sequence $ (t^j_n)_{n \geq 1} $ has a limit $ T^j $, possibly $ \pm \infty $. If $ T^j $ is finite one may assume that $ t^j_n \equiv T^j =0 $ by replacing $ \phi^j $ by $ e^{T^j \partial_x^3} \phi^j $ and by absorbing $ e^{- (t^j_n-T^j) \partial_x^3} \text{Re}\phi^j - \text{Re}\phi^j $ into the remainder term $ w^J_n $. One defines:
\begin{itemize}
\item When $ t^j_n \equiv 0 $, let $ v^j $ be the the maximal-lifespan solution to \eqref{1.1} with $ v^j(0)=\text{Re}\phi^j $.
\item If $ t^j_n \to \pm \infty $, let $ v^j $ be the the maximal-lifespan solution to \eqref{1.1} which scatters forward/backward in time to $ e^{- t \partial_x^3} \text{Re}\phi^j $.
\end{itemize}
Due to the smallness property \eqref{smallness}, each $ v^j $ is global and $ S_{\mathbb{R}} (v^j) \lesssim M[ \text{Re}\phi^j] $.
The nonlinear profiles are defined by
$$
v^j_n(t) \vcentcolon= T_{g^j_n}[ v^j(\cdot + t^j_n)](t), \qquad j \in \overline{2,J_0}, \ n \geq 1,
$$
so that $ v^j_n : \mathbb{R} \times \mathbb{R} \to \mathbb{R} $ with $ v^j_n(0)=g^j_n v^j(t_n) $.
B) For $ j \in \overline{J_0+1,J} $ the reordering satisfies $ \xi_n^j \lambda_n^j \to \infty $.
For $ n $ sufficiently large, the solution to \eqref{1.1} with data
$$
\tilde{v}^j_n (t^j_n) = e^{-t^j_n \partial^3_x} \text{Re}[e^{ix \xi_n^j \lambda_n^j } \phi^j ]
$$
is global and small. Moreover, by applying the Riemann-Lebesgue lemma to
$$
2 \vn{\text{Re}[e^{ix \xi_n^j \lambda_n^j } \phi^j ] }_{L^2}^2 = M(\phi^j) +\int_{\mathbb{R}}
\text{Re}[e^{i 2 x \xi_n^j \lambda_n^j } \phi^j(x)^2 ] \,\mathrm{d} x
$$
to obtain a bound on $ M(\phi^j) $,
one has the approximation given by Theorem \ref{oscilatory.profiles}
(since one can insure, using a diagonal argument, that $ (t^j_n \xi^j_n \lambda^j_n)_{n \geq 1} $ has a limit).
Again, one transforms these solutions to obtain $ v^j_n : \mathbb{R} \times \mathbb{R} \to \mathbb{R} $ by
$$
v^j_n(t) \vcentcolon= T_{g^j_n}[ \tilde{v}^j_n(\cdot + t^j_n)](t), \qquad j \in \overline{J_0+1, J}, \ n \gg 1.
$$
\
For both cases A) and B) Lemma \ref{decoupling} and Theorem \ref{oscilatory.profiles} give the decoupling property
\begin{equation} \label{decouplingg}
\lim_{n \to \infty} \vn{v^j_n v^k_n}_{L^{\frac{5}{2}}_x L^5_t (I \times \mathbb{R})} =0 \qquad \qquad \forall \ 1 \leq j <k
\end{equation}
where for $ j=1 $ we denote $ v^1_n=v^1 $.
Moreover, due to the smallness and the invariance of the scattering norm one has
\begin{equation} \label{vjn.bound}
S_{\mathbb{R}} (v^j_n) \lesssim \vn{\text{Re}[e^{ix \xi_n^j \lambda_n^j } \phi^j ]}_{L^2_x}^2, \qquad j \geq 2, \ n \gg_j 1.
\end{equation}
\
{\bf Step 3.} (Construct approximate solutions and bound the difference)
For any $ J \geq 2 $ construct the approximate solution, defined on $ I $ for $ n \gg_J 1 $ by
$$ \tilde{u}_n^J(t) \vcentcolon= v^1(t)+\sum_{j=2}^J v^j_n(t) + e^{-t \partial^3_x} w^J_n. $$
and define the remainders $ r_n^J $ on $ I \cap I_n $ by
$$ u_n(t)= \tilde{u}_n^J(t) + r_n^J(t). $$
From the way the $ v^j_n $ were constructed we obtain
\begin{equation} \label{rn0}
\vn{r_n^J(0)} _{L^2}= \vn{u_n(0)- \tilde{u}_n^J(0)}_{L^2} \overset{n}{\longrightarrow} 0, \qquad \forall J \geq 2.
\end{equation}
Next we bound the scattering size on any interval $ \tilde{I} $, using \eqref{smallness},\eqref{vjn.bound} and using the decoupling \eqref{decouplingg} after having raised the sum to the power 5:
\begin{align} \nonumber
\limsup_{n \to \infty} S_{\tilde{I}}( \tilde{u}_n^J ) & \lesssim \limsup_{n \to \infty} S_{\tilde{I}}( \sum_{j=1}^J v^j_n ) + \limsup_{n \to \infty} S_{\mathbb{R} }(e^{-t \partial^3_x} w^J_n) \\
\nonumber
& \lesssim S_{\tilde{I}}(v^1) + \limsup_{n \to \infty} \sum_{j=2}^J S_{\mathbb{R} }(v^j_n ) + \delta \\
\label{Scatteringlimit}
& \lesssim S_{\tilde{I}}(v^1) + \limsup_{n \to \infty} \sum_{j=2}^J \vn{\text{Re}[e^{ix \xi_n^j \lambda_n^j } \phi^j ]}_{L^2_x}^2+ \delta \lesssim S_{\tilde{I}}(v^1) + \delta.
\end{align}
\
In the remainder of this step we prove
\begin{equation} \label{limit.intervals}
I \subseteq \bigcup_{N \geq 1} \bigcap_{n \geq N} I_n
\end{equation}
and that for any $ t \in I $ one has
\begin{equation} \label{rnt}
\lim_{J \to \infty} \limsup_{n \to \infty} \vn{r_n^J(t)} _{L^2}=0.
\end{equation}
Suppose $ t > 0 $. Divide $ [0,t] $ into intervals $[t_k, t_{k+1}] $, $ k \in \overline{1,N} $, $ t_1 = 0 $ such that
\begin{equation} \label{v1.ep0}
\vn{v^1}_{L^5_x L^{10}_t ( [t_k, t_{k+1}] \times \mathbb{R} ) } \simeq \varepsilon_0 \qquad \forall \ k \in \overline{1,N-1}
\end{equation}
where $ \varepsilon_0=\varepsilon_0(M_Q,1)>0 $ is the universal constant given by Lemma \ref{short.time.stability}. Then Lemma \ref{num.intervals} gives a bound on the number of intervals $ N $.
We begin with \eqref{rn0} and do an inductive argument to show that if $ t_k \in I_n $ for $ n \gg_k 1 $ and
\eqref{rnt} holds at $ t=t_k $, then $ t_{k+1} \in I_n $ holds for $ n \gg_{k+1} 1 $ and
$$
\lim_{J \to \infty} \limsup_{n \to \infty} \vn{r_n^J} _{L^{\infty}_t L^2_x ([t_k,t_{k+1}] \times \mathbb{R})}=0.
$$
These facts follow from the short-time stability Lemma \ref{short.time.stability} applied with $ u_n $ and $ \tilde{u}_n^J $, provided we check:
\begin{align}
\label{sclim}
\limsup_{n \to \infty} \vn{ \tilde{u}_n^J}_{L^5_x L^{10}_t( [t_k, t_{k+1}] \times \mathbb{R} ) } \leq \frac{ \varepsilon_0}{2} \qquad \forall \ J \geq 2, \ k \in \overline{1,N} \\
\label{l1l2lim}
\lim_{J \to \infty} \limsup_{n \to \infty} \vn{ \vm{\partial_x}^{-1} [ (\partial_t + \partial_x^3)\tilde{u}_n^J- \partial_x ( \tilde{u}_n^J)^5 ] }_{L^1_x L^2_t ([t_k, t_{k+1}] \times \mathbb{R})} = 0.
\end{align}
The first bound \eqref{sclim} follows from \eqref{Scatteringlimit} by appropriately choosing the implicit constant in \eqref{v1.ep0} and choosing $ \delta $ small enough.
The asymptotic solution bound \eqref{l1l2lim} is proved in Lemma \ref{approx.sol} below.
This completes the proof of \eqref{limit.intervals} and \eqref{rnt}. Moreover, by summing over intervals and recalling that $ \varepsilon_0 $ is fixed, this argument and Lemma \ref{short.time.stability} give the uniform bound
\begin{equation} \label{Scat.norm.unif}
\vn{u_n}_{L^5_x L^{10}_t([0,t]) } \lesssim N \varepsilon_0 \leq C \big( \vn{v^1}_{L^5_x L^{10}_t([0,t]) } \big), \qquad n \gg_t 1.
\end{equation}
\
{\bf Step 4.} (Show that $ v^j_n(t) $ converges weakly to $ 0 $)
Fix $ t \in \mathbb{R} $ and $ j \geq 2 $. Recall that $\Gamma_n^j =(\lambda_n^j, \xi_n^j, x_n^j, t_n^j ) \to \infty $ in the sense of Definition \ref{asym.orth}.
A) We first assume $ \xi^j_n \equiv 0 $. Then
$$ v^j_n(t) = g^j_n v^j \Big( t^j_n + \frac{t}{(\lambda^j_n)^3} \Big). $$
By passing to a subsequence, we may assume
$$ t^j_n + \frac{t}{(\lambda^j_n)^3} \to T_j \in [-\infty,\infty]. $$
If $ T_j $ is finite, in either case $ t^j_n \equiv 0 $ or $ t^j_n \to \infty $ we have $ g^j_n \to \infty $ and the claim reduces to $ g^j_n v^j(T_j) \rightharpoonup 0 $, which follows from \eqref{gamma.weak.conv}.
If $ T_j \to \pm \infty $ we use scattering to replace $ v^j_n(t) $ by $ g^j_n e^{-\big( t^j_n + \frac{t}{(\lambda^j_n)^3} \big) \partial^3_x} v_{\pm} $. Then we can approximate by bump functions and apply the dispersive estimate.
B) It remains to consider the case $ \xi^j_n \lambda^j_n \to \infty $. This implies in particular that $ \xi^j_n + \lambda^j_n \to \infty $. Fix $ t \in \mathbb{R} $, $ \epsilon >0 $, $ j \geq J_0 +1 $ and $ \varphi \in C_c^{\infty}(\mathbb{R}) $. We will use the approximation involving NLS solutions from Theorem \ref{oscilatory.profiles} to show
$$
\vm{ \langle v_n^j (t), \varphi \rangle } < \epsilon
$$
for $ n $ large enough. Since
$$
v_n^j (t) = g_n^j \tilde{v}_n^j \Big(t_n^j + \frac{t}{(\lambda_n^j)^3} \Big)
$$
we can use the approximation \eqref{unifapproxu} to reduce to
$$
\vm{ \langle g_n^j \tilde{u}_n^T \Big( t_n^j + \frac{t}{(\lambda_n^j)^3} \Big) , \varphi \rangle } < \frac{\epsilon}{2}
$$
for a fixed large $ T $, where the $ \tilde{u}_n^T $ are defined by \eqref{unt} in terms of NLS solutions $ V_n $.
By passing to a subsequence we may assume that all the $ t_n^j + \frac{t}{(\lambda_n^j)^3} $ are in $ [ -\frac{T}{3 \xi_{n} \lambda_n}, \frac{T}{3 \xi_{n} \lambda_n} ] $ or in $ [ \frac{T}{3 \xi_{n} \lambda_n} , \infty) $ or in $ (-\infty, -\frac{T}{3 \xi_{n} \lambda_n}] $ and that in the first case we have a limit
$$ T_1 \vcentcolon= \lim_{n \to \infty} 3 \xi_{n} \lambda_n \Big( t_n^j + \frac{t}{(\lambda_n^j)^3} \Big) \in [-T, T].
$$
In the other two cases we define $ T_1 \vcentcolon= \pm T $. Using \eqref{VnV}, \eqref{unt} and $ V \in C_t L_x^2 $ we approximate
$$
\vn{ \tilde{u}_n^T \Big( t_n^j + \frac{t}{(\lambda_n^j)^3} \Big) - f_n(T_1) }_{L^2} < \epsilon^2, \qquad n \gg 1
$$
where we denote
$$
f_n (T_1) \vcentcolon= e^{-s_n \partial^3_x} \text{Re} [e^{ ix \xi_n \lambda_n } e^{i c_n} V(T_1, x-y_n) ]
$$
for some values $ s_n, c_n, y_n $.
Therefore, denoting $ W $ to be either $ V $ or $ \bar{V} $, we reduce to showing
$$
e^{\pm i \theta_n} g_{x_n^j, \lambda_n^j } g_{y_n, 1} e^{-s_n \partial^3_x} [e^{ \pm ix \xi_n \lambda_n } W(T_1) ] \rightharpoonup 0,
$$
for some $ \theta_n $'s. This follows from Lemma \ref{weakgconv} because $ g_{x_n^j, \lambda_n^j } g_{y_n, 1} = g_{z_n, \lambda_n^j} $ for some $ z_n $ and we have $\xi^j_n + \lambda^j_n \to \infty $.
\
From A) and B) we conclude
\begin{equation} \label{vjn.weak.conv}
v^j_n(t) \rightharpoonup 0, \qquad \forall \ t \in \mathbb{R}, \ j \geq 2.
\end{equation}
\
{\bf Step 5.} (Prove that $ v^1 $ is $ \delta $-close to $ Q $)
Fix an arbitrary $ t \in I $, where we recall that $ I $ is the maximal lifespan of $ v^1 $. Then, by \eqref{limit.intervals} we have $ t \in I_n $ for $ n $ large enough. We expand
\begin{equation} \label{expand}
\delta^2 \geq \vn{u_n(t) - g_n(t)^{-1} Q }_{L^2}^2 = \vn{v^1(t)-g_n(t)^{-1} Q }_{L^2}^2 + A_n^J(t) + B_n^J(t)
\end{equation}
with the terms
\begin{align*}
A_n^J(t) & \vcentcolon= \vn{ \sum_{j=2}^J v^j_n(t) + e^{-t \partial^3_x} w^J_n+ r_n^J(t)}_{L^2}^2 \\
B_n^J (t) & \vcentcolon= 2 \langle v^1(t)-g_n(t)^{-1} Q \ , \sum_{j=2}^J v^j_n(t) + e^{-t \partial^3_x} w^J_n+ r_n^J(t) \rangle.
\end{align*}
Due to the uniform bound \eqref{Scat.norm.unif}, Lemma \ref{group.compactness} provides the existence of a compact set $ K_t $ such that $ g_n(t) \in K $ for $ n $ large enough. We extract a subsequence such that $ g_n(t) $ converges to some $ g(t) \in G $ in the strong operator topology. Then also $ g_n(t)^{-1} \to g(t)^{-1} $, so we may replace $ g_n(t)^{-1} Q $ by $ g(t)^{-1} Q $ when we use \eqref{vjn.weak.conv}, \eqref{wnJ.weaklimit} and \eqref{rnt} to obtain
$$
\lim_{J \to \infty} \limsup_{n \to \infty} B_n^J (t)=0.
$$
We use this together with $ A_n^J(t) \geq 0 $ to pass to the limit in \eqref{expand} and conclude
$$ \vn{g(t) v^1(t)-Q }_{L^2} \leq \delta \qquad \forall t \in I.
$$
This means $ v^1 \in S(\delta) $ with $ M(v^1) < m_0(\delta) $, a contradiction.
\end{proof}
\
It remains to verify the asymptotic solution bound \eqref{l1l2lim}.
\begin{lemma} \label{approx.sol}
Suppose $ w_n^J \in L^2_x(\mathbb{R}) $, $ J \geq 1,\ n \geq 1 $ and that
$ v^j_n \in L^5_x L^{10}_t (\tilde{I} \times \mathbb{R} ) $ are solutions to \eqref{1.1} such that for any $ 1 \leq j <k $
$$
\lim_{n \to \infty} \vn{v^j_n v^k_n}_{L^{\frac{5}{2}}_x L^5_t (\tilde{I} \times \mathbb{R})} =0, \qquad \lim_{J \to \infty} \limsup_{n \to \infty} \vn{e^{-t \partial^3_x} w_n^J }_{L^5_x L^{10}_t (\tilde{I} \times \mathbb{R} )}=0.
$$
Then, assuming the $ \tilde{u}_n^J $ are uniformly bounded in $ L^5_x L^{10}_t (\tilde{I} \times \mathbb{R} ) $, defined by
$$ \tilde{u}_n^J(t) \vcentcolon= \sum_{j=1}^J v^j_n(t) + e^{-t \partial^3_x} w^J_n, $$
one has
$$
\lim_{J \to \infty} \limsup_{n \to \infty} \vn{ \vm{\partial_x}^{-1} [ (\partial_t + \partial_x^3)\tilde{u}_n^J- \partial_x ( \tilde{u}_n^J)^5 ] }_{L^1_x L^2_t ( \tilde{I} \times \mathbb{R})} = 0.
$$
\end{lemma}
\begin{proof}
This is proved in \cite[Lemma 5.3]{killip2009mass}. We review the argument for the sake of completeness. One writes
$$
\left(\partial_{t}+\partial_{x}^3\right) \tilde{u}_{n}^{J}=\sum_{1 \leq j \leq J} \partial_{x}\left(v_{n}^{j}\right)^{5}.
$$
Thus it suffices to estimate $ (\tilde{u}_{n}^{J})^{5}-\sum_{1 \leq j \leq J} (v_{n}^{j})^{5} $ as follows:
$$
\vn{ \left(\tilde{u}_{n}^{J}-e^{-t \partial_{x}^{3}} w_{n}^{J}\right)^{5}-\left(\tilde{u}_{n}^{J}\right)^{5}}_{L^1_x L^2_t ( \tilde{I} \times \mathbb{R})} \lesssim
\vn{(e^{-t \partial_{x}^{3}} w_{n}^{J})^{5} }_{L^1_x L^2_t ( \tilde{I} \times \mathbb{R})} +
\vn{ (e^{-t \partial_{x}^{3}} w_{n}^{J}) \left|\tilde{u}_{n}^{J}\right|^{4}}_{L^1_x L^2_t ( \tilde{I} \times \mathbb{R})},
$$
then one uses Holder's inequality and pass to the limit. Secondly,
$$
\vn{ \big( \sum_{1 \leq j \leq J} v_{n}^{j}\big)^{5}-\sum_{1 \leq j \leq J}\left(v_{n}^{j}\right)^{5} }_{L^1_x L^2_t ( \tilde{I} \times \mathbb{R})} \lesssim
\sum_{i_{1}, i_{2}, i_{3}=1}^{J} \sum_{1 \leq j \neq k \leq J} \vn{ v_{n}^{i_{1}} v_{n}^{i_{2}} v_{n}^{i_{3}} ( v_{n}^{j} v_{n}^{k} )}_{L^1_x L^2_t ( \tilde{I} \times \mathbb{R})},
$$
and one uses Holder's inequality again to pass to the limit.
This completes the proof.
\end{proof}
\section{Reductions of an almost periodic solution}
Having proved Theorem \ref{almost_periodic_prop}, we have reduced the main result, Theorem \ref{main.thm}, to the case of almost periodic solutions.
The remainder of the paper is devoted to this case, i.e. proving Theorem \ref{t1.1}.
We begin with studying $ N(t) $ from Definition \ref{d1.1}. In this section we prove
\begin{theorem}\label{t3.3}
If there exists an almost periodic solution to $(\ref{1.1})$ with $\| u_0 \|_{L^{2}} < \| Q \|_{L^{2}}$, then there exists an almost periodic solution to \eqref{1.1} satisfying \eqref{1.4} on a maximal interval $I$ with $N(t) \geq 1$ on $I$, and
\begin{equation}\label{3.33}
\int_{I} N(t)^{2} dt = \infty.
\end{equation}
Moreover, if the initial solution is $ \delta $-close to $ Q $, then the solution we obtain is also $ \delta $-close to $ Q $.
\end{theorem}
\begin{proof}[{\bf Proof of Theorem \ref{t3.3}}]
Using elementary reductions (see \cite{killip2009mass}) it suffices to consider an almost periodic solution to $(\ref{1.1})$ that satisfies $N(t) \leq 1$ for $t \in [0, \infty)$. Such a solution will satisfy one of two properties:
\begin{equation}\label{3.1}
\lim_{T \rightarrow \infty} \inf_{t \in [0, T]} N(t) > 0,
\end{equation}
or
\begin{equation}\label{3.2}
\lim_{T \rightarrow \infty} \inf_{t \in [0, T]} N(t) = 0.
\end{equation}
{\bf 1)} Begin with scenario $(\ref{3.1})$, $N(t) \sim 1$ for any $t \in [0, \infty)$. Thus, there exists a function $x(t) : [0, \infty) \rightarrow \mathbb{R}$ such that
\begin{equation}\label{3.3}
\{ u(t, x - x(t)) : t \in [0, \infty) \}
\end{equation}
lies in a precompact subset of $L^{2}(\mathbb{R})$. Therefore, taking $t_{n} \rightarrow +\infty$ and possibly after passing to a subsequence,
\begin{equation}\label{3.4}
u(t_{n}, x - x(t_{n})) \rightarrow u_{0} \qquad \text{in} \qquad L^{2}(\mathbb{R}),
\end{equation}
and moreover, $u_{0}$ is the initial data for a solution to $(\ref{1.1})$ satisfying
\begin{equation}\label{3.5}
\{ u(t, x - x(t)) : t \in \mathbb{R} \}
\end{equation}
lies in a precompact subset of $L^{2}(\mathbb{R})$.\medskip
{\bf 2)} Now consider scenario $(\ref{3.2})$. Split this scenario into two separate cases:
\begin{equation}\label{3.6}
\limsup_{T \rightarrow \sup(I)} \frac{\sup_{t \in [t_{0}(T), T]} N(t)}{N(t_{0}(T))} < \infty,
\end{equation}
or
\begin{equation}\label{3.7}
\limsup_{T \rightarrow \sup(I)} \frac{\sup_{t \in [t_{0}(T), T]} N(t)}{N(t_{0}(T))} = \infty.
\end{equation}
where
$$
t_{0}(T)=\inf \big\{t \in[0, T]: N(t)=\inf _{t \in[0, T]} N(t)\big\}
$$
Following \cite{dodson2017global}, for any $k \in \mathbb{Z}$, let
\begin{equation}\label{3.8}
t_{k} = \inf \{ t \in [0, T] : N(t) = 2^{-k} \}.
\end{equation}
Since $N(t)$ is a continuous function of time and $(\ref{3.2})$ holds, $t_{k}$ is well-defined.
{\bf 2A)} When $(\ref{3.6})$ holds, there exists $C < \infty$ such that $N(t) \leq C 2^{-k}$ for any $t \geq t_{k}$.
\begin{lemma}\label{t3.1}
Suppose $(\ref{3.2})$ and $(\ref{3.6})$ hold.
Then the sequence $(t_{k + 1} - t_{k}) \cdot 2^{-3k}$ is unbounded as $k \rightarrow +\infty$.
\end{lemma}
\noindent \emph{Proof:} Suppose that there exists a constant $C_{0}$ such that
\begin{equation}\label{3.9}
(t_{k + 1} - t_{k}) \cdot 2^{-3k} \leq C_{0}.
\end{equation}
Then for any $k \in \mathbb{Z}$,
\begin{equation}\label{3.10}
t_{k} \lesssim C_{0} 2^{3k}.
\end{equation}
Meanwhile, as in the scaling symmetry implies
\begin{equation}\label{3.11}
t_{k} \gtrsim 2^{3k}.
\end{equation}
Therefore, for any $k$,
\begin{equation}\label{3.12}
N(t_{k}) \sim t_{k}^{-1/3}.
\end{equation}
As in \cite{dodson2017global}, $(\ref{3.12})$ implies that after passing to another subsequence, we have a solution $u$ to $(\ref{1.1})$ satisfying $N(t) \sim t^{-1/3}$ for any $t \geq 0$. Moreover, following the exact arguments in Section five of \cite{dodson2017global} shows that the self similar solution $u(t,x)$ satisfies the estimate
\begin{equation}\label{3.13}
E(u) \lesssim 1.
\end{equation}
However, by the Gagliardo-Nirenberg inequality, this contradicts $N(t) \nearrow +\infty$ as $t \searrow 0$. $\Box$\medskip
Now take a sequence $t_{k} \rightarrow \infty$ such that
\begin{equation}\label{3.14}
(t_{k + 1} - t_{k}) \cdot 2^{-3k} \rightarrow +\infty.
\end{equation}
In this case, $(\ref{3.6})$ guarantees that $N(t) \sim 2^{-k}$ for any $t_{k} < t < t_{k + 1}$. Choose the sequence of times $t_{k}' = \frac{t_{k} + t_{k + 1}}{2}$. After passing to a subsequence,
\begin{equation}\label{3.15}
2^{k/2} u(t_{k}', 2^{k}(x - x(t_{k}'))) \rightarrow u_{0}, \qquad \text{in} \qquad L^{2}(\mathbb{R}),
\end{equation}
and furthermore, $u_{0}$ is the initial data of a solution to $(\ref{1.1})$ satisfying
\begin{equation}\label{3.16}
\{ u(t, x - x(t)) : t \in \mathbb{R} \}
\end{equation}
lies in a precompact subset of $L^{2}(\mathbb{R})$.\medskip
{\bf 2B)} Finally, consider the case when $(\ref{3.2})$ and $(\ref{3.7})$ hold. In this case, possibly after passing to a subsequence,
\begin{equation}\label{3.17}
2^{k/2} u(t_{k}, 2^{k}(x - x(t_{k}))) \rightarrow u_{0}, \qquad \text{in} \qquad L^{2}(\mathbb{R}),
\end{equation}
where $u_{0}$ is the initial data of a solution to $(\ref{1.1})$ on an interval $I$ such that
\begin{equation}\label{3.18}
\{ N(t)^{-1/2} u(t, N(t)^{-1} x + x(t))) : t \in I \}
\end{equation}
lies in a precompact subset of $L^{2}(\mathbb{R})$, and moreover, $N(t) \geq 1$ for all $t \in I$.
\begin{proposition}\label{t3.2}
If $u$ is an almost periodic solution to $(\ref{1.1})$ with $\| u \|_{L^{2}} < \| Q \|_{L^{2}}$ on a maximal interval $I \subset \mathbb{R}$ that satisfies $N(t) \geq 1$ for all $t \in I$, and $N(0) = 1$, then
\begin{equation}\label{3.19}
\int_{I} N(t)^{2} dt = \infty.
\end{equation}
\end{proposition}
\noindent \emph{Proof:} Again following \cite{dodson2017global}, suppose
\begin{equation}\label{3.20}
\int_{I} N(t)^{2} dt = R_{0} < \infty.
\end{equation}
Translating in space so that $x(0) = 0$, define the Morawetz potential
\begin{equation}\label{3.21}
M(t) = \int \psi(\frac{x}{R}) u(t,x)^{2} dx,
\end{equation}
where
\begin{equation}\label{3.22}
\psi(x) = \int_{0}^{x} \phi(t) dt,
\end{equation}
where $\phi$ is a smooth, even function, $\phi(x) = 1$ for $-1 \leq x \leq 1$, and $\phi$ is supported on $|x| \leq 2$.\medskip
Since $N(t) \geq 1$ and $\int_{I} N(t)^{2} dt < \infty$, $I$ is necessarily a finite interval. Therefore, $N(t) \nearrow +\infty$ as $t \rightarrow \sup(I)$ or $t \rightarrow \inf(I)$. Combining this with the fact that $|\dot{x}(t)| \lesssim N(t)^{2}$,
\begin{equation}\label{3.23}
\sup_{t \in I} |M(t)| \lesssim R_{0},
\end{equation}
with implicit constant independent of $R$. Moreover, by direct computation,
\begin{equation}\label{3.24}
\frac{d}{dt} M(t) = -3 \int \phi(\frac{x}{R}) u_{x}(t,x)^{2} dx + \frac{1}{R^{2}} \int \phi''(\frac{x}{R}) u(t,x)^{2} dx + \frac{5}{3} \int \phi(\frac{x}{R}) u(t,x)^{6} dx.
\end{equation}
Therefore, by the fundamental theorem of calculus,
\begin{equation}\label{3.25}
\int_{I} \int \phi(\frac{x}{R}) u_{x}(t,x)^{2} dx dt \lesssim R_{0} + \frac{|I|}{R^{2}} + \int_{I} \int u(t,x)^{6} dx dt.
\end{equation}
We have already demonstrated that the first two terms on the right hand side are uniformly bounded for any $R \geq 1$. So it remains to control the third term.\medskip
Partition $I$ into consecutive intervals
\begin{equation}\label{3.26}
I = \cup_{k} J_{k},
\end{equation}
where
\begin{equation}\label{3.27}
\int_{J_{k}} \int u(t,x)^{8} dx dt \sim 1.
\end{equation}
Using standard perturbation arguments, for any fixed $J_{k}$ with $t_{1}, t_{2} \in J_{k}$
\begin{equation}\label{3.28}
N(t_{1}) \sim N(t_{2}), \qquad \text{and} \qquad |t_{1} - t_{2}| \lesssim N(t_{1})^{-3}.
\end{equation}
Therefore, by H{\"o}lder's inequality,
\begin{equation}\label{3.29}
\int_{J_{k}} \int u(t,x)^{6} dx dt \lesssim |J_{k}|^{1/3} \| u \|_{L_{t}^{\infty} L_{x}^{2}}^{2/3} \| u \|_{L_{t,x}^{8}(J_{k} \times \mathbf{R})}^{16/3} \lesssim |J_{k}|^{1/3} \lesssim \int_{J_{k}} N(t)^{2} dt.
\end{equation}
Therefore,
\begin{equation}\label{3.30}
\int_{I} \int \phi(\frac{x}{R}) u_{x}(t,x)^{2} dx dt \lesssim R_{0} + \frac{|I|}{R^{2}}.
\end{equation}
Taking $R \rightarrow \infty$,
\begin{equation}\label{3.31}
\int_{I} \int u_{x}(t,x)^{2} dx dt \lesssim R_{0}.
\end{equation}
Therefore, by the Gagliardo-Nirenberg inequality, when $\| u \|_{L^{2}} < \| Q \|_{L^{2}}$, by conservation of energy,
\begin{equation}\label{3.32}
\int_{I} E(u(t)) dt = |I| E(u_{0}) \lesssim R_{0}.
\end{equation}
However, when $\| u_{0} \|_{L^{2}} < \| Q \|_{L^{2}}$, conservation of energy combined with $(\ref{3.32})$ contradicts the fact that $N(t)$ is unbounded on $I$, which completes the proof of Proposition $\ref{t3.2} $. $\Box$\medskip
Since the subsequence in the above analysis always converges strongly in $L^{2}$ to $u_{0}$, if we begin with an $ \delta $-close to $ Q $ solution, then the solution that we obtain is also $ \delta $-close to $ Q $.
This completes the proof of Theorem \ref{t3.3}.
\end{proof}
\section{Decomposition of the solution near a soliton}
Since after rescaling and translation, $u$ is close to $Q$, we can use a decomposition lemma of \cite{martel2002stability}. This lemma was proved when $u$ was close to $Q$ in $H^{1}$ norm, however, it is possible to prove a slightly weaker result when $u$ is merely close in $L^{2}$ norm.
\begin{lemma}\label{l2.2}
There exists $\delta > 0$ such that if
\begin{equation}\label{2.5}
\| u - \lambda_{0}(t)^{-1/2} Q(\frac{x - x_{0}(t)}{\lambda_{0}(t)}) \|_{L^{2}} < 2 \delta,
\end{equation}
then there exist $x(t)$ and $\lambda(t)$ such that
\begin{equation}\label{2.6}
\epsilon(t, y) := \lambda(t)^{1/2} u(t, \lambda(t) y + x(t)) - Q(y)
\end{equation}
satisfies
\begin{equation}\label{2.7}
(y Q_{y}, \epsilon) = (y(\frac{Q}{2} + y Q_{y}), \epsilon) = 0.
\end{equation}
Moreover,
\begin{equation}\label{2.8}
|\frac{\lambda_{0}(t)}{\lambda(t)} - 1| + |\frac{x_{0}(t) - x(t)}{\lambda(t)}| + \| \epsilon \|_{L^{2}} \lesssim \delta.
\end{equation}
\end{lemma}
\begin{remark}\label{r2.2}
Observe that by $(\ref{2.8})$, almost periodicity (according to Definition $\ref{d1.1}$) is maintained with the new $x(t)$ and $N(t) = \frac{1}{\lambda(t)}$.
\end{remark}
\noindent \emph{Proof:} Use the implicit function theorem. For $\delta > 0$, let
\begin{equation}\label{2.9}
U_{\delta} = \{ u \in L^{2} : \| u - Q \|_{L^{2}} < 2 \delta \},
\end{equation}
and for $u \in L^{2}(\mathbb{R})$, $\lambda_{1} > 0$, $x_{1} \in \mathbb{R}$, define
\begin{equation}\label{2.10}
\epsilon_{\lambda_{1}, x_{1}}(y) = \lambda_{1}^{1/2} u(\lambda_{1} y + x_{1}) - Q.
\end{equation}
Define the functionals
\begin{equation}\label{2.11}
\rho_{\lambda_{1}, x_{1}}^{1}(u) = \int \epsilon_{\lambda_{1}, x_{1}} (y Q_{y}) dy, \qquad \rho_{\lambda_{1}, x_{1}}^{2}(u) = \int \epsilon_{\lambda_{1}, x_{1}} (y (\frac{Q}{2} + y Q_{y})) dy.
\end{equation}
Then by direct computation,
\begin{equation}\label{2.12}
\frac{\partial \epsilon_{\lambda_{1}, x_{1}}}{\partial x_{1}} = \lambda_{1}^{1/2} u_{x}(\lambda_{1} y + x_{1}),
\end{equation}
and
\begin{equation}\label{2.13}
\frac{\partial \epsilon_{\lambda_{1}, x_{1}}}{\partial \lambda_{1}} = \frac{1}{2} \lambda_{1}^{-1/2} u(\lambda_{1} y + x_{1}) + \lambda_{1}^{1/2} y u_{x}(\lambda_{1} y + x_{1}).
\end{equation}
Integrating by parts,
\begin{equation}\label{2.14}
\frac{\partial \rho_{\lambda_{1}, x_{1}}^{1}}{\partial x_{1}} = \int \lambda_{1}^{1/2} u_{x}(\lambda_{1} y + x_{1}) (y Q_{y})(y) dy = -\int \lambda_{1}^{-1/2} u(\lambda_{1} y + x_{1}) (y Q_{yy} + Q_{y})(y) dy,
\end{equation}
\begin{equation}\label{2.15}
\frac{\partial \rho_{\lambda_{1}, x_{1}}^{2}}{\partial x_{1}} = \int \lambda_{1}^{1/2} u_{x}(\lambda_{1} y + x_{1}) (\frac{y}{2} Q + y^{2} Q_{y})(y) dy = - \int \lambda_{1}^{-1/2} u(\lambda_{1} y + x_{1}) (\frac{Q}{2} + \frac{5y}{2} Q_{y} + y^{2} Q_{yy})(y) dy,
\end{equation}
\begin{equation}\label{2.16}
\aligned
\frac{\partial \rho_{\lambda_{1}, x_{1}}^{1}}{\partial \lambda_{1}} = \int [\frac{1}{2} \lambda_{1}^{-1/2} u(\lambda_{1} y + x_{1}) + \lambda_{1}^{1/2} y u_{x}(\lambda_{1} y + x_{1})] (y Q_{y})(y) dy \\ = \int \frac{1}{2} \lambda_{1}^{-1/2} u(\lambda_{1} y + x_{1}) y Q_{y}(y) dy -\int \lambda_{1}^{-1/2} u(\lambda_{1} y + x_{1}) y^{2} Q_{yy}(y) dy \\ -2 \int \lambda_{1}^{-1/2} u(\lambda_{1} y + x_{1}) y Q_{y}(y) dy,
\endaligned
\end{equation}
and
\begin{equation}\label{2.17}
\aligned
\frac{\partial \rho_{\lambda_{1}, x_{1}}^{2}}{\partial \lambda_{1}} = \int [\frac{1}{2} \lambda_{1}^{-1/2} u(\lambda_{1} y + x_{1}) + \lambda_{1}^{1/2} y u_{x}(\lambda_{1} y + x_{1})] (\frac{y}{2} Q + y^{2} Q_{y})(y) dy \\ = \frac{1}{2} \int \lambda_{1}^{-1/2} u(\lambda_{1} x + x_{1}) (\frac{y}{2} Q + y^{2} Q_{y})(y) dy - \int \lambda_{1}^{-1/2} u(\lambda_{1} y + x_{1}) (y Q + \frac{7}{2} y Q_{y} + y^{3} Q_{y})(y) dy.
\endaligned
\end{equation}
This implies that $(\rho_{\lambda_{1}, x_{1}}^{1}, \rho_{\lambda_{1}, x_{1}}^{2})$ are $C^{1}$ functions of $(\lambda_{1}, x_{1})$.\medskip
Also,
\begin{equation}\label{2.18}
\aligned
\frac{\partial \rho_{\lambda_{1}, x_{1}}^{1}}{\partial x_{1}}|_{\lambda_{1} = 1, x_{1} = 0, u = Q} = \int Q_{y} \cdot y Q_{y} dy = 0, \\
\frac{\partial \rho_{\lambda_{1}, x_{1}}^{1}}{\partial \lambda_{1}}|_{\lambda_{1} = 1, x_{1} = 0, u = Q} = \int Q_{y} \cdot y(\frac{Q}{2} + y Q_{y}) = \int (\frac{Q}{2} + y Q_{y})^{2} dy > 0, \\
\frac{\partial \rho_{\lambda_{1}, x_{1}}^{2}}{\partial \lambda_{1}}|_{\lambda_{1} = 1, x_{1} = 0, u = Q} = \int (\frac{Q}{2} + y Q_{y}) y Q_{y} = \int (\frac{Q}{2} + y Q_{y})^{2} dy > 0, \\
\frac{\partial \rho_{\lambda_{1}, x_{1}}^{2}}{\partial x_{1}}|_{\lambda_{1} = 1, x_{1} = 0, u = Q} = \int (\frac{Q}{2} + y Q_{y}) y(\frac{Q}{2} + y Q_{y}) dy = 0.
\endaligned
\end{equation}
Therefore, by the implicit function theorem, if
\begin{equation}\label{2.18.1}
\| u(x) - Q(x) \|_{L^{2}} < 2 \delta,
\end{equation}
then there exist $\lambda$, $x$ such that
\begin{equation}\label{2.19}
|\lambda - 1| + |x| + \| \epsilon \|_{L^{2}} \lesssim \| u - Q \|_{L^{2}} < 2 \delta,
\end{equation}
satisfying
\begin{equation}\label{2.20}
(\epsilon, y Q_{y}) = (\epsilon, y(\frac{Q}{2} + y Q_{y})) = 0.
\end{equation}
Now take a general $\lambda_{0}(t)$ and $x_{0}(t)$ such that
\begin{equation}\label{2.25}
\| u(y) - \lambda_{0}^{-1/2} Q(\frac{y - x_{0}}{\lambda_{0}}) \|_{L^{2}} < 2 \delta.
\end{equation}
Then after translation and rescaling,
\begin{equation}\label{2.26}
\| \lambda_{0}^{1/2} u(\lambda_{0} y + x_{0}) - Q(y) \|_{L^{2}} < 2 \delta.
\end{equation}
Then there exist $|\tilde{x}| + |\tilde{\lambda}| \lesssim \| u - Q \|_{L^{2}}$ such that
\begin{equation}\label{2.27}
\aligned
(\lambda_{0}^{1/2} (1 - \tilde{\lambda})^{1/2} u(\lambda_{0} (1 - \tilde{\lambda}) y + \lambda_{0} \tilde{x} + x_{0})&, x Q_{x}) = 0, \\
(\lambda_{0}^{1/2} (1 - \tilde{\lambda})^{1/2} u(\lambda_{0} (1 - \tilde{\lambda}) y + \lambda_{0} \tilde{x} + x_{0})&, x(\frac{Q}{2} + x Q_{x})) = 0.
\endaligned
\end{equation}
Since $|\tilde{\lambda}| \lesssim \delta$, $|\frac{\lambda_{0} (1 - \tilde{\lambda})}{\lambda_{0}}| \lesssim \delta$. Also, $|\lambda_{0} \tilde{x}| \lesssim \lambda_{0} \delta$, so $\frac{|x_{0}' - x_{0}|}{\lambda_{0}} \lesssim \delta$. This completes the proof of Lemma $\ref{l2.2}$. $\Box$\medskip
Introduce the variable
\begin{equation}\label{2.28}
s = \int_{0}^{t} \frac{dt'}{\lambda(t')^{3}}, \qquad \text{equivalently} \qquad \frac{ds}{dt} = \frac{1}{\lambda^{3}}.
\end{equation}
\begin{lemma}[Properties of the decomposition]\label{l2.3}
$(1)$ The function $\epsilon(s, y)$ satisfies the equation
\begin{equation}\label{2.30}
\aligned
\epsilon_{s} = (L \epsilon)_{y} + \frac{\lambda_{s}}{\lambda} (\frac{Q}{2} + y Q_{y}) + (\frac{x_{s}}{\lambda} - 1) Q_{y}
+ \frac{\lambda_{s}}{\lambda} (\frac{\epsilon}{2} + y \epsilon_{y}) + (\frac{x_{s}}{\lambda} - 1) \epsilon_{y} - (R(\epsilon))_{y},
\endaligned
\end{equation}
where
\begin{equation}\label{2.31}
L \epsilon = -\epsilon_{xx} + \epsilon - 5 Q^{4} \epsilon, \qquad \text{and} \qquad R(\epsilon) = 10 Q^{3} \epsilon^{2} + 10 Q^{2} \epsilon^{3} + 5 Q \epsilon^{4} + \epsilon^{5}.
\end{equation}
$(2)$ $\lambda$ and $x$ are $C^{1}$ functions of $s$ and
\begin{equation}\label{2.32}
\aligned
\frac{\lambda_{s}}{\lambda} (\int (\frac{Q}{2} + y Q_{y})^{2} dy - \int (2 y Q_{y} + y^{2} Q_{yy}) \epsilon dy) - (\frac{x_{s}}{\lambda} - 1) \int (y Q_{yy} + Q_{y}) \epsilon dy \\ = \int L(y Q_{yy} + Q_{y}) \cdot \epsilon dy - \int R(\epsilon) (y Q_{y}) dy,
\endaligned
\end{equation}
and
\begin{equation}\label{2.33}
\aligned
-\frac{\lambda_{s}}{\lambda} \int \epsilon(y Q + \frac{7y^{2}}{2} Q_{y} + y^{3} Q_{yy}) dy + (\frac{x_{s}}{\lambda} - 1) (\int (\frac{Q}{2} + y Q_{y})^{2} dy - \int (\frac{Q}{2} + \frac{5y}{2} Q_{y} + y^{2} Q_{yy}) \epsilon dy) \\ = \int L(\frac{Q}{2} + \frac{5y}{2} Q_{y} + y^{2} Q_{yy}) \cdot \epsilon dy - \int (\frac{Q}{2} + \frac{5y}{2} Q_{y} + y^{2} Q_{yy}) R(\epsilon) dy.
\endaligned
\end{equation}
\end{lemma}
\emph{Proof:} See \cite{martel2002stability}. $\Box$\medskip
This lemma has an important corollary.
\begin{corollary}\label{c2.4}
For all $s \in \mathbb{R}$,
\begin{equation}\label{2.35}
|\frac{\lambda_{s}}{\lambda}| + |\frac{x_{s}}{\lambda} - 1| \lesssim \| \epsilon \|_{L^{2}} + \| \epsilon \|_{L^{2}} \| \epsilon \|_{L^{8}}^{4}.
\end{equation}
\end{corollary}
\begin{proof}
First observe that by H{\"o}lder's inequality and the boundedness of $Q$,
\begin{equation}\label{2.36}
\int L(y Q_{yy} + Q_{y}) \cdot \epsilon dy - \int R(\epsilon) (y Q_{y}) dy \lesssim \| \epsilon \|_{L^{2}} + \| \epsilon \|_{L^{2}} \| \epsilon \|_{L^{8}}^{4},
\end{equation}
and
\begin{equation}\label{2.37}
\int L(\frac{Q}{2} + \frac{5y}{2} Q_{y} + y^{2} Q_{yy}) \cdot \epsilon dy - \int (\frac{Q}{2} + \frac{5y}{2} Q_{y} + y^{2} Q_{yy}) R(\epsilon) dy \lesssim \| \epsilon \|_{L^{2}} + \| \epsilon \|_{L^{2}} \| \epsilon \|_{L^{8}}^{4}.
\end{equation}
Since $\int (\frac{Q}{2} + y Q_{y})^{2} dy > 0$,
\begin{equation}\label{2.38}
\aligned
|\frac{\lambda_{s}}{\lambda}| (1 + O(\| \epsilon \|_{L^{2}})) + |\frac{x_{s}}{\lambda} - 1| O(\| \epsilon \|_{L^{2}}) \lesssim \| \epsilon \|_{L^{2}} + \| \epsilon \|_{L^{2}} \| \epsilon \|_{L^{8}}^{4}, \\
|\frac{\lambda_{s}}{\lambda}| O(\| \epsilon \|_{L^{2}}) + |\frac{x_{s}}{\lambda} - 1| (1 + O(\| \epsilon \|_{L^{2}})) \lesssim \| \epsilon \|_{L^{2}} + \| \epsilon \|_{L^{2}} \| \epsilon \|_{L^{8}}^{4},
\endaligned
\end{equation}
so after doing some algebra,
\begin{equation}\label{2.39}
|\frac{\lambda_{s}}{\lambda}| + |\frac{x_{s}}{\lambda} - 1| \lesssim \| \epsilon \|_{L^{2}} + \| \epsilon \|_{L^{2}} \| \epsilon \|_{L^{8}}^{4}.
\end{equation}
\end{proof}
Next, by Strichartz estimates, rescaling, and perturbation theory, for any $k \in \mathbb{Z}$,
\begin{equation}\label{2.40}
\| u(s, y) \|_{L_{s,x}^{8}([k, k + 1] \times \mathbb{R})} \lesssim \| u_{0} \|_{L^{2}} < \| Q \|_{L^{2}}.
\end{equation}
Therefore, by the triangle inequality,
\begin{equation}\label{2.42}
\| \epsilon \|_{L_{s,x}^{8}([k, k + 1] \times \mathbb{R})} \lesssim \| Q \|_{L^{8}} + \| u \|_{L_{s,x}^{8}} \lesssim 1.
\end{equation}
Also, by perturbative arguments, for $\| \epsilon_{0} \|_{L^{2}}$ sufficiently small, if $\lambda(k) = 1$ and $x(k) = 0$,
\begin{equation}\label{2.41}
\| u(t,x) - Q(x - t) \|_{L_{t}^{\infty} L_{x}^{2}([k, k + 1] \times \mathbf{R})} \lesssim \| \epsilon_{0} \|_{L^{2}}.
\end{equation}
Thus using scaling and translation symmetries, along with Strichartz estimates,
\begin{equation}\label{2.41.1}
\| \epsilon \|_{L_{s}^{\infty} L_{y}^{2}([k, k + 1] \times \mathbb{R})} + \| \epsilon \|_{L_{s,y}^{8}([k, k + 1] \times \mathbb{R})} \lesssim \| \epsilon(k) \|_{L^{2}}.
\end{equation}
Combining $(\ref{2.40})$, $(\ref{2.41.1})$, Lemma $\ref{l2.2}$, and the fact that $\| Q \|_{L^{8}}$ is uniformly bounded, along with choosing $\| \epsilon_{0} \|_{L^{2}}$ to be the infimum of $\| \epsilon \|_{L^{2}}$ on the interval $[k, k + 1]$,
\begin{equation}\label{2.43}
\int_{k}^{k + 1} |\frac{\lambda_{s}}{\lambda}|^{2} + |\frac{x_{s}}{\lambda} - 1|^{2} ds \lesssim \int_{k}^{k + 1} \| \epsilon \|_{L^{2}}^{2} ds + \| \epsilon \|_{L_{t}^{\infty} L_{x}^{2}([k, k + 1] \times \mathbb{R})}^{2} \int_{k}^{k + 1} \| \epsilon \|_{L^{8}}^{8} ds \lesssim \int_{k}^{k + 1} \| \epsilon \|_{L^{2}}^{2} ds.
\end{equation}
\section{Exponential decay estimates of $u$}
Having obtained a decomposition of $u$ close to the soliton, the next step is to prove exponential decay of a solution that stays close to $Q$ in the case when $N(t) \geq 1$ and $\int_{I} N(t)^{2} dt = \infty$. The proof follows a similar argument in \cite{merle2001existence} and utilizes the fact that $u$ is close to a soliton, and the soliton moves to the right while a dispersive solution moves to the left.\medskip
Recall that
\begin{equation}\label{re1.8}
\sup_{t \in I} \| \epsilon(t) \|_{L^{2}(\mathbb{R})} =
\sup_{t \in I} \| u(t,x) - \frac{1}{\lambda(t)^{1/2}} Q(\frac{x - x(t)}{\lambda(t)}) \|_{L^{2}(\mathbb{R})} \lesssim \delta
\end{equation}
Observe that $N(t) \geq 1$ implies $\lambda(t) \lesssim 1$, where $\lambda(t) $ is given by Lemma \ref{l2.2}. It is convenient to rescale so that $\lambda(t) \leq 1$ for all $t \in I$. Note that after rescaling $N(t) \geq 1$. See Remark $\ref{r2.2}$.
\begin{lemma}[Exponential decay to the left of the soliton]\label{l4.1}
There exists some $a_{0}$ such that for $x_{0} \geq 10 a_{0}$, if $u$ satisfies Theorem $\ref{t3.3}$, $(\ref{1.4})$ and $ \| u_{0} \|_{L^{2}} < \| Q \|_{L^{2}}$, then
\begin{equation}\label{4.1}
\| u(t, x + x(t)) \|_{L^{2}(x \leq -x_{0})}^{2} \leq 10 c_{1} e^{-\frac{x_{0}}{6}}.
\end{equation}
\end{lemma}
\noindent \textbf{Remark:} It is important to note that $a_{0}$ does not depend on the $\delta > 0$ in $(\ref{re1.8})$.\medskip
\begin{proof}
Suppose there exists some $t_{0} \in \mathbb{R}$ and $x_{0} \geq 10 a_{0}$ such that
\begin{equation}\label{4.2}
\int_{x \leq -x_{0}} u(t_{0}, x + x(t_{0}))^{2} dx > 10 c_{1} e^{-\frac{x_{0}}{6}}.
\end{equation}
Let $K = 3 \sqrt{2}$, and let
\begin{equation}\label{4.3}
\phi(x) = c Q(\frac{x}{K}),
\end{equation}
where
\begin{equation}\label{4.4}
c = \frac{1}{K \int_{-\infty}^{\infty} Q(x) dx}.
\end{equation}
Define
\begin{equation}\label{4.5}
\psi(x) = \int_{-\infty}^{x} \phi(y) dy.
\end{equation}
Then,
\begin{equation}\label{4.6}
\lim_{x \rightarrow -\infty} \psi(x) = 0, \qquad \lim_{x \rightarrow +\infty} \psi(x) = 1.
\end{equation}
Next, define a modification of $x(s)$, $\tilde{x}(s)$, such that $x(k) = \tilde{x}(k)$ for all $k \in \mathbb{Z}$, and for any $s \in \mathbb{R}$, and for any $k < s < k + 1$, $\tilde{x}(s)$ is the linear interpolation between $\tilde{x}(k)$ and $\tilde{x}(k + 1)$. Then by $(\ref{2.43})$,
\begin{equation}\label{4.14}
\aligned
\tilde{x}(k + 1) - \tilde{x}(k) = x(k + 1) - x(k) = \int_{k}^{k + 1} x_{s}(s) ds \\ = \int_{k}^{k + 1} \lambda(s) + (\sup_{k \leq s \leq k + 1} \lambda(s)) \cdot (\int_{k}^{k + 1} \| \epsilon \|_{L^{2}} ds)
= (\int_{k}^{k + 1} \lambda(s) ds) \cdot (1 + O(\delta)).
\endaligned
\end{equation}
The last estimate follows from the fact that $\lambda(s) \sim \lambda(k)$ for any $k \leq s \leq k + 1$. It also follows from $(\ref{4.14})$ that for any $k \leq s \leq k + 1$,
\begin{equation}\label{4.15}
|\tilde{x}(s) - x(s)| \leq |\tilde{x}(s) - \tilde{x}(k)| + |x(s) - x(k)| \lesssim (\int_{k}^{k + 1} \lambda(s) ds).
\end{equation}
For technical reasons, it is useful to consider two cases separately. First, suppose that
\begin{equation}\label{4.15.1}
\int_{0}^{\sup(I)} N(t)^{2} dt = \int_{\inf(I)}^{0} N(t)^{2} dt = +\infty.
\end{equation}
In this case, suppose without loss of generality that $t_{0} = 0$, where $t_{0}$ is given by $(\ref{4.2})$. Then,
\begin{equation}\label{4.2.1}
\int_{x \leq -x_{0}} u(0, x + x(0))^{2} dx > 10 c_{1} e^{-\frac{x_{0}}{6}}.
\end{equation}
Define the function
\begin{equation}\label{4.7}
I(t) = \int u(t,x)^{2} \psi(x - \tilde{x}(0) + x_{0} - \frac{1}{4} (\tilde{x}(t) - \tilde{x}(0))) dx.
\end{equation}
Then by $(\ref{4.2})$, since $\tilde{x}(0) = x(0)$,
\begin{equation}\label{4.8}
I(0) \leq \int u(0, x)^{2} dx - \frac{1}{2} \int_{x \leq -x_{0} + \tilde{x}(0)} u(0, x)^{2} dx \leq \int u(0,x)^{2} dx - 5 c_{1} e^{-\frac{x_{0}}{K}}.
\end{equation}
Integrating by parts,
\begin{equation}\label{4.17}
\aligned
I'(t) = -3 \int u_{x}(t,x)^{2} \phi(x - \tilde{x}(0) + x_{0} - \frac{1}{4} (\tilde{x}(t) - \tilde{x}(0))) dx \\
+ \int u(t,x)^{2} \phi''(x - \tilde{x}(0) + x_{0} - \frac{1}{4}(\tilde{x}(t) - \tilde{x}(0))) dx \\
+ \frac{5}{3} \int u(t,x)^{6} \phi(x - \tilde{x}(0) + x_{0} - \frac{1}{4}(\tilde{x}(t) - \tilde{x}(0))) dx \\
- \frac{\dot{\tilde{x}}(t)}{4} \int u(t,x)^{2} \phi(x - \tilde{x}(0) + x_{0} - \frac{1}{4}(\tilde{x}(t) - \tilde{x}(0))) dx.
\endaligned
\end{equation}
Following \cite{merle2001existence}, observe that
\begin{equation}\label{4.10}
\frac{\dot{\tilde{x}}(t)}{4} = \frac{\tilde{x}_{s}}{4 \lambda^{3}} = \frac{1}{4} \frac{1}{\lambda^{2}} \frac{\tilde{x}_{s}}{\lambda} = \frac{1}{4 \lambda^{2}} (1 + O(\delta)).
\end{equation}
Also observe that
\begin{equation}\label{4.11}
\phi''(x) = \frac{c}{K^{2}} Q_{xx}(\frac{x}{K}) \leq \frac{c}{K^{2}} Q(\frac{x}{K}) = \frac{1}{K^{2}} \phi(x) = \frac{1}{18} \phi(x).
\end{equation}
Since $\lambda(t) \leq 1$,
\begin{equation}\label{4.12}
-\frac{\dot{\tilde{x}}(t)}{4 \lambda^{2}} \phi(x) + \frac{1}{K^{2}} \phi(x) \leq -\frac{1}{18}.
\end{equation}
Therefore,
\begin{equation}\label{4.18}
\aligned
I'(t) \leq -3 \int u_{x}(t,x)^{2} \phi(x - \tilde{x}(0) + x_{0} - \frac{1}{4} (\tilde{x}(t) - \tilde{x}(0))) dx \\
+ \frac{5}{3} \int u(t,x)^{6} \phi(x - \tilde{x}(0) + x_{0} - \frac{1}{4}(\tilde{x}(t) - \tilde{x}(0))) dx \\
- \frac{1}{18} \int u(t,x)^{2} \phi(x - \tilde{x}(0) + x_{0} - \frac{1}{4}(\tilde{x}(t) - \tilde{x}(0))) dx.
\endaligned
\end{equation}
Next, using Lemma $6$ from \cite{merle2001existence} and H{\"o}lder's inequality,
\begin{equation}\label{4.19}
\aligned
\int_{|x - \tilde{x}(t)| > a_{0}} u(t,x)^{6} \phi(x - \tilde{x}(0) + x_{0} - \frac{1}{4}(\tilde{x}(t) - \tilde{x}(0))) dx \leq \| u^{2} \phi^{1/2} \|_{L^{\infty}(|x - \tilde{x}(t)| > a_{0}}^{2} (\int_{|x - \tilde{x}(t)| > a_{0}} u(t,x)^{2} dx) \\
\lesssim (\int_{|x - \tilde{x}(t)| > a_{0}} u(t,x)^{2} dx)^{2} (\int u_{x}(t,x)^{2} \phi(x - \tilde{x}(0) + x_{0} - \frac{1}{4} (\tilde{x}(t) - \tilde{x}(0))) dx \\ + \int u(t,x)^{2} \phi(x - \tilde{x}(0) + x_{0} - \frac{1}{4} (\tilde{x}(t) - \tilde{x}(0))) dx).
\endaligned
\end{equation}
Since $\lambda(t) \leq 1$ and $|x - \tilde{x}(t)| \lesssim 1$,
\begin{equation}\label{4.19.1}
\int_{|x - \tilde{x}(t)| > a_{0}} \lambda(t)^{-1} Q(\frac{x - x(t)}{\lambda(t)})^{2} dx \lesssim e^{-2 a_{0}},
\end{equation}
and by $(\ref{re1.8})$,
\begin{equation}\label{4.19.2}
\int_{|x - \tilde{x}(t)| > a_{0}} \lambda(t)^{-1} \epsilon(t, \frac{x - x(t)}{\lambda(t)})^{2} dx \leq \delta^{2}.
\end{equation}
Therefore, for $a_{0}$ sufficiently large, plugging $(\ref{4.19.2})$ into $(\ref{4.18})$,
\begin{equation}\label{4.20}
I'(t) \leq \int_{|x - \tilde{x}(t)| \leq a_{0}} u(t,x)^{6} \phi(x - \tilde{x}(0) + x_{0} - \frac{1}{4}(\tilde{x}(t) - \tilde{x}(0))) dx.
\end{equation}
By direct computation,
\begin{equation}\label{4.21}
\phi(x) \leq c e^{-\frac{1}{K} |x - \tilde{x}(0) + x_{0} - \frac{1}{4}(\tilde{x}(t) - \tilde{x}(0))|} = c e^{-\frac{1}{K} |x - \tilde{x}(t) + \frac{3}{4}(\tilde{x}(t) - \tilde{x}(0)) + x_{0}|}.
\end{equation}
Since $\tilde{x}(t) \geq \tilde{x}(0)$ and $|x - \tilde{x}(t)| \leq a_{0}$,
\begin{equation}\label{4.22}
= c e^{-\frac{1}{K}(x - \tilde{x}(t) + \frac{3}{4} (\tilde{x}(t) - \tilde{x}(0)) + x_{0})}.
\end{equation}
Therefore, since from $(\ref{4.10})$, $\dot{\tilde{x}}(t) \geq \frac{1}{2 \lambda^{2}}$, so
\begin{equation}\label{4.23}
I'(t) \leq C e^{\frac{-x_{0}}{K}} e^{-\frac{3}{4K}(\tilde{x}(t) - \tilde{x}(0))} \dot{\tilde{x}}(t) \int \lambda(t)^{2} u(t,x)^{6} dx.
\end{equation}
Making a change of variables, for any $T > 0$,
\begin{equation}\label{4.24}
\int_{0}^{T} I'(t) dt \leq \sum_{k \geq 0} C e^{\frac{-x_{0}}{K}} \int_{k}^{k + 1} x_{s}(s) e^{-\frac{3}{4K}(\tilde{x}(s) - \tilde{x}(0))} \int \lambda(s)^{2} u(t(s), x)^{6} dx ds.
\end{equation}
Then by $(\ref{2.40})$, conservation of mass, and a change of variables,
\begin{equation}\label{4.25}
(\ref{4.24}) \lesssim C K e^{\frac{-x_{0}}{K}}.
\end{equation}
However, by the fundamental theorem of calculus, $(\ref{4.8})$, the fact that by concentration compactness,
\begin{equation}\label{4.26.1}
I(t) \nearrow \int u(0,x)^{2} dx, \qquad \text{as} \qquad t \nearrow \sup(I),
\end{equation}
and $K = 3 \sqrt{2} > 6$ gives a contradiction for $a_{0}$ sufficiently large.\medskip
Proving $(\ref{4.26.1})$ is the only place where $(\ref{4.15.1})$ is used. (Since $[0, t_{0}]$ is a compact set for any $t_{0} \in I$, and $N(t)$ is a continuous function, $(\ref{4.15.1})$ would also hold when $0$ is replaced by any $t_{0} \in I$.) Then by $(\ref{4.10})$, for any $T > 0$, $T \in I$,
\begin{equation}\label{4.26.2}
\tilde{x}(T) - \tilde{x}(0) = \int_{0}^{T} \dot{\tilde{x}}(t) dt \geq \int_{0}^{T} \frac{1}{2 \lambda^{2}} dt \sim \int_{0}^{T} N(t)^{2} dt \rightarrow +\infty,
\end{equation}
as $T \nearrow \sup(I)$. This proves $(\ref{4.26.1})$.
\end{proof}
Now prove exponential decay to the right.
\begin{lemma}[Exponential decay to the right of the soliton]\label{l4.2}
For $x_{0} \geq 10 a_{0}$,
\begin{equation}\label{4.27}
\| u(t, x + x(t)) \|_{L^{2}(x \geq x_{0})}^{2} \leq 10 c_{1} e^{-\frac{x_{0}}{6}}.
\end{equation}
\end{lemma}
\begin{proof}
In this case, observe that if $u(t,x)$ solves $(\ref{1.1})$, then so does $v(t,x) = u(-t, -x)$. Once again assume without loss of generality that $(\ref{4.27})$ fails at $t_{0} = 0$. Define the function
\begin{equation}\label{4.28}
I(t) = \int v(t,x)^{2} \psi(x + \tilde{x}(0) + x_{0} + \frac{1}{4}(\tilde{x}(-t) - \tilde{x}(0))) dx.
\end{equation}
If $(\ref{4.27})$ fails at $t_{0} = 0$ for some $x_{0}$, then
\begin{equation}\label{4.28.1}
I(0) \leq \int u(t,x)^{2} dx - 5 c_{1} e^{-\frac{x_{0}}{6}}.
\end{equation}
Again by direct calculation,
\begin{equation}\label{4.29}
\aligned
I'(t) = -3 \int v_{x}(t,x)^{2} \phi(x + \tilde{x}(0) + x_{0} + \frac{1}{4}(\tilde{x}(-t) - \tilde{x}(0))) dx \\
+ \int v(t,x)^{2} \phi''(x + \tilde{x}(0) + x_{0} + \frac{1}{4}(\tilde{x}(-t) - \tilde{x}(0))) dx \\
+ \frac{5}{3} \int v(t,x)^{6} \phi(x + \tilde{x}(0) + x_{0} + \frac{1}{4}(\tilde{x}(-t) - \tilde{x}(0))) dx \\
- \frac{\dot{\tilde{x}}(-t)}{4} \int v(t,x)^{2} \phi(x + \tilde{x}(0) + x_{0} + \frac{1}{4}(\tilde{x}(-t) - \tilde{x}(0))) dx.
\endaligned
\end{equation}
Making the same argument as in Lemma $\ref{l4.1}$ and making a change of variables
\begin{equation}\label{4.30}
\aligned
I'(t) \leq \int_{|x - \tilde{x}(t)| \leq a_{0}} u(-t, -x)^{6} \phi(x + \tilde{x}(0) + x_{0} + \frac{1}{4}(\tilde{x}(-t) - \tilde{x}(0)) dx \\ = \int_{|x - \tilde{x}(t)| \leq a_{0}} u(-t, x)^{6} \phi(-x + \tilde{x}(0) + x_{0} + \frac{1}{4}(\tilde{x}(-t) - \tilde{x}(0)) dx.
\endaligned
\end{equation}
Then
\begin{equation}\label{4.31}
\phi(x) \leq C e^{-\frac{1}{K} (-x + \tilde{x}(-t) + \frac{3}{4} (\tilde{x}(0) - \tilde{x}(-t)) + x_{0})} \leq C e^{-\frac{3}{4K} (\tilde{x}(0) - \tilde{x}(-t)) - \frac{x_{0}}{K}}
\end{equation}
Therefore, as in Lemma $\ref{l4.1}$, we can show that
\begin{equation}\label{4.32}
\int_{0}^{T} I'(t) dt \lesssim CK e^{-\frac{x_{0}}{K}}.
\end{equation}
This proves $(\ref{4.27})$.
\end{proof}
\noindent \textbf{Remark:} Once again $K = 3 \sqrt{2}$.\medskip
It only remains to prove
\begin{theorem}\label{t4.3}
There does not exist an almost periodic solution to $(\ref{1.1})$ that satisfies $N(t) \geq 1$ for all $t \in I$,
\begin{equation}\label{4.33}
\int_{0}^{\sup(I)} N(t)^{2} dt = \infty,
\end{equation}
and
\begin{equation}\label{4.34}
\int_{\inf(I)}^{0} N(t)^{2} dt < \infty.
\end{equation}
\end{theorem}
\begin{proof}
By $(\ref{4.26.2})$ and $(\ref{4.33})$, exponential decay to the left must hold for such a solution. That is,
\begin{equation}\label{4.35}
\| u(t, x + x(t)) \|_{L^{2}(x \leq -x_{0})} \leq 10 c_{1} e^{-\frac{x_{0}}{K(u)}}.
\end{equation}
Now let $\chi$ be a smooth function such that $\chi(x) = 0$ for $x \leq 1$ and $\chi(x) = 1$ when $x > 2$. Then define the functional
\begin{equation}\label{4.36}
M(t) = \int \chi(\frac{x}{x_{0}}) u(t,x + x(0))^{2} dx.
\end{equation}
The fact that $N(t) \geq 1$ combined with $(\ref{4.34})$ implies $\inf(I) > -\infty$. This fact implies that $N(t) \nearrow \infty$ as $t \searrow \inf(I)$, so $(\ref{4.14})$ combined with almost periodicity imply that
\begin{equation}\label{4.37}
\lim_{t \searrow \inf(I)} M(t) = 0.
\end{equation}
Then integrating by parts,
\begin{equation}\label{4.38}
\aligned
\frac{d}{dt} M(t) = -\frac{3}{x_{0}} \int \chi'(\frac{x}{x_{0}}) u_{x}(t, x + x(0))^{2} + \frac{5}{3 x_{0}} \int \chi'(\frac{x}{x_{0}}) u(t, x + x(0))^{6} dx + \frac{1}{x_{0}^{3}} \int \chi'''(\frac{x}{x_{0}}) u(t, x + x(0))^{2} dx \\
\leq \frac{5}{3 x_{0}} \int \chi'(\frac{x}{x_{0}}) u(t, x + x(0))^{6} dx + \frac{1}{x_{0}^{3}} \int \chi'''(\frac{x}{x_{0}}) u(t, x + x(0))^{2} dx.
\endaligned
\end{equation}
Then by $(\ref{2.40})$,
\begin{equation}\label{4.39}
\int_{\inf(I)}^{0} \frac{d}{dt} M(t) dt \lesssim \frac{5}{3 x_{0}} \int_{\inf(I)}^{0} N(t)^{2} dt - \frac{1}{x_{0}^{3}} \inf(I) \lesssim \frac{1}{x_{0}} (\int_{\inf(I)}^{0} N(t)^{2} dt).
\end{equation}
This implies that for any $t \in (\inf(I), 0]$,
\begin{equation}\label{4.40}
\int \chi(\frac{x}{x_{0}}) u(t, x + x(0))^{2} dx \lesssim \frac{1}{x_{0}} (\int_{\inf(I)}^{0} N(t)^{2} dt).
\end{equation}
Since $|\chi'(\frac{x}{x_{0}})| \leq \chi(\frac{2x}{x_{0}})$, plugging $(\ref{4.40})$ back in to $(\ref{4.39})$,
\begin{equation}\label{4.41}
\aligned
\int_{\inf(I)}^{0} \frac{d}{dt} M(t) dt \lesssim \| \chi(\frac{2 x}{x_{0}}) u \|_{L^{2}}^{2/3} \| u(t) \|_{L^{8}}^{16/3} dt \\ \lesssim \frac{1}{x_{0}^{1/3}}(\int_{\inf(I)}^{0} N(t)^{2} dt)^{1/3} \cdot \frac{1}{x_{0}} (\int_{\inf(I)}^{0} N(t)^{2} dt) = \frac{1}{x_{0}^{4/3}} (\int_{\inf(I)}^{0} N(t)^{2} dt)^{4/3}.
\endaligned
\end{equation}
Therefore, since $\int_{\inf(I)}^{0} N(t)^{2} dt = R < \infty$,
\begin{equation}\label{4.42}
\int_{x \geq 0} u(t, x + x(0))^{2} x dx < \infty,
\end{equation}
which combined with $(\ref{4.36})$ implies
\begin{equation}\label{4.43}
\int |x| u(t, x + x(0))^{2} dx < \infty.
\end{equation}
Then following the proof of Proposition $\ref{t3.2}$,
\begin{equation}\label{4.44}
\int_{\inf(I)}^{0} \int u_{x}(t,x)^{2} dx dt < \infty.
\end{equation}
By the Sobolev embedding theorem, $E(u) < \infty$. Then by conservation of energy and the Gagliardo-Nirenberg inequality, the solution to $(\ref{1.1})$ cannot blow up in finite time, which gives a contradiction.
\end{proof}
The proof that there does not exist a solution satisfying
\begin{equation}\label{4.45}
\int_{0}^{\sup(I)} N(t)^{2} dt < \infty, \qquad \int_{\inf(I)}^{0} N(t)^{2} dt = \infty,
\end{equation}
is identical.
\section{Virial identities}
Next, use the virial identity from \cite{martel2001instability} to show that, on average, the inner product $(\epsilon, Q)$ is bounded by $\| \epsilon \|_{L^{2}}^{2}$.
\begin{theorem}\label{t6.1}
For any $T > 0$,
\begin{equation}\label{6.1}
|\int_{0}^{T} \lambda(s)^{1/2} \int \epsilon(s,x) Q(x) dx ds| \lesssim C(u) + \int_{0}^{T} \lambda(s)^{1/2} \| \epsilon(s) \|_{L^{2}}^{2} ds.
\end{equation}
\end{theorem}
\begin{proof}
Define the quantity,
\begin{equation}\label{6.2}
J(s) = \lambda(s)^{1/2} \int \epsilon(s, x) \int_{-\infty}^{x} (\frac{Q}{2} + z Q_{z}) dz dx - \lambda(s)^{1/2} \kappa,
\end{equation}
where $\kappa = \frac{1}{4} (\int Q)^{2}$. By rescaling, Lemmas $\ref{l4.1}$ and $\ref{l4.2}$, and the fact that $\lambda(s) \leq 1$,
\begin{equation}\label{6.3}
\sup_{s \in \mathbb{R}} J(s) < \infty.
\end{equation}
Then compute
\begin{equation}\label{6.4}
\frac{d}{ds} J(s) = \lambda(s)^{1/2} \int \epsilon_{s}(s, x) \int_{-\infty}^{x} (\frac{Q}{2} + z Q_{z}) dz dx + \frac{\lambda_{s}}{2 \lambda^{1/2}} \int \epsilon(s, x) \int_{-\infty}^{x} (\frac{Q}{2} + z Q_{z}) dz dx - \frac{\lambda_{s}}{2 \lambda^{1/2}} \kappa.
\end{equation}
Then taking the expression of $\epsilon_{s}$ given by $(\ref{2.30})$, and integrating by parts,
\begin{equation}\label{6.5}
-\int R(\epsilon)_{y} \int_{-\infty}^{y} (\frac{Q}{2} + z Q_{z}) dz dy = \int R(\epsilon) (\frac{Q}{2} + y Q_{y}) dy \lesssim \| \epsilon \|_{L^{2}}^{2} + \| \epsilon \|_{L^{2}} \| \epsilon \|_{L^{8}}^{4}.
\end{equation}
Next, integrating by parts, by $(\ref{2.35})$,
\begin{equation}\label{6.7}
(\frac{x_{s}}{\lambda} - 1) \int \epsilon_{y} \int_{-\infty}^{y} \frac{Q}{2} + z Q_{z} dz dy = -(\frac{x_{s}}{\lambda} - 1) \int \epsilon (\frac{Q}{2} + y Q_{y}) dy \lesssim \| \epsilon \|_{L^{2}}^{2} + \| \epsilon \|_{L^{2}}^{2} \| \epsilon \|_{L^{8}}^{4}.
\end{equation}
Next, integrating by parts and using $\epsilon \perp y(\frac{Q}{2} + y Q_{y})$,
\begin{equation}\label{6.8}
\aligned
\frac{\lambda_{s}}{\lambda} \int (\frac{\epsilon}{2} + y \epsilon_{y}) \int_{-\infty}^{y} (\frac{Q}{2} + z Q_{z}) dz dy = -\frac{1}{2} \frac{\lambda_{s}}{\lambda} \int \epsilon(s, y) \int_{-\infty}^{y} (\frac{Q}{2} + z Q_{z}) dz dy \\
- \frac{\lambda_{s}}{\lambda} \int \epsilon(s,y) y (\frac{Q}{2} + y Q_{y}) dy = -\frac{1}{2} \frac{\lambda_{s}}{\lambda} \int \epsilon(s, y) \int_{-\infty}^{y} (\frac{Q}{2} + z Q_{z}) dz dy.
\endaligned
\end{equation}
By direct calculation,
\begin{equation}\label{6.9}
(\frac{x_{s}}{\lambda} - 1) \int Q_{y} \int_{-\infty}^{y} \frac{Q}{2} + z Q_{z} dz dx = -(\frac{x_{s}}{\lambda} - 1) \int Q(\frac{Q}{2} + y Q_{y}) = 0.
\end{equation}
Also, since $Q$ is an even function,
\begin{equation}\label{6.10}
\frac{\lambda_{s}}{\lambda} \int (\frac{Q}{2} + y Q_{y}) \int_{-\infty}^{y} \frac{Q}{2} + z Q_{z} dz dx = \frac{\lambda_{s}}{\lambda} \frac{1}{2} (\int \frac{Q}{2} + y Q_{y} dy)^{2} = \frac{\lambda_{s}}{\lambda} \kappa.
\end{equation}
Finally, since $L$ is a self-adjoint operator,
\begin{equation}\label{6.11}
\int (L \epsilon)_{y} \int_{-\infty}^{y} \frac{Q}{2} + z Q_{z} dz = -\int (L \epsilon) (\frac{Q}{2} + y Q_{y}) dy = -\int \epsilon \cdot L(\frac{Q}{2} + y Q_{y}) dy.
\end{equation}
Now, by direct computation,
\begin{equation}\label{6.12}
\aligned
L(\frac{Q}{2} + x Q_{x}) = -\frac{Q_{xx}}{2} + \frac{Q}{2} - \frac{5}{2} Q^{5} - x Q_{xxx} - 2 Q_{xx} - 5 x Q^{4} Q_{x} + x Q_{x} \\ = x \partial_{x}(-Q_{xx} - Q^{5} + Q) - \frac{5}{2} (Q_{xx} + Q^{5}) + \frac{Q}{2} = -2 Q.
\endaligned
\end{equation}
Plugging this into $(\ref{6.11})$,
\begin{equation}\label{6.13}
(\ref{6.11}) = 2 \int Q \epsilon.
\end{equation}
Therefore, we have proved,
\begin{equation}\label{6.14}
\frac{d}{ds} J(s) = 2 \lambda(s)^{1/2} \int Q(y) \epsilon(s, y) dy + O(\lambda(s) \| \epsilon \|_{L^{2}}^{2}) + O(\lambda(s) \| \epsilon \|_{L^{2}} \| \epsilon \|_{L^{8}}^{4}).
\end{equation}
Using $(\ref{2.41})$ to estimate $\| \epsilon \|_{L_{s,y}^{8}}$ proves the theorem.
\end{proof}
We are now ready to finish the proof of the main result.
\begin{proof}[{\bf Proof of Theorem $\ref{t1.1}$} ]
Theorem $\ref{t1.1}$ may now be proved using a second virial identity. Let
\begin{equation}\label{5.1}
M(s) = \frac{1}{2} \lambda(s) \int y \epsilon(s, y)^{2} dy.
\end{equation}
Lemmas $\ref{l4.1}$ and $\ref{l4.2}$ imply that $(\ref{5.1})$ is uniformly bounded for all $s \in \mathbb{R}$.
Now, by the product rule,
\begin{equation}\label{2.48}
\frac{d}{ds} M(s) = \lambda(s) \int y \epsilon(s, y) \epsilon_{s}(s, y) dy + \frac{1}{2} \lambda_{s}(s) \int y \epsilon(s,y)^{2} dy.
\end{equation}
Again use $(\ref{2.30})$ to compute $\epsilon_{s}$. Integrating by parts,
\begin{equation}\label{2.49}
\aligned
\int y \epsilon (L \epsilon)_{y} dy = \int y \epsilon (-\epsilon_{yyy} + \epsilon_{y} - 20 Q^{3} Q_{y} \epsilon - 5 Q^{4} \epsilon_{y}) dy \\ = -\frac{3}{2} \int \epsilon_{y}^{2} dy - \frac{1}{2} \int \epsilon^{2} dy - 10 \int Q^{3} Q_{y} y \epsilon^{2} dy - \frac{5}{2} \int Q^{4} \epsilon^{2} dy =: H(\epsilon, \epsilon).
\endaligned
\end{equation}
Next, since $\epsilon \perp y Q_{y}$ and $\epsilon \perp y(\frac{Q}{2} + y Q_{y})$ for all $s \in \mathbb{R}$,
\begin{equation}\label{6.20}
\frac{\lambda_{s}}{\lambda} \int y \epsilon (\frac{Q}{2} + y Q_{y}) dy = (\frac{x_{s}}{\lambda} - 1) \int y \epsilon Q_{y} dy = 0.
\end{equation}
Next, integrating by parts and using $(\ref{2.35})$,
\begin{equation}\label{2.51}
(\frac{x_{s}}{\lambda} - 1) \int y \epsilon \epsilon_{y} = -(\frac{x_{s}}{\lambda} - 1) \int \epsilon^{2} dy \lesssim \| \epsilon \|_{L^{2}}^{3} (1 + \| \epsilon \|_{L^{8}}^{4}) \lesssim \| \epsilon \|_{L^{2}}^{3} + \| \epsilon \|_{L^{2}}^{11/2} \| \epsilon_{y} \|_{L^{2}}^{3/2}.
\end{equation}
Also,
\begin{equation}\label{2.52}
\aligned
-\int R(\epsilon)_{y} \epsilon(s,y) y dy = -\int y \epsilon (10 Q^{3} \epsilon^{2} + 10 Q^{2} \epsilon^{3} + 5 Q \epsilon^{4} + \epsilon^{5})_{y} dy = \frac{20}{3} \int Q^{3} \epsilon^{3} - 10 \int Q^{2} Q_{y} y \epsilon^{3} \\ -5 \int Q Q_{y} y \epsilon^{4} + \frac{15}{2} \int Q^{2} \epsilon^{4} + 4 \int Q \epsilon^{5} - \int Q_{y} y \epsilon^{5} + \frac{5}{6} \epsilon^{6} \\ \lesssim \| \epsilon \|_{L^{2}}^{3/2} \| \epsilon \|_{L^{6}}^{3/2} + \| \epsilon \|_{L^{6}}^{6} \lesssim \| \epsilon \|_{L^{2}}^{5/2} \| \epsilon_{y} \|_{L^{2}}^{1/2} + \| \epsilon \|_{L^{2}}^{4} \| \epsilon_{y} \|_{L^{2}}^{2}.
\endaligned
\end{equation}
Finally, integrating by parts,
\begin{equation}\label{2.53}
\frac{\lambda_{s}}{\lambda} \int y \epsilon (\frac{\epsilon}{2} + y \epsilon_{y}) = -\frac{\lambda_{s}}{2 \lambda} \int y \epsilon^{2} = -\frac{\lambda_{s}}{ \lambda} M(s).
\end{equation}
Multiplying $(\ref{2.49})$--$(\ref{2.53})$ by $\lambda(s)$ and plugging in to $(\ref{2.48})$,
\begin{equation}\label{6.24}
\aligned
\int_{0}^{T} \lambda(s) H(\epsilon, \epsilon) ds \lesssim C(u) + \int_{0}^{T} \lambda(s) \| \epsilon \|_{L^{2}}^{3} + \lambda(s) \| \epsilon \|_{L^{2}}^{11/2} \| \epsilon_{y} \|_{L^{2}}^{3/2} ds \\ + \int_{0}^{T} \lambda(s) \| \epsilon \|_{L^{2}}^{5/2} \| \epsilon_{y} \|_{L^{2}}^{1/2} + \lambda(s) \| \epsilon \|_{L^{2}}^{4} \| \epsilon_{y} \|_{L^{2}}^{2} ds \lesssim C(u) + \delta \int_{0}^{T} \lambda(s) \| \epsilon \|_{L^{2}}^{2} ds + \delta \int_{0}^{T} \lambda(s) \| \epsilon_{y} \|_{L^{2}}^{2} ds.
\endaligned
\end{equation}
The last inequality follows from $(\ref{re1.8})$.\medskip
Now then, take
\begin{equation}\label{6.25}
\epsilon_{1} = \epsilon - \frac{(\epsilon, Q)}{\| Q \|_{L^{2}}^{2}} Q = \epsilon - a Q.
\end{equation}
Since $Q \perp x(\frac{Q}{2} + x Q_{x})$, $\epsilon_{1} \perp Q$ and $\epsilon_{1} \perp x(\frac{Q}{2} + x Q_{x})$. Therefore, from \cite{martel2000liouville}, there exists some $\delta_{1} > 0$ such that
\begin{equation}\label{6.26}
H(\epsilon_{1}, \epsilon_{1}) \geq \delta_{1} \| \epsilon_{1} \|_{H^{1}}^{2}.
\end{equation}
Also, integrating by parts,
\begin{equation}\label{6.27}
2 \lambda(s) H(\epsilon_{1}, aQ) + \lambda(s) H(aQ, aQ) \lesssim \lambda(s)^{1/2} |a| \cdot \lambda(s)^{1/2} \| \epsilon_{1} \|_{L^{2}} + \lambda(s) a^{2}.
\end{equation}
Therefore, $(\ref{6.24})$ and $(\ref{6.25})$ imply
\begin{equation}\label{6.27.1}
\aligned
\delta_{1} \int_{0}^{T} \lambda(s) \| \epsilon_{1} \|_{H^{1}}^{2} ds \lesssim C(u) + \delta \int_{0}^{T} \lambda(s) \| \epsilon \|_{H^{1}}^{2} ds + \int_{0}^{T} \lambda(s) a(s)^{2} ds + \int_{0}^{T} \lambda(s) a(s) \| \epsilon_{1} \|_{L^{2}} ds \\
\lesssim C(u) + \delta \int_{0}^{T} \lambda(s) \| \epsilon_{1} \|_{H^{1}}^{2} ds + \int_{0}^{T} \lambda(s) a(s)^{2} ds + \int_{0}^{T} \lambda(s) a(s) \| \epsilon_{1} \|_{L^{2}} ds.
\endaligned
\end{equation}
Furthermore, for $\delta \ll \delta_{1}$, absorbing $ \delta \int_{0}^{T} \lambda(s) \| \epsilon_{1} \|_{H^{1}}^{2} ds$ into the left hand side,
\begin{equation}\label{6.27.2}
\aligned
\frac{\delta_{1}}{2} \int_{0}^{T} \lambda(s) \| \epsilon_{1} \|_{H^{1}}^{2} ds \lesssim C(u) + \int_{0}^{T} \lambda(s) a(s)^{2} ds + \int_{0}^{T} \lambda(s) a(s) \| \epsilon_{1} \|_{L^{2}} ds.
\endaligned
\end{equation}
Also, by the Cauchy-Schwarz inequality,
\begin{equation}\label{6.28}
\frac{\delta_{1}}{4} \int_{0}^{T} \lambda(s) \| \epsilon_{1} \|_{H^{1}}^{2} ds \lesssim C(u) + \frac{1}{\delta_{1}} \int_{0}^{T} \lambda(s) (\epsilon, Q)^{2} ds.
\end{equation}
Also, since
\begin{equation}\label{6.29}
\| \epsilon \|_{H^{1}}^{2} \lesssim \| \epsilon_{1} \|_{H^{1}}^{2} + (\epsilon, Q) \| Q \|_{H^{1}}^{2},
\end{equation}
\begin{equation}\label{6.30}
\frac{\delta_{1}}{4} \int_{0}^{T} \lambda(s) \| \epsilon \|_{H^{1}}^{2} ds \lesssim C(u) + \frac{1}{\delta_{1}} \int_{0}^{T} \lambda(s) (\epsilon, Q)^{2} ds.
\end{equation}
Next, by conservation of mass and scaling invariance of the $L^{2}$ norm,
\begin{equation}\label{6.31}
\frac{1}{2} \| u_{0} \|_{L^{2}}^{2} = \frac{1}{2} \| Q + \epsilon \|_{L^{2}}^{2} = \frac{1}{2} \| Q \|_{L^{2}}^{2} + (\epsilon, Q) + \frac{1}{2} \| \epsilon \|_{L^{2}}^{2},
\end{equation}
and therefore, after doing some algebra,
\begin{equation}\label{6.32}
-(\epsilon, Q) = \frac{1}{2} \| Q \|_{L^{2}}^{2} - \frac{1}{2} \| u_{0} \|_{L^{2}}^{2} + \frac{1}{2} \| \epsilon \|_{L^{2}}^{2}.
\end{equation}
Since $\frac{1}{2} \| Q \|_{L^{2}}^{2} - \frac{1}{2} \| u_{0} \|_{L^{2}}^{2} > 0$ is a conserved quantity, it is convenient to label this quantity
\begin{equation}\label{6.33}
M = \frac{1}{2} \| Q \|_{L^{2}}^{2} - \frac{1}{2} \| u_{0} \|_{L^{2}}^{2}.
\end{equation}
Plugging $(\ref{6.32})$ into the right hand side of $(\ref{6.30})$,
\begin{equation}\label{6.34}
\frac{\delta_{1}}{4} \int_{0}^{T} \lambda(s) \| \epsilon \|_{H^{1}}^{2} ds \lesssim C(u) + \frac{M^{2}}{\delta_{1}} \int_{0}^{T} \lambda(s) ds + \frac{1}{\delta_{1}} \int_{0}^{T} \lambda(s) \| \epsilon \|_{L^{2}}^{4}.
\end{equation}
Since $\| \epsilon \|_{L^{2}} \lesssim \delta$, the second term in the right hand side may be absorbed into the left hand side, so
\begin{equation}\label{6.35}
\frac{\delta_{1}}{8} \int_{0}^{T} \lambda(s) \| \epsilon \|_{H^{1}}^{2} \lesssim C(u) + \frac{M^{2}}{\delta_{1}} \int_{0}^{T} \lambda(s) ds.
\end{equation}
Likewise, by Theorem $\ref{t6.1}$ and $(\ref{6.32})$,
\begin{equation}\label{6.36}
M \int_{0}^{T} \lambda(s)^{1/2} ds \lesssim C(u) + \int_{0}^{T} \lambda(s)^{1/2} \| \epsilon \|_{L^{2}}^{2} ds.
\end{equation}
Letting
\begin{equation}\label{6.37}
K = \int_{0}^{T} \lambda(s) ds, \qquad \text{and} \qquad R = \int_{0}^{T} \lambda(s)^{1/2} ds,
\end{equation}
combining $(\ref{6.35})$ and $(\ref{6.36})$,
\begin{equation}\label{6.38}
\frac{\delta_{1}}{8} \int_{0}^{T} \lambda(s) \| \epsilon \|_{H^{1}}^{2} ds \lesssim \frac{M K}{R \delta_{1}} \int_{0}^{T} \lambda(s)^{1/2} \| \epsilon \|_{L^{2}}^{2} ds + C(u) + \frac{MK}{R \delta_{1}} C(u).
\end{equation}
If it were the case that $\lambda(s) = 1$ for all $s \in \mathbb{R}$, (as in \cite{martel2001instability}), the proof would be complete, since in that case, $K = R = T$ and $M \lesssim \| \epsilon \|_{L^{2}} \leq \delta$, so for $\delta > 0$ sufficiently small, $(\ref{6.38})$ along with the fact that
\begin{equation}\label{6.39}
\lim_{T \nearrow \infty} \int_{0}^{T} \lambda(s) ds = \lim_{T \nearrow \infty} \int_{0}^{T} \lambda(s)^{1/2} ds = \infty,
\end{equation}
would imply that there exists a sequence $s_{n} \rightarrow +\infty$ such that
\begin{equation}\label{6.39.1}
\| \epsilon(s_{n}) \|_{H^{1}} \rightarrow 0,
\end{equation}
as $n \rightarrow \infty$. However, this would contradict the fact that $\| u_{0} \|_{L^{2}} < \| Q \|_{L^{2}}$.\medskip
In the general case, the proof will make use of the fact that $\lambda(s) \leq 1$ for all $s \in \mathbb{R}$ along with the fact that conservation of energy gives a lower bound (depending on $M$) on $\lambda(s)$.\medskip
Expanding out the energy,
\begin{equation}\label{6.40}
\aligned
E(Q + \epsilon) = \frac{1}{2} \int Q_{x}^{2} + \int Q_{x} \epsilon_{x} + \frac{1}{2} \int \epsilon_{x}^{2} \\ - \frac{1}{6} \int Q^{2} - \int Q^{5} \epsilon - \frac{5}{2} \int Q^{4} \epsilon^{2} - \frac{10}{3} \int Q^{3} \epsilon^{3} - \frac{5}{2} \int Q^{2} \epsilon^{4} - \int Q \epsilon^{5} - \frac{1}{6} \int \epsilon^{6}.
\endaligned
\end{equation}
First, note that
\begin{equation}\label{6.41}
E(Q) = \frac{1}{2} \int Q_{x}^{2} - \frac{1}{6} \int Q^{6} = 0.
\end{equation}
Next, integrating by parts, by $(\ref{6.32})$,
\begin{equation}\label{6.42}
\int Q_{x} \epsilon_{x} - \int Q^{5} \epsilon = -\int \epsilon(Q_{xx} + Q^{5}) = -\int \epsilon Q = M + \frac{1}{2} \int \epsilon^{2}.
\end{equation}
Therefore, by H{\"o}lder's inequality and the Sobolev embedding theorem,
\begin{equation}\label{6.43}
E(Q + \epsilon) = M + \frac{1}{2} \int \epsilon_{x}^{2} + \frac{1}{2} \int \epsilon^{2} - \frac{5}{2} \int Q^{4} \epsilon^{2} + O(\| \epsilon \|_{L^{2}}^{5/2} \| \epsilon \|_{H^{1}}^{1/2} + \| \epsilon \|_{L^{2}}^{4} \| \epsilon \|_{H^{1}}^{2}).
\end{equation}
Also, scaling symmetry implies
\begin{equation}\label{6.44}
E(Q + \epsilon) = \lambda(s)^{2} E_{0}.
\end{equation}
Recalling $(\ref{6.25})$ and $(\ref{6.26})$,
\begin{equation}\label{6.45}
\aligned
\frac{1}{2} \int \epsilon_{x}^{2} + \frac{1}{2} \int \epsilon^{2} - \frac{5}{2} \int Q^{4} \epsilon^{2} \geq \delta_{1} \| \epsilon_{1} \|_{H^{1}}^{2} - \frac{1}{\delta_{1}} (\epsilon, Q)^{2} \geq \delta_{1} \| \epsilon \|_{H^{1}}^{2} - \frac{2}{\delta_{1}} (\epsilon, Q)^{2} \\
\geq \delta_{1} \| \epsilon \|_{H^{1}}^{2} - \frac{2}{\delta_{1}} M^{2} - \frac{2}{\delta_{1}} \| \epsilon \|_{L^{2}}^{4} \geq \frac{\delta_{1}}{2} \| \epsilon \|_{H^{1}}^{2} - O(\frac{M^{2}}{\delta_{1}}).
\endaligned
\end{equation}
Since $M \lesssim \delta$ and $\| \epsilon \|_{L^{2}} \lesssim \delta$, for $\delta > 0$ sufficiently small,
\begin{equation}\label{6.46}
\lambda(s)^{2} E_{0} \geq \frac{\delta_{1}}{4} \| \epsilon \|_{H^{1}}^{2} + \frac{M}{2}.
\end{equation}
Since $E_{0}$ and both of the terms on the right hand side are positive, $(\ref{6.46})$ implies
\begin{equation}\label{6.47}
M \lesssim \lambda(s)^{2} E_{0},
\end{equation}
and therefore,
\begin{equation}\label{6.48}
\frac{M}{E_{0}} \lesssim \lambda(s)^{2}, \qquad \text{which implies} \qquad \lambda(s)^{-1/2} \lesssim (\frac{E_{0}}{M})^{1/4}.
\end{equation}
Plugging this into $(\ref{6.38})$,
\begin{equation}\label{6.49}
\frac{\delta_{1}}{8} \int_{0}^{T} \lambda(s) \| \epsilon \|_{H^{1}}^{2} ds \lesssim \frac{M^{3/4} E_{0}^{1/4} K}{R} \int_{0}^{T} \lambda(s) \| \epsilon \|_{L^{2}}^{2} ds + \frac{M K}{R \delta_{1}} C(u) + C(u).
\end{equation}
Since $\lambda(s) \leq 1$, $K \leq R$, so
\begin{equation}\label{6.50}
\frac{\delta_{1}}{8} \int_{0}^{T} \lambda(s) \| \epsilon \|_{H^{1}}^{2} ds \lesssim M^{3/4} E_{0}^{1/4} \int_{0}^{T} \lambda(s) \| \epsilon \|_{L^{2}}^{2} ds + C(u).
\end{equation}
Assuming for a moment that $E_{0} \lesssim 1$, $M \lesssim \delta$ and $(\ref{6.39})$ imply that $(\ref{6.39.1})$ must hold in this case as well, obtaining a contradiction.\medskip
The fact that $E_{0} \lesssim 1$ is a straightforward consequence of Lemmas $\ref{l4.1}$ and $\ref{l4.2}$. Suppose without loss of generality that
\begin{equation}\label{6.51}
\lambda(0) \geq \frac{1}{2} = \frac{1}{2} \sup_{s \in \mathbb{R}} \lambda(s).
\end{equation}
Lemmas $\ref{l4.1}$ and $\ref{l4.2}$ imply that
\begin{equation}\label{6.52}
\lambda(s) \int y \epsilon(s, y)^{2} dy \lesssim 1,
\end{equation}
with implicit constant independent of $u$, so long as $u$ satisfies $(\ref{re1.8})$. Then by $(\ref{6.30})$,
\begin{equation}\label{6.53}
\int_{0}^{1} \lambda(s) \| \epsilon \|_{H^{1}}^{2} ds \lesssim 1 + \frac{1}{\delta_{1}} \int_{0}^{1} \lambda(s) \| \epsilon \|_{L^{2}}^{2} ds.
\end{equation}
Since $(\ref{2.39})$ guarantees that $\lambda(s) \sim 1$ on $[0, 1]$,
\begin{equation}\label{6.54}
\int_{0}^{1} \| \epsilon \|_{H^{1}}^{2} ds \lesssim 1 + \frac{1}{\delta_{1}} \int_{0}^{1} \| \epsilon \|_{L^{2}}^{2} ds \lesssim 1.
\end{equation}
The last inequality follows from $(\ref{re1.8})$. Therefore, the proof that $E_{0} \lesssim 1$ is complete.
\end{proof}
\nocite{*}
|
\section{Introduction}
Category theory provides a means through which many far-ranging fields of mathematics can be related by their similar structure. In a paper by Robinson \cite{Robinson1981}, this interconnectivity afforded by categorical perspectives allowed for the realization of torsion products as the homotopy groups of a topological space, which is itself constructed for this express purpose. However, even stating this result formally requires a multitude of preliminaries in algebra, topology, and category theory.
The goal of this document is to present a self-contained guide to the fundamental concepts and results, with few proofs, required to do work with this kind of mathematics in hopes of making the field of homotopical algebra more accessible. We only assume familiarity with topological spaces and groups, so it is approachable from an undergraduate level. This project culminates in a discussion of the result of Robinson mentioned above along with a computation as a proof of concept.
\newpage
\section{Categories}
\subsection{Basic Notions}
\begin{defin}
A \textit{category} $\Cs{C}$ consists of the listed data:
\begin{itemize}
\item a class of \textit{objects} $\mathrm{Obj}(\Cs{C});$
\item for any two objects $A,B \in \mathrm{Obj}(\Cs{C})$ there is a set $\Mor{\Cs{C}}{A}{B}$ of \textit{morphisms} from $A$ to $B$ such that
\[
\Mor{\Cs{C}}{A}{B} \cap \Mor{\Cs{C}}{A'}{B'} = \varnothing,
\]
if $A \ne A'$ or $B \ne B';$
\item for any objects $A,B,C \in \ope{Obj}(\Cs{C})$, there is defined a \textit{composition} operation for morphisms defined by
\[
\circ:
\begin{cases}
\Mor{\Cs{C}}{B}{C} \times \Mor{\Cs{C}}{A}{B} \to \Mor{\Cs{C}}{A}{C} \\
(g,f) \mapsto g \circ f
\end{cases}
\]
with the condition that this operation is associative and that for any object $A \in \mathrm{Obj}(\Cs{C})$ there is a morphism $\mathds{1}_A \in \Mor{\Cs{C}}{A}{A}$ called the \textit{identity} such that for any object $B,$ $f \in \Mor{\Cs{C}}{A}{B},$ and $g \in \Mor{\Cs{C}}{B}{A}$:
\[
f \circ \mathds{1}_A = f, \qquad \mathds{1}_A \circ g = g.
\]
\end{itemize}
\end{defin}
\begin{rem}
In general $\mathrm{Obj}(\Cs{C})$ is a class, not necessarily a set; whenever $\mathrm{Obj}(\Cs{C})$ is a set, $\Cs{C}$ is called a \textit{small category}. Small categories may be visualized as graphs in which vertices correspond to objects and edges correspond to morphisms. For notational simplicity, it is common to write $X \in \Cs{C}$ in place of $X \in\mathrm{Obj}(\Cs{C})$ whenever it does not cause confusion. In addition, it is common to write $\Cs{C}(A,B)$ or even $(A,B)$ in place of $\Mor{\Cs{C}}{A}{B}$ for $A,B \in \Cs{C}$ when the category is understood.
\end{rem}
\begin{defin}
Let $\Cs{C}$ be a category. Define an \textit{opposite category} $\Cs{C}^{op}$ as having the same objects---but denoting object $A$ by $A^{op}$---and $$\Mor{\Cs{C}^{op}}{A^{op}}{B^{op}} = \Mor{\Cs{C}}{B}{A}$$ with the composition rule $$f^{op} \circ g^{op} = (g \circ f)^{op}.$$
\end{defin}
\begin{theo}[Duality Principle]
If $P$ is a true statement within a category $\Cs{C}$, then the dual statement $P^{op}$ obtained by reversing all arrows is true about category $\Cs{C}^{op}$.
\end{theo}
\begin{exam}
The following are examples of some common categories:
\begin{itemize}
\item category $\Cb{Set}$ of sets and functions;
\item category $\Cb{Top}$ of topological spaces and continuous functions;
\item category $\Cb{Top}_*$ of pointed topological spaces with base point-preserving continuous functions;
\item category $\Cb{Ab}$ of abelian groups and group homomorphisms;
\item category $\Cb{Ring}$ of unital rings and ring homomorphisms;
\item category $\Cb{R\text{-}Mod}$ of left $R$-modules and homomorphisms of left $R$-modules.
\end{itemize}
\end{exam}
\begin{exam} \label{pos}
A partially ordered set (poset) $(I,\leq)$ may be viewed as a category $\Cs{P}$ by regarding
\begin{itemize}
\item objects $\mathrm{Obj}(\Cs{P}) = \{i\ |\ i \in I\}$,
\item morphisms $\Mor{\Cs{P}}{i}{j} =
\begin{cases}
f_{ij}:i \to j, & \text{if }i \leq j \\
\varnothing, & \text{otherwise.}
\end{cases}$
\end{itemize}
An especially important poset is $\Delta[n] = \{0 \leq 1 \leq \cdots \leq n\}$, whose category representation may be illustrated diagrammatically as
\[
\begin{tikzcd}
0 \ar[r] & 1 \ar[r] & \cdots \ar[r] & n,
\end{tikzcd}
\]
after omitting the arrows derived from transitivity and identities. Furthermore, $\Delta[n]$ represents the ordinal with $n+1$ elements, which will become important later on in the discussion of simplicial sets.
\end{exam}
\begin{defin}
A morphism $f$ is called a \textit{monomorphism} (monic) if it is left cancellable; that is for any morphisms $g_1,g_2$, $$fg_1 = fg_2 \Longrightarrow g_1 = g_2.$$ Dually, a morphism $f$ is called an \textit{epimorphism} (epic) if it is right cancellable; that is for any morphisms $h_1,h_2$, $$h_1f = h_2f \Longrightarrow h_1 = h_2.$$ A morphism which is both epic and monic is called a \textit{bimorphism}.
\end{defin}
\begin{rem}
The definition of a monomorphism is based on the condition for a function between sets to be injective, and in fact monomorphisms do coincide with injective functions in the category of sets. On the other hand, epimorphisms are related to surjections, but do not coincide with them in all categories.
\end{rem}
\begin{prop}
Epimorphisms are surjections in both categories $\Cb{Set}$ and $\Cb{R\text{-}Mod},$ but they are not necessarily surjective in $\Cb{Ring}.$
\end{prop}
\begin{proof}
Let us start in $\Cb{Set}.$ Suppose that $f:X \to Y$ is an epimorphism and that $g_1,g_2:Y \to Z$ are any functions such that $g_1f = g_2f.$ This indicates that $g_1$ and $g_2$ coincide on the image of $f,$ but we can say no more because elements of $Y$ outside the image of $f$ may be mapped to anywhere. But our hypothesis implies that $g_1 = g_2$ on all of $Y,$ which is only possible if there are no elements of $Y$ not in the image of $f.$ Hence, $f$ must be surjective.
Now let us investigate the category $\Cb{Ring}.$ It will suffice to give a counterexample, so consider the inclusion $i:\Z \to \Q.$ Again, let $g_1,g_2$ be any ring homomorphisms with source $\Q$ such that $g_1i = g_2i.$ This tells us that $g_1$ coincides with $g_2$ on the integers. Because ring homomorphisms preserve the multiplicative structure, we can deduce where all the reciprocals of the integers are mapped to. Extending this further, we see that every rational is a product of an integer and the reciprocal of another integer, so any homomorphism from $\Q$ is completely determined by where it sends $\Z.$ Therefore, $g_1i = g_2i \Longrightarrow g_1 = g_2,$ but clearly this is not a surjection.
Finally, we examine $\Cb{R\text{-}Mod}.$ Let $f:X \to Y$ be an epimorphism. Now consider the natural projection map $\pi:Y \to Y/\Im(f)$ and the zero map $z:Y \to Y/\Im(f).$ Clearly $\pi f = zf,$ so $\pi = z$ by hypothesis. Projection maps are always surjective, so this implies that $\Im(f) = Y.$ Therefore, $f$ is a surjection.
\end{proof}
\begin{defin} \label{iso}
A morphism $f$ is called a \textit{retraction} if it is right invertible while a morphism $g$ is called a \textit{section} if it is left invertible; that is, if $$fg = \mathds{1},$$ then $f$ is a retraction, and $g$ is a section. Notice that retractions and sections come in pairs. A morphism $h:A \to B$ which is both a retraction and a section is called an \textit{isomorphism}. In this case, the objects $A$ and $B$ are said to be isomorphic, and we write $A\cong B$.
\end{defin}
\begin{lem}
Let $\Cs{C}$ be a category. All the following maps are morphisms within $\Cs{C}$ with sources and targets implicit from the allowed compositions.
\begin{enumerate}[label = \emph{(\roman*)}]
\item If $s_1$ and $s_2$ are sections, then $s_2s_1$ is a section.
\item If $r_1$ and $r_2$ are retractions, then $r_1r_2$ is a retraction.
\item If $s_2s_1$ is a section, then $s_1$ is a section.
\item If $r_1r_2$ is a retraction, then $r_1$ is a retraction.
\end{enumerate}
\end{lem}
\begin{proof}
The proofs for sections will be formally stated but not for retractions, as they are very similar.
\begin{enumerate}[label = (\roman*)]
\item Because $s_1,s_2$ are sections, there exist maps $r_1,r_2$ such that $$r_1s_1 = \mathds{1}, ~~~~ r_2s_2 = \mathds{1}.$$ Because we may form the composition $s_2s_1,$ the composition $r_1r_2$ may be formed since the source of $r_1$ is the target of $r_2.$ Then we find that $$(r_1r_2)(s_2s_1) = r_1(r_2s_2)s_1 = r_1\mathds{1}s_1 = r_1s_1 = \mathds{1},$$ so $r_1r_2$ is a left inverse for $s_2s_1.$ Therefore, $s_2s_1$ is a section.
\item This proof follows from switching the roles of the hypothesis and constructed maps in (i).
\item Because $s_2s_1$ is a section, there exists a map $r$ such that $rs_2s_1 = \mathds{1}.$ But then $rs_2$ is a left inverse for $s_1,$ so $s_1$ is a section.
\item Similar to (iii).
\end{enumerate}
\end{proof}
\begin{defin}
Consider a category $\Cs{C}$. An object $I \in \Cs{C}$ is called an \textit{initial object} if there is exactly one morphism from $I$ to any object in the category. Dually, an object $T\in\Cs{C}$ is called a \textit{terminal object} if there is exactly one morphism from any object in the category into $T$. An object which is both initial and terminal is called a \textit{zero object}, which is commonly denoted by 0. A category $\Cs{C}$ with a zero object is called a \textit{pointed category}.
\end{defin}
\begin{rem}
Once cokernels are introduced, it will become evident that the proof from $\Cb{R\text{-}Mod}$ may be recycled for most all pointed categories which have cokernels to show that epimorphisms coincide with surjections. On the other hand, the notions were not equivalent in $\Cb{Ring}$ by exploiting that there exist subrings which generate their field of fractions. As such, it is not difficult to see that epimorphisms are not surjections for commutative monoids.
\end{rem}
\begin{exam}
Outlined below are some categories with initial, terminal, and zero objects indicated when applicable.
\begin{itemize}
\item In $\Cb{Set},$ the empty set is an initial object while any singleton is a terminal object. The same is true in $\Cb{Top}$ when the empty set and singleton are regarded as topological spaces.
\item In the poset category $(I,\leq),$ the bottom (resp. top) element is an initial (resp. terminal) object, should it exist.
\item In $\Cb{Ring},$ $\Z$ is the initial object while the zero ring is a terminal object. In fact, the image of $\Z$ in a ring $R$ is called the \textit{prime ring} of $R$ and is used to define the characteristic of a ring.
\item $\Cb{Ab}$ is a pointed category with zero object the trivial group.
\item $\Cb{Top}_*$ is a pointed category with zero objects being any singleton.
\end{itemize}
\end{exam}
\begin{rem}
Initial, terminal, and zero objects are unique up to unique isomorphism if they exist. Notice that initial objects in $\Cs{C}$ are terminal objects in $\Cs{C}^{op}$ and vice versa. In a pointed category $\Cs{C}$, between any two objects $A,B \in \Cs{C}$ there exists a unique morphism $0_{A,B}:A \to B$, which is defined as the composition $$A \to 0 \to B.$$ Because the zero map is unique, the subscripts in $0_{A,B}$ are usually omitted so that we may instead use 0 to denote both the zero object and the zero map.
\end{rem}
\begin{defin}
Let $\Cs{C}$ and $\Cs{D}$ be categories. A \textit{covariant functor} $F:\Cs{C}\to\Cs{D}$ is a pair of correspondences between classes of objects and morphisms in $\Cs{C}$ and $\Cs{D}$ defined by
\[
F:
\begin{cases}
\mathrm{Obj}(\Cs{C}) \to \mathrm{Obj}(\Cs{D}) \\
A \mapsto F(A)
\end{cases} ~~
F_{A,B}:
\begin{cases}
\Mor{\Cs{C}}{A}{B} & \to \Mor{\Cs{D}}{F(A)}{F(B)} \\
(f: A \to B) & \mapsto (F(f):F(A) \to F(B))
\end{cases}
\]
such that for any composable morphisms $f$ and $g$ and any object $A \in \Cs{C}$: $$F(fg) = F(f)F(g),$$ $$F(\mathds{1}_A) = \mathds{1}_{F(A)}.$$
\end{defin}
\begin{rem}
Covariant functors preserve the direction of morphisms, and they can be thought of as mappings that respect the category structure (i.e. preserving the identities and compositions of morphisms). Covariant functors are simply referred to as functors.
\end{rem}
\begin{defin}
A \textit{diagram} in a category $\Cs{C}$ is a functor $D:\Cs{D} \to \Cs{C}$, where $\Cs{D}$ is a small category. Objects are represented by vertices, and morphisms are represented by arrows.
\end{defin}
\begin{exam}
Recalling that posets may be viewed as categories, consider the poset
\[
I = \left\{
\begin{tikzcd}
& \bullet \ar[d] \\
\bullet \ar[r] & \bullet
\end{tikzcd}
\right\}
\]
Let $\Cs{C}$ be a category and $F:I \to \Cs{C}$ be a functor. Then $F$ is a diagram in $\Cs{C}$ which may be represented as
\begin{center}
\begin{tikzcd}
& A \ar[d,"f"] \\
B \ar[r,"g"'] & C,
\end{tikzcd}
\end{center}
where $A,B,C \in \Cs{C}$ and $f,g$ are morphisms.
\end{exam}
\begin{exam}
A very important and commonly encountered type of functor is the covariant representable functor. Let $\Cs{C}$ be a category and $f:A \to B$ be a morphism in $\Cs{C}.$ Define
\[
\Mor{\Cs{C}}{X}{-}:
\begin{cases}
\Cs{C} \to \Cb{Set} \\
A \mapsto (X,A) \\
(A \stackrel{f}{\to} B) \mapsto ((X,A) \stackrel{f_*}{\to} (X,B))
\end{cases}
\]
where the map $f_*$ is defined as
\[
\Mor{\Cs{C}}{X}{f} = f_*:
\begin{cases}
(X,A) \to (X,B) \\
\alpha \mapsto f\alpha
\end{cases}
\]
Such defined correspondence is a covariant functor, and it is usually called a covariant representable functor. If it is clear what the source category is, then we write this functor as $(X,-)$. Notice that the definition of $f_*$ as $\alpha \mapsto f\alpha$ is forced by the sources and targets of $f$ and $\alpha$.
\end{exam}
\begin{defin}
Let $\Cs{C}$ and $\Cs{D}$ be categories. Then a \textit{contravariant functor} $F:\Cs{C}\to\Cs{D}$ is a pair of correspondences between classes of objects and morphisms in $\Cs{C}$ and $\Cs{D}$ defined by
\[
F:
\begin{cases}
\mathrm{Obj}(\Cs{C}) \to \mathrm{Obj}(\Cs{D}) \\
A \mapsto F(A)
\end{cases} ~~
F_{A,B}:
\begin{cases}
\Mor{\Cs{C}}{A}{B} & \to \Mor{\Cs{D}}{F(B)}{F(A)} \\
(f:A \to B) & \mapsto (F(f):F(B) \to F(A))
\end{cases}
\]
such that for any composable morphisms $f$ and $g$ and any object $A \in \Cs{C}$: $$F(fg) = F(g)F(f),$$ $$F(\mathds{1}_A) = \mathds{1}_{F(A)}.$$
\end{defin}
\begin{rem}
A contravariant functor $F:\Cs{C} \to \Cs{D}$ can be considered as a covariant functor $F':\Cs{C}^{op} \to \Cs{D}$. Therefore, contravariant functors are covariant functors acting either to or from an opposite category. Recalling the remark about covariant functors preserving category structure, the involvement of an opposite category produces a sort of reversing of arrows.
Within a diagram of a category, this translates to all arrows being turned backwards while the objects are left stationary. Again, the term functor usually refers to covariant functor unless it is specified that it is a contravariant functor.
\end{rem}
\begin{defin}
Given functors $F:\Cs{C} \to \Cs{D}$ and $G:\Cs{D} \to \Cs{E}$, we define the \textit{composition of functors} $G \circ F$ as $$(G \circ F)(X) = G(F(X)),$$ $$(G \circ F)(f) = G(F(f)),$$ where $f$ is a morphism in $\Cs{C}$ and $X \in \Cs{C}$. Note that these functors may be of opposite variance.
\end{defin}
\begin{exam}
Let's consider the contravariant representable functor. Let $\Cs{C}$ be a category and $X$ be an object in $\Cs{C}.$ Define
\[
\Mor{\Cs{C}}{-}{X}:
\begin{cases}
\Cs{C} \to \Cb{Set} \\
A \mapsto (A,X) \\
(A \stackrel{f}{\to} B) \mapsto ((B,X) \stackrel{f^*}{\to} (A,X))
\end{cases}
\]
where $f^*$ is defined as,
\[
\Mor{\Cs{C}}{f}{X} = f^*:
\begin{cases}
\Mor{\Cs{C}}{B}{X} \to \Mor{\Cs{C}}{A}{X} \\
\beta \mapsto \beta f
\end{cases}
\]
It is easy to see that it is a contravariant functor. Again if it is clear what the source category is, then we write this functor as $(-,X)$. Although less straightforward, the action of this functor on morphisms is still the only way that makes sense for correct composition.
\end{exam}
\begin{exam}
Here we observe two simple but important functors with no distinction between contravariant and covariant. Let $\Cs{C}$ and $\Cs{D}$ be categories, $A \in \Cs{C}$ be any object in $\Cs{C}$, $f$ be a morphism in $\Cs{C}$, and $D \in \Cs{D}$ be some object in $\Cs{D}$. The identity functor $\mathds{1}_\Cs{C}$ on $\Cs{C}$ defined as $$\mathds{1}_\Cs{C}(A) = A, ~~ \mathds{1}_\Cs{C}(f) = f,$$ i.e. it is the functor which fixes all objects and morphisms.
The constant functor with the value $D$ defined as
\[
D:
\begin{cases}
\Cs{C} & \to \Cs{D} \\
A & \mapsto D \\
f & \mapsto \mathds{1}_D
\end{cases}
\]
is the functor which sends all objects to $D$ and all morphisms to the identity morphism of $D$.
\end{exam}
\subsection{Natural Transformations and Yoneda Lemma}
\begin{defin}
Given categories $\Cs{C}$ and $\Cs{D}$ and covariant functors $F,G:\Cs{C} \to \Cs{D}$, a \textit{natural transformation} $\eta:F \to G$ is a one-parameter family of morphisms in $\Cs{D}$ $$\eta = \{\eta_X:F(X) \to G(X)\}_{X \in \Cs{C}},$$ such that the following diagram commutes for every morphism $f \in \Mor{\Cs{C}}{X}{Y}$,
\begin{center}
\begin{tikzcd}
F(X) \ar[d,"F(f)"'] \ar[r,"\eta_X"] & G(X) \ar[d,"G(f)"] \\
F(Y) \ar[r,"\eta_Y"'] & G(Y).
\end{tikzcd}
\end{center}
\end{defin}
A natural transformation is a means of relating functors. Natural transformations also exist for contravariant functors, and the definition is the same except the commuting diagram has its vertical arrows pointing up. For now, we describe when a natural transformation provides an equivalence.
\begin{defin} \label{nequ}
Let $F,G:\Cs{C} \to \Cs{D}$ be functors. A natural transformation $\eta:F \to G$ is called a \textit{natural isomorphism} if $\eta_X$ is an isomorphism in $\Cs{D}$ for each $X \in \Cs{C}$. In this case the functors $F$ and $G$ are said to be isomorphic, and we write $F \approx G$.
\end{defin}
\begin{defin}
Given natural transformations $\eta:F \to G$ and $\tau:G \to H$ of the functors $F,G,H:\Cs{C} \to \Cs{D}$, we may define the \textit{composition of natural transformations} $$\tau\eta = \{\tau_X \eta_X:F(X) \to H(X)\}_{X \in \Cs{C}}$$ as the composition of the component morphisms $\tau_X$ and $\eta_X$ for any object $X$.
\end{defin}
\begin{defin}
Given categories $\Cs{C}$ and $\Cs{D}$, a functor $F:\Cs{C}\to\Cs{D}$ is said to be an \textit{equivalence} if there exists a functor $G:\Cs{D}\to\Cs{C}$ such that $$FG \approx \mathds{1}_\Cs{D}, ~~ GF \approx \mathds{1}_\Cs{C}.$$ In this case the categories $\Cs{C}$ and $\Cs{D}$ are said to be equivalent, and we write $\Cs{C} \approx \Cs{D}$. It is clear that $G$ would then also be an equivalence of the categories $\Cs{C}$ and $\Cs{D}$.
\end{defin}
Given a functor $F$, it may not be easy to demonstrate a compatible functor $G$ to establish that $F$ is an equivalence. The following theorem provides a means for checking whether a functor is an equivalence based on its own properties.
\begin{theo}
A functor $F:\Cs{C} \to \Cs{D}$ is an equivalence of categories if and only if it is:
\begin{itemize}
\item faithful (injective on morphisms): $F_{X,Y}:\Mor{\Cs{C}}{X}{Y} \to \Mor{\Cs{D}}{F(X)}{F(Y)}$ is an injective function;
\item full (surjective on morphisms): $F_{X,Y}:\Mor{\Cs{C}}{X}{Y} \to \Mor{\Cs{D}}{F(X)}{F(Y)}$ is a surjective function;
\item essentially surjective (surjective on objects): there exists $C \in \Cs{C}$ such that $F(C) = D$ for each $D \in \Cs{D}$.
\end{itemize}
\end{theo}
For the next lemma, we must introduce a piece of notation. For functors $F,G:\Cs{C}\to\Cs{D}$, let $(F,G)$ be the collection of all natural transformations between the functors $F$ and $G$.
\begin{lem}[Yoneda Lemma]
Let $\Cs{C}$ be a category and $C \in \Cs{C}$ be any object. If $F:\Cs{C} \to \Cb{Set}$ is a functor, then there is a bijection
\[
y:
\begin{cases}
((C,-),F) \to F(C) \\
\eta \mapsto \eta_C(\mathds{1}_C).
\end{cases}
\]
\end{lem}
\begin{lem}[Contravariant Yoneda Lemma]
Let $\Cs{C}$ be a category and $C \in \Cs{C}$ be any object in $C$. If $F:\Cs{C} \to \Cb{Set}$ is a contravariant functor, then there is a bijection
\[
y:
\begin{cases}
((-,C),F) \to F(C) \\
\eta \mapsto \eta_C(\mathds{1}_C).
\end{cases}
\]
\end{lem}
With these two versions of the Yoneda lemma, we obtain the following corollary.
\begin{cor} \label{bij}
For any $A,B\in\Cs{C}$, there are bijections $$((A,-),(B,-)) \leftrightarrow (B,A),$$ $$((-,A),(-,B)) \leftrightarrow (A,B).$$
\end{cor}
To obtain the above bijections, we apply the Yoneda Lemma with the functor $F$ being $(B,-)$ and the contravariant Yoneda Lemma with the functor $F$ being $(-,B).$
\begin{defin}
Let $\Cs{C}$ be a small category and $\Cs{D}$ be any category. Then the \textit{functor category} $\Cs{D}^\Cs{C},$ also denoted $(\Cs{C},\Cs{D}),$ is the category whose objects are functors and morphisms are natural transformations, and the composition is defined as a composition of natural transformations:
\begin{itemize}
\item $\mathrm{Obj}(\Cs{D}^\Cs{C}) = \{F:\Cs{C} \to \Cs{D}\ |\ F \text{ is a functor}\};$
\item $\Mor{\Cs{D}^\Cs{C}}{F}{G} = (F,G)$.
\end{itemize}
\end{defin}
Now with the necessary terminology in place, we may proceed to the following very important corollary. In essence, it asserts that any small category is a subcategory of $\Cb{Set}$ for which all morphism sets are left completely intact. Said differently, small categories are full subcategories of $\Cb{Set}$.
\begin{cor}[Yoneda Embedding]
If $\Cs{C}$ is a small category, then the functor
\[
Y:
\begin{cases}
\Cs{C}^{op} \to \Cb{Set}^\Cs{C} \\
C \mapsto (C,-)
\end{cases}
\]
is injective on objects, full, and faithful. This functor is called the Yoneda Embedding.
\end{cor}
\begin{proof}
To show that $Y$ is injective on objects, consider the fact that for objects $A,B \in \Cs{C}$ such that $A \neq B$, $(A,-)$ and $(B,-)$ will have disjoint images by the definition of a category. As such, $(A,-) \neq (B,-)$ so that $Y$ is injective on objects. Now recall from \thref{bij} that there is a bijection between $\Mor{\Cs{C}}{A}{B} = \Mor{\Cs{C}^{op}}{B}{A}$ and $((B,-),(A,-))$. Because natural transformations are morphisms in $\Cb{Set}^\Cs{C}$, this shows that $Y$ is full and faithful.
\end{proof}
\begin{cor}
Let $A,B \in \Cs{C}$ be objects of $\Cs{C}$. If for every $C \in \Cs{C}$ there is a bijection $$(A,C) \leftrightarrow (B,C),$$ then $A \cong B$ in $\Cs{C}$. The converse also holds.
\end{cor}
Because bijections are isomorphisms in $\Cb{Set}$, this corollary amounts to saying that if $Y(A) \approx Y(B)$, then $A \cong B$. Seeing as $\approx$ indicates isomorphism in $\Cb{Set}^\Cs{C}$ and that $Y$ is injective on objects, this corollary closely resembles the condition for injectivity of functions.
The reader may have noted the many contexts where the notation $(-,-)$ appears. For example replacing the dashes with objects gives a collection of morphisms, categories a collection of functors, functors (with identical source and target) a collection of natural transformations, and functors (with identical target) a collection of comma categories. While it may seem the notation is overused, the usage is quite similar across the cases because they all take structures and return a collection of things which relate those structures. Care need only be taken in paying attention to what the input structures are.
\subsection{Limits and Colimits}
\begin{defin}
Let $\Cs{C},\Cs{D},\Cs{E}$ be categories and $F:\Cs{C} \to \Cs{E},G:\Cs{D} \to \Cs{E}$ be functors. Then a \textit{comma category} $F/G $ is a category whose:
\begin{comment}
The notation $(F,G),$ while antiquated, also makes sense along with its other uses. There is a sentence later on which mentions this explicitly that Alex really likes but is unsure what to do with.
\end{comment}
\begin{itemize}
\item objects are triples $(C,D,f)$ where $C \in \Cs{C}$, $D \in \Cs{D}$, and $f:F(C) \to G(D)$ is a morphism in $\Cs{E}$;
\item morphisms between $(C_1,D_1,f_1)$ and $(C_2,D_2,f_2)$ are pairs $(\alpha,\beta)$, where
\[
\alpha:F(C_1) \to F(C_2), \quad \beta:G(D_1) \to G(D_2)
\]
are morphisms in $\Cs{E}$ such that $f_2\alpha = \beta f_1$. It can be expressed diagrammatically as
\begin{center}
\begin{tikzcd}
F(C_1) \ar[d,"f_1"'] \ar[r,"\alpha"] & F(C_2) \ar[d,"f_2"] \\
G(D_1) \ar[r,"\beta"'] & G(D_2).
\end{tikzcd}
\end{center}
\end{itemize}
\end{defin}
\begin{defin}
In a category $\Cs{C}$, consider a family of objects and morphisms between them $$(\{C_i\},\{f_{ij}:C_i \to C_j\})_{i,j \in I}.$$ A \textit{cone} over this family is a pair $(C,\{f_i:C \to C_i\}_{i \in I})$ such that the following diagram commutes for any $i,j \in I$:
\begin{center}
\begin{tikzcd}
& \ar[ld,"f_i"'] C \ar[rd,"f_j"] & \\
C_i \ar[rr,"f_{ij}"'] && C_j.
\end{tikzcd}
\end{center}
Similarly, a \textit{cocone} below this family is a pair $(C,\{f_i:C_i \to C\}_{i \in I})$ such that the following diagram commutes for any $i,j \in I$:
\begin{center}
\begin{tikzcd}
C_i \ar[rd,"f_i"'] \ar[rr,"f_{ij}"] && \ar[ld,"f_j"] C_j \\
& C. &
\end{tikzcd}
\end{center}
The maps $f_i$ associated with the cone and cocone are called \textit{structural maps}.
\end{defin}
We can define a preorder between cones that will be useful in stating the primary structure of this section.
\begin{defin}
A cone $(C,\{f_i\}_{i \in I})$ is \textit{dominated} by a cone $(C',\{f'_i\}_{i \in I})$ if there exists a \textit{domination map} $g:C' \to C$ such that $$f'_i = f_ig$$ for all $i \in I$. Dually, a cocone $(C,\{f_i\}_{i \in I})$ is \textit{dominated} by a cocone $(C',\{f_i'\}_{i \in I})$ if there exists a \textit{domination map} $g':C \to C'$ such that $$f'_i = g'f_i$$ for all $i \in I$.
\end{defin}
Below is a remark with diagrams to help demonstrate cones and domination visually. The diagrams will illustrate domination for cones, and the corresponding diagrams for cocones may be obtained by reversing all structural maps.
\begin{rem}
Let the family of objects and morphisms make up the base of a geometric cone. Then the cone (resp. cocone) is the object at the tip of the geometric cone, resulting in a diagram with the following form:
\begin{center}
\begin{tikzcd}
\ar[d] C \ar[rdd] \ar[rrd] & & \\
\bullet \ar[rr] & & \ar[ld] \bullet \\
& \ar[lu] \bullet &
\end{tikzcd}
\end{center}
When a cone $C'$ dominates a cone $C$, a new tip is introduced with an arrow $g$ pointing from $C'$ to $C$ (resp. from $C$ to $C'$), and $C',$ and the new tip becomes the highest point. Now we form a new cone (resp. cocone) from the new base and tip by letting the descending arrows from tip to base (resp. base to tip) be the hypotenuses of the right triangle with legs $g$ and $f_i$, resulting in a diagram with the following form:
\begin{center}
\begin{tikzcd}
& & \ar[lldd,dashed] \ar[lld,"g"'] \ar[lddd,dashed] C' \ar[dd,dashed] \\
\ar[d] C \ar[rdd] \ar[rrd] & & \\
\bullet \ar[rr] & & \ar[ld] \bullet \\
& \ar[lu] \bullet &
\end{tikzcd}
\end{center}
\end{rem}
\begin{defin}
A \textit{limit} of a functor $F \in \Cs{C}^I$, where $(I,\leq)$ is a poset category (refer to \ref{pos}), is a cone $(C,\{f_i\}_{i \in I})$ over the family $(\{F(i)\},\{F(i) \to F(j)\})_{i \leq j \in I}$ which is dominated by every other cone, such that the domination maps are unique. We write $C = \varprojlim F$. So if $(C',\{f'_i\}_{i \in I})$ is a cone over the family, then
\begin{center}
\begin{tikzcd}
& \ar[ldd,bend right,"f'_i"'] \ar[d,dashed,"\exists!"] C' \ar[rdd,bend left,"f'_j"] & \\
& \ar[ld,"f_i"'] C \ar[rd,"f_j"] & \\
F(i) \ar[rr,"f_{ij}"] & & F(j)
\end{tikzcd}
\end{center}
commutes. If the domination maps are not unique, it is called a \textit{weak limit}. Limits are also called projective limits or inverse limits.
Similarly, a \textit{colimit} of a functor $F \in \Cs{C}^I$, where $(I,\leq)$ is a poset, is a cocone $(C,\{f_i\}_{i \in I})$ below the family $(\{F(i)\},\{F(i) \to F(j)\})_{i,j \in I}$ which is dominated by every other cocone, such that the domination maps are unique. We write $C = \varinjlim F$. So if $(C',\{f'_i\}_{i \in I})$ is a cocone under the family, then
\begin{center}
\begin{tikzcd}
F(i) \ar[rd,"f_i"'] \ar[rdd,bend right,"f'_i"'] \ar[rr,"f_{ij}"'] & & \ar[ld,"f_j"] \ar[ldd,bend left,"f'_j"] F(j) \\
& \ar[d,dashed,"\exists!"] C & \\
& C' &
\end{tikzcd}
\end{center}
commutes. If the domination maps are not unique, it is called a \textit{weak colimit}. Colimits are also called injective limits or direct limits.
\end{defin}
\begin{defin}
A category $\Cs{C}$ is said to be (\textit{co})\textit{complete} if it admits all small (co)limits (i.e. there exists a (co)limit for any diagram $F$ in $\Cs{C}$). Similarly, $\Cs{C}$ is said to be \textit{finitely (co)complete} if it admits all finite (co)limits. That is, for any finite poset $\Cs{D}$ and functor $F:\Cs{D} \to \Cs{C}$, there exists a (co)limit of $F$.
\end{defin}
\subsection{Adjunction of Functors}
Here we observe how a relation between functors resulting from simple changes of perspective produce a litany of useful and important information about the functors. One such motivating example originates from a function of 2-variables
\[
f:
\left\{
\begin{array}{rcl}
X \times Y & \longrightarrow & Z \\
(x,y) & \longmapsto & z,
\end{array}
\right.
\]
which we may write alternatively as $f(x,y) = z.$ It is common in multivariable calculus courses to parameterize such multivariable functions, each time increasing the number of functions but simultaneously decreasing the number of variables of these functions. In particular, we may instead write $f_x(y) = f(x,y)$ to define a family of maps, one for each value $x.$ Formally, this function may be written
\[
f_x:
\left\{
\begin{array}{rcl}
Y & \longrightarrow & Z \\
y & \longmapsto & z
\end{array}
\right.
.
\]
What we find is that every 2-variable function corresponds to a family of 1-variable functions, expressed symbolically as
\[
(X \times Y,Z) \longleftrightarrow (X,(Y,Z)).
\]
By defining mappings $L = - \times Y$ and $R = (Y,-),$ we may rewrite the above bijection as
\[
(L(X),Z) \longleftrightarrow (X,R(Z)),
\]
which no longer explicitly mentions $Y.$
There is yet another motivating example which instead comes from the scalar product of vectors. Let $u,v \in \mathbb{R}^n$ be any vectors and $O \in (\mathbb{R}^n,\mathbb{R}^n)$ be some linear operator (equivalently an $n \times n$ matrix). The scalar product of $x$ and $y$ is given by $x\cdot y = x^Ty,$ which is nothing more than the familiar dot product from linear algebra. Introducing the linear transformation, we notice that
\[
x \cdot Oy = x^TOy = x^T(O^T)^Ty = (O^Tx)^Ty = O^Tx \cdot y.
\]
Intuitively, the scalar product measures the ``$x$ character of $y,$'' which most often is geometrically interpreted as how parallel $y$ is to $x.$ By operating on $y$ before taking the scalar product, we actually measure how parallel the image of $y$ (under $O$) is to $x.$ The important change of perspective demonstrated by the above chain of equalities is that this measurement is equivalent to first transforming $x$ in a ``reverse'' fashion and then measuring it against $y,$ where reverse is used very loosely to designate how transpose matrices act geometrically. If we change the notation to, $(x,y) = x \cdot y,$ then the equality may be written suggestively as
\[
(x,Oy) = (O^Tx,y)
\]
so the operator may now go without being explicitly mentioned. In the special case where the operator is orthogonal (geometrically representing a rotation), operating on $y$ and then measuring gives the same result as operating on $x$ in reverse and then measuring. Alternatively, it is like thinking of a rotation not as a transformation of vectors but instead as a reverse transformation of the space within which the vectors sit while leaving the vectors to fall in their new spots.
\begin{defin}
Let $\Cs{C},\Cs{D}$ be categories and $L:\Cs{C} \to \Cs{D},R:\Cs{D} \to \Cs{C}$ be functors. If for any $C \in \Cs{C}, D \in \Cs{D}$ there is a bijection
\[
\varphi_{C,D}:\Mor{\Cs{D}}{L(C)}{D} \to \Mor{\Cs{C}}{C}{R(D)}
\]
which is natural in both $C$ and $D,$ then we say that the functors $L$ and $R$ are \textit{adjoint}, $(L,R)$ is an \textit{adjoint pair}, $L$ is \textit{left adjoint} to $R,$ or $R$ is \textit{right adjoint} to $L;$ we write this symbolically as $L \dashv R$ or as
\[
\begin{tikzcd}
\Cs{C} \ar[r, "L"{name = L}, bend left = 25] &\Cs{D}. \ar[l, "R"{name = R}, bend left = 25] \ar[phantom, from = L, to = R, "\dashv" rotate = -90]
\end{tikzcd}
\]
\end{defin}
\begin{rem}
What is meant by saying that $\varphi_{C,D}$ is natural in $C$ and $D$ is that for all $C_1,C_2 \in \Cs{C},$ $D_1,D_2 \in \Cs{D},$ $f \in \Mor{\Cs{C}}{C_1}{C_2},$ and $g \in \Mor{\Cs{D}}{D_1}{D_2},$ the diagram
\[
\begin{tikzcd}
\ar[d,"\Mor{\Cs{D}}{Lf}{g} = g \circ \Box \circ Lf"'] \Mor{\Cs{D}}{L(C_1)}{D_1} \ar[r,"\varphi_{C_1,D_1}"] & \ar[d, "\Mor{\Cs{C}}{f}{Rg} = Rg \circ \Box \circ f"] \Mor{\Cs{C}}{C_1}{R(D_1)} \\
\Mor{\Cs{D}}{L(C_2)}{D_2} \ar[r,"\varphi_{C_2,D_2}"'] & \Mor{\Cs{C}}{C_2}{R(D_2)}
\end{tikzcd}
\]
commutes. The word ``natural'' does in fact relate to natural transformations, and it is useful to stop and think about explicitly how these tie together.
\end{rem}
Intuitively, given a pair of objects in separate categories, a pair of adjoint functors provides rules for finding another pair of category-separated objects such that there is a one-to-one correspondence between the morphisms for objects in the same category. As we will see more in later discussions, bijections between sets of morphisms often indicate some type of universal arrow mechanism; this holds true here as well, which is made precise in the following theorems.
\begin{theo}
Let $L:\Cs{C} \to \Cs{D}$ be a functor. Then $L$ has a right adjoint if and only if there exists a functor $R:\Cs{D} \to \Cs{C}$ such that $L/D$ has terminal object $(R(D),D,L(R(D)) \stackrel{\varepsilon_D}{\longrightarrow} D)$ for any $D \in \Cs{D}.$ Under such conditions, $L \dashv R.$
\end{theo}
\begin{theo}
Let $R:\Cs{D} \to \Cs{C}$ be a functor. Then $R$ has a left adjoint if and only if there exists a functor $L:\Cs{C} \to \Cs{D}$ such that $C/R$ has initial object $(C,L(C),C \stackrel{\mu_C}{\longrightarrow} R(L(C)))$ for any $C \in \Cs{C}.$ Under such conditions, $L \dashv R.$
\end{theo}
The universal arrows in these theorems arise from the unique arrows originating from initial objects (resp. ending at terminal objects). These universal arrows, denoted by $\varepsilon_D$ and $\mu_C$ in the above theorems, are in fact component maps of natural transformations which form the foundation of the next equivalent formulation of adjunction.
\begin{defin}
Let $F:\Cs{C} \to \Cs{D}$ and $G:\Cs{D} \to \Cs{C}$ be a pair of functors. The natural transformations (if they exist)
\[
\mu:\mathds{1}_{\Cs{C}} \to GF, \quad \varepsilon:FG \to \mathds{1}_{\Cs{D}}
\]
are called \textit{unit} and \textit{counit}, respectively.
\end{defin}
\begin{theo}
Let $F:\Cs{C} \to \Cs{D}$ and $G:\Cs{D} \to \Cs{C}$ be a pair of functors. There exists a unit-counit pair
\[
\mu:\mathds{1}_{\Cs{C}} \to GF, \quad \varepsilon:FG \to \mathds{1}_{\Cs{D}}
\]
satisfying the ``counit-unit equations''
\begin{itemize}
\item $\mathds{1}_L = \varepsilon L \circ L\mu,$
\item $\mathds{1}_R = R\varepsilon \circ \mu R,$
\end{itemize}
if and only if $L \dashv R.$
\end{theo}
\begin{cor}
Adjoints are defined up to natural isomorphism, and the composition of adjunctions is an adjunction.
\end{cor}
\begin{exam} \label{adj}
While the reader may not be familiar with all of the examples to be stated here, there will be brief descriptions of functors which are not discussed further in this document. Some adjunctions are discussed at length elsewhere and so will be omitted from this list. Regardless, some of these examples will help to demonstrate just how ubiquitous adjunctions are.
\begin{itemize}
\item $(-)_1:\Cb{Rng} \dashv \Cb{Ring}:F,$ where $F$ forgets about the role of the identity and $(-)_1$ adjoins the identity. If $R$ is a rng, then $(R)_1 = R \times \Z$ with the operation $(r,n)(s,m) = (rs+rm+ns,nm)$ and identity element $(0,1);$
\item $K_0:\Cb{CMon} \dashv \Cb{Ab}:F,$ where $K_0$ adjoins inverses through a particular quotient construction (study Grothendieck group for reference) and $F$ forgets about the role of inverses;
\item Group ring: $\Cb{Grp} \dashv \Cb{Ring}:$ Group of units, where a group is sent to its group ring over $\Z$ and a ring is sent to its group of units;
\item $Frac(-):$ Domains $\dashv$ Fields: $F,$ where $Frac(R)$ takes the domain $R$ to its field of fractions and $F$ forgets the role of multiplicative inverses;
\item $\Sigma:\Cb{Top}_* \dashv \Cb{Top}_*:\Omega,$ where $\Sigma$ and $\Omega$ produce the reduced suspension and loop space of a pointed topological space, respectively;
\item $\beta:\Cb{Top}_2 \dashv \Cb{KHaus}:\iota,$ where $\beta$ is the Stone-\v{C}ech compactification, i.e. smallest compact space in which we may embed the given Hausdorff space, and $\iota$ is the inclusion of compact Hausdorff spaces into Hausdorff spaces;
\item If $F:\Cs{C} \stackrel{\approx}{\longrightarrow} \Cs{D}:G$ is an equivalence of categories, then $F \dashv G,$ $F \vdash G,$ and $\varepsilon,\mu$ are natural isomorphisms.
\end{itemize}
\end{exam}
Lastly, there is a very desirable property of adjoint pairs.
\begin{defin}
A functor $F:\Cs{C} \to \Cs{D}$ is called \textit{continuous} if
\[
F(\varprojlim_iC_i) \cong \varprojlim_iF(C_i)
\]
and cocontinuous if
\[
F(\varinjlim_iC_i) \cong \varinjlim_iF(C_i).
\]
\end{defin}
\begin{theo}
If $L \dashv R$ is an adjoint pair of functors, then $L$ is cocontinuous and $R$ is continuous.
\end{theo}
\newpage
\section{Rings and Modules}
\subsection{Fundamentals of Rings and Modules}
Although rings are common knowledge to many mathematicians, we start by setting straight what conventions will be followed throughout this document.
\begin{defin}
A nonempty set $R$ with two binary operations
\[
+:
\left\{
\begin{array}{rcl}
R \times R & \longrightarrow & R \\
(a,b) & \longmapsto & a+b
\end{array}
\right.
, \qquad \qquad
\cdot:
\left\{
\begin{array}{rcl}
R \times R & \longrightarrow & R \\
(a,b) & \longmapsto & ab
\end{array}
\right.
\]
subject to the conditions
\begin{enumerate}[label = (\roman*)]
\item $(R,+)$ is an abelian group (with identity element 0);
\item there exists an element $1 \in R$ such that $1a = a1 = a$ for any $a \in R$ (existence of multiplicative identity);
\item $(ab)c = a(bc)$ for any $a,b,c \in R$ (associativity of multiplication);
\item $a(b+c) = ab+ac$ for any $a,b,c \in R$ (left distributivity);
\item $(a+b)c = ac+bc$ for any $a,b,c \in R$ (right distributivity);
\end{enumerate}
is called a unital \textit{ring}. The operations are referred to as \textit{addition} ($+$) and multiplication ($\cdot$). Imposing only conditions (ii) and (iii) on $\cdot$ makes $(R,\cdot)$ a \textit{monoid}, i.e. a group without the condition of having inverses for every element.
\end{defin}
\begin{defin}
By adding (or removing) further axioms, we may define other related structures. Let $R$ be a ring.
\begin{itemize}
\item If there is not a multiplicative identity, then $R$ is called a \textit{rng}.
\item If $ab = ba$ for any $a,b \in R,$ then $R$ is called a \textit{commutative ring}.
\item If there are non-zero elements $a,b \in R$ such that $ab = 0,$ then we call $a$ a \textit{left zero-divisor} and $b$ a \textit{right zero-divisor}. If $R$ has no left and right zero-divisors, we call $R$ a \textit{domain}.
\item If $R$ is a commutative domain, it is called an \textit{integral domain}.
\item If for every $a \in R-\{0\}$ there exists an element $a^{-1} \in R$ such that $aa^{-1} = a^{-1}a = 1,$ then $R$ is called a \textit{division ring}. Clearly, $R$ must be a unital ring first in order for 1 to exist. We call $a^{-1}$ the \textit{multiplicative inverse} of $a.$ Informally, a division ring is a ring for which every non-zero element has a multiplicative inverse.
\item If $R$ is a commutative division ring, then it is a \textit{field}.
\end{itemize}
\end{defin}
The multiplicative identity is usually denoted by 1, and it is often called the \textit{unity} for a ring. As such, rngs are then just rings without unity. For those uncomfortable with the idea of zero-divisors, consider that non-zero matrices can multiply to the zero matrix and that a zero-divisor of this kind is dependent on which side multiplication occurs. Also, the reader may have noticed that there is no mention of zero-divisors with respect to fields, which leads us to this simple and illustrative result.
\begin{lem}
If $R$ is a field, then $R$ is an integral domain.
\end{lem}
\begin{proof}
Suppose that $R$ is a field and that there are non-zero elements $a,b \in R$ such that $ab = 0.$ Then we have that
\[
ab = 0 \Longrightarrow a^{-1}ab = a*0 \Longrightarrow 1*b = 0 \Longrightarrow b = 0,
\]
but this contradicts our assumption that $b$ is non-zero.
\end{proof}
There are many more properties which may be obtained from the axioms for any of these structures, but these may be found easily from many other sources. We will now provide some other basic but important definitions.
\begin{defin}
Let $R$ be a ring. Then a subset $I \subseteq R$ is a \textit{left ideal} if it satisfies the following axioms:
\begin{enumerate}[label = (\roman*)]
\item if $i_1,i_2 \in I,$ then $i_1-i_2 \in I.$
\item for any $r \in R$ and $i \in I,$ $ri \in I$ ($RI \subseteq I$).
\end{enumerate}
\end{defin}
\begin{rem}
Axiom (i) establishes that $I$ is a subgroup of $(R,+)$ while axiom (ii) establishes that $I$ is closed under multiplication by ring elements.
\end{rem}
There are also right ideals and two-sided ideals. In commutative rings, all ideals are clearly two-sided, so mention of sidedness disappears. We now mention how to generate ideals because it will later mimic how to generate modules. Although this definition is stated for commutative rings, the definitions are similar for non-commutative rings but with the inclusion of handedness.
\begin{defin}
Let $S \subseteq R$ be a subset of a commutative ring, then the \textit{ideal generated by $S$}, denoted by $\langle S\rangle$, is defined as
\[
R\langle S\rangle = \langle S\rangle = \left\{\left.\sum_{i = 1}^n r_i s_i\right|r_i \in R,s_i \in S, n \geq 1\right\}.
\]
If $S$ is finite, then $\langle S\rangle$ is said to be a \textit{finitely generated ideal}. If $S = \{s\}$, then $\langle S\rangle = sR$ is a \textit{principal ideal} generated by $s$.
\end{defin}
\begin{rem}
With groups, there were distinguished subgroups (normal subgroups) for which the set of cosets of the subgroup took on a group structure. With all subrings (subsets which are rings sharing the same identity as the ambient ring), there is a group structure on these cosets, but not necessarily a ring structure. Ideals are of particular importance because the cosets of ideals have a well-defined ring structure.
\end{rem}
\begin{defin}
Let $R$ and $S$ be rings. A function $\varphi:R \to S$ satisfying
\begin{itemize}
\item $\varphi(r_1+r_2) = \varphi(r_1)+\varphi(r_2),$
\item $\varphi(r_1r_2) = \varphi(r_1)\varphi(r_2),$
\item $\varphi(1) = 1,$
\end{itemize}
is called a \textit{ring homomorphism}. A bijective ring homomorphism is called a \textit{ring isomorphism}.
\end{defin}
It may seem odd that it is included as a separate axiom that the multiplicative identity is preserved, but it is not implied by the other axioms due to the lack of multiplicative inverses. In addition, this is a very important condition to impose as it entirely changes the category of study if it is not present, which will be seen in later sections. Now we observe some distinguished subsets constructed from ring homomorphisms.
\begin{defin}
Let $\varphi:R \to S$ be a ring homomorphism. The \textit{kernel} of $\varphi,$ denoted by $\Ker(\varphi)$ is defined as
\[
\Ker(\varphi) = \{r \in R\ |\ \varphi(r) = 0 \in S\}.
\]
Furthermore, the \textit{image} of $\varphi,$ denoted by $\Im(\varphi)$ is defined as
\[
\Im(\varphi) = \{\varphi(r) \in S\ |\ r \in R\}.
\]
\end{defin}
\begin{defin}
Let $S$ be a ring. A subset $T \subseteq S$ is called a \textit{subring} if there exists a ring $R$ and ring homomorphism $\varphi:R \to S$ such that $T = \Im(\varphi).$
\end{defin}
\begin{rem}
Kernels of ring homomorphisms are two-sided ideals. Furthermore, images of ideals are not necessarily ideals unless the ring homomorphism is surjective.
\end{rem}
\begin{defin}
Let $I \subseteq R$ be a left ideal of the ring $R.$ Then the \textit{quotient ring} $R/I$ is defined as the set of cosets
\[
R/I = \{r+I\ |\ r \in R\}
\]
with addition and multiplication given by
\[
(r+I)+(r'+I) = (r+r')+I, \qquad (r+I)(r'+I) = rr'+I,
\]
for any $r,r' \in R$ and $r \in R.$
\end{defin}
Vector spaces are familiar objects of study from linear algebra courses, and they have a litany of useful properties and applications. To recall, a vector space is an abelian group with an operation defined involving a field, which we call scalar multiplication. If we relax these conditions so that scalars instead originate from a ring, we obtain an important structure which is less well-behaved than vector spaces in that there is no longer a strict notion of basis.
\begin{defin}
Let $R$ be a ring. A \textit{left $R$-module} is an abelian group $(M,+)$ with an operation
\[
\cdot:
\left\{
\begin{array}{rcl}
R \times M & \to & M \\
(r,m) & \mapsto & rm
\end{array}
\right.
\]
subject to the conditions
\begin{itemize}
\item $1m = m$ for all $m \in M;$
\item $(r_1r_2)m = r_1(r_2m)$ for all $r_1,r_2 \in R$ and $m \in M;$
\item $r(m_1+m_2) = rm_1+rm_2$ for all $r \in R$ and $m_1,m_2 \in M;$
\item $(r_1+r_2)m = r_1m+r_2m$ for all $r_1,r_2 \in R$ and $m \in M.$
\end{itemize}
We call this operation \textit{scalar multiplication}. An additive subgroup $S$ of $(M,+)$ is called a \textit{submodule} if it is closed under scalar multiplication, i.e. $rs \in S$ for any $r \in R$ and $s \in S,$ and we denote this by $S \leq M.$ We call $S$ a \textit{proper submodule} if $S \neq M.$
\end{defin}
A right $R$-module just has multiplication on the right as opposed to on the left. When the ring is commutative, we dispense with any mention of handedness. If the handedness of a module over a non-commutative ring is not mentioned, assume that it applies to the left (resp. right) case. Note that $R$-modules are not necessarily isomorphic to some direct sum of copies of $R$ as in the case of vector spaces. Those that do satisfy this property constitute an important family of $R$-modules.
\begin{defin}
Let $X$ be a nonempty set and $R$ be a ring. A left $R$-module
\[
R\ev{X} = \left\{ \left. \sum_{i = 1}^nr_ix_i \right| n \geq 0,r_i \in R,x_i \in X \right\},
\]
consisting of formal linear combinations, is called a \textit{free left $R$-module generated by $X.$} $X$ is called the \textit{basis} of $R\ev{X},$ and its cardinality is called the \textit{rank} of $R\ev{X}.$
\end{defin}
The notation for a free right $R$-module generated by a set $X$ is $\ev{X}R.$ In many cases, it is unnecessary to mention handedness because it is apparent from the handedness of other modules.
\begin{rem}
Considering that sets are bijective (isomorphic) if and only if they have the same cardinality, it is reasonable to wonder what relation there is between free $R$-modules generated by bijective sets. Harkening back to vector spaces, we may alternatively define
\[
R\ev{X} := \bigoplus_{x \in X}R_x,
\]
where $R_x = R$ for all $x \in X.$ From this definition, it is evident that if sets $X,Y$ are bijective, then $R\ev{X} \cong R\ev{Y}.$ If $X$ is finite of cardinality $n,$ then we represent the free $R$-module generated by $X$ by $R^n.$
\end{rem}
Instead of constructing a free module, we may go backwards and define when an existing module is a free module. Note that the sum mentioned is an actual sum within an abelian group as opposed to a formal sum.
\begin{defin}
A left $R$-module $F$ is said to be \textit{free} if there exists a set $X \subseteq F$ such that for any $a \in F,$ there exists a unique collection $\{r_x\}_{x \in X}$ of scalars such that
\[
a = \sum_{x \in X}r_xx,
\]
where almost all (i.e. all but finitely many) $r_x = 0.$ If it exists, $X$ is called a \textit{basis}.
\end{defin}
In linear algebra, the action of a linear transformation was defined entirely by where it sent the basis vectors. Free modules behave similarly, and a formal description of this attribute also describes an important universal property.
\begin{defin}
Let $X$ be a nonempty set, $R$ be a ring, and $\iota:X \hookrightarrow R\ev{X}$ be the inclusion. For any function $f:X \to Y,$ where $Y$ is a left $R$-module, there is a unique $R$-homomorphism $g:R\ev{X} \to Y$ making the diagram
\[
\begin{tikzcd}
X \ar[rd,hook,"\iota"'] \ar[rr,"f"] & & Y \\
& R\ev{X} \ar[ru,dashed,"\exists!g"'] &
\end{tikzcd}
\]
commute. The universal arrow $g$ is obtained by \textit{extending $f$ by linearity}. Conversely, any left $R$-module $F$ with map $\iota':X \to F$ satisfying the universal property described above is called a \textit{free $R$-module generated by $X.$}
\end{defin}
Extending by linearity allows us to construct a canonical isomorphism between free $R$-modules generated by equinumerous sets. As such, we may talk about \emph{the} free $R$-module generated by a set $X,$ up to isomorphism. Before leaving the notion of freeness, we would be remiss to not mention how this notion generalizes outside of modules.
\begin{defin}
Let $S:\Cs{C} \to \Cs{D}$ be a forgetful functor between appropriate categories $\Cs{C}$ and $\Cs{D}$ and $X \in \Cs{D}$ be any object. A \textit{free $\Cs{C}$ object} on $X$ is a pair $(Y,\iota),$ where $Y \in \Cs{C}$ and $\iota:X \to S(Y),$ such that for any $Z \in \Cs{C}$ and morphism $f:X \to S(Z),$ there exists a unique morphism $g:Y \to Z$ making the diagram
\[
\begin{tikzcd}
X \ar[rd,"\iota"'] \ar[rr,"f"] & & S(Z) \\
& S(Y) \ar[ru,dashed,"\exists!S(g"'] &
\end{tikzcd}
\]
commute. In other words, a \textit{free $\Cs{C}$ object} on $X$ is an initial object of the comma category $X/S,$ where $X:\Cs{D} \to \Cs{D}$ is the constant functor. If $S$ has a left adjoint $F:\Cs{D} \to \Cs{C},$ then we call $F$ a \textit{free functor}, and we could alternatively define a \textit{free $\Cs{C}$ object} on $X$ to be $F(X)$ with a suitable morphism $\iota:X \to S(F(X)).$
\end{defin}
\begin{rem}
An \textit{endomorphism} of a group $G$ is a homomorphism from $G$ to itself. The set of all endomorphisms of a group form a ring under the operations of point-wise addition and composition of functions; we call this ring the \textit{endomorphism ring of $G$} and denote it by $\ope{End}(G).$
Scalar multiplication may be equivalently defined through an alternative construction mirroring the permutation representation of a group action. Let $R$ be a ring. A scalar multiplication in an $R$-module $M$ may then be defined as a ring homomorphism $R \to \ope{End}(M).$
\end{rem}
\begin{exam}
Here are some examples of familiar structures which may be viewed as modules.
\begin{itemize}
\item Abelian groups may be regarded as $\Z$-modules, with submodules being subgroups.
\item Vector spaces over a field $k$ are $k$-modules, with submodules being subspaces.
\item Any ring $R$ may be considered as (left or right) $R$-module with scalar multiplication as ring multiplication, with submodules being ideals.
\end{itemize}
\end{exam}
By virtue of the fact that we call modules as left/right $R$-modules, it is redundant to explicitly say that $R$ is a ring. As such, we will resort to naming the ring through its relation to a specific module. Unless stated otherwise, lack of a mention of handedness when considering non-commutative rings means that inserting either ``left'' or ``right'' will leave a true statement. Now we mention very simple methods of constructing submodules.
\begin{defin}
Let $S \subseteq M$ be a subset of a left $R$-module, then the \textit{submodule generated by $S$}, denoted by $\langle S\rangle$, is defined as
\[
R\ev{S} = \langle S\rangle = \left\{\left.\sum_{i = 1}^n r_i s_i\right|r_i \in R,s_i \in S, n \geq 1\right\}.
\]
If $S = \{m\}$ for some $m \in M,$ then $\langle S\rangle$ is called the \textit{cyclic submodule generated by $m.$}
\end{defin}
\begin{defin}
Let $S,T \leq M$ be submodules of a left $R$-module $M.$ Then the \textit{sum of submodules} $S$ and $T,$ denoted by $S+T,$ is defined as
\[
S+T = \{s+t\ |\ s \in S, t \in T\}.
\]
Furthermore, the \textit{intersection of submodules} $S$ and $T,$ denoted by $S \cap T,$ is defined as
\[
N = S \cap T = \{n\ |\ n \in S, n \in T\}.
\]
\end{defin}
Sums and intersections of additive subgroups which are not submodules often are not submodules because subgroups are not necessarily closed under scalar multiplication. Now we present a means of imposing a module structure on a set constructed from a given submodule. This type of construction is fundamental within much of algebra.
\begin{defin}
Let $N \leq M$ be a submodule of of the left $R$-module $M.$ Then the \textit{quotient module} $M/N$ is defined as the set of cosets
\[
M/N = \{m+N\ |\ m \in M\}
\]
with addition and scalar multiplication given by
\[
(m+N)+(m'+N) = (m+m')+N, \qquad r(m+N) = rm+N,
\]
for any $m,m' \in M$ and $r \in R.$
\end{defin}
\begin{defin}
Let $M,N$ be left $R$-modules. A function $\varphi:M \to N$ satisfying
\begin{itemize}
\item $\varphi(m_1+m_2) = \varphi(m_1)+\varphi(m_2),$
\item $\varphi(rm) = r\varphi(m),$
\end{itemize}
for any $r \in R$ and $m,m_1,m_2 \in M$ is called an \textit{$R$-homomorphism}. A bijective $R$-homomorphism is called an \textit{$R$-isomorphism}.
\end{defin}
It may be apparent that there is a lot of parallelism to these definitions. Explicitly, there is some type of object with structure along with functions which relate these objects by preserving structure. More importantly, these maps relating structured objects give rise to important substructures.
\begin{defin}
Let $\varphi:M \to N$ be an $R$-homomorphism of left $R$-modules. The following distinguished subsets are, in fact, submodules.
\begin{itemize}
\item The \textit{kernel} of $\varphi$ is denoted by $\Ker(\varphi)$ and defined as $\Ker(\varphi) = \{m \in M\ |\ \varphi(m) = 0 \in N\}.$
\item The \textit{image} of $\varphi$ is denoted by $\Im(\varphi)$ and defined as $\Im(\varphi) = \{\varphi(m)\ |\ m \in M\}.$
\item The \textit{cokernel} of $\varphi$ is denoted by $\Coker(\varphi)$ and defined as $\Coker(\varphi) = N/\Im(\varphi).$
\end{itemize}
\end{defin}
Now, we should discuss a means of combining modules. There are two ways of doing this which may seem different but which are inherently tied to one another.
\begin{defin}
If $M,N$ are left $R$-modules, then their \textit{external direct sum}, denoted by $M \boxplus N,$ is their cartesian product $M \times N$ with coordinate-wise operations
\begin{align*}
(m,n)+(m',n') & = (m+m',n+n'), \\
r(m,n) & = (rm,rn),
\end{align*}
where $r \in R,$ $m,m' \in M,$ and $n,n' \in N.$ Under these operations, $T = M \boxplus N$ becomes a left $R$-module.
\end{defin}
\begin{prop} \label{summ}
The following statements are equivalent for left $R$-modules $M_k$ and $T,$ where $1 \leq k \leq n$ for some $n \in \mathbb{N}:$
\begin{enumerate}[label = \emph{(\roman*)}]
\item $M_1 \boxplus M_2 \boxplus \ldots \boxplus M_n \cong T;$
\item there exist injective $R$-homomorphisms $\iota_k:M_k \to T$ such that
\[
T = \sum_{k = 1}^n \Im(\iota_k) \quad\text{and}\quad \Im(\iota_k) \cap \Im(\iota_j) = 0,
\]
where $1 \leq k \neq j \leq n;$
\item there exist $R$-homomorphisms $\iota_k:M_k \to T$ such that, for every $t \in T,$ there are unique $m_k \in M_k$ such that
\[
t = \sum_{k = 1}^n \iota_k(m_k);
\]
\item there exist $R$-homomorphisms $\iota_k:M_k \to T$ (called injections) and $\pi_k:T \to M_k$ (called projections) such that
\begin{itemize}
\item $\pi_k\iota_k = \mathds{1}_{M_k},$
\item $\sum_{k = 1}^n \iota_k\pi_k = \mathds{1}_T,$
\item $\pi_k\iota_j = 0$ for $1 \leq k \neq j \leq n.$
\end{itemize}
\item there exist $R$-homomorphisms $\pi_k:T \to M_k,$ such that the map
\[
\psi:
\begin{cases}
T \to M_1 \boxplus M_2 \boxplus \ldots \boxplus M_n \\
t \to (\pi_1(t),\pi_2(t),\ldots,\pi_n(t))
\end{cases}
\]
is an isomorphism.
\end{enumerate}
\end{prop}
\begin{defin}
If $M,N$ are submodules of a left $R$-module $T$ such that each $t \in T$ has a unique expression of the form $t = m+n$ for $m \in M$ and $n \in N,$ then we call $T$ the \textit{internal direct sum} of $M$ and $N,$ denoted as $T = M \oplus N.$
\end{defin}
\begin{cor}
Let $T$ be a left $R$-module having submodules $M_k,$ where $1 \leq k \leq n$ for some $n \in \mathbb{N}.$ Then
\[
T = \bigoplus_{k = 1}^nM_k
\]
if and only if $\sum_{k = 1}^n M_k = T$ and $M_k \cap M_j = 0$ for $1 \leq k \neq j \leq n.$ In particular,
\[
\bigoplus_{k = 1}^n M_k \cong M_1 \boxplus M_2 \boxplus \ldots \boxplus M_n.
\]
\end{cor}
This corollary tells us that the external direct sum of modules is in fact an internal direct sum of them (up to isomorphism) when regarded as submodules of the resulting external direct sum. Consequently, we will dispense with the notation $\boxplus$ and let $\oplus$ represent both internal and external direct sums. Now we present a condition for when a ``summing'' collection of submodules constitute a direct sum.
\begin{prop} \label{spa}
Let $M = S_1+\ldots+S_n$ in which $S_i \leq M$ for all $i.$ Then $M = S_1 \oplus \ldots \oplus S_n$ if and only if
\[
S_i \cap \left(S_1+\ldots+\hat{S_i}+\ldots+S_n\right) = 0
\]
for all $i,$ where $\hat{S_i}$ indicates that $S_i$ is omitted from the sum.
\end{prop}
\begin{exam}
Let $B$ be a 2-dimensional vector space over field $k$ with basis $\{x,y\}.$ It is then a $k$-module and may be written as the direct sum $V = \ev{x} \oplus \ev{y}.$ Observe that
\[
\ev{x+y} \cap \ev{x} = 0 = \ev{x+y} \cap \ev{y},
\]
but it is not true that $V = \ev{x+y} \oplus \ev{x} \oplus \ev{y}$ because $\ev{x+y} \cap \ev{\ev{x}+\ev{y}} \neq 0.$ But note that we may write $V = \ev{x+y} \oplus \ev{x} = \ev{x+y} \oplus \ev{y},$ which goes to show that bases are not unique and that alternative bases may be easily constructed from the standard basis.
\end{exam}
\begin{exam}
If $V$ is an $n$-dimensional vector space over field $k$ with basis $\{x_1,\ldots,s_n\},$ then $V = \ev{x_1} \oplus \ldots \oplus \ev{x_n}.$ This is the familiar statement from linear algebra that every vector may be represented uniquely by a linear combination of the basis vectors. Generalizing the previous example, we may not add any other basis vectors because we would end up representing 0 in multiple ways, implying that we cannot have any fewer basis vectors. Consequently, $V$ is a direct sum of $n$ 1-dimensional vector spaces if and only if $\ope{dim}(V) = n.$
\end{exam}
\begin{defin}
A submodule $M$ of a left $R$-module $T$ is a \textit{direct summand} of $T$ if there exists a submodule $N$ for which $M \oplus N = T.$ If such $N$ exists, it is called the \textit{complement} of $M$ in $T.$ To denote a module as a direct summand, we write $M \leq_\oplus T.$
\end{defin}
It is obvious that inclusion of submodules is a transitive relation, but it is also true that direct summands are also transitive. If $S \leq_\oplus N$ and $N \leq_\oplus M$ be $R$-modules, then $S \leq_\oplus M$ (this can be shown by writing $M$ as a direct sum of $N$ and its complement before writing $N$ as the direct sum of $S$ and its complement).
\begin{rem}
For a direct summand $S$ of a left $R$-module $M,$ complements of $S$ are not necessarily unique. Let $V$ be a vector space over the field $k$ with basis $\{a,b\}.$ Then $\langle\alpha a+b\rangle$ is a complement of $\langle a\rangle$ for any $\alpha \in k.$ While they may not be unique, all complements are isomorphic. This may seem strange, but the definition of internal direct sum involves an equality, not an isomorphism.
\end{rem}
There exists a special type of homomorphism which aids in detecting direct summands. This was mentioned earlier on in the category theory section and also in the splitting lemma. A proof will be provided since the result of the following lemma will be used to simplify the proof of the splitting lemma.
\begin{defin}
A submodule $M$ of a left $R$-module $T$ is a \textit{retract} of $T$ if there exists an $R$-homomorphism $r:T \to M,$ called a \textit{retraction}, such that $r(m) = m$ for all $m \in M.$ Equivalently, $r$ is a retraction if and only if the inclusion $i:M \to T$ composes with $r$ to the identity, i.e. $ri = \mathds{1}_M.$
\end{defin}
\begin{lem} \label{ret}
A submodule $M$ of a left $R$-module $T$ is a direct summand if and only if there exists a retraction $r:T \to M.$
\end{lem}
\begin{proof}
Suppose that $M \leq T$ are $R$-modules such that there is a retraction $r:T \to M;$ in particular, $ri = \mathds{1}_M$ for $i:M \hookrightarrow T$ the inclusion. We claim that $T = K \oplus M,$ where $K = \Ker(r).$ We may show this using statement (ii) from \thref{summ}. Let $j:K \hookrightarrow T$ be the inclusion. Because $ri = \mathds{1}_M,$
\[
\Im(i) \cap \Im(j) = 0
\]
because otherwise there would be multiple elements of $M$ mapped to zero by $r$ by the definition of the kernel. On the other hand, we may write any element $t \in T$ as $t = (t-r(t))+r(t).$ Notice that $r(t) \in M$ and that $r(t) = i(r(t)).$ Furthermore $r^2 = r,$ so $t-r(t) \in K$ since $K$ is the kernel of $r.$ Similarly, we may then say that $t-r(t) = j(t-r(t)),$ so
\[
T = \Im(i) \oplus \Im(j)
\]
and we are done.
Suppose that $T = M \oplus K$ is a direct sum of $R$-modules. Then for any $t \in T,$ there exist unique $m \in M$ and $k \in K$ such that $t = m+k.$ We then define
\[
r:
\begin{cases}
T \longrightarrow M \\
m+k \mapsto m
\end{cases}.
\]
For $m \in M,$ we may write $m = m+0,$ so $r(m) = m.$ As such, $r$ is a retraction.
\end{proof}
\begin{cor}
Let $T = S \oplus N.$
\begin{enumerate}[label = \emph{(\roman*)}]
\item If there exists some submodule $M$ such that $S \leq M \leq T,$ then $M = S \oplus (N \cap M).$
\item If there exists some submodule $S' \leq S,$ then $T/S' = S/S' \oplus (N+S')/S'.$
\end{enumerate}
\end{cor}
\begin{proof}
\thref{ret} allows us to demonstrate suitable retractions to produce the desired decompositions into direct summands.
\begin{enumerate}[label = (\roman*)]
\item Because we may write each $t \in T$ uniquely as a sum $t = s+n$ for $s \in S$ and $n \in N,$ we may define a map $r:T \to S$ defined by $s+n \mapsto s.$ This is clearly a retraction and remains so even if we restrict the domain to obtain the map $r|_M:M \to S.$ Restricting the domain similarly restricts the kernel, so we obtain that $\Ker(r|_M) = N \cap M,$ and we are done.
\item Consider the map on quotient modules $\hat{r}:T/S' \to S/S'$ defined similarly to $r$ but with elements being representatives of cosets rather than individual elements. This is clearly a retraction, so we have only left to determine the kernel. Note that $\Ker(\hat{r})$ consists of all representatives $t+S'$ such that
\[
\hat{r}(t+S') = r(t)+S' = 0+S' \Longleftrightarrow r(t) \in S'.
\]
From this it is clear that all elements of the kernel take the form $(n+s')+S'$ for some $s' \in S'$ and $n \in N,$ so $\Ker(\hat{r}) = (N+S')/S'$ and we are done.
\end{enumerate}
\end{proof}
Thus far, we have only been considering methods of combining finitely many modules together to form a new module. When our family of modules is infinite, there are two primary methods for combining the modules.
\begin{defin}
Let $\{S_i\}_{i \in I}$ be an indexed family of sets. An \textit{$I$-tuple} is a function $f:I \to \bigcup_{i \in I} S_i$ where $f(i) \in S_i$ for all $i \in I.$ The \textit{cartesian product} $\prod_{i \in I} S_i$ is the set of all $I$-tuples.
\end{defin}
\begin{defin}
Let $\{M_i\}_{i \in I}$ be an indexed family of left $R$-modules. The \textit{direct product} $\prod_{i \in I} M_i$ is the cartesian product $\prod_{i \in I} M_i$ with component-wise addition and scalar multiplication. The \textit{direct sum} $\bigoplus_{i \in I} M_i$ consists of all those $I$-tuples having 0 for almost all (i.e. all but finitely many) components, also with component-wise addition and scalar multiplication. The \textit{$k$-th projection} is the map
\[
\pi_k:
\left\{
\begin{array}{rcl}
\prod_{i \in I} M_i & \longrightarrow & M_k \\
f & \longmapsto & f(k)
\end{array}
\right.,
\]
and the \textit{$k$-th injection} is the map
\[
\iota_k:
\left\{
\begin{array}{rcl}
M_k & \longrightarrow & \bigoplus_{i \in I} M_i \\
m_k & \longmapsto & f_k
\end{array}
\right.,
\]
where $f_k(k) = m_k$ and 0 otherwise.
\end{defin}
\begin{rem}
The direct sum is always a submodule of the direct product since it is a subset. Direct sums and direct products of finitely indexed families of modules are isomorphic, motivating our decision to omit mention of direct products up until this point. Unsurprisingly, direct sums over finitely indexed families coincide with the finite direct sums discussed earlier in this subsection.
\end{rem}
There is another result which generalizes \thref{spa} by using submodules generated by unions as opposed to sums. In particular, it asserts that a collection of submodules whose union spans the ambient module constitutes a direct sum precisely when it is impossible to write elements of a particular submodule as linear combinations of elements from other submodules.
\begin{prop}
Let $\{M_i\}_{i \in I}$ be a family of submodules of a left $R$-module $T.$ If $T = \ev{\bigcup_{i \in I} S_i},$ then the following are equivalent:
\begin{enumerate}[label = \emph{(\roman*)}]
\item $T = \bigoplus_{i \in I} M_i;$
\item every $t \in T$ has a unique expression of the form $t = m_{i_1}+m_{i_2}+\cdots+m_{i_n},$ where $m_{i_j} \in M_{i_j};$
\item $M_i \cap \ev{\bigcup_{i \neq j} M_j} = 0$ for each $i \in I.$
\end{enumerate}
\end{prop}
\subsection{Correspondence and Isomorphism Theorems}
Correspondence theorems provide a bijective correspondence between ``containers'' of a modded out submodule and the submodules of the resulting quotient object.
\begin{theo}[Correspondence Theorem]
Suppose that $M \leq T$ is a submodule of the left $R$-module $T.$ Let
\[
\mathcal{S} := \{S \leq T\ |\ M \leq S\}, \qquad \mathcal{S}/\mathcal{M} := \{S/M \leq T/M\ |\ M \leq S\}
\]
be the sets of intermediate submodules in $T$ and in $T/M,$ respectively. Then the function
\[
\phi:
\begin{cases}
\mathcal{S} \longrightarrow \mathcal{S}/\mathcal{M} \\
S \longmapsto S/M
\end{cases}
\]
is a bijection.
\end{theo}
The correspondence theorem for rings (resp. abelian groups) is just a corollary if you replace $T$ by a ring viewed as a left $T$-module (resp. group viewed as a $\Z$-module) and $M$ by a left ideal (resp. subgroup). We now present the primary isomorphism theorem which ties together quotient objects, kernels, and images of homomorphisms.
\begin{theo}[First Isomorphism Theorem] \label{firso}
If $\varphi:M \to N$ is an $R$-homomorphism of left $R$-modules, then the function
\[
\overline{\varphi}:
\left\{
\begin{array}{rcl}
M/\Ker(\varphi) & \longrightarrow & \Im(\varphi) \\
m+\Ker(\varphi) & \longmapsto & \varphi(m)
\end{array}
\right.
\]
is an $R$-isomorphism.
\end{theo}
\begin{rem}
This theorem not only provides an isomorphism, but it gives a means of factoring $\varphi,$ namely as the composition $\varphi = \iota\overline{\varphi}\pi,$ where $\iota:\Im(\varphi) \hookrightarrow N$ is the inclusion and $\pi:M \to M/\Ker(\varphi)$ is the canonical projection. In particular, the projection map is determined by the kernel, so $\varphi$ is determined by its kernel.
\end{rem}
This theorem also allows us to prove a statement about all $R$-modules which serves to highlight a similarity with groups, specifically how they are all derived from free groups by adding relations.
\begin{theo} \label{res}
Every left or right $R$-module is a quotient module of some free $R$-module.
\end{theo}
\begin{proof}
Let $M$ be a left $R$-module and $\varphi:R\ev{M} \to M$ be the extension of the identity function $\mathds{1}:M \hookrightarrow M$ by linearity. Explicitly, $\varphi$ sends formal linear combinations in $R\ev{M}$ to actual linear combinations in $M.$ Clearly $\varphi$ is surjective, so \thref{firso} states that $M \cong F/\Ker(\varphi).$
\end{proof}
The proof shows a particular way to write $M$ as a quotient of a free module, but any surjective $R$-homomorphism $F \to M,$ for $F$ a free $R$-module, will suffice to prove the result. The following two isomorphism theorems follow from applying the first isomorphism theorem to certain maps.
\begin{cor}[Second Isomorphism Theorem]
Suppose that $M,N \leq T$ are submodules of a left $R$-module $T.$ If $\pi:T \to T/M$ is the canonical projection, let $\pi|N:N \to T/M$ be the restriction to the submodule $N.$ Applying \thref{firso} to $\pi|N$ produces the $R$-isomorphism
\[
\overline{\pi|N}:N/(N \cap M) \longrightarrow (N+M)/M.
\]
\end{cor}
\begin{cor}[Third Isomorphism Theorem]
If $S \leq M \leq T$ is a chain of submodules of a left $R$-module $T,$ then applying \thref{firso} to the map $f:T/S \to T/M$ produces the $R$-isomorphism
\[
\overline{f}:(T/S)/(M/S) \longrightarrow T/M.
\]
\end{cor}
Again, there are analogues of all the isomorphism theorems for rings as well as abelian groups (and general groups) obtained by identifying the module structure of rings and abelian groups.
\subsection{Exact Sequences}
\begin{defin}
A sequence of left $R$-modules and $R$-module homomorphisms
\[
\begin{tikzcd}
\cdots \ar[r] & M_{n+1} \ar[r,"f_{n+1}"] & M_n \ar[r,"f_n"] & M_{n-1} \ar[r] & \cdots
\end{tikzcd}
\]
is called an \textit{exact sequence} if $\Im(f_{n+1}) = \Ker(f_n)$ for all $n.$
\end{defin}
There are some specific examples of sequences which may be considered exact based on simple criteria.
\begin{prop}
Let $A,B,C$ be left $R$-modules and $f,g,h$ be $R$-module homomorphisms.
\begin{itemize}
\item A sequence $0 \to A \stackrel{f}{\to} B$ is exact if and only if $f$ is a monomorphism.
\item A sequence $B \stackrel{g}{\to} C \to 0$ is exact if and only if $g$ is an epimorphism.
\item A sequence $0 \to A \stackrel{h}{\to} B \to 0$ is exact if and only if $h$ is an isomorphism.
\end{itemize}
\end{prop}
\begin{defin}
A \textit{short exact sequence} of left $R$-modules is an exact sequence of the form
\[
\begin{tikzcd}
0 \ar[r] & A \ar[r,"f"] & B \ar[r,"g"] & C \ar[r] & 0,
\end{tikzcd}
\]
i.e. $f$ is monic, $g$ is epic, and $\Im(f) = \Ker(g).$
\end{defin}
Short exact sequences may be very useful for stating many definitions and theorems, as will become apparent throughout this document. For now, we will state the definitions of finitely generated, finitely presented, and coherent modules as well as restate the isomorphism theorems for modules.
\begin{defin} \label{pres}
Let $M$ be a left $R$-module. A \textit{presentation} of $M$ is an exact sequence
\[
\begin{tikzcd}
\bigoplus_{\alpha \in A}R \ar[r] & \bigoplus_{\beta \in B}R \ar[r] & M \ar[r] & 0,
\end{tikzcd}
\]
where $A,B$ are index sets.
\end{defin}
\begin{rem}
Such a presentation always exists on account of \thref{res}.
\end{rem}
\begin{defin}
With reference to the exact sequence in definition \ref{pres}, a left $R$-module $M$ is
\begin{enumerate}
\item \textit{finitely generated} if there exists a presentation in which $B$ is finite;
\item \textit{finitely presented} if there exists a presentation in which both $A$ and $B$ are finite;
\item \textit{coherent} if $M$ is finitely generated and if every finitely generated submodule $N \leq M$ is finitely presented. Alternatively $M$ is coherent if, for any submodule $N \leq M,$ $B$ being finite implies that $A$ is finite in any presentation of $N.$
\end{enumerate}
\end{defin}
\begin{rem}
It should be clear that $3 \Longrightarrow 2 \Longrightarrow 1$ from the definitions. It is also interesting to note the parallelism to presentations of groups. The direct sum indexed by $B$ corresponds to the generators of the free group while the direct sum indexed by $A$ corresponds to the relations which generate a normal subgroup to be modded out by.
\end{rem}
\begin{theo}
Let all modules be $R$-modules and all maps be $R$-module homomorphisms.
\begin{itemize}
\item If
\[
\begin{tikzcd}
0 \ar[r] & A \ar[r,"f"] & B \ar[r,"g"] & C \ar[r] & 0
\end{tikzcd}
\]
is a short exact sequence, then $A \cong \Im(f)$ and $B/\Im(f) \cong C.$
\item If $A \subseteq B \subseteq C$ is a chain of submodules, then there is a short exact sequence
\[
\begin{tikzcd}
0 \ar[r] & B/A \ar[r] & C/A \ar[r] & C/B \ar[r] & 0.
\end{tikzcd}
\]
\end{itemize}
\end{theo}
We know in general that the last module in a short exact sequence is isomorphic to the quotient of the second module by the image of the first module ($C \cong B/A$ based on the short exact sequence in the above definition). This brings us to a special class of short exact sequences.
\begin{defin}
A short exact sequence
\[
\begin{tikzcd}
0 \ar[r] & A \ar[r] & B \ar[r] & C \ar[r] & 0.
\end{tikzcd}
\]
is called \textit{split exact} if $B \cong A \oplus C.$
\end{defin}
Split exact sequences often make computations more manageable, so it would be nice to be able to identify them. Luckily, there is a lemma which gives us some criterion for determining when a short exact sequence is split exact.
\begin{lem}[Splitting Lemma]
Let $R$ be a ring. For a short exact sequence
\[
\begin{tikzcd}
0 \ar[r] & A \ar[r,"f"] & B \ar[r,"g"] & C \ar[r] & 0.
\end{tikzcd}
\]
of $R$-modules, the following statements are equivalent:
\begin{enumerate} [label = \emph{(\roman*)}]
\item $f$ is a section;
\item $g$ is a retraction;
\item there is an isomorphism $B \cong A \oplus C$ making a commutative diagram as below, where the maps are inclusions and projections.
\end{enumerate}
\[
\begin{tikzcd}
& & B \ar[dd,"\cong"] \ar[rd,"g"] & & \\
0 \ar[r] & A \ar[ru,"f"] \ar[rd,"\iota_1"'] & & C \ar[r] & 0 \\
& & A \oplus C \ar[ru,"\pi_2"'] & &
\end{tikzcd}
\]
\end{lem}
\begin{proof}
This proof will be completed in 3 parts.
\begin{itemize}
\item (i) $\Longrightarrow$ (iii): Because $f$ is a section, there exists a section $p:B \to A$ such that $pf = \mathds{1}_A.$ Define
\[
h:
\begin{cases}
B & \to A \oplus C \\
b & \mapsto (p(b),g(b)).
\end{cases}
\]
The map $p$ is clearly epic as a left divisor of an epimorphism. Because both its components are epimorphisms, $h$ is an epimorphism. Additionally,
\[
\Ker(h) = \Ker(p) \cap \Ker(g) = \Ker(p) \cap \Im(f).
\]
Choose some $a \in A$ such that $f(a) \in \Ker(p).$ Because $pf = \mathds{1}_A,$ $p(f(a)) = 0$ if and only if $a = 0.$ Therefore, $\Ker(h) = \{f(0)\} = 0$ so that $h$ is a monomorphism. Therefore $h:B \to A \oplus C$ is an isomorphism. It is then easy to check that the triangles commute:
\[
\pi_2 h(b) = \pi_2(p(b),g(b)) = g(b),
\]
\[
hf(a) = (pf(a),gf(a)) = (a,0) = \iota_1(a).
\]
\item (iii) $\Longrightarrow$ (i) Let $h:B \to A \oplus C$ be an isomorphism satisfying $\iota_1 = hf,$ which is guaranteed by hypothesis. Define
\[
p:
\begin{cases}
B & \to A \\
b & \mapsto \pi_1h(b),
\end{cases}
\]
where $\pi_1:A \oplus C \to A$ is the projection onto the first component. Then observe that
\[
pf(a) = \pi_1(hf(a)) = \pi_1\iota_1(a) = \pi_1(a,0) = a,
\]
so $p$ is a retraction.
\item (ii) $\iff$ (iii): Because $g$ is a retraction, \thref{ret} states that
\[
B \cong C \oplus \Ker(g) = C \oplus \Im(f) \cong C \oplus A,
\]
where the equality is due to exactness and the last isomorphism follows from \thref{firso}. There must exist a section $s:C \to B$ such that $gs = \mathds{1}_C$ because $g$ is a retraction. Similar to the proof of \thref{ret}, we may write $b = (b-sg(b))+sg(b),$ where $b-sg(b) \in \Ker(g) = \Im(f).$ We then define
\[
h:
\begin{cases}
B & \to A \oplus C \\
b & \mapsto (f^{-1}(b),g(b)),
\end{cases}
\]
and it is not difficult to show that the triangles commute:
\[
hf(a) = (f^{-1}f(a),gf(a)) = (a,0) = \iota_1(a),
\]
\[
\pi_2h(b) = \pi_2(f^{-1}(b),g(b)) = g(b).
\]
Conversely, start with $h$ as defined, which satisfies (iii), and $\iota_2:C \to A \oplus C$ the inclusion. Note that
\[
gh^{-1}\iota_2 = \pi_2hh^{-1}\iota_2 = \pi_2\iota_2 = \mathds{1}_C,
\]
so $g$ is a retraction.
\end{itemize}
\end{proof}
Whenever there is a sequence of objects and morphisms within a particular category $\Cs{C},$ we may apply any functor with source $\Cs{C}$ and observe how the properties of the sequence in the image differ from those of the original sequence. In particular, loss or preservation of exactness is often of importance. Here we classify some ways in which exactness may be lost or preserved within the setting of short exact sequences because it extends to exact sequences of arbitrary length.
\begin{defin}
Let
\[
\begin{tikzcd}
0 \ar[r] & A \ar[r] & B \ar[r] & C \ar[r] & 0
\end{tikzcd}
\]
be a short exact sequence in $\Cb{R\text{-}Mod}.$ A functor $F:\Cb{R\text{-}Mod} \to \Cb{Ab}$ is called
\begin{itemize}
\item \textit{right exact} if
\[
\begin{tikzcd}
F(A) \ar[r] & F(B) \ar[r] & F(C) \ar[r] & 0
\end{tikzcd}
\]
is exact,
\item \textit{left exact} if
\[
\begin{tikzcd}
0 \ar[r] & F(A) \ar[r] & F(B) \ar[r] & F(C)
\end{tikzcd}
\]
is exact,
\item \textit{half exact} if
\[
\begin{tikzcd}
F(A) \ar[r] & F(B) \ar[r] & F(C)
\end{tikzcd}
\]
is exact,
\item \textit{exact} if
\[
\begin{tikzcd}
0 \ar[r] & F(A) \ar[r] & F(B) \ar[r] & F(C) \ar[r] & 0
\end{tikzcd}
\]
is exact.
\end{itemize}
\end{defin}
Unsurprisingly, there are corresponding notions for contravariant functors.
\begin{defin}
A contravariant functor $F:\Cb{R\text{-}Mod} \to \Cb{Ab}$ is called
\begin{itemize}
\item \textit{right exact} if
\[
\begin{tikzcd}
F(C) \ar[r] & F(B) \ar[r] & F(A) \ar[r] & 0
\end{tikzcd}
\]
is exact,
\item \textit{left exact} if
\[
\begin{tikzcd}
0 \ar[r] & F(C) \ar[r] & F(B) \ar[r] & F(A)
\end{tikzcd}
\]
is exact,
\item \textit{half exact} if
\[
\begin{tikzcd}
F(C) \ar[r] & F(B) \ar[r] & F(A)
\end{tikzcd}
\]
is exact,
\item \textit{exact} if
\[
\begin{tikzcd}
0 \ar[r] & F(C) \ar[r] & F(B) \ar[r] & F(A) \ar[r] & 0
\end{tikzcd}
\]
is exact.
\end{itemize}
\end{defin}
\begin{rem}
It is fairly easy to keep track of what exact and half exact functors are in the two cases, but it is a bit more tricky with left and right exact functors. The simplest way to remember is to recognize that the ``functored'' exact sequence, when written with arrows going from left to right, is exact on the side suggested by the name. However, there is another interpretation which gets more to the heart of the matter.
For functors, left exact refers to preservation of injectivity; on the other hand (pun intended), right exact refers to preservation of surjectivity. Dually, left exact contravariant functors convert epimorphisms into monomorphisms while the right exact ones convert monomorphisms into epimorphisms. Yet another means of describing this distinction involves kernels and cokernels, and this may be deduced easily by the reader by considering what kernels of monomorphisms and cokernels of epimorphisms are.
\end{rem}
In the next subsections, we will discuss some important functors which provide important examples of how exactness is lost or preserved.
\subsection{Hom}
For $R$-modules, Mor sets are instead called Hom sets because morphisms are $R$-module homomorphisms. In fact categories of groups, rings, or other familiar algebraic structures have the same name replacement because morphisms are just different types of homomorphisms. Let $M,N$ be left $R$-modules. Then we write $f \in \Hom{R}{M}{N}$ to mean that $f:M \to N$ is a homomorphism for which $f(rm) = rf(m)$ for any $r \in R$ and $m \in M$, so the subscript indicates from which ring scalars are taken for multiplication by elements of $M$. Something similar may be said when $M,N$ are right $R$-modules. It is interesting to note that $\Hom{R}{M}{N}$ has additional structure when $M,N$ are modules over $R$.
\begin{theo}
Let $R$ be a ring and $M,N$ be left $R$-modules. Suppose that $f,g \in \Hom{R}{M}{N}$ are any $R$-module homomorphisms, then $\Hom{R}{M}{N}$ forms an abelian group under the binary operation $(f+g)(m) = f(m)+g(m)$. The same is also true when $M,N$ are right $R$-modules.
\end{theo}
\begin{rem}
Associativity and commutativity follow from those properties in $N$ while the identity is the zero homomorphism (as $\Cb{R\text{-}Mod}$ is a pointed category). For a map $f$, its inverse $-f$ may be defined by $-f(m) = f((-m))$ and clearly is an $R$-module homomorphism. Therefore, Hom is just a specific kind of representable functor, the same notation will be used when scalars are unimportant or obvious.
\end{rem}
What this theorem tells us is that Hom is a functor from $\Cb{R\text{-}Mod}$ to $\Cb{Ab}.$ Even more, this functor has some useful properties which are worth discussing, especially considering that these properties manifest in other situations. But before going onto new properties, let us briefly return to something familiar.
\begin{theo}
Both the functor $(X,-)$ and contravariant functor $(-,Y)$ are left exact.
\end{theo}
The proof is just a routine diagram chase. However, showing that these functors are not right exact requires counterexamples.
\begin{exam}
Consider the projection map $\pi:\Z \to \Z/n\Z,$ which is clearly a surjection. Apply the functor $(\Z/n\Z,-)$ to obtain the induced map $(\Z/n\Z,\pi):(\Z/n\Z,\Z) \to (\Z/n\Z,\Z/n\Z).$ Because $\Z$ has no non-zero elements of finite order, $(\Z/n\Z,\Z) \cong 0.$ Clearly then $(\Z/n\Z,\pi)$ cannot be a surjection without also asserting that $(\Z/n\Z,\Z/n\Z) \cong 0,$ which is clearly false considering that the identity map is a non-zero element. Therefore, $(X,-)$ does not preserve surjections and therefore cannot be right exact.
Consider the map $f:\Z \to \Z$ defined by $n \mapsto 2n,$ which is clearly injective. Apply $(-,\Z)$ to obtain the induced map $(f,\Z):(\Z,\Z) \to (\Z,\Z).$ Consider an element $\beta$ of the target $(\Z,\Z)$ defined as multiplication by 3. For there to exist some element $\alpha$ of the source $(\Z,\Z)$ such that $(f,\Z)(\alpha) = \beta,$ there would need to be an integer solution to the equation $3 = 2x,$ which is clearly a contradiction.
\end{exam}
\begin{defin}
Let $F:\Cb{R\text{-}Mod} \to \Cb{Ab}$ be a covariant or contravariant functor. $F$ is called \textit{pointed} if $F(0) = 0.$
\end{defin}
\begin{exam}
Let $M$ be any $R$-module. Because $(M,0)$ is the trivial abelian group, the functor $(M,-)$ is pointed. Similarly, $(-,M)$ is pointed. However, $M \oplus -$ is not a pointed functor because $M \oplus 0 \cong M \ncong 0,$ where the isomorphisms are between the underlying abelian groups.
\end{exam}
\begin{defin} \label{adf}
Let $F:\Cb{R\text{-}Mod} \to \Cb{Ab}$ be a covariant or contravariant functor. $F$ is called \textit{additive} if for any $X,Y \in \Cb{R\text{-}Mod}$ and $f,g \in \Hom{R}{X}{Y},$ $F(f+g) = F(f)+F(g),$ i.e. $F_{X,Y}:\Hom{R}{X}{Y} \to \Hom{R}{F(X)}{F(Y)}$ is a group homomorphism.
\end{defin}
\begin{theo}
Let $F:\Cb{R\text{-}Mod} \to \Cb{Ab}$ be a covariant or contravariant functor. The following are equivalent:
\begin{enumerate}[label = \emph{(\roman*)}]
\item $F$ is additive;
\item for any $X,Y \in \Cb{R\text{-}Mod},$ $F(X \oplus Y) \cong F(X) \oplus F(Y);$
\item $F$ preserves split short exact sequences.
\end{enumerate}
\end{theo}
\begin{rem}
Clearly Hom is an additive functor according to how its induced maps are defined. Furthermore, there is a hierarchy between all these properties of functors. It is given by exact $\Longrightarrow$ left/right exact $\Longrightarrow$ half exact $\Longrightarrow$ additive $\Longrightarrow$ pointed.
\end{rem}
So Hom sets for $R$-modules have an abelian group structure, but it is interesting to investigate when Hom sets can have greater structure. This leads us to the following definition and theorem.
\begin{defin}
Let $R,S$ be rings. An abelian group $M$ is said to be an $(R,S)$\textit{-bimodule} if
\begin{itemize}
\item $M \in \Cb{R\text{-}Mod}$,
\item $M \in \Cb{Mod\text{-}S}$,
\item $(\forall m \in M,\forall s \in S,\forall r \in R)$, $(rm)s = r(ms).$
\end{itemize}
That $M$ is an $(R,S)$-bimodule may also be conveyed by writing $_RM_S$, $M \in \Cb{R\text{-}Mod\text{-}S}$, or $M \in \Cb{(R,S)\text{-}Mod}$.
\end{defin}
\begin{exam}
Every ring $R$ can be considered as an $(R,R)$-bimodule. Scalar multiplication on either side may simply defined as ring multiplication on either side. In this case, the compatibility condition between left and right scalar multiplication is satisfied by the associativity of ring multiplication.
\end{exam}
\begin{theo}
Let $R,S,T$ be rings.
\begin{itemize}
\item $\Hom{R}{_RM}{{}_RN_S} \in \Cb{Mod\text{-}S}$ where $(f*s)(rm) = f(rm)s$.
\item $\Hom{R}{_RM_S}{{}_RN} \in \Cb{S\text{-}Mod}$ where $(s*f)(rm) = f(rms)$.
\item $\Hom{S}{M_S}{{}_RN_S} \in \Cb{R\text{-}Mod}$ where $(r*f)(ms) = rf(ms)$.
\item $\Hom{S}{_RM_S}{N_S} \in \Cb{Mod\text{-}R}$ where $(f*r)(ms) = f(rms)$.
\item $\Hom{S}{_RM_S}{{}_TN_S} \in \Cb{T\text{-}Mod\text{-}R}$ where $(t*f*r)(ms) = tf(rms)$.
\item $\Hom{R}{_RM_S}{{}_RN_T} \in \Cb{S\text{-}Mod\text{-}T}$ where $(s*f*t)(rm) = f(rms)t$.
\end{itemize}
\end{theo}
This theorem outlines how various combinations of bimodules can be used to impose additional structure on Hom sets. But thinking categorically, this provides a means for creating functors between module and bimodule categories. This also provides motivation for speaking about the covariant and contravariant Hom functors, which are nothing more than important instances of the covariant and contravariant representable functors.
\begin{exam}
Let $R,S$ be rings. Then we have the following functors:
\begin{itemize}
\item $\Hom{R}{-}{{}_RN_S}:\Cb{R\text{-}Mod} \to \Cb{Mod\text{-}S}$
\item $\Hom{R}{_RM_S}{-}:\Cb{R\text{-}Mod} \to \Cb{S\text{-}Mod}$
\item $\Hom{S}{-}{{}_RN_S}:\Cb{S\text{-}Mod} \to \Cb{R\text{-}Mod}$
\item $\Hom{S}{_RM_S}{-}:\Cb{Mod\text{-}S} \to \Cb{Mod\text{-}R}$
\end{itemize}
\end{exam}
\subsection{Tensor Product}
\begin{defin}
Let $R$ be a ring, $A \in \Cb{Mod\text{-}R}$, $B \in \Cb{R\text{-}Mod}$, and $G$ be any abelian group. A function $f:A \times B \to G$ between abelian groups is called
\begin{itemize}
\item \textit{biadditive} if $$f(a+a',b) = f(a,b)+f(a',b),$$ $$f(a,b+b') = f(a,b)+f(a,b')$$ for any $a,a' \in A$ and $b,b' \in B$;
\item \textit{balanced} if $$f(ar,b) = f(a,rb)$$ for any $r \in R$, $a \in A$, and $b \in B$.
\end{itemize}
\end{defin}
\begin{defin} \label{tens}
Let $R$ be a ring, $A_R$ and $_R B$ be right and left $R$-modules, respectively. Then their \textit{tensor product} is a pair $(A \otimes_R B,\tau)$ where $A \otimes_R B$ is an abelian group and $\tau:A \times B \to A \otimes_R B$ a biadditive and balanced function such that for any abelian group $X$ and anyc biadditive and balanced function $f:A \times B \to X$, there exists a unique homomorphism of abelian groups $f':A \otimes_R B \to X$ making the following diagram commute,
\[
\begin{tikzcd}
A \times B \ar[rr,"f"] \ar[rd,"\tau"'] && X \\
& A \otimes_R B. \ar[ru,dashed,"\exists! f'"'] &
\end{tikzcd}
\]
\end{defin}
The elements of the tensor product take the form of finite sums of the expressions $a \otimes b = \tau(a,b)$, where $a \in A$ and $b \in B$. The tensor product seems to resemble a colimit of sorts, but it is not because the morphisms are special functions but not necessarily group homomorphisms. A colimit would have structural maps which are morphisms within the category of the diagram. We show that the tensor product of any two right and left $R$-modules exists and is unique up to isomorphism of abelian groups.
\begin{theo}
Let $A$ and $B$ be right and left $R$-modules, respectively. Then $A \otimes_R B$ exists.
\end{theo}
\begin{proof}
Let $R$ be a ring, $A_R$ and $_R B$ be right and left $R$-modules, respectively. Let $F$ be the free abelian group with basis $A \times B$. We define a subgroup $S \subseteq F$ consisting of all elements of the type $$(a,b+b')-(a,b)-(a,b'),$$ $$(a+a',b)-(a,b)-(a',b),$$ $$(ar,b)-(a,rb),$$ where $a,a' \in A$, $b,b' \in B$, and $r \in R$. This particular subgroup was selected because taking a quotient by this subgroup will ensure that all elements of those forms will become equivalent to 0. With this said, we define $A \otimes_R B = F/S$ and
\[
\tau:
\begin{cases}
A \times B \to A \otimes_R B \\
(a,b) \mapsto a \otimes b.
\end{cases}
\]
Making elements of the form found in $S$ equivalent to 0 then forces $\tau$ to be biadditive and balanced. Now consider the following diagram:
\begin{center}
\begin{tikzcd}
A\times B \ar[rr,"f"] \ar[rd,hook,"i"] \ar[rdd,"\tau"'] & & X \\
& F \ar[d,two heads,"\nu"] \ar[ru,dashed,"h"] & \\
& A \otimes_R B. \ar[ruu,dashed,"f'"'] &
\end{tikzcd}
\end{center}
Here $i:A \times B \to F$ is the inclusion, $\nu:F \to A \otimes_R B$ the natural map from a group into its quotient, $X$ an abelian group, and $f:A \times B \to X$ an arbitrary biadditive and balanced function. Because the domain of $f$ is the set of generators of $F$, we may extend $f$ by linearity to induce a map $h:F \to X$. We may define a map $f':A \otimes_R B \to X$ by $$f'(a \otimes b) = h(a,b) = f(a,b).$$ Since $f$ is biadditive and balanced, $S \subseteq \Ker(h)$ so that $f'$ is well-defined. Note that $$h(a,b) = f'(a \otimes b) = f'((a,b) + S) = f'\nu(a,b),$$ so $h = f'\nu$. Lastly, by the definition of $\nu$, $\nu(a,b) = a \otimes b = \tau(a,b)$, so we obtain that the outer triangle is commutative, as desired.
\end{proof}
\begin{theo}
For $A$ and $B$ right and left $R$-modules, respectively, $A \otimes_R B$ is unique up to isomorphism.
\end{theo}
\begin{proof}
Suppose that there exists a pair $(Y,g)$ where $Y$ is an abelian group and $g:A \times B \to Y$ a biadditive and balanced function also satisfying definition \ref{tens}. Because $g$ is biadditive and balanced, there exists a unique group homomorphism $g':A \otimes_R B \to Y$, since $A \otimes_R B$ is a tensor product, making the following diagram commute:
\begin{center}
\begin{tikzcd}
A \times B \ar[rd,"\tau"'] \ar[rr,"g"] & & Y \\
& A \otimes_R B. \ar[ur,dashed,"\exists! g'"'] &
\end{tikzcd}
\end{center}
By a similar argument, we have a unique group homomorphism $\tau':Y \to A \otimes_R B$ making the following diagram commute:
\begin{center}
\begin{tikzcd}
A \times B \ar[rd,"g"'] \ar[rr,"\tau"] & & A \otimes_R B \\
& Y. \ar[ur,dashed,"\exists! \tau'"'] &
\end{tikzcd}
\end{center}
By ``gluing'' the triangles together, we obtain another diagram. This diagram commutes because its component triangles individually commute. This commuting allows us to simplify the diagram as such:
\[
\begin{tikzcd}
& Y & \\
A \times B \ar[ru,"g"] \ar[rd,"g"'] \ar[rr,"\tau"] & & \ar[lu,dashed,"g'"'] A \otimes_R B \\
& Y \ar[ru,dashed,"\tau'"'] &
\end{tikzcd}
\Longrightarrow
\begin{tikzcd}
A \times B \ar[rd,"g"'] \ar[rr,"g"] & & Y \\
& Y. \ar[ur,dashed,"g'\tau'"'] &
\end{tikzcd}
\]
But this diagram is just another tensor product triangle, and the induced map between a tensor product and itself is just the identity. Consequently, $g'\tau' = \mathds{1}_Y$ so that $g'$ is a retraction and $\tau'$ is a section. By gluing the initial triangles together across $g$ instead, we find that $\tau' g' = \mathds{1}_{A \otimes_R B}$ so that $\tau'$ is a retraction and $g'$ is a section. Then by \thref{iso}, both $\tau'$ and $g'$ are isomorphisms of abelian groups, so we are done.
\end{proof}
\begin{rem}
Just as with Hom, we can give tensor products additional structure when we include bimodules. Let $R,S,T$ be rings. Using an obvious multiplication definition, we obtain
\[
_RM_S \otimes_S {}_SN_T = {}_R(M \otimes_S N)_T,
\]
which implies that
\begin{itemize}
\item $({}_RM_S \otimes_S -):\Cb{S\text{-}Mod} \to \Cb{R\text{-}Mod},$
\item $(- \otimes_S {}_SN_T):\Cb{Mod\text{-}S} \to \Cb{Mod\text{-}T}.$
\end{itemize}
\end{rem}
There are some other ways in which we may uncover new structure which are slightly less obvious.
\begin{defin}
Let $f:R \to S$ be a ring homomorphism, and let $M$ be a left $S$-module. Then $M$ is a left $R$-module with multiplication defined as
\[
rm = f(r)m.
\]
We say that $_RM$ is obtained from $_SM$ through \textit{restriction of scalars}.
\end{defin}
\begin{defin}
Let $R,S$ be rings, $f:R \to S$ be a ring homomorphism, and $M$ be a right $R$-module. We define $S$ to have the structure of a left $R$-module according to the rule
\[
rs = f(r)s \Longrightarrow M_R \otimes_R {}_RS_S \in \Cb{Mod\text{-}S},
\]
so we obtain a right $S$-module structure from an $R$-module and say that $(M \otimes_R S)_S$ is obtained from $M$ through \textit{extension of scalars}.
\end{defin}
\begin{rem}
If $f:M_R \to M'_R,g:{}_RN \to {}_RN'$ are $R$-homomorphisms, then we may define a homomorphism of abelian groups:
\[
f \otimes g:M \otimes_R N \to M' \otimes_R N'
\]
if we using the biadditive and balanced maps guaranteed by the definition of a tensor product like so
\[
\begin{tikzcd}
A \times B \ar[rr,"\tau'"] \ar[rd,"\tau"'] && M' \otimes_R N' \\
& A \otimes_R B. \ar[ru,dashed,"\exists! f \otimes g'"'] &
\end{tikzcd}
\]
\end{rem}
Now that we have explained what tensor products and Hom are and why they are significant, it is fitting to conclude with a deep connection between the two functors which strongly resembles one of the examples which motivated the concept of adjunction.
\begin{theo}[Hom-Tensor Adjunction]
Let $R,S$ be rings and $M_R,{}_RX_S,N_S$ be modules. Then there is a natural isomorphism of abelian groups:
\[
\begin{tikzcd}
\Hom{S}{M \otimes_R X}{N} \ar[r,"\tau_{M,X,N}"] & \Hom{R}{M}{\Hom{S}{X}{N}}.
\end{tikzcd}
\]
In particular,
\[
\tau_{M,X,N}(M \otimes_R X \stackrel{f}{\longrightarrow} N) =
\left\{
\begin{array}{rcl}
M & \longrightarrow & \Hom{S}{X}{N} \\
m &\longmapsto & f(m \otimes -)
\end{array}
\right.
\]
As such, there is the adjunction:
\[
\begin{tikzcd}
\Cb{Mod\text{-}R} \ar[r, "(- \otimes_R X_S)"{name = L}, bend left = 25] &\Cb{Mod\text{-}S}. \ar[l, "\Hom{S}{_RX_S}{-}"{name = R}, bend left = 25] \ar[phantom, from = L, to = R, "\dashv" rotate = -90]
\end{tikzcd}
\]
\end{theo}
\subsection{Limits and Colimits in the Category of Modules}
In this section, we will go over notable limits. This will be accomplished by varying the form of the family (by changing the poset) over (resp. beneath) which the limit takes form. Because $\Cb{R\text{-}Mod}$ is finitely complete, these examples will also be mentioned. When they exist, limits provide solutions to universal mapping problems, which is to say that the limit is a cone through which all other cones factor uniquely. The dual approach, obtained by adding co- wherever possible in the above statements, will also be discussed.
\subsubsection{Product and Coproduct}
Let $\Cs{C}$ be a category and $A,B \in \Cs{C}$.
\begin{defin}
The (binary) \textit{product} of $A$ and $B$ is a triple $(A \sqcap B,\pi_A,\pi_B)$ in which $A \sqcap B \in \Cs{C}$ and $\pi_A:A \sqcap B \to A$, $\pi_B:A \sqcap B \to B$ are the \textit{projection} morphisms such that for any $X \in \Cs{C}$ and any pair of morphisms $g_A:X \to A$ and $g_B:X \to B$, there exists a unique $h:X\to C$ making the following diagram commute:
\begin{center}
\begin{tikzcd}
& \ar[ld,"g_A"'] X \ar[d,dashed,"\exists! h"] \ar[rd,"g_B"] \\
A & \ar[l,"\pi_A"] A \sqcap B \ar[r,"\pi_B"'] & B.
\end{tikzcd}
\end{center}
\end{defin}
\begin{defin}
The (binary) \textit{coproduct} of $A$ and $B$ is a triple $(A \sqcup B,\iota_A,\iota_B)$ in which $A \sqcup B \in \Cs{C}$ and $\iota_A:A \to A \sqcup B$, $\iota_B:B \to A \sqcup B$ are the \textit{injection} morphisms such that for any $Y \in \Cs{C}$ and any pair of morphisms $g_A:A \to Y$ and $g_B:B \to Y$, there exists a unique $h:A \sqcup B \to Y$ making the following diagram commute:
\begin{center}
\begin{tikzcd}
A \ar[r,"\iota_A"] \ar[rd,"g_A"'] & A \sqcup B \ar[d,dashed,"\exists! h"] & \ar[l,"\iota_B"'] \ar[ld,"g_B"] B \\
& Y &
\end{tikzcd}
\end{center}
\end{defin}
The definitions of (binary) product and coproduct may be obtained much more succinctly by means of the limit and colimit constructions, respectively. Let $I = \{\bullet,*\}$ be a poset with two incomparable elements and $F:I \to \Cs{C}$ be a functor such that $F(\bullet) = A$ and $F(*) = B$. Then the pair $(A,B)$ can be identified with the image of the functor $F$ so that $$A \sqcap B = \varprojlim F,$$ $$A \sqcup B = \varinjlim F.$$ Additionally, products and coproducts may be generalized to families of more than just two objects by considering a trivial poset with arbitrarily many elements.
\begin{exam}
In $\Cb{R\text{-}Mod}$, the (co)product is called the direct product (sum).
\end{exam}
\subsubsection{Pullback and Pushout}
Consider posets $I$ and $J$ defined as,
\[
I = \left\{
\begin{tikzcd}
& \bullet \ar[d] \\
\bullet \ar[r] & \bullet
\end{tikzcd}
\right\}
~~
J = \left\{
\begin{tikzcd}
\ar[d] \bullet \ar[r] & \bullet \\
\bullet &
\end{tikzcd}
\right\}.
\]
Let $F:I \to \Cs{C}$ and $G:J \to \Cs{C}$ be functors whose images are defined, respectively, by
\begin{tabu} to \textwidth {XX}
\begin{equation} \label{one}
\begin{tikzcd}[ampersand replacement = \&]
\& A \ar[d,"f"] \\
B \ar[r,"g"'] \& C
\end{tikzcd}
\end{equation}
&
\begin{equation} \label{two}
\begin{tikzcd}[ampersand replacement = \&]
\ar[d,"g'"'] C' \ar[r,"f'"] \& A' \\
B' \& .
\end{tikzcd}
\end{equation}
\end{tabu}
An image of a functor $F:I \to \Cs{C}$ is called a \textit{cospan}, and an image of a functor $G:J \to \Cs{C}$ is called a \textit{span}. In general, $I$ and $J$ are called cospan and span diagrams, respectively.
\begin{defin}
Given a diagram of objects and morphisms in $\Cs{C}$ of the form \eqref{one}, its \textit{pullback} (or \textit{fibered product}) is a triple $(D,\alpha,\beta)$ where $D \in \Cs{C}$, $\alpha:D \to A$, and $\beta:D \to B$ such that
\begin{equation} \label{thre}
f\alpha = g\beta.
\end{equation}
For every similar triple $(D',\alpha',\beta')$ also satisfying \eqref{thre} (with appropriate substitutions), there exists a unique $h:D' \to D$ making the diagram
\begin{equation} \label{fou}
\begin{tikzcd}
\ar[rdd,bend right,"\beta'"'] D' \ar[rd,dashed,"\exists! h"] \ar[rrd,bend left, "\alpha'"] && \\
& \ar[d,"\beta"'] D \ar[r,"\alpha"] & A \ar[d,"f"] \\
& B \ar[r,"g"'] & C.
\end{tikzcd}
\end{equation}
commute. This diagram is referred to as a \textit{pullback square}.
\end{defin}
\begin{defin}
Given a diagram of objects and morphisms in $\Cs{C}$ of the form \eqref{two}, its \textit{pushout} (or \textit{fibered sum}) is a triple $(D,\alpha,\beta)$ where $D \in \Cs{C}$, $\alpha:A \to D$, and $\beta:B \to D$ such that
\begin{equation} \label{fiv}
\alpha f = \beta g.
\end{equation}
For every similar triple $(D',\alpha',\beta')$ also satisfying \eqref{fiv} (with substitutions), there exists a unique $h:D \to D'$ making the diagram
\begin{center}
\begin{tikzcd}
\ar[d,"g"'] C \ar[r,"f"] & \ar[d,"\alpha"] A \ar[rdd,bend left,"\alpha'"] & \\
B \ar[r,"\beta"'] \ar[rrd,bend right,"\beta'"'] & D \ar[rd,dashed,"\exists! h"] & \\
& & D'
\end{tikzcd}
\end{center}
commute. This diagram is referred to as a \textit{pushout square}.
\end{defin}
As with products and coproducts, limits and colimits make the definitions of pullback and pushout, respectively, much more compact. In particular, pullbacks are limits of cospans ($\varprojlim F$). Dually, pushouts are colimits of spans ($\varinjlim G$). Additionally, pullbacks (resp. pushouts) may be generalized to families with more than just two legs on the cospan (resp. span).
\begin{theo}
In $\Cb{R\text{-}Mod}$, pullbacks and pushouts exist.
\end{theo}
\begin{proof}
The pullback case will be proven, but the pushout solution will be provided for verification by the reader. Let $A,B$ be $R$-modules, $f:A \to C$ and $g:B \to C$ be $R$-module homomorphisms. Define $$D = \{(a,b) \in A \oplus B\ |\ f(a) = g(b)\}.$$ Let $\alpha:D \to A$ be the projection $(a,b) \mapsto a$ restricted to domain $D$, and $\beta:D \to B$ to be the projection $(a,b) \mapsto b$ restricted to domain $D$. Let $d = (a_0,b_0)\in D$ be an element of $D$, then $$f\alpha(d) = f(a_0) = g(b_0) = g\beta(d),$$ so $(D,\alpha,\beta)$ makes the inner pullback square commute.
Consider some $(D',\alpha',\beta')$ with sources and targets as indicated in \eqref{fou} such that the outer square commutes. Define a map $h:D' \to D$ by $$x \mapsto (\alpha'(x),\beta'(x)).$$ Then $$\alpha h(x) = \alpha(\alpha'(x),\beta'(x)) = \alpha'(x)\text{ and }\beta h(x) = \beta(\alpha'(x),\beta'(x)) = \beta'(x),$$ and so $(D,\alpha,\beta)$ is a pullback. Note that the conditions $\alpha' = \alpha h$ and $\beta' = \beta h$ uniquely determine the definition of $h$.
Before writing the pushout, define $$X = \{(f(c),-g(c))\ |\ c \in C\}.$$ Then the pushout $D = (A \oplus B)/X$.
\end{proof}
\subsubsection{Kernel and Cokernel}
Let $\Cs{C}$ be a pointed category i.e. $\Cs{C}$ has a zero object. The kernel and cokernel are actually special cases of the pullback and pushout, respectively.
\begin{defin}
If one of the legs of a cospan is a zero morphism, then the pullback is called the \textit{kernel} of the remaining morphism. In particular, the kernel of a map $f:A \to C$ is a pair $(\Ker(f),\ker(f))$ such that $$f \circ \ker(f) = 0,$$ and if $\alpha:X \to A$ is a map satisfying $$f \circ \alpha = 0,$$ then there exists a unique map $h:X \to \Ker(f)$ such that the following diagram commutes:
\begin{center}
\begin{tikzcd}
\ar[rdd,bend right,"0"'] X \ar[rd,dashed,"\exists! h"] \ar[rrd,bend left, "\alpha"] && \\
& \ar[d,"0"'] \Ker(f) \ar[r,"\ker(f)"] & A \ar[d,"f"] \\
& 0 \ar[r,"0"'] & C.
\end{tikzcd}
\end{center}
\end{defin}
\begin{defin}
If one of the legs of a span is a zero morphism, then the pushout is called the \textit{cokernel} of the remaining morphism. In particular, the cokernel of a map $f:C \to A$ is a pair $(\Coker(f),\coker(f))$ such that $$\coker(f) \circ f = 0,$$ and if $\alpha:A \to Y$ is a map satisfying $$\alpha \circ f = 0,$$ then there exists a unique map $h:\Coker(f) \to Y$ such that the following diagram commutes:
\begin{center}
\begin{tikzcd}
\ar[d,"0"'] C \ar[r,"f"] & \ar[d,"\coker(f)"'] A \ar[rdd,bend left,"\alpha"] & \\
0 \ar[r,"0"'] \ar[rrd,bend right,"0"'] & \Coker(f) \ar[rd,dashed,"\exists! h"] & \\
& & Y
\end{tikzcd}
\end{center}
\end{defin}
\begin{exam}
In $\Cb{R\text{-}Mod}$, this definition of kernels and cokernels yields the familiar submodules taught in algebra courses. In particular, $\Ker(f)$ is the submodule of $A$ mapped to 0 by $f$, and $\Coker(f)$ is the submodule of $A$ formed by the quotient $A/\Im(f)$. For a proof of these facts, alter the constructions of the pullback and pushout by incorporating the zero module and zero map.
\end{exam}
\subsubsection{Equalizer and Coequalizer}
\begin{defin}
Let $f,g:A \to B$ be a pair of morphisms in $\Cs{C}$. The \textit{equalizer} is a pair $(\ope{Eq}(f,g),\alpha)$ where $\ope{Eq}(f,g) \in \Cs{C}$ and $\alpha:\ope{Eq}(f,g) \to A$ such that $$f\alpha = g\alpha.$$ For every similar pair $(X,\alpha')$, there exists a unique $h:X \to \ope{Eq}(f,g)$ making the following diagram commute:
\begin{center}
\begin{tikzcd}
& \ar[ld,dashed,"\exists! h"'] X \ar[d,"\alpha'"] & \\
\ope{Eq}(f,g) \ar[r,"\alpha"'] & A \ar[r,shift left,"f"] \ar[r, shift right,"g"'] & B.
\end{tikzcd}
\end{center}
\end{defin}
\begin{defin}
Let $f,g:A \to B$ be a pair of parallel morphisms in $\Cs{C}$. The \textit{coequalizer} is a pair $(\ope{Coeq}(f,g),\beta)$ where $\ope{Coeq}(f,g) \in \Cs{C}$ and $\beta:B \to \ope{Coeq}(f,g)$ such that $$\beta f = \beta g.$$ For every similar pair $(Y,\beta')$, there exists a unique $h:\ope{Coeq}(f,g) \to Y$ making the following diagram commute:
\begin{center}
\begin{tikzcd}
A \ar[r,shift left,"f"] \ar[r,shift right,"g"'] & \ar[d,"\beta'"'] B \ar[r,"\beta"] & \ar[ld,dashed,"\exists! h"] \ope{Coeq}(f,g). \\
& Y &
\end{tikzcd}
\end{center}
\end{defin}
We may again retrieve the definitions concisely through the (co)limit construction. Consider the poset $I$ defined as,
\[
I = \{
\begin{tikzcd}
\bullet \ar[r,shift left] \ar[r,shift right] & \bullet
\end{tikzcd}
\}.
\]
Let $F:I \to \Cs{C}$ be a functor whose image is defined by \eqref{three}.
\begin{equation} \label{three}
\begin{tikzcd}
A \ar[r,shift left,"f"] \ar[r,shift right,"g"'] & B
\end{tikzcd}
\end{equation}
Then we obtain $$\ope{Eq}(f,g) = \varprojlim F,$$ $$\ope{Coeq}(f,g) = \varinjlim F.$$ In addition, equalizers and coequalizers may be generalized to families with more than just two parallel morphisms between $A$ and $B$.
\begin{exam}
In $\Cb{R\text{-}Mod}$, there is a simple description of the equalizer and coequalizer. Namely, $$\ope{Eq}(f,g) = \{a \in A\ |\ f(a) = g(a)\} = \Ker(f-g).$$ For the coequalizer, define $$D = \{a \in A\ |\ f(b) = a = g(b)\text{ for some }b \in B\};$$ then $\ope{Coeq}(f,g) = A/D$.
\end{exam}
Now that the most common manifestations of limits and colimits have been treated, it is worthwhile to investigate when they even exist. Recalling the definition of (co)completeness, it would be useful if there was a simpler means for checking whether a category is (co)complete. The following theorem provides an alternate method for checking finite (co)completeness.
\begin{theo} \label{comp}
The following are equivalent for a category $\Cs{C}$:
\begin{enumerate}[label = \emph{(\roman*)}]
\item $\Cs{C}$ is finitely (co)complete;
\item $\Cs{C}$ has a terminal (resp. initial) object and binary pullbacks (resp. pushouts);
\item $\Cs{C}$ has a terminal (resp. initial) object, binary (co)products, and (co)equalizers.
\end{enumerate}
\end{theo}
To prove this theorem, one might proceed to show that the second two criteria imply the first since the first clearly implies everything else. A proof would then assume either of the second two criteria and show that you can construct a (co)limit from an arbitrary finite diagram. When binary pullbacks exist, the unique maps to the terminal object set up the conditions for an initial pullback square. When products and equalizers exist, products are the first step before potentially introducing equalizers in case there are morphisms between the objects in the diagram. The dual approach would proceed similarly.
\newpage
\section{Special Modules and Purity}
\subsection{Projective Modules}
Recall that the covariant Hom functor $(X,-)$ in general is only left exact, so it does not preserve epimorphisms. This may make the reader curious if there is ever a case in which $(X,-)$ is an exact functor. Answering this question distinguishes a certain collection of modules.
\begin{defin} \label{proje}
A left $R$-module $P$ is a \textit{projective module} if for any left modules $M$ and $N$, any epimorphism $f:M \to N$, and any homomorphism $g:P \to N$, there exists a homomorphism $h:P \to M$ making the following diagram commute:
\[
\begin{tikzcd}
& \ar[ld,dashed,"h"'] P \ar[d,"g"] & \\
M \ar[r,"f"'] & N \ar[r] & 0.
\end{tikzcd}
\]
Said differently, a module is projective if every homomorphism into a quotient module $N \cong M/\Ker(f)$ \textit{lifts} to a homomorphism into the base module $M.$
\end{defin}
Upon examination of this definition, it seems evident that it was formulated with the condition of making covariant Hom an exact functor in mind. For example, the bottom row of the diagram is essentially the right portion of an exact sequence. Then surjectivity of the Hom sets follows immediately because of the induced map. This, along with other equivalent formulations, is made more formal in the following theorem.
\begin{theo} \label{proj}
For a left $R$-module $P$, the following statements are equivalent:
\begin{enumerate}[label = \emph{(\roman*)}]
\item $P$ is a projective module;
\item the covariant Hom functor $\Hom{R}{P}{-}$ is exact:
\[
\begin{tikzcd}[cramped,sep = small]
0 \ar[r] & A \ar[r] & B \ar[r] & C \ar[r] & 0
\end{tikzcd} ~~
\Longrightarrow ~~
\begin{tikzcd}[cramped,sep = small]
0 \ar[r] & (P,A) \ar[r] & (P,B) \ar[r] & (P,C) \ar[r] & 0;
\end{tikzcd}
\]
\item any short exact sequence
\[
\begin{tikzcd}[cramped,sep = small]
0 \ar[r] & A \ar[r] & B \ar[r] & P \ar[r] & 0
\end{tikzcd}
\]
splits, i.e. the map $B \to P$ is a retraction so that $B \cong A \oplus P;$
\item there exists a free left $R$-module $F$ and submodule $Q \subseteq F$ such that $F = P \oplus Q;$ informally, every projective module is a direct summand of a free module;
\item there exists a system $(\{e_i\},\{f_i\})_{i \in I}$ of elements $e_i \in P$ and $f_i \in P^* = \Hom{R}{P}{R}$ such that, for any $x \in P,$ only finitely many of the $f_i(x)$ are nonzero and $$x = \sum_i f_i(x)e_i.$$
\end{enumerate}
\end{theo}
\begin{cor}
From the equivalent definitions of projective modules, the following facts follow immediately.
\begin{itemize}
\item Direct summands of projective modules are projective.
\item Direct sums of projective modules are projective.
\item All free modules are projective.
\item Because every module is a quotient of some free module, every module is a quotient of some projective module.
\end{itemize}
\end{cor}
\begin{defin}
The system $(\{e_i\},\{f_i\})_{i \in I}$ mentioned in \thref{proj} is called a \textit{projective basis of $P.$} The $e_i$ are referred to as \textit{basis vectors} and the $f_i$ as \textit{coordinate functions}. Notice that if $P$ is generated by $\{e_i\}$ then $P^* = \Hom{}{P}{R}$ is generated by $\{f_i\}.$ In accordance with general modules, a projective module whose set of basis vectors is finite is \textit{finitely generated}.
\end{defin}
\begin{exam}
We will consider a case where we have a finitely generated module which is not projective as well as a projective module which is not free. The same abelian group will be projective in one case and not in another based on the selection of ring.
\begin{itemize}
\item If we let our ring $R = \Z$, then a free module would be of the form $\Z^n.$ Multiplying by a positive number gives rise to the exact sequence
\[
\begin{tikzcd}
0 \ar[r] & \Z \ar[r,"*5"] & \Z \ar[r] & \Z/5\Z \ar[r] & 0.
\end{tikzcd}
\]
Notice this exact sequence demonstrates that $\Z/5\Z$ is finitely generated. If $\Z/5\Z$ was projective, then this exact sequence would split, which would imply that $\Z \cong \Z \oplus \Z/5\Z,$ but this is clearly not true. As a result, $\Z/5\Z$ is not projective as a $\Z$-module.
\item If we let our ring $R = \Z/10\Z$, then a free module would be of the form $(\Z/10\Z)^n.$ Because $$\Z/5\Z \oplus \Z/2\Z \cong \Z/10\Z,$$ $\Z/5\Z$ is projective as a direct summand of a free module. However, $\Z/5\Z$ cannot be free because it has only 5 elements, where free modules would need to have $10^n$ elements. As such, $\Z/5\Z$ is projective but not free.
\end{itemize}
\end{exam}
The next important theorem makes us of countably generated projective modules, which are defined in much a similar fashion to finitely generated projective modules.
\begin{theo}[Kaplansky]
Every projective module is a direct sum of countably generated projective modules.
\end{theo}
\begin{lem}[Schanuel's Lemma]
Let $R$ be a ring. If $M$ is a module, $P$ and $P'$ are projective modules, and
\[
\begin{tikzcd}
0 \ar[r] & K \ar[r] & P \ar[r] & M \ar[d,equal] \ar[r] & 0 \\
0 \ar[r] & K' \ar[r] & P' \ar[r] & M \ar[r] & 0,
\end{tikzcd}
\]
is a diagram with exact rows. Then $K \oplus P' \cong K' \oplus P.$
\end{lem}
We will now see that a presentation of a projective module indicating that it is finitely generated also indicates that it is finitely presented.
\begin{cor}
Every finitely generated projective module $P$ is finitely presented.
\end{cor}
\begin{proof}
Because $P$ is finitely generated, we denote its basis vector set by $\{e_i\}_{i \leq n}$. We may construct a surjective map $\varphi:R^n \to P$ defined by sending the basis elements $\varepsilon_i \in R^n$ to the basis vectors of $P$ as in $\varepsilon_i \mapsto e_i$. This gives rise to the exact sequence
\[
\begin{tikzcd}
0 \ar[r] & \Ker(\varphi) \ar[r] & R^n \ar[r,"\varphi"] & P \ar[r] & 0.
\end{tikzcd}
\]
Since $P$ is a projective module, this is a split exact sequence. As a result, $R^n \cong \Ker(\varphi) \oplus P$. Because $R^n$ is finitely generated, $\Ker(\varphi)$ must also be finitely generated. In fact, $\Ker(\varphi)$ is projective as a direct summand of a free $R$-module. Therefore, $P$ is finitely presented.
\end{proof}
There is an additional fact about projective modules which is interesting to note. If $P$ is a finitely generated projective right $R$-module, then there exists a square matrix $E = E^2 \in M_n(R)$ such that $$P \cong R^n/ER^n = \Coker(E) = \{x + \Im(E)\ |\ x \in R^n\}.$$
\begin{defin}
Let $M$ be an $R$-module and $K \subseteq M$ be any submodule $K \neq M$. If $N \subseteq M$ is a submodule such that $N \oplus K \neq M$ for all $K$, then $N$ is called \textit{small in $M$} or \textit{superfluous}. We denote this by $N \ll M.$
\end{defin}
\begin{defin}
Let $M$ be a left $R$-module. If there is an exact sequence
\[
\begin{tikzcd}
0 \ar[r] & \Ker(\pi) \ar[r] & P \ar[r,"\pi"] & M \ar[r] & 0,
\end{tikzcd}
\]
where $P$ is a projective $R$-module and $\Ker(\pi) \ll P,$ then the pair $(P,\pi)$ is called a \textit{projective cover of $M.$}
\end{defin}
\begin{defin}
If every finitely generated module over $R$ has a projective cover, then $R$ is called \textit{semiperfect}. If every module $R/aR$ has a projective cover, then $R$ is called \textit{semiregular}.
\end{defin}
\subsection{Injective Modules}
Just as projective modules were motivated by the condition of making $(X,-)$ exact, there is a class of modules which make $(-,X)$ exact. Not only does this definition satisfy the same condition for the dual of $(X,-)$, but the definition itself is exactly the dual of that for projective modules.
\begin{defin}
A left $R$-module $E$ is an \textit{injective module} if for any left modules $A$ and $B$, any monomorphism $f:A \to B$, and any homomorphism $g:A \to E$, there exists a homomorphism $h:B \to E$ making the following diagram commute.
\[
\begin{tikzcd}
0 \ar[r] & A \ar[d,"g"'] \ar[r,"f"] & \ar[ld,dashed,"h"] B \\
& E &
\end{tikzcd}
\]
Said differently, a module $E$ is injective if every homomorphism from a submodule to $E$ can be \textit{extended} to a homomorphism from the whole module.
\end{defin}
It should be easy to see that this does indeed make $(-,X)$ exact by reasoning dual to that used for projective modules. Injective modules also have equivalent formulations which are important.
\begin{theo}
For a left $R$-module $E$, the following statements are equivalent:
\begin{enumerate}[label = \emph{(\roman*)}]
\item $E$ is an injective module;
\item the contravariant Hom functor $\Hom{R}{-}{E}$ is exact:
\[
\begin{tikzcd}[cramped,sep = small]
0 \ar[r] & A \ar[r] & B \ar[r] & C \ar[r] & 0
\end{tikzcd} ~~
\Longrightarrow ~~
\begin{tikzcd}[cramped,sep = small]
0 \ar[r] & (C,E) \ar[r] & (B,E) \ar[r] & (A,E) \ar[r] & 0;
\end{tikzcd}
\]
\item any short exact sequence
\[
\begin{tikzcd}[cramped,sep = small]
0 \ar[r] & E \ar[r] & A \ar[r] & B \ar[r] & 0
\end{tikzcd}
\]
splits, i.e. the map $E \to A$ is a section so that $A \cong E \oplus B;$
\item any short exact sequence
\[
\begin{tikzcd}[cramped,sep = small]
0 \ar[r] & E \ar[r] & A \ar[r] & C \ar[r] & 0,
\end{tikzcd}
\]
where $C$ is a cyclic module, splits;
\item Baer Criterion: for any left ideal $I$ of $R$ and any homomorphism $f: I \to E$, there exists a homomorphism $\hat{f}:R \to E$ making the following diagram commute.
\[
\begin{tikzcd}
0 \ar[r] & I \ar[d,"f"'] \ar[r] & \ar[ld,dashed,"\hat{f}"] R \\
& E &
\end{tikzcd}
\]
\end{enumerate}
\end{theo}
One of the examples of injective modules requires the following important definition.
\begin{defin}
Let $R$ be an integral domain. Consider the equivalence relation $\sim$ on $R \times R$ defined by
\[
(n,m) \sim (x,y) \text{ iff } ny = mx.
\]
Then the \textit{field of fractions} $\ope{Frac}(R)$ is
\[
\ope{Frac}(R) := R \times R/\sim.
\]
Informally the ordered pairs represent fractions, making the equivalence relation equate fractions which are considered equivalent by cross multiplication.
\end{defin}
\begin{exam}
The following are injective modules:
\begin{itemize}
\item $\Q$ as a $\Z$-module,
\item $\Q/\Z$ as a $\Z$-module,
\item $\ope{Frac}(R)$ as an $R$-module, when $R$ is an integral domain,
\item any vector space over the field of fractions $\ope{Frac}(R)$ of an integral domain $R$ when viewed as an $R$-module.
\end{itemize}
\end{exam}
\begin{rem}
There are some useful properties enjoyed by injective modules similar to those for projective modules, which in many cases are dual those for projective modules.
\begin{itemize}
\item Arbitrary direct products of injective modules are injective.
\item Finite direct sums of injective modules are injective.
\item Injective modules are direct summands of all modules containing them.
\item Direct summands of injective modules are injective.
\item Every module is a submodule of some injective module.
\end{itemize}
\end{rem}
The dual of Schanuel's Lemma also holds.
\begin{lem}[Schanuel's Lemma]
Let $R$ be a ring. If $M$ is a module, $E$ and $E'$ are injective modules, and the rows in the diagram
\[
\begin{tikzcd}
0 \ar[r] & M \ar[d,equal] \ar[r] & E \ar[r] & L \ar[r] & 0 \\
0 \ar[r] & M \ar[r] & E' \ar[r] & L' \ar[r] & 0
\end{tikzcd}
\]
are exact. Then $E \oplus L' \cong E' \oplus L.$
\end{lem}
And unsurprisingly, there is a dual notion to projective covers.
\begin{defin}
Let $M$ be a nonzero $R$-module. A nonzero submodule $N \subseteq M$ is said to be \textit{essential in $M$} (denoted as $N \triangleleft M$) if it has nonzero intersection with each nonzero submodule of $M.$
\end{defin}
\begin{defin}
Let $M$ and $E$ be $R$-modules and $\alpha:M \to E$ be a homomorphism such that $\Im(\alpha) \triangleleft E.$ Then we say that $E$ is an \textit{essential extension of $M.$} If $E$ is also injective, then $E$ is called an \textit{injective envelope of $M.$} We denote this by $\ope{Env}(M).$
\end{defin}
\begin{theo}[Eckmann-Sch\"opf]
Let $M$ be a left $R$-module. Then $\ope{Env}(M)$ always exists. Additionally if $E,E'$ are two injective envelopes of $M,$ then there exists an isomorphism $\eta:E \to E'$ for which $\eta|_M = \mathds{1}_M.$
\end{theo}
\subsection{Flat Modules}
Projective and injective modules were discovered in trying to find modules which made Hom functors exact. We saw earlier that the tensor product functor is right exact but not left exact, so it is natural to wonder if there exists some class of modules for which tensor products preserve exactness. This leads us to the following definition.
\begin{defin}
A right $R$-module $A$ is called \textit{flat} if the exactness of the sequence of left $R$-modules
\[
\begin{tikzcd}
0 \ar[r] & X \ar[r] & Y \ar[r] & Z \ar[r] & 0
\end{tikzcd}
\]
implies the exactness of the sequence
\[
\begin{tikzcd}
0 \ar[r] & A \otimes_R X \ar[r] & A \otimes_R Y \ar[r] & A \otimes_R Z \ar[r] & 0.
\end{tikzcd}
\]
\end{defin}
\begin{theo}
For a right $R$-module $A$, the following are equivalent:
\begin{enumerate}[label = \emph{(\roman*)}]
\item $A$ is a flat module;
\item $A \otimes_R -$ is an exact functor;
\item
\begin{tikzcd}[cramped]
A \otimes_R I \ar[r,"\mathds{1}_A \otimes i"] & A \otimes_R R
\end{tikzcd}
is a monomorphism for any left ideal $I$ of $R;$
\item
\begin{tikzcd}[cramped]
A \otimes_R I \ar[r,"\mathds{1}_A \otimes i"] & A \otimes_R R
\end{tikzcd}
is a monomorphism for any finitely generated left ideal $I$ of $R;$
\item for any $n$-tuple $a^T = (a_1,a_2,\ldots,a_n)$ of elements in $A$ and any $X \in M_{n,d}(R)$ such that $$a^T X = 0^T,$$ there exists an $m$-tuple $\alpha^T = (\alpha_1,\alpha_2,\ldots,\alpha_m)$ of elements of $A$ and $Y \in M_{m,n}(R)$ such that
\[
\alpha^T Y = a^T, ~~ YX = 0.
\]
\end{enumerate}
\end{theo}
It should be clear that $R$ as a right $R$-module is flat because tensoring over $R$ with $R$ produces an isomorphic right $R$-module. While arbitrary direct sums of projective modules are projective if and only if each summand is projective, there is a similar statement for flat modules.
\begin{theo} \label{fla}
A direct sum $M = \bigoplus_{k \in K} M_k$ of right $R$-modules is flat if and only if each $M_k$ is flat.
\end{theo}
\begin{cor}
Every projective right $R$-module is flat.
\end{cor}
\begin{proof}
Any projective right $R$-module $P$ is isomorphic to a direct summand of a free $R$-module $R^{(I)}$ for $I$ some set. Because $R$ is flat, $R^{(I)}$ must be flat by \thref{fla}. But $R^{(I)}$ can only be flat if all its direct summands is flat, so $P$ must also be flat.
\end{proof}
The reader may wonder whether the converse is true. The following example shows that it is not.
\begin{exam}
If we let our ring be $R = \Z$, then we know that tensor products with $\Q$ over $\Z$ are trivial. For example, tensoring any exact sequence over $\Z$ by $\Q/\Z$ will annihilate the entire sequence, which is trivial case of an exact sequence. Consequently, $\Q/\Z$ is a flat $\Z$-module. Because it is a quotient, there is an obvious short exact sequence that we may form, namely
\[
\begin{tikzcd}
0 \ar[r] & \Z \ar[r] & \Q \ar[r] & \Q/\Z \ar[r] & 0.
\end{tikzcd}
\]
If $\Q/\Z$ was projective, then we would have that $$\Q/\Z \oplus \Z \cong \Q.$$ But because every nonzero element of $\Q/\Z$ has finite order, $(q,0) \in \Q/\Z \oplus \Z$ would have finite order whenever $q \neq 0.$ But $\Q$ has no elements of finite order, so $\Q/\Z$ cannot be projective.
\end{exam}
\begin{theo}
Let $M$ be a right $R$-module. If every finitely generated submodule of $M$ is flat, then $M$ is flat.
\end{theo}
\begin{theo}
Let $M$ be a finitely generated right $R$-module. Then $M$ is projective if and only if $M$ is finitely presented and flat.
\end{theo}
There are also connections between flat and injective modules. Unsurprisingly, this comes out of taking a sort of dual.
\begin{defin}
If $M$ is a right $R$-module, we define its \textit{character module} $M^+$ as the left $R$-module $$M^+ = \Hom{\Z}{M}{\Q/\Z}.$$
\end{defin}
\begin{lem}
A sequence of right $R$-modules
\[
\begin{tikzcd}
A \ar[r,"f"] & B \ar[r,"g"] & C
\end{tikzcd}
\]
is exact if and only if the induced sequence of character modules
\[
\begin{tikzcd}
C^+ \ar[r,"g^+"] & B^+ \ar[r,"f^+"] & A^+
\end{tikzcd}
\]
is also exact.
\end{lem}
\begin{theo}[Lambek]
A right $R$-module $M$ is flat if and only if its character module $M^+$ is an injective left $R$-module.
\end{theo}
\subsection{Pure Exact Sequences}
In looking at projective, injective, and flat modules, our motivation was to make certain functors exact. However, we could ask the reversed question of what short exact sequences make the functors exact always. In particular, there is a class of exact sequences which make $X \otimes_R -$ exact for all right $R$-modules $X.$
\begin{defin}
A short exact sequence
\[
\begin{tikzcd}
0 \ar[r] & A \ar[r,"f"] & B \ar[r] & C \ar[r] & 0
\end{tikzcd}
\]
of left $R$ modules is \textit{pure exact} if
\[
\begin{tikzcd}
0 \ar[r] & X \otimes_R A \ar[r,"\mathds{1}_X \otimes f"] & X \otimes_R B \ar[r] & X \otimes_R C \ar[r] & 0
\end{tikzcd}
\]
is exact for all right $R$-modules $X.$ In this case, we call $f(A) \subseteq B$ a \textit{pure submodule}.
\end{defin}
The following theorem formalizes the relationship between purity and flatness.
\begin{theo}
A left $R$-module $Z$ is flat if and only if every exact sequence
\[
\begin{tikzcd}
0 \ar[r] & X \ar[r] & Y \ar[r] & Z \ar[r] & 0
\end{tikzcd}
\]
of left $R$-modules is pure exact.
\end{theo}
\begin{theo}
Every split short exact sequence is pure exact.
\end{theo}
\begin{proof}
Let
\[
\begin{tikzcd}
0 \ar[r] & A \ar[r,"i"] & B \ar[r,"p"] & C \ar[r] & 0
\end{tikzcd}
\]
be a split short exact sequence. Then there exists a map $q:B \to A$ such that
\[
qi = \mathds{1}_A.
\]
Applying the functor $M \otimes_R -$ to the short exact sequence, we obtain
\[
\begin{tikzcd}
M \otimes_R A \ar[r,"\mathds{1}_M \otimes i"] & M \otimes_R B \ar[r,"\mathds{1}_M \otimes p"] & M \otimes_R C \ar[r] & 0.
\end{tikzcd}
\]
The composition
\[
(\mathds{1}_M \otimes q)(\mathds{1}_M \otimes i) = \mathds{1}_M \otimes qi = \mathds{1}_M \otimes \mathds{1}_A = \mathds{1}_{M \otimes_R A}
\]
shows that $\mathds{1}_M \otimes i$ is a section, so it is a monomorphism as the right divisor of an injective map (the identity). Then the tensored exact sequence is also exact, so the original exact sequence is pure.
\end{proof}
And the following theorem gives a condition for a submodule to be pure.
\begin{theo}
Let $f:X \to Y$ be an injection of left $R$-modules. Then $f(X)$ is a pure submodule of $Y$ if and only if, given any commutative diagram with $F_0$ and $F_1$ finitely generated free $R$-modules, there is a map $F_1 \to X$ making the upper triangle commute.
\[
\begin{tikzcd}
& \ar[d] F_0 \ar[r] & \ar[d] \ar[dl,dashed] F_1 \\
0 \ar[r] & X \ar[r,"f"] & Y
\end{tikzcd}
\]
\end{theo}
The following lemma ties in exactness of $(X,-)$ with purity.
\begin{lem}
Let
\[
\begin{tikzcd}
0 \ar[r] & X \ar[r,"\iota"] & Y \ar[r,"\pi"] & Z \ar[r] & 0
\end{tikzcd}
\]
be a pure exact sequence. From exactness, $Z \cong Y/X,$ so $\pi:Y \to Z$ is the canonical projection. If $M$ is a finitely presented left $R$-module, then $\pi_*:(M,Y) \to (M,Z)$ is surjective.
\end{lem}
\begin{rem}
This lemma tells us that $(M,-)$ is an exact functor on pure exact sequences when $M$ is finitely presented.
\end{rem}
\begin{cor}
Let
\[
\begin{tikzcd}
0 \ar[r] & X \ar[r] & Y \ar[r] & Z \ar[r] & 0
\end{tikzcd}
\]
be an exact sequence. If $Z$ is finitely presented, then this sequence is pure exact if and only if it is split.
\end{cor}
And lastly, we have a more lax condition for checking whether a sequence is pure exact.
\begin{theo}
An exact sequence of left $R$-modules
\[
\begin{tikzcd}
0 \ar[r] & X \ar[r] & Y \ar[r] & Z \ar[r] & 0
\end{tikzcd}
\]
is pure exact if and only if it remains exact after tensoring by any finitely presented right $R$-module.
\end{theo}
\subsection{Torsion Product Functor}
\begin{defin}
A pointed category $\Cs{C}$ satisfying the conditions:
\begin{itemize}
\item $\Hom{}{A}{B}$ is an abelian group for all $A,B \in \Cs{C},$
\item given morphisms in $\Cs{C}$
\[
\begin{tikzcd}
A \ar[r,"a"] & B \ar[r,shift left,"b"] \ar[r,shift right,"c"'] & C \ar[r,"d"] & D,
\end{tikzcd}
\]
the distributive laws hold:
\[
d(b+c) = db+dc \quad (b+c)a = ba+ca,
\]
\item $\Cs{C}$ admits all finite products and coproducts,
\end{itemize}
is called an \textit{additive category}. A covariant or contravariant functor $F:\Cs{C} \to \Cs{D}$ between additive categories which respects the additive structure of Hom sets is called an \textit{additive functor} (cf. definition \ref{adf}).
\end{defin}
\begin{rem}
In additive categories, finite products and coproducts coincide (up to isomorphism), so they are both written as the direct sum $\oplus.$ Conveniently, the direct sum has projection and injection maps along with conditions on their compositions in all additive categories, just as discussed in depth with modules. In addition, additive functors preserve direct sums.
\end{rem}
\begin{defin}
An additive category $\Cs{C}$ is an \textit{abelian category} if
\begin{itemize}
\item all morphisms have kernels and cokernels,
\item every monomorphism is a kernel and every epimorphism is a cokernel.
\end{itemize}
\end{defin}
\begin{rem}
By applying \thref{comp}, we find that all abelian categories are in fact both finitely complete and cocomplete, i.e. all finite limits and colimits exist. This makes abelian categories very convenient.
\end{rem}
\begin{exam}
Here are some examples (and non-examples) of abelian categories.
\begin{itemize}
\item For any ring $R,$ both $\Cb{R\text{-}Mod}$ and $\Cb{Mod\text{-}R}$ are abelian categories. Notably, $\Cb{\Z\text{-}Mod} = \Cb{Ab}$ is the prototypical example of an abelian category.
\item The full subcategory (i.e. restricted objects with preserved Hom sets) of $\Cb{Ab}$ consisting of all finitely generated abelian groups is abelian along with the full subcategory of all torsion abelian groups.
\item The categories $\Cb{Grp}$ and $\Cb{ComRing}$ of groups and commutative rings, respectively, are both not abelian categories; in fact, neither of them are even additive.
\item The full subcategory of $\Cb{Ab}$ consisting of all torsion-free abelian groups is additive but not abelian because some maps do not have cokernels, e.g. the inclusion $2\Z \hookrightarrow \Z$ has Cokernel $\Z/2\Z,$ which is not torsion-free.
\end{itemize}
\end{exam}
This field of study heavily involves both homological and homotopical algebra, both of which grew out of algebraic topology. For the time being, we will focus on the homological aspect, which assigned a sequence of free abelian groups to a topological space that encoded specific information about the space. More generally, the sequence may be composed of free $R$-modules with little to no alteration to the subsequent theory. The goal of homological algebra in this respect was to identify and study functors that could extract invariants from these sequences, which could then be associated to a topological space and its characteristics.
The motivation for mentioning abelian categories prior to this lies in the fact that abelian categories are modeled after $\Cb{Ab},$ and in fact the Freyd-Mitchell embedding theorem asserts that any small abelian category is a full subcategory of $\Cb{R\text{-}Mod}$ for some ring $R,$ somewhat resembling the Cayley theorem of group theory which asserts that any group may be viewed as a group of symmetries. In other words, we may think of objects and morphisms in abelian categories as $R$-modules and $R$-homomorphisms. Consequently, we may study the objects in any small abelian category using the methods of homological algebra as long as we have a means of encoding information in similar sequences, which leads us to now present these ideas more rigorously.
\begin{defin}
Let $\Cs{C}$ be an additive category. A sequence $(C_i,d_i)_{i \in \Z}$ of objects and morphisms in $\Cs{C}$ (called \textit{boundary maps})
\[
(\Cb{C},d):
\begin{tikzcd}
\cdots \ar[r] & C_{i+1} \ar[r,"d_{i+1}"] & C_{i} \ar[r,"d_{i}"] & C_{i} \ar[r] & \cdots
\end{tikzcd}
\]
is called a \textit{chain complex} if
\[
\Im(d_{i+1}) \subseteq \Ker(d_i)
\]
for all $i \in \Z.$ This is equivalent to saying that $d_id_{i+1} = 0$ for all $i \in \Z.$ If the boundary maps are obvious, we may notate the complex by just $\Cb{C}.$ The category of all chain complexes in $\Cs{C}$ is $\Cb{Comp}(\Cs{C}).$ Elements of $C_n$ are called \textit{$n$-chains}; elements of $\Ker(d_n)$ are called \textit{$n$-cycles}, and elements of $\Im(d_{n+1})$ are called \textit{$n$-boundaries}.
\end{defin}
\begin{prop}
If $\Cs{C}$ is an abelian category, then $\Cb{Comp}(\Cs{C})$ is also an abelian category.
\end{prop}
\begin{rem}
In algebraic topology, these chain complexes arise naturally from building free abelian groups from bases composed of continuous images of $n$-simplices. Finite linear combinations of simplices then may be interpreted as compiling a chain of simplices to traverse with particular orientations on each. The boundary maps are defined so that they produce lower dimensional chains which represent the boundaries of anything input into them. Most importantly, a chain which mapped to 0 through the boundary map had no boundary because it had formed a closed loop, or cycle. The condition that boundary maps compose to the zero map then has the topological significance that an $(n+1)$-simplex has a boundary which is a closed cycle and thus has no boundary of its own.
\end{rem}
\begin{defin}
Let $(\Cb{C},d)$ and $(\Cb{C}',d')$ be chain complexes in $\Cs{C}.$ A family of morphisms $f = \{f_i|C_i \to C'_i\}_{i \in \Z}$ satisfying the relation
\[
f_{i-1}d_i = d'_if_i,
\]
which is to say that the maps commute with the boundary maps, is called a \textit{chain map}. Such a collection of maps constitutes an element $f \in \Mor{\Cb{Comp}(\Cs{C})}{\Cb{C}}{\Cb{C}'}.$
\end{defin}
\begin{rem}
Because we may assign chain complexes to topological spaces and because topological spaces can have continuous maps between them, it is of interest how these maps manifest on the level of chain complexes. The continuous map induces a family of homomorphisms due to the continuous mapping of simplices from one space to another, and these maps commute with the boundary maps. As such, the motivation for the definition of chain maps also derives from algebraic topology.
\end{rem}
\begin{defin}
Let $(\Cb{C},d)$ be a chain complex in $\Cs{C}.$ The quotient module
\[
H_n(\Cb{C}) := \frac{\Ker(d_n)}{\Im(d_{n+1})} := \frac{Z_n}{B_n}
\]
is called the \textit{$n$-th homology group} of $\Cb{C}.$
\end{defin}
\begin{prop}
If $\Cs{C}$ is an abelian category, then $H_n:\Cb{Comp}(\Cs{C}) \to \Cs{C}$ is an additive functor for all $n \in \Z.$ In particular, if $f:(\Cb{C},d) \to (\Cb{C}',d')$ is a chain map, define
\[
H_n(f) = f_* :=
\left\{
\begin{array}{rcl}
H_n(\Cb{C}) & \longrightarrow & H_n(\Cb{C}') \\
z_n+B_n & \longmapsto & f_nz_n+B'_n
\end{array}.
\right.
\]
\end{prop}
\begin{rem}
Topologically, homology groups indicate when there is a cycle which is not itself a boundary of a higher dimensional simplex, which is in essence a hole in the topological space. Algebraically, homology groups are a measure of deviation from exactness. Notice that if the complex were in fact an exact sequence, then all the homology groups would be trivial.
\end{rem}
\begin{theo}[Long Exact Sequence in Homology]
Let $\Cs{C}$ be an abelian category. If
\[
\begin{tikzcd}
0 \ar[r] & \Cb{C}_1 \ar[r,"\iota"] & \Cb{C}_2 \ar[r,"\pi"] & \Cb{C}_3 \ar[r] & 0
\end{tikzcd}
\]
is a short exact sequence in $\Cb{Comp}(\Cs{C}),$ then there is an exact sequence in $\Cs{C}$
\[
\begin{tikzcd}
\cdots \ar[r] & H_{n+1}(\Cb{C}_3) \ar[r,"\partial_{n+1}"] & H_n(\Cb{C}_1) \ar[r,"\iota_*"] & H_n(\Cb{C}_2) \ar[r,"\pi_*"] & H_n(\Cb{C}_3) \ar[r,"\partial_n"] & H_{n-1}(\Cb{C}_1) \ar[r] & \cdots.
\end{tikzcd}
\]
The morphisms $\partial$ are called \textit{connecting morphisms}.
\end{theo}
If there isn't enough information using just the definition of homology to carry out computations but the object of study has morphisms with better-understood objects (such as a quotient object and the corresponding Kernel), then the long exact sequence becomes an indispensable tool in completing computations. It is also useful to note that the connecting morphisms are natural with respect to complexes, which we state precisely now.
\begin{theo}
Let $\Cs{C}$ be an abelian category. Given a commutative diagram in $\Cb{Comp}(\Cs{C})$ with exact rows,
\[
\begin{tikzcd}
0 \ar[r] & \ar[d,"f"] \Cb{C}_1 \ar[r,"\iota"] & \ar[d,"g"] \Cb{C}_2 \ar[r,"\pi"] & \ar[d,"h"] \Cb{C}_3 \ar[r] & 0 \\
0 \ar[r] & \Cb{A}_1 \ar[r,"\iota'"] & \Cb{A}_2 \ar[r,"\pi'"] & \Cb{A}_3 \ar[r] & 0,
\end{tikzcd}
\]
there is a commutative diagram in $\Cs{C}$ with exact rows,
\[
\begin{tikzcd}
\cdots \ar[r] & \ar[d,"h_*"] H_{n+1}(\Cb{C}_3) \ar[r,"\partial_{n+1}"] & \ar[d,"f_*"] H_n(\Cb{C}_1) \ar[r,"\iota_*"] & \ar[d,"g_*"] H_n(\Cb{C}_2) \ar[r,"\pi_*"] & \ar[d,"h_8"] H_n(\Cb{C}_3) \ar[r,"\partial_n"] & \ar[d,"f_*"] H_{n-1}(\Cb{C}_1) \ar[r] & \cdots \\
\cdots \ar[r] & H_{n+1}(\Cb{A}_3) \ar[r,"\partial'_{n+1}"] & H_n(\Cb{A}_1) \ar[r,"\iota'_*"] & H_n(\Cb{A}_2) \ar[r,"\pi'_*"] & H_n(\Cb{A}_3) \ar[r,"\partial'_n"] & H_{n-1}(\Cb{A}_1) \ar[r] & \cdots
\end{tikzcd}
\]
\end{theo}
Now with the necessary terminology, we may go on to presenting a particularly useful tool for studying objects in abelian categories.
\begin{defin}
Let $\Cs{C}$ be an abelian category and $C \in \Cs{C}$ be any object. An exact sequence of the form
\[
(\Cb{P},d):
\begin{tikzcd}
\cdots \ar[r] & P_2 \ar[r,"d_2"] & P_1 \ar[r,"d_1"] & P_0 \ar[r,"\varepsilon"] & C \ar[r] & 0,
\end{tikzcd}
\]
where each $P_n$ is projective (cf. definition \ref{proje}), is called a \textit{projective resolution} of $C.$ The map $\varepsilon:P_0 \to C$ is called an \textit{augmentation}. Given a projective resolution $\Cb{P}$ of $C,$ the complex obtained by removing the augmentation
\[
(\Cb{P}_A,d):
\begin{tikzcd}
\cdots \ar[r] & P_2 \ar[r,"d_2"] & P_1 \ar[r,"d_1"] & P_0 \ar[r] & 0
\end{tikzcd}
\]
is called a \textit{deleted projective resolution} of $C.$ When $\Cs{C} = \Cb{R\text{-}Mod},$ strengthening the condition that each $P_n$ is projective by requiring that each $P_n$ be free or flat results in \textit{free resolutions} and \textit{flat resolutions} of $C,$ respectively.
\end{defin}
\begin{rem}
No information is lost in removing the augmentation because $C \cong \Coker(d_1)$ and $\varepsilon = \coker(d_1).$ We may think of resolutions as enhanced presentations, which contributes intuition to the proof of the following proposition.
\end{rem}
\begin{prop}
Every $R$-module has a free resolution, which is necessarily a projective resolution and a flat resolution.
\end{prop}
\begin{proof}
Let $C$ be an $R$-module. According to \thref{res}, there exists a free $R$-module $F_0$ and a short exact sequence
\[
\begin{tikzcd}
0 \ar[r] & \Ker(\pi_0) \ar[r,hook,"\iota_0"] & F_0 \ar[r,two heads,"\pi_0"] & M \ar[r] & 0.
\end{tikzcd}
\]
Additionally, there is a free $R$-module $F_1$ and a short exact sequence
\[
\begin{tikzcd}
0 \ar[r] & \Ker(\pi_1) \ar[r,hook,"\iota_1"] & F_1 \ar[r,two heads,"\pi_1"] & \Ker(\pi_0) \ar[r] & 0.
\end{tikzcd}
\]
Clearly $\Ker(\iota_0\pi_1) = \Im(\iota_1),$ so we may join the short exact sequences to form an exact sequence
\[
\begin{tikzcd}
0 \ar[r] & \Ker(\pi_1) \ar[r,hook,"\iota_1"] & F_1 \ar[r,"\iota_0\pi_1"] & F_0 \ar[r,"\pi_0"] & M \ar[r] & 0.
\end{tikzcd}
\]
From here, we may create another short exact sequence for $\Ker(\pi_1)$ and adjoin it in a similar way, and then continue to repeat the process indefinitely. To see that other free resolutions exist, we may instead use presentations of modules and adjoin those (albeit with a bit more work to preserve exactness).
\end{proof}
\begin{defin}
A category $\Cs{C}$ has \textit{enough projectives} if, for each $X \in \Cs{C},$ there exists a projective object $P$ and an epimorphism $P \twoheadrightarrow X.$
\end{defin}
\begin{cor}
If $\Cs{C}$ is an abelian category with enough projectives, then every $X \in \Cs{C}$ has a projective resolution.
\end{cor}
\begin{proof}
Copy the prior proof but replace free modules with projective objects guaranteed by the hypothesis that $\Cs{C}$ has enough projectives.
\end{proof}
For this final part of this section, we will return to $\Cb{R\text{-}Mod}$ because we will need tensor products. This does mean that abelian categories with an analogue to tensor products can make use of the upcoming notions and definitions.
\begin{defin}
Let $A$ and $B$ be right and left $R$-modules, respectively. Given a particular projective resolution $(\Cb{P},d)$ of $A,$ define
\[
\Tor{n}{R}{A}{B} := H_n(\Cb{P}_A \otimes_R B) := \frac{\Ker(d_n \otimes \mathds{1}_B)}{\Im(d_{n+1} \otimes \mathds{1}_B)}.
\]
\end{defin}
\begin{rem}
Although Tor is defined with reference to a particular projective resolution, we will arrive at the same computation for Tor regardless of what projective resolution is used. In particular, this means that flat resolutions may be used since they are a specific type of projective resolution. Furthermore, fixing $A$ and computing Tor with a resolution of $B$ coincides with the definition.
\end{rem}
\begin{cor}
If $A$ and $B$ be right and left $R$-modules, then
\[
\Tor{0}{R}{A}{B} \cong A \otimes_R B.
\]
\end{cor}
\begin{exam}
Here are a couple of sample computations of torsion products.
\item $\Tor{1}{R}{R/I}{R/J} \cong \frac{I \cap J}{IJ}.$
\item $\Tor{1}{R}{R/aR}{R/Rb} \cong \frac{aR \cap bR}{aRb}.$
\end{exam}
\begin{theo}
If a right $R$-module $F$ is flat, then $\Tor{n}{R}{F}{M} = 0$ for all left $R$-modules $M$ and for all $n \geq 1.$ Conversely, if $\Tor{1}{R}{F}{M} = 0$ for all $M,$ then $F$ is flat.
\end{theo}
\begin{theo}
If $A$ and $B$ are right and left $R$-modules, respectively, then
\[
\Tor{n}{R}{A}{B} \cong \Tor{n}{R^{op}}{B}{A}
\]
for all $n \geq 0.$ Consequently, if $R$ is a commutative ring, then
\[
\Tor{n}{R}{A}{B} \cong \Tor{n}{R}{B}{A}
\]
for all $n \geq 0.$
\end{theo}
Just as with homology, there is a long exact sequence for Tor as well.
\begin{theo}
If
\[
\begin{tikzcd}
0 \ar[r] & A_1 \ar[r] & A_2 \ar[r] & A_3 \ar[r] & 0
\end{tikzcd}
\]
is a short exact sequence of right $R$-modules, then there is an exact sequence for every left $R$-module $B.$
\[
\begin{tikzcd}[cramped,sep = small]
\cdots \ar[r] & \Tor{n+1}{R}{A_3}{B} \ar[r] & \Tor{n}{R}{A_1}{B} \ar[r] & \Tor{n}{R}{A_2}{B} \ar[r] & \Tor{n}{R}{A_3}{B} \ar[r] & \Tor{n-1}{R}{A_1}{B} \ar[r] & \cdots.
\end{tikzcd}
\]
\end{theo}
Additionally, Tor preserves some important constructions.
\begin{prop}
If $\{B_k\}_{k \in K}$ is a family of left $R$-modules then there are natural isomorphisms for all $n \geq 0,$
\[
\Tor{n}{R}{A}{\bigoplus_{k \in K} B_k} \cong \bigoplus_{k \in K}\Tor{n}{R}{A}{B_k}.
\]
The same again holds if the positions of $A$ and $B$ are swapped.
\end{prop}
\begin{prop}
If $\{B_i,\varphi_j^i\}$ is a direct system of left $R$-modules over a directed index set $I,$ then there is an isomorphism for all $n \geq 0,$
\[
\Tor{n}{R}{A}{\varinjlim B_k} \cong \varinjlim \Tor{n}{R}{A}{B_k}.
\]
The same again holds if the roles of $A$ and $B$ are swapped.
\end{prop}
Now let us observe what happens when the coefficient ring is actually a domain.
\begin{lem}
Let $R$ be a domain.
\begin{itemize}
\item If $A$ is a right torsion $R$-module, then $\Tor{1}{R}{\ope{Frac}(R)/R}{A} \cong A.$
\item For every right $R$-module $A,$ $\Tor{n}{R}{\ope{Frac}(R)/R}{A} = 0$ for all $n \geq 2.$
\item If $A$ is a right torsion-free $R$-module, then $\Tor{1}{R}{\ope{Frac}(R)/R}{A} \cong 0.$
\end{itemize}
\end{lem}
\begin{theo}
If $R$ is a domain, then $\Tor{n}{R}{A}{B}$ is a torsion module for all $A,B$ and all $n \geq 1.$
\end{theo}
Finally, we may approach Tor through an axiomatic framework so that, instead of constructing it, we can show that other constructions coincide (or differ) from it based on whether the axioms are met. Towards the end of this document, the primary result by Robinson relies on these axioms to show that he constructed something which was isomorphic to Tor.
\begin{theo}
Let $(T_n:\Cb{R\text{-}Mod} \to \Cb{Ab})_{n \geq 0}$ be a sequence of additive covariant functors. If
\begin{itemize}
\item for every short exact sequence
\[
\begin{tikzcd}
0 \ar[r] & A \ar[r] & B \ar[r] & C \ar[r] & 0
\end{tikzcd}
\]
of right $R$-modules, there is a long exact sequence with natural connecting homomorphisms
\[
\begin{tikzcd}
\cdots \ar[r] & T_{n+1}(C) \ar[r,"\Delta_{n+1}"] & T_n(A) \ar[r] & T_n(B) \ar[r] & T_n(C) \ar[r,"\Delta_n"] & T_{n-1}(A) \ar[r] & \cdots,
\end{tikzcd}
\]
\item $T_0$ is naturally isomorphic to $- \otimes_R M$ for some left $R$-module $M,$
\item $T_n(P) = 0$ for all projective right $R$-modules $P$ and all $n \geq 1,$
\end{itemize}
then $T_n$ is naturally isomorphic to $\Tor{n}{R}{-}{M}$ for all $n \geq 0.$
\end{theo}
\newpage
\section{Homotopy Theory}
\subsection{Compactly Generated Spaces}
The goal of this section is to construct a category of topological spaces such that, for any $X \in \Cb{Top},$ a functor $(- \times X)$ has right adjoint, i.e. the space $C(X,Y)$ has a ``nice'' topology.
\begin{defin}
For $X,Y \in \Cb{Top},$ a function $f:X \to Y$ between their underlying sets is called \textit{$k$-continuous} if, for all compact Hausdorff spaces $C$ and all continuous maps $t:C \to X,$ the composition $tf$ is continuous.
\end{defin}
\begin{defin}
A topological space $X$ is \textit{weak Hausdorff} if all continuous images of compact Hausdorff spaces $C$ in
$X$ are closed in $X.$ Succinctly,
\[
f:C \to X \text{ continuous } \Longrightarrow \Im(f) \subseteq X \text{ closed.}
\]
\end{defin}
\begin{theo} \label{ks}
Let $X$ be a topological space. Then the following are equivalent:
\begin{enumerate}[label = \emph{(\roman*)}]
\item for all spaces $Y$ and functions $f:X \to Y,$ $f$ is continuous if and only if $f$ is $k$-continuous;
\item the previous statement holds for some family of compact Hausdorff spaces;
\item $X$ is a quotient space of some disjoint union of compact Hausdorff spaces;
\item any continuous map $g:C \to X$ from a compact Hausdorff space $C$ is open;
\item $A \subseteq X$ is closed if and only if $A \cap K \subseteq K$ is closed for any compact $K \subseteq X.$
\end{enumerate}
\end{theo}
\begin{defin}
A space $X$ satisfying any of the equivalent conditions of \thref{ks} is called a \textit{$k$-space}. If $X$ is also weak Hausdorff, then we say that $X$ is \textit{compactly generated}.
\end{defin}
\begin{rem}
The continuous image of a compact Hausdorff space $C$ inside a compactly generated space $X$ must be clopen (i.e. closed and open), so the existence of a non-surjective continuous map $g:C \to X$ implies that $X$ has more than one connected component. With regard to the separation axioms, $T_2 \Longrightarrow \text{weak Hausdorff} \Longrightarrow T_1.$ Recall that $T_2$ spaces are identically Hausdorff and that $T_1$ spaces are defined by the property that any pair of points $x,y$ have open sets $U,V$ such that $x \in U,$ $y \in V,$ $x \notin V,$ and $y \notin U.$
\end{rem}
\begin{exam}
Some examples of compactly generated spaces include
\begin{itemize}
\item compact Hausdorff spaces,
\item locally compact Hausdorff spaces,
\item CW-complexes,
\item first countable spaces, i.e. spaces where each point has a countable base.
\end{itemize}
\end{exam}
\begin{theo}
We denote the category of $k$-spaces with continuous maps between them by $\Cb{kTop}.$ Additionally, we denote the category of all topological spaces with $k$-continuous maps between them by $\Cb{Top_k}.$ Then we have
\[
\begin{tikzcd}
\Cb{kTop} \ar[rd,"\approx"'] \ar[r,hook] & \ar[d] \Cb{Top} \\
& \Cb{Top_k};
\end{tikzcd}
\]
recall that $\approx$ denotes an equivalence of categories.
\end{theo}
\begin{defin}
For $X \in \Cb{Top},$ we define the \textit{kaonization} (\textit{k-ification}) of $X$ to be a topological space $k(X)$ having the same underlying set as $X$ but with the condition that a subset $A \subseteq k(X)$ is closed if and only if it is closed in any compact subspace $K \subseteq X.$ Succinctly,
\[
A \subseteq k(X) \text{ closed } \iff A \cap K \subseteq K \text{ closed, } \forall K \subseteq X \text{ compact.}
\]
\end{defin}
\begin{rem}
In finding the kaonization of a topological space$(X,\tau),$ we are enriching $\tau$ by adding more closed sets, and hence open sets, in order to generate the coarsest topology $\tau'$ relative to $\tau$ such that $\tau'$ makes $(X,\tau')$ a $k$-space (which is in fact even compactly generated).
\end{rem}
\begin{prop}
For any space $X:$
\begin{itemize}
\item $k(X) \stackrel{\mathds{1}}{\longrightarrow} X$ is continuous;
\item $k(X)$ is compactly generated;
\item $k(X) = X$ if and only if $X$ is compactly generated;
\item if $Y$ is compactly generated, then
\begin{align*}
Y \to k(X) \text{ continuous } & \iff Y \to X \text{ continuous}, \\
\Mor{\Cb{kTop}}{Y}{k(X)} & \longleftrightarrow \Mor{\Cb{Top}}{Y}{X}, \\
(f:Y \to k(X)) & \longmapsto (\mathds{1}f:Y \to k(X) \to X);
\end{align*}
\item if $\iota:\Cb{CGTop} \to \Cb{Top}$ is an inclusion functor with source the category of all compactly generated spaces with continuous maps between them, then $\iota/X$ has final object $\mathds{1}:k(X) \to X;$
\item $k(X)$ and $X$ have the same compact subsets;
\item $k(X)$ and $X$ are weakly homotopy equivalent, i.e. they have isomorphic homotopy groups. Moreover, if $Y = \Delta^n$ is an $n$-simplex, then singular simplices are the ``same.'' Therefore, \emph{(}co\emph{)}homology theories of $k(X)$ and $X$ coincide.
\end{itemize}
\end{prop}
\begin{rem}
The identity map $k(X) \to X,$ which you may recall is continuous, acts as the counit map for an adjunction of the inclusion and kaonization functors. More specifically, the functors
\[
\iota:\Cb{kTop} \hookrightarrow \Cb{Top} \text{ and } k:\Cb{Top} \longrightarrow \Cb{kTop}
\]
comprise the adjunction $\iota \dashv k.$ Consequently, $k$ preserves limits and $\iota$ preserves colimits. In particular, $k$ preserves products. If we define $X \times_k Y := k(X \times Y),$ then we obtain that
\[
k(X) \times k(Y) \cong X \times_k Y.
\]
In fact, if $X$ is locally compact and $Y$ is compactly generated, then
\[
X \times_k Y = X \times Y.
\]
\end{rem}
\begin{defin}
A category $\Cs{C}$ with finite products and terminal object is said to be \textit{cartesian closed} if the functor
\[
- \sqcap X:\Cs{C} \to \Cs{C}
\]
has right adjoint, which is called the \textit{exponentiation functor} and denoted by $(-)^X.$ The terminal object $T \in \Cs{C}$ has the property that $A \sqcap T \cong A$ for any $A \in \Cs{C}.$ This further implies that $A^T \cong A,$ which may be shown. Adjunction produces the bijection
\[
(A \sqcap T,B) \leftrightarrow (A,B^T) \Longrightarrow (A,B) \leftrightarrow (A,B^T) \Longrightarrow B \cong B^T,
\]
where the last implication follows from the Yoneda Lemma.
\end{defin}
\begin{defin}
Let $X,Y$ be topological spaces. We imbue the set $C(X,Y)$ of continuous maps from $X$ to $Y$ with a topology whose basis consists of the sets
\[
U(K,V) := \{f:X \to Y \text{ continuous }|f(K) \subseteq V\},
\]
where $K \subseteq X$ compact and $V \subseteq Y$ open. We call this topology the \textit{compact-open topology}.
\end{defin}
\begin{rem}
$\Cb{kTop} \approx \Cb{Top_k}$ are cartesian closed. Exponential objects are $kC(X,Y),$ i.e. all $k$-continuous (and thereby continuous since $X$ is a $k$-space) maps with the compact-open topology.
\end{rem}
\begin{defin}
Let $X = \varinjlim X_n$ in $\Cb{Top}.$ We say that $A \subseteq X$ open if $A \cap f_n(X_n) \subseteq f_n(X_n)$ open for each $n,$ where $f_n:X_n \to X$ are structural maps. This defines the \textit{weak topology} on $X.$
\end{defin}
\subsection{Geometric Realization}
The category $\Delta$ generalizes the topological notion of simplicial complexes and provides a completely algebraic method for developing homotopy and homology theories, which is a common theme in abstract homotopy theory.
\begin{defin}
Let $\Delta$ be a category of finite ordinals $[n]$ (cf. $\Delta[n]$ in \thref{pos}) with order-preserving (equivalently non-decreasing) maps $f:[n] \to [m].$ A contravariant functor $F:\Delta \to \Cb{Set}$ is called a \textit{simplicial set}. From this we obtain a category $\Cb{sSet} = (\Delta^{op},\Cb{Set})$ of simplicial sets with simplicial maps between them.
\end{defin}
Here we should note well that $\Cb{sSet}$ is a functor category, whose objects are simplicial sets. Harkening back to the definition of a functor category, this means that simplicial maps are in fact natural transformations between simplicial sets.
\begin{defin}
The map $d_i^{n+1}:[n] \to [n+1]$ defined by
\[
d_i^{n+1}(j):
\begin{cases}
j, & 0 \leq j < i \\
j+1, & i \leq j \leq n
\end{cases}
\]
is called a \textit{coface map}. On the other hand, the map $s_i^{n}:[n+1] \to [n]$ defined by
\[
s_i^{n}(j):
\begin{cases}
j, & 0 \leq j \leq i \\
j-1, & i < j \leq n+1
\end{cases}
\]
is called a \textit{codegeneracy map}.
\end{defin}
The coface map $d_i$ fixes all elements below $i$ and raises those elements at least $i,$ effectively leaving a blank in the position that $i$ used to be in. This indicates $[n]$ as the $i$-th face of $[n+1].$ The codegeneracy map $s_i$ fixes all elements at most $i$ and lowers those elements above $i,$ leading to consolidation of the element $(i+1)$ into $i.$ Considering that $i+1 = i$ indicates a degenerate simplex, $s_i$ acts as though it is converting a degenerate $[n+1]$ simplex into its corresponding $[n]$ simplex.
\begin{theo} \label{simp}
The coface and codegeneracy maps satisfy the following relations:
\begin{align*}
d_i^{n+1}d_j^{n} &= d_{j+1}^{n+1}d_i^n \quad\ \text{ for } i \leq j, \\
s_j^{n-1}s_i^n &= s_i^{n-1}s_j^n \quad\ \text{ for } i \leq j, \\
s_j^{n-1}d_i^n &=
\begin{cases}
d_i^{n-1}s_{j-1}^{n-2}, & i < j \\
\mathds{1}_{[n-1]}, & i = j \text{ or } i = j+1 \\
d_{i-1}^{n-1}s_j^{n-2}, & i > j+1
\end{cases}
\end{align*}
Moreover, any order-preserving map $f:[n] \to [m]$ may be written uniquely as an expression of the form
\[
f = d_{i_k}^m\ldots d_{i_1}^{n-h+1}s_{j_1}^{n-h}\ldots s_{j_h}^{n-1},
\]
where $k$ and $h$ are non-negative integers such that
\begin{itemize}
\item $n+k-h = m,$
\item $0 \leq i_1 < \ldots < i_k \leq m,$
\item $0 \leq j_1 < \ldots < j_h \leq n,$
\item $i_1,\ldots,i_k$ are the elements of $[m]$ not in the image of $f,$
\item $j_1,\ldots,j_h$ are the elements of $[n]$ at which $f$ does not increase (as in $f(j) = f(j+1)$).
\end{itemize}
\end{theo}
\begin{rem}
Here it is important to note that, while the functions above are for $\Delta,$ a simplicial set is a contravariant functor, so there are dual maps to the coface and codegeneracy maps which manifest in simplicial sets in a manner that may be familiar to those who have worked with simplicial complexes.
\end{rem}
\begin{defin}
Explicitly, a \textit{simplicial set} $X \in (\Delta^{op},\Cb{Set})$ is a sequence of sets $\{X_k = X[k]\}_{k \in \mathbb{N}}$ together with the maps
\[
\partial_i^n = X(d_i^n), \qquad \sigma_j^n = X(s_j^n)
\]
which are called \textit{face} and \textit{degeneracy} maps, respectively. The elements of $X_0$ are called \textit{vertices}, of $X_1$ are called \textit{edges}, of $X_n$ are called \textit{$n$-simplices of $X.$}
\end{defin}
By the duality principle, the face and degeneracy maps must satisfy relations dual to those in \thref{simp}. This theorem also implies that simplicial maps may be written uniquely as a composition of degeneracy and face maps.
\begin{rem}
The process described in this section is reversible. That is, given a sequence of sets $X_0,X_1,\ldots,X_n,\ldots$ and maps $\partial_i^n,\sigma_j^n$ satisfying the condition of \thref{simp}, it is possible to construct a functor $X:\Delta^{op} \to \Cb{Set}$ such that $X[i] = X_i,$ $X(d_i^n) = \partial_i^n,$ and $X(s_j^n) = \sigma_j^n$ for $i,j \geq 0.$
\end{rem}
There is a method of ``topologizing'' categories by means of the following construction.
\begin{rem}
Given a category $\Cs{C}$, its \textit{classifying space} is a CW-complex $B\Cs{C}$ constructed in the following manner:
\begin{enumerate}
\item The 0-cells (vertices) are objects of $\Cs{C}$.
\item The 1-cells (edges) are morphisms of $\Cs{C}$ (excluding the identities) which attach at the source and target of the morphism.
\item For each ordered pair $(f,g)$ of composable non-identity morphisms in $\Cs{C}$, attach a 2-simplex $\Delta^2$ as indicated by the following diagram:
\begin{center}
\begin{tikzcd}
\bullet \ar[rr,"gf"] \ar[rd,"f"'] & & \bullet \\
& \bullet \ar[ru,"g"'] &
\end{tikzcd}
\end{center}
\item Continue this process inductively by attaching an $n$-simplex $\Delta^n$ when given an $n$-tuple of composable non-identity morphisms.
\item $B\Cs{C}$ is the union of all these spaces with the weak topology, which is the topology in which a subset is open if and only if its intersection with each of the subspaces above is open.
\end{enumerate}
\end{rem}
\begin{exam}
Some illustrative examples of classifying spaces include the following.
\begin{itemize}
\item $B\{\bullet\}$ is a point.
\item $B\{0 \leq 1\} = [0,1]$.
\item $B\{0 \leq 1 \leq 2\} = \Delta^2$.
\item $B\Delta[n] = \Delta^n$.
\item $B\{\bullet \rightrightarrows \bullet\} = S^1$.
\end{itemize}
\end{exam}
\subsection{Homotopy Groups}
Topological spaces formalize and generalize continuity as we intuitively conceive of it. From this, we are able to talk about what it means for two spaces to be topologically equivalent, which is referred to as homeomorphism. As a reminder, a homeomorphism is a continuous bijection with the extra condition that its inverse is also continuous. It turns out this is a fairly strong requirement for something that is colloquially referred to as ``rubber-sheet geometry.'' For example, homeomorphic spaces must share properties such as dimension, cardinality, compactness, and separability even though a rubber sheet should allow us to squish an open ball to a closed ball contained within or even to a point in an obvious fashion, leading to a change in compactness or even dimension.
So what we seek is a weaker means of distinguishing spaces which preserves properties consistent with this rubber sheet perspective. Here we begin with the notion of homotopy and build up to a suitable means of distinguishing spaces. But first it is necessary to outline precisely what types of topological spaces are necessary for homotopy to be well-defined; in essence, we must set out by outlining the operative category.
\begin{defin}
A \textit{pointed topological space} is a pair $(X,x_0),$ where $X$ is a space and $x_0 \in X$ is a distinguished point. The base point is often not explicitly written when it is clear that the space is based, so we may refer to $(X,x_0)$ as just $X.$ Then a continuous map $f:X \to Y$ between based spaces is a \textit{pointed map} if $f(x_0) = y_0.$ The category of pointed spaces with pointed (continuous) maps is notated as $\Cb{Top_*}.$
\end{defin}
For the remainder of this section, assume that all spaces and maps between them belong within $\Cb{Top_*}$ unless indicated otherwise. Furthermore, we will use $I$ to denote the closed interval $[0,1].$
\begin{defin}
Let $f,g:X \to Y$ be pointed maps between the spaces $X$ and $Y.$ A continuous map $h:X \times I \to Y$ satisfying the conditions
\[
h(x,0) = f(x) \quad\quad\quad h(x,1) = g(x) \quad\quad\quad h(x_0,t) = y_0
\]
is called a \textit{homotopy} from $f$ to $g.$ If there exists a homotopy from $f$ to $g,$ we say that $f$ is \textit{homotopic} to $g,$ notated as $f \simeq g.$ A map which is homotopic to the constant map, whose image must be $\qty{y_0}$ in $\Cb{Top_*},$ is called \textit{nullhomotopic}. If the identity $\mathds{1}_X$ is nullhomotopic, then we say that $X$ is \textit{contractible}.
\end{defin}
We may conceive of a homotopy $h:f \simeq g$ between pointed maps $f,g:X \to Y$ as a map which assigns to each point of $x \in X$ a path $p_x:I \to Y$ beginning in the image of $f$ and ending in the image of $g$ with the condition that $p_{x_0}$ be the constant path. With this insight, it becomes clear that path-connectedness is a very convenient property when dealing with homotopy. For one, having path-connectedness makes the choice of base point trivial, and it is also fairly easy to see that contractible spaces must always be path-connected.
\begin{rem}
The third condition listed for a map to be a homotopy serves the purpose of ensuring that for any choice of $t \in I,$ that the intermediate map $h(-,t):X \to Y$ is a pointed map. However, the map $h(x,-):I \to Y$ need not be pointed unless $x = x_0.$
Also, the alternative view of homotopy as a path-assigning function should remind us of a collection of parameterized univariate maps. Equipped with the motivation given for adjunction, we can reasonably predict that some kind of adjunction is at play, and in fact there is (c.f. loop space/reduced suspension adjunction from \thref{adj}).
\end{rem}
\begin{exam}
Provided here are some examples in real space which serve to develop geometric intuition for the definitions just given.
\begin{itemize}
\item In $\R^n$ with arbitrary base point, any two paths $f,g:I \to \R^n$ are homotopic by way of the ``straight-line'' homotopy, which is given by
\[
h:
\left\{
\begin{array}{rcl}
I \times I &\to & \R^n \\
(x,t) &\mapsto & (1-t)f(x)+tg(x).
\end{array}
\right.
\]
In some multivariable calculus courses, it may have been mentioned that spaces with this property are called simply connected. This will be defined more precisely further on, but this example serves as a useful heuristic.
\item The straight-line homotopy on $\R^n$ may actually be used to show that any map is nullhomotopic and, in particular, that $\R^n$ is contractible. As we will soon see, this suggests the curious result that $\R^n$ behaves like a point with respect to homotopy. And because $\C^n \cong \R^{2n}$ are homeomorphic, anything statement about homotopy in real space extends to complex space.
\item The seemingly similar space $\R^n-\vb{0}$ with arbitrary base point actually exhibits very different properties for $n = 1,$ $n = 2,$ and $n \geq 3.$
\begin{itemize}
\item For $n = 1,$ the space consists of 2 connected components (which are also path-components). In fact, $\R-\vb{0}$ is homeomorphic to two copies of $\R$ by way of the maps $e^x$ and $e^{-x}$ defined on each of the copies of $\R.$ Consequently, $\R-\vb{0}$ behaves like 2 points up to homotopy since each copy of $\R$ is contractible.
\item For $n = 2,$ the space is path-connected, and many maps are nullhomotopic. However, consider the map
\[
f:x \mapsto (\cos(2\pi x),\sin(2\pi x))
\]
on $I.$ It is a fact---intuitive but not obvious to show---that this map is not nullhomotopic and that every loop (path with matching endpoints) is either nullhomotopic or homotopic to $f.$ With that said, it is clear then that $\R^2$ is no longer contractible or simply connected if you remove a point from it.
\item For $n \geq 3,$ the space is again path-connected. In contrast with $\R^2-\vb{0},$ this space is simply connected, but that does not mean that all maps are nullhomotopic as it is not hard to convince ourselves intuitively that this space is still not contractible. The distinction is between a generic map (with arbitrary domain) and a path (a map with domain $I$).
\end{itemize}
\end{itemize}
\end{exam}
Homotopies may be defined in $\Cb{Top}$ by omitting the condition that base points be sent to base points since we do not have base points to speak of. However, we will see further on that we run into problems when we ignore the base point entirely. Luckily enough, it is straightforward to show that homotopies of maps $X \to Y$ induce equivalence relations both in the presence and absence of a base point.
\begin{cor} \label{homt}
The relation $\simeq$ is an equivalence relation on the sets $\Mor{\Cb{Top}}{X}{Y}$ and $\Mor{\Cb{Top_*}}{X}{Y}.$
\end{cor}
\begin{proof}
It suffices to confirm this within $\Cb{Top}$ since maps between pointed spaces are particular cases of maps between general spaces. Let $f,f',f'':X \to Y$ be continuous maps.
\begin{itemize}
\item Reflexivity: The map $h(x,t) = f(x)$ is a homotopy $f \simeq f.$
\item Symmetry: Let $f \simeq f'$ and $h:f \simeq f'$ be a homotopy between them. Then $h'(x,t) = h(x,1-t)$ is a homotopy $f' \simeq f.$
\item Transitivity: Let $f \simeq f'$ and $f' \simeq f''$ with corresponding homotopies $h:f \simeq f'$ and $h':f' \simeq f''.$ Then the concatenation function
\[
h^\dagger(x,t) =
\begin{cases}
h(x,2t) & \text{if } t \in [0,\frac{1}{2}] \\
h'(x,2t-1)) & \text{if } t \in [\frac{1}{2},1]
\end{cases}
\]
is continuous by the pasting lemma because $h(x,1) = h'(x,0)$ and thereby constitutes a homotopy $f \simeq f''.$
\end{itemize}
\end{proof}
\begin{defin}
Let $f:X \to Y$ be a pointed map. The equivalence class
\[
[f] := \qty{g:X \to Y\ |\ g \simeq f}
\]
is referred to as the \textit{homotopy class} of $f,$ and the collection of all such classes forms the \textit{homotopy set} $[X,Y].$ Alternatively, the homotopy set may be obtained as the quotient
\[
[X,Y] := \frac{\Mor{\Cb{Top}_*}{X}{Y}}{\simeq}.
\]
Pointed topological spaces with homotopy classes of maps between them form a category $\Cb{HoTop}_*.$
\end{defin}
\begin{rem}
Categories require compositions laws to combine their morphisms. If $f,g$ are composable maps (composition given by $gf$) in $\Cb{Top}_*,$ then the composition $[g][f]$ is given by $[gf].$ As with any representable functor, $[X,-]$ and $[-,Y]$ are covariant and contravariant functors, respectively, from $\Cb{Top_*}$ to $\Cb{Set},$ and the induced maps are defined in the same fashion as with the representable functors.
\end{rem}
Recall that a homeomorphism is a continuous bijection with a continuous inverse. What we mean by inverse is that there exists another function going in the opposite direction which composes on either side to produce the relevant identity map. At the outset of this subsection, we cited a desire to relax the conditions for a homeomorphism, and homotopy seems to be a good fit. However, homotopy is a relation between maps, not spaces. By switching one equivalence relation for another, we arrive at what we sought.
\begin{defin}
Let $f:X \to Y$ be a continuous map (not necessarily pointed). We call $f$ a \textit{homotopy equivalence} if there exists some $g:Y \to X$ such that
\[
fg \simeq \mathds{1}_Y \quad \quad \quad gf \simeq \mathds{1}_X,
\]
in which case $X$ and $Y$ are said to be \textit{homotopy equivalent}, denoted by $X \simeq Y.$ If such a $g$ exists, it is also a homotopy equivalence which is called the \textit{homotopy inverse} of $f.$
\end{defin}
To show that homotopy equivalences are closed under composition, it is sufficient to construct the homotopy inverse ($(g'g)(ff') \simeq \mathds{1}$). From this, homotopy equivalence is an equivalence relation on all pointed (and non-pointed) topological spaces, where reflexivity follows from the identity map, symmetry from the existence of a homotopy inverse, and transitivity from closure under compositions.
\begin{rem}
When it is said that some statement/construction is valid for a topological space $X$ up to homeomorphism, what is really being said is that these statements/constructions are invariants of homeomorphism. In the same way, there are homotopy invariants, which are statements/constructions valid up to homotopy equivalence. Now we are prepared to present the titular construction of this subsection, an important homotopy invariant.
\end{rem}
\begin{defin}
Let $(X,x_0)$ be a pointed topological space. The homotopy set $[S^1,X]$ forms a group under the associative binary operation of concatenation (c.f. transitivity of $\simeq$ in \thref{homt}), where the inverse of a loop $[f]$ is given by the loop $[f]^{-1}$ going in the opposite direction. Explicitly, a loop $f:I \to X$ has inverse
\[
f^{-1}(t) = f(1-t).
\]
This group is called the \textit{fundamental group} and is denoted by $\pi_1(X,x_0).$
When $X$ is path-connected, it may be shown that fundamental groups with different base points are isomorphic, so we may instead write $\pi_1(X).$ If $\pi_1(X)$ is trivial, then $X$ is called \textit{simply connected}.
\end{defin}
The fundamental group is an algebraic construct which allows us to detect portions of the topological space which may not be deformed to a point and their ``positions'' relative to each other. However, it is often not very simply to compute, and in many cases is not abelian. Leaving that aside for the time being, the formulation of the fundamental group submits to fairly straightforward generalization.
\begin{comment}
Add stuff about how homotopies assign paths to each point of the domain and how this leads to the importance of path-connectedness (or components) within homotopy theory. Use this to conclude that contractible spaces are path-connected (converse not true in general). Then go onto homotopy equivalence before defining homotopy groups.
\end{comment}
\begin{defin}
Let $(X,x_0)$ be a pointed topological space. Then
\[
\pi_n(X,x_0) = [S^n,X]
\]
is the \textit{$n$-th homotopy group}, where $[S^n,X]$ is the set of homotopy classes of base point-preserving continuous functions $S^n \to X.$ It is an Abelian group for $n \geq 2$ with addition defined by
\[
f+g \colon
\begin{tikzcd}
S^n \ar[r,"eq"] & S^n \vee S^n \ar[r,"f \vee g"] & X,
\end{tikzcd}
\]
where $eq \colon S^n \to S^n \vee S^n$ is the quotient map which collapses the equator of $S^n$---any copy of $S^{n-1}$ containing the base point---to the wedge point of the two resulting spheres. However for $n = 0,$ there is no canonical group operation in general, so $\pi_0(X)$ is a set in bijection with the connected components of $X.$
\end{defin}
\begin{lem} \label{pizc}
Let $\Cs{C}$ be a category and $\sim$ be an equivalence relation on $\ope{Obj}(\Cs{C})$ generated by all pairs of objects $(A,B)$ such that there is a morphism $A \to B$. Then $$\pi_0(B\Cs{C}) = \ope{Obj}(\Cs{C})/\sim$$
\end{lem}
This lemma tells us that two objects $A$ and $B$ are equivalent mod $\sim$ if and only if there is a \emph{mixed path} between them. There are four possible forms a mixed path can take:
\begin{align*}
&
\begin{tikzcd}[ampersand replacement = \&]
\bullet \ar[r] \& \bullet \& \ar[l] \cdots \& \ar[l] \bullet \ar[r] \& \bullet
\end{tikzcd}
&
&
\begin{tikzcd}[ampersand replacement = \&]
\bullet \ar[r] \& \bullet \& \ar[l] \cdots \ar[r] \& \bullet \& \ar[l] \bullet
\end{tikzcd}
\\
&
\begin{tikzcd}[ampersand replacement = \&]
\bullet \& \ar[l] \bullet \ar[r] \& \cdots \& \ar[l] \bullet \ar[r] \& \bullet
\end{tikzcd}
&
&
\begin{tikzcd}[ampersand replacement = \&]
\bullet \& \ar[l] \bullet \ar[r] \& \cdots \ar[r] \& \bullet \& \ar[l] \bullet
\end{tikzcd}
\end{align*}
\newpage
\section{Torsion Products as Homotopy Groups}
\subsection{Category $\Tor{}{R}{A}{B}$}
If $_RP$ is a left $R$-module, then $P_R^* = \Hom{R}{P}{R}$ is a right $R$-module. If $f:P_1\to P_2$ is a homomorphism of left $R$-modules, then $f^*:P_2^* \to P_1^*$ is a homomorphism of right $R$-modules.
\begin{defin}
Let $R$ be a ring, $A_R$ and $_RB$ be right and left $R$-modules, respectively. Define the \textit{torsion category} $\mathrm{Tor}^R(A,B)$ of modules $A$ and $B$ as a category whose
\begin{itemize}
\item objects are triples $(P,\varepsilon,\eta)$, where $P$ is a finitely generated projective right $R$-module with homomorphisms $\varepsilon:P \to A$ and $\eta:P^* \to B$,
\item morphisms between triples $(P_1,\varepsilon_1,\eta_1)$ and $(P_2,\varepsilon_2,\eta_2)$ are given by the homomorphisms $f:P_1 \to P_2$ such that the following diagrams commute:
\[
\begin{tikzcd}
P_1 \ar[rr,"f"] \ar[rd,"\varepsilon_1"'] & & \ar[ld,"\varepsilon_2"] P_2 \\
& A &
\end{tikzcd}
~~
\begin{tikzcd}
P_1^* \ar[rd,"\eta_1"'] & & \ar[ld,"\eta_2"] \ar[ll,"f^*"'] P_2^* \\
& B, &
\end{tikzcd}
\]
\item composition of morphisms of triples is given by the composition of morphisms between the finitely generated projective modules within the triple,
\item identity on the triple $(P,\varepsilon,\eta)$ is the identity homomorphism on $P$.
\end{itemize}
\end{defin}
\begin{defin}
Define a bifunctor $\Box:\Tor{}{R}{A}{B} \times \Tor{}{R}{A}{B} \to \Tor{}{R}{A}{B}$ as $$(P_1,\varepsilon_1,\eta_1) \ \Box\ (P_2,\varepsilon_2,\eta_2) = (P_1 \oplus P_2,(\varepsilon_1,\varepsilon_2),(\eta_1,\eta_2)^T)$$ for objects. Let $p:(P_1,\varepsilon_1,\eta_1) \to (P_2,\varepsilon_2,\eta_2)$ and $q:(Q_1,\varepsilon_1',\eta_1') \to (Q_2,\varepsilon_2',\eta_2')$ be morphisms in $\Tor{}{R}{A}{B}$. Then $p \ \Box\ q:P_1 \oplus Q_1 \to P_2 \oplus Q_2$ makes the following diagrams commute:
\[
\begin{tikzcd}
P_1 \oplus Q_1 \ar[rr,"p \ \Box\ q"] \ar[rd,"{(\varepsilon_1,\varepsilon_1')}"'] & & \ar[ld,"{(\varepsilon_2,\varepsilon_2')}"] P_2 \oplus Q_2 \\
& A &
\end{tikzcd}
~~
\begin{tikzcd}
P_1^* \oplus Q_1^* \ar[rd,"{(\eta_1,\eta_1')}"'] & & \ar[ld,"{(\eta_2,\eta_2')}"] \ar[ll,"p^* \ \Box\ q^*"'] P_2^* \oplus Q_2^* \\
& B, &
\end{tikzcd}
\]
where
\[
p \ \Box\ q =
\begin{pmatrix}
p & 0 \\
0 & q
\end{pmatrix}.
\]
The object $(0,0,0)$ is an identity for this operation.
\end{defin}
\subsection{Homotopic Relationship Between Torsion Products and Category $\Tor{}{R}{A}{B}$}
The primary result of Robinson's paper \cite{Robinson1981} is as follows.
\begin{theo}
Let $R$ be a ring, $A_R$ a right $R$-module, and $_RB$ a left $R$-module. Then
\[
\pi_n(B\mathrm{Tor}^R(A,B))\cong \Tor{n}{R}{A}{B}
\]
as abelian groups.
\end{theo}
This may be used to carry out a computation of tensor products since $\Tor{0}{n}{A}{B} \cong A \otimes_R B.$ We will use the theorem to show that $$\Z/4\Z \otimes_\Z\Z/6\Z \cong \Z/2\Z.$$
\begin{proof}
The outline of this computation is as follows. Note that projective $\Z$-modules are precisely free abelian groups.
\begin{enumerate}
\item Determine generators: Both $P_1$ and $P_2$ may be considered as free $\Z$ modules and so may be written as $\Z^n$ and $\Z^m$, respectively. The maps may then be written as $$\varepsilon_1 = (x_1,x_2,\ldots,x_n) ~~ \text{ and } ~~ \eta_1 = (y_1,y_2,\ldots,y_n)^T$$ for $P_1$. Because $\Hom{\Z}{\Z^n}{\Z} \cong \Z^n$, the maps for $P_1^*$ will be similar (but transposed from $\varepsilon_1$ and $\eta_1$). With this, we may write $$(\Z^n,(x_1,x_2,\ldots,x_n),(y_1,y_2,\ldots,y_n)^T) = \Box_{i,j}x_i y_j(\Z,1,1),$$ so the generators we obtain are of the form $(\Z,1,1)$. The same argument applies to $\Z^m$.
\item Establish path-connectedness within category: Because all objects are generated by $(\Z,1,1)$ through multiplication, we have the following triangles to satisfy in order to establish path-connectedness to the zero module object $(0,0,0)$ for any object. We will omit the subscripts on the maps $\varepsilon$ and $\eta$ as a result of working with only one generator.
\begin{align*}
&
\begin{tikzcd}[ampersand replacement = \&]
0 \ar[rd] \ar[rr] \& \& \Z \ar[ld,"\varepsilon"] \\
\& \Z/4\Z \&
\end{tikzcd}
&
&
\begin{tikzcd}[ampersand replacement = \&]
\Z \ar[rd,"\varepsilon"'] \ar[rr] \& \& 0 \ar[ld] \\
\& \Z/4\Z \&
\end{tikzcd}
\\
&
\begin{tikzcd}[ampersand replacement = \&]
0 \ar[rd] \& \& \ar[ll] \Z \ar[ld,"\eta"] \\
\& \Z/6\Z \&
\end{tikzcd}
&
&
\begin{tikzcd}[ampersand replacement = \&]
\Z \ar[rd,"\eta"'] \& \& \ar[ll] 0 \ar[ld] \\
\& \Z/6\Z \&
\end{tikzcd}
\end{align*}
Then the conditions for a morphism existing between $(\Z,\varepsilon,\eta)$ and $(0,0,0)$ are only satisfied when $\varepsilon = 0$ or $\eta = 0$. Therefore, only objects of the form $(\Z^x,0,\eta)$ or $(\Z^x,\varepsilon,0)$ are connected by a path to $(0,0,0)$.
\item Determine order of generators: We now multiply the generator by specific integers in order to find objects path-connected to $(0,0,0)$: $$4(\Z,1,1) = (\Z,0,1) \sim (0,0,0)\text{ and }6(\Z,1,1) = (\Z,1,0) \sim (0,0,0).$$ Because $2 = \ope{gcd}(4,6)$, there exist integers $p,q \in \Z$ such that $4p+6q = 2$ (namely $q = 1$ and $p = -1$), which implies $$(6-4)(\Z,1,1) = 2(\Z,2,2) \sim (0,0,0).$$ Because $(\Z,1,1)$ is not connected to $(0,0,0)$ by a morphism, we see that $B\Tor{}{\Z}{\Z/4\Z}{\Z/6\Z}$ consists of two path-components. By \ref{pizc}, $$\pi_0(B\Tor{}{\Z}{\Z/4\Z}{\Z/6\Z) \cong (\Z,1,1}/2(\Z,1,1) \cong \Z/2\Z.$$ This aligns with what is obtained by direct computations of the tensor product.
\end{enumerate}
\end{proof}
\section*{Acknowledgements}
I would like to thank my mentor Alex Sorokin for his time and attention in both preparing me for and working with me on this project. I learned a great deal from him about math and beyond, and his constant patience with me was very much appreciated. Under his guidance, I made significant progress towards being able to identify and postulate about important and ``good'' questions. I would not be upset to be half the math educator he is when I pursue graduate studies.
In addition, I would like to thank Dr. Robin Walters for his advice and feedback on the progress of the project and for providing further opportunities for me to integrate what I learned through meaningful discussions during my progress reports. And I would remiss to not mention Professor Anthony (Tony) Iarrobino for his comments on my presentations, writing, and aspirations which seemed to perfectly balance generality with detail. And overall, I would like to thank the Northeastern RTG group for supporting this very informative and stimulating experience.
|
\section{Introduction}
\vspace{-4mm}
Humans are adaptive by nature. While it is very natural for humans to ask clarifying questions and immediately correct course whenever there is a misunderstanding, interactive learning using self-supervision has been a holy grail area for Conversational AI. Most Conversational AI systems rely on knowledge-based or machine-learning based (or hybrid) understanding components under the hood to understand the intents of the users’ and their arguments (usually called as slots), such as “{\em set an alarm for 7 am}”, where the intent can be {\em set\_alarm}, and {\em 7 am} can be the time slot. These systems typically do not have the capability to interact with end-users during a live conversation to seek explanations to improve the AI system's understanding of entities, intents or other constructs. When users speak an utterance outside the comprehension of these systems, the AI agent responds with a dead-end response such as ‘Sorry I don’t know that’.
In this paper we describe a teachable AI system towards enabling users to directly teach a conversational AI agent via a live interactive teaching session using natural language explanations. Learning concept definitions using interactive voice is a challenging problem due to a number of reasons, including wide vocabulary of words that users could use to describe definitions, user distractions during teaching sessions, grounding of related entities, fuzziness in the re-use of previously taught concepts across domains, etc. The method described in this paper augments goal-oriented AI agents with an interactive teaching capability, to improve task completion rates by asking questions to fill gaps in the AI agent’s understanding, hence making the agent more personal, conversational and adaptive.
More specifically, the AI agent will be checking the utterances whether it is “Teachable” using a deep learning based parser, which not only classifies the utterance, but also identifies the entity or intent which needs to be learnt from the users. If so, the “Classroom” sub-dialogue kicks in, where the users are asked what they mean by these concepts. This sub-dialogue is driven by a dedicated dialogue manager policy model which incorporates a definition understanding model to interpret users' responses. For example, if the utterance is "{\em set an alarm for my baseball practice}", the system would ask "{\em when is your baseball practice?}".
Despite a few academic papers towards learning directly from users as presented in Section~\ref{sec:rw}, to the best of our knowledge, there is no work that covers all of these novel contributions:
\begin{itemize}
\itemsep -0.5ex
\item A dedicated multi-turn domain-agnostic dialogue system specifically tailored for interactive teaching, that augments an existing Conversational AI System to learning explanations from user in real-time.
\item A multi-task neural Concept Parser that automatically identify gaps in an AI agent’s understanding, using a multi-task model that incorporates semantic role labeling and constituency parser.
\item A neural Definition Understanding system along with a policy model to aid in conducting robust teaching sessions with the user.
\end{itemize}
The following sections are organized as follows. Section~\ref{sec:rw} describes the related work in the area of interactive learning methods to learn explanations, and predicting gaps in the AI agent’s understanding for a given user input; Section~\ref{sec:approach} describes our architecture and modeling methods, and Section~\ref{sec:exp} describes our experimental evaluations.
\section{Related Work}
\vspace{-2mm}
\label{sec:rw}
The related work on learning via user-in-the-loop for Conversational AI systems can be analyzed in two categories. The higher level interactive learning approaches, and utterance parsing based gap prediction studies.
\subsection{Interactive Learning}
\vspace{-2mm}
Existing work on interactive learning methods can be grouped into 3 categories:
\begin{itemize}
\item Game learning using limited/no vocabulary: Several methods have been proposed for learning a game (such as Hanoi) using interactive sessions with the user Kirk~{\em et al.}~\cite{kirk-19} Wang~{\em et al.}\cite{wang-17}. Such methods represent the game scene as a symbolic representation of attributes (such as shape, color) which is used to learn the state progression, to learn the game.
\item Neural code synthesis methods: Methods like Yin~{\em et al.}~\cite{yin-2020} aim at converting user utterances directly into a coding language like Python. To our knowledge, such syntax-driven code generation methods are still in their early stages, and not yet ready for integration into mainstream conversational AI agents that cover a wide variety of domains.
\item Learning by GUI demonstrations: GUI based methods use semantic parsers to learn conditional expressions and their associated actions. One notable study is by Allen~{\em et al.} using a web browser to teach tasks such as buying a book from Amazon~\cite{plow}.
\end{itemize}
\subsection{NLU gap prediction}
\vspace{-2mm}
The problem of predicting segments of an utterance which a base NLU system could not interpret, was tackled by Jia {\em et al.}~\cite{jia-2017}, by using a set of post-processing rules on top of the base NLU's slot-filling model. The rules were built on heuristics from the base NLU model, such as confidence score threshold, out-of-vocabulary flag, and span-expansion based on syntactic dependency tree. This paper also introduced a new dataset for this problem that is publicly available, which is used in our work as one of the datasets for evaluating model performances. In \cite{Kobayashi-2018,kobayashi2019slot} the authors proposed augmenting the training data for the slot-filling model, by injecting noise tokens into the regular slot values of training data, to force the model to learn about the context of the respective slots. This augmentation is done to improve the model robustness in predicting the correct span, on unseen slot values. One variant of such approach uses negative sampling to train the model to identify unknown slot values together with a joint slot tagging and slot classifier~\cite{Hou-2019}. Another line of work focuses on different model structures such as pointer-networks, to better copy unknown slot values from the utterance into final slot value prediction~\cite{xu-2018,Zhao-2018,Yang-2020}. This line of work usually considers the problem under a dialogue state tracking setting, therefore only one slot value is predicted each time.
\subsection{Definition Understanding}
\vspace{-2mm}
Our task of understanding the definition of concepts from users' explanations, generally falls into the categories of reading comprehension and slot filling. For reading comprehension, there is a particular thread that formulates question answering as span prediction. Those works typically learn a representation for the question and passage, and then predict the start and end of the answer with attention mechanism between question, passage and answer \cite{chen2017reading,chen2018neural}. The slot filling thread work focuses on sequential labeling of an answer\cite{mesnil2014using}, with most recent works utilizing neural slot-fillers with joint intent classifiers \cite{liu2016attention,goo2018slot,xu2013convolutional,zhang2018joint,chen2019bert}.
\subsection{Dialog Policy for Teaching sessions}
\vspace{-2mm}
While there is relatively rare work for dialog policy models specifically meant for teaching sessions, a dialog policy in general can be learned using a) supervised policy learning, and (b) reinforcement policy learning. Existing supervised learning approaches usually considers it as an intent prediction problem. Recent work usually takes a joint prediction for both intents and slot prediction \cite{liu2016attention,goo2018slot,xu2013convolutional,zhang2018joint,chen2019bert}. Reinforcement learning approaches formulate the policy learning as sequential decision making process, and learns optimal policy from a large number of user interactions \cite{liu2017iterative,takanobu2019guided,zhou2017end,zhao2016towards}.
\section{Methodology}
\vspace{-2mm}
\label{sec:approach}
In this work, we incorporate a dedicated dialogue system called Teachable Dialogue System, which is specifically tailored for conducting interactive teaching sessions with users, for learning definitions of concepts that were previously not understood by the AI system. This dialogue system consists of three components:
\begin{itemize}[leftmargin=*]
\itemsep -0.5ex
\item Concept parser - that uses pre-trained embeddings and signals from the Conversational AI system to predict gaps in NLU’s understanding for interpreting a given utterance from the user,
\item Definition understanding model that extracts and interprets explanations from the user and maps the unknown concept phrases to the learnt definitions, and
\item Dialog policy model that drives teaching conversations with a goal-completion objective of extracting necessary explanations from the user and ground the respective unknown concepts.
\end{itemize}
\begin{figure}[t]
\centering
\includegraphics[width=0.9\textwidth]{overview_v3.png}
\caption{Architecture of AI system using a Teachable AI Dailog system}
\label{fig:main}
\vspace{-1.5em}
\end{figure}
The teachable dialogue system acts as a subsidiary to the multi-domain Conversational AI system, and springs into action when the AI agent is not able to successfully interpret/execute a user’s input utterance. This way, the parent Conversational AI system remains decoupled from the Teachable Dialogue System, staying independent of user-specific interpretation or modeling. Concept parser identifies the sequence of tokens (called concept phrases) that are not understood by the parent NLU system, and helps in initiating a teaching session with the user with the question - "{\em Can you teach me what you mean by <concept phrase>?}" Once a teaching session has been initiated with the user by the dialogue system, the dialog-policy model helps in predicting the right clarification question to ask the user, based on the context of the original utterance and the interactions with the user during the teaching session. The answers from the user are then processed by the definition understanding component that works along with the policy models to conduct multi-turn conversations with the user to learn, clarify and ground the respective concepts in the user’s utterance. When the Teachable Dialogue System deems a teaching session as successful, the taught actions are subsequently executed by the AI system, and the respective definitions and artifacts are then stored for future re-use. Figure~\ref{fig:main} describes the architecture of this Teachable AI system.
\vspace{-2mm}
\subsection{Concept Parser}
\vspace{-2mm}
Identifying gaps in NLU interpretation of an utterance, is a challenging task by itself. This task involves accurately localizing the segments within an utterance that an NLU system is unable to comprehend, and needed to execute the relevant action. In addition to using a slot-tagger objective in our Concept Parser model, we added an auxiliary task of semantic-chunking to make sure the model is sensitive to segments that might be neglected by a token-level slot tagger. For example, the chunking model should be able to segment the sentence: "[show] [me] [navigation] [to] [where we go camping every year]" while the slot tagger may not correctly tag "where we go camping every year" as "Location" concept.
\begin{figure}[tb]
\centering
\includegraphics[width=0.8\textwidth]
{cp_model_structure.png}
\caption{
\small
Concept Parser Model Structure: Transformer-based shared layer with Multi-Task Heads
}
\label{fig:cp}
\vspace{-1.0em}
\end{figure}
Additionally, there are several real-world challenges that a Concept Parser needs to address, such as: a) Tentative user behavior - such as "set the lights to, never mind" - where the user intends to cancel the request, rather than mean to set the lights to a user defined value, b) Unsupported actions such as "set the light to fifty degrees" where NLU interpretation is successful, but the Action Execution module cannot execute the respective action - this could be because of an unintentional user error, or due to an error in Automatic Speech Recognition (ASR). Concept Parser needs to treat the above scenarios as 'non teachable' instances, and hence should prevent the initiation of teaching sessions to learn the respective definitions from users. In addition to the task of identifying the concept phrase from a given utterance, Concept Parser also generates 'relevance scores' that indicates how confidence score for the concept phrase to be teachable.
To address the above challenges, we implemented a Concept Parser using a multi-task transformer model which is end-to-end trainable using the following task objectives: a) Slot tagging b) Semantic chunking, and c) Concept Relevance. The model structure is depicted in Figure \ref{fig:cp} , and is composed of the following sub-components:
\textbf{BERT layer}. It is a multi-layer multi-head self-attention Bert model \cite{bert}. The input sentence is first tokenized with WordPiece \cite{schuster2012japanese}. Then positional embedding and segment embedding are added to the token embedding as final input embedding $X = {x_1,x_2,...,x_L}$. The final output of bert layer $H = {h_1,h_2,...,h_L}$ is a contextual representation for each input token.
\textbf{Slot tagging head}. Given the bert output $H = {h_1,h_2,...,h_L}$, the slot tagging head feeds the output into a multi-layer perceptron followed by a softmax to predict slot label for each token.
\begin{equation}
z_i = f (W \cdot h_i + b)
\end{equation}
\begin{equation}
\hat{y}_{ij}(st) = \frac{e^{z_{ij}}}{\sum_{k=1}^{N_1} e^{z_{ik}} }
\end{equation}
Where $N_1$ is the number of slot classes and $L$ is the length of the utterance. Then the slot tagging loss $L_{st}$ will be a average cross-entropy loss across all tokens.
\begin{equation}
L_{st} = - \frac{1}{L \cdot N_1 } \sum_{i=1}^{L } \sum_{j=1}^{N_1} y_{ij}(st) \cdot log (\hat{y}_{ij}(st))
\end{equation}
\textbf{Chunking tagging head}. The chunking head has exact same structure as the slot tagging head. The only difference is the different $W$ and $b$ parameters, as well as the supervised chunking labels $y_i$. The loss $L_{ck}$ can be calculated as follows:
\begin{equation}
L_{ck} = - \frac{1}{L \cdot N_2 } \sum_{i=1}^{L } \sum_{j=1}^{N_2} y_{ij}(ck) \cdot log (\hat{y}_{ij}(ck))
\end{equation}
Where $N_2$ is the number of chunking classes and $L$ is the length of the utterance. $y_{ij}(ck)$ is ground-truth chunking labels, and $\hat{y_{ij}(ck)}$ is the predicted chunking labels. We also add an interweaving loss to enforce the "synchronization" between the slot tagging head and chunking head, by imposing a KL-divergence loss on the output of the two heads.
\begin{equation}
L_{kl} = - \frac{1}{L \cdot N}\sum_{i=1}^{L} \sum_{j=1}^{N} \hat{z}_{ij}(ck) \cdot log (\frac {\hat{z}_{ij}(ck)}{\hat{z}_{ij}(st)})
\end{equation}
\textbf{Relevance scoring head}. The relevance scoring head takes the bert output embedding of the ${[CLS]}$ token as the input, and feeds it into multi-layer perceptron followed by a binary cross-entropy loss:
\begin{equation}
z_{[CLS]} = f (W \cdot h_{[CLS]} + b)
\end{equation}
\begin{equation}
\hat{y}_{[CLS],i} = \frac{e^{z_{[CLS],i}}}{\sum_{j=1}^{2} e^{z_{[CLS],j} }}
\end{equation}
\begin{equation}
L_{rel} = - \frac{1}{2} \sum_{i=1}^{2} y_i(rel) \cdot log (\hat{y}_{[CLS],i})
\end{equation}
\textbf{Final loss objective}. The final loss objective $L_{CP}$ is a weighted sum of all four losses mentioned above, with $\alpha_1$, $\alpha_2$, $\alpha_3$ and $\alpha_4$ as the weights.
\begin{equation}
L_{CP} = \alpha_1 \cdot L_{st} + \alpha_2 \cdot L_{ck} + \alpha_3 \cdot L_{kl} + (1 -\alpha_1 - \alpha_2 - \alpha_3 ) \cdot L_{rel}
\end{equation}
\subsection{Definition Understanding}
\vspace{-2mm}
Once Concept Parser has detected a teachable phrase, a teaching session is initiated by requesting the user to provide a suitable definition of the identified concept. There are several real-world challenges involved in accurately extracting the appropriate definition from the user's answer, such as: a) Verbose answers: When posed with an impromptu definition of a concept, users generally tend to use verbose and indirect answers such as ``yeah i mean red color or may be just orange would do'' or ``i meant red not blue''; b) Distracted users: It is quite likely that users do not intend to answer the definition question, but would like to move on with another new request to the AI assistant, e.g. ``Whats the weather outside'', which is not really an answer to the question; c) Incomplete answers: Users might have provided an answer but its not yet sufficient to fully ground the respective definition and needs more clarification questions; d) Complicated answers: Users could provide a definition that sounds logical to another human, but the respective definition could still be not be understood by the parent AI system and hence cannot be grounded to the equivalent actions; e) Contextual references: such as ``make it brighter''.
\begin{figure}[t]
\centering
\includegraphics[width=0.75\textwidth]
{final_au_model_updated_final.png}
\vspace{-1.0em}
\caption{Definition Understanding model}
\label{fig:au}
\vspace{-1.5em}
\end{figure}
We developed a Definition Understanding model that can generically address the above scenarios. The model uses a Transformer-based joint intent and span detection framework that takes in the dialogue history ($\bm{z}$ = $z_1$, $z_2$, $\dots$, $z_{H}$), user answer ($\bm{x} = x_1, x_2, \dots, x_{A}$) and the slot type ($\bm{m} = m_1, m_2, \dots, m_{S}$), illustrated in Figure~\ref{fig:au}. We first concatenate user answer and dialogue history and pass the concatenated vector through a Transform-based encoder (e.g. BERT~\cite{bert}) and extract the encoder's last layer hidden states, denoted as $\bm{H} = (\bm{h}_{[CLS]}, \bm{h}_{x, 1}, \dots, \bm{h}_{x, A}, \bm{h}_{[SEP]}, \bm{h}_{z, 1}, \bm{h}_{z, 2}, \dots, \bm{h}_{z, H}, \bm{h}_{[SEP]})$. Separately, we obtain a slot type embedding, $\bm{e}_s$ with some pre-trained model (e.g. BERT). The contextual representation of each token is then fused with the slot type embedding
\begin{equation}
\bm{H} \oplus \bm{e}_s = (\bm{h}_{[CLS]} \oplus \bm{e}_s, \dots, \bm{h}_{x, i} \oplus \bm{e}_s, \dots \bm{h}_{z, j} \oplus \bm{e}_s, \dots, \bm{h}_{[SEP]} \oplus \bm{e}_s)
\end{equation}
and further gets passed through a set of post-Transformer encoder layers, e.g. feed-forward and CRF layers, to produce the final representation of each input token, $\bm{H}^O = (\bm{h}_{[CLS]}^O, \dots, \bm{h}_{x, i}^O, \dots \bm{h}_{z, j}^O, \dots, \bm{h}_{[SEP]}^O)$. Intent and span classification are performed with separate output layers and described below in details.
\textbf{Intent classification}. We adopt a single fully connected layer followed by a softmax layer on top of the final representation of ${[CLS]}$ token to perform the intent classification.
\begin{equation}
\hat{y}_i^{intent} = {Softmax}(f(\bm{W}^{intent} \cdot \bm{h}_{[CLS]}^O + \bm{b}^{intent}))
\end{equation}
We use cross-entropy to calculate the intent classification loss for $C$ intents.
\begin{equation}
L_{intent} = - \frac{1}{C} \sum_{i=1}^C y_i^{intent} \cdot \log({\hat{y}_i^{intent}})
\end{equation}
\textbf{Span classification}. We use the same layer structure to perform span classification as intent classification except that the fully connected layer is 3-way (i.e. ``B'', ``I'', ``O'').
\begin{equation}
\hat{y}_{i,k}^{span} = {Softmax}(f(\bm{W}_i^{intent} \cdot \bm{h}_{x, i}^O + \bm{b}_i^{span}))
\end{equation}
The span classification loss is also calculated using cross-entropy:
\begin{equation}
L_{span} = - \frac{1}{3A} \sum_{i=1}^A \sum_{k=1}^3 y_{i,k}^{span} \cdot \log({\hat{y}_{i,k}^{span}}) - \frac{1}{3H} \sum_{j=1}^H \sum_{k=1}^3 y_{j,k}^{span} \cdot \log({\hat{y}_{j,k}^{span}})
\end{equation}
\textbf{Joint loss}. The final loss objective $L_{DU}$ is a weighted sum of intent and span classification losses. The hyperparameter of relative intent loss proportion $\alpha_{intent}$ is furthered tuned on a validation dataset.
\begin{equation}
L_{DU} = \alpha_{intent} \cdot L_{intent} + (1 - \alpha_{intent}) \cdot L_{span}
\end{equation}
\subsection{Dialogue Policy}
\vspace{-2mm}
For the Teachable Dialogue system described in this paper, a Dialogue Policy predicts the next action in a multi-turn teaching dialogue session with a goal-completion target of extracting all the necessary definitions from the user during the teaching. A successful completion of the teaching session is one where the concept phrase in the first-turn utterance has been grounded to an equivalent action that is executable by the parent AI system.
The Dialogue Policy uses a Transformer-based contextual model to predict the next action during a teaching session. The action-space for our Dialogue Policy model includes a) Ask or repeat a clarification question, b) Guardrail the conversations to channelize users back into the teaching conversations, c) Identify Out-Of-Domain (OOD) turns during a teaching session, d) Ground the extracted definitions with the parent NLU system, and e) deem a teaching session as successful or unsuccessful, or decide to end a teaching session. The dialogue policy model takes in utterances from contextual dialogue history, predicted definition spans and confidence scores from Definition Understanding module, and the set of recognized/resolved slots from the parent NLU model and passes the input through a Transformer-based encoder to get the last layer output of ${[CLS]}$ token as a contextual representation of the inputs, $\bm{h}_{[CLS]}$. Then we apply a single feed-forward layer and a softmax layer to obtain a probability distribution over all the possible actions.
\begin{equation}
\hat{y}^{action} = {Softmax}(f(\bm{W}^{action} \cdot \bm{h}_{[CLS]} + \bm{b}^{action}))
\end{equation}
\section{Experiments and Results}
\label{sec:exp}
\vspace{-3mm}
\textbf{Datasets:} We evaluate our results on two datasets: 1) The dataset published by \cite{jia-2017} which consists of both personal concepts and generic slot values for five slot types namely date, time, location, people and restaurant-name. While the dataset in \cite{jia-2017} pertains to the problem that we have attempted to solve in this paper, we find that it does not include challenging real-world scenarios such as the ones described in Section~\ref{sec:approach} which are addressed by our models. We hence used a second internal dataset collected from crowd-sourcing which addressed the challenges of realistic multi-turn teaching sessions. Crowd workers provided both cooperative and non-cooperative inputs for two tasks:
\begin{itemize}[leftmargin=*]
\itemsep -0.5ex
\item Generate first-turn utterances containing both personal and non-personal concept phrases for performing actions supported by the AI assistant spanning multiple domains. Each utterance was annotated with the ground-truth concept phrase. For the "not teachable" class, we synthesize a wide variety of cases, including utterances without any concept, out-of-domain utterance, ill-grammar and incomplete utterances.
\item Answers to clarification questions within teaching sessions. Answers were annotated with the concept definition phrases along with additional information like direct-answer, new-request, etc.
\end{itemize}
\textbf{Concept parser experiment setting}: For concept parser model, a pre-trained BERT-base model is used as the base layer, and it is fine-tuned together with all the head layers during training \cite{bert}. We used a 300 dimensional hidden-layer in MLP, a RELU activation function \cite{nair2010rectified}, and a learning rate of 1e-5. For the public dataset, the model is simultaneously trained on $L_{st}$, $L_{ck}$ and $L_{kl}$ losses without relevance scoring for 20 epochs, where weights are set to 0.5/0.5/2.0 empirically. For our internal dataset, the model is first trained on the $L_{ck}$ loss for 2 epochs, and then fine-tuned on $L_{st}$ and $L_{rel}$ losses with 1.0/1.0 weights for 2 epochs. For the auxiliary chunking label, we use a fusion of semantic role labeling model(SRL) and constituency parsing model (CPM) models from AllenNLP \footnote{https://allennlp.org/}.
\textbf{Definition understanding experiment setting}: We experimented with a few pre-trained transformer-based models, namely BERT~\cite{bert}, RoBERTa~\cite{roberta}, ALBERT~\cite{albert} as the base encoder. We used a single feed-forward layer with hidden state set to 100 and RELU \cite{nair2010rectified} activation function followed by a CRF layer for post-Transformer encoder. The initial learning rate is set to $1 \times 10^{-4}$ and reduces to half at loss plateau, i.e. no loss reduction within last consecutive 10 epochs. We used Adam optimizer~\cite{adam} with weight decay enabled.
\subsection{Performance evaluations}
\vspace{-2mm}
\subsubsection{Concept Parser results}
\vspace{-2mm}
\textbf{Public Datasets}: Table~\ref{tab1} shows the performance of our concept parser on personalized test set from \cite{jia-2017} using phrase-level metrics on slots with unknown concepts. From Table~\ref{tab1} we observe the following: a) When there is no personalized training data available (Zero-shot), there is an absolute improvement in F1-score of 8.6\% for single-task model, and 15\% using Multi-task model. This supports our idea that an end-to-end model with built-in objectives for both slot tagging and semantic-chunking can be more robust in this zero-shot unknown concept prediction setting. b) When personalized concept phrases are added to the training data, the model performance jumps to 70\% and 76\% for single-task and multi-task models respectively. Some source of errors still remains due to annotation inconsistencies ("after lunch" or "lunch" as concept).
\begin{table}
\begin{minipage}[t]{0.5\linewidth}
\begin{tabular}[t]{ lcccc }
\cmidrule[\heavyrulewidth]{1-4}
\textbf{Model} & \textbf{Precision} & \textbf{Recall} & \textbf{F1}\\[0.2em]
\cmidrule[\heavyrulewidth]{1-4}
\textbf{Zero-shot} \\
Jia {\em et al.}~\cite{jia-2017} & 41.1\% & 40.0\% & 40.3\%\\
\cmidrule{2-5}
Single-Task BERT & 44.0\% & 55.1\% & 48.9\%\\
\cmidrule{2-5}
Multi-Task BERT & \textbf{50.5}\% & \textbf{61.3}\% & \textbf{55.3}\%\\
\cmidrule[\heavyrulewidth]{1-4}
\textbf{Supervised} \\
Jia {\em et al.}~\cite{jia-2017} & N/A & N/A & N/A\\[0.2em]
\cmidrule{2-5}
Single-Task BERT & 62.2\% & 80.0\% & 70.0\%\\[0.2em]
\cmidrule{2-5}
Multi-Task BERT & \textbf{69.1}\% & \textbf{86.0}\% & \textbf{76.6}\%\\[0.2em]
\cmidrule[\heavyrulewidth]{1-4}
\end{tabular}
\caption{Concept Parser performance on unknown concepts on Personalized Test.}
\label{tab1}
\end{minipage}%
\hspace{3em}
\begin{minipage}[t]{0.45\linewidth}
\begin{tabular}[t]{ p{1.8cm}p{1cm}p{1cm}p{1cm} }
\cmidrule[\heavyrulewidth]{1-4}
\textbf{\space\space\space Model} & \textbf{Precision} & \textbf{Recall}& \textbf{F1}\\[0.2em]
\cmidrule{1-4}
\cmidrule{1-4}
Single-Task BERT & 90.86 & 91.72 & 91.29 \\[0.2em]
\cmidrule{1-4}
Multi-Task BERT & 93.24 & 84.86 & 88.86 \\[0.2em]
\cmidrule[\heavyrulewidth]{1-4}
\end{tabular}
\caption{Concept Parser performance on internal dataset using models with additional Relevance head}
\label{tab2}
\end{minipage}
\end{table}
\textbf{Internal Datasets}: Table~\ref{tab2} reports the performance of our concept parser on the internal dataset. From the table we can observe the following. Our single-task and multi-task models, both trained on personalized data, achieve competitive performance on the synthesized evaluation data. When personalized data are available in training data, single-task model performs already decently (91.29\% and 88.86\%); yet multi-task model could still achieve higher precision. We would like to highlight here that our internal evaluation dataset is also more challenging than the public dataset, including a wide variety of "not teachable" examples including out-of-domain requests, regular requests without concept and ill-grammar and incomplete utterances.
\vspace{-0.75em}
\subsubsection{Definition understanding results}
\vspace{-0.5em}
\textbf{Public Datasets}: Table~\ref{tab3} shows the performance of our definition understanding model on personalized test set from \cite{jia-2017} with three Transformer-based encoders. From the results, we observe that model with RoBERTa~\cite{roberta} encoder achieves the best performance in terms of precision, recall and F1-score. BERT~\cite{bert} and ALBERT~\cite{albert} have similar performance and both are lightly behind RoBERTa~\cite{roberta}. We also did an error analysis and found that nearly half of the span prediction errors are due to a mismatch on preposition (e.g. "on Sept 15" vs "Sept 15"), which is not unexpected since span annotations in \cite{jia-2017} are not consistent on the inclusion of prepositions.
\textbf{Internal Datasets}: The performance of our concept parser on the internal dataset is reported in Table~\ref{tab4}. Again, we observe that model with RoBERTa~\cite{roberta} encoder slightly outperforms model with BERT~\cite{bert} and ALBERT~\cite{albert}.
\begin{table}
\begin{minipage}[t]{0.5\linewidth}
\centering
\begin{tabular}{lccc}
\cmidrule[\heavyrulewidth]{1-4}
\textbf{Model} & \textbf{Precision} & \textbf{Recall}& \textbf{F1-score}\\[0.2em]
\cmidrule{1-4}
RoBERTa & \textbf{91.80} & \textbf{90.89} & \textbf{91.34}\\[0.2em]
\cmidrule{1-4}
BERT & 90.71 & 90.01 & 90.36 \\[0.2em]
\cmidrule{1-4}
ALBERT & 89.93 & 90.53 & 90.23\\[0.2em]
\cmidrule[\heavyrulewidth]{1-4}
\end{tabular}
\caption{Definition Understanding results on Jia {\em et al.} \cite{jia-2017} dataset}
\label{tab3}
\end{minipage}%
\hspace{2em}
\begin{minipage}[t]{0.5\linewidth}
\centering
\begin{tabular}{lccc}
\cmidrule[\heavyrulewidth]{1-4}
\textbf{Model} & \textbf{Precision} & \textbf{Recall}& \textbf{F1-score}\\[0.2em]
\cmidrule{1-4}
RoBERTa & \textbf{95.92} & \textbf{96.81} & \textbf{96.36}\\[0.2em]
\cmidrule{1-4}
BERT & 95.57 & 96.03 & 95.80\\[0.2em]
\cmidrule{1-4}
ALBERT & 95.41 & 95.08 & 95.24\\[0.2em]
\cmidrule[\heavyrulewidth]{1-4}
\end{tabular}
\centering
\caption{Definition Understanding results on internal dataset}
\vspace{-1.2em}
\label{tab4}
\end{minipage}
\end{table}
\subsubsection{Policy Model results}
\vspace{-2mm}
Table~\ref{tab5} shows the performance of our Transformer-based policy model along with heuristics based on NLU grounding results, on an internal dataset with annotated ground-truth actions. We observe that model with RoBERTa~\cite{roberta} encoder achieves the best performance in terms of precision, recall and F1-score. ALBERT~\cite{albert} showed comparable performance to BERT~\cite{bert} but ran the fastest.
\begin{table}[h]
\centering
\begin{tabular}{lccccccccc}
\cmidrule[\heavyrulewidth]{1-5}
\textbf{\space\space\space Model} & \textbf{Precision} & \textbf{Recall}& \textbf{F1}& \textbf{mins/epoch}\\[0.2em]
\cmidrule{1-5}
RoBERTa & \textbf{98.78} & \textbf{97.71} & \textbf{98.24} & 6.12\\[0.2em]
\cmidrule{1-5}
BERT & 98.73 & 96.26 & 97.48 & 5.11\\[0.2em]
\cmidrule{1-5}
ALBERT & 97.98 & 96.12 & 97.04 & 4.56\\[0.2em]
\cmidrule[\heavyrulewidth]{1-5}
\end{tabular}
\caption{Policy Model performance evaluation results on internal dataset}
\vspace{-1.2em}
\label{tab5}
\end{table}
\vspace{-3mm}
\section{Conclusion}
\vspace{-3mm}
We have presented a teachable dialogue system that uses neural models for gap identification, definition understanding and dialogue policy prediction to conduct interactive teaching sessions with the user to learn and re-use definitions of concepts that are unknown to a conversational AI system. This Teachable Dialogue system helps in automatically improving the understanding capabilities of the AI system to hold more natural conversations with the end-users, and progressively improve the agents’ understanding of users' parlance to enable more natural ways of interaction with a conversational AI system. We believe this is an effort towards building truly interactive learning systems, and plan to extend the scope to higher level capabilities such as intent teaching.
\bibliographystyle{plain}
\section{Introduction}
\vspace{-4mm}
Humans are adaptive by nature. While it is very natural for humans to ask clarifying questions and immediately correct course whenever there is a misunderstanding, interactive learning using self-supervision has been a holy grail area for Conversational AI. Most Conversational AI systems rely on knowledge-based or machine-learning based (or hybrid) understanding components under the hood to understand the intents of the users’ and their arguments (usually called as slots), such as “{\em set an alarm for 7 am}”, where the intent can be {\em set\_alarm}, and {\em 7 am} can be the time slot. These systems typically do not have the capability to interact with end-users during a live conversation to seek explanations to improve the AI system's understanding of entities, intents or other constructs. When users speak an utterance outside the comprehension of these systems, the AI agent responds with a dead-end response such as ‘Sorry I don’t know that’.
In this paper we describe a teachable AI system towards enabling users to directly teach a conversational AI agent via a live interactive teaching session using natural language explanations. Learning concept definitions using interactive voice is a challenging problem due to a number of reasons, including wide vocabulary of words that users could use to describe definitions, user distractions during teaching sessions, grounding of related entities, fuzziness in the re-use of previously taught concepts across domains, etc. The method described in this paper augments goal-oriented AI agents with an interactive teaching capability, to improve task completion rates by asking questions to fill gaps in the AI agent’s understanding, hence making the agent more personal, conversational and adaptive.
More specifically, the AI agent will be checking the utterances whether it is “Teachable” using a deep learning based parser, which not only classifies the utterance, but also identifies the entity or intent which needs to be learnt from the users. If so, the “Classroom” sub-dialogue kicks in, where the users are asked what they mean by these concepts. This sub-dialogue is driven by a dedicated dialogue manager policy model which incorporates a definition understanding model to interpret users' responses. For example, if the utterance is "{\em set an alarm for my baseball practice}", the system would ask "{\em when is your baseball practice?}".
Despite a few academic papers towards learning directly from users as presented in Section~\ref{sec:rw}, to the best of our knowledge, there is no work that covers all of these novel contributions:
\begin{itemize}
\itemsep -0.5ex
\item A dedicated multi-turn domain-agnostic dialogue system specifically tailored for interactive teaching, that augments an existing Conversational AI System to learning explanations from user in real-time.
\item A multi-task neural Concept Parser that automatically identify gaps in an AI agent’s understanding, using a multi-task model that incorporates semantic role labeling and constituency parser.
\item A neural Definition Understanding system along with a policy model to aid in conducting robust teaching sessions with the user.
\end{itemize}
The following sections are organized as follows. Section~\ref{sec:rw} describes the related work in the area of interactive learning methods to learn explanations, and predicting gaps in the AI agent’s understanding for a given user input; Section~\ref{sec:approach} describes our architecture and modeling methods, and Section~\ref{sec:exp} describes our experimental evaluations.
\section{Related Work}
\vspace{-2mm}
\label{sec:rw}
The related work on learning via user-in-the-loop for Conversational AI systems can be analyzed in two categories. The higher level interactive learning approaches, and utterance parsing based gap prediction studies.
\subsection{Interactive Learning}
\vspace{-2mm}
Existing work on interactive learning methods can be grouped into 3 categories:
\begin{itemize}
\item Game learning using limited/no vocabulary: Several methods have been proposed for learning a game (such as Hanoi) using interactive sessions with the user Kirk~{\em et al.}~\cite{kirk-19} Wang~{\em et al.}\cite{wang-17}. Such methods represent the game scene as a symbolic representation of attributes (such as shape, color) which is used to learn the state progression, to learn the game.
\item Neural code synthesis methods: Methods like Yin~{\em et al.}~\cite{yin-2020} aim at converting user utterances directly into a coding language like Python. To our knowledge, such syntax-driven code generation methods are still in their early stages, and not yet ready for integration into mainstream conversational AI agents that cover a wide variety of domains.
\item Learning by GUI demonstrations: GUI based methods use semantic parsers to learn conditional expressions and their associated actions. One notable study is by Allen~{\em et al.} using a web browser to teach tasks such as buying a book from Amazon~\cite{plow}.
\end{itemize}
\subsection{NLU gap prediction}
\vspace{-2mm}
The problem of predicting segments of an utterance which a base NLU system could not interpret, was tackled by Jia {\em et al.}~\cite{jia-2017}, by using a set of post-processing rules on top of the base NLU's slot-filling model. The rules were built on heuristics from the base NLU model, such as confidence score threshold, out-of-vocabulary flag, and span-expansion based on syntactic dependency tree. This paper also introduced a new dataset for this problem that is publicly available, which is used in our work as one of the datasets for evaluating model performances. In \cite{Kobayashi-2018,kobayashi2019slot} the authors proposed augmenting the training data for the slot-filling model, by injecting noise tokens into the regular slot values of training data, to force the model to learn about the context of the respective slots. This augmentation is done to improve the model robustness in predicting the correct span, on unseen slot values. One variant of such approach uses negative sampling to train the model to identify unknown slot values together with a joint slot tagging and slot classifier~\cite{Hou-2019}. Another line of work focuses on different model structures such as pointer-networks, to better copy unknown slot values from the utterance into final slot value prediction~\cite{xu-2018,Zhao-2018,Yang-2020}. This line of work usually considers the problem under a dialogue state tracking setting, therefore only one slot value is predicted each time.
\subsection{Definition Understanding}
\vspace{-2mm}
Our task of understanding the definition of concepts from users' explanations, generally falls into the categories of reading comprehension and slot filling. For reading comprehension, there is a particular thread that formulates question answering as span prediction. Those works typically learn a representation for the question and passage, and then predict the start and end of the answer with attention mechanism between question, passage and answer \cite{chen2017reading,chen2018neural}. The slot filling thread work focuses on sequential labeling of an answer\cite{mesnil2014using}, with most recent works utilizing neural slot-fillers with joint intent classifiers \cite{liu2016attention,goo2018slot,xu2013convolutional,zhang2018joint,chen2019bert}.
\subsection{Dialog Policy for Teaching sessions}
\vspace{-2mm}
While there is relatively rare work for dialog policy models specifically meant for teaching sessions, a dialog policy in general can be learned using a) supervised policy learning, and (b) reinforcement policy learning. Existing supervised learning approaches usually considers it as an intent prediction problem. Recent work usually takes a joint prediction for both intents and slot prediction \cite{liu2016attention,goo2018slot,xu2013convolutional,zhang2018joint,chen2019bert}. Reinforcement learning approaches formulate the policy learning as sequential decision making process, and learns optimal policy from a large number of user interactions \cite{liu2017iterative,takanobu2019guided,zhou2017end,zhao2016towards}.
\section{Methodology}
\vspace{-2mm}
\label{sec:approach}
In this work, we incorporate a dedicated dialogue system called Teachable Dialogue System, which is specifically tailored for conducting interactive teaching sessions with users, for learning definitions of concepts that were previously not understood by the AI system. This dialogue system consists of three components:
\begin{itemize}[leftmargin=*]
\itemsep -0.5ex
\item Concept parser - that uses pre-trained embeddings and signals from the Conversational AI system to predict gaps in NLU’s understanding for interpreting a given utterance from the user,
\item Definition understanding model that extracts and interprets explanations from the user and maps the unknown concept phrases to the learnt definitions, and
\item Dialog policy model that drives teaching conversations with a goal-completion objective of extracting necessary explanations from the user and ground the respective unknown concepts.
\end{itemize}
\begin{figure}[t]
\centering
\includegraphics[width=0.9\textwidth]{overview_v3.png}
\caption{Architecture of AI system using a Teachable AI Dailog system}
\label{fig:main}
\vspace{-1.5em}
\end{figure}
The teachable dialogue system acts as a subsidiary to the multi-domain Conversational AI system, and springs into action when the AI agent is not able to successfully interpret/execute a user’s input utterance. This way, the parent Conversational AI system remains decoupled from the Teachable Dialogue System, staying independent of user-specific interpretation or modeling. Concept parser identifies the sequence of tokens (called concept phrases) that are not understood by the parent NLU system, and helps in initiating a teaching session with the user with the question - "{\em Can you teach me what you mean by <concept phrase>?}" Once a teaching session has been initiated with the user by the dialogue system, the dialog-policy model helps in predicting the right clarification question to ask the user, based on the context of the original utterance and the interactions with the user during the teaching session. The answers from the user are then processed by the definition understanding component that works along with the policy models to conduct multi-turn conversations with the user to learn, clarify and ground the respective concepts in the user’s utterance. When the Teachable Dialogue System deems a teaching session as successful, the taught actions are subsequently executed by the AI system, and the respective definitions and artifacts are then stored for future re-use. Figure~\ref{fig:main} describes the architecture of this Teachable AI system.
\vspace{-2mm}
\subsection{Concept Parser}
\vspace{-2mm}
Identifying gaps in NLU interpretation of an utterance, is a challenging task by itself. This task involves accurately localizing the segments within an utterance that an NLU system is unable to comprehend, and needed to execute the relevant action. In addition to using a slot-tagger objective in our Concept Parser model, we added an auxiliary task of semantic-chunking to make sure the model is sensitive to segments that might be neglected by a token-level slot tagger. For example, the chunking model should be able to segment the sentence: "[show] [me] [navigation] [to] [where we go camping every year]" while the slot tagger may not correctly tag "where we go camping every year" as "Location" concept.
\begin{figure}[tb]
\centering
\includegraphics[width=0.8\textwidth]
{cp_model_structure.png}
\caption{
\small
Concept Parser Model Structure: Transformer-based shared layer with Multi-Task Heads
}
\label{fig:cp}
\vspace{-1.0em}
\end{figure}
Additionally, there are several real-world challenges that a Concept Parser needs to address, such as: a) Tentative user behavior - such as "set the lights to, never mind" - where the user intends to cancel the request, rather than mean to set the lights to a user defined value, b) Unsupported actions such as "set the light to fifty degrees" where NLU interpretation is successful, but the Action Execution module cannot execute the respective action - this could be because of an unintentional user error, or due to an error in Automatic Speech Recognition (ASR). Concept Parser needs to treat the above scenarios as 'non teachable' instances, and hence should prevent the initiation of teaching sessions to learn the respective definitions from users. In addition to the task of identifying the concept phrase from a given utterance, Concept Parser also generates 'relevance scores' that indicates how confidence score for the concept phrase to be teachable.
To address the above challenges, we implemented a Concept Parser using a multi-task transformer model which is end-to-end trainable using the following task objectives: a) Slot tagging b) Semantic chunking, and c) Concept Relevance. The model structure is depicted in Figure \ref{fig:cp} , and is composed of the following sub-components:
\textbf{BERT layer}. It is a multi-layer multi-head self-attention Bert model \cite{bert}. The input sentence is first tokenized with WordPiece \cite{schuster2012japanese}. Then positional embedding and segment embedding are added to the token embedding as final input embedding $X = {x_1,x_2,...,x_L}$. The final output of bert layer $H = {h_1,h_2,...,h_L}$ is a contextual representation for each input token.
\textbf{Slot tagging head}. Given the bert output $H = {h_1,h_2,...,h_L}$, the slot tagging head feeds the output into a multi-layer perceptron followed by a softmax to predict slot label for each token.
\begin{equation}
z_i = f (W \cdot h_i + b)
\end{equation}
\begin{equation}
\hat{y}_{ij}(st) = \frac{e^{z_{ij}}}{\sum_{k=1}^{N_1} e^{z_{ik}} }
\end{equation}
Where $N_1$ is the number of slot classes and $L$ is the length of the utterance. Then the slot tagging loss $L_{st}$ will be a average cross-entropy loss across all tokens.
\begin{equation}
L_{st} = - \frac{1}{L \cdot N_1 } \sum_{i=1}^{L } \sum_{j=1}^{N_1} y_{ij}(st) \cdot log (\hat{y}_{ij}(st))
\end{equation}
\textbf{Chunking tagging head}. The chunking head has exact same structure as the slot tagging head. The only difference is the different $W$ and $b$ parameters, as well as the supervised chunking labels $y_i$. The loss $L_{ck}$ can be calculated as follows:
\begin{equation}
L_{ck} = - \frac{1}{L \cdot N_2 } \sum_{i=1}^{L } \sum_{j=1}^{N_2} y_{ij}(ck) \cdot log (\hat{y}_{ij}(ck))
\end{equation}
Where $N_2$ is the number of chunking classes and $L$ is the length of the utterance. $y_{ij}(ck)$ is ground-truth chunking labels, and $\hat{y_{ij}(ck)}$ is the predicted chunking labels. We also add an interweaving loss to enforce the "synchronization" between the slot tagging head and chunking head, by imposing a KL-divergence loss on the output of the two heads.
\begin{equation}
L_{kl} = - \frac{1}{L \cdot N}\sum_{i=1}^{L} \sum_{j=1}^{N} \hat{z}_{ij}(ck) \cdot log (\frac {\hat{z}_{ij}(ck)}{\hat{z}_{ij}(st)})
\end{equation}
\textbf{Relevance scoring head}. The relevance scoring head takes the bert output embedding of the ${[CLS]}$ token as the input, and feeds it into multi-layer perceptron followed by a binary cross-entropy loss:
\begin{equation}
z_{[CLS]} = f (W \cdot h_{[CLS]} + b)
\end{equation}
\begin{equation}
\hat{y}_{[CLS],i} = \frac{e^{z_{[CLS],i}}}{\sum_{j=1}^{2} e^{z_{[CLS],j} }}
\end{equation}
\begin{equation}
L_{rel} = - \frac{1}{2} \sum_{i=1}^{2} y_i(rel) \cdot log (\hat{y}_{[CLS],i})
\end{equation}
\textbf{Final loss objective}. The final loss objective $L_{CP}$ is a weighted sum of all four losses mentioned above, with $\alpha_1$, $\alpha_2$, $\alpha_3$ and $\alpha_4$ as the weights.
\begin{equation}
L_{CP} = \alpha_1 \cdot L_{st} + \alpha_2 \cdot L_{ck} + \alpha_3 \cdot L_{kl} + (1 -\alpha_1 - \alpha_2 - \alpha_3 ) \cdot L_{rel}
\end{equation}
\subsection{Definition Understanding}
\vspace{-2mm}
Once Concept Parser has detected a teachable phrase, a teaching session is initiated by requesting the user to provide a suitable definition of the identified concept. There are several real-world challenges involved in accurately extracting the appropriate definition from the user's answer, such as: a) Verbose answers: When posed with an impromptu definition of a concept, users generally tend to use verbose and indirect answers such as ``yeah i mean red color or may be just orange would do'' or ``i meant red not blue''; b) Distracted users: It is quite likely that users do not intend to answer the definition question, but would like to move on with another new request to the AI assistant, e.g. ``Whats the weather outside'', which is not really an answer to the question; c) Incomplete answers: Users might have provided an answer but its not yet sufficient to fully ground the respective definition and needs more clarification questions; d) Complicated answers: Users could provide a definition that sounds logical to another human, but the respective definition could still be not be understood by the parent AI system and hence cannot be grounded to the equivalent actions; e) Contextual references: such as ``make it brighter''.
\begin{figure}[t]
\centering
\includegraphics[width=0.75\textwidth]
{final_au_model_updated_final.png}
\vspace{-1.0em}
\caption{Definition Understanding model}
\label{fig:au}
\vspace{-1.5em}
\end{figure}
We developed a Definition Understanding model that can generically address the above scenarios. The model uses a Transformer-based joint intent and span detection framework that takes in the dialogue history ($\bm{z}$ = $z_1$, $z_2$, $\dots$, $z_{H}$), user answer ($\bm{x} = x_1, x_2, \dots, x_{A}$) and the slot type ($\bm{m} = m_1, m_2, \dots, m_{S}$), illustrated in Figure~\ref{fig:au}. We first concatenate user answer and dialogue history and pass the concatenated vector through a Transform-based encoder (e.g. BERT~\cite{bert}) and extract the encoder's last layer hidden states, denoted as $\bm{H} = (\bm{h}_{[CLS]}, \bm{h}_{x, 1}, \dots, \bm{h}_{x, A}, \bm{h}_{[SEP]}, \bm{h}_{z, 1}, \bm{h}_{z, 2}, \dots, \bm{h}_{z, H}, \bm{h}_{[SEP]})$. Separately, we obtain a slot type embedding, $\bm{e}_s$ with some pre-trained model (e.g. BERT). The contextual representation of each token is then fused with the slot type embedding
\begin{equation}
\bm{H} \oplus \bm{e}_s = (\bm{h}_{[CLS]} \oplus \bm{e}_s, \dots, \bm{h}_{x, i} \oplus \bm{e}_s, \dots \bm{h}_{z, j} \oplus \bm{e}_s, \dots, \bm{h}_{[SEP]} \oplus \bm{e}_s)
\end{equation}
and further gets passed through a set of post-Transformer encoder layers, e.g. feed-forward and CRF layers, to produce the final representation of each input token, $\bm{H}^O = (\bm{h}_{[CLS]}^O, \dots, \bm{h}_{x, i}^O, \dots \bm{h}_{z, j}^O, \dots, \bm{h}_{[SEP]}^O)$. Intent and span classification are performed with separate output layers and described below in details.
\textbf{Intent classification}. We adopt a single fully connected layer followed by a softmax layer on top of the final representation of ${[CLS]}$ token to perform the intent classification.
\begin{equation}
\hat{y}_i^{intent} = {Softmax}(f(\bm{W}^{intent} \cdot \bm{h}_{[CLS]}^O + \bm{b}^{intent}))
\end{equation}
We use cross-entropy to calculate the intent classification loss for $C$ intents.
\begin{equation}
L_{intent} = - \frac{1}{C} \sum_{i=1}^C y_i^{intent} \cdot \log({\hat{y}_i^{intent}})
\end{equation}
\textbf{Span classification}. We use the same layer structure to perform span classification as intent classification except that the fully connected layer is 3-way (i.e. ``B'', ``I'', ``O'').
\begin{equation}
\hat{y}_{i,k}^{span} = {Softmax}(f(\bm{W}_i^{intent} \cdot \bm{h}_{x, i}^O + \bm{b}_i^{span}))
\end{equation}
The span classification loss is also calculated using cross-entropy:
\begin{equation}
L_{span} = - \frac{1}{3A} \sum_{i=1}^A \sum_{k=1}^3 y_{i,k}^{span} \cdot \log({\hat{y}_{i,k}^{span}}) - \frac{1}{3H} \sum_{j=1}^H \sum_{k=1}^3 y_{j,k}^{span} \cdot \log({\hat{y}_{j,k}^{span}})
\end{equation}
\textbf{Joint loss}. The final loss objective $L_{DU}$ is a weighted sum of intent and span classification losses. The hyperparameter of relative intent loss proportion $\alpha_{intent}$ is furthered tuned on a validation dataset.
\begin{equation}
L_{DU} = \alpha_{intent} \cdot L_{intent} + (1 - \alpha_{intent}) \cdot L_{span}
\end{equation}
\subsection{Dialogue Policy}
\vspace{-2mm}
For the Teachable Dialogue system described in this paper, a Dialogue Policy predicts the next action in a multi-turn teaching dialogue session with a goal-completion target of extracting all the necessary definitions from the user during the teaching. A successful completion of the teaching session is one where the concept phrase in the first-turn utterance has been grounded to an equivalent action that is executable by the parent AI system.
The Dialogue Policy uses a Transformer-based contextual model to predict the next action during a teaching session. The action-space for our Dialogue Policy model includes a) Ask or repeat a clarification question, b) Guardrail the conversations to channelize users back into the teaching conversations, c) Identify Out-Of-Domain (OOD) turns during a teaching session, d) Ground the extracted definitions with the parent NLU system, and e) deem a teaching session as successful or unsuccessful, or decide to end a teaching session. The dialogue policy model takes in utterances from contextual dialogue history, predicted definition spans and confidence scores from Definition Understanding module, and the set of recognized/resolved slots from the parent NLU model and passes the input through a Transformer-based encoder to get the last layer output of ${[CLS]}$ token as a contextual representation of the inputs, $\bm{h}_{[CLS]}$. Then we apply a single feed-forward layer and a softmax layer to obtain a probability distribution over all the possible actions.
\begin{equation}
\hat{y}^{action} = {Softmax}(f(\bm{W}^{action} \cdot \bm{h}_{[CLS]} + \bm{b}^{action}))
\end{equation}
\section{Experiments and Results}
\label{sec:exp}
\vspace{-3mm}
\textbf{Datasets:} We evaluate our results on two datasets: 1) The dataset published by \cite{jia-2017} which consists of both personal concepts and generic slot values for five slot types namely date, time, location, people and restaurant-name. While the dataset in \cite{jia-2017} pertains to the problem that we have attempted to solve in this paper, we find that it does not include challenging real-world scenarios such as the ones described in Section~\ref{sec:approach} which are addressed by our models. We hence used a second internal dataset collected from crowd-sourcing which addressed the challenges of realistic multi-turn teaching sessions. Crowd workers provided both cooperative and non-cooperative inputs for two tasks:
\begin{itemize}[leftmargin=*]
\itemsep -0.5ex
\item Generate first-turn utterances containing both personal and non-personal concept phrases for performing actions supported by the AI assistant spanning multiple domains. Each utterance was annotated with the ground-truth concept phrase. For the "not teachable" class, we synthesize a wide variety of cases, including utterances without any concept, out-of-domain utterance, ill-grammar and incomplete utterances.
\item Answers to clarification questions within teaching sessions. Answers were annotated with the concept definition phrases along with additional information like direct-answer, new-request, etc.
\end{itemize}
\textbf{Concept parser experiment setting}: For concept parser model, a pre-trained BERT-base model is used as the base layer, and it is fine-tuned together with all the head layers during training \cite{bert}. We used a 300 dimensional hidden-layer in MLP, a RELU activation function \cite{nair2010rectified}, and a learning rate of 1e-5. For the public dataset, the model is simultaneously trained on $L_{st}$, $L_{ck}$ and $L_{kl}$ losses without relevance scoring for 20 epochs, where weights are set to 0.5/0.5/2.0 empirically. For our internal dataset, the model is first trained on the $L_{ck}$ loss for 2 epochs, and then fine-tuned on $L_{st}$ and $L_{rel}$ losses with 1.0/1.0 weights for 2 epochs. For the auxiliary chunking label, we use a fusion of semantic role labeling model(SRL) and constituency parsing model (CPM) models from AllenNLP \footnote{https://allennlp.org/}.
\textbf{Definition understanding experiment setting}: We experimented with a few pre-trained transformer-based models, namely BERT~\cite{bert}, RoBERTa~\cite{roberta}, ALBERT~\cite{albert} as the base encoder. We used a single feed-forward layer with hidden state set to 100 and RELU \cite{nair2010rectified} activation function followed by a CRF layer for post-Transformer encoder. The initial learning rate is set to $1 \times 10^{-4}$ and reduces to half at loss plateau, i.e. no loss reduction within last consecutive 10 epochs. We used Adam optimizer~\cite{adam} with weight decay enabled.
\subsection{Performance evaluations}
\vspace{-2mm}
\subsubsection{Concept Parser results}
\vspace{-2mm}
\textbf{Public Datasets}: Table~\ref{tab1} shows the performance of our concept parser on personalized test set from \cite{jia-2017} using phrase-level metrics on slots with unknown concepts. From Table~\ref{tab1} we observe the following: a) When there is no personalized training data available (Zero-shot), there is an absolute improvement in F1-score of 8.6\% for single-task model, and 15\% using Multi-task model. This supports our idea that an end-to-end model with built-in objectives for both slot tagging and semantic-chunking can be more robust in this zero-shot unknown concept prediction setting. b) When personalized concept phrases are added to the training data, the model performance jumps to 70\% and 76\% for single-task and multi-task models respectively. Some source of errors still remains due to annotation inconsistencies ("after lunch" or "lunch" as concept).
\begin{table}
\begin{minipage}[t]{0.5\linewidth}
\begin{tabular}[t]{ lcccc }
\cmidrule[\heavyrulewidth]{1-4}
\textbf{Model} & \textbf{Precision} & \textbf{Recall} & \textbf{F1}\\[0.2em]
\cmidrule[\heavyrulewidth]{1-4}
\textbf{Zero-shot} \\
Jia {\em et al.}~\cite{jia-2017} & 41.1\% & 40.0\% & 40.3\%\\
\cmidrule{2-5}
Single-Task BERT & 44.0\% & 55.1\% & 48.9\%\\
\cmidrule{2-5}
Multi-Task BERT & \textbf{50.5}\% & \textbf{61.3}\% & \textbf{55.3}\%\\
\cmidrule[\heavyrulewidth]{1-4}
\textbf{Supervised} \\
Jia {\em et al.}~\cite{jia-2017} & N/A & N/A & N/A\\[0.2em]
\cmidrule{2-5}
Single-Task BERT & 62.2\% & 80.0\% & 70.0\%\\[0.2em]
\cmidrule{2-5}
Multi-Task BERT & \textbf{69.1}\% & \textbf{86.0}\% & \textbf{76.6}\%\\[0.2em]
\cmidrule[\heavyrulewidth]{1-4}
\end{tabular}
\caption{Concept Parser performance on unknown concepts on Personalized Test.}
\label{tab1}
\end{minipage}%
\hspace{3em}
\begin{minipage}[t]{0.45\linewidth}
\begin{tabular}[t]{ p{1.8cm}p{1cm}p{1cm}p{1cm} }
\cmidrule[\heavyrulewidth]{1-4}
\textbf{\space\space\space Model} & \textbf{Precision} & \textbf{Recall}& \textbf{F1}\\[0.2em]
\cmidrule{1-4}
\cmidrule{1-4}
Single-Task BERT & 90.86 & 91.72 & 91.29 \\[0.2em]
\cmidrule{1-4}
Multi-Task BERT & 93.24 & 84.86 & 88.86 \\[0.2em]
\cmidrule[\heavyrulewidth]{1-4}
\end{tabular}
\caption{Concept Parser performance on internal dataset using models with additional Relevance head}
\label{tab2}
\end{minipage}
\end{table}
\textbf{Internal Datasets}: Table~\ref{tab2} reports the performance of our concept parser on the internal dataset. From the table we can observe the following. Our single-task and multi-task models, both trained on personalized data, achieve competitive performance on the synthesized evaluation data. When personalized data are available in training data, single-task model performs already decently (91.29\% and 88.86\%); yet multi-task model could still achieve higher precision. We would like to highlight here that our internal evaluation dataset is also more challenging than the public dataset, including a wide variety of "not teachable" examples including out-of-domain requests, regular requests without concept and ill-grammar and incomplete utterances.
\vspace{-0.75em}
\subsubsection{Definition understanding results}
\vspace{-0.5em}
\textbf{Public Datasets}: Table~\ref{tab3} shows the performance of our definition understanding model on personalized test set from \cite{jia-2017} with three Transformer-based encoders. From the results, we observe that model with RoBERTa~\cite{roberta} encoder achieves the best performance in terms of precision, recall and F1-score. BERT~\cite{bert} and ALBERT~\cite{albert} have similar performance and both are lightly behind RoBERTa~\cite{roberta}. We also did an error analysis and found that nearly half of the span prediction errors are due to a mismatch on preposition (e.g. "on Sept 15" vs "Sept 15"), which is not unexpected since span annotations in \cite{jia-2017} are not consistent on the inclusion of prepositions.
\textbf{Internal Datasets}: The performance of our concept parser on the internal dataset is reported in Table~\ref{tab4}. Again, we observe that model with RoBERTa~\cite{roberta} encoder slightly outperforms model with BERT~\cite{bert} and ALBERT~\cite{albert}.
\begin{table}
\begin{minipage}[t]{0.5\linewidth}
\centering
\begin{tabular}{lccc}
\cmidrule[\heavyrulewidth]{1-4}
\textbf{Model} & \textbf{Precision} & \textbf{Recall}& \textbf{F1-score}\\[0.2em]
\cmidrule{1-4}
RoBERTa & \textbf{91.80} & \textbf{90.89} & \textbf{91.34}\\[0.2em]
\cmidrule{1-4}
BERT & 90.71 & 90.01 & 90.36 \\[0.2em]
\cmidrule{1-4}
ALBERT & 89.93 & 90.53 & 90.23\\[0.2em]
\cmidrule[\heavyrulewidth]{1-4}
\end{tabular}
\caption{Definition Understanding results on Jia {\em et al.} \cite{jia-2017} dataset}
\label{tab3}
\end{minipage}%
\hspace{2em}
\begin{minipage}[t]{0.5\linewidth}
\centering
\begin{tabular}{lccc}
\cmidrule[\heavyrulewidth]{1-4}
\textbf{Model} & \textbf{Precision} & \textbf{Recall}& \textbf{F1-score}\\[0.2em]
\cmidrule{1-4}
RoBERTa & \textbf{95.92} & \textbf{96.81} & \textbf{96.36}\\[0.2em]
\cmidrule{1-4}
BERT & 95.57 & 96.03 & 95.80\\[0.2em]
\cmidrule{1-4}
ALBERT & 95.41 & 95.08 & 95.24\\[0.2em]
\cmidrule[\heavyrulewidth]{1-4}
\end{tabular}
\centering
\caption{Definition Understanding results on internal dataset}
\vspace{-1.2em}
\label{tab4}
\end{minipage}
\end{table}
\subsubsection{Policy Model results}
\vspace{-2mm}
Table~\ref{tab5} shows the performance of our Transformer-based policy model along with heuristics based on NLU grounding results, on an internal dataset with annotated ground-truth actions. We observe that model with RoBERTa~\cite{roberta} encoder achieves the best performance in terms of precision, recall and F1-score. ALBERT~\cite{albert} showed comparable performance to BERT~\cite{bert} but ran the fastest.
\begin{table}[h]
\centering
\begin{tabular}{lccccccccc}
\cmidrule[\heavyrulewidth]{1-5}
\textbf{\space\space\space Model} & \textbf{Precision} & \textbf{Recall}& \textbf{F1}& \textbf{mins/epoch}\\[0.2em]
\cmidrule{1-5}
RoBERTa & \textbf{98.78} & \textbf{97.71} & \textbf{98.24} & 6.12\\[0.2em]
\cmidrule{1-5}
BERT & 98.73 & 96.26 & 97.48 & 5.11\\[0.2em]
\cmidrule{1-5}
ALBERT & 97.98 & 96.12 & 97.04 & 4.56\\[0.2em]
\cmidrule[\heavyrulewidth]{1-5}
\end{tabular}
\caption{Policy Model performance evaluation results on internal dataset}
\vspace{-1.2em}
\label{tab5}
\end{table}
\vspace{-3mm}
\section{Conclusion}
\vspace{-3mm}
We have presented a teachable dialogue system that uses neural models for gap identification, definition understanding and dialogue policy prediction to conduct interactive teaching sessions with the user to learn and re-use definitions of concepts that are unknown to a conversational AI system. This Teachable Dialogue system helps in automatically improving the understanding capabilities of the AI system to hold more natural conversations with the end-users, and progressively improve the agents’ understanding of users' parlance to enable more natural ways of interaction with a conversational AI system. We believe this is an effort towards building truly interactive learning systems, and plan to extend the scope to higher level capabilities such as intent teaching.
\bibliographystyle{plain}
|
\section{Introduction}
Magnetohydrodynamic (MHD) turbulence has been studied extensively for the purpose of understanding macroscopic plasma turbulence as well as understanding universal physics of turbulence which should be shared with neutral fluid turbulence.
Numerous theoretical, numerical, and observational works have been devoted to understand MHD turbulence over the past decades\cite{Iroshnikov1964,Kraichnan1965,SridharGoldreich1994,GoldreichSridhar1997,Stawicki2001,ChoLazarian2002,Bale2005,Biskamp2003,Horbury2008,Alexandrova2009,Kiyani2009,Servidio2009,Davidson2013,Bruno2016}.
More recently, a study of sub-ion-scale effects on turbulence has gathered wide attention\cite{Mahajan2000,Matthaeus2003,Smith2004,Ren2005,Dmitruk2006,Galtier2007,Mininni2007,Galtier2008,HoriMiura2008,Miura2013,Miura2014a}.
The Hall term represents effects of the ion-electron separation in the electric field, the scale of which can be characterized by the ratio of the ion skin-depth to the system size (Hall parameter).
The Hall term can change turbulence dynamics at the small, near-dissipation ion-skin-depth scale, especially through its modified frozen-in condition of the magnetic field to the velocity field.
The Hall effects have been discussed as modifying the scaling-law of turbulence to $k^{-7/3}$ or $k^{-11/3}$ ($k$ is the wave number)\cite{Galtier2008,HoriMiura2008,Galtier2012,Miura2013} and coherent turbulent structures\cite{Dmitruk2006,Miura2014a,Miura2019}.
The Hall term appears also in studies of basic plasma instabilities\cite{Huba1996,Winske1996,Zhu2008,Gotor2014,Gotor2015,Umeda2017} as well as of nuclear fusion\cite{Park1999,Sugiyama2000,Sovinec2004,Miura2014b} which are studied by the use of extended MHD equations\cite{Braginskii1965,HazeltineMeiss1991,Schnack2006}.
Since plasma dynamics in a small scale can be closely coupled with larger scales through advection and other nonlinear terms, introducing the Hall or two-fluid effects at the small ion skin-depth scales can influence dynamics in larger scales.
In fact, our earlier studies in Ref.\cite{Miura2014a,Miura2019} have revealed that the Hall term can induce a structure transition of vortices involving scales larger than the ion skin depth, making the turbulent field be thoroughly different from that of non-Hall MHD (or single-fluid MHD) turbulence qualitatively.
Furthermore, the Hall term can affect large-scale magnetic field dynamics through influencing local magnetic reconnection events in turbulent field\cite{Servidio2009} or through other kinetic effects\cite{Umeda2017}.
The Hall term not only enriches physics such as whistler waves but also increases a numerical difficulty in simulations.
A time step width $\Delta t$ in a simulation is limited strictly as $\Delta t < 1 / k_{\parallel} k$ ($k$ and $k_{\parallel}$ are the wave number and wave number parallel to magnetic field, respectively) in a Hall MHD simulation because of the whistler waves.
In addition, the Hall term can enhance some instabilities at small scales, obliging us using either a large number of grid points or a hyper-diffusivity in numerical simulations\cite{Huba1996,Winske1996,Zhu2008,Gotor2014,Gotor2015}.
Thus Direct Numerical Simulations (DNS) of Hall MHD turbulence are much more expensive than DNS of non-Hall MHD equations.
Furthermore, with respect to a real application such as numerical simulations of instabilities in a magnetically confined nuclear fusion, an excessively large computational cost is not very fruitful because large-scale motions are of main interest.
Nevertheless, since Hall effects can influence a scale larger than the ion skin depth, we must take the Hall effects into large-scale dynamics while suppressing the computational cost.
For studies of large-scale turbulence, so-called Large Eddy Simulations (LES) can be carried out at a much smaller computational cost than DNS\cite{Durbin2001,Garnier2009}.
In LES, we operate a low-pass filter to a set of governing equations, and replace effects of the scales smaller than the filter width by a phenomenological model while retaining the scales larger than the filter width (or the Grid Scale, GS).
Such a phenomenological model is called the Sub-Grid-Scale(SGS) model.
Spatial and temporal resolutions of LES of Hall MHD turbulence can be much coarser than DNS, with an expense of precise dynamics near to or smaller than the filter width.
While LES of MHD turbulence have been studied in some earlier works\cite{Kobayashi2006,Petrosyan2006,Hamba2010,David2012} (see also Refs.\cite{Petrosyan2014,Miesch2015} and references therein), studies of LES of Hall MHD turbulence are not very popular.
Recently, we have shown in Ref.\cite{Miura2016} that the LES approach can be applicable to homogeneous and isotropic Hall MHD turbulence as well as homogeneous turbulence under a constant and uniform magnetic field by applying the SGS model developed in Ref.\cite{Hamba2010}.
We have also applied the LES approach to numerical simulations of nonlinear evolution of ballooning instability in a heliotron-type device for magnetically confined nuclear fusion experiments\cite{Miura2017} by adopting the same type of SGS model as in Ref.\cite{Miura2016} to Braginskii-type extended MHD equations.
More recently, Camporeale et al.\cite{Camporeale2018} analyzed coherent structures in 2D Hall MHD turbulence simulations by the use of the filtering technique, mentioning the LES approach explicitly.
These numerical works have shown the effectiveness of the LES approach of Hall MHD/extended MHD simulations.
In spite of the success in Ref.\cite{Miura2016,Miura2017}, there remains a subject to be studied further on SGS modeling of the Hall term.
We have taken the Smagorinsky constants of the SGS models relatively large, with an expense of a reduced numerical resolution (a steep decay of the energy spectra), in order to reproduce the coherent or local structures of turbulence such as vortex tubes and current sheets clearly.
Since reproduction of both the energy spectra and the local structures are important, we need to find a better compromise between them.
The purpose of this paper is to develop a new SGS model for Hall MHD equations in order to improve the compromise and study natures of the model for homogeneous and isotropic turbulence.
This paper is organized as follows.
In \S.2, Hall MHD equations, the grid-scale (GS) equations of the Hall MHD model, and SGS model for Hall MHD turbulence are introduced.
In \S.3, LES of homogeneous and isotropic Hall MHD turbulence are carried out.
Results of the LES are compared to DNS results.
\S.4 is for summary.
\section{Homogeneous Hall MHD turbulence}
The incompressible Hall MHD equations can be expressed as
\begin{eqnarray}
{\partial u_i \over \partial t}
& = & - {\partial \over \partial x_j}\left[
\left( u_i u_j -B_i B_j\right)
+ \left(p + {1 \over 2} B_k B_k \right) \delta_{ij}\right]
+ \nu {\partial S_{ij} \over \partial x_j},
\label{eq:HallMHD1}
\\
{\partial u_k \over \partial x_k } & = & 0,
\label{eq:HallMHD2}
\\
S_{ij} & =& {\partial u_i \over \partial x_j}
+ {\partial u_j \over \partial x_i},
\label{eq:HallMHD3}
\\
{\partial B_i \over \partial t}
& = & - \epsilon_{ijk} {\partial E_k \over \partial x_j},
\label{eq:HallMHD4}
\\
{\partial B_k \over \partial x_k } & = & 0,
\label{eq:HallMHD5}
\\
E_i & =& -\epsilon_{ijk}
\left( u_j - \epsilon_H J_j \right) B_k + \eta J_i.
\label{eq:HallMHD6}
\end{eqnarray}
The three symbols $B_i$, $J_i = \epsilon_{ijk} \partial_j B_k$, and $u_i$ represent the $i$-th components of the magnetic field, the current density, and velocity field vectors, respectively.
Tensor symbols $\delta_{ij}$ and $\epsilon_{ijk}$ are the Kronecker's delta and the Levi-Civita's anti-symmetric tensor, respectively.
The sum of $1$, $2$, and $3$ is taken for repeated suffixes of the vector and tensor variables.
The equations (\ref{eq:HallMHD1})-(\ref{eq:HallMHD6}) are already normalized by a reference length scale $L_0$, magnetic field strength $B_0$, mass density $\rho_0 = m_0 n_0$ ($m_0$ and $n_0$ are the ion mass and the number density, respectively), and Alfv{\'e}n velocity $V_A=\sqrt{B^2/\mu_0 \rho_0}$ where $\mu_0$ is the vacuum permeability.
The symbols $\epsilon_H$, $\eta$, and $\nu$ are the Hall parameter (the ratio of the ion skin depth $d_i = \sqrt{m_0/n_0 \mu_0 e^2}$ to the length scale $L_0$, where $e$ is the electric charge), magnetic diffusivity, and shear viscosity, respectively.
We can understand $1/\nu$ and $1/\eta$ as the reference Reynolds number and the Lundquist number, respectively.
These parameters control near-dissipation scales of turbulence in simulations.
The equations (\ref{eq:HallMHD1})-(\ref{eq:HallMHD6}) are operated by a low-pass filter
\begin{eqnarray}
{\overline {v}} & := & \int G(x-s) v(s) ds
\end{eqnarray}
to obtain the GS equations of the Hall MHD model, where $v(x,t)$ is a dependent variable.
The filter $G$ is assumed to be a sharp and isotropic in the Fourier space, with the cut-off wave number $k_{cut}$.
The GS Hall MHD equations are described as
\begin{eqnarray}
{\partial {\overline u}_i \over \partial t}
& = & - {\partial \over \partial x_j}\left[
\left( {\overline u}_i {\overline u}_j
-{\overline B}_i {\overline B}_j\right)
+ \left ({\overline p}
+ {1 \over 2}{\overline B}_k {\overline B}_k \right) \delta_{ij}
\right]
+ \nu {\partial {\overline S}_{ij} \over \partial x_j}
- {\partial {\overline \tau}_{ij} \over \partial x_j},
\label{eq:GSHallMHD1} \\
{\overline \tau}_{ij}
& = & \left[
\left( {\overline {u_i u_j} } -{\overline {B_i B_j}}\right)
+ {1 \over 2} {\overline {B_k B_k}} \delta_{ij} \right]
- \left[
\left( {\overline u}_i {\overline u}_j -{\overline B}_i {\overline B}_j\right)
+ {1 \over 2}{\overline B}_k {\overline B}_k \delta_{ij} \right],
\label{eq:GSHallMHD2} \\
{\overline S}_{ij} & =& {\partial {\overline u}_i \over \partial x_j}
+ {\partial {\overline u}_j \over \partial x_i},
\label{eq:GSHallMHD4}
\\
{\partial {\overline u}_k \over \partial x_k } & = & 0,
\label{eq:GSHallMHD3} \\
%
{\partial {\overline B}_i \over \partial t}
& = & - \epsilon_{ijk} {\partial \overline{E}_k \over \partial x_j},
\label{eq:GSHallMHD5} \\
{\overline {E}}_i
& =& -\epsilon_{ijk} \left( \overline {u}_j
- \epsilon_H {\overline J}_j \right) \overline {B}_k
+ \eta \overline{J}_i
- {\overline E}^M_i - {\overline E}^H_i,
\label{eq:GSHallMHD6} \\
\overline {E}^M_i
& = & -\epsilon_{ijk} \left(
\overline{ u_j B_k } - \overline {u}_j \overline
{B}_k\right),
\label{eq:GSHallMHD7} \\
\overline {E}^H_i
& = & - \epsilon_H \epsilon_{ijk} \left(
-\overline{ J_j B_k } + \overline {J}_j \overline
{B}_k\right),
\label{eq:GSHallMHD8} \\
{\partial \overline {B}_k \over \partial x_k} & = & 0.
\label{eq:GSHallMHD9}
\end{eqnarray}
The GS equations (\ref{eq:GSHallMHD1})-(\ref{eq:GSHallMHD8}) are closed when the SGS terms ${\overline \tau}_{ij}$, $\overline{E}^M_i$, and $\overline{E}^H_i$ are replaced by SGS models expressed only by the GS variables $\overline{u}_i$ and $\overline{B}_i$.
In Ref.\cite{Miura2016}, the SGS terms $\overline{\tau}_{ij}$, $\overline {E}^M_i$ and $\overline {E}^H_i$ have been modelled as
\begin{eqnarray}
\overline{\tau}_{ij} & = & - \nu_{{}_{SGS}} \overline{S}_{ij}, \label{eq:SGS1}
\\
\overline{E}^M_i +\overline{E}^H_i & = & - \eta_{{}_{SGS}} \overline{J}_i, \label{eq:SGS2}
\\
\nu_{{}_{SGS}}
& =& C_{\nu}\Delta^2 \left({1 \over 2} C_{\nu} \overline{S}_{ij}^2 + C_{\eta} {\overline J}_i^2 \right)^{1/2}, \label{eq:SGS3}
\\
\eta_{{}_{SGS}} & =& C_{\eta}\Delta^2 \left({1 \over 2} C_{\nu} \overline{S}_{ij}^2 + C_{\eta} {\overline J}_i^2 \right)^{1/2}, \label{eq:SGS4}
\end{eqnarray}
based on the studies in Refs.\cite{Hamba2010,Miura2013,Miura2014a}.
The symbol $\Delta$ is the filter width.
In this model, we have two Smagorinsky constants $C_{\nu}$ and $C_{\eta}$.
They have been given originally in Ref.\cite{Hamba2010} as $C_{\nu}=0.046$ and $C_{\eta}=C_{\nu}\times (5/7)$.
However, in Ref.\cite{Miura2016}, we have chosen coefficients $C_{\nu}=C_{\eta}=0.345$ for $\Delta = \Delta_0 = 1/k_{cut}$ to reproduce local structures, because the turbulent field is too noisy when the Smagorinsky constants are small.
Consequently, the numerical resolution in the LES has been reduced effectively.
In order to find a better compromise between the reproduction of the energy spectra and spatial structures, we model $\overline{E}^M_i$ and $\overline{E}^H_i$ as
\begin{eqnarray}
\overline{E}^M_i + \overline{E}^H_i
& = & - \eta_{{}_{SGS}} \overline{J}_i \nonumber \\
& & + \epsilon_H {\partial \over \partial x_j}
\left( \nu_{{}_{SGS}} \sigma_{{}_{H,1}} \overline{S}_{ij} \right)
- \epsilon_H^2 {\partial \over \partial x_j}
\left[ \nu_{{}_{SGS}} \sigma_{{}_{H,2}} \overline{S}_{ij}
\left( {\partial {\overline J}_i \over \partial x_j}
+ {\partial {\overline J}_j \over \partial x_i}\right)
\right],
\label{eq:SGS8}
\end{eqnarray}
keeping $\overline{\tau}_{ij}$ the same as eq.(\ref{eq:SGS1}).
We do not separate $\overline{E}^M_i$ and $\overline{E}^H_i$ because a forward-transfer part of the Hall term in the Fourier space can be combined as the diffusive part of $\overline{E}^H_i$ together with the $\overline{E}^M_i$.
See Appendix for more detailed description on the model.
The contribution of this SGS model to the magnetic energy budget ${\partial_t}({1 \over 2} B_i B_i)$ is
\begin{eqnarray}
-\epsilon_{MSH}
& = & -\overline{E}^H_i \overline{J}_i \nonumber \\
& = & \epsilon_H
\left( \nu_{{}_{SGS}} \sigma_{{}_{H,1}} \overline{S}_{ij} \right)
\left( {\partial {\overline J}_i \over \partial x_j} \right)
- \epsilon_H^2
\nu_{{}_{SGS}} \sigma_{{}_{H,2}}
\left( {\partial {\overline J}_i \over \partial x_j}
+ {\partial {\overline J}_j \over \partial x_i}\right)
\left( {\partial {\overline J}_i \over \partial x_j} \right) \nonumber \\
&& - {\partial \over \partial x_j}\left[
\epsilon_H
\left( \nu_{{}_{SGS}} \sigma_{{}_{H,2}} \overline{S}_{ij} \right)
{\overline J}_i
- \epsilon_H^2
\nu_{{}_{SGS}} \sigma_{{}_{H,2}}
\left( {\partial {\overline J}_i \over \partial x_j}
+ {\partial {\overline J}_j \over \partial x_i}\right)
\overline{J}_i \right].
\end{eqnarray}
While the second term works only for the energy dissipation and the third term vanishes by the volume integral over the computational domain, a behavior of the first term depends on detailed dynamics in the nonlinear evolution.
\section{LES of decaying homogeneous and isotropic Hall MHD turbulence}
\subsection{Outline of LES}
LES of freely decaying homogeneous and isotropic Hall MHD turbulence for ${\epsilon}_H = 0.05$ and ${\epsilon}_H=0.025$ are carried out in this section.
The LES results are compared to the GS components of DNS data reported in Ref.\cite{Miura2014a}, with the same viscosity, resistivity, and Hall parameter, and the same initial condition as the LES.
The initial velocity and magnetic fields are given by the energy spectrum $E(k,t) \propto k^2 \exp(-k^2/k_0^2)$ ($k_0=2$) and random phases of the Fourier components.
Equations~(\ref{eq:GSHallMHD1})-(\ref{eq:GSHallMHD8}) are coupled with with the SGS model in eqs.~(\ref{eq:SGS1}), (\ref{eq:SGS3}), (\ref{eq:SGS4}), and (\ref{eq:SGS8}), and solved numerically by the pseudo-spectral method and the Runge-Kutta-Gill scheme.
Aliasing errors are removed by the $2/3$-truncation in spectral space.
Note that the implementation of the $2/3$-truncation is different between this paper and Ref.\cite{Miura2016}.
The $2/3$-truncation has been implemented in Ref.\cite{Miura2016} as anisotropic low-pass filter (rectangular in the spectral space), in order to study not only homogeneous and isotropic turbulence but also homogeneous turbulence under a constant and uniform magnetic field.
In this paper, to the contrary, an isotropic filter (spherical in the spectral space) is adopted for de-aliasing because we focus on homogeneous and isotropic turbulence.
An advantage of a Hall MHD simulation to a (non-Hall) MHD simulation with respect to LES is that the forward energy transfer of the magnetic energy can be dominant at a large wave number region in the spectral space, because the Hall term makes the magnetic field (induction) equations be quadratic to the magnetic field\cite{Miura2013,Miura2014a}.
In order to make use of this advantage, we study firstly LES of $k_{max} > k_H = 1/\epsilon_H$ so that the GS Hall MHD equations retain the scales comparable to or smaller than the ion skin depth $\epsilon_H$, where $k_{max}$ is the maximum wave number in a simulation.
We will also examine LES with $k_{max} \simeq k_H$ below.
We refer to the DNS data with the number of grid points $N^3=1024^3$ and the parameters $\eta = \nu = 5 \times 10^{-4}$, $\epsilon_H = 0.05$, and $0.025$ reported in \cite{Miura2014a} as reference data.
Since we adopt the isotropic low-pass filter for de-aliasing, the maximum wave number in an LES is $k_{max} = 42$ for $N^3=128^3$ and $k_{max} = 21$ for $N^3=128^3$, while the wave number associated with the ion skin depth is $k_H = 20$ for $\epsilon_H=0.05$ and $k_H = 40$ for $\epsilon_H=0.025$.
All the parameters needed for the SGS model studied in this paper are shown in Table \ref{tab:params}.
We place a higher priority on verifying applicability of our new SGS model to Hall MHD turbulence, and on clarifying the nature of the model, rather than on calibrating model constants included in the SGS model to reproduce some typical quantities precisely.
While we have changed $C_{\nu}$ and $C_{\eta}$ by keeping $\Delta=\Delta_0$, we can understand that the $C_{\nu}$ and $C_{\eta}$ values are given for a larger $\Delta$ (this means that the GS is coarser than that of $\Delta=\Delta_0$).
Thus we show $C_{\nu}$ and $C_{\eta}$ for $\Delta_0$, $1.5 \Delta_0$, $2 \Delta_0$, and $3 \Delta_0$.
Different combinations of $\left(C_{\nu}, C_{\eta}, \Delta\right)$ in each of parameter sets (000, 001, ...) in Table \ref{tab:params} are equivalent to each other and give the same results for the SGS model in eqs.(\ref{eq:SGS1})-(\ref{eq:SGS4}).
\subsection{LES with the previous SGS model}
Firstly, we carry out LES with $N^3=128^3$ and $\epsilon_H=0.05$, and thus $k_{max} < k_H$, for parameter sets 000-004, in order to clarify the subject to be studied in this paper.
We emphasize here again that the parameter sets 000-004 are the same as those studied in our earlier work\cite{Miura2016} but the de-aliasing filter is different in the LES.
Both $\sigma_{{}_{H,1}}$ and $\sigma_{{}_{H,2}}$ are null in the parameter sets 000-004, and thus the Hall SGS term which is developed newly in this paper is switched off in the runs with the parameter sets 000-004 (hereafter referred to as runs 000-004).
In Fig.\ref{fig:LESstats01}, (a) the mean kinetic energy $E_K = \left< u_i u_i\right>/2 $, (b) the mean magnetic energy $E_M = \left< B_i B_i\right>/2$, (c) the enstrophy $Q = \left< \omega_i \omega_i\right>/2$ where $\omega_i = \epsilon_{ijk} \partial_j u_k$ is the $i$-th component of the vorticity, and (d) the mean current $J = \left < J_i J_i \right> /2$ are shown for the LES with the parameter set 000-004, together with the GS component of $E_K$ obtained from the reference DNS data filtered by $k_{cut}=42$ low-pass filter.
Hereafter, we refer to DNS data filtered as DNS GS data.
The filter width is either $k_{cut}=42$ or $21$ low-pass filter corresponding to $k_{max}$ of LES that are to be compared with the DNS GS data.
The symbol $\left< \cdot \right>$ indicates the volume average.
In Fig.\ref{fig:LESstats01}(a), the parameter set 000 gives the best fitting of $E_K$ to that of the DNS GS data (thick line) among the parameter sets 000-004.
Run 002 overestimates $E_K$ while the other runs underestimate the time evolution of $E_K$.
In Fig.\ref{fig:LESstats01}(b), the parameter set 002 gives the best fit to the time evolution of $E_M$ in the GS, while 000 and the other parameter sets underestimate $E_M$.
The enstrophy $Q$ and mean current $J$ in the GS are in between those given by the parameter sets 000 and 003 in Fig.\ref{fig:LESstats01}(c) and (d), respectively.
The quantities $Q$ and $J$ in the runs 000 and 002 can be twice as large as the DNS GS data.
In Fig.\ref{fig:LESspec01}, (a) the kinetic energy spectra $E_K(k,t) = \sum_{\left[k\right]} {\widetilde u}_i(k,t) \widetilde {u}_i^{*}(k,t)$ and (b) the magnetic energy spectra $E_M(k,t) = \sum_{\left[k\right]} {\widetilde B}_i(k,t) {\widetilde B}_i^{*}(k,t)$ at $t=1$ are shown, where the symbols $\sum_{\left[k\right]}$, ${\widetilde {\cdot}}$, and ${}^{*}$ represent the shell average in the spectral space, the Fourier coefficient, and the complex conjugate of the Fourier coefficients, respectively.
In Fig.\ref{fig:LESspec01}(a), $E_K(k,t)$ of run 002 piles up slightly at $k \geq 30$, suggesting the dissipation by the SGS viscosity $\nu_{{}_{SGS}}$ being too small.
The pile-up can be also seen $E_M(k,t)$ of run 002 in Fig.\ref{fig:LESspec01}(b).
These pile-ups are responsible for the large overestimate of $Q$ and $J$ in Fig.\ref{fig:LESstats01}(c) and (d).
The parameter set 000 gives a reasonable reproduction of $E_K(k,t)$ and $E_M(k,t)$, and other parameter sets, of which $C_{\nu}$ and $C_{\eta}$ are larger than those of 002, give a steeper decay of the spectra at $k \geq 30$.
In Fig.\ref{fig:LESiso00}, isosurfaces of the enstrophy density $q$ and the current density $I$ of (a) runs 000 and (b) 003 are shown at $t=1$.
The variable $q$ is drawn in blue-green colors and $I$ in gray throughout this article.
In each panel, thresholds of $q$ and $I$ are changed depending on its mean ($m$) and deviation ($\sigma$) values so that we can see spatial structures of the two quantities clearly.
Although we expect some specific structures such as sheets and tubes in the isosurfaces of $q$ and $I$, the isosurfaces in (a) are cloudy and structures therein are not clear.
Detailed structures are difficult to distinguish from the cloudy isosurfaces even when we lower or raise the thresholds except for a few tubular vortices.
The isosurfaces of run 002, which is not shown here, does not show a clearer spatial structures than run 000.
The isosurfaces in (b) show clear vortex tubes and current sheets.
However, run 003 underestimates $E_M$ in Fig.\ref{fig:LESstats01}(b) too much, and the decay of the energy spectra at $k > 30$ is also steep in Fig.\ref{fig:LESspec01}.
Furthermore, the Smagorinsky constants of the parameter set 003 are very large, and equivalent to $C_{\nu} = 0.046$ (which has been given theoretically in Ref.\cite{Hamba2010}) and $\Delta \simeq 3.34 \Delta_0$.
This means that the essential resolution in the LES is three times ore more coarser than that in the parameter set 002.
This is why we need to make a compromise on the spatial structures and the other quantities.
\subsection{LES for $k_{max} > k_H$ with the new SGS model}
Now we move on to LES with the new SGS model (\ref{eq:SGS8}), with the parameter sets 100-109.
In the parameter sets 100-109, we set $C_{\eta} = C_{\nu} \times (5/7)$ as in Ref.\cite{Hamba2010}.
The parameters $\sigma_{{}_{H,1}}$ and/or $\sigma_{{}_{H,2}}$ are non-zero in the parameter sets 100-109.
These two parameters are new in the SGS model (\ref{eq:SGS8}).
Figure \ref{fig:LESstats02} is (a) $E_K$, (b) $E_M$, (c) $Q$, and (d) $J$ of runs 100-109.
In Fig.\ref{fig:LESstats02}(a), all the LES give the time evolution of $E_K$ very close to that DNS GS data.
Though the parameter sets 102 and 109 give relatively large deviation of $E_K$ from that of DNS GS data, the deviation is smaller than those in Fig.\ref{fig:LESstats01}.
In Fig.\ref{fig:LESstats02}(b), $E_M$s of all the LES are smaller than that of DNS GS data.
The Smagorinsky constant $C_{\eta}$ in the parameter sets 100 and 102 are $5/7$-times of $C_{\eta}$ in the parameter sets 000 and 002, respectively.
In spite of the smaller $C_{\eta}$, $E_M$ in runs 100 and 102 underestimate $E_M$ considerably in contrast to runs 000 and 002.
Although the new model causes underestimation of $E_M$, $Q$ and $J$ are reproduced more precisely by the new model.
In Fig.\ref{fig:LESstats02}(c) and (d), the LES either overestimate or underestimate $Q$ and $J$, respectively, depending on the parameters $C_{\eta}$, $C_{\nu}$, $\sigma_{{}_{H,1}}$ and $\sigma_{{}_{H,2}}$.
We can compare Figs.\ref{fig:LESstats01} and \ref{fig:LESstats02} to figure out a difference due to these new parameters.
For example, parameter sets 003, 103, and 109 have the same $C_{\nu} = 0.23$ and different $C_{\eta}$ and $\sigma_{{}_{H,1}}$.
We also compare the results of runs 100 to 000, and those of runs 101 to 001, respectively.
The deviation of $Q$, and $J$ of runs 100, 101, 103, and 109 from the DNS GS data are smaller than those of runs 000 and 003, primarily because of the smaller $C_{\eta}$.
The parameter sets 101 and 105 appear to give the best fitting of $Q$ and $J$ among the parameters 100-109, although the decay rate of $Q$ at $t > 0.5$ appears too slow in comparison to that of the GS component of DNS.
These comparisons suggest that the new SGS model can give a better compromise between the reproduction of the GS DNS data of $E_K$, $Q$ and $J$ in exchange for the underestimation of $E_M$.
In Fig.\ref{fig:LESspec02}, the energy spectra (a)$E_K(k,t)$ and (b)$E_M(k,t)$ at $t=1$ of runs 100-104 are shown, while (c) and (d) are for the two spectra of runs 105-109, respectively.
The two spectra $E_K(k,t)$ and $E_M(k,t)$ of run 100 collapse with those of runs of 106 and 108, respectively, when they are plotted together.
The two spectra of runs 101 and 107 also collapse with those of runs 106 and 108 except a small difference at $k > 30$.
The energy spectra of the three runs 100, 106, and 108 appear going along $E_K(k,t)$ of DNS better than those of 002, without showing pile-up as in run 000 in Fig.\ref{fig:LESspec01}.
The energy spectra $E_K(k,t)$ and $E_M(k,t)$ of runs 102, 103, and 109 also collapse each other, showing a steeper decay at $k > 30$ of $E_M(k,t)$ than that of runs 100, 101, 106, 107, and 108.
The steepest decay of $E_M(k,t)$ is obtained by the parameter sets 104 and 105, in which $\sigma_{{}_{H,2}}=1$ provides a strong hyper-diffusivity.
Among the parameter sets 100-109, the parameter set 108 and 109 have the largest $\sigma_{{}_{H,1}}=8$.
We pay attention to run 108 because this parameter set gives not only a good energy spectrum profiles in Fig.\ref{fig:LESspec02}, but also the best fitting to $E_K$ and $E_M$ in Fig.\ref{fig:LESstats02} (a) and (b).
Although $Q$ and $J$ are overestimated in Fig.\ref{fig:LESstats02} (c) and (d), respectively, the difference from the DNS GS data is much smaller in comparison to those in the runs 000 and 002.
Additionally, the two spectra $E_K(k,t)$ and $E_M(k,t)$ of run 108 can be compared to those of run 104 because the parameters of the two runs are the same but $\sigma_{{}_{H,1}}$.
Obviously, $E_K(k,t)$ and $E_M(k,t)$ of run 108 in Fig.\ref{fig:LESspec02} fit well with the spectra of DNS better than those of run 104.
The comparison shows that the new term with the coefficient $\sigma_{{}_{H,1}}$ can improve the spectral nature of the LES.
Based on the observation in the above, we hereafter focus on runs 105-109.
In order to investigate the LES results further, we see spatial structures of the enstrophy density $q=\omega_i \omega_i/2$ and the current density $I=j_i j_i/2$.
Two aspects of the two quantities are of interest here.
The first aspect is that high-intensity regions of $q$ and $I$ should be well separated from each other in Hall MHD turbulence while they often stay together in single-fluid MHD turbulence\cite{Miura2014a}.
The Hall term, representing a separation effect of the ion and electron motions, changes the frozen-in condition between the velocity field and the magnetic field, and, consequently, the spatial distributions of high-intensity regions of $q$ and $I$, too.
Thus we can see an effect of the Hall term visually through the separation of $q$ and $I$.
The second aspect is that the vortex structures are transformed from sheets to tubes as the consequence of the introduction of the Hall term, and the structure transition can be observed not only at the scale of the ion skin depth but also at a larger scale.
Since these two aspects are essential change from single-fluid to Hall-MHD turbulence, we require LES of Hall MHD turbulence to reproduce these two aspects.
Fig.\ref{fig:LESiso01} is isosurfaces of $q$ and $I$ of runs (a)105, (b)106, (c)107, and (d)108 at $t=1$.
In Fig.\ref{fig:LESiso01}(a), the isosurfaces of $q$ and $I$ are unclear, their shapes being neither sheets nor tubes, and distinguishing isosurfaces of $q$ from those of $I$ is not very simple.
In contrast, vortex tubes and sheets are clearly observed, and isosurfaces of $q$ and $I$ can be distinguished easily in Fig.\ref{fig:LESiso01}(b)-(d).
The visualizations in Fig.\ref{fig:LESiso01}(b)-(d) suffice for the two aspects described in the previous paragraph.
Among the three panels, Fig.\ref{fig:LESiso01}(d) shows the clearest vortex tubes/sheets and current sheets, and looks the most favorable.
Here we remark on how we give the thresholds of $q$ and $I$.
In Fig.\ref{fig:LESiso01}(a), the thresholds of the two quantities are given by $m + 4\sigma$.
In Fig.\ref{fig:LESiso01}(b) and (d), the thresholds are given by $m + 3\sigma$, whereas they are $m + 2\sigma$ in (c).
The threshold $m + 2\sigma$ gives about $5\%$ of the volume if the scalar field (either $q$ or $I$) has the Gaussian distribution.
The fact that we need to change the thresholds in the visualizations (a)-(d) suggests that distributions of $q$ and $I$ are qualitatively changed.
In order to clarify the qualitative change, we observe the Probability Density Functions (PDFs) of the vorticity and current density of runs 105-108.
In Fig.\ref{fig:LESpdf01}, PDF of the first component of (a) the vorticity and (b) the current density at $t=1$ are shown.
The two thick lines represent PDFs by DNS, operated by the low-pass filter of $k_{cut}=42$ and $21$.
In Fig.\ref{fig:LESpdf01}(a), the PDFs of all of the four runs are scattered near the plot of DNS GS data with $k_{cut} = 42$, and none of them are plotted near the DNS GS data with $k_{cut}=21$.
These plots show that the four runs reproduce vorticity distribution fairly well for the resolution of $N^3=128^3$.
In Fig.\ref{fig:LESpdf01}(b), in contrast to (a), the PDFs of the current density of the four runs do not collapse.
While the plots of the PDFs of the runs 106, 107, and 108 collapse near the PDF of DNS GS data of $k_{cut}=42$, the PDF of the run 105 collapses with DNS GS data of $k_{cut}=21$.
The latter indicates that the run 105 has the resolution comparable to $N^3=64^3$ grid points.
This observation can be interesting because a nature of the two new terms in the Hall SGS model (\ref{eq:SGS8}), multiplied by $\sigma_{{}_{H,1}}$ and $\sigma_{{}_{H,2}}$ appears very well.
The two Smagorinsky constants $C_{\nu}$ and $C_{\eta}$ in the run 105 are the same as those of the run 108.
A difference in the parameters between the runs 105 and 107 is in $\sigma_{{}_{H,1}}$ and $\sigma_{{}_{H,2}}$.
Since $\sigma_{{}_{H,2}}=1$ in the run 105, the fourth-order hyper-diffusivity can reduce the effective numerical resolution of the magnetic field from $N^3=128^3$ to $N^3=64^3$.
However, we set $\sigma_{{}_{H,2}}=1$ in the run 108, too.
Nevertheless, the PDF of $j_1$ of run 108 is scattered near the PDF of DNS GS data of $k_{cut}=42$, without showing the reduction of the effective numerical resolution.
This indicates that the second term in the right-hand-side (RHS) of eq.(\ref{eq:SGS8}), multiplied with $\sigma_{{}_{H,1}}=8$ in the run 108, keeps the effective numerical resolution against the fourth-order hyper-diffusivity in eq.(\ref{eq:SGS8}).
It should be remembered here that $E_M(k,t)$ of run 108 fits better to that of DNS in comparison to run 105 as we have seen in Fig.\ref{fig:LESspec02}, indicating that the second term in the RHS of eq.(\ref{eq:SGS8}) can improve the spectra property as well.
From these numerical results, we consider that our new SGS model in eq.({eq:SGS8}) improves spectral and statistical natures of LES of Hall MHD turbulence well, especially when the parameter set 108 is adopted.
The Smagorinsky coefficient $C_{\nu}=0.092$ for the filter width $\Delta=\Delta_0$ is equivalent to $C_{\nu} = 0.046$ and $\Delta=1.68 \Delta_0$.
This means that the numerical resolution in the LES with the parameter set 108 is two times finer than that with the parameter set 003.
\subsection{LES for $k_{max} \simeq k_H$ }
We have studied LES of $k_{max} \simeq 2 k_H > k_H$ in the previous subsection.
However, we wish to set $k_{max} \leq k_H$ in an LES because resolving scales $k_{max} \simeq 2 k_H$ is still expensive for applications of the Hall MHD equations.
Though making a numerical resolution too coarse in an LES can change nonlinear dynamics in GS qualitatively from the original governing equation, it is still worth examining the marginal region $k_{max} \simeq k_H$.
In order to study influences of the ion skin depth scale on LES of Hall MHD turbulence, LES with the parameter sets 105-108 are carried out, firstly with the number of grid points $N^3=128^3$ and the Hall parameter $\epsilon_H=0.025$ ($k_{max}=42$ and $k_{H}=40$), and secondly with the number of grid points $N^3=64^3$ and the Hall parameter $\epsilon_H=0.05$ ($k_{max}=21$ and $k_{H}=20$).
In both cases, the relation between the maximum resolution of LES and the ion skin depth is $k_{max} > k_H$.
In Fig.\ref{fig:LESstats03}, time evolution of (a) $E_K$, (b) $E_M$, (c) $Q$, and (d) $J$ for LES with the parameter sets 105-108, $N^3=128^3$, and $\epsilon_H{=0.05}$ are shown.
DNS GS data is also for $\epsilon_H = 0.025$.
The relation between the LES and DNS results are quite similar to that in Fig.\ref{fig:LESstats02}: the LES predict $E_K$ fairly well in (a), underestimate $E_M$ in (b), moderately reproduce $Q$ in (c) and $J$ in (d).
In Fig.\ref{fig:LESspec03}, (a) $E_K(k,t)$ and (b) $E_M(k,t)$ at $t=1$ of runs 105-108 are shown.
Again, the energy spectra in the figure are quite similar to those in Fig.\ref{fig:LESspec02}.
These observations indicate that we can apply our SGS model for LES of $k_{max} \simeq k_H$.
We recall that the LES results of single-fluid MHD turbulence with the parameter sets 000-004 show a large deviation from the DNS result at $k=1$, the largest scale, in Ref.\cite{Miura2016}.
Since single-fluid MHD can be understood as the limit of $\epsilon_H \rightarrow 0$ ($k_H \rightarrow \infty$ equivalently), the LES results are considered to be worse for $k_H > k_{cut}$.
In Fig.\ref{fig:LESstats04}, time evolution of (a) $E_K$, (b) $E_M$, (c) $Q$, and (d) $J$ for LES with the parameter sets 105-108 with $N^3=64^3$ and $\epsilon_H=0.05$ ($k_{max}=21$ and $k_H=20$) are shown, together with the DNS GS data of $k_{cut}=21$.
While the LES give time evolution of $E_K$ in (a) very close to the DNS GS data with $k_{cut}=21$, the LES show a clear tendency of overestimation of $E_K$.
The overestimation is clearer in (c) on $Q$.
On the other hand, time evolution of $E_M$ in (b), especially of runs 106 and 108 which collapse each other completely, is very close to $E_M$ of the DNS GS data in the initial stage of the time evolution at $t \leq 0.5$ (the time of the peak of $Q$ and $J$).
At this point, the LES of $N^3=64^3$ give a better agreement with DNS GS data than the LES of $N^3=128^3$ and $\epsilon_H=0.05$.
At $t > 0.5$, however, $E_M$ of LES are not along the time evolution of that of the DNS GS data.
This shows a contrast with the time evolution of $E_M$ in Fig.\ref{fig:LESstats02}(b).
This separation of the $E_M$ in the LES from the DNS GS data may come from the lack of a sufficient number of Fourier coefficients rather than from the relative relation of $k_{cut}$ and $k_{H}$.
Fig.\ref{fig:LESspec04} is for (a)$E_K(k,t)$ and (b)$E_M(k,t)$.
A clear change in Fig.\ref{fig:LESspec04} from Fig.\ref{fig:LESspec02} is that the LES with $N^3=64^3$ apparently underestimate $E_M(k,t)$ except $k=1$.
Though the mean magnetic energy is well reproduced by the LES, the energy spectrum $E_M(k,t)$ in Fig.\ref{fig:LESspec04} may not be acceptable for some applications because of clear underestimation of $E_M(k,t)$ at $k > 1$.
Considering observations in Figs.\ref{fig:LESstats04} and \ref{fig:LESspec04} together, the Fourier coefficients of $21 \leq k \leq 42$ which are eliminated from a $N^3=64^3$ simulation are coupled closely with low-$k$ coefficients, and play significant roles in the reproduction of the spontaneous magnetic energy spectrum and long-term behavior of the mean magnetic energy.
This point should be studied further in relation with the cascade/inverse-cascade nature of turbulence in a future paper.
In summary, the results of LES in this section show that we can apply our new SGS model successfully to homogeneous and isotropic Hall MHD turbulence within the range of $k_H \simeq k_{cut}$, but the reliability of the LES can depend on the number of grid points.
\section{Concluding Remarks}
We have developed a new sub-grid-scale model for a large eddy simulation of freely-decaying homogeneous and isotropic Hall Magnetohydrodynamic turbulence.
Natures of the new model have been studied on time evolution of the energies, enstrophy, total current, energy spectra of the kinetic and magnetic field, and PDFs of the vorticity and the current density.
Some combinations of the Smagorinsky constants have been assessed from the points of view of these quantities, with the help of the visualization of local structures as well as the probability density functions of the vorticity and current components.
LES of Hall MHD turbulence have shown that our new model improves spectral properties of LES, and provides reproduction of the DNS GS data.
While the previous model is also acceptable if we focus on the kinetic and magnetic energies, the new model achieves not only good spectral properties but also reproduces the kinetic energy, enstrophy, current, and local structure visualization in exchange for an underestimate of the magnetic energy.
Analysis for combinations of the numerical resolution and the Hall parameter shows that we can carry out LES not only for $k_{max} > k_H$ but also for $k_{max} \simeq k_H$.
By the use of the new SGS, we can double an essential resolution of LES of Hall MHD turbulence in comparison to the previous SGS model.
Our new SGS model enables numerical simulations of homogeneous and isotropic magnetohydrodynamic turbulence with Hall effects with a small computational cost, retaining the ion-electron separation effects by the Hall term in the grid scale.
This research was partially supported by JSPS KAKENHI Grant Number 17K05734 and MEXT KAKENHI Grant Number 15H02218, Japan.
The numerical simulations were performed on the FUJITSU FX100 supercomputer {\it Plasma Simulator} of NIFS with the support and under the auspices of the NIFS Collaboration Research program (NIFS15KNSS053, NIFS15KNTS038), and partially on the NEC SX-ACE supercomputer of Tohoku University as well as on the FUJITSU Oakforest-PACS supercomputer of the University of Tokyo, being partially supported by ``Joint Usage/Research Center for Interdisciplinary Large-scale Information Infrastructures'' in Japan.
\clearpage
\section*{APPENDIX: MODELING THE HALL TERM}
In the appendix we describe the derivation of the model expression for the Hall term appearing in (\ref{eq:SGS8}).
The SGS Hall electromotive force $\overline E_i^H$ defined in (\ref{eq:GSHallMHD8}) can be rewritten as
\begin{equation}
\overline E_i^H = {\epsilon _H}\frac{\partial }{{\partial {x_j}}}(\overline {{B_i}{B_j}} - {\overline B_i}{\overline B_j}) - {\epsilon _H}\frac{1}{2}\frac{\partial }{{\partial {x_i}}}(\overline {{B_j}{B_j}} - {\overline B_j}{\overline B_j}) .
\end{equation}
The second part on the right-hand side does not contribute to $ - \nabla \times {{\bf{\overline E}}^H}$ in the induction equation given by (12).
We need to model the magnetic field correlation $\overline {{B_i}{B_j}} - {\overline B_i}{\overline B_j}$.
Here, we use the Markovianized two-scale method for inhomogeneous turbulence modeling proposed by Yoshizawa \cite{yosh98}.
For simplicity, we assume that $\overline {\overline f} = \overline f$ and $\overline {f'} = 0$ where $f' = f - \overline f$ and $f$ is a quantity such as the velocity and the magnetic field.
The magnetic field correlation is then written as $\overline {{B'_i}{B'_j}} $.
Using (\ref{eq:HallMHD4})-(\ref{eq:HallMHD6}) we can derive the equation for the fluctuating magnetic field ${B'_i}$ as follows:
\[
\frac{{\partial {B'_i}}}{{\partial t}} + \frac{\partial }{{\partial {x_j}}}{{\overline u}_j}{B'_i} + \frac{\partial }{{\partial {x_j}}}({u'_j}{B'_i} - \overline {{u'_j}{B'_i}} ) - \eta \frac{{{\partial ^2}{B'_i}}}{{\partial {x_j}\partial {x_j}}}
\]
\begin{equation} \label{eq:a2}
= - {u'_j}\frac{{\partial {{\overline B}_i}}}{{\partial {x_j}}} + \frac{{\partial {u'_i}}}{{\partial {x_j}}}{{\overline B}_j} + {B'_j}\frac{{\partial {{\overline u}_i}}}{{\partial {x_j}}} + {\epsilon _H}\left( {\frac{{\partial {B'_i}}}{{\partial {x_j}}}{{\overline J}_j} + {J'_j}\frac{{\partial {{\overline B}_i}}}{{\partial {x_j}}} - \frac{{\partial {J'_i}}}{{\partial {x_j}}}{{\overline B}_j} - {B'_j}\frac{{\partial {{\overline J}_i}}}{{\partial {x_j}}}} \right) ,
\end{equation}
where some insignificant terms are omitted. Only the terms on the left-hand side remain for homogeneous isotropic turbulence, while the terms on the right-hand side represent the anisotropic and inhomogeneous effects caused by the mean field. We expand fluctuating quantities appearing in (\ref{eq:a2}) as
\begin{equation} \label{eq:a3}
{u'_i} = {u'_{0i}} + {u'_{1i}} + \cdots ,\ \ \
{B'_i} = {B'_{0i}} + {B'_{1i}} + \cdots ,\ \ \
{J'_i} = {J'_{0i}} + {J'_{1i}} + \cdots ,
\end{equation}
where the 0th-order terms denote the homogeneous isotropic turbulent field and the higher-order terms reflect the anisotropic and inhomogeneous effects. We substitute (\ref{eq:a3}) into (\ref{eq:a2}) and solve the turbulent field iteratively. We approximate that the Green's function for the operator on the left-hand side of (\ref{eq:a2}) can be replaced by the turbulent time scale $\tau$ \cite{yosh98}. We then obtain the expression for ${B'_{1i}}$ as follows:
\[
{B'_{1i}} = \tau \left( { - {u'_{0j}}\frac{{\partial {{\overline B}_i}}}{{\partial {x_j}}} + \frac{{\partial {u'_{0i}}}}{{\partial {x_j}}}{{\overline B}_j} + {B'_{0j}}\frac{{\partial {{\overline u}_i}}}{{\partial {x_j}}}} \right)
\]
\begin{equation} \label{eq:a4}
+ {\epsilon _H}\tau \left( {\frac{{\partial {B'_{0i}}}}{{\partial {x_j}}}{{\overline J}_j} + {J'_{0j}}\frac{{\partial {{\overline B}_i}}}{{\partial {x_j}}} - \frac{{\partial {J'_{0i}}}}{{\partial {x_j}}}{{\overline B}_j} - {B'_{0j}}\frac{{\partial {{\overline J}_i}}}{{\partial {x_j}}}} \right) .
\end{equation}
The magnetic field correlation is expanded as
\begin{equation} \label{eq:a5}
\overline {{B'_i}{B'_j}} = \overline {{B'_{0i}}{B'_{0j}}} + \overline {{B'_{0i}}{B'_{1j}}} + \overline {{B'_{1i}}{B'_{0j}}} + \cdots .
\end{equation}
Substituting (\ref{eq:a4}) into (\ref{eq:a5}) and considering the isotropy of the 0th-order quantities, we have
\[
\overline {{B'_i}{B'_j}} = \frac{1}{3}\overline {B'^2_{0k}} {\delta _{ij}} + \frac{\tau }{3}\overline {B'^2_{0k}} {\overline S_{ij}} - \frac{{{\epsilon _H}\tau }}{3}\overline {B'^2_{0k}} \left( {\frac{{\partial {{\overline J}_i}}}{{\partial {x_j}}} + \frac{{\partial {{\overline J}_j}}}{{\partial {x_i}}}} \right)
\]
\begin{equation} \label{eq:a6}
- \frac{\tau }{3}\left( {\overline {{u'_{0k}}{B'_{0k}}} - {\epsilon _H}\overline {{B'_{0k}}{J'_{0k}}} } \right)\left( {\frac{{\partial {{\overline B}_i}}}{{\partial {x_j}}} + \frac{{\partial {{\overline B}_j}}}{{\partial {x_i}}}} \right) .
\end{equation}
The cross helicity $\overline {{u'_{0k}}{B'_{0k}}}$ and the current helicity $\overline {{B'_{0k}}{J'_{0k}}}$ show non-zero values for special kind of MHD turbulence only, whereas the magnetic energy $\overline {B'^2_{0k}}$ always exists.
Therefore, we keep the first three terms on the right-hand side of (\ref{eq:a6}).
The magnetic field correlation also appears in the SGS Reynolds stress defined in (9), for which we adopt the eddy viscosity model given by (\ref{eq:SGS1}).
Comparing the second term in (\ref{eq:a6}) with (\ref{eq:SGS1}), we can see that $(\tau /3)\overline {B'^2_{0k}} \propto {\nu _{SGS}}$.
Therefore, we obtain the model expression for $\overline {{B'_i}{B'_j}} $ as follows:
\begin{equation}
\overline {{B'_i}{B'_j}} = \frac{1}{3}\overline {B'^2_k} {\delta _{ij}} + {\nu _{SGS}}{\sigma _{H,1}}{\overline S_{ij}} - {\epsilon _H}{\nu _{SGS}}{\sigma _{H,2}}\left( {\frac{{\partial {{\overline J}_i}}}{{\partial {x_j}}} + \frac{{\partial {{\overline J}_j}}}{{\partial {x_i}}}} \right) ,
\end{equation}
where ${\sigma _{H,1}}$ and ${\sigma _{H,2}}$ are non-dimensional model constants.
Because the first term on the right-hand side does not contribute to $ - \nabla \times {{\bf{\overline E}}^H}$, the model expression for the SGS Hall electromotive force can be written as follows:
\begin{equation}
\overline E_i^H = {\epsilon _H}\frac{\partial }{{\partial {x_j}}}({\nu _{SGS}}{\sigma _{H,1}}{\overline S_{ij}}) - \epsilon _H^2\frac{\partial }{{\partial {x_j}}}\left[ {{\nu _{SGS}}{\sigma _{H,2}}\left( {\frac{{\partial {{\overline J}_i}}}{{\partial {x_j}}} + \frac{{\partial {{\overline J}_j}}}{{\partial {x_i}}}} \right)} \right] .\end{equation}
\clearpage
\section*{References}
|
\section{Introduction}
Social networks have become increasingly important in our daily life. Commonly, people join multiple social networks to enjoy different types of services at the same time, e.g., Facebook and Twitter. Users usually have separate accounts in different social networks. They can act as bridges connecting the networks. The problem of User Identity Linkage (UIL), which aims to link the identities of the same natural person across different social platforms, has become increasingly crucial among national security, public opinion supervision, and business recommendation all over the world. Developing a highly accurate UIL model contributes a lot to construct a comprehensive view of users' characteristics. An increasing number of applications, including friend recommendation \cite {274, 71}, cross-network information diffusing prediction \cite {273, 72}, link prediction \cite {260}, and network dynamics analysis \cite {73}, have recognized the necessity and benefit of UIL.
Early studies address the UIL problem by leveraging self-reported user attributes, including user profiles \cite {64, 69, 78} (e.g., username, gender, location), user-generated content \cite {258, 256} (e.g., tweets, posts, publications) and user behaviors \cite {247, 79}(e.g., tagging, habits). Many of these attribute-based solutions are heuristic, handling with particular string patterns or similarity comparisons. A study by Zafarani et al. (2009) finds a set of specific rules based on several hypotheses created from empirical observations of username string patterns \cite {78}. The model proposed by Kong et al. (2013) converts each user's posts into a bag-of-words vector that is weighted by the technology of Term Frequency-Inverse Document Frequency (TF-IDF) in multiple heterogeneous social networks. The inner product similarity and the cosine similarity between these vectors can, therefore, be used to find matched user identity pairs \cite {256}. These approaches are sensitive to the similarity measures or the string patterns of user profiles, which share the following two significant shortcomings:
\begin{enumerate}[(i)]
\item \textbf{Lack of a general approach for dealing with multi-types of attributes.}\\
The difference in attribute types creates difficulty in developing general methods. A perfect example can be found in the following three types of user attributes. Username often has a large proportion of customized words (e.g., Tom008, Tom{\&}Jerry). Affiliation is usually a phrase that consists of some regular words (e.g., professor, university). User posts may include several paragraphs and high-level semantic features (e.g., post topic, user opinion). Consequently, traditional heuristic approaches can only cover some types, but never all. There is no general way to handle attributes of multi-types.
\item \textbf{Challenge to mine the hidden high-level semantic connections in texts.}\\
It is required to catch the hidden connections between different attributes. The case can be seen in the two words ``teacher" and ``professor" who appeared in someone's affiliation. String alignment rules cannot match them. However, ``teacher" and ``professor" are highly related in semantics. Thus, traditional heuristic approaches will not apply in such a scenario.
\end{enumerate}
Alternatively, the structural information of the social networks can be directly exploited for the UIL problem. For example, Singh et al. (2008) \cite{261} investigate the model IsoRank on the basis of a pair-wise vertex similarity calculated with spectral graph theory to find identical vertices. Liu et al. (2016) \cite{55} present to learn a network embedding by representing the follower-ship/followee-ship of users as input/output context vectors. According to network structure parameters, a seminal study by Feng et al. (2018) \cite{3} designs three similarity metrics (CPS, CCS, and CPSC). Then, they present a two-stage iterative algorithm CPCC for the UIL problem. However, the scale of real social networks is usually extremely massive and changing dynamically. It is not easy to obtain the complete structural information. Therefore, such methods present challenges to the integration of user attributes and highlight the requirement for joint analysis across multiple attribute categories.
Some recent attention has focused on the application of incorporating both user attributes and network structure information.
Heimann et al. (2018) \cite {18} have shown how to capture the node's structural identity from the degrees of its neighbors and to detect the attribute identity by using an embedding approach. Zhong et al. (2018) \cite{52} suggest incorporating an attribute-based model and a relation-based model into the co-training style, making them reinforce each other iteratively. Many of the methods ignore the apparent differences in text composition and the higher semantic features, e.g., topics. They use one single representation process, e.g., bag-of-word or TF-IDF, to embed all the text attributes into a low-dimensional feature vector. The multi-view attribute embedding mechanism is supported by Zhang et al. (2018) \cite{15}, who adopt each attribute at both the character view and word view. The work still neither distinguish between different types of attributes, nor addresses higher-level semantics such as text topic categories. Li et al. (2020) \cite{419} use the similarity of k-hop neighbors to present the information redundancy of the network structures, which can be applied to perform user identifications by combining name-based user attributes.
This paper addresses to study the semi-supervised UIL problem from a multi-view perspective. The basic idea is to recognize the user features on two target social networks as different viewpoints (feature sets) of the same natural person. The concept of isomorphism \cite {54} shows that the same natural person's social identity has a high correlation in different social feature spaces. Hence, the idea requires to create separate feature spaces for the target social networks. If we can find mappings from the views to the same feature space and maximize the correlations between their projected features, the user identities of the same person can be identified through existing similarity or distance measures.
Previous investigators \cite {61, 50, 51} have examined the popular of the assumption that the correlation between the projections of the two views is linear. Based on the assumption, this study finds that the Canonical Correlation Analysis (CCA) is quite suitable to be employed for the UIL problem. CCA aims to find pairs of projections for different views so that the correlations between them are maximized. It has been successfully used in many research areas, such as computer vision, information retrieval, and natural language processing. This study explores the solution for the UIL task under the semi-supervised manner when we don't have enough label data, or the ability to get more data for producing an accurate model. Classical CCA assumes that feature dimensions are smaller than the number of observed samples. This assumption does not apply in semi-supervised situations. Therefore, we introduce the Regularized CCA (RCCA) \cite{287} that is well adapted to annotate constrained scenarios in the UIL problem.
This study proposes a novel semi-supervised model, called Multi-level Attribute embedding for semi-supervised User Identity Linkage (MAUIL). The task of MAUIL is to connect user identities between two arbitrary social networks, which involves two components: multi-level user attribute embedding and RCCA based projection. First of all, the text attributes for each network are divided into three types: character-level, word-level, and topic-level attributes. In order to capture higher-level semantic features and merge multiple types of attributes without the reliance on any text annotation, three unsupervised methods are exercised to generate three types of vectors, respectively. User relationships in social networks are considered as the fourth user feature type, which are also attached to the feature vector set. Finally, the total four vectors are combined to form the final representation of the users to be linked. The significant advantage of MAUIL is that combining user attributes and network structure increases the model robustness in the cases of attribute missing or the dynamic change of network structure.
On the other hand, this work introduces RCCA to find mappings from social networks to feature spaces. In this study, the mappings can project the UIL networks into a common correlated space for user identity linkage. When a user requires to find his/her shared identity, MAUIL can select his/her matched candidates by using spatial distance comparison. A smaller distance in the common space indicates that the two candidates have a more notable chance to be the same natural person. The benefit of RCCA is that both user attribute embedding and network structure embedding are unsupervised, thus avoiding the dependence on traditional text annotations. We summarize the main contributions as follows:
\begin{enumerate}[(i)]
\item One of the more significant contributions to emerge from this study is that we recommend a multi-level attribute embedding method to deal with multiple user attribute types. User attributes and network structures are integrated to realize a more promising representation of social network users. The MAUIL method is suitable for the unfortunate situation of user attribute missing or sparse network structure, which has strong robustness. These contributions make it possible to study the UIL problem in a semi-supervised way. It remains efficient to capture various feature types and high-level semantic features in text attributes without relying on any text annotations.
\item This paper proposes a novel UIL model MAUIL, which consists of multi-level attribute embedding and RCCA based linear projection. Unlike existing methods, MAUIL represents the UIL problem from a multi-view perspective and realizes the projections
from social networks to shared feature space by maximizing the correlations between these views.
\item This study extensively evaluates MAUIL on two real-world data sets of social networks and coauthor networks. The results show the superior performance of the proposed model by a comprehensive comparison with state-of-the-art baseline methods.
\end{enumerate}
The rest of this paper is organized as follows. Section \ref{sec:Related Work} reviews and summarizes the related works. Section \ref{sec:Definition} formally defines the studied problem. Section \ref{sec:model} details the proposed model. Section \ref{sec:experiment} presents the experimental results. Finally, we conclude this work in Section \ref{sec:conclusion}.
\section{Related Work}
\label{sec:Related Work}
The task of User Identity Linkage (UIL) \cite {78} was initially defined as connecting identical user identities across communities. The problem is also known as user alignment or anchor link prediction, which aims to link the social identities belonging to the same natural person across different social platforms. Formally, the online social communication platforms that share common users are called aligned social networks. These shared users who act like anchors aligning the networks are called
matched user pairs.
Existing UIL approaches can be roughly categorized into three groups, supervised, semi-supervised, and unsupervised methods. Most of the available literatures are supervised, aiming to learn a ranking model or a binary classifier to identify user identities \cite {260, 64, 256, 247, 3, 15, 61,50, 277, 12,1,47}. Similarly, Man et al. (2016) \cite{61} employ network embedding technologies to keep significant structural regularities of networks by supervised pre-matched links. Likewise, Mu et al. (2016) \cite{50} have attempted to project the user identities of multiple networks into the same latent user space. The authors first introduce space distance optimization to solve the UIL problem. The distance between the user identities of the same person (res. different persons) in multiple social networks is minimized (res. maximized) according to distance optimizations. A graph attention embedding model is proposed by Liu et al. (2019) \cite{1}, which exploits the social structures by painting the weighted contribution probabilities between follower-ships and followee-ships. Higher-level feature representations formed the central focus of the study by Qiao et al. (2019) \cite{47} in which a Siamese neural network is proposed to learn the high-level representation of users web-browsing behaviors.
In all the supervised studies reviewed here, the behavior of collecting enough aligned users as initial annotations is inevitable and high labor costs. Hence, some unsupervised approaches are proposed to address the UIL problem without any labeled data \cite{69, 52, 18, 51, 68, 63, 70, 6}. As noted by Liu et al. (2013) \cite{69}, the N-gram model does not rely on any annotated operations, which automatically generates training data by estimating the rarity of usernames. Consequently, a powerful binary classifier is trained to link user identities. To find a new research path, Lacoste-Julien et al. (2013) \cite{68} have highlighted the function of heuristic string similarities and propose a valuable greedy model SiGMa to align the user attributes. Some researchers have mainly been interested in feature mappings or distance optimizations. Nie et al. (2016) \cite{63} propose a dynamic core interest mapping algorithm (DCIM) that combines network structure and user article contents to link identities across networks. Li et al. (2018) \cite{51} found that the earth mover's distance can be utilized to measure the abstract distance between the user identity distributions in two aligned social networks. Zhou et al. (2018) \cite{6} work on the UIL problem complemented by introducing friend relationship in social networks, who propose a user identification algorithm without prior knowledge.
Unsupervised approaches have brought convenience to avoid relying on labeled data. However, the very technologies have also created some potential problems almost every work is exposed. Unsupervised methods usually suffer from lower performances compared with supervised ones.
Therefore, several semi-supervised approaches have recently emerged
to solve the UIL problem. The typical characteristics of a semi-supervised approach is the common use of unlabeled samples along with a few annotations. As a result, unlabeled identity pairs can be predicted more accurately during the learning process \cite {55, 65, 66, 7, 5, 16,75, 249}.
Recent studies have benefited a lot from semi-supervised models. Semi-supervised idea exploits the advantages of both the annotation behavior and the unlabeled samples to capture the shape of underlying data distribution. Semi-supervised solutions for the UIL problem are more encouraging to link social identities within double `E' (effectively and efficiently) pattern. An energy-based model (COSNET) by Zhang et al. (2015) \cite{66} reports that jointly consideration of local user matching, network structure, and global consistency is helpful to increase the accuracy of UIL. Similarly, the approach of embedding the follower-ship/followee-ship of users into input/output context vector reputations was studied extensively by Liu et al. (2016) \cite{55}. In the embedding process, a few annotations are used to learn a semi-supervised classifier. Zhou et al. (2016) \cite{7} discovered that user identities could be iteratively linked by calculating network structural matching degrees. Li et al. (2019) \cite{16, 75} present an adversarial learning framework to minimize the distances between
user distributions of multiple social networks.
Graph neural network frameworks are available for UIL problem. Zhang et al. (2019) \cite{249} present a graph embedding model to capture both local and global information in the user-user social graph.
The above methods share the same weakness, which can neither cover the entire set of different attribute features, nor capture higher-level semantic features in attribute text. This paper proposes a novel semi-supervised model MAUIL to link user identities
between two arbitrary social networks.
\section{Problem Definition}
\label{sec:Definition}
This section proposes the basics of the studied problem. Table \ref{Tab:Notations} summarizes the notations involved in this paper.
A social network is a three-tuple $G=(V,E,A)$, where $V=\{v_1,v_2,\ldots,v_n\}$ and $E=\{e_{ij}=(v_i,v_j)|v_i,v_j\in V\}$ represent the set of $n$ users and the set of undirected edges between users, respectively. The edge $e_{ij}$ denotes the binary status of the link between user $v_i$ and $v_j$, e.g., friendships on Facebook. $A=(A_c,A_w,A_t)$ is a set of user attributes, e.g., names, affiliations, and education experiences, which consists of three subsets:(1) character-level attributes $A_c$, (2) word-level attributes $A_w$, and (3) topic-level attributes $A_t$.
\begin{table}[h!]
\caption{Main notations used in this paper}\label{Tab:Notations}
\centering
\begin{small}
\setlength{\tabcolsep}{3pt}
\renewcommand{\arraystretch}{1.2}
\begin{tabular}{p{60pt}<{\centering}|p{180pt}}
\toprule[1.0pt]
\textbf{Notation}&\textbf{Description}\\
\hline
$G^X/G^Y$ & Networks $X$ and $Y$ that participate in UIL. \\
$V$ & Set of users in a social network.\\
$E$ & Set of edges in a social network.\\
$A$ & Set of user attributes in a social network.\\
$\mathbb{R}$ & Set of real numbers.\\
$A_c,P_c$ & Set of character-level attributes and its feature matrix.\\
$A_w,P_w$ & Set of word-level attributes and its feature matrix.\\
$A_t,P_t$ & Set of topic-level attributes and its feature matrix.\\
$P_s$ & Structure feature matrix of a social network.\\
$d_c,d_w,d_t,d_s$ & Dimensions of character-level, word-level, topic-level, and structure embedding.\\
$X/Y$ & Combined feature matrices of $G^X/G^Y$.\\
$H/M$ & Canonical matrices of a pair of networks.\\
$\mathbf{h}_i/\mathbf{m}_i$ & $i$-th pair of canonical vectors.\\
$C_{XY}$ & Covariance matrix of $X$ and $Y$.\\
$\hat{C}_{XX}$& Regularized covariance matrix of network $X$.\\
$r^X/r^Y$ & Regularization coefficients of $C_{XX}/C_{YY}$\\
$Z^X/Z^Y$ & Projected feature matrices of $X/Y$.\\
\bottomrule[1.0pt]
\end{tabular}
\end{small}
\end{table}
This paper considers such a scenario: some users in two different social networks are the same nature person that forms the anchor link across the networks. We refer to one of the networks as $G^X=(V^X,E^X,A^X)$ and the other as $G^Y=(V^Y,E^Y,A^Y)$.
This investigation intends to identify, if any, the counterpart in $G^Y$ (resp. $G^X$) for each node in $G^X$ (resp. $G^Y$). The user identity linkage problem can be formalized as follows.
\newdefinition{definition}{Definition}
\begin{definition}
\textbf{User Identity Linkage (UIL)}: Given two arbitrary social networks $G^X$ and $G^Y$, and a few pre-matched user pairs $S=\{(v_i,v_j)|v_i\in V^X,v_j\in V^Y \}$, the UIL problem is to find the other hidden matched identity pairs $L=\{(v_i,v_j)|v_i\in V^X,v_j\in V^Y,(v_i,v_j)\notin S\}$, where $v_i$ and $v_j$ belong to a same natural person.
\end{definition}
This study proposes a novel semi-supervised model called MAUIL (Multi-level Attribute embedding for semi-supervised User Identity Linkage) to solve the UIL problem. In this model, the networks $G^X$ and $G^Y$ are embedded into two latent feature spaces $P^X\in \mathbb{R}^{d\times n}$ and $P^Y\in \mathbb{R}^{d\times n}$, independently, where $d$ and $n$ denote the number of dimensions and that of users, respectively. Besides, $G^X$ and $G^Y$ are mapped into a common correlated space by using two projections $H\in \mathbb{R}^{d\times k}$ and $M\in \mathbb{R}^{d\times k}$, respectively, where superscript $k$ represents the number of projection vectors.
\section{Model Specification}
\label{sec:model}
\subsection{Overview}
\begin{figure*}[!h]
\begin{center}
\includegraphics[width=0.8\textwidth]{example.pdf}
\caption{\small A case of text attributes.}\label{Fig:case_textattrubutes}
\end{center}
\end{figure*}
This section presents the details of MAUIL. The model has two components: multi-level attribute embedding and RCCA based projection. We utilize MAUIL to detect the potential user identity links covering target networks $G^X$ and $G^Y$. First, the text attributes of each network are divided into three types: character-level attributes $A_c$, word-level attributes $A_w$, and topic-level attributes $A_t$. The example of the text attributes can be seen in the case of Figure \ref{Fig:case_textattrubutes}. Second, three unsupervised methods (Discussed in sections \ref{sec:char}, \ref{sec:word}, and \ref{sec:topic}) are employed to generate three corresponding feature matrices $P_c$, $P_w$, and $P_t$, respectively. Besides, the user's relationship in social networks has been proved to be beneficial for UIL tasks. In this study, user relationships are evolved as a particular type of user feature $P_s$ (Discussed in section \ref{sec:network}).
As a result, a total of four feature matrices $P_c$, $P_w$, $P_t$, and $P_s$ are combined to form the final representation of the target social networks $G^X$ and $G^Y$.
On the other hand, RCCA is usually used to explore the correlation between two multi-variables (vectors). In this study, we introduce RCCA to find the mappings that can project the networks $G^X$ and $G^Y$ into a common correlated space (Discussed in section \ref{sec:CCA}).
\begin{figure*}[!h]
\begin{center}
\includegraphics[width=0.9\textwidth]{model.pdf}
\caption{The overview of MAUIL model, including two components: multi-level attribute embedding and RCCA based projection.}\label{Fig:MAUIL}
\end{center}
\end{figure*}
For any user belonging to network $G^X$ or $G^Y$, we identify the most likely counterpart in the other network by comparing their distance in the common correlation space. A closer distance indicates that the two users are more likely to be the same natural person. The overview of MAUIL is presented in Figure \ref{Fig:MAUIL}.
\subsection{Multi-level Attribute Embedding}
\label{sec:emb}
In this section, different social networks are embedded in the same way. Hence, we use the same notations $G$ without distinguishing $G^X$ and $G^Y$.
\subsubsection{Character-level attribute embedding}
\label{sec:char}
This section intends to preserve the text similarity of similar users through the character-level attribute embedding.
The character-level attribute of user $v_i$ in a social network $G$ with $n$ users is denoted by $a_{i}^c$, which is specified as the username string that may contain part of the characters, spaces, and special symbols. Attribute $a_{i}^c$ can be divided into a series of total $m$ unique tokens $\mathbf{w}={w_1,w_2,\ldots,w_k,\ldots,w_m}$, including characters, numbers, and q-grams \cite {81}. The vector form of $a_{i}^c$ refers to a count-weighted expression $\overrightarrow{\mathbf{x}_{i}^c}=[\mathbf{x}_{w_1},\mathbf{x}_{w_2},\ldots,\mathbf{x}_{w_k},\ldots,\mathbf{x}_{w_m}]\in \mathbb{R}^{m\times 1}$, where for any $k\in\{1,2,\ldots,m\}$, $\mathbf{x}_{w_k}$ is the count number of the corresponding piece $w_k$ in $\mathbf{w}$.
This study achieves the count-weighted expression by performing a simple bag-of-word vectorization calculation.
The above formalizations can be illustrated by the user $v_i$ shown in Figure \ref{Fig:case_textattrubutes}. In terms of the username, we have the character-level attribute $a_{i}^c=$``Yoshua Bengio'' that can be divided in to a series of tokens $\mathbf{w}=$`a':1,`b':1,`c':0,$\ldots$,`o':2,$\ldots$, others:0.
Their count numbers follow each token in the text. Finally, the corresponding count-weighted vector is
$\overrightarrow{\mathbf{x}_{i}^c}=[1,1,0,\ldots,2,\ldots]$.
Having discussed how to represent the character-level attribute of a user, the follow description addresses ways of considering all the $n$ users in a social network. The notation $A_c=\{a_{1}^c,a_{2}^c,\ldots,a_{i}^c,\ldots,a_{n}^c\}$ is used here to refer to the set of character-level attributes for $n$ users. Similarly, the vectorization of set $A_c$ can be represented as the count-weighted vector sequences $X_c=[\overrightarrow{\mathbf{x}_{1}^c},\overrightarrow{\mathbf{x}_{2}^c},\ldots,\overrightarrow{\mathbf{x}_{i}^c},\ldots,\overrightarrow{\mathbf{x}_{n}^c}]^{T}$
that is also called the character-level count-weighted matrix.
Data dimensionality reduction technologies are initially applied in deep image learning, which have gradually turned to the NLP area and achieved excellent results. Autoencoders are necessarily served as an effective way to reduce the dimension of data and reduce the burden of deep learning. This study employs a one-layer Autoencoder to embed the count-weighted vectors into distributed representations. First, the Autoencoder maps an input vector $\overrightarrow{\mathbf{x}}$ to a hidden representation $\overrightarrow{\mathbf{z}}$ through a deterministic function $\overrightarrow{\mathbf{z}}=f(\overrightarrow{\mathbf{x}})=W\overrightarrow{\mathbf{x}}+b$, where $W\in \mathbb{R}^{d_c\times m}$ and $b\in \mathbb{R}^{d_c\times 1}$ are the weight matrix and a bias vector, respectively. Second, the resulting latent representation $\overrightarrow{\mathbf{z}}$ is mapped back to a ``reconstructed'' vector $\overrightarrow{\mathbf{y}}$ in input space $\overrightarrow{\mathbf{y}} = g (\overrightarrow{\mathbf{z}})= W^*\overrightarrow{\mathbf{z}} + b^*$, where $W^*\in \mathbb{R}^{d_c\times m}$ is the weight matrix of the reverse mapping function $g(.)$.
This study optimizes the parameters to minimize the average reconfiguration loss with:
\begin{equation}\label{AE_loss}
L=\frac{1}{n}\sum\|x_i-y_i\|^2
\end{equation}
After optimization, we can obtain the character-level feature matrix $P_c=[\overrightarrow{\mathbf{p}_{1}^c},\overrightarrow{\mathbf{p}_{2}^c},\ldots,\overrightarrow{\mathbf{p}_{i}^c},\ldots,\overrightarrow{\mathbf{p}_{n}^c}]^{T}$
of the character-level attribute $A_c$ with $d_c$ dimensions by:
\begin{equation}\label{char_emb}
P_c= WX_c+b
\end{equation}
where $W$ and $b$ are the weight matrix and the bias of Autoencoder. For any $i\in\{1,2,\ldots,n\}$, $\overrightarrow{\mathbf{p}_{i}^c}$ is the character-level feature vector of the count-weighted vector $\overrightarrow{\mathbf{x}_{i}^c}$. Algorithm \ref{alg:char_emb} presents the entire procedure of character-level attribute embedding by a function, called CharLevelEmb.
\renewcommand{\algorithmicrequire}{ \textbf{Input:}}
\renewcommand{\algorithmicensure}{ \textbf{Output:}}
\begin{algorithm}[h!]
\caption{Character-level attribute embedding.}
\label{alg:char_emb}
\begin{small}
\begin{algorithmic}[1]
\Require The character-level attribute set $A_c$ and the character-level embedding dimension $d_c$.
\Ensure The character-level attribute feature matrix $P_c$.
\Function{CharLevelEmb}{$A_c,d_c$}
\State Initialize $X_c$
\For {$\forall a_{i}^c$ in $A_c$}
\State $\mathbf{w}=CharTokenize(a_{i}^c)$
\State $\overrightarrow{\mathbf{x}_{i}^c}=CountVector(\mathbf{w})$
\State $X_c.append(\overrightarrow{\mathbf{x}_{i}^c})$
\EndFor
\State Initialize $W,b$
\Repeat
\For {$\forall \overrightarrow{\mathbf{x}}$ in $X_c$}
\State $\overrightarrow{\mathbf{z}}=f(\overrightarrow{\mathbf{x}})=W\overrightarrow{\mathbf{x}}+b$
\State $\overrightarrow{\mathbf{y}}=g(\overrightarrow{\mathbf{z}})=W\overrightarrow{\mathbf{z}}+b$
\State update $W,b$ based on Eq. $\ref{AE_loss}$
\EndFor
\Until Convergence
\State $P_c \gets$ Eq. $\ref{char_emb}$
\State \textbf{return} $P_c$
\EndFunction
\end{algorithmic}
\end{small}
\end{algorithm}
\subsubsection{Word-level attribute embedding}
\label{sec:word}
Word2vec \cite {290} is one of the most common procedures for transforming texts to feature vectors. This part applies the Word2vec approach to capture the word-level characteristics of user attributes.
The word-level attribute of user $v_i$ in a social network is represented as $a_{i}^w$, which consists of the phrases or short sentences such as gender, location, affiliations, and education experiences. The attribute $a_{i}^w$ can be split into a sequence of total $m$ words $\mathbf{w}_i=w_{i1},w_{i2},\ldots,w_{ik},\ldots,w_{im}$, where $w_{ik}$ is the vocabulary expression of the $k$-th word in $a_{i}^w$. A more detailed of the above formalization is given in the example of the user ``Yoshua Bengio'' that is shown in Figure \ref{Fig:case_textattrubutes}. His word-level attribute is $a_{i}^w=$``University of Montréal, Department of Computer Science and Operations Research, QC, Canada'', which is tokenized into a sequence of words followed by its count number as below:
$\mathbf{w}_i=$``University'':1,``of'':2,``Montréal'':1, ``Department'':1, $\ldots$,others:0. Each word is later represented by a unique number as its ID in the vocabulary.
We use $A_w=\{a_{1}^w,a_{2}^w,\ldots,a_{i}^w,\ldots,a_{n}^w\}$ to denote the set of word-level attributes in a social network with $n$ users. Each $\mathbf{w}_i$ of the attribute $a_{i}^w$ is treated as a word-level document for deep learning. Then, all the $n$ documents constitute a corpus. We plan to train the word vectors in the corpus by using the language model CBOW \cite {290}. In the training process, for any $k\in\{1,2,\ldots,m\}$, the word vector of $w_{ik}$ in document $\mathbf{w}_i$ is a $d_w$-dimensional vector $\overrightarrow{\mathbf{x}_{ik}}\in R^{d_w\times 1}$. Hence, the word-level attribute vector of user $v_i$ can be denoted as $\overrightarrow{\mathbf{p}_i^w}$ and calculated by summing up the word vectors of all words in document $\mathbf{w}_i$:\\
\begin{equation}\label{word_emb_sum}
\overrightarrow{\mathbf{p}_i^w}=\sum_{w_{ik}\in \mathbf{w}_i} \overrightarrow{\mathbf{x}_{ik}}
\end{equation}
There is a particular problem with the practice of word-level attributes. Word-level attributes compared with character-level ones may be missing or not distinguishable enough to provide sufficient evidence. The solution is to consider the additional evidence offered by their neighbors according to the homophily principle \cite {252}. Thus, we smooth the embedding of each user by their neighbor embeddings that are regularized by a real number parameter $\lambda\in [0,1]$:\\
\begin{equation}\label{word_emb_ave}
\overrightarrow{\mathbf{p}_i^w}=(1-\lambda) \sum_{w_{ik}\in \mathbf{w}_i}{\overrightarrow{\mathbf{x}_{ik}}} +\lambda \frac{1}{s_i} \sum_{j\in \mathcal{N}_{i}}\sum_{w_{jk}\in \mathbf{w}_j}{\overrightarrow{\mathbf{x}_{jk}}}
\end{equation}
where $\mathcal{N}_{i}=\{v_j|(v_i,v_j)\in E\}$ is the neighbor set of user $v_i$ and $s_i=|\mathcal{N}_{i}|$ is the number of the neighbors.
As a result, the word-level attributes $A_w$ of all users can be represented by a word-level feature matrix $P_w=[\overrightarrow{\mathbf{p}_{1}^w},\overrightarrow{\mathbf{p}_{2}^w},\ldots,\overrightarrow{\mathbf{p}_{i}^w},\ldots,\overrightarrow{\mathbf{p}_{n}^w}]^T$. Algorithm \ref{alg:word_emb} presents the detailed procedure of word-level attribute embedding by a function, called WordLevelEmb.
\begin{algorithm}[h!]
\caption{Word-level attribute embedding.}
\label{alg:word_emb}
\begin{small}
\begin{algorithmic}[1]
\Require The word-level attribute set $A_w$, the word-level embedding dimension $d_w$, the regularization real number parameter $\lambda$, and the network edge set $E$.
\Ensure The word-level attribute feature matrix $P_w$.
\Function{WordLevelEmb}{$A_w,d_w,\lambda,E$}
\State $wordvectors = Word2Vec(A_w,dim=d_w)$
\State Initialize $P_w \gets \mathbf{0}$.
\For {$\forall a_{i}^w$ in $A_w$}
\State $\mathbf{w}_i=WordTokenize(a_{i}^w)$
\State Initialize $\overrightarrow{\mathbf{p}_{i}^w} \gets \mathbf{0}$.
\For {$\forall w_{ik}$ in $\mathbf{w}_i$}
\State $\overrightarrow{\mathbf{x}_{ik}}=LookUpEmbeddings(wordvectors,w_{ik})$
\State $\overrightarrow{\mathbf{p}_{i}^w} += \overrightarrow{\mathbf{x}_{ik}}$ (Eq. $\ref{word_emb_sum}$)
\State $P_w.append(\overrightarrow{\mathbf{p}_{i}^w})$
\EndFor
\EndFor
\State $P^*_w = copy(P_w)$
\For {$\forall a_{i}^w$ in $A_w$}
\State Initialize $(\overrightarrow{\mathbf{p}_{i}^{w}})^* \gets \mathbf{0}$.
\State $\mathcal{N}_i =GetNeighbors(E,v_i)$
\For {$\forall v_j$ in $\mathcal{N}_i$}
\State $\overrightarrow{\mathbf{p}_{j}^w} = P^*_w[j]$
\State $(\overrightarrow{\mathbf{p}_{i}^{w}})^* += \overrightarrow{\mathbf{p}_{j}^w}$
\EndFor
\State $\overrightarrow{\mathbf{p}_{i}^w} = (1-\lambda)\overrightarrow{\mathbf{p}_{i}^w}+\frac{\lambda}{s_i}(\overrightarrow{\mathbf{p}_{i}^w})^*$ (Eq. $\ref{word_emb_ave}$)
\State Update $P_w \gets \overrightarrow{\mathbf{p}_{i}^w}$.
\EndFor
\State \textbf{return} $P_w$
\EndFunction
\end{algorithmic}
\end{small}
\end{algorithm}
\subsubsection{Topic-level attribute embedding}
\label{sec:topic}
This study employs Latent Dirichlet Allocation (LDA) model \cite {100} to extract the topic-level features of user attributes.
The attribute texts of user $v_i$ in a social network, consisting of the paragraphs or articles such as posts, blogs, and publications, are combined to form the topic-level attribute $a_{i}^t$. In a similar case shown in Figure \ref{Fig:case_textattrubutes}, the texts of user ``Yoshua Bengio'' include three publications. Their titles or full texts are combined to form the $a_{i}^t$. Hence, we have the set of topic-level attributes $A_t=\{a_{1}^t,a_{2}^t,\ldots,a_{i}^t,\ldots,a_{n}^t\}$ for all users in the social network DBLP. In this study, the $i$-th topic-level attribute $a_{i}^t$ in $A_t$ is treated as the topic-level document $\mathbf{w}_i$ for deep learning.
LDA is a kind of document-topic generation model. A document can contain multiple topics, and each word in the document is generated by one of the topics. LDA can present the topic of each document in a document set as a probability distribution. For example, a document $\mathbf{w}_i$ can be modelled by a topic distribution $\theta_{\mathbf{w}_i}$. As a result, LDA extracts a few words in the vocabulary to describe each topic. In this study, we consider the topic distributions as the topic-level features of user attributes for the UIL problem.
Let $\mathbf{z}=\{1,2,\ldots,d_t\}$ be a set of topic indexes. Notation $z_i\in \mathbf{z}$ refers to the topic index of document $\mathbf{w}_i$. A latent topic with index $z_i$ in LDA is characterized by exactly the word distribution $\phi_{z_i}$. Topic distributions provide concise representations of documents. The process of LDA is as follows:
\\
\par\setlength{\parindent}{1em}1. For each topic index $z\in \mathbf{z}$:
\par\setlength{\parindent}{2em}(a) Choose a word distribution $\phi_z\sim Dirichlet(\beta)$.
\par\setlength{\parindent}{1em}2. For each document $\mathbf{w}_i$ in a corpus:
\par\setlength{\parindent}{2em}(a) Choose a topic distribution $\theta_{\mathbf{w}_i} \sim Dirichlet(\alpha)$.
\par\setlength{\parindent}{2em}(b) For each word $w_i$ of the $\overline{\mathcal{N}}$ words in the model vocabulary:
\par\setlength{\parindent}{3em}a. Choose a topic with the index $z_i\in \mathbf{z}\sim Multinomial(\theta_{\mathbf{w}_i})$.
\par\setlength{\parindent}{3em}b. Choose a word $w_i\sim Multinomial(\phi_{z_i})$.\\
\setlength{\parindent}{0.8em}
\\
where $Dirichlet(\cdot)$ and $Multinomial(\cdot)$ denote the Dirichlet distribution and Multinomial distribution, respectively. $\alpha$ and $\beta$ are hyper-parameters of Dirichlet distribution, specifying the property of the priors on topic and word distributions. $\overline{\mathcal{N}}$ represents the total number of words in the vocabulary used in the model.
In general, let $\mathbf{w}$ be a document, $\mathbf{z}$ be a set of topic indexes that identify topics, $\theta$ be a topic distribution, $\phi$ be a word distribution, and $\phi_z$ be a word distribution with respect to topic index $z_i\in\mathbf{z}$. LDA results in the following joint distribution:
\begin{equation}\label{LDA_p}
p(\mathbf{w},\mathbf{z},\theta,\phi|\alpha,\beta) =p(\phi|\beta)p(\theta|\alpha)p(\mathbf{z}|\theta)p(\mathbf{w}|\phi_z)
\end{equation}
Then, the posterior distribution $p(\mathbf{z|\mathbf{w}})$ can be calculated by Gibbs Sampling \cite {279} as:\\
\begin{equation}\label{LDA_Gibbs}
\begin{split}
p(z_i=j|\mathbf{z}_{-i},\mathbf{w})&\propto \frac{WT_{-i,j}^{(w_i) }+\beta}{\sum_{i=1}^{N}{WT_{-i,j}^{(w_i)}} +\overline{\mathcal{N}}\beta}\\
&\cdot \frac{DT_{-i,j}^{(\mathbf{w})}+\alpha}{\sum_{k=1}^{d_t}{DT_{-i,k}^{(\mathbf{w})} +d_t \alpha}}
\end{split}
\end{equation}
where $z_i=j$ represents the topic that generates the word $w_i$, which is just being indexed by $j\in\mathbf{z}$. Notation $\mathbf{z}_{-i}$ denotes all other topics except for $z_i$. In a corpus, i.e., a set of documents, a word may appear more than once in different documents.
Hence, the count relation between words and topics can be represented as a matrix $WT$. An entry of $WT$, denoted by $WT_j^{(w_i)}$, means the number of word $w_i$ assigned to topic $j$. Similarly, the number of word $w_i$ assigned to topic $j$ and appeared in a corpus without including current $w_i$ itself is depicted by the notation $WT_{-i,j}^{(w_i)}$. Correspondingly, we have the count relation matrix $DT$ between documents and topics. An entry of $DT$, denoted by $DT_j^{(\mathbf{w})}$, means the total number of all words in the document $\mathbf{w}$ assigned to topic $j$. Similarly, the total number of all words in the document $\mathbf{w}$, which have been assigned to topic $j$ without including current $w_i$ itself is depicted by the notation $DT_{-i,j}^{(\mathbf{w})}$.
For a numbered document $\mathbf{w}_i$, the estimate of a topic distribution $\theta_{\mathbf{w}_i}$ over topic $j$ can be calculated as:
\begin{equation}\label{LDA_theta}
\hat{\theta}_j^{(\mathbf{w}_i)}=\frac{DT_j^{\mathbf{w}_i}+\alpha}{\sum_{k=1}^{d_t}{DT_k^{(\mathbf{w}_i)}+d_t\alpha}}
\end{equation}
Thus, we can obtain the topic probability vector of each user $v_i$ as $\overrightarrow{\mathbf{p}_{i}^t}=\{[\hat{\theta}_j^{(\mathbf{w}_i}])\}_{j=1}^{d_t}$, which is the proposed topic-level attribute feature vector. Consequently, the topic-level attributes $A_t$ of all users can be represented by a topic-level feature matrix $P_t=[\overrightarrow{\mathbf{p}_{1}^t},\overrightarrow{\mathbf{p}_{2}^t},\ldots,\overrightarrow{\mathbf{p}_{i}^t},\ldots,\overrightarrow{\mathbf{p}_{n}^t}]^T$.
The pseudo-code function, called TopicLevelEmb, of topic-level attribute embedding is shown in Algorithm \ref{alg:topic_emb}.
\begin{algorithm}[h!]
\caption{Topic-level attribute embedding.}
\label{alg:topic_emb}
\begin{small}
\begin{algorithmic}[1]
\Require The topic-level attribute set $A_t$, the topic-level embedding dimension $d_t$, and the hyper-parameters $\alpha$ and $\beta$ of Dirichlet distribution.
\Ensure The topic-level attribute feature matrix $P_t$.
\Function{TopicLevelEmb}{$A_t,d_t,\alpha,\beta$}
\State Initialize topic assignments $\mathbf{z}$, and counters $DT, TW$
\For{each iteration}
\State update $\mathbf{z},DT,TW$ based on Gibbs Sampling Eq. $\ref{LDA_Gibbs}$
\EndFor
\State Initialize $P_t\gets \mathbf{0}$.
\For{$\forall\mathbf{w}_i$ in $A_t$}
\State Initialize $\overrightarrow{\mathbf{p}_{i}^t}\gets \mathbf{0}$.
\For {$j=1:d_t$}
\State $ \hat{\theta}_j^{(\mathbf{w}_i)} \gets$ Eq. $\ref{LDA_theta}$
\State $\overrightarrow{\mathbf{p}_{i}^t}.append( \hat{\theta}_j^{(\mathbf{w}_i)})$
\EndFor
\State $P_t.append(\overrightarrow{\mathbf{p}_{i}^t})$
\EndFor
\State \textbf{return} $P_t$
\EndFunction
\end{algorithmic}
\end{small}
\end{algorithm}
\subsubsection{Network Structure Embedding and Feature Combining}
\label{sec:network}
Using network structure embedding has been well accepted for its distinct advantages on the UIL problem. Network structures have been playing an increasingly important role in helping researchers find similar user relations. The main idea of utilizing structure information is to map the network into a latent space, such that the users with similar structural roles are positioned close to each other. Several successful methods currently exist for the network embedding. This study empirically chooses LINE \cite {312} to embed the network into the feature matrix $P_s$ with $d_s$ dimensions.
Consequently, we combine the developed three attribute features $P_c$, $P_w$, and $P_t$ with the structure features $P_s$ by a concatenate operation to obtain the final representation of the social network $G^X/G^Y$, denoted as:
\begin{equation}\label{NE_comb}
\begin{split}
X&=[P_c^X;P_w^X;P_t^X;P_s^X ]\in \mathbb{R}^{d\times n}\\
Y&=[P_c^Y;P_w^Y;P_t^Y;P_s^Y ]\in \mathbb{R}^{d\times n}
\end{split}
\end{equation}
where $d=d_c+d_w+d_t+d_s$ is the dimension of the combined feature matrix. To bring these different features onto the same scale, we further add standardization on the feature matrices by centering each row at mean 0 with standard deviation 1. The standardization of feature matrices can establish different features on the same scale. In this paper, each row of a matrix's standardization center is set to mean 0 with standard deviation 1.
\subsection{RCCA based projection}
\label{sec:CCA}
Section \ref{sec:emb} has shown how to establish the combined feature metrics $X$ and $Y$ using the text attributes and the structures of a pre-matched network pair $G^X/G^Y$. This section considers the method to practice user identity links by proposed two feature metrics. First of all, we follow the widespread assumption \cite {61, 50, 51} that the correlations between the projections of linked networks are linear. On top of that, Regularized Canonical Correlation Analysis (RCCA) \cite {285} is particularly useful in maximizing their correlations. Hence, this study adopts an RCCA approach to investigate the identity links between similar users in the UIL problem.
RCCA approaches mostly defined the canonical matrices as $H=[\mathbf{h} _1,\mathbf{h} _2,\ldots,$ $\mathbf{h}_i,\ldots,\mathbf{h}_k]\in \mathbb{R}^{d\times k}$ and $M=[\mathbf{m} _1,\mathbf{m} _2,\ldots,$ $\mathbf{m}_j,\ldots,\mathbf{m}_k]\in \mathbb{R}^{d\times k}$, including
$k$ pairs of linear projections. The canonical matrices of the UIL problem are solved by mapping the vector reputation $X/Y$ of the associated social networks $G^X/G^Y$ into a common correlated space $Z$ through a series of well-designed linear projections to maximize the correlation $\rho$ between $H^TX$ and $M^TY$. Then, the potential user identity links between $G^X$ and $G^Y$ can be estimated by comparing the distance of their vectorization features in $Z$. The detailed steps of RCCA based projection are specified below.
First of all, for each pair of canonical vectors $\mathbf{h}_i\in \mathbb{R}^{d\times 1}$ of $X$, and $\mathbf{m}_j\in \mathbb{R}^{d\times 1}$ of $Y$, the purpose of RCCA is to maximize the correlation $\rho$ between $\mathbf{h}_i^TX$ and $\mathbf{m}_j^TY$ \cite {285}, i.e.:
\begin{equation}\label{rho}
\begin{split}
\rho &=\max corr(\mathbf{h}_i^TX, \mathbf{m}_j^TY) \\
&=\max \frac{cov(\mathbf{h}_i^TX, \mathbf{m}_j^TY)}{\sqrt{var(\mathbf{h}_i^TX)var(\mathbf{m}_j^TY)}} \\
&=\max \frac{\mathbf{h}_i^TC_{XY}\mathbf{m}_j^T}{\sqrt{(\mathbf{h}_i^TC_{XX}\mathbf{h}_i)(\mathbf{m}_j^TC_{YY}\mathbf{m}_j)}}
\end{split}
\end{equation}
where the superscript $T$ of $\mathbf{h}_i^T$ or $\mathbf{m}_j^T$ is the transpose of the vector $\mathbf{h}_i$ or $\mathbf{m}_j$. $C_{XY}$, $C_{XX}$, and $C_{YY}$ are the covariance matrices involving feature vectors $X$ and $Y$ of social networks. Vector centralization can avoid the adverse effects of outliers and extreme values. The vectors $X$ and $Y$ obtained in the previous section are zero-centered. Hence, we can calculate their covariance matrices by $C_{XY}=\frac{1}{n}XY^T$, $C_{XX}=\frac{1}{n}XX^T$, and $C_{YY}=\frac{1}{n}YY^T$, respectively.
The objective of Eq. $\ref{rho}$ is equivalent to the following constrained optimization problem since any canonical vector pair $\mathbf{h}_i$ and $\mathbf{m}_j$ are scale-independent.
\begin{equation}\label{CCAobjective}
\begin{split}
\max \ & \mathbf{h}_i^TXY \mathbf{m}_j \\
s.t. \ & \mathbf{h}_i^TC_{XX}\mathbf{h}_i=1 \\
& \mathbf{m}_j^TC_{YY}\mathbf{m}_j=1
\end{split}
\end{equation}
However, when we have a higher feature dimension $d$ and a relatively smaller number $\mathcal{T}$ of training samples, especially $d>\mathcal{T}$, covariance matrices $C_{XX}$ and $C_{YY}$ are singular. The problem is remedied by adding the regularization \cite {287} to the covariance matrices as:\\
\begin{equation}\label{CCAreg}
\begin{split}
\hat{C}_{XX} &= C_{XX}+r^XI \\
\hat{C}_{YY} &= C_{YY}+r^YI
\end{split}
\end{equation}
where $r^X$ and $r^Y$ are non-negative regularization coefficients, and $I$ is the identity matrix. The canonical matrices $H$ and $M$ can be obtained through solving a generalized eigenvalue decomposition problem \cite {285}:\\
\begin{equation}\label{CCAsovle}
\begin{bmatrix}
\mathbf{0} & C_{XY} \\
C_{YX} & \mathbf{0}
\end{bmatrix}
\begin{bmatrix}
H \\
M
\end{bmatrix}
=
\lambda
\begin{bmatrix}
\hat{C}_{XX} & \mathbf{0} \\
\mathbf{0} &\hat{C}_{YY}
\end{bmatrix}
\begin{bmatrix}
H\\
M
\end{bmatrix}
\end{equation}
On top of that, projecting the linked social networks $G^X$ and $G^Y$ to a new common correlated space is available using the canonical matrices $H$ and $M$. We can link users' identities by calculating the distances of their projected features obtained by $Z^X=H^TX$ and $Z^Y=M^TY$ for $G^X$ and $G^Y$, respectively. This study takes the Euclidean distance as the measure. The distance of users $u_i$ and $u_j$ separately from the network $G^X$ and $G^Y$ in the common space $Z$ can be calculated by the square of a binorm:
\begin{equation}\label{Eq.Dist}
D(z_i^X,z_j^Y )=\|z_i^X-z_j^Y\|_2^2
\end{equation}
A closer distance $D$ of two users represents that they have a greater chance of becoming a same natural person. Consequently, the overall procedure of MAUIL can be given in Algorithm \ref{alg:MAUIL}.
\begin{algorithm}[h!]
\caption{Overall procedure of MAUIL.}
\label{alg:MAUIL}
\begin{spacing}{1.2}
\begin{small}
\begin{algorithmic}[1]
\Require Two social networks $G^X=(V^X,E^X,A^X)$ and $G^Y=(V^Y,E^Y,A^Y)$ , a set of pre-matched user pairs $S$, the dimension $d_c,d_w,d_t$, and $d_s$, the number $k$ of projection vectors.
\Ensure A set of matched user pairs $L$.
\State $A^X_c,A^X_w,A^X_t,A^Y_c,A^Y_w,A^Y_t=Preprocess(A^X,A^Y)$
\State $P_c^X / P_c^Y=\Call{CharLevelEmbed}{A^X_c / A^Y_c}$
\State $P_w^X / P_w^Y=\Call{WordLevelEmbed}{A^X_w / A^Y_w}$
\State $P_t^X / P_t^Y=\Call{TopicLevelEmbed}{A^X_t / A^Y_t}$
\State $P_s^X / P_s^Y=\Call{NetworkEmbed}{E^X / E^Y}$
\State $X,Y \gets$ Eq. $\ref{NE_comb}$
\State Standardize each column of $X$ and $Y$ to 0-mean, 1-std.
\State $X_{train},Y_{train},X_{test},Y_{test}\gets DataSpliting(X,Y,S)$
\State $H,M=\Call{LearnProjections}{X_{train},Y_{train}}$
\State $Z^{X}_{test}=H^{T}X_{test}, Z^{Y}_{test}=M^{T}Y_{test}$
\State $Similarities=EuclideanDistance(Z^X_{test},Z^X_{test})$
\State $L\gets argmax(Similarities)$
\State \textbf{return} L
\Statex
\Function{LearnProjections}{$X,Y$}
\State $C_{XX}=\frac{1}{n}XX^T, C_{XY}=\frac{1}{n}XY^T$
\State $C_{YX}=\frac{1}{n}YX^T, C_{YY}=\frac{1}{n}YY^T$
\State $\hat{C}_{XX}, \hat{C}_{YY} \gets$ Eq. $\ref{CCAreg}$
\State $H,M \gets$ Eq. $\ref{CCAsovle}$
\State \textbf{return} $H, M$
\EndFunction
\end{algorithmic}
\end{small}
\end{spacing}
\end{algorithm}
\section{Experiments}
\label{sec:experiment}
This section discusses the data sets, experimental settings, and the experimental analysis of some significant baseline models systematically. We also perform the sensitivity study on three primary parameters.
\subsection{Data sets}
To evaluate the performance of the model MAUIL, the survey data sets were collected from the Internet, including two social networks and two academic coauthor networks.
\textbf{Social networks}: The data set, denoted as Weibo-Douban (WD), refers to two Chinese popular social platforms: Sina Weibo\footnote{https://weibo.com} and Douban\footnote{https://www.douban.com}. We provide an accessible method to build cross-network data set WD according to the following steps. First, a small number of users on the Douban platform have posted their Weibo accounts on their homepages. These users on distinguish two platforms are real user identity links, which can be used as pre-aligned user pairs in the UIL task. Second, the original data is prepared by crawling users' information pages, including users' attributes and their follower/followee relations. A clear benefit of data crawling in the Weibo platform could not be directly identified in this step. Weibo allows only a small part (two hundred) of follower/followee relations to be returned by crawlers. Hence, the relations that come from traditional Weibo crawling methods are quite incomplete. On the other hand, the size of Weibo network is enormous. The empirical treatment is to extract a subnet with common characteristics from the original Weibo network. We repeatedly remove the nodes with the degrees less than 2 or more than 1000. Then, the community discovery Algorithm reported in \cite{280} is performed to find the subnets with the typical social network characteristics, including the approximate power-law degree distribution (Figures \ref{Fig:degree}a and \ref{Fig:degree}b) and the high aggregation coefficient. Similar operations are carried out on the Douban network.
\textbf{Coauthor networks}: DBLP is a classic computer science bibliography network. Its data is publicly available\footnote{https://dblp.org/xml/release/}. Each author in DBLP has a unique key, which can be used as the ground truth for the UIL problem. In this study, the DBLP network backups of different periods, i.e.,2017-12-1 and 2018-12-1, were used as the aligned subjects in the UIL experiments. Following previous work \cite {16}, we select the Turing Award winner \textit{Yoshua Bengio} as the center node in each network, and then delete any nodes more than three steps away from the center.
Besides, the size of two DBLPs is reduced by discovering network communities and repeatedly deleting leaf nodes. The final DBLPs also enjoy the characteristics of the power-law distribution (Figure \ref{Fig:degree}c,\ref{Fig:degree}d) and the high aggregation coefficient. The statistics of WD and DBLP are displayed in Table \ref{tab:datasets}.
\begin{table*}[h!]
\setlength{\tabcolsep}{1pt}
\renewcommand{\arraystretch}{1.7}
\caption{The statistics of the data sets used in the experiments.}\label{tab:datasets}
\begin{footnotesize}
\begin{center}
\begin{tabularx}{0.7\textwidth}{p{1.5cm}<{\centering}p{1.4cm}<{\centering}p{1.2cm}<{\centering}p{1.6cm}<{\centering}
p{1.0cm}<{\centering}p{1.0cm}<{\centering}p{1.0cm}<{\centering}p{1.0cm}<{\centering}p{1.8cm}<{\centering}}
\toprule[1.0pt]
\textbf{Datasets}&\textbf{Networks}&\textbf{\#Users}&\textbf{\#Relations}&\textbf{Min. degree}&
\textbf{Ave. degree}&\textbf{Max. degree}&\textbf{Ave. coeff.}&\textbf{\#Matched pairs}\\
\hline
Social &Weibo &9,714 &117,218 &2 &12.1 &607 &0.112 &\multirow{2}{*}{1,397}\\
\cline{2-8}
networks &Douban &9,526 &120,245 &2 &12.6 &608 &0.101 & \\
\hline
coauthor &DBLP17 &9,086 &51,700 &2 &5.7 &144 &0.280 &\multirow{2}{*}{2,832}\\
\cline{2-8}
networks &DBLP19 &9,325 &47,775 &2 &5.1 &138 &0.322 & \\
\bottomrule[1.0pt]
\end{tabularx}
\end{center}
\end{footnotesize}
\setlength{\tabcolsep}{3pt}
\renewcommand{\arraystretch}{1.2}
\end{table*}
\begin{figure}[!h]
\begin{center}
\includegraphics[width=3.2 in]{degree.pdf}
\caption{The degree distributions of the networks in the data sets.}\label{Fig:degree}
\end{center}
\end{figure}
\subsection{Experimental Settings}
\subsubsection{Baseline Methods}
We select the following baseline methods to assess the comparative performance of the proposed model:
\begin{itemize}
\item \textbf{IONE}\cite {55}: IONE is a semi-supervised UIL model depending on network structures, in which the followership /followee-ship of users are modeled as input/output context vector. This model sets the weight unification among users to 1.
\item \textbf{PALE-LINE}\cite {61}: PALE-LINE is a supervised UIL model based on PALE \cite {61} and LINE \cite{312}. A typical operation is that the matched networks are embedded by the LINE approach. PALE-LINE provides an Multi-Layer Perceptron (MLP) that can project the source embeddings to the target space for The UIL task.
\item \textbf{ABNE}\cite {1}: ABNE is a supervised UIL model using graph attentions, which exploits the social structure by modeling the weighted contribution probabilities between follower-ships and followee-ships.
\item \textbf{REGAL}\cite {18}: REGAL is an embedding based unsupervised UIL model that employs both structures and attributes to capture the node similarities. The experiment of REGAL in our comparison adopts TF-IDF vectorization and cosine distance measure to deal with the node attributes and node similarity, respectively.
\item \textbf{TADW-MLP}: TADW-MLP is a composite extension of model TADW \cite{313} and MLP. TADW is a popular network embedding model, which incorporates text attributes under the framework of matrix factorization. We apply TADW to embed networks and employ a three-layer MLP that maps source embeddings to the target space for the UIL problem. Besides, the TF-IDF is also adopted to encode the node attributes.
\end{itemize}
\textbf{Variants of the proposed MAUIL}: This section also examines three MAUIL variants to evaluate the effectiveness of model components, including:
\begin{itemize}
\item \textbf{MAUIL-a}: MAUIL-a ignores the components of network structure embedding of the MAUIL to access the component effect of multi-level attribute embedding. Attributes are treated as unique features to deal with the UIL problem.
\item \textbf{MAUIL-s}: MAUIL-s plans to compare MAUIL-a by retaining only structural features and ignoring attribute ones to clarify the component performance of network structure embedding for solving the UIL problem.
\item \textbf{MAUIL-v}: MAUIL-v directly encodes the feature vectors generated by the MAUIL component of multi-level attribute embedding, which realizes the UIL by comparing Euclidean distance of the vectors. The model avoids the use of correlated common space and runs in an unsupervised manner.
\end{itemize}
\subsubsection{Evaluation Metric}
This study follows the suggestions of the previous work \cite {50} to deliver the evaluation indicators. A typical evaluation metric Hit-Precision is adopt to compare the top-k candidates for the identity linkage. The score is computed as follows:\\
\begin{equation}
h(x)=\frac{k-(hit(x)-1)}{k}
\end{equation}
where $hit(x)$ is the position of a correctly linked user in the returned list of the top-k candidates. Let $n$ be the number of tested user pairs. The Hit-Precision can be calculated by the average score of the successfully matched user pairs: $\frac{1}{n}\sum_{i=1}^{i=n}h(x_i)$. This study adopt $k=3$ for all the experiment instances unless otherwise stated.
\subsubsection{Implementation Details}
This study implements the proposed model by the programming language version Python 3.7 and runs the codes on a Linux Server with Intel(R) Xeon(R) CPU (E5-2620 v4) and GeForce TITAN X GPU (12GB memory).
The experiment involved two types of datasets: the social network WD and the coauthor networks DBLPs. First, we directly takes the network usernames of WD as character-level attributes and convert the Chinese characters into Pinyin characters (Chinese phonetic alphabet). The geographical locations and recent posts of users are considered as the word-level and topic-level attributes, respectively. Second, the authors in DBLPs may have several variants of names, e.g., \textit{Nicholas Drivalos Matragkas, Nicholas Matragkas}, and \textit{Nikolaos Drivalos}. We randomly chose one of the author's names as character-level attributes. In a similar way to WD, the authors' affiliations in DBLPs are selected as word-level attributes. We pick a random number of articles (at most 100) for each author and take their titles as the topic-level attributes. To illustrate the model's robustness, we add some noise by randomly removing some attribute text items with probability $p=0.2$.
Data pre-processing consists of converting all letters of attributes to lowercase, removing any tonal marks above the letters, eliminating the rare words that occur less that 10 times, excluding the stop words like `the' and `with', and extracting text stems by the tool NLTK\footnote{https://www.nltk.org}. We perform the text segmentation of original texts by adding q-gram with $q=2$ and $q=3$.
There are many ways of expression in Chinese, such as simplified, traditional, and archaic. The experiment required all Chinese characters to be converted into simplified ones \footnote{https://pypi.org/project/zhconv/}. It is also demanded to split the texts into separated words by using the tool Jieba \footnote{https://pypi.org/project/jieba/}. The procedure of the word-level embedding includes retaining the neighbor's weight parameter with $\lambda=0.1$. Besides, the word vectors of Chinese texts are trained according to the Chinese wiki corpus\footnote{https://dumps.wikimedia.org/zhwiki/}. Finally, we set both the hyper-parameters $\alpha$ and $\beta$ are $1/d_t$ for the topic-level embedding component of the model MAUIL and its variants.
\subsubsection{Parameter Setup}
This section outlines the model parameters of training MAUIL, including multi-level attribute embedding, RCCA based projection, and training parameter setting.
\begin{enumerate}[(i)]
\item The dimensions of all features in the multi-level attribute embedding and the structural embedding are the same without loss of generality,
i.e., $D=d_c=d_w=d_t=d_s=100$.
\item The regulation parameters $r^X$ and $r^Y$ for covariance matrices $C_{XX}$ and $C_{YY}$ are fixed to be $R=r^X=r^Y$. The number of projection vectors is empirically set to $k=25$ and $k=80$ for WD and DBLPs, respectively. Correspondingly, regulation parameter $R=10^5$ and $R=10^3$ is considered for WD and DBLPs, respectively.
\item Any experiment instances randomly select $N_{tr}$ and $N_{te}$ matched identity pairs as the training seeds and test data. $N_{tr}$ varies between 100 and 800. $N_{te}$ is set to be a constant value 500. The comparative results among baseline models are expressed in Hit-Precision scores with the training setting $N_{tr}=200$.
\end{enumerate}
Each experiment instance is run 10 times independently. Their average performances are considered as the final report. The parameters of the comparison models come from the defaults in their papers.
\subsection{Experimental Results}
\subsubsection{Overall Performance}
All the comparison methods can be treated under three groups according to model component categories. The goal of category testing is to explore the combinatorial correlations between user attributes and network structures in model performances. The first set of analyses examine the UIL by using purely network structures, including IONE, PALE-LINE, ABNE, and MAUIL-s. The second one only leverages user attributes for solving the UIL problem, i.e., MAUIL-a. The third category combines user attributes and network structures to training, including REGAL, TADW-MAUIL, MAUIL-v, and MAUIL. Table \ref{tab:comparison} summarizes the overall results of all compared methods on the two datasets WD and DBLPs.
The MAUIL and its variants are successful because they are able to reveal potential user identity links. The details of performance comparison are described below.
First, the experimental results of Group 1 are presented. It is clear from Table \ref{tab:comparison} that the quality of datasets leads to the differences in model performances.
The methods of IONE, PALE-LINE, ABNE, and MAUIL-s perform better on DBLPs than that on WD. For example, INOE enjoy higher Hit-precision (k=3) by 9.2\%, PALE-LINE by 10.5\%, ABNE by 11.4\%, and MAUIL-s by 16.4\% on DBLPs. The fact is mainly because of the lower data structure integrity exist in real social networks WD that cannot provide valid clues for recognizing user identities. However, the four models' most prominent is the highest score of the proposed variant MAUIL-s that obtains at least 5.6\% higher than the others.
Figures \ref{Fig:Seeds_wd} and \ref{Fig:Seeds_dblp} also shows the effect of the training seeds parameter $N_{tr}$ on the comparative models. The four comparative models improve their performance with the increasing of training seed number $N_{tr}$ from 0 to 800. For instance, IONE's performance grows up from 11.4\% to 50.1\% on DBLP and ABNE from 18\% to 58.1\%. Consequently, more training data did contribute to the precision increase. However, there is also a significant difference between the four models' performance improvement with the increase in training data. Following the performance scores of PALE-LINE and MAUIL-s shown in Figures \ref{Fig:Seeds_wd} and \ref{Fig:Seeds_dblp}, the higher score of MAUIL-s was recorded, although both of them depend on the structure embedding method of LINE. As a result, MAUIL-s performs 4.7\% and 10.6\% better than PALE-LINE on WD and DBLPs, respectively. MAUIL and its variants provide a linear projection that maps the two networks to the common space. The precision advantages of MAUIL-s prove that the established mapping on the MAUIL model series does indeed benefit the UIL task since MAUIL and its variants possess the same model component to deal with network structure embedding.
\begin{table}
\setlength{\tabcolsep}{1pt}
\caption{Comparison with the baseline methods (Hit -Precision score).}\label{tab:comparison}
\begin{footnotesize}
\begin{center}
\renewcommand{\arraystretch}{1.5}
\begin{tabularx}{8.5cm}{p{2cm}<{\centering}|
p{1.0cm}<{\centering}p{1.0cm}<{\centering}p{1.0cm}<{\centering}|
p{1.0cm}<{\centering}p{1.0cm}<{\centering}p{1.0cm}<{\centering}}
\toprule[1.2pt]
\multirow{2}{*}{Method}&\multicolumn{3}{c}{Weibo-Douban} &\multicolumn{3}{c}{DBLP17-DBLP19} \\
\cline{2-7}
& k=1 & k=3 & k=5 & k=1 & k=3 & k=5 \\
\hline
IONE & 0.013 & 0.022& 0.031 & 0.076 & 0.114 & 0.140\\
PALE-LINE & 0.013 & 0.025& 0.036 & 0.085 & 0.130 & 0.165\\
ABNE & 0.049 & 0.066& 0.077 & 0.133 & 0.180 & 0.211\\
REGAL & 0.016 & 0.030& 0.041& 0.411 & 0.436 & 0.452\\
TADW-MLP & 0.004 & 0.007& 0.011 & 0.188 & 0.261 & 0.313\\
\hline
MAUIL-v & 0.238 & 0.280& 0.306 & 0.628 & 0.655 & 0.695\\
MAUIL-s & 0.044 & 0.072& 0.095& 0.159 & 0.236 & 0.289\\
MAUIL-a &\textbf{0.313}&0.368&0.400&\textbf{0.661}& 0.690 & 0.706\\
MAUIL&\textbf{0.313}&\textbf{0.373}&\textbf{0.408}&0.660&\textbf{0.702}&\textbf{0.725}\\
\bottomrule[1.2pt]
\end{tabularx}
\end{center}
\end{footnotesize}
\setlength{\tabcolsep}{3pt}
\renewcommand{\arraystretch}{1.2}
\end{table}
Second, the experimental results of Group 2 involved MAUIL-a are exhibited, which is only concerned with user attributes. MAUIL-a achieves the Hit-Precision scores of 0.368 and 0.69 on WD and DBLPs, respectively. The more adequate and cleaner user attribute information in DBLPs compared with WD creates a significant 0.322 difference in model precision. Actually, the attributes of WD consist of the users' blogs and posts, etc., which parameters are normally inconstant and vary a lot in content, number of posts, and publishing frequency. What stands out in Table \ref{tab:comparison} is that MAUIL-a performs significantly better than the structure-based models by at least 23.7\% on WD and at least 45.5\% on DBLPs. The abundant attribute information in our datasets has an apparent positive influence on the model effect.
Third, Group 3 incorporates both attribute and structure features to investigate the UIL task. Both REGAL and TADW-MLP can extract attribute features. However, they do not distinguish attribute types and the higher semantic features level. A comparison of the scores among those models in Group 3 confirms that the utilization and differentiation of multi-type user attributes are beneficial to improve model performance. For example, MAUIL-v performs considerably better than REGAL by 25\% in dataset WD, which also recorded the more excellent score than TADW-MLP by 27.3\%.
The character-level attributes (user names) of social networks are the strongest of all attribute types, while the word-level locations and topic-level posts are more diverse and noisy. Consequently, character-level attributes dominate model performance.
\begin{figure}[!h]
\begin{center}
\includegraphics[width=3.0 in]{Ntrain_wd.pdf}
\caption{Hit-precision performance on social networks w.r.t the increase of training seeds $N_{tr}$.}\label{Fig:Seeds_wd}
\end{center}
\end{figure}
\subsubsection{The effect of different training seeds}
The performance trend tests were used to analyze the comparison models' Hit-precision with the increase of training seeds parameter $N_{tr}$.
First, the test results of all the comparison models in dataset WD are shown in Figure \ref{Fig:Seeds_wd}. Real social networks are often sparse in structure. Hence, the performance improvement of the simple structure methods, i.e., IONE, PALE-LINE, ABNE, and MAUIL-s, is very limited. Figure \ref{Fig:Seeds_wd} provides an example of the above statement. The test scores of the purely structural methods on WD rise slowly with the increase of $N_{tr}$ from 100 to 800. The comparison of their scores with the proposed MAUIL confirms that combining network structures and user attributes can reach a peak and has an advantage of at least 25.8\% over all structural baselines when $N_{tr}=800$.
\begin{figure}[!h]
\begin{center}
\includegraphics[width=3.0 in]{Ntrain_dblp.pdf}
\caption{Hit-precision performance on coauthor networks w.r.t the increase of training seeds $N_{tr}$.}\label{Fig:Seeds_dblp}
\end{center}
\end{figure}
On top of that, we consider a similar test on dataset DBLPs. The result is shown in Figure \ref{Fig:Seeds_dblp}. The performance of all methods except for REGAL and MAUIL-v present similar upward trend curves.
ABNE and IONE increase faster than MAUIL-s, TADW-MAL, and PALE-LINE. However, MAUIL achieves the best when $N_{tr}$ reaches 800. For example, MAUIL beats ABNE and IONE by 19.7\% and 27.7\%.
Finally, what can be clearly seen in Figures \ref{Fig:Seeds_wd} and \ref{Fig:Seeds_dblp} are the red straight lines referring to Hit-precision of MAUIL-v with the seed increase on both datasets. The unsupervised working mode of MAUIL-v is responsible for this unexpected stable performance.
Under the semi-supervised setting, MAUIL can effectively combine supervised annotations with unsupervised information, e.g., structures and attributes, to ensure its superiority. Therefore, MAUIL is expected to be well adapted to annotate constrained scenarios.\\
\begin{figure}[!h]
\begin{center}
\includegraphics[width=3.2 in]{components.pdf}
\caption{Hit-precision performance on two data sets w.r.t variants of MAUIL.}\label{Fig:components}
\end{center}
\end{figure}
\subsubsection{The contribution of different components of MAUIL.}
As seen from Figures \ref{Fig:Seeds_wd} and \ref{Fig:Seeds_dblp}, MAUIL always outperforms its three variants, i.e., MAUIL-s, MAUIL-a, and MAUIL-v, across all training seeds $N_{tr}$. This section takes $N_{tr}=800$ as an example to access the contributions of each MAUIL component.
The columnar analysis shown in Figure \ref{Fig:components} illustrates that MAUIL wins 24\% higher scores than MAUIL-v on WD and 12.4\% higher on DBLP. The priorities demonstrate the power of MAUIL when the model adds a step of projecting the linked networks into the common correlated space.
On the other hand, MAUIL performs better than MAUIL-s by 39.7\% on WD and 42.7\% on DBLP. The advantages are cased by incorporating attributes information. Besides, MAUIL also beats MAUIL-a by 3.4\% on WD and 3.8\% on DBLP. Utilizing network structural information indeed benefits the UIL tasks. However, the role of network structures is not as strong as that of attribute information in our datasets. Thus, MAUIL is quite encouraging when attribute information is sufficient.
\subsubsection{Parametric Sensitivity Analysis}
This section investigates the parameter sensitivity of the proposed MAUIL on three primary parameters: (1) the feature dimensions $D$, (2) the number of projection vectors $K$, and (3) the regularization coefficient $R$.
\begin{figure}[!h]
\begin{center}
\includegraphics[width=3.2 in]{dim.pdf}
\caption{Hit-precision performance w.r.t the increase of the feature dimension $D$.}\label{Fig:dim}
\end{center}
\end{figure}
Figure \ref{Fig:dim} portrays the Hit-precision variation concerning the value change of feature dimension parameter $D$. We can observe that the MAUIL performance scores increase at the beginning and then keep stable as the dimension rises on both datasets. The trend means that a more top dimensional feature space helps to preserve the user information better, which further contributes the performance improvement of the UIL models.
In the actual operation process, too high dimension will increase the consumption of computing resources. Hence, we select $D=100$ to a trade-off between effectiveness and efficiency.
\begin{figure}[!h]
\begin{center}
\includegraphics[width=3.2 in]{K.pdf}
\caption{Hit-precision performance w.r.t the increase pf projection vectors $k$.}\label{Fig:K}
\end{center}
\end{figure}
Figure \ref{Fig:K} depicts the sensitivity of model performances when the value of the projection vector $k$ changes. MAUIL presents a similar trend in both of the datasets. Its performance can get optimal when $k$ fluctuates from 25 to 35 on WD and from 60 to 110 on DBLP. Therefore, if $k$ changes within a reasonable range, MAUIL will remain stable.
\begin{figure}[!h]
\begin{center}
\includegraphics[width=3.2 in]{reg.pdf}
\caption{Hit-precision performance w.r.t the increase of regulation parameter $R$.}\label{Fig:reg}
\end{center}
\end{figure}
Figures \ref{Fig:reg} illustrates the performance with different setting of regulation parameter $R$. We set the range from $0$ to $10^6$ for $R$ across the two datasets. The Hit-Precision increases gradually and keeps steady after $R>10^4$ on WD. However, the Hit-Precision slightly drops after $R$ reaches $10^3$ on DBLP. One possible reason is that the parameter's variation is exponential, and the stable interval of $R$ is slightly smaller than that of the exponential change. Hence, it is required to choose $R$ in a small range around the locus $10^3$.
\section{Conclusion}
\label{sec:conclusion}
The purpose of the present study was to investigate the problem of user identity linkages between multiple social networks. A high-precision method for user identity discovery can significantly improve the reliability of AI applications such as recommender systems, search engines, information diffusion predictions, cyber identity, and criminal stalking, etc. Hence, this work has constructed a novel semi-supervised model, namely MAUIL, to seek the potential user identity between two arbitrary social networks. Semi-supervision insight greatly reduces the model's dependence on data annotation. Compared with other models, MAUIL synthesizes three kinds of user attributes, i.e., character-level, word-level, and topic-level attributes, and network structure features to obtain the best Hit-precision scores for the UIL problem. The experimental analysis shows that the combination of different text features in social networks can significantly improve the resolution accuracy of the problem. On the other hand, the RCCA based linear projections designed in MAUIL revealed that maximizing the correlations between available feature vectors from different views benefits to find the same natural persons across multiple social networks. The method of feature correlation space is likely to be a potential means to break the link accuracy points required for AI applications.
This study extensively evaluates the proposed model on two real-world data sets of social networks and coauthor networks. The results show the superior performance of MAUIL by comprehensive comparison with state-of-the-art baseline methods. A limitation of this study is that the data set WD in this paper is merely a subset of the current social networks with the similar structure distribution. We don't have the privacy rights to get the full data. However, the more data in the deep learning domain, the higher the score. Finally, this paper exposes the complete program code and data sets as a complement. We recommend the suggested UIL problem-solving framework and technical update directions for other researchers.
\input{mybibfile.bbl}
\end{document}
|
\section{Introduction}
\label{sec:introduction}
\IEEEPARstart{M}{illimeter-wave} (mmWave) communication technology presents great opportunities in a wide range of data-intensive applications.
Owing to the wider available bandwidth in the mmWave band, this greatly supports the 5G new radio (NR) access technology to compensate for the scarcity of spectrum in the microwave bands, thereby meeting the demand created by the increasing amount of mobile traffic\cite{niu2015survey, xiao2017millimeter}.
Apart from the 5G NR, the recent IEEE 802.11ay standard built upon the ratified IEEE 802.11ad standard proceeds beyond multi-gigabit-per-second connectivities to provide a data rate of 100\,Gbit/s\cite{chen2019millimeter, aldubaikhy2020mmwave,ghasempour2017ieee}.
This not only enables indoor high-speed Wi-Fi but also makes the concept of fiber-like connectivities such as outdoor mmWave wireless backhaul\cite{niu2015survey, rangan2014millimeter, ge20145g} or wireless-to-the-home systems\cite{aldubaikhy2020mmwave} a reality.
However, attempts to deploy mmWave nodes densely in outdoor environments could be thwarted by the limited available space for placing the nodes because of the line-of-sight (LOS) requirement.
In mmWave communications, a link budget is severely penalized by obstacles such as buildings and roadside trees\cite{rappaport2013millimeter, rappaport2015wideband}.
To prevent blockages by obstacles such as these and to reach the full potential of mmWave communications, it is important to position mmWave nodes such that obstacles do not obstruct the LOS between nodes that need to communicate.
Typically, owing to this LOS requirement, mmWave nodes would have to be placed at higher altitudes to prevent obstacles from obstructing LOS paths.
This limits the physical spaces for placing mmWave nodes to surfaces of buildings or telephone poles, and accordingly, the density of the mmWave nodes may depend on the density of suitably located buildings and telephone poles.
\begin{figure}[tb]
\centering
\includegraphics[width=0.9\columnwidth]{system_model_ver3}
\caption{
Scenario showing the on-wire deployment of mmWave nodes in a wireless-to-the-building use case.
The beam should be tracked against various perturbations such as wind perturbations and other disturbances caused by impulsive forces to the wire.
We investigate whether non-RF sensing information (e.g., position and velocity of several points on the wire) is useful or not to enhance the accuracy of the beam tracking against such perturbations specific to on-wire deployment.
}
\label{fig:system_model}
\end{figure}
To address this shortcoming, we explore the new possibility of installing mmWave nodes on overhead messenger wires as shown in Fig.~\ref{fig:system_model}.
The placement of mmWave nodes not only on buildings or on telephone poles but also on, for example, splice closures \textit{between telephone poles} offers flexibility with respect to the physical placement of nodes, thereby facilitating ensuring LOS connections.
In the course of this exploration, the key concern that remained is that on-wire deployment introduces dynamics in mmWave nodes owing to unprecedented root causes such as wind perturbations and other disturbances caused by impulsive forces to the messenger wire, which cause mmWave beam misalignment.
Beam misalignment is problematic because mmWave communications generally require directional antennas to compensate for the large path-loss in mmWave communications and to align the receiver and transmitter beams \cite{li2009spectrum}.
A naive approach to address beam misalignment is to let the transmitter and receiver scan their beam directions to determine the direction of maximum power periodically.
In a similar approach, in the IEEE 802.11ad standard, transmitter and receiver steering is decoupled using an omnidirectional antenna \cite{11ad, zhou2012efficient}.
Following the aforementioned on-wire dynamic conditions, search-based methods would require frequent beam-search to repair beam misalignment.
Consequently, the increases the overhead generated by the beam search, which deteriorates the spectral efficiency.
To combat the inefficiency resulting from the periodical beam-search, non-radio frequency (non-RF) information such as the positions and velocities of the mmWave nodes can be leveraged\cite{wang2018mmwave, wang2019mmwave, va2019online, arvinte2019beam, abdelreheem2016millimeter}.
However, acquisition of the positions and velocities is not necessarily synchronous to the beam-tracking operation, and, as a result, information about the current position of a mmWave node is not available.
In this case, we would have to predictively determine an appropriate beam steering angle while referring to the past positions of mmWave nodes.
In the case of the on-wire deployment of a mmWave node, our key question is:
\textbf{In view of the complicated perturbations in the on-wire deployment, would past information about the position and velocity of a mmWave node be useful to feasibly track directional beams?}
This work attempts to answer this question by proposing a feasible system for tracking directional beams based on historical values of the position and velocity under an on-wire mmWave node deployment.
Specifically, for the examination of the feasibility, we leverage deep reinforcement learning (DRL) to determine beam steering angles based on past position/velocity information of a mmWave node.
The reason for the choice of DRL is discussed in Sections~\ref{subsec:contribution} and \ref{subsec:problem_formulation} in this paper.
It should be noted that many studies have applied DRL to mmWave communication\cite{khan2019reinforcement, feng2019dealing, mismar2019deep, wang2020precodernet}, among which one of the studies \cite{mismar2019deep} already proposed DRL-based mmWave beam tracking, yet considered a simpler scenario of on-ground mobile environments.
However, this emergence of DRL is precisely why it is important to proceed beyond simple on-ground mobile environments and to explore the feasibility of targeting more complicated dynamic environments such as those in which mmWave nodes are placed on messenger wires.
Importantly, another benefit of examining the aforementioned feasibility is that it is expected to shed light on mmWave node deployments in other unstable surroundings.
These surroundings are exemplified by lampposts and traffic lights.
These locations are subject to perturbations similar to those of messenger wires such as swaying due to wind and rapid displacements due to forces applied to the poles caused by airborne objects or ground vibrations from road traffics.
The deployment of mmWave nodes in such places greatly supports not only the wireless-to-the-building use cases in Fig.~\ref{fig:system_model}, but also, though not limited to, upcoming mmWave vehicle-to-everything (V2X)\cite{choi2016millimeter, sakaguchi2020towards}.
This is because such placements allow vehicles to communicate with mmWave nodes regarded as roadside units in the proximity of them, enabling to exchange large volume sensory data crucial to support autonomous driving (e.g., camera images, light detection and ranging data, or other radar data)\cite{sakaguchi2020towards}.
Hence, via examining the beam tracking in an on-wire mmWave node placement, this work attempts to open up the opportunities to deploy mmWave nodes in various unstable placements towards diverse mmWave use cases.
\subsection{Related Work}
MmWave communications necessitate the use of directional antennas to compensate for the large path loss in the mmWave spectrum.
Accordingly, considering the dynamics of mmWave nodes, beam tracking is required to maintain the received power; consequently, the design of efficient beam tracking methods has attracted considerable research interest.
This section provides an overview of previous work and discusses the differences among the proposed solutions.
\vspace{.5em}\noindent \textbf{Search-based Beam Tracking.}\quad
To perform beam tracking in mobile environments, a typical approach entails periodically searching for an appropriate beam pair between transmitters and receivers.
The most naive approach is to exhaustively scan all possible beam pairs during the beam search.
In the IEEE 802.11ad standard, transmitter and receiver steering is decoupled using an omnidirectional antenna\cite{11ad}, which is more efficient than exhaustive beam search.
More efficient methods capable of further reducing the search spaces are exemplified by subspace sampling in hierarchial codebooks\cite{hur2013millimeter, li2012efficient, tsang2011coding}, multi-stage search \cite{wang2009beam}, agile-link leveraging multi-armed beams\cite{hassanieh2018fast}, parallel-adaptive beam training\cite{de2017millimeter}, the use of location information\cite{abdelreheem2016millimeter}, and hybrid beam forming with compressing sensing\cite{alkhateeb2014channel}.
In general, this type of approach requires channel observation of the scanned beam pairs, which generates overhead proportionate to the number of beam pairs in terms of spectrum utilization.
Unlike these solutions, our approach does not require beam scanning by learning appropriate beam angles from past experience and applying the beam-tracking policy afterward.
\vspace{.5em}\noindent \textbf{Learning-based Beam Tracking.}\quad
To avoid overhead in terms of spectrum utilization, several studies proposed learning-based beam tracking.
Beam scanning procedures can be eliminated by learning an appropriate beam-tracking policy from past experiences and applying the learned policy afterward.
In this approach, the design of input information for learned machine learning (ML) models to be associated with beam steering angles is divided into two categories: RF information or non-RF information.
\begin{enumerate}
\item \textbf{RF Information.}\quad
Many studies devoted to developing ML models to determine the appropriate beam steering angles leveraged RF information such as the channel state information or signal-to-noise ratio as the input\cite{alkhateeb2018deep, zhou2018deep, araujo2019beam, chang2019learning, jeong2020online}.
In \cite{alkhateeb2018deep}, beam tracking based on supervised learning was proposed to learn achievable data rates corresponding to possible beam pairs while accepting input information of pilot signals in coordinated BSs.
Once the relationship between the data rates of possible beam pairs and pilot signals is learned, appropriate beam pairs can be determined without scanning possible beams.
In \cite{zhou2018deep}, multiple transmitter/receiver pairs were considered, wherein a beam-tracking policy that avoids interference was learned.
An RL-based approach was also proposed\cite{araujo2019beam, jeong2020online}, wherein a channel observation and current beam pair are regarded as a state, and the beam steering angle is regarded as an action.
In \cite{jeong2020online}, the beam steering angle and beam width were jointly optimized via DRL.
However, this approach still requires channel observations to obtain RF information to determine the steering angle, which may incur overhead particularly in a mobile environment in which the beam frequently needs to be fixed.
Unlike these studies, we leverage non-RF information as input to our ML models.
\item \vspace{.3em}\textbf{Non-RF Information.}\quad
Leveraging non-RF information as the input enables the elimination of channel observation to determine the beam steering angles.
Supervised learning-based beam tracking that learns the relationship between user locations and appropriate beam steering angles was proposed \cite{arvinte2019beam}.
In another study, DRL was proposed by regarding the positions of mobile users as a state, which is the input of the ML model \cite{mismar2019deep}.
The positions of cars were utilized as state information in RL-based beam tracking \cite{klautau20185g}.
In addition to the tracked vehicles, the positions of surrounding vehicles were also utilized to determine an appropriate path \cite{wang2018mmwave, wang2019mmwave, va2019online}.
In recent work \cite{alrabeiah2020millimeter}, visual information was leveraged to monitor vehicles.
However, this assumes simpler two-dimensional on-ground dynamics of tracked mobile devices and the availability of the current positions of the devices.
Contrary to these studies, we investigate the feasibility of learning under more complicated three-dimensional on-wire dynamics in a mmWave node while relaxing the assumption that synchronous non-RF information is acquired.
\end{enumerate}
To summarize, the main difference between this work and reports in the literature is \textit{1)} the consideration of dynamics specific to on-wire deployments and \textit{2)} using only historical information about the position and velocity of a mmWave node.
In on-wire deployment, the dynamics of a mmWave node essentially differ from those of on-ground mobile environments in the sense that the dynamics include oscillations due to the tensile force of wires and rapid position validations as a result of impulsive forces to wires.
Hence, the literature does not provide straightforward answers as to whether an appropriate beam-tracking policy can be feasibly learned based on past information about the position and velocity.
\subsection{Contributions and Organization of This Paper}
\label{subsec:contribution}
The salient contributions of this paper are summarized as follows:
\begin{itemize}
\item
We demonstrate the feasibility of tracking directional beams of a mmWave node placed on an overhead messenger wire and that is subject to complicated on-wire dynamics, which is an unexplored scenario in the wireless communications research area.
Specifically, motivated by the aforementioned question, we examine whether DRL could be leveraged to predictively determine beam steering angles based on past information about the position and velocity of the mmWave node.
The choice of DRL is because of its powerful capability to learn complicated relationships between state information (i.e., historical information about the position and velocity of the mmWave node in our case) and the appropriate actions (i.e., beam steering angles) by using neural networks (NNs).
Although DRL has previously been used for beam tracking \cite{mismar2019deep}, to the best of our knowledge, the feasibility of tracking beams under on-wire dynamics based on past information about the position and velocity has not yet been investigated.
\item
Examining the feasibility of the beam tracking in view of impulsive forces to a messenger wire, we find it is useful to take advantage of the positional interaction among several points on the wire, which are specific characteristics of on-wire deployments.
More specifically, as state information in DRL, we leverage the information pertaining to the position and velocity of several points on the wire and demonstrate that it is possible to avoid beam misalignment that could not be avoided based only on the position/velocity of the mmWave node.
This is because the variation in the position of points at which the impulsive force applied propagates to the mmWave node via the tensile force.
Hence, the positions/velocities of several points on the wire not only provide information about the position of the mmWave node, they also indicate the way in which the position of the mmWave node could be expected to vary in the future, leading to more accurate beam tracking.
\end{itemize}
It should be noted that the main objective of this work is to answer the aforementioned question, which is again stated here: Is historical information about the position and velocity of mmWave nodes useful to feasibly track directional beams under perturbations specific to on-wire deployments?
To focus on this, we assume that past information about the position and velocity of a mmWave node is available beforehand, and we consider an in-depth discussion of the acquisition of such information to be beyond the scope of this paper.
This gives us the aforementioned insight, which is sufficiently helpful to explore the possibility of on-wire mmWave node placements.
In addition, compared with the preliminary version of this work presented at IEEE VTC-fall 2020\cite{shinzaki2020deep}, this work addresses the aforementioned question more comprehensively by investigating various characteristics of messenger wires.
Moreover, this work considers not only wind perturbations, but also impulsive forces to wires, whereas the preliminary version in \cite{shinzaki2020deep} considered only wind perturbations.
Accordingly, this work examines the usage of multiple position/velocity sensors to combat against the disturbances caused by impulsive forces, and this insight was not provided in the previous work.
The remainder of this paper is organized as follows.
In Section \ref{sec:system_model_problem_formulation}, we present the system model and problem formulation.
In Section \ref{sec:DeepRL}, we provide the DRL-based beam-tracking framework to solve the formulated problem.
In Section \ref{sec:simulation}, we describe the numerical evaluation of the DRL-based beam-tracking framework by reproducing the on-wire dynamics subject to the aforementioned perturbations.
Finally, we provide our concluding remarks in Section \ref{sec:conclusion}.
\section{System Model and Problem Formulation}
\label{sec:system_model_problem_formulation}
\subsection{System Model}
\label{subsec:system_model}
\noindent\textbf{On-wire mmWave Node Deployment.}\quad
We consider an on-wire mmWave node deployment wherein a single mmWave node serves another node mounted on a vertical building surface as shown in Fig.~\ref{fig:system_model}.
The former mmWave node is attached to the messenger wire, with its placement being motivated by the necessity to ensure a LOS connection.
The on-wire mmWave node relays data from the node attached to the building, forwarding indoor data traffic into core networks via the messenger wire, and vice-versa.
The position of the on-wire node is perturbed by wind and impulsive forces to the messenger wire due to airborne objects exemplified by birds and unmanned aerial vehicles.
Against these perturbations, the mmWave directional beam is periodically steered by leveraging non-RF information exchanged in the messenger wire.
To investigate the way in which non-RF information aids beam tracking, we consider that, as an example, the positions and velocities at certain points on the messenger wire are available with multiple sensors.
These sensors obtain the positions and velocities in an absolute coordinates in an error-free manner with constant intervals.
This assumption is motivated by our focus on the perturbations specific to the on-wire, and hence, we eliminate other root causes of beam-misalignments such as sensory errors.
Meanwhile, motivated by the question in Section~\ref{sec:introduction}, the acquisition of the sensor information does not need to be synchronous to the beam-tracking operations because of, for example, inherent sensor delay or the delay for transmitting the sensory information to beam-tracking agents.
To ensure the generality, we do not specify the root cause of the time-difference between the sensory acquisition and the beam-tracking operations by just denoting the difference as $T$, which is termed look-back time and is detailed in Section~\ref{subsec:problem_formulation}.
\vspace{.5em}\noindent \textbf{Dynamics of Overhead Messenger Wire.}\quad
Points on messenger wire interact with adjacent ones via tensile force.
To imitate such positional interaction, we consider the messenger wire as a chain of points wherein adjacent points are interconnected with a spring\cite{maor1975discrete}.
Specifically, the messenger wire is considered as $N$ material points (see. Fig.~\ref{fig:dynamics_wire} in the next section) with the mass of $m/N$, where the adjacent points are physically connected by springs with spring constant $k_0$.
Note that $m$ denotes the mass of the entire messenger wire.
The considered $N$ points encompass two endpoints of the messenger wire, and we term the endpoints $\mathrm{P}1$ and $\mathrm{P}N$.
The other points are termed $\mathrm{P2}, \dots, \mathrm{P}N - 1$ in the order of proximity from $\mathrm{P}1$.
To obtain the position of $\mathrm{P}1, \dots, \mathrm{P}N$ affected by the aforementioned interaction with proximity points, wind perturbations, and impulsive forces, we consider the following dynamics model.
For $i = 1, 2, \dots, N$, let $\bm{a}_i(t), \bm v_{i}(t), \bm x_i(t)\in\mathbb{R}^3$ denote the acceleration, velocity, and position of $\mathrm{P}i$ in a three-dimensional Euclidean space at time instant $t$, respectively.
Without loss of generality, we assume that the origin $o$ is located at the mid-point of the messenger wire under the stable status without wind perturbations and impulsive forces.
The components of the acceleration $\bm{a}_i(t)$ are denoted by $a_{i, \mathrm{X}}(t), a_{i, \mathrm{Y}}(t), a_{i, \mathrm{Z}}(t)\in\mathbb{R}$, where the superscript $\mathrm{X}$ and $\mathrm{Y}$ denote the horizontal component, and $\mathrm{Z}$ denotes the vertical component.
This notation is also applied to $\bm v_{i}(t), \bm x_i(t)$.
The endpoints $\mathrm{P}1$ and $\mathrm{P}N$ are fixed, and hence $\bm a_1(t) = \bm a_N(t) = \bm v_1(t) = \bm v_N(t) = \bm 0$.
For $i=2,3,\dots,N-1$, $\bm a_i(t)$ modeled as:
\begin{multline}
\label{eq:dynamics_acceleration}
\bm a_i(t) = \bm g + \underbrace{\frac{k_0N}{m}\bigl[\bm x_{i+1}(t) + \bm x_{i-1}(t) - 2\bm x_i(t)\bigr]}_{\text{From tensile force yielding positional interaction}}
\\ + \underbrace{\frac{N}{m} \bm{F}_i(t)}_{\substack{\text{From impulsive force}}},
\end{multline}
where $\bm g\in\mathbb R^3$ denotes the gravitational acceleration.
As an example of wind dynamics, we consider an Ornstein-Uhlenbeck process with an average wind velocity $\bm v_0(t)\in\mathbb{R}^3$\cite{shiri2019massive}.
Given the above acceleration, the derivatives of velocity and position $\mathrm d\bm v_i(t)$ and $\mathrm d\bm r_i(t)$ for $i=2,3,\dots,N-1$ are given by:
\begin{align}
\mathrm d\bm v_i(t) & = \bm a_i(t)\,\mathrm dt \underbrace{-c_0\bigl[\bm v_i(t)-\bm v_\mathrm o(t)\bigr]\,\mathrm dt + \bm V_\mathrm o \mathrm d\bm W_i(t)}_{\text{From wind perturbation}}, \label{eq:dynamics_velocity} \\
\mathrm d\bm x_i(t) & = \bm v_i(t)\,\mathrm dt, \label{eq:dynamics_position}
\end{align}
where $c_0$, $\bm V_\mathrm o\in\mathbb R^{3\times 3}$, $\bm W_i(t)\in\mathbb R^3$ denote the drag constant, the covariance matrix of the wind speed, and the standard Wiener process, respectively.
\vspace{.5em}\noindent\textbf{Radio Wave Propagation and Directional Antenna Model.}\quad
Owing to the placement to ensure LOS connectivity between on-wire and on-building nodes, we assume that the the direct waves are dominant.
Accordingly, we only consider the path loss in the channel model.
Given this, the received power denoted as $P_{\mathrm{RX}}$ is given by:
\begin{align}
P_{\mathrm{RX}} = P_\mathrm{TX} G_\mathrm{TX}G_\mathrm{RX}\beta r^{-\alpha},
\label{eq:channel}
\end{align}
where $P_\mathrm{TX}$, $\beta$, $\alpha$ $r$ denote the transmit power, path loss at the unit distance, path loss exponent, and distance between the on-wire and on-building nodes, respectively.
For subscript $j\in\{\mathrm{TX}, \mathrm{RX}\}$, $G_j$ denotes the transmitter/receiver antenna gain, which is discussed next in detail.
Because we consider the three-dimensional dynamics in the on-wire mmWave node, we model the antenna radiation pattern following the 3GPP antenna radiation model with three-dimensional uniform planar antennas\cite{3gpp}.
Following the model, the antenna radiation pattern $A_j(\theta,\phi, \theta_{\mathrm{s}, j}, \phi_{\mathrm{s}, j})$ is given by
\begin{align}
\label{eq:antenna_gain}
A_j(\theta, \phi, \theta_{\mathrm{s}, j}, \phi_{\mathrm{s}, j}) = A_\mathrm{E}(\theta, \phi) + \mathit{AF}(\theta, \phi, \theta_{\mathrm{s}, j}, \phi_{\mathrm{s}, j}),
\end{align}
where $(\theta, \phi)$ are the zenith and azimuth angles relative to reference angles $(\theta_{\mathrm{s}, j}, \phi_{\mathrm{s}, j})$ exhibiting the maximum directivity gain, which is termed main lobe steering direction.
In \eqref{eq:antenna_gain}, $A_\mathrm{E}(\theta, \phi)$ and $\mathit{AF}(\theta, \phi, \theta_{\mathrm{s}, j}, \phi_{\mathrm{s}, j})$ denote the element radiation pattern and the array factor, respectively, which are precisely given in \cite{3gpp}.
It should be noted that $G_\mathrm{TX}$ and $G_{\mathrm{RX}}$ in \eqref{eq:channel} are given by substituting the angles of arrival and angles of departure into $A_{\mathrm{TX}}(\theta, \phi, \theta_{\mathrm{s}, \mathrm{TX}}, \phi_{\mathrm{s}, \mathrm{TX}})$ and $A_{\mathrm{RX}}(\theta, \phi, \theta_{\mathrm{s}, \mathrm{RX}}, \phi_{\mathrm{s}, \mathrm{RX}})$, respectively.
More specifically, $G_\mathrm{TX}$ and $G_{\mathrm{RX}}$ are given by:
\begin{align}
& G_{\mathrm{TX}} = A_\mathrm{TX}(\theta_{\mathrm{AoD}}, \phi_{\mathrm{AoD}}, \theta_{\mathrm{s}, \mathrm{TX}}, \phi_{\mathrm{s}, \mathrm{TX}}), \\
& G_{\mathrm{RX}} = A_\mathrm{RX}(\theta_{\mathrm{AoA}}, \phi_{\mathrm{AoA}}, \theta_{\mathrm{s}, \mathrm{RX}}, \phi_{\mathrm{s} , \mathrm{RX}}),
\end{align}
where $(\theta_{\mathrm{AoD}}, \phi_{\mathrm{AoD}})$ and $(\theta_{\mathrm{AoA}}, \phi_{\mathrm{AoA}})$ denote the angle of departure and angle of arrival relative to the aforementioned reference angles, respectively.
Beam tracking is generally performed by configuring the main lobe steering direction $(\theta_{\mathrm{s}, j}, \phi_{\mathrm{s}, j})$ in the array factor, and hence, we note the mathematical expression of the array factor in detail.
For the sake of simplicity, we omit the subscript $j$ indicating transmitter or receiver.
The array factor for an array of $n=n_{\mathrm{v}}\times n_{\mathrm{h}}$ elements is expressed as
\begin{align}
\mathit{AF}(\theta, \phi, \theta_{\mathrm{s}}, \phi_{\mathrm{s}}) = 10\log_{10}\Bigl[1+\rho\Bigl(\left|\bm a\cdot\bm w^{\mathrm{T}}\right|^{2}-1\Bigr)\Bigr],
\label{eq:array_factor}
\end{align}
where $n_{\mathrm{v}}$, $n_{\mathrm{h}}$, $\rho$ denote the number of vertical elements, the number of horizontal elements, and correlation coefficient, respectively.
In \eqref{eq:array_factor}, $\bm a\in\mathbb C^{n_{\mathrm{v}}n_{\mathrm{h}}}$ and $\bm w\in\mathbb C^{n_{\mathrm{v}}n_{\mathrm{h}}}$ denote the amplitude vector and beamforming vector, respectively, and all elements of $\bm a$ are generally set as $1/n_{\mathrm{v}}n_{\mathrm{h}}$.
Hence, the problem of tracking beams falls into setting the beamforming vector $\bm w$ such that an appropriate main lobe steering direction $(\theta_{\mathrm{s}}, \phi_{\mathrm{s}})$ is configured.
The relationship between $\bm w$ and $(\theta_{\mathrm{s}}, \phi_{\mathrm{s}})$ is given by
\begin{align*}
\bm w & = \left[w_{1,1}, w_{1,2}, \ldots, w_{n_{\mathrm{v}}, n_{\mathrm{h}}}\right]^{\mathrm{T}}, \notag \\
w_{p,r} & = \mathrm e^{\mathrm j2\pi[(p-1)\Delta_{\mathrm{V}}\Psi_p/\lambda + (r-1)\Delta_{\mathrm{H}}\Psi_r]/\lambda}, \notag \\
\Psi_p & = \cos(\theta + \theta_{\mathrm{s}}) - \cos \theta_{\mathrm{s}}, \notag \\
\Psi_r & = \sin(\theta + \theta_{\mathrm{s}})\sin(\phi + \phi_{\mathrm{s}}) - \sin\theta_{\mathrm{s}}\sin\phi_{\mathrm{s}},
\end{align*}
where $\Delta_{\mathrm{V}}$ and $\Delta_{\mathrm{H}}$ denote the spacing distances between the vertical and horizontal elements of the uniform planar array, respectively.
\subsection{Problem Formulation}
\label{subsec:problem_formulation}
Our objective is to obtain a strategy for refining the beam direction against the aforementioned perturbations.
Without loss of generality, we only consider the beam tracking of the on-wire node by employing isotropic antennas in the other mmWave node, i.e., $G_{\mathrm{RX}} = 0$\,dBi over every angle.
We consider the problem of sequentially determining whether the main lobe steering direction is varied with a predefined angle difference $A$.
Considering a finite time duration $t\in [0, T_{\mathrm{d}}]$, where $T_{\mathrm{d}}\in\mathbb{R}$, we let superscript $(k)$ denote the time step, where $k \coloneqq \lfloor t/\tau \rfloor$, and $\tau$ is the period for which beam refinement is performed.
Given this, for a sufficiently large integer $K_{\mathrm{d}} \coloneqq \lfloor T_{\mathrm{d}}/\tau \rfloor$, the objective problem is formulated as follows:
\begin{maxi!}[1]
{\substack{\bigl(a_{\theta}^{(k)}\!, \ a_{\phi}^{(k)}\bigr)}}
{\sum_{k = 0}^{K_{\mathrm{d}}} P^{(k)}_{\mathrm{RX}}}{}{}
\addConstraint{\eqref{eq:dynamics_acceleration},}{\eqref{eq:dynamics_velocity},\eqref{eq:dynamics_position}}{}\label{eq_const:dynamics}
\addConstraint{r^{(k)}}{=\left\Vert\bm x_{i_{\mathrm{TX}}}(k\tau) - \bm x_{\mathrm{RX}}\right\Vert}{}
\addConstraint{\theta^{(k)}_{\mathrm{AoD}}}{=\arccos \frac{ x_{\mathrm{RX}, \mathrm{Z}} - x_{i_{\mathrm{TX}},\mathrm{Z}}(k\tau)}{r^{(k)}}} - \theta^{(k)}_{\mathrm{s}, \mathrm{TX}}{} \label{eq_const:AOItheta}
\addConstraint{\phi^{(k)}_{\mathrm{AoD}}}{=\arctan \frac{x_{\mathrm{RX}, \mathrm{Y}} - x_{i_{\mathrm{TX}}, \mathrm{Y}}(k\tau)}{x_{\mathrm{RX}, \mathrm{X}} - x_{i_{\mathrm{TX}}, \mathrm{X}}(k\tau)}} - \phi^{(k)}_{\mathrm{s}, \mathrm{TX}}{}\label{eq_const:AOIphi}
\addConstraint{\theta^{(k)}_{\mathrm{s}, \mathrm{TX}}}{=\,\theta^{(k - 1)}_{\mathrm{s}, \mathrm{TX}}+\Delta_{\theta}^{(k)}}{}
\addConstraint{\phi^{(k)}_{\mathrm{s}, \mathrm{TX}}}{=\,\phi^{(k - 1)}_{\mathrm{s}, \mathrm{TX}}+\Delta_{\phi}^{(k)}}{}
\addConstraint{\Delta_{\theta}^{(k)}, \Delta_{\phi}^{(k)}}{\in\,\{-A, A, 0\}}{}.\label{eq_const:action}
\end{maxi!}
In the above problem, $i_{\mathrm{TX}}\in\{1, 2, \dots, N\}$ denotes the point at which the on-wire node is attached, and $\bm x_{\mathrm{RX}} = [x_{\mathrm{RX}, \mathrm{X}}, x_{\mathrm{RX}, \mathrm{Y}}, x_{\mathrm{RX}, \mathrm{Z}}]^\mathrm{T}$ is the position at which the fixed node is mounted.
The constraint \eqref{eq_const:dynamics} determines the dynamics of the messenger wire and the attached on-wire node affected by wind perturbations and impulsive forces, which is an essential difference from the problem formulations in the literature regarding beam-tracking methods.
The constraints \eqref{eq_const:AOItheta} and \eqref{eq_const:AOIphi} determine the effect of the dynamics of the messenger wire and main lobe steering angles on the dynamics of the relative angles of departure, which can be obtained by a simple geometric calculation.
The above problem can be easily solved by employing delay-free position information, that is, when determining the main lobe steering direction at each time instant $k\tau$ for $k = 1, 2, \dots$, the positions of the on-wire node at the same time instant, i.e., $\bm x_{i_{\mathrm{TX}}}(k\tau)$, are available.
Indeed, given this, the optimal strategy is to configure the main lobe steering direction such that the main lobe points at the fixed node.
However, as discussed above, we concern that it may not be easy to obtain such perfect position information in terms of time because of the processing delay or the transmission delay of the information or the difference between the interval of beam tracking and that of the position acquisition.
In this case, it is necessary to \textit{look back} at past positions of the on-wire node and determine an appropriate main lobe steering direction while predicting the variation in the on-wire node position within the look-back time.
Let the look-back time be denoted by $T$ hereinafter.
This motivated us to implement deep learning-based beam tracking.
Deep learning successfully builds prediction models in a data-driven manner by leveraging NNs, even when the relationship between inputs and outputs cannot be given in an explicit form owing to its complexity.
Hence, applying deep learning to our problem, we learn an appropriate relationship between past look-back positions of the on-wire node and beam direction even when, within the look-back time $T$, the position of the node is subject to complicated on-wire dynamics.
The approach we employ for deep learning-based beam tracking is to use deep RL to find the optimal beam-tracking policy based on the look-back position information on the on-wire node.
In the next section, we detail the DRL-based beam tracking that is implemented in this work.
\section{Implemented DRL-Based Beam Tracking}
\label{sec:DeepRL}
DRL learns an NN model that relates observable information in a considered environment termed the \textit{state} to an appropriate \textit{action} that maximizes a long-term \textit{reward}.
In this section, we detail the state, action, and reward we used to solve the aforementioned beam-tracking problem.
Subsequently, we discuss the importance of leveraging not only the information about the on-wire node but also information about other points on the messenger wire, focusing on the positional interaction specific to the on-wire deployment.
\subsection{State, Action, and Reward}
\vspace{.3em}\noindent \textbf{State.}\quad
The state is defined as the look-back positions and velocities of the on-wire node and main lobe steering direction.
More specifically, the state at each time instant $s_k$ for $k\in\mathbb{N}$ includes:
\begin{itemize}
\item look-back positions of the on-wire node, i.e., $\bm{x}_{i_{\mathrm{TX}}}(k\tau - T)$.
\item look-back velocities of the on-wire node, i.e., $\bm{v}_{i_{\mathrm{TX}}}(k\tau - T)$.
\item The current main lobe steering direction $\bm b^{(k)}$, which is given by
\begin{multline}
\label{eq:beam_direction}
\bm b^{(k)} \coloneqq
\\
\left[\sin\theta_{\mathrm{s}, \mathrm{TX}}^{(k)}\cos\phi_{\mathrm{s}, \mathrm{TX}}^{(k)}, \sin\theta_{\mathrm{s}, \mathrm{TX}}^{(k)}\sin\phi_{\mathrm{s}, \mathrm{TX}}^{(k)}, \cos\theta_{\mathrm{s}, \mathrm{TX}}^{(k)}\right]^{\mathrm{T}}.
\end{multline}
\end{itemize}
\vspace{.3em}\noindent \textbf{Action.}\quad
The action determines whether the main lobe is steered at a pre-defined angle $A$ or not.
Specifically, each action at time instant $k\tau$ for $k\in\mathbb{N}$ is defined as:
\begin{align}
\mathit{action}_k\coloneqq \Bigl(\Delta_{\theta}^{(k)}, \Delta_{\phi}^{(k)}\Bigr),
\end{align}
where the $\Delta_{\theta}^{(k)}$ and $\Delta_{\phi}^{(k)}$ correspond to the optimized direction of the main lobe in the aforementioned problem and are subject to the constraint \eqref{eq_const:action}.
\vspace{.3em}\noindent\textbf{Reward.}\quad
The immediate reward is defined as the instantaneous received signal power.
The reward at each time instant $r_k$ is defined as
\begin{align}
r_k \coloneqq P_{\mathrm{RX}}^{(k)}.
\end{align}
However, applying the above reward definition may result in an unstable learning performance owing to the large variation of the reward values.
Hence, following a previous study \cite{mnih2015human}, we use a proxy reward that is scaled via clipping.
Specifically, the proxy reward $\hat{r}_k$ is given by
\begin{align}
\label{eq:clipping}
\hat{r}_k =
\begin{cases}
1, & (r_k-b_\mathrm c)/d_\mathrm c >1; \\
(r_k-b_\mathrm c)/d_\mathrm c, & -1\leq(r_k-b_\mathrm c)/d_\mathrm c\leq 1; \\
-1, & (r_k-b_\mathrm c)/d_\mathrm c < -1,
\end{cases}
\end{align}
where $b, d_{\mathrm{c}}\in\mathbb{R}$ denote the offset and scale of the clipping, respectively.
Note that the use of the proxy reward may violate our initial objective to solve the aforementioned problem because of the difference between $r_k$ and $\hat{r}_k$.
However, an appropriate setting of the offset and scale values leads to a feasible beam-tracking policy, which is validated in the numerical evaluation.
\subsection{State Expansion for Leveraging Positional Interaction}
\label{subsec:state_expansion}
\begin{figure}[t]
\centering
\subfigure[$t = 0$\,ms]{\includegraphics[width=.45\columnwidth]{collision_200_ver2.pdf}}
\subfigure[$t = 50$\,ms]{\includegraphics[width=.45\columnwidth]{collision_205_ver2.pdf}}
\subfigure[$t = 150$\,ms]{\includegraphics[width=.45\columnwidth]{collision_215_ver2.pdf}}
\subfigure[$t = 250$\,ms]{\includegraphics[width=.45\columnwidth]{collision_225_ver2.pdf}}
\subfigure[$t = 350$\,ms]{\includegraphics[width=.45\columnwidth]{collision_235_ver2.pdf}}
\subfigure[$t = 450$\,ms]{\includegraphics[width=.45\columnwidth]{collision_245_ver2.pdf}}
\caption{Dynamics of overhead messenger wire after impulsive force with $N = 21$.}
\label{fig:dynamics_wire}
\end{figure}
In this section, we expand the state such that the beam-forming agent leverages not only the positions and velocities of the on-wire node but also those of the points distributed on the messenger wire.
This is motivated by the fact the points on the messenger wire interact with adjacent points via the tensile force.
Therein, the look-back positional information of the other on-wire points is also informative to predict the current positions of the on-wire node, thereby enhancing the quality of the beam steering.
To illustrate such positional interaction on messenger wires, in Fig.~\ref{fig:dynamics_wire}, we plot the 2D positions of the on-wire points after impulsive forces are injected.
This dynamics is consistent with \eqref{eq:dynamics_acceleration}, \eqref{eq:dynamics_velocity}, and \eqref{eq:dynamics_position}.
As an example, we set the parameters as $k = 1000\,\mathrm{Nm^{-1}}$, $m = 10\,\mathrm{kg}$, and $N = 21$.
At $t = 0$, an impulsive force is exerted at $\mathrm{P}4$ with the force of 470\,N, i.e., for $i = 4$, $\bm{F}_i(0) = [0, 0, 470]^{\mathrm{T}}$, and $\bm{F}_i(0) = [0, 0, 0]^{\mathrm{T}}$, otherwise.
As shown in Fig.~\ref{fig:dynamics_wire}, the positions of the points on the messenger wire correlate with the past positions of adjacent points, wherein the points are displaced in accordance with the adjacent ones and form a wave as a whole.
Hence, by looking back not only at the positions/velocities of the on-wire node, but also those of the other points, the beam-forming agent predicts the current positions of the on-wire node more accurately, thereby achieving beam steering policies nearer to the optimal solution of the aforementioned problem.
\vspace{.3em}\noindent\textbf{Expanded State.}\quad
Hence, the expanded state includes the positions and velocities of the points distributed on the messenger wire.
Specifically, the expanded state at each time instant $s_{\mathrm{ex}, k}$ is given as follows.
\begin{itemize}
\item look-back positions \textit{of multiple points on the messenger wire}, i.e., $(\bm{x}_i(k\tau - T))_{i\in\mathcal{N}_{\mathrm{sense}}}$, where $\mathcal{N}_{\mathrm{sense}}\subseteq\{1, 2, \dots, N\}$.
\item look-back velocities of the same points on the messenger wire, i.e., $(\bm{v}_i(k\tau - T))_{i\in\mathcal{N}_{\mathrm{sense}}}$.
\item The current main lobe steering direction $\bm b^{(k)}$, which is given by \eqref{eq:beam_direction}.
\end{itemize}
A schematic overview of the DRL-based beam steering discussed above is presented in Fig.~\ref{fig:RL_overview}.
Note that, because our focus is on the state, action, and reward design to solve the aforementioned problem, we do not detail the specific algorithm to train the NN shown in Fig.~\ref{fig:RL_overview} and apply the algorithm detailed in the previous study \cite{mnih2015human}.
Hence, for the particulars of the algorithm, readers are encouraged to refer to the paper by Mnih \textit{et al.} \cite{mnih2015human}.
\begin{figure}[t]
\centering
\includegraphics[width=\columnwidth]{system_model_vs_delay_ver3.pdf}
\caption{
Implemented DRL-based beam tracking.
At each time instant $k\tau$, the DRL-agent looks back sensory information from the on-wire sensors with look-back time $T$ and determines the beam steering directions while predicting the displacement of the on-wire node within $T$ by leveraging the NN.
}
\label{fig:RL_overview}
\end{figure}
\section{Numerical Evaluation}
\label{sec:simulation}
\subsection{Setup}
\vspace{.3em}\noindent\textbf{Beam-Tracking Environment Setting.}\quad
The parameters for identifying the beam-tracking environment are summarized in Table~\ref{tbl:SimurationParam}.
The endpoints of the messenger wire and the fixed node are placed at the same height of 5\,m.
The messenger wire comprises 21 points, and the on-wire node is installed at the center of the wire, i.e., $i_{\mathrm{TX}} = 10$.
The wind speed $\bm v_\mathrm o(t)$ is given by
\begin{align}
\label{eq:wind_speed}
\bm v_\mathrm o(t) = \left[5\sin\frac{2\pi t}{4}, 5\sin\frac{2\pi t}{6}, 5\sin\frac{2\pi t}{8}\right]^{\mathrm{T}},
\end{align}
which follows the prior version of this work\cite{shinzaki2020deep}.
An impulsive force to the wire is occasionally applied at $\mathrm{P}4$ with the force of 470\,N, i.e., for $i = 4$, $\bm{F}_i(t_{\mathrm{force}}) = [0, 0, 470]^{\mathrm{T}}$, and $\bm{F}_i(t) = [0, 0, 0]^{\mathrm{T}}$, otherwise, where $t_{\mathrm{force}}$ is the time at which the force is applied.
When training the NN, we repeat the simulation of the on-wire dynamics for $t\in[0, 3]$, and within the duration, the impulsive force is applied once at the time determined uniformly and randomly from 1\,s, 2\,s, and 3\,s.
\begin{table}[t]
\centering
\caption{Parameters for Beam-Tracking Environment}
\begin{tabular}{cc}
\toprule
Distance between endpoints $d_w$ & \multirow{2}{*}{$10\,\mathrm{m}$}\\
(i.e., Distance between poles in Fig.~\ref{fig:system_model}) & \\
Distance between wire and building $d_r$ & \multirow{2}{*}{$5\,\mathrm{m}$}\\
(i.e., Road width in Fig.~\ref{fig:system_model}) & \\
Transmit power $P_{\mathrm{TX}}$ & $23\,\mathrm{dBm}$ \\
Wavelength of radio waves $\lambda$ & $5\,\mathrm{mm}$ \\
Receive antenna gain $G_{\mathrm{RX}}$ & $8\,\mathrm{dBi}$ \\
Gravitational acceleration $\bm{g}$ & $[0, 0, -9.8]\,\mathrm{ms^{-2}}$ \\
Drag constant $c_0$ & $1\,\mathrm{s}^{-1}$ \\
Number of points $N$ & 21 \\
Mass of wire $m$ & $10\,\mathrm{kg}$ \\
Covariance matrix of the wind speed $\bm{V}_0$ & $0.1\bm{I}$ \\
Number of vertical elements $n_{\mathrm{v}}$ & 32 \\
Number of horizontal elements $n_{\mathrm{h}}$ & 8 \\
Correlation coefficient $\rho$ & 1 \\
Vertical spacing distance $\Delta_{\mathrm{V}}$ & $2.5\,\mathrm{mm}$ \\
Horizontal spacing distance $\Delta_{\mathrm{H}}$ & $2.5\,\mathrm{mm}$ \\
Point installed on-wire node & $\mathrm{P}10$ \\
Interval between successive time instant $\tau$ & $10\,\mathrm{ms}$ \\
Refinement angles $A$ & $1^\circ$ \\
\bottomrule
\end{tabular}
\label{tbl:SimurationParam}
\end{table}
\vspace{.3em}\noindent\textbf{Beam-Tracking Agent Setting.}\quad
We detail the parameter setting for the beam-tracking agent.
Following the DRL algorithm \cite{mnih2015human}, the beam-tracking agent trains an NN that outputs an estimator of the optimal action-value $Q^{\star}(s, a)$, which is defined by the expectation of the cumulative sum of the discounted reward following the optimal policy, i.e.,
\begin{align}
\label{eq:action_value}
Q^{\star}(s, a)\coloneqq\max_{\pi}\mathbb{E}_{\pi}\left[\sum_{k' = 0} ^{\infty} \gamma^{k'} r_{k + k' + 1}\,\middle|\,s_k = s, a_k = a\right],
\end{align}
where $\gamma$ is the discount factor, which is assigned a value of 0.99.
The term $\pi:\mathcal{S}\to\mathcal{A}$ denotes the arbitral policy, where $\mathcal{S}$ and $\mathcal{A}$ are the state and action sets, respectively.
Note that the optimal policy $\pi^{\star}$ is determined by taking an action that maximizes the optimal action-value, i.e., by taking an action $\argmax_{a\in\mathcal{A}} Q^{\star}(s, a)$.
Hence, to obtain the optimal actions given the state information, it is sufficient to learn the optimal action-value as accurately as possible.
Let the NN that estimates the optimal action-value be denoted by $Q(s, a; \bm \theta)$ in the sequel.
The NN model $Q(s, a; \bm \theta)$ in this evaluation is multi-layer-perceptron that accepts the input of state information and outputs the action-values corresponding each action.
The network consists of three hidden layers having $128$ units.
The activation function of the hidden layers is the rectified linear unit (ReLU) $R(x)$ given by $R(x) \coloneqq \max \{x, 0\}$.
This network architecture is just a preliminary example, and however is shown to be sufficient to achieve our objective, i.e., feasibly learning an appropriate beam-tracking policies in the on-wire dynamics.
In the DRL algorithm, the NN is trained such that the difference between $Q(s_k, a_k; \bm \theta)$ and $r_k + \gamma\max_{a'} Q(s_{k + 1}, a'; \bm \theta)$ is minimized.
To quantify the difference, we leverage the following Huber Loss\cite{varga2018deeprn} as an example:
\begin{align}
L(\bm \theta) =
\begin{cases}
x^2/2, & x \leq 1; \\
|x| - 0.5, & \text{otherwise},
\end{cases}
\end{align}
where $x \coloneqq Q(s_k, a_k; \bm \theta) - \bigl(r_k + \gamma\max_{a'} Q\bigl(s_{k + 1}, a_{k + 1}; \hat{\bm \theta}\bigr)\bigr)$.
Therein, $Q\bigl(\cdot, \cdot; \hat{\bm \theta}\bigr)$ is the proxy NN, which is known as the ``target network'' and is updated periodically to stabilize the training\cite{mnih2015human}.
To minimize the above loss function, we utilize the well-known Adam algorithm\cite{sutskever2013importance} with a learning rate of $10^{-4}$ per 300 time steps.
This parameter update with the Adam optimizer is performed over the 2048 past experiences of $(s_k, a_k, r_k, s_{k + 1})$ randomly sampled from a replay buffer with the mini-batch size of 32 and the number of epochs of 8, which is iterated by four times while varying the sampled experiences.
The target network is updated each 3000 time steps.
The process of training the network required 100,000 time steps in total, and the policy determined via the trained NN was tested while running 300 time steps after every update of the NN.
During training, DRL requires the exploration of unexplored actions.
To do this, we employ an $\epsilon$-greedy policy that select the action that maximizes $Q(s, a; \bm \theta)$ with the probability of $1 - \epsilon$, selecting actions uniformly randomly otherwise.
In this experiment, $\epsilon$ is set as 0.2 as an example.
Meanwhile, while testing the learned policy, we set $\epsilon$ as 0, which corresponds to select an action in a greedy fashion according to the learned optimal action values.
\subsection{Beam-Tracking Performance without Knowing Positional Interaction}
We evaluate the learned beam-tracking policy without knowing the positional interaction on the wire, i.e., without the state expansion discussed in Section~\ref{subsec:state_expansion}.
These results showed that a longer look-back time $T$ results in frequent beam misalignment particularly in the case of disturbances caused by impulsive forces to the wire, which seamlessly emphasizes the importance of the aforementioned state expansion in leveraging the positional interaction.
\begin{figure}[t]
\centering
\subfigure[Instantaneous received power.]{\includegraphics[width=0.7\columnwidth]{received_power_quick.pdf}}\vspace{.5em}
\subfigure[Difference from optimal angle.]{\includegraphics[width=0.7\columnwidth]{angle_difference_quick.pdf}}
\caption{Time-varying received power and difference from optimal policy in terms of the steering angle when the position of the on-wire node is affected only by wind perturbation ($m = 10\,\mathrm{kg}$, $k_0 = 1000\,\mathrm{Nm^{-1}}$, and $T = 20\,\mathrm{ms}$).}
\label{fig:quick_analysis}
\end{figure}
\vspace{.3em}\noindent\textbf{Robustness Against Wind Perturbations.}\quad
First, we confirm that our DRL-based beam tracking learns an appropriate beam-tracking policy for a simple case in which the on-wire node is perturbed only by the wind.
Fig.~\ref{fig:quick_analysis} shows the instantaneous angle difference from the optimal angles and received power in an exemplary situation where $m = 10\,\mathrm{kg}$, $k_0 = 1000\,\mathrm{Nm^{-1}}$, and $T = 20\,\mathrm{ms}$.
As a reference, we plot the results from the beam fixed at an initial angle to allow ones to capture how seriously wind perturbations affect the received power.
As shown in Fig.~\ref{fig:quick_analysis}(a), with the fixed beam, the received power is drastically degraded, which indicates the importance of beam tracking in the on-wire mmWave node.
Meanwhile, the received power in the DRL-based beam tracking comes close to the optimal values, and the difference between them is within 2\,dB.
This indicates that the angle of departure relative to the main lobe steering angle is within half-power beam width, and in this sense, the beam is appropriately steered against the wind perturbations.
Note that the near-optimal solution can also be achieved by conventional exhaustive search for short periods in which possible beam directions are rapidly scanned to find the beam steering direction with maximum received power.
Nevertheless, the importance of this result lies in the fact that \textbf{an appropriate beam-tracking policy can be feasibly learned based on the look-back information about the position and velocity} without scanning channel information.
In Fig.~\ref{fig:wind_one_point_various_parameters}, we investigate the feasibility of learning appropriate beam-tracking policies for various wire characteristics.
Specifically, the dependency of the mass of the messenger wire $m$ is shown in Fig.~\ref{fig:wind_one_point_various_parameters}(a) whereas that of the spring constant $k_0$ is shown in Fig.~\ref{fig:wind_one_point_various_parameters}(b).
Both figures show that the received powers of the DRL-based beam-tracking come close to the optimal values, with the difference between them being less than 0.5\,dB, which is a much smaller difference than without beam tracking.
Hence, against wind perturbations (without impulsive forces to the wire), we can feasibly learn an appropriate beam-tracking policy for various wire characteristics.
\begin{figure}[t]
\centering
\subfigure[Average received power vs. mass of overhead messenger wire $m$ ($k_0 = 1000\,\mathrm{Nm^{-1}}$).]{\includegraphics[width=0.66\columnwidth]{avg_rss_vs_mass_ver2.pdf}}
\subfigure[Average received power vs. spring constant $k_0$ ($m = 10\,\mathrm{kg}$).]{\includegraphics[width=0.66\columnwidth]{avg_rss_spring_ver2.pdf}}\vspace{-0.5em}
\caption{Dependency of wire characteristics when the position of the on-wire node is affected only by wind perturbation (Look-back time $T = 80\,\mathrm{ms}$).}
\label{fig:wind_one_point_various_parameters}
\end{figure}
\begin{figure}[t]
\centering
\subfigure[Instantaneous received power.]{\includegraphics[width=0.7\columnwidth]{received_power_one_point_shock.pdf}}\vspace{1em}
\subfigure[Difference from optimal angle.]{\includegraphics[width=0.7\columnwidth]{angle_difference_one_point_shock.pdf}}
\caption{Time-varying received power and difference from optimal policy in terms of steering angle in the position of the on-wire node affected by both wind perturbations and impulsive forces to the wire ($m = 10\,\mathrm{kg}$, $k_0 = 1000\,\mathrm{Nm^{-1}}$, and $T = 80\,\mathrm{ms}$).}
\label{fig:collsision_missalignment}
\end{figure}
\vspace{.3em}\noindent\textbf{Robustness Against Displacements From Impulsive Force.}\quad
We evaluate the performance ot the learned beam-tracking policy in a situation in which the position of the on-wire node is affected by both wind perturbations and impulsive forces.
Fig.~\ref{fig:collsision_missalignment} shows the instantaneous received power and the angle differences from the optimal angles in the same parameter setting as Fig.~\ref{fig:quick_analysis}.
In Fig.~\ref{fig:collsision_missalignment}(a), we can observe that the received power is degraded by approximately 10\,dB after the impulsive force is applied at 2\,s.
This degradation in received power is due to beam misalignment, which is confirmed in Fig.~\ref{fig:collsision_missalignment}(b), in which the difference from the optimal policy is larger in terms of the steering angle.
This beam misalignment is attributed to the fact that the look-back position of the on-wire node lacks informative features for predicting the current on-wire node position.
Indeed, the variation in the position of the on-wire node caused by the impulsive force is more rapid than that caused by wind perturbations.
During such a rapid variation, the look-back positions of the on-wire node and the current positions are decorrelated, and hence, using only the look-back positions, it would not be possible for the agent to predict the current on-wire node positions, resulting in beam misalignment.
In the next section, we demonstrate that this beam misalignment can be prevented by utilizing the aforementioned expanded state, that is, by utilizing the positions and velocities of the points distributed on the messenger wire as state information.
\begin{figure}[t]
\centering
\includegraphics[width=0.8\columnwidth]{avg_rss_vs_delay_w_wo_bird.pdf}
\caption{Received power averaged for 300\,ms after an impulsive force is applied to the wire in the use of only the position and veocity of the on-wire node.}
\label{fig:avg_rss_delay_w_wo_bird}
\end{figure}
During the rapid displacement of the on-wire node as a result of the impulsive force, the look-back positions of the on-wire node become less informative to predict current positions, and the average received power decreases accordingly.
We confirm this in Fig.~\ref{fig:avg_rss_delay_w_wo_bird}, which shows the received power averaged for 300\,ms after the impulsive force is applied.
As a reference, we also plotted the average received power within the same time duration for a situation in which no impulsive forces are applied.
Fig.~\ref{fig:avg_rss_delay_w_wo_bird} clearly shows that the learned beam-tracking policy performs worse in terms of the average received power as the look-back time becomes longer, which can be explained by the aforementioned reasons.
\subsection{Beam-Tracking Performance with Knowing Positional Interaction}
In this section, we confirm that the aforementioned beam misalignment can be avoided by utilizing the expanded state.
This allows the beam-training agent to be informed not only of the dynamics of the on-wire node but also that of the multiple points of the messenger wire.
We consider the impulsive force as the cause of the perturbation that affects the position of the on-wire node because, under such conditions, the beam misalignment was shown to be more severe (refer the previous section).
As an example, we use the position and velocity of nine points on the messenger wire.
The indices of these points are $2, 4, 6, 8, 10, 12, 14, 16, 18$, i.e., $\mathcal{N}_{\mathrm{sense}} = \{\,2n\,|\, n\in\mathbb{N}, 1\leq n\leq 9 \,\}$.
The results in Fig.~\ref{fig:tiemseries_rss_angle_1p_9p} confirm that the beam misalignment in Fig.~\ref{fig:collsision_missalignment} can be feasibly avoided by leveraging the expanded state that utilizes the position and velocity of nine points on the wire.
Fig.~\ref{fig:tiemseries_rss_angle_1p_9p} compares the situation with and without the state expansion in terms of the instantaneous received power and difference from the optimal steering angle.
Note that the result in Fig.~\ref{fig:tiemseries_rss_angle_1p_9p} without the state expansion is same as that in Fig.~\ref{fig:collsision_missalignment}.
Fig.~\ref{fig:tiemseries_rss_angle_1p_9p}(a) shows that the received power for the beam tracking with the state expansion does not decrease as severely as that without the state expansion.
This is because beam misalignment is avoided, as shown in Fig.~\ref{fig:tiemseries_rss_angle_1p_9p}(b), which shows the less difference from the optimal steering angle.
These results are understood as discussed in Section~\ref{subsec:state_expansion}: the looked-back positions and velocities of multiple points on the messenger wire include an informative feature that can be used to predict sudden variation in the current position of the on-wire node as a result of the impulsive force.
Hence, the appropriate beam-tracking policy could be learnable even when the position of the on-wire node varies suddenly because of the impulsive force.
\begin{figure}[t]
\centering
\subfigure[Instantaneous received power.]{\includegraphics[width=0.83\columnwidth]{timeseries_rss_1p_9p_ver2.pdf}}\hspace{1em}
\subfigure[Difference from optimal angle.]{\includegraphics[width=0.83\columnwidth]{timeseries_angle_diff_1p9p.pdf}}
\caption{Comparison of the state definition (use of only the position and velocity of the on-wire node vs. that for nine points on the messenger wire) in terms of the time-varying received power and difference from the optimal angle ($m = 10\,\mathrm{kg}$, $k_0 = 1000\,\mathrm{Nm^{-1}}$, and $T = 80\,\mathrm{ms}$).}
\label{fig:tiemseries_rss_angle_1p_9p}
\end{figure}
We also confirm that the degradation of the average received power because of the increase in the look-back time $T$ can be mitigated by using the expanded state.
In Fig.~\ref{fig:avg_rss_vs_delay_1p_9p}, we show the received power averaged for 300\,ms after the impulsive force is applied.
Again, the results without state expansion shown in Fig.~\ref{fig:avg_rss_vs_delay_1p_9p} are the same as those in Fig.~\ref{fig:avg_rss_delay_w_wo_bird}.
The results in Fig.~\ref{fig:avg_rss_vs_delay_1p_9p} show that the average received power for the beam tracking with the state expansion is higher than that without.
This is attributable to the aforementioned avoidance of the beam-misalignment achieved by knowing the positional interaction on the wire.
\begin{figure}[t]
\centering
\includegraphics[width=0.8\columnwidth]{avg_rss_vs_delay_1p_9p.pdf}
\caption{Comparison of the state definition (use of only the position and velocity of the on-wire node vs. that for nine points on the messenger wire) in terms of the average received power for 300\,ms after the impulsive force is applied to the wire.}
\label{fig:avg_rss_vs_delay_1p_9p}
\end{figure}
\section{Conclusion}
\label{sec:conclusion}
In this paper, we investigated the feasibility of tracking mmWave directional beams under different types of perturbations specific to on-wire placement based on historical information about the position and velocity.
This beam-tracking problem presents a challenge owing to the complicated dynamics of messenger wires subject to a tensile force, wind perturbations, and impulsive forces applied to the wire.
We implemented a DRL-based beam-tracking framework to learn the relationship between these position and velocity values of a mmWave node and an appropriate beam steering angle.
Numerical evaluations to reproduce the wire-like dynamics demonstrated the feasibility of the DRL-based beam-tracking framework.
It was shown that, against wind perturbations, beam misalignment is feasibly avoided based on the historical positions and velocities of the mmWave node.
Moreover, it was shown that beam misalignment resulting from an impulsive force to the wire cannot necessarily be avoided by using only historical positions and velocities; however, it can be feasibly avoided by leveraging the positions and velocities of several points on the messenger wire.
It should finally be noted that this study was conducted with the objective of allowing ones to make the most of mmWave radios including 5G NR by exploring a new installation place potentially beneficial to ensure LOS connectivities.
This sheds light on gaining the flexibility in the physical node deployment, and in this sense, we believe that the insight provided by this study contributes to the future on-going network deployment of the 5G and beyond.
\bibliographystyle{IEEEtran}
|
\section{Introduction}
Network data is among the most common types of relational data.
As a fundamental task in network data analysis \cite{wasserman1994social,goldenberg2010survey},
\emph{community detection}
refers to the problem of partitioning the nodes of a network into clusters so that intra-cluster nodes are connected in a different way from inter-cluster nodes, usually more densely.
\emph{Stochastic block model} (SBM) \cite{holland1983stochastic}
is a canonical model for studying community detection.
In an SBM, $n$ nodes are partitioned into $k$ disjoint subsets.
Each unordered pair of nodes are connected independently with probability $p$ if they come from the same community and with a different probability $q$ otherwise.
The observed connection pattern is encoded in an $n\times n$ symmetric adjacency matrix $A$.
Here the goal of community detection is to, upon observing $A$, estimate the partitioning of nodes.
\iffalse
In the two block case,
the model is also known as the planted partition model, and the community structure among the $n$ nodes is parameterized by an unknown binary vector
$\mathbf{z}\in\{\pm 1\}^n$.
Pairs of nodes are connected independently with probability $p$ if they come from the same community and with probability $q$ otherwise.
Such a data generating procedure gives rise to a symmetric adjacency matrix $A\in\{0, 1\}^{n\times n}$, whose diagonal entries are zeros and the off-diagonal entries satisfy
\begin{equation}
\label{eq:sbm}
A_{ij} = A_{ji} \stackrel{ind}{\sim} \text{Bern}(p)\cdot \mathbf{1}{\mathbf{z}_i = \mathbf{z}_j} + \text{Bern}(q) \cdot \mathbf{1}{\mathbf{z}_i \neq \mathbf{z}_j}, ~~ \forall i\neq j\in[n],
\end{equation}
where $\text{Bern}(\cdot)$ is a Bernoulli random variable and $\mathbf{1}{\cdot}$ is the indicator function.
The goal of the data analyst is to, upon observing $A$, estimate the vector $\mathbf{z}$ up to a relabeling.
\fi
The stochastic block model, albeit simple, has found its success in many fields of science (see, e.g., \cite{jackson2010social,lu2011link,jackson2011overview,garcia2018applications}).
It has also undergone a plethora of theoretical investigations \cite{abbe2017community}.
With joint efforts from mathematics, statistics, and computer science,
we not only have a large algorithmic toolbox for detecting communities in SBMs, but also know the {information-theoretic limits} of this task as well as
{which algorithms are optimal}.
Despite its popularity,
stochastic block model focuses only on
a {single} adjacency matrix.
This is in sharp contrast to the widely recognized fact that real world networks are often superpositions of multiple networks (layers), each encoding a potentially different but correlated interaction pattern among the same set of nodes \cite{kivela2014multilayer,boccaletti2014structure}.
For example, in social network data, different layers could correspond to different types of relationship that link the social entities, and the information presented in the friendship network, compared to that in the professional network, is different, but not completely unrelated \cite{dickison2016multilayer}.
Another example is given by the network representation of human brains, where each layer corresponds to an individual person's functional brain network.
It is well known that the parcellation of brain regions into different functional units are different but strongly correlated among human beings \cite{de2017multilayer}.
A natural attempt at generalizing SBMs to multilayer networks is as follows.
Let us focus on the symmetric two block case where in each layer all nodes are partitioned into two blocks of roughly equal sizes.
Instead of observing a single adjacency matrix, the data analyst is now presented with a collection of $L$ adjacency matrices $\{\AAA{\ell}\}_{\ell=1}^L$.
To model that ``communities in different layers are different but correlated'', we take a hierarchical modeling approach.
A \emph{global} community assignment vector $\mathbf{z}^\star \in \{\pm 1\}^n$ is introduced in our model.
To ensure symmetry, we require $\mathbf{z}^\star$ to have roughly equal numbers of $1$'s and $-1$'s.
We let the \emph{individual} community assignments $\{\zzz{\ell}\}_1^L$ be independent samples from the following distribution:
\begin{equation}
\label{eq:label_flip}
\forall\ell\in[L], i\in[n], \qquad \zzz{\ell}_i ~{\stackrel{ind}{\sim}}~ \mathbf{z}^\star_i \times \big[2 \text{Bern}(1-\rho) - 1\big].
\end{equation}
Here, $\text{Bern}(\cdot)$ is a Bernoulli random variable.
That is, in a specific layer $\ell\in[L]$, with probability $1-\rho$, the community membership of the $i$-th node agrees with the global one, $\mathbf{z}^\star_i$, and with probability $\rho$, it ``flips'' to the opposite side $-\mathbf{z}^\star_i$. The parameter $\rho$ controls the level of \emph{inhomogeneity} across layers. When $\rho = 0$, all layers share the same community structure, whereas when $\rho=1/2$, the community structures across layers are mutually uninformative.
Upon realizations of $\zzz{\ell}$'s, the adjacency matrices are independently generated by
\begin{equation}
\label{eq:layerwise_sbm}
\AAA{\ell}_{ij} = \AAA{\ell}_{ji} \stackrel{ind}{\sim} \text{Bern}(p_\ell)\cdot \mathbf{1}{\zzz{\ell}_i = \zzz{\ell}_j} + \text{Bern}(q_\ell) \cdot \mathbf{1}{\zzz{\ell}_i \neq \zzz{\ell}_j}, ~~ \forall i\neq j\in[n],\,\ell\in [L],
\end{equation}
where $\mathbf{1}{\cdot}$ is the indicator function, and all diagonal entries are zeros.
In other words, the $\ell$-th layer network is generated by an SBM with community partitioning specified by $\zzz{\ell}$, intra-community connection probability $p_\ell$ and inter-community connection probability $q_\ell$.
Connection probabilities across different layers are not linked in any way.
The foregoing generalization, to the best of our knowledge, was first introduced by \citet{paul2018random}, which they termed as the \emph{random effects stochastic block model}.
In \cite{paul2018random}, the random effects \eqref{eq:label_flip} can take other forms. Hence, to avoid confusion, we term the model in \eqref{eq:label_flip}--\eqref{eq:layerwise_sbm} as the \emph{inhomogeneous multilayer stochastic block model} (IMLSBM).
Clearly, the model can be generalized in obvious ways to include more than two communities and unequal community sizes.
However, the present manuscript shall focus on the foregoing symmetric two block case as it is the simplest nontrivial model that reveals key new phenomena of community detection in inhomogeneous multilayer networks.
The goal of community detection in an IMLSBM~is now two-fold---upon observing $\{\AAA{\ell}\}_1^L$, we are interested in:
\begin{enumerate}
\item \emph{Global~estimation.} Estimating the global community assignment $\mathbf{z}^\star$;
\item \emph{Individualized~estimation.} Estimating each of the individual assignments $\{\zzz{\ell}\}_1^L$.
\end{enumerate}
Global estimation needs to aggregate connection patterns across layers to better infer the global consensus structure, an instance of \emph{integrative data analysis} \cite{curran2009integrative}. On the other hand, individualized~estimation~requires borrowing information from different layers to better estimate the layer-wise community structure, an example of \emph{multi-task learning} \cite{caruana1997multitask}.
Theoretical understanding of community detection in an IMLSBM~is lacking. Partial results exist in the homogeneous\footnote{By ``homogeneity'' we mean the layer-wise community structures are shared across all layers. The connecting probabilities are allowed to differ.} case ($\rho=0$) where global~estimation~and individualized~estimation~coincide. Under such a setup, it has been proved by \citet{paul2016consistent} that the minimax rates for expected proportion of misclustered nodes\footnote{Under homogeneity, \citet{paul2016consistent} established the rates for a more general setting than the symmetric two block case considered in this paper.} scales as
\begin{equation}
\label{eq:rho_zero_rate}
\exp\bigg\{ -\big(1+o(1)\big) \cdot \frac{n}{2}\sum_{\ell\in[L]} (\sqrt{p_\ell} - \sqrt{q_\ell})^2 \bigg\},
\end{equation}
provided that the exponent diverges to infinity as $n$ tends to infinity.
Later, a polynomial-time algorithm that achieves this rate was proposed by \citet{xu2020optimal}. Nonetheless, it is unclear how to generalize their results to the inhomogeneous setting.
From an algorithmic perspective, spectral clustering \cite{bhattacharyya2018spectral,lei2020tail,paul2020spectral} and least-square estimators \cite{lei2020consistent,wang2019multiway} have been proposed and justified to be consistent (i.e., achieving an $o(1)$ misclustering proportion with high probability) under homogeneity.
However, it is unknown whether any of these methods attains the information-theoretic limit \eqref{eq:rho_zero_rate}.
Although the homogeneous case ($\rho=0$) is interesting in its own right, it is the inhomogeneous case ($\rho > 0$) that characterizes our inductive bias --- ``layers are different but correlated''.
In \cite{paul2018random}, a few heuristic algorithms were introduced and their performances were assessed by simulations.
To the best of our limited knowledge, in the inhomogeneous regime, no algorithm with provable guarantee for either global or individualized estimation is known in the literature, let alone any optimality statement.
\subsection{Main Contributions}
The main contributions of the present manuscript are two-fold.
First, we give precise characterization of information-theoretic limits of both global and individualized community detection in a {symmetric two block} IMLSBM~when $\rho =o(1)$;
Moreover, we provide a polynomial-time algorithm that simultaneously attains information-theoretic limits for both global and individualized~estimation~under mild conditions.
We reiterate that results in the present manuscript are obtained under the symmetric two block setting which has already posed highly nontrivial theoretical and algorithmic challenges.
We leave extensions to more general settings for future work.
To provide an overview of our main results, we start with several key information-theoretic quantities that will appear throughout this paper.
For any $\ell\in[L]$ and $t\in[0, 1]$, define
\begin{align}
\label{eq:layerwise_info}
\III{\ell}_t & : = -\log\big[p_\ell^{1-t}q_\ell^t + (1-p_\ell)^{1-t}(1-q_\ell)^t\big]\big[p_\ell^tq_\ell^{1-t} + (1-p_\ell)^t(1-q_\ell)^{1-t}\big].
\end{align}
The quantity $\III{\ell}_t$ can be regarded as the \emph{signal strength} of the $\ell$-th layer.
When $p_\ell\asymp q_\ell =o(1)$, one can show that
$
\III{\ell}_{1/2} = \left(1+o(1)\right) (\sqrt{p_\ell} - \sqrt{q_\ell})^2,
$
and hence the minimax rate for community detection in an SBM with community assignment $\zzz{\ell}$, intra-community connection probability $p_\ell$ and the inter-community connection probability $q_\ell$ dervied in \citet{zhang2016minimax} can be equivalently written as
$
e^{-\left(1+o(1)\right) \III{\ell}_{1/2}}$,
as long as the exponent tends to infinity.
For any collection of layers $S\subseteq[L]$, let
\begin{equation}
\label{eq:cum_gen_fun}
\psi_S(t) := -\frac{n}{2}\sum_{\ell \in S} \III{\ell}_t, \qquad \psi_S^\star(a) := \sup_{0\leq t \leq 1} at-\psi_S(t).
\end{equation}
The function $\psi_S^\star$
characterizes the \emph{collective signal strength for layers in $S$}.
Indeed, from the definition of $\III{\ell}_t$, one readily checks that
$
\psi_S^\star(0) = -\psi_S(1/2) =(n/2) \sum_{\ell\in S}\III{\ell}_{1/2},
$
and thus the minimax rate \eqref{eq:rho_zero_rate} for community detection in a homogeneous multilayer SBM can be expressed equivalently as
$
e^{-\left(1+o(1)\right) \psi_{[L]}^\star(0)}.
$
Intuitively, inhomogeneity ($\rho > 0$)
introduces additional noises.
To characterize the \emph{noise level}, define
\begin{equation}
\label{eq:J_rho}
{J_\rho} := -\log 2\sqrt{\rho(1-\rho)}
\end{equation}
Since $J_0= \infty$ and $J_{1/2} = 0$, one can effectively think of $J_\rho$ as a measure of proximity of the individual layer community assignments $\{\zzz{\ell}\}_1^L$ to the global assignment $\mathbf{z}^\star$.
Note that both $\III{\ell}_t$ and $J_\rho$ can be written as convex combinations of R\'enyi divergences \cite{renyi1961measures}
between Bernoulli distributions.
Specifically, we have
$\III{\ell}_t = (1-t) D_t(p_\ell\| q_\ell) + t D_{1-t}(p_\ell\| q_\ell)$, and $J_\rho = \frac{1}{2}D_{1/2}(\rho \| 1-\rho)$, where $D_{t}(p\| q)$ is the \emph{R\'enyi divergence of order $t$} between $\textnormal{Bern}(p)$ and $\textnormal{Bern}(q)$.
\paragraph{Global estimation error.}
With the foregoing definitions,
we first show that, under certain regularity conditions, the minimax rate for global~estimation, measured in terms of proportion of misclustered nodes, is given by
\begin{equation}
\label{eq:resbm_glob_rate}
\exp\big\{-\left(1+o(1)\right) \min_{S\subseteq[L]} \mathcal{I}_S\big\},
\end{equation}
where $\mathcal{I}_{S}$ represents the \emph{signal-to-noise ratio (SNR) for global~estimation} for layers in $S$, defined as
\begin{equation}
\label{eq:globinfo}
\mathcal{I}_S :=
\begin{cases}
|S^c| J_\rho + \psi_S^\star(0) & \textnormal{ if } |S^c| \textnormal{ is even},\\
(|S^c| + 1) J_\rho + \psi^\star_S(-2 J_\rho) & \textnormal{ if } |S^c| \textnormal{ is odd}.
\end{cases}
\end{equation}
The minimax error rate \eqref{eq:resbm_glob_rate} exhibits two intriguing properties. First, when $\rho = 0$, we have $J_\rho = \infty$, and thus the only way to make $\mathcal{I}_S$ finite is to choose $S = [L]$, which gives
$
\min_{S\subseteq[L]} \mathcal{I}_S = \mathcal{I}_{[L]} = \psi_{[L]}^\star(0).
$
As the result, \eqref{eq:resbm_glob_rate} recovers the minimax rate in a homogeneous multilayer SBM given in \eqref{eq:rho_zero_rate}. Second, the SNR for layers in $S$ takes different forms according to the \emph{parity} of $S$, a phenomenon induced by inhomogeneity across layers.
\paragraph{Individualized estimation error.}
In correspondence, the minimax rate for individualized~estimation~for the $\ell$-th layer, measured by proportion of misclustered nodes, is given by
\begin{equation}
\label{eq:resbm_ind_rate}
\exp\big\{-\left(1+o(1)\right) \min_{S\subseteq[L]\setminus\{\ell\}} \mathcal{I}_{S\cup\{\ell\}}\big\} + \exp\big\{ -\left(1+o(1)\right)\mathcal{J}_{\{\ell\}}\big\} ,
\end{equation}
where $\mathcal{J}_{\{\ell\}}$ is a suitably defined quantity that measures the \emph{SNR for individualized~estimation} for the $\ell$-th layer.
We briefly mention here that similar to
\eqref{eq:resbm_glob_rate}, the last display can recover the minimax rate in a homogeneous multilayer SBM by setting $\rho = 0$, and it crucially depends on the parity of the ``most informative'' set $S$ as well.
We refer readers to Sections \ref{sec:lower_bound} and \ref{sec:minimax_rate} for details.
\paragraph{Algorithm.}
We propose an algorithm that achieves the optimal rates in \eqref{eq:resbm_glob_rate} and \eqref{eq:resbm_ind_rate} simultaneously under mild conditions.
The idea stems from \emph{maximum a posteriori} (MAP) estimation.
Note that IMLSBM~is a hierarchical model where individual community assignments $\{\zzz{\ell}\}_1^L$ are drawn from the prior distribution \eqref{eq:label_flip}.
It is thus tempting to write out the posterior of $\{\zzz{\ell}\}_1^L$ given the observed data $\{\AAA{\ell}\}_1^L$ and maximize the posterior density with respect to the parameters $(\mathbf{z}^\star, \{\zzz{\ell}\}_1^L)$.
A naive implementation of this strategy is doomed to fail, due to the fact that the MAP objective function gives rise to a combinatorial optimization problem, whose search space has cardinality $2^{n(L+1)}$.
To bypass the combinatorial search, we adopt a two-stage ``warm-start'' MAP algorithm. In the first stage, an initial estimator of $\mathbf{z}^\star$ is obtained using spectral clustering on a trimmed version of the weighted average of layer-wise adjacency matrices.
In the second stage, a refined estimator of $\mathbf{z}^\star$ and estimators of $\{\zzz{\ell}\}$ are simultaneously obtained by optimizing a ``decoupled'' MAP objective function, which can be computed in linear time (in $n$ and $L$).
Although the formal definition of the MAP refinement step requires knowledge of $\rho$, our numerical experiments later show that the outcome is not sensitive to misspecification of $\rho$.
While similar ``spectral clustering + refinement'' procedures have appeared in community detection in \textnormal{SBM} s and variants,
our algorithm has novelties in both stages, especially in their technical analysis.
For Stage \RN{1}, compared to existing analyses of spectral clustering for homogeneous multilayer SBMs \cite{bhattacharyya2018spectral,lei2020tail,paul2020spectral},
our analysis is novel in that we establish a \emph{stability} result, asserting that inhomogeneity hurts spectral clustering error rate by at most an additive factor of $\textnormal{poly}(\rho)$.
The proof is based on a new concentration inequality on the spectral norm of a weighted average of Bernoulli random matrices, which is derived via a nontrivial generalization of the graph decomposition approach in \citet{le2017concentration} to multilayer networks.
The concentration inequality improves the ones used in the existing work (e.g., \cite{bhattacharyya2018spectral,paul2020spectral}) and could be of independent interest.
For Stage \RN{2}, due to presence of multiple layers, devising a refinement scheme with time complexity that is polynomial in the number of layers presents new challenges.
In addition, due to inhomogeneity, the analysis
is considerably more involved.
A key step towards establishing matching upper bounds lies in a novel application of Sion's minimax theorem \cite{sion1958general}.
\subsection{Related Work}
The past decade has witnessed a venerable line of work on the theoretical development of community detection for SBMs.
Optimal algorithms have been developed under various criteria, including
(1) weak recovery, where the best achievable goal is to cluster the nodes better than random guess \cite{decelle2011asymptotic,mossel2012stochastic,mossel2018proof,bordenave2015non,mossel2015reconstruction,chen2014statistical,neeman2014non,montanari2015finding,abbe2018proof};
(2) exact recovery, where the requirement is to reconstruct from data the ground truth up to relabeling \cite{mossel2015consistency,abbe2015exact,abbe2015community};
and more related to our formulation,
(3) almost exact recovery, where the goal is to output a community assignment with vanishing misclustering error \cite{yun2014community,mossel2015consistency,abbe2015exact,gao2017achieving}.
The study under the minimax framework was initiated by \cite{zhang2016minimax,abbe2015community,yun2016optimal} and was later extended to more general settings such as \cite{gao2018community,xu2020optimal}.
The above list of work is by no means exhaustive and we refer the readers to the review papers \cite{abbe2017community,li2018convex,gao2018minimax} for a more systematic account.
In comparison, study of community detection in multilayer networks is still in its early stage.
Initial works in this area have focused on algorithmic developments (see, e.g., \cite{mucha2010community,kumar2011co,de2013mathematical,de2015identifying,peixoto2015inferring}), and most theoretical studies are restricted to the homogeneous case where all layers share the same community structure \cite{paul2016consistent,bhattacharyya2018spectral,paul2018random,paul2020spectral,lei2020consistent,lei2020tail,zhang2020flexible}.
There are a few exceptions, such as \cite{stanley2016clustering,le2018estimating,arroyo2019inference,jing2020community}, where consistency has been established under several inhomogeneity-aware variants of SBMs while optimality results are missing.
In addition to estimating community structures, there is a related line of work aiming at testing whether community structures across layers are indeed correlated or equivalent.
We refer interested readers to recent papers \cite{gao2019testing,gao2020clusterings,gao2019testingEquiv} and references therein.
\subsection{Paper Organization}
The rest of the paper is organized as follows. In Section \ref{sec:lower_bound}, we provide our construction of minimax lower bounds for both global and individualized~estimation. We present the two-stage algorithm in Section \ref{sec:alg}, and its theoretical analysis is given in Section \ref{sec:minimax_rate}. We conduct comprehensive experiments to corroborate our theoretical results in Section \ref{sec:exp}.
For brevity, most of the proofs are deferred to the appendix.
\smallskip
{\noindent \bf Notations.} We conclude this section by introducing some notations that will be used throughout this paper. For a positive integer $n$, we write $[n]:=\{1, \hdots, n\}$.
Given $a, b \in {\mathbb{R}}$, we denote $a\lor b := \max\{a, b\}$ and $a\land b := \min\{a, b\}$.
For a set $S$, we let $\mathbf{1}_S$ be its indicator function and we use $\# S$ and $|S|$ interchangeably to denote its cardinality.
For two positive sequences $\{a_n\}$ and $\{b_n\}$, we write
$a_n \lesssim b_n$ or $a_n = {\mathcal{O}}(b_n)$ to denote $\limsup a_n/b_n < \infty$,
and we let $a_n \gtrsim b_n$ or $a_n = \Omega(b_n)$ to denote $b_n \lesssim a_n$.
Meanwhile, the notation $a_n\asymp b_n$ or $a_n = \Theta(b_n)$ means $a_n \lesssim b_n$ and $a_n \gtrsim b_n$ simultaneously.
Moreover, we write $a_n\ll b_n$ to mean $b_n/a_n\to\infty$ and $a_n\gg b_n$ to mean $b_n \ll a_n$.
For a vector $x$, we let $\|x\|_p$ denote its $\ell_p$ norm, and we write $\|x\|_2 = \|x\|$ when there is no ambiguity.
For a matrix $A$, we let $\|A\|_F$ be its Frobenius norm and $\|A\|_{p\to q}$ be its $\ell_p$ to $\ell_q$ operator norm. We will write $\|A\|_{2\to 2} = \|A\|_2 = \|A\|$ when there is no ambiguity.
\section{Fundamental Limits and Costs of Inhomogeneity}\label{sec:lower_bound}
In this section, we present minimax lower bounds for estimating both $\mathbf{z}^\star$ and individual $\zzz{\ell}$'s. To start with, let us recall that a two-block IMLSBM~parameterized by $(\mathbf{z}^\star, \rho, \{p_\ell\}_1^L, \{q_\ell\}_1^L)$ is a probability measure on a multilayer network, whose adjacency matrices $\{\AAA{\ell}\}_1^L$
are generated according to \eqref{eq:label_flip}--\eqref{eq:layerwise_sbm}.
\paragraph{Parameter space and the loss function.}
Let $n^\star_\pm(\mathbf{z}^\star) = \sum_{i\in[n]}\mathbf{1}{\mathbf{z}^\star = \pm1}$ be the sizes of the positive and negative clusters of $\mathbf{z}^\star$, respectively.
We propose to consider the following collection of IMLSBM s:
\begin{align}
\label{eq:param_space}
{\mathcal{P}}_n(\rho, \{p_\ell\}_1^L, \{q_\ell\}_1^L, \beta):= \bigg\{ \textnormal{IMLSBM}(\mathbf{z}^\star, \rho, \{p_\ell\}_1^L, \{q_\ell\}_1^L) : \ \frac{n}{2\beta} \leq n^\star_\pm(\mathbf{z}^\star) \leq \frac{n\beta}{2}, p_\ell > q_\ell \ \forall \ell \bigg\}.
\end{align}
As we focus on the symmetric case, the constant $\beta\geq 1$ is taken to be $1+o(1)$ as $n\to\infty$. The rest of the quantities appearing above, namely $L, \rho, \{p_\ell\}_1^L, \{q_\ell\}_1^L$, are all allowed to scale with $n$. For an estimator $\mathbf{\hat z}^\star$ of $\mathbf{z}^\star$, we evaluate its performance by the misclustering proportion, defined as
\begin{align}
\label{eq:loss_func}
{\mathcal{L}}(\mathbf{\hat z}^\star, \mathbf{z}^\star) : = \frac{d_\textnormal{H}(\mathbf{\hat z}^\star, \mathbf{z}^\star) \land d_\textnormal{H}(-\mathbf{\hat z}^\star, \mathbf{z}^\star)}{n},
\end{align}
where $d_\textnormal{H}(\mathbf{z}, \mathbf{z}') = \sum_{i\in[n]}\mathbf{1}{\mathbf{z}_i\neq \mathbf{z}_i'}$ is the Hamming distance between $\mathbf{z}$ and $\mathbf{z}'$, and the minimum is taken because $\mathbf{\hat z}^\star$ and $-\mathbf{\hat z}^\star$ give rise to the same partitioning of nodes.
Similarly, for an estimator $\hzzz{\ell}$ of $\zzz{\ell}$, we evaluate its performance by ${\mathcal{L}}(\hzzz{\ell}, \zzz{\ell})$.
\paragraph{An idealized setup.}
To characterize the information-theoretic limits in estimating $\mathbf{z}^\star$ and $\zzz{\ell}$'s, we will consider an idealized setup as follows. Suppose $n = 2m+1$ and the nodes are labeled as $0 , 1, \hdots, 2m$. Consider a global assignment vector $\mathbf{z}^\star$ whose value on $i\in\{1, \hdots, 2m\}$ is known to us:
\begin{equation}
\label{eq:idealized_setup}
\mathbf{z}^\star_{i} =
\begin{cases}
+1 & 1\leq i\leq m\\
-1 & m+1\leq i\leq 2m.
\end{cases}
\end{equation}
We further observe $\{\AAA{\ell}\}_1^L \sim \textnormal{IMLSBM}(\mathbf{z}^\star, \rho, \{p_\ell\}_1^L, \{q_\ell\}_1^L)$. And our goal is to estimate $\mathbf{z}^\star_0$ as well as $\zzz{\ell}_0$'s, a substantially simplified task compared to the original one.
Such a strategy of ``reducing'' the task of doing inference for the whole parameter vector to doing inference for each coordinate is an instance of the celebrated Assouad's method \cite{assouad1983deux}, and has been successfully used in many recent works on characterizing the fundamental limits of community detection in SBMs and variants (see, e.g., \cite{zhang2016minimax,gao2017achieving,gao2018community}).
Our discussion in the rest of this section will largely rely on the intuitions built upon this idealized setup, and we refer the readers to Appendix \ref{append:prf_lower_bound} for fully rigorous proofs.
\subsection{Minimax Lower Bound for Global~Estimation}\label{subsec:lower_bound_z_star}
Consider the task of estimating $\mathbf{z}^\star$. Under the idealized setup \eqref{eq:idealized_setup}, our goal is to differentiate between $\mathbf{z}^\star_0 = +1$ and $\mathbf{z}^\star_0 = -1$ based on the data $\{\AAA{\ell}\}_1^L$, which gives rise to a binary hypothesis testing problem:
\begin{equation}
\label{eq:hypothesis_z_star_intractable}
H_0: \mathbf{z}^\star_0 = +1 \ \ \ \textnormal{v.s.} \ \ \ H_1: \mathbf{z}^\star_0 = -1.
\end{equation}
By Neyman--Pearson lemma, in principle, we can characterize the difficulty of the above testing problem by calculating the error made by the likelihood ratio test. However, due to the complicated structure of the likelihood function, this strategy is analytically intractable, calling for further simplifications.
\paragraph{The fundamental testing problem.}
If $\zzz{\ell}_0$ is actually observed by us, then since $\zzz{\ell}_0\sim2 \textnormal{Bern}(1-\rho)-1$ under $H_0$ and $\zzz{\ell}_0\sim 2\textnormal{Bern}(\rho)-1$ under $H_1$, deciding the value of $\mathbf{z}^\star_0$ from $\zzz{\ell}_0$ is equivalent to the problem of differentiating between
\begin{equation}
\label{eq:hypothesis_z_star_observed_zl}
H_0: 2 \textnormal{Bern}(1-\rho)-1 \ \ \ \textnormal{v.s.} \ \ \ H_1: 2\textnormal{Bern}(\rho)-1.
\end{equation}
In reality, we need to estimate each $\zzz{\ell}_0$ from its corresponding $\AAA{\ell}_0$. This is the community detection problem in a vanilla two-block SBM, whose fundamental difficulty is characterized by the following testing problem \citep{gao2017achieving}:
\begin{equation}
\label{eq:hypothesis_z_star_vanilla_sbm}
H_0: \bigotimes_{i\in[m]} \textnormal{Bern}(p_\ell) \otimes \bigotimes_{i\in[m]} \textnormal{Bern}(q_\ell) \ \ \ \textnormal{v.s.} \ \ \ H_1: \bigotimes_{i\in[m]} \textnormal{Bern}(q_\ell) \otimes \bigotimes_{i\in[m]} \textnormal{Bern}(p_\ell).
\end{equation}
Intuitively, if the signal strength in the $\ell$-th layer is strong enough, then we are close to the case of known $\zzz{\ell}_0$, in which the error for testing \eqref{eq:hypothesis_z_star_intractable} is mainly captured by that of testing \eqref{eq:hypothesis_z_star_observed_zl}. On the other hand, if we have barely any signal in the $\ell$-th layer, then we are in the unknown $\zzz{\ell}_0$ case, in which the error for testing \eqref{eq:hypothesis_z_star_intractable} mainly comes from testing \eqref{eq:hypothesis_z_star_vanilla_sbm}. This intuition is formalized in the following lemma.
\begin{lemma}[The fundamental testing problem for global~estimation]
\label{lemma:testing_problem_z_star}
Assume $L \leq n^c$ for some $c>0$. Then for any sequence $\delta_n=o(1)$, there exists another sequence $\delta'_n = o(1)$ satisfying $(1+\delta_n')n/2\in {\mathbb{N}}$, such that for any $S\subseteq [L]$,
\begin{equation}
\label{eq:lower_bound_by_testing_problem_z_star}
\inf_{\mathbf{\hat z}^\star} \sup_{\mathbf{z}^\star\in {\mathcal{P}}_n} {\mathbb{E}} {\mathcal{L}}(\mathbf{\hat z}^\star, \mathbf{z}^\star) \gtrsim \delta_n \inf_{\phi}\bigg({\mathbb{E}}_{H_0}[\phi] + {\mathbb{E}}_{H_1}[1-\phi] \bigg),
\end{equation}
where $\phi$ is a testing function of the following problem:
\begin{align}
& H_0: \bigg(\bigotimes_{\ell\in S^c} [2 \textnormal{Bern}(1-\rho) - 1]\bigg) \otimes \bigg(\bigotimes_{\ell\in S} \bigotimes_{i = 1}^{(1+\delta'_n)n/2} \textnormal{Bern}(p_\ell)\otimes \textnormal{Bern}(q_\ell)\bigg)\nonumber\\
\label{eq:fund_testing_problem_z_star}
& \qquad \textnormal{v.s. } H_1: \bigg(\bigotimes_{\ell\in S^c} [2 \textnormal{Bern}(\rho) - 1]\bigg) \otimes \bigg(\bigotimes_{\ell\in S} \bigotimes_{i = 1}^{(1+\delta'_n)n/2} \textnormal{Bern}(q_\ell)\otimes \textnormal{Bern}(p_\ell)\bigg).
\end{align}
\end{lemma}
\begin{proof}
See Appendix \ref{subappend:prf_testing_problem_z_star}.
\end{proof}
\begin{remark}
\label{rmk:opt_S_lower_bound_z_star}
Though the lower bound \eqref{eq:lower_bound_by_testing_problem_z_star} holds for an arbitrary $S\subseteq[L]$, by our previous intuition, it is the tightest if we choose $S$ such that $S^c$ corresponds to layers with high SNRs, where \eqref{eq:hypothesis_z_star_observed_zl} kicks in, and $S$ corresponds to layers with low SNRs, where the error from \eqref{eq:hypothesis_z_star_vanilla_sbm} dominates.
\end{remark}
\paragraph{Optimal testing error and parity of $\boldsymbol{|S^c|}$.}
By Neyman--Pearson lemma, the test that gives the optimal Type-I plus Type-II error is the likelihood ratio test with a cutoff of $1$. For the problem \eqref{eq:fund_testing_problem_z_star}, it is not hard to show that the optimal error is asymptotically proportional to
\begin{align}
\label{eq:error_of_LR_test_z_star}
{\mathbb{P}}\bigg(\sum_{\ell\in S} \log \bigg(\frac{q_\ell(1-p_\ell)}{p_\ell(1-q_\ell)}\bigg) \cdot \sum_{i=1}^{(1+\delta'_n)n/2} (\XXX{\ell}_i-\YYY{\ell}_i) \geq \log \bigg(\frac{1-\rho}{\rho}\bigg) \cdot \sum_{\ell\in S^c}\ZZZ{\ell}\bigg),
\end{align}
where
\begin{equation}
\label{eq:def_of_rvs_in_testing}
\{\XXX{\ell}_i\}_{i=1}^n \overset{\textnormal{i.i.d.}}{\sim} \textnormal{Bern}(p_\ell), \ \ \ \{\YYY{\ell}_i\}_{i=1}^n \overset{\textnormal{i.i.d.}}{\sim} \textnormal{Bern}(q_\ell), \ \ \ \{\ZZZ{\ell}\}_{\ell=1}^L \overset{\textnormal{i.i.d.}}{\sim} 2\textnormal{Bern}(1-\rho) - 1,
\end{equation}
all of which are mutually independent.
Readers with an expertise in {large deviation principles} may have noticed that \eqref{eq:error_of_LR_test_z_star} is the tail probability of a sum of independent random variables, and a tight characterization of this probability should involve the cumulant generating functions (CGFs) of those random variables as well as their rate functions (i.e., the Legendre transforms of CGFs).
This intuition explains the appearance of the two key information-theoretic quantities, namely $\psi_S^\star$ and $J_\rho$, in the definition of $\mathcal{I}_S$ in \eqref{eq:globinfo}. As one can check, $\psi_S(t)$ is precisely the CGF of the random variable that appears to the left of the ``$\geq$'' sign in \eqref{eq:error_of_LR_test_z_star}, and $-J_\rho$ is the CGF of $\log\big(\rho/(1-\rho)\big)\cdot \ZZZ{\ell}$ evaluated at $1/2$.
The following lemma gives the asymptotically optimal testing error for \eqref{eq:fund_testing_problem_z_star}.
\begin{lemma}[Optimal testing error for global~estimation]
\label{lemma:opt_test_error_z_star}
Assume $\rho = o(1)$ and $p_\ell \asymp q_\ell = o(1) , \forall\ell\in[L]$. Then there exists a sequence $\delta''_n=o(1)$ which is independent of $S$, such that the probability \eqref{eq:error_of_LR_test_z_star} is lower bounded by
\begin{equation}
\label{eq:opt_test_error_z_star}
C\cdot \exp\big\{- (1+\delta_n'')\mathcal{I}_S\big\},
\end{equation}
where $C>0$ is an absolute constant and ${\mathcal{I}}_S$ is defined in \eqref{eq:globinfo}.
\end{lemma}
\begin{proof}
See Appendix \ref{subappend:prf_opt_test_error_z_star}.
\end{proof}
\begin{remark}
\label{rmk:parity}
The SNR ${\mathcal{I}}_S$ for global~estimation, which appears on the exponent in the optimal testing error \eqref{eq:opt_test_error_z_star}, takes different forms according to the \emph{parity} of $|S^c|$. There is a fundamental reason for this.
It happens that the dominating term in the probability \eqref{eq:error_of_LR_test_z_star} is given by the part with $\sum_{\ell\in S^c}\ZZZ{\ell}$ being \emph{non-positive and closest to zero}. Since $\ZZZ{\ell}$'s are $\{\pm 1\}$-valued, such a requirement translates to $\sum_{\ell\in S^c}\ZZZ{\ell} = 0$ when $|S^c|$ is even, and gives $\sum_{\ell\in S^c}\ZZZ{\ell} = -1$ when $|S^c|$ is odd.
\end{remark}
\paragraph{Minimax lower bound for global~estimation.} Equipped with Lemmas \ref{lemma:testing_problem_z_star} and \ref{lemma:opt_test_error_z_star}, we are ready to present the main result in this subsection.
\begin{theorem}[Minimax lower bound for global~estimation]
\label{thm:lower_bound_z_star}
Assume $\rho = o(1)$, $p_\ell\asymp q_\ell = o(1) , \forall \ell\in[L]$ and $L\lesssim n^c$ for some $c \geq 0$. If
$
\min_{S\subseteq[L]} \mathcal{I}_S \to \infty,
$
then there exists a sequence $\underline{\delta}_n = o(1)$ such that
\begin{align}
\label{eq:lower_bound_z_star_vanishing}
& \inf_{\mathbf{\hat z}^\star} \sup_{\mathbf{z}^\star\in {\mathcal{P}}_n} {\mathbb{E}} {\mathcal{L}}(\mathbf{\hat z}^\star, \mathbf{z}^\star) \geq \exp\big\{-(1+\underline{\delta}_n) \min_{S\subseteq[L]}\mathcal{I}_S\big\}.
\end{align}
On the other hand, if $\min_{S\subseteq[L]} \mathcal{I}_S={\mathcal{O}}(1)$, then there exists some $c' >0$ such that
\begin{equation}
\label{eq:lower_bound_z_star_constant}
\inf_{\mathbf{\hat z}^\star} \sup_{\mathbf{z}^\star\in {\mathcal{P}}_n} {\mathbb{E}} {\mathcal{L}}(\mathbf{\hat z}^\star, \mathbf{z}^\star) \geq c'.
\end{equation}
\end{theorem}
\begin{proof}
If $\min_{S\subseteq[L]} \mathcal{I}_S\to\infty$, then we can find a sequence $\delta_n = o(1)$ such that
$
\log (\delta_n^{-1}) \ll
\min_{S\subseteq[L]} \mathcal{I}_S.
$
Then, invoking Lemmas \ref{lemma:testing_problem_z_star} and \ref{lemma:opt_test_error_z_star}, there exists $\delta_n' = o(1)$ such that
\begin{align*}
& \inf_{\mathbf{\hat z}^\star} \sup_{\mathbf{z}^\star\in {\mathcal{P}}_n} {\mathbb{E}} {\mathcal{L}}(\mathbf{\hat z}^\star, \mathbf{z}^\star)
\gtrsim \delta_n \exp\big\{-(1+\delta_n')\min_{S\subseteq[L]} \mathcal{I}_S\big\}
= \exp\big\{-(1+{\delta_n' + \delta_n''}) \min_{S\subseteq[L]} \mathcal{I}_S \big\},
\end{align*}
where
$
\delta_n'' = {\log({\delta_n}^{-1})}/{\min_{S\subseteq[L]} \mathcal{I}_S} = o(1)
$
by construction. Choosing $\underline{\delta}_n = \delta_n' + \delta_n''$ gives \eqref{eq:lower_bound_z_star_vanishing}.
On the other hand, if $\min_{S\subseteq[L]} \mathcal{I}_S={\mathcal{O}}(1)$, then repeating the above arguments gives $\inf_{\mathbf{\hat z}^\star} \sup_{\mathbf{z}^\star\in {\mathcal{P}}_n} {\mathbb{E}} {\mathcal{L}}(\mathbf{\hat z}^\star, \mathbf{z}^\star) \gtrsim \delta_n$ for any $o(1)$ sequence $\delta_n$.
If $\inf_{\mathbf{\hat z}^\star} \sup_{\mathbf{z}^\star\in {\mathcal{P}}_n} {\mathbb{E}} {\mathcal{L}}(\mathbf{\hat z}^\star, \mathbf{z}^\star)$ is itself $o(1)$, then we would have
$$
\inf_{\mathbf{\hat z}^\star} \sup_{\mathbf{z}^\star\in {\mathcal{P}}_n} {\mathbb{E}} {\mathcal{L}}(\mathbf{\hat z}^\star, \mathbf{z}^\star) \gtrsim \sqrt{\inf_{\mathbf{\hat z}^\star} \sup_{\mathbf{z}^\star\in {\mathcal{P}}_n} {\mathbb{E}} {\mathcal{L}}(\mathbf{\hat z}^\star, \mathbf{z}^\star)},
$$
a contradiction. Hence \eqref{eq:lower_bound_z_star_constant} follows.
\end{proof}
As an immediate corollary, we have the following result for the homogeneous case $(\rho=0)$.
\begin{corollary}[Minimax lower bound for global~estimation~under homogeneity]
\label{cor:lower_bound_z_star_hom}
Under the setup of Theorem \ref{thm:lower_bound_z_star}, assume in addition that $\rho=0$. If
$
n\sum_{\ell\in[L]} \III{\ell}_{1/2} \to \infty,
$
then there exists a sequence $\underline{\delta}_n = o(1)$ such that
\begin{align}
\label{eq:lower_bound_z_star_vanishing_hom}
& \inf_{\mathbf{\hat z}^\star} \sup_{\mathbf{z}^\star\in {\mathcal{P}}_n} {\mathbb{E}} {\mathcal{L}}(\mathbf{\hat z}^\star, \mathbf{z}^\star) \geq \exp\big\{-(1+\underline{\delta}_n) \frac{n}{2}\sum_{\ell\in[L]} \III{\ell}_{1/2}\big\}.
\end{align}
On the other hand, if $n\sum_{\ell\in[L]} \III{\ell}_{1/2} = {\mathcal{O}}(1)$, then there exists some $c' >0$ such that
\begin{equation}
\label{eq:lower_bound_z_star_constant_hom}
\inf_{\mathbf{\hat z}^\star} \sup_{\mathbf{z}^\star\in {\mathcal{P}}_n} {\mathbb{E}} {\mathcal{L}}(\mathbf{\hat z}^\star, \mathbf{z}^\star) \geq c'.
\end{equation}
\end{corollary}
\begin{proof}
This follows from the fact that if $\rho = 0$, then $J_\rho=\infty$ and thus the set $S$ that minimizes $\mathcal{I}_S$ is $S=[L]$.
\end{proof}
\subsection{Minimax Lower Bound for Individualized~Estimation}
We now derive minimax lower bound for estimating individual $\zzz{\ell}$'s. Let us again consider the idealized setup \eqref{eq:idealized_setup}.
\paragraph{Two testing problems from two sources of errors.} Suppose that we additionally know the value of $\mathbf{z}^\star_0$, say $\mathbf{z}^\star_0 = +1$.
Since $\zzz{\ell}_0$ is independent of $\zzz{-\ell}_0:= \{\zzz{r}_0: r\neq \ell\}$, the only information that's useful in determining $\zzz{\ell}_0$ comes from the following ``label sampling'' model:
\begin{equation}
\label{eq:rho_flipping_model}
\zzz{\ell}_i \sim \mathbf{z}^\star_i\times \big[2\text{Bern}(1-\rho)-1\big], \qquad \AAA{\ell}\sim \textnormal{SBM}(\zzz{\ell}, p_\ell, q_\ell).
\end{equation}
Now, for any estimator $\hzzz{\ell}_0$ of $\zzz{\ell}_0$, the error probability reads
$$
{\mathbb{P}}(\hzzz{\ell}_0 \neq \zzz{\ell}_0) = (1-\rho) \cdot {\mathbb{P}}(\hzzz{\ell}_0 = -1 ~|~ \zzz{\ell}_0 = +1) + \rho \cdot {\mathbb{P}}(\hzzz{\ell}_0 = +1 ~|~ \zzz{\ell}_0 = -1),
$$
which can be regarded as the $(1-\rho)\times \textnormal{type-I error} + \rho \times \textnormal{type-II error}$ of testing $H_0: \zzz{\ell}_0 = +1$ v.s. $H_1: \zzz{\ell}_0 = -1$ in a vanilla two-block SBM, which is almost equivalent to \eqref{eq:hypothesis_z_star_vanilla_sbm}. This intuition is formalized by the following lemma.
\begin{lemma}[The fundamental testing problem for individualized~estimation, Part \RN{1}]
\label{lemma:testing_problem_z_l_part_I}
Assume $\rho =o(1)$ and $L \lesssim n^c$ for some $c\geq0$. Then there exists sequence $\delta_n = o(1)$ satisfying $(1+\delta_n)n/2 \in {\mathbb{N}}$, such that for any $\ell\in[L]$,
\begin{equation}
\label{eq:lower_bound_by_testing_problem_z_l_part_I}
\inf_{\hzzz{\ell}} \sup_{\mathbf{z}^\star\in {\mathcal{P}}_n} {\mathbb{E}} {\mathcal{L}}(\hzzz{\ell}, \zzz{\ell}) \gtrsim \inf_{\phi}\bigg((1-\rho)\cdot{\mathbb{E}}_{H_0}[\phi] + \rho\cdot{\mathbb{E}}_{H_1}[1-\phi] \bigg),
\end{equation}
where $\phi$ is a testing function of the following problem:
\begin{align}
\label{eq:fund_testing_problem_z_l_part_I}
& H_0: \bigotimes_{i = 1}^{(1+\delta_n)n/2} \textnormal{Bern}(p_\ell)\otimes \textnormal{Bern}(q_\ell)
\ \ \
\textnormal{v.s. } \ \ \
H_1: \bigotimes_{i = 1}^{(1+\delta_n)n/2} \textnormal{Bern}(q_\ell)\otimes \textnormal{Bern}(p_\ell).
\end{align}
\end{lemma}
\begin{proof}
See Appendix \ref{subappend:prf_testing_problem_z_l_part_I}.
\end{proof}
Recall that Lemma \ref{lemma:testing_problem_z_l_part_I} reflects the situation when $\mathbf{z}^\star_0$ is known to us. In practice, we need to estimate $\mathbf{z}^\star_0$ from the data, giving rise to a testing problem similar to the one presented in Lemma \ref{lemma:testing_problem_z_star}.
\begin{lemma}[The fundamental testing problem for individualized~estimation, Part \RN{2}]
\label{lemma:testing_problem_z_l_part_II}
Assume $\rho = o(1)$ and $L \lesssim n^c$ for some $c\geq0$. Then for any sequence $\delta_n = o(1)$, there exists another sequence $\delta_n' = o(1)$ satisfying $(1+\delta_n)n/2 \in {\mathbb{N}}$, such that for any $\ell\in[L]$ and any $S\subseteq[L]\setminus\{\ell\}$, we have
\begin{equation}
\label{eq:lower_bound_by_testing_problem_z_l_part_II}
\inf_{\hzzz{\ell}} \sup_{\mathbf{z}^\star\in {\mathcal{P}}_n} {\mathbb{E}} {\mathcal{L}}(\hzzz{\ell}, \zzz{\ell}) \gtrsim \delta_n \inf_{\phi}\bigg({\mathbb{E}}_{H_0}[\phi] + {\mathbb{E}}_{H_1}[1-\phi] \bigg),
\end{equation}
where $\phi$ is a testing function of the following problem:
\begin{align}
& H_0: \bigg(\bigotimes_{r\in (S\cup\{\ell\})^c} [2 \textnormal{Bern}(1-\rho) - 1]\bigg) \otimes \bigg(\bigotimes_{r\in S\cup\{\ell\}} \bigotimes_{i = 1}^{(1+\delta'_n)n/2} \textnormal{Bern}(p_\ell)\otimes \textnormal{Bern}(q_\ell)\bigg)\nonumber\\
\label{eq:fund_testing_problem_z_l_part_II}
& \qquad \textnormal{v.s. } H_1: \bigg(\bigotimes_{r\in (S\cup\{\ell\})^c} [2 \textnormal{Bern}(\rho) - 1]\bigg) \otimes \bigg(\bigotimes_{r\in S\cup\{\ell\}} \bigotimes_{i = 1}^{(1+\delta'_n)n/2} \textnormal{Bern}(q_\ell)\otimes \textnormal{Bern}(p_\ell)\bigg).
\end{align}
\end{lemma}
\begin{proof}
See Appendix \ref{subappend:prf_testing_problem_z_l_part_II}.
\end{proof}
The testing problem in \eqref{eq:fund_testing_problem_z_l_part_II} differs from the one in \eqref{eq:fund_testing_problem_z_star} in that the layer $\ell$ is never involved in the term regarding $\textnormal{Bern}(1-\rho)$ and $\textnormal{Bern}(\rho)$. This makes sense, because according to our intuition in Section \ref{subsec:lower_bound_z_star}, this term reflects the case when $\zzz{\ell}$ is (nearly) known to us, which can never happen since $\zzz{\ell}$ itself is the estimating target.
To characterize the optimal testing error for the two testing problems given in Lemmas \ref{lemma:testing_problem_z_l_part_I} and \ref{lemma:testing_problem_z_l_part_II}, apart from the SNR for global~estimation~$\mathcal{I}_S$ defined in \eqref{eq:globinfo}, we additionally define the corresponding \emph{SNR for individualized~estimation}:
\begin{equation}
\label{eq:indinfo}
\mathcal{J}_S:=
\begin{cases}
|S| J_\rho + \psi^\star_S(0) & \textnormal{ if } |S| \textnormal{ is even}, \\
(|S| + 1) J_\rho + \psi^\star_S(-2J_\rho) & \textnormal{ if } |S| \textnormal{ is odd}.
\end{cases}
\end{equation}
A careful analysis on the error incurred by the likelihood ratio test gives the following result.
\begin{lemma}[Optimal testing error for individualized~estimation]
\label{lemma:opt_test_error_z_l}
Assume $\rho = o(1)$ and $p_\ell \asymp q_\ell = o(1) , \forall\ell\in[L]$. Then there exists a sequence $\delta''_n=o(1)$ such that for any $\ell\in[L], S \subseteq[L]$, the optimal $(1-\rho)\times \textnormal{type-I error} + \rho\times \textnormal{type-II error}$ of the testing problem in \eqref{eq:fund_testing_problem_z_l_part_I} is lower bounded by
\begin{equation}
\label{eq:opt_test_error_z_l_part_I}
C\cdot \exp\big\{ - (1+\delta_n'')\mathcal{J}_{\{\ell\}}\big\} ,
\end{equation}
and the optimal type-I plus type-II error of the testing problem in \eqref{eq:fund_testing_problem_z_l_part_II} is lower bounded by
\begin{equation}
\label{eq:opt_test_error_z_l_part_II}
C\cdot \exp\big\{-(1+\delta_n'')\mathcal{I}_{S\cup\{\ell\}}\big\} ,
\end{equation}
where $C>0$ is an absolute constant.
\end{lemma}
\begin{proof}
See Appendix \ref{subappend:prf_opt_test_error_z_l}.
\end{proof}
For the same reason as explained in Remark \ref{rmk:parity}, the optimal testing error for \eqref{eq:fund_testing_problem_z_l_part_II} depends on the parity of $|(S\cup\{\ell\})^c|$.
\paragraph{Minimax lower bound for individualized~estimation.} We are now ready to present the main result in this subsection.
\begin{theorem}[Minimax lower bound for individualized~estimation]
\label{thm:lower_bound_z_l}
Under the same setup as Theorem \ref{thm:lower_bound_z_star},
if for a fixed $\ell\in[L]$, it holds that
$
\min_{S\subseteq[L]\setminus\{\ell\}} \mathcal{I}_{S\cup\{\ell\}} \land \mathcal{J}_{\{\ell\}} \to \infty,
$
then there exists a sequence $\underline{\delta}_n = o(1)$, independent of $\ell$, such that
\begin{align}
\label{eq:lower_bound_z_l_vanishing}
& \inf_{\hzzz{\ell}} \sup_{\mathbf{z}^\star\in {\mathcal{P}}_n} {\mathbb{E}} {\mathcal{L}}(\hzzz{\ell}, \zzz{\ell}) \geq \exp\big\{ -(1+\underline{\delta}_n) \min_{S\subseteq[L]\setminus\{\ell\}} \mathcal{I}_{S\cup\{\ell\}} \big\} + \exp\big\{-(1+\underline{\delta}_n) \mathcal{J}_{\{\ell\}}\big\}.
\end{align}
On the other hand, if $\min_{S\subseteq[L]\setminus\{\ell\}} \mathcal{I}_{S\cup\{\ell\}} \land \mathcal{J}_{\{\ell\}}={\mathcal{O}}(1)$, then there exists $c'>0$ such that
\begin{equation}
\label{eq:lower_bound_z_l_constant}
\inf_{\hzzz{\ell}} \sup_{\mathbf{z}^\star\in {\mathcal{P}}_n} {\mathbb{E}} {\mathcal{L}}(\hzzz{\ell}, \zzz{\ell})\geq c'.
\end{equation}
\end{theorem}
\begin{proof}
Given Lemma \ref{lemma:testing_problem_z_l_part_I}, \ref{lemma:testing_problem_z_l_part_II} and \ref{lemma:opt_test_error_z_l}, the proof is essentially the same as the proof of Theorem \ref{thm:lower_bound_z_star}, and we omit the details.
\end{proof}
Under homogeneity, we have $\zzz{\ell}=\mathbf{z}^\star ,\forall\ell\in[n]$, and the lower bound in the above theorem should coincide with \eqref{eq:lower_bound_z_star_vanishing_hom}.
Indeed,
when $\rho = 0$, the only way to make the exponent finite is to choose $S = [L]\setminus\{\ell\}$, in which case we have
$$
\min_{S\subseteq [L]\setminus\{\ell\}}\mathcal{I}_{S\cup\{\ell\}} \land \mathcal{J}_{\{\ell\}} = \mathcal{I}_{[L]}\land \mathcal{J}_{\{\ell\}} = \psi_{[L]}^\star(0)\land \infty = \psi_{[L]}^\star(0),
$$
and hence Corollary \ref{cor:lower_bound_z_star_hom} can be alternatively derived from Theorem \ref{thm:lower_bound_z_l}.
Based on the intuitions built from Lemmas \ref{lemma:testing_problem_z_l_part_I} and \ref{lemma:testing_problem_z_l_part_II}, the interpretations of the two terms in the lower bound \eqref{eq:lower_bound_z_l_vanishing} should be clear: $\mathcal{J}_{\{\ell\}}$ is the error incurred by the label sampling model \eqref{eq:rho_flipping_model}, which we cannot avoid even if we know the ground truth $\mathbf{z}^\star$, whereas $\min_{S\subseteq[L]\setminus\{\ell\}}\mathcal{I}_{S\cup\{\ell\}}$ represents the error incurred by empirically estimating $\mathbf{z}^\star$.
\section{A Two-Stage Algorithm}\label{sec:alg}
Recall that the IMLSBM~is a hierarchical model, where the individual assignments $\zzz{\ell}$'s are independent realizations from the ``prior'' distribution \eqref{eq:label_flip} which is parametrized by $\mathbf{z}^\star$ and $\rho$.
We start by writing down the posterior density, which is proportional to
\begin{align}
& \prod_{\ell\in[L]} \prod_{i\in[n]} (1-\rho)^{\indicator{\zzz{\ell}_i = \mathbf{z}^{\star}_i}} \rho^{\indicator{\zzz{\ell}_i \neq \mathbf{z}^{\star}_i}} \nonumber\\
\label{eq:posterior}
& ~~\times \prod_{\ell\in[L]} \prod_{i< j} \bigg(p_\ell^{\AAA{\ell}_{ij}} (1-p_\ell)^{1 - \AAA{\ell}_{ij}} \cdot \indicator{\zzz{\ell}_i = \zzz{\ell}_j} + q_\ell^{\AAA{\ell}_{ij}} (1-q_\ell)^{1-\AAA{\ell}_{ij}} \cdot \indicator{\zzz{\ell}_i \neq \zzz{\ell}_j} \bigg).
\end{align}
Computing the vanilla MAP estimator requires searching over a discrete set with cardinality $2^{n(L+1)}$, a hopeless task for even moderately-sized $n$ and $L$.
Now, supposed that for a fixed $i\in[n]$, we are given a collection of estimators $\{\tzzz{\ell}_{j}: \ell\in[n], j\neq i\}$ for the individual assignments $\{\zzz{\ell}_j: \ell\in[n], j\neq i\}$. On the event that $\tzzz{\ell}_j$'s ($j\neq i$) agree with the ground truth parameters, the posterior density given in \eqref{eq:posterior}, as a function of $(\mathbf{z}^\star_i, \{\zzz{\ell}_i\}_{\ell=1}^L)$, reduces to a constant multiple of
\begin{align}
& \prod_{\ell\in[L]} (1-\rho)^{\indicator{\zzz{\ell}_i = \mathbf{z}^{\star}_i}} \rho^{\indicator{\zzz{\ell}_i \neq \mathbf{z}^{\star}_i}} \nonumber\\
\label{eq:posterior_reduced}
& ~~\times \prod_{\ell\in[L]} \prod_{j\neq i} \bigg(p_\ell^{\AAA{\ell}_{ij}} (1-p_\ell)^{1 - \AAA{\ell}_{ij}} \cdot \indicator{\zzz{\ell}_i = \tzzz{\ell}_j} + q_\ell^{\AAA{\ell}_{ij}} (1-q_\ell)^{1-\AAA{\ell}_{ij}} \cdot \indicator{\zzz{\ell}_i \neq \tzzz{\ell}_j} \bigg).
\end{align}
With some algebra, one finds that maximizing the above display over $(\zzz{\star}_i, \{\zzz{\ell}_i\}_{\ell=1}^L)$ is equivalent to maximizing the following objective function:
\begin{align}
\label{eq:log_posterior_reduced}
\sum_{\ell\in[L]} \bigg\{\log\bigg(\frac{1-\rho}{\rho}\bigg) \cdot \indicator{\zzz{\ell}_i = \mathbf{z}^{\star}_i} + \sum_{\substack{j\neq i :\tzzz{\ell}_j = \zzz{\ell}_i}} \bigg[ \log\bigg(\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg) \AAA{\ell}_{ij} + \log \bigg(\frac{1-p_\ell}{1-q_\ell}\bigg)\bigg]\bigg\}.
\end{align}
This is already simpler than the original one of maximizing \eqref{eq:posterior}, because the search space now has cardinality $2^{L+1} \ll 2^{n(L+1)}$.
A closer look at \eqref{eq:log_posterior_reduced} reveals that this function can be maximized in linear (in $L$) time.
Indeed, if we fix $\mathbf{z}^{\star}_i$, the problem of seeking for optimal $\zzz{\ell}$'s is \emph{decoupled} into $L$ subproblems. That is, it suffices to maximize
\begin{align}
\label{eq:log_posterior_decoupled}
\log\bigg(\frac{1-\rho}{\rho}\bigg) \cdot \indicator{\zzz{\ell}_i = \zzz{\star}_i} + \sum_{\substack{j\neq i :\tzzz{\ell}_j = \zzz{\ell}_i}}\bigg[ \log\bigg(\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg) \AAA{\ell}_{ij} + \log \bigg(\frac{1-p_\ell}{1-q_\ell}\bigg)\bigg]
\end{align}
for each $\ell\in[L]$. Note that each subproblem can be efficiently solved, since one only needs to search over a space with cardinality two (i.e., $\zzz{\ell}_i\in\{\pm 1\}$). Thus, to obtain the global maximizer of \eqref{eq:log_posterior_reduced}, one can proceed as follows:
\begin{enumerate}
\item Solve $L$ subproblems \eqref{eq:log_posterior_decoupled} with $\mathbf{z}^{\star}_i = + 1$, and record the objective value of \eqref{eq:log_posterior_reduced};
\item Repeat Step 1 with $\mathbf{z}^{\star}_i = -1$;
\item Obtain the global maximizer of \eqref{eq:log_posterior_reduced} by comparing the two objective values in the previous two steps.
\end{enumerate}
The foregoing discussion shows that the MAP estimator of $(\mathbf{z}^{\star}_i, \{\zzz{\ell}_i\}_{\ell = 1}^L)$ can be efficiently computed, provided \emph{the remaining parameters are given}.
This observation motivates the main algorithm of this paper, which is a two-stage procedure that first obtains initial estimators of $\{\zzz{\ell}\}_{1}^L$ via spectral clustering, and then refines the initial estimators in a node-wise fashion using MAP estimation.
\subsection{Stage \texorpdfstring{\RN{1}}{I}: Initialization via Spectral Clustering}
While our analysis in Section \ref{sec:minimax_rate} reveals that any consistent initialization would work, we will focus on a specific initialization scheme in this subsection: spectral clustering.
If $\rho$ is of order $o(1)$, then the proportion of flips in $\zzz{\ell}$ from $\mathbf{z}^{\star}$ will also be of order $o(1)$ with high probability.
Hence, as long as a consistent estimator $\mathbf{\tilde z}^{\star}$ of the global assignment $\mathbf{z}^{\star}$ is given, consistent individualized~estimation~is automatic by setting $\tzzz{\ell} = \mathbf{\tilde z}^{\star}$. In the rest of this subsection, we restrict ourselves to global~estimation.
Let $\boldsymbol{\omega} = (\omega_1, \hdots, \omega_L)$ be an arbitrary positive (i.e., $\omega_\ell > 0,\forall \ell\in[L]$) weight vector, and let us consider the following weighted adjacency matrix
\begin{equation}
\label{eq:wted_adjacency}
\bar A := \sum_{\ell\in [L]} \omega_\ell \AAA{\ell}.
\end{equation}
In the case of $\rho = 0$, one readily checks that all the information in $\mathbf{z}^{\star}$ is contained in the top two eigenvectors of ${\mathbb{E}}[\bar A]$.
A natural proposal is then to take top two eigenvectors of $\bar A$, and apply $k$-means clustering to them.
In the case of a small $\rho > 0$, we expect spectral clustering to continue to work well for estimating $\mathbf{z}^{\star}$, provided it exhibits a certain level of \emph{stability} to the additional ``noise'' induced by $\rho$. Our later analysis in Section \ref{subsec:init} shows that this is indeed the case.
\begin{algorithm}[t]
\KwIn{Adjacency matrices $\{\AAA{\ell}\}_1^L$, weight vector $\boldsymbol{\omega}$, intra-cluster connecting probabilities $\{{p_{\ell}}\}_1^L$, trimming parameter $\gamma > 1$}
\KwOut{Initial global estimator $\mathbf{\tilde z}^{\star}$}
Identify nodes $I\subseteq[n]$ such that $\sum_{j\in[n]}{\bar{A}}_{ij} > \gamma n \sum_{\ell\in[L]}\omega_\ell p_\ell , \forall i\in I$\;
For any $i\in (I\times[n])\cup ([n]\times I)$, set ${\bar{A}}_{ij}$ to zero and call the resulting matrix $\tau(\bar A)$\;
Compute $U\in\mathbb{R}^{n\times2}$, the first two eigenvectors of $\tau(\bar A)$\;
Solve the $(1+\varepsilon)$-approximate $k$-means objective \eqref{equ: approx kmeans} on $U$ to get $\hat Z$ and set
$$
\mathbf{\tilde z}^\star_i = \mathbf{1}\{\hat Z_{i,1} = 1\} - \mathbf{1}\{\hat Z_{i,2} = 1\}
$$
for any $i\in[n]$\;
\Return{$\mathbf{\tilde z}^\star$}\;
\caption{Stage \RN{1}: initialization via spectral clustering}
\label{alg: specc}
\end{algorithm}
The overall initialization scheme is detailed in Algorithm \ref{alg: specc}. There are two subtleties in this algorithm. First, instead of applying spectral clustering to ${\bar{A}}$, we apply it to a {trimmed version}, $\tau({\bar{A}})$, which is obtained by setting the ``larger-than-average'' entries of ${\bar{A}}$ to zero. As shown in Section \ref{subsec:init}, such a trimming operation can significantly improve the concentration of $\bar A$, especially when the signal-to-noise ratio is low. The second subtlety is a computational one: since exactly solving the $k$-means objective is NP-hard, we instead find the solution of an $(1+\varepsilon)$-approximation of it \citep{kumar2004simple}. Specifically, letting $U\in {\mathbb{R}}^{n\times 2}$ be the top two eigenvectors of $\tau(\bar A)$, we seek for $(\hat Z, \hat X)$ such that
\begin{align}
\|\hat Z\hat X- U\|_F^2\leq(1+\varepsilon)\min_{Z,\, X}\|ZX-U\|_F^2,
\label{equ: approx kmeans}
\end{align}
where the the minimum is over all $n\times 2$ assignment matrix $Z$ (i.e., each row of $Z$ is a canonical basis of ${\mathbb{R}}^2$) and all $2\times 2$ matrix $X$. The initial estimator $\mathbf{\tilde z}^{\star} = \tzzz{\ell}$ is then taken to be the clustering induced by $\hat Z$.
\subsection{Stage \texorpdfstring{\RN{2}}{II}: Node-Wise Refinement via MAP Estimation}\label{subsec:refine}
According to our discussion in the previous subsection, once an initial global estimator $\mathbf{\tilde z}^{\star}$ is given, we can also take that to be the initial individualized estimator. Now, in view of the MAP objective functions \eqref{eq:posterior}---\eqref{eq:log_posterior_decoupled}, we propose to solve
\begin{equation}
\label{eq:joint_refinement}
(\mathbf{\hat z}^\star_i, \hzzz{1}_i, ..., \hzzz{L}_i) = \underset{\substack{s_\star\in\{\pm 1\}, \\ s_\ell \in \{\pm 1\} ~\forall\ell\in[L]}}{\arg\max} \sum_{\ell\in[L]} f^{(\ell)}_i(s_\star, s_\ell, \mathbf{\tilde z}^{\star}),
\end{equation}
where
\begin{equation}
\label{eq:local_map_objective}
f^{(\ell)}_i(s_\star, s_\ell, \mathbf{\tilde z}^{\star}) = \log\bigg(\frac{1-\rho}{\rho}\bigg) \cdot \indicator{s_\ell = s_\star} + \sum_{j\neq i: \mathbf{\tilde z}^{\star}_j = s_\ell} \bigg[\log\bigg(\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg) \AAA{\ell}_{ij} + \log \bigg(\frac{1-p_\ell}{1-q_\ell}\bigg)\bigg],
\end{equation}
By our discussion at the beginning of this section, the above optimization problem can be solved in linear time. A detailed description is given in Algorithm \ref{alg: generic_refinement}.
\begin{algorithm}[t]
\SetNoFillComment
\KwIn{Initial global estimator $\mathbf{\tilde z}^{\star}$, adjacency matrices $\{\AAA{\ell}\}_1^L$, connecting probabilities $\{(p_\ell, q_\ell)\}_1^L$, flipping probability $\rho$}
\KwOut {Global estimator $\mathbf{\hat z}^\star$, individualized estimator $\{\hzzz{\ell}\}_1^L$}
\For{$i = 1, \hdots, n$}{
\For{$\ell = 1, \hdots, L$}{
${\mathsf{z}}(\ell, +1) \leftarrow \mathop{\rm argmax}_{s\in\{\pm1\}} f^{(\ell)}_{i}(+1, s, \mathbf{\tilde z}^{\star})$\tcp*[r]{$f^{(\ell)}_i$ is defined in \eqref{eq:local_map_objective}}
${\mathsf{z}}(\ell, -1)\leftarrow \mathop{\rm argmax}_{s\in\{\pm1\}} f^{(\ell)}_{i}(-1, s, \mathbf{\tilde z}^{\star})$\;
}
$\mathbf{\hat z}^\star_i \leftarrow \mathop{\rm argmax}_{s_\star\in\{\pm 1\}} \sum_{\ell\in [L]} f^{(\ell)}_i\big(s_\star, {\mathsf{z}}(\ell, s_\star), \mathbf{\tilde z}^{\star}\big)$\tcp*[r]{final global estimator}
\For{$\ell=1, \hdots, L$}{
$\hzzz{\ell}_i \leftarrow {\mathsf{z}}(\ell, \mathbf{\hat z}^\star_i)$\tcp*[r]{final individualized estimators}
}
}
\Return{$\mathbf{\hat z}^\star, \{\hzzz{\ell}\}_1^L$}\;
\caption{Stage \RN{2}: Node-wise refinement via MAP estimation}
\label{alg: generic_refinement}
\end{algorithm}
We conclude this section by remarking that our proposed algorithm is naturally a {distributed} one: the two for loops in Algorithm \ref{alg: generic_refinement} can be easily parallelized.
\section{Performance of the Two-Stage Algorithm}
\label{sec:minimax_rate}
In this section, we present theoretical results on the two-stage algorithm introduced in Section \ref{sec:alg}. Specifically, the performance of spectral clustering is presented in Section \ref{subsec:init}, followed by an analysis of MAP-based refinement in Section \ref{subsec:theory_refine}. The minimax optimality of the two-stage algorithm is proved in Section \ref{subsec:minimax_optimality}.
Throughout this section, the high probability error bounds are uniform with respect to probability measures defined in the parameter space \eqref{eq:param_space}. In particular, the ``${\mathbb{P}}$'' symbol represents the probability after marginalizing over the realizations of the $\zzz{\ell}$'s.
\input{0401-spectralClustering}
\input{0402-refinement}
\input{0403-optimality}
\subsection{Performance of Spectral Clustering}\label{subsec:init}
In this subsection, we analyze theoretical properties of Algorithm \ref{alg: specc}.
An important degree of freedom in Algorithm \ref{alg: specc} is the choice of the weight vector $\boldsymbol{\omega}$, and it is restricted by the following assumption.
\begin{assump}[Balanced weights across layers]
\label{assump:balanced_wts}
Assume $\omega_\ell > 0 ,\forall \ell\in[L]$ and $\sum_{\ell\in[L]}\omega_\ell = 1$. Moreover, assume that there exist two absolute constants $c_0 > 0$ and $c_1 \geq 1$ such that the following two inequalities hold:
\begin{align}
\max_{\ell\in[L]} \{\omega_\ell\} \cdot \sum_{\ell\in[L]} \omega_\ell p_\ell \leq c_0 \sum_{\ell\in[L]} \omega_\ell^2 p_\ell , \qquad \max_{\ell\in[L]} \{\omega_\ell\} \cdot \sum_{\ell\in[L]} p_\ell \leq c_1 \sum_{\ell} \omega_\ell q_\ell.
\end{align}
\end{assump}
The above assumption essentially states that $\boldsymbol{\omega}$ should be relatively balanced across layers. In particular, if $\omega_\ell \asymp 1/L ,\forall \ell\in[L]$, then this assumption holds.
We are now ready to state the main theorem of this subsection. We emphasize that the following theorem does not require $\rho = o(1)$ or $\beta = 1+o(1)$.
\begin{theorem}[Performance of spectral clustering]
\label{thm:specc}
Let Assumption \ref{assump:balanced_wts} hold with $c_0>0, c_1\geq 1$. Let the input to Algorithm \ref{alg: specc} be an instance generated by an $\textnormal{IMLSBM}\in {\mathcal{P}}_n(\rho, \{p_\ell\}_1^L, \{q_\ell\}_1^L, \beta)$ with $L \lesssim n^{c}$ and
\begin{equation}
\label{eq:spectral_gap}
\frac{\bar p}{\bar p - \bar q} \leq c'\cdot \frac{2(1-2\rho)^2}{\beta - \beta^{-1}} + 2(\rho - \rho^2),
\end{equation}
where $c\geq 0, c'\in[0, 1)$ are two absolute constants and $\bar p = \sum_{\ell\in[L]} \omega_\ell p_\ell, \bar q = \sum_{\ell\in[L]} \omega_\ell q_\ell$ are the weighted averages of connecting probabilities.
Fix any $r\geq 1$ and choose the regularization parameter to be $\gamma > e^{c_1}$. Then, there exist constants $c_2 = c_2(\gamma)$ and $C = C(c, c', c_0 , c_1, r)$ such that with probability at least $1 - 11n^{-r} - c_2^{-n}$, the output $\mathbf{\tilde z}^{\star}$ of Algorithm \ref{alg: specc} satisfies
\begin{equation}
\label{eq:error_specc}
{\mathcal{L}}(\mathbf{\tilde z}^{\star}, \mathbf{z}^{\star}) \leq \frac{C(2+\varepsilon) (\Delta_1 + \Delta_2 + \Delta_3)}{n^2(1-2\rho)^4 (\bar p - \bar q)^2} ,
\end{equation}
where
\begin{align}
\label{eq:error_specc_1}
\Delta_1& = \big[\bar p -2\rho(1-\rho)(\bar p - \bar q)\big]^2, \qquad \Delta_2 = n\sum_{\ell\in[L]} \omega_\ell^2 p_\ell\\
\label{eq:error_specc_2}
\Delta_3& = \max_{\ell\in[L]} \{\omega_\ell^2 (p_\ell-q_\ell)^2\} \cdot (L^2\rho^2 + n^2L\rho + n\log n) + \rho^2(1-\rho)^2 (\bar p - \bar q)^2.
\end{align}
\end{theorem}
\begin{proof}
See Appendix \ref{prf:thm:specc}.
\end{proof}
\begin{remark}
Under our working assumption that $\beta = 1+o(1)$, the first term on the righthand side of \eqref{eq:spectral_gap} tends to infinity, and so the inequality holds if $p_\ell / (p_\ell - q_\ell)$ is uniformly bounded for all $\ell\in[L]$.
\end{remark}
In the upper bound \eqref{eq:error_specc}, the three terms $\Delta_{1}, \Delta_2, \Delta_3$ come from different sources. The first term, $\Delta_1$, is induced by our model assumption that the diagonal entries of the adjacency matrices $\{A^{(\ell)}\}_{\ell=1}^L$ are all zero (i.e., there are no self-loops). The latter two terms come from the fact that in our proof, we relate the misclustering error to the deviation (in spectral norm) of the trimmed weighted adjacency matrix $\tau(\bar A)$ from the expectation of ${\bar{A}}$. More explicitly, $\Delta_2$ is induced by the concentration of $\tau(\bar A)$ around ${\mathbb{E}}[{\bar{A}} ~|~ \{\zzz{\ell}\}_1^L]$, the \emph{conditional mean} of ${\bar{A}}$, conditioning on the realization of $\zzz{\ell}$'s, whereas $\Delta_3$ is induced by the concentration of ${\mathbb{E}}[\bar A ~|~ \{\zzz{\ell}\}_1^L]$ around the \emph{marginal mean} ${\mathbb{E}}[\bar A]$.
It turns out that bounding $\|\tau({\bar{A}}) - {\mathbb{E}} [{\bar{A}}]\|$ is closely related to bounding $\|\tau(\bar B) - {\mathbb{E}}[\bar B]\|$, where $\bar B = \sum_{\ell\in[L]}\omega_\ell \BBB{\ell}$ and $\BBB{\ell}$'s are $n\times n$ independent Bernoulli random matrices with independent $\text{Bern}(\ppp{\ell}_{ij})$ entries.
In order to have a tight control of $\|\bar B - {\mathbb{E}}[\bar B]\|$, we give a non-trivial generalization of the results in \cite{le2017concentration} to the multilayer setup in Appendix \ref{append:concentration}, which roughly states the following: if the weight vector is sufficiently ``balanced'', then with high probability, for the trimmed version of $\bar B$, we have
\begin{align}
\label{eq:main_concentration_reg}
\|\tau(\bar B) - {\mathbb{E}}[\bar B]\| \lesssim \sqrt{n\sum_{\ell\in[L]}\omega_\ell^2 \max_{i, j}\ppp{\ell}_{ij}},
\end{align}
and without any trimming operation, we have
\begin{align}
\label{eq:main_concentration_no_reg}
\|\bar B - {\mathbb{E}}[\bar B]\| \lesssim \sqrt{n\sum_{\ell\in[L]}\omega_\ell^2 \max_{i, j}\ppp{\ell}_{ij}} + \max_{\ell\in[L]}\{\omega_\ell\} \cdot \sqrt{\log n}.
\end{align}
If $\rho = 0$, then the conditional mean ${\mathbb{E}}[\bar A~|~\{\zzz{\ell}\}_1^L]$ coincides with the marginal mean ${\mathbb{E}}[{\bar{A}}]$, and thus \eqref{eq:error_specc} holds with $\Delta_3 = 0$.
Curious readers may wonder why the expression of $\Delta_3$ given in \eqref{eq:error_specc_2} does not vanish as $\rho$ tends to zero. In particular, there is an additive term of $n\log n$. This is related to an interesting phenomenon regarding the concentration of Bernoulli random matrices. It happens that the problem of bounding $\big\|{\mathbb{E}}[\bar A~|~\{\zzz{\ell}\}_1^L]- {\mathbb{E}}[{\bar{A}}]\big\|$ can be related to bounding $\|B - {\mathbb{E}}[B]\|$, where $B\in\{0, 1\}^{n\times n}$ has i.i.d. $\text{Bern}(\rho)$ entries. The expression of $\Delta_3$ in \eqref{eq:error_specc_2} is based on that
$
\|B - {\mathbb{E}} [B]\| \lesssim \sqrt{n\rho} + \sqrt{\log n}
$
with high probability, which does not vanish as $\rho$ tends to zero.
In fact, such a ``discontinuity at zero'' is unavoidable: it has been shown in \cite{krivelevich2003largest} that if $\rho \lesssim 1/n$, then with probability tending to one,
$
\|B - {\mathbb{E}}[B]\| = \left(1+o(1)\right) \sqrt{{\log n}/{\log\log n}},
$
which diverges as $n$ tends to infinity.
To have a better understanding on the magnitude of the bound \eqref{eq:error_specc}, let us choose $\omega_\ell = 1/L , \forall \ell\in[L]$ and consider the following scaling of the connecting probabilities:
\begin{equation}
\label{eq:hom_scaling_connecting_prob}
p_\ell = \frac{a\log n}{nL}, \qquad q_\ell = \frac{b\log n}{nL}, \qquad \forall \ell\in[L],
\end{equation}
where $a>b>0$ are two constants. With some algebra, it follows that
\begin{align*}
\Delta_1 \asymp \frac{a^2(\log n)^2}{n^2L^2}, ~~~~ \Delta_2 = \frac{a\log n}{L^2}, ~~~~ \Delta_3 \asymp \frac{(a-b)^2 (\log n)^2}{n^2L^2} \cdot \bigg(\rho^2 + \frac{n^2\rho}{L} + \frac{n\log n}{L^2}\bigg).
\end{align*}
It is clear that $\Delta\asymp \frac{\log n}{L^2} \cdot (1+ \frac{\rho \log n}{L})$, and from \eqref{eq:error_specc} we arrive at
\begin{equation}
\label{eq:error_specc_hom_scaling}
{\mathcal{L}}(\mathbf{\tilde z}^\star, \mathbf{z}^\star) \lesssim \frac{1}{(1-2\rho)^4 (a-b)^2} \cdot \bigg(\frac{1}{\log n} + \frac{\rho}{L}\bigg).
\end{equation}
In summary, consistent estimation of $\mathbf{z}^\star$ by spectral clustering is possible when the connecting probabilities are as small as $\Omega\big(\frac{\log n}{nL}\big)$.
Since ${\mathcal{L}}(\mathbf{\tilde z}^{\star}, \zzz{\ell}) \leq {\mathcal{L}}(\mathbf{\tilde z}^{\star}, \mathbf{z}^{\star}) + {\mathcal{L}}(\mathbf{z}^{\star}, \zzz{\ell})$, we can obtain performance guarantees of Algorithm \ref{alg: specc} for individualized~estimation~by bounding the number of flips at each layer, as detailed in the following corollary.
\begin{corollary}[Spectral clustering for individualized~estimation]
\label{cor:specc_ind_est}
Under the setup of Theorem \ref{thm:specc}, for any $\rho' \in(0, 1-\rho)$, with probability at least $1 - 11n^{-r} - c_2^{-n} - L e^{-n D_\textnormal{KL}(\rho+\rho'\|\rho)}$, we have
\begin{equation}
\label{eq:error_specc_ind_est}
\max_{\ell\in[L]} {\mathcal{L}}(\mathbf{\tilde z}^{\star}, \zzz{\ell}) \leq \frac{C(2+\varepsilon) (\Delta_1 + \Delta_2 + \Delta_3)}{n^2(1-2\rho)^4 (\bar p - \bar q)^2} + \rho + \rho'.
\end{equation}
where $D_\textnormal{KL}(p\|q)$ is the Kullback-Leibler divergence between $\textnormal{Bern}(p)$ and $\textnormal{Bern}(q)$.
\end{corollary}
\begin{proof}
Classical Chernoff--Hoeffding bound \cite{chernoff1952measure,hoeffding1963probability} gives that for each fixed $\ell\in[L]$,
\begin{equation}
\label{eq:large_dev_bern_sum}
{\mathbb{P}}\bigg({\mathcal{L}}(\mathbf{z}^{\star}, \zzz{\ell}) \geq \rho + \rho'\bigg) \leq e^{-n D_\textnormal{KL}(\rho + \rho' \| \rho)},
\end{equation}
and the desired result follows by taking a union bound over $\ell\in[L]$.
\end{proof}
It is well-known that \eqref{eq:large_dev_bern_sum}, which is obtained by computing the rate function of Bernoulli random variables, is asymptotically tight (see, e.g, \cite{varadhan1984large}). If the goal is merely to ensure consistency (i.e., $\rho' = o(1)$), then we can use the following standard weakening of \eqref{eq:large_dev_bern_sum}:
$$
{\mathbb{P}}\bigg( {\mathcal{L}}(\mathbf{z}^\star, \zzz{\ell}) \geq \rho + \rho'\bigg) \leq e^{-2n(\rho')^2}.
$$
Thus, \eqref{eq:error_specc_ind_est} holds with $\rho' = n^{-1/4}$ with probability at least $1-11n^{-r} - c_2^{-n} - L e^{-2\sqrt{n}}$, and one can bound $Le^{-2\sqrt{n}}\leq e^{-c_3 \sqrt{n}}$ for some absolute constant $c_3 > 0$ when $L\lesssim n^{c}$.
\subsection{Performance of MAP-Based Refinement}\label{subsec:theory_refine}
The refinement procedure as introduced in Section \ref{subsec:refine}, in its current form (Algorithm \ref{alg: generic_refinement}), is highly flexible in that no assumption is imposed on the initial estimator $\mathbf{\tilde z}^{\star}$ other than consistency.
While such a flexibility is favored in practice, it brings some unnecessary complications to its theoretical analysis.
In addition, for a fixed $i$, the initial estimators $\mathbf{\tilde z}^{\star}_{-i}$ may have arbitrary dependence structures with $\{\AAA{\ell}_{ij}: j\neq i, \ell\in[L]\}$, which makes the analysis intractable.
\begin{algorithm}[t]
\SetNoFillComment
\SetKwFunction{Initialize}{Initialize}
\KwIn{Adjacency matrices $\{\AAA{\ell}\}_1^L$, connection probabilities $\{(p_\ell, q_\ell)\}_1^L$, flipping probability $\rho$}
\KwOut {Global estimator $\mathbf{\hat z}^\star$, individualized estimator $\{\hzzz{\ell}\}_1^L$}
\tcc{\textcolor{blue}{Stage \RN{1}: Leave-one-out initialization}}
\For{$i = 1, \hdots, n$}{
$\tzzz{\star, -i} \leftarrow \boldsymbol{0}_n$\;
$\tzzz{\star, -i}_{-i}\leftarrow$ \Initialize\big($\{\AAA{\ell}_{-i, -i}, p_\ell, q_\ell\}_{\ell=1}^L, \rho$\big)\tcp*[r]{initial estimator of $\mathbf{\hat z}^\star_{-i}$}
}
Set $\tzzz{\ell, -i} \leftarrow \tzzz{\star,-i}, \forall \ell\in[L]$\;
\tcc{\textcolor{blue}{Stage \RN{2}: MAP-based refinement}}
\For{$i = 1, \hdots, n$}{
\For{$\ell = 1, \hdots [L]$}{
${\mathsf{z}}(\ell, +1) \leftarrow \mathop{\rm argmax}_{s\in\{\pm1\}} f^{(\ell)}_{i}(+1, s, \tzzz{\star, -i})$\;
${\mathsf{z}}(\ell, -1) \leftarrow \mathop{\rm argmax}_{s\in\{\pm1\}} f^{(\ell)}_{i}(-1, s, \tzzz{\star, -i})$\;
}
\tcp{final estimator of $\mathbf{\hat z}^\star_i$, not aligned}
$\tzzz{\star, -i}_{i} \leftarrow \mathop{\rm argmax}_{s_\star\in\{\pm 1\}} \sum_{\ell\in [L]} f^{(\ell)}_i \big(s_\star, {\mathsf{z}}(\ell, s_\star), \tzzz{\star, -i}\big)$\;
\For{$\ell = 1, \hdots, [L]$}{
$\tzzz{\ell, -i}_i \leftarrow {\mathsf{z}}(\ell, \tzzz{\star, -i}_i)$\tcp*[r]{final estimator of $\zzz{\ell}_i$, not aligned}
}
}
\tcc{\textcolor{blue}{Stage \RN{3}: Alignment}}
Set $\mathbf{\hat z}^{\star}_1 \leftarrow \tzzz{\star, -1}_{1}$ and $\hzzz{\ell}_1 \leftarrow \tzzz{\ell, -1}_i , \forall \ell\in[L]$\;
\For{$i =2,\hdots, n$}{
$\mathbf{\hat z}^\star_i = \mathop{\rm argmax}_{s_\star\in\{\pm 1\}} \# \bigg\{ \{j\in[n]: \tzzz{\star, -1}_j = s_\star\} \bigcap \{j\in[n]: \tzzz{\star, -i}_j = \tzzz{\star, -i}_i\} \bigg\}$\;
\For{$\ell = 1, \hdots ,L$}{
$\hzzz{\ell}_i = \mathop{\rm argmax}_{s\in\{\pm 1\}} \# \bigg\{\{j\in[n]: \tzzz{\ell, -1}_j = s\} \bigcap \{j\in[n]: \tzzz{\ell, -i}_j = \tzzz{\ell, -i}_i\} \bigg\}$\;
}
}
\Return{$\mathbf{\hat z}^\star, \{\hzzz{\ell}\}_1^L$}\;
\caption{A provable version of Algorithm \ref{alg: generic_refinement}}
\label{alg: provable_refinement}
\end{algorithm}
To facilitate the analysis, we propose a modified version as shown in Algorithm \ref{alg: provable_refinement}. Instead of taking an arbitrary initial estimator as input (as done in Algorithm \ref{alg: generic_refinement}), we consider a {leave-one-out} initialization scheme.
In Stage \RN{1}, for each fixed $i$, the initial estimator $\tzzz{\star, -i}$ of $\mathbf{z}^{\star}_{-i}$ are computed using only $\{\AAA{\ell}_{-i}\}_1^L:=\{\AAA{\ell}_{jk}: j, k\neq i, \ell \in[L]\}$, which ensures the conditional (on the realization of $\{\zzz{\ell}\}_1^L$) independence between $\{\tzzz{\ell, -i}\}_1^L$ and $\{\AAA{\ell}_{ij}: j\neq i, \ell\in[L]\}$, thus simplifying the analysis, though the final analysis still turns out to be highly nontrivial.
In Stage \RN{2}, for each $i\in[n]$, we conduct MAP-based refinement using the initial estimators $\{\tzzz{\ell, -i}_{-i}\}_{\ell=1}^L$ (which are all equal to $\tzzz{\star, -i}_{-i}$) of $\zzz{\ell}_{-i}$, and the ``diagonal slots'' $(\tzzz{\star, -i}_i, \{\tzzz{\ell, -i}_i\}_{\ell=1}^L)$ are all zeros before Stage \RN{2} by our construction.
These ``diagonal slots'' are then filled in by the refined estimators of $(\mathbf{z}^\star_i, \{\zzz{\ell}_i\}_{\ell=1}^L)$.
After Stage \RN{2}, it is temping to directly output $\mathbf{\hat z}^\star_i = \tzzz{\star, -i}_i$ and $\hzzz{\ell}_i = \tzzz{\ell, -i}_i$ as the final estimators.
However, a subtlety arises due to the leave-one-out initialization. Since the initial estimators $\{\tzzz{\star ,-i}_{-i}\}_{i=1}^n$ are not necessarily {aligned}, the refined estimators $\{\tzzz{\star ,-i}_{i}\}_{i=1}^n$ and $\{\tzzz{\ell, -i}_{i}: i\in[n], \ell\in[L]\}$ can have different {orientations}.
For example, it could happen that $\tzzz{\star, -1}_1$ is estimating $\mathbf{z}^{\star}_1$, but $\tzzz{\star, -2}_2$ is estimating $-\mathbf{z}^{\star}_{2}$.
This is where the extra Stage \RN{3} of Algorithm \ref{alg: provable_refinement} comes into play. By using an {alignment} procedure, all coordinates of $\mathbf{\hat z}^\star$ and $\{\hzzz{\ell}\}_1^L$ will have the same orientation with high probability.
We shall remark that Algorithm \ref{alg: provable_refinement} is mostly of theoretical interest, and similar strategies have appeared in \cite{gao2017achieving,gao2018community}.
Our simulation in Section \ref{sec:exp} indicates that the estimation accuracy of Algorithm \ref{alg: generic_refinement} is indistinguishable from that of Algorithm \ref{alg: provable_refinement}, while Algorithm \ref{alg: generic_refinement} is much faster in speed.
Such a near perfect match in accuracy between the two algorithms is itself an interesting phenomenon, which we leave for future work\footnote{A promising approach for analyzing Algorithm \ref{alg: generic_refinement} is the ``leave-one-out'' analysis such as that used \cite{ma2018implicit}.}.
Before we present the main result of this subsection, we introduce the following assumption on consistent initialization.
\begin{assump}[Consistent initialization]
\label{assump:consistent_init}
Assume the \textnormal{\texttt{Initialize}} procedure used in Algorithm \ref{alg: provable_refinement} takes an instance generated by an $\textnormal{IMLSBM} \in {\mathcal{P}}_n(\rho, \{p_\ell\}_1^L, \{q_\ell\}_1^L, \beta)$ as its input and outputs a $\mathbf{\tilde z}^{\star}$ satisfying
\begin{equation}
\label{eq:consistent_init}
{\mathbb{P}}\big({\mathcal{L}}(\mathbf{\tilde z}^\star, \mathbf{z}^\star) \geq \delta_{{\texttt{init}},n} \big) \lesssim n^{-(1+\varepsilon_{\texttt{init}})}
\end{equation}
for some $\delta_{{\texttt{init}},n}=o(1)$ and $\varepsilon_{\texttt{init}} > 0$.
\end{assump}
\paragraph{Performance for global~estimation.}
The performance of Algorithm \ref{alg: provable_refinement} for global~estimation~is given by the following theorem.
\begin{theorem}[Performance of MAP-based refinement for global~estimation]
\label{thm:refine_glob}
Let the input to Algorithm \ref{alg: provable_refinement} be an instance generated by an $\textnormal{IMLSBM} \in {\mathcal{P}}_n(\rho, \{p_\ell\}_1^L, \{q_\ell\}_1^L, \beta)$ satisfying $\rho = o(1)$, $p_\ell\asymp q_\ell = o(1) , \forall \ell\in[L]$, $\beta = 1+o(1)$ and $L\leq n^c$ for some $c > 0$.
Let Assumption \ref{assump:consistent_init} hold and assume that for any $\delta_n = o(1)$, the following holds:
\begin{align}
\label{eq:diverging_glob_snr_sum}
\lim_{n\to\infty } \sum_{S\subseteq[L]} e^{-(1-\delta_n) \mathcal{I}_S} =0,
\end{align}
where $\mathcal{I}_S$ is defined in \eqref{eq:globinfo}.
Then, there exist two sequences $\overline{\delta}_n, \overline{\delta}_n' = o(1)$ such that
\begin{equation}
\label{eq:refine_glob}
\lim_{n\to \infty} \inf_{\mathbf{z}^\star \in {\mathcal{P}}_n} {\mathbb{P}}\bigg[ {\mathcal{L}}(\mathbf{\hat z}^\star , \mathbf{z}^\star) \leq \bigg(\sum_{S\subseteq[L]} e^{-(1-\overline{\delta}_n)\mathcal{I}_S}\bigg)^{1-\overline{\delta}_n'} \bigg] = 1.
\end{equation}
\end{theorem}
\begin{proof}
See Appendix \ref{prf:thm:refine_glob}.
\end{proof}
Note that the lower bound given in \eqref{eq:lower_bound_z_star_vanishing} takes the form of the {maximum} of $2^L$ terms indexed by $S\subseteq[L]$, whereas the upper bound given in \eqref{eq:refine_glob} is a {summation} of $2^L$ terms. Our later analysis in Section \ref{subsec:minimax_optimality} shows that under slightly stronger conditions on the SNR, the upper and lower bounds match asymptotically.
Under homogeneity ($\rho = 0$), we have $\mathcal{I}_S = \infty$ for every $S$ but $S = [L]$. So we have
\begin{equation}
\label{eq:refine_glob_hom_informal}
\sum_{S\subseteq[L]} e^{-(1-\delta_n)\mathcal{I}_S} = e^{-(1-\delta_n)\mathcal{I}_{[L]}} = \exp\big\{-(1-\delta_n) \frac{n}{2}\sum_{\ell\in[L]} \III{\ell}_{1/2}\big\},
\end{equation}
and the upper bound in \eqref{eq:refine_glob} matches the lower bound provided by Corollary \ref{cor:lower_bound_z_star_hom}.
However, the derivation of \eqref{eq:refine_glob_hom_informal} is not fully rigorous, because the layer-wise objective function $\fff{\ell}_i$ defined in \eqref{eq:local_map_objective} becomes infinity when $\rho = 0$, which makes the optimization problem in \eqref{eq:joint_refinement} ill-defined.
To address this issue, let us note that when $\rho = 0$, the ``regularization term'' in $\fff{\ell}_i$, namely $\log\big((1-\rho)/\rho\big)\cdot \mathbf{1}{s_\ell = s_\star}$, essentially requires $s_\ell$ to \emph{exactly agree} with $s_\star$. Thus, we can shift from solving \eqref{eq:joint_refinement} to solving the following problem:
\begin{align}
\label{eq:mle_refine}
\mathbf{\hat z}^\star_i = \mathop{\rm argmax}_{s_\star \in\{\pm 1\}} \sum_{\ell\in[L]} \sum_{j\neq i: \mathbf{\tilde z}^{\star}_j = s_\star} \bigg[\log\bigg(\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg) \AAA{\ell}_{ij} + \log \bigg(\frac{1-p_\ell}{1-q_\ell}\bigg)\bigg].
\end{align}
With the above optimization formulation, Algorithm \ref{alg: provable_refinement} can be modified in a \emph{mutatis mutandis} fashion, and the upper bound in \eqref{eq:refine_glob_hom_informal} can be made rigorous, as detailed in the following corollary.
\begin{corollary}[Performance of MAP-based refinement for global~estimation~under homogeneity]
\label{cor:refine_glob_hom}
Consider again Algorithm \ref{alg: provable_refinement}, except that we change its Stage \RN{2} from MAP-based refinement \eqref{eq:joint_refinement} to maximum-likelihood-based refinement \eqref{eq:mle_refine}.
Let the input be an instance generated by an $\textnormal{IMLSBM} \in {\mathcal{P}}_n(\rho, \{p_\ell\}_1^L, \{q_\ell\}_1^L, \beta)$ satisfying $\rho = 0$, $p_\ell\asymp q_\ell = o(1) , \forall \ell\in[L]$, $\beta = 1+o(1)$ and $L\leq n^c$ for some $c > 0$.
Let Assumption \ref{assump:consistent_init} hold and assume $n\sum_{\ell\in[L]}\III{\ell}_{1/2} \to \infty$.
Then, there exists a sequence $\overline{\delta}_n= o(1)$ such that
\begin{equation}
\label{eq:refine_glob_hom}
\lim_{n\to \infty} \inf_{\mathbf{z}^\star\in{\mathcal{P}}_n}{\mathbb{P}}\bigg( {\mathcal{L}}(\mathbf{\hat z}^\star , \mathbf{z}^\star) \leq \exp\big\{-(1-\overline{\delta}_n) \frac{n}{2}\sum_{\ell\in[L]} \III{\ell}_{1/2}\big\} \bigg) = 1.
\end{equation}
\end{corollary}
\begin{proof}
The proof is a straightforward adaptation of the proof of Theorem \ref{thm:refine_glob}, and we omit the details.
\end{proof}
Minimax optimal algorithms for community detection in a homogeneous MLSBM~have appeared in the literature \cite{paul2016consistent,xu2020optimal}. The procedure in \cite{paul2016consistent} is based on exactly solving the maximum likelihood objective, which is computationally infeasible. The algorithm in \cite{xu2020optimal} is computable in polynomial-time and it operates on a variant of \textnormal{SBM}, called {weighted \textnormal{SBM}}, of which the homogeneous multilayer SBM is a special case. The corollary above gives another polynomial-time minimax optimal algorithm for community detection in homogeneous multilayer SBMs.
\paragraph{Performance for individualized~estimation.}
The performance guarantee of Algorithm \ref{alg: provable_refinement} for individualized~estimation~is given by the following theorem.
\begin{theorem}[Performance of MAP-based refinement for individualized~estimation]
\label{thm:refine_ind}
Let the input to Algorithm \ref{alg: provable_refinement} be an instance generated by an $\textnormal{IMLSBM} \in {\mathcal{P}}_n(\rho, \{p_\ell\}_1^L, \{q_\ell\}_1^L, \beta)$ satisfying $\rho = o(1)$, $p_\ell\asymp q_\ell = o(1) , \forall \ell\in[L]$, $\beta = 1+o(1)$ and $L\leq n^c$ for some $c > 0$.
Let Assumption \ref{assump:consistent_init} hold and assume that for a fixed $\ell\in[L]$ and for any $\delta_n = o(1)$, the following holds:
\begin{align}
\label{eq:diverging_ind_snr_sum}
\lim_{n\to\infty } \sum_{S\subseteq[L]\setminus\{\ell\}} \Big( e^{-(1-\delta_n) \mathcal{I}_{S\cup\{\ell\}}} + e^{-(1-\delta_n)\mathcal{J}_{S\cup\{\ell\}}} \Big) =0,
\end{align}
where $\mathcal{I}_{S\cup\{\ell\}}$ and $\mathcal{J}_{S\cup\{\ell\}}$ are defined in \eqref{eq:globinfo} and \eqref{eq:indinfo}, respectively.
Then, there exist two sequences $\overline{\delta}_n, \overline{\delta}_n' = o(1)$, independent of $\ell$, such that
\begin{equation}
\label{eq:refine_ind}
\lim_{n\to \infty}\inf_{\mathbf{z}^\star\in{\mathcal{P}}_n}{\mathbb{P}}\bigg[ {\mathcal{L}}(\hzzz{\ell} , \zzz{\ell}) \leq \bigg(\sum_{S\subseteq[L]\setminus\{\ell\}} \Big[ e^{-(1-\overline{\delta}_n)\mathcal{I}_{S\cup\{\ell\}}} + e^{-(1-\overline{\delta}_n) \mathcal{J}_{S\cup\{\ell\}}} \Big]
\bigg)^{1-\overline{\delta}_n'}
\bigg] = 1.
\end{equation}
\end{theorem}
\begin{proof}
See Appendix \ref{prf:thm:refine_ind}.
\end{proof}
In this upper bound, the terms involving $\mathcal{J}_{S\cup\{\ell\}}$'s come from estimating $\zzz{\ell}$ {given the knowledge of $\mathbf{z}^\star$} (i.e., error from the label sampling model defined in \eqref{eq:rho_flipping_model}), whereas the terms involving $\mathcal{I}_{S\cup\{\ell\}}$'s come from {empirically estimating $\mathbf{z}^\star$}.
Similar to Theorem \ref{thm:refine_glob}, the bound given in this theorem is a {summation} of $2\times 2^{L-1}$ terms, whereas the corresponding lower bound in \eqref{eq:lower_bound_z_l_vanishing} is the {maximum} of $2^{L-1} + 1$ terms. We will show in Section \ref{subsec:minimax_optimality} that the two bounds asymptotically coincide under slightly stronger assumptions on both global and individualized SNRs.
If $\rho = 0$, then $\mathcal{I}_{S\cup\{\ell\}}$ is infinity for every $S$ but $S = [L]\setminus \{\ell\}$. On the other hand, $\mathcal{J}_{S\cup\{\ell\}}$ is infinity for any $S \subseteq[L]\setminus\{\ell\}$. It follows that the upper bound in \eqref{eq:refine_ind} becomes
$$
e^{-\big(1-o(1)\big)\mathcal{I}_{[L]}} = \exp\big\{-\big(1-o(1)\big) \frac{n}{2}\sum_{\ell\in[L]} \III{\ell}_{1/2}\big\},
$$
which agrees with the upper bound in \eqref{eq:refine_glob_hom}.
This makes sense as global and individualized~estimation~coincide under homogeneity.
\subsection{Minimax Optimality}\label{subsec:minimax_optimality}
Recall that the two upper bounds in Theorems \ref{thm:refine_glob} and \ref{thm:refine_ind} are both {summations} of {exponentially many} (in $L$) terms indexed by some subset $S\subseteq[L]$, whereas the corresponding lower bounds in Theorems \ref{thm:lower_bound_z_star} and \ref{thm:lower_bound_z_l} are both {maxima} of that many terms. Thus, a priori, there is no reason to believe that the upper and lower bounds should match, especially when $L$ tends to infinity with $n$.
However, in this subsection, we show that this is indeed the case under mild regularity conditions, establishing {asymptotic minimaxity} of Algorithm \ref{alg: provable_refinement} for both global and individualized estimation.
\paragraph{Minimax optimality for global~estimation.}
Based on \eqref{eq:refine_glob}, a naive argument would upper bound $\sum_{S\subseteq[L]}e^{-(1-\overline{\delta}_n)\mathcal{I}_S}$ by $2^L \exp\{-(1-\overline{\delta}_n)\min_{S\subseteq[L]}\mathcal{I}_S\}$.
In order to match the lower bound \eqref{eq:lower_bound_z_star_vanishing}, we need to assume that
$
L \ll \min_{S\subseteq[L]}\mathcal{I}_S.
$
It turns out such a requirement on the growth rate of $L$ can be substantially relaxed, as detailed in the next theorem.
\begin{theorem}[Minimax optimality of MAP-based refinement for global~estimation]
\label{thm:opt_glob_est}
Consider the setup of Theorem \ref{thm:refine_glob}, except that instead of assuming \eqref{eq:diverging_glob_snr_sum}, we now assume
\begin{equation}
\label{eq:req_for_opt_glob_est}
\log L + L e^{-c'J_\rho} \ll \min_{S\subseteq[L]} \mathcal{I}_S \to \infty \qquad \textnormal{as } n\to\infty
\end{equation}
for some $c'\in(0, 1)$. Then, there exists a sequence $\overline{\delta}_n = o(1)$ such that
\begin{equation}
\label{eq:tight_upper_bound_z_star}
\lim_{n\to \infty}\inf_{\mathbf{z}^\star\in{\mathcal{P}}_n}{\mathbb{P}}\bigg( {\mathcal{L}}(\mathbf{\hat z}^\star , \mathbf{z}^\star) \leq \exp\big\{-(1-\overline{\delta}_n) \min_{S\subseteq[L]}\mathcal{I}_S\big\} \bigg) = 1.
\end{equation}
\end{theorem}
\begin{proof}
See Appendix \ref{prf:thm:opt_glob_est}.
\end{proof}
Consider the case where $\rho = n^{-c''}$ for some constant $c''>0$. In this case, we have $J_\rho = \left(1+o(1)\right) \frac{1}{2}\log \frac{1}{\rho} = \left(1+o(1)\right) \frac{c''}{2}\log n$, and thus the requirement in \eqref{eq:req_for_opt_glob_est} becomes $\log L + L n^{-\left(1+o(1)\right) c'c''/2}\ll \min_{S\subseteq[L]}\mathcal{I}_S$, a vast improvement of over the naive requirement of $L \ll \min_{S\subseteq[L]}\mathcal{I}_S$.
In the homogeneous case of $\rho = 0$, the requirement in \eqref{eq:req_for_opt_glob_est} becomes $\log L \ll \mathcal{I}_{[L]} = \frac{n}{2}\sum_{\ell\in[L]} \III{\ell}_{1/2}$.
In the proof, we need to identify the optimal $S\subseteq[L]$ such that $\mathcal{I}_S$ is minimized. While it is easy to do so when $\rho = 0$, this task turns out to be challenging in the presence of inhomogeneity, and our proof is based on a nontrivial application of a generalization of Von Neumann's minimax theorem \cite{neumann1928theorie} due to \citet{sion1958general}.
\paragraph{Minimax optimality for individualized~estimation.}
Similar to the case of global~estimation, in order to prove the tightness of \eqref{eq:refine_ind}, a naive argument would require $L \ll \big(\min_{S\subseteq[L]\setminus\{\ell\}}\mathcal{I}_{S\cup\{\ell\}}\big)\land \big(\min_{S\subseteq[L]\setminus\{\ell\}}\mathcal{J}_{S\cup\{\ell\}}\big)$, and this is relaxed in the following theorem.
\begin{theorem}[Minimax optimality of MAP-based refinement for individualized~estimation]
\label{thm:opt_ind_est}
Consider the setup of Theorem \ref{thm:refine_ind}, except that instead of assuming \eqref{eq:diverging_ind_snr_sum}, we now assume
\begin{equation}
\label{req_for_ind_est}
\log L + L e^{-c' J_\rho} \ll
\min_{S\subseteq[L]\setminus\{\ell\}} \mathcal{I}_{S\cup\{\ell\}}
\land \mathcal{J}_{\{\ell\}}\to \infty \qquad \textnormal{as } n \to \infty
\end{equation}
for some $c'\in(0, 1)$. Then, there exists a sequence $\overline{\delta}_n = o(1)$ such that
\begin{align}
\label{eq:tight_upper_bound_z_l}
\lim_{n\to \infty}\inf_{\mathbf{z}^\star\in{\mathcal{P}}_n}{\mathbb{P}}\bigg( {\mathcal{L}}(\hzzz{\ell} , \zzz{\ell}) \leq \exp\big\{-(1-\overline{\delta}_n) \min_{S\subseteq[L]\setminus\{\ell\}}\mathcal{I}_{S\cup\{\ell\}}\big\} + \exp\big\{-(1-\overline{\delta}_n)\mathcal{J}_{\{\ell\}}\big\}
\bigg) = 1.
\end{align}
\end{theorem}
\begin{proof}
See Appendix \ref{prf:thm:opt_ind_est}.
\end{proof}
In the proof, we identify that the optimal $S$ that minimizes $\mathcal{J}_{S\cup\{\ell\}}$ is precisely the empty set. On the other hand, the identification of the set that minimizes $\mathcal{I}_{S\cup\{\ell\}}$ is done in a similar fashion as in the proof of Theorem \ref{thm:opt_glob_est}.
\section{Numerical Experiments}\label{sec:exp}
In this section, we conduct simulation studies to corroborate our theoretical results. Since we focus on the symmetric case, we can without loss of generality assume $\mathbf{z}^\star_i = +1$ for $1\leq i\leq \lfloor n/2\rfloor$ and $\mathbf{z}^\star_i = -1$ for $\lfloor n/2\rfloor + 1\leq i\leq n$.
The $L$ layers are divided into three disjoint groups:
\begin{enumerate}
\item \emph{Weak layers.} For $1\leq \ell\leq \floor{0.3L}$, we let $p_\ell = {c}/(nL), q_\ell = 1/(nL)$ for some constant $c$ that controls the amount of information. Since $\III{\ell}_{1/2} \asymp 1/(nL)$, in view of the lower bound \eqref{eq:lower_bound_z_star_vanishing}, these layers, even when pooled together, cannot consistently estimate $\mathbf{z}^\star$.
\item \emph{Intermediate layers.} For $\floor{0.3L}+1\leq \ell \leq \floor{0.95 L}$, we let $p_\ell = c(\log n)/(nL), q_\ell = (\log n)/(nL)$ where $c$ is the same constant as that appears in the weak layers. Note that $\III{\ell}_{1/2} \asymp \log(n)/(nL)$. Thus, while each individual layer does not contain sufficient information for consistent estimation of its own $\zzz{\ell}$, consistent estimation of $\mathbf{z}^\star$ becomes possible if information is aggregated across these layers.
\item \emph{Strong layers.} For $\floor{0.95 L} + 1\leq \ell \leq L$, we let $p_\ell = c\log(n)/n, q_\ell = \log(n)/n$ again for the same $c$ as above. As $\III{\ell}_{1/2}\asymp \log(n)/n$, these layers are capable of consistently estimating their $\zzz{\ell}$'s, even when treated individually without aggregation.
\end{enumerate}
The rationale behind the above partition is to simulate the behaviors that are likely to appear in real world multilayer networks. For example, let us consider the case where layers are distinct
``participants''
collaborating with each other, with the hope that they can borrow information from others to better estimate their own $\zzz{\ell}$'s. Such a setup is also known as ``federated learning'' in the machine learning literature \cite{kairouz2019advances}. The intermediate layers are participants with the most incentive in the collaboration, as ``united they stand, divided they fall''. In comparison, the weak layers may not be as incentivized as the intermediate layers, because they would ``fall even when united''. Nevertheless, they may still want to participate as ``hitchhikers''. Finally, the strong layers are participants that would ``stand even when divided'', and the only reason for them to participate is the hope for even more accurate estimation of $\zzz{\ell}$'s. A small proportion of strong layers (0.05 in our case) reflects our belief that the strong layers are relatively scarce.
In this section, we will always use Algorithm \ref{alg: specc} as the initialization scheme. Our experiment in Appendix \ref{subappend:weight_choice} shows that using uniform weights (i.e., $\omega_\ell = 1/L,\forall \ell\in[L]$) and setting the trimming threshold $\gamma = 5$ work well in a wide range of scenarios, and we will always use such a hyperparameter choice in the following discussion.
\begin{figure}[t]
\begin{subfigure}{.5\textwidth}
\centering
\includegraphics[width=1\linewidth]{plot/symm_compare_algs_c=2.pdf}
\end{subfigure}
\begin{subfigure}{.5\textwidth}
\centering
\includegraphics[width=1\linewidth]{plot/symm_compare_algs_c=5.pdf}
\end{subfigure}
\caption{\small Comparison between estimation accuracies of the generic version (Algorithm \ref{alg: generic_refinement}) and the provable version (Algorithm \ref{alg: provable_refinement}) when $c = 2$ (left) and $c = 5$ (right). In each panel, the four groups from left to right correspond to global estimation and individualized estimation in weak, intermediate, and strong layers, respectively.}
\label{fig:compare_algs}
\end{figure}
\paragraph{Comparison between Algorithms \ref{alg: generic_refinement} and \ref{alg: provable_refinement}.}
Recall that we have developed two versions of the same algorithm: Algorithm \ref{alg: generic_refinement} is fast but we were not able to establish any theoretical guarantees, whereas Algorithm \ref{alg: provable_refinement} is slower but provably optimal.
We have argued that these two versions should perform similarly in terms of estimation accuracy, and we now empirically justify this claim.
We set $n = 200, L = 100, \rho = 0.1$ and let $c$ be either $2$ or $5$.
We then run the two algorithms over $500$ instances of the model (assuming $\{p_\ell\}, \{q_\ell\}, \rho$ are known) and record the misclustering proportions for global~estimation~(\texttt{Layer=Global}), individualized~estimation~in weak layers (\texttt{Layer=Weak}), intermediate layers (\texttt{Layer=Intermediate}) and strong layers (\texttt{Layer=Strong}).
The results are presented in Figure \ref{fig:compare_algs}. We see that the performances of the two versions are indeed similar, and they even become indistinguishable when $c=5$. Thus, in the rest of this section, we always use Algorithm \ref{alg: generic_refinement}.
\begin{figure}[t]
\centering
\begin{subfigure}{.45\textwidth}
\centering
\includegraphics[width=1.1\linewidth]{plot/symm_vary_rho.pdf}
\end{subfigure}
\begin{subfigure}{.45\textwidth}
\centering
\includegraphics[width=1.1\linewidth]{plot/symm_vary_rho_PC20.pdf}
\end{subfigure}
\vskip -0.33\baselineskip
\begin{subfigure}{.45\textwidth}
\centering
\includegraphics[width=1.1\linewidth]{plot/symm_vary_signal.pdf}
\caption{Our method: Algorithm \ref{alg: specc}+\ref{alg: generic_refinement}}
\label{fig:our_method}
\end{subfigure}
\begin{subfigure}{.45\textwidth}
\centering
\includegraphics[width=1.1\linewidth]{plot/symm_vary_signal_PC20.pdf}
\caption{Co-regularized spectral clustering}
\label{fig:PC20}
\end{subfigure}
\caption{\small Performance of our method (left panel) and co-regularized spectral clustering (right panel) \cite{kumar2011co,paul2020spectral} under different SNRs. The top two figures consider the setup where $c = 2$ is fixed and $\log(1/\rho)$ varies, and the bottom two figures consider the setup where $\rho=0.1$ is fixed and $c$ varies.
}
\label{fig:effect_of_snr}
\end{figure}
\paragraph{Effects of SNRs and comparison with co-regularized spectral clustering.}
Recall that the minimax rates for global~estimation~and individualized~estimation~both rely on two information theoretic quantities: $J_\rho = \left(1+o(1)\right) \cdot\frac{1}{2}\log(1/\rho)$ and $\III{\ell}_t = \left(1+o(1)\right)\cdot (c^t -1)(c^{1-t}-1)\times(\textnormal{scaling of } q_\ell)$. We now experiment on how these two quantities influence the performance of our proposed algorithm. We set $n = 1000, L = 100$, and we either fix $\rho = 0.1$ and vary $c$, or fix $c = 2$ and vary $\log(1/\rho)$.
We run Algorithm \ref{alg: generic_refinement} over $500$ instances of the model (again assuming $\{p_\ell\}, \{q_\ell\}, \rho$ are known) and record the misclustering proportions for both global~estimation~and individualized~estimation.
As a comparison, we implement the \emph{co-regularized spectral clustering} algorithm, a popular algorithm for clustering in multilayer networks originally proposed by \cite{kumar2011co} and later shown to be consistent in the $\rho=0$ case by \cite{paul2020spectral}.
Since co-regularized spectral clustering requires running multiple ``coordinate ascent'' steps, each involving computing the eigen-decomposition of an $n\times n$ matrix, it is substantially slower than our method, and we only run this algorithm for $100$ instances of the model.
We refer the readers to Appendix \ref{subappend:co_reg} for details on this algorithm.
Figure \ref{fig:effect_of_snr} shows the results of this simulation. We see that our method significantly outperforms co-regularized spectral clustering in all scenarios considered.
By the top-left plot in Figure \ref{fig:effect_of_snr}, the misclustering proportions of our method for both global~estimation~and individualized~estimation~tend to zero if we fix $c=2$ and increase the value of $J_\rho$.
In contrast, by the bottom-left plot in Figure \ref{fig:effect_of_snr},
for fixed $\rho = 0.1$, misclustering proportions for $\mathbf{z}^\star$ tend to zero as $c$ becomes large, which is as expected.
However, for individualized~estimation, while errors of strong layers still tend to zero,
errors of intermediate and weak layers both tend to $\rho = 0.1$.
This behavior actually is well explained by our theory.
Note that the minimax rate \eqref{eq:tight_upper_bound_z_l} for individualized~estimation~consists of two terms, where the first term represents the error from label sampling \eqref{eq:rho_flipping_model} and scales as $e^{-(1+o(1))\mathcal{J}_{\{\ell\}}} = \rho^{1+o(1)}\times e^{-(1+o(1))\psi^\star_{\{\ell\}} (-2J_\rho)}$, whereas the second term comes from empirically estimating $\mathbf{z}^\star$, which tends to zero much faster than the first term in the current setting.
Different behaviors of individualized estimation errors in strong layers and in intermediate/weak layers occur since $\psi^\star_{\{\ell\}}(-2J_\rho)$ is quite large when the $\ell$-th layer is strong while it is nearly zero when the layer only has intermediate or weak signal.
\begin{figure}[t]
\centering
\includegraphics[width=0.6\textwidth]{plot/symm_sens2params.pdf}
\caption{\small Average misclustering proportions against $\log(1/\rho_{\textnormal{input}})$ when $\{p_\ell\}, \{q_\ell\}$ are estimated from data.
The black dashed vertical line represents the location of the true $\rho$ that generates the data, and the horizontal lines represent the errors made by Algorithm \ref{alg: generic_refinement} when $\{p_\ell\}, \{q_\ell\}, \rho$ are all known.
}
\label{fig:sens2params}
\end{figure}
\paragraph{Sensitivity to inexact parameter specifications.}
The optimality of our proposed algorithm has been established assuming knowledge of the true $\{p_\ell\}, \{q_\ell\}$ and $\rho$.
In practice, we need to estimate them from data.
Estimating $\{p_\ell\}$ and $\{q_\ell\}$ is relatively easy --- we could first obtain a crude estimate of $p_\ell$'s using method of moment (see Appendix \ref{subappend:est_deg} for details), then input that to Algorithm \ref{alg: specc} to obtain an initial estimator $\tilde \mathbf{z}$ of $\mathbf{z}^\star$, and finally compute the \emph{intra-cluster} and \emph{inter-cluster} average of edges in each layer, which will be our final estimator of $\{p_\ell\}$ and $\{q_\ell\}$.
Estimating $\rho$ is, however, a nontrivial task.
Alternatively, we could treat the input $\rho$ (denoted as $\rho_\textnormal{input}$) to Algorithm \ref{alg: generic_refinement} as a hyperparameter. In this simulation, we examine the sensitivity of the algorithm to the estimated (hence inexact) $\{p_\ell\}, \{q_\ell\}$ and $\rho$. We again set $n = 1000, L = 100$, and we fix $c = 2, \rho = 0.1$.
We run Algorithm \ref{alg: generic_refinement} with estimated $\{p_\ell\}$ and $\{q_\ell\}$ over $500$ instances of the model with different (misspecified) input values of $\rho$ and plot misclustering proportions for both global~estimation~and individualized~estimation~in Figure \ref{fig:sens2params}.
From Figure \ref{fig:sens2params}, we see that our algorithm is robust to inexact parameters.
As long as we do not set $\rho_{\textnormal{input}}$ to be too large, the performance of our algorithm with estimated $\{p_\ell\}$ and $\{q_\ell\}$ only slightly degrades compared to when true parameter values are used.
\section{Discussion}\label{sec:discuss}
\section{Proofs of Lower Bounds}\label{append:prf_lower_bound}
\input{0700-proof-lower-bound}
\section{Proofs of Upper Bounds}\label{append:prf_upper_bound}
\input{0700-proof-specClust}
\input{0700-proof-refinement}
\input{0700-proof-optimality}
\section{Concentration and Regularization of Multilayer Networks}\label{append:concentration}
\input{0700-concentration}
\section{Properties of Key Information-Theoretic Quantities
\input{0700-properties}
\section{More Details on Experiments}\label{append:more_exp}
\input{0700-additional-simulations}
\end{appendices}
\begin{comment}
\section{Current cemetery}
\subsection{Simulation}
See Figure \ref{fig: compare clusterings}, \ref{fig: compare weights} for a comparison of different weighted average spectral clustering.
\subsection{Concentration of weighted sum of Bernoulli}
\label{sec: weighted sum bernoulli}
Suppose $\PP{X_i=1-p_1}=p_1,\PP{X_i=-p_i}=1-p_i$ for $1\leq i\leq n_1$, $\PP{X_i=1-p_2}=p_2,\PP{X_i=-p_2}=1-p_2$ for $n_1+1\leq i\leq n_1+n_2$. Let $\omega_1,\omega_2>0$ such that $\omega_1n_1+\omega_2n_2=1$. We want to prove the concentration inequality for $Y=\omega_1\sum_{i=1}^{n_1}X_i+\omega_2\sum_{i=n_1+1}^{n_1+n_2}X_i$. Note that $\forall \lambda>0$,
\begin{align*}
\PP{Y\geq a}&\leq e^{-\lambda a}\EE{e^{-\lambda Y}}=\Exp{-\lambda a-\lambda(n_1\omega_1p_1+n_2\omega_2p_2)}\cdot(p_1e^{\lambda\omega_1}+1-p_1)^{n_1}\cdot(p_2e^{\lambda\omega_2}+1-p_2)^{n_2}.
\end{align*}
Choose $\omega_1,\omega_2,\lambda$ to minimize
\begin{align*}
f(\omega_1,\omega_2,\lambda)&=\Exp{-\lambda a-\lambda(n_1\omega_1p_1+n_2\omega_2p_2)}\cdot(p_1e^{\lambda\omega_1}+1-p_1)^{n_1}\cdot(p_2e^{\lambda\omega_2}+1-p_2)^{n_2}.
\end{align*}
The logarithm can be written as
\begin{align*}
\log f(\omega_1,\omega_2,\lambda)&=-\lambda(a+n_1\omega_1p_1+n_2\omega_2p_2)+n_1\log(p_1e^{\lambda\omega_1}+1-p_1)+n_2\log(p_2e^{\lambda\omega_2}+1-p_2)\\
&= -\lambda(a+n_1\omega_1p_1+n_2\omega_2p_2)+n_1\left[\lambda p_1\omega_1(1+\frac{1}{2}\lambda\omega_1)-\frac{1}{2}(\lambda p_1\omega_1)^2(1+\frac{1}{2}\lambda\omega_1)^2\right]+\\
&\qquad n_2\left[\lambda p_2\omega_2(1+\frac{1}{2}\lambda\omega_2)-\frac{1}{2}(\lambda p_2\omega_2)^2(1+\frac{1}{2}\lambda\omega_2)^2\right]+o(\lambda^2 (p_1\omega_1^2+p_2\omega_2^2))\\
&\leq -\lambda a+\frac{\lambda^2}{2}\left[\omega_1^2n_1p_1(1-p_1)+\omega_2^2n_2p_2(1-p_2)
\right].
\end{align*}
By Cauchy-Schwarz inequality, RHS is minimized when $\frac{\omega_1}{\omega_2}=\frac{p_2(1-p_2)}{p_1(1-p_1)}$. So we should let
\begin{align*}
&\omega_1=\frac{1/(p_1-p_1^2)}{n_1/(p_1-p_1^2)+n_2/(p_2-p_2^2)},\\
&\omega_2=\frac{1/(p_2-p_2^2)}{n_1/(p_1-p_1^2)+n_2/(p_2-p_2^2)},\\
&\lambda=a(\frac{n_1}{p_1-p_1^2}+\frac{n_2}{p_2-p_2^2}).
\end{align*}
\SL{In this case,
$$
\lambda^2(p_1\omega_1^2+p_2\omega_2^2)=a^2(\frac{1}{1-p_1}+\frac{1}{1-p_2}).
$$
Need $a=o(1)$.
}
Then
\begin{align*}
\log f(\omega_1,\omega_2,\lambda)&\leq-\frac{a^2}{2}(\frac{n_1}{p_1-p_1^2}+\frac{n_2}{p_2-p_2^2}).
\end{align*}
Thus
\begin{align*}
\PP{Y\geq a}\leq\Exp{-\frac{a^2}{2}(\frac{n_1}{p_1-p_1^2}+\frac{n_2}{p_2-p_2^2})}
\end{align*}
More generally, consider ${\bar{A}}=\sum_{\ell=1}^L\omega_\ell A^{(\ell)}$, where $\sum\omega_\ell=1$ and $\omega_\ell\propto(p^{(\ell)}-(p^{(\ell)})^2)^{-1}$. Let $e(I,J)=\sum_{\ell}\omega_\ell e^{(\ell)}(I,J)$, $\bar{\mu}(I,J)=(\sum_{\ell}\omega_\ell p^{(\ell)})|I||J|$. Want bounded degree and bounded discrepancy.
\begin{align*}
\sum_{\ell}d_i^{(\ell)}\leq Cn\sum_\ell p_\ell.
\end{align*}
\begin{figure}
\centering
\begin{subfigure}{.45\textwidth}
\includegraphics[scale=0.4]{n=400,L=20_4_weighted.png}
\end{subfigure}
\begin{subfigure}{.45\textwidth}
\includegraphics[scale=0.4]{n=400,rho=01_4_weighted.png}
\end{subfigure}
\caption{Weighted mean spectral clustering}
\label{fig: compare clusterings}
\end{figure}
\begin{figure}
\centering
\begin{subfigure}{.45\textwidth}
\includegraphics[scale=.4]{compare_weights_n=400,L=10}
\end{subfigure}
\begin{subfigure}{.45\textwidth}
\includegraphics[scale=.4]{compare_weights_n=400,rho=01}
\end{subfigure}
\caption{Comparing different weights}
\label{fig: compare weights}
\end{figure}
\subsection{Gilbert-Varshamov bound}
Write ${\mathcal{H}}_n=\{\pm1\}^n$. For any $m>1$, let ${\mathcal{V}}_m$ be the maximal subset of ${\mathcal{H}}_n$ such that for any $z,z'\in{\mathcal{V}}_m$, $\|z-z'\|_1> n/m$. For $z\in{\mathcal{V}}_m$, let $B(z, n/m)=\{x\in{\mathcal{H}}_n:\|x-z\|_1\leq n/m\}$. Thus $\cup_{z\in{\mathcal{V}}_m}B(z,n/m)={\mathcal{H}}_n$, and
$$
|{\mathcal{V}}_m|\cdot|B(z,n/m)|\geq 2^n.
$$
Let $b_1,\ldots,b_n\stackrel{iid}{\sim}\text{Bern}(1/2)$. By the uniformity,
\begin{align*}
2^{-n}|B(z,n/m)|&=\PP{b_1+\ldots+b_n\leq \frac{n}{2m}}\\
&=\PP{b_1+\ldots+b_n\geq \frac{n(2m-1)}{2m}}\\
&\leq \Exp{-\lambda n\frac{2m-1}{2m}}\EE{\Exp{\lambda(b_1+\ldots+b_n)}}.
\end{align*}
Take $\lambda=\log(2m-1)$, we get
\begin{align*}
|{\mathcal{V}}_m|\geq\Exp{n(\frac{2m-1}{2m}\log(2m-1)-\log m)}.
\end{align*}
(When $m\rightarrow\infty$, RHS approximates $\Exp{n\log 2}=2^n$. When $m=2$, $|{\mathcal{V}}_m|\geq\exp{n(\frac{3}{4}\log 3-\log(2))}>\Exp{n/8}$)
\subsection{Local initialization given mean assignment}
Consider a fixed $\ell\in[L]$ and let $A=A^{(\ell)}$. Let $z\in\{\pm1\}^n$ be the assignment of $A$ and $Z$ be the $n\times 2$ matrix with $Z_{i\cdot}=(1,0)$ if $z_i=1$ and $Z_{i\cdot}=(0,1)$ if $z_i=-1$. Let $z^*$ be the mean assignment, such that $z^*_i=1$ for $1\leq i\leq n/2$ and $z^*_i=-1$ for $(n/2)< i\leq n$. $Z^*$ is similarly defined. Let $J=[n/2]$ and $J^c=\{\frac{n}{2}+1,\ldots,n\}$, $p=\frac{a\log n}{n},q=\frac{b\log n}{n}$. Write $P^*=Z^*B(Z^*)^\intercal=\lambda_1^*u_1^*(u_1^*)^\intercal+\lambda_2^*u_2^*(u_2^*)^\intercal$, where
\begin{gather*}
\lambda_1^*=\frac{n(p-q)}{2},\quad\lambda_2^*=\frac{n(p-q)}{2}\\
u_1^*=\frac{1}{\sqrt{n}}(1,\ldots,1)^\intercal,\quad u_2^*=\frac{1}{\sqrt n}(1,\ldots,1,-1,\ldots,-1)^\intercal
\end{gather*}
Note that for any fixed $i\in[n]$,
\begin{align*}
\sqrt{n}z_i(\frac{Au_2^*}{\lambda_2^*})&=\frac{2z_i}{(a-b)\log n}[\sum_{j\in J}A_{ij}-\sum_{j\in J^c}A_{ij}]
\end{align*}
Goal: prove that there exists a constant $\eta>0$, such that
\begin{align*}
\min_{i\in[n]}z_i(\frac{\sqrt{n}Au_2^*}{\lambda_2^*})\geq\eta
\end{align*}
with high probability. That means, the sign of $Au_2^*$ is a consistent estimator of $z$.
\begin{theorem}[Exact recovery]
If $\sqrt{(1-\rho)a+\rho b}-\sqrt{(1-\rho)b+\rho a}\geq\sqrt{2}$, then there exists a constant $\eta=\eta(a,b,\rho)>0$, such that
\begin{align*}
\min_{i\in[n]}z_i\frac{\sqrt{n}(Au_2^*)_i}{\lambda_2^*}\geq\eta
\end{align*}
with probability at least $1-o(1)$.
\end{theorem}
\begin{proof}
Note that
\begin{align*}
z_i\frac{\sqrt{n}(Au_2^*)_i}{\lambda_2^*}&=\frac{2z_i}{(a-b)\log n}[\sum_{j=1}^{n/2}A_{ij}-\sum_{j=(n/2)+1}^nA_{ij}].
\end{align*}
Assume $z_i=1$. For any $\varepsilon\in\mathbb{R}$ and $\lambda<0$, consider
\begin{align*}
\PP{\sum_{j=1}^{n/2}A_{ij}-\sum_{j=(n/2)+1}^nA_{ij}\leq\varepsilon\log n}&=\PP{\Exp{\lambda[\sum_{j=1}^{n/2}A_{ij}-\sum_{j=(n/2)+1}^nA_{ij}]}\geq e^{\lambda\varepsilon\log n}}\\
&\leq n^{-\lambda\varepsilon}\EE{\Exp{\lambda[\sum_{j=1}^{n/2}A_{ij}-\sum_{j=(n/2)+1}^nA_{ij}]}}
\end{align*}
Let $n_1=\#\{j:1\leq j\leq n/2:z_j=-1\}$ and $n_2=\#\{j:(n/2)< j\leq n:z_j=1\}$, then
\begin{align*}
&\PP{\sum_{j=1}^{n/2}A_{ij}-\sum_{j=(n/2)+1}^nA_{ij}\leq\varepsilon\log n}\leq\\
&\qquad n^{-\lambda\varepsilon}[1+(e^\lambda-1)p]^{n/2-n_1}[1+(e^{\lambda}-1)q]^{n_1}[1+(e^\lambda-1)q]^{n/2-n_2}[1+(e^{\lambda}-1)p]^{n_2},
\end{align*}
thus
\begin{align*}
&\log\PP{\sum_{j=1}^{n/2}A_{ij}-\sum_{j=(n/2)+1}^nA_{ij}\leq\varepsilon\log n}\\
&\qquad\leq-\lambda\varepsilon\log n+(e^\lambda-1)\left[(\frac{1}{2}-\frac{n_1}{n})a+\frac{n_1}{n}b\right]\log n+(e^{-\lambda}-1)\left[(\frac{1}{2}-\frac{n_2}{n})b+\frac{n_2}{n}a\right]\log n\\
&\qquad =C\log n,
\end{align*}
where $C=-\lambda\varepsilon+(e^\lambda-1)[(\frac{1}{2}-\frac{n_1}{n})a+\frac{n_1}{n}b]+(e^{-\lambda}-1)[(\frac{1}{2}-\frac{n_2}{n})b+\frac{n_2}{2}a]$. Write $\tilde a=(\frac{1}{2}-\frac{n_1}{n})a+\frac{n_1}{n}b$ and $\tilde b=(\frac{1}{2}-\frac{n_2}{n})b+\frac{n_2}{n}a$, and choose $\lambda$ such that $e^{\lambda}=\sqrt{\tilde{b}/\tilde{a}}$. Then
\begin{align*}
\log\PP{\sum_{j=1}^{n/2}A_{ij}-\sum_{j=(n/2)+1}^nA_{ij}\leq\varepsilon\log n\bigg\vert z}\leq [\frac{\varepsilon}{2}\log(\frac{\tilde a}{\tilde b})-(\sqrt{\tilde a}-\sqrt{\tilde b})^2]\cdot\log n.
\end{align*}
If $\sqrt{\tilde a}-\sqrt{\tilde b}>1$, then we can choose $\varepsilon$ such that
\begin{align*}
\sum_{j=1}^{n/2}A_{ij}-\sum_{j=(n/2)+1}^nA_{ij}\geq\varepsilon\log n
\end{align*}
with probability at least $1-o(n^{-1})$. Hence by union bound,
\begin{align*}
\min_{i\in[n]}z_i\frac{\sqrt{n}(Au_2^*)_i}{\lambda_2^*}\geq\frac{2\varepsilon}{a-b}:=\eta,
\end{align*}
with probability at least $1-o(1)$.
It only remains to show that $\sqrt{\tilde a}-\sqrt{\tilde b}>1$ with probability at least $1-o(1)$. Note that $n_1,n_2$ are iid $\mathrm{Binom}(\rho,\frac{n}{2})$ random variables.
\begin{lemma}
If $n_1,n_2$ are iid $\mathrm{Binom}(\rho,m)$, and $\sqrt{(1-\rho)a+\rho b}-\sqrt{(1-\rho)b+\rho a}=\sqrt{2}+\delta_0$ where $\delta_0>0$, then
$$
\sqrt{(1-\frac{n_1}{m})a+\frac{n_1}{m} b}-\sqrt{(1-\frac{n_2}{m})b+\frac{n_2}{m} a}>\sqrt{2}
$$
with probability at least $1-cm^{-1}$ for some $c=c(\rho,a,b,\delta_0)$.
\end{lemma}
\begin{proof}
Note that
\begin{align*}
&\PP{\sqrt{(1-\frac{n_1}{m})a+\frac{n_1}{m} b}-\sqrt{(1-\frac{n_2}{m})b+\frac{n_2}{m} a}\leq\sqrt{2}}\\
&\qquad=\PP{(\sqrt{(1-\frac{n_1}{m})a+\frac{n_1}{m} b}-\sqrt{(1-\rho)a+\rho b})-(\sqrt{(1-\frac{n_2}{m})b+\frac{n_2}{m} a}-\sqrt{(1-\rho)b+\rho a})\leq-\delta_0}\\
&\qquad\leq\PP{\sqrt{(1-\frac{n_1}{m})a+\frac{n_1}{m} b}-\sqrt{(1-\rho)a+\rho b}\leq-\frac{\delta_0}{2}}+\PP{\sqrt{(1-\frac{n_2}{m})b+\frac{n_2}{m} a}-\sqrt{(1-\rho)b+\rho a}\geq\frac{\delta_0}{2}}\\
&\qquad=\PP{\sqrt{1-\frac{n_1}{m}(1-\frac{b}{a})}-\sqrt{1-\rho(1-\frac{b}{a})}\leq-\frac{\delta_0}{2\sqrt{a}}}+\PP{\sqrt{1+\frac{n_2}{m}(\frac{a}{b}-1)}-\sqrt{1+\rho(\frac{a}{b}-1)}\geq\frac{\delta_0}{2\sqrt{b}}}.
\end{align*}
When $0<x<y<1-\frac{b}{a}$, $\sqrt{1-x}-\sqrt{1-y}\leq\sqrt{\frac{a}{4b}}(y-x)$, thus
\begin{align*}
\PP{\sqrt{1-\frac{n_1}{m}(1-\frac{b}{a})}-\sqrt{1-\rho(1-\frac{b}{a})}\leq-\frac{\delta_0}{2\sqrt{a}}}&\leq\PP{\sqrt{\frac{a}{4b}}|(1-\frac{b}{a})(\frac{n_1}{m}-\rho)|\geq\frac{\delta_0}{2\sqrt{a}}}\\
&=\PP{|\frac{n_1}{m}-\rho|\geq\frac{\delta_0\sqrt{b}}{a-b}}\\
&\leq(\frac{a-b}{\delta_0\sqrt{b}})^2\mathrm{Var}{\frac{n_2}{m}}\\
&=\frac{1}{m}\frac{\rho(1-\rho)(a-b)^2}{\delta_0^2b}
\end{align*}
and
\begin{align*}
\PP{\sqrt{1+\frac{n_2}{m}(\frac{a}{b}-1)}-\sqrt{1+\rho(\frac{a}{b}-1)}\geq\frac{\delta_0}{2\sqrt{b}}}&\leq\PP{\sqrt{\frac{b}{4a}}(\frac{a}{b}-1)|\frac{n_2}{m}-\rho|\geq\frac{\delta_0}{2\sqrt{b}}}\\
&=\PP{|\frac{n_2}{m}-\rho|\geq\frac{\delta_0\sqrt{a}}{a-b}}\\
&\leq\frac{1}{m}\frac{\rho(1-\rho)(a-b)^2}{\delta_0^2 a}
\end{align*}
Therefore
$$
\sqrt{(1-\frac{n_1}{m})a+\frac{n_1}{m} b}-\sqrt{(1-\frac{n_2}{m})b+\frac{n_2}{m} a}>\sqrt{2}
$$
with probability at least $1-cm^{-1}$ for some $c=c(\rho,a,b,\delta_0)$.
\end{proof}
\end{proof}
\end{comment}
\subsection{Spectral Clustering and Choice of Weights}\label{subappend:weight_choice}
Recall that our Algorithms \ref{alg: generic_refinement} and \ref{alg: provable_refinement} both require an initialization scheme, which by default is set to spectral clustering on the trimmed weighted adjacency matrix $\bar A = \sum_{\ell\in[L]}{\omega_\ell}\AAA{\ell}$ (i.e., Algorithm \ref{alg: specc}). In this experiment, we set the trimming threshold $\gamma = 5$, and we explore three choices of weights: (1) $\omega_\ell \propto 1$ (uniform weight), (2) $\omega_\ell \propto 1/p_\ell$ (scale by variance), and (3) $\omega_\ell \propto 1/\sqrt{p_\ell}$ (scale by standard deviation), where $p_\ell$'s are either known or estimated from the data using the method of moment (see Appendix \ref{subappend:est_deg} for the detailed estimation procedure). We consider the setup in Section \ref{sec:exp}, and we set $n =1000, L = 100, \rho = 0.1$ and vary $c$.
\begin{figure}[t]
\centering
\includegraphics[width=0.6\textwidth]{plot/symm_compare_wts.pdf}
\caption{\small Average misclustering proportion against the signal strength $c$ for spectral clustering under different choices of weights. A check mark means the ground truth $p_\ell$'s are used, and a cross mark means $p_\ell$'s are estimated from the data.
}
\label{fig:compare_wts}
\end{figure}
Figure \ref{fig:compare_wts} shows the misclustering proportion (for estimating $\mathbf{z}^\star$) over $500$ simulations. We see that overall, uniform weight performs the best, regardless of whether $p_\ell$'s are known or not. Scaling by $\sqrt{p_\ell}$ performs slightly better than uniform weight when the signal strength $c$ is large, but is significantly worse when $c$ is small. Scale by $p$ is the worst among the three. In the rest of the experiments, the initialization scheme is always set to spectral clustering with uniform weight.
\subsection{Co-Regularized Spectral Clustering}\label{subappend:co_reg}
Co-regularized spectral clustering is based on the following optimization problem:
\[
[\hat U^{(1)},\ldots,\hat U^{(L)},\hat U^\star]=\mathop{\rm argmax}_{\substack{{U^{(\ell)}}^\intercal U^{(\ell)}=I,\forall\ell\in[L] \\ {U^\star}^\intercal U^\star=I }}\sum_{\ell=1}^L\tr{{U^{(\ell)}}^\intercal \AAA{\ell}U^{(\ell)} } + \gamma_\ell \tr{{U^{\star}}^\intercal U^{(\ell)}{U^{(\ell)}}^\intercal U^{\star} }.
\]
In our implementation, the regularization parameter is set to be $\gamma_\ell=\|\AAA{\ell}\|_2$ as suggested by \cite{paul2020spectral}, and we solve the problem by alternating between optimizing $U^{(\ell)}$ and $U^\star$ via eigen-decomposition. The maximum number of iterations is 20. We then apply $k$-means on $\hat U^{\star}$ and $\hat U^{(\ell)}$ to get the global assignment $\mathbf{\hat z}^{\star}$ and individual assignments $\{\mathbf{\hat z}^{(\ell)}\}$. We emphasize that \cite{paul2020spectral} only proposed to use $\hat U^\star$ to get the global assignment and proved the consistency, and did not propose to use $\hat U^{(\ell)}$ to get the individual assignments.
\subsection{Estimating \texorpdfstring{$p_\ell$}{pl}'s}\label{subappend:est_deg}
The inputs to Algorithm \ref{alg: specc} include $\{p_\ell\}$. To estimate them, we begin by recalling that $E[\AAA{\ell}_{ij}] =\tilde p_\ell = p_\ell - 2(p_\ell-q_\ell)(\rho-\rho^2) \approx p_\ell$ if $\mathbf{z}^\star_i = \mathbf{z}^\star_j$ and $E[\AAA{\ell}_{ij}] =\tilde q_\ell= q_\ell + 2(p_\ell-q_\ell)(\rho-\rho^2) \approx q_\ell$ otherwise.
Assuming $\tilde p_\ell = C\tilde q_\ell$, we then have
$$
\sum_{i<j} \mathbb{E}[A^{\ell}_{ij}] = \tilde{p}_\ell\times \frac{n^2 - n - 2(1-C^{-1})n^\star_+n^\star_-}{2}.
$$
If $n^\star_+ = C'n$, then we have
$$
\sum_{i<j} \mathbb{E}[\AAA{\ell}_{ij}] = \tilde{p}_\ell\times \frac{[1-2(1-C^{-1})C'(1-C')]n^2 - n}{2}.
$$
Specifically, if $C' \approx 1/2$ (i.e., the two clusters are approximately symmetric), we have
$$
\sum_{i<j} \mathbb{E}[\AAA{\ell}_{ij}] \approx \tilde{p}_\ell\times \frac{(0.5+0.5C^{-1})n^2 - n}{2} \geq \tilde{p}_\ell\times \frac{0.5n^2 - n}{2}.
$$
So a conservative estimator for $p_\ell$ is given by
$$
\hat{p}_\ell = \frac{2\sum_{i<j} \mathbb{E}[\AAA{\ell}_{ij}]}{0.5n^2 -n}.
$$
\subsection{Step \texorpdfstring{\RN{1}}{I}: Concentration on a Big Block}
We first introduce a technical tool called \emph{Grothendieck-Pietsch factorization}, which allows us to ``upgrade'' an $\ell_\infty$-to-$\ell_2$ norm bound to an $\ell_2$-to-$\ell_2$ norm bound.
\begin{lemma}[Grothendieck-Pietsch factorization, Theorem 3.2 of \cite{le2017concentration}]
\label{lemma: grothendieck_pietsch_factorization}
Let $B\in {\mathbb{R}}^{k\times m}$ and $\delta > 0$. Then there exists $J\subseteq [m]$ with $|J|\geq (1-\delta)m$ such that the following holds:
$$
\|B_{[k]\times J}\| \leq \frac{2 \|B \|_{\infty\to 2}}{\sqrt{\delta m}}.
$$
\end{lemma}
With the above lemma at hand, the strategy now is to first establish a concentration result in $\ell_\infty\rightarrow\ell_2$ norm and then to upgrade it to the operator norm using Lemma~\ref{lemma: grothendieck_pietsch_factorization}.
\begin{lemma}[Concentration in $\infty$-to-$2$ norm]
\label{lemma: concentration_in_infty_to_two_norm}
Assume \eqref{eq: assump_on_weights_general_bernoulli_matrix} holds with $c_0 >0$.
For any $r \geq 1$, the following holds with probability at least $1-n^{-r}$:
uniformly for any $m\in[n]$ and any block $I\times J\subseteq [n]\times[n]$ with $|I|=|J|=m$,
if we let $I'$ be the indices of rows of ${\bar{A}}_{I\times J}$ whose $\ell_1$-norm is bounded above by $\alpha \sum_{\ell\in[L]}\omega_\ell d_\ell$, where $\alpha$ is any number satisfying $\alpha\geq m/n$,
then we have
\begin{equation}
\label{eq:concentration_inf_to_two_norm}
\| ({\bar{A}} - {\mathbb{E}} {\bar{A}})_{I'\times J} \|_{\infty \to 2} \leq C\sqrt{\alpha \cdot (\sum_{\ell\in[L]} \omega_\ell^2 d_\ell) \cdot mr \log (\frac{en}{m})},
\end{equation}
where $C= C(c_0)$ is an absolute constant.
\end{lemma}
\begin{proof}[Proof of Lemma \ref{lemma: concentration_in_infty_to_two_norm}]
Let us fix any $m\in[n]$, $\alpha \geq m/n$,
and take any block $I\times J\subseteq [n]\times[n]$ with $|I| = |J| = m$.
By definition, we have
\begin{align*}
\|({\bar{A}} - {\mathbb{E}} {\bar{A}})_{I'\times J}\|_{\infty \to 2}^2 &= \sup_{\|x\|_\infty \leq 1} \|Ax\|^2_2.
\end{align*}
Since the right-hand side is the supremum of a convex function over a convex set, the supremum is attained at the boundary. Hence we have
\begin{align*}
\|({\bar{A}} - {\mathbb{E}} {\bar{A}})_{I'\times J}\|_{\infty \to 2}^2 = \max_{x\in\{\pm 1\}^m} \sum_{i\in I^\prime} \bigg(\sum_{j\in J} ({\bar{A}}_{ij} - {\mathbb{E}} {\bar{A}}_{ij})x_j\bigg)^2
:= \max_{x\in\{\pm 1\}^m} \sum_{i\in I} (X_i \xi_i)^2,
\end{align*}
where we let
\begin{align*}
X_i & := \sum_{j\in J} ({\bar{A}}_{ij}-{\mathbb{E}} {\bar{A}}_{ij})x_j = \sum_{\ell\in[L]} \sum_{j\in J} \omega_\ell(\AAA{\ell}_{ij}-{\mathbb{E}} \AAA{\ell}_{ij})x_j, \\
\xi_i &:= \mathbf{1}{i\in I'} = \mathds{1}\bigg\{\sum_{j\in J} {\bar{A}}_{ij}\leq \alpha \sum_{\ell\in[L]} \omega_\ell d_\ell \bigg\} = \mathds{1} \bigg\{\sum_{\ell\in[L]} \sum_{j\in J} \omega_\ell \AAA{\ell}_{ij}\leq \alpha \sum_{\ell\in[L]}\omega_\ell d_\ell \bigg\}.
\end{align*}
Note that $X_i$ has mean zero and its variance satisfies
\[
\mathrm{Var}{(X_i)}=\sum_{\ell\in[L]}\sum_{j\in J}\omega_\ell^2\ppp{\ell}_{ij}(1-\ppp{\ell}_{ij})\leq\sum_{\ell\in[L]}\sum_{j\in J}\omega_\ell^2\ppp{\ell}_{ij}\leq
\frac{m}{n}\sum_{\ell\in[L]}\omega_\ell^2d_\ell
\]
since $d_\ell=\max_{i,j} np_{ij}^{(\ell)}$.
Meanwhile, we have $|\omega_\ell(\AAA{\ell}_{ij}-{\mathbb{E}} \AAA{\ell}_{ij})x_j|\leq\|\boldsymbol{\omega}\|_\infty$.
Invoking Bernstein's inequality, we get
\begin{align*}
{\mathbb{P}}(|X_i\xi_i| \geq tm)\leq {\mathbb{P}}(|X_i|\geq tm) & \leq 2\exp\bigg\{\frac{-m^2t^2/2}{ (m/n)\cdot \sum_{\ell\in[L]} \omega_\ell^2 d_\ell + \frac{1}{3}mt\|\boldsymbol{\omega}\|_\infty} \bigg\}.\addtocounter{equation}{1}\tag{\theequation}\label{equ: bernstein X_i}
\end{align*}
Note that
\[
|X_i\xi_i| \leq \sum_{\ell\in[L]} \sum_{j\in J} \omega_\ell (\AAA{\ell}_{ij}\xi_i + {\mathbb{E}} \AAA{\ell}_{ij}) \leq \sum_{\ell\in [L]} \sum_{j\in J} \omega_\ell \AAA{\ell}_{ij}\xi_i + \frac{m}{n} \sum_{\ell\in[L]} \omega_{\ell} d_\ell\leq (\alpha+\frac{m}{n})\sum_{\ell\in[L]}\omega_\ell d_\ell,
\]
where the last inequality is by the definition of $\xi_i$.
Since $m/n\leq\alpha$, the above display translates to $|X_i\xi_i|\leq 2\alpha\sum_{\ell\in[L]}\omega_\ell d_\ell$. This means that if $tm > 2\alpha\sum_{\ell\in[L]}\omega_t d_\ell$, then the probability in the left-hand side of \eqref{equ: bernstein X_i} is zero. On the other hand, if we we assume $tm\leq 2\alpha\sum_{\ell\in[L]}\omega_\ell d_\ell$, then we can further bound the right-hand side of \eqref{equ: bernstein X_i} by
\begin{align*}
{\mathbb{P}}(|X_i\xi_i|\geq tm)
&\leq 2\exp\bigg\{\frac{-m^2t^2/2}{ \alpha\cdot \sum_{\ell\in[L]} \omega_\ell^2 d_\ell + \frac{2\alpha}{3} \|\boldsymbol{\omega}\|_\infty \sum_{\ell\in[L]}\omega_\ell d_\ell} \bigg\}\\
&\leq 2\exp\bigg\{\frac{-m^2t^2/2}{ (1+\frac{2c_0}{3}) \cdot \alpha\sum_{\ell\in[L]} \omega_\ell^2 d_\ell} \bigg\},
\end{align*}
where the last inequality is due to \eqref{eq: assump_on_weights_general_bernoulli_matrix}. Combining the two cases, we conclude that the above display holds for all choices of $tm > 0$.
This means that $X_i\xi_i$ has sub-Gaussian norm $\lesssim \sqrt{\alpha\sum_{\ell\in[L]}\omega^2_\ell d_\ell}$ (see, e.g., Lemma 5.5 of \cite{vershynin2010introduction}), and hence $(X_i\xi_i)^2$ has sub-exponential norm $\lesssim\alpha \sum_{\ell\in[L]} \omega_\ell^2 d_\ell$ (see, e.g., Lemma 5.14 of \cite{vershynin2010introduction}). Invoking Corollary 5.17 of \cite{vershynin2010introduction}, we have
$$
{\mathbb{P}}\bigg(\sum_{i\in I}(X_i \xi_i)^2 > \varepsilon m {\alpha} \sum_{\ell\in[L]} \omega_\ell^2 d_\ell \bigg) \leq 2\exp\bigg\{ -c (\varepsilon^2 \land \varepsilon) m \bigg\}
$$
for some constant $c$ only depending on $c_0$. Choosing $\varepsilon = (10/c)r \log(en/m)$ for some constant $r\geq 1$, we deduce that with probability at least $1-(en/m)^{-5rm}$, we have
$$
\sum_{i\in I} (X_i\xi_i)^2 \leq (10/c) r \log (en/m) \cdot m\alpha \sum_{\ell\in[L]} \omega_\ell^2 d_\ell.
$$
Taking a union bound over all possible configurations of $m\in[n], x\in\{\pm 1\}^m$, and $I, J$ with $|I|=|J|=m$, the conclusion of the lemma holds with probability at least
\begin{equation}
\label{eq:concentration_linf_l2_norm_prob}
1 - \sum_{m=1}^n 2^m \binom{n}{m}^2 \bigg(\frac{en}{m}\bigg)^{-5rm}\geq 1- \sum_{m=1}^n 2^m \bigg(\frac{en}{m}\bigg)^{-(5r-2)m}\geq 1- \sum_{m=1}^n \bigg(\frac{en}{m}\bigg)^{-(5r-3)m},
\end{equation}
where the first inequality is by $\binom{n}{m}\leq (en/m)^m$ and the second inequality is by $2\leq en/m$.
We claim that
$
(en/m)^{-(5r-3)m} \geq \big(en/(m+1)\big)^{-(5r-3)(m+1)}
$
for any $m\in[n]$.
Indeed, with some algebra, this claim is equivalent to
\[
(m+1)\log(m+1)-m\log n\leq \log(en)
\]
which holds for any $1\leq m\leq n$. Now, the right-hand side of \eqref{eq:concentration_linf_l2_norm_prob} can be further lower bounded by
$$
1 - n \cdot (en)^{-(5r-3)} \geq 1-n^{-r},
$$
where we have used $r\geq 1$, and this is exactly the desired result.
\end{proof}
The above lemma, along with Lemma \ref{lemma: grothendieck_pietsch_factorization} (with $\delta =1/4$), gives the following result.
\begin{lemma}[Concentration in spectral norm]
\label{lemma: concentration_in_spectral_norm}
Assume \eqref{eq: assump_on_weights_general_bernoulli_matrix} holds with $c_0 >0$.
Then for any $r \geq 1$, the following holds with probability at least $1-n^{-r}$:
uniformly for any $m\in[n]$ and any block $I\times J \subseteq[n]\times [n]$ with $|I|=|J|=m$, if we let $I'$ be the indices of rows of $A_{I\times J}$ whose $\ell_1$-norm is bounded above by $\alpha \sum_{\ell\in[L]}\omega_\ell d_\ell$, where $\alpha$ an arbitrary (but fixed) number satisfying $\alpha \geq m/n$, then there exists a subset $J'\subseteq J$ with $|J'|\geq 3m/4$ such that
\begin{equation}
\label{eq:concentration_in_spectral_norm}
\|({\bar{A}} - {\mathbb{E}} {\bar{A}})_{I'\times J'}\| \leq C\sqrt{\alpha \cdot (\sum_{\ell\in[L]} \omega_\ell^2 d_\ell) \cdot r \log (\frac{en}{m})},
\end{equation}
where $C=C(c_0)$ is an absolute constant.
\end{lemma}
\subsection{Step \texorpdfstring{\RN{2}}{II}: Restricted \texorpdfstring{$\ell_1$}{l1} Norm}
The following lemma shows that {most of the rows} of ${\bar{A}}$ have $\ell_1$ norm bounded from above by a constant multiple of $r\alpha\sum_{\ell\in[L]}\omega_{\ell\in[L]} d_\ell$.
\begin{lemma}[Degree of subgraphs]
\label{lemma: deg_of_subgraphs}
Assume \eqref{eq: another_assump_on_weights_general_bernoulli_matrix} holds $c_1 \geq 1$.
Then for any $r\geq 1$, the following holds with probability at least $1-n^{-r}$:
uniformly for any $m\in[n]$ and any block $I\times J\subseteq [n]\times[n]$ with $|I|=|J|=m$, all but $m\|\boldsymbol{\omega}\|_\infty/(\alpha \sum_{\ell\in[L]}\omega_\ell d_\ell)$ rows of ${\bar{A}}_{I\times J}$ have $\ell_1$-norm bounded above by $Cr\alpha \sum_{\ell\in[L]}\omega_\ell d_\ell$, where $\alpha$ is an arbitrary (but fixed) number satisfying $\alpha \geq\sqrt{m/n}$ and $C = C(c_1)$ is an absolute constant.
\end{lemma}
The proof of this lemma relies on the following concentration inequality for the weighted average of Bernoulli random variables, which is a generalization of the classical concentration inequality for the sum of independent Bernoulli random variables proved in \cite{hoeffding1963probability}.
\begin{lemma}[Concentration inequality for weighted Bernoulli sum]
\label{lemma: concentration_of_weighted_bernoulli_sum}
Let $\{X_i\}_{1\leq i\leq n}$ be independent random variables, each distributed as $X_i\sim \textnormal{\text{Bern}}(p_i)$. Let $\boldsymbol{\omega} = \{\omega_i\}_{1\leq i\leq n}$ be a weight vector such that $\omega_i>0 ,\forall i\in[n]$. Assume there exists a constant $c_1\geq 1$ such that
$$
\|\boldsymbol{\omega} \|_\infty \sum_{i\in[n]} p_i \leq c_1 \sum_{i\in[n]} \omega_i p_i.
$$
Then for any $t\geq\sum_{i\in[n]}\omega_i p_i$, we have
$$
{\mathbb{P}}(\sum_{i\in[n]}\omega_i X_i \geq t) \leq \bigg(\frac{C \sum_{i\in[n]} \omega_i p_i}{t}\bigg)^{t/\|\boldsymbol{\omega}\|_\infty},
$$
where $C = e^{c_1}$.
\end{lemma}
\begin{proof}[Proof of Lemma \ref{lemma: concentration_of_weighted_bernoulli_sum}]
For any $\lambda > 0$ we have
\begin{align*}
{\mathbb{P}}(\sum_{i\in[n]}\omega_i X_i \geq t) & \leq \exp\{-\lambda t\} \prod_{i\in[n]} {\mathbb{E}} \exp\{\lambda \omega_i X_i\} \\
& \leq \exp\{-\lambda t\} \bigg(\frac{\sum_{i\in[n]}{\mathbb{E}}\exp\{\lambda\omega_iX_i\}}{n}\bigg)^n,
\end{align*}
where the second inequality is due to the inequality of arithmetic and geometric means. Since $\exp\{\lambda \omega_i x\}$ is convex in $x$, its graph for $x\in[0, 1]$ is dominated by the line segment connecting the two points $(0, 1)$ and $(1, e^{\lambda\omega_i})$ in ${\mathbb{R}}^2$. Hence we have
$
e^{\lambda \omega_i X_i} \leq (e^{\lambda \omega_i } -1) X_i + 1.
$
Taking expectation on both sides, we get
$
{\mathbb{E}} e^{\lambda \omega_i X_i} \leq (e^{\lambda\omega_i} -1) p_i + 1.
$
This gives
\begin{align*}
{\mathbb{P}}(\sum_{i\in[n]} \omega_i X_i \geq t) \leq \exp\{-\lambda t\} \bigg(\frac{\sum_{i\in[n]}\big( p_i e^{\lambda \omega_i}+ (1- p_i)\big)}{n}\bigg)^n.
\end{align*}
Taking $\lambda = \|\boldsymbol{\omega}\|_\infty^{-1}\log ({t}/{\sum_{i\in[n]}\omega_ip_i})$, the right-hand side above is equal to
\begin{align*}
& \bigg(\frac{\sum_{i\in[n]}\omega_i p_i}{t}\bigg)^{t/\|\boldsymbol{\omega}\|_\infty} \bigg(1 + \frac{\sum_{i\in[n]}p_i (t/\sum_{i\in[n]} \omega_i p_i)^{w_i/\|\boldsymbol{\omega}\|_\infty} }{n} - \frac{\sum_{i\in[n]} p_i}{n}\bigg)^n \\
& \overset{(1)}{\leq} \bigg(\frac{\sum_{i\in[n]}\omega_i p_i}{t}\bigg)^{t/\|\boldsymbol{\omega}\|_\infty} \exp\bigg\{ \sum_{i\in[n]}p_i (t/\sum_{i\in[n]} \omega_i p_i)^{w_i/\|\boldsymbol{\omega}\|_\infty}\bigg\} \\
& \overset{(2)}\leq \bigg(\frac{ \sum_{i\in[n]} \omega_i p_i}{t}\bigg)^{t/\|\boldsymbol{\omega}\|_{\infty}} \exp\bigg\{ \sum_{i\in[n]}p_i (t/\sum_{i\in[n]} \omega_i p_i)\bigg\} \\
& = \bigg(\frac{ \sum_{i\in[n]} \omega_i p_i}{t}\bigg)^{t/\|\boldsymbol{\omega}\|_{\infty}} \exp\bigg\{t \cdot \frac{\sum_{i\in[n]} p_i}{\sum_{i\in[n]}\omega_i p_i}\bigg\} \\
& \overset{(3)}\leq \bigg(\frac{ \sum_{i\in[n]} \omega_i p_i}{t}\bigg)^{t/\|\boldsymbol{\omega}\|_{\infty}} (e^{c_1})^{t/\|\boldsymbol{\omega}\|_\infty} \\
& = \bigg(\frac{ e^{c_1}\sum_{i\in[n]} \omega_i p_i}{t}\bigg)^{t/\|\boldsymbol{\omega}\|_{\infty}}
\end{align*}
where the (1) is by $1+x \leq e^x$ for any $x\in{\mathbb{R}}$, (2) is by our assumption that $t \geq \sum_{i\in[n]}\omega_i p_i$, and (3) is by our assumption on the weight vector.
\end{proof}
We now present the proof of Lemma \ref{lemma: deg_of_subgraphs}.
\begin{proof}[Proof of Lemma \ref{lemma: deg_of_subgraphs}]
Let the $\ell_1$-norm of the $i$-th row of ${\bar{A}}_{I\times J}$ be
$$
D_i = \sum_{j\in J} {\bar{A}}_{ij} = \sum_{\ell \in [L]}\sum_{j\in J} \omega_\ell \AAA{\ell}_{ij}.
$$
We have
$$
{\mathbb{E}} D_i =\sum_{\ell\in[L]} \sum_{j\in J} \omega_\ell \ppp{\ell}_{ij} \leq \frac{m}{n} \sum_{\ell\in[L]} \omega_\ell d_\ell \leq \alpha \sum_{\ell\in[L]} \omega_\ell d_\ell.
$$
Using \eqref{eq: another_assump_on_weights_general_bernoulli_matrix}, for any $J\subseteq[n]$, we have
\[
\|\boldsymbol{\omega}\|_\infty \sum_{\ell\in[L]}\sum_{j\in J}\ppp{\ell}_{ij}\leq c_1\sum_{\ell\in[L]}\sum_{j\in J}\omega_\ell\ppp{\ell}_{ij}.
\]
Thus we can invoke Lemma \ref{lemma: concentration_of_weighted_bernoulli_sum} to conclude that for a large enough $C'$,
\begin{align*}
{\mathbb{P}}(D_i > C'r\alpha \sum_{\ell\in[L]}\omega_\ell d_\ell)
& \leq \bigg(\frac{C (m/n)\sum_{\ell\in[L]}\omega_\ell d_\ell}{ C'r\alpha \sum_{\ell\in[L]}\omega_\ell d_\ell }\bigg)^{C' r\alpha \sum_{\ell\in[L]}\omega_\ell d_\ell /\|\omega \|_\infty} \\
& \leq \bigg(\frac{C'\alpha n}{Cm }\bigg)^{-C'r\alpha \sum_{\ell\in[L]}\omega_\ell d_\ell /\|\omega \|_\infty} \\
& =: \mu.
\end{align*}
Let $S$ be the number of rows $i \in I$ such that $D_i >C'r\alpha \sum_{\ell\in[L]}\omega_\ell d_\ell$. Then $S$ is a sum of $m = |I|$ independent Bernoulli random variables, each having head probability at most $\mu$. So invoking Lemma \ref{lemma: concentration_of_weighted_bernoulli_sum} again (with $c_1 = 1$, $\omega_i=1$ for all $i$), we have
$$
{\mathbb{P}}\bigg( S > \frac{m\|\boldsymbol{\omega}\|_\infty}{\alpha \sum_{\ell\in[L]} \omega_\ell d_\ell} \bigg) \leq \bigg(\frac{em\mu}{m\|\boldsymbol{\omega}\|_\infty/(\alpha \sum_{\ell\in[L]}\omega_\ell d_\ell)}\bigg)^{m\|\boldsymbol{\omega}\|_\infty/(\alpha \sum_{\ell\in[L]}\omega_\ell d_\ell)}.
$$
We claim that the right-hand side above is at most
$
\mu^{m\|\boldsymbol{\omega}\|_\infty/(2\alpha\sum_{\ell\in[L]}\omega_\ell d_\ell)}
$
for $C'$ large enough.
Indeed, this claim is equivalent to
$$
\frac{e\alpha\sum_{\ell\in[L]}\omega_\ell d_\ell}{\|\boldsymbol{\omega}\|_\infty} \leq \mu^{-1/2} = \bigg(\frac{C'\alpha n}{Cm}\bigg)^{ \frac{C'r\alpha \sum_{\ell\in[L]}\omega_\ell d_\ell}{2\|\boldsymbol{\omega}\|_\infty}}.
$$
Since $\alpha n/m \geq 1$, it is true if
$$
e \cdot {\alpha\sum_{\ell\in[L]}\omega_\ell d_\ell}/{\|\boldsymbol{\omega}\|_\infty} \leq \bigg((C'/C)^{C'r/2}\bigg)^{{\alpha\sum_{\ell\in[L]}\omega_\ell d_\ell}/{\|\boldsymbol{\omega}\|_\infty}}.
$$
For a given constant $C$ (which only depends on $c_1$), we can choose $C'$ large enough such that the above inequality holds. Hence, we have
$$
{\mathbb{P}}\bigg(S > m\| \omega\|_\infty/(\alpha \sum_{\ell\in[L]}\omega_\ell d_\ell)\bigg) \leq \mu^{m\|\boldsymbol{\omega}\|_\infty/(2\alpha\sum_{\ell\in[L]}\omega_\ell d_\ell)} = \bigg(\frac{C'\alpha n}{C m}\bigg)^{-C'rm/2}\leq \bigg(\frac{(C')^2n}{C^2m}\bigg)^{-C'rm/4},
$$
where the last inequality is due to $\alpha^2n/m\geq1$.
Taking a union bound over all possible $m\in[n]$ and $I, J$ with $|I|=|J|=m$, we know that $S > m\| \omega\|_\infty/(\alpha\sum_{\ell\in[L]}\omega_\ell d_\ell)$ with probability at least
$$
1 - \sum_{m=1}^n \binom{n}{m}^2 \bigg(\frac{(C')^2n}{C^2m}\bigg)^{-C'rm/4}
\geq 1 - \sum_{m=1}^n \bigg(\frac{en}{m}\bigg)^{2m} \bigg(\frac{(C')^2n}{C^2m}\bigg)^{-C'rm/4}
\geq 1 - \sum_{m=1}^n \bigg(\frac{({C'})^2n}{{C^2}m}\bigg)^{-(\frac{C'r}{4}-2)m},
$$
where the last inequality holds by choosing a large enough $C'$. Similar to the proof of Lemma \ref{lemma: concentration_in_infty_to_two_norm}, one readily checks that among the summands in the right-hand side above, the one with $m=1$ is the dominating term, and thus the right-hand side above can be further lower bounded by
\begin{equation}
\label{eq:deg_of_subgraphs_prob}
1 - n\cdot \bigg(\frac{({C'})^2 n}{{C}^2}\bigg)^{-(\frac{C'r}{4} - 2)} \geq 1-n^{-r}
\end{equation}
for $C'$ large enough, and this concludes the proof of Lemma \ref{lemma: deg_of_subgraphs}.
\end{proof}
The following lemma shows that if a block has a small number of rows, then {most of its columns} has small $\ell_1$-norm.
\begin{lemma}[More on degrees of subgraphs]
\label{lemma: more_on_deg_of_subgraphs}
Assume \eqref{eq: another_assump_on_weights_general_bernoulli_matrix} holds for $c_1 \geq 1$.
Then for any $r \geq 1$, the following holds with probability at least $1-n^{-r}$:
uniformly for any $m\in[n]$, any
$$
k\leq \bigg(m \|\boldsymbol{\omega}\|_\infty/(\alpha \sum_{\ell\in[L]}\omega_\ell d_\ell)\bigg) \land m
$$
where $\alpha$ is an arbitrary (but fixed) number satisfying $\alpha \geq \sqrt{m/n}$,
and any block $I \times J \subseteq [n]\times[n]$ with $|I|=k, |J|=m$, all but $m/4$ columns of ${\bar{A}}_{I\times J}$ have $\ell_1$-norm bounded above by $Cr\|\boldsymbol{\omega}\|_\infty$, where $C = C(c_1)$ is an absolute constant.
\end{lemma}
\begin{proof}[Proof of Lemma \ref{lemma: more_on_deg_of_subgraphs}]
We define the $\ell_1$ norm of the $j$-th column of the matrix ${\bar{A}}_{I\times J}$ as
$$
D_j = \sum_{i\in I} {\bar{A}}_{ij} =\sum_{\ell \in[L]} \sum_{i\in I} \omega_\ell \AAA{\ell}_{ij}.
$$
Now, $D_j$ is a weighted sum of Bernoulli random variables with
$$
{\mathbb{E}} D_j = \sum_{\ell\in[L]}\sum_{i\in I} \omega_\ell \ppp{\ell}_{ij} \leq \frac{k}{n} \sum_{\ell\in[L]}\omega_\ell d_\ell \leq \frac{m\|\boldsymbol{\omega}\|_\infty}{\alpha n},
$$
where the last inequality is due to $k\leq m\|\boldsymbol{\omega}\|_\infty/(\alpha \sum_{\ell\in[L]}\omega_\ell d_\ell)$.
By Lemma \ref{lemma: concentration_of_weighted_bernoulli_sum}, for large enough $C'$ we have
$$
{\mathbb{P}}(D_j > C'r\|\boldsymbol{\omega}\|_\infty) \leq \bigg(\frac{C m\|\boldsymbol{\omega}\|_\infty/(\alpha n)}{C'r \|\boldsymbol{\omega}\|_\infty}\bigg)^{C'r} = \bigg(\frac{C' r \alpha n}{Cm}\bigg)^{-C'r} =: \mu.
$$
Let $S$ be the number of columns $j\in J$ with $D_j > C'r\|\boldsymbol{\omega}\|_\infty$. Then $S$ is a sum of independent Bernoulli random variables, each having success probability at most $\mu$. Applying Lemma \ref{lemma: concentration_of_weighted_bernoulli_sum} (with $c_1 = 1$) gives
$
{\mathbb{P}}(S > m/4) \leq (4e\mu)^{m/4}.
$
We claim that the above probability is at most $\mu^{m/6}$. This claim is equivalent to
$
4e < \mu^{-1/3} = \big(\frac{C' r\alpha n }{Cm}\big)^{C'r/3}.
$
Since $\alpha n/m\geq 1$ and $r \geq 1$, it suffices to require
$
4e < \big(\frac{C' }{C}\big)^{C'r/3},
$
which is true for $C'$ large enough. Hence
$$
{\mathbb{P}}(S > m/4) \leq \mu^{m/6} = \bigg(\frac{C'r\alpha n}{Cm}\bigg)^{-C'rm/6} \leq \bigg(\frac{C'\alpha n}{Cm}\bigg)^{-C'rm/6} \leq \bigg(\frac{(C')^2n }{C^2 m}\bigg)^{-C'rm/12},
$$
where the last inequality is by $\alpha \geq \sqrt{m/n}$.
We now take a union bound over all $m, k$ and $I, J$ with $|I| = k, |J| = m$. Note that it suffices to consider the largest possible $k$, which is at most $m$. So $S > m/4$ happens with probability at least
$$
1-\sum_{m=1}^n \binom{n}{m}^2 \bigg(\frac{(C')^2n }{C^2 m}\bigg)^{-C'rm/12}\leq 1-n^{-r},
$$
where the last inequality holds by choosing a large $C'$ and using similar arguments as those which lead to \eqref{eq:deg_of_subgraphs_prob}.
\end{proof}
\subsection{Step \texorpdfstring{\RN{3}}{III}: Graph Decomposition}
The main idea in \cite{le2017concentration} is to seek for a partition of the set $[n]\times [n]$ into three blocks ${\mathcal{N}}$, ${\mathcal{R}}$ and ${\mathcal{C}}$, where ${\mathcal{N}}$ is a big block with good concentration behavior, and the rows of ${\mathcal{R}}$ and the columns of ${\mathcal{C}}$ have small $\ell_1$ norm. This \emph{graph decomposition} is implemented below.
\begin{proposition}[Graph decomposition]
\label{prop:graph_decomp}
Let Assumption \ref{assump:wts_for_concentration} hold with constants $c_0 > 0, c_1 \geq 1$. For any $r\geq 1$, with probability at least $1-3n^{-r}$, we can decompose $[n]\times [n]$ into three classes ${\mathcal{N}}, {\mathcal{R}}, {\mathcal{C}}$ so that the following holds:
\begin{itemize}
\item The matrix ${\bar{A}}$ concentrates well on ${\mathcal{N}}$ in the sense that
$$
\|({\bar{A}}-{\mathbb{E}} {\bar{A}})_{\mathcal{N}}\|\leq C_1 r^{3/2}\sqrt{\sum_{\ell\in[L]} \omega_\ell^2 d_\ell},
$$
where $C_1 = C_1(c_0, c_1)$ is an absolute constant;
\item Each row of ${\bar{A}}_{\mathcal{R}}$ and each column of ${\bar{A}}_{\mathcal{C}}$ has $\ell_1$-norm bounded above by $C_2r\|\boldsymbol{\omega}\|_\infty$, where $C_2 = C_2(c_0, c_1)$ is another absolute constant;
\item Moreover, ${\mathcal{R}}$ intersects at most $n\|\boldsymbol{\omega}\|_\infty/{\sum_{\ell\in[L]} \omega_\ell d_\ell}$ columns and ${\mathcal{C}}$ intersects at most $n\|\boldsymbol{\omega}\|_\infty/{\sum_{\ell\in[L]} \omega_\ell d_\ell}$ rows of $[n]\times[n]$.
\end{itemize}
\end{proposition}
The proof of the above result is based on iterative applications of the following lemma.
\begin{lemma}[Decomposition of one block]
\label{lemma: decomp_of_one_block}
Let Assumption \ref{assump:wts_for_concentration} hold with constants $c_0 > 0, c_1 \geq 1$.
Then for $r \geq 1$, the following holds with probability at least $1-3n^{-r}$:
uniformly for any $m\in[n]$, any block $I\times J \subseteq [n]\times[n]$ with $|I|=|J| = m$ and an arbitrary (but fixed) number $\alpha$ satisfying $\alpha \geq \sqrt{m/n}$, there exists a sub-block $I_1\times J_1 \subseteq I \times J$ with $|I_1|, |J_1|\leq m/2$, such that the remaining part of the block, namely $(I\times J)\setminus (I_1 \times J_1)$, can be decomposed into three parts, ${\mathcal{R}}\subseteq (I\setminus I_1)\times J$, ${\mathcal{C}}\subseteq I \times (J\setminus J_1)$, and $\cal N$, so that the following holds:
\begin{itemize}
\item The matrix ${\bar{A}}$ concentrates well on ${\mathcal{N}}$ in the sense that
$$
\|({\bar{A}}-{\mathbb{E}} {\bar{A}})_{\mathcal{N}}\|\leq C_1 r^{3/2}\sqrt{\alpha (\sum_{\ell\in[L]} \omega_\ell^2 d_\ell) \log (\frac{en}{m})},
$$
where $C_1 = C_1(c_0, c_1)$ is an absolute constant;
\item Each row of ${\bar{A}}_{\mathcal{R}}$ and each column of ${\bar{A}}_{\mathcal{C}}$ has $\ell_1$-norm bounded above by $C_2r\|\boldsymbol{\omega}\|_\infty$, where $C_2 = C_2(c_0, c_1)$ is another absolute constant;
\item Moreover, ${\mathcal{R}}$ intersects at most $m\|\boldsymbol{\omega}\|_\infty/(\alpha \sum_{\ell\in[L]}\omega_\ell d_\ell)$ columns and ${\mathcal{C}}$ intersects at most $m\|\boldsymbol{\omega}\|_\infty/(\alpha \sum_{\ell\in[L]}\omega_\ell d_\ell)$ rows of $I\times J$.
\end{itemize}
\end{lemma}
\begin{proof}
The proof is an adaptation of arguments in the proof of Lemma 3.7 of \cite{le2017concentration}. We fix a realization of ${\bar{A}}$ such that Lemmas \ref{lemma: concentration_in_spectral_norm}, \ref{lemma: deg_of_subgraphs}, and \ref{lemma: more_on_deg_of_subgraphs} hold. Note that this event happens with probability at least $1- 3n^{-r}$.
\begin{figure}[t]
\centering
\includegraphics[width=\textwidth]{plot/graph_decomp.pdf}
\caption{\small A pictorial illustration of one step of graph decomposition in Lemma \ref{lemma: decomp_of_one_block}. In the leftmost figure, we construct from $\bar A_{I\times J}$ the red shaded block which concentrates well, and a block ${\mathcal{C}}$ which satisfies the requirements in Lemma \ref{lemma: decomp_of_one_block}. In the middle figure, we apply the same construction on ${\bar{A}}_{I\times J}^\top$ to get the blue shaded block which concentrates well, and a block ${\mathcal{R}}$ which satisfies the requirements in Lemma \ref{lemma: decomp_of_one_block}. In the rightmost figure, we combine the two constructions to obtain the desired decomposition in Lemma \ref{lemma: decomp_of_one_block}. The union of the red and blue shaded block is precisely ${\mathcal{N}}$.}
\label{fig:graph_decomp}
\end{figure}
We first construct the ``bad columns'' $J_1$. Fix some $\alpha \geq \sqrt{m/n}$.
By Lemma \ref{lemma: deg_of_subgraphs}, all but $m\|\boldsymbol{\omega}\|_\infty/(\alpha \sum_{\ell\in[L]}\omega_\ell d_\ell)$ rows of ${\bar{A}}_{I\times J}$ have $\ell_1$ norm bounded above by $C r\alpha \sum_{\ell\in[L]}\omega_\ell d_\ell$, where $C$ only depends on $c_1$.
Let $I'\subseteq I$ be the indices of those rows whose $\ell_1$ norm are bounded above by $Cr\alpha \sum_{\ell\in[L]} \omega_\ell d_\ell$ and $|I\setminus I'|\leq m\|\boldsymbol{\omega}\|_\infty/(\alpha \sum_{\ell\in[L]}\omega_\ell d_\ell)$.
By Lemma \ref{lemma: concentration_in_spectral_norm} (with $\alpha$ replaced by $Cr\alpha$), we know that there exists a subset $J'\subseteq J$ with $|J'|\geq 3m/4$ such that
$$
\|({\bar{A}} - {\mathbb{E}} {\bar{A}})_{I'\times J'} \| \leq C' r \sqrt{\alpha (\sum_{\ell\in[L]}\omega_\ell^2 d_\ell) \log(\frac{en}{m})},
$$
where $C'$ depends on both $c_0$ and $c_1$.
For rows in $I\setminus I'$, whose cardinality is bounded above by both $m\|\boldsymbol{\omega}\|_\infty/(\alpha \sum_{\ell\in[L]}\omega_\ell d_\ell)$ and $m$, we use Lemma \ref{lemma: more_on_deg_of_subgraphs} to deduce that, all but $m/4$ columns of ${\bar{A}}_{(I\setminus I')\times J}$ have $\ell_1$ norm bounded above by $C''r\|\boldsymbol{\omega}\|_\infty$, where $C''$ only depends on $c_1$. Let $J_1$ be the union of columns in $J\setminus J'$ and the columns of $A_{(I\setminus I')\times J}$ whose $\ell_1$ norm is larger than $C'' r \|\boldsymbol{\omega}\|_\infty$. Note that $|J_1|\leq m/4 + m/4 = m/2$ by construction. In summary, we have found row indices $I'$ and column indices $J_1$, such that:
\begin{itemize}
\item The block $({\bar{A}} - {\mathbb{E}} {\bar{A}})_{I' \times (J\setminus J_1)}$ satisfies the concentration inequality in the last display;
\item The block ${\mathcal{C}} := (I\setminus I')\times (J\setminus J_1)$ satisfies the property specified in the lemma, i.e., columns of ${\bar{A}}_{\mathcal{C}}$ has $\ell_1$ norm bounded above by $C''r\|\boldsymbol{\omega}\|_\infty$, and it intersects at most $m\|\boldsymbol{\omega}\|_\infty/(\alpha \sum_{\ell\in[L]}\omega_\ell d_\ell)$ rows of $I\times J$;
\item The size of $J_1$ is at most $m/2$.
\end{itemize}
See the leftmost part of Figure \ref{fig:graph_decomp} for a pictorial illustration.
Now, we apply the same arguments to $A^\top$, which allows us to find row indices $I_1$ and column indices $J''$, such that
\begin{itemize}
\item The block $({\bar{A}} - {\mathbb{E}} {\bar{A}})_{(I\setminus I_1) \times J''}$ satisfies the concentration inequality in the last display;
\item The block ${\mathcal{R}} := (I\setminus I_1)\times (J\setminus J'')$ satisfies the property specified in the lemma, i.e., rows of $A_{\mathcal{R}}$ has $\ell_1$ norm bounded above by $C'''r\|\boldsymbol{\omega}\|_\infty$ for some constant $C'''$ only depending on $c_0, c_1$, and it intersects at most $m\|\boldsymbol{\omega}\|_\infty/(\alpha \sum_{\ell\in[L]}\omega_\ell d_\ell)$ columns of $I\times J$;
\item The size of $I_1$ is at most $m/2$.
\end{itemize}
See the middle part of Figure \ref{fig:graph_decomp} for a pictorial illustration.
To this end, we let ${\mathcal{N}} = (I'\times (J\setminus J_1)) \cup ((I\setminus I_1)\times J'')$. See the rightmost part of Figure \ref{fig:graph_decomp} for a pictorial illustration. It is clear that $(A-{\mathbb{E}} A)_{\mathcal{N}}$ satisfies the concentration inequality required by the lemma, which completes the proof.
\end{proof}
\begin{proof}[Proof of Proposition \ref{prop:graph_decomp}]
Proposition \ref{prop:graph_decomp} follows by iteratively applying Lemma \ref{lemma: decomp_of_one_block} for ${\mathcal{O}}(\log n)$ times. Such arguments are nearly identical to arguments used in the proof of Theorem 2.6 in \cite{le2017concentration}, so we omit the details.
\end{proof}
\subsection{Proof of Theorem \ref{thm:concentration_reg_mat}}
First, let us observe that bounded row and column $\ell_1$ norms leads to bounded operator norm.
\begin{lemma}[Lemma 2.7 of \cite{le2017concentration}]
\label{lemma: relation_between_spectral_norm_and_l1_norm}
Consider a matrix $B$ in which each row has $\ell_1$-norm at most $a$, and each column has $\ell_1$-norm at most $b$. Then $\|B\|\leq \sqrt{ab}$.
\end{lemma}
Following the route taken by Theorem 2.1 in \cite{le2017concentration}, we start with the following decomposition:
$$
\tau({\bar{A}}) -{\mathbb{E}} {\bar{A}} = (\tau({\bar{A}}) - {\mathbb{E}} {\bar{A}})_{\mathcal{N}} + (\tau({\bar{A}}) - {\mathbb{E}} {\bar{A}})_{\mathcal{R}} + (\tau({\bar{A}}) - {\mathbb{E}} {\bar{A}})_{\mathcal{C}}.
$$
We take a realization of ${\bar{A}}$ such that the conclusions in Proposition \ref{prop:graph_decomp} hold, which happens with probability at least $1-3n^{-r}$.
For the well-behaved part, we write
$$
(\tau({\bar{A}}) - {\mathbb{E}} {\bar{A}})_{\mathcal{N}} = ({\bar{A}} - {\mathbb{E}} {\bar{A}})_{\mathcal{N}} - ({\bar{A}} - \tau({\bar{A}}))_{\mathcal{N}}.
$$
By Proposition \ref{prop:graph_decomp}, we have
$$
({\bar{A}} - {\mathbb{E}} {\bar{A}})_{\mathcal{N}} \leq Cr^{3/2}\sqrt{\sum_{\ell\in[L]}\omega_\ell^2 d_\ell}.
$$
On the other hand, since $\tau$ only takes effect on the elements in ${\mathcal{E}}$, we have ${\bar{A}}_{{\mathcal{E}}^c} = \tau({\bar{A}})_{{\mathcal{E}}^c}$, and hence
$$
\|({\bar{A}} - \tau({\bar{A}}))_{\mathcal{N}}\|=\|({\bar{A}}-\tau({\bar{A}}))_{{\mathcal{N}}\cap {\mathcal{E}}}\| \leq \|{\bar{A}}_{{\mathcal{N}} \cap {\mathcal{E}}}\| \leq \|({\bar{A}} - {\mathbb{E}} {\bar{A}})_{{\mathcal{N}} \cap{\mathcal{E}}}\| + \|{\mathbb{E}} {\bar{A}}_{{\mathcal{N}} \cap {\mathcal{E}}}\|,
$$
where the first inequality is due to $0\leq {\bar{A}} - \tau({\bar{A}}) \leq {\bar{A}}$ entry-wise (recall that we only do down-weighting in the regularization process). Since ${\mathcal{E}}$ is a disjoint union two of rectangular blocks, we have
$$
\| ({\bar{A}} - {\mathbb{E}} {\bar{A}})_{{\mathcal{N}} \cap {\mathcal{E}}}\| \leq 2\|({\bar{A}} - {\mathbb{E}} {\bar{A}})_{\mathcal{N}}\| \leq Cr^{3/2}\sqrt{\sum_{\ell\in[L]}\omega_\ell^2 d_\ell}.
$$
Moreover, since the spectral norm of a non-negative matrix can only reduce by restricting onto any subset of $[n]\times[n]$, we get
$$
\|{\mathbb{E}} {\bar{A}}_{{\mathcal{N}}\cap{\mathcal{E}}}\|\leq \|{\mathbb{E}} {\bar{A}}_{\mathcal{E}}\| \leq \|{\mathbb{E}} {\bar{A}}_{I \times [n]}\| + \|{\mathbb{E}} {\bar{A}}_{I^c\times I}\|,
$$
where we recall that $I$ contains the the vertices we choose to regularize. By construction, the $\ell_1$-norm of all rows of ${\mathbb{E}} {\bar{A}}_{I\times[n]}$ is bounded above by $\sum_{\ell\in[L]}\omega_\ell d_\ell$.
Meanwhile, by definition, $|I|\leq cn\|\boldsymbol{\omega}\|_\infty/(\sum_{\ell\in[L]}\omega_\ell d_\ell)$, and thus the columns of ${\mathbb{E}} {\bar{A}}_{I\times[n]}$ have $\ell_1$ norm bounded above by $c\|\omega \|_\infty$. By Lemma \ref{lemma: relation_between_spectral_norm_and_l1_norm}, we have
$$
\|{\mathbb{E}} {\bar{A}}_{I\times[n]}\| \leq \sqrt{c \|\omega \|_\infty \sum_{\ell\in[L]}\omega_\ell d_\ell} \leq \sqrt{c c_0 \sum_{\ell\in[L]}\omega_\ell^2 d_\ell},
$$
where the last inequality is by our assumption on the weights \eqref{eq: assump_on_weights_general_bernoulli_matrix}.
A similar bound holds for ${\mathbb{E}} {\bar{A}}_{I^c\times I}$, which gives
$$
\|{\mathbb{E}} {\bar{A}}_{{\mathcal{N}} \cap {\mathcal{E}}}\|\lesssim \sqrt{c_0 \sum_{\ell\in[L]}\omega_\ell^2 d_\ell}.
$$
Hence we arrive at
$$
(\tau({\bar{A}}) - {\mathbb{E}} {\bar{A}})_{\mathcal{N}} \leq C_1 r^{3 / 2} \sqrt{\sum_{\ell\in[L]} \omega_\ell^2 d_\ell},
$$
where $C_1$ only depending on $c_0, c_1$.
Now we deal with the block ${\mathcal{R}}$. We have
$$
\|(\tau({\bar{A}}) - {\mathbb{E}} {\bar{A}})_{\mathcal{R}}\| \leq \|\tau({\bar{A}})_{\mathcal{R}}\| + \|{\mathbb{E}} {\bar{A}}_{\mathcal{R}}\|.
$$
Recall that $0\leq \tau({\bar{A}})_{\mathcal{R}} \leq {\bar{A}}_{\mathcal{R}}$ entry-wise because we only do down-weighting. By Proposition \ref{prop:graph_decomp}, each row of ${\bar{A}}_{\mathcal{R}}$, and thus of $\tau({\bar{A}})_{\mathcal{R}}$, has $\ell_1$-norm at most $Cr\|\boldsymbol{\omega}\|_\infty$. Moreover, by our regularization process, each column of $\tau({\bar{A}})$, and thus of $\tau({\bar{A}})_{\mathcal{R}}$, has $\ell_1$-norm at most $d_\tau$. Then, Lemma \ref{lemma: relation_between_spectral_norm_and_l1_norm} gives
$$
\|\tau({\bar{A}})_{\mathcal{R}} \| \leq \sqrt{C r \|\boldsymbol{\omega}\|_\infty d_\tau}.
$$
For ${\mathbb{E}} {\bar{A}}_{\mathcal{R}}$, by Proposition \ref{prop:graph_decomp}, each row of it has at most $n\|\boldsymbol{\omega}\|_\infty/{\sum_{\ell\in[L]} \omega_\ell d_\ell}$ non-zero entries, and all entries are bounded above by $\sum_{\ell\in[L]}\omega_\ell d_\ell/n$. Hence each row of ${\mathbb{E}} {\bar{A}}_{\mathcal{R}}$ has $\ell_1$-norm at most $\|\boldsymbol{\omega}\|_\infty$. Meanwhile, each column of ${\mathbb{E}} {\bar{A}}_{\mathcal{R}}$ has $\ell_1$-norm at most $\sum_{\ell\in[L]}\omega_\ell d_\ell$. An application of Lemma \ref{lemma: relation_between_spectral_norm_and_l1_norm} and \eqref{eq: assump_on_weights_general_bernoulli_matrix} gives
$$
\|{\mathbb{E}} {\bar{A}}_{\mathcal{R}}\| \leq \sqrt{c_0 \sum_{\ell\in[L]}\omega_\ell^2 d_\ell},
$$
and hence
$$
\|(\tau({\bar{A}})-{\mathbb{E}} {\bar{A}} )_{\mathcal{R}}\| \leq \sqrt{C r \|\omega \|_\infty d_\tau} + \sqrt{c_0 \sum_{\ell\in[L]}\omega_\ell^2 d_\ell}.
$$
A nearly identical argument gives
$$
\|(\tau({\bar{A}})-{\mathbb{E}} {\bar{A}} )_{\mathcal{C}}\| \leq \sqrt{C r \| \omega\|_\infty d_\tau} + \sqrt{c_0 \sum_{\ell\in[L]}\omega_\ell^2 d_\ell}.
$$
Finally, we combine the bounds above to conclude
$$
\|\tau({\bar{A}}) - {\mathbb{E}} {\bar{A}}\| \leq C_1 r^{3/2}\sqrt{\sum_{\ell\in[L]} \omega_\ell^2 d_\ell} + 2 \sqrt{C_2 r\|\boldsymbol{\omega}\|_\infty d_\tau}\\ \leq C r^{3/2}\left( \sqrt{\sum_{\ell\in[L]} \omega_\ell^2 d_\ell}+\sqrt{\|\boldsymbol{\omega}\|_\infty} d_\tau \right),
$$
where $C$ only depends on $c_0,c_1, c$.
\subsection{Proof of Corollary \ref{cor:concentration_trimmed_mat}}
Since $\gamma > e^{c_1} \geq e$, by Lemma \ref{lemma: concentration_of_weighted_bernoulli_sum}, we have
$$
{\mathbb{P}}(\sum_{j\in[n]} {\bar{A}}_{ij} > \gamma \sum_{\ell\in[L]} \omega_\ell d_\ell)
\leq \bigg(\frac{e^{c_1} \sum_{\ell\in[L]}\sum_{j\in[n]}\omega_\ell p_{ij}^{(\ell)}}{\gamma\sum_{\ell\in[L]}\omega_\ell d_\ell}\bigg)^{\gamma \sum_{\ell\in[L]}\omega_\ell d_\ell / \| \omega\|_\infty} \leq (e^{c_1}/\gamma)^{\gamma \sum_{\ell\in[L]}\omega_\ell d_\ell / \| \omega\|_\infty} =: \mu.
$$
Invoking Lemma \ref{lemma: concentration_of_weighted_bernoulli_sum} again, for large enough $c$ we have
$$
{\mathbb{P}}(|I| \geq cn\|\boldsymbol{\omega}\|_\infty/\sum_{\ell\in[L]} \omega_\ell d_\ell) \leq \bigg(\frac{en\mu}{cn\| \omega\|_\infty /\sum_{\ell\in[L]}\omega_\ell d_\ell}\bigg)^{cn\| \omega\|_\infty /\sum_{\ell\in[L]}\omega_\ell d_\ell}.
$$
We claim that, we can specify $c$ based only on $c_1$ and $\gamma$, so that the above probability is less than or equal to $\mu^{cn\|\boldsymbol{\omega}\|_\infty/2\sum_{\ell\in[L]}\omega_\ell d_\ell}$. Indeed, this claim is equivalent to
$$
\frac{e \sum_{\ell\in[L]}\omega_\ell d_\ell}{c\|\boldsymbol{\omega}\|_\infty} \leq \mu^{-1/2} = \bigg(\frac{\gamma}{e^{c_1}}\bigg)^{\gamma \sum_{\ell\in[L]}\omega_\ell d_\ell / 2\| \omega\|_\infty}.
$$
Since we've chosen $\gamma$ such that $\gamma/e^{c_1}> 1$, it suffices to choose $c$ such that
$$
\frac{ex}{c} \leq \bigg((\gamma/e^{c_1})^{\gamma/2}\bigg)^x , \qquad \forall x > 0.
$$
This can be done by choosing $c$ only based on $c_1$ and $\gamma$. Hence, with such a choice of $c$, we have
$$
{\mathbb{P}}(|I| \geq cn\|\boldsymbol{\omega}\|_\infty/\sum_{\ell\in[L]} \omega_\ell d_\ell ) \leq \mu^{cn\|\boldsymbol{\omega}\|_\infty/2\sum_{\ell\in[L]}\omega_\ell d_\ell} = (\gamma/e^{c_1})^{-\gamma c n /2}.
$$
Thus, an application of Theorem \ref{thm:concentration_reg_mat} and \eqref{eq: assump_on_weights_general_bernoulli_matrix} gives the desired result.
\subsection{Proof of Corollary \ref{cor:concentration_w/o_reg}}
Recall that in the proof of Corollary \ref{cor:concentration_trimmed_mat}, we have established:
$$
{\mathbb{P}}(\sum_{j\in[n]} {\bar{A}}_{ij} > \gamma \sum_{\ell\in[L]} \omega_\ell d_\ell) \leq (e^{c_1}/\gamma)^{\gamma \sum_{\ell\in[L]}\omega_\ell d_\ell / \| \omega\|_\infty},
$$
where $\gamma$ is some constant satisfying $\gamma > e^{c_1}$. If we are in the not-too-sparse regime, i.e., $\sum_{\ell\in[L]}\omega_\ell d_\ell \geq c_2 \|\omega \|_\infty\log n$ for some constant $c_2 > 0$, then the above probability can be bounded by
$$
n^{-c_2 \gamma \log (\gamma/e^{c_1})}.
$$
Applying a union bound over $[n]$, we conclude that in this regime, every row of $A$ has $\ell_1$ norm bounded above by $\gamma \sum_{\ell\in[L]}\omega_\ell d_\ell$ with high probability. So without any regularization, we obtain the following guarantee:
$$
\|A - {\mathbb{E}} A \| \leq C r^{3/2} \sqrt{\sum_{\ell\in[L]} \omega_\ell^2 d_\ell}
$$
with probability at least $1-3n^{-r} - n^{-c_3}$, where $c_3$ only depends on $c_1, c_2$
On the other hand, if we are in the very sparse regime, i.e., $\sum_{\ell\in[L]}\omega_\ell d_\ell \leq c_2 \|\omega \|_\infty\log n$, by Lemma \ref{lemma: concentration_of_weighted_bernoulli_sum} we have
\begin{align*}
{\mathbb{P}}(\sum_{j\in[n]} {\bar{A}}_{ij} > \gamma \| \omega\|_\infty \log n) & \leq \bigg(\frac{e^{c_1} \sum_{j}\sum_{\ell\in[L]} \omega_\ell p_{ij}^{(\ell)}}{\gamma \| \omega\|_\infty \log n}\bigg)^{\gamma \| \omega\|_\infty \log (n) / \| \omega \|_\infty} \\
& \leq (c_2 e^{c_1}/\gamma)^{\gamma \log n} \\
& = n^{-\gamma \log (\gamma / c_2e^{c_1})}.
\end{align*}
By choosing $\gamma$ and taking a union bound over $[n]$, we conclude that in this regime, every row of $A$ has $\ell_1$ norm bounded above by $\gamma \|\omega \|_\infty\log n$ with high probability. Hence, invoking Theorem \ref{thm:concentration_reg_mat}, without regularization, we obtain the following guarantee:
$$
\|A - {\mathbb{E}} A \| \leq Cr^{3/2} \sqrt{\sum_{\ell\in[L]}\omega_\ell^2 d_\ell} + C_2 \| \omega\|_\infty \sqrt{\log n},
$$
with probability at least $1 - 3n^{-r} - n^{-c_3}$, where $c_3$ and $C_2$ only depend on $c_1$. The desired result follows by combining the two regimes.
\subsection{Proof of Lemma \ref{lemma:testing_problem_z_star}}\label{subappend:prf_testing_problem_z_star}
Fix any $\mathbf{\tilde z}^\star \in {\mathcal{P}}_n$ with $n_+^\star(\mathbf{\tilde z}^\star) = \lfloor n/2 \rfloor$ and $n_-^\star(\mathbf{\tilde z}^\star) = n - \lfloor n/2\rfloor$. Let
$$
C_+(\mathbf{\tilde z}^\star):= \{i\in[n]: \mathbf{\tilde z}^\star_i = +1\} ,\qquad C_-(\mathbf{\tilde z}^\star):= \{i\in[n]: \mathbf{\tilde z}^\star_i = -1\}.
$$
Now fix any $\delta_n = o(1)$ and choose any $\tilde C_+ \subset C_+(\mathbf{\tilde z}^\star), \tilde C_- \subset C_-(\mathbf{\tilde z}^\star)$ such that
$$
|\tilde C_+| = |\tilde C_-| = \lfloor n/2\rfloor - \lceil \delta_n n/2 \rceil.
$$
Denoting $T = \tilde C_+ \cup \tilde C_-$, we define
\begin{equation}
\label{eq:clusters_in_sub_param_space_z_star}
{\mathcal{Z}}_T := \{\mathbf{z}\in \{\pm 1\}^n: \mathbf{z}_i = \mathbf{\tilde z}^\star_i , \forall i \in T\}.
\end{equation}
Further, we define
\begin{equation}
\label{eq:sub_param_space_z_star}
{\mathcal{P}}_n^0 := \bigg\{ \textnormal{IMLSBM}(\mathbf{z}^\star, \rho, \{p_\ell\}_1^L, \{q_\ell\}_1^L): \mathbf{z}^\star \in {\mathcal{Z}}_T, p_\ell > q_\ell ~ \forall \ell\in[L]\bigg\} \subseteq {\mathcal{P}}_n.
\end{equation}
Note that $\inf_{\mathbf{\hat z}^*} \sup_{\mathbf{z}^* \in {\mathcal{P}}_n} {\mathbb{E}} {\mathcal{L}}(\mathbf{\hat z}^*, \mathbf{z}^*) \geq \inf_{\mathbf{\hat z}^*} \sup_{\mathbf{z}^* \in {\mathcal{P}}_n^0} {\mathbb{E}} {\mathcal{L}}(\mathbf{\hat z}^*, \mathbf{z}^*)$. For any two $\mathbf{z}, \mathbf{z}' \in {\mathcal{P}}_n^0$, we have
\begin{align*}
{\mathcal{L}}(\mathbf{z}, \mathbf{z}') = \frac{1}{n} \sum_{i\in[n]} \mathbf{1}{\mathbf{z}_i \neq \mathbf{z}'_i},
\end{align*}
because
$$
\frac{1}{n} \sum_{i\in[n]} \mathbf{1}{\mathbf{z}_i \neq \mathbf{z}'_i} \leq \frac{n - |T|}{n} = \frac{n - 2 \lfloor n/2 \rfloor + 2 \lceil \delta_n n/2 \rceil}{n} \leq \frac{4}{n} + \delta_n < 1/2
$$
for large $n$. Hence, we can proceed by
\begin{align*}
\inf_{\mathbf{\hat z}^*} \sup_{\mathbf{z}^* \in {\mathcal{P}}_n} {\mathbb{E}} {\mathcal{L}}(\mathbf{\hat z}^*, \mathbf{z}^*) & \geq \inf_{\mathbf{\hat z}^*} \sup_{\mathbf{z}^* \in {\mathcal{Z}}_T} \frac{1}{n} \sum_{i\in[n]} {\mathbb{P}}(\mathbf{\hat z}^*_i \neq \mathbf{z}^*_i)\\
& = \frac{|T^c|}{n} \inf_{\mathbf{\hat z}^*} \sup_{\mathbf{z}^* \in {\mathcal{Z}}_T} \frac{1}{|T^c|} \sum_{i\in T^c} {\mathbb{P}}(\mathbf{\hat z}^*_i \neq \mathbf{z}^*_i) \\
& \geq \frac{1}{n} \sum_{i\in T^c} \inf_{\mathbf{\hat z}^*_i} \underset{\mathbf{z}^*\in {\mathcal{Z}}_T}{\textnormal{ave}} \ {\mathbb{P}}(\mathbf{\hat z}^*_i \neq \mathbf{z}^*_i),
\end{align*}
where
``$\textnormal{ave}$'' denotes the expectation if we assume $\mathbf{z}^\star$ has a uniform distribution over ${\mathcal{Z}}_T$. We are to show that all $\inf_{\mathbf{\hat z}^*_i} \underset{\mathbf{z}^*\in {\mathcal{Z}}_T}{\textnormal{ave}} \ {\mathbb{P}}(\mathbf{\hat z}^*_i \neq \mathbf{z}^*_i)$'s are lower bounded by the same quantity, which is the type-I plus type-II error of the testing problem \eqref{eq:fund_testing_problem_z_star}, so that for any $i\in T^c$ we would have
\begin{equation}
\label{eq:same_error_across_i_lower_bound_z_star}
\inf_{\mathbf{\hat z}^*} \sup_{\mathbf{z}^* \in {\mathcal{P}}_n} {\mathbb{E}} {\mathcal{L}}(\mathbf{\hat z}^*, \mathbf{z}^*) \geq \frac{|T^c|}{n}\cdot \inf_{\mathbf{\hat z}^*_i} \underset{\mathbf{z}^*\in {\mathcal{Z}}_T}{\textnormal{ave}} \ {\mathbb{P}}(\mathbf{\hat z}^*_i \neq \mathbf{z}^*_i) \gtrsim \delta_n \cdot \inf_{\mathbf{\hat z}^*_i} \underset{\mathbf{z}^*\in {\mathcal{Z}}_T}{\textnormal{ave}} \ {\mathbb{P}}(\mathbf{\hat z}^*_i \neq \mathbf{z}^*_i),
\end{equation}
which is the desired result.
Now let us fix any $i\in T^c$ and $S\subseteq[L]$. We then have
\begin{align*}
& \inf_{\mathbf{\hat z}^\star_i} \underset{\mathbf{z}^\star\in {\mathcal{Z}}_T}{\textnormal{ave}} \ {\mathbb{P}}(\mathbf{\hat z}^\star_i \neq \mathbf{z}^\star_i) \\
& = \inf_{\mathbf{\hat z}^\star_i} \underset{\mathbf{z}^\star_{-i}}{\textnormal{ave}} \ \underset{\mathbf{z}^\star_{i}}{\textnormal{ave}}\sum_{\xi\in \{\pm 1\}^S} {\mathbb{P}}( \mathbf{\hat z}^\star_i \neq \mathbf{z}^\star \ | \ \zzz{\ell}_i = \xi_\ell \mathbf{z}_i^\star \ \forall \ell\in S) \cdot {\mathbb{P}}(\zzz{\ell}_i = \xi_\ell \mathbf{z}_i^\star \ \forall \ell\in S) \\
& \geq \underset{\mathbf{z}^\star_{-i}}{\textnormal{ave}}\inf_{\mathbf{\hat z}^\star_i} \sum_{\xi\in \{\pm 1\}^S} \underset{\mathbf{z}^\star_{i}}{\textnormal{ave}} \ {\mathbb{P}}( \mathbf{\hat z}^\star_i \neq \mathbf{z}^\star \ | \ \zzz{\ell}_i = \xi_\ell \mathbf{z}_i^\star \ \forall \ell\in S) \cdot {\mathbb{P}}(\zzz{\ell}_i = \xi_\ell \mathbf{z}_i^\star \ \forall \ell\in S) \\
& = \frac{1}{2} \cdot \underset{\mathbf{z}^\star_{-i}}{\textnormal{ave}}\inf_{\mathbf{\hat z}^\star_i} \sum_{\xi\in \{\pm 1\}^S} \bigg({\mathbb{P}}( \mathbf{\hat z}^\star_i =-1 \ | \ \mathbf{z}_i^\star = 1, \zzz{\ell}_i = \xi_\ell \mathbf{z}_i^\star \ \forall \ell\in S) \cdot {\mathbb{P}}(\zzz{\ell}_i = \xi_\ell \mathbf{z}_i^\star \ \forall \ell\in S \ | \ \mathbf{z}_i^\star = 1) \\
& \qquad \qquad \qquad + {\mathbb{P}}( \mathbf{\hat z}^\star_i =1 \ | \ \mathbf{z}_i^\star = -1, \zzz{\ell}_i = \xi_\ell \mathbf{z}_i^\star \ \forall \ell\in S) \cdot {\mathbb{P}}(\zzz{\ell}_i = \xi_\ell \mathbf{z}_i^\star \ \forall \ell\in S \ | \ \mathbf{z}_i^\star = -1) \bigg).
\end{align*}
Because
\begin{align*}
{\mathbb{P}}(\zzz{\ell}_i = \xi_\ell \mathbf{z}_i^\star \ \forall \ell\in S \ | \ \mathbf{z}_i^\star = 1) & = {\mathbb{P}}(\zzz{\ell}_i = \xi_\ell \mathbf{z}_i^\star \ \forall \ell\in S \ | \ \mathbf{z}_i^\star = 1) = {\mathbb{P}}(\zzz{\ell}_i = \xi_\ell \mathbf{z}_i^\star \ \forall \ell\in S),
\end{align*}
we can write
\begin{align*}
& \inf_{\mathbf{\hat z}^\star_i} \underset{\mathbf{z}^\star}{\textnormal{ave}} \ {\mathbb{P}}(\mathbf{\hat z}^\star_i \neq \mathbf{z}^\star_i) \\
& \geq \frac{1}{2} \cdot \underset{\mathbf{z}^\star_{-i}}{\textnormal{ave}}\sum_{\xi\in \{\pm 1\}^S} {\mathbb{P}}(\zzz{\ell}_i = \xi_\ell \mathbf{z}_i^\star \ \forall \ell\in S)\\
& \qquad \times \inf_{\mathbf{\hat z}^\star_i}\bigg({\mathbb{P}}( \mathbf{\hat z}^\star_i =-1 \ | \ \mathbf{z}_i^\star = 1, \zzz{\ell}_i = \xi_\ell \mathbf{z}_i^\star \ \forall \ell\in S) + {\mathbb{P}}( \mathbf{\hat z}^\star_i =1 \ | \ \mathbf{z}_i^\star = -1, \zzz{\ell}_i = \xi_\ell \mathbf{z}_i^\star \ \forall \ell\in S) \bigg).
\end{align*}
Note that the term inside the parenthesis is the type-I plus type-II error of the following binary hypothesis testing problem:
\begin{equation}
\label{eq: test_without_knowing_community_in_each_layer}
H_0: \mathbf{z}^\star_i = 1 \ \ \ \textnormal{v.s.} \ \ \ H_1: \mathbf{z}^\star_i = -1,
\end{equation}
under the \emph{conditional law} of $\{\AAA{\ell}\} \ | \ \{\zzz{\ell}_i = \xi_\ell z_i^\star \ \forall \ell\in S\}$. Hence, this probability can be written as (with $d_{TV}$ denoting the total variation distance)
\begin{align*}
& 1 - d_{TV} \bigg[ \textnormal{law}\bigg(\{\AAA{\ell}\} \ \bigg| \{\mathbf{z}^\star_i = 1, \zzz{\ell}_i = \xi_\ell \mathbf{z}_i^\star \ \forall \ell\in S\}\bigg), \textnormal{law}\bigg(\{\AAA{\ell}\} \ \bigg| \{\mathbf{z}^\star_i = -1, \zzz{\ell}_i = \xi_\ell \mathbf{z}_i^\star \ \forall \ell\in S\}\bigg)\bigg] \\
& \geq 1 - d_{TV} \bigg[ \textnormal{law}\bigg(\{\AAA{\ell}, \zzz{\ell}\} \ \bigg| \{\mathbf{z}^\star_i = 1, \zzz{\ell}_i = \xi_\ell \mathbf{z}_i^\star \ \forall \ell\in S\}\bigg) , \textnormal{law}\bigg(\{\AAA{\ell}, \zzz{\ell}\} \ \bigg| \{\mathbf{z}^\star_i = -1, \zzz{\ell}_i = \xi_\ell \mathbf{z}_i^\star \ \forall \ell\in S\}\bigg)\bigg],
\end{align*}
where the inequality is by data-processing inequality for total variation: for any random variables $X, Y, X', Y'$, we have
$$
d_{TV} \bigg(\textnormal{law}(X), \textnormal{law}(X')\bigg) \leq d_{TV} \bigg(\textnormal{law}(X, Y), \textnormal{law}(X', Y')\bigg).
$$
Now the lower bound is the type-I plus type-II error of the same binary hypothesis testing problem, but under the \emph{joint law} of $\{\AAA{\ell}, \zzz{\ell}\} \ | \ \{\zzz{\ell}_i = \xi_\ell \mathbf{z}_i^\star \ \forall \ell\in S\}$.
We now begin to lower bound the optimal testing error of \eqref{eq: test_without_knowing_community_in_each_layer} under the joint law. For $j\neq i$, either (1) $j\in T$, so that $\mathbf{z}^\star_j = \pm 1$ depending on their membership of ${\mathcal{C}}_{\pm}$; or (2) $j\in T^c$, so that $\mathbf{z}^\star_j$ is unknown to us. Let $m_\pm$ be the number of positive (resp. negative) nodes apart from $i$. By construction, most of the positive nodes come from ${\mathcal{C}}_+$ and most of the negative nodes come from ${\mathcal{C}}_-$, where the word ``most'' is justified by $m_- = (1+o(1))n/2$ and $m_+ = (1+o(1))n/2$. Let $m = m_- \lor m_+ = (1+o(1))n/2$. By data-processing inequality for total variation, the testing error of \eqref{eq: test_without_knowing_community_in_each_layer} is further lower bounded by the testing error of the same test, but \emph{with $m$ positive nodes and $m$ negative nodes} (excluding $i$).
For notational simplicity, we consider the following equivalent setup: we have $2m +1$ nodes in total, where nodes $1, \hdots, m$ are labeled as $+1$, nodes $m +1,\hdots, 2m+ 1$ are labeled as $-1$, and the node labeled as $0$ (which is originally labeled as $i$) is the node whose community is to be decided. This is exactly the idealized setup in \eqref{eq:idealized_setup}.
Under the current notations, the joint density of $\{\AAA{\ell}, \zzz{\ell}\}$ is given by
\begingroup
\allowdisplaybreaks
\begin{align*}
&(1-\rho)^{\#\{\ell\in[L], 0\leq i\leq 2m: \zzz{\ell}_i = \mathbf{z}^\star_i\}} \rho^{nL - \#\{\ell\in[L], 0\leq i\leq 2m: \zzz{\ell}_i = \mathbf{z}^\star_i\}} \\
& \qquad \times \prod_{\ell\in[L]} \prod_{i\neq j} p_\ell^{\AAA{\ell}_{ij}} (1-p_\ell)^{1 - \AAA{\ell}_{ij}} \mathbf{1}{\zzz{\ell}_i = \zzz{\ell}_j} + q_\ell^{\AAA{\ell}_{ij}} (1-q_\ell)^{1 - \AAA{\ell}_{ij}} \mathbf{1}{\zzz{\ell}_i \neq \zzz{\ell}_j} \\
& = \rho^{nL} \bigg(\frac{1-\rho}{\rho}\bigg)^{\#\{\ell\in[L], 1\leq i\leq 2m: \zzz{\ell}_i = \mathbf{z}^\star_i\}} \cdot \bigg(\frac{1-\rho}{\rho}\bigg)^{\#\{\ell\in S: \zzz{\ell}_0 = \mathbf{z}^\star_0\} + \#\{\ell\in S^c: \zzz{\ell}_0 = \mathbf{z}^\star_0\}} \\
& \qquad \times \prod_{\ell\in[L]} \prod_{\substack{i\neq j \\ i\neq 0 \\ j\neq 0}} p_\ell^{\AAA{\ell}_{ij}} (1-p_\ell)^{1 - \AAA{\ell}_{ij}} \mathbf{1}{\zzz{\ell}_i = \zzz{\ell}_j} + q_\ell^{\AAA{\ell}_{ij}} (1-q_\ell)^{1 - \AAA{\ell}_{ij}} \mathbf{1}{\zzz{\ell}_i \neq \zzz{\ell}_j} \\
& \qquad \times \prod_{\ell\in S} \prod_{\substack{j\neq 0 \\ \zzz{\ell}_j = \zzz{\ell}_0 }} p_\ell^{\AAA{\ell}_{0j}} (1-p_\ell)^{1 - \AAA{\ell}_{0j}} \prod_{\substack{j\neq 0 \\ \zzz{\ell}_j \neq \zzz{\ell}_0 }} q_\ell^{\AAA{\ell}_{0j}} (1-q_\ell)^{1 - \AAA{\ell}_{0j}}\\
& \qquad \times \prod_{\ell\in S^c} \prod_{\substack{j\neq 0 \\ \zzz{\ell}_j = \zzz{\ell}_0 }} p_\ell^{\AAA{\ell}_{0j}} (1-p_\ell)^{1 - \AAA{\ell}_{0j}} \prod_{\substack{j\neq 0 \\ \zzz{\ell}_j \neq \zzz{\ell}_0 }} q_\ell^{\AAA{\ell}_{0j}} (1-q_\ell)^{1 - \AAA{\ell}_{0j}}.
\end{align*}
\endgroup
Hence, conditional on $\{\zzz{\ell}_0 = \xi_\ell z_0^\star \ \forall \ell\in S\}$, the density of $\{\AAA{\ell}, \zzz{\ell}\}$ becomes
\begingroup
\allowdisplaybreaks
\begin{align*}
& \rho^{nL} \bigg(\frac{1-\rho}{\rho}\bigg)^{\#\{\ell\in[L], 1\leq i\leq 2m: \zzz{\ell}_i = \mathbf{z}^\star_i\}} \cdot \bigg(\frac{1-\rho}{\rho}\bigg)^{\#\{\ell\in S: \xi_\ell = 1\} + \#\{\ell\in S^c: \zzz{\ell}_0 = \mathbf{z}^\star_0\}} \\
& \qquad \times \prod_{\ell\in[L]} \prod_{\substack{i\neq j \\ i\neq 0 \\ j\neq 0}} p_\ell^{\AAA{\ell}_{ij}} (1-p_\ell)^{1 - \AAA{\ell}_{ij}} \mathbf{1}{\zzz{\ell}_i = \zzz{\ell}_j} + q_\ell^{\AAA{\ell}_{ij}} (1-q_\ell)^{1 - \AAA{\ell}_{ij}} \mathbf{1}{\zzz{\ell}_i \neq \zzz{\ell}_j} \\
& \qquad \times \prod_{\ell\in S} \prod_{\substack{j\neq 0 \\ \zzz{\ell}_j = \xi_\ell \mathbf{z}^\star_0 }} p_\ell^{\AAA{\ell}_{0j}} (1-p_\ell)^{1 - \AAA{\ell}_{0j}} \prod_{\substack{j\neq 0 \\ \zzz{\ell}_j = - \xi_\ell \mathbf{z}^\star_0}} q_\ell^{\AAA{\ell}_{0j}} (1-q_\ell)^{1 - \AAA{\ell}_{0j}}\\
& \qquad \times \prod_{\ell\in S^c} \prod_{\substack{j\neq 0 \\ \zzz{\ell}_j = \zzz{\ell}_0 }} p_\ell^{\AAA{\ell}_{0j}} (1-p_\ell)^{1 - \AAA{\ell}_{0j}} \prod_{\substack{j\neq 0 \\ \zzz{\ell}_j \neq \zzz{\ell}_0 }} q_\ell^{\AAA{\ell}_{0j}} (1-q_\ell)^{1 - \AAA{\ell}_{0j}}.
\end{align*}
\endgroup
Let $L_0$ be the likelihood under the null and let $L_1$ be the likelihood under the alternative. The likelihood ratio is given by
\begin{align*}
\frac{L_0}{L_1} & = \bigg(\frac{1-\rho}{\rho}\bigg)^{\#\{\ell\in S^c: \zzz{\ell}_0 = 1 \} - \#\{\ell\in S^c: \zzz{\ell}_0 = -1 \}} \\
& \qquad \times \prod_{\ell\in S} \prod_{\substack{j\neq 0 \\ \zzz{\ell}_j = \xi_\ell }} p_\ell^{\AAA{\ell}_{0j}} (1-p_\ell)^{1 - \AAA{\ell}_{0j}} \prod_{\substack{j\neq 0 \\ \zzz{\ell}_j = - \xi_\ell }} q_\ell^{\AAA{\ell}_{0j}} (1-q_\ell)^{1 - \AAA{\ell}_{0j}} \\
& \qquad \times \bigg(\prod_{\ell\in S} \prod_{\substack{j\neq 0 \\ \zzz{\ell}_j = -\xi_\ell }} p_\ell^{\AAA{\ell}_{0j}} (1-p_\ell)^{1 - \AAA{\ell}_{0j}} \prod_{\substack{j\neq 0 \\ \zzz{\ell}_j = \xi_\ell }} q_\ell^{\AAA{\ell}_{0j}} (1-q_\ell)^{1 - \AAA{\ell}_{0j}}\bigg)^{-1}.
\end{align*}
By Neyman-Pearson lemma,
\begin{align*}
& \inf_{\mathbf{\hat z}^\star_0}\bigg({\mathbb{P}}( \mathbf{\hat z}^\star_0 =-1 \ | \ \mathbf{z}_0^\star = 1, \zzz{\ell}_0 = \xi_\ell \mathbf{z}_0^\star \ \forall \ell\in S) + {\mathbb{P}}( \mathbf{\hat z}^\star_0 =1 \ | \ \mathbf{z}_0^\star = -1, \zzz{\ell}_0 = \xi_\ell \mathbf{z}_0^\star \ \forall \ell\in S) \bigg) \\
& = {\mathbb{P}}\bigg( \frac{L_0}{L_1}\leq 1 \ \bigg| \ \mathbf{z}_0^\star = 1, \zzz{\ell}_0 = \xi_\ell \mathbf{z}_0^\star \ \forall \ell\in S\bigg) + {\mathbb{P}}\bigg( \frac{L_0}{L_1}\geq 1 \ | \ \mathbf{z}_0^\star = -1, \zzz{\ell}_0 = \xi_\ell \mathbf{z}_0^\star \ \forall \ell\in S\bigg).
\end{align*}
By symmetry, the two terms in the right-hand side above are equal to each other, so we focus on the first term. For notational simplicity, we let ${\mathbb{P}}_{H_0, S, \xi}$ to denote the conditional law of $\{\AAA{\ell}, \zzz{\ell}\} \ | \ \{\mathbf{z}_0^\star = 1, \zzz{\ell}_0 = \xi_\ell \mathbf{z}_0^\star \ \forall \ell\in S\}$. We then have
\begingroup
\allowdisplaybreaks
\begin{align*}
{\mathbb{P}}_{H_0, S, \xi} \bigg(\frac{L_0}{L_1} \leq 1\bigg)
& = {\mathbb{P}}_{H_0, S, \xi}\bigg[ \log\bigg(\frac{\rho}{1-\rho}\bigg)\cdot \sum_{\ell \in S^c} \bigg(\mathbf{1}{\zzz{\ell}_0 = 1} - \mathbf{1}{\zzz{\ell}_0 = -1} \bigg)\\
& \qquad\qquad \qquad + \sum_{\ell \in S} \sum_{\substack{j\neq 0 \\ \zzz{\ell}_j = \xi_\ell}} \AAA{\ell}_{0j} \log \bigg(\frac{q_\ell(1-p_\ell)}{p_\ell(1-q_\ell)}\bigg) + \log \bigg(\frac{1-q_\ell}{1-p_\ell}\bigg)\\
& \qquad\qquad \qquad + \sum_{\ell \in S} \sum_{\substack{j\neq 0 \\ \zzz{\ell}_j = -\xi_\ell}} \AAA{\ell}_{0j} \log \bigg(\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg) + \log \bigg(\frac{1-p_\ell}{1-q_\ell}\bigg)\geq 0\bigg] \\
& = {\mathbb{E}}_{\{\zzz{\ell}_{-0}: \ell\in S\}} \bigg\{{\mathbb{P}}_{H_0, S, \xi}\bigg[ \log\bigg(\frac{\rho}{1-\rho}\bigg)\cdot \sum_{\ell \in S^c}\bigg( \mathbf{1}{\zzz{\ell}_0 = 1} - \mathbf{1}{\zzz{\ell}_0 = -1} \bigg)\\
& \qquad \qquad \qquad + \sum_{\ell \in S} \sum_{\substack{j\neq 0 \\ \zzz{\ell}_j = \xi_\ell}} \AAA{\ell}_{0j} \log \bigg(\frac{q_\ell(1-p_\ell)}{p_\ell(1-q_\ell)}\bigg) + \log \bigg(\frac{1-q_\ell}{1-p_\ell}\bigg)\\
& \qquad\qquad \qquad + \sum_{\ell \in S} \sum_{\substack{j\neq 0 \\ \zzz{\ell}_j = -\xi_\ell}} \AAA{\ell}_{0j} \log \bigg(\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg) + \log \bigg(\frac{1-p_\ell}{1-q_\ell}\bigg)\geq 0 \ \bigg| \ \{\zzz{\ell}_{-0}\}\bigg]\bigg\}.
\end{align*}
\endgroup
The conditional probability ${\mathbb{P}}_{H_0, S, \xi}(\cdot \ | \ \{\zzz{\ell}_{-0}: \ell\in S\})$ in the right-hand side above is then equal to
\begin{align}
& {\mathbb{P}}\bigg( \log\bigg(\frac{\rho}{1-\rho}\bigg)\cdot \sum_{\ell \in S^c} \ZZZ{\ell} + \sum_{\ell \in S} \sum_{i = 1}^{\mmm{\ell}_1} \XXX{\ell}_i \cdot \log \bigg(\frac{q_\ell(1-p_\ell)}{p_\ell(1-q_\ell)}\bigg) + \sum_{\ell \in S}\sum_{i = 1}^{\mmm{\ell}_2} \YYY{\ell}_i \cdot \log \bigg(\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg) \nonumber\\
\label{eq:prf_lr_test_z_star}
&\qquad - \sum_{\ell \in S} (\mmm{\ell}_1- \mmm{\ell}_2) \cdot \log \bigg(\frac{1-p_\ell}{1-q_\ell}\bigg) \geq 0\bigg),
\end{align}
where
\begin{align*}
& \XXX{\ell}_i \overset{\textnormal{i.i.d.}}{\sim} \text{Bern}(p_\ell), \ \ \ \YYY{\ell}_i \overset{\textnormal{i.i.d.}}{\sim} \text{Bern}(q_\ell), \ \ \ \ZZZ{\ell} \overset{\textnormal{i.i.d.}}{\sim} 2\text{Bern}(1-\rho)-1, \\
& \mmm{\ell}_1 = \#\{j\neq 0: \zzz{\ell}_j = \xi_\ell\} , \ \ \ \mmm{\ell}_2= \#\{j\neq 0: \zzz{\ell}_j = -\xi_\ell\},
\end{align*}
and $\XXX{\ell}_i$'s, $\YYY{\ell}_i$'s and $\ZZZ{\ell}$'s are jointly independent.
Note that $\mmm{\ell}_1, \mmm{\ell}_2$ are treated as fixed when we condition on $\{\zzz{\ell}_{-0}: \ell\in S\}$.
With some algebra, one recognizes that \eqref{eq:prf_lr_test_z_star} is the type-I plus type-II error incurred by the likelihood ratio test of the following binary hypothesis testing problem:
\begin{align*}
H_0' & : W \sim \bigg(\bigotimes_{\ell \in S^c} 2\text{Bern}(1-\rho) -1 \bigg) \otimes \bigg(\bigotimes_{\ell \in S} \bigotimes_{i = 1}^{\mmm{\ell}_1}\text{Bern}(p_\ell) \bigg) \otimes \bigg(\bigotimes_{\ell \in S} \bigotimes_{i = 1}^{\mmm{\ell}_2}\text{Bern}(q_\ell) \bigg) \\
H_1' & : W \sim \bigg(\bigotimes_{\ell \in S^c} 2\text{Bern}(\rho) -1 \bigg) \otimes \bigg(\bigotimes_{\ell \in S} \bigotimes_{i = 1}^{\mmm{\ell}_1}\text{Bern}(q_\ell) \bigg) \otimes \bigg(\bigotimes_{\ell \in S} \bigotimes_{i = 1}^{\mmm{\ell}_2}\text{Bern}(p_\ell) \bigg) .
\end{align*}
By data processing equality for the total variation distance, we know that the optimal testing error of $H_0'$ v.s. $H_1'$ is lower bounded by the optimal testing error of the following testing problem:
\begin{align*}
H_0'' & : W \sim \bigg(\bigotimes_{\ell \in S^c} 2\text{Bern}(1-\rho) -1 \bigg) \otimes \bigg(\bigotimes_{\ell \in S} \bigotimes_{i = 1}^{\mmm{\ell}}\text{Bern}(p_\ell) \bigg) \otimes \bigg(\bigotimes_{\ell \in S} \bigotimes_{i = 1}^{\mmm{\ell}}\text{Bern}(q_\ell) \bigg) \\
H_1'' & : W \sim \bigg(\bigotimes_{\ell \in S^c} 2\text{Bern}(\rho) -1 \bigg) \otimes \bigg(\bigotimes_{\ell \in S} \bigotimes_{i = 1}^{\mmm{\ell}}\text{Bern}(q_\ell) \bigg) \otimes \bigg(\bigotimes_{\ell \in S} \bigotimes_{i = 1}^{\mmm{\ell}}\text{Bern}(p_\ell) \bigg),
\end{align*}
where we have let
$$
\mmm{\ell} = \mmm{\ell}_1 \lor \mmm{\ell}_2.
$$
This means that the probability in \eqref{eq:prf_lr_test_z_star} can be lower bounded by
\begin{align}
& {\mathbb{P}}\bigg( \log\bigg(\frac{\rho}{1-\rho}\bigg)\cdot \sum_{\ell \in S^c} \ZZZ{\ell} + \sum_{\ell \in S} \sum_{i = 1}^{\mmm{\ell}} \XXX{\ell}_i \cdot \log \bigg(\frac{q_\ell(1-p_\ell)}{p_\ell(1-q_\ell)}\bigg) + \sum_{\ell \in S}\sum_{i = 1}^{\mmm{\ell}} \YYY{\ell}_i \cdot \log \bigg(\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg) \geq 0\bigg)\nonumber,
\end{align}
Thus, we can lower bound $\inf_{\mathbf{\hat z}^\star_0} \textnormal{ave}_{\mathbf{z}^\star} {\mathbb{P}}(\mathbf{\hat z}^\star_0 \neq \mathbf{z}^\star_0)$ by taking the expectation w.r.t. $\{\zzz{\ell}: \ell \in S\}$:
\begingroup
\allowdisplaybreaks
\begin{align*}
\inf_{\mathbf{\hat z}^\star_0} \underset{\mathbf{z}^\star}{\textnormal{ave}}~ {\mathbb{P}}(\mathbf{\hat z}^\star_0 \neq \mathbf{z}^\star_0)
& \geq {\mathbb{E}}_{\{\zzz{\ell}_0: \ell \in S\}} {\mathbb{E}}_{\{\zzz{\ell}_{-0}: \ell \in S\}} \bigg[ {\mathbb{P}}\bigg( \log\bigg(\frac{\rho}{1-\rho}\bigg)\cdot \sum_{\ell \in S^c} \ZZZ{\ell} + \sum_{\ell \in S} \sum_{i = 1}^{\mmm{\ell}_1} \XXX{\ell}_i \cdot \log \bigg(\frac{q_\ell(1-p_\ell)}{p_\ell(1-q_\ell)}\bigg) \\
& \qquad \qquad + \sum_{\ell \in S}\sum_{i = 1}^{\mmm{\ell}_2} \YYY{\ell}_i \cdot \log \bigg(\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg)
- \sum_{\ell \in S} (\mmm{\ell}_1- \mmm{\ell}_2) \cdot \log \bigg(\frac{1-p_\ell}{1-q_\ell}\bigg) \geq 0\bigg) \bigg] \\
& \geq {\mathbb{E}}_{\{\zzz{\ell}_0: \ell \in S\}} {\mathbb{E}}_{\{\zzz{\ell}_{-0}: \ell \in S\}} \bigg[ {\mathbb{P}}\bigg( \log\bigg(\frac{\rho}{1-\rho}\bigg)\cdot \sum_{\ell \in S^c} \ZZZ{\ell} + \sum_{\ell \in S} \sum_{i = 1}^{\mmm{\ell}} \XXX{\ell}_i \cdot \log \bigg(\frac{q_\ell(1-p_\ell)}{p_\ell(1-q_\ell)}\bigg) \\
& \qquad \qquad + \sum_{\ell \in S}\sum_{i = 1}^{\mmm{\ell}} \YYY{\ell}_i \cdot \log \bigg(\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg)
\geq 0\bigg) \bigg] \\
& \geq {\mathbb{E}}_{\{\zzz{\ell}: \ell \in S\}} \bigg[\mathbf{1}_{E_\varepsilon} \cdot {\mathbb{P}}\bigg( \log\bigg(\frac{\rho}{1-\rho}\bigg)\cdot \sum_{\ell \in S^c} \ZZZ{\ell} + \sum_{\ell \in S} \sum_{i = 1}^{\mmm{\ell}} \XXX{\ell}_i \cdot \log \bigg(\frac{q_\ell(1-p_\ell)}{p_\ell(1-q_\ell)}\bigg) \\
& \qquad \qquad + \sum_{\ell \in S}\sum_{i = 1}^{\mmm{\ell}} \YYY{\ell}_i \cdot \log \bigg(\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg)
\geq 0\bigg) \bigg],
\end{align*}
\endgroup
where the event $E_\varepsilon$ is defined as
\begin{align*}
E_\varepsilon:= \bigg\{ \bigg|\frac{\mmm{\ell}_1}{m} - 1\bigg| \lor \bigg|\frac{\mmm{\ell}_2}{m} - 1\bigg| \leq \varepsilon \ \forall \ell \in S \bigg\}.
\end{align*}
Note that for a fixed $\ell \in [L]$, if $\zzz{\ell}_0 = \mathbf{z}^\star_0 = 1$ (i.e., $\xi_\ell = 1$), then we have
\begin{align*}
\mmm{\ell}_1 = \sum_{j = 1}^m \mathbf{1}{\zzz{\ell}_j = 1} + \sum_{j = m+1}^{2m}\mathbf{1}{\zzz{\ell}_j = 1},
\end{align*}
where for $1\leq j \leq m$, we have $\mathbf{1}{\zzz{\ell}_j = 1} \overset{\textnormal{i.i.d.}}{\sim} \text{Bern}(1-\rho)$, and for $m+1\leq j\leq 2m$, we have $\mathbf{1}{\zzz{\ell}_j = 1} \overset{\textnormal{i.i.d.}}{\sim} \text{Bern}(\rho)$. This gives
$
{\mathbb{E}}[\mmm{\ell}_1] = m,
$
and by Chernoff bound, we have
$$
{\mathbb{P}}( |\mmm{\ell}_1 - m |\geq m\varepsilon) \leq 2e^{-\varepsilon^2 m}.
$$
Since $\mmm{\ell}_1 + \mmm{\ell}_2 = 2m$, if $|\mmm{\ell}_1 - m |\leq m\varepsilon$, then we automatically have $|\mmm{\ell}_2 - m |\leq m\varepsilon$. Hence, we get
$$
{\mathbb{P}}( |\mmm{\ell}_1 - m | \lor |\mmm{\ell}_2 - m | \leq m\varepsilon) \geq 1- 2e^{-\varepsilon^2 m}.
$$
The above inequality also holds for the case of $\xi_\ell = -1$. Now, taking a union bound over $L$ and recalling $L\leq n^c$, we have
\begin{equation}
\label{eq:concentration_of_number_of_pos_and_neg_nodes}
{\mathbb{P}}(E_\varepsilon) \geq 1- 2L e^{-\varepsilon^2 m} \geq 1- 2e^{-\varepsilon^2 m + c \log n}.
\end{equation}
Choosing $\varepsilon = n^{-1/4}$, we have ${\mathbb{P}}(E_\varepsilon) = 1-e^{-c'\sqrt{n}}$ for some $c' > 0$. That is, with high probability, we have $\mmm{\ell} = (1+o(1))m$ uniformly over $\ell \in S$. Hence, by further applying data-processing inequality if necessary, we can lower bound $\inf_{\mathbf{\hat z}^\star_0} \textnormal{ave}_{\mathbf{z}^\star} {\mathbb{P}}(\mathbf{\hat z}^\star_0 \neq \mathbf{z}^\star_0)$ by
\begin{align*}
&{\mathbb{E}}_{\{\zzz{\ell}: \ell \in S\}} \bigg[\mathbf{1}_{E_\varepsilon} \cdot {\mathbb{P}}\bigg( \log\bigg(\frac{\rho}{1-\rho}\bigg)\cdot \sum_{\ell \in S^c} \ZZZ{\ell} + \sum_{\ell \in S} \sum_{i = 1}^{m'} \XXX{\ell}_i \cdot \log \bigg(\frac{q_\ell(1-p_\ell)}{p_\ell(1-q_\ell)}\bigg) \\
& \qquad \qquad + \sum_{\ell \in S}\sum_{i = 1}^{m'} \YYY{\ell}_i \cdot \log \bigg(\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg)
\geq 0\bigg) \bigg],
\end{align*}
where $m' = \max_{\ell\in S} \mmm{\ell} = (1+o(1))m = (1+o(1))n/2$. In the above display, the randomness of $\{\zzz{\ell}: \ell\in S\}$ only appears in the event $E_\varepsilon$. So the above display is equal to
\begin{align*}
&{\mathbb{P}}(E_\varepsilon) \cdot {\mathbb{P}}\bigg( \log\bigg(\frac{\rho}{1-\rho}\bigg)\cdot \sum_{\ell \in S^c} \ZZZ{\ell} + \sum_{\ell \in S} \sum_{i = 1}^{m'} \XXX{\ell}_i \cdot \log \bigg(\frac{q_\ell(1-p_\ell)}{p_\ell(1-q_\ell)}\bigg) + \sum_{\ell \in S}\sum_{i = 1}^{m'} \YYY{\ell}_i \cdot \log \bigg(\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg) \geq 0\bigg) \\
& \gtrsim {\mathbb{P}}\bigg( \log\bigg(\frac{\rho}{1-\rho}\bigg)\cdot \sum_{\ell \in S^c} \ZZZ{\ell} + \sum_{\ell \in S} \sum_{i = 1}^{m'} \XXX{\ell}_i \cdot \log \bigg(\frac{q_\ell(1-p_\ell)}{p_\ell(1-q_\ell)}\bigg) + \sum_{\ell \in S}\sum_{i = 1}^{m'} \YYY{\ell}_i \cdot \log \bigg(\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg) \geq 0\bigg),
\end{align*}
where the inequality is by ${\mathbb{P}}(E_\varepsilon) = 1-o(1)$. Finally, we conclude the proof by recalling \eqref{eq:same_error_across_i_lower_bound_z_star} and noting that the right-hand side above is the error incurred by the likelihood ratio test for the testing problem \eqref{eq:fund_testing_problem_z_star} with $m' = (1+\delta_n')n/2$.
\subsection{Proof of Lemma \ref{lemma:opt_test_error_z_star}}\label{subappend:prf_opt_test_error_z_star}
We will give a lower bound for
\begin{align}
\label{eq:target_of_lower_bound_z_star}
{\mathbb{P}}\bigg(\sum_{\ell\in S} \log \bigg(\frac{q_\ell(1-p_\ell)}{p_\ell(1-q_\ell)}\bigg) \cdot \sum_{i=1}^{m} (\XXX{\ell}_i-\YYY{\ell}_i) \geq \log \bigg(\frac{1-\rho}{\rho}\bigg) \cdot \sum_{\ell\in S^c}\ZZZ{\ell}\bigg),
\end{align}
where $m = (1+\delta'_n)n/2 = (1+o(1))n/2$. We begin by decomposing the above probability as
\begin{align*}
{\mathbb{E}}_{\ZZZ{S^c}}\bigg[{\mathbb{P}}\bigg(\sum_{\ell \in S} \sum_{i = 1}^{m} \XXX{\ell}_i \cdot \log \bigg(\frac{q_\ell(1-p_\ell)}{p_\ell(1-q_\ell)}\bigg) + \YYY{\ell}_i \cdot \log \bigg(\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg) \geq \log\bigg(\frac{1-\rho}{\rho}\bigg)\cdot \sum_{\ell\in S^c}\ZZZ{\ell} \ \bigg| \ \ZZZ{S^c} \bigg)\bigg].
\end{align*}
Consider the moment generating function
\begin{align*}
\phi_S(t) & = {\mathbb{E}}\bigg[\exp\bigg\{ t \times \sum_{\ell \in S} \sum_{i = 1}^{m} \XXX{\ell}_i \cdot \log \bigg(\frac{q_\ell(1-p_\ell)}{p_\ell(1-q_\ell)}\bigg) + \YYY{\ell}_i \cdot \log \bigg(\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg) \bigg\}\bigg]\\
& = \prod_{\ell \in S} \prod_{i\in[m]} \bigg[ p_\ell \bigg(\frac{q_\ell(1-p_\ell)}{p_\ell (1-q_\ell)}\bigg)^t + (1-p_\ell)\bigg] \bigg[q_\ell \bigg(\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg)^t + (1-q_\ell)\bigg] \\
\label{eq:mgf_opt_test_z_star}
& = \prod_{\ell\in S} \bigg(p_\ell^{1-t}q_\ell^t + (1-p_\ell)^{1-t}(1-q_\ell)^{t}\bigg)^m\bigg(p_\ell^t q_\ell^{1-t} + (1-p_\ell)^t (1-q_\ell)^{1-t}\bigg)^m.\addtocounter{equation}{1}\tag{\theequation}
\end{align*}
The information-theoretic quantity $\psi_S(t)$ defined in \eqref{eq:cum_gen_fun} is the corresponding cumulant generating function (with $m$ replaced by $n/2$):
\begin{align}
\label{eq:cgf_opt_test_z_star}
\psi_S(t) & = \sum_{\ell \in S} \frac{n}{2} \cdot \bigg[\log\bigg(p_\ell^{1-t}q_\ell^t + (1-p_\ell)^{1-t}(1-q_\ell)^{t}\bigg) + \log\bigg(p_\ell^t q_\ell^{1-t} + (1-p_\ell)^t (1-q_\ell)^{1-t}\bigg)\bigg].
\end{align}
And we recall that
\begin{align*}
\psi_S^\star(a) = \sup_{0\leq t\leq 1} \{ at - \psi_S(t)\}.
\end{align*}
Now, treating $\ZZZ{S^c}$ as fixed, we let
\begin{align}
\label{eq:original_law}
\mu &= \textnormal{Law} \bigg[ \sum_{\ell \in S} \sum_{i = 1}^{m} \XXX{\ell}_i \cdot \log \bigg(\frac{q_\ell(1-p_\ell)}{p_\ell(1-q_\ell)}\bigg) + \YYY{\ell}_i \cdot \log \bigg(\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg) \bigg] \\
\label{eq:tilted_law}
\tilde \mu_t & = \Conv_{\ell\in S} \Conv_{i = 1}^{m} \textnormal{Law} (\tXXX{\ell}_t \circledast \tYYY{\ell}_t),
\end{align}
where $\circledast$ denotes the convolution of random variables and
\begin{equation}
\label{eq:tilted_rvs}
\tXXX{\ell}_t + \tYYY{\ell}_t =
\begin{cases}
\log \frac{q_\ell(1-p_\ell)}{p_\ell(1-q_\ell)} & \textnormal{w.p. } \frac{[p_\ell(1-q_\ell)]^{1-t} [q_\ell(1-p_\ell)]^t }{p_\ell q_\ell + (1-p_\ell)(1-q_\ell) + [p_\ell(1-q_\ell)]^{1-t} [q_\ell(1-p_\ell)]^t + [p_\ell(1-q_\ell)]^{t} [q_\ell(1-p_\ell)]^{1-t}} \\
\log \frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)} & \textnormal{w.p. } \frac{[p_\ell(1-q_\ell)]^{t} [q_\ell(1-p_\ell)]^{1-t} }{p_\ell q_\ell + (1-p_\ell)(1-q_\ell) + [p_\ell(1-q_\ell)]^{1-t} [q_\ell(1-p_\ell)]^t + [p_\ell(1-q_\ell)]^{t} [q_\ell(1-p_\ell)]^{1-t}}\\
0 & \textnormal{w.p. } \frac{p_\ell q_\ell + (1-p_\ell)(1-q_\ell)}{p_\ell q_\ell + (1-p_\ell)(1-q_\ell) + [p_\ell(1-q_\ell)]^{1-t} [q_\ell(1-p_\ell)]^t + [p_\ell(1-q_\ell)]^{t} [q_\ell(1-p_\ell)]^{1-t}}.
\end{cases}
\end{equation}
Here, the random variables $\tXXX{\ell}_t$ and $\tYYY{\ell}_t$ are exponentially tilted version of $\XXX{\ell}_i \cdot \log \frac{q_\ell(1-p_\ell)}{p_\ell (1-q_\ell)}$ and $\YYY{\ell}_i \cdot \log \frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}$, respectively. As discussed in Remark \ref{rmk:parity}, the lower bound for \eqref{eq:target_of_lower_bound_z_star} depends on the parity of $|S^c|$.
\subsubsection{The case of even \texorpdfstring{$|S^c|$}{Sc}}
Assume $|S^c|$ is even. If $S^c = \varnothing$, then we have
\begin{align*}
&{\mathbb{E}}_{\ZZZ{S^c}}\bigg[{\mathbb{P}}\bigg(\sum_{\ell \in S} \sum_{i = 1}^{m} \XXX{\ell}_i \cdot \log \bigg(\frac{q_\ell(1-p_\ell)}{p_\ell(1-q_\ell)}\bigg) + \YYY{\ell}_i \cdot \log \bigg(\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg) \geq \log\bigg(\frac{1-\rho}{\rho}\bigg)\cdot \sum_{\ell\in S^c}\ZZZ{\ell} \ \bigg| \ \ZZZ{S^c} \bigg)\bigg]\\
& = {\mathbb{P}}\bigg(\sum_{\ell \in S} \sum_{i = 1}^{m} \XXX{\ell}_i \cdot \log \bigg(\frac{q_\ell(1-p_\ell)}{p_\ell(1-q_\ell)}\bigg) + \YYY{\ell}_i \cdot \log \bigg(\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg) \geq 0 \bigg).
\end{align*}
If $S^c \neq \varnothing$, we can write
\begin{align*}
&{\mathbb{E}}_{\ZZZ{S^c}}\bigg[{\mathbb{P}}\bigg(\sum_{\ell \in S} \sum_{i = 1}^{m} \XXX{\ell}_i \cdot \log \bigg(\frac{q_\ell(1-p_\ell)}{p_\ell(1-q_\ell)}\bigg) + \YYY{\ell}_i \cdot \log \bigg(\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg) \geq \log\bigg(\frac{1-\rho}{\rho}\bigg)\cdot \sum_{\ell\in S^c}\ZZZ{\ell} \ \bigg| \ \ZZZ{S^c} \bigg)\bigg]\\
& = \sum_{x \in \{-|S^c|+ 2j: 0\leq j\leq |S^c|\}}\binom{|S^c|}{\frac{|S^c| + x}{2}} (1-\rho)^{\frac{|S^c| + x}{2}} \rho^{\frac{|S^c| - x}{2}}\\
& \qquad \times {\mathbb{P}}\bigg(\sum_{\ell \in S} \sum_{i = 1}^{m} \XXX{\ell}_i \cdot \log \bigg(\frac{q_\ell(1-p_\ell)}{p_\ell(1-q_\ell)}\bigg) + \YYY{\ell}_i \cdot \log \bigg(\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg) \geq \log\bigg(\frac{1-\rho}{\rho}\bigg)\cdot x \bigg)\\
& \geq \binom{|S^c|}{\frac{|S^c|}{2}} \bigg((1-\rho)\rho\bigg)^{\frac{|S^c|}{2}}\times {\mathbb{P}}\bigg(\sum_{\ell \in S} \sum_{i = 1}^{m} \XXX{\ell}_i \cdot \log \bigg(\frac{q_\ell(1-p_\ell)}{p_\ell(1-q_\ell)}\bigg) + \YYY{\ell}_i \cdot \log \bigg(\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg) \geq 0\bigg)\\
& \geq 2^{\frac{|S^c|}{2}} \cdot \exp\bigg\{-(1+o(1))|S^c|J_\rho \bigg\}\times {\mathbb{P}}\bigg(\sum_{\ell \in S} \sum_{i = 1}^{m} \XXX{\ell}_i \cdot \log \bigg(\frac{q_\ell(1-p_\ell)}{p_\ell(1-q_\ell)}\bigg) + \YYY{\ell}_i \cdot \log \bigg(\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg) \geq 0\bigg)\\
& \geq \exp\bigg\{-(1+o(1))|S^c|J_\rho \bigg\}\times {\mathbb{P}}\bigg(\sum_{\ell \in S} \sum_{i = 1}^{m} \XXX{\ell}_i \cdot \log \bigg(\frac{q_\ell(1-p_\ell)}{p_\ell(1-q_\ell)}\bigg) + \YYY{\ell}_i \cdot \log \bigg(\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg) \geq 0\bigg),
\end{align*}
where we have used the fact that $|S^c|$ is even, $\binom{n}{k}\geq (n/k)^k$, and
$$
J_\rho := -\log2\sqrt{\rho(1-\rho)} = -(1+o(1))\log\sqrt{\rho(1-\rho)}
$$
when $\rho = o(1)$. Thus, in both cases, \eqref{eq:target_of_lower_bound_z_star} can be lower bounded by
$$
\exp\bigg\{-(1+o(1))|S^c|J_\rho \bigg\}\times {\mathbb{P}}\bigg(\sum_{\ell \in S} \sum_{i = 1}^{m} \XXX{\ell}_i \cdot \log \bigg(\frac{q_\ell(1-p_\ell)}{p_\ell(1-q_\ell)}\bigg) + \YYY{\ell}_i \cdot \log \bigg(\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg) \geq 0\bigg).
$$
Hence, we focus on lower bounding the following probability:
\begin{equation}
\label{eq:target_of_lower_bound_z_star_even}
{\mathbb{P}}\bigg(\sum_{\ell \in S} \sum_{i = 1}^{m} \XXX{\ell}_i \cdot \log \bigg(\frac{q_\ell(1-p_\ell)}{p_\ell(1-q_\ell)}\bigg) + \YYY{\ell}_i \cdot \log \bigg(\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg) \geq 0\bigg) = \mu([0, \infty)).
\end{equation}
By a standard exponential tilting argument (also known as the Cram\'er-Chernoff argument, a technique commonly used in proving large deviation principles), we have
\begin{align*}
\mu([ 0, \infty))& = e^{-(1+o(1))\psi_S(t)} \cdot {\mathbb{E}}_{W \sim \tilde\mu_t}\bigg[e^{-tW} \cdot \mathbf{1}{W \geq 0}\bigg]\\
& \geq \exp\bigg\{-(1+o(1))\psi_S(t)-t\xi\bigg\} \times {\mathbb{P}}_{W\sim \tilde\mu_t}(0\leq W\leq \xi),
\end{align*}
where the $(1+o(1))$ term comes from $m = (1+o(1))n/2$.
Choosing $t = 1/2$, we arrive at the following lower bound:
\begin{align*}
\mu([ 0, \infty))& \geq \exp\bigg\{-(1+o(1))\sum_{\ell\in S}m\III{\ell}_{1/2}-\xi/2\bigg\} \times {\mathbb{P}}_{W\sim \tilde\mu_{1/2}}(0\leq W\leq \xi).
\end{align*}
We need the following lemma.
\begin{lemma}
\label{lemma: exp_and_var_of_tilted_bernoulli}
Assume $p_\ell \asymp q_\ell = o(1)$ for all $\ell\in[L]$. Then for any $t\in[0, 1]$, we have
\begin{align*}
{\mathbb{E}}[\tXXX{\ell}_t + \tYYY{\ell}_t] &\asymp \III{\ell}_{1/2} \cdot \frac{p_\ell^t q_\ell^{1-t} - p_\ell^{1-t}q_\ell^t}{p_\ell - q_\ell},\\
\mathrm{Var}(\tXXX{\ell}_t + \tYYY{\ell}_t) & \asymp \III{\ell}_{1/2}.
\end{align*}
\end{lemma}
\begin{proof}
we have
\begin{align*}
{\mathbb{E}}[\tXXX{\ell}_t + \tYYY{\ell}_t] & = \bigg(\log \frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg) \\
& \qquad \times \frac{[p_\ell(1-q_\ell)]^{t} [q_\ell(1-p_\ell)]^{1-t} - [p_\ell(1-q_\ell)]^{1-t} [q_\ell(1-p_\ell)]^t }{p_\ell q_\ell + (1-p_\ell)(1-q_\ell) + [p_\ell(1-q_\ell)]^{1-t} [q_\ell(1-p_\ell)]^t + [p_\ell(1-q_\ell)]^{t} [q_\ell(1-p_\ell)]^{1-t}}.
\end{align*}
Recall that
$$
\log\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)} = \log \bigg(1 + \frac{p_\ell- q_\ell}{q_\ell}\bigg) + \log \bigg(1 + \frac{p_\ell - q_\ell}{1-p_\ell}\bigg) \asymp \frac{p_\ell -q_\ell}{p_\ell}.
$$
Also, note that for any $t\in [0, 1]$, we have
\begin{align*}
& \frac{[p_\ell(1-q_\ell)]^{t} [q_\ell(1-p_\ell)]^{1-t} - [p_\ell(1-q_\ell)]^{1-t} [q_\ell(1-p_\ell)]^t }{p_\ell q_\ell + (1-p_\ell)(1-q_\ell) + [p_\ell(1-q_\ell)]^{1-t} [q_\ell(1-p_\ell)]^t + [p_\ell(1-q_\ell)]^{t} [q_\ell(1-p_\ell)]^{1-t}} \\
& = (1+o(1)) \cdot( p_\ell^{t} q_\ell^{1-t} - p_\ell^{1-t} q_\ell^t).
\end{align*}
This gives
\begin{align*}
{\mathbb{E}}[\tXXX{\ell}_t + \tYYY{\ell}_t] & \asymp \frac{(p_\ell-q_\ell)^2}{p_\ell} \cdot \frac{( p_\ell^{t} q_\ell^{1-t} - p_\ell^{1-t} q_\ell^t)}{p_\ell - q_\ell}\\
& \asymp \III{\ell}_{1/2} \cdot \frac{( p_\ell^{t} q_\ell^{1-t} - p_\ell^{1-t} q_\ell^t)}{p_\ell - q_\ell},
\end{align*}
where the last line is by \eqref{eq:asymp_equiv_of_I_1/2}.
We then compute its second moment:
\begin{align*}
{\mathbb{E}}[(\tXXX{\ell}_t + \tYYY{\ell}_t)^2] & = \bigg(\log\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg)^2 \cdot \frac{[p_\ell(1-q_\ell)]^t[q_\ell(1-p_\ell)]^{1-t} + [p_\ell(1-q_\ell)]^{1-t}[q_\ell(1-p_\ell)]^{t}}{1+o(1)}\\
& \asymp \bigg(\frac{p_\ell - q_\ell}{p_\ell}\bigg)^2 \cdot p_\ell
= \frac{(p_\ell-q_\ell)^2}{p_\ell}
\asymp \III{\ell}_{1/2},
\end{align*}
which gives
$$
\mathrm{Var}(\tXXX{\ell}_t + \tYYY{\ell}_t) \lesssim \III{\ell}_{1/2}.
$$
On the other hand, note that
\begin{align*}
\frac{p_\ell^{t} q_\ell^{1-t} - p_\ell^{1-t} q_\ell^t}{p_\ell - q_\ell} & = \frac{(p_\ell/q_\ell)^t - (p_\ell/q_\ell)^{1-t}}{p_\ell/q_\ell - 1}\in [-1, 1].
\end{align*}
Hence,
$$
\mathrm{Var}(\tXXX{\ell}_t + \tYYY{\ell}_t) \gtrsim \III{\ell}_{1/2} - (\III{\ell}_{1/2})^2 \asymp \III{\ell}_{1/2},
$$
which gives the desired result.
\end{proof}
Now, by lemma \ref{lemma: exp_and_var_of_tilted_bernoulli} and Chebyshev's inequality, we have
$$
{\mathbb{P}}_{W\sim \tilde\mu_{1/2}} \bigg(|W|>c \sqrt{\frac{\sum_{\ell\in S} m\III{\ell}_{1/2}}{\delta}}\bigg) \leq \delta,
$$
where $c$ is some absolute constant. Note that with $t = 1/2$, the random variable $\tXXX{\ell}_{1/2} + \tXXX{\ell}_{1/2}$ is actually \emph{symmetric about zero}. This gives
$$
{\mathbb{P}}_{W\sim \tilde \mu_{1/2}}\bigg(0\leq W\leq c \sqrt{\frac{\sum_{\ell\in S} m \III{\ell}_{1/2}}{\delta}}\bigg) \geq \frac{1-\delta}{2}.
$$
Hence, by properly choosing $\delta$, we have
\begin{align*}
\mu([0, \infty)) \gtrsim \bigg\{-(1+o(1)) \bigg(\sum_{\ell \in S} m\III{\ell}_{1/2} - c \sqrt{\sum_{\ell \in S}m\III{\ell}_{1/2}}\bigg)\bigg\}.
\end{align*}
This gives the following lower bound for \eqref{eq:target_of_lower_bound_z_star}:
\begin{align*}
& C \cdot \exp\bigg\{-\bigl(1+o(1)\bigr) \cdot \bigg(|S^c| J_\rho + \sum_{\ell \in S}m\III{\ell}_{1/2} + c\sqrt{\sum_{\ell\in S}m\III{\ell}_{1/2}}\bigg) \bigg\}
\end{align*}
Consider the following three cases.
\begin{enumerate}
\item In this case, we assume $S^c = \varnothing$. Now, if $\sum_{\ell \in S}m\III{\ell}_{1/2} = \sum_{\ell \in [L]}m\III{\ell}_{1/2}\to \infty$, then the lower bound for \eqref{eq:target_of_lower_bound_z_star} becomes
$$
C\cdot\exp\bigg\{-\bigl(1+o(1)\bigr) \cdot \sum_{\ell \in [L]}m\III{\ell}_{1/2}\bigg\}.
$$
On the other hand, if $\sum_{\ell \in [L]}m\III{\ell}_{1/2} = O(1)$, then the corresponding lower bound is again
$$
C\cdot\exp\bigg\{-\bigl(1+o(1)\bigr) \cdot \sum_{\ell \in [L]}m\III{\ell}_{1/2}\bigg\},
$$
because $\sqrt{\sum_{\ell\in[L]} m\III{\ell}_{1/2}} = O(1)$.
\item In this case, we assume $S^c \neq \varnothing$ and $\sum_{\ell \in S} m\III{\ell}_{1/2} \geq J_\rho$. Since $J_\rho \to\infty$, $\sum_{\ell\in S} m\III{\ell}_{1/2}\to\infty$. Now the lower bound for \eqref{eq:target_of_lower_bound_z_star} becomes
\begin{align*}
& c\cdot \exp\bigg\{-\bigl(1+o(1)\bigr)|S^c| J_\rho + \bigg(1+o(1) + \frac{c}{\sqrt{\sum_{\ell\in S}m\III{\ell}_{1/2}}}\bigg)\cdot \sum_{\ell \in S}m\III{\ell}_{1/2}\bigg] \bigg\}\\
& =c\cdot \exp\bigg\{-\bigl(1+o(1)\bigr) \cdot \bigg(|S^c| J_\rho + \sum_{\ell \in S}m\III{\ell}_{1/2}\bigg) \bigg\}.
\end{align*}
\item If $S^c \neq \varnothing$ and $\sum_{\ell \in S}m\III{\ell}_{1/2} < J_\rho$, then the lower bound for \eqref{eq:target_of_lower_bound_z_star} becomes
\begin{align*}
& C\cdot \exp\bigg\{-\bigg(1+o(1) + \frac{c\sqrt{\sum_{\ell\in S}m\III{\ell}_{1/2}}}{J_\rho}\bigg)|S^c| J_\rho + \bigl(1+o(1)\bigr) \cdot \sum_{\ell \in S}m\III{\ell}_{1/2}\bigg] \bigg\}\\
& = C\cdot \exp\bigg\{-\bigl(1+o(1)\bigr) \cdot \bigg(|S^c| J_\rho + \sum_{\ell \in S}m\III{\ell}_{1/2} \bigg) \bigg\}.
\end{align*}
\end{enumerate}
Hence, for any $|S^c|$ even, we have the following lower bound for \eqref{eq:target_of_lower_bound_z_star}:
\begin{align*}
& C\cdot \exp\bigg\{-\bigl(1+o(1)\bigr) \cdot \bigg(|S^c| J_\rho + \sum_{\ell \in S}mI_\ell \bigg) \bigg\} = C\cdot\exp\bigg\{-\bigl(1+o(1)\bigr) \cdot \bigg(|S^c| J_\rho + \psi_S^\star(0)\bigg) \bigg\},
\end{align*}
where we remark that the $o(1)$ term does not depend on $S$, thus proving the first part of \eqref{eq:opt_test_error_z_star}.
\subsubsection{The case of odd \texorpdfstring{$|S^c|$}{Sc}}\label{subsubappend:prf_opt_test_error_z_star_odd_Sc}
Now consider the case where $|S^c|$ is odd. Similar to the previous case, we start by writing
\begingroup
\allowdisplaybreaks
\begin{align*}
&{\mathbb{E}}_{\ZZZ{S^c}}\bigg[{\mathbb{P}}\bigg(\sum_{\ell \in S} \sum_{i = 1}^{m} \XXX{\ell}_i \cdot \log \bigg(\frac{q_\ell(1-p_\ell)}{p_\ell(1-q_\ell)}\bigg) + \YYY{\ell}_i \cdot \log \bigg(\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg) \geq \log\bigg(\frac{1-\rho}{\rho}\bigg)\cdot \sum_{\ell\in S^c}\ZZZ{\ell} \ \bigg| \ \ZZZ{S^c} \bigg)\bigg]\\
& \geq \sum_{x \in \{-|S^c|+ 2j: 0\leq j\leq |S^c|\}}\binom{|S^c|}{\frac{|S^c| + x}{2}} (1-\rho)^{\frac{|S^c| + x}{2}} \rho^{\frac{|S^c| - x}{2}}\\
& \qquad \times {\mathbb{P}}\bigg(\sum_{\ell \in S} \sum_{i = 1}^{m} \XXX{\ell}_i \cdot \log \bigg(\frac{q_\ell(1-p_\ell)}{p_\ell(1-q_\ell)}\bigg) + \YYY{\ell}_i \cdot \log \bigg(\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg) \geq \log\bigg(\frac{1-\rho}{\rho}\bigg)\cdot x \bigg)\\
& \geq \binom{|S^c|}{\frac{|S^c + 1|}{2}} \bigg(\rho(1-\rho)\bigg)^{\frac{|S^c|}{2}} \cdot \bigg(\frac{1-\rho}{\rho}\bigg)^{-1/2} \\
& \qquad \times {\mathbb{P}}\bigg(\sum_{\ell \in S} \sum_{i = 1}^{m} \XXX{\ell}_i \cdot \log \bigg(\frac{q_\ell(1-p_\ell)}{p_\ell(1-q_\ell)}\bigg) + \YYY{\ell}_i \cdot \log \bigg(\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg) \geq \log\bigg(\frac{\rho}{1-\rho}\bigg) \bigg)\\
& \geq \exp\bigg\{- \bigl(1+o(1)\bigr) \bigg(|S^c| + 1\bigg) J_\rho \bigg\}\\
& \qquad \times {\mathbb{P}}\bigg(\sum_{\ell \in S} \sum_{i = 1}^{m} \XXX{\ell}_i \cdot \log \bigg(\frac{q_\ell(1-p_\ell)}{p_\ell(1-q_\ell)}\bigg) + \YYY{\ell}_i \cdot \log \bigg(\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg) \geq \log\bigg(\frac{\rho}{1-\rho}\bigg) \bigg).
\end{align*}
\endgroup
where the second inequality is by choosing $x = -1$ (note that we can do so because $|S^c|$ is odd).
Hence, we focus on lower bounding the following probability:
\begin{equation}
\label{eq:target_of_lower_bound_z_star_odd}
{\mathbb{P}}\bigg(\sum_{\ell \in S} \sum_{i = 1}^{m} \XXX{\ell}_i \cdot \log \bigg(\frac{q_\ell(1-p_\ell)}{p_\ell(1-q_\ell)}\bigg) + \YYY{\ell}_i \cdot \log \bigg(\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg) \geq \log\bigg(\frac{\rho}{1-\rho}\bigg) \bigg).
\end{equation}
With some algebra, we have
\begin{align*}
\sum_{\ell\in S} \sum_{i=1}^{m}{\mathbb{E}}[\tXXX{\ell}_0 + \tYYY{\ell}_0] = -\sum_{\ell\in S} m(p_\ell - q_\ell) \cdot \log\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)} \asymp -\sum_{\ell\in S}m \III{\ell}_{1/2},
\end{align*}
where the asymptotic equivalence is by Lemma \ref{lemma: exp_and_var_of_tilted_bernoulli}. We divide our discussion into three cases.
~\\
{\bf Case A.}
In this case, we assume
$$
\sum_{\ell\in S} m (p_\ell - q_\ell) \cdot \log\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)} \leq \sqrt{J_\rho}.
$$
Hence, we have
$$
\sum_{\ell \in S} m\III{\ell}_{1/2} \lesssim \sqrt{J_\rho}.
$$
We now have
\begingroup
\allowdisplaybreaks
\begin{align*}
& {\mathbb{P}}\bigg(\sum_{\ell \in S} \sum_{i = 1}^{m} \XXX{\ell}_i \cdot \log \bigg(\frac{q_\ell(1-p_\ell)}{p_\ell(1-q_\ell)}\bigg) + \YYY{\ell}_i \cdot \log \bigg(\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg) \geq \log\bigg(\frac{\rho}{1-\rho}\bigg) \bigg)\\
& = 1 - {\mathbb{P}}\bigg(\sum_{\ell \in S} \sum_{i = 1}^{m} \XXX{\ell}_i \cdot \log \bigg(\frac{q_\ell(1-p_\ell)}{p_\ell(1-q_\ell)}\bigg) + \YYY{\ell}_i \cdot \log \bigg(\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg) <\log\bigg(\frac{\rho}{1-\rho}\bigg) \bigg)\\
& = 1 - {\mathbb{P}}\bigg(-\sum_{\ell \in S} \sum_{i = 1}^{m}\bigg(\tXXX{\ell}_0 + \tYYY{\ell}_0\bigg) > \log\bigg(\frac{1-\rho}{\rho}\bigg) \bigg)\\
& \geq 1- \exp\bigg\{- \frac{1}{2}\log \bigg(\frac{1-\rho}{\rho}\bigg)\bigg\} \cdot \prod_{\ell\in S} \prod_{i=1}^{m} {\mathbb{E}} \exp\bigg\{ -\frac{\tXXX{\ell}_0 + \tYYY{\ell}_0}{2}\bigg\} \\
& = 1 - e^{-\bigl(1+o(1)\bigr)J_\rho} \cdot \prod_{\ell\in S} \prod_{i = 1}^{m} \frac{\bigl(p_\ell(1-q_\ell)\bigr)^{3/2}}{\bigl(q_\ell(1-p_\ell)\bigr)^{1/2}} + \frac{\bigl(q_\ell(1-p_\ell)\bigr)^{3/2}}{\bigl(p_\ell(1-q_\ell)\bigr)^{1/2}} + 1 - p_\ell(1-q_\ell) - q_\ell(1-p_\ell),
\end{align*}
\endgroup
where the fourth line is by Markov's inequality.
Note that
\begin{align*}
& \frac{\bigl(p_\ell(1-q_\ell)\bigr)^{3/2}}{\bigl(q_\ell(1-p_\ell)\bigr)^{1/2}} + \frac{\bigl(q_\ell(1-p_\ell)\bigr)^{3/2}}{\bigl(p_\ell(1-q_\ell)\bigr)^{1/2}} + 1 - p_\ell(1-q_\ell) - q_\ell(1-p_\ell) \\
& = 1 - p_\ell - q_\ell + 2p_\ell q_\ell + \frac{\bigl(p_\ell(1-q_\ell)\bigr)^{3/2}}{\bigl(q_\ell(1-p_\ell)\bigr)^{1/2}} + \frac{\bigl(q_\ell(1-p_\ell)\bigr)^{3/2}}{\bigl(p_\ell(1-q_\ell)\bigr)^{1/2}}\\
& = 1 - \bigl(\sqrt{p_\ell} -\sqrt{q_\ell}\bigr)^2 - 2 \sqrt{p_\ell q_\ell} + 2p_\ell q_\ell + \frac{\bigl(p_\ell(1-q_\ell)\bigr)^{3/2}}{\bigl(q_\ell(1-p_\ell)\bigr)^{1/2}} + \frac{\bigl(q_\ell(1-p_\ell)\bigr)^{3/2}}{\bigl(p_\ell(1-q_\ell)\bigr)^{1/2}}.
\end{align*}
Now,
\begingroup
\allowdisplaybreaks
\begin{align*}
& \log\bigg(\frac{\bigl(p_\ell(1-q_\ell)\bigr)^{3/2}}{\bigl(q_\ell(1-p_\ell)\bigr)^{1/2}} + \frac{\bigl(q_\ell(1-p_\ell)\bigr)^{3/2}}{\bigl(p_\ell(1-q_\ell)\bigr)^{1/2}} + 1 - p_\ell(1-q_\ell) - q_\ell(1-p_\ell)\bigg) \\
& \sim - \bigl(\sqrt{p_\ell} -\sqrt{q_\ell}\bigr)^2 - 2 \sqrt{p_\ell q_\ell} + 2p_\ell q_\ell + \frac{\bigl(p_\ell(1-q_\ell)\bigr)^{3/2}}{\bigl(q_\ell(1-p_\ell)\bigr)^{1/2}} + \frac{\bigl(q_\ell(1-p_\ell)\bigr)^{3/2}}{\bigl(p_\ell(1-q_\ell)\bigr)^{1/2}}\\
& = - \bigl(1+o(1)\bigr)\III{\ell}_{1/2} - 2\sqrt{p_\ell q_\ell} + 2\sqrt{p_\ell q_\ell} \cdot \bigg[ \sqrt{p_\ell q_\ell} + \frac{1+o(1)}{2} \cdot \bigg(\frac{p_\ell}{q_\ell}+\frac{q_\ell }{p_\ell}\bigg)\bigg] \\
& \lesssim -2\sqrt{p_\ell q_\ell} + \sqrt{p_\ell q_\ell} \cdot \bigg(\frac{p_\ell}{q_\ell}+\frac{q_\ell }{p_\ell}\bigg)\\
& = -2\sqrt{p_\ell q_\ell} + \sqrt{p_\ell q_\ell} \cdot \bigg( 2+ \frac{p_\ell - q_\ell}{q_\ell} - \frac{p_\ell - q_\ell}{p_\ell}\bigg) \\
& = \frac{(p_\ell - q_\ell)^2}{\sqrt{p_\ell q_\ell}}\\
& \lesssim\III{\ell}_{1/2},
\end{align*}
\endgroup
where we have used the asymptotic equivalence of \eqref{eq:asymp_equiv_of_I_1/2}.
Hence, for some constant $c, c'>0$ we have
\begingroup
\allowdisplaybreaks
\begin{align*}
& {\mathbb{P}}\bigg(\sum_{\ell \in S} \sum_{i = 1}^{m} \XXX{\ell}_i \cdot \log \bigg(\frac{q_\ell(1-p_\ell)}{p_\ell(1-q_\ell)}\bigg) + \YYY{\ell}_i \cdot \log \bigg(\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg) \geq \log\bigg(\frac{\rho}{1-\rho}\bigg) \bigg)\\
& = 1 - e^{-\bigl(1+o(1)\bigr)J_\rho} \cdot \prod_{\ell\in S} \prod_{i = 1}^{m} e^{c\III{\ell}_{1/2}}\\
& = 1 - \exp\bigg\{- \bigl( 1 + o(1)\bigr)\cdot \bigg(J_\rho - c\sum_{\ell\in S}m \III{\ell}_{1/2}\bigg)\bigg\} \\
& \geq 1 - \exp\bigg\{- \bigl( 1 + o(1)\bigr)\cdot \bigg(J_\rho - c' \sqrt{J_\rho}\bigg)\bigg\} \\
& = 1 - e^{-\bigl(1+o(1)\bigr)J_\rho}\\
& \geq 1/2,
\end{align*}
\endgroup
where the third line is by our assumption that $\sum_{\ell\in S}m \III{\ell}_{1/2} \lesssim \sqrt{J_\rho}$ and the last line is by $J_\rho \to\infty$.
This means that \eqref{eq:target_of_lower_bound_z_star} can be lower bounded by
\begin{align*}
\frac{1}{2}\cdot \exp\bigg\{- \bigl(1+o(1)\bigr) \bigg(|S^c| + 1\bigg) J_\rho \bigg\}
& = \exp\bigg\{- \bigl(1+o(1)\bigr) \bigg(|S^c| + 1\bigg) J_\rho \bigg\}\\
& \geq \exp\bigg\{- \bigl(1+o(1)\bigr) \bigg(\bigl(|S^c| + 1\bigr) J_\rho + \psi_S^\star(-2J_\rho)\bigg) \bigg\},
\end{align*}
where the last inequality is by $\psi_S^\star(-2J_\rho) \geq 0$, as shown in \eqref{eq:ineq_for_cvx_conjugate}.
~\\
{\bf Case B.}
In this case, we assume
$$
\sqrt{J_\rho} < \sum_{\ell\in S} m (p_\ell - q_\ell) \cdot \log\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)} \leq \log \frac{1-\rho}{\rho} = \bigl(1+o(1)\bigr) 2J_\rho.
$$
By Lemma \ref{lemma: exp_and_var_of_tilted_bernoulli}, we have
$$
\sqrt{J_\rho}\lesssim \sum_{\ell\in S} m\III{\ell}_{1/2} \to \infty,
$$
and
$$
\sum_{\ell\in S} m\III{\ell}_{1/2} \asymp \mathrm{Var} \bigg(\sum_{\ell \in S}\sum_{i = 1}^{m} \tXXX{\ell}_0 + \tYYY{\ell}_0\bigg) \to \infty.
$$
We now have
\begin{align*}
& {\mathbb{P}}\bigg(\sum_{\ell \in S} \sum_{i = 1}^{m} \XXX{\ell}_i \cdot \log \bigg(\frac{q_\ell(1-p_\ell)}{p_\ell(1-q_\ell)}\bigg) + \YYY{\ell}_i \cdot \log \bigg(\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg) \geq \log\bigg(\frac{\rho}{1-\rho}\bigg) \bigg)\\
& = {\mathbb{P}}\bigg(\sum_{\ell\in S}\sum_{i=1}^{m} \tXXX{\ell}_0 + \tYYY{\ell}_0 - {\mathbb{E}}[\tXXX{\ell} + \tYYY{\ell}] \geq \log \bigg(\frac{\rho}{1-\rho}\bigg) + \sum_{\ell\in S} m (p_\ell - q_\ell) \cdot \log\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg) \\
& \geq {\mathbb{P}}\bigg(\sum_{\ell\in S}\sum_{i=1}^{m} \tXXX{\ell}_0 + \tYYY{\ell}_0 - {\mathbb{E}}[\tXXX{\ell} + \tYYY{\ell}] \geq 0\bigg).
\end{align*}
Note that
\begin{align*}
\bigg(\tXXX{\ell}_0 + \tYYY{\ell}_0 - {\mathbb{E}}[\tXXX{\ell} + \tYYY{\ell}] \bigg)^2 \leq 4 \bigg(\log \frac{p_\ell(1-q_\ell)}{q_\ell (1-p_\ell)}\bigg)^2 = O(1).
\end{align*}
Hence, by Lindeberg-Feller CLT, we have
$$
\frac{\sum_{\ell\in S}\sum_{i=1}^{m} \tXXX{\ell}_0 + \tYYY{\ell}_0 - {\mathbb{E}}[\tXXX{\ell} + \tYYY{\ell}]}{\mathrm{Var} \bigg(\sum_{\ell\in S}\sum_{i=1}^{m} \tXXX{\ell}_0 + \tYYY{\ell}_0 - {\mathbb{E}}[\tXXX{\ell} + \tYYY{\ell}]\bigg)} \Rightarrow {\mathcal{N}}(0, 1).
$$
This gives
\begin{align*}
& {\mathbb{P}}\bigg(\sum_{\ell \in S} \sum_{i = 1}^{m} \XXX{\ell}_i \cdot \log \bigg(\frac{q_\ell(1-p_\ell)}{p_\ell(1-q_\ell)}\bigg) + \YYY{\ell}_i \cdot \log \bigg(\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg) \geq \log\bigg(\frac{\rho}{1-\rho}\bigg) \bigg) \geq C
\end{align*}
for large enough $n$ where $C > 0$ is some absolute constant. Hence, in this case, we have the following lower bound for \eqref{eq:target_of_lower_bound_z_star}:
\begin{align*}
C\cdot \exp\bigg\{- \bigl(1+o(1)\bigr) \bigg(|S^c| + 1\bigg) J_\rho \bigg\}
& \geq C\cdot\exp\bigg\{- \bigl(1+o(1)\bigr) \bigg(\bigl(|S^c| + 1\bigr) J_\rho + \psi_S^\star(-2J_\rho)\bigg) \bigg\}.
\end{align*}
~\\
{\bf Case C.}
In this case, we assume
$$
\sum_{\ell\in S} m (p_\ell - q_\ell) \cdot \log\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)} > \log \frac{1-\rho}{\rho} = \bigl(1+ o(1)\bigr)2J_\rho.
$$
We let
$$
\tilde \psi_S(t) = -m\sum_{\ell\in S} \III{\ell}_t = \frac{\psi_S(t)}{n/2} \cdot m = (1+o(1)) \psi_S(t).
$$
That is, we replace $n/2$ in the definition of $\psi_S(t)$ with $m = (1+o(1))n/2$. Meanwhile, we let $t^\star$ be the maximizer of
$$
\sup_{t\in[0, 1]} -t \log\frac{1-\rho}{\rho} - \tilde\psi_S(t).
$$
By construction, we have
$$
\sum_{\ell\in S} \sum_{i=1}^{m} {\mathbb{E}}[\tXXX{\ell}_{t} + \tYYY{\ell}_{t}] = \frac{d}{dt} \tilde\psi_S(t).
$$
and $\tilde\psi_S(t)$ is a convex function in $t$ (see Lemma \ref{eq:convexity_of_cum_gen_func} for details).
By convexity, the optimal $t^\star$ is such that
\begin{align*}
-\log \frac{1-\rho}{\rho} & = \sum_{\ell\in S} \sum_{i=1}^{m} {\mathbb{E}}[\tXXX{\ell}_{t^\star} + \tYYY{\ell}_{t^\star}],
\end{align*}
and the right-hand side is an increasing function in $t$.
This gives
$$
- \sum_{\ell\in S} m (p_\ell - q_\ell) \cdot \log\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)} \leq \sum_{\ell\in S} \sum_{i=1}^{m} {\mathbb{E}}[\tXXX{\ell}_{t} + \tYYY{\ell}_{t}] \leq 0
$$
for $t \in [0, 1/2]$, where the left-hand side above is attained at $t = 0$ and right-hand side is attained at $t = 1/2$. Since $\sum_{\ell\in S}\sum_{i\in[m]}{\mathbb{E}}[\tXXX{\ell}_t + \tYYY{\ell}_t]$ is continuous and increasing in $t$ for $t\in[0, 1/2]$, it attains every value between the two sides in the above display. On the other hand, the optimal $t^\star$ is such that the expectation is exactly equal to $-\log\frac{1-\rho}{\rho}$, which is between the two sides in the above display under our current assumption. As a result, we have
$$
t^\star \in [0, 1/2].
$$
Now note that
$$
\sup_{t\in[0, 1]} -t \log\frac{1-\rho}{\rho} - \tilde\psi_S(t) = \sup_{t\in[0, 1]} (1+o(1))\bigg(-2tJ_\rho - \psi_S(t)\bigg) = (1+o(1))\psi_S^\star(-2J_\rho).
$$
Hence, with a standard tilting argument, we get
\begin{align*}
\mu\bigg(\bigg[ \log\frac{\rho}{1-\rho}, \infty\bigg)\bigg)& \geq \exp\bigg\{- (1+o(1))\psi_S^\star(-2J_\rho) - t^\star\xi\bigg\} \\
& \qquad \times {\mathbb{P}}_{W\sim \tilde\mu_{t^\star}}\bigg( \log\frac{\rho}{1-\rho}\leq W\leq \xi + \log\frac{\rho}{1-\rho}\bigg).
\end{align*}
By Lemma \ref{lemma: exp_and_var_of_tilted_bernoulli}, $\mathrm{Var}(W)\asymp \sum_{\ell\in S}m\III{\ell}_{1/2}\to \infty$, and with a similar argument as the previous case, an application of Lindeberg-Feller CLT gives
$$
\frac{W - \log\frac{\rho}{1-\rho}}{\sqrt{\mathrm{Var}(W)}} \Rightarrow {\mathcal{N}}(0 , 1).
$$
Hence, choosing $\xi = \sqrt{\mathrm{Var}(W)} \asymp \sqrt{\sum_{\ell\in S}m\III{\ell}_{1/2}}$, we get
\begin{align*}
\mu\bigg(\bigg[ \log\frac{\rho}{1-\rho}, \infty\bigg)\bigg)& \geq C \cdot \exp\bigg\{- (1+o(1))\psi_S^\star(-2J_\rho) - ct^\star\sqrt{\sum_{\ell\in S}m\III{\ell}_{1/2}}\bigg\},
\end{align*}
and we arrive at the following lower bound for \eqref{eq:target_of_lower_bound_z_star}:
\begin{align*}
C\cdot \exp\bigg\{- \bigl(1+o(1)\bigr) \bigg(\bigl(|S^c| + 1\bigr) J_\rho + \psi_S^\star(-2J_\rho) + ct^\star\sqrt{\sum_{\ell\in S}m\III{\ell}_{1/2}}\bigg) \bigg\}.
\end{align*}
By the inequality in \eqref{eq:ineq_for_cvx_conjugate}, we have
$$
\bigl(|S^c| + 1\bigr) J_\rho + \psi_S^\star(-2J_\rho) \geq |S^c| J_\rho + \sum_{\ell \in S} m\III{\ell}_{1/2} \gg \sqrt{\sum_{\ell \in S}m\III{\ell}_{1/2}}.
$$
This gives
$$
\bigl(|S^c| + 1\bigr) J_\rho + \psi_S^\star(-2J_\rho) + ct^\star\sqrt{\sum_{\ell\in S}mI_\ell} = \bigl(1 + o(1)\bigr) \cdot \bigg(\bigl(|S^c| + 1\bigr) J_\rho + \psi_S^\star(-2J_\rho)\bigg).
$$
As a result, we get the following lower bound for \eqref{eq:target_of_lower_bound_z_star}:
\begin{align*}
C\cdot \exp\bigg\{- \bigl(1+o(1)\bigr) \bigg(\bigl(|S^c| + 1\bigr) J_\rho + \psi_S^\star(-2J_\rho)\bigg) \bigg\}.
\end{align*}
~\\
{\bf Summary.}
Combining the above three cases and noting that the $o(1)$ term does not depend on $S$ proves the second part of \eqref{eq:opt_test_error_z_star}. Thus the proof of Lemma \ref{lemma:opt_test_error_z_star} is concluded.
\subsection{Proof of Lemma \ref{lemma:testing_problem_z_l_part_I}}\label{subappend:prf_testing_problem_z_l_part_I}
Fix $\mathbf{\tilde z}^\star \in {\mathcal{P}}_n$ with $n^\star_+(\mathbf{\tilde z}^\star) = \lfloor n/2 \rfloor$ and $n_-^\star(\mathbf{\tilde z}^\star) = n - \lfloor n/2 \rfloor$. Consider the following parameter space, which consists of a single clustering vector:
\begin{equation}
\label{eq:sub_param_space_z_star_known}
{\mathcal{P}}_n^{1}:= \bigg\{ \textnormal{IMLSBM}(\mathbf{z}^\star, \rho, \{p_\ell\}_1^L, \{q_\ell\}_1^L): \mathbf{z}^\star = \mathbf{\tilde z}^\star, p_\ell > q_\ell \ \forall \ell\in[L] \bigg\}.
\end{equation}
We then have
\begin{align*}
\inf_{\hzzz{\ell}} \sup_{\mathbf{z}^\star \in {\mathcal{P}}_n} {\mathbb{E}}{\mathcal{L}}(\hzzz{\ell}, \zzz{\ell}) & \geq \inf_{\hzzz{\ell}} \sup_{\mathbf{z}^\star \in {\mathcal{P}}_n^1} {\mathbb{E}}{\mathcal{L}}(\hzzz{\ell}, \zzz{\ell}) \geq \inf_{\hzzz{\ell}} \sup_{\mathbf{z}^\star \in {\mathcal{P}}_n^1} {\mathbb{E}}[{\mathcal{L}}(\hzzz{\ell}, \zzz{\ell})\cdot \mathbf{1}{E}],
\end{align*}
where the event $E$ is defined as
$$
E = \bigg\{\frac{1}{n}\sum_{i\in[n]} \mathbf{1}{\mathbf{\tilde z}^\star_i \neq \zzz{\ell}_i} \leq \frac{1}{2} ~\forall \ell \in[L]\bigg\}.
$$
Since $\mathbf{\tilde z}^\star$ is known to us, it is a legitimate estimator of $\zzz{\ell}$. Hence, for the optimal estimator $\hzzz{\ell}$ of $\zzz{\ell}$, on the event $E$, we necessarily have
$$
\frac{1}{n}\sum_{i\in[n]} \mathbf{1}{\hzzz{\ell}_i \neq \zzz{\ell}_i} \leq \frac{1}{n}\sum_{i\in[n]} \mathbf{1}{\mathbf{\tilde z}^\star_i \neq \zzz{\ell}_i} \leq \frac{1}{2}.
$$
This gives
\begin{align*}
\inf_{\hzzz{\ell}} \sup_{\mathbf{z}^\star \in {\mathcal{P}}_n} {\mathbb{E}}{\mathcal{L}}(\hzzz{\ell}, \zzz{\ell}) & \geq \inf_{\hzzz{\ell}} {\mathbb{E}}[ \frac{1}{n}\sum_{i\in[n]} \mathbf{1}{\hzzz{\ell}_i \neq \zzz{\ell}_i}\cdot \mathbf{1}{E}] \geq \frac{1}{n}\sum_{i\in[n]} \inf_{\hzzz{\ell}_i} {\mathbb{P}}(\hzzz{\ell}_i \neq \zzz{\ell}_i , E).
\end{align*}
Let us define the event
$$
E_i := \bigg\{\frac{1}{n}\sum_{j\neq i} \mathbf{1}{\mathbf{\tilde z}^\star_j \neq \zzz{\ell}_j}\leq \frac{1}{3}~\forall \ell \in [L]\bigg\} \subseteq E,
$$
where the inclusion is by
$$
\frac{1}{n} \sum_{j\in[n]}\mathbf{1}{\mathbf{\tilde z}^\star_j \neq \zzz{\ell}_j} \leq \frac{1}{n} \sum_{j\neq i}\mathbf{1}{\mathbf{\tilde z}^\star_j \neq \zzz{\ell}_j} + \frac{1}{n} \leq \frac{1}{3} + \frac{1}{n} \leq \frac{1}{2}
$$
for large $n$. In addition, let us define
$$
F_i := \bigg\{ \bigg|\#\{j\neq i: \zzz{\ell}_j = \mathbf{z}^\star_i\} - \frac{n}{2}\bigg| \lor\bigg|\#\{j\neq i: \zzz{\ell}_j = -\mathbf{z}^\star_i\} - \frac{n}{2}\bigg| \leq n\cdot n^{-1/4} ~\forall \ell\in[L]\bigg\}.
$$
By the inclusion, we have
\begin{align*}
\inf_{\hzzz{\ell}} \sup_{\mathbf{z}^\star \in {\mathcal{P}}_n} {\mathbb{E}}{\mathcal{L}}(\hzzz{\ell}, \zzz{\ell}) & \geq \frac{1}{n}\sum_{i\in[n]} \inf_{\hzzz{\ell}_i} {\mathbb{P}}(\hzzz{\ell}_i \neq \zzz{\ell}_i, E_i\cap F_i).
\end{align*}
Fix any $i\in[n]$. Without loss of generality we assume $\mathbf{\tilde z}^\star_i = +1$. Since the event $E_i\cap F_i$ only depends on $\zzz{\ell}_{-i} := \{\zzz{\ell}_j: j\neq i\}$, we have
\begin{align*}
&\inf_{\hzzz{\ell}_i} {\mathbb{P}}(\hzzz{\ell}_i \neq \zzz{\ell}_i, E_i\cap F_i) \\
& \geq \inf_{\hzzz{\ell}_i} \sum_{\xi\in\{\pm 1\}^{n-1}} {\mathbb{P}}(\zzz{\ell}_j = \xi_j \mathbf{z}^\star_j ~\forall j\neq i) \cdot \mathbf{1}\bigg\{\frac{1}{n}\sum_{j\neq i} \mathbf{1}{\xi_j = -1} \leq \frac{1}{3}, |\mmm{\ell}_+ - n/2| \lor |\mmm{\ell}_- - n/2| \leq n\cdot n^{-1/4}\bigg\} \\
& \qquad \times \bigg((1-\rho)\cdot {\mathbb{P}}(\hzzz{\ell}_i =-1 ~|~ \zzz{\ell}_i = +1, \zzz{\ell}_j = \xi_j \mathbf{z}^\star_j ~\forall j\neq i)
+ \rho \cdot {\mathbb{P}}(\hzzz{\ell}_i = +1 ~|~ \zzz{\ell}_i = -1, \zzz{\ell}_j = \xi_j \mathbf{z}^\star_j ~\forall j\neq i) \bigg) \\
& \geq \sum_{\xi\in\{\pm 1\}^{n-1}} {\mathbb{P}}(\zzz{\ell}_j = \xi_j \mathbf{z}^\star_j ~\forall j\neq i) \cdot \mathbf{1}\bigg\{\frac{1}{n}\sum_{j\neq i} \mathbf{1}{\xi_j = -1} \leq \frac{1}{3}, |\mmm{\ell}_+ - n/2| \lor |\mmm{\ell}_- - n/2| \leq n\cdot n^{-1/4}\bigg\} \\
& \qquad \times \inf_{\hzzz{\ell}_i} \bigg((1-\rho)\cdot {\mathbb{P}}(\hzzz{\ell}_i =-1 ~|~ \zzz{\ell}_i = +1, \zzz{\ell}_j = \xi_j \mathbf{z}^\star_j ~\forall j\neq i) + \rho \cdot {\mathbb{P}}(\hzzz{\ell}_i = +1 ~|~ \zzz{\ell}_i = -1, \zzz{\ell}_j = \xi_j \mathbf{z}^\star_j ~\forall j\neq i) \bigg),
\end{align*}
where
$$
\mmm{\ell}_+ = \#\{j\neq i: \xi_j\mathbf{z}^\star_j = + 1\}, \ \ \ \mmm{\ell}_- = \#\{j\neq i: \xi_j\mathbf{z}^\star_j = -1\}.
$$
Since we know $\mathbf{z}^\star = \mathbf{\tilde z}^\star$, by independence, we can without loss of generality restrict ourself to $\hzzz{\ell}_i$'s that are only functions of $\AAA{\ell}$ alone. Thus, the ``$\inf$'' term in the right-hand side above can be regarded as the $(1-\rho)\times \textnormal{type-I error} + \rho \times \textnormal{type-II error}$ of the following binary hypothesis testing problem:
$$
H_0: \zzz{\ell}_i = +1 \ \ \ \textnormal{v.s.} \ \ \ H_1: \zzz{\ell}_i = -1,
$$
where the data is a single adjacency matrix $\AAA{\ell}$ sampled from a vanilla two-block SBM.
We now focus on lower bounding this ``$\inf$'' term. By Lemma \ref{lemma: optimal_test}, the optimal test is given by the likelihood ratio test with cutoff being $\rho/(1-\rho)$. Let $L_0$ and $L_1$ be the likelihood function under $H_0$ and $H_1$, respectively. With some algebra, we have
\begin{align*}
\frac{L_0}{L_1} & = \prod_{\substack{j\neq i\\ \xi_j\mathbf{z}^\star_j = 1}} p_\ell^{\AAA{\ell}_{ij}} (1-p_\ell)^{1-\AAA{\ell}_{ij}}\prod_{\substack{j\neq i\\ \xi_j\mathbf{z}^\star_j = -1}} q_\ell^{\AAA{\ell}_{ij}} (1-q_\ell)^{1-\AAA{\ell}_{ij}} \\
& \qquad \times \bigg(\prod_{\substack{j\neq i\\ \xi_j\mathbf{z}^\star_j = -1}} p_\ell^{\AAA{\ell}_{ij}} (1-p_\ell)^{1-\AAA{\ell}_{ij}}\prod_{\substack{j\neq i\\ \xi_j\mathbf{z}^\star_j= 1}} q_\ell^{\AAA{\ell}_{ij}} (1-q_\ell)^{1-\AAA{\ell}_{ij}}\bigg)^{-1}.
\end{align*}
Thus, the type-I error of the optimal test is given by
\begin{align*}
& {\mathbb{P}}\bigg(\sum_{i=1}^{\mmm{\ell}_+} \XXX{\ell}_i \cdot \log \frac{q_\ell(1-p_\ell)}{p_\ell(1-q_\ell)} + \sum_{i = 1}^{\mmm{\ell}_-} \YYY{\ell}_i \cdot \log\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\geq \log \frac{1-\rho}{\rho}\bigg),
\end{align*}
and the type-II error of the optimal test is given by
\begin{align*}
& {\mathbb{P}}\bigg(\sum_{i=1}^{\mmm{\ell}_+} \XXX{\ell}_i \cdot \log \frac{q_\ell(1-p_\ell)}{p_\ell(1-q_\ell)} + \sum_{i = 1}^{\mmm{\ell}_-} \YYY{\ell}_i \cdot \log\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\geq -\log \frac{1-\rho}{\rho}\bigg),
\end{align*}
where $\XXX{\ell}_i, \YYY{\ell}_i$'s are defined as
\begin{align*}
& \XXX{\ell}_i \overset{\textnormal{i.i.d.}}{\sim} \textnormal{Bern}(p_\ell), \ \ \ \YYY{\ell}_i \overset{\textnormal{i.i.d.}}{\sim} \textnormal{Bern}(q_\ell).
\end{align*}
So overall, the weighted testing error is given by
\begin{align*}
& {\mathbb{P}}\bigg(\sum_{i=1}^{\mmm{\ell}_+} \XXX{\ell}_i \cdot \log \frac{q_\ell(1-p_\ell)}{p_\ell(1-q_\ell)} + \sum_{i = 1}^{\mmm{\ell}_-} \YYY{\ell}_i \cdot \log\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\geq \ZZZ{\ell}\log \frac{1-\rho}{\rho}\bigg),
\end{align*}
where $\ZZZ{\ell}\sim 2\textnormal{Bern}(1-\rho)-1$, which is independent of $\XXX{\ell}_i$ and $\YYY{\ell}_i$'s.
By Lemma \ref{lemma: general_data_processing}, the above probability can be lower bounded by
\begin{align*}
& {\mathbb{P}}\bigg(\sum_{i=1}^{m} \XXX{\ell}_i \cdot \log \frac{q_\ell(1-p_\ell)}{p_\ell(1-q_\ell)} + \sum_{i = 1}^{m} \YYY{\ell}_i \cdot \log\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\geq \ZZZ{\ell}\log \frac{1-\rho}{\rho}\bigg),
\end{align*}
for $m = \mmm{\ell}_+ \lor \mmm{\ell}_-$, which is $(1+n^{-1/4})n/2$ under $F_i$.
In summary, we have
\begin{align*}
& \inf_{\hzzz{\ell}} \sup_{\mathbf{z}^\star \in {\mathcal{P}}_n} {\mathbb{E}}{\mathcal{L}}(\hzzz{\ell}, \zzz{\ell}) \\
& \geq \sum_{\xi\in\{\pm 1\}^{n-1}} {\mathbb{P}}(\zzz{\ell}_j = \xi_j \mathbf{z}^\star_j ~\forall j\neq i) \cdot \mathbf{1}\bigg\{\frac{1}{n}\sum_{j\neq i} \mathbf{1}{\xi_j = -1} \leq \frac{1}{3}, |\mmm{\ell}_+ - n/2| \lor |\mmm{\ell}_- - n/2| \leq n\cdot n^{-1/4}\bigg\} \\
& \qquad \times {\mathbb{P}}\bigg(\sum_{i=1}^{m} \XXX{\ell}_i \cdot \log \frac{q_\ell(1-p_\ell)}{p_\ell(1-q_\ell)} + \sum_{i = 1}^{m} \YYY{\ell}_i \cdot \log\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\geq \ZZZ{\ell}\log \frac{1-\rho}{\rho}\bigg)\\
& = {\mathbb{P}}(E_i\cap F_i) \cdot {\mathbb{P}}\bigg(\sum_{i=1}^{m} \XXX{\ell}_i \cdot \log \frac{q_\ell(1-p_\ell)}{p_\ell(1-q_\ell)} + \sum_{i = 1}^{m} \YYY{\ell}_i \cdot \log\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\geq \ZZZ{\ell}\log \frac{1-\rho}{\rho}\bigg) \\
& \gtrsim {\mathbb{P}}\bigg(\sum_{i=1}^{m} \XXX{\ell}_i \cdot \log \frac{q_\ell(1-p_\ell)}{p_\ell(1-q_\ell)} + \sum_{i = 1}^{m} \YYY{\ell}_i \cdot \log\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\geq \ZZZ{\ell}\log \frac{1-\rho}{\rho}\bigg),
\end{align*}
where the last line is by an application of Chernoff bound and union bound (similar to the arguments that lead to \eqref{eq:concentration_of_number_of_pos_and_neg_nodes}). We finish the proof by noting that the right-hand side above is the $(1-\rho)\times \textnormal{type-I error} + \rho \times \textnormal{type-II error}$ of the testing problem \eqref{eq:fund_testing_problem_z_l_part_I} with $(1+\delta_n)n/2 = m$ where $\delta_n = n^{-1/4} = o(1)$.
\subsection{Proof of Lemma \ref{lemma:testing_problem_z_l_part_II}}\label{subappend:prf_testing_problem_z_l_part_II}
The proof has a similar flavor to the proof of Lemma \ref{lemma:testing_problem_z_star}. Recall the sub-parameter space ${\mathcal{P}}_n^0$ defined in \eqref{eq:sub_param_space_z_star}, which consists of clustering vectors that agree with $\mathbf{\tilde z}^\star$ on $T$. We have
\begin{align*}
\inf_{\hzzz{\ell}} \sup_{\mathbf{z}^\star \in {\mathcal{P}}_n} {\mathbb{E}}{\mathcal{L}}(\hzzz{\ell}, \zzz{\ell}) & \geq \inf_{\hzzz{\ell}} \sup_{\mathbf{z}^\star \in {\mathcal{P}}_n^0} {\mathbb{E}}{\mathcal{L}}(\hzzz{\ell}, \zzz{\ell})
\end{align*}
Since $\mathbf{\tilde z}^\star$ is known to us, it is a legitimate estimator of $\zzz{\ell}$. Thus, for the optimal estimator $\hzzz{\ell}$ of $\zzz{\ell}$, it necessarily satisfies
$$
\frac{1}{n} \sum_{i\in[n]} \mathbf{1}{\hzzz{\ell}_i \neq \zzz{\ell}_i} \leq \frac{1}{n} \sum_{i\in[n]} \mathbf{1}{\mathbf{\tilde z}^\star_i \neq \zzz{\ell}_i} \leq \frac{1}{n} \sum_{i\in T} \mathbf{1}{\mathbf{z}^\star_i \neq \zzz{\ell}_i} + \frac{|T^c|}{n} \leq \frac{1}{n} \sum_{i\in T} \mathbf{1}{\mathbf{z}^\star_i \neq \zzz{\ell}_i} + \frac{4}{n} + \delta_n.
$$
Define the event $E$ to be
$$
E:= \bigg\{\frac{1}{n}\sum_{i\in [n]} \mathbf{1}{\mathbf{z}^\star_i \neq \zzz{\ell}_i} \leq \frac{1}{3} ~ \forall \ell\in[L] \bigg\}.
$$
On the event $E$, for large $n$ we have
$$
\frac{1}{n} \sum_{i\in[n]} \mathbf{1}{\hzzz{\ell}_i \neq \zzz{\ell}_i} \leq \frac{1}{3} + \frac{4}{n} + \delta_n \leq \frac{1}{2},
$$
and hence
$$
{\mathcal{L}}(\hzzz{\ell}, \zzz{\ell}) = \frac{1}{n} \sum_{i\in[n]} {\mathbb{P}}(\hzzz{\ell}_i \neq \zzz{\ell}_i).
$$
Let us define
$$
E_i = \bigg\{ \frac{1}{n}\sum_{j\neq i} \mathbf{1}{\mathbf{z}^\star_i \neq \zzz{\ell}_i} \leq \frac{1}{4} ~\forall \ell\in[L]\bigg\} \subseteq E
$$
for large $n$. In addition, we let
$$
F_i := \bigg\{ \bigg|\#\{j\neq i: \zzz{\ell}_j = \mathbf{z}^\star_i\} - \frac{n}{2}\bigg| \lor\bigg|\#\{j\neq i: \zzz{\ell}_j = -\mathbf{z}^\star_i\} - \frac{n}{2}\bigg| \leq n\cdot n^{-1/4} ~\forall\ell\in[L] \bigg\}.
$$
We then proceed by
\begin{align*}
\inf_{\hzzz{\ell}} \sup_{\mathbf{z}^\star \in {\mathcal{P}}_n} {\mathbb{E}}{\mathcal{L}}(\hzzz{\ell}, \zzz{\ell}) & \geq \inf_{\hzzz{\ell}} \sup_{\mathbf{z}^\star \in {\mathcal{P}}_n^0} {\mathbb{E}}[{\mathcal{L}}(\hzzz{\ell}, \zzz{\ell})\cdot \mathbf{1}{E}]\\
& = \inf_{\hzzz{\ell}} \sup_{\mathbf{z}^\star \in {\mathcal{P}}_n^0} \frac{1}{n} \sum_{i\in[n]}{\mathbb{P}}(\hzzz{\ell}_i \neq \zzz{\ell}_i, E)\\
& = \inf_{\hzzz{\ell}} \sup_{\mathbf{z}^\star \in {\mathcal{P}}_n^0} \frac{1}{n} \sum_{i\in[n]}{\mathbb{P}}(\hzzz{\ell}_i \neq \zzz{\ell}_i, E_i\cap F_i)\\
& \geq \frac{|T^c|}{n} \inf_{\hzzz{\ell}} \sup_{\mathbf{z}^\star \in {\mathcal{Z}}_T} \frac{1}{|T^c|} \sum_{i\in T^c} {\mathbb{P}}(\hzzz{\ell}_i \neq \zzz{\ell}_i, E_i\cap F_i)\\
& \geq \frac{1}{n} \sum_{i\in T^c} \inf_{\hzzz{\ell}_i} \underset{\mathbf{z}^\star\in {\mathcal{Z}}_T}{\textnormal{ave}}~ {\mathbb{P}}(\hzzz{\ell}_i \neq \zzz{\ell}_i, E_i\cap F_i).
\end{align*}
We are to show that all the summands in the right-hand side above are lower bounded by the same quantity, which is the type-I plus type-II error of the testing problem \eqref{eq:fund_testing_problem_z_l_part_II}, so that for any $i\in T^c$ we would have
$$
\inf_{\hzzz{\ell}} \sup_{\mathbf{z}^\star \in {\mathcal{P}}_n} {\mathbb{E}}{\mathcal{L}}(\hzzz{\ell}, \zzz{\ell}) \gtrsim \delta_n \inf_{\hzzz{\ell}_i} \underset{\mathbf{z}^\star\in {\mathcal{Z}}_T}{\textnormal{ave}}~ {\mathbb{P}}(\hzzz{\ell}_i \neq \zzz{\ell}_i, E_i\cap F_i),
$$
which is the desired result.
In the following discussion, we without loss of generality assume $\ell = 1$.
Now for any $i\in T^c$ and $S \subseteq\{2, \hdots, L\}$, we have
\begin{align*}
& \inf_{\hzzz{1}_i} \underset{\mathbf{z}^\star\in {\mathcal{Z}}_T}{\textnormal{ave}}~ {\mathbb{P}}(\hzzz{1}_i \neq \zzz{1}_i, E_i\cap F_i) \\
& = \inf_{\hzzz{1}_i} \underset{\mathbf{z}^\star_{-i}}{\textnormal{ave}}~\underset{\mathbf{z}^\star_i}{\textnormal{ave}}~ \sum_{\xi\in \{\pm 1\}^S} {\mathbb{P}}(\hzzz{1}_i \neq \zzz{1}_i , E_i\cap F_i~|~ \zzz{\ell}_i = \xi_\ell \mathbf{z}^\star_i ~\forall \ell\in S)\cdot {\mathbb{P}}(\zzz{\ell}_i = \xi_\ell \mathbf{z}^\star_i ~\forall \ell\in S)\\
& \geq \underset{\mathbf{z}^\star_{-i}}{\textnormal{ave}}~ \sum_{\xi\in \{\pm 1\}^S}{\mathbb{P}}(\zzz{\ell}_i = \xi_\ell \mathbf{z}^\star_i ~\forall \ell\in S) \cdot \inf_{\hzzz{1}_i}\underset{\mathbf{z}^\star_i}{\textnormal{ave}}~{\mathbb{P}}(\hzzz{1}_i \neq \zzz{1}_i, E_i\cap F_i ~|~ \zzz{\ell}_i = \xi_\ell \mathbf{z}^\star_i ~\forall \ell\in S),
\end{align*}
where in the last inequality we can pull ${\mathbb{P}}(\zzz{\ell}_i = \xi_\ell \mathbf{z}^\star_i ~\forall \ell\in S)$ in front of $\textnormal{ave}_{\mathbf{z}^\star_i}$ because
$$
{\mathbb{P}}(\zzz{\ell}_i = \xi_\ell \mathbf{z}^\star_i ~\forall \ell\in S) = {\mathbb{P}}(\zzz{\ell}_i = \xi_\ell \mathbf{z}^\star_i ~\forall \ell\in S ~|~ \mathbf{z}^\star_i = +1) = {\mathbb{P}}(\zzz{\ell}_i = \xi_\ell \mathbf{z}^\star_i ~\forall \ell\in S ~|~ \mathbf{z}^\star_i = -1).
$$
Since $E_i\cap F_i$ only depends on $\{\zzz{\ell}_{-i}\}$, we can decompose the error probability according to whether $\zzz{1}_j$ is flipped or not:
\begingroup
\allowdisplaybreaks
\begin{align*}
& \inf_{\hzzz{1}_i} \underset{\mathbf{z}^\star_i}{\textnormal{ave}} \ {\mathbb{P}} ( \hzzz{1}_i \neq \zzz{1}_i, E_i\cap F_i \ | \ \zzz{\ell}_i = \xi_\ell \mathbf{z}^\star_i \ \forall \ell \in S ) \\
& = \frac{1}{2} \inf_{\hzzz{1}_i} {\mathbb{P}} ( \hzzz{1}_i \neq \zzz{1}_i , E_i\cap F_i\ | \ \mathbf{z}^\star_i = 1, \zzz{\ell}_i = \xi_\ell \mathbf{z}^\star_i \ \forall \ell \in S ) + {\mathbb{P}} ( \hzzz{1}_i \neq \zzz{1}_i, E_i\cap F_i \ | \ \mathbf{z}^\star_i = -1, \zzz{\ell}_i = \xi_\ell \mathbf{z}^\star_i \ \forall \ell \in S ) \\
& \geq \frac{1}{2} \inf_{\hzzz{1}_0} \sum_{\zeta\in\{\pm 1\}^{n-1}} {\mathbb{P}}(\zzz{1}_j = \zeta_j \mathbf{z}^\star_j ~\forall j\neq i) \cdot \mathbf{1}\bigg\{ \frac{1}{n}\sum_{j\neq i}\mathbf{1}{\zeta_j = -1} \leq \frac{1}{4}, |\mmm{1}_1- n/2|\lor |\mmm{1}_2 - n/2|\leq n\cdot n^{-1/4}\bigg\} \\
& \qquad \times\bigg( (1-\rho) \cdot {\mathbb{P}} ( \hzzz{1}_i \neq \zzz{1}_i \ | \ \mathbf{z}^\star_i = 1, \zzz{1}_i = \mathbf{z}^\star_i, \zzz{1}_j = \zeta_j \mathbf{z}^\star_j ~\forall j\neq i, \zzz{\ell}_i = \xi_\ell \mathbf{z}^\star_i \ \forall \ell \in S ) \\
& \qquad\qquad \qquad + \rho \cdot {\mathbb{P}} ( \hzzz{1}_i \neq \zzz{1}_i\ | \ \mathbf{z}^\star_i = 1, \zzz{1}_i = -\mathbf{z}^\star_i, \zzz{1}_j = \zeta_j \mathbf{z}^\star_j ~\forall j\neq i, \zzz{\ell}_i = \xi_\ell \mathbf{z}^\star_i \ \forall \ell \in S ) \\
& \qquad\qquad \qquad + (1-\rho) \cdot {\mathbb{P}} ( \hzzz{1}_i \neq \zzz{1}_i \ | \ \mathbf{z}^\star_i = -1, \zzz{1}_i = \mathbf{z}^\star_i, \zzz{1}_j = \zeta_j \mathbf{z}^\star_j ~\forall j\neq i, \zzz{\ell}_i = \xi_\ell \mathbf{z}^\star_i \ \forall \ell \in S ) \\
& \qquad\qquad \qquad + \rho \cdot {\mathbb{P}} ( \hzzz{1}_i \neq \zzz{1}_i \ | \ \mathbf{z}^\star_i = -1, \zzz{1}_i = -\mathbf{z}^\star_i, \zzz{1}_j = \zeta_j \mathbf{z}^\star_j ~\forall j\neq i, \zzz{\ell}_i = \xi_\ell \mathbf{z}^\star_i \ \forall \ell \in S ) \bigg),
\end{align*}
\endgroup
where
$$
\mmm{1}_1 = \#\{j\neq i: \zeta_j\mathbf{z}^\star_j = + 1\}, \ \ \ \mmm{1}_2 = \#\{j\neq i: \zeta_j\mathbf{z}^\star_j = -1\}.
$$
Here $\zeta_j$ indicates whether $\zzz{1}_j$ is flipped or not. We then have
\begin{align*}
& \inf_{\hzzz{1}_i} \underset{\mathbf{z}^\star_i}{\textnormal{ave}} \ {\mathbb{P}} ( \hzzz{1}_i \neq \zzz{1}_i, E_i\cap F_i \ | \ \zzz{\ell}_i = \xi_\ell \mathbf{z}^\star_i \ \forall \ell \in S ) \\
& \geq \sum_{\zeta\in\{\pm 1\}^{n-1}} {\mathbb{P}}(\zzz{1}_j = \zeta_j \mathbf{z}^\star_j ~\forall j\neq i) \cdot \mathbf{1}\bigg\{ \frac{1}{n}\sum_{j\neq i}\mathbf{1}{\zeta_j = -1} \leq \frac{1}{4}, |\mmm{1}_1 - n/2|\lor |\mmm{1}_2 - n/2|\leq n\cdot n^{-1/4}\bigg\}\\
& \qquad \times \frac{1-\rho}{2} \cdot \inf_{\hzzz{1}_i} \bigg\{{\mathbb{P}} ( \hzzz{1}_i = - 1 \ | \ \mathbf{z}^\star_i = 1, \zzz{1}_i = \mathbf{z}^\star_i, \zzz{1}_j = \zeta_j \mathbf{z}^\star_j ~\forall j\neq i, \zzz{\ell}_i = \xi_\ell \mathbf{z}^\star_i \ \forall \ell \in S ) \\
& \qquad \qquad\qquad\qquad+ {\mathbb{P}} ( \hzzz{1}_i = +1\ | \ \mathbf{z}^\star_i = -1, \zzz{1}_i = \mathbf{z}^\star_i, \zzz{1}_j = \zeta_j \mathbf{z}^\star_j ~\forall j\neq i, \zzz{\ell}_i = \xi_\ell \mathbf{z}^\star_i \ \forall \ell \in S ) \bigg\} \\
& \qquad + \frac{\rho}{2} \cdot \inf_{\hzzz{1}_i} \bigg\{{\mathbb{P}} ( \hzzz{1}_i = +1 \ | \ \mathbf{z}^\star_i = 1, \zzz{1}_i = -\mathbf{z}^\star_i, \zzz{1}_j = \zeta_j \mathbf{z}^\star_j ~\forall j\neq i, \zzz{\ell}_i = \xi_\ell \mathbf{z}^\star_i \ \forall \ell \in S ) \\
& \qquad \qquad \qquad \qquad + {\mathbb{P}} ( \hzzz{1}_i = -1 \ | \ \mathbf{z}^\star_i = -1, \zzz{1}_i = -\mathbf{z}^\star_i, \zzz{1}_j = \zeta_j \mathbf{z}^\star_j ~\forall j\neq i, \zzz{\ell}_i = \xi_\ell \mathbf{z}^\star_i \ \forall \ell \in S )\bigg\}.
\end{align*}
Let $\RN{1}, \RN{2}$ be the first and the second ``$\inf$'' term in the right-hand side above, respectively.
We first deal with term $\RN{1}$. Note that in $\RN{1}$, $\hzzz{1}_i$ can be regarded as the testing function for
$$
H_0: \mathbf{z}^\star_i = 1, \qquad H_1: \mathbf{z}^\star_i = -1
$$
under the conditional law of $\{\AAA{\ell}\} \ | \ \{\zzz{1}_i = \mathbf{z}^\star_i, \zzz{1}_j = \zeta_j \mathbf{z}^\star_j ~\forall j\neq i, \zzz{\ell}_i = \xi_\ell \mathbf{z}^\star_i \ \forall \ell \in S\}$. By the same arguments as those in the proof of Lemma \ref{lemma:testing_problem_z_star}, $\RN{1}$ is lower bounded by the testing error of $H_0$ v.s. $H_1$ under the joint law of $\{\AAA{\ell}, \zzz{\ell}\} \ | \ \{\zzz{1}_i = \mathbf{z}^\star_i,\zzz{1}_j = \zeta_j \mathbf{z}^\star_j ~\forall j\neq i, \zzz{\ell}_i = \xi_\ell \mathbf{z}^\star_i \ \forall \ell \in S\}$ with $m$ positive nodes and $m$ negative nodes (excluding node $i$), where $m = (1+o(1))n/2$.
For notational simplicity, we again consider the following equivalent setup: we have $2m+1$ nodes in total, where nodes $1, \hdots, m$ are labeled as $+1$, nodes $m+1, \hdots, 2m$ are labeled as $-1$, and the node labeled as $0$ (which is originally labeled as $i$) is the node whose community is to be decided.
Under the current notations, the density of $\{\AAA{\ell}, \zzz{\ell}\} \ | \ \{\zzz{1}_0 = \mathbf{z}^\star_0, \zzz{1}_j = \zeta_j \mathbf{z}^\star_j ~\forall j\neq 0, \zzz{\ell}_0 = \xi_\ell \mathbf{z}^\star_0 \ \forall \ell \in S\}$ is given by
\begingroup
\allowdisplaybreaks
\begin{align*}
& \rho^{nL} \bigg(\frac{1-\rho}{\rho}\bigg)^{\#\{\ell\in[L], 1\leq i\leq 2m: \zzz{\ell}_i = \mathbf{z}^\star_i\}} \cdot \bigg(\frac{1-\rho}{\rho}\bigg)^{1+ \#\{\ell\in S: \xi_{\ell} = 1\} + \#\{\ell\notin S\cup\{1\}: \zzz{\ell}_0 = \mathbf{z}^\star_0\}} \\
& \qquad \times \prod_{\ell\in[L]} \prod_{\substack{i\neq j \\ i\neq 0 \\ j\neq 0}} p_\ell^{\AAA{\ell}_{ij}} (1-p_\ell)^{1 - \AAA{\ell}_{ij}} \mathbf{1}{\zzz{\ell}_i = \zzz{\ell}_j} + q_\ell^{\AAA{\ell}_{ij}} (1-q_\ell)^{1 - \AAA{\ell}_{ij}} \mathbf{1}{\zzz{\ell}_i \neq \zzz{\ell}_j} \\
& \qquad \times \prod_{\substack{j\neq 0 \\ \zeta_j\mathbf{z}^\star_j = \mathbf{z}^\star_0 }} p_1^{\AAA{1}_{0j}} (1-p_1)^{1 - \AAA{1}_{0j}} \prod_{\substack{j\neq 0 \\ \zeta_j \mathbf{z}^\star_j \neq \mathbf{z}^\star_0 }} q_1^{\AAA{1}_{0j}} (1-q_1)^{1 - \AAA{1}_{0j}} \\
& \qquad \times \prod_{\ell\in S} \prod_{\substack{j\neq 0 \\ \zzz{\ell}_j =\xi_\ell\mathbf{z}^\star_0 }} p_\ell^{\AAA{\ell}_{0j}} (1-p_\ell)^{1 - \AAA{\ell}_{0j}} \prod_{\substack{j\neq 0 \\ \zzz{\ell}_j \neq \xi_\ell \mathbf{z}^\star_0 }} q_\ell^{\AAA{\ell}_{0j}} (1-q_\ell)^{1 - \AAA{\ell}_{0j}}\\
& \qquad \times \prod_{\ell\notin S\cup\{1\}} \prod_{\substack{j\neq 0 \\ \zzz{\ell}_j = \zzz{\ell}_0 }} p_\ell^{\AAA{\ell}_{0j}} (1-p_\ell)^{1 - \AAA{\ell}_{0j}} \prod_{\substack{j\neq 0 \\ \zzz{\ell}_j \neq \zzz{\ell}_0 }} q_\ell^{\AAA{\ell}_{0j}} (1-q_\ell)^{1 - \AAA{\ell}_{0j}}.
\end{align*}
\endgroup
The likelihood ratio is then given by
\begingroup
\allowdisplaybreaks
\begin{align*}
\frac{L_0}{L_1} & = \bigg(\frac{1-\rho}{\rho}\bigg)^{\#\{\ell\notin S\cup\{1\}: \zzz{\ell}_0 = 1\} - \#\{\ell\notin S\cup\{1\}: \zzz{\ell}_0 = -1\}} \\
& \qquad \times \prod_{\substack{j\neq 0 \\ \zeta_j\mathbf{z}^\star_j = 1 }} p_1^{\AAA{1}_{0j}} (1-p_1)^{1 - \AAA{1}_{0j}} \prod_{\substack{j\neq 0 \\ \zeta_j\mathbf{z}^\star_j =-1 }} q_1^{\AAA{1}_{0j}} (1-q_1)^{1 - \AAA{1}_{0j}} \\
& \qquad \times \bigg( \prod_{\substack{j\neq 0 \\ \zeta_j\mathbf{z}^\star_j= -1 }} p_1^{\AAA{1}_{0j}} (1-p_1)^{1 - \AAA{1}_{0j}} \prod_{\substack{j\neq 0 \\ \zeta_j\mathbf{z}^\star_j =1 }} q_1^{\AAA{1}_{0j}} (1-q_1)^{1 - \AAA{1}_{0j}} \bigg)^{-1}\\
& \qquad \times \prod_{\ell\in S} \prod_{\substack{j\neq 0 \\ \zzz{\ell}_j =\xi_\ell }} p_\ell^{\AAA{\ell}_{0j}} (1-p_\ell)^{1 - \AAA{\ell}_{0j}} \prod_{\substack{j\neq 0 \\ \zzz{\ell}_j = - \xi_\ell }} q_\ell^{\AAA{\ell}_{0j}} (1-q_\ell)^{1 - \AAA{\ell}_{0j}} \\
& \qquad \times \bigg( \prod_{\ell\in S} \prod_{\substack{j\neq 0 \\ \zzz{\ell}_j = - \xi_\ell }} p_\ell^{\AAA{\ell}_{0j}} (1-p_\ell)^{1 - \AAA{\ell}_{0j}} \prod_{\substack{j\neq 0 \\ \zzz{\ell}_j = \xi_\ell }} q_\ell^{\AAA{\ell}_{0j}} (1-q_\ell)^{1 - \AAA{\ell}_{0j}} \bigg)^{-1}.
\end{align*}
\endgroup
By Neyman-Pearson lemma, the term $\RN{1}$ is given by
\begin{align*}
& {\mathbb{P}}\bigg(\frac{L_0}{L_1} \leq 1 \ \bigg| \ \mathbf{z}^\star_0 = 1, \zzz{1}_0 = \mathbf{z}^\star_0, \zzz{1}_j = \zeta_j \mathbf{z}^\star_j ~\forall j\neq 0, \zzz{\ell}_0 = \xi_\ell \mathbf{z}^\star_0 \ \forall \ell \in S \bigg) \\
& \qquad + {\mathbb{P}}\bigg(\frac{L_0}{L_1} \geq 1 \ \bigg| \ \mathbf{z}^\star_0 = -1, \zzz{1}_0 = \mathbf{z}^\star_0, \zzz{1}_j = \zeta_j \mathbf{z}^\star_j ~\forall j\neq 0, \zzz{\ell}_0 = \xi_\ell \mathbf{z}^\star_0 \ \forall \ell \in S \bigg).
\end{align*}
By symmetry, the two terms above are equal to each other, and we calculate the first term. With a slight abuse of notation, let ${\mathbb{P}}_{H_0, S,\xi, \zeta}$ be the conditional law of $\{\AAA{\ell}, \zzz{\ell}\} \ | \ \{\mathbf{z}^\star_0 = 1, \zzz{1}_0 = \mathbf{z}^\star_0,\zzz{1}_j = \zeta_j \mathbf{z}^\star_j ~\forall j\neq 0, \zzz{\ell}_0 = \xi_\ell \mathbf{z}^\star_0 \ \forall \ell \in S\}$. Then the quantity of interest is
\begingroup
\allowdisplaybreaks
\begin{align*}
&{\mathbb{P}}_{H_0, S, \xi, \zeta} \bigg(\frac{L_0}{L_1} \leq 1\bigg)\\
& = {\mathbb{P}}_{H_0, S, \xi,\zeta} \bigg[ \log \bigg(\frac{\rho}{1-\rho}\bigg) \cdot \sum_{\ell \notin S\cup\{1\}} \mathbf{1}{\zzz{\ell}_0 = 1} - \mathbf{1}{\zzz{\ell}_0 = -1} \\
& \qquad \qquad + \sum_{\substack{j\neq 0 \\ \zeta_j\mathbf{z}^\star_j= 1}} \AAA{1}_{0j} \log \bigg(\frac{q_1 (1-p_1)}{p_1(1-q_1)}\bigg) + \log \bigg(\frac{1-q_1}{1-p_1}\bigg)\\
& \qquad \qquad + \sum_{\substack{j\neq 0 \\ \zeta_j\mathbf{z}^\star_j = -1}} \AAA{1}_{0j} \log \bigg(\frac{p_1 (1-q_1)}{q_1(1-p_1)}\bigg) + \log \bigg(\frac{1-p_1}{1-q_1}\bigg) \\
& \qquad \qquad + \sum_{\ell \in S}\sum_{\substack{j\neq 0 \\ \zzz{\ell}_j = \xi_\ell}} \AAA{\ell}_{0j} \log \bigg(\frac{q_\ell (1-p_\ell)}{p_\ell(1-q_\ell)}\bigg) + \log \bigg(\frac{1-q_\ell}{1-p_\ell}\bigg)\\
& \qquad \qquad + \sum_{\ell \in S}\sum_{\substack{j\neq 0 \\ \zzz{\ell}_j = -\xi_\ell}} \AAA{\ell}_{0j} \log \bigg(\frac{p_\ell (1-q_\ell)}{q_\ell(1-p_\ell)}\bigg) + \log \bigg(\frac{1-p_\ell}{1-q_\ell}\bigg) \geq 0 \bigg] \\
& = {\mathbb{E}}_{\{\zzz{\ell}_{-0}: \ell\in S\}}\bigg\{ {\mathbb{P}}_{H_0, S, \xi, \zeta} \bigg[ \log \bigg(\frac{\rho}{1-\rho}\bigg) \cdot \sum_{\ell \notin S\cup\{1\}} \mathbf{1}{\zzz{\ell}_0 = 1} - \mathbf{1}{\zzz{\ell}_0 = -1} \\
& \qquad \qquad\qquad\qquad + \sum_{\substack{j\neq 0 \\ \zeta_j\mathbf{z}^\star_j = 1}} \AAA{1}_{0j} \log \bigg(\frac{q_1 (1-p_1)}{p_1(1-q_1)}\bigg) + \log \bigg(\frac{1-q_1}{1-p_1}\bigg)\\
& \qquad \qquad\qquad\qquad + \sum_{\substack{j\neq 0 \\ \zeta_j\mathbf{z}^\star_j = -1}} \AAA{1}_{0j} \log \bigg(\frac{p_1 (1-q_1)}{q_1(1-p_1)}\bigg) + \log \bigg(\frac{1-p_1}{1-q_1}\bigg) \\
& \qquad \qquad\qquad\qquad + \sum_{\ell \in S}\sum_{\substack{j\neq 0 \\ \zzz{\ell}_j = \xi_\ell}} \AAA{\ell}_{0j} \log \bigg(\frac{q_\ell (1-p_\ell)}{p_\ell(1-q_\ell)}\bigg) + \log \bigg(\frac{1-q_\ell}{1-p_\ell}\bigg)\\
& \qquad \qquad\qquad\qquad + \sum_{\ell \in S}\sum_{\substack{j\neq 0 \\ \zzz{\ell}_j = -\xi_\ell}} \AAA{\ell}_{0j} \log \bigg(\frac{p_\ell (1-q_\ell)}{q_\ell(1-p_\ell)}\bigg) + \log \bigg(\frac{1-p_\ell}{1-q_\ell}\bigg) \geq 0 \ \bigg| \ \{\zzz{\ell}_{-0}: \ell\in S\}\bigg] \bigg\}.
\end{align*}
\endgroup
The conditional probability above is equal to
\begingroup
\allowdisplaybreaks
\begin{align*}
& {\mathbb{P}}\bigg( \log\bigg(\frac{\rho}{1-\rho}\bigg)\cdot \sum_{\ell \notin S\cup\{1\}} \ZZZ{\ell} + \sum_{\ell \in S\cup\{1\}} \sum_{i = 1}^{\mmm{\ell}_1} \XXX{\ell}_i \cdot \log \bigg(\frac{q_\ell(1-p_\ell)}{p_\ell(1-q_\ell)}\bigg) + \sum_{\ell \in S\cup\{1\}}\sum_{i = 1}^{\mmm{\ell}_2} \YYY{\ell}_i \cdot \log \bigg(\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg) \\
&\qquad - \sum_{\ell \in S\cup\{1\}} (\mmm{\ell}_1- \mmm{\ell}_2) \cdot \log \bigg(\frac{1-p_\ell}{1-q_\ell}\bigg) \geq 0\bigg),
\end{align*}
\endgroup
where
\begin{align*}
& \XXX{\ell}_i \overset{\textnormal{i.i.d.}}{\sim} \text{Bern}(p_\ell), \ \ \ \YYY{\ell}_i \overset{\textnormal{i.i.d.}}{\sim} \text{Bern}(q_\ell), \ \ \ \ZZZ{\ell} \overset{\textnormal{i.i.d.}}{\sim} 2\text{Bern}(1-\rho)-1, \\
& \mmm{\ell}_1 = \#\{j\neq 0: \zzz{\ell}_j = \xi_\ell\} , \ \ \ \mmm{\ell}_2= \#\{j\neq 0: \zzz{\ell}_j = -\xi_\ell\} \ \ \ \forall \ell \in S,\\
& \mmm{1}_1 = \#\{j\neq 0: \zeta_j\mathbf{z}^\star_j = 1\}, \ \ \ \mmm{1}_2 = \#\{j\neq 0: \zeta_j \mathbf{z}^\star_j = -1\}.
\end{align*}
and $\XXX{\ell}_i$'s, $\YYY{\ell}_i$'s and $\ZZZ{\ell}$'s are jointly independent. By data-processing inequality for total variation, with $\mmm{\ell}:= \mmm{\ell}_1 \lor \mmm{\ell}_2$, the above probability can be further lower bounded by
\begin{align*}
& {\mathbb{P}}\bigg( \log\bigg(\frac{\rho}{1-\rho}\bigg)\cdot \sum_{\ell \notin S\cup\{1\}} \ZZZ{\ell} + \sum_{\ell \in S\cup\{1\}} \sum_{i = 1}^{\mmm{\ell}} \XXX{\ell}_i \cdot \log \bigg(\frac{q_\ell(1-p_\ell)}{p_\ell(1-q_\ell)}\bigg) + \sum_{\ell \in S\cup\{1\}}\sum_{i = 1}^{\mmm{\ell}} \YYY{\ell}_i \cdot \log \bigg(\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg) \geq 0\bigg).
\end{align*}
Thus, we have
\begin{align*}
\RN{1} & \geq {\mathbb{E}}_{\{\zzz{\ell}_{-0}: \ell\in S\}}{\mathbb{P}}\bigg( \log\bigg(\frac{\rho}{1-\rho}\bigg)\cdot \sum_{\ell \notin S\cup\{1\}} \ZZZ{\ell} + \sum_{\ell \in S\cup\{1\}} \sum_{i = 1}^{\mmm{\ell}} \XXX{\ell}_i \cdot \log \bigg(\frac{q_\ell(1-p_\ell)}{p_\ell(1-q_\ell)}\bigg) \\
&\qquad \qquad \qquad \qquad + \sum_{\ell \in S\cup\{1\}}\sum_{i = 1}^{\mmm{\ell}} \YYY{\ell}_i \cdot \log \bigg(\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg) \geq 0\bigg).
\end{align*}
The term $\RN{2}$ is treated similarly.
In summary, by further taking expectation w.r.t. $\{\zzz{\ell}_0: \ell\in S\cup\{1\}\}$ (note that the expression $(1-\rho)\cdot\RN{1} + \rho \cdot\RN{2}$ is taking the expectation w.r.t. $\zzz{1}_0$), we have
\begin{align*}
& \inf_{\hzzz{1}_0} \underset{\mathbf{z}^\star\in {\mathcal{Z}}_T}{\textnormal{ave}}~ {\mathbb{P}}(\hzzz{1}_0 \neq \zzz{1}_0, E_0\cap F_0) \\
& \geq {\mathbb{E}}_{\{\zzz{\ell}: \ell\in S\cup\{1\}\}} \bigg[ \mathbf{1}{E_0\cap F_0} \cdot {\mathbb{P}}\bigg( \log\bigg(\frac{\rho}{1-\rho}\bigg)\cdot \sum_{\ell \notin S\cup\{1\}} \ZZZ{\ell} + \sum_{\ell \in S\cup\{1\}} \sum_{i = 1}^{\mmm{\ell}} \XXX{\ell}_i \cdot \log \bigg(\frac{q_\ell(1-p_\ell)}{p_\ell(1-q_\ell)}\bigg) \\
&\qquad \qquad \qquad \qquad \qquad + \sum_{\ell \in S\cup\{1\}}\sum_{i = 1}^{\mmm{\ell}} \YYY{\ell}_i \cdot \log \bigg(\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg) \geq 0\bigg)\bigg].
\end{align*}
The rest of the proof is exactly the same as the proof of Lemma \ref{lemma:testing_problem_z_star}. We can find some $m' = (1+o(1))n/2$ such that the right-hand side above is lower bounded by
\begin{align*}
{\mathbb{P}}\bigg( \log\bigg(\frac{\rho}{1-\rho}\bigg)\cdot \sum_{\ell \notin S\cup\{1\}} \ZZZ{\ell} + \sum_{\ell \in S\cup\{1\}} \sum_{i = 1}^{m'} \XXX{\ell}_i \cdot \log \bigg(\frac{q_\ell(1-p_\ell)}{p_\ell(1-q_\ell)}\bigg)+ \sum_{\ell \in S\cup\{1\}}\sum_{i = 1}^{m'} \YYY{\ell}_i \cdot \log \bigg(\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg) \geq 0\bigg),
\end{align*}
and the above probability is the error incurred by the likelihood ratio test for the testing problem \eqref{eq:fund_testing_problem_z_l_part_II} with $m' = (1+\delta_n')n/2$.
\subsection{Proof of Lemma \ref{lemma:opt_test_error_z_l}}\label{subappend:prf_opt_test_error_z_l}
The optimal testing error for \eqref{eq:fund_testing_problem_z_l_part_II} (i.e., \eqref{eq:opt_test_error_z_l_part_II}) follows directly from Lemma \ref{lemma:opt_test_error_z_star}.
On the other hand, as shown in the proof of Lemma \ref{lemma:testing_problem_z_l_part_I}, the optimal (weighted) testing error for $\eqref{eq:fund_testing_problem_z_l_part_I}$ is given by
\begin{align*}
& {\mathbb{P}}\bigg(\sum_{i=1}^{m} \XXX{\ell}_i \cdot \log \frac{q_\ell(1-p_\ell)}{p_\ell(1-q_\ell)} + \sum_{i = 1}^{m} \YYY{\ell}_i \cdot \log\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\geq \ZZZ{\ell}\log \frac{1-\rho}{\rho}\bigg)\\
& \geq \rho \cdot {\mathbb{P}}\bigg(\sum_{i=1}^{m} \XXX{\ell}_i \cdot \log \frac{q_\ell(1-p_\ell)}{p_\ell(1-q_\ell)} + \sum_{i = 1}^{m} \YYY{\ell}_i \cdot \log\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\geq \log \frac{\rho}{1-\rho}\bigg).
\end{align*}
The probability in the right-hand side above is calculated in the proof of Lemma \ref{lemma:opt_test_error_z_star} (see Appendix \ref{subsubappend:prf_opt_test_error_z_star_odd_Sc}, \eqref{eq:target_of_lower_bound_z_star_odd} with $S = \{\ell\}$). Now recognizing $\rho = e^{-(1+o(1))2J_\rho}$ gives \eqref{eq:opt_test_error_z_l_part_I}.
\subsection{Optimal Tests for Minimizing Weighted Type-\texorpdfstring{\RN{1}}{I} and \texorpdfstring{\RN{2}}{II} Errors}
The following lemma is a version of Neyman-Pearson lemma, and it allows us to characterize the optimal test that minimizes the \emph{weighted average} of type-\RN{1} and type-\RN{2} errors.
\begin{lemma}[Neyman--Pearson lemma]
\label{lemma: optimal_test}
Consider testing $H_0: X\sim {\mathbb{P}}_0$ against $H_1: X\sim {\mathbb{P}}_1$, where ${\mathbb{P}}_0$ and ${\mathbb{P}}_1$ have densities $p_0(x), p_1(x)$ respectively against some dominating measure $\mu(x)$. For any $\omega_0, \omega_1 > 0$, the optimal test that minimizes $\omega_0 \cdot \textnormal{Type-I error} + \omega_1 \cdot \textnormal{Type-II error}$ is given by rejecting $H_0$ when
$$
\frac{p_0(x)}{p_1(x)} \leq \frac{\omega_1}{\omega_0}.
$$
\end{lemma}
\begin{proof}
This is a restatement of Problem 3.10 in \cite{lehmann2006testing}, and we provide a proof here for completeness.
It suffices to consider $w_0 = 1-\rho, w_1 = \rho$ for some $\rho \in (0, 1)$.
The optimal error is given by
\begin{align*}
\inf_{\psi_0, \psi_1: \psi_0 + \psi_1 = 1} \int (1-\rho) \cdot p_0(x) \psi_1(x) d\mu(x) + \rho \cdot p_1(x)\psi_0(x) d\mu(x).
\end{align*}
When $(1-\rho)\cdot p_0(x)\geq \rho \cdot p_1(x)$, the integrand is lower bounded by
$$
\rho \cdot p_1(x) = (1-\rho) \cdot p_0(x) \land \rho \cdot p_1(x).
$$
When $(1-\rho)\cdot p_0(x)\leq \rho \cdot p_1(x)$, the integrand is lower bounded by
$$
(1-\rho) \cdot p_0(x) = (1-\rho) \cdot p_0(x) \land \rho \cdot p_1(x).
$$
This gives
\begin{align*}
& \inf_{\psi_0, \psi_1: \psi_0 + \psi_1 = 1} \int (1-\rho) \cdot p_0(x) \psi_1(x) d\mu(x) + \rho \cdot p_1(x)\psi_0(x) d\mu(x) \geq \int (1-\rho) \cdot p_0(x) \land \rho \cdot p_1(x) d\mu(x).
\end{align*}
On the other hand, if we take $\psi_1(x) = \mathbf{1}{(1-\rho)\cdot p_0(x)\leq \rho \cdot p_1(x)} $, then it is easy to check that the integrand is exactly equal to $(1-\rho) \cdot p_0(x) \land \rho \cdot p_1(x)$. This gives the desired result.
\end{proof}
The following lemma asserts that the testing error can only be smaller when extra information is present, and can be regarded as an instance of ``data-processing inequalities''.
\begin{lemma}
\label{lemma: general_data_processing}
Under the setup of Lemma \ref{lemma: optimal_test}, let us additionally consider testing $H'_0: (X, Y)\sim {\mathbb{Q}}_0$ against $H'_1: (X, Y)\sim {\mathbb{Q}}_1$, where
the marginal of the first coordinate of ${\mathbb{Q}}_0$ (resp. ${\mathbb{Q}}_1$) agrees with ${\mathbb{P}}_0$ (resp. ${\mathbb{Q}}_1$), and there exist densities $q_0(x, y)$ and $q_1(x, y)$ such that
$$
d {\mathbb{Q}}_0(x, y) = q_0(x, y) d\mu(x)d\mu(y), \ \ \ d {\mathbb{Q}}_1(x, y) = q_1(x, y) d\mu(x)d\mu(y).
$$
Then we have
$$
\inf_{\psi}~ w_0 \cdot {\mathbb{E}}_{H_0} \psi + w_1 \cdot {\mathbb{E}}_{H_1} [1-\psi] \geq \inf_{\psi'}~ w_0 \cdot {\mathbb{E}}_{H'_0} \psi' + w_1 \cdot {\mathbb{E}}_{H'_1} [1-\psi'],
$$
where $\psi, \psi'$ are testing functions for $H_0$ v.s. $H_1$ and $H_0'$ v.s. $H_1'$, respectively.
\end{lemma}
\begin{proof}
If $\omega_0 = \omega_1$, this lemma immediately follows from the data-processing inequality for total variation distances. But we need more work for the general case.
We again without loss of generality assume $w_0 = 1-\rho$, $w_1 = \rho$ for some $\rho \in (0, 1)$. By the proof of Lemma \ref{lemma: optimal_test} and the existence of densities, it suffices to show
$$
\int (1-\rho) \cdot p_0 (x) \land \rho \cdot p_1(x) d\mu(x) \geq \iint (1-\rho)\cdot q_0(x, y) \land \rho \cdot q_1(x, y) d\mu(x) d\mu(y),
$$
which is implied by
$$
(1-\rho) \cdot p_0(x) \land \rho \cdot p_1(x) \geq \int (1-\rho) \cdot q_0(x, y) \land \rho \cdot q_1(x, y) d\mu(y).
$$
Note that
\begin{align*}
\int (1-\rho) \cdot q_0(x, y) \land \rho \cdot q_1(x, y) d\mu(y) & \leq \int (1-\rho) \cdot q_0(x, y) d\mu(y) \land \int \rho \cdot q_1(x, y) d\mu(y) \\
& = (1-\rho) \cdot p_0(x) \land \rho \cdot p_0(x),
\end{align*}
which is the desired result.
\end{proof}
\section{Proofs in Section \ref{sec:minimax_rate}}\label{append:prf_minimax_rate}
\subsection{Proof of Theorem \ref{thm:opt_glob_est}}\label{prf:thm:opt_glob_est}
The desired result follows from the following two propositions.
\begin{proposition}
\label{prop:opt_glob_even}
For any $\overline{\delta}_n = o(1)$ and any $c\in(0, 1)$, there exists another $\overline{\delta}_n'=o(1)$ such that
\begin{align*}
& \sum_{S: |S^c|\textnormal{ even}} \exp\big\{- (1-\overline{\delta}_n) \big(|S^c|J_\rho + \psi_S^\star(0)\big)\big\} \\
& \leq 2\exp\big\{-(1-\overline{\delta}_n') \min_{S:|S^c|\textnormal{ even}}\big(|S^c|J_\rho + \psi_S^\star(0)\big) + \log L + L e^{-(1-\overline{\delta}_n) cJ_\rho}\big\}.
\end{align*}
\end{proposition}
\begin{proof}
See Appendix \ref{prf:prop:opt_glob_even}.
\end{proof}
\begin{proposition}
\label{prop:opt_glob_odd}
For any $\overline{\delta}_n = o(1)$ and any $c\in(0, 1)$, there exists another $\overline{\delta}_n' = o(1)$ such that
\begin{align*}
& \sum_{S: |S^c|\textnormal{ odd}} \exp\big\{- (1-\overline{\delta}_n) \big(|S^c|J_\rho + \psi_S^\star(-2J_\rho)\big)\big\} \\
& \leq 2 \exp\big\{-(1-\overline{\delta}_n') \min_{S:|S^c|\textnormal{ odd}}\big(|S^c|J_\rho + \psi_S^\star(-2J_\rho)\big) + \log L + Le^{-(1-\overline{\delta}_n)cJ_\rho}\big\}.
\end{align*}
\end{proposition}
\begin{proof}
See Appendix \ref{prf:prop:opt_glob_odd}.
\end{proof}
\subsubsection{Proof of Proposition \ref{prop:opt_glob_even}}\label{prf:prop:opt_glob_even}
For any $t\in[0, 1]$, let us define
\begin{align}
\label{eq:sets}
S_{\geq , t}:= \{\ell\in[L]: n\III{\ell}_t/2 \geq J_\rho\}, \qquad S_{<, t}:= \{\ell\in[L]: n\III{\ell}_t/2 < J_\rho\}.
\end{align}
For notational simplicity we let $m = n/2$. In this proof we will assume both $S_{\geq, 1/2}$ and $S_{<, 1/2}$ are non-empty. The proof when one of them is empty is nearly identical. We begin by noting that
\begin{align*}
&\sum_{S\subseteq[L]: |S^c|\textnormal{ even}} \exp\bigg\{ -(1-\overline{\delta}_n)\cdot \bigg( |S^c|J_\rho + \psi_S^\star(0)\bigg) \bigg\} + \sum_{S\subseteq[L]: |S^c|\textnormal{ odd}} \exp\bigg\{ -(1-\overline{\delta}_n)\cdot \bigg( |S^c|J_\rho + \psi_S^\star(0)\bigg) \bigg\}\\
\label{eq:opt_glob_even_summation}
& = \prod_{\ell\in[L]} \bigg(e^{-(1-\overline{\delta}_n) J_\rho} + e^{-(1-\overline{\delta}_n) m\III{\ell}_{1/2}}\bigg)\addtocounter{equation}{1}\tag{\theequation},
\end{align*}
and that
\begin{align*}
&\sum_{S\subseteq[L]: |S^c|\textnormal{ even}} \exp\bigg\{ -(1-\overline{\delta}_n)\cdot \bigg( |S^c|J_\rho + \psi_S^\star(0)\bigg) \bigg\} - \sum_{S\subseteq[L]: |S^c|\textnormal{ odd}} \exp\bigg\{ -(1-\overline{\delta}_n)\cdot \bigg( |S^c|J_\rho + \psi_S^\star(0)\bigg) \bigg\}\\
\label{eq:opt_glob_even_difference}
& = \prod_{\ell\in[L]} \bigg(-e^{-(1-\overline{\delta}_n) J_\rho} + e^{-(1-\overline{\delta}_n) m\III{\ell}_{1/2}}\bigg). \addtocounter{equation}{1}\tag{\theequation}
\end{align*}
We split the discussion into two cases.
~\\
{\noindent \bf Case A: $|\boldsymbol{S_{\geq, 1/2}}|$ is even.} In this case, by \eqref{eq:opt_glob_even_summation}, we have
\begin{align*}
& \sum_{S: |S^c|\textnormal{ even}} \exp\big\{- (1-\overline{\delta}_n) \big(|S^c|J_\rho + \psi_S^\star(0)\big)\big\} \\
& \leq \prod_{\ell\in[L]} \bigg(e^{-(1-\overline{\delta}_n) J_\rho} + e^{-(1-\overline{\delta}_n) m\III{\ell}_{1/2}}\bigg)\\
& = \exp\bigg\{-(1-\overline{\delta}_n) \bigg(|S_{\geq , 1/2}|J_\rho + \sum_{\ell\in S_{<, 1/2}}m\III{\ell}_{1/2}\bigg)\bigg\}\\
& \qquad \times \prod_{\ell\in S_{\geq, 1/2}} \bigg(1 + e^{-(1-\overline{\delta}_n)(m\III{\ell}_{1/2} - J_\rho)}\bigg)
\cdot \prod_{\ell\in S_{<, 1/2}} \bigg(1 + e^{-(1-\overline{\delta}_n)(J_\rho-m\III{\ell}_{1/2})}\bigg)\\
& \leq\exp\bigg\{-(1-\overline{\delta}_n) \bigg(|S_{\geq , 1/2}|J_\rho + \sum_{\ell\in S_{<, 1/2}}m\III{\ell}_{1/2}\bigg)\bigg\}\\
& \qquad \times \exp\bigg\{\sum_{\ell\in S_{\geq, 1/2}} e^{-(1-\overline{\delta}_n)(m \III{\ell}_{1/2}-J_\rho)} + \sum_{\ell\in S_{<, 1/2}} e^{-(1-\overline{\delta}_n)(J_\rho - m\III{\ell}_{1/2})}\bigg\}\\
& {\leq} \exp\bigg\{-(1-\delta_{1,n }) \bigg(|S_{\geq , 1/2}|J_\rho + \sum_{\ell\in S_{<, 1/2}}m\III{\ell}_{1/2}\bigg) + \sum_{\ell\in S_{<, 1/2}} e^{-(1-\overline{\delta}_n)(J_\rho - m\III{\ell}_{1/2})}\bigg\},
\end{align*}
where the last inequality is by $e^{-(1-\overline{\delta}_n)(m\III{\ell}_{1/2}-J_\rho)}\leq 1\ll J_\rho$ for any $\ell \in S_{\geq ,1/2}$, and $\delta_{1, n}=o(1)$.
Let
$$
S_{<, 1/2}^{-} := \{\ell\in S_{<, 1/2}: m\III{\ell}_{1/2}\leq (1-c)J_\rho\}, \qquad S^+_{<, 1/2} := \{\ell\in S_{<, 1/2}: m\III{\ell}_{1/2}> (1-c)J_\rho\}.
$$
For $\ell\in S_{<, 1/2}^-$, we have $e^{-(1-\overline{\delta}_n)(J_\rho - m\III{\ell}_{1/2})}\leq e^{-(1-\overline{\delta}_n)cJ_\rho}$ and for $\ell\in S_{<,1/2}^+$, we have $e^{-(1-\overline{\delta}_n)(J_\rho - m\III{\ell}_{1/2})}\leq 1$.
Thus we have
\begin{align*}
& \sum_{S: |S^c|\textnormal{ even}} \exp\big\{- (1-\overline{\delta}_n) \big(|S^c|J_\rho + \psi_S^\star(0)\big)\big\} \\
& \overset{(*)}{\leq} \exp\bigg\{-(1-\delta_{1,n})\bigg(|S_{\geq, 1/2}| J_\rho + \sum_{\ell\in S_{<, 1/2}^+}m\III{\ell}_{1/2} \big(1- \frac{1}{m\III{\ell}_{1/2}} \big) + \sum_{\ell\in S_{<,1/2}^-} m\III{\ell}_{1/2}\bigg) + L e^{-(1-\overline{\delta}_n)cJ_\rho}\bigg\}\\
& \overset{(**)}{\leq} \exp\bigg\{-(1-\delta_{1,n})\bigg(|S_{\geq, 1/2}| J_\rho + \sum_{\ell\in S_{<, 1/2}^+}m\III{\ell}_{1/2} \big(1- \frac{1}{(1-c)J_\rho} \big) + \sum_{\ell\in S_{<,1/2}^-} m\III{\ell}_{1/2}\bigg) + L e^{-(1-\overline{\delta}_n)cJ_\rho}\bigg\}\\
\label{eq:opt_glob_even_caseA}
& \overset{(***)}{\leq} \exp\bigg\{-(1-\delta_{2,n }) \min_{S: |S^c|\textnormal{ even}}\bigg(|S^c|J_\rho + \sum_{\ell\in S}m\III{\ell}_{1/2}\bigg) + Le^{-(1-\overline{\delta}_n) cJ_\rho} \bigg\}\addtocounter{equation}{1}\tag{\theequation},
\end{align*}
where $(*)$ is by $e^{-(1-\overline{\delta}_n)(J_\rho - m\III{\ell}_{1/2})}\leq 1$ for $\ell\in S_{<, 1/2}^+$ and $e^{-(1-\overline{\delta}_n)(J_\rho - m\III{\ell}_{1/2})}\leq e^{-(1-\overline{\delta}_n)cJ_\rho}$ for $\ell\in S_{<, 1/2}^-$, $(**)$ is by the definition of $S_{<, 1/2}^+$, $(***)$ is by our assumption that $|S_{\geq, 1/2}|$ is even, and $\delta_{2, n} = o(1)$.
~\\
{\noindent \bf Case B: $|\boldsymbol{S_{\geq ,1/2}}|$ is odd.} Invoking \eqref{eq:opt_glob_even_summation} and \eqref{eq:opt_glob_even_difference}, we have
\begin{align*}
& \sum_{S: |S^c|\textnormal{ even}} \exp\big\{- (1-\overline{\delta}_n) \big(|S^c|J_\rho + \psi_S^\star(0)\big)\big\} \\
& = \frac{1}{2} \exp\bigg\{-(1-\overline{\delta}_n)\bigg(|S_{\geq, 1/2}|J_\rho + \sum_{\ell\in S_{<, 1/2}}m\III{\ell}_{1/2}\bigg)\bigg\} \\
& \qquad \times \bigg[
\prod_{\ell\in S_{\geq ,1/2}} \bigg(1+ e^{-(1-\overline{\delta}_n)(m\III{\ell}_{1/2}-J_\rho)}\bigg)
\prod_{\ell\in S_{<, 1/2}} \bigg(1+ e^{-(1-\overline{\delta}_n)(J_\rho - m\III{\ell}_{1/2})}\bigg)\\
& \qquad \qquad -
\prod_{\ell\in S_{\geq ,1/2}} \bigg(1- e^{-(1-\overline{\delta}_n)(m\III{\ell}_{1/2}-J_\rho)}\bigg)
\prod_{\ell\in S_{<, 1/2}} \bigg(1- e^{-(1-\overline{\delta}_n)(J_\rho - m\III{\ell}_{1/2})}\bigg)
\bigg]\\
& = \frac{1}{2} \exp\bigg\{-(1-\overline{\delta}_n)\bigg(|S_{\geq, 1/2}|J_\rho + \sum_{\ell\in S_{<, 1/2}}m\III{\ell}_{1/2}\bigg)\bigg\}\cdot \prod_{\ell\in S_{\geq, 1/2}}\bigg(1+ e^{-(1-\overline{\delta}_n)(m\III{\ell}_{1/2}-J_\rho)}\bigg) \\
& \qquad \times
\bigg[ \prod_{\ell\in S_{<, 1/2}}\bigg(1+ e^{-(1-\overline{\delta}_n)(J_\rho - m\III{\ell}_{1/2})}\bigg) - \prod_{\ell\in S_{<, 1/2}}\bigg(1- e^{-(1-\overline{\delta}_n)(J_\rho - m\III{\ell}_{1/2})}\bigg) \bigg]\\
& \qquad + \frac{1}{2} \exp\bigg\{-(1-\overline{\delta}_n)\bigg(|S_{\geq, 1/2}|J_\rho + \sum_{\ell\in S_{<, 1/2}}m\III{\ell}_{1/2}\bigg)\bigg\}
\cdot \prod_{\ell\in S_{<, 1/2}}\bigg(1- e^{-(1-\overline{\delta}_n)(J_\rho - m\III{\ell}_{1/2})}\bigg) \\
& \qquad \qquad \times
\bigg[ \prod_{\ell\in S_{\geq, 1/2}}\bigg(1 + e^{-(1-\overline{\delta}_n)(m\III{\ell}_{1/2}- J_\rho )}\bigg) - \prod_{\ell\in S_{\geq, 1/2}}\bigg(1- e^{-(1-\overline{\delta}_n)(m\III{\ell}_{1/2}-J_\rho )}\bigg) \bigg].
\end{align*}
We need the following estimate.
\begin{lemma}
\label{lemma: diff_of_products}
Let $\{x_i: i\in[n]\}, \{y_i:i\in[n]\}$ be two collections of complex numbers. Assume $x_i$'s are outside of the unit disk and $y_i$'s are inside the unit disk, then
$$
\bigg|\prod_{i\in[n]} x_i - \prod_{i\in[n]} y_i\bigg| \leq \bigg(\sum_{i\in[n]}|x_i - y_i|\bigg) \cdot \prod_{i\in[n]} |x_i|.
$$
\end{lemma}
\begin{proof}
We first prove the following algebraic identity:
$$
\prod_{i\in[n]} x_i - \prod_{i\in[n]} y_i = \sum_i (x_i - y_i) \cdot \bigg(\prod_{j=1}^{i-1} x_j \bigg) \cdot \bigg(\prod_{j=i+1}^n y_i\bigg)
$$
with the convention that $\prod_{i = j}^k x_i = 1$ if $k<j$, and then the desired result follows from triangle inequality. We induct on $n$. The case of $n = 1$ is trivial. Assume the identity holds for $n = k$. Now for $n = k+1$, we have
\begin{align*}
\prod_{i=1}^{k+1} x_i - \prod_{i=1}^{k+1} y_i & = x_{k+1} \prod_{i=1}^k x_i - y_{k+1} \prod_{i=1}^k y_i\\
& = x_{k+1} \prod_{i=1}^k x_i - y_{k+1} \prod_{i=1}^k x_i + y_{k+1} \prod_{i=1}^k x_i - y_{k+1} \prod_{i=1}^k y_i\\
& = (x_{k+1}-y_{k+1}) \prod_{i=1}^k x_i + y_{k+1} \cdot \sum_{i=1}^k(x_i - y_i)\cdot \prod_{j=1}^{i-1} x_j \cdot \prod_{j=i+1}^k y_i \\
& = (x_{k+1}-y_{k+1}) \prod_{i=1}^k x_i + \sum_{i=1}^k(x_i - y_i)\cdot \prod_{j=1}^{i-1} x_j \cdot \prod_{j=i+1}^{k+1} y_i\\
& = \sum_{i=1}^{k+1}(x_i - y_i)\cdot \prod_{j=1}^{i-1} x_j \cdot \prod_{j=i+1}^{k+1} y_i,
\end{align*}
which finishes the proof.
\end{proof}
By Lemma \ref{lemma: diff_of_products}, we have
\begin{align*}
& \sum_{S: |S^c|\textnormal{ even}} \exp\big\{- (1-\overline{\delta}_n) \big(|S^c|J_\rho + \psi_S^\star(0)\big)\big\} \\
& \leq \exp\bigg\{-(1-\overline{\delta}_n)\bigg(|S_{\geq, 1/2}|J_\rho + \sum_{\ell\in S_{<, 1/2}}m\III{\ell}_{1/2}\bigg)\bigg\}
\cdot \prod_{\ell\in S_{\geq, 1/2}} \bigg(1+ e^{-(1-\overline{\delta}_n)(m\III{\ell}_{1/2}-J_\rho)}\bigg)\\
& \qquad \times \bigg(\sum_{\ell\in S_{<, 1/2}} e^{-(1-\overline{\delta}_n)(J_\rho - m\III{\ell}_{1/2})}\bigg) \cdot \prod_{\ell\in S_{<, 1/2}} \bigg(1+ e^{-(1-\overline{\delta}_n)(J_\rho - m\III{\ell}_{1/2})}\bigg)\\
& \qquad +
\exp\bigg\{-(1-\overline{\delta}_n)\bigg(|S_{\geq, 1/2}|J_\rho + \sum_{\ell\in S_{<, 1/2}}m\III{\ell}_{1/2}\bigg)\bigg\}\\
& \qquad \qquad \times \bigg(\sum_{\ell\in S_{\geq, 1/2}} e^{-(1-\overline{\delta}_n)(m \III{\ell}_{1/2}-J_\rho)} \bigg)
\cdot \prod_{\ell\in S_{\geq ,1/2}} \bigg(1+ e^{-(1-\overline{\delta}_n)(J_\rho - m \III{\ell}_{1/2})}\bigg)\\
& \leq \exp\bigg\{-(1-\delta_{2, n})\bigg(|S_{\geq, 1/2}|J_\rho + \sum_{\ell\in S_{<, 1/2}}m\III{\ell}_{1/2}\bigg)\bigg\}\\
& \qquad \times \bigg(\sum_{\ell\in S_{<, 1/2}} e^{-(1-\overline{\delta}_n)(J_\rho - m\III{\ell}_{1/2})}\bigg) \cdot \prod_{\ell\in S_{<, 1/2}} \bigg(1+ e^{-(1-\overline{\delta}_n)(J_\rho - m\III{\ell}_{1/2})}\bigg)\\
& \qquad +
\exp\bigg\{-(1-\delta_{2, n})\bigg(|S_{\geq, 1/2}|J_\rho + \sum_{\ell\in S_{<, 1/2}}m\III{\ell}_{1/2}\bigg)\bigg\}
\cdot \bigg(\sum_{\ell\in S_{\geq, 1/2}} e^{-(1-\overline{\delta}_n)(m \III{\ell}_{1/2}-J_\rho)} \bigg),
\end{align*}
where $\delta_{2,n} = o(1)$ and the last inequality is by $e^{-(1-\overline{\delta}_n)(m\III{\ell}_{1/2}-J_\rho)}\leq 1 \ll J_\rho$ for any $\ell\in S_{\geq, 1/2}$.
Defining
$$
\ell_{<, 1/2} := \mathop{\rm argmax}_{\ell\in S_{<, 1/2}} \frac{m\III{\ell}_{1/2}}{J_\rho}, \qquad \ell_{\geq , 1/2} := \mathop{\rm argmin}_{\ell\in S_{\geq, 1/2}} \frac{m\III{\ell}_{1/2}}{J_\rho},
$$
we have
\begin{align*}
& \sum_{S: |S^c|\textnormal{ even}} \exp\big\{- (1-\overline{\delta}_n) \big(|S^c|J_\rho + \psi_S^\star(0)\big)\big\} \\
& \leq \exp\bigg\{-(1-{\delta_{2, n}})\bigg(|S_{\geq, 1/2}|J_\rho + (J_\rho - m\III{\ell_{<, 1/2}}_{1/2}) + \sum_{\ell\in S_{<, 1/2}}m\III{\ell}_{1/2}\bigg)\bigg\} \\
& \qquad \times |S_{<, 1/2}| \exp\bigg\{\sum_{\ell \in S_{<, 1/2}} e^{-(1-\overline{\delta}_n)(J_\rho - m\III{\ell}_{1/2})}\bigg\}\\
& \qquad + \exp\bigg\{-(1-{\delta_{2,n}})\bigg(|S_{\geq, 1/2}|J_\rho + (m\III{\ell_{\geq, 1/2}}_{1/2} - J_\rho ) + \sum_{\ell\in S_{<, 1/2}}m\III{\ell}_{1/2}\bigg)\bigg\} \times |S_{\geq, 1/2}|,
\end{align*}
Using similar argument as those in Case A, we can proceed by
\begin{align*}
& \sum_{S: |S^c|\textnormal{ even}} \exp\big\{- (1-\overline{\delta}_n) \big(|S^c|J_\rho + \psi_S^\star(0)\big)\big\} \\
& \leq \exp\bigg\{-(1-{\delta_{3, n}})\bigg(|S_{\geq, 1/2}|J_\rho + (J_\rho - m\III{\ell_{<, 1/2}}_{1/2}) + \sum_{\ell\in S_{<, 1/2}}m\III{\ell}_{1/2}\bigg) + \log L + Le^{-(1-\overline{\delta}_n) c J_\rho}\bigg\} \\
& \qquad + \exp\bigg\{-(1-{\delta_{2,n}})\bigg(|S_{\geq, 1/2}|J_\rho + (m\III{\ell_{\geq, 1/2}}_{1/2} - J_\rho ) + \sum_{\ell\in S_{<, 1/2}}m\III{\ell}_{1/2}\bigg) + \log L\bigg\
\end{align*}
for some $\delta_{3,n}=o(1)$.
Since $|S_{\geq, 1/2}\cup\{\ell_{<, 1/2}\}|$ and $|S_{\geq, 1/2}\setminus \{\ell_{\geq, 1/2}\}|$ are both even, we conclude that
\begin{align*}
& \sum_{S: |S^c|\textnormal{ even}} \exp\big\{- (1-\overline{\delta}_n) \big(|S^c|J_\rho + \psi_S^\star(0)\big)\big\}\\
\label{eq:opt_glob_even_caseB}
& \leq 2\exp\bigg\{-(1-\delta_{4,n }) \min_{S: |S^c|\textnormal{ even}}\bigg(|S^c|J_\rho + \sum_{\ell\in S}m\III{\ell}_{1/2}\bigg) + \log L + Le^{-(1-\overline{\delta}_n)cJ_\rho}\bigg\} \addtocounter{equation}{1}\tag{\theequation}
\end{align*}
for some $\delta_{4, n}=o(1)$
The proof is concluded by combining \eqref{eq:opt_glob_even_caseA} and \eqref{eq:opt_glob_even_caseB}.
\subsubsection{Proof of Proposition \ref{prop:opt_glob_odd}}\label{prf:prop:opt_glob_odd}
Recall the definition of $S_{\geq, t}$ and $S_{<, t}$ in \eqref{eq:sets}. Let $m = n/2$. For any $S\subseteq[L]$, define
$$
t_S:= \mathop{\rm argmax}_{t\in[0, 1]} -2t J_\rho + \sum_{\ell \in S}m\III{\ell}_{t}.
$$
Fix an arbitrary constant $t\in [0, 1]$.
Similar to the proof of Proposition \ref{prop:opt_glob_even}, we will assume $S_{\geq, t}$ and $S_{< , t}$ are both non-empty, and the case of one of them being empty is treated similarly.
We have
\begin{align}
& \sum_{S\subseteq[L]: |S^c|\textnormal{ odd}} \exp\bigg\{ -(1-\overline{\delta}_n)\cdot \bigg( (|S^c|+1)J_\rho + \psi_S^\star(-2J_\rho)\bigg) \bigg\}\nonumber\\
& =\sum_{S\subseteq[L]: |S^c|\textnormal{ odd}} \exp\bigg\{ -(1-\overline{\delta}_n)\cdot \bigg( |S^c|J_\rho + (1-2t_S) J_\rho + \sum_{\ell \in S} m\III{\ell}_{t_S}\bigg) \bigg\}\nonumber\\
\label{eq:opt_glob_even_general_t}
& \leq \sum_{S\subseteq[L]: |S^c|\textnormal{ odd}} \exp\bigg\{ -(1-\overline{\delta}_n)\cdot \bigg( |S^c|J_\rho + (1-2t) J_\rho + \sum_{\ell \in S} m\III{\ell}_{t}\bigg) \bigg\}.
\end{align}
Similar to \eqref{eq:opt_glob_even_summation} and \eqref{eq:opt_glob_even_difference}, we have
\begin{align*}
& \sum_{S\subseteq[L]: |S^c|\textnormal{ odd}} \exp\bigg\{ -(1-\overline{\delta}_n)\cdot \bigg( |S^c|J_\rho + (1-2t) J_\rho + \sum_{\ell \in S} m\III{\ell}_{t}\bigg) \bigg\} \\
& \qquad +\sum_{S\subseteq[L]: |S^c|\textnormal{ even}} \exp\bigg\{ -(1-\overline{\delta}_n)\cdot \bigg( |S^c|J_\rho + (1-2t) J_\rho + \sum_{\ell \in S} m\III{\ell}_{t}\bigg) \bigg\} \\
\label{eq:opt_glob_odd_summation}
& = e^{-(1-\overline{\delta}_n) (1-2t)J_\rho} \prod_{\ell\in[L]} \bigg(e^{-(1-\overline{\delta}_n)J_\rho} + e^{-(1-\overline{\delta}_n) m\III{\ell}_{t}}\bigg),\addtocounter{equation}{1}\tag{\theequation}
\end{align*}
and
\begin{align*}
& -\sum_{S\subseteq[L]: |S^c|\textnormal{ odd}} \exp\bigg\{ -(1-\overline{\delta}_n)\cdot \bigg( |S^c|J_\rho + (1-2t) J_\rho + \sum_{\ell \in S} m\III{\ell}_{t}\bigg) \bigg\} \\
& \qquad +\sum_{S\subseteq[L]: |S^c|\textnormal{ even}} \exp\bigg\{ -(1-\overline{\delta}_n)\cdot \bigg( |S^c|J_\rho + (1-2t) J_\rho + \sum_{\ell \in S} m\III{\ell}_{t}\bigg) \bigg\} \\
\label{eq:opt_glob_odd_difference}
& = e^{-(1-\overline{\delta}_n) (1-2t)J_\rho} \prod_{\ell\in[L]} \bigg(-e^{-(1-\overline{\delta}_n)J_\rho} + e^{-(1-\overline{\delta}_n) m\III{\ell}_{t}}\bigg).\addtocounter{equation}{1}\tag{\theequation}
\end{align*}
With the above two equations at hand, using similar arguments as those in the proof of Proposition \ref{prop:opt_glob_even}, we arrive at
\begin{align*}
& \sum_{S\subseteq[L]: |S^c|\textnormal{ odd}} \exp\bigg\{ -(1-\overline{\delta}_n)\cdot \bigg( (|S^c|+1)J_\rho + \psi_S^\star(-2J_\rho)\bigg) \bigg\}\\
\label{eq:opt_glob_odd_summary_two_cases}
& {\leq} 2\exp\bigg\{-(1-\delta_{1,n})\min_{S: |S^c|\textnormal{ odd}}\bigg(|S^c|J_\rho + (1-2t)J_\rho + \sum_{\ell\in S} m\III{\ell}_t\bigg)+ \log L + Le^{-(1-\overline{\delta}_n) cJ_\rho}\bigg\},\addtocounter{equation}{1}\tag{\theequation}
\end{align*}
where $\delta_{1,n}=o(1)$.
We claim that
\begin{equation}
\label{eq:minimax_equality_z_star}
\sup_{t\in[0, 1]}\min_{\substack{S\subseteq[L]\\ |S^c|\textnormal{ odd}}} \big(|S^c|J_\rho + (1-2t)J_\rho + \sum_{\ell\in S} m\III{\ell}_{t} \big)= \min_{\substack{S\subseteq[L]\\ |S^c|\textnormal{ odd}}} \big(\sup_{t\in[0, 1]} |S^c|J_\rho + (1-2t)J_\rho + \sum_{\ell\in S} m\III{\ell}_{t}\big).
\end{equation}
If this claim holds, then plugging the optimal $t$ to \eqref{eq:opt_glob_odd_summary_two_cases} gives the desired result.
To prove \eqref{eq:minimax_equality_z_star}, we need the following theorem.
\begin{theorem}[Sion's minimax theorem]
\label{thm:minimax}
Let $X$ be a compact convex subset of a topological vector space and $Y$ be a convex subset of a topological vector space. If $f$ is a real-valued function on $X\times Y$ such that
\begin{enumerate}
\item for each $x\in X$, $f(x, \cdot)$ is upper semi-continuous (usc) and quasi-concave on $Y$,
\item for each $y\in Y$, $f(\cdot, y)$ is lower semi-continuous (lsc) and quasi-convex on $X$,
\end{enumerate}
then
$$
\min_{x\in X} \sup_{y\in Y} f(x, y) = \sup_{y\in Y} \min_{x\in X} f(x, y).
$$
\end{theorem}
To use the above theorem, we let
$$
{\mathcal{V}} := \bigg\{v\in\{0, 1\}^L: \#\{\ell\in[L]: v_\ell = 0\} \textnormal{ is odd}\bigg\}.
$$
Then the right-hand side of \eqref{eq:minimax_equality_z_star} is
\begin{align*}
\sup_{t\in[0, 1]} \min_{v\in{\mathcal{V}}}\ f(v, t),
\end{align*}
where
\begin{align*}
f(v, t) & = \langle J_\rho \mathbf{1}_L, \mathbf{1}_L - v \rangle + (1-2t)J_\rho + \langle v, m I_{*, t} \rangle\\
& = LJ_\rho + (1-2t)J_\rho + \langle v, mI_{*, t} - J_\rho \mathbf{1}_L\rangle,
\end{align*}
where we let $I_{*, t}$ be the $L$-dimensional vector whose $\ell$-th entry is $\III{\ell}_{t}$. It suffices to show
$$
\min_{v\in{\mathcal{V}}}\sup_{t\in[0, 1]}\ f(v, t) = \sup_{t\in[0, 1]} \min_{v\in{\mathcal{V}}}\ f(v, t).
$$
We are to invoke a version of minimax theorem, but an immediate difficulty is that ${\mathcal{V}}$ is non-convex. Fortunately we have the following lemma.
\begin{lemma}
\label{lemma: hidden_convexity_of_the_vertex_set}
For any $t\in[0, 1]$, we have
$$
\min_{v\in {\mathcal{V}}} f(v, t) = \min_{v\in \textnormal{conv}({\mathcal{V}})} f(v, t),
$$
where $\textnormal{conv}({\mathcal{V}})$ is the convex hull of ${\mathcal{V}}$.
\end{lemma}
\begin{proof}
This follows from the fact that the convex hull of ${\mathcal{V}}$ is a polytope.
\end{proof}
By Lemma \ref{lemma: hidden_convexity_of_the_vertex_set}, we conclude that
$$
\sup_{t\in[0, 1]} \min_{v\in{\mathcal{V}}}\ f(v, t) = \sup_{t\in[0, 1]} \min_{v\in\textnormal{conv}({\mathcal{V}})}\ f(v, t)
$$
and
$$
\min_{v\in{\mathcal{V}}}\sup_{t\in[0, 1]} \ f(v, t) = \min_{v\in\textnormal{conv}({\mathcal{V}})}\sup_{t\in[0, 1]}\ f(v, t).
$$
Hence, it suffices to show
$$
\min_{v\in\textnormal{conv}({\mathcal{V}})}\sup_{t\in[0, 1]}\ f(v, t) = \sup_{t\in[0, 1]} \min_{v\in\textnormal{conv}({\mathcal{V}})}\ f(v, t).
$$
We make a few observations:
\begin{enumerate}
\item $\textnormal{conv}({\mathcal{V}})$ is a compact convex subset of the topological vector space ${\mathbb{R}}^L$ (with the Euclidean topology);
\item $[0, 1]$ is a convex subset of the topological vector space ${\mathbb{R}}$ (again with the Euclidean topology);
\item For each $v\in \textnormal{conv}({\mathcal{V}})$, the function $f(v, \cdot)$ is continuous and concave (and hence quasi-concave) in $t$;
\item For each $t\in[0, 1]$, the function $f(\cdot, t)$ is linear (and hence quasi-convex) in $v$.
\end{enumerate}
Thus, invoking Theorem \ref{thm:minimax}, we obtain the desired equality.
\subsection{Proof of Theorem \ref{thm:opt_ind_est}}\label{prf:thm:opt_ind_est}
This theorem is a consequence of the following two propositions.
\begin{proposition}
\label{prop:opt_ind_given_glob}
For any $\overline{\delta}_n = o(1)$, there exists another $\overline{\delta}_n'=o(1)$ such that for any $\ell\in[L]$, we have
\begin{align*}
& \sum_{\substack{S\subseteq L\setminus\{\ell\} }} e^{-(1-\overline{\delta}_n)\mathcal{J}_{S\cup\{\ell\}}}\leq 2 \exp\big\{-(1-\overline{\delta}_n)\mathcal{J}_{\{\ell\}} + \log L + Le^{-(1-\overline{\delta}_n)J_\rho}\big\}.
\end{align*}
\end{proposition}
\begin{proof}
See Appendix \ref{prf:prop:opt_ind_given_glob}.
\end{proof}
\begin{proposition}
\label{prop:opt_ind_est_glob}
For any $\overline{\delta}_n = o(1)$ and any $c\in(0, 1)$, there exists another $\overline{\delta}_n'=o(1)$ such that for any $\ell\in[L]$, we have
\begin{align*}
&\sum_{\substack{S\subseteq [L]\setminus\{\ell\}\\ |(S\cup\{\ell\})^c|\textnormal{ even}}} \exp\big\{-(1-\overline{\delta}_n) \big(|(S\cup\{\ell\})^c| J_\rho + \psi_{S\cup\{\ell\}}^\star(0)\big)\big\}\\
& \leq 2\exp\big\{ -(1-\overline{\delta}_n') \min_{\substack{S\subseteq[L]\setminus\{\ell\} \\ |(S\cup\{\ell\})^c| \textnormal{ even}}} \big( |(S\cup\{\ell\})^c| J_\rho + \psi_{S\cup\{\ell\}}^\star(0) \big) + \log L + Le^{-(1-\overline{\delta}_n)c J_\rho}\big\},
\end{align*}
and
\begin{align*}
& \sum_{\substack{S\subseteq[L]\setminus\{\ell\} \\ |(S\cup\{\ell\})^c| \textnormal{ odd}}} \exp\big\{- (1-\overline{\delta}_n) \big(|(S\cup\{\ell\})^c|J_\rho + \psi_{S\cup\{\ell\}}^\star(-2J_\rho)\big)\big\} \\
& \leq 2 \exp\big\{-(1-\overline{\delta}_n') \min_{\substack{S\subseteq[L]\setminus\{\ell\}\\ |(S\cup\{\ell\})^c| \textnormal{ odd}}}\big(|(S\cup\{\ell\})^c|J_\rho + \psi_{S\cup\{\ell\}}^\star(-2J_\rho)\big) + \log L + Le^{-(1-\overline{\delta}_n)cJ_\rho}\big\}.
\end{align*}
\end{proposition}
\begin{proof}
The proof is nearly identical to that of Propositions \ref{prop:opt_glob_even} and \ref{prop:opt_glob_odd}, so we omit the details.
\end{proof}
\subsubsection{Proof of Proposition \ref{prop:opt_ind_given_glob}}\label{prf:prop:opt_ind_given_glob}
Without loss of generality we consider the first layer, i.e., the $\ell$ in the statement of this proposition is $1$.
~\\
{\bf The even terms.} We first consider the terms whose $|S\cup\{1\}|$'s are even.
Note that
\begingroup
\allowdisplaybreaks
\begin{align*}
& \sum_{\substack{S\subseteq\{2, \hdots, L\}\\ |S\cup\{1\}|\textnormal{ even}}} \exp\bigg\{ -(1-\overline{\delta}_n)\bigg(|S\cup\{1\}|J_\rho + \sum_{\ell \in S\cup\{1\}}m\III{\ell}_{1/2}\bigg) \bigg\} \\
& \qquad + \sum_{\substack{S\subseteq\{2, \hdots, L\}\\ |S\cup\{1\}|\textnormal{ odd}}} \exp\bigg\{ -(1-\overline{\delta}_n)\bigg(|S\cup\{1\}|J_\rho + \sum_{\ell \in S\cup\{1\}}m\III{\ell}_{1/2}\bigg) \bigg\} \\
& = e^{-(1-\overline{\delta}_n)(J_\rho + m\III{1}_{1/2})} \times \prod_{\ell \neq 1} \bigg(1 + e^{-(1-\overline{\delta}_n)(J_\rho + m\III{\ell}_{1/2})}\bigg).
\end{align*}
\endgroup
On the other hand, we have
\begingroup
\allowdisplaybreaks
\begin{align*}
& \sum_{\substack{S\subseteq\{2, \hdots, L\}\\ |S\cup\{1\}|\textnormal{ even}}} \exp\bigg\{ -(1-\overline{\delta}_n)\bigg(|S\cup\{1\}|J_\rho + \sum_{\ell \in S\cup\{1\}}m\III{\ell}_{1/2}\bigg) \bigg\} \\
& \qquad - \sum_{\substack{S\subseteq\{2, \hdots, L\}\\ |S\cup\{1\}|\textnormal{ odd}}} \exp\bigg\{ -(1-\overline{\delta}_n)\bigg(|S\cup\{1\}|J_\rho + \sum_{\ell \in S\cup\{1\}}m\III{\ell}_{1/2}\bigg) \bigg\} \\
& = -e^{-(1-\overline{\delta}_n)(J_\rho + m\III{1}_{1/2})} \prod_{\ell\neq 1} \bigg(1 - e^{-(1-\overline{\delta}_n)(J_\rho + m\III{\ell}_{1/2})}\bigg).
\end{align*}
\endgroup
Hence, we have
\begin{align*}
& \sum_{\substack{S\subseteq\{2, \hdots, L\}\\ |S\cup\{1\}|\textnormal{ even}}} \exp\bigg\{ -(1-\overline{\delta}_n)\bigg(|S\cup\{1\}|J_\rho + \sum_{\ell \in S\cup\{1\}}m\III{\ell}_{1/2}\bigg) \bigg\} \\
& = \frac{1}{2} e^{-(1-\overline{\delta}_n)(J_\rho + m\III{1}_{1/2})} \times \bigg[ \prod_{\ell\neq 1} \bigg(1 + e^{-(1-\overline{\delta}_n)(J_\rho + m\III{\ell}_{1/2})}\bigg) - \prod_{\ell\neq 1} \bigg(1 - e^{-(1-\overline{\delta}_n)(J_\rho + m\III{\ell}_{1/2})}\bigg)\bigg]\\
& \leq e^{-(1-\overline{\delta}_n)(J_\rho + m\III{1}_{1/2})} \cdot \bigg(\sum_{i\neq 1} e^{-(1-\overline{\delta}_n)(J_\rho + m\III{\ell}_{1/2})}\bigg) \cdot \prod_{\ell\neq 1} \bigg(1+ e^{-(1-\overline{\delta}_n)(J_\rho + m\III{\ell}_{1/2})}\bigg)\\
& \leq e^{-(1-\overline{\delta}_n)(2J_\rho + m\III{1}_{1/2})}\cdot L \cdot \bigg(1+e^{-(1-\overline{\delta}_n)J_\rho}\bigg)^{L}\\
\label{eq:opt_ind_given_glob_even}
& \leq e^{-(1-\overline{\delta}_n)\mathcal{J}_{\{1\}}} \cdot L \exp\big\{ L e^{-(1-\overline{\delta}_n)J_\rho}\big\},\addtocounter{equation}{1}\tag{\theequation}
\end{align*}
where the third line is by Lemma \ref{lemma: diff_of_products} and the last inequality is by Lemma \ref{lemma:ineq_for_cvx_conjugate}.
~\\
{\bf\noindent The odd terms.} We now consider the terms whose $|S\cup\{1\}|$'s are odd. For any fixed $t\in[0, 1]$, we have
\begin{align*}
&\sum_{\substack{S\subseteq\{2, \hdots, L\} \\ |S\cup\{1\}| \textnormal{ odd}}}\exp\bigg\{- (1-\overline{\delta}_n) \bigg(\bigl(|(S\cup\{1\})| + 1\bigr) J_\rho + \psi_{S\cup\{1\}}^\star(-2J_\rho)\bigg) \bigg\} \\
& \leq \sum_{\substack{S\subseteq\{2, \hdots, L\} \\ |S\cup\{1\}| \textnormal{ odd}}}\exp\bigg\{- (1-\overline{\delta}_n) \bigg(\bigl(|(S\cup\{1\})| + 1\bigr) J_\rho + (1-2t)J_\rho + \sum_{\ell \in S\cup\{1\}}m\III{\ell}_{t}\bigg) \bigg\} \\
& = \exp\bigg\{-(1-\overline{\delta}_n)\bigg(J_\rho + (1-2t)J_\rho + m\III{1}_{t}\bigg)\bigg\} \times \sum_{\substack{S\subseteq\{2, \hdots, L\} \\ |S\cup\{1\}| \textnormal{ odd}}} \exp\bigg\{-(1-\overline{\delta}_n)\bigg(|S|J_\rho + \sum_{\ell\in S} m\III{\ell}_{t}\bigg)\bigg\}.
\end{align*}
Note that
\begin{align*}
& \sum_{\substack{S\subseteq\{2, \hdots, L\} \\ |S\cup\{1\}| \textnormal{ odd}}} \exp\bigg\{-(1-\overline{\delta}_n)\bigg(|S|J_\rho + \sum_{\ell\in S} m\III{\ell}_{t}\bigg)\bigg\} + \sum_{\substack{S\subseteq\{2, \hdots, L\} \\ |S\cup\{1\}| \textnormal{ even}}} \exp\bigg\{-(1-\overline{\delta}_n)\bigg(|S|J_\rho + \sum_{\ell\in S} m\III{\ell}_{t}\bigg)\bigg\} \\
& = \prod_{\ell\neq 1} \bigg(1+ e^{-(1-\overline{\delta}_n)(J_\rho + m\III{\ell}_{t})}\bigg).
\end{align*}
Meanwhile, we have
\begin{align*}
& \sum_{\substack{S\subseteq\{2, \hdots, L\} \\ |S\cup\{1\}| \textnormal{ odd}}} \exp\bigg\{-(1-\overline{\delta}_n)\bigg(|S|J_\rho + \sum_{\ell\in S} m\III{\ell}_{t}\bigg)\bigg\} - \sum_{\substack{S\subseteq\{2, \hdots, L\} \\ |S\cup\{1\}| \textnormal{ even}}} \exp\bigg\{-(1-\overline{\delta}_n)\bigg(|S|J_\rho + \sum_{\ell\in S} m\III{\ell}_{t}\bigg)\bigg\} \\
& = \prod_{\ell\neq 1} \bigg(1- e^{-(1-\overline{\delta}_n)(J_\rho + m\III{\ell}_{t})}\bigg).
\end{align*}
Hence, the odd terms are bounded above by
\begin{align*}
\exp\bigg\{-(1-\overline{\delta}_n)\bigg(J_\rho + (1-2t)J_\rho + m\III{1}_t\bigg)\bigg\} \times \frac{1}{2} \bigg[ \prod_{\ell\neq 1} \bigg(1+e^{-(1-\overline{\delta}_n)(J_\rho + m\III{\ell}_{t})}\bigg) + \prod_{\ell\neq 1} \bigg(1-e^{-(1-\overline{\delta}_n)(J_\rho + m\III{\ell}_{t})} \bigg)\bigg].
\end{align*}
Choosing $t\in[0, 1]$ such that it maximizes $-2tJ_\rho + m\III{1}_t$, and using
\begin{align*}
& \frac{1}{2} \bigg[ \prod_{\ell\neq 1} \bigg(1+e^{-(1-\overline{\delta}_n)(J_\rho + m\III{\ell}_{t})}\bigg) + \prod_{\ell\neq 1} \bigg(1-e^{-(1-\overline{\delta}_n)(J_\rho + m\III{\ell}_{t})} \bigg)\bigg] \leq \bigg(1 + e^{-(1-\overline{\delta}_n) J_\rho}\bigg)^{L-1},
\end{align*}
we get
\begin{align*}
&\sum_{\substack{S\subseteq\{2, \hdots, L\} \\ |S\cup\{1\}| \textnormal{ odd}}}\exp\bigg\{- (1-\overline{\delta}_n) \bigg(\bigl(|(S\cup\{1\})| + 1\bigr) J_\rho + \psi_{S\cup\{1\}}^\star(-2J_\rho)\bigg) \bigg\} \\
& \leq \exp\bigg\{-(1-\overline{\delta}_n)\bigg(2J_\rho + \psi_{\{1\}}^\star(-2J_\rho)\bigg) + L e^{-(1-\overline{\delta}_n)J_\rho}\bigg\} \\
\label{eq:opt_ind_given_glob_odd}
& =\exp\bigg\{-(1-\overline{\delta}_n)\mathcal{J}_{\{1\}} + L e^{-(1-\overline{\delta}_n)J_\rho}\bigg\}\addtocounter{equation}{1}\tag{\theequation}.
\end{align*}
The proof is concluded by combining \eqref{eq:opt_ind_given_glob_even} and \eqref{eq:opt_ind_given_glob_odd}.
\subsection{Proof of Theorem \ref{thm:refine_glob}} \label{prf:thm:refine_glob}
The high-level idea of this proof is that we can bound node-wise errors separately due to the additive form of the loss function. While such an idea has appeared in \cite{gao2017achieving,gao2018community}, the implementations of this idea is considerably more complicated in our case due to the combinatorial structure induced by the presence of inhomogeneity across layers.
Note that in Stage \RN{2} of Algorithm \ref{alg: provable_refinement}, we modify the $i$-th coordinate of $\tzzz{\star, -i}$ and $\tzzz{\ell, -i}$. To avoid confusions, we let $\tzzz{\star, -i} = \tzzz{\ell, -i}$ be the initial estimators computed in Stage \RN{1} (but before Stage \RN{2}), whose $i$-th coordinates are zero by construction, and we let $\barzzz{\star, -i}, \barzzz{\ell, -i}$ be the estimators computed in Stage \RN{2}, whose $i$-th coordinates satisfy
\begin{equation}
\label{eq:provable_refinement_obj_func}
(\barzzz{\star, -i}_i, \barzzz{1, -i}_i, \hdots, \barzzz{L, -i}_i) = \mathop{\rm argmax}_{\substack{s_\star \in\{\pm 1\} \\ s_\ell \in \{\pm 1\}~\forall \ell\in[L] }} \sum_{\ell\in[L]} \fff{\ell}_i(s_\star, s_\ell, \tzzz{\star, -i}),
\end{equation}
and whose rest of the coordinates agree with $\tzzz{\star, -i}_{-i}$.
We start by presenting two preliminary results.
\begin{lemma}
\label{lemma:align}
Fix any $\mathbf{z}, \mathbf{z}'\in\{\pm 1\}^n$ and assume there exists a constant $C\geq 1$ such that
$$
\min_{s\in\{\pm 1\}} \# \{i\in[n]: \mathbf{z}_i = s\} \geq \frac{n}{2C}, ~~~ \min_{s\in\{\pm 1\}} \# \{i\in[n]: \mathbf{z}_i' = s\} \geq \frac{n}{2C}, ~~~ \min_{\pi\in\{\pm 1\}} d_\textnormal{H}(\mathbf{z},\pi \mathbf{z}') \leq \frac{1}{2C}.
$$
Define $\vartheta: \{\pm 1\} \to \{\pm 1\}$ by
\begin{equation}
\label{eq:align_map}
\vartheta(r) = \mathop{\rm argmax}_{s\in\{\pm 1\}} \#\bigg\{\{j\in[n]: \mathbf{z}_j = s\} \bigcap \{j\in[n]: \mathbf{z}_j' = r\}\bigg\}.
\end{equation}
Then $\vartheta$ is a bijection and hence can be identified by $\vartheta\in\{\pm 1\}$ with $+ 1$ being the identity map. Moreover, we have
$$
d_\textnormal{H}(\mathbf{z}, \vartheta \mathbf{z}') = \min_{\pi\in\{\pm 1\}} d_\textnormal{H}(\mathbf{z}, \pi \mathbf{z}').
$$
\end{lemma}
\begin{proof}
This is Lemma 4 in \cite{gao2017achieving}.
\end{proof}
\begin{proposition}
\label{prop:refine_glob_nodewise}
Assume $\rho = o(1)$, $p_\ell\asymp q_\ell = o(1) , \forall \ell\in[L]$, $\beta= 1+o(1)$, and $L\lesssim n^c$ for some $c > 0$. Assume there exists a sequence $\delta_n=o(1)$ and constants $\varepsilon_{{\texttt{init}}}>0, C>0$ such that $\forall i\in[n], \exists \pi_i\in\{\pm 1\}$ which makes the following holds:
\begin{equation}
\label{eq:ind_consistent_init}
{\mathbb{P}}\bigg(d_\textnormal{H}(\pi_i \tzzz{\star, -i}_{-i} ,\zzz{\ell}_{-i})\leq n\delta_n ~ \forall \ell\in[n]\bigg) \geq 1 - C n^{-(1+\varepsilon_{{\texttt{init}}})},
\end{equation}
Then there exists another sequence $\delta_n'=o(1)$ and an absolute constant $C'>0$ such that for any $i\in[n]$, we have
\begin{align}
\label{eq:upper_bound_z_star}
{\mathbb{P}}(\pi_i\barzzz{\star, -i}_i \neq \mathbf{z}^\star_i)
& \leq C'n^{-(1+\varepsilon_{{\texttt{init}}})}+ \sum_{S\subseteq[L]}e^{-(1-\delta_n') \mathcal{I}_S}
\end{align}
\end{proposition}
\begin{proof}
See Appendix \ref{prf:prop:refine_glob_nodewise}.
\end{proof}
By Assumption \ref{assump:consistent_init}, for any $i\in[n]$, there exists $\pi_i\in\{\pm 1\}$ such that
\begin{equation}
\label{eq:consist_init_leave_i_out}
{\mathbb{P}}\bigg(d_\textnormal{H}(\pi_i \tzzz{\star, -i}_{-i}, \mathbf{z}^\star_{-i}) \leq n\eta_{{\texttt{init}}, n-1}\bigg) \geq 1- Cn^{-(1+\varepsilon_{{\texttt{init}}})}.
\end{equation}
Since $d_\textnormal{H}(\mathbf{z}^\star, \zzz{\ell})$ is the i.i.d. sum of $n$ Bernoulli random variables, an application of Chernoff bound gives $d_\textnormal{H}(\mathbf{z}^\star, \zzz{\ell})\geq n\rho + nt$ with probability at most $2e^{-c_1nt^2}$ for any $t > 0$ and some constant $c_1 > 0$. Choosing $t = n^{-1/4}$ and invoking a union bound over all layers, we conclude that
\begin{equation}
\label{eq:vanishing_flips}
{\mathbb{P}}\bigg(d_\textnormal{H}(\mathbf{z}^\star, \zzz{\ell}) \leq n\cdot (\rho + n^{-1/4}) ~ \forall \ell\in[L]\bigg) \geq 1 - C_1n^{-(1+\varepsilon_{{\texttt{init}}})}
\end{equation}
for another constant $C'>0$. In particular, on the union of the two high probability events in \eqref{eq:consist_init_leave_i_out} and \eqref{eq:vanishing_flips}, for any fixed $i\in[n]$ and uniformly over $\ell\in[L]$, we have
$$
d_\textnormal{H}(\pi_i\tzzz{\star, -i}_{-i}, \zzz{\ell}_{-i}) \leq d_\textnormal{H}(\pi_i \tzzz{\star, -i}_{-i}, \mathbf{z}^\star_{-i}) + d_\textnormal{H}(\mathbf{z}^\star_{-i}, \zzz{\ell}_{-i}) \leq n \cdot(\eta_{{\texttt{init}}, n-1} + \rho + n^{-1/4} + n^{-1}).
$$
Hence, we can invoke Proposition \ref{prop:refine_glob_nodewise} to conclude that for any $i\in[n]$, \eqref{eq:upper_bound_z_star} holds.
In the rest of the proof, we assume $\pi_1 = +1$ without loss of generality.
Now, for each $i\in[n]\setminus\{1\}$, we define the map $\vartheta^\star_i:\{\pm 1\}\to \{\pm 1\}$ as in \eqref{eq:align_map} with $\mathbf{z} = \barzzz{\star, -1}$ and $\mathbf{z}' = \barzzz{\star, -i}$. By construction we have
$$
\mathbf{\hat z}^\star_i = \vartheta^\star_i (\barzzz{\star, -i}_i).
$$
Thus we have
\begin{align*}
{\mathbb{P}}(\mathbf{\hat z}^\star_i\neq \mathbf{z}^\star_i) & = {\mathbb{P}}(\vartheta^\star_i(\barzzz{\star, -i}_i) \neq \mathbf{z}^\star_i)\\
\label{eq:nodewise_error_alignment}
& \leq {\mathbb{P}}(\pi_i \barzzz{\star, -i}_i \neq \mathbf{z}^\star_i) + {\mathbb{P}}(\vartheta^\star_i(\barzzz{\star, -i})\neq \pi_i \barzzz{\star, -i}_i).\addtocounter{equation}{1}\tag{\theequation}
\end{align*}
On the following event:
\begin{equation}
\label{eq:Ei_cap_E1}
\bigg\{ d_\textnormal{H}(\pi_i \barzzz{\star, -i}_{-i}, \mathbf{z}^\star_{-i})\leq n\eta_{{\texttt{init}}, n-1} \bigg\} \bigcap \bigg\{ d_\textnormal{H}(\barzzz{\star, -1}_{-1}, \mathbf{z}^\star_{-1})\leq n\eta_{{\texttt{init}}, n-1}\bigg\},
\end{equation}
we have
\begin{equation}
\label{eq:two_layer_alignment}
d_\textnormal{H}(\barzzz{\star, -1}, \barzzz{\star, -i})\leq n\cdot (2\eta_{{\texttt{init}}, n-1} + 2n^{-1}).
\end{equation}
Invoking Lemma \ref{lemma:align} gives that on the above event, $\vartheta^\star_i$ is a bijection, and
$$
\vartheta^\star_i = \mathop{\rm argmin}_{\pi\in\{\pm 1\}} d_\textnormal{H}(\barzzz{\star, -1}, \pi \barzzz{\star ,-i}) = \pi_i
$$
where we have regarded $\vartheta^\star_i$ as a $\{\pm 1\}$-valued scaler, with $+ 1$ representing the identity map, and the last equality follows from \eqref{eq:two_layer_alignment}. In particular, we know that $\vartheta^\star_i(\barzzz{\star, -i}) = \pi_i \barzzz{\star, -i}_i$ on the event \eqref{eq:Ei_cap_E1}. Since this event happens with probability at least $1-2Cn^{1+\varepsilon_{{\texttt{init}}}}$, from \eqref{eq:nodewise_error_alignment} we get
\begin{align*}
{\mathbb{P}}(\mathbf{\hat z}^\star_i \neq \mathbf{z}^\star_i) & \leq C_2 n^{-(1+\varepsilon_{{\texttt{init}}})}+ \sum_{S\subseteq[L]} e^{-(1-\overline{\delta}_n) \mathcal{I}_S},
\end{align*}
where $C_2>0$ is an absolute constant and $\overline{\delta}_n=o(1)$. Let us set
$$
\overline{\delta}_n' = \bigg[\log \bigg(\frac{1}{\sum_{S\subseteq [L]} \exp\{-(1-\overline{\delta}_n\mathcal{I}_S)\}}\bigg)\bigg]^{-1/2},
$$
which tends to zero as $n$ tends to infinity by \eqref{eq:diverging_glob_snr_sum}. By Markov's inequality, we have
\begin{align*}
& {\mathbb{P}}\bigg({\mathcal{L}}(\mathbf{\hat z}^\star, \mathbf{z}^\star) \geq \big(\sum_{S\subseteq[L]} e^{-(1-\overline{\delta}_n) \mathcal{I}_S}\big)^{1-\overline{\delta}_n'}\bigg)\\
& \leq \big(\sum_{S\subseteq[L]} e^{-(1-\overline{\delta}_n) \mathcal{I}_S}\big)^{\overline{\delta}_n'-1} \cdot \frac{1}{n}\sum_{i\in[n]} {\mathbb{P}}(\mathbf{\hat z}^\star_i \neq \mathbf{z}^\star_i) \\
& \leq \big(\sum_{S\subseteq[L]} e^{-(1-\overline{\delta}_n) \mathcal{I}_S}\big)^{\overline{\delta}_n'} + C_2 n^{-(1+\varepsilon_{{\texttt{init}}})} \big(\sum_{S\subseteq[L]} e^{-(1-\overline{\delta}_n) \mathcal{I}_S}\big)^{\overline{\delta}_n'-1}\\
& = \exp\bigg\{- \sqrt{\log \bigg(\frac{1}{\sum_{S\subseteq[L]} e^{-(1-\overline{\delta}_n) \mathcal{I}_S} }\bigg)}\bigg\}
+ C_2 {n^{-(1+\varepsilon_{{\texttt{init}}})}} \cdot \exp\bigg\{ (1-\overline{\delta}_n')\log\bigg(\frac{1}{\sum_{S\subseteq[L]} e^{-(1-\overline{\delta}_n)\mathcal{I}_S}}\bigg) \bigg\}
\end{align*}
If
$$
\exp\bigg\{ (1-\overline{\delta}_n')\log\bigg(\frac{1}{\sum_{S\subseteq[L]} e^{-(1-\overline{\delta}_n)\mathcal{I}_S}}\bigg) \bigg\} \leq n^{1+\frac{\varepsilon_{{\texttt{init}}}}{2}},
$$
then we get
$$
{\mathbb{P}}\bigg({\mathcal{L}}(\mathbf{\hat z}^\star, \mathbf{z}^\star) \geq \big(\sum_{S\subseteq[L]} e^{-(1-\overline{\delta}_n) \mathcal{I}_S}\big)^{1-\overline{\delta}_n'}\bigg) \leq \exp\bigg\{- \sqrt{\log \bigg(\frac{1}{\sum_{S\subseteq[L]} e^{-(1-\overline{\delta}_n) \mathcal{I}_S} }\bigg)}\bigg\}
+ C_2 {n^{-\varepsilon_{{\texttt{init}}}/2}} \to 0
$$
as $n\to \infty$. Otherwise, we can proceed by
\begin{align*}
{\mathbb{P}}\bigg({\mathcal{L}}(\mathbf{\hat z}^\star, \mathbf{z}^\star) \geq \big(\sum_{S\subseteq[L]} e^{-(1-\overline{\delta}_n) \mathcal{I}_S}\big)^{1-\overline{\delta}_n'}\bigg)
& \leq {\mathbb{P}}\bigg({\mathcal{L}}(\mathbf{\hat z}^\star, \mathbf{z}^\star) > 0\bigg) \\
& \leq \sum_{i\in[n]} {\mathbb{P}}(\mathbf{\hat z}^\star_i \neq \mathbf{z}^\star_i)\\
& \leq C_2n^{-\varepsilon_{{\texttt{init}}}} + n \exp\bigg\{-\log\bigg( \frac{1}{\sum_{S\subseteq[L]} e^{-(1-\overline{\delta}_n) \mathcal{I}_S}}\bigg)\bigg\}\\
& \leq C_2n^{-\varepsilon_{{\texttt{init}}}} + n \exp\bigg\{-(1-\overline{\delta}_n')\log\bigg( \frac{1}{\sum_{S\subseteq[L]} e^{-(1-\overline{\delta}_n) \mathcal{I}_S}}\bigg)\bigg\}\\
& \leq C_2 n^{-\varepsilon_{{\texttt{init}}}} + n^{-\varepsilon_{{\texttt{init}}}/2}\\
& \to 0.
\end{align*}
Thus, in either case, \eqref{eq:refine_glob} holds, and the proof is concluded.
\subsubsection{Proof of Proposition \ref{prop:refine_glob_nodewise}}\label{prf:prop:refine_glob_nodewise}
Fix $i\in[n]$ and we without loss of generality assume $\pi_i = 1$.
Let
\begin{align}
\label{eq:Ei}
E_i & := \bigg\{d_\textnormal{H}(\tzzz{\star, -i}_{-i}, \zzz{\ell}_{-i})\leq n\delta_n ~ \forall\ell\in[L]\bigg\},
\end{align}
which happens with probability at least $1- Cn^{-(1+\varepsilon_{{\texttt{init}}})}$ by assumption. For $\nnn{\ell, -i}_{\pm}:= \sum_{j\neq i}\mathbf{1}{\zzz{\ell}_j = \pm 1}$, we define
\begin{equation}
\label{eq:Fi}
F_i := \bigg\{\bigg|(\nnn{\ell, -i}_+ - \nnn{\ell, -i}_-) - (n^{\star}_+ - n^{\star}_-)\bigg| \leq n\cdot (n^{-1/4} + n^{-1} + 2\rho) ~ \forall \ell\in[L]\bigg\}.
\end{equation}
We claim that the event $F_i$ also happens with high probability.
\begin{lemma}
\label{lemma:high_prob_Fi}
If $L\lesssim n^{c}$, then there exists an absolute constant $c'>0$ such that
$$
{\mathbb{P}}\big(\bigcap_{i\in[n]} F_i\big) \geq 1 - e^{-c'\sqrt{n}}.
$$
\end{lemma}
\begin{proof}
Recall that
$$
\nnn{\ell}_+ - \nnn{\ell}_- = \sum_{i\in[n]} \mathbf{1}{\zzz{\ell}_i = +1} - \mathbf{1}{\zzz{\ell}_i = -1} = \sum_{i: \mathbf{z}^\star_i = +1} \textnormal{Rad}(1-\rho) + \sum_{i: \mathbf{z}^\star_i = -1} \textnormal{Rad}(\rho),
$$
where $\textnormal{Rad}(p)$ is a Rademacher random variable with positive probability $p$. Since the above display has mean $(1-2\rho)(n^{\star}_+ - n^{\star}_-)$, by Hoeffding's inequality, for any $t>0$ we have
$$
{\mathbb{P}}\bigg((\nnn{\ell}_+ - \nnn{\ell}_-) - (1-2\rho)(n^{\star}_+ - n^{\star}_-) \geq nt\bigg) \leq 2 e^{-c_1nt^2}
$$
where $c_1>0$ is some absolute constant. Setting $t = n^{-1/4}$ and using a union bound over all $\ell\in[L]$, we get
$$
{\mathbb{P}}\bigg((\nnn{\ell}_+ - \nnn{\ell}_-) - (1-2\rho)(n^{\star}_+ - n^{\star}_-) \leq n\cdot n^{-1/4} ~ \forall \ell\in[L]\bigg) \geq 1-2Le^{-c_1\sqrt{n}} \geq 1-e^{-c_2 \sqrt{n}}
$$
for another constant $c_2$, where the last inequality is by $L\lesssim n^{c}$. Note that by construction, we have
$$
\bigg|(\nnn{\ell, -i}_+ - \nnn{\ell, -i}_-) - (\nnn{\ell}_+ - \nnn{\ell}_-)\bigg| = 1
$$
for any $i\in[n]$. An application of the triangle inequality gives
$$
\bigg|(\nnn{\ell, -i}_+, \nnn{\ell, -i}_-) - (n^{\star}_+ - n^{\star}_-)\bigg| \leq 2\rho|n^{\star}_+ - n^{\star}_-| + 1 + n\cdot n^{-1/4} \leq n\cdot (n^{-1/4} + n^{-1} + 2\rho)
$$
with probability at least $1-e^{-c_2\sqrt{n}}$, and this is the desired result.
\end{proof}
The above lemma, along our assumption \eqref{eq:ind_consistent_init}, gives
\begin{equation}
\label{eq:high_prob_Ei_Fi}
{\mathbb{P}}(E_i\cap F_i) \geq 1 - Cn^{-(1+\varepsilon_{{\texttt{init}}})} - e^{-c'\sqrt{n}} \geq 1- C' n^{-(1+\varepsilon_{{\texttt{init}}})}
\end{equation}
for some constant $C'>0$. Thus, we have
\begin{equation}
\label{eq:glob_nodewise_err_decomp}
{\mathbb{P}}(\barzzz{\star, -i}_i\neq \mathbf{z}^\star_i) = {\mathbb{P}}(\barzzz{\star, -i}_i = -\mathbf{z}^\star_i) \leq {\mathbb{P}}(\barzzz{\star, -i}_i = -\mathbf{z}^\star_i, E_i\cap F_i) + C'n^{-(1+\varepsilon_{{\texttt{init}}})}.
\end{equation}
We can decompose the probability via
\begin{align}
\label{eq:glob_nodewise_err_decomp_all_subsets}
{\mathbb{P}}(\barzzz{\star, -i}_i = -\mathbf{z}^\star_i, E_i\cap F_i) & = \sum_{S \subseteq [L]} {\mathbb{P}}(\barzzz{\star, -i}_i = -\mathbf{z}^\star_i , \hzzz{S}_i = -\zzz{S}_i, \hzzz{S^c}_i = \zzz{S^c}_i, E_i\cap F_i),
\end{align}
where we used the shorthand notation $\zzz{S}_i:= \{\zzz{\ell}_i: \ell \in S\}$ for any $S \subseteq[L]$. Occurrence of the event in the right-hand side above implies that
$$
\sum_{\ell\in S} \fff{\ell}_i(-\mathbf{z}^\star_i, -\zzz{\ell}_i, \tzzz{\star, -i}) + \sum_{\ell \in S^c} \fff{\ell}_i(-\mathbf{z}^\star_i, \zzz{\ell}_i , \tzzz{\star, -i}) \geq \sum_{\ell\in [L]} \fff{\ell}_i(\mathbf{z}^\star_i, \zzz{\ell}_i, \tzzz{\star, -i}).
$$
Hence, we have
\begin{align*}
& {\mathbb{P}}(\barzzz{\star, -i}_i = -\mathbf{z}^\star_i, E_i\cap F_i) \\
& \leq \sum_{S \subseteq [L]}
{\mathbb{P}}\bigg\{ \log\bigg(\frac{1-\rho}{\rho}\bigg) \cdot \bigg(\#\{\ell\in S^c: \mathbf{z}^{\star}_i = -\zzz{\ell}_i\} - \#\{\ell\in S^c: \mathbf{z}^{\star}_i = \zzz{\ell}_i\}\bigg) \\
& \qquad + \sum_{\ell\in S}\sum_{\substack{j\neq i: \tzzz{\star, -i}_j = -\zzz{\ell}_i}} \bigg[\log\bigg(\frac{p_\ell (1-q_\ell)}{q_\ell(1-p_\ell)}\bigg) \AAA{\ell}_{ij} + \log\bigg(\frac{1-p_\ell}{1-q_\ell}\bigg) \bigg] \\
& \qquad - \sum_{\ell\in S} \sum_{j\neq i : \tzzz{\star, -i}_j = \zzz{\ell}_i} \bigg[\log\bigg(\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg) \AAA{\ell}_{ij} + \log\bigg(\frac{1-p_\ell}{1-q_\ell}\bigg)\bigg] \geq 0 ~ \textnormal{and } E_i\cap F_i \bigg\}.
\end{align*}
Note that $E_i$ and $F_i$ are both independent of $\{\zzz{\ell}_i\}_{\ell=1}^L$. So we can decompose the above probability by conditioning on the value of $\#\{\ell\in S^c: \mathbf{z}^\star_i = -\zzz{\ell}_i\} - \#\{\ell\in S^c: \mathbf{z}^\star_i = \zzz{\ell}_i\}$:
\begin{align*}
& {\mathbb{P}}(\barzzz{\star, -i}_i = -\mathbf{z}^\star_i, E_i\cap F_i) \\
& \leq \sum_{S \subseteq [L]} \sum_{x \in\{ -|S^c|+2k: 0\leq k \leq |S^c|\}} \binom{|S^c|}{\frac{|S^c| + x}{2}} \bigg(\frac{1-\rho}{\rho}\bigg)^{{x}/{2}}\bigg(\rho(1-\rho)\bigg)^{|S^c|/2}\\
& \qquad \times {\mathbb{P}}\bigg\{ -x\log\bigg(\frac{1-\rho}{\rho}\bigg)
+ \sum_{\ell\in S}\sum_{\substack{j\neq i: \tzzz{\star, -i}_j = -\zzz{\ell}_i}} \bigg[\log\bigg(\frac{p_\ell (1-q_\ell)}{q_\ell(1-p_\ell)}\bigg) \AAA{\ell}_{ij} + \log\bigg(\frac{1-p_\ell}{1-q_\ell}\bigg)\bigg] \\
& \qquad \qquad - \sum_{\ell\in S} \sum_{j\neq i : \tzzz{\star, -i}_j = \zzz{\ell}_i} \bigg[\log\bigg(\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg) \AAA{\ell}_{ij} + \log\bigg(\frac{1-p_\ell}{1-q_\ell}\bigg)\bigg] \geq 0 ~ \textnormal{and } E_i\cap F_i \bigg\}.
\end{align*}
We further decompose the above probability according to the orientations of $\zzz{\ell}_i$'s for $\ell\in S$:
\begin{align*}
& {\mathbb{P}}(\barzzz{\star, -i}_i = -\mathbf{z}^\star_i, E_i\cap F_i) \\
& \leq \sum_{S \subseteq [L]} \sum_{x \in\{ -|S^c|+2k: 0\leq k \leq |S^c|\}} \binom{|S^c|}{\frac{|S^c| + x}{2}} \bigg(\frac{1-\rho}{\rho}\bigg)^{{x}/{2}}\bigg(\rho(1-\rho)\bigg)^{|S^c|/2} \sum_{\xi\in\{\pm 1\}^S}{\mathbb{P}}(\zzz{S}_i = \mathbf{z}^\star_i \xi )\\
& \qquad \times {\mathbb{P}}\bigg\{ -x\log\bigg(\frac{1-\rho}{\rho}\bigg)
+ \sum_{\ell\in S}\sum_{\substack{j\neq i: \tzzz{\star, -i}_j = -\xi_\ell \mathbf{z}^\star_i}} \bigg[\log\bigg(\frac{p_\ell (1-q_\ell)}{q_\ell(1-p_\ell)}\bigg) \AAA{\ell}_{ij} + \log\bigg(\frac{1-p_\ell}{1-q_\ell}\bigg) \bigg]\\
& \qquad \qquad - \sum_{\ell\in S} \sum_{j\neq i : \tzzz{\star, -i}_j = \xi_\ell\mathbf{z}^\star_i} \bigg[\log\bigg(\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg) \AAA{\ell}_{ij} + \log\bigg(\frac{1-p_\ell}{1-q_\ell}\bigg)\bigg]
\geq 0 ~ \textnormal{and } E_i\cap F_i \bigg\} \\
& = \sum_{S \subseteq [L]} \sum_{x \in\{ -|S^c|+2k: 0\leq k \leq |S^c|\}} \binom{|S^c|}{\frac{|S^c| + x}{2}} \exp\bigg\{ -|S^c|\log\frac{1}{\sqrt{\rho(1-\rho)}} + x\log \sqrt{\frac{1-\rho}{\rho}}\bigg\} \\
& \qquad \times \sum_{\xi\in\{\pm 1\}^S}{\mathbb{P}}(\zzz{S}_i = \mathbf{z}^\star_i \xi ) \cdot {\mathbb{E}}_{\{\zzz{\ell}_{-i}\}_{\ell=1}^L}\bigg[{\mathbb{P}}\bigg(\mathscr{E}_{S, x, \xi} ~\bigg|~ \{\zzz{\ell}_{-i}\}_{\ell=1}^L \bigg)\bigg],
\end{align*}
where
\begin{align*}
\mathscr{E}_{S, x, \xi} &:= \bigg\{-x\log\bigg(\frac{1-\rho}{\rho}\bigg)
+ \sum_{\ell\in S}\sum_{\substack{j\neq i: \tzzz{\star, -i}_j = -\xi_\ell \mathbf{z}^\star_i}} \bigg[\log\bigg(\frac{p_\ell (1-q_\ell)}{q_\ell(1-p_\ell)}\bigg) \AAA{\ell}_{ij} + \log\bigg(\frac{1-p_\ell}{1-q_\ell}\bigg)\bigg] \\
& \qquad - \sum_{\ell\in S} \sum_{j\neq i : \tzzz{\star, -i}_j = \xi_\ell\mathbf{z}^\star_i} \bigg[\log\bigg(\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg) \AAA{\ell}_{ij} + \log\bigg(\frac{1-p_\ell}{1-q_\ell}\bigg)\bigg] \geq 0 ~ \textnormal{and } E_i\cap F_i\bigg\}.
\end{align*}
Invoking Markov's inequality, for $t = t_{S, x, \xi} \in [0, 1]$ whose value will be specified later, we can bound the conditional probability by
\begin{align*}
& {\mathbb{P}}\bigg(\mathscr{E}_{S, x, \xi}~\bigg|~ \{\zzz{\ell}_{-i}\}_{\ell=1}^L\bigg)\\
& \leq {\mathbb{E}}\bigg\{ \exp\bigg\{ -t x\log \bigg(\frac{1-\rho}{\rho}\bigg)
+ t\sum_{\ell\in S}\sum_{\substack{j\neq i: \tzzz{\star, -i}_j = -\xi_\ell \mathbf{z}^\star_i}} \bigg[\log\bigg(\frac{p_\ell (1-q_\ell)}{q_\ell(1-p_\ell)}\bigg) \AAA{\ell}_{ij} + \log\bigg(\frac{1-p_\ell}{1-q_\ell}\bigg)\bigg]\\
& \qquad -t \sum_{\ell\in S} \sum_{j\neq i : \tzzz{\star, -i}_j = \xi_\ell\mathbf{z}^\star_i} \bigg[\log\bigg(\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg) \AAA{\ell}_{ij} + \log\bigg(\frac{1-p_\ell}{1-q_\ell}\bigg)\bigg]\bigg\} \cdot \mathbf{1}_{E_i\cap F_i}~ \bigg| ~ \{\zzz{\ell}_{-i}\}_{\ell=1}^L\bigg\}
\end{align*}
Let us define
\begin{align}
\label{eq:ml-i-}
\mmm{\ell, -i}_- &:= \#\{j\neq i: \tzzz{\star, -i}_j = -\xi_\ell \mathbf{z}^\star_i\} ,\\
\label{eq:tml-i-}
\tmmm{\ell, -i}_- & := \# \{j\neq i: \tzzz{\star, -i}_j = -\xi_\ell \mathbf{z}^\star_i, \tzzz{\star, -i}_j = \zzz{\ell}_j\} \\
\label{eq:ml-i+}
\mmm{\ell, -i}_+ & := \#\{j\neq i: \tzzz{\star, -i}_j = \xi_\ell \mathbf{z}^\star_i\} \\
\label{eq:tml-i+}
\tmmm{\ell, -i} & := \#\{j\neq i: \tzzz{\star, -i}_j = \xi_\ell \mathbf{z}^\star_i , \tzzz{\star, -i}_j = \zzz{\ell}_j\}.
\end{align}
Then, we have
\begin{align*}
& {\mathbb{P}}\bigg(\mathscr{E}_{S, x, \xi}~\bigg|~ \{\zzz{\ell}_{-i}\}_{\ell=1}^L\bigg)\\
& \leq \exp\bigg\{-tx \log\bigg(\frac{1-\rho}{\rho}\bigg)\bigg\} \\
& ~~\times {\mathbb{E}}\bigg[
\prod_{\ell \in S} \exp\bigg\{ t \log\bigg(\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg)\\
& \qquad\qquad\qquad\qquad \times \bigg(
\sum_{\substack{j\neq i\\ \tzzz{\star, -i}_j = -\xi_\ell \mathbf{z}^\star_i \\ \tzzz{\star, -i}_{j} = \zzz{\ell}_j}} \AAA{\ell}_{ij}
+ \sum_{\substack{j\neq i\\ \tzzz{\star, -i}_j = -\xi_\ell \mathbf{z}^\star_i \\ \tzzz{\star, -i}_{j} = -\zzz{\ell}_j}} \AAA{\ell}_{ij}
- \sum_{\substack{j\neq i\\ \tzzz{\star, -i}_j = \xi_\ell \mathbf{z}^\star_i \\ \tzzz{\star, -i}_{j} = \zzz{\ell}_j}} \AAA{\ell}_{ij}
- \sum_{\substack{j\neq i\\ \tzzz{\star, -i}_j = \xi_\ell \mathbf{z}^\star_i \\ \tzzz{\star, -i}_{j} = -\zzz{\ell}_j}} \AAA{\ell}_{ij}
\bigg)\\
& \qquad\qquad\qquad\qquad + t \log\bigg(\frac{1-p_\ell}{1-q_\ell}\bigg)(\mmm{\ell, -i}_- - \mmm{\ell, -i}_+)\bigg\} \cdot \mathbf{1}_{E_i \cap F_i}
~\bigg|~ \{\zzz{\ell}_{-i}\}_{\ell=1}^L\bigg]\\
& = \exp\bigg\{-tx \log\bigg(\frac{1-\rho}{\rho}\bigg)\bigg\} \\
& ~~ \times \prod_{\ell\in S}{\mathbb{E}}\bigg\{ \bigg[q_\ell \bigg(\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg)^t + (1-q_\ell)\bigg]^{\tmmm{\ell, -i}_-}
\cdot \bigg[p_\ell\bigg(\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg)^t + (1-p_\ell)\bigg]^{\mmm{\ell, -i}_- - \tmmm{\ell, -i}_-} \\
& \qquad \times \bigg[p_\ell\bigg(\frac{q_\ell(1-p_\ell)}{p_\ell(1-q_\ell)}\bigg)^t + (1-p_\ell)\bigg]^{\tmmm{\ell, -i}_+}
\cdot \bigg[ q_\ell\bigg(\frac{q_\ell(1-p_\ell)}{p_\ell(1-q_\ell)}\bigg)^t + (1-q_\ell) \bigg]^{\mmm{\ell, -i}_+ - \tmmm{\ell, -i}_+}\\
& \qquad \times \bigg( \frac{1-p_\ell}{1-q_\ell} \bigg)^{t(\mmm{\ell, -i}_- - \mmm{\ell, -i}_+)}
\cdot \mathbf{1}_{E_i \cap F_i}
~\bigg|~ \{\zzz{\ell}_{-i}\}_{\ell=1}^L\bigg\}\\
& = \exp\bigg\{-tx \log\bigg(\frac{1-\rho}{\rho}\bigg)\bigg\} \\
& ~~ \times \prod_{\ell\in S} {\mathbb{E}}\bigg\{
\bigg(\frac{1-p_\ell}{1-q_\ell}\bigg)^{t(\mmm{\ell, -i}_- - \mmm{\ell,-i}_+)}
\cdot \bigg[ q_\ell \bigg(\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg)^t + (1-q_\ell) \bigg]^{\mmm{\ell, -i}_-}\\
& \qquad \times \bigg[ p_\ell\bigg(\frac{q_\ell(1-p_\ell)}{p_\ell(1-q_\ell)}\bigg)^t + (1-p_\ell) \bigg]^{\mmm{\ell, -i}_+}
\cdot \bigg[ \frac{p_\ell \bigg(\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg)^t + (1-p_\ell)}{q_\ell \bigg(\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg)^t + (1-q_\ell)} \bigg]^{\mmm{\ell, -i}_- - \tmmm{\ell, -i}_-}\\
& \qquad \times \bigg[ \frac{q_\ell \bigg(\frac{q_\ell(1-p_\ell)}{p_\ell(1-q_\ell)}\bigg)^t + (1-q_\ell)}{p_\ell \bigg(\frac{q_\ell(1-p_\ell)}{p_\ell(1-q_\ell)}\bigg)^t + (1-p_\ell)}\bigg]^{\mmm{\ell, -i}_+ - \tmmm{\ell, -i}_+}
\cdot \mathbf{1}_{E_i\cap F_i}
~\bigg|~ \{\zzz{\ell}_{-i}\}_{\ell=1}^L\bigg\}\\
& = \exp\bigg\{-tx \log\bigg(\frac{1-\rho}{\rho}\bigg)\bigg\}
\times \prod_{\ell\in S}{\mathbb{E}}\bigg\{ \mathscr{T}_{1, \ell} \times \mathscr{T}_{2, \ell} \times \mathscr{T}_{3, \ell}
\cdot \mathbf{1}_{E_i\cap F_i}
~\bigg|~ \{\zzz{\ell}_{-i}\}_{\ell=1}^L\bigg\},
\end{align*}
where
\begin{align*}
\mathscr{T}_{1, \ell} & := \bigg(\frac{1-p_\ell}{1-q_\ell}\bigg)^{t(\mmm{\ell, -i}_- - \mmm{\ell, -i}_+)}
\times \bigg[q_\ell\bigg(\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg)^t + (1-q_\ell)\bigg]^{\frac{\mmm{\ell, -i}_- - \mmm{\ell, -i}_+}{2}} \\
& \qquad
\times \bigg[p_\ell \bigg(\frac{q_\ell(1-p_\ell)}{p_\ell(1-q_\ell)}\bigg)^t + (1-p_\ell)\bigg]^{\frac{\mmm{\ell, -i}_+ - \mmm{\ell, -i}_-}{2}} \\
\mathscr{T}_{2, \ell} & := \bigg[ q_\ell\bigg(\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg)^t + (1-q_\ell) \bigg]^{\frac{\mmm{\ell, -i}_- + \mmm{\ell, -i}_+}{2}}
\times \bigg[p_\ell\bigg(\frac{q_\ell(1-p_\ell)}{p_\ell(1-q_\ell)}\bigg)^t + (1-p_\ell)\bigg]^{\frac{\mmm{\ell, -i}_+ + \mmm{\ell, -i}_-}{2}} \\
\mathscr{T}_{3, \ell} & := \bigg[ \frac{p_\ell \bigg(\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg)^t + (1-p_\ell)}{q_\ell \bigg(\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg)^t + (1-q_\ell)} \bigg]^{\mmm{\ell, -i}_- - \tmmm{\ell, -i}_-}
\times \bigg[ \frac{q_\ell \bigg(\frac{q_\ell(1-p_\ell)}{p_\ell(1-q_\ell)}\bigg)^t + (1-q_\ell)}{p_\ell \bigg(\frac{q_\ell(1-p_\ell)}{p_\ell(1-q_\ell)}\bigg)^t + (1-p_\ell)}\bigg]^{\mmm{\ell, -i}_+ - \tmmm{\ell, -i}_+}.
\end{align*}
In summary, we arrive at
\begin{align}
& {\mathbb{P}}(\barzzz{\star, -i}_i = -\mathbf{z}^\star_i , E_i\cap F_i) \nonumber\\
& \leq \sum_{S \subseteq [L]}
\sum_{x \in\{ -|S^c|+2k: 0\leq k \leq |S^c|\}}
\binom{|S^c|}{\frac{|S^c| + x}{2}}
\exp\bigg\{ -|S^c|\log\frac{1}{\sqrt{\rho(1-\rho)}} + x(1-2t)\log \sqrt{\frac{1-\rho}{\rho}}\bigg\} \nonumber\\
\label{eq:glob_nodewise_err_decom_3terms}
& ~~ \times \sum_{\xi\in\{\pm 1\}^S}{\mathbb{P}}(\zzz{S}_i = \mathbf{z}^\star_i \xi )
\cdot {\mathbb{E}}_{\{\zzz{\ell}_{-i}\}_{\ell=1}^L} \bigg\{
\prod_{\ell\in S}{\mathbb{E}}\bigg[ \mathscr{T}_{1, \ell}\times \mathscr{T}_{2, \ell} \times \mathscr{T}_{3, \ell} \cdot \mathbf{1}_{E_i\cap F_i} ~\bigg|~ \{\zzz{\ell}_{-i}\}_{\ell=1}^L \bigg]\bigg\},
\end{align}
We now bound the three terms $\mathscr{T}_{1, \ell}, \mathscr{T}_{2, \ell}$ and $\mathscr{T}_{3, \ell}$ separately.
~\\
{\noindent \bf Bounding the first term.} We can write
\begin{align*}
\mathscr{T}_{1, \ell} & = \bigg[ \frac{(1-p_\ell)^{2t}}{(1-q_\ell)^{2t}}
\times \frac{q_\ell \bigg(\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg)^t + (1-q_\ell)}{p_\ell\bigg( \frac{q_\ell(1-p_\ell)}{p_\ell(1-q_\ell)} \bigg)^t + (1-p_\ell)} \bigg]^{\frac{\mmm{\ell, -i}_+ - \mmm{\ell, -i}_-}{2}} \\
& = \bigg[ \frac{p_\ell^t q_\ell^{1-t} + (1-p_\ell)^t (1-q_\ell)^{1-t}}{p_\ell^{1-t}q_\ell^t + (1-p_\ell)^{1-t}(1-q_\ell)^t} \bigg]^{\frac{\mmm{\ell, -i}_+ - \mmm{\ell, -i}_-}{2}} \\
& = \exp\bigg\{\frac{\mmm{\ell, -i}_+ - \mmm{\ell, -i}_-}{2}
\times \log \bigg(\frac{p_\ell^t q_\ell^{1-t} + (1-p_\ell)^t (1-q_\ell)^{1-t}}{p_\ell^{1-t}q_\ell^t + (1-p_\ell)^{1-t}(1-q_\ell)^t}\bigg)\bigg\}.
\end{align*}
We need the following two lemmas.
\begin{lemma}
\label{lemma:glob_nodewise_err_decomp_1st_term_part1}
Under the setup of Proposition \ref{prop:refine_glob_nodewise}, on the event $E_i$, for any $\ell\in S$, we have
$$
\bigg|(\mmm{\ell, -i}_+ - \mmm{\ell, -i}_-) - (\nnn{\ell, -i}_+ - \nnn{\ell, -i}_{-})\bigg| \leq 2n\delta_n
$$
\end{lemma}
\begin{proof}
Note that for $\ell\in S$, we have $\xi_\ell \mathbf{z}^\star_i = \zzz{\ell}_i$. By definition \eqref{eq:ml-i+}, we have
\begin{align*}
\mmm{\ell, -i}_+ - \nnn{\ell, -i}_+ & = \sum_{j\neq i}\bigg(\mathbf{1}{\tzzz{\star, -i}_j = \zzz{\ell}_i} - \mathbf{1}{\zzz{\ell}_j = \zzz{\ell}_i}\bigg)\\
& = \sum_{j\neq i} \bigg(\mathbf{1}{\tzzz{\star, -i}_j = \zzz{\ell}_i, \tzzz{\star , -i}_j = \zzz{\ell}_j} + \mathbf{1}{\tzzz{\star, -i}_j = \zzz{\ell}_i , \tzzz{\star, -i} \neq \zzz{\ell}_j} \\
& \qquad - \mathbf{1}{\zzz{\ell}_j = \zzz{\ell}_i , \tzzz{\star, -i}_j = \zzz{\ell}_j } - \mathbf{1}{\zzz{\ell}_j = \zzz{\ell}_i , \tzzz{\star, -i}_j \neq \zzz{\ell}_j} \bigg)\\
& = \sum_{j\neq i} \bigg(\mathbf{1}{\tzzz{\star, -i}_j = \zzz{\ell}_i, \tzzz{\star , -i}_j = \zzz{\ell}_j} - \mathbf{1}{\zzz{\ell}_j = \zzz{\ell}_i , \tzzz{\star, -i}_j \neq \zzz{\ell}_j} \bigg)\\
& \leq \sum_{j\neq i}\mathbf{1}{\tzzz{\star, -i}_j = \zzz{\ell}_i, \tzzz{\star , -i}_j = \zzz{\ell}_j} \\
& \leq \#\{j\neq i: \tzzz{\ell, -i}_j \neq \zzz{\ell}_j\} \\
& \leq n\delta_n
\end{align*}
where the last inequality holds on the event $E_i$. On the other hand, we have
\begin{align*}
\mmm{\ell, -i}_+ - \nnn{\ell, -i}_+ & = \sum_{j\neq i} \bigg(\mathbf{1}{\tzzz{\star, -i}_j = \zzz{\ell}_i, \tzzz{\star , -i}_j = \zzz{\ell}_j} - \mathbf{1}{\zzz{\ell}_j = \zzz{\ell}_i , \tzzz{\star, -i}_j \neq \zzz{\ell}_j}\bigg)\\
& \geq -\sum_{j\neq i} \mathbf{1}{\zzz{\ell}_j = \zzz{\ell}_i , \tzzz{\star, -i}_j \neq \zzz{\ell}_j} \\
& \geq -\# \{j\neq i: \tzzz{\star, -i}_j \neq \zzz{\ell}_j\} \geq -n\delta_n,
\end{align*}
where the last inequality again holds on the event $E_i$. Thus, we arrive at
$$
|\mmm{\ell, -i}_+ - \nnn{\ell, -i}_+| \leq n\delta_n
$$
on $E_i$. A similar argument shows that $|\mmm{\ell, -i}_- - \nnn{\ell, -i}_-|\leq n\delta_n$ on $E_i$, and the proof is concluded by invoking the triangle inequality.
\end{proof}
\begin{lemma}
\label{lemma:glob_nodewise_err_decomp_1st_term_part2}
Under the setups of Proposition \ref{prop:refine_glob_nodewise}, for any $t\in[0, 1]$, we have
$$
\log \bigg|\frac{p_\ell^t q_\ell^{1-t} + (1-p_\ell)^t (1-q_\ell)^{1-t}}{p_\ell^{1-t}q_\ell^t + (1-p_\ell)^{1-t}(1-q_\ell)^t}\bigg| \leq \III{\ell}_{t},
$$
where $\III{\ell}_t$ is defined in \eqref{eq:layerwise_info}.
\end{lemma}
\begin{proof}
We first show the numerator $p_\ell^t q_\ell^{1-t} + (1-p_\ell)^t (1-q_\ell)^{1-t}\leq 1 $. To do this, we take the derivative w.r.t. $t$:
$$
\frac{\partial}{\partial t} \bigg(p_\ell^t q_\ell^{1-t} + (1-p_\ell)^t (1-q_\ell)^{1-t} \bigg)= q_\ell \bigg(\frac{p_\ell}{q_\ell}\bigg)^t \log \frac{p_\ell}{q_\ell} + (1-q_\ell) \bigg(\frac{1-p_\ell}{1-q_\ell}\bigg)^t \log \frac{1-p_\ell}{1-q_\ell}.
$$
Note that since $p_\ell > q_\ell$, the right-hand side above is an increasing function in $t$. So the numerator is a convex function in $t$. This means that its maximum must occur at the boundary, which is at either $t = 0$ or $t = 1$, both of which gives $p_\ell^t q_\ell^{1-t} + (1-p_\ell)^t (1-q_\ell)^{1-t} = 1$. By symmetry, the denominator also satisfies $p_\ell^{1-t}q_\ell^t + (1-p_\ell)^{1-t}(1-q_\ell)^t \leq 1$. Thus, we can proceed by
\begin{align*}
& \log \bigg|\frac{p_\ell^t q_\ell^{1-t} + (1-p_\ell)^t (1-q_\ell)^{1-t}}{p_\ell^{1-t}q_\ell^t + (1-p_\ell)^{1-t}(1-q_\ell)^t}\bigg| \\
& = \bigg|\log\bigg(p_\ell^t q_\ell^{1-t} + (1-p_\ell)^t (1-q_\ell)^{1-t}\bigg) - \log\bigg(p_\ell^{1-t}q_\ell^t + (1-p_\ell)^{1-t}(1-q_\ell)^t\bigg)\bigg| \\
& \leq - \log \bigg(p_\ell^t q_\ell^{1-t} + (1-p_\ell)^t (1-q_\ell)^{1-t}\bigg) - \log \bigg(p_\ell^{1-t}q_\ell^t + (1-p_\ell)^{1-t}(1-q_\ell)^t\bigg) \\
& = \III{\ell}_t,
\end{align*}
which is the desired result.
\end{proof}
Invoking Lemma \ref{lemma:glob_nodewise_err_decomp_1st_term_part1}, we know the on the event $E_i \cap F_i$,
\begin{align*}
& \bigg|(\mmm{\ell, -i}_+ - \mmm{\ell, -i}_-) - (n^{\star}_+ - n^{\star}_-)\bigg| \\
& \leq \bigg|(\mmm{\ell, -i}_+ - \mmm{\ell, -i}_-) - (\nnn{\ell, -i}_+ - \nnn{\ell, -i}_-)\bigg| + \bigg|(\nnn{\ell, -i}_+ - \nnn{\ell, -i}_-) - (n^{\star}_+ - n^{\star}_-)\bigg|\\
& \leq n\cdot ( 2\delta_n + n^{-1/4} + n^{-1} + 2\rho).
\end{align*}
Combining the above inequality with Lemma \ref{lemma:glob_nodewise_err_decomp_1st_term_part2}, and recalling that $\zzz{\ell}_i = \xi_\ell \mathbf{z}^\star_i$ for $\ell\in[L]$, we get
\begin{align*}
& \mathscr{T}_{1, \ell}\cdot \mathbf{1}_{E_i\cap F_i} \\
& \leq \exp\bigg\{\frac{n^{\star}_{\mathbf{z}^\star_i} - n^{\star}_{-\mathbf{z}^\star_i}}{2}\cdot \bigg(\mathbf{1}{\xi_\ell = 1} - \mathbf{1}{\xi_\ell = -1}\bigg)\times \log \bigg(\frac{p_\ell^t q_\ell^{-1-t} + (1-p_\ell)^t (1-q_\ell)^{1-t}}{p_\ell^{1-t} q_\ell^t + (1-p_\ell)^{1-t}(1-q_\ell)^t}\bigg)\\
& \qquad \qquad + \frac{ n\cdot ( 2\delta_n + n^{-1/4} + n^{-1} + 2\rho)}{2} \cdot \III{\ell}_t\bigg\}\addtocounter{equation}{1}\tag{\theequation} \\
& \leq \exp\bigg\{ \frac{n\III{\ell}_t}{2} \cdot (\beta - \beta^{-1} + n^{-1/4} + n^{-1} + 2\rho)\bigg\} \nonumber\\
\label{eq:glob_nodewise_err_decomp_1st_term}
& = \exp\{o(1)\cdot n \III{\ell}_t\}. \addtocounter{equation}{1}\tag{\theequation}
\end{align*}
~\\
{\noindent \bf Bounding the second term.} Since $\mmm{\ell, -i}_+ + \mmm{\ell, -i}_- = n-1$, we have
\begin{align*}
\mathscr{T}_{2, \ell} & = \exp\bigg\{ \frac{n-1}{2}\cdot \log\bigg(p_\ell q_\ell + (1-p_\ell)(1-q_\ell) + [q_\ell(1-p_\ell)]^{1-t}[p_\ell(1-q_\ell)]^t + [q_\ell(1-p_\ell)]^t [p_\ell(1-q_\ell)]^{1-t}\bigg)\bigg\} \\
\label{eq:glob_nodewise_err_decomp_2ed_term}
& = \exp\{-(n-1)\III{\ell}_t/2\} \addtocounter{equation}{1}\tag{\theequation}
\end{align*}
~\\
{\noindent \bf Bounding the third term.} With some algebra, one can show that
\begin{align*}
\frac{p_\ell \bigg(\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg)^t + (1-p_\ell)}{q_\ell \bigg(\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg)^t + (1-q_\ell)}
& = 1 + \frac{(p_\ell - q_\ell)\bigg[\bigg(1 + \frac{p_\ell - q_\ell}{q_\ell - p_\ell q_\ell}\bigg)^t - 1\bigg]}{1-q_\ell + q_\ell\bigg(1 + \frac{p_\ell - q_\ell}{q_\ell - p_\ell q_\ell}\bigg)^t}.
\end{align*}
Since $(1+ \frac{p_\ell-q_\ell}{q_\ell-p_\ell q_\ell})^t - 1\leq \frac{p_\ell-q_\ell}{q_\ell-p_\ell q_\ell}$ when $t \in[0, 1]$, along with the assumption that $p_\ell\asymp q_\ell = o(1)$, the right-hand side above can be bounded by
$$
\frac{p_\ell \bigg(\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg)^t + (1-p_\ell)}{q_\ell \bigg(\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg)^t + (1-q_\ell)}
\leq 1 + {\mathcal{O}}\bigg(\frac{(p_\ell-q_\ell)^2}{p_\ell}\bigg) \leq \exp\bigg\{{\mathcal{O}}\bigg(\frac{(p_\ell-q_\ell)^2}{p_\ell}\bigg)\bigg\}.
$$
A similar argument shows that
$$
\frac{q_\ell \bigg(\frac{q_\ell(1-p_\ell)}{p_\ell(1-q_\ell)}\bigg)^t + (1-q_\ell)}{p_\ell \bigg(\frac{q_\ell(1-p_\ell)}{p_\ell(1-q_\ell)}\bigg)^t + (1-p_\ell)} \leq \exp\bigg\{{\mathcal{O}}\bigg(\frac{(p_\ell-q_\ell)^2}{p_\ell}\bigg)\bigg\}.
$$
Now, on the event $E_i$, we have
\begin{align*}
|\mmm{\ell, -i}_- - \tmmm{\ell, -i}_-| & = \bigg|\sum_{j\neq i}\bigg(\mathbf{1}{\tzzz{\star, -i}_j = -\zzz{\ell}_i} - \mathbf{1}{\tzzz{\star, -i}_j = -\zzz{\ell}_i, \tzzz{\star, -i}_j = \zzz{\ell}_j}\bigg)\bigg|\\
& \leq \sum_{j\neq i} \mathbf{1}{\tzzz{\star, -i}_j = -\zzz{\ell}_j} \\
& \leq n\delta_n,
\end{align*}
and the same bound holds for $|\mmm{\ell, -i}_+ - \tmmm{\ell, -i}_+|$. Thus, we get
\begin{align*}
\label{eq:glob_nodewise_err_decomp_3ed_term}
\mathscr{T}_{3, \ell}\cdot \mathbf{1}_{E_i} \leq \exp\bigg\{{\mathcal{O}}\bigg(\frac{n\delta_n (p_\ell-q_\ell)^2}{p_\ell}\bigg)\bigg\} = \exp\bigg\{ o(1) \cdot n\III{\ell}_{1/2} \bigg\}, \addtocounter{equation}{1}\tag{\theequation}
\end{align*}
where the last inequality is by Lemma \ref{lemma:asymp_equiv_of_I_t}.
~\\
{\bf \noindent Summarizing the three terms.} Plugging \eqref{eq:glob_nodewise_err_decomp_1st_term}, \eqref{eq:glob_nodewise_err_decomp_2ed_term} and \eqref{eq:glob_nodewise_err_decomp_3ed_term} to \eqref{eq:glob_nodewise_err_decom_3terms}, we get
\begin{align*}
& {\mathbb{P}}(\barzzz{\star, -i}_i = -\mathbf{z}^\star_i , E_i\cap F_i) \nonumber\\
& \leq \sum_{S \subseteq [L]}
\sum_{x \in\{ -|S^c|+2k: 0\leq k \leq |S^c|\}}
\binom{|S^c|}{\frac{|S^c| + x}{2}}
\exp\bigg\{ -|S^c|\log\frac{1}{\sqrt{\rho(1-\rho)}} + x(1-2t_{S, x, \xi})\log \sqrt{\frac{1-\rho}{\rho}}\bigg\} \nonumber\\
& ~~ \times \sum_{\xi\in\{\pm 1\}^S}{\mathbb{P}}(\zzz{S}_i = \mathbf{z}^\star_i \xi )
\cdot {\mathbb{E}}_{\{\zzz{\ell}_{-i}\}_{\ell=1}^L} \bigg\{
\prod_{\ell\in S}{\mathbb{E}}\bigg[ \exp\big\{ -n\III{\ell}_{t_{S, x, \xi}}/2 + o(1)\cdot n(\III{\ell}_{t_{S, x , \xi}} + \III{\ell}_{1/2}) \big\} ~\bigg|~ \{\zzz{\ell}_{-i}\}_{\ell=1}^L \bigg]\bigg\} \\
& \leq \sum_{S \subseteq [L]}
\sum_{x \in\{ -|S^c|+2k: 0\leq k \leq |S^c|\}}
\binom{|S^c|}{\frac{|S^c| + x}{2}}
\exp\bigg\{ -|S^c|\log\frac{1}{\sqrt{\rho(1-\rho)}} + x(1-2t)\log \sqrt{\frac{1-\rho}{\rho}}\bigg\} \nonumber\\
\label{eq:nodewise_err_decomp_3terms_simplified}
& ~~ \times \sum_{\xi\in\{\pm 1\}^S}{\mathbb{P}}(\zzz{S}_i = \mathbf{z}^\star_i \xi )
\cdot \exp\bigg\{ -\frac{n}{2}\sum_{\ell\in S}\III{\ell}_{t_{S, x, \xi}}+ o(1)\cdot n\sum_{\ell\in S}\III{\ell}_{1/2}\bigg\},\addtocounter{equation}{1}\tag{\theequation}
\end{align*}
where the last inequality is by $\III{\ell}_t\leq \III{\ell}_{1/2}$ for any $t\in[0, 1]$, proved in Lemma \ref{lemma:monotonicity_of_I_t}. This is a summation over all $2^L$ subsets of $[L]$, and we now carefully choose $t_{x, S, \xi}$ to make each summand as tight as possible.
~\\
{\bf \noindent Case A: either $\boldsymbol{|S^c|}$ is even, or $\boldsymbol{\log e|S^c| \geq \sqrt{J_\rho}}$.} In this case, we choose $t_{S, x, \xi} = 1/2$. For any fixed $|S^c|$ falling into this case, the corresponding summand in the right-hand side of \eqref{eq:nodewise_err_decomp_3terms_simplified} becomes
\begin{align*}
& \sum_{x \in\{ -|S^c|+2k: 0\leq k \leq |S^c|\}}
\binom{|S^c|}{\frac{|S^c| + x}{2}}
\exp\bigg\{ -|S^c|\log\frac{1}{\sqrt{\rho(1-\rho)}}
\bigg\}
\times \exp\bigg\{-\frac{(1+o(1))n}{2}\sum_{\ell\in S} \III{\ell}_{1/2}\bigg\}\\
& = \exp\bigg\{-(1+o(1)) \bigg( |S^c|J_\rho + \sum_{\ell\in S}\III{\ell}_{1/2} \bigg)\bigg\}\\
& = \exp\bigg\{-(1+o(1))\bigg(|S^c| J_\rho + \psi^\star_S(0)\bigg)\bigg\},
\end{align*}
where the last equality is by Lemma \ref{lemma:ineq_for_cvx_conjugate}. We now make the following claim.
\begin{lemma}
\label{lemma:nodewise_err_caseA_asymp_equiv}
Under the setup of Proposition \ref{prop:refine_glob_nodewise}, if $\log e|S^c| \geq \sqrt{J_\rho}$, then we have
\begin{equation}
\label{eq:nodewise_err_caseA_asymp_equiv}
1 - \frac{1}{e^{\sqrt{J_\rho}-1}}\leq \frac{|S^c|J_\rho + \psi^\star_S(0)}{(|S^c|+1)J_\rho + \psi_S(-2J_\rho)} \leq 1.
\end{equation}
\end{lemma}
\begin{proof}
By Lemma \ref{lemma:ineq_for_cvx_conjugate}, the denominator in the left-hand side of \eqref{eq:nodewise_err_caseA_asymp_equiv} lies between
$
-J_\rho + \sum_{\ell\in S} \III{\ell}_{1/2}
$
and
$
\sum_{\ell \in S} \III{\ell}_{1/2}.
$
Thus, the left-hand side of \eqref{eq:nodewise_err_caseA_asymp_equiv} is between
$$
\bigg[\frac{|S^c| J_\rho + \sum_{\ell\in S} \III{\ell}_{1/2}}{(|S^c| + 1) J_\rho + \sum_{\ell\in S} \III{\ell}_{1/2}}~ , ~1\bigg].
$$
Note that the lower bound in the above display can be further lower bounded by
$$
1 - \frac{J_\rho}{(|S^c|+ 1)J_\rho + \sum_{\ell\in S}\III{\ell}_{1/2}} \geq 1 - \frac{1}{|S^c|} \geq 1 - \frac{1}{e^{\sqrt{J_\rho} - 1}},
$$
where the last inequality is by our assumption that $\log e|S^c| \geq \sqrt{J_\rho}$.
\end{proof}
The above lemma tells that under Case A, each summand (for a fixed $S$) in the right-hand side of \eqref{eq:nodewise_err_decomp_3terms_simplified} can be upper bounded by
\begin{align*}
\begin{cases}
\exp\bigg\{-(1+o(1)) \bigg(|S^c| J_\rho + \psi^\star_S(0)\bigg)\bigg\} & \textnormal{ if } |S^c| \textnormal{ is even},\\
\exp\bigg\{-(1+o(1)) \bigg((|S^c| + 1) J_\rho + \psi^\star_S(-2J_\rho)\bigg)\bigg\} & \textnormal{ if } |S^c| \textnormal{ is odd}.
\end{cases}
\end{align*}
~\\
{\bf \noindent Case B: $\boldsymbol{|S^c|}$ is odd and $\boldsymbol{\log e|S^c| \leq \sqrt{J_\rho}}$.} With the requirement that $t_{S, x , \xi} = t_{S, x}$ (i.e., $t$ is independent of $\xi$), each summand (for a fixed $S$) in the right-hand side of \eqref{eq:nodewise_err_decomp_3terms_simplified} becomes
\begin{align*}
& \sum_{x \in\{ -|S^c|+2k: 0\leq k \leq |S^c|\}}
\binom{|S^c|}{\frac{|S^c| + x}{2}}
\exp\bigg\{ -|S^c|\log\frac{1}{\sqrt{\rho(1-\rho)}} + x(1-2t_{S, x})\log \sqrt{\frac{1-\rho}{\rho}}\bigg\} \\
& \qquad \times \exp\bigg\{ - \frac{n}{2}\sum_{\ell\in S} \III{\ell}_{t_{S, x}} + o(1)\cdot n\sum_{\ell\in S} \III{\ell}_{1/2}\bigg\} \\
& \leq
\sum_{x \in\{ -|S^c|+2k: 0\leq k \leq |S^c|\}}
\exp\bigg\{ |S^c| \log (e|S^c|) - |S^c| \log \frac{1}{\sqrt{\rho(1-\rho)}} + x(1- 2t_{S, x}) \log \sqrt{\frac{1-\rho}{\rho}}\bigg\}\\
& \qquad \times \exp\bigg\{ - \frac{n}{2}\sum_{\ell\in S} \III{\ell}_{t_{S, x}} + o(1)\cdot n\sum_{\ell\in S} \III{\ell}_{1/2}\bigg\} \\
& \leq
\sum_{x \in\{ -|S^c|+2k: 0\leq k \leq |S^c|\}}
\exp\bigg\{|S^c| \bigg(\sqrt{J_\rho} - \log \frac{1}{\sqrt{\rho(1-\rho)}}\bigg) + x(1- 2t_{S, x}) \log \sqrt{\frac{1-\rho}{\rho}}\bigg\}\\
\label{eq:nodewise_err_decomp_3terms_simplified_caseB}
& \qquad \times \exp\bigg\{ - \frac{n}{2}\sum_{\ell\in S} \III{\ell}_{t_{S, x}} + o(1)\cdot n \sum_{\ell\in S} \III{\ell}_{1/2}\bigg\},\addtocounter{equation}{1}\tag{\theequation}
\end{align*}
where the first inequality is by $\binom{n}{k}\leq (en/k)^k$ for any values of $n, k$ such that $1\leq k\leq n$, and the second inequality is by our assumption that $\log e |S^c|\leq \sqrt{J_\rho}$.
Let us require $t_{x, S}$ to be symmetric about $1/2$:
$$
\mathbf{1}{x\geq 0}\cdot t_{S, x} - \frac{1}{2} = \frac{1}{2} - \mathbf{1}{x\leq 0} \cdot t_{S, x} .
$$
Under such a requirement, using the fact that $\III{\ell}_t$ is also symmetric about $1/2$ (i.e., $\III{\ell}_{1/2-\delta} = \III{\ell}_{1/2 + \delta}$ for any $\delta\in[0, 1/2]$), the right-hand side of \eqref{eq:nodewise_err_decomp_3terms_simplified_caseB} becomes
\begin{align*}
& 2 \times \sum_{x\in\{-|S^c| + 2k: 0\leq k \leq (|S^c|-1)/2\}}
\exp\bigg\{|S^c| \bigg(\sqrt{J_\rho} - \log \frac{1}{\sqrt{\rho(1-\rho)}}\bigg) - |x|(1- 2t_{S, x}) \log \sqrt{\frac{1-\rho}{\rho}}\bigg\}\\
& \qquad \times \exp\bigg\{ - \frac{n}{2}\sum_{\ell\in S} \III{\ell}_{t_{S, x}} + o(1)\cdot n\sum_{\ell\in S} \III{\ell}_{1/2}\bigg\} \\
& \leq (|S^c| + 1) \exp\bigg\{|S^c| \bigg(\sqrt{J_\rho} - \log \frac{1}{\sqrt{\rho(1-\rho)}}\bigg) - (1- 2t_{S, x}) \log \sqrt{\frac{1-\rho}{\rho}} - \frac{n}{2}\sum_{\ell\in S} \III{\ell}_{t_{S, x}} + o(1)\cdot n\sum_{\ell\in S}\III{\ell}_{1/2}\bigg\},
\end{align*}
where the inequality is because the minimum value that $|x|$ can take is $1$, a consequence of $|S^c|$ being odd. Rearranging terms and using $\log e|S^c|\leq \sqrt{J_\rho}$, the right-hand side above can be further upper bounded by
\begin{align*}
& \exp\bigg\{ |S^c|\cdot {\mathcal{O}}(\sqrt{J_\rho}) + o(1)\cdot n\sum_{\ell\in S}\III{\ell}_{1/2} \\
& \qquad - |S^c|\log \frac{1}{\sqrt{\rho(1-\rho)}} - \log \sqrt{\frac{1-\rho}{\rho}} - 2t_{S, x}\cdot \log\sqrt{\frac{1-\rho}{\rho}} - \frac{n}{2}\sum_{\ell\in S}\III{\ell}_{t_{S, x}}\bigg\}\\
& \leq \exp\bigg\{ |S^c|\cdot {\mathcal{O}}(\sqrt{J_\rho}) + o(1)\cdot n\sum_{\ell\in S}\III{\ell}_{1/2} -(1+o(1))\bigg( (|S^c|+1)J_\rho - 2t_{S, x}J_\rho - \psi_S(t_{S, x}) \bigg) \bigg\},
\end{align*}
where the inequality is by $\rho =o(1)$. Now the optimal choice of $t_{S, x}$ is clear. For $x < 0$, choosing
$
t_{S, x} = \mathop{\rm argmax}_{0\leq t\leq 1} -2t J_\rho - \psi_S(t)
$
gives the following upper bound for the right-hand side of \eqref{eq:nodewise_err_decomp_3terms_simplified_caseB}:
\begin{align*}
& \exp\bigg\{ |S^c|\cdot {\mathcal{O}}(\sqrt{J_\rho}) + o(1)\cdot n\sum_{\ell\in S}\III{\ell}_{1/2}- (1+o(1))\bigg( (|S^c| + 1) J_\rho + \psi_S^\star(-2J_\rho)\bigg)\bigg\} \\
\label{eq:glob_nodewise_err_caseB}
& \leq \exp\bigg\{- (1+o(1))\bigg( (|S^c| + 1) J_\rho + \psi_S^\star(-2J_\rho)\bigg)\bigg\},\addtocounter{equation}{1}\tag{\theequation}
\end{align*}
where the last inequality is by $\sqrt{J_\rho}\ll J_\rho$ and $\frac{n}{2}\cdot \sum_{\ell\in S}\III{\ell}_{1/2}\leq \psi_S^\star(-2J_\rho) + J_\rho$, the latter of which is proved in Lemma \ref{lemma:ineq_for_cvx_conjugate}.
~\\
{\bf \noindent Finishing the proof of Proposition \ref{prop:refine_glob_nodewise}.} Summarizing Case A and Case B, we have
\begin{align*}
{\mathbb{P}}(\barzzz{\star, -i}_i = -\mathbf{z}^\star_i, E_i\cap F_i)
& \leq \sum_{S\subseteq [L]: |S^c|\textnormal{ even}} \exp\bigg\{- (1+o(1)) \bigg(|S^c| J_\rho + \psi^\star_S(0)\bigg)\bigg\} \\
& \qquad + \sum_{S\subseteq[L]: |S^c|\textnormal{ odd}} \exp\bigg\{ -(1+o(1))\bigg((|S^c|+1) J_\rho + \psi_S^\star(-2J_\rho)\bigg)\bigg\},
\end{align*}
where we emphasize that the $o(1)$ terms are independent of $S$. The proof is concluded by plugging the above inequality to \eqref{eq:glob_nodewise_err_decomp}.
\subsection{Proof of Theorem \ref{thm:refine_ind}}\label{prf:thm:refine_ind}
We use the same notations as those in the proof of Theorem \ref{thm:refine_glob}.
The proof is based on the following counterpart to Proposition \ref{prop:refine_glob_nodewise}.
\begin{proposition}
\label{prop:refine_ind_nodewise}
Fix $\ell\in[L]$. Under the setup of Proposition \ref{prop:refine_glob_nodewise}, there exists a sequence $\delta_n'=o(1)$ and an absolute constant $C'>0$ such that for any $i\in[n]$, we have
\begin{align}
\label{eq:upper_bound_z_l}
{\mathbb{P}}(\pi_i\barzzz{\ell, -i}_i \neq \zzz{\ell}_i)
& \leq C'n^{-(1+\varepsilon_{{\texttt{init}}})}+ \sum_{S\subseteq[L]\setminus\{\ell\}} \bigg(e^{-(1-\delta_n) \mathcal{I}_{S\cup\{\ell\}}} + e^{-(1-\delta_n)\mathcal{J}_{S\cup\{\ell\}}}\bigg),
\end{align}
where $\mathcal{I}_S, \mathcal{J}_S$ are defined in \eqref{eq:globinfo} and \eqref{eq:indinfo} respectively.
\end{proposition}
\begin{proof}
See Appendix \ref{prf:prop:refine_ind_nodewise}.
\end{proof}
With Lemma \ref{lemma:align} and the above proposition at hand, the rest of the proof is nearly identical to the proof of Theorem \ref{thm:refine_glob}, and we omit the details.
\subsubsection{Proof of Proposition \ref{prop:refine_ind_nodewise}}\label{prf:prop:refine_ind_nodewise}
Without loss of generality we consider the first layer and we assume $\pi_i = +1$. We start by computing
\begin{align*}
{\mathbb{P}}(\barzzz{1, -i}_i\neq \zzz{1}_i) = {\mathbb{P}}(\barzzz{1, -i}_i = -\zzz{1}_i) \leq {\mathbb{P}}(\barzzz{1, -i}_i = -\zzz{1}_i, E_i\cap F_i) + C' n^{-(1+\varepsilon_{{\texttt{init}}})},
\end{align*}
where $E_i, F_i$ are defined in \eqref{eq:Ei} and \eqref{eq:Fi} respectively, and the last inequality is by \eqref{eq:high_prob_Ei_Fi}. We now proceed by
\begin{align*}
& {\mathbb{P}}(\barzzz{1, -i}_i = -\zzz{1}_i, E_i\cap F_i) \\
& = \sum_{S\subseteq \{2, \hdots, L\}} \bigg({\mathbb{P}}(\barzzz{\star, -i}_i = -\mathbf{z}^\star_i, \barzzz{1, -i}_i = -\zzz{1}_i, \barzzz{S, -i}_i = -\zzz{S}_i, \barzzz{S^c, -i}_i = \zzz{S^c}_i, E_i\cap F_i) \\
\label{eq:ind_nodewise_err_decomp_all_subsets}
& \qquad + {\mathbb{P}}(\barzzz{\star, -i}_i = \mathbf{z}^\star_i, \barzzz{1, -i}_i = -\zzz{1}_i, \barzzz{S, -i}_i = -\zzz{S}_i, \barzzz{S^c, -i}_i = \zzz{S^c}_i, E_i\cap F_i)\bigg) ,\addtocounter{equation}{1}\tag{\theequation}
\end{align*}
where we denoted $\barzzz{S, -i}_i = \{\barzzz{\ell, -i}_i: \ell\in S\}$ and $\zzz{S}_i=\{\zzz{\ell}_i: \ell\in S\}$. The right-hand side above is the superposition of two terms, the first of which has already been calculated in the proof of Proposition \ref{prop:refine_glob_nodewise} (see \eqref{eq:glob_nodewise_err_decomp_all_subsets}):
\begin{align*}
& \sum_{S\subseteq\{2, \hdots, L\}}{\mathbb{P}}(\barzzz{\star, -i}_i = -\mathbf{z}^\star_i, \barzzz{1, -i}_i = -\zzz{1}_i, \barzzz{S, -i}_i = -\zzz{S}_i, \barzzz{S^c, -i}_i = \zzz{S^c}_i, E_i\cap F_i)\\
\label{eq:ind_nodewise_err_globinfo}
& \leq \sum_{S\subseteq\{2, \hdots, L\}} e^{-(1+o(1))\mathcal{I}_{S\cup\{1\}}},\addtocounter{equation}{1}\tag{\theequation}
\end{align*}
where the $o(1)$ term is independent of $S$. For the second term in the right-hand side of \eqref{eq:ind_nodewise_err_decomp_all_subsets}, we have
\begin{align*}
&\sum_{S\subseteq\{2, \hdots, L\}} {\mathbb{P}}(\barzzz{\star, -i}_i = \mathbf{z}^\star_i, \barzzz{S\cup\{1\}, -i}_i = -\zzz{S\cup\{1\}}_i, \barzzz{(S\cup\{1\})^c, -i}_i = \zzz{(S\cup\{1\})^c}_i, E_i\cap F_i)\\
& \leq \sum_{S\subseteq\{2, \hdots, L\}}
{\mathbb{P}}\bigg\{ \log\bigg(\frac{1-\rho}{\rho}\bigg) \cdot \bigg( \#\{\ell\in S\cup\{1\}: \mathbf{z}^\star_i = -\zzz{\ell}_i\}-\#\{\ell\in S\cup\{1\}: \mathbf{z}^\star_i = \zzz{\ell}_i\} \bigg) \\
& \qquad\qquad \qquad \qquad + \sum_{\ell\in S\cup\{1\}} \sum_{j\neq i: \tzzz{\star, -i}_j = -\zzz{\ell}_i} \bigg[\log\bigg(\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg) \AAA{\ell}_{ij} + \log \bigg(\frac{1-p_\ell}{1-q_\ell}\bigg)\bigg] \\
&\qquad\qquad \qquad \qquad - \sum_{\ell\in S\cup\{1\}} \sum_{j\neq i: \tzzz{\star, -i}_j = \zzz{\ell}_i} \bigg[\log\bigg(\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg) \AAA{\ell}_{ij} + \log \bigg(\frac{1-p_\ell}{1-q_\ell}\bigg)\bigg] \geq 0 \textnormal{ and } E_i\cap F_i
\bigg\},
\end{align*}
where the inequality is by the fact that the occurrence of the event in each summand of the left-hand side above implies
$$
\sum_{\ell\in S\cup\{1\}} \fff{\ell}_i(\mathbf{z}^{\star}_i, -\zzz{\ell}_i, \tzzz{\star, -i}) + \sum_{\ell \in (S\cup\{1\})^c} \fff{\ell}_i(\mathbf{z}^{\star}_i, \zzz{\ell}_i , \tzzz{\star, -i}) \geq \sum_{\ell\in [L]} \fff{\ell}_i(\mathbf{z}^{\star}_i, \zzz{\ell}_i, \tzzz{\star, -i}).
$$
Since $E_i, F_i$ are both independent of $\{\zzz{\ell}_i\}_{\ell=1}^L$, we can do the following decomposition:
\begin{align*}
&\sum_{S\subseteq\{2, \hdots, L\}} {\mathbb{P}}(\barzzz{\star, -i}_i = \mathbf{z}^\star_i, \barzzz{S\cup\{1\}, -i}_i = -\zzz{S\cup\{1\}}_i, \barzzz{(S\cup\{1\})^c, -i}_i = \zzz{(S\cup\{1\})^c}_i, E_i\cap F_i)\\
& \leq \sum_{S\subseteq\{2, \hdots, L\}}
\sum_{\xi\in\{\pm 1\}^{S\cup\{1\}}} {\mathbb{P}}(\zzz{S\cup\{1\}}_i = \mathbf{z}^\star_i \xi)
\times {\mathbb{P}} \big(\mathscr{F}_{S, \xi}\big)\\
& = \sum_{S\subseteq\{2, \hdots, L\}}
\sum_{\xi\in\{\pm 1\}^{S\cup\{1\}}} {\mathbb{P}}(\zzz{S\cup\{1\}}_i = \mathbf{z}^\star_i \xi) \times {\mathbb{E}}_{\{\zzz{\ell}_{-i}\}_{\ell=1}^L}\bigg[{\mathbb{P}}\bigg( \mathscr{F}_{S, \xi} ~\bigg|~ \{\zzz{\ell}_{-i}\}_{\ell=1}^L \bigg)\bigg],
\end{align*}
where
\begin{align*}
\mathscr{F}_{S, \xi} &:= \bigg\{ \log\bigg(\frac{1-\rho}{\rho}\bigg) \cdot \bigg( \#\{\ell\in S\cup\{1\}: \xi_\ell = -1\}-\#\{\ell\in S\cup\{1\}: \xi_\ell = 1\} \bigg) \\
& \qquad \qquad + \sum_{\ell\in S\cup\{1\}} \sum_{j\neq i: \tzzz{\star, -i}_j = -\xi_\ell \mathbf{z}^\star_i} \bigg[\log\bigg(\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg) \AAA{\ell}_{ij} + \log \bigg(\frac{1-p_\ell}{1-q_\ell}\bigg)\bigg] \\
&\qquad\qquad - \sum_{\ell\in S\cup\{1\}} \sum_{j\neq i: \tzzz{\star, -i}_j = \xi_\ell\mathbf{z}^\star_i} \bigg[\log\bigg(\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg) \AAA{\ell}_{ij} + \log \bigg(\frac{1-p_\ell}{1-q_\ell}\bigg)\bigg] \geq 0 \textnormal{ and } E_i\cap F_i \bigg\}.
\end{align*}
By Markov's inequality, we have
\begin{align*}
&{\mathbb{P}}\bigg( \mathscr{F}_{S, \xi} ~\bigg|~ \{\zzz{\ell}_{-i}\}_{\ell=1}^L \bigg) \\
& \leq {\mathbb{E}}\bigg[ \exp\bigg\{t \log \bigg(\frac{1-\rho}{\rho}\bigg)
\cdot \bigg(\# \{\ell\in S\cup\{1\}: \xi_\ell = -1\} - \#\{\ell\in S\cup\{1\}: \xi_\ell = 1\}\bigg)\\
& \qquad + t \sum_{\ell\in S\cup\{1\}} \sum_{j\neq i: \tzzz{\star, -i}_j = -\xi_\ell \mathbf{z}^\star_i} \bigg[\log\bigg(\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg) \AAA{\ell}_{ij} + \log \bigg(\frac{1-p_\ell}{1-q_\ell}\bigg)\bigg]\\
& \qquad - t \sum_{\ell\in S\cup\{1\}} \sum_{j\neq i: \tzzz{\star, -i}_j = \xi_\ell\mathbf{z}^\star_i}\bigg[ \log\bigg(\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg) \AAA{\ell}_{ij} + \log \bigg(\frac{1-p_\ell}{1-q_\ell}\bigg)\bigg]
\bigg\} \cdot \mathbf{1}_{E_i\cap F_i}
~\bigg|~ \{\zzz{\ell}_{-i}\}_{\ell=1}^L \bigg],
\end{align*}
where the value of $t$ will be determined later. Using the same arguments as those that give rise to \eqref{eq:nodewise_err_decomp_3terms_simplified}, we get
\begin{align*}
& {\mathbb{E}}\bigg[ \exp\bigg\{ t \sum_{\ell\in S\cup\{1\}} \sum_{j\neq i: \tzzz{\star, -i}_j = -\xi_\ell \mathbf{z}^\star_i} \bigg[\log\bigg(\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg) \AAA{\ell}_{ij} + \log \bigg(\frac{1-p_\ell}{1-q_\ell}\bigg)\bigg]\\
& \qquad - t \sum_{\ell\in S\cup\{1\}} \sum_{j\neq i: \tzzz{\star, -i}_j = \xi_\ell\mathbf{z}^\star_i} \bigg[\log\bigg(\frac{p_\ell(1-q_\ell)}{q_\ell(1-p_\ell)}\bigg) \AAA{\ell}_{ij} + \log \bigg(\frac{1-p_\ell}{1-q_\ell} \bigg\}\bigg]
\cdot \mathbf{1}_{E_i\cap F_i}
~\bigg|~ \{\zzz{\ell}_{-i}\}_{\ell=1}^L\bigg] \\
&\leq \prod_{\ell\in S\cup\{1\}} \exp\bigg\{-\frac{n}{2}\cdot \III{\ell}_t + o(1)\cdot n\III{\ell}_{1/2}\bigg\}.
\end{align*}
Hence, we arrive at
\begin{align*}
&\sum_{S\subseteq\{2, \hdots, L\}} {\mathbb{P}}(\barzzz{\star, -i}_i = \mathbf{z}^\star_i, \barzzz{S\cup\{1\}, -i}_i = -\zzz{S\cup\{1\}}_i, \barzzz{(S\cup\{1\})^c, -i}_i = \zzz{(S\cup\{1\})^c}_i, E_i\cap F_i)\\
& \leq \sum_{S\subseteq \{2, \hdots, L\}} \sum_{\xi\in\{\pm 1\}^{S\cup\{1\}}} {\mathbb{P}}(\zzz{S\cup\{1\}}_i = \mathbf{z}^\star_i \xi)\\
&\qquad \times \exp\bigg\{t_{S, \xi} \log \bigg(\frac{1-\rho}{\rho}\bigg)
\cdot \bigg(\# \{\ell\in S\cup\{1\}: \xi_\ell = -1\} - \#\{\ell\in S\cup\{1\}: \xi_\ell = 1\}\bigg) \bigg\}\\
\label{eq:ind_nodewise_err_decomp_indinfo}
&\qquad \times \exp\bigg\{-\frac{n}{2}\sum_{\ell\in S\cup\{1\}} \III{\ell}_{t_{S, \xi}} + o(1)\cdot n\sum_{\ell\in S\cup\{1\}}\III{\ell}_{1/2} \bigg\}.\addtocounter{equation}{1}\tag{\theequation}
\end{align*}
We discuss according to two cases.
~\\
{\noindent \bf Case A: either $\boldsymbol{|S\cup\{1\}|}$ is even, or $\boldsymbol{\log e |S\cup\{1\}| \geq \sqrt{J_\rho}}$.} In this case, we choose $t_{S, \xi} = 1/2$. Then for any fixed $S$ satisfying the assumptions made in Case A, the corresponding summand in the right-hand side of \eqref{eq:ind_nodewise_err_decomp_indinfo} becomes
\begin{align*}
&{\mathbb{E}}_{\{\zzz{\ell}_i: \ell\in S\cup\{1\}\}} \bigg[\exp\bigg\{\log\sqrt{\frac{1-\rho}{\rho}}\cdot \bigg( \# \{\ell\in S\cup\{1\}: \zzz{\ell}_i = -\mathbf{z}^\star_i\} - \#\{\ell\in S\cup\{1\}: \zzz{\ell}_i = \mathbf{z}^\star_i\} \bigg)\bigg\}\bigg] \\
& \qquad \times \exp\bigg\{-\frac{(1+o(1))\cdot n}{2}\sum_{\ell\in S\cup\{1\}}\III{\ell}_{1/2} \bigg\} \\
\label{eq:ind_nodewise_err_caseA}
& = \exp\bigg\{ -|S\cup\{1\}|J_\rho - \frac{(1+o(1)) \cdot n}{2}\sum_{\ell\in S\cup\{1\}}\III{\ell}_{1/2}\bigg\}\addtocounter{equation}{1}\tag{\theequation}
\end{align*}
If $|S\cup\{1\}|$ is even, then the right-hand side above is
$$
\exp\bigg\{-(1+o(1))\bigg( |S\cup\{1\}| J_\rho - \psi^\star_{S\cup\{1\}}(0)\bigg)\bigg\} = e^{-(1+o(1))\mathcal{J}_{S\cup\{1\}}}.
$$
If $|S\cup\{1\}|$ is odd but $\log e|S\cup\{1\}| \geq \sqrt{J_\rho}$, we have
\begin{align*}
& (|S\cup\{1\}| + 1) J_\rho + \sup_{0\leq t\leq 1}\bigg\{-2tJ_\rho + \frac{n}{2}\sum_{\ell\in S\cup\{1\}} \III{\ell}_t\bigg\} \\
&\leq (|S\cup\{1\}| + 1) J_\rho + \frac{n}{2} \sum_{\ell\in S\cup\{1\}} \III{\ell}_{1/2}\\
& \leq |S\cup\{1\}| J_\rho \cdot ( 1 + e^{-\sqrt{J_\rho} +1} )+ \frac{n}{2}\sum_{\ell\in S\cup\{1\}} \III{\ell}_{1/2} \\
& \leq (1+o(1)) \bigg(|S\cup\{1\}|J_\rho + \frac{n}{2}\sum_{\ell\in S\cup\{1\}} \III{\ell}_{1/2}\bigg),
\end{align*}
where the first inequality is by Lemma \ref{lemma:monotonicity_of_I_t}, the second inequality is by $\log e|S\cup\{1\}|\geq \sqrt{J_\rho}$ and the third inequality is by $J_\rho \to \infty$ as $n\to\infty$. So we can upper bound the right-hand side of \eqref{eq:ind_nodewise_err_caseA} by
$$
\exp\bigg\{ (1+o(1))\bigg( (|S\cup\{1\}| + 1) J_\rho + \psi^\star_{S\cup\{1\}}(-2J_\rho)\bigg)\bigg\} = e^{-(1+o(1))\mathcal{J}_{S\cup\{1\}}}.
$$
In summary, for any fixed $S$ satisfying the assumptions made in Case A, the corresponding summand in the right-hand side of \eqref{eq:ind_nodewise_err_decomp_indinfo} can be upper bounded by
$
e^{-(1+o(1))\mathcal{J}_{S\cup\{1\}}}.
$
~\\
{\noindent \bf Case B: $\boldsymbol{|S\cup\{1\}|}$ is odd, and $\boldsymbol{\log e|S\cup\{1\}| \leq \sqrt{J_\rho}}$.} In this case, we re-write each summand in the right-hand side of \eqref{eq:ind_nodewise_err_decomp_indinfo} by
\begin{align*}
&\sum_{\xi\in\{\pm 1\}^{S\cup\{1\}}} {\mathbb{P}}(\zzz{S\cup\{1\}}_i = \mathbf{z}^\star_i \xi)\\
&\qquad \times \exp\bigg\{t_{S, \xi} \log \bigg(\frac{1-\rho}{\rho}\bigg)
\cdot \bigg(\# \{\ell\in S\cup\{1\}: \xi_\ell = -1\} - \#\{\ell\in S\cup\{1\}: \xi_\ell = 1\}\bigg) \bigg\}\\
&\qquad \times \exp\bigg\{-\frac{n}{2}\sum_{\ell\in S\cup\{1\}} \III{\ell}_{t_{S, \xi}} + o(1)\cdot n\sum_{\ell\in S\cup\{1\}}\III{\ell}_{1/2} \bigg\}\\
& = \sum_{x\in\{-|S\cup\{1\}| + 2k: 0\leq k\leq |S\cup\{1\}|\}} \binom{|S\cup\{1\}|}{\frac{|S\cup\{1\}| + x}{2}} (1-\rho)^{(|S\cup\{1\}| + x)/2} \rho^{(|S\cup\{1\}| - x)/2} \\
& \qquad \times \exp\bigg\{ -2x t_{x, S}\log \sqrt{\frac{1-\rho}{\rho}}- \frac{n}{2}\sum_{\ell\in S\cup\{1\}} \III{\ell}_{t_{x, S}} + o(1)\cdot n\sum_{\ell\in S\cup\{1\}} \III{\ell}_{1/2}\bigg\}.
\end{align*}
Using similar arguments as those that give rise to \eqref{eq:glob_nodewise_err_caseB}, we can bound the right-hand side above by
$$
\exp\bigg\{ -(1+o(1))\bigg((|S\cup\{1\}| + 1) J_\rho + \psi^\star_{S\cup\{1\}}(-2J_\rho)\bigg)\bigg\} = e^{-(1+o(1))\mathcal{J}_{S\cup\{1\}}}.
$$
~\\
{\bf \noindent Finishing the proof of Proposition \ref{prop:refine_ind_nodewise}}. The proof is concluded by combining Case A and B above.
\subsection{Proof of Theorem \ref{thm:specc}}\label{prf:thm:specc}
We start by stating a structural lemma, which relates the misclustering error to the deviation of the trimmed weighted adjacency matrix $\tau(\bar A)$ from the expectation of ${\bar{A}}$.
\begin{lemma}\label{lem: structure lem}
Assume there exists a constant $c\in[0, 1)$ such that
\begin{equation}
\label{eq:spectral_gap_structral_lemma}
\frac{\bar p}{\bar p - \bar q} \leq c\cdot \frac{2(1-2\rho)^2}{\beta - \beta^{-1}} + 2(\rho - \rho^2).
\end{equation}
Then, for any instance generated by an $\textnormal{IMLSBM}\in {\mathcal{P}}_n(\rho, \{p_\ell\}_1^L, \{q_\ell\}_1^L, \beta)$, the output of Algorithm \ref{alg: specc}, $\mathbf{\tilde z}^\star$, will satisfy
\begin{equation}
\label{eq: mistakes_of_spec_clustering}
{\mathcal{L}}(\mathbf{\tilde z}^\star, \mathbf{z}^\star)
\leq C\cdot \frac{ (2+\varepsilon)}{n^2 (1-2\rho)^4} \cdot \frac{(\|\tau(\bar A) - {\mathbb{E}} {\bar{A}} \|_2+\Delta)^2}{(\bar p-\bar q) ^2},
\end{equation}
where $\Delta=\bar p-2\rho(1-\rho)(\bar p-\bar q)$ and $C$ is an absolute constant only depending on $c$.
\end{lemma}
The proof of the above lemma uses the following result, which relates the misclustering error to the geometry of the point cloud.
\begin{lemma
\label{lemma: geometry}
Let $\mathbf{z}^{\star}\in\{\pm1\}^n$ be the global parameter for an IMLSBM~in the parameter space defined in \eqref{eq:param_space}. Suppose there exists a matrix $V\in{\mathbb{R}}^{n\times 2}$ (to be thought as the ``ground truth Euclidean embedding'' of the nodes) and a constant $b>0$ such that
\[
\min_{\mathbf{z}^{\star}_i\neq \mathbf{z}^{\star}_j}\|V_{i\bigcdot}-V_{j\bigcdot}\|\geq 2b,
\]
where $V_{i\bigcdot}\in{\mathbb{R}}^2$ is the $i$-th row of $V$.
Then, for any estimator $\mathbf{\tilde z}^{\star}$, any $\{\tilde v_1, \tilde v_2\} \subset {\mathbb{R}}^2$ (to be thought as the ``estimated centroids'' of the nodes), and any $\tilde V\in {\mathbb{R}}^{n\times 2}$ satisfying $\tilde V_{i\bigcdot} = \tilde v_{\mathbf{\tilde z}^{\star}_i}$ (to be thought as the ``estimated Euclidean embedding'' of the nodes), we have
\begin{align*}
d_\textnormal{H}(\mathbf{\tilde z}^\star, \mathbf{z}^\star) \land d_\textnormal{H}(-\mathbf{\tilde z}^\star, \mathbf{z}^\star) \leq C \cdot \# {\mathcal{S}},
\end{align*}
where ${\mathcal{S}}:=\{i\in[n]:\|\tilde{V}_{i\bigcdot}-V_{i\bigcdot}\|\geq b \}$, $\# {\mathcal{S}}$ is the cardinality of ${\mathcal{S}}$, and $C$ is an absolute constant.
\end{lemma}
\begin{proof}
This is Lemma 5 in \cite{gao2018community}. See also the proof of Theorem 2 in \cite{chen2018convexified}.
\end{proof}
We now present the proof of Lemma \ref{lem: structure lem}.
\begin{proof}[Proof of Lemma \ref{lem: structure lem}]
Note that for $\mathbf{z}_i^\star=\mathbf{z}_j^\star$, marginally we have ${\mathbb{P}}({A^{(\ell)}_{ij}=1})=[\rho^2+(1-\rho)^2]{p_{\ell}}+2\rho(1-\rho){q_{\ell}}={p_{\ell}}-2({p_{\ell}}-{q_{\ell}})(\rho-\rho^2)$. Similarly, for $\mathbf{z}_i^\star\neq\mathbf{z}_j^\star$, we have ${\mathbb{P}}({A^{(\ell)}_{ij}=1})={q_{\ell}}+2({p_{\ell}}-{q_{\ell}})(\rho-\rho^2)$.
So marginally, we have
$$
A^{(\ell)}_{ij}\sim\text{Bern}({\tilde p}_{\ell}\mathbf{1}{\mathbf{z}_i^\star=\mathbf{z}_j^\star}+{\tilde q}_{\ell}\mathbf{1}{\mathbf{z}_i^\star\neq\mathbf{z}_j^\star}),
$$
where
$$
{\tilde p_{\ell}} = {p_{\ell}}-2({p_{\ell}}-{q_{\ell}})(\rho-\rho^2) , \qquad {\tilde q_{\ell}} = {q_{\ell}}+2({p_{\ell}}-{q_{\ell}})(\rho-\rho^2).
$$
Let $Z$ be the $n\times 2$ assignment matrix such that $Z_{i,1}=\indicator{\mathbf{z}^\star_i=+1}$ and $Z_{i, 2} = \indicator{\mathbf{z}^\star_i = -1}$. Then $Z^\intercal Z=\mathrm{diag}(n^{\star}_+, n^{\star}_-)$, where $n^{\star}_\pm \in[\frac{n}{2\beta}, \frac{n\beta}{2}]$ are the sizes of the two communities. Define
\begin{align*}
B^{(\ell)}=\begin{pmatrix}
{\tilde p_{\ell}} & {\tilde q_{\ell}} \\ {\tilde q_{\ell}} & {\tilde p_{\ell}}
\end{pmatrix}.
\end{align*}
One readily checks that ${\mathbb{E}} \AAA{\ell}=Z B^{(\ell)} Z^\intercal - \mathrm{diag}({\tilde p_{\ell}})$, and thus ${\mathbb{E}}\bar A=Z \bar B Z^\intercal-\mathrm{diag}(\sum_\ell\omega_\ell{\tilde p_{\ell}})$, where $\bar B=\sum_\ell\omega_\ell B^{(\ell)}$. We let $P=Z\bar B Z^\intercal$, and $G=Z(Z^\intercal Z)^{-1/2}$. Then $G$ is orthonormal and $P=G X G^\intercal$, where $X= (Z^\intercal Z)^{1/2} \bar B (Z^\intercal Z)^{1/2}$. Let $X=WDW^\intercal$ be the eigen-decomposition of $X$. Then $P$ has the eigen-decomposition $P=VDV^\intercal$ with $V=GW=Z(Z^\intercal Z)^{-1/2}W$. Note that $W$ is a $2\times 2$ orthogonal matrix and $(Z^\intercal Z)^{-1/2}=\mathrm{diag}[(n^{\star}_+)^{-1/2},(n^{\star}_-)^{-1/2}]$. So if $Z_{i\bigcdot}\neq Z_{j\bigcdot}$, we have $\|V_{i\bigcdot}-V_{j\bigcdot}\|=\|(n^{\star}_+)^{-1/2}W_{1\bigcdot}-(n^{\star}_-)^{-1/2}W_{2\bigcdot}\|=\sqrt{(n^{\star}_+)^{-1}+(n^{\star}_-)^{-1}}$.
We now invoke Lemma \ref{lem: structure lem} with the ``estimated embedding'' $\tilde V$ being the output of Algorithm \ref{alg: specc} (i.e., the solution of the $(1+\varepsilon)$-approximate $k$-means clustering), and with the ``ground truth'' embedding being $V O$, where $O\in{\mathbb{R}}^{2\times 2}$ is an arbitrary orthonormal matrix. Since $\|(VO)_{i\bigcdot}-(VO)_{j\bigcdot}\|=\sqrt{(n^{\star}_+)^{-1}+(n^{\star}_-)^{-1}}$, we can take $b=\frac{1}{2}\sqrt{(n^{\star}_+)^{-1}+(n^{\star}_-)^{-1}}$.
We bound $\# {\mathcal{S}}=\#\{i\in[n]:\|\tilde{V}_{i\bigcdot}-(VO)_{i\bigcdot}\|\geq b \}$ as follows:
\begin{align*}
\# {\mathcal{S}}&\leq \frac{1}{b^2} \sum_{i\in{\mathcal{S}}}\|\tilde{V}_{i\bigcdot}-(VO)_{i\bigcdot}\|^2 \leq \frac{1}{b^2}\sum_{i=1}^n\|\tilde{V}_{i\bigcdot}-(VO)_{i\bigcdot}\|^2=\frac{1}{b^2}\|\tilde{V}-VO\|_F^2
\end{align*}
Since $\tilde V$ solves the $(1+\varepsilon)$ $k$-means clustering objective, the above display can be further bounded by
\begin{equation}
\label{equ: bound S}
\# {\mathcal{S}} \leq \frac{2}{b^2}(\|\tilde V-U\|_F^2+\|U-VO\|_F^2) \leq \frac{2}{b^2}(2+\varepsilon)\|U-VO\|_F^2=\frac{8(2+\varepsilon)}{(n^{\star}_+)^{-1}+(n^{\star}_-)^{-1}}\|U-VO\|_F^2,
\end{equation}
where we recall that $U$ consists of the top two eigenvectors of the trimmed weighted adjacency matrix $\tau({\bar{A}})$.
Since $V$ consists of the two leading eigenvectors of $P$, we can invoke Davis-Kahan Sin-$\Theta$ theorem (specifically, the version proved in Lemma 5.1 of \cite{lei2015consistency}) to conclude that, if the smallest eigenvalue $\gamma_n$ of $P$ is strictly positive, we will have
\begin{align*}
\# {\mathcal{S}}\leq \frac{8(2+\varepsilon)}{(n^{\star}_+)^{-1}+(n^{\star}_-)^{-1}}\|U-VO\|_F^2\leq 32(2+\varepsilon)\frac{1}{(n^{\star}_+)^{-1}+(n^{\star}_-)^{-1}}\frac{\|\tau(\bar{A})-P\|^2}{\gamma_n^2}
\addtocounter{equation}{1}\tag{\theequation}
\label{equ: bound Frobenius norm}
\end{align*}
Note that $\gamma_n$ is also the smallest singular value of
\[
X=(Z^\intercal Z)^{1/2}\bar B(Z^\intercal Z)^{1/2}=\begin{pmatrix}
n^{\star}_+\bar B_{11} & \sqrt{n^{\star}_+ n^{\star}_-}\bar B_{12}\\ \sqrt{n^{\star}_+ n^{\star}_-}\bar B_{21} & n^{\star}_-\bar B_{22}
\end{pmatrix}.
\]
Solving the characteristic polynomial, one finds that
\begin{align*}
2\gamma_n & = {n \sum_{\ell}\omega_\ell \tilde p_\ell - \sqrt{(n^{\star}_+ - n^{\star}_-)^2 (\sum_{\ell}\omega_\ell \tilde p_\ell)^2 + 4 n^{\star}_+ n^{\star}_- (\sum_{\ell} \omega_\ell \tilde q_\ell)^2}} \\
& \geq n \sum_{\ell}\omega_\ell \tilde p_\ell - |n^{\star}_+ - n^{\star}_-| \sum_{\ell}\omega_\ell \tilde p_\ell - n\sum_\ell \omega_\ell \tilde q_\ell,
\end{align*}
where the last inequality follows from $\sqrt{a+b}\leq \sqrt{a}+\sqrt{b}$ for $a, b\geq 0$ and $n^{\star}_+n^{\star}_-\leq n^2/4$. Since $n^{\star}_\pm \in [\frac{n}{2\beta}, \frac{\beta n}{2}]$, we can further lower bound the right-hand side above by
\begin{align*}
2\gamma_n & \geq n \sum_\ell \omega_\ell (\tilde p_\ell - \tilde q_\ell) - \frac{n}{2} (\beta - \beta^{-1}) \sum_{\ell} \omega_\ell \tilde p_\ell.
\end{align*}
We claim that if \eqref{eq:spectral_gap_structral_lemma} holds, then
$$
\frac{n}{2}(\beta -\beta^{-1}) \sum_{\ell} \omega_\ell \tilde p_\ell \leq c n \sum_{\ell} \omega_\ell (\tilde p_\ell - \tilde q_\ell).
$$
Indeed, since
$$
\sum_{\ell} \omega_\ell(\tilde p_\ell - \tilde q_\ell) = (\bar p - \bar q) - 4(\bar p - \bar q)(\rho - \rho^2) = (1-2\rho)^2 (\bar p - \bar q),
$$
one readily checks that
$$
(\beta - \beta^{-1}) (\bar p - 2(\bar p - \bar q) (\rho-\rho^2)) \leq 2c(\bar p - \bar q) (1 -2\rho)^{2},
$$
and this is exactly \eqref{eq:spectral_gap_structral_lemma} by rearranging terms. Thus, we arrive at
\begin{equation}
\label{eq:smallest_eigenval}
\gamma_n \geq (1-c) n\sum_{\ell} \omega_\ell (\tilde p_\ell - \tilde q_\ell) = (1-c) n (1-2\rho)^2(\bar p - \bar q).
\end{equation}
By triangle inequality, we have
\begin{align*}
\|\tau({\bar{A}})-P\|&\leq \|\tau({\bar{A}})-{\mathbb{E}} {\bar{A}}\|+\|{\mathbb{E}}{\bar{A}}-P\| = \|\tau({\bar{A}})-{\mathbb{E}} {\bar{A}}\|+ \bar p - 2\rho(1-\rho)(\bar p -\bar q).
\end{align*}
Plugging the above display and \eqref{eq:smallest_eigenval} to \eqref{equ: bound Frobenius norm}, we get
$$
{\mathcal{L}}(\mathbf{\tilde z}^{\star}, \mathbf{z}^{\star}) \lesssim \frac{(2+\varepsilon)\big(\|\tau(\bar A) - {\mathbb{E}} \bar A\| + \bar p - 2\rho(1-\rho)(\bar p - \bar q)\big)^2}{(1-c)^2 n^2 (1-2\rho)^4 (\bar p - \bar q)^2},
$$
from which \eqref{eq: mistakes_of_spec_clustering} follows.
\end{proof}
In view of Lemma \ref{lem: structure lem}, what is left is to upper bound the deviation of $\tau({\bar{A}})$ from ${\mathbb{E}} \bar A$. Consider the following decomposition:
\begin{equation}
\label{eq: decomp_spec_clustering}
\| \tau(\bar A) - {\mathbb{E}}{\bar{A}} \|_2 \leq \|\tau(\bar A) - {\mathbb{E}}^{(1:L)} [\bar A] \|_2 + \| {\mathbb{E}}^{(1:L)}[\bar A] - {\mathbb{E}}[\bar A]\|_2,
\end{equation}
where we let ${\mathbb{E}}$ be the marginal expectation, and ${\mathbb{E}}^{(1:L)}$ be the expectation \emph{conditional} on the realization of $\{\zzz{\ell}\}_1^L$.
The right-hand side of \eqref{eq: decomp_spec_clustering} is the superposition of two terms: the first term is the deviation of $\tau(\bar A)$ from the \emph{conditional mean} of $\bar A$, and the second term is the deviation induced by label sampling.
~\\
{\bf \noindent Bounding the deviation from the conditional mean.}
Conditional on the realization of $\zzz{\ell}$'s, the layer-wise adjacency matrices become symmetric Bernoulli random matrices with independent entries (on the upper-triangular part). Thus, we can invoke the technical tools we developed in Appendix \ref{append:concentration} to get the following result.
\begin{lemma}[Bounding the deviation from the conditional mean]
\label{lemma:dev_from_cond_mean}
Let Assumption \ref{assump:balanced_wts} hold with $c_0 > 0, c_1 \geq 1$ and fix two constants $r\geq 1, \gamma > e^{c_1}$. Define $I:= \{i\in[n]: \sum_{j\in[n]} {\bar{A}}_{ij} > \gamma n \sum_{\ell\in[L]} \omega_\ell p_\ell\}$. We trim the entries of $\bar A$ in ${\mathcal{E}} = (I\times[n])\cup ([n]\times I)$, so that the resulting matrix $\tau({\bar{A}})$ is zero on ${\mathcal{E}}$. Then with probability at least $1-3n^{-r} - c_2^{-n}$, we have
$$
\|\tau({\bar{A}}) - {\mathbb{E}}^{(1:L) }\bar A\| \leq C \cdot \sqrt{n \sum_{\ell\in[L]} \omega_\ell^2 p_\ell},
$$
where $c_2$ only depends on $\gamma$, and $C$ only depends on $c_0, c_1, r$.
\end{lemma}
\begin{proof}
This follows from Corollary \ref{cor:concentration_trimmed_mat} with $d_\ell = np_\ell$.
\end{proof}
~\\
{\bf \noindent Bounding the deviation due to label sampling.} We have the following lemma.
\begin{lemma}[Bounding the deviation due to label sampling]
\label{lemma:cluster_flips}
Assume $L\lesssim n^{c'}$ for some $c'\geq 0$. For any $r\geq 1, c>0$, with probability at least $1-6n^{-r} - 2n^{-c}$, we have
\begin{equation*}
\| {\mathbb{E}}^{(1:L)} \bar A - {\mathbb{E}} {\bar{A}}\| \leq C \cdot \bigg[ \max_{\ell\in[L]}\{\omega_\ell (p_\ell - q_\ell)\} \cdot \bigg( L\rho + n\sqrt{L\rho} + \sqrt{n\log n}\bigg) + \rho(1-\rho) \sum_{\ell} \omega_\ell(p_\ell-q_\ell)\bigg],
\end{equation*}
where $C$ is an absolute constant only depending on $r, c, c'$.
\end{lemma}
\begin{proof}
Let
$
\MMM{\ell} := {\mathbb{E}}^{1:L}[\AAA{\ell}] - {\mathbb{E}} [\AAA{\ell}].
$
We are interested in bounding the spectral norm of
$
\bar M := \sum_{\ell\in[L]} \omega_\ell \MMM{\ell}.
$
Note that the diagonal element of $\MMM{\ell}$ is zero. Meanwhile, for $i\neq j$, we have
\begin{align*}
\MMM{\ell}_{ij}& = (p_\ell - \tilde p_\ell) \mathbf{1}{\zzz{\ell}_i = \zzz{\ell}_j, \mathbf{z}^\star_i = \mathbf{z}^\star_j} + (q_\ell - \tilde q_\ell)\mathbf{1}{\zzz{\ell}_i\neq \zzz{\ell}_j, \mathbf{z}^\star_i \neq \mathbf{z}^\star_j} \\
& \qquad + (p_\ell - \tilde q_\ell) \mathbf{1}{\zzz{\ell}_i= \zzz{\ell}_j, \mathbf{z}^\star_i \neq \mathbf{z}^\star_j} + (q_\ell - \tilde p_\ell) \mathbf{1}{\zzz{\ell}_i \neq \zzz{\ell}_j, \mathbf{z}^\star_i = \mathbf{z}^\star_j}\\
& = (p_\ell-q_\ell) (2\rho - 2\rho^2) \mathbf{1}{\zzz{\ell}_i = \zzz{\ell}_j, \mathbf{z}^\star_i = \mathbf{z}^\star_j} + (p_\ell - q_\ell)(-2\rho + 2\rho^2) \mathbf{1}{\zzz{\ell}_i\neq \zzz{\ell}_j, \mathbf{z}^\star_i \neq \mathbf{z}^\star_j}\\
& \qquad + (p_\ell-q_\ell) (1-2\rho+2\rho^2)\mathbf{1}{\zzz{\ell}_i= \zzz{\ell}_j, \mathbf{z}^\star_i \neq \mathbf{z}^\star_j} + (p_\ell - q_\ell)(-1+2\rho - 2\rho^2) \mathbf{1}{\zzz{\ell}_i \neq \zzz{\ell}_j, \mathbf{z}^\star_i = \mathbf{z}^\star_j} \\
& = (p_\ell-q_\ell) \bigg[\mathbf{1}{\mathbf{z}^\star_i = \mathbf{z}^\star_j} \cdot \bigg(2\rho(1-\rho) - \mathbf{1}{\zzz{\ell}_i\neq \zzz{\ell}_j}\bigg) - \mathbf{1}{\mathbf{z}^\star_i\neq \mathbf{z}^\star_j} \cdot \bigg( 2\rho(1-\rho) - \mathbf{1}{\zzz{\ell}_i = \zzz{\ell}_j} \bigg) \bigg].
\end{align*}
Thus, one readily checks that
$$
\MMM{\ell} = \frac{1}{2} (p_\ell - q_\ell) \bigg(\zzz{\ell}{\zzz{\ell}}^\top - (1-2\rho)^2\mathbf{z}^\star {\mathbf{z}^\star}^\top - 4\rho(1-\rho) I_n\bigg),
$$
where the $4\rho(1-\rho)I_n$ offsets the diagonal entries so that $\mathop{\text{diag}}(\MMM{\ell}) = 0$. We then have
\begin{align*}
\|\bar M\| & = \bigg\|\frac{1}{2} \sum_{\ell\in[L]} \omega_\ell(p_\ell - q_\ell) \bigg(\zzz{\ell}{\zzz{\ell}}^\top - (1-2\rho)^2\mathbf{z}^\star {\mathbf{z}^\star}^\top - 4\rho(1-\rho)I_n\bigg)\bigg\| \\
& \leq \frac{1}{2}\max_{\ell\in[L]} \{\omega_\ell (p_\ell -q_\ell)\} \cdot \bigg\| \sum_{\ell\in[L]} \zzz{\ell}{\zzz{\ell}}^\top - (1-2\rho)^2\mathbf{z}^\star {\mathbf{z}^\star}^\top\bigg\| + 2\rho(1-\rho) \sum_{\ell\in[L]} \omega_\ell(p_\ell -q_\ell).
\end{align*}
With some algebra, one can show that
\begin{align*}
& \bigg\| \sum_{\ell\in[L]} \zzz{\ell}{\zzz{\ell}}^\top - (1-2\rho)^2\mathbf{z}^\star {\mathbf{z}^\star}^\top\bigg\| \\
& = \bigg\|\sum_{\ell\in[L]} (\zzz{\ell}-{\mathbb{E}} \zzz{\ell})(\zzz{\ell}-{\mathbb{E}} \zzz{\ell})^\top + (\zzz{\ell}-{\mathbb{E}} \zzz{\ell}){\mathbb{E}}[\zzz{\ell}]^\top + {\mathbb{E}}[\zzz{\ell}](\zzz{\ell}-{\mathbb{E}} \zzz{\ell})^\top \bigg\| \\
& \leq \underbrace{\bigg\|\sum_{\ell\in[L]} (\zzz{\ell}-{\mathbb{E}} \zzz{\ell})(\zzz{\ell}-{\mathbb{E}} \zzz{\ell})^\top\bigg\|}_{\RN{1}} + \underbrace{2\sqrt{n} \bigg\|\sum_{\ell\in[L]} \zzz{\ell}- {\mathbb{E}}\zzz{\ell}\bigg\|}_{\RN{2}},
\end{align*}
where in the last inequality we have used $\|\zzz{\ell}\|= \sqrt{n}$. We now bound the two terms in the right-hand side above separately.
To bound Term \RN{1}, let us introduce
\[
Z :=
\begin{pmatrix}
(\zzz{1}-{\mathbb{E}} \zzz{1})^\top \\
(\zzz{2}-{\mathbb{E}} \zzz{2})^\top \\
\vdots \\
(\zzz{L}-{\mathbb{E}} \zzz{L})^\top
\end{pmatrix} \in {\mathbb{R}}^{L\times n}.
\]
Then, we can write Term \RN{1} as $\|Z^\top Z\|$.
Let $\BBB{\ell}_i = (1-\zzz{\ell}_i\mathbf{z}^\star_i)/2$, which is distributed as $\BBB{\ell}_i \sim \text{Bern}(\rho)$. Under current notations, we have
\[
\zzz{\ell}_i - {\mathbb{E}} \zzz{\ell}_i = -2 \mathbf{z}^\star_i (\BBB{\ell}_i - {\mathbb{E}} \BBB{\ell}_i).
\]
Let $B$ denote the $L\times n$ matrix with $B_{\ell,i}=B_i^{(\ell)}-{\mathbb{E}} B_i^{(\ell)}$. Then $B$ is a matrix with i.i.d. centered $\text{Bern}(\rho)$ entries. Because $Z_{\ell,i}=(\zzz{\ell}_i-{\mathbb{E}} \zzz{\ell}_i)=-2B_{\ell, i}\mathbf{z}^\star_i $, we have
$
-Z/2= B \cdot \mathrm{diag}(\mathbf{z}^\star).
$
Hence, we have
$$
\| Z\| \leq 2 \|B-{\mathbb{E}} B \|.
$$
We define the $(n+L)\times(n+L)$ matrix ${\tilde{B}}$ as
$$
\tilde B =
\begin{pmatrix}
\mathbf{0}_{L\times L} & B \\
B^\top & \mathbf{0}_{n\times n}
\end{pmatrix}.
$$
It is clear that $\|B - {\mathbb{E}} B \| = \| \tilde B - {\mathbb{E}}\tilde B \|$. Since ${\tilde{B}}$ is a symmetric square matrix with i.i.d. Bernoulli entries in its upper-triangular part, by Corollary \ref{cor:concentration_w/o_reg}, for any $r\geq 1$ and $c>0$, we have
$$
\|\tilde B - {\mathbb{E}} \tilde B \| \leq C \bigg(\sqrt{(n+L)\rho} + \sqrt{\log(n+L)}\bigg)
$$
with probability at least $1-3n^{-r} - n^{-c}$, where $C$ is a constant only depending on $r$ and $c$.
Thus, on the same high probability event, we have
$$
\RN{1} \leq \|Z\|^2 \leq C' \bigg((n+L)\rho + \log (n+L)\bigg)
$$
for another $C'$ which only depends on $r$ and $c$.
For Term \RN{2}, we use a similar trick:
$$
\|\sum_{\ell} \zzz{\ell}-{\mathbb{E}} \zzz{\ell} \|_2
=
\bigg\|
\begin{pmatrix}
\mathbf{0}_{n\times n} & \sum_{\ell} (\zzz{\ell}-{\mathbb{E}} \zzz{\ell}) \\
\sum_{\ell} (\zzz{\ell}-{\mathbb{E}} \zzz{\ell})^\top & \mathbf{0}_{1\times 1}
\end{pmatrix} \bigg\|,
$$
and the right-hand side above is bounded above by
$$
2 \cdot \|\hat B \| := 2 \cdot \bigg\|
\begin{pmatrix}
\mathbf{0}_{n\times n} & \sum_{\ell} (\BBB{\ell}-{\mathbb{E}} \BBB{\ell}) \\
\sum_{\ell} (\BBB{\ell}-{\mathbb{E}} \BBB{\ell})^\top & \mathbf{0}_{1\times 1}
\end{pmatrix} \bigg\|,
$$
where again $\BBB{\ell}_i$'s are i.i.d. $\text{Bern}(\rho)$ random variables. By Corollary \ref{cor:concentration_w/o_reg}, we have
$$
\|\hat B \| \leq C'' \bigg(\sqrt{n L\rho } + \sqrt{\log n}\bigg)
$$
with probability at least $1 - 3n^{-r} - n^{-c}$, where $C''$ only depends on $r$ and $c$. This means that on the same high probability event, we have
$$
\RN{2} \leq 4 C''\sqrt{n} \cdot (\sqrt{nL\rho} + \sqrt{\log n}).
$$
Combining the bound on $\RN{1}$ and $\RN{2}$, we conclude that with probability at least $1 - 6n^{-r} - 2n^{-c}$,
\begin{align*}
\|\bar M\| & \lesssim \max_{\ell\in[L]} \{\omega_\ell(p_\ell-q_\ell)\} \cdot (\RN{1} + \RN{2}) + \rho(1-\rho) \sum_{\ell\in[L]} \omega_\ell(p_\ell-q_\ell) \\
& \lesssim \max_{\ell\in[L]} \{\omega_\ell(p_\ell-q_\ell)\} \cdot \bigg((n+L)\rho + n\sqrt{L\rho} + \log(n+L) + \sqrt{n \log n}\bigg) + \rho(1-\rho) \sum_{\ell\in[L]} \omega_\ell(p_\ell-q_\ell).
\end{align*}
The desired result follows by noting that $n\rho \leq n\sqrt{L\rho}$ and $\log(n+L) \ll \sqrt{n\log n}$ when $L \lesssim n^{c'}$ for some $c'\geq 0$.
\end{proof}
~\\
{\bf \noindent Finishing the proof of Theorem \ref{thm:specc}.}
Theorem \ref{thm:specc} is a direct consequence of Lemma \ref{lem: structure lem}, \ref{lemma:dev_from_cond_mean} and \ref{lemma:cluster_flips}.
\subsection{#1}}
\newcommand{\sssec}[1]{\subsubsection{#1}}
\newcommand{\noindent \rule[0pt]{\textwidth}{2pt}}{\noindent \rule[0pt]{\textwidth}{2pt}}
\newcommand{\pth}[1]{\left( #1 \right)}
\newcommand{\qth}[1]{\left[ #1 \right]}
\newcommand{\sth}[1]{\left\{ #1 \right\}}
\newcommand{\bpth}[1]{\Bigg( #1 \Bigg)}
\newcommand{\bqth}[1]{\Bigg[ #1 \Bigg]}
\newcommand{\bsth}[1]{\Bigg\{ #1 \Bigg\}}
\newcommand{\bfloor}[1]{\Bigg\lfloor #1 \Bigg\rfloor}
\newcommand{\fracp}[2]{\frac{\partial #1}{\partial #2}}
\newcommand{\fracpk}[3]{\frac{\partial^{#3} #1}{\partial #2^{#3}}}
\newcommand{\fracd}[2]{\frac{{\rm d} #1}{{\rm d} #2}}
\newcommand{\fracdk}[3]{\frac{{\rm d}^{#3} #1}{{\rm d} #2^{#3}}}
\newcommand{\pushf}[2]{#1_{\sharp} #2}
\newcommand{\KL}[2]{D(#1 \, || \, #2)}
\newcommand{\Fisher}[2]{I(#1 \, || \, #2)}
\newcommand{{\sf TV}}{{\sf TV}}
\newcommand{{\sf LC}}{{\sf LC}}
\newcommand{d_{\sf H}}{d_{\sf H}}
\newcommand{\ell_{\max}}{\ell_{\max}}
\newcommand{\texttt{bits}\xspace}{\texttt{bits}\xspace}
\newcommand{{\texttt{nats}}}{{\texttt{nats}}}
\newcommand{{\mathsf{snr}}}{{\mathsf{snr}}}
\newcommand{{\sqrt{\mathsf{snr}}}}{{\sqrt{\mathsf{snr}}}}
\newcommand{{\mathsf{SNR}}}{{\mathsf{SNR}}}
\newcommand{{\mathsf{INR}}}{{\mathsf{INR}}}
\newcommand{{\mathsf{SIR}}}{{\mathsf{SIR}}}
\newcommand{\todolong}[1]{\todo[inline,color=red!40]{#1}}
\newcommand{\textbf{xxx}\xspace}{\textbf{xxx}\xspace}
\newcommand{A^{(\ell)}}{A^{(\ell)}}
\newcommand{\Rom}[1]{\uppercase\expandafter{\romannumeral #1\relax}}
\newcommand{\rom}[1]{\lowercase\expandafter{\romannumeral #1\relax}}
\newtheorem{assumpA}{A.}
\newtheorem{assumpAprime}{A$^\prime$}
|
\section{Introduction}
\label{sec:intro}
In recent years, the outskirts of galaxy clusters have emerged as one of the new frontiers for cosmology and astrophysics \citep[see][for review]{walker_etal19}.
Recent theoretical advances revealed that the physical outer boundary for a dark matter (DM) halo can be defined using by the ``splashback'' radius based on the DM density profile drop \citep[e.g.,][]{diemer_kravtsov2014,adhikari_etal2014,more_etal2015}, the aspherical splashback surface \citep{mansfield_etal17,Mansfield2020}, or the edge radius of the DM phase space structure \citep{Aung2020}, with various definitions encompass varying fraction of orbiting DM particles \citep{diemer_etal17}. Observationally, the outer boundaries of the DM haloes have recently been detected using weak-lensing \citep[e.g.,][]{chang_etal18}, galaxy number density \citep[e.g.,][]{more_etal2016,Baxter2017,Shin2019,Zurcher_More2019,Murata2020} and phase space structure \citep{Tomooka2020}. Upcoming multiwavelength surveys (such as CMB-S4 in microwave and DESI, Rubin, PFS in optical) will provide unprecedented insight on the outer boundaries of massive DM haloes and promise to shed new insight into cosmology and non-linear structure formation of the Universe.
Gas accreting at cluster outskirts provides an alternative probe of cluster boundary. However, the dynamics of the collisional gas is fundamentally different from that of collisionless DM. Unlike the collisionless DM particles which orbit within the DM halos, the collisional gas is shock heated during its first infall, resulting in a high Mach number ($\mathcal{M}>100$) cosmic accretion shock marked by the prominent entropy jump. The secondary infall model predicts that the location of the accretion shock coincides with the splashback radius \citep{bertschinger1985, shi2016b}. Commonly referred to as ``external shock'' in the literature \citep[e.g.,][]{Miniati2000,ryu_etal2003,Skillman2008, molnar_etal2009}, the accretion shock arises from the infall of low density pristine gas in the void regions onto the cluster potential (in contrast to ``internal shocks'' which occurs within the virialization region of DM haloes due to mergers and penetrating filaments). The external accretion shock thus defines a physical boundary of the hot collapsed gas in DM haloes, which is also dependent on their mass accretion rate (MAR) \citep{lau_etal2015}.
In this work we investigate the locations of shock and splashback radii by analyzing the {\em Omega500} hydrodynamical cosmological simulations. We find that the accretion shock radius defined using the drop in the gas entropy is larger than all definitions of the splashback radius in the literature by $20-100\%$, in contrast to the prediction of the self-similar models. Specifically, we find that the accretion shock radius is larger by $\approx 1.89$ relative to the splashback radius and $\approx 1.2$ relative to the edge radius of the DM phase space structure. Furthermore, we find that the ratios of the shock and splashback/edge radii are independent of halo mass and redshift, but dependent on their MAR.
We describe our simulations and analysis methods in \S\ref{sec:sim}. Results and discussions are presented in \S\ref{sec:results} and \S\ref{sec:discussions}, respectively. Conclusions are summarized in \S\ref{sec:conclusions}.
\section{Simulations}
\label{sec:sim}
We analyze the clusters from the {\em Omega500} simulation \citep{nelson_etal2014}, a high-resolution hydrodynamical simulation of a large cosmological volume
with the comoving box size of $500\,h^{-1}\,{\rm \, Mpc}$.
The simulation is performed using the Adaptive Refinement Tree (ART) $N$-body+gas-dynamics code \citep{kravtsov1999, kravtsov_etal2002, rudd_etal2008},
which is an Eulerian code that uses adaptive refinement in space and time,
and non-adaptive refinement in mass \citep{klypin_etal2001}
to achieve the dynamic ranges to resolve the cores of haloes formed in self-consistent cosmological simulations in a flat $\Lambda$CDM model with WMAP 5 years cosmological parameters:
$\Omega_m = 1 - \Omega_{\Lambda} = 0.27$, $\Omega_b = 0.0469$,
$h = 0.7$ and $\sigma_8 = 0.82$, where the Hubble constant is defined as
$100\,h\;{\rm \, km \, s}^{-1}\,{\rm \, Mpc}^{-1}$ and $\sigma_8$ is the mass variance within
spheres of radius $8\,h^{-1}{\rm \, Mpc}$.
Haloes are identified in the simulation using a spherical overdensity halo finder described in \cite{nelson_etal2014}. We select 65 haloes with mass $M_{\rm 500c} \geq 3\times10^{14}\,h^{-1}M_{\odot}$ at $z=0$ and re-simulate the box with the higher resolution DM particles in regions of the selected haloes, resulting in an effective mass resolution of $1.09 \times 10^9\, h^{-1}M_{\odot}$, which corresponds to $2048^{3}$ DM particles. We built the merger tree by tracking the most massive progenitors of haloes over time using the merger tree code presented in \citet{yu_etal2015}. This is done by following the $10\%$ most bound DM particles at each snapshot. We define the start of the merging process at the epoch when $R_{\rm 500c}$ of the two haloes start to overlap with each other. Following \citet{diemer_kravtsov2014}, we compute the MAR of DM haloes as
\begin{equation}
\Gamma_{\rm 200m} = \frac{\log_{10}M_{\rm 200m} (z) - \log_{10}M_{\rm 200m} (z')}{\log_{10}a(z)-\log_{10}a(z')} \; ,
\end{equation}
where $a$ is the expansion factor and $z=0$, $z'=0.5$, and $M_{\rm 200m}$ is the mass enclosed within the radius $R_{\rm 200m}$ such that the density enclosed is 200 times the mean density of the universe, $M_{\rm 200m} = 200\rho_m \times 4\pi R_{\rm 200m}^3/3$.
The simulation is performed on a uniform $512^3$ grid with 8 levels of mesh refinement, implying a maximum comoving spatial resolution of $3.8\,h^{-1} {\rm \, kpc}$. The spatial resolution is controlled by the density of the cells, and the maximum comoving resolution is only achieved at the centre of the haloes. However, the spatial resolution near the shock radius is between $0.03\,h^{-1}{\rm \, Mpc}$ and $0.12\,h^{-1}{\rm \, Mpc}$, which is sufficient to determine the locations of the edge and shock radii (which is typically of order several Mpc in size) with the accuracy better than 5\%. With a typical number density of $n_e\approx 10^{-3}{\rm \, cm}^{-3}$ and temperature of $T\approx 10^6{\rm \, K}$, the mean free path of electron is much smaller than the resolution.
Since the effects of non-gravitational baryonic physics (such as gas cooling and energy feedback from supernova and black holes) are small in cluster outskirts compared to cluster cores, we focus on analyzing the outputs of the non-radiative simulation for simplicity.
For completeness, we also checked the effects of baryonic physics by comparing the results to those of runs with cooling and star formation and AGN feedback.
\section{Results}
\label{sec:results}
\begin{figure}
\begin{center}
\includegraphics[scale=0.40]{./figures/shock_splashback_map_2.pdf}
\caption{Gas entropy (top left), DM density (top right), gas temperature (bottom left) and gas pressure (bottom right) maps of the simulated cluster (CL135) extracted from the non-radiative {\em Omega500} hydrodynamical cosmological simulation. The images are 15.625Mpc/h wide with the projection depth of 3.90625 Mpc/h. The inner dashed lines indicate the splashback shell computed using the method from \citet{mansfield_etal17}, whereas the outer white lines indicate the shock shell found by the discontinuous jump in entropy as well as pressure. Note that several low-entropy gas streams have penetrated inside of the accretion shock radius along the filaments without getting shock heated.
}
\label{fig:rsp_rsh_map}
\end{center}
\end{figure}
\subsection{Identifying Splashback and Shock Surface}
We determine the location of the splashback radius using the \textsc{Shellfish}\ code \citep{mansfield_etal17}. For each halo, the code draws $10^5$ random sight lines from the halo centre and samples the DM density along each line-of-sight (LOS). The splashback radius $R_{\rm sp}$ is defined as the radius of a spherical surface that encompasses the same volume as that enclosed by the surface of sharp DM density jumps in all LOS \citep[see][for more details]{mansfield_etal17}.
We determine the accretion shock radius in a similar manner using the \textsc{Shellfish}\ code. Namely, we draw $786$ LOS according to HEALPix \citep{Healpix} pixels from the halo centre and sample the gas entropy profile along each LOS. The 786 LOS chosen here corresponds to the fourth level resolution of HEALPix, and at the shock radius, each pixel corresponds to about $0.26\,h^{-2}{\rm \, Mpc}^2$, with a length resolution of $\approx 0.5\,h^{-1}{\rm \, Mpc}$, about 4 times the simulation resolution in the region. Measuring the shock radius with the fifth level resolution with 3072 LOS leads to a less than $3\%$ difference in shock measurements.
For each LOS, we select all the gas cells that each LOS passes through, and sample the gas entropy profile along each LOS. The profile is then smoothed with a Savitzky-Golay (SG) filter with window-length of 9 equally spaced logarithmic radial bins and a polynomial order of 5; we checked that the results are robust to the variation in the parameter for SG filter (in the window-length from 5 to 11 bins) and polynomial orders (in the range of 2 to 7). We define the location of the accretion shock as the radius of the minimum in the logarithmic entropy slope. We remove LOS where the entropy jump is less than a factor of 50 (corresponding to Mach number of $\mathcal{M} \sim 20$; the results are unchanged for $\mathcal{M}=10$ to $50$), as these directions host substructures and filamentary gas streams.
After these removals, the covering fraction of the shock surface of the total spherical area is approximately $80\%$ for all clusters.
We fit Penna-Dines function with $K=1,I=J=2$ \citep[corresponding to the first and second order expansion of $\cos\phi$ and $\sin\phi$ of polar angle $\phi$,][]{Penna_Dines2007}:
\begin{equation}\label{eq:penna_dines}
r(\theta,\phi) = \sum_{i,j,k=0}^{I,J,K} c_{\rm ijk} \sin^{i+j}\phi \cos^k\phi \sin^j\theta \cos^i\theta,
\end{equation}
to the shock position along each LOS and calculate the enclosed volume inside the fitted (non-spherical) surface. We then define the shock "radius" as the radius of a sphere that encompasses the same amount of volume as that enclosed within the fitted surface. We note that the total volume inside the shock radius is larger than the total volume of shock heated gas as it also includes volume of unshocked gas residing in filaments.
Figure~\ref{fig:rsp_rsh_map} shows the map of one of the haloes indicating the splashback and shock shells identified using the above algorithms. The splashback shell identified encompasses the DM structure, whereas the shock shell encompasses a much larger, extended area, where the entropy and pressure shows significant decline. We also note that there are several unshocked, low-entropy gas streams that have penetrated inside of the accretion shock radius of the halo along the filaments.
\subsection{Phase Space Structures of DM and Gas}\label{sec:phase_space}
\label{sec:dynamics}
\begin{figure}
\begin{center}
\includegraphics[width=0.49\textwidth,trim={10 1 0 1},clip]{./figures/phase_space.pdf}
\caption {Phase-space diagrams of DM ({\em top} panels) and gas ({\em bottom} panels) for a relaxed cluster CL135 ({\em left} panels) with low MAR ($\Gamma = 0.5$), and a merging cluster CL77 ({\em right} panels) with high MAR ($\Gamma = 2.9$). The solid, dashed, dotted, and dot-dashed lines represent $R_{\rm 200m}$, $R_{\rm sp}$ (splashback radius), $R_{\rm max,sp}$ (maximum splashback radius along any line-of-sight), and $R_{\rm sh}$ (shock radius), respectively. The colour represents the relative mass fraction of DM (top panels) and gas (bottom panels), with deeper colour indicating higher mass fraction at a given radius. The phase space structure of virialized DM haloes extends past $R_{\rm sp}$, reaching nearly $R_{\rm sh}$. Gas follows DM at $r\gtrsim R_{\rm sh}$, while gas dynamics differs significantly from that of collisionless DM at $r\lesssim R_{\rm sh}$.}
\label{fig:phase_diagram}
\end{center}
\end{figure}
Figure~\ref{fig:phase_diagram} shows phase-space densities of DM and gas for their radial velocity components as a function of radius, for a relaxed cluster (CL135) and a merging cluster (CL77), respectively. The average radial velocity of gas and DM is negative outside the shock radius ($R_{\rm sh}$) as they fall onto the cluster potential. The two trace each other as the gas pressure is low, rendering gas to behave similarly to collisionless DM. Gas infalling from the void is shock heated at the shock radius, causing gas to lose its kinetic energy into heating of the gas. Filamentary accretion can bring infalling gas inside of the virialization region of the DM haloes and form shocks with small Mach number ($\mathcal{M}<3$) before gas loses their kinetic energy. The DM particles, on the other hand, exchange energy through gravitational interactions as they orbit through the interior of the DM halo.
Within the splashback radius ($R_{\rm sp}$), a DM halo exhibits a typical virialized phase-space structure, where the splashback and orbital motions of DM particles produce a dispersion with zero mean radial velocity. The phase space structure of the DM halo can extend out to radii larger than the volume-averaged splashback radius as not all particles within the splashback surface are expected to lie within the volume-averaged splashback radius $R_{\rm sp}$ due to asphericity \citep{mansfield_etal17,diemer_etal17,Aung2020}. Even when using $R_{\rm max,sp}$ (the maximum radius of the splashback surface defined using the Penna-Dines approximation surface), a small amount (upto $1\%$) of orbiting particles can still exist beyond $R_{\rm max,sp}$ \citep{Mansfield2020} as shown in Figure~\ref{fig:phase_diagram}.
Due to the collisional nature of the gas, however, the phase-space distribution of gas differs significantly within the interior of the cluster. The radial velocity dispersion of gas is considerably smaller because, as the gas is shock heated through the accretion shock, where most of the gas kinetic energy is converted to thermal energy. Thus, inside $R_{\rm 200m}$, the level of gas motions inside the accretion shock remains small in absence of external disruption by mergers for CL135, while gas motions induced by mergers comprise of most of the velocity dispersion in CL77. The position of the shock radius is closer to the edge of DM phase space than the splashback radius.
\subsection{Shock Radii Determined from Profiles}
\begin{figure}
\begin{center}
\includegraphics[width=0.5\textwidth]{./figures/profiles_individual.pdf}
\caption{Spherically averaged gas density, DM density, volume-weighed entropy and pressure as a function of radius for the two different clusters: a relaxed cluster CL135 with low MAR ($\Gamma = 0.5$), and a merging cluster CL77 ({\em right} panels) with high MAR ($\Gamma = 2.9$). The vertical lines indicate splashback and shock radii based on the \textsc{Shellfish}\ and our method, respectively. The gas pressure and entropy profiles show significant decrease near the shock radii, while the gas density and DM density decreases at the splashback radii. The gas density slope is shallower than the DM density slope. The slower accreting halo (CL135) also has a smoother jump and larger shock and splashback radii than the fast accreting CL77.}
\label{fig:profiles}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=0.49\textwidth]{./figures/profiles_slope.pdf}
\caption{The logarithmic slope of different gas and DM median profiles as a function of radius. The radius where the minimum of gas entropy, temperature and pressure slope is defined as the shock radius $R_{\rm sh,sph}$ of the spherically averaged gas profile, denoted by dotted line. The radius where the minimum of DM density slope is defined as the splashback radius $R_{\rm sp,sph}$ of the spherically averaged profile, denoted by dashed line, and coincides with the minimum of the gas density slope. }
\label{fig:profiles_slope}
\end{center}
\end{figure}
Figure~\ref{fig:profiles} shows the spherically averaged DM density, gas density, and volume-weighed entropy and pressure profiles for two representative clusters in the sample, CL135 a relaxed cluster with low $\Gamma$, and CL77, a merging cluster with high $\Gamma$. We also overplot their splashback radii estimated from \textsc{Shellfish}\ $R_{\rm sp}$, and their shock radii from the shock surface $R_{\rm sh}$. Both $R_{\rm sp}$ and $R_{\rm sh}$ are smaller for the high $\Gamma$ cluster CL77 than than the low $\Gamma$ cluster CL135. In both clusters, there are sharp drops in the pressure and entropy profiles at the accretion shock, and they are particularly prominent in the high $\Gamma$ cluster CL77. However, the corresponding decrease in gas density profiles are small at the accretion shock. This is because the density contrast across a shock is intrinsically smaller than their counterparts in pressure and entropy, and is capped at a maximum value of $4$, as expected from the Rankine -- Hugoniot shock jump condition.
Figure~\ref{fig:profiles_slope} shows the logarithmic slope of the spherically averaged median DM density, gas density, temperature, pressure, and entropy profiles for all clusters at $z=0$. The entropy profile is increasing at all radii before the shock radius, and within $0.1\leq r/R_{\rm 200m}\leq 0.6$, it is consistent with previous findings of entropy slope of $1.1$ \citep[e.g.,][]{voit_etal2005} as expected from self-similar cluster growth. At the shock radius, entropy decreases sharply indicating a strong shock. The shock front, however, is wider and is not as abrupt as the shock front in LOS profile due to smoothing over aspherical shock fronts and variations among clusters. The pressure profile has the minimum slope about $-7$ at the shock radius. Inside the shock radius, the pressure profile is also rapidly decreasing, consistent with the universal pressure profile characterized by the generalized NFW profile \citep[e.g.,][]{nagai_etal2007a,arnaud_etal2010}. Similarly, the temperature profile has the minimum slope of about $-2$ at the shock radius.
The DM density slope follows NFW profile \citep{nfw1996} closely, where the slope is $-1$ in the inner region and slowly decreases to $-3$ in outer region before hitting minimum at the splashback radius. The gas density is much flatter in inner region starting with slope of $\approx 0$, but approaches NFW and follows DM profile at outer radii. In fact, gas density slope becomes minimum at DM splashback radius, while only showing mild decrease at the shock radius. The smaller decrease in density slope at the shock radius compared to other thermodynamic quantities is expected, as the density jump for a shock wave in ideal gas is capped at $4$, while there is no upper bound for the jumps in temperature, pressure or entropy.
For DM, the splashback radius identified as the steepest point in the spherically averaged density profile $R_{\rm sp, sph}$ is smaller than the splashback radius estimated from \textsc{Shellfish}\ $R_{\rm sp}$ \citep[see also][]{mansfield_etal17}.
For the accretion shock radius, the radius computed from spherically averaged profile $R_{\rm sh, sph}$, is the same as that identified from the volume-averaged shock surface $R_{\rm sh}$.
Figure~\ref{fig:mar_slope} shows the profile slopes for different $\Gamma$ samples. Here the splashback and shock radii are identified as the steepest jumps in DM density and gas entropy profiles, respectively. Both radii decrease for larger $\Gamma$. The DM density slope is steeper for larger $\Gamma$, consistent with the previous result \citep{diemer_kravtsov2014,more_etal2015}. The pressure jump is larger for larger $\Gamma$, indicating a stronger shock.
\begin{figure}
\begin{center}
\includegraphics[width=0.51\textwidth]{./figures/slope_mar.pdf}
\caption{The logarithmic slope of different gas and DM median profiles for different MAR. The MAR is split at the 33 and 66-percentile of all haloes, which results in the low $\Gamma_{\rm 200m}$ sample with $\Gamma_{\rm 200m}<1.5$ (dashed line) and high $\Gamma_{\rm 200m}$ sample with $\Gamma_{\rm 200m}>2.7$ (solid line). The splashback radius is where the total matter (DM+gas) density slopes is the smallest, and the shock radius where the minimum of entropy and pressure slope is smaller for the higher MAR haloes. }
\label{fig:mar_slope}
\end{center}
\end{figure}
\subsection{Offsets between Splashback and Shock Radii}
\begin{figure}
\begin{center}
\includegraphics[width=0.49\textwidth]{./figures/L500_rsp_gamma_updated}
\caption{The splashback radius $R_{\rm sp}$ (red points) and the accretion shock radius $R_{\rm sh}$ (blue points), normalized by the halo radius $R_{\rm 200m}$, plotted as a function of the MAR of the cluster-size DM haloes, $\Gamma_{\rm 200m}$.
The solid line is the $R_{\rm sp}-\Gamma_{\rm 200m}$ relation from DM-only cosmological simulation defined by \textsc{Shellfish}\ \citep[M17,][]{mansfield_etal17} and $R_{\rm sp,87}$ computing using SPARTA \citep{diemer_etal17}. The dashed lines represent the best-fitting $R_{\rm sp}-\Gamma_{\rm 200m}$ relation times the average $R_{\rm sp}/R_{\rm sp}$ ratios for three different baryonic simulations (magenta for AGN feedback simulation, cyan for cooling and star forming simulation, and yellow for non-radiative, embedded behind cyan line). Also shown is $r_{\rm edge}\approx 1.6R_{\rm sp,87}$, marking the edge of the DM phase space structure \citep{Aung2020}.
}
\label{fig:rsp_rsh}
\end{center}
\end{figure}
Figure~\ref{fig:rsp_rsh} shows the splashback radius $R_{\rm sp}$ normalized by the halo radius $R_{\rm 200m}$, plotted as a function of the MAR ($\Gamma_{\rm 200m}$) of haloes extracted from the non-radiative {\it Omega500} simulation. The splashback radius decreases with increasing MAR, confirming previous numerical \citep{diemer_kravtsov2014,mansfield_etal17} and analytic results \citep{adhikari_etal2014, shi2016a}.
The $R_{\rm sp}$ from our hydrodynamic simulation agrees well with the best-fitting relation from the DM-only simulation.
In the same panel, we show that $R_{\rm sh}/R_{\rm 200m}$ decreases with $\Gamma_{\rm 200m}$ in a similar manner to the $R_{\rm sp}$--$\Gamma_{\rm 200m}$ relation. The average ratio between the radii is $R_{\rm sh}/R_{\rm sp} = 1.89 \pm 0.16$ (based on the yellow dashed line in Figure~\ref{fig:rsp_rsh}, where the error indicates 1$\sigma$ scatter) at $z=0$, and it is only weakly dependent on MAR for the range probed here. At $z=1$, $R_{\rm sh}/R_{\rm sp} = 2.03 \pm 0.32$, and at $z=3$, $R_{\rm sh}/R_{\rm sp} = 2.12 \pm 0.35$ which is consistent with no evolution with redshift. We also find that the scatter in $R_{\rm sh}$ is larger than that of $R_{\rm sp}$.
We repeated the same analysis for the same sets of haloes with baryonic physics that include radiative cooling and star formation, and AGN feedback. The differences in $R_{\rm sp}$ and $R_{\rm sh}$ between simulations with different baryonic physics is $\lesssim 1$\%, thus both radii remain essentially unchanged in the presence of baryonic physics.
We note that the phase space structure of the DM halo can extend out to radii larger than the splashback radius, because not all particles within the splashback surface are expected to be enclosed within the volume-averaged $R_{\rm sp}$. In fact, the splashback radius from \textsc{Shellfish}\ only contains about $87\%$ of the particle apocentre \citep{diemer_etal17}. The edge radius, which marks the end of DM phase space, corresponds to the radius where the fraction of orbiting subhaloes is greater than 99\% (denoted as $R_{\rm sp,99}$), which is approximately $1.6$ times larger than the splashback radius measured with \textsc{Shellfish}\ (denoted as $R_{\rm sp,87}$) \citep{Aung2020}. This edge radius lies in the region in between the accretion shock and splashback surfaces. Specifically, the ratio of the shock and edge radius is $R_{\rm sh}/1.6R_{\rm sp,87} \approx 1.2$ for $1\leq \Gamma_{\rm 200m} \leq 4$, indicating that the shock radius is about 20$\%$ larger than the edge radius, on average.
\subsection{Shapes of Shock and Splashback Shells}\label{sec:apshericity}
\begin{figure}
\begin{center}
\includegraphics[width=0.49\textwidth]{./figures/rmaxmin_r200.pdf}
\caption{\change{The normalized distribution of maximum (solid) and minimum (dotted) radius of shock (blue) and splashback (red) surface with respect to $R_{\rm 200m}$. The shock and splashback shells are highly aspherical, causing the maximum and minimum radius along lines-of-sight to differ by more than $65\%$. Note that the overlap in the distributions of the minimum shock and maximum splashback radii is due to the scatter in $R_{\rm sh}/R_{\rm sp}$ ratio. For individual haloes, the shock radius is always larger than the splashback radius along every line-of-sight.
}}
\label{fig:rmaxmin}
\end{center}
\end{figure}
In practice, DM splashback and accretion shock are aspherical, because haloes form through merger and accretion of materials through cosmic web of filaments that are inherently aspherical, as shown in Figure~\ref{fig:rsp_rsh_map}. Figure~\ref{fig:rmaxmin} shows the distribution of the maximum ($R_{\rm max}$) and minimum ($R_{\rm min}$) splashback and shock shells. We find that the $R_{\rm max,sh}$ and $R_{\rm min,sh}$ of the shock shell can be $1.65$ times larger and $0.7$ times smaller compared to the volume-averaged shock radius, respectively, while the maximum and minimum distance of the splashback shell ranges between $(0.8-1.42)$ times the volume-averaged splashback radius. This shows that the accretion shock shells are generally more aspherical than the splashback shells, which cause the spherically averaged gas profiles to appear smoother than the actual accretion shocks. \change{Even though the minimum shock radii and maximum splashback radii overlap in Figure~\ref{fig:rmaxmin}, we emphasize that the largest radius of the splashback shell is still smaller than the smallest radius of the shock shell of the same halo for the individual haloes in our sample. The minimum shock to splashback radius ratio is found in the directions perpendicular to the axis of filament and merger (see the panels in the second and third rows of Figure~ \ref{fig:rsh_rsp_merge_map} where the ratio is as small as $1.06$). }
Major mergers are also responsible for the aspherical shapes in the accretion shock and splashback shells. Even though mergers are not directly responsible for the formation of the accretion shock, the `run-away' shocks generated from mergers \citep{Zhang2019} can overtake and power the accretion shock, thus affecting the shape of the accretion shock more than the shape of the DM splashback shell (see Section~\ref{sec:mergers} for the impact of mergers on accretion shock radius).
In Figure~\ref{fig:rsh_rsp_merge_map}, we show the evolution of gas entropy (the most apparent feature in the accretion shock as seen in Figure~\ref{fig:rsp_rsh_map}) and DM density maps as the cluster undergoes a major merger. Before the merger, the splashback shell encompasses the two merging sub-clusters, while the accretion shock encloses the shock-heated gas associated with these two subclusters. After the merger, the splashback shell decreases rapidly as the collisionless DM of the two clusters overlap with each other. The accretion shock radius, on the other hand, decreases more slowly with time as the gas lags behind (middle two panels in Figure~\ref{fig:rsh_rsp_merge_map}). Thus, towards the end of the relaxation period, the ratio of shock to splashback radii is slightly larger than that before the merger, $R_{\rm sh}/R_{\rm sp}$ by $\approx 10\%$ (bottom panel in Figure~\ref{fig:rsh_rsp_merge_map}).
Note that the transient internal shocks driven by mergers have much lower Mach number compared to the external accretion shock, leading to much smaller entropy jumps compared to those produced by the accretion shocks.
\begin{figure}
\begin{center}
\includegraphics[scale=0.5]{./figures/merger_2.pdf}
\caption {Maps of gas entropy ({\em left} panels) and DM ({\em right} panels) of a cluster (CL21) undergoing an almost equal mass (mass ratio = 0.83) merger. The solid line and the dotted line show the accretion shock and splashback shells respectively.
The panels from {\em top} to {\em bottom} show the cluster at different merging stages: $t_{\rm merge}= -0.4{\rm \, Gyr}, +0.4{\rm \, Gyr},
+1.25{\rm \, Gyr}, +2{\rm \, Gyr}$, where $t_{\rm merge}$ is the merging time defined as when $R_{\rm 500c}$ of the two merging haloes first touches. During the merger, the splashback and accretion shock shells continue to evolve. After $2 {\rm \, Gyr}$, the DM splashback shell becomes more spherical, while the accretion shock is still elongated along the axis of merger and filament.
}
\label{fig:rsh_rsp_merge_map}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=0.49\textwidth]{./figures/rsh_rsp_lastmerger_m1.pdf}
\caption{The shock to splashback radius ratio as a function of time since major merger for different impact parameters, indicated with solid and dashed lines. The lines and shaded regions indicate median and the 16-84 percentile of the ratio. The green line shows the most extreme major merger in CL21, whose gas entropy and DM density maps are shown in Figure~\ref{fig:rsh_rsp_merge_map}. The $R_{\rm sh}/R_{\rm sp}$ ratios increase temporarily within $1-2{\rm \, Gyr}$ after major mergers, and they all stay within the range of 1.6 and 2.4 throughout the mergers.
}
\label{fig:rsh_rsp_merger}
\end{center}
\end{figure}
\subsection{Effects of Mergers on Accretion Shock Radius}\label{sec:mergers}
Mergers can increase the accretion shock radius, boosting the ratio of the shock to splashback radii ($R_{\rm sh}/R_{\rm sp}$).
A merger event can generate a merger shock in the ICM, which accelerates to outer radii through a ``run-away'' shock with high Mach number \citep{Zhang2019}. When a merger shock runs into the accretion shock, the merger shock accelerates the accretion shock and pushes it out temporarily, whose duration depends on the MAR of the halo and the Mach number $\mathcal{M}$ of the shock \citep{Zhang2020}. Higher MAR leads to smaller increase in the accretion shock radius. For example, for a merger shock with $\mathcal{M} = 1.5$, in a halo with low MAR of $\Gamma_{\rm 200m}=1$, the accretion shock radius can grow to twice as predicted by self-similar model, before receding back to self-similar predicted value $3$~Gyr later. However, for the same Mach number merger shock in a halo with a higher MAR of $\Gamma_{\rm 200m}=3$, the accretion shock radius can only grow to 1.5 times the self-similar predicted shock radius, before receding back to the self-similar value after $1.5$~Gyr. Physically, the higher MAR of the halo implies larger ram pressure of the infalling gas, which pushes the run-away shock farther back towards the cluster center, leading to the smaller accretion shock radius.
Figure~\ref{fig:rsh_rsp_merger} shows $R_{\rm sh}/R_{\rm sp}$ before and after major merger events from $z=4$ onward for different impact parameters in the simulation, where a major merger is defined as the merger event where the mass ratio is less than 1:4. Impact parameter $b$ is defined as the perpendicular distance between the paths of the mergers, where the direction of the path is determined by the velocities of the haloes in the last snapshot when $R_{\rm 500c}$ of two haloes do not overlap.
It shows a temporary increase in $R_{\rm sh}/R_{\rm sp}$ right after the merger peaking after 1~Gyr and lasting around 2~Gyr.
The average $R_{\rm sh}/R_{\rm sp}$ before and after merger remains relatively constant, but can increase by $10\%$ within $2-3{\rm \, Gyr}$ after the merger. In extreme cases where there is equal mass merger and almost head-on collision (with the mass ratio of $0.83$ and impact parameter $b=0.49$~Mpc for CL21), we find that after 2 Gyr of the merger the $R_{\rm sh}/R_{\rm sp}$ ratio temporarily increases up to 2.4, corresponding to an increase of about $30\%$, before receding back to 1.8 after 4 Gyr of the merger. In fact, the ratio ranges between 1.6 to 2.4 during mergers of all haloes in our sample.
Our results indicate that mergers cannot account for all of the offsets between the volume-averaged $R_{\rm sh}$ and $R_{\rm sp}$, which is almost a factor of 2. When undergoing major merger, the ratio of volume-averaged radii $R_{\rm sh}/R_{\rm sp}$ of CL21 remains above 1.6 before and after the merger (Figure~\ref{fig:rsh_rsp_merger}). However, in Figure~\ref{fig:rsh_rsp_merge_map}, we found the instance where the shock to splashback ratio becomes very close to 1 and rapidly increases as the merger shock propagates away from the merger axis. This indicates that whether the merger-accelerated shocks can explain the offset between the shock and splashback radii depends on which line-of-sight we choose to compare. We leave further exploration of this idea to a future work.
\section{Discussions \& Future Work}
\label{sec:discussions}
The present work focuses on the massive DM haloes where the accretion shock is growing and the cluster outskirt is not affected significantly by galaxy formation physics. However, the accretion shocks are expected to behave differently for lower-mass group and galaxy scales. Radiative cooling can significantly reduce the pressure support of the galactic haloes which can lead to the collapse of the accretion shock into the inner region \citep{Birnboim03}. In addition, at lower halo masses of $M<10^{12}M_{\odot}$, the DM phase space structure of the orbiting halo can extend out to turnaround radii \citep{Prada2006}, which leads to much larger ratio of $R_{\rm edge/}R_{\rm 200c}\approx4$ \citep{Ludlow2009,Wang2009}. Future studies should extend the current work to lower mass haloes. Additionally, higher mass and spatial resolution simulations will be required to resolve the shock radius of these low mass haloes.
\section{Conclusions}
\label{sec:conclusions}
In this work, we investigate the relation between the splashback of dark matter (DM) and the accretion shock of gas in the outskirts of cluster-size DM haloes, using the {\it Omega500} cosmological hydrodynamical simulations. Our main findings are summarized below:
\begin{enumerate}
\item The accretion shock radius is located farther from the cluster centre than the DM splashback radius (Figure~\ref{fig:rsp_rsh_map}). The phase space structures of DM and gas follow each other outside the accretion shock where the gas thermal energy is small compared to its kinetic energy. Inside the accretion shock, gas is thermalized with relatively small radial infall velocities, while DM particles orbit with large velocity dispersion within the interior of DM haloes. The phase space structures for both the orbiting DM and the thermalized gas extend beyond DM splashback radius (Figure~\ref{fig:phase_diagram}).
\item The ratio between the two radii depends on the definitions of the splashback and the shock radii.
Specifically, the accretion shock radius defined by the entropy drop is larger than all definitions of the splashback radius in the literature. The accretion shock radius defined using the steepest drop in entropy is approximately $\approx 1.89$ times larger than the splashback radius defined by the steepest slope in the DM density profile, and it is $\approx1.2$ times larger than the edge of the DM phase-space structure (Figure~\ref{fig:rsp_rsh}).
\item The accretion shock radius of gas decreases with mass accretion rate (MAR) of the halo, similar to the MAR dependence in the splashback radius (Figures~\ref{fig:profiles} and \ref{fig:mar_slope}). The resulting ratio is fairly independent of redshift and baryonic physics for the cluster-size DM haloes (Figure~\ref{fig:rsp_rsh}).
\item The gas density follows DM density at the outer radii with steepest density decrease near the DM splashback radius. The steepest decrease in other thermodynamic (temperature, pressure, and entropy) profiles, however, occurs at the location of the accretion shock (Figure~\ref{fig:profiles_slope}). We do not find significant gas density jump in radial profiles at the accretion shock as the ratio of the post- and pre-shock density is capped at a maximum value of $4$ in the strong shock limit, which is further smeared out by azimuthally averaging around aspherical shocks.
\item Mergers account for on average $\approx 10\%$, and upto $30\%$ increase in the shock to splashback ratio for the duration of $\approx 1$~Gyr right after the merger. After that, the merger-accelerated shock recedes back to the pre-merger accretion shock (Figure~\ref{fig:rsh_rsp_merger}).
\end{enumerate}
Our results have broad implications for using galaxy clusters as probes of astrophysics and cosmology. First, the apparent spatial offset between splashback and the more extended shock radii indicates that the transition boundaries between the 1-halo and 2-halo terms are different between DM and gas, leaving imprints in this transition region as probed by optical and SZ surveys. Second, the extended gas distribution beyond DM splashback must be taken into account when quantifying baryon fraction associated with DM haloes and cosmic web filaments as well as modeling galaxy quenching in the outskirts of clusters. Finally, the low density shock-heated gas between the accretion shock and the splashback radii may contribute to a significant fraction of the missing baryons and may play a role in early quenching of infalling galaxies. Investigating the differential dynamics of collisionless DM and collisional gas in the outskirts of galaxy clusters will be important for using the edges of galaxy clusters as laboratories for cosmology and astrophysics in the era of multi-wavelength cosmological surveys.
\section*{Acknowledgement}
We acknowledge Benedikt Diemer, Andrey Kravtsov, Philip Mansfield, and Xun Shi for useful comments on the draft. We used the publicly available python packages, \textsc{COLOSSUS} \citep{colossus} and \textsc{yt} \citep{yt} as part of our analysis pipeline. This work was supported in part by NSF AST-1412768 and the facilities and staff of the Yale Center for Research Computing.
\section*{Data Availability}
The {\em Omega500} simulations and the analysis code used in this paper are available upon request.
\bibliographystyle{mnras}
\section{Introduction}
\label{sec:intro}
In recent years, the outskirts of galaxy clusters have emerged as one of the new frontiers for cosmology and astrophysics \citep[see][for review]{walker_etal19}.
Recent theoretical advances revealed that the physical outer boundary for a dark matter (DM) halo can be defined using by the ``splashback'' radius based on the DM density profile drop \citep[e.g.,][]{diemer_kravtsov2014,adhikari_etal2014,more_etal2015}, the aspherical splashback surface \citep{mansfield_etal17,Mansfield2020}, or the edge radius of the DM phase space structure \citep{Aung2020}, with various definitions encompass varying fraction of orbiting DM particles \citep{diemer_etal17}. Observationally, the outer boundaries of the DM haloes have recently been detected using weak-lensing \citep[e.g.,][]{chang_etal18}, galaxy number density \citep[e.g.,][]{more_etal2016,Baxter2017,Shin2019,Zurcher_More2019,Murata2020} and phase space structure \citep{Tomooka2020}. Upcoming multiwavelength surveys (such as CMB-S4 in microwave and DESI, Rubin, PFS in optical) will provide unprecedented insight on the outer boundaries of massive DM haloes and promise to shed new insight into cosmology and non-linear structure formation of the Universe.
Gas accreting at cluster outskirts provides an alternative probe of cluster boundary. However, the dynamics of the collisional gas is fundamentally different from that of collisionless DM. Unlike the collisionless DM particles which orbit within the DM halos, the collisional gas is shock heated during its first infall, resulting in a high Mach number ($\mathcal{M}>100$) cosmic accretion shock marked by the prominent entropy jump. The secondary infall model predicts that the location of the accretion shock coincides with the splashback radius \citep{bertschinger1985, shi2016b}. Commonly referred to as ``external shock'' in the literature \citep[e.g.,][]{Miniati2000,ryu_etal2003,Skillman2008, molnar_etal2009}, the accretion shock arises from the infall of low density pristine gas in the void regions onto the cluster potential (in contrast to ``internal shocks'' which occurs within the virialization region of DM haloes due to mergers and penetrating filaments). The external accretion shock thus defines a physical boundary of the hot collapsed gas in DM haloes, which is also dependent on their mass accretion rate (MAR) \citep{lau_etal2015}.
In this work we investigate the locations of shock and splashback radii by analyzing the {\em Omega500} hydrodynamical cosmological simulations. We find that the accretion shock radius defined using the drop in the gas entropy is larger than all definitions of the splashback radius in the literature by $20-100\%$, in contrast to the prediction of the self-similar models. Specifically, we find that the accretion shock radius is larger by $\approx 1.89$ relative to the splashback radius and $\approx 1.2$ relative to the edge radius of the DM phase space structure. Furthermore, we find that the ratios of the shock and splashback/edge radii are independent of halo mass and redshift, but dependent on their MAR.
We describe our simulations and analysis methods in \S\ref{sec:sim}. Results and discussions are presented in \S\ref{sec:results} and \S\ref{sec:discussions}, respectively. Conclusions are summarized in \S\ref{sec:conclusions}.
\section{Simulations}
\label{sec:sim}
We analyze the clusters from the {\em Omega500} simulation \citep{nelson_etal2014}, a high-resolution hydrodynamical simulation of a large cosmological volume
with the comoving box size of $500\,h^{-1}\,{\rm \, Mpc}$.
The simulation is performed using the Adaptive Refinement Tree (ART) $N$-body+gas-dynamics code \citep{kravtsov1999, kravtsov_etal2002, rudd_etal2008},
which is an Eulerian code that uses adaptive refinement in space and time,
and non-adaptive refinement in mass \citep{klypin_etal2001}
to achieve the dynamic ranges to resolve the cores of haloes formed in self-consistent cosmological simulations in a flat $\Lambda$CDM model with WMAP 5 years cosmological parameters:
$\Omega_m = 1 - \Omega_{\Lambda} = 0.27$, $\Omega_b = 0.0469$,
$h = 0.7$ and $\sigma_8 = 0.82$, where the Hubble constant is defined as
$100\,h\;{\rm \, km \, s}^{-1}\,{\rm \, Mpc}^{-1}$ and $\sigma_8$ is the mass variance within
spheres of radius $8\,h^{-1}{\rm \, Mpc}$.
Haloes are identified in the simulation using a spherical overdensity halo finder described in \cite{nelson_etal2014}. We select 65 haloes with mass $M_{\rm 500c} \geq 3\times10^{14}\,h^{-1}M_{\odot}$ at $z=0$ and re-simulate the box with the higher resolution DM particles in regions of the selected haloes, resulting in an effective mass resolution of $1.09 \times 10^9\, h^{-1}M_{\odot}$, which corresponds to $2048^{3}$ DM particles. We built the merger tree by tracking the most massive progenitors of haloes over time using the merger tree code presented in \citet{yu_etal2015}. This is done by following the $10\%$ most bound DM particles at each snapshot. We define the start of the merging process at the epoch when $R_{\rm 500c}$ of the two haloes start to overlap with each other. Following \citet{diemer_kravtsov2014}, we compute the MAR of DM haloes as
\begin{equation}
\Gamma_{\rm 200m} = \frac{\log_{10}M_{\rm 200m} (z) - \log_{10}M_{\rm 200m} (z')}{\log_{10}a(z)-\log_{10}a(z')} \; ,
\end{equation}
where $a$ is the expansion factor and $z=0$, $z'=0.5$, and $M_{\rm 200m}$ is the mass enclosed within the radius $R_{\rm 200m}$ such that the density enclosed is 200 times the mean density of the universe, $M_{\rm 200m} = 200\rho_m \times 4\pi R_{\rm 200m}^3/3$.
The simulation is performed on a uniform $512^3$ grid with 8 levels of mesh refinement, implying a maximum comoving spatial resolution of $3.8\,h^{-1} {\rm \, kpc}$. The spatial resolution is controlled by the density of the cells, and the maximum comoving resolution is only achieved at the centre of the haloes. However, the spatial resolution near the shock radius is between $0.03\,h^{-1}{\rm \, Mpc}$ and $0.12\,h^{-1}{\rm \, Mpc}$, which is sufficient to determine the locations of the edge and shock radii (which is typically of order several Mpc in size) with the accuracy better than 5\%. With a typical number density of $n_e\approx 10^{-3}{\rm \, cm}^{-3}$ and temperature of $T\approx 10^6{\rm \, K}$, the mean free path of electron is much smaller than the resolution.
Since the effects of non-gravitational baryonic physics (such as gas cooling and energy feedback from supernova and black holes) are small in cluster outskirts compared to cluster cores, we focus on analyzing the outputs of the non-radiative simulation for simplicity.
For completeness, we also checked the effects of baryonic physics by comparing the results to those of runs with cooling and star formation and AGN feedback.
\section{Results}
\label{sec:results}
\begin{figure}
\begin{center}
\includegraphics[scale=0.40]{./figures/shock_splashback_map_2.pdf}
\caption{Gas entropy (top left), DM density (top right), gas temperature (bottom left) and gas pressure (bottom right) maps of the simulated cluster (CL135) extracted from the non-radiative {\em Omega500} hydrodynamical cosmological simulation. The images are 15.625Mpc/h wide with the projection depth of 3.90625 Mpc/h. The inner dashed lines indicate the splashback shell computed using the method from \citet{mansfield_etal17}, whereas the outer white lines indicate the shock shell found by the discontinuous jump in entropy as well as pressure. Note that several low-entropy gas streams have penetrated inside of the accretion shock radius along the filaments without getting shock heated.
}
\label{fig:rsp_rsh_map}
\end{center}
\end{figure}
\subsection{Identifying Splashback and Shock Surface}
We determine the location of the splashback radius using the \textsc{Shellfish}\ code \citep{mansfield_etal17}. For each halo, the code draws $10^5$ random sight lines from the halo centre and samples the DM density along each line-of-sight (LOS). The splashback radius $R_{\rm sp}$ is defined as the radius of a spherical surface that encompasses the same volume as that enclosed by the surface of sharp DM density jumps in all LOS \citep[see][for more details]{mansfield_etal17}.
We determine the accretion shock radius in a similar manner using the \textsc{Shellfish}\ code. Namely, we draw $786$ LOS according to HEALPix \citep{Healpix} pixels from the halo centre and sample the gas entropy profile along each LOS. The 786 LOS chosen here corresponds to the fourth level resolution of HEALPix, and at the shock radius, each pixel corresponds to about $0.26\,h^{-2}{\rm \, Mpc}^2$, with a length resolution of $\approx 0.5\,h^{-1}{\rm \, Mpc}$, about 4 times the simulation resolution in the region. Measuring the shock radius with the fifth level resolution with 3072 LOS leads to a less than $3\%$ difference in shock measurements.
For each LOS, we select all the gas cells that each LOS passes through, and sample the gas entropy profile along each LOS. The profile is then smoothed with a Savitzky-Golay (SG) filter with window-length of 9 equally spaced logarithmic radial bins and a polynomial order of 5; we checked that the results are robust to the variation in the parameter for SG filter (in the window-length from 5 to 11 bins) and polynomial orders (in the range of 2 to 7). We define the location of the accretion shock as the radius of the minimum in the logarithmic entropy slope. We remove LOS where the entropy jump is less than a factor of 50 (corresponding to Mach number of $\mathcal{M} \sim 20$; the results are unchanged for $\mathcal{M}=10$ to $50$), as these directions host substructures and filamentary gas streams.
After these removals, the covering fraction of the shock surface of the total spherical area is approximately $80\%$ for all clusters.
We fit Penna-Dines function with $K=1,I=J=2$ \citep[corresponding to the first and second order expansion of $\cos\phi$ and $\sin\phi$ of polar angle $\phi$,][]{Penna_Dines2007}:
\begin{equation}\label{eq:penna_dines}
r(\theta,\phi) = \sum_{i,j,k=0}^{I,J,K} c_{\rm ijk} \sin^{i+j}\phi \cos^k\phi \sin^j\theta \cos^i\theta,
\end{equation}
to the shock position along each LOS and calculate the enclosed volume inside the fitted (non-spherical) surface. We then define the shock "radius" as the radius of a sphere that encompasses the same amount of volume as that enclosed within the fitted surface. We note that the total volume inside the shock radius is larger than the total volume of shock heated gas as it also includes volume of unshocked gas residing in filaments.
Figure~\ref{fig:rsp_rsh_map} shows the map of one of the haloes indicating the splashback and shock shells identified using the above algorithms. The splashback shell identified encompasses the DM structure, whereas the shock shell encompasses a much larger, extended area, where the entropy and pressure shows significant decline. We also note that there are several unshocked, low-entropy gas streams that have penetrated inside of the accretion shock radius of the halo along the filaments.
\subsection{Phase Space Structures of DM and Gas}\label{sec:phase_space}
\label{sec:dynamics}
\begin{figure}
\begin{center}
\includegraphics[width=0.49\textwidth,trim={10 1 0 1},clip]{./figures/phase_space.pdf}
\caption {Phase-space diagrams of DM ({\em top} panels) and gas ({\em bottom} panels) for a relaxed cluster CL135 ({\em left} panels) with low MAR ($\Gamma = 0.5$), and a merging cluster CL77 ({\em right} panels) with high MAR ($\Gamma = 2.9$). The solid, dashed, dotted, and dot-dashed lines represent $R_{\rm 200m}$, $R_{\rm sp}$ (splashback radius), $R_{\rm max,sp}$ (maximum splashback radius along any line-of-sight), and $R_{\rm sh}$ (shock radius), respectively. The colour represents the relative mass fraction of DM (top panels) and gas (bottom panels), with deeper colour indicating higher mass fraction at a given radius. The phase space structure of virialized DM haloes extends past $R_{\rm sp}$, reaching nearly $R_{\rm sh}$. Gas follows DM at $r\gtrsim R_{\rm sh}$, while gas dynamics differs significantly from that of collisionless DM at $r\lesssim R_{\rm sh}$.}
\label{fig:phase_diagram}
\end{center}
\end{figure}
Figure~\ref{fig:phase_diagram} shows phase-space densities of DM and gas for their radial velocity components as a function of radius, for a relaxed cluster (CL135) and a merging cluster (CL77), respectively. The average radial velocity of gas and DM is negative outside the shock radius ($R_{\rm sh}$) as they fall onto the cluster potential. The two trace each other as the gas pressure is low, rendering gas to behave similarly to collisionless DM. Gas infalling from the void is shock heated at the shock radius, causing gas to lose its kinetic energy into heating of the gas. Filamentary accretion can bring infalling gas inside of the virialization region of the DM haloes and form shocks with small Mach number ($\mathcal{M}<3$) before gas loses their kinetic energy. The DM particles, on the other hand, exchange energy through gravitational interactions as they orbit through the interior of the DM halo.
Within the splashback radius ($R_{\rm sp}$), a DM halo exhibits a typical virialized phase-space structure, where the splashback and orbital motions of DM particles produce a dispersion with zero mean radial velocity. The phase space structure of the DM halo can extend out to radii larger than the volume-averaged splashback radius as not all particles within the splashback surface are expected to lie within the volume-averaged splashback radius $R_{\rm sp}$ due to asphericity \citep{mansfield_etal17,diemer_etal17,Aung2020}. Even when using $R_{\rm max,sp}$ (the maximum radius of the splashback surface defined using the Penna-Dines approximation surface), a small amount (upto $1\%$) of orbiting particles can still exist beyond $R_{\rm max,sp}$ \citep{Mansfield2020} as shown in Figure~\ref{fig:phase_diagram}.
Due to the collisional nature of the gas, however, the phase-space distribution of gas differs significantly within the interior of the cluster. The radial velocity dispersion of gas is considerably smaller because, as the gas is shock heated through the accretion shock, where most of the gas kinetic energy is converted to thermal energy. Thus, inside $R_{\rm 200m}$, the level of gas motions inside the accretion shock remains small in absence of external disruption by mergers for CL135, while gas motions induced by mergers comprise of most of the velocity dispersion in CL77. The position of the shock radius is closer to the edge of DM phase space than the splashback radius.
\subsection{Shock Radii Determined from Profiles}
\begin{figure}
\begin{center}
\includegraphics[width=0.5\textwidth]{./figures/profiles_individual.pdf}
\caption{Spherically averaged gas density, DM density, volume-weighed entropy and pressure as a function of radius for the two different clusters: a relaxed cluster CL135 with low MAR ($\Gamma = 0.5$), and a merging cluster CL77 ({\em right} panels) with high MAR ($\Gamma = 2.9$). The vertical lines indicate splashback and shock radii based on the \textsc{Shellfish}\ and our method, respectively. The gas pressure and entropy profiles show significant decrease near the shock radii, while the gas density and DM density decreases at the splashback radii. The gas density slope is shallower than the DM density slope. The slower accreting halo (CL135) also has a smoother jump and larger shock and splashback radii than the fast accreting CL77.}
\label{fig:profiles}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=0.49\textwidth]{./figures/profiles_slope.pdf}
\caption{The logarithmic slope of different gas and DM median profiles as a function of radius. The radius where the minimum of gas entropy, temperature and pressure slope is defined as the shock radius $R_{\rm sh,sph}$ of the spherically averaged gas profile, denoted by dotted line. The radius where the minimum of DM density slope is defined as the splashback radius $R_{\rm sp,sph}$ of the spherically averaged profile, denoted by dashed line, and coincides with the minimum of the gas density slope. }
\label{fig:profiles_slope}
\end{center}
\end{figure}
Figure~\ref{fig:profiles} shows the spherically averaged DM density, gas density, and volume-weighed entropy and pressure profiles for two representative clusters in the sample, CL135 a relaxed cluster with low $\Gamma$, and CL77, a merging cluster with high $\Gamma$. We also overplot their splashback radii estimated from \textsc{Shellfish}\ $R_{\rm sp}$, and their shock radii from the shock surface $R_{\rm sh}$. Both $R_{\rm sp}$ and $R_{\rm sh}$ are smaller for the high $\Gamma$ cluster CL77 than than the low $\Gamma$ cluster CL135. In both clusters, there are sharp drops in the pressure and entropy profiles at the accretion shock, and they are particularly prominent in the high $\Gamma$ cluster CL77. However, the corresponding decrease in gas density profiles are small at the accretion shock. This is because the density contrast across a shock is intrinsically smaller than their counterparts in pressure and entropy, and is capped at a maximum value of $4$, as expected from the Rankine -- Hugoniot shock jump condition.
Figure~\ref{fig:profiles_slope} shows the logarithmic slope of the spherically averaged median DM density, gas density, temperature, pressure, and entropy profiles for all clusters at $z=0$. The entropy profile is increasing at all radii before the shock radius, and within $0.1\leq r/R_{\rm 200m}\leq 0.6$, it is consistent with previous findings of entropy slope of $1.1$ \citep[e.g.,][]{voit_etal2005} as expected from self-similar cluster growth. At the shock radius, entropy decreases sharply indicating a strong shock. The shock front, however, is wider and is not as abrupt as the shock front in LOS profile due to smoothing over aspherical shock fronts and variations among clusters. The pressure profile has the minimum slope about $-7$ at the shock radius. Inside the shock radius, the pressure profile is also rapidly decreasing, consistent with the universal pressure profile characterized by the generalized NFW profile \citep[e.g.,][]{nagai_etal2007a,arnaud_etal2010}. Similarly, the temperature profile has the minimum slope of about $-2$ at the shock radius.
The DM density slope follows NFW profile \citep{nfw1996} closely, where the slope is $-1$ in the inner region and slowly decreases to $-3$ in outer region before hitting minimum at the splashback radius. The gas density is much flatter in inner region starting with slope of $\approx 0$, but approaches NFW and follows DM profile at outer radii. In fact, gas density slope becomes minimum at DM splashback radius, while only showing mild decrease at the shock radius. The smaller decrease in density slope at the shock radius compared to other thermodynamic quantities is expected, as the density jump for a shock wave in ideal gas is capped at $4$, while there is no upper bound for the jumps in temperature, pressure or entropy.
For DM, the splashback radius identified as the steepest point in the spherically averaged density profile $R_{\rm sp, sph}$ is smaller than the splashback radius estimated from \textsc{Shellfish}\ $R_{\rm sp}$ \citep[see also][]{mansfield_etal17}.
For the accretion shock radius, the radius computed from spherically averaged profile $R_{\rm sh, sph}$, is the same as that identified from the volume-averaged shock surface $R_{\rm sh}$.
Figure~\ref{fig:mar_slope} shows the profile slopes for different $\Gamma$ samples. Here the splashback and shock radii are identified as the steepest jumps in DM density and gas entropy profiles, respectively. Both radii decrease for larger $\Gamma$. The DM density slope is steeper for larger $\Gamma$, consistent with the previous result \citep{diemer_kravtsov2014,more_etal2015}. The pressure jump is larger for larger $\Gamma$, indicating a stronger shock.
\begin{figure}
\begin{center}
\includegraphics[width=0.51\textwidth]{./figures/slope_mar.pdf}
\caption{The logarithmic slope of different gas and DM median profiles for different MAR. The MAR is split at the 33 and 66-percentile of all haloes, which results in the low $\Gamma_{\rm 200m}$ sample with $\Gamma_{\rm 200m}<1.5$ (dashed line) and high $\Gamma_{\rm 200m}$ sample with $\Gamma_{\rm 200m}>2.7$ (solid line). The splashback radius is where the total matter (DM+gas) density slopes is the smallest, and the shock radius where the minimum of entropy and pressure slope is smaller for the higher MAR haloes. }
\label{fig:mar_slope}
\end{center}
\end{figure}
\subsection{Offsets between Splashback and Shock Radii}
\begin{figure}
\begin{center}
\includegraphics[width=0.49\textwidth]{./figures/L500_rsp_gamma_updated}
\caption{The splashback radius $R_{\rm sp}$ (red points) and the accretion shock radius $R_{\rm sh}$ (blue points), normalized by the halo radius $R_{\rm 200m}$, plotted as a function of the MAR of the cluster-size DM haloes, $\Gamma_{\rm 200m}$.
The solid line is the $R_{\rm sp}-\Gamma_{\rm 200m}$ relation from DM-only cosmological simulation defined by \textsc{Shellfish}\ \citep[M17,][]{mansfield_etal17} and $R_{\rm sp,87}$ computing using SPARTA \citep{diemer_etal17}. The dashed lines represent the best-fitting $R_{\rm sp}-\Gamma_{\rm 200m}$ relation times the average $R_{\rm sp}/R_{\rm sp}$ ratios for three different baryonic simulations (magenta for AGN feedback simulation, cyan for cooling and star forming simulation, and yellow for non-radiative, embedded behind cyan line). Also shown is $r_{\rm edge}\approx 1.6R_{\rm sp,87}$, marking the edge of the DM phase space structure \citep{Aung2020}.
}
\label{fig:rsp_rsh}
\end{center}
\end{figure}
Figure~\ref{fig:rsp_rsh} shows the splashback radius $R_{\rm sp}$ normalized by the halo radius $R_{\rm 200m}$, plotted as a function of the MAR ($\Gamma_{\rm 200m}$) of haloes extracted from the non-radiative {\it Omega500} simulation. The splashback radius decreases with increasing MAR, confirming previous numerical \citep{diemer_kravtsov2014,mansfield_etal17} and analytic results \citep{adhikari_etal2014, shi2016a}.
The $R_{\rm sp}$ from our hydrodynamic simulation agrees well with the best-fitting relation from the DM-only simulation.
In the same panel, we show that $R_{\rm sh}/R_{\rm 200m}$ decreases with $\Gamma_{\rm 200m}$ in a similar manner to the $R_{\rm sp}$--$\Gamma_{\rm 200m}$ relation. The average ratio between the radii is $R_{\rm sh}/R_{\rm sp} = 1.89 \pm 0.16$ (based on the yellow dashed line in Figure~\ref{fig:rsp_rsh}, where the error indicates 1$\sigma$ scatter) at $z=0$, and it is only weakly dependent on MAR for the range probed here. At $z=1$, $R_{\rm sh}/R_{\rm sp} = 2.03 \pm 0.32$, and at $z=3$, $R_{\rm sh}/R_{\rm sp} = 2.12 \pm 0.35$ which is consistent with no evolution with redshift. We also find that the scatter in $R_{\rm sh}$ is larger than that of $R_{\rm sp}$.
We repeated the same analysis for the same sets of haloes with baryonic physics that include radiative cooling and star formation, and AGN feedback. The differences in $R_{\rm sp}$ and $R_{\rm sh}$ between simulations with different baryonic physics is $\lesssim 1$\%, thus both radii remain essentially unchanged in the presence of baryonic physics.
We note that the phase space structure of the DM halo can extend out to radii larger than the splashback radius, because not all particles within the splashback surface are expected to be enclosed within the volume-averaged $R_{\rm sp}$. In fact, the splashback radius from \textsc{Shellfish}\ only contains about $87\%$ of the particle apocentre \citep{diemer_etal17}. The edge radius, which marks the end of DM phase space, corresponds to the radius where the fraction of orbiting subhaloes is greater than 99\% (denoted as $R_{\rm sp,99}$), which is approximately $1.6$ times larger than the splashback radius measured with \textsc{Shellfish}\ (denoted as $R_{\rm sp,87}$) \citep{Aung2020}. This edge radius lies in the region in between the accretion shock and splashback surfaces. Specifically, the ratio of the shock and edge radius is $R_{\rm sh}/1.6R_{\rm sp,87} \approx 1.2$ for $1\leq \Gamma_{\rm 200m} \leq 4$, indicating that the shock radius is about 20$\%$ larger than the edge radius, on average.
\subsection{Shapes of Shock and Splashback Shells}\label{sec:apshericity}
\begin{figure}
\begin{center}
\includegraphics[width=0.49\textwidth]{./figures/rmaxmin_r200.pdf}
\caption{\change{The normalized distribution of maximum (solid) and minimum (dotted) radius of shock (blue) and splashback (red) surface with respect to $R_{\rm 200m}$. The shock and splashback shells are highly aspherical, causing the maximum and minimum radius along lines-of-sight to differ by more than $65\%$. Note that the overlap in the distributions of the minimum shock and maximum splashback radii is due to the scatter in $R_{\rm sh}/R_{\rm sp}$ ratio. For individual haloes, the shock radius is always larger than the splashback radius along every line-of-sight.
}}
\label{fig:rmaxmin}
\end{center}
\end{figure}
In practice, DM splashback and accretion shock are aspherical, because haloes form through merger and accretion of materials through cosmic web of filaments that are inherently aspherical, as shown in Figure~\ref{fig:rsp_rsh_map}. Figure~\ref{fig:rmaxmin} shows the distribution of the maximum ($R_{\rm max}$) and minimum ($R_{\rm min}$) splashback and shock shells. We find that the $R_{\rm max,sh}$ and $R_{\rm min,sh}$ of the shock shell can be $1.65$ times larger and $0.7$ times smaller compared to the volume-averaged shock radius, respectively, while the maximum and minimum distance of the splashback shell ranges between $(0.8-1.42)$ times the volume-averaged splashback radius. This shows that the accretion shock shells are generally more aspherical than the splashback shells, which cause the spherically averaged gas profiles to appear smoother than the actual accretion shocks. \change{Even though the minimum shock radii and maximum splashback radii overlap in Figure~\ref{fig:rmaxmin}, we emphasize that the largest radius of the splashback shell is still smaller than the smallest radius of the shock shell of the same halo for the individual haloes in our sample. The minimum shock to splashback radius ratio is found in the directions perpendicular to the axis of filament and merger (see the panels in the second and third rows of Figure~ \ref{fig:rsh_rsp_merge_map} where the ratio is as small as $1.06$). }
Major mergers are also responsible for the aspherical shapes in the accretion shock and splashback shells. Even though mergers are not directly responsible for the formation of the accretion shock, the `run-away' shocks generated from mergers \citep{Zhang2019} can overtake and power the accretion shock, thus affecting the shape of the accretion shock more than the shape of the DM splashback shell (see Section~\ref{sec:mergers} for the impact of mergers on accretion shock radius).
In Figure~\ref{fig:rsh_rsp_merge_map}, we show the evolution of gas entropy (the most apparent feature in the accretion shock as seen in Figure~\ref{fig:rsp_rsh_map}) and DM density maps as the cluster undergoes a major merger. Before the merger, the splashback shell encompasses the two merging sub-clusters, while the accretion shock encloses the shock-heated gas associated with these two subclusters. After the merger, the splashback shell decreases rapidly as the collisionless DM of the two clusters overlap with each other. The accretion shock radius, on the other hand, decreases more slowly with time as the gas lags behind (middle two panels in Figure~\ref{fig:rsh_rsp_merge_map}). Thus, towards the end of the relaxation period, the ratio of shock to splashback radii is slightly larger than that before the merger, $R_{\rm sh}/R_{\rm sp}$ by $\approx 10\%$ (bottom panel in Figure~\ref{fig:rsh_rsp_merge_map}).
Note that the transient internal shocks driven by mergers have much lower Mach number compared to the external accretion shock, leading to much smaller entropy jumps compared to those produced by the accretion shocks.
\begin{figure}
\begin{center}
\includegraphics[scale=0.5]{./figures/merger_2.pdf}
\caption {Maps of gas entropy ({\em left} panels) and DM ({\em right} panels) of a cluster (CL21) undergoing an almost equal mass (mass ratio = 0.83) merger. The solid line and the dotted line show the accretion shock and splashback shells respectively.
The panels from {\em top} to {\em bottom} show the cluster at different merging stages: $t_{\rm merge}= -0.4{\rm \, Gyr}, +0.4{\rm \, Gyr},
+1.25{\rm \, Gyr}, +2{\rm \, Gyr}$, where $t_{\rm merge}$ is the merging time defined as when $R_{\rm 500c}$ of the two merging haloes first touches. During the merger, the splashback and accretion shock shells continue to evolve. After $2 {\rm \, Gyr}$, the DM splashback shell becomes more spherical, while the accretion shock is still elongated along the axis of merger and filament.
}
\label{fig:rsh_rsp_merge_map}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=0.49\textwidth]{./figures/rsh_rsp_lastmerger_m1.pdf}
\caption{The shock to splashback radius ratio as a function of time since major merger for different impact parameters, indicated with solid and dashed lines. The lines and shaded regions indicate median and the 16-84 percentile of the ratio. The green line shows the most extreme major merger in CL21, whose gas entropy and DM density maps are shown in Figure~\ref{fig:rsh_rsp_merge_map}. The $R_{\rm sh}/R_{\rm sp}$ ratios increase temporarily within $1-2{\rm \, Gyr}$ after major mergers, and they all stay within the range of 1.6 and 2.4 throughout the mergers.
}
\label{fig:rsh_rsp_merger}
\end{center}
\end{figure}
\subsection{Effects of Mergers on Accretion Shock Radius}\label{sec:mergers}
Mergers can increase the accretion shock radius, boosting the ratio of the shock to splashback radii ($R_{\rm sh}/R_{\rm sp}$).
A merger event can generate a merger shock in the ICM, which accelerates to outer radii through a ``run-away'' shock with high Mach number \citep{Zhang2019}. When a merger shock runs into the accretion shock, the merger shock accelerates the accretion shock and pushes it out temporarily, whose duration depends on the MAR of the halo and the Mach number $\mathcal{M}$ of the shock \citep{Zhang2020}. Higher MAR leads to smaller increase in the accretion shock radius. For example, for a merger shock with $\mathcal{M} = 1.5$, in a halo with low MAR of $\Gamma_{\rm 200m}=1$, the accretion shock radius can grow to twice as predicted by self-similar model, before receding back to self-similar predicted value $3$~Gyr later. However, for the same Mach number merger shock in a halo with a higher MAR of $\Gamma_{\rm 200m}=3$, the accretion shock radius can only grow to 1.5 times the self-similar predicted shock radius, before receding back to the self-similar value after $1.5$~Gyr. Physically, the higher MAR of the halo implies larger ram pressure of the infalling gas, which pushes the run-away shock farther back towards the cluster center, leading to the smaller accretion shock radius.
Figure~\ref{fig:rsh_rsp_merger} shows $R_{\rm sh}/R_{\rm sp}$ before and after major merger events from $z=4$ onward for different impact parameters in the simulation, where a major merger is defined as the merger event where the mass ratio is less than 1:4. Impact parameter $b$ is defined as the perpendicular distance between the paths of the mergers, where the direction of the path is determined by the velocities of the haloes in the last snapshot when $R_{\rm 500c}$ of two haloes do not overlap.
It shows a temporary increase in $R_{\rm sh}/R_{\rm sp}$ right after the merger peaking after 1~Gyr and lasting around 2~Gyr.
The average $R_{\rm sh}/R_{\rm sp}$ before and after merger remains relatively constant, but can increase by $10\%$ within $2-3{\rm \, Gyr}$ after the merger. In extreme cases where there is equal mass merger and almost head-on collision (with the mass ratio of $0.83$ and impact parameter $b=0.49$~Mpc for CL21), we find that after 2 Gyr of the merger the $R_{\rm sh}/R_{\rm sp}$ ratio temporarily increases up to 2.4, corresponding to an increase of about $30\%$, before receding back to 1.8 after 4 Gyr of the merger. In fact, the ratio ranges between 1.6 to 2.4 during mergers of all haloes in our sample.
Our results indicate that mergers cannot account for all of the offsets between the volume-averaged $R_{\rm sh}$ and $R_{\rm sp}$, which is almost a factor of 2. When undergoing major merger, the ratio of volume-averaged radii $R_{\rm sh}/R_{\rm sp}$ of CL21 remains above 1.6 before and after the merger (Figure~\ref{fig:rsh_rsp_merger}). However, in Figure~\ref{fig:rsh_rsp_merge_map}, we found the instance where the shock to splashback ratio becomes very close to 1 and rapidly increases as the merger shock propagates away from the merger axis. This indicates that whether the merger-accelerated shocks can explain the offset between the shock and splashback radii depends on which line-of-sight we choose to compare. We leave further exploration of this idea to a future work.
\section{Discussions \& Future Work}
\label{sec:discussions}
The present work focuses on the massive DM haloes where the accretion shock is growing and the cluster outskirt is not affected significantly by galaxy formation physics. However, the accretion shocks are expected to behave differently for lower-mass group and galaxy scales. Radiative cooling can significantly reduce the pressure support of the galactic haloes which can lead to the collapse of the accretion shock into the inner region \citep{Birnboim03}. In addition, at lower halo masses of $M<10^{12}M_{\odot}$, the DM phase space structure of the orbiting halo can extend out to turnaround radii \citep{Prada2006}, which leads to much larger ratio of $R_{\rm edge/}R_{\rm 200c}\approx4$ \citep{Ludlow2009,Wang2009}. Future studies should extend the current work to lower mass haloes. Additionally, higher mass and spatial resolution simulations will be required to resolve the shock radius of these low mass haloes.
\section{Conclusions}
\label{sec:conclusions}
In this work, we investigate the relation between the splashback of dark matter (DM) and the accretion shock of gas in the outskirts of cluster-size DM haloes, using the {\it Omega500} cosmological hydrodynamical simulations. Our main findings are summarized below:
\begin{enumerate}
\item The accretion shock radius is located farther from the cluster centre than the DM splashback radius (Figure~\ref{fig:rsp_rsh_map}). The phase space structures of DM and gas follow each other outside the accretion shock where the gas thermal energy is small compared to its kinetic energy. Inside the accretion shock, gas is thermalized with relatively small radial infall velocities, while DM particles orbit with large velocity dispersion within the interior of DM haloes. The phase space structures for both the orbiting DM and the thermalized gas extend beyond DM splashback radius (Figure~\ref{fig:phase_diagram}).
\item The ratio between the two radii depends on the definitions of the splashback and the shock radii.
Specifically, the accretion shock radius defined by the entropy drop is larger than all definitions of the splashback radius in the literature. The accretion shock radius defined using the steepest drop in entropy is approximately $\approx 1.89$ times larger than the splashback radius defined by the steepest slope in the DM density profile, and it is $\approx1.2$ times larger than the edge of the DM phase-space structure (Figure~\ref{fig:rsp_rsh}).
\item The accretion shock radius of gas decreases with mass accretion rate (MAR) of the halo, similar to the MAR dependence in the splashback radius (Figures~\ref{fig:profiles} and \ref{fig:mar_slope}). The resulting ratio is fairly independent of redshift and baryonic physics for the cluster-size DM haloes (Figure~\ref{fig:rsp_rsh}).
\item The gas density follows DM density at the outer radii with steepest density decrease near the DM splashback radius. The steepest decrease in other thermodynamic (temperature, pressure, and entropy) profiles, however, occurs at the location of the accretion shock (Figure~\ref{fig:profiles_slope}). We do not find significant gas density jump in radial profiles at the accretion shock as the ratio of the post- and pre-shock density is capped at a maximum value of $4$ in the strong shock limit, which is further smeared out by azimuthally averaging around aspherical shocks.
\item Mergers account for on average $\approx 10\%$, and upto $30\%$ increase in the shock to splashback ratio for the duration of $\approx 1$~Gyr right after the merger. After that, the merger-accelerated shock recedes back to the pre-merger accretion shock (Figure~\ref{fig:rsh_rsp_merger}).
\end{enumerate}
Our results have broad implications for using galaxy clusters as probes of astrophysics and cosmology. First, the apparent spatial offset between splashback and the more extended shock radii indicates that the transition boundaries between the 1-halo and 2-halo terms are different between DM and gas, leaving imprints in this transition region as probed by optical and SZ surveys. Second, the extended gas distribution beyond DM splashback must be taken into account when quantifying baryon fraction associated with DM haloes and cosmic web filaments as well as modeling galaxy quenching in the outskirts of clusters. Finally, the low density shock-heated gas between the accretion shock and the splashback radii may contribute to a significant fraction of the missing baryons and may play a role in early quenching of infalling galaxies. Investigating the differential dynamics of collisionless DM and collisional gas in the outskirts of galaxy clusters will be important for using the edges of galaxy clusters as laboratories for cosmology and astrophysics in the era of multi-wavelength cosmological surveys.
\section*{Acknowledgement}
We acknowledge Benedikt Diemer, Andrey Kravtsov, Philip Mansfield, and Xun Shi for useful comments on the draft. We used the publicly available python packages, \textsc{COLOSSUS} \citep{colossus} and \textsc{yt} \citep{yt} as part of our analysis pipeline. This work was supported in part by NSF AST-1412768 and the facilities and staff of the Yale Center for Research Computing.
\section*{Data Availability}
The {\em Omega500} simulations and the analysis code used in this paper are available upon request.
\bibliographystyle{mnras}
|
\section{Introduction}
Let $T$ be a weighted tree. The tree metric induced by $T$, denoted $M_T$, is the complete graph on the vertices of $T$ where the weight of each edge $(u, v)$ is the weight of the path connecting $u$ and $v$ in $T$. For $t\ge 1$, a $t$-spanner for a metric $(V, d)$ is a subgraph $H$ of the complete graph on $V$ such that every pair of distinct points $u, v\in V$ is connected by a path in $H$ of total weight at most $t\cdot d(u, v)$. We refer to such paths as $t$-spanner paths. A $t$-spanner has diameter $\Lambda$ if every pair of points is connected by a $t$-spanner path consisting of at most $\Lambda$ edges. Typically, $t$-spanners are designed to be sparse, often with a linear number of edges. The lightness of a graph is the ratio of its weight to the weight of its minimum spanning tree. Solomon and Elkin \cite{tree-spanner} define a 1-spanner for tree metrics. Given an $n$ vertex weighted tree of maximum degree $\Delta$, the 1-spanner has $O(n)$ edges, $O(\log n)$ diameter, $O(\log n)$ lightness and maximum degree bounded by $\Delta + O(1)$. While being an interesting construction in its own right, this tree metric 1-spanner has been used in a series of papers as a tool for reducing the diameter of various Euclidean and doubling metric spanner constructions \cite{short-thin-lanky,doubling-spanner,optimal-euclidean-spanners,tree-spanner,partitions-to-covers}.
Once a spanner has been constructed, it becomes important to find these short paths efficiently. A local routing algorithm for a weighted graph $G$ is a method by which a message can be sent from any vertex in $G$ to a given destination vertex. The successor to each vertex $u$ on the path traversed by the routing algorithm must be determined using only knowledge of the destination vertex, the neighbourhood of $u$ and possibly some extra information stored at $u$. The efficiency of a routing algorithm is measured by the distance a message needs to travel through a network before reaching its destination as well as by the storage requirements for each vertex. There is a great deal of work on local routing algorithms in the literature. The difficulty of designing a good local routing algorithm clearly depends on the properties of the underlying network. Some authors have designed algorithms for very general classes of networks. For example, the algorithm of Chan et al. \cite{heirarchical-routing-in-doubling-metrics} works in any network although its quality depends on the induced doubling dimension. Many authors have focused on highly efficient algorithms for specific classes of networks. For example, there has been a line of research into routing algorithms for various classes of planar graphs \cite{convex-subdivisions,routing-in-delaunay}. Support for efficient local routing algorithms is a desirable feature in a spanner and in some recent papers, researchers have designed spanners which simultaneously achieve support for efficient local routing with other properties. See, for example, the work of Ashvinkumar et al. \cite{local-routing-in-sparse-and-light}.
There appears to be little work in the literature in designing spanners which achieve both support for local routing as well as low diameter. Given a spanner with diameter $\Lambda$, it is natural to require that a local routing algorithm on this spanner match this diameter, i.e., the algorithm should be guaranteed to terminate after at most $\Lambda$ hops. Abraham and Malkhi \cite{compact-routing-in-euclidean-metrics} give a construction, for any $\epsilon > 0$, of a $(1+\epsilon)$-spanner for points in two dimensional Euclidean space with an accompanying routing algorithm. The diameter of the routing algorithm is $O(\log D)$ with high probability where $D$ is the quotient of the largest and smallest inter-point distances. The routing algorithm has routing ratio $O(\log n)$, with high probability, for $n$ points on a uniform grid.
In this paper, we demonstrate that the 1-spanner construction of Solomon and Elkin \cite{tree-spanner} supports a local routing algorithm with $O(\log n)$ diameter in the worst case. In Section~\ref{app:doubling}, we show that this routing algorithm can be adapted to a class of doubling metric $(1+\epsilon)$-spanners which employ the 1-spanner construction to achieve low diameter while retaining low weight and low degree.
\section{The Model}
A local routing algorithm for a weighted graph $G$ is a distributed algorithm in which each vertex is an independent processor. At the beginning of a round, a node may find that it has received a message. If a message is received, the algorithm decides to which neighbour the message should be forwarded. The following information is available to each vertex $v$ in $G$:
\begin{enumerate}
\item A header contained in the message.
\item The label of $v$ as well as labels of neighbouring vertices.
\end{enumerate}
The message header stores the label of the destination vertex as well as other information if required by the routing algorithm. The labels are used not only as unique identifiers of vertices but may also store additional information if required. The labels of vertices are computed in a pre-processing step before the algorithm is run. Our model does not consider the running time of computation performed at each vertex in each round and so we do not specify the type of data structure used for headers and labels. The routing decision made at each vertex is deterministic. A routing algorithm is evaluated on the basis of the following quality measures.
\begin{itemize}
\item \textit{Routing Ratio}. Given two vertices $u$, $v$ of $G$, let $d_G(u, v)$ denote the shortest path distance from $u$ to $v$ and let $d_{route}(u, v)$ denote the total length of the path traversed by the routing algorithm when routing from $u$ to $v$. The routing ratio is defined to be $\max_{u,v \in V} \left\{ \frac{d_{route}(u, v)}{d_G(u, v)}\right\}$.
\item \textit{Diameter}. A routing algorithm is said to have diameter $\Lambda$ if a message is guaranteed to reach its destination after traversing at most $\Lambda$ edges.
\item \textit{Storage}. A bound on the number of bits stored at vertices and in message headers.
\end{itemize}
\section{Local Routing in Tree Metrics}
In this section we present a slightly modified version of the tree metric 1-spanner of Solomon and Elkin \cite{tree-spanner}. We then present our routing algorithm for this spanner and show that it has a routing ratio of 1.
\subsection{The Spanner}
In this section we describe the tree metric 1-spanner construction of Solomon and Elkin \cite{tree-spanner}.
We first define some notation used in this section. $T$ will denote a weighted, rooted tree and $wt(T)$ will denote its weight. Given a graph $G$, $V(G)$ and $E(G)$ denote its vertex and edge sets respectively. The root of $T$ is denoted $rt(T)$ and $ch(v)$ denotes the number of children of $v$. As in the notation of \cite{tree-spanner}, the children of a vertex $v$ are denoted $c_1(v),...,c_{ch(v)}(v)$ and $p(v)$ denotes the parent of a vertex $v$. The lowest common ancestor of two vertices $u,v\in V(T)$ will be denoted by $lca(u, v)$. Given $u, v\in V(T)$, $P(u, v)$ denotes the unique path from $u$ to $v$ in $T$. Given $v\in V(T)$, $T_v$ will denote the subtree of $T$ rooted at $v$.
The shortcutting procedure selects a constant number of cut vertices in the tree whose removal results in a forest of trees which are at most a constant fraction of the size of the input tree. The spanner is obtained by building the complete graph on these cut vertices and recursively applying the procedure to all subtrees obtained by removing the cut vertices. We note that the original construction appearing due to Solomon and Elkin \cite{tree-spanner} actually builds a low diameter spanner on the cut vertices rather than the complete graph.
We first outline the method by which the cut vertices are selected.
We assume that among all subtrees rooted at children of a vertex $v$, the subtree rooted at the leftmost vertex is the largest. That is, $|T_{c_1(v)}| \ge |T_{c_i(v)}|$, for all $i > 1$.
For an integer $d$, we call a vertex $v$ $d$-balanced if $|T_{c_1(v)}| \le |T| - d$.
We label an edge $(u, v)$ of $T$ as leftmost if $u=c_1(v)$ or $v=c_1(u)$. Let $P(v)$ denote the path of maximum length from $v$ to some descendant of $v$ which includes only leftmost edges. We say the last vertex on $P(v)$ is the leftmost vertex in $T_v$ and we denote it by $l(v)$. We define $l(T):=l(rt(T))$. The construction we describe involves taking subtrees of an input tree. These subtrees inherit the `leftmost' labelling of the input tree so it may be the case that $l(T)=rt(T)$. If there is a $d$-balanced vertex along $P(v)$, we denote the first such vertex by $b_d(v)$. Otherwise, $b_d(v)=NULL$.
Given a rooted tree $T$ and a positive integer $d$, we define a set of vertices $CV(T, d)$ as follows. Set $v:=rt(T)$. If $b_d(v)=NULL$, $CV(T, d)$ is defined to be $\emptyset$.
Otherwise, $$CV(T, d):=\{b_d(v)\}\cup\left(\bigcup_{i=1}^{ch(b_d(v))}CV(T_{c_i(b_d(v))}, d)\right).$$
Let $k$ be a positive integer. We define a set of vertices
$$
C_{T} :=
\begin{cases}
V(T) \mbox{ if } k \ge n/2 - 1, \\
CV(T, d)\cup \{l(T), rt(T)\} \mbox{ otherwise.}
\end{cases}
$$
where $d:=n/k$. (See Figure~\ref{fig:canonical-forest} for an example.) The spanner is constructed via the following recursive procedure which takes as input a tree $T$ and an integer parameter $k \ge 4$. Initialize the spanner as $G=T$. Compute the set $C_{T}$, with respect to $k$, and add the edges of the complete graph on $C_{T}$ to $G$. Denote by $T\setminus C_{T}$ the forest obtained by removing the vertices in $C_{T}$, along with all incident edges, from $T$. Recursively run the algorithm on all trees in the forest $T\setminus C_{T}$ and add the resulting edges to $G$. Note that the parameter $k$ is passed down to recursive calls of the algorithm while the parameter $d$ is recomputed based on the size of the subtree on which the algorithm is called. The following lemmas are established by Solomon and Elkin \cite{tree-spanner}:
\begin{lemma}\label{size-of-cut-set}
For $k\ge2$, the set $C_T$ contains at most $k+1$ vertices.
\end{lemma}
\begin{lemma}\label{elkin-solomon-lemma}
Each tree in the forest $T\setminus C_T$ has size at most $2n/k$.
\end{lemma}
In particular, Lemma \ref{elkin-solomon-lemma} implies that the recursion depth of the spanner construction algorithm is $O(\log n)$ for $k\ge 4$.
\begin{figure}
\centering
\includegraphics[width=30mm,scale=0.15]{figures/canonical_forest.eps}
\caption{A depiction of the set $C_T$ for a tree with $n=17$ vertices and the parameter $k$ set to $k=5$. The vertices of $C_T$ are inside the dashed circles. The vertices and edges of the forest $T\setminus C_{T}$ are shown in bold.}
\label{fig:canonical-forest}
\end{figure}
Solomon and Elkin~\cite{tree-spanner} show that the graph resulting from a slightly more elaborate version of this shortcutting scheme has weight $O(\log n)\cdot wt(T)$. Their scheme differs from what we have presented in that instead of building the complete graph on the set of cut vertices $C_T$, they build a certain 1-spanner with $O(k)$ edges and diameter $O(\alpha(k))$ where $\alpha$ is the inverse Ackermann function. Since we consider the parameter $k$ to be constant, this modification does not affect the weight bound of the construction.
\begin{theorem}[\cite{tree-spanner}]\label{spanner-theorem}
Let $T$ be a weighted rooted tree and let $k$ be a positive integer, $4\le k \le n/2 - 1$.
The graph $G$ obtained by applying the algorithm described above to $T$ using the parameter $k$ is a 1-spanner for the tree metric induced by $T$. Moreover, $G$ has diameter bounded by $O(\log_k n)$, weight bounded by $O(k^2\cdot \log_k n)\cdot wt(T)$ and maximum degree bounded by $\Delta + O(k)$ where $\Delta$ is the maximum degree of $T$.
\end{theorem}
Let $G$ be the spanner resulting from running the algorithm described above on some tree $T$ with respect to the parameter $k$. We define \textit{canonical subtrees} of $T$ with respect to $k$ to be the subtrees computed during the course of the construction of $G$. Canonical subtrees are defined recursively as follows. As a base for the recursive definition, the input tree $T$ is considered to be a canonical subtree with respect to $T$ and $k$. Suppose $T'$ is a canonical subtree with respect to $T$ and $k$. Then each tree in the forest $T'\setminus C_{T'}$ is a canonical subtree with respect to $T$ and $k$. We speak of canonical subtrees without reference to the parameters $T$ and $k$ when they are clear from the context. Given a vertex $v$ of $T$, we denote by $T^v$ the canonical subtree for which $v\in C_{T^v}$. When a canonical subtree $T'$ is small enough, $C_{T'}=V(T')$ and so it is clear that $C_{T^v}$ is well defined for each $v\in V(T)$. We say that $T^v$ is the canonical subtree of $v$ and that $v$ is a cut vertex of $T^v$.
We establish some technical properties of canonical subtrees which will be of use in the following section. We reword statement 4 in Corollary 2.17 from the paper of Solomon and Elkin \cite{tree-spanner} in Lemma~\ref{at-most-two-entrances} and prove it using our terminology for completeness.
\begin{lemma}\label{at-most-two-entrances}
Let $T'$ be a canonical subtree of $T$ and let $T''$ be the canonical subtree such that $T' \in T''\setminus C_{T''}$. There are at most two edges in $T$ with one endpoint in $T'$ and the other outside $T'$. Moreover, any vertex of $T'$ incident to a vertex outside $T'$ must be $rt(T')$ or $l(T')$.
\end{lemma}
\begin{proof}
We first need to establish the following claim.\\
Claim:
Let $u, v \in C_{T'}$ for some canonical subtree $T'$. Suppose $v$ is a descendant of $u'$ where $u'$ is some child of $u$. If $w$ is the first $d$-balanced vertex on the path from $u'$ to the left most vertex in $T'_{u'}$, then $v=w$ or $v$ is a descendant of $w$.\\
Proof of claim:
By definition of the cut vertex procedure, $v$ is a member of the set $CV(T'_{u'}, d)$. If $v=w$, we are done so suppose $v\ne w$. Every cut vertex in $CV(T'_{u'}, d)$ other than $w$ is contained in some $T'_{w'}$ for some child $w'$ of $w$. The claim follows.\\
We note that any edge with one endpoint in $T'$ and the other outside $T'$ must have its other endpoint in $C_{T''}$. In particular, the parent of $x:=rt(T')$ must be an element of $C_{T''}$. Let $v_1$ be the parent of $x$. Suppose there is a second edge from a vertex $v_2 \in C_{T''}$ to a vertex $w$ in $V(T')$. Note that $v_2$ must be a child of $w$ for otherwise $T$ would contain a cycle. Then $v_2$ is a descendant of $x$. Let $x'$ be the first $d$-balanced vertex on the path from $x$ to the left most vertex in $T''_{x}$. Note that $x'\in C_{T''}$. By the claim, $v_2$ either coincides with, or is a descendant of, $x'$. If the latter holds, the parent of $v_2$ cannot be a vertex in $T'$ and so $v_2=x'$. Suppose there exists a third vertex $v_3\in C_{T''}$ which has a parent in $T'$. Then $v_3$ is a member of $C_{T''}$ which is a descendant of $x$ and neither equal to, nor a descendant of, the vertex $x'$ which is impossible given the claim. Since $v_2$ is the first $d$-balanced vertex on the path from $x$ to the left most vertex in $T''_x$, $v_2$ must be the leftmost child of its parent. It follows that the parent of $v_2$ is $l(T')$.
\end{proof}
\begin{lemma}\label{root-of-canonical-subtree}
Let $T'$ and $T''$ be canonical subtrees such that $T''\in T'\setminus C_{T'}$ and let $v$ be some vertex of $T''$. Let $X$ be the set of vertices in $C_{T'}$ which are ancestors of $v$. Let $x$ be the element of $X$ deepest in $T$ and let $x'$ be the child of $x$ which is an ancestor of $v$. Then $x'=rt(T'')$. (See Figure \ref{fig:root-of-canonical-subtree}.)
\end{lemma}
\begin{proof}
Consider the path $P(rt(T'), v)$. Note that all vertices in $X$ lie on this path. It is easy to see that this path enters $T''$ through $rt(T'')$. The vertex on this path preceding $rt(T'')$ is a member of $C_{T'}$, and an ancestor of $v$ and therefore a member of $X$. Since the predecessor of $rt(T'')$ is clearly deeper than all other vertices in $X$, the lemma follows.
\end{proof}
\begin{figure}
\centering
\includegraphics[width=50mm,scale=0.5]{figures/root-of-canonical-subtree.eps}
\caption{Lemma \ref{root-of-canonical-subtree}. The set $X$ consists of the red vertices.}
\label{fig:root-of-canonical-subtree}
\end{figure}
In the spanner, a vertex $u$ is connected to all vertices in $C_{T^u}$. The following lemma ensures that in the routing algorithm described in the next section, under certain conditions, it is safe to make a `greedy' choice from a subset of vertices in $C_{T^u}$.
\begin{lemma}\label{u-tree-in-v-tree-u-ancestor}
Let $u$ and $v$ be vertices in $G$ such that $v$ is not a vertex of $T^u$. Let $X$ be the set of vertices in $C_{T^u}$ which are ancestors of $v$. Suppose $X\ne \emptyset$ and let $x$ be the last vertex on the path from $u$ to $v$ which is contained in $T^u$. Then $x\in X$. Moreover, $x$ is the deepest vertex in $X$. (See Figure \ref{fig:u-tree-in-v-tree-u-ancestor}.)
\end{lemma}
\begin{proof}
By Lemma \ref{at-most-two-entrances}, $x$ is either $rt(T^u)$ or $l(T^u)$. Since $X\ne \emptyset$, $rt(T^u)$ is an ancestor $v$ which also implies $l(T^u)$ is an ancestor of $v$. Then $x$ is an ancestor of $v$ and it remains to be shown that $x$ is the deepest vertex in $X$. Let $x^*$ be the deepest vertex in $X$ and suppose $x^* \ne x$. The path from $x^*$ to $v$ must leave $T^u$ through either $rt(T^u)$ or $l(T^u)$. Since we also have that $x\in \{rt(T^u), l(T^u)\}$ and since $x^*\ne x$, we must have $x=rt(T^u)$ and $x^*=l(T^u)$. Then there would be two distinct paths from $rt(T^u)$ to $v$, a contradiction since $T$ is a tree. The lemma follows.
\end{proof}
\begin{figure}[h!]
\centering
\includegraphics[width=50mm,scale=0.5]{figures/u-tree-in-v-tree-u-ancestor.eps}
\caption{Lemma \ref{u-tree-in-v-tree-u-ancestor}. The set $X$ consists of the red vertices.}
\label{fig:u-tree-in-v-tree-u-ancestor}
\end{figure}
\begin{lemma}\label{u-descendant-of-v}
Let $u$ and $v$ be vertices of $T$ such that $u$ is a descendant of $v$ and $T^v$ is contained in $T^u$. Let $T'$ be the canonical subtree in the forest $T^u\setminus C_{T^u}$ which contains $v$ and let $X$ be the set of vertices in $C_{T^u}$ which are descendants of $v$ and ancestors of $u$. Let $x$ be the element of $X$ which is highest in $T$. Then either $rt(T')$ or $l(T')$ is the parent of $x$. (See Figure \ref{fig:u-descendant-of-v}.)
\end{lemma}
\begin{proof}
Note that $X=P(u, v)\cap C_{T^u}$. Let $w$ be the first vertex on $P(u, v)$ which is contained in $T'$. By Lemma \ref{at-most-two-entrances}, $w\in \{rt(T'), l(T')\}$. Let $x'$ be the predecessor of $w$ on $P(u, v)$. Since $v$ is an ancestor of $u$, $x'$ is a child of $w$. Since $x'$ is a vertex outside $T'$ connected to a vertex in $T'$, $x'\in C_{T^u}$ and hence $x'\in X$. Clearly $x'$ is also the highest vertex in $X$ and so $x'=x$ and the lemma follows.
\end{proof}
\begin{figure}[h!]
\centering
\includegraphics[width=50mm,scale=0.5]{figures/u-descendant-of-v.eps}
\caption{Lemma \ref{u-descendant-of-v}. The set $X$ consists of the red vertices.}
\label{fig:u-descendant-of-v}
\end{figure}
\subsection{Routing Algorithm}
In this section, we describe a local routing algorithm for the spanner defined above. Throughout what follows, let $G$ denote the graph obtained when the algorithm of the previous section is applied to a weighted, rooted tree $T$ using a parameter $k\ge 4$.
We first define the labels $label(v)$ for vertices $v\in V(G)$.
We make use of the interval labelling scheme of Santaro and Khatib \cite{santoro-khatib}.
Let $rank(v)$ denote the rank of $v$ in a post-order traversal of $T$. We define $$L(v):=\min\{rank(w):w\in V(T_v)\}.$$
We define the label of $v$ to be $label(v)=[L(v), rank(v)]$. The observation used in the routing algorithm of Santaro and Khatib \cite{santoro-khatib} is that a vertex $w$ is a descendant of $v$ if and only if $rank(w)\in[L(v), rank(v)]$. Note that the label of each vertex can be computed in linear time in a single traversal of the tree.
\begin{lemma}\label{small-data}
In the labelling scheme outlined above, each vertex of $G$ stores $O((\Delta + k)\log n)$ bits of information.
\end{lemma}
\begin{proof}
Since $G$ has $n$ vertices, for each $v\in V(G)$, $rank(v)$ and $L(v)$ are both integers in the interval $[1,...,n]$ and therefore require $O(\log n)$ bits to be represented. By Theorem \ref{spanner-theorem}, each vertex of $G$ has at most $\Delta + O(1)$ neighbours in $G$. Then, for any $v\in V(G)$, we require $O((\Delta + k)\log n)$ bits to store $rank(w)$ and $L(w)$ for each neighbour $w$ of $v$.
\end{proof}
For our routing algorithm, no auxiliary data structure is required at each vertex and so the total storage requirement per vertex is $O((\Delta + k)\log n)$ bits.
Our routing algorithm considers a number of cases depending on the labels of the current vertex $u$ and destination vertex $v$.
For convenience of analysis, in each case we specify two routing steps. For ease of exposition, we consider a vertex to be both a descendant and ancestor of itself.
\textbf{Case 0:} If $v$ is a neighbour of $u$, route to $v$.\\
\textbf{Case 1:}
$u$ is an ancestor of $v$ in $T$. Let $X$ be the set of vertices in $C_{T^u}$ which are ancestors of $v$. Let $x$ be the deepest element of $X$. Route first to $x$ and then to the child of $x$ which is an ancestor of $u$. \\
\textbf{Case 2:}
$u$ is a descendant of $v$ in $T$. Let $X$ be the set of vertices in $C_{T^u}$ which are descendants of $v$ and ancestors of $u$. Let $x$ be the highest vertex in $X$. Route first to $X$ and then to its parent.\\
\textbf{Case 3:} $u$ is not an ancestor or descendant of $v$. Let $X$ be the set of vertices in $C_{T^u}$ which are ancestors of $v$ and not ancestors of $u$. If $X\ne \emptyset$, we define $x$ to be the deepest vertex in $X$ and define $x'$ to be the child of $x$ which is an ancestor of $v$. Let $Y$ be the set of vertices in $C_{T^u}$ which are ancestors of $u$ but not ancestors of $v$. We define $y$ to be the highest vertex in $Y$.\\
\indent \textbf{Case 3 a):} $X$ is empty. Route first to $y$ and then to the parent of $y$.\\
\\
\indent \textbf{Case 3 b):} $X$ is non-empty. Route first to $x$ and then to $x'$.\\
The routing algorithm uses a greedy strategy. Given the destination vertex and neighbours due to tree edges and shortcut edges, the algorithm simply selects the edge that appears to make the most progress. It is not obvious that this strategy gives the desired $O(\log n)$ diameter of the routing algorithm. Indeed, this bound would not hold if the shortcuts were arbitrary edges and hinges on the particular structure of the 1-spanner.
The arguments we make in our analysis will make use of certain integer sequences we assign to vertices of $G$. Note that these sequences are used only for the analysis of the algorithm and are not part of the labelling scheme. We define a unique integer sequence for each canonical subtree computed during the course of the spanner construction. Each vertex $v$ will be assigned the sequence corresponding to the tree $T^v$.
The integer sequence for each canonical subtree is defined recursively as follows. The input tree $T$ is given the empty sequence. Suppose $T'$ is a canonical subtree which has already been associated with some sequence $S$. Consider the forest $T'\setminus C_{T'} = \{T_1,...,T_p\}$. Each tree $T_j \in T'\setminus C_{T'}$ is associated with the sequence obtained by appending $j$ to $S$. It is clear that every vertex of $T$ appears in $C_{T'}$ for exactly one canonical subtree $T'$. Let $S_v$ denote the sequence assigned to the vertex $v$. We refer to $S_v$ as the canonical sequence of $v$.
Observe that if for two vertices $u, v\in V(G)$ we have that $S_u = S_v$, by definition of the spanner construction algorithm, $u$ and $v$ must be cut vertices of the same canonical subtree of $T$ and are therefore connected by an edge in $G$. The routing algorithm works by choosing a successor vertex so as to incrementally transform the canonical sequence of the current vertex into the canonical sequence of the destination vertex.
\begin{lemma}\label{always deeper after case 1}
Let $u$ and $v$ be vertices of $G$ such that $u$ is an ancestor of $v$. Consider the two vertices visited after executing the routing steps of Case 1 when routing from $u$ to $v$. These vertices are on the path from $u$ to $v$ in $T$. Moreover, these vertices are visited in the order they appear on this path.
\end{lemma}
\begin{lemma}\label{always higher after case 2}
Let $u$ and $v$ be vertices of $G$ such that $u$ is a descendant of $v$. Consider the two vertices visited after executing the routing steps of Case 2 when routing from $u$ to $v$. These vertices are on the path from $u$ to $v$ in $T$. Moreover, these vertices are visited in the order they appear on this path.
\end{lemma}
Lemma \ref{always deeper after case 1} and Lemma \ref{always higher after case 2} are immediate from the definition of the routing steps.
\begin{lemma}\label{case-3-a)-lemma}
Let $u$ and $v$ be vertices of $G$ such that $u$ is not an ancestor or a descendant of $v$. Suppose the set $X$, as defined in Case 3 of the routing algorithm, is empty. Consider the vertices visited after executing Case 3 a) of the routing algorithm. These vertices are on the path from $u$ to the $v$. Moreover, these vertices are visited by the routing algorithm in the order they appear on this path.
\end{lemma}
\begin{proof}
Let $w_1$ be the first vertex visited in Case 3 a) and let $w_2$ be the second. Then $w_1$ is the vertex $y$ and $w_2$ is its parent. Since $y$ is, by definition, not an ancestor of $v$, it lies on the path from $u$ to $lca(u, v)$ and is not equal to $lca(u, v)$. Since $w_2$ is the parent of $y$, it is clearly the next vertex on $P(u, lca(u, v))$.
\end{proof}
\begin{lemma}\label{case-3-b)-lemma}
Let $u$ and $v$ be vertices of $G$ such that $u$ is not an ancestor or descendant of $v$. Suppose the set $X$, as defined in Case 3 of the routing algorithm, is non-empty. Consider the vertices visited after executing Case 3 b) of the routing algorithm. These vertices are on the path from $lca(u, v)$ to $v$. Moreover, these vertices are visited in the order they appear on this path.
\end{lemma}
\begin{proof}
Let $w_1$ be the first vertex visited in Case 3 b) and let $w_2$ be the second. $w_1$ is an ancestor of $v$ which is not an ancestor of $u$. It follows that $w_1$ is a descendant of $lca(u,v)$. By definition of Case 3 b), $w_2$ is the child of $w_1$ which is an ancestor of $v$. It is clear that $w_2$ is the successor to $w_1$ on the path $P(lca(u, v), v)$. The lemma follows.
\end{proof}
Lemmas \ref{always deeper after case 1}, \ref{always higher after case 2}, \ref{case-3-a)-lemma} and \ref{case-3-b)-lemma} imply the following:
\begin{lemma}\label{edge-on-path}
Let $u$ and $v$ be vertices of $G$ and let $P(u, v)=(u=x_1,...,x_p=v)$. Suppose the routing steps of a single case of the routing algorithm are executed and vertices $w_1$ and $w_2$ are visited. Then there are indices $1\le i_1\le i_2 \le p$ such that $w_1 = x_{i_1}$ and $w_2=x_{i_2}$.
\end{lemma}
Using Lemma \ref{edge-on-path}, we can show the routing algorithm has a routing ratio of 1.
\begin{theorem}\label{routing-ratio}
Let $u$ and $v$ be vertices of $G$. Let $\delta_{T}(u, v)$ denote the length of the path from $u$ to $v$ in $T$. The routing algorithm described above is guaranteed to terminate after a finite number of steps and the length of the path traversed is exactly $\delta_{T}(u, v)$.
\end{theorem}
\begin{proof}
Lemma \ref{edge-on-path} implies that the vertices visited when any case of the routing algorithm is executed are on the path from the current vertex to the destination and they are explored in the order they appear on the path. Each case leads to at least one new vertex being explored. If $P=(u=x_1,x_2,...,x_p=v)$ is the path from $u$ to $v$ in $T$, the arguments made above imply that the path traversed by the routing algorithm is of the form $P'=(x_{i_1}, x_{i_2},...,x_{i_l})$ where $1\le i_1<i_2<...<i_l\le p$, i.e., $P'$ is a sub-sequence of $P$. In the tree metric $M_{T}$, the weight of the edge $(x_{i_j}, x_{i_{j+1}})$ is by definition equal to the weight of the path $(x_{i_j}, x_{i_j+1}, ..., x_{i_{j+1}})$ and so the theorem follows.
\end{proof}
We now argue that the routing algorithm is guaranteed to terminate after traversing $O(\log n)$ edges. To this end, we first prove the following lemma.
\begin{lemma}\label{argument-about-canonical-sequences-in-case-1-and-case-2}
Let $u$ and $v$ be vertices of $T$ such that $u$ is either an ancestor or a descendant of $v$. Let $u'$ be the vertex reached after executing the routing steps of either Case 1 or Case 2 when routing from $u$ to $v$. Then the following statements hold:
\begin{enumerate}
\item If $S_u$ is a prefix of $S_v$, then $|S_{u'}| > |S_u|$. Moreover, either $S_{u'} = S_v$ or $S_{u}'$ is a prefix of $S_v$.
\item If $S_v$ is a prefix of $S_u$, then $|S_{u'}| < |S_u|$. Moreover, either $S_{u'} = S_v$ or $S_v$ is a prefix of $S_{u'}$.
\item Suppose $S_u$ and $S_v$ share a common prefix $S$ of length $m < \min\{|S_u|, |S_v|\}$. Then $|S_{u'}| < |S_u|$. Moreover, either $S_{u'} = S$ or $S$ is a prefix of $S_{u'}$
\end{enumerate}
\end{lemma}
\begin{proof}
We begin with the first statement. Suppose $S_u$ is a prefix of $S_v$. Then $T^u$ contains $T^v$. Let $T'$ be the canonical subtree in the forest $T^u\setminus C_{T^u}$ which contains $T^v$. By definition, the canonical sequence of any cut vertex of $T'$ can be obtained by appending some integer to $S_u$. Since $T^v$ is contained in $T'$, the canonical sequence associated to $T'$ is a prefix of $S_v$. Then it is sufficient to show that $u'\in C_{T'}$. Suppose $u$ is an ancestor of $v$ so that the algorithm executes the routing steps of Case 1. Recall that in Case 1 the set $X$ is defined as the set of vertices in $C_{T^u}$ which are ancestors of $v$ and the vertex $x$ is defined as the deepest vertex in $X$. Then by definition of Case 1, $u'$ is the child of $x$ which is an ancestor of $v$. By Lemma \ref{root-of-canonical-subtree}, $u'=rt(T')$. Since $rt(T')\in C_{T'}$, the statement of the lemma follows in this case. Now suppose $u$ is a descendant of $v$ so that the algorithm executes the routing steps of Case 2. Recall that in Case 2, $X$ is defined to be the set of vertices in $C_{T^u}$ which are descendants of $v$ and ancestors of $u$ and $x$ is defined to be the highest vertex in $X$. By definition of Case 2, the algorithm routes to $x$ and then to the parent of $x$. Then $u'$ is the parent of $x$. By Lemma \ref{u-descendant-of-v}, $u'\in \{l(T'), rt(T')\}$. Since $\{l(T'), rt(T')\}\subseteq C_{T'}$, the first statement of the lemma holds in this case.
We now address the second and third statements of the lemma. Suppose $S_u$ is not a prefix of $S_v$. Let $S$ be the longest common prefix of $S_u$ and $S_v$. Then either $S=S_v$ or $S$ is a prefix of $S_v$. Let $T''$ be the canonical subtree corresponding to the canonical sequence $S$. Observe that $T''$ contains $T^u$. Let $T'$ be the canonical subtree such that $T^u\in T'\setminus C_{T'}$. Then either $T''=T'$ or $T''$ contains $T'$. Note that the canonical sequence of any cut vertex of $T'$ is a prefix of $S_u$. Moreover, for any canonical sequence $S'$ of a cut vertex in $T'$, either $S=S'$ or $S$ is a prefix of $S'$. We claim that $u'\in C_{T'}$. When $S_v$ is a prefix of $S_u$, we see this implies the second statement. When $S_u$ and $S_v$ share a prefix of length $m < \min\{|S_u|, |S_v|\}$, we see that our claim implies the third statement.
We now prove the claim. Suppose that $u$ is an ancestor of $v$ so that the algorithm executes the routing steps of Case 1. Let $X$ and $x$ be as defined in Case 1. Then $u'$ is the child of $x$ which is an ancestor of $v$. By Lemma \ref{u-tree-in-v-tree-u-ancestor}, $x$ is the last vertex on the path from $u$ to $v$ which is contained in $T$. Since $x$ is an ancestor of $v$ and $u'$ is both a child of $x$ and ancestor of $v$, it is clear that $u'$ is the next vertex on $P(u, v)$. Since $u$ is a vertex outside $T'$ connected to a vertex in $T^u$, we see that $u'\in C_{T'}$ and so the claim holds in this case. Suppose now that $u$ is a descendant of $v$ so that the algorithm executes the routing steps of Case 2. Let $X$ and $x$ be as defined in Case 2. Note that since $u$ is a descendant of $v$, $rt(T^u)$ must also be a descendant of $v$. Then $rt(T^u)\in X$. Since $rt(T^u)$ must be the highest vertex in $X$, we see that $x=rt(T^u)$. Since the parent of $rt(T^u)$ is clearly a member of $C_{T'}$, the claim holds. This completes the proof of the lemma.
\end{proof}
Note that if $u$ is an ancestor (resp. descendant) of $v$, then by Lemma \ref{edge-on-path} the vertex reached after executing the steps of Case 1 (resp. Case 2) will also be an ancestor (resp. descendant) of $v$.
\begin{lemma}\label{logn-steps-when-above-or-below}
Suppose $u$ and $v$ in $G$ are such that $u$ is an ancestor or descendant of $v$ in $T$. Then, when routing from $u$ to $v$, the routing algorithm reaches $v$ after traversing $O(\log n)$ edges.
\end{lemma}
\begin{proof}
Suppose that $u$ is an ancestor of $v$ (the argument in the case where $u$ is a descendant of $v$ is symmetric) and let $K$ be the maximum length of the canonical sequence of any vertex of $G$. Note that by Lemma \ref{elkin-solomon-lemma}, each canonical subtree is at most $2/k$ times the size of its smallest containing canonical subtree and so the recursion depth is at most $O(\log n)$ which, by definition of canonical sequences, implies $K=O(\log n)$. If $S_u$ is a prefix of $S_v$, by Lemma \ref{argument-about-canonical-sequences-in-case-1-and-case-2}, the routing steps executed in a single iteration of Case 1 traverse at most two edges and lead to a vertex $u'$ which is an ancestor of $v$ and for which $S_{u'}$ is a prefix of $S_v$ strictly longer than $S_u$. Then after traversing at most $2\cdot K$ edges, the routing algorithm will reach $v$. An analogous argument can be made in the case where $S_v$ is a prefix of $S_u$.
\\
Assume that $S_u$ and $S_v$ have a common prefix of length $m < \min\{|S_u|, |S_v|\}$. By Lemma \ref{argument-about-canonical-sequences-in-case-1-and-case-2}, after traversing at most $2\cdot K$ edges, the algorithm reaches a vertex $u'$ which is an ancestor of $v$ and is such that $S_{u'}$ is a prefix of $S_v$. By the argument made above, after traversing at most another $2\cdot K$ edges, the routing algorithm reaches $v$. Then in all cases, after traversing at most $4\cdot K$ edges, the algorithm reaches its destination. This completes the proof of the lemma.
\end{proof}
Consider the case where $u$ is neither an ancestor nor a descendant of $v$. The following lemma shows that in this case, the algorithm either routes to a vertex on the path $P(lca(u, v), v)$ or it follows the routing steps that would be executed if the algorithm were routing from $u$ to $lca(u, v)$.
\begin{lemma}\label{case-2-simulation}
Let $u$ and $v$ be vertices of $G$ such that $lca(u, v)\notin \{u, v\}$. Suppose that the set $X$ as defined in Case 3 is empty so that the algorithm executes the routing steps of Case 3 a) when routing from $u$ to $v$. The same steps would be performed when routing from $u$ to $lca(u, v)$.
\end{lemma}
\begin{proof}
Recall the set $Y$ in Case 3 is defined as the set of ancestors of $u$ which are not ancestors of $v$. Consider Case 2 when routing from $u$ to $lca(u, v)$. The set $X$ in Case 2 is defined as the set of ancestors of $u$ which are descendants of $lca(u, v)$. Since an ancestor of $u$ is a descendant of $lca(u, v)$ if and only if it is not an ancestor of $v$, we see that $X=Y$. In Case 3 a) the algorithm routes to the highest vertex in $Y$ and then its parent. In Case 2 the algorithm routes to the highest vertex in $X$ and then its parent. We see the algorithm executes the same routing steps in both cases and so the lemma follows.
\end{proof}
Using Lemmas \ref{logn-steps-when-above-or-below} and \ref{case-2-simulation}, we establish the main result of this section.
\begin{theorem}\label{main-theorem}
Let $u$ and $v$ be vertices in $G$. The routing algorithm reaches $v$ when routing from $u$ after traversing at most $O(\log n)$ edges.
\end{theorem}
\begin{proof}
Lemma \ref{logn-steps-when-above-or-below} implies the theorem when $u$ is an ancestor or descendant of $v$. Then we need only consider the case where $u$ is neither an ancestor nor descendant of $v$. Suppose the algorithm executes the steps of Case 3 a) at some point when routing from $u$ to $v$. Let $u'$ be the current vertex after the steps of Case 3 a) are executed. Then $u'$ is an ancestor of $v$ and so the algorithm will reach $v$ after traversing another $O(\log n)$ edges. It remains to argue that $O(\log n)$ edges are traversed before the current vertex is an ancestor of $v$. By Lemma \ref{case-2-simulation} and Lemma \ref{logn-steps-when-above-or-below}, the algorithm either executes the steps of Case 3 a) or reaches $lca(u, v)$ after traversing $O(\log n)$ edges. This completes the proof of the theorem.
\end{proof}
\section{Routing in a Doubling Metric Spanner}
\label{app:doubling}
The 1-spanner of the previous section was designed to be a tool for reducing the diameter of Euclidean spanners without compromising bounds on degree and weight. It has been applied in a series of papers on various constructions of spanners of doubling metrics. The doubling dimension of a metric $(M, d)$ is the smallest number $\lambda$ such that, for any $r > 0$, any ball of radius $r$ can be covered by at most $2^\lambda$ balls of radius $r/2$. An $n$-point metric is said to be doubling if the doubling dimension is constant in $n$. The class of doubling metrics contains the class of Euclidean metrics as $m$-dimensional Euclidean metrics can be shown to have doubling dimension $\Theta(m)$. The spanners for doubling and Euclidean metrics to which the 1-spanner shortcutting scheme has been applied are all based on some variety of hierarchical decomposition of the input point set. In particular, spanners based on the dumbbell trees of Arya et al. \cite{short-thin-lanky} and the net tree approach \cite{def-spanner-and-app,fast-construction-of-nets} have been proposed. We will show that the routing algorithm described in the previous section can be extended to a routing algorithm on a low diameter doubling metric $(1+\epsilon)$-spanner based on the net tree.
\subsection{A net tree based spanner}
The spanner we describe here is a simplified version of the spanner described by Chan et al. \cite{doubling-spanner}. The starting point of their construction is based on the spanner of Gao et al. \cite{def-spanner-and-app}.\\
In this section, $(M, d)$ will denote an $n$-point doubling metric. Let $\dim(M)$ denote the doubling dimension of $M$. We assume without loss of generality that the distance between the closest pair of points in $M$ is 1.
\begin{definition}
For $r>0$, a subset $N\subseteq M$ is said to be an $r$-net for $M$ if \begin{itemize}
\item $d(x_1, x_2) > r$ for all distinct pairs of points $x_1, x_2\in N$.
\item For all $y\in M$, there exists $x\in N$ such that $d(x, y)\le r$.
\end{itemize}
\end{definition}
The first step of the construction is to compute a certain sequence of nets. Specifically, we compute a nested sequence of sets $M=N_0\supseteq N_1\supseteq ... \supseteq N_{\log D}$ where $N_i$ is a $2^i$-net of $N_{i-1}$ and $D$ is the largest inter-point distance among pairs of points in $M$. Note that $N_{\log D}$ consists of a single point. An efficient method for computing such a sequence of nets is given by Har-Peled and Mendel \cite{fast-construction-of-nets}. Next, we describe the so-called net tree which we denote $\mathcal{T}$. The net tree has a node at level-$i$ for each point in the set $N_i$. If a level-$i$ node corresponds to a point $p\in N_i$, we say that $p$ is the representative of $v$ and write $rep(v)=p$. Note that a given point in $p$ can be the representative of multiple nodes in $\mathcal{T}$. The root of the tree corresponds to the singleton set $N_{\log D}$ and the leaves of $\mathcal{T}$ are in one-to-one correspondence with the points of $M$. The parent of each node $v$ at level-$i$ is a node $w$ at level-($i+1$) for which $d(rep(v), rep(w)) \le 2^{i+1}$. The constant doubling dimension of $M$ immediately implies that nodes in $\mathcal{T}$ have a constant number of children. The tree induces a graph $H$ on $M$ in the natural way, i.e., each edge $(u, v)\in E(\mathcal{T})$ corresponds to the edge $(rep(u), rep(v))\in E(H)$. For a point $p\in M$, we also use $p$ to denote the leaf of $\mathcal{T}$ which has $p$ as its representative. Given a leaf $p$, we denote by $p^{(i)}$ the level-$i$ ancestor of $p$ in $\mathcal{T}$. Given two nodes $u, v\in V(\mathcal{T})$, we define $d(u, v):=d(rep(u), rep(v))$. The graph $H$ as currently defined is not a $(1+\epsilon)$-spanner for $M$. To guarantee a $1+\epsilon$ spanning ratio, certain cross edges are added at each level. Fix a constant $\gamma > 4$. For each pair of level-$i$ points $u, v\in V(\mathcal{T})$ for which $d(rep(u), rep(v))\le \gamma\cdot2^i$, we add the cross edge $(u, v)$ to $\mathcal{T}$. If we make a sufficiently large choice of $\gamma = O(1/ \epsilon)$, the graph $H$ resulting from the addition of these cross edges to $\mathcal{T}$ can be shown to be a $(1+\epsilon)$-spanner for $M$. The following theorem is established in the proof of Theorem 3.2 in~\cite{def-spanner-and-app}.
\begin{theorem}\label{structure-of-spanning-paths}
Let $p, q$ be points in $M$ and let $i$ be the smallest integer such that the level-$i$ ancestors of $p$ and $q$ are connected by a cross edge. Consider the path in $\mathcal{T}$ obtained by climbing from $p$ to its level-$i$ ancestor, taking the cross edge to the level-$i$ ancestor of $q$ and descending to $q$. Then the path in $H$ corresponding to this path has total length $(1+\epsilon)\cdot d(p, q)$.
\end{theorem}
The maximum degree of the spanner as described above is $O(\epsilon^{-\dim(M)}\log D)$ as some points in $M$ may represent many nodes in $\mathcal{T}$. Gottlieb and Roddity~\cite{improved-algorithms-for-fully-dynamic} devised a technique to reduce the degree bound to $O(\epsilon^{-dim(M)})$ by carefully reassigning representatives.
The depth of the net tree is $O(\log D)$. In particular, if $D$ is not bounded by a polynomial in $n$, the net tree may have linear depth and so the diameter of the spanner may be linear. We say a subtree of $\mathcal{T}$ is \textit{light} if it has height $\log(\frac{D}{n})$. Chan et al. \cite{doubling-spanner} show that by applying the tree metric construction of the previous section to all light subtrees of $\mathcal{T}$, it is possible to reduce the diameter of the spanner to $O(\log n)$ without increasing the asymptotic bound on the weight. Chan et al. \cite{doubling-spanner} show that the weight of the spanner, after shortcuts are added, is at most $O(\epsilon^{-dim(M)}\log n)\cdot wt(MST(M))$, where $MST(M)$ is a minimum spanning tree of $M$. Chan et al. \cite{doubling-spanner} make additional modifications to ensure fault tolerance although we will not work with this version of the spanner.
To summarise, the spanner we work with has $(1+\epsilon)$-stretch, $O(\epsilon^{-\dim(M)})$ degree, $O(\epsilon^{-dim(M)}\log n)\cdot wt(MST(M))$ weight and $O(\log n)$ diameter.
For the remainder of this section, $\mathcal{T}^*$ will denote the net tree with cross edges and shortcuts. $H$ denotes the constant degree spanner of $(M, d)$ induced by $\mathcal{T}^*$.
Our routing algorithm will make use of the following property of the spanner~$H$.
\begin{lemma}\label{interval-lemma}
For $p, q\in M$, the smallest integer $i$ such that the level-$i$ ancestors of $p$ and $q$ are guaranteed to be joined by a cross edge lies in the interval $$\left[\left\lfloor \log\left(\frac{d(p, q)}{\gamma + 4}\right) \right\rfloor, \left\lceil \log\left(\frac{d(p, q)}{\gamma - 4}\right) \right\rceil \right].$$
\end{lemma}
\begin{proof}
Lemma 4.1 in the paper of Gao et al. \cite{def-spanner-and-app} implies that if $(p^{(i)}, q^{(i)})$ is a cross edge, then $(p^{(j)}, q^{(j)})$ is a cross edge for all $j\ge i$. Then $i$ is the unique integer for which $(p^{(i)}, q^{(i)})$ is a cross edge and $(p^{(i-1)}, q^{(i-1)})$ is not. Observe that for $j<\log\left( \frac{d(p, q)}{\gamma + 4} \right)$, we have $d(p, q) > (\gamma + 4)\cdot2^j$.
By definition of the net tree, $d(p^{(j)}, p) \le \sum_{k=1}^{j}2^k \leq 2\cdot2^j$. By the triangle inequality we have $d(p, q) \le d(p^{(j)}, q^{(j)}) + d(p,p^{(j)}) + d(q, q^{(j)}) \leq d(p^{(j)}, q^{(j)}) + 4\cdot2^j$ and so $d(p^{(j)}, q^{(j)}) \ge d(p, q) - 4\cdot2^j > \gamma\cdot2^j.$ Then by the construction of the spanner, there is no cross edge between $p^{(j)}$ and $q^{(j)}$. A similar argument shows that for $j > \left\lceil \log\left(\frac{d(p, q)}{\gamma - 4}\right) \right\rceil$, $d(p^{(j)}, q^{(j)}) \le \gamma \cdot 2^j$ and so there is guaranteed to be a cross edge between $p^{(j)}$ and $q^{(j)}$. Then since $(p^{(j)}, q^{(j)})$ is not a cross edge for $j=\left\lfloor \log\left(\frac{d(p, q)}{\gamma + 4}\right) \right\rfloor-1$ and is guaranteed to be a cross edge for $j=\left\lceil \log\left(\frac{d(p, q)}{\gamma - 4}\right) \right\rceil$, the lemma follows.
\end{proof}
Note that the length of the interval computed in Lemma \ref{interval-lemma} is $\log\left( \frac{\gamma+4}{\gamma-4} \right) = O(1)$. This fact is needed to establish a bound on the diameter of the routing algorithm we present in the next section.
\subsection{Labelling Scheme and Routing Algorithm}
Gottlieb and Roddity \cite{improved-algorithms-for-fully-dynamic} show it is possible to route in the net tree in their construction while taking care to use the first available cross edge. This algorithm has a routing ratio of $1+\epsilon$ and the required labels are short as a result of the degree bound. We apply their method to the version of the spanner have described although we make use of our 1-spanner routing algorithm whenever the message passes through short subtrees so as to take advantage of the available shortcuts.
We first describe the labels of $\mathcal{T}^{*}$. Since each point in $p\in M$ is the representative of multiple nodes in $\mathcal{T}^{*}$, the label of $p$ will contain the labels of all nodes $v\in V(\mathcal{T}^{*})$ for which $p=rep(v)$. Note that there is a constant number of such nodes in $\mathcal{T}^{*}$ for each $p\in M$.
Initially, each node $v\in V(\mathcal{T}^{*})$ is labelled with the interval $[L(v), rank(v)]$, the same labelling scheme used in our 1-spanner routing algorithm. Next, for each light subtree $\mathcal{T}'$ of $\mathcal{T}$, we label the vertices of $\mathcal{T}'$ with the labels required for the 1-spanner routing algorithm on $\mathcal{T}'$.
When routing from a point $p$ to a point $q$ in $M$, in order to obtain a $(1+\epsilon)$ routing ratio, the algorithm must make use of the first cross edge connecting ancestors of $p$ and $q$. Since we make use of shortcuts, there is a danger the algorithm may `overshoot' the correct level. Therefore, the algorithm requires a means to estimate the level of the tree containing a viable cross edge, i.e., the algorithm should be able to locally compute the interval of Lemma \ref{interval-lemma}. If $M$ is a Euclidean metric, the distance $d(p, q)$ can be computed if the labels of the points store their coordinates. However, if $M$ is an arbitrary doubling metric, it is not obvious how to achieve this goal. For $\delta > 0$, a $(1+\delta)$-ADLS is a labeling of points in a metric space that allows approximation of pairwise distances to within a factor of $1+\delta$. That is, given the labels of points $x$ and $y$ in a $(1+\delta)$-ADLS, we are able to compute a value $\tilde{d}(x, y)$ such that $$d(x, y) \le \tilde{d}(x, y) \le (1+\delta)d(x, y).$$ Har-Peled and Mendel \cite{fast-construction-of-nets} give a $(1+\delta)$-ADLS for doubling metrics in which each label consists of at most $\min\{\delta^{-O(\dim(M))}\cdot\log D, \delta^{-O(\dim(M))}\cdot \log n(\log n + \log\log D)\}$ bits. If we work with a $(1+\delta)$-ADLS rather than exact distances, the interval of Lemma~\ref{interval-lemma} becomes $$\left[\left\lfloor \log\left(\frac{\tilde{d}(p, q)}{(\gamma + 4)(1+\delta)}\right) \right\rfloor, \left\lceil \log\left(\frac{\tilde{d}(p, q)}{(\gamma - 4)}\right) \right\rceil \right].$$ Note that the length of this interval is still $O(1)$.
To summarize, the label of each point $p\in M$ contains:
\begin{itemize}
\item The label of $p$ in a $(1+\delta)$-ADLS for $M$.
\item The label of each node $v\in V(\mathcal{T}^{*})$ for which $p=rep(v)$.
\end{itemize}
Since the degree of the spanner is bounded by $O(\epsilon^{-dim(M)})$, we obtain the following bound on the storage of our labelling scheme.
\begin{lemma}\label{size-of-doubling-spanner-labels}
In the labelling scheme for the spanner $H$ defined above, each label requires at most $O(\epsilon^{-dim(M)}\log n) + \eta$ bits of storage, where $\eta = \min\{\delta^{-O(dim(M))}\cdot\log D, \delta^{-O(dim(M))}\cdot \log n (\log n + \log\log D)\}$.
\end{lemma}
We now describe the routing algorithm for the spanner $H$. Let $p$ be the current vertex and $q$ be the destination. Since each point of $H$ is the representative of multiple vertices in $\mathcal{T}^{*}$, the message header stores a $O(\log n)$ bit variable to keep track of the current position in $\mathcal{T}^{*}$. Suppose the current position is a vertex $v\in \mathcal{T}^{*}$ with representative $p$. When we say the algorithm routes to a neighbour $w$ of $v$, we mean the message is forwarded to a neighbour $p'$ of $p$ in $H$ such that $rep(w)=p'$. Throughout the following, $u$ will denote the current position of the algorithm in $\mathcal{T}^{*}$ and $v$ will denote the leaf in $\mathcal{T}^{*}$ corresponding to the destination vertex in~$H$.\\
The algorithm operates in the following distinct states.
\textbf{Initialization:} Using the ADLS labels stored in the labels of $p$ and $q$, a $(1+\delta)$-approximation of $d(p, q)$ is computed. Let $\tilde{d}(p, q)$ denote this approximation. The integer $i:=\left\lfloor \log\left(\frac{\tilde{d}(p, q)}{(1 + \delta)(\gamma + 4)}\right) \right\rfloor$ is stored in the message header as the cross edge target level.\\
\textbf{Ascending: } In this state, the current position $u$ of the message in $\mathcal{T}^{*}$ is at a level lower than $i$. Suppose $u$ is in a light subtree. Then we use our 1-spanner routing algorithm to route towards the level-$i$ ancestor of $u$. Otherwise, we simply route to the parent of $u$.\\
\textbf{Searching for a cross edge: } In this state, $u$ is at level $i$ or higher in $\mathcal{T}^{*}$. If $u$ is connected by a cross edge to an ancestor of $v$, the algorithm routes to this ancestor. Otherwise, the algorithm routes to the parent of $u$.\\
\textbf{Descending: } In this state, $u$ is an ancestor of $v$ in $\mathcal{T}^{*}$. If $u$ is in a light subtree, the algorithm routes towards $v$ using our 1-spanner routing algorithm. Otherwise, the algorithm routes to the child of $u$ which is an ancestor of $v$.
\begin{theorem}
The routing ratio of the algorithm defined above has routing ratio $1+\epsilon$ and diameter $O(\log n)$.
\end{theorem}
\begin{proof}
Observe that the path traversed when routing from $p$ to $q$ by the algorithm differs from the $(1+\epsilon)$ spanning path described in Theorem \ref{structure-of-spanning-paths} only by the shortcuts taken on sections of the path which pass through light subtrees. Since these shortcuts cannot increase the length of the path, the first statement of the theorem follows.
To prove the second statement, we bound the number of vertices visited in each state. In the ascending state, at most $O(\log n)$ vertices are visited in a light subtree since the 1-spanner routing algorithm has $O(\log n)$ diameter. Since there are at most $O(\log n)$ levels of the net tree higher than level $\log(\frac{D}{n})$, at most $O(\log n)$ vertices are visited by the algorithm outside a light subtree. The same arguments hold for the descending state.
Note that the level at which to begin the search for a cross edge determined in the initialization state is at most a constant number of levels lower that the lower endpoint of the interval of Lemma~\ref{interval-lemma}. Since the length of the interval in Lemma \ref{interval-lemma} has constant length, the second statement of the theorem follows.
\end{proof}
\section{Conclusion}
We have shown that a simplified version of the logarithmic diameter tree metric 1-spanner of Solomon and Elkin \cite{tree-spanner} supports a local routing algorithm of routing ratio 1 and logarithmic diameter. We have also shown that this algorithm can be used to route in a spanner for doubling metrics which uses the shortcutting scheme to achieve a low diameter while maintaining low weight and low degree.
\bibliographystyle{plain}
|
\section*{ACKNOWLEDGMENTS}
The 1.85-m telescope project is promoted on a lot of contributions of people engaged in this project, so the authors would like to thank all those people, Takafumi Kojima, Taku Nakajima, Toshihisa Tsutsumi, Ryoko Amari, Kozuki Yuto, Noriaki Arima, Ryosuke Kiridoshi, Takao Matsumoto, Shigeki Osaki, Touga Shiori, Minato Kozu, Yuya Ota, Yoshiharu Kojima, Akio Hashizume, Akihito Minami, Tsubasa Sakaguchi, Hidetoshi Tsuji, Tetsuya Katase, Shimpei Yashima, Masato Kunizane, Masahiro Minowa, Yukimasa Takenaka, Shinpei Yashima, Hirotaka Kurimoto, Atsushi Ezaki, Kazutoshi Maruyama, Hirofumi Okuno, Takashi Nohara, Kiyoko Tsuji, Yoshihide Tokko, Kazuki Toki, Tatsuro Nakayama, and Jun Korogi.
We are also grateful to Shigeru Fuji, Akira Mori and Hiroyuki Iwashita, and the entire staff of the Nobeyama Radio Observatory for their useful support.
This work was supported by JSPS KAKENHI Grant Numbers JP18H05440, JP15K05025, JP26247026, JP14J12320, JP22244014, and JP15071205, by the Mitsubishi Foundation and the Toray Science Foundation.
\section{INTRODUCTION}
\label{sec:intro}
The rotational transition line of CO is one of the most important emission line in astrophysics to investigate the distribution, physical properties, and kinematics of the interstellar medium (ISM).
Especially, the lines of CO $J=$1--0 and its isotopes have been widely used as a probe of the mass of molecular contents of the ISM, because of its high abundance, low critical density for excitation, and relatively stable chemical properties.
The distribution of the line in the Galaxy has been explored with relatively small-aperture telescopes, such as the 1.2-m telescopes at the Harvard-Smithsonian Center for Astrophysics (CfA) and Cerro Tololo Inter-American Observatory in Chile\cite{1985ApJ...297..751D, 1987ApJ...322..706D, 2001ApJ...547..792D}, as well as the 4-m telescope in Nagoya and the NANTEN telescope in Chile\cite{1998IAUS..179..165F, 2004dimg.conf..203O, 2004ASPC..317...59M, 2020SPIE.nishimura.nasco}.
Although the $J=$1--0 lines of CO are powerful tools to investigate the mass of the molecular cloud, the other transitions with different critical densities for the excitation are needed to investigate the local density and the temperature of the ISM.
However, such observations targeting the higher transition lines of CO were conduced only at coarse angular resolutions such as a survey of 60-cm telescope\cite{1994ApJ...425..641S, 1995ApJS..100..125S, 2001ApJ...551..794S, 2010PASJ...62.1277Y} or only toward small regions such as a survey of JCMT/HARP\cite{2010MNRAS.401..204B, 2012MNRAS.422..521B}, at the time of early 2010s, mainly because the development of sensitive receivers at high frequencies has been very difficult and also because the opacity of the Earth's atmosphere is high at low altitude site.
The 1.85-m mm-submm telescope was developed to realize a large-scale CO $J=$2--1 survey.
The first science operation was started on November 2011 by using a sideband-separating (2SB) SIS receiver, and the CO survey was continued several years.
The developments of receiver was continued during the operation, and the 1.85-m telescope was also used as a test bench of the new receiver concepts.
In this paper, we describe the history of the developments and observations of the 1.85-m telescope, current status of the project and the future plan in detail.
\section{DEVELOPMENT OF RADIO INSTRUMENTS}
\label{sec:inst}
The 1.85-m telescope\cite{2013PASJ...65...78O} has a Cassegrain reflector antenna with Nasmyth beam-waveguide feed, and is mounted on an azimuth-elevation rotating structure (Fig \ref{fig:telescope}).
The main reflector is made of carving of one piese of aluminium, and its surface accuracy was measured to be 19 $\mu$m by a 3-D coordinate measuring machine, SNK MM-3500.
The radome is equipped to protect the telescope from wind, precipitation, and sunlight, that was originally used for the 1.2-m telescope installed at the summit of Mt. Fuji\cite{2000SPIE.4015..185S}.
The path length modulator\cite{2002.JIMRW.Mizuno} (PLM) was installed in front of the RF window of the receiver to reduce standing waves between the RF window and the calibration unit or the Al frames of the radome.
For more detail on the antenna and optics, see Ref. \citenum{2013PASJ...65...78O}.
\begin{figure} [ht]
\begin{center}
\begin{tabular}{c}
\includegraphics[width=9cm, bb=0 0 480 410]{fig/1p85.jpg}
\end{tabular}
\end{center}
\caption[example]
{ \label{fig:telescope}
Photograph of the 1.85-m telescope at the Nobeyama Radio Observatory.
}
\end{figure}
The construction of the telescope at the Nobeyama Radio Observatory (NRO) was started on 2007, and the first simultaneous observations of the $^{12}$CO, $^{13}$CO, and C$^{18}$O of $J=$2--1 transition lines were achieved in September 2009.
The beam size was measured to be 2.7$'$ at 230 GHz band.
Fig. \ref{fig:history} summarizes the receiver system and control system developed and used for the 1.85-m telescope.
\begin{figure} [ht]
\begin{center}
\begin{tabular}{c}
\includegraphics[width=10cm, bb=0 0 956 508]{fig/history.png}
\end{tabular}
\end{center}
\caption[example]
{ \label{fig:history}
History of the installed receiver and control system of the 1.85-m telescope.
}
\end{figure}
\subsection{Receiver System}
\label{sec:rx}
\subsubsection{Season 2009--2012}
The first light receiver for the 1.85-m telescope was a balanced-type sideband-separating (2SB) SIS receiver.
The basic design is the same as Ref. \citenum{2007PASJ...59.1005N} which was developed for the 60-cm AMANOGAWA telescope.
The target LO frequency is $\sim225$ GHz, and the receiver outputs two IF signals with a frequency of 4--8 GHz.
In order to achieve simultaneous observations of $^{12}$CO $J=$2--1 (230.538 GHz), $^{13}$CO $J=$2--1 (220.399 GHz), and C$^{18}$O $J=$2--1 (219.560 GHz) by using one spectrometer with a bandwidth of 1 GHz, Acqiris AC240\cite{2005A&A...442..767B}, two output signals of the SIS receiver (1st IF) were split to 3 signals, then each of the signal was further down-converted to different frequency in the 2nd IF band, and finally those 3 signals merged and one 0--1 GHz signal which contains the 3 different RF bands was generated in the IF system\cite{2013PASJ...65...78O}.
In the observation season of 2012, an Orthomode Transducer\cite{2009.JIMRW.Asayama, 2010.JIMRW.Kamikura} (OMT) for dual-polarization observations was developed and installed to the telescope.
The measured characteristics of the OMT were an insertion loss of less than 0.5 dB, return loss of better than 20 dB, and cross polarization coupling of better than 20 dB across 210--250 GHz\cite{2013PASJ...65...78O}.
Using the OMT with two balanced-type 2SB SIS receivers, observation efficiency was improved by a factor of 2 than previous seasons.
\subsubsection{Season 2013--2015}
Although the balanced-type 2SB SIS mixer receivers are very powerful to observe wide frequency ranges simultaneously, its image rejection ratio (IRR) is relatively low\cite{2009.JIMRW.Nakajima, 2016SPIE.9914E..1ZM}, typically $10$ dB, and the stability of the IRR is not enough especially for the long-term observations\cite{2016EP&S...68...34O}.
The 2SB-Filter\cite{2015JIMTW..36..445A} (2SBF) receiver has a high IRR typically more than $20$ dB, and the long-term stability.
In the 2SBF receiver, the cutout frequency of the lower sideband (LSB) and upper sideband (USB) are determined at the design of the filter implemented in waveguide.
The target emission lines of the 1.85-m telescope was CO $J=$2--1, and other lines were not planned to observe the season, thus it was not a problem even if the LO frequency was fixed.
We developed a 2SBF receiver optimized for $^{12}$CO, $^{13}$CO, and C$^{18}$O $J=$2--1 lines observations\cite{2017PASJ...69...91H}.
The IRR of the receiver was measured to be better than $25$ dB.
The spectrometer was replaced on the season of 2013 with a bandwidth of 2 GHz model, RPG XFFTS\cite{2012A&A...542L...3K}.
We installed four XFFTS boards, and use each board for each receiver output: one board for the $^{12}$CO band and another board for the $^{13}$CO and C$^{18}$O band.
In the observation season of 2014, we replaced the OMT to a stepped septum-type waveguide circular polarizer\cite{2017.JIMRW.Hasegawa} (SST-CP).
By this replacement, the 1.85-m telescope was ready to use as a base of VLBI observations.
\subsubsection{Season 2016--2019}
By the season of 2015, the survey observations of CO $J=$2--1 was almost completed for the Galactic plane on the northern hemisphere (see also Section \ref{sec:co}), so we changed the science goal to survey weaker emission line in the 220 GHz band such as SO, CS, and CH$_{3}$OH.
In order to achieve the new theme, we developed the new filter system based on 2SBF (220/230/240 Rx).
In the system, the polarization of the received radio wave was firstly separated by the SST-CP, then signals of each of polarization are derived to the two 2SBFs which were tuned for the different RF frequency (Fig. \ref{fig:rx2016}): one for 220 and 230 GHz, and another for 230 and 240 GHz.
The typical system noise temperature ($T_{\rm sys}$) of the 220/230/240 Rx was measured to be 80, 90, 100, and 150 K for 220, 230 (LHCP), 230 (RHCP), and 240 GHz band, respectively, measured with the 1.85-m telescope at NRO.
\begin{figure} [tb]
\begin{center}
\begin{tabular}{c}
\includegraphics[width=15cm, bb=0 0 883 452]{fig/Rx2016.jpg}
\end{tabular}
\end{center}
\caption[example]
{ \label{fig:rx2016}
(Left) Schematic diagram of the frequency separation in the 220/230/240 receiver.
(Right) Photograph of the 220/230/240 receiver.
}
\end{figure}
\subsection{Telescope Control System}
The telescope and various equipment including the spectrometers are controlled and monitored on a Linux PC system via TCP/UDP/IP socket connections.
The whole system consists of the following components: 1, Hardware driver layer consisting of programs for controlling each of the instrument (e.g., telescope drives, encoders, receiver equipment); 2, Communication layer consisting of programs which command the instruments in cooperation to realize the observations; 3, Database layer consisting of programs that collect and store the data including the spectrometer output, environmental parameters, and operation log.
\subsubsection{Season 2009--2017}
\begin{figure} [b]
\begin{center}
\begin{tabular}{c}
\includegraphics[width=15cm, bb=0 0 1015 189]{fig/queue-obs.png}
\end{tabular}
\end{center}
\caption[example]
{ \label{fig:queue}
Flowchart of the queue based semi-automatic observation implemented in the observation system on the 1.85-m telescope.
}
\end{figure}
\begin{figure} [tb]
\begin{center}
\begin{tabular}{c}
\includegraphics[width=12cm, bb=0 0 869 531]{fig/pipeline.png}
\end{tabular}
\end{center}
\caption[example]
{ \label{fig:pipeline}
Schematic diagram of the web-based Q-look system.
}
\end{figure}
The control system for the 1.85-m telescope was started with a following composition.
The hardware driver layer was mostly written in C language to control the PCI boards, and written as a TCP server.
The communication layer was written in Python, and the database was implemented integrally on MySQL database system.
The detail of the implementation of control system is described in Ref. \citenum{2013PASJ...65...78O}.
Based on the framework, the queue based semi-automatic operation was realized which reduced burden for the telescope operator and the time loss due to the careless mistake, and hence improved the actual observation efficiency (Fig. \ref{fig:queue}).
Furthermore, we implemented the automatic data analyzing pipeline that provides observation results graphically via the web interface soon after the finishing each observation (Fig. \ref{fig:pipeline}).
\subsubsection{Season 2018--2019}
From the observation season of 2018, we started to develop a new control system based on ROS\cite{2009.ROS} (robot operating system).
The ROS is one of the most prevailing open source framework to control robots and provides a very useful communication module (called Topic) which realizes many-to-many communications between processes and/or computers.
The reliability and scalability on the communication layer was improved by introducing the Topic based system.
In addition, aiming to realize further efficiency on developing time, we developed the fully python based hardware drivers, {\tt pypci}\footnote{https://github.com/ars096/pypci} and {\tt pyinterface}\footnote{https://github.com/ogawa-ros/pyinterface}.
The detail of the ROS based control system NECST is described in Ref. \citenum{2020SPIE.kondo}, and the performance evaluations of the new control system is ongoing.
\section{OBSERVATIONS}
\label{sec:obs}
The science operation was started on 2011.
Usually, the science observations were conducted during the winter season (from November to next April), and the telescope was shut-down during the summer season because the high humidity condition reduces observation efficiency from the ground.
Figure \ref{fig:tau} shows the distribution of the optical depth and the system noise temperature ($T_{\rm sys}$) including the atmosphere measured toward the zenith with a frequency band of 230 GHz at NRO.
About the 60\% of the observation seasons show a better sky condition than the threshold of $T_{\rm sys} < 400$ K and the optical depth of $<0.4$.
For the seasons of 2010--2014 (Nov. 2010 to Apr. 2015), we performed the large scale survey of CO $J=$2--1, for the seasons of 2015--2016 were devoted to the demonstration experiments for the new instruments described in Section \ref{sec:rx}, and we conducted the observations of multiple line mapping in the season 2017 (Nov. 2017 to Apr. 2018).
\subsection{Survey Observations of CO J=2--1 Lines}
\label{sec:co}
Fig. \ref{fig:co} shows the survey coverage of the 1.85-m observations and some examples of obtained CO $J=$2--1 maps.
Totally 1800 deg$^{2}$ area (corresponding to 4.4\% of all-sky) was observed in the seasons of 2010, 2011, 2012, 2013, and 2014, and the observed areas for each season were 323, 160, 729, 303, and 307 deg$^{2}$, respectively.
All observations were conducted by the On-the-Fly (OTF) mapping mode with an angular spacing of 1$'$ grid.
The number of obtained spectra is 6,480,000.
The spatial resolution after the data reduction was $\sim3'$, and the typical noise level was $T_{\rm rms} \sim 0.8$ K with a velocity resolution of $\sim$0.1 km s$^{-1}$.
Using the CO data, studies on the star formation\cite{2013ApJ...768...72S, 2015ApJS..216...18N, 2016ApJ...818...59D, 2016A&A...589A..80H, 2016A&A...590A...2S, 2017ApJ...837..154N, 2017ApJ...845..105D, 2017A&A...608A..21S, 2017ApJ...844..138K, 2018AJ....156...84K}, and on the ISM\cite{2018A&A...609A.127S, 2019A&A...628A..44F} were promoted.
From the series of CO $J=$2--1 observations toward the GMCs in the Galaxy, we found that $^{12}$CO $J=$2--1 is easily becoming optically thick at the inner part of the GMC where the column density is higher than $\sim 3 \times 10^{21}$ cm$^{-2}$, and hence the ratio of $^{12}$CO $J=$2--1/$^{12}$CO $J=$1--0, $R^{12}_{2-1/1-0}$, is observed as $\sim 0.7$ which is a value expected to be observed in the condition of the local thermodynamic equilibrium (LTE) by the numerical simulations\cite{2017MNRAS.465.2277P}.
On the other hand, we found that the ratio of $^{13}$CO $J=$2--1/$^{13}$CO $J=$1--0, $R^{13}_{2-1/1-0}$, is a good tracer of the volume density even toward the inner region of the GMC.
For the case of Orion GMCs, the observed $R^{13}_{2-1/1-0}$ is in the range of 0.2--2.0, and its variation is clearly seen even toward the most inner part of the GMC where the column density is $\sim 10^{22}$ cm$^{-2}$\cite{2015ApJS..216...18N}.
The volume density of 800--2000 cm$^{-3}$ is estimated in the region including the inner part of the GMC by using LVG analysis.
\subsection{Spectral Line Surveys for 230 GHz Band}
\begin{figure} [b]
\begin{center}
\begin{tabular}{c}
\includegraphics[width=15cm, bb=0 0 580 286]{fig/line2016.png}
\end{tabular}
\end{center}
\caption[example]
{ \label{fig:line2016}
Spectra observed toward the Orion-KL region.
}
\end{figure}
In the early 2018, the test observations of the 2SBF receiver was conducted.
The receiver has four windows for observations: 219--221 GHz band including $^{13}$CO, C$^{18}$O, SO, and CH$_{3}$OH lines; 229.5--231.5 GHz band of the RHCP side including $^{12}$CO and CH$_{3}$OH lines; 230--232 GHz band of the LHCP side including $^{12}$CO line; and 240.5--242.5 GHz band including C$^{34}$S, SO$_{2}$, and CH$_{3}$OH lines.
Fig. \ref{fig:line2016} shows the spectra observed toward the Ori-KL region.
All spectra listed above were simultaneously detected.
Fig. \ref{fig:map2016} shows maps of $^{12}$CO, $^{13}$CO, C$^{18}$O, SO, and CH$_{3}$OH observed toward the Ori-KL region.
These results demonstrate the capability for the wide-band observations of the 2SBF receiver.
\begin{figure} [t]
\begin{center}
\begin{tabular}{c}
\includegraphics[width=15cm, bb=0 0 593 357]{fig/map2016.png}
\end{tabular}
\end{center}
\caption[example]
{ \label{fig:map2016}
Maps of $^{12}$CO, $^{13}$CO, C$^{18}$O, SO, and CH$_{3}$OH observed toward the Ori-KL region.
}
\end{figure}
\subsection{Technology Verification on VLBI Observations in 230 GHz Band}
The experiment on VLBI observations in 230 GHz band was conducted on April 2015 between two bases using the 1.85-m telescope and the SPART telescope which is a single dish operation mode of Nobeyama millimeter array.
The length of the baseline was 150 m, and the instruments for the VLBI observations (e.g., a frequency standard generator, samplers) were brought to the telescope site in temporarily.
By observing the moon edge, VLBI fringe was successfully detected by the software correlator.
This success of the experiment led to the realization of the next VLBI experiment among the SPART 10-m, SRAO 6-m in Korea, and GLT 12-m in Greenland with a frequency of 230 GHz.
\section{FUTURE PLAN}
\label{sec:future}
In order to extend the observation frequency band further wide, and to achieve simultaneous observations of 230 and 345 GHz band by sharing same single beam\cite{2020SPIE.masui, 2020SPIE.yamasaki}, we are planing to relocate the 1.85-m telescope to San Pedro de Atacama (SPdA), Chile (alt. 2400 m).
Although, the altitude of SPdA is not high as the Atacama high site where NANTEN2 and ASTE are installed, its weather condition would be almost sufficient for observations of 230 and 345 GHz band.
On the contrary, SPdA would be better choice in terms of the cost of construction and running.
The 1.85-m telescope will be installed in the base facility of the TAO project as part of the cooperative research with the university of Tokyo.
Fig. \ref{fig:tao} shows a photograph of the planned site of the 1.85-m telescope installation and a CAD drawing of the new telescope base which will be constructed to the site.
By this relocation, we will be able to perform the verification tests for the higher frequency instruments and to observe the higher frequency bands, as well as to observe southern sky including the Large/Small Magellanic Clouds.
\begin{figure} [t]
\begin{center}
\begin{tabular}{c}
\includegraphics[width=15cm, bb=0 0 500 193]{fig/tao-site.jpg}
\end{tabular}
\end{center}
\caption[example]
{ \label{fig:tao}
Left: Photograph of the planned location of the 1.85-m telescope installation in the base facility of the TAO project.
Persons, Prof. Hideo Ogawa, Mr. Yasumasa Yamasaki, and Mr. Takeru Matsumoto from left to right, indicate the size of the telescope base shown in the right figure.
Right: CAD drawing of the model of telescope base which will be constructed at the telescope site in SPdA.
}
\end{figure}
\section{SUMMARY}
\label{sec:summary}
In the paper, we reviewed and summarized the activities of developments and observations on the 1.85-m telescope installed at Nobeyama Radio Observatory (NRO), Japan.
The construction of the telescope at NRO was started on 2007.
The 2SB receiver with a RF of 230 GHz was installed, and its first light was achieved on September 2009 by observing the $^{12}$CO, $^{13}$CO, and C$^{18}$O $J=$2--1 lines simultaneously.
The science observation operations were started on November 2011, and the large-scale mapping observations of the $^{12}$CO, $^{13}$CO, and C$^{18}$O $J=$2--1 lines were continued to the observation season of 2014.
Finally, an area of 1800 deg$^{2}$ was observed and a number of 6,480,000 spectra was obtained.
During the period of science operations, the developments of the receivers and control systems were also continued.
Those developed systems were installed to the telescope in the summer season, when the science observations were stopped, and were tested for verification, and then, were used for scientific observations.
To date, 18 of refereed journal papers, 3 of doctor theses, 23 of master theses, and 31 of graduation theses were published related to the telescope or by using the data obtained as the CO survey (Fig. \ref{fig:pub}).
\begin{figure} [H]
\begin{center}
\begin{tabular}{c}
\includegraphics[width=15cm, bb=0 0 792 216]{fig/pub1p85.png}
\end{tabular}
\end{center}
\caption[example]
{ \label{fig:pub}
Bar chart indicating numbers of publications related to the 1.85-m telescope project.
}
\end{figure}
|
\section{\@startsection{section}{1}%
\z@{.7\linespacing\@plus\linespacing}{.5\linespacing}%
{\normalfont\scshape}
\makeatother
\title[Compactness of Free Boundary CMC Surfaces]{Compactness of the Space of Free Boundary CMC Surfaces with Bounded Topology}
\author{Nicolau S. Aiex and Han Hong}
\date{\today}
\address{University of Auckland \\ Department of Mathematics \\ Auckland 1010 \\ New Zealand}
\email{<EMAIL>}
\address{University of British Columbia \\ Department of Mathematics \\ Vancouver BC V6T 1Z2 \\ Canada}
\email{<EMAIL>}
\begin{document}
\begin{abstract}
We prove that the space of free boundary CMC surfaces of bounded topology, bounded area and bounded boundary length is compact in the $C^k$ graphical sense away from a finite set of points.
This is a CMC version of a result for minimal surfaces by Fraser-Li \cite{fraser.a-li.m2014}.
\end{abstract}
\maketitle
\section{Introduction}
A Constant Mean Curvature (CMC) surface $\Sigma$ is a critical point of the area functional with respect to variations that preserve enclosed volume.
As a consequence of the first variation of area, the scalar mean curvature has to be constant for a given choice of normal direction.
If $\Sigma$ is an immersed surface with boundary immersed on $N$ with boundary and $\partial \Sigma \subset \partial N$, we say the surface is free boundary CMC if it is a critical point with respect to variations that are in addition tangent along $\partial N$.
Although most results about minimal surfaces have an equivalent version for CMC surfaces, there are many distinctions between their behaviour.
For example, when the mean curvature is non-zero, the mean curvature vector defines a trivialization of the normal bundle.
That is, every CMC surface on a $3$-manifold is $2$-sided.
On the other hand, CMC surfaces may have tangential self-touching points as long as the mean curvature vector points at opposite directions on those points.
The goal of this article is to prove the CMC equivalent to Fraser-Li's result \cite{fraser.a-li.m2014} for free boundary minimal surfaces.
We prove:
\begin{maintheorem}{\ref{compactness fbcmc}}
Let $N$ be a compact $3$-dimensional manifold with boundary.
Suppose $H_{\partial N}\geq H_0$ and let $\Sigma_i$ be a sequence of free boundary embedded CMC surfaces with mean curvature $H_i$, genus $g_i$ and number of ends $r_i$ satisfying:
\begin{enumerate}[(a)]
\item $|H_i|\leq H_0$;
\item $g_i\leq g_0$;
\item $r_i\leq r_0$;
\item $\textup{area}(\Sigma_i)\leq A_0$ and
\item $\textup{length}(\partial \Sigma_i)\leq L_0$.
\end{enumerate}
Then there exists a smooth properly almost embedded CMC surface $\Sigma\subset N$ and a finite set $\Gamma\subset \Sigma$ such that, up to a subsequence, $\Sigma_i$ converges to $\Sigma$ locally graphically in the $C^k$ topology on compact sets of $N\setminus \Gamma$ for all $k\geq 2$.
Moreover, if $\Sigma$ is minimal then it is properly embedded.
If in addition $(N,\partial N)$ satisfies either $Ric_N > 0$ and $A_{\partial N}\geq 0$ or $Ric_N \geq 0$ and $A_{\partial N} > 0$, then:
\begin{enumerate}[(i)]
\item when $H_\Sigma=0$ the convergence is at most $2$-sheeted;
\item when $H_\Sigma\neq 0$, then the convergence is $1$-sheeted away from $\Gamma$.
\end{enumerate}
\end{maintheorem}
Let us highlight the main differences that justify the extra hypotheses and the weaker compactness for embedded surfaces.
Firstly, we mention that there is no Steklov eigenvalue estimates for free boundary CMC surfaces.
Consequently, the hypothesis of length bound remain crucial.
Secondly, there is no suitable isoperimetric inequality that allows us to remove the bound on the area.
One could exchange the length bound condition by stability of the surface (see \cite{mendes2018}) but as seen on \cite{aiex-hong2021}, in particular, stable free boundary CMC surfaces have bounded topology so this condition would be topologically restrictive.
Thirdly, even under positive Ricci curvature of the ambient space a sequence of free boundary CMC surfaces may have a neck-pinching phenomenum where the norm of the second fundamental form blows-up at a point in the limit.
Naturally the convergence is not smooth along these points where curvature is accumulating.
Finally, the maximum principle for CMC surfaces only apply when their mean curvature vectors point in the same direction.
That is, a sequence of embedded free boundary CMC surfaces may touch tangentially in the limit as long as the limiting surface is not minimal.
Despite the lack of certain useful technical results, the proof of the theorem rely on the same main ideas, each of which is proved using a less optimal technique to compensate for the above limitations.
These are: $L^2$-curvature bounds from the topological bound, improvement to local pointwise curvature estimates, a removable singularity theorem for interior and boundary points and construction of positive eigenfunctions for the Jacobi operator to study convergence under curvature condition on the ambient space.
Let us briefly address our approach to each of the above tools.
The integral curvature bounds follow directly from Gauss-Bonnet Theorem, from which the fact the mean curvature is non-zero introduces an extra term of area and the geodesic curvature along the boundary brings in a term of boundary length.
The local pointwise curvature estimate comes from a blow-up argument as in \cite{white1987} and Schauder estimates.
It becomes relevant that the blow-up of a CMC surface around a point is a minimal surface in $\real^3$.
Unlike \cite{fraser.a-li.m2014}*{Theorem 4.1} we do not know whether a CMC surface is conformally equivalent to a punctured Riemann surface so we do not have the branch point structure to prove the removable singularity theorem.
This will follow from a blow-up argument to prove that tangent cones are in fact planes (or half-planes) and a local free boundary CMC foliation argument to prove that it is also unique.
These are the same ideas as in \cite{white1987} together with the methods in \cite{ambrozio-carlotto-sharp2018.3} to deal with boundary singularity points.
In fact, most of the calculations are done in the latter reference and only minor adaptations are necessary for the non-minimal case.
Finally, the contruction of positive eigenfunctions for the Jacobi operator on the limiting surface is a well known method (see \cites{simon1987,colding-minicozzi2000,sharp2017}) and again, most of the necessary calculations are done in \cite{ambrozio-carlotto-sharp2018.3}.
The extra condition on the mean curvature of $\partial N$ is only necessary to apply the maximum principle for CMC surfaces and ensure that the interior of the limiting surface is properly immersed in $N$.
Removing this condition would allow for interior tangential touching points between $\partial N$ and the limit surface.
Finally, we mention that the corresponding result for closed CMC surfaces was proved by Sun in \cite{sun.a2020}.
Although the core ideas are the same, the approach to certain steps is different and we focus on the behaviour at the boundary.
This article is divided as follows.
Section $2$ establishes notation, necessary definitions and we prove the geometric version of Schauder estimates.
In section $3$ we prove the local pointwise curvature estimates from uniform integral curvature bounds.
The Removable Singularity Theorem is proved in section $4$ and we write the proof for the specific case of a singularity along the boundary.
Section $5$ is dedicated to prove the Compactness Theorem.
\textit{
Acknowledgements: The majority of the work on this article took place when the first author was a postdoctoral fellow at PIMS-University of British Columbia.
The authors would like to thank Professors Jingyi Chen and Ailana Fraser for useful discussions and support.
}
\section{Preliminaries}
In this section we establish notation and preliminary results that will be used throughout this article.
Let $(N^3,\partial N, g)$ be a $3$-dimensional manifold with non-empty boundary and a smooth Riemannian metric $g$.
\begin{definition}
We say that an immersed surface $\Sigma\subset N$ with non-empty boundary is properly almost embedded in an open set $U\subset N$ if $(\Sigma\setminus\partial \Sigma)\cap U$ is properly immersed in $(N\setminus \partial N)\cap U$, $\partial \Sigma\subset \partial N\cap U$ and there exists a set $\mathcal{S} \subset\Sigma$ such that
\begin{enumerate}[(a)]
\item $\Sigma\setminus\mathcal{S}$ is embedded;
\item for each $p\in\mathcal{S}$ there exists a neighbourhood $V$ of $p$ in $U$ such that $\Sigma\cap V$ is a union of connected components $W_j$, $j=1,\ldots,l_p$, each $W_j$ is embedded, for each $j\neq j'$ we have $W_j$ lying to one side of $W_{j'}$ and $W_j\cap W_{j'}\subset \mathcal{S}\cap V$.
\end{enumerate}
The set $\mathcal{S}$ is called the self-touching set of $\Sigma$.
We say $\Sigma$ is free boundary if in addition $\Sigma$ is orthogonal to $\partial N$.
\end{definition}
The following lemma is a standard application of Schauder estimates in geometry.
It will allow us to improve from $C^{1,\alpha}$ to $C^{2,\alpha}$ graphical convergence of surfaces, given pointwise curvature estimates.
\begin{lemma}\label{schauder estimates}
Let $N$ be a three dimensional compact manifold, $A>0$ and $f\in C^{0,\alpha}(N)$ with $|f|_{0,\alpha}<H_0$.
There exists $r_0(N, A)>0$ and $C_0(N, A, H_0)>0$ such that for any $C^2$ immersed surface with free boundary $\Sigma\subset N$ with $H_{\Sigma}=f$ and $x_0\in \Sigma$ satisfying
\begin{equation*}
|A_\Sigma(x)|\leq A \text{ on } \{x\in \Sigma: d_\Sigma(x,x_0)< r_0\}
\end{equation*}
there exists a function $u$ defined on a subset of $T_{x_0}\Sigma$ with the following properties:
\begin{enumerate}[(i)]
\item $\{x\in \Sigma: d_\Sigma(x,x_0)< r_0\}$ is the graph of $u$ over the exponential function;
\item the domain of $u$ contains a ball centered at $0$ and radius $\frac{r_0}{2}$ and
\item $\|u|_{B_{\frac{r_0}{4}}(0)}\|_{2,\alpha}<C_0$.
\end{enumerate}
\end{lemma}
\begin{proof}
We may assume, that $r_0>0$ is smaller than the convexity radius of $N$ so that $\{x\in \Sigma: d_\Sigma(x,x_0)<r_0\}$ is contained in a geodesic ball in $N$.
Hence, without loss of generality we may assume that $\{x\in \Sigma: d_\Sigma(x,x_0)<r_0\}$ is contained in a small ball in $\real^3$ with a metric $g$, $x_0=0$ and $T_{x_0}\Sigma=\real^2\times\{0\}$.
We may further assume that $g$ is equivalent to the Euclidean metric $g_0$, that is, $c^{-1}g_0\leq g\leq c g_0$ where the constant $c > 0$ depends only on the geometry of $N$.
\begin{claim}
If $r_0$ is sufficiently small, depending only on the geometry of $N$ and $A$, then for all $x\in\{x\in \Sigma: d_\Sigma(x,0)<r_0\}$ we have $|\nu(x)-\nu(0)|_{g_0}\leq CAd_\Sigma(x,0)$ for some $C > 0$ depending only on the geometry of $N$.
\end{claim}
Indeed, let $\gamma\subset\Sigma$ be a curve parametrised by arc-length joining $0$ to $x$.
If we denote by $\nabla^0$ the Euclidean connection, then it follows that
\begin{equation*}
|\nu(x)-\nu(0)|_{g_0}\leq\sup_{t}|\nabla^0_{\dot\gamma}\nu|_{g_0} l(\gamma)
\end{equation*}
Since we are working in geodesic coordinates then $\nabla^0_{\dot\gamma} = \nabla_{\dot\gamma} + O(|\gamma|)$, hence
\begin{equation*}
|\nabla^0_{\dot\gamma}\nu|_{g_0}\leq c|\nabla^0_{\dot\gamma}\nu|_g\leq c|\nabla_{\dot\gamma}\nu|_g+|\gamma| C\leq c A + r_0 C,
\end{equation*}
for some $C > 0$ depending only on the geometry of $N$.
If we pick $r_0 C < c A$ we obtain
\begin{equation*}
|\nu(x)-\nu(0)|_{g_0}\leq 2cA l(\gamma).
\end{equation*}
Taking the limit as the length of $\gamma$ tends to $d_\Sigma(x,0)$ we conclude the claim as desired.
Now, suppose that $C Ar_0\leq 1$, then $\{x\in \Sigma: d_\Sigma(x,0)\leq r_0\}$ is the graph of a function $u$, otherwise we would have $\nu(x)$ perpendicular to $\nu(0)$ for some $x$ hence $|\nu(x)-\nu(0)|>1$.
Furthermore, it follows that $\sup|\nabla u|\leq 2 C A r_0$ as long as $C^2A^2r_0^2 < \frac{3}{4}$.
Let us denote by $\Omega\subset T_0 \Sigma$ the domain of $u$ and take $\delta>0$ the largest radius such that $B_{\delta}(0)\subset\Omega$.
In particular there exists $\bar y \in \partial B_{\delta}(0)$ such that $d_\Sigma((\bar y,u(\bar y)),(0,0))=r_0$.
\begin{claim}
For any pair $y_1,y_2\in B_\delta(0)$ we have $d_\Sigma((y_1,u(y_1)),(y_2,u(y_2)))\leq (1+(2CAr_0)^2)^{\frac{1}{2}}|y_1-y_2|$ and, if $4CAr_0<1$ then $\frac{r_0}{2}\leq \delta$.
\end{claim}
We have $y_1 + t(y_2-y_1)\in B_\delta(0)\subset \Omega$ for $t\in[0,1]$, from which follows that $d_\Sigma((y_1,u(y_1)), (y_2,u(y_2)))\leq \int_0^1(1+|\nabla u(y_1 + t(y_2-y_1))|^2)^{\frac{1}{2}}|y_1-y_2|dt$ and it proves the inequality.
For the second part observe that if $4CAr_0<1$ then $d_\Sigma((\bar y,u(\bar y)),(0,0))\leq \frac{3}{2}|\bar y|$.
Suppose that $\delta<\frac{r_0}{2}$, then $|\bar y|<\frac{r_0}{2}$ and $d_\Sigma((\bar y,u(\bar y)),0)\leq \frac{3r_0}{4}$ which is a contradiction and it proves the claim.
\begin{claim}
If $4CAr_0<1$ then $\nabla u$ restricted to $B_{\frac{r_0}{2}}(0)$ is a Lipschitz function of Lipschitz constant less than $\frac{3}{2}A$.
\end{claim}
The second fundamental form is a multiple of the Hessian of $u$, thus $|A_\Sigma|\leq A$ implies $|Hess u|\leq (1+|\nabla u|^2)^{\frac{1}{2}} A < 2A$ whenever $4CA\varepsilon_0<1$.
It follows that for any $y_1,y_2\in\Omega$, $|\nabla u(y_1)-\nabla u(y_2)|\leq 2A d_\Sigma((y_1,u(y_1)),(y_2,u(y_2)))$.
From the previous claim we get that $d_\Sigma((y_1,u(y_1)),(y_2,u(y_2)))\leq (1+(2CAr_0)^2)^{\frac{1}{2}}|y_1-y_2|\leq \frac{3}{2}|y_1-y_2|$, thus proving the claim.
Finally, on $B_{\frac{r_0}{2}}(0)$ the function $u$ satisfies the equation $\divergent\left( \frac{\nabla u}{\sqrt{1+|\nabla u|^2}}\right)=f$.
The coefficients of the equation have $C^{2,\alpha}$ norm depending on the metric $g$ and the Lipschitz constant of $\nabla u$.
In particular all coefficients have $C^{2,\alpha}$ norm depending only on $N$ and $A$.
Let us consider the case without boundary components on $\{x\in \Sigma: d_\Sigma(x,x_0)<r_0\}$.
Then Schauder estimates \cite{gilbarg-trudinger}*{Corollary $6.3$} implies that there exists a constant $C=C(N,A)$ such that $\|u|_{B_{\frac{r_0}{4}}(0)}\|_{2,\alpha}<C(\|u|_{B_{\frac{r_0}{2}}(0)}\|_0 +\|f\|_{0,\alpha})$.
Note that $|u(y)-u(0)|<2Ar_0|y|<\frac{2Ar_0^2}{2}$ and $u(0)=0$.
That is, $\|u|_{B_{\frac{r_0}{4}}(0)}\|_{2,\alpha}<C_0$ for $C_0=C_0(N,A,H_0)$.
In case there are boundary components of $\Sigma$ on $\{x\in \Sigma: d_\Sigma(x,x_0)<r_0\}$ then Schauder estimates (\cite{agmon-douglis-nirenberg1959}*{Theorem 7.1}) gives us the same inequality with an extra term depending on the $C^{1,\alpha}$ norm of the inward normal derivative along the boundary.
Since $\Sigma$ is free boundary, $u$ satisfies homogenous Neumann boundary conditions in which case the extra term vanishes and we obtain the same result.
\end{proof}
In the following we make precise the notion of graphical convergence of surfaces (see also \cite{sharp2017}).
\begin{definition}
Let $N$ be a $3$-manifold, with or without boundary, $\Sigma_i$ a sequence of smooth surfaces in $N$ and $\Sigma$ a smooth surface in $N$.
Pick $p\in\Sigma$ and $r>0$ sufficiently small so that we may identify $\nball_r(p)$ with an Euclidean ball (or half-ball if $p\in\partial N$) with the same metric as $N$.
We say that $\Sigma_i$ converges locally graphically in the $C^k$ ($C^{k,\alpha}$) topology to $\Sigma$ at $p$ if for $r>0$ sufficiently small we have:
\begin{enumerate}[(a)]
\item $\Sigma\cap\nball_r(p)$ is the graph of a $C^k$ ($C^{k,\alpha}$) function $u$ defined on $B_r(0)\subset T_p\Sigma$;
\item $\Sigma_i\cap\nball_r(p)$ is the graph of $C^k$ ($C^{k,\alpha}$) functions $u_i^1,\ldots u_i^L$ defined on $B_r(0)\subset T_p\Sigma$ for $i$ sufficiently large and
\item $u_i^j$ converges to $u$ in the $C^k$ ($C^{k,\alpha}$) topology as $i\rightarrow\infty$.
\end{enumerate}
If $L$ is constant for sufficiently large $i$, we say that the convergence is $L$-sheeted.
\end{definition}
\begin{remark}
Let $U\subset N$ be an open set.
If $\Sigma_i$ is a sequence of embedded surfaces converging locally graphically with $L$-sheets to a two-sided surface $\Sigma$ on a compacts sets of $U$ then for each compact set $\Omega\subset \Sigma\cap U$, $r>0$ sufficiently small and $i$ sufficiently large we may write $\Sigma_i\cap \nball_r(\Omega)$ as the graph of functions $u_{i,j}:\Omega\rightarrow \real$, $j=1,\ldots, L$, under the exponential map in the normal direction of $\Sigma$.
In other words, each connected component $\Sigma_{i,j}$ of $\Sigma_i\cap \nball_r(\Omega)$ can be written as $\{\exp_x(u_{i,j}(x)N_\Sigma(x)):x\in\Omega\}$ for each $j=1,\ldots, L$.
In addition, $u_{i,j}$ must tend to $0$ for each $j=1,\ldots, L$.
\end{remark}
\section{Curvature estimate}
In this section we prove an improvement from uniformly small total curvature estimate to uniform local pointwise curvature estimate.
The proof is inspired by \cite{white1987}*{Theorem 1} and we focus on the local estimates around a boundary point.
We point out as well that the same proof holds even if the surface is not CMC but has uniformly bounded $C^{0,\alpha}$ mean curvature.
\begin{theorem}\label{curvature estimates}
Let $N$ be a compact $3$-manifold with boundary. There exists a small enough $r_0>0$ such that the following holds: whenever $\Sigma$ is a properly immersed CMC surface in $N$, $Q\in\bar\Sigma$, $\partial\Sigma\cap \nball_{r_0}(Q)$ is either empty or free boundary in $\partial N\cap \nball_{r_0}(Q)$ and the mean curvature of $\Sigma$ satisfies $H_{\Sigma}\leq H_0$.
Then there exists $\varepsilon_0>0$ depending on $\nball_{r_0}(Q)$ and $H_0$ such that if $\int_{\Sigma\cap \nball_{r_0}(Q)} |A|^2\leq \varepsilon_0$, then
\begin{equation*}
\max_{0\leq \sigma\leq r_0}\left(\sigma^2\sup_{\Sigma\cap \nball_{r_0-\sigma}(Q)}|A|^2\right)\leq C_0
\end{equation*}
where the constant $C_0$ only depends on geometry of $\nball_{r_0}(Q)$ and $H_0$.
\end{theorem}
\begin{proof}
Suppose false, that is, for $r_n\rightarrow 0$ and $\varepsilon_n\rightarrow 0$ there exist free boundary CMC surfaces $\Sigma_n\subset N$ and $Q_n\in\Sigma_n$ satisfying:
\begin{enumerate}[(i)]
\item $H_n\leq H_0$;
\item $\int_{\Sigma_n\cap \nball_{r_n}(Q_n)} |A_n|^2\leq \varepsilon_n$ and
\item $\max_{0\leq \sigma\leq r_n}\left(\sigma^2\sup_{\Sigma\cap \nball_{r_n-\sigma}(Q_n)}|A_n|^2\right) > n$.
\end{enumerate}
Pick $0<\sigma_n< r_n$ such that
\begin{equation*}
\sigma_n^2\sup_{\Sigma\cap \nball_{r_n-\sigma_n}(Q_n)}|A_n|^2=\max_{0\leq \sigma\leq r_n}\left(\sigma^2\sup_{\Sigma\cap \nball_{r_n-\sigma}(Q_n)}|A_n|^2\right)
\end{equation*}
and write $\lambda_n^2 = \sup_{\Sigma\cap \nball_{r_n-\sigma_n}(Q_n)}|A_n|^2$.
For each $n$ there exists $z_n\in\Sigma\cap \nball_{r_n-\sigma_n}(Q_n)$ such that $|A_n(z_n)|>\frac{\lambda_n}{2}$.
By taking a subsequence we have $Q_n\rightarrow Q$ and $\nball_{r_n}(Q_n)$ contained in a geodesic ball of $N$.
Without loss of generality we may assume that $\Sigma_n\cap \nball_{r_n}(Q_n)\subset \real^3$ with a metric $g$.
Henceforth we denote by $B_r(p)$ the ball in $\real^3$ with respect to the metric $g$.
Now, define $\tilde\Sigma_n=\lambda_n(\Sigma_n-z_n)$, it satisfies:
\begin{enumerate}[(a)]
\item $|\tilde A_n(\tilde x_n)|\leq 2$ for all $\tilde x_n\in\tilde \Sigma_n\cap B_1(0)$ and $n$ sufficiently large;
\item $|\tilde A_n(0)|>\frac{1}{2}$ and
\item $\int_{\tilde\Sigma_n} |\tilde A_n|^2\leq \varepsilon_n$
\end{enumerate}
Indeed, if $\tilde x_n\in B_1(0)$ then $\tilde x_n = \lambda_n(x_n-z_n)$ with $x_n\in \Sigma_n\cap B_{\frac{1}{\lambda_n}}(z_n)$.
It follows that $x_n\in \Sigma_n\cap B_{r_n-(\sigma_n-\frac{1}{\lambda_n})}(Q_n)$.
Since
\begin{equation*}
\left(\sigma_n-\frac{1}{\lambda_n}\right)^2\sup_{\Sigma\cap \nball_{r_n-(\sigma_n-\frac{1}{\lambda_n})}(Q_n)}|A_n|^2\leq \max_{0\leq \sigma\leq r_n}\left(\sigma^2\sup_{\Sigma\cap \nball_{r_n-\sigma}(Q_n)}|A_n|^2\right)=\sigma_n^2\lambda_n^2,
\end{equation*}
it implies that
\begin{equation*}
|A_n(x_n)|^2<\left(\frac{1}{1-\frac{1}{\sigma_n\lambda_n}}\right)^2\lambda_n^2.
\end{equation*}
We know that $\sigma_n^2\lambda_n^2 > n$ thus $|A_n(x_n)|<2\lambda_n$ for $n$ sufficiently large, which proves (a).
Property (b) follows from rescaling and (c) holds because the total curvature is scale invariant.
Let $\hat \Sigma_n$ denote the connected component of $\tilde \Sigma_n\cap B_1(0)$ containing $0$.
We may further assume, after an ambient rotation and translation that $T_{0}\hat \Sigma_n=\{x_3=0\}$.
Using property (a), it follows from Lemma \ref{schauder estimates} that there exists $\hat r_0$ independent of $n$ such that $\hat \Sigma_n\cap B_{\hat r_0}(0)$ is the graph of a function $u_n$ satisfying $\|u_n|_{B_{\frac{\hat r_0}{4}}(0)}\|_{2,\alpha}<\hat C_0$, where $\hat C_0$ is independent of $n$.
If $0$ is a boundary point then the domain of $u_n$ is a half ball but Lemma \ref{schauder estimates} remains true.
Finally, $u_n$ converges, up to a subsequence, in the $C^2$ topology to a function $u_\infty$.
If we denote its graph by $\hat{\Sigma}_\infty$ then it satisfies $|\hat A_\infty(0)|\geq\frac{1}{2}$ from (b) and $\int_{\hat\Sigma_\infty}|\hat A_\infty|^2=0$ from (c), which is a contradiction and completes the proof of the Theorem.
\end{proof}
\section{Removable singularities}
As we will see later, the compactness result does not give us smooth convergence everywhere.
The points in which we do not have sufficient curvature estimates are potential singularities either because of a neckpinching phenomenum where the curvature may blow up, or self-touching points where the convergence is not single sheeted.
However, we are still able to prove that these are removable singularities so the limiting surface is still a smooth object.
We are going to prove that if the total curvature is bounded on a CMC surface, then isolated singularities are removable.
This is an adaptation of \cite{white1987}*{Theorem 2} and the arguments are the same except for the foliation argument to prove uniqueness of the tangent cone.
We are going to focus on the case in which the singularity is along the boundary, but the same result holds for interior singularities and the proof follows the exact same arguments.
The idea of the proof is to improve the integral curvature bound to a pointwise curvature decay near the singularity to show that the tangent cones are totally geodesic.
By adapting the foliation argument of White \cite{white1987} we prove that the tangent cone is unique from which we can show that near the singularity the surface is indeed the graph of a $C^1$ function.
We can then improve it further using elliptic regularity.
Firstly, let us prove the existence of a local CMC foliation with free boundary which will be needed later.
This is a straightforward adaptation of \cite{ambrozio-carlotto-sharp2018.3}*{Section $3$} together with White's approach to deal with a family of functionals \cite{white1987}*{Appendix}.
Let $\theta\in(0,\frac{\pi}{4})$ and define $D_\theta = \{ x \in\real^2 : (x_1+a)^2+x_2^2 \leq 1 \text{ and } x_1\geq 0\}$, where $a=\cos^{-1}(\theta)\in(\frac{1}{\sqrt{2}},1)$.
This is the part of the disk of radius $1$ centered on the $x_1$-axis that intersects the line $x_1=0$ at angle $\theta$.
Its boundary components are denoted by $\partial_0D_\theta=\partial D_\theta\cap\{x\in\real^2:x_1=0\}$ and $\partial_+D_\theta=\overline{ \partial{D_\theta\setminus \partial_0D_\theta}}$.
The regular cylinder over $D_\theta$ in $\real^3$ is denoted by $C_\theta=D_\theta\times\real$, with corresponding boundary components $\partial_0C_\theta=\partial_0 D_\theta\times\real$ and $\partial_+C_\theta=\partial_+ D_\theta\times\real$.
\begin{center}
\input{half-cylinder-nosphere}
\end{center}
Given a function $f\in C^{2,\alpha}(D_\theta)$, we define $N_g^+(f)$ to be the normal vector over $\textup{graph}(f)$ with respect to $g$ pointing in the positive direction of the $x_3$-axis, that is, $g(N_g^+(f),\frac{\partial}{\partial x_3})>0$.
We write $H_g^+(f)=g(\vec{H}_g(f),N_g^+(f))$ as the scalar mean curvature with respect to $N_g^+(f)$.
In particular, $\vec{H}_g(f)$ points in the positive direction of the $x_3$-axis when $H_g^+(f) > 0$ and in the negative direction otherwise.
Let us denote by $X$ the space of $C^{2,\alpha}$ metrics on $C_\theta$ and define the map
\begin{equation*}
\Phi:\real\times\real\times X \times C^{2,\alpha}(\partial_+D_\theta) \times C^{2,\alpha}(D_\theta)\rightarrow C^{0,\alpha}(D_\theta)\times C^{1,\alpha}(\partial_0 D_\theta)\times C^{2,\alpha}(\partial_+ D_\theta)
\end{equation*}
by
\begin{equation*}
\Phi(h,t,g,w,u)=\left(H^+_g(t+u)-h,\frac{\partial}{\partial\eta_g}(t+u),u_{|_{\partial_{+}D_{\theta}}}-w\right),
\end{equation*}
where $\eta_g$ is the inward conormal vector along $\partial_0D_\theta$.
\begin{proposition}[{\cite{ambrozio-carlotto-sharp2018.3}*{Proposition $21$}}]\label{proposition foliation}
For every $t_0\in\mathbb{R}$, there exist a neighbourhood $U_{t_0}$ of the Euclidean metric $\delta$ in $X$, $\varepsilon_{t_0}>0$ and
\begin{equation*}
u:(-\varepsilon_{t_0},\varepsilon_{t_0})\times(t_0-\varepsilon_{t_0},t_0+\varepsilon_{t_0})\times U_{t_0}\times B^{C^{2,\alpha}(\partial_+D_\theta)}_{\varepsilon_{t_0}}(0)\rightarrow C^{2,\alpha}(D_\theta)
\end{equation*}
so that $t\mapsto \textup{graph}(t+u(h,t,g,w))$ defines a $C^{2,\alpha}$ foliation of $D_\theta\times[t_0-\frac{\varepsilon_{t_0}}{2},t_0+\frac{\varepsilon_{t_0}}{2}]$ by surfaces with constant mean curvature $h$ with respect to the metric $g$, free boundary along $\partial_0 C_\theta$ and $(t+u)_{|_{\partial_+D_\theta}}=t+w$.
Furthermore, if $h>0$ then $\vec{H}_g(t+u)$ points in the positive direction of the $x_3$-axis and in the negative direction when $h<0$.
\end{proposition}
\begin{proof}
Observe that $\Phi$ defined above is a $C^1$ function and $D_5\Phi(0,t_0,\delta,0,0)$ defines the same isomorphism as in \cite{ambrozio-carlotto-sharp2018.3}*{Appendix B}.
The result then follows from the Implicit Function Theorem.
\end{proof}
Let us denote by $B^+_1=\{x\in\real^3 : \|x\|\leq 1, x_1\geq 0\}$ the upper half ball in $\real^3$ and $\partial_0B^+_1=\partial B^+_1\cap\{x\in\real^3:x_1=0\}$.
\begin{theorem}\label{removable singularity boundary}
Let $g$ be a Riemannian metric on $B^+_1$ and $\Sigma$ be a smooth, properly embedded, CMC surface in $B^+_1\setminus\{0\}$, $\partial \Sigma\subset \partial B^+_1$, free boundary on $\partial_0 B^+_1\setminus\{0\}$ and $0\in\overline{\partial\Sigma}$.
Suppose $\int_\Sigma|A_\Sigma|^2\leq C$ then $\Sigma\cup\{0\}$ is a smooth properly embedded CMC surface in $B_1^+$.
\end{theorem}
\begin{proof}
Let $r_0>0$ and $\varepsilon_0>0$ be as in Theorem \ref{curvature estimates}.
Pick $\delta>0$ sufficiently small so that $\int_{\Sigma\cap B^+_\delta}|A_\Sigma|^2\leq \varepsilon_0$.
It follows from Theorem \ref{curvature estimates} that
\begin{equation*}
|A_\Sigma(x)|d_g(x,0)\leq C_0,
\end{equation*}
whenever $d_g(x,0)<\delta$.
\begin{claim}
Every tangent cone of $\Sigma$ at $0$ is a union of half-planes in $\real^3\setminus\{0\}$.
\end{claim}
Let $r_i\rightarrow\infty$ be any sequence and $\Sigma_i=r_i\Sigma$ its corresponding blow-up around $0$.
Observe that the curvature estimate above is scale invariant, so $\Sigma_i$ satisfies the same curvature bounds whenever $d_g(x,0)<r_i\delta$.
It follows that, up to a subsequence, $\Sigma_i$ converges locally graphically in the $C^{1,\alpha}$ topology on compact sets to a complete surface $\Sigma_\infty$ in $T_0N$, which we identify with $\real^3$ with the Euclidean metric.
Lemma \ref{schauder estimates} implies that $\Sigma_i$ in fact converges locally graphically in the $C^{2,\alpha}$ topology on compact sets of $\real^3\setminus\{0\}$.
In particular, $\Sigma_\infty$ has free boundary on $\{x\in\real^3:x_1=0\}\setminus\{0\}$ and for any compact set $K\subset\real^3\setminus\{0\}$ we have
\begin{equation*}
\int_{K\cap\Sigma_\infty}|A_{\infty}|^2 = \lim_{i\rightarrow\infty}\int_{K\cap \Sigma_i}|A_i|^2= \lim_{i\rightarrow\infty}\int_{(r_i^{-1}K)\cap \Sigma}|A_\Sigma|^2=0
\end{equation*}
That is, $A_\infty=0$.
Hence $\Sigma_\infty$ is an union of half-planes perpendicular to $\{x\in\real^3:x_1=0\}$.
\begin{claim}
If $\delta>0$ is sufficently small then $\Sigma\cap B^+_\delta\setminus\{0\}$ is topologically a finite and disjoint union of disks, half disks or half-disks punctured at $0$ with free boundary on $\partial_0 B^+_\delta\setminus\{0\}$.
\end{claim}
Firstly, we improve the curvature estimates.
Fix $y\in\Sigma_i\cap B^+_{r_i\delta}$ and put $x=r_i^{-1}y\in\Sigma\cap B^+_{\delta}$.
Then $|A_\Sigma(x)|d_g(x,0)=|A_i(y)|d_g(y,0)\rightarrow 0$ as $i\rightarrow\infty$ from the previous claim.
Thus $\lim_{x\rightarrow 0}|A_\Sigma(x)|d_g(x,0)=0$.
Secondly, we use a standard Morse Theory argument.
Let $f:\Sigma\cap B_\delta^+\rightarrow \real$ be defined by $f(x)=\frac{1}{2}d_g(x,0)^2$.
We can write its Hessian at a point $x$ and direction $v$ as $\textup{Hess}f(v,v)_x=A_\Sigma(v,v)_x g(N_\Sigma(x),\dot\gamma(1))_x+Q_x(v,v)$, where $N_\Sigma$ is the normal vector field on $\Sigma$, $\gamma$ is the minimizing geodesic in $B_1^+$ from $0$ to $x$ and $Q$ is a quadratic form satisfying $Q\geq\frac{1}{2}g$ as long as $d_g(x,0)<\delta$ is sufficiently small.
Hence, any critical point of $f$ in a small neighbourhood of $0$ is a strict local minimum, even if the critical point is along the boundary.
It follows from Morse Theory for manifolds with boundary \cite{jankowski-rubinsztein1972} that every connected component of $\Sigma\cap B^+_\delta$ is a disk with a single critical point, a free boundary half disk with a single critical point or a free boundary half disk punctured at $0$ without critical points.
Since critical points cannot accumulate, it must be finite for $\delta$ sufficiently small.
This proves the claim.
Next, we shall prove that for each punctured half-disk its tangent cone is unique.
Pick $\hat\Sigma\subset\Sigma\cap B^+_\delta\setminus\{0\}$ a connected component that corresponds to a punctured half-disk.
Let $\hat\Sigma_i$ be the corresponding dilation by $r_i\rightarrow\infty$ and $\hat\Sigma_\infty\subset\Sigma_\infty$ its tangent cone at $0$.
As we have seen above, $\hat\Sigma_\infty$ is a half-plane perpendicular to $\{x\in\real^3:x_1=0\}$.
\begin{claim}
The tangent cone $\hat\Sigma_\infty$ is unique, that is, it is independent of the blow-up sequence $r_i\rightarrow\infty$.
\end{claim}
Without loss of generality let us identify $\hat\Sigma_\infty$ with the half-plane $P_+=\{x\in\real^3: x_3=0\}\cap\{x\in\real^3: x_1\geq 0\}$.
We may also assume, possibly taking another subsequence, that the mean curvature vector of $\hat\Sigma_i$ points in the positive direction of the $x_3$-axis for $i$ sufficiently large.
Let $D_\theta$ and $C_\theta$ be as in Proposition \ref{proposition foliation}
Since $\hat\Sigma_i$ is embedded, then for $i$ sufficiently large $\hat\Sigma_i\cap\partial_+C_{\theta}$ is the graph of a function $w_i=w_{r_i}$ over $\partial_+D_{\theta}$.
We know that $\hat\Sigma_i$ converges in $C^{2,\alpha}$ graphical sense to $P_+$ away from zero, from which follows that $\|w_i\|_{2,\alpha}\rightarrow 0$.
The mean curvature of $\hat\Sigma_i$ is given by $\hat{H}_i={r_i}^{-1}H_\Sigma$ which is constant at each $i$ and tends to $0$ as $i$ tends to infinity.
If we denote by $g_i$ the corresponding blow up of the metric $g$ in a neighbourhood of $0$ in $B^+_1$, it follows from Proposition \ref{proposition foliation} that for all $i$ sufficiently large there exists $\varepsilon_0>0$ and a unique function $u_{i,t}=u(\hat{H}_i,t,g_i,w_i):D_{\theta}\rightarrow\real$ for each $|t|<\varepsilon_0$ such that $u_{i,t}=t+w_i$ on $\partial_+D_{\theta}$, the graph of $u_{i,t}$ over $D_{\theta}$ meets $\partial_0C_{\theta}$ orthogonally, it has constant mean curvature equal to $\hat{H}_i$ and its mean curvature vector points in the same direction as the mean curvature vector of $\hat\Sigma_i$.
Furthermore, $u_{i,t}$ varies smoothly on $t$ and it defines a foliation of a region $D_{\theta}\times[-c,c]$ for some $c>0$ independent of $i$.
Let $t_i\in(-\varepsilon_0,\varepsilon_0)$ be such that $u_{i,t_i}(0)=0$.
Since their mean curvature vector points in the same direction, we may apply the maximum principle to the graph of $u_{i,t_i}$ and $\hat{\Sigma}_i$ to conclude that $\hat\Sigma_i$ must be contained entirely to one side of the graph of $u_{i,t_i}$.
The side itself will depend on whether $t_i$ is positive or negative.
In case the mean curvature vector of $\hat\Sigma_i$ were to point in the negative direction of the $x_3$-axis, we would have used the same argument with $u(-\hat{H}_i,t,g_i,w_i)$ instead.
Finally, pick another sequence $r_i'\rightarrow \infty$ and let $\hat\Sigma_i'$, $P_+'$ be its corresponding blow-up and tangent cone respectively.
For each $k$ we pick $i_k$ so that $r'_{i_k}>k r_k$.
Observe that $\hat\Sigma_{i_k}=\frac{r_{i_k}'}{r_k}\hat\Sigma_k$ which is contained to one side of the graph of $\frac{r_{i_k}'}{r_k}u_{k,t_k}$.
Since each $u_{k,t_k}$ is regular at $0$, $\frac{r_{i_k}'}{r_k}u_{k,t_k}$ must converge to a unique tangent cone, that is, $P_+$.
If $P_+'$ were different from $P_+$, then it would imply that $\frac{r_{i_k}'}{r_k}\hat\Sigma_k$ contains points on both sides of the graph of $\frac{r_{i_k}'}{r_k}u_{k,t_k}$.
Notice that $P_+'$ must also be a half plane with free boundary on $\{x\in\real^3: x_1= 0\}$.
This concludes the claim.
It follows from uniqueness of the tangent cone that $\hat\Sigma\cup\{0\}$ is the graph of a $C^1$ function $u$ around a neighbourhood of $0$ over the tangent cone.
Since $u$ is also a solution to the CMC equation with smooth coefficients, then by elliptic regularity it must also be smooth.
To conclude the proof, now that we know that $\hat\Sigma\cup\{0\}$ is regular, it follows from the maximum principle that it is the unique connected component containing $0$.
Thus $\Sigma\cup\{0\}$ is properly embedded.
\end{proof}
We now state the equivalent result for removable interior singularities without repeating the proof.
\begin{theorem}\label{removable singularity interior}
Let $g$ be a Riemannian metric on $B_1$ and $\Sigma$ be a smooth, properly embedded, CMC surface in $B_1\setminus\{0\}$ with $H_\Sigma\leq H_0$, and $0\in\overline{\Sigma}$.
Suppose $\int_\Sigma|A_\Sigma|^2\leq C$ then $\Sigma\cup\{0\}$ is a smooth properly embedded CMC surface in $B_1$.
\end{theorem}
\begin{remark}
The proof is exactly the same with a minor modification on the construction of the foliation.
That is, by dropping the Neumann component on the definition of $\Phi$ in Proposition \ref{proposition foliation}.
\end{remark}
\section{Compactness Theorem}
In this section we will prove our main theorem for free boundary embedded CMC surfaces.
As we shall see, the limiting surface may not be embedded because CMC surfaces may have tangential self-intersection as long as the normal vector points at opposite directions.
\begin{theorem}\label{compactness fbcmc}
Let $N$ be a compact $3$-dimensional manifold with boundary.
Suppose $H_{\partial N}\geq H_0$ and let $\Sigma_i$ be a sequence of free boundary embedded CMC surfaces with mean curvature $H_i$, genus $g_i$ and number of ends $r_i$ satisfying:
\begin{enumerate}[(a)]
\item $|H_i|\leq H_0$;
\item $g_i\leq g_0$;
\item $r_i\leq r_0$;
\item $\textup{area}(\Sigma_i)\leq A_0$ and
\item $\textup{length}(\partial \Sigma_i)\leq L_0$.
\end{enumerate}
Then there exists a smooth properly almost embedded CMC surface $\Sigma\subset N$ and a finite set $\Gamma\subset \Sigma$ such that, up to a subsequence, $\Sigma_i$ converges to $\Sigma$ locally graphically in the $C^k$ topology on compact sets of $N\setminus \Gamma$ for all $k\geq 2$.
Moreover, if $\Sigma$ is minimal then it is properly embedded.
If in addition $(N,\partial N)$ satisfies either $Ric_N > 0$ and $A_{\partial N}\geq 0$ or $Ric_N \geq 0$ and $A_{\partial N} > 0$, then:
\begin{enumerate}[(i)]
\item when $H_\Sigma=0$ the convergence is at most $2$-sheeted;
\item when $H_\Sigma\neq 0$, then the convergence is $1$-sheeted away from $\Gamma$.
\end{enumerate}
\end{theorem}
\begin{proof}
Let us dente by $A_i$ the second fundamental form of $\Sigma_i$.
Given $x\in \Sigma$, it follows from Gauss equation that $|A_i|^2(x) = H_i^2 + 2K_N(T_x\Sigma) - 2K_i(x)$, where $K_N, K_i$ are the sectional curvatures of $N$ and $\Sigma_i$ respectively.
From Gauss-Bonnet theorem we have
\begin{equation*}
\int_{\Sigma_i}|A_i|^2 = H_i^2\textup{area}(\Sigma_i) + 2\int_{\Sigma_i}K_N(T_xN) +2\int_{\partial\Sigma_i}\kappa_g + 4\pi(2g_i+r_i -2),
\end{equation*}
where $\kappa_g$ denotes the geodesic curvature of $\partial\Sigma_i$.
Because $\Sigma_i$ is free boundary, we have that that $\kappa_g = A_{\partial N}(\tau_{\partial\Sigma},\tau_{\partial\Sigma})$, where $A_{\partial N}$ is the second fundamental form of $\partial N$ with respect to the inward unit normal vector and $\tau_{\partial \Sigma}$ is the unit tangent vector of $\partial \Sigma$.
Since $N$ is compact, there exists a constant $C=C(N)>0$ such that
\begin{equation*}
\int_{\Sigma_i}|A_i|^2\leq C(H_i^2\textup{area}(\Sigma_i)+g_i+r_i+\textup{area}(\Sigma_i)+\textup{length}(\partial\Sigma_i)).
\end{equation*}
Hence, from hypotheses $(a)$-$(e)$ we have that the total curvature is uniformly bounded by a constant $C_0=C_0(N,H_0,g_0,r_0,A_0,L_0)>0$.
Denote by $\mu_i$ the Radon measure on $N$ defined by $\mu_i(U)=\int_{\Sigma_i\cap U}|A_i|^2$, for a subset $U\subset N$.
It follows from the above that there exists a Radon measure $\mu$ in $N$ such that, up to a subsequence, $\mu_i$ converges weakly to $\mu$.
Furthermore, the set $\Gamma=\{p\in N: \mu(\{p\})\geq 1\}$ has at most $C_0$ elements.
For each $x\in N\setminus\Gamma$ there exists $r>0$ such that $\mu(\nball_r(x))<1$.
Hence, for each $i$ sufficiently large $\mu_i(\nball_r(x))<1$, that is, $\int_{\Sigma_i\cap\nball_r(x)}|A_i|^2<1$.
By possibly choosing a smaller value of $r$, it follows from Theorem \ref{curvature estimates} that
\begin{equation*}
\sup_{\Sigma_i\cap \nball_{\frac{r}{2}}}|A_i|^2\leq C,
\end{equation*}
for some constant $C>0$ independent of $i$.
Let $r<r_0$ as in Lemma \ref{schauder estimates} and suppose that, up to a subsequence, $\Sigma_i \cap \nball_{\frac{r}{4}}(x)$ is non-empty for all $i$ sufficiently large.
Since $\Sigma_i$ is embedded then $\Sigma_i \cap \nball_{\frac{r}{4}}(x)$ is the union of disjoint embedded connected components $\Sigma_{i,1},\ldots,\Sigma_{i,L}$ each of which is the graph of a function defined on an open ball of fixed radius on $T_{y_{i,j}}\Sigma_{i,j}$ for some $y_{i,j}\in\Sigma_{i,j}$, j=1,\ldots, L.
Because $\Sigma_i$ is compact, the number of sheets $L$ must be finite and thus constant for $i$ sufficiently large.
Hence, under the appropriate identifications we may further assume that $\Sigma_{i,j}\cap\nball_{r'}(x)$ is the graph of a function $u_{i,j}$ defined on $B_{r'}(0)\subset T_{y_i}\Sigma_i$, for some $0<r'<\frac{r}{8}$ depending only on $L$ and a fixed $y_i\in\Sigma_i\cap\nball_{r'}(x)$.
Furthermore $u_{i,j}$ has uniform $C^{2,\alpha}$ bounds as in Lemma \ref{schauder estimates}.
We may assume that, up to a subsequence, $y_{i}$ converges to $y'$ and $T_{y_{i}}\Sigma_i$ converges to a plane $P\subset T_{y'}N$.
In which case, under further identifications, we have that $\Sigma_{i,j}\cap\nball_{\frac{r'}{2}}(y')$ is the graph of a function $u_{i,j}'$ defined on an open ball on $P$ (or half-ball in case $y'$ is on the boundary of $N$) and uniform $C^{2,\alpha}$ estimates for all $i$ sufficiently large, and each $j=1,\ldots, L$.
Hence, up to a subsequence $u_{i,j}'$ converges to a function $u_j'$ in the $C^{2,\beta}$ topology for all $\beta<\alpha$ and $\Sigma_{i} \cap \nball_{\frac{r'}{2}}(y')$ converges to $\Sigma_{j}'=\textup{graph}( u_j')$ for each $j=1,\ldots, L$.
From which follows that $y'\in\Sigma'=\cup_j\Sigma_j'$ and $P$ is in fact $T_{y'}\Sigma'$.
Now, given any compact set $K\subset N\setminus \Gamma$ we cover it by finitely many open balls $\{\nball_{\frac{r}{4}}(x_k)\}_{k=0,\ldots,m}$ as above so that, up to a subsequence, $\Sigma_i\cap(\cup_k \nball_{\frac{r}{4}}(x_k))$ converges to a surface $\Sigma'\subset N\setminus\Gamma$ locally graphically in the $C^{2,\beta}$ topology on $K$.
Taking a countable exhaustion by compact sets and using a diagonal argument we have that, up to a subsequence, $\Sigma_i$ converges to $\Sigma'$ locally graphically in the $C^{2,\beta}$ topology on compact sets of $N\setminus\Gamma$.
Smooth convergence away from $\Gamma$ follows from Allard's regularity Theorem \cites{allard1972, allard1975, gruter-jost1986}.
Since $\Sigma_i$ is a properly embedded CMC surface with free boundary along $\partial N\setminus \Gamma$, then $\Sigma'$ is a properly almost embedded CMC surface in $N\setminus \Gamma$ with free boundary along $\partial N \setminus \Gamma$.
Observe that on a neighbourhood of any self-touching point of $\Sigma'$ the surface can be written as connected components that lie to one side of one another.
A transversal self-intersection is an open condition so it would contradict the fact that $\Sigma_i$ is embedded.
Define $\Sigma$ to be the closure of $\Sigma'$, so $\Gamma=\Sigma\setminus\Sigma'$.
Since $\Gamma$ is finite, there exists $r > 0$ so that $\nball_r(p)\setminus\{p\}$ contains no points of $\Gamma$.
From graphical convergence on compact sets of $\nball_r(p)\setminus\{p\}$, each sheet must converge to an embedded component of $\Sigma\cap \nball_r(p)\setminus\{p\}$.
Since there are only finitely many sheets, we may pick $r>0$ sufficiently small so that every component of $\Sigma\cap \nball_r(p)\setminus\{p\}$ contains $p$ in its closure.
\begin{claim}
The limit surface $\Sigma$ is regular.
\end{claim}
We know that $\int_{\Sigma\cap (\nball_r(p)\setminus\{p\})}|A_\Sigma|^2\leq C_0$.
Thus we may apply the Removable Singularity Theorems \ref{removable singularity boundary} or \ref{removable singularity interior} to each embedded component of $\Sigma\cap (\nball_r(p)\setminus\{p\})$ depending on whether $p$ belongs to the boundary of $N$ or to the interior.
Hence $\Sigma=\Sigma'\cup\Gamma$ is a regular surface.
Now, suppose that $H_\Sigma=0$ then the Maximum Principle for minimal surfaces implies that there are no self-touching points so $\Sigma$ is embedded.
We now prove properties (i) and (ii).
Henceforth, let us assume that either $Ric_N>0$ and $A_{\partial N}\geq 0$ or $Ric_N\geq 0$ and $A_{\partial N}>0$.
The following argument is the same as in \cite{sharp2017} and \cite{ambrozio-carlotto-sharp2018.3} and we include the main idea without repeating the calculations.
\begin{claim}
If $H_\Sigma\neq 0$ then the convergence is $1$-sheeted away from $\Gamma$.
\end{claim}
Suppose by contradiction that $\Sigma_i$ converges to $\Sigma$ with at least $2$ sheets.
Since the convergence is graphical over compact sets of $N\setminus\Gamma$, for any compact set $\Omega\subset \Sigma\setminus\Gamma$ and a sufficiently small tubular neighbourhood $V_r(\Omega)$ of $\Omega$, $\Sigma_i\cap V_r(\Omega)$ contains at least $2$ connected components $\Sigma_i^0$ and $\Sigma_i^1$ each of which can be written as follows, for $i$ sufficiently large:
\begin{equation*}
\Sigma_i^\nu=\{exp_p(u_i^\nu(p)N_\Sigma(p)):p\in \Omega \},
\end{equation*}
for $u_i^\nu:\Omega\rightarrow \real$, $\nu=0,1$.
We may assume without loss of generality that $u_i^1 > u_i^0$ for all $i$ sufficiently large.
Let us denote $v_i(t)=u_i^0+t(u_i^1-u_i^0)$ and $\Phi_i(p,t)=exp_p(v_i(t)(p)N_\Sigma(p))$ for $p\in\Omega$.
Consider the variation $\Sigma_i(t)=\{\Phi_i(p,t):p\in\Omega\}$ so that $\Sigma_i(\nu)=\Sigma_i^\nu$, $\nu=0,1$.
Take a function $w\in C^\infty_c(\Omega)$ with compact support and define $w^\nu\in C^\infty_c(\Sigma_i^\nu)$ by $w^\nu(\Phi(p,\nu))=u(p)$ for each $\nu=0,1$.
Now, consider any compactly supported vector field $X$ in $V_r(\Omega)$ such that $g(X,N_{\Sigma_i^\nu})= w^\nu$ along $\Sigma_i^\nu$, for each $\nu=0,1$, and its
associated flow $\Psi(q,s)$ on $V_r(\Omega)$.
Finally we consider the following $2$-parameter variation:
\begin{equation*}
\Sigma_i(t,s)=(\Psi_s)_\sharp\Sigma_i(t).
\end{equation*}
From the first variation formula it follows that
\begin{equation*}
\frac{\partial}{\partial s}{|_{s=0}} area(\Sigma_i(\nu,s))= - H_i\int_{\Sigma_i^\nu}w^\nu,
\end{equation*}
for each $\nu=0,1$.
Since the mean curvature vectors of each sheet converge to $\vec H_\Sigma$, we have that their normal vectors point in the same direction for $i$ sufficiently large.
That is, their scalar mean curvature has the same sign, so we have
\begin{equation*}
\frac{\partial}{\partial s}{|_{s=0}} area(\Sigma_i(1,s))-\frac{\partial}{\partial s}{|_{s=0}} area(\Sigma_i(0,s))=- H_i\int_\Sigma (J\Phi_i^1-J\Phi_i^0)w,
\end{equation*}
where for each $\nu=0,1$, $\Phi_i^\nu(p)=\Phi_i(p,\nu)$ and $J\Phi_i^\nu$ its corresponding Jacobian.
Following the same idea and calculations as in \cite{ambrozio-carlotto-sharp2018.3}, using the Mean Value Theorem and taking the second variation of area for some $t_i\in(0,1)$, we obtain that $\tilde h_i=u_i^1-u_i^0$ satisfies
\begin{equation*}
-\int_\Omega w \left(L_i\tilde h_i+H_iF_i(\tilde h_i)\right)+\int_{\partial N\cap \Omega} B_i(w,\tilde h_i)= 0.
\end{equation*}
Where $L_i$ is an elliptic operator in divergence form, $L_i$ converges uniformly to the Jacobian operator $L_\Sigma$, of $\Sigma$ as $i\rightarrow\infty$, $F_i(\tilde h_i)=(J\Phi_i^1-J\Phi_i^0)$ and $B_i(w_1,w_2)$ converges uniformly to $w_1(\frac{\partial w_2}{\partial \eta }+A_{\partial N}(N_\Sigma,N_\Sigma)w_2)$, where $\eta$ denotes the outward co-normal vector on $\partial \Sigma$ and $A_{\partial N}$ the second fundamental form with respect to the outward co-normal of $\partial N$.
Since $u_i^\nu$ tends to $0$ at least $C^2$, then $\Phi_i^\nu$ tends to the identity uniformly, that is, $F_i(\tilde h_i)$ tends to $0$.
Fix a point $q_0\in\Omega$ and define $h_i(p)=\frac{\tilde h_i(p)}{h_i(q_0)}$.
Using Harnack estimates and a blow-up contradiction argument we have that $h_i$ converges smoothly to $h$ on $\Omega$ satisfying
\begin{equation*}
\left\{
\begin{aligned}
& L_\Sigma h = 0, \textup{ on }\Omega;\\
& \frac{\partial h}{\partial\eta} +A_{\partial N}(N_\Sigma,N_\Sigma)h=0, \textup{ on } \partial\Sigma\cap\Omega,\\
\end{aligned}
\right.
\end{equation*}
Since the blow up of the elliptic equation for $h_i$ is the one above, the proof is the same as in \cite{ambrozio-carlotto-sharp2018.3}*{Claim $1$, p.20}.
It follows from the Maximum Principle for elliptic equations that $h>0$ on $\Omega$.
By taking an exhaustion of compact sets over $\Sigma\setminus\Gamma$ and using a diagonal sequence argument, we have that $h_i$ converges to $h>0$ on compact sets of $\Sigma\setminus\Gamma$.
Again, it follows from the exact same argument as in \cite{ambrozio-carlotto-sharp2018.3} that $h$ is uniformly bounded on $\Sigma\setminus\Gamma$ thus it extends to a smooth function over $\Sigma$, which is positive by the maximum principle.
Finally, we observe that the conditions $Ric_N>0$ and $A_{\partial N}\geq 0$ or $Ric_N\geq 0$ and $A_{\partial N}>0$ imply that the quadratic form
\begin{equation*}
Q_\Sigma(w_1,w_2)=-\int_\Sigma w_1 L_\Sigma w_2+\int_{\partial \Sigma} w_1\left(\frac{\partial w_2}{\partial \eta }+A_{\partial N}(N_\Sigma,N_\Sigma)w_2\right)
\end{equation*}
is positive definite.
Hence its first eigenvalue is positive.
However, $h$ is a positive eigenfunction so it must correspond to the first eigenvalue, which is a contradiction since the corresponding eigenvalue is $0$.
We conclude that the convergence must be $1$-sheeted away from $\Gamma$.
The case $H_\Sigma=0$ follows the exact same arguments as above.
Unlike the previous case, the mean curvature vector of each graph converges to $0$ so they may have opposite orientation.
As a consequence, the scalar mean curvature may have opposite sign, hence the function $F_i$ would not tend to $0$ as desired.
\begin{claim}
If $H_\Sigma=0$ then the convergence is at most $2$-sheeted.
\end{claim}
Suppose the convergence is at least $3$-sheeted.
Then at least $2$ of which, say $u^1_i, u^0_i$ have the same orientation, that is, the normal vector along their graphs points in the same direction.
Following the same argument as in the previous case, the function $F_i$ will tend to $0$ so we can construct a positive eigenfunction $h>0$ for the Jacobi operator $L_\Sigma$ on $\Sigma$, corresponding to the eigenvalue $0$.
This is a contradiction because $L_\Sigma$ is a positive operator under either of the conditions $Ric_N>0$ and $A_{\partial N}\geq 0$ or $Ric_N\geq 0$ and $A_{\partial N}>0$.
This completes the proof.
\end{proof}
\bibliographystyle{plain}
|
\section{Introduction}\label{sec:intro}
\begin{figure*}[!t]
\centering
\includegraphics[width=\linewidth, page=1]{./figs/utils/ris_nonterrestrial}
\caption{RISs can enable low-cost and energy-efficient non-terrestrial communication links including HAPS, satellites, and even deep space communications systems like the lunar relays. Furthermore, the mechanical conformity of RISs makes light-weight and high-capacity communication subsystems in satellites possible.}
\label{fig:general_view}
\end{figure*}
It is widely accepted that user-centric and ubiquitous connectivity, which are desired by both end-users and operators \textcolor{black}{for 6G and beyond}, can be achieved through unique orchestration of terrestrial and \acp{NTN} in next-generation communication systems~\cite{tekbiyik2020holistic}. This vision is also described by the 3rd Generation Partnership Project (3GPP) in TR~38.821 for the operation of New Radio (NR) in \acp{NTN}. An \ac{NTN} basically consists of unmanned aerial vehicles, \ac{HAPS} systems, and satellite deployments. Among these, \ac{LEO} satellites and \ac{HAPS} systems are considered to be the key enablers for \acp{NTN} due to their unique features, which include longer operating times and wider coverage areas~\cite{kurt2020vision}. HAPS systems, positioned at altitudes of $20-50$ km possibly in a quasi-stationary manner, can provide significant advantages while connecting space and terrestrial networks~\cite{kurt2020vision}. LEO satellites, for their part, present a remarkable potential for \acp{DSN}, serving as a densely deployed near-Earth network to enable connectivity to and from the Earth.
The number of space studies conducted by various organizations led by NASA is increasing. It is aimed that humanity will again reach the Moon by starting space travels in 2024. There has also been an increase in studies of Mars, and ways to establish a second world for humanity are being sought. \textcolor{black}{For instance, a relay network composed of multiple orbiters was used during the landing of Perseverance and the transmission of data on the red planet, which is a step that NASA has recently carried out as part of its Mars mission.} Although the communication systems installed in space for this purpose are advanced technology, their service quality is far below that of terrestrial communication technology due to limitations for \ac{SWaP} and cost associated with operating in deep space. \textcolor{black}{For example, the link between Perseverance and the Mars orbiter can only provide a data rate of $2$ Mbps~\cite{Communic97online}.}
As illustrated in \FGR{fig:general_view}, \ac{LEO} satellites and \ac{HAPS} systems have the potential to operate as key relays between \acp{DSN} and terrestrial networks due to the seamless and ubiquitous connectivity they provide. In the coming years, data traffic both in HAPS systems and LEO networks and \ac{LEO} networks is expected to increase because of proliferation of satellites, and high data demand from space missions. However, due to the increased transmission distances,the transmission power required will also increase to compensate for the high path loss. As \ac{HAPS} systems and \ac{LEO} satellites are strictly limited by the \ac{SWaP} requirements, they can employ fewer RF chains. \acp{RIS}, which have been recently proposed to maximize the \ac{SNR} at the receiver by adjusting the incident wave phase with only a single RF chain~\cite{huang2019reconfigurable}, can be employed in \acp{NTN} to improve energy efficiency~\cite{giordani2020non}. Due to drawbacks with current antenna array systems in both NTNs and DSNs, RISs could improve communication performance in a passive and energy-efficient way. The seminal works (e.g.,~\cite{basar2019wireless, huang2019reconfigurable}) on RISs demonstrate that they can offer better communication performance while consuming less power.
Previously, the use of a \ac{DTN} was intended to tolerate delays caused by the intermittent capture of the direct-link so that communication between remote stations on Mars or other planets and Earth would be reliable~\cite{burleigh2003delay}. \acp{DTN} produce solutions in upper layers instead of the physical layer, but they cannot increase the achievable communication duration between stations. By leveraging the beamforming and steering capabilities of RISs, it is possible to track near Earth satellites and orbiters for a longer time. To do this, reflectarrays and phased arrays are currently used in deep space systems. Although reflectarrays can produce a sharp beam, they cannot steer. Besides, phase-arrays suffer from both cost and \ac{SWaP} drawbacks as they contain multiple active RF elements. RISs therefore appear to be a promising solution for both NTNs and DSNs, since RISs include only a single RF chain, reduces the cost and size. This study addresses possible use-cases and challenges of \acp{RIS} in \acp{NTN} and \acp{DSN} considering the unique characteristics of these networks while highlighting the potential of \ac{DL} techniques to address challenging channel environments.
The rest of the paper is organized as follows. In~\SEC{sec:ris_motivation}, we discuss the motivation behind RISs in non-terrrestrial and deep space communications. \SEC{sec:use_cases} introduces possible use cases for \ac{RIS}-assisted \acp{NTN}. \SEC{sec:challenges} lists the challenges and introduces possible solutions, which are supported with numerical results. Open issues for the realization of \ac{RIS}-assisted \acp{NTN} and \acp{DSN} are discussed in~\SEC{sec:open_issues}. \SEC{sec:conclusion} concludes the study.
\section{Why \acl{RIS}s?}\label{sec:ris_motivation}
In addition to being a candidate technology for 6G, \ac{RIS} communications present a remarkable potential for future wireless communications. There are numerous attractive features of utilizing \acp{RIS} in \acp{NTN}, but three of them stand out: energy efficiency, compatibility with \ac{SWaP} requirements, and affordability.~\cite{basar2019wireless}.
\subsection{Energy Efficient System Architecture}
While \ac{HAPS} and satellite systems are expected to yield significant gains in terms of capacity and coverage for wireless communications, we should not forget that these systems have energy constraints~\cite{kurt2020vision}. Therefore, it is necessary to carefully consider energy management in system designs. Along this line, solar-powered \ac{HAPS} systems can be utilized for wireless communications, but this does not mean that the energy of a HAPS or satellite is unlimited. The capacity of fuel cells is one of the lead determinants of a satellite's operational life. At this point, it should be noted that the ability of \acp{RIS} to configure the phases of meta-atoms to maximize the \ac{SNR} at the receiver can provide a considerable amount of gain for an energy-efficient communication system.
Owing to the virtual \ac{MIMO} structure and single RF chain of RISs, when they are used as transmitters, they require less operational and transmission energy than traditional \ac{MIMO} systems that contain many active circuit elements~\cite{basar2019wireless}. In the context of Lunar relay systems and other relays in DSNs, \acp{RIS} can enable low-power relay systems in deep space due to their capacity to maintain almost the same \ac{EIRP} with a significantly reduced power consumption~\cite{huang2019reconfigurable}.
\subsection{Compatibility with SWaP Requirements}\label{sec:swap}
Non-terrestrial systems must strictly obey their \ac{SWaP} constraints, since these constraints directly impact the OPEX and CAPEX of non-terrestrial communication ensembles. For example, besides the production cost of satellites, the cost of launching satellites into orbit is also quite high. One of the main factors affecting this cost is the weight of a satellite system. Although space transport today costs less than in the past, launches still require satellite operators to have a significant budget. To illustrate , the average launch between the years 1970 and 2000 cost $\$16000$ per kg. but this has decreased to $\$6000$ per kg as of 2018~\cite{jones2018recent}. Recently proposed reusable launch systems have reduced the cost of space missions, but it is still quite high. A reduction in the weight of satellites and \ac{HAPS} systems can reduce the cost per operation significantly. Furthermore, aircraft and spacecraft (e.g., \ac{HAPS}, satellites, orbiters) can carry a limited payload due to their aerodynamics.
\acp{RIS} can be smaller in size than phased array structures while still providing extensive coverage~\cite{huang2020holographic}. In order to serve ground users with small antennas, satellite antennas are designed to be as large as possible. This situation increases the risk of collisions due to the increasing density of satellite deployments. The main factors in the deployment cost of a \ac{HAPS} are weight, power consumption, and flight duration. Utilizing \ac{RIS} can reduce operational costs and extend flight duration of \acp{HAPS}, since \acp{RIS} are both lightweight and low-power consuming devices compared to traditional phased-arrays. It is expected that the OPEX and CAPEX can be gradually reduced by utilizing \acp{RIS} in \acp{NTN}.
Given that \ac{HAPS} systems and satellites are the first point of contact between Earth and deep space, they can be considered to increase the signal power received or transmitted by using \acp{RIS} which enhance the \ac{QoS} and energy efficiency.
\subsection{Relatively Affordable Alternative Solution}
In addition to the lower operational costs discussed above, another important factor that makes \acp{RIS} attractive is that they can be produced at a relatively lower cost compared to conventional multi-antenna systems. The main reason for this is that \acp{RIS} include a single RF chain which is the most costly part of the transceivers. For example, it has been reported that the first prototype introduced in~\cite{dai2020reconfigurable} reduced hardware costs compared to those of phased array systems.
\section{Use Cases}\label{sec:use_cases}
Although \ac{NTN} standardization is underway by 3GPP in Rel. 16 in TR~38.821, the use cases extend well beyond densely deployed mega-constellations. This section focuses on RIS use cases in high altitude systems and space networks.
\paragraph{\ac{RIS}-assisted \ac{HAPS}}
Here, we consider scenarios involving RISs on HAPS systems. Ultra-dense deployment of \ac{HAPS} systems can enhance ground coverage and eliminate coverage gaps~\cite{kurt2020vision}. Furthermore, it is possible to provide seamless connectivity by handling the handoffs of \ac{LEO} satellites. By making use of the large surfaces of a \ac{HAPS} for the placement of \acp{RIS}, both coverage and energy efficiency can be improved. In addition to serving rural areas, HAPS systems can also increase the overall throughput of the entire network. In addition, as the payload weight and the amount of energy consumed by the communication payload decrease with the use of \acp{RIS}, the average flight duration of a \ac{HAPS} is expected to increase.
By enabling \ac{RIS}-assisted \ac{HAPS}-satellite integrated networks, a multi-\ac{RIS} scheme~\cite{yildirim2020modeling} can be used to increase the error performance and achievable data rate. Furthermore, the prediction that dense \ac{LEO} satellite deployments will be an important enabler of next generation communication systems increases interest in investing in space technology. The use of \acp{RIS} in \ac{LEO} satellite networks shows that both error performance and achievable data rates can be significantly improved~\cite{tekbiyik2020reconfigurable}. \textcolor{black}{Since it is expected that \ac{HAPS} systems will bridge the space and terrestrial networks, outfitting them with \acp{RIS} introduces a significant performance advantage.}
\paragraph{Integration with \acp{DSN} towards Human Settlement in Space}
Communication and navigation systems will be indispensable subsystems in future space missions. For instance, NASA plans to land on the Moon's south pole by 2024 as part of the Artemis project. This landing is connected with the broader objective of building permanent structures, settling, and ultimately sending humans to Mars. However, some communication and navigation issues must be handled carefully since space agencies now aims to establish a living station like a Lunar Gateway. Thus, a reliable and secure communication link between a control station on Earth and a Lunar Gateway will be indispensable.
Furthermore, this communication system will need to support high data rates in order to transmit high-resolution camera and sensor data for mapping celestial bodies and analyzing soil in interplanetary reconnaissance. In terms of manned reconnaissance, it can be clearly predicted that communication security and continuity will be vital. For example, it takes up to $20$ hours for NASA to receive $250$ megabits of data transmitted directly from Mars to the Earth~\cite{mars_comm}. The main reasons for this are that the rover can be in\ac{LOS} with the Earth for only a few hours a day, and the rover's transmission power is limited. By utilizing \acp{RIS} with non-terrestrial (i.e., \acp{DSN}), the transmission time can be extended by employing multi-hop relays with steering capabilities. When a direct LOS is blocked, it will still be possible to transmit information over consecutive multi-\ac{RIS} satellites and improve error performance, as shown in~\cite{tekbiyik2020reconfigurable}. Also, a rover can direct beams towards a target by utilizing passive beamforming supported by \ac{RIS}. \textcolor{black}{Likewise, Perseverance uses a high gain steerable antenna for the same aim~\cite{Communic97online}.}
A relay station-mounted \ac{RIS} can be deployed on Mars to extend the \ac{LOS} duration by leveraging the \ac{RIS}'s beamforming capabilities. \textcolor{black}{The reflect arrays have currently been utilized in \acp{DSN}. Without essential modifications on the system, the performance can be improved by only adjusting the phase shifts~\cite{basar2019wireless}.} By providing continuous connectivity, \ac{IP}-based networks would start to take the place of \acp{DTN} operating with the store-and-forward paradigm. Especially in cases where remotely controlled vehicles and rovers are employed during the operation, this type of communication protocol is very likely to cause failure of the operations. Supporting extra-terrestrial cellular communication systems with \acp{V-HetNet}~\cite{tekbiyik2020holistic} together with RISs would be a game-changer for space communications. It is expected that such extra-terrestrial cellular communication systems will be installed on the Moon before the realization of manned research projects and/or human settlements.
\section{Operational Challenges}\label{sec:challenges}
Although RISs have many attractive features, they are still at an early stage of being integrated into non-terrestrial systems and there are many challenging tasks still to be addressed. Due to long distance and short durations of scheduled direct communication, the transmitted data cannot be received properly via \ac{IP}. A physical layer boost involving an \ac{SNR} improvement will certainly be useful to the connectivity of these links. In this section, we focus on the challenges related to this physical layer, which are expected to have a significant effect on the link performance, and we also numerically highlight the potential of using \acp{RIS} to increase the \ac{SNR}.
\subsection{\ac{RIS} Deployment on NTN Systems}
The first and most fundamental problem in implementing RIS-powered aerial systems is undoubtedly the integration of smart surfaces in aerial platforms. Considering the flight aerodynamics and mechanical structures of a \ac{HAPS} or satellites, it is clear that a serious design problem needs to be addressed: namely that depending on changes in position and elevation angles some or all meta-atoms are likely to be shaded by the system itself.
We believe that conformal metasurfaces can be used in the solution of this problem depending on the development of their structures. Conformal metasurfaces can allow the coating of irregular surfaces or arbitrarily shaped 3-D objects, thus allowing the coating of spherical and elliptical surfaces with \acp{RIS}. Thus, surfaces more suitable for flight aerodynamics can be obtained. By utilizing artificial impedance surfaces, the generation of any desired radiation pattern becomes possible when the surface is designed to integrate antennas into complex-shaped objects. An \ac{RIS} can be deployed on the surface of tethered balloons and on the lower and upper surfaces of a \ac{HAPS}. \acp{RIS} can also be used to replace the antenna subsystems already on satellites.
Solar-powered RIS-based relay stations can thus be built to support deep space missions by expanding the coverage areas and increasing capacity. It should be noted that different application-specific \ac{RIS} designs may be needed in terms of material and conformal coating, due to the different deep-space environments and atmospheric weather conditions.
\subsection{Solar Scintillation}\label{sec:scintillation}
\begin{figure}[!t]
\centering
\includegraphics[width=\linewidth, page=2]{./figs/utils/ris_nonterrestrial}
\caption{Solar scintillation is a phenomenon caused by changes in plasma density in solar winds ejected from the Sun. Solar scintillation crucially degrades the performance of space communications \textcolor{black}{by affecting the power and phase of the transmitted signal due to diffraction and refraction.}}
\label{fig:solar_scintillation}
\end{figure}
\begin{figure}[!t]
\centering
\includegraphics[width=\linewidth]{./figs/results/ber_vs_scintillation_pt_no.eps}
\caption{\ac{BER} performance analysis of an \ac{RIS}-assisted \acl{ISL} of a Starlink-like constellation whose inter-satellite distance is $945.4$ km, which indicates that the solar scintillation can be diminished by increasing the number of meta-atoms. The number of meta-atoms, scintillation index, and transmit power-to-noise ratio are denoted by $N$, $m$, and $P_{t}/N_{0}$, respectively.}
\label{fig:ber_scintillation}
\end{figure}
It is vital to establish robust and reliable communication links between the Earth and remote space stations, reconnaissance robots, and even astronauts for deep space missions. In order to provide reliable communications, the space environment should be accurately modeled. It should be noted that many factors --some of which are not yet known-- may disrupt the structure of the transmitted signals in space.
Solar scintillation, is caused by irregularities in plasma density in solar winds as illustrated in~\FGR{fig:solar_scintillation}. Solar scintillation substantially decreases communication performance. The scintillation effect reaches its peak during the solar conjunction, which refers to the alignment of the Earth, Sun, and another communication node on the same line. The \ac{SEP} angle is defined to show the solar conjunction state. This angle helps us understand how close celestial bodies are to the solar conjunction.
To illustrate, the SEP angle between the Earth and Mars is below $20^{\circ}$ for $170$ days over a $780$-day period~\cite{xu2019effects}. This means that solar scintillation would have a serious effect for much of this $780$-day period. To address this, we investigate the impact of solar scintillation on an \ac{RIS}-assisted \ac{ISL} with \ac{BPSK} signaling. In accordance with the \ac{ISL} frequency given in ITU-R S.1591, the link between two satellites in the same orbit is simulated under the assumption of the Starlink constellation at $23$ GHz. The distance between two nearby satellites is approximately $945.4$ km~\cite{tekbiyik2020reconfigurable}. It is therefore assumed that the satellites are operating under solar scintillations with varying strengths. The \ac{BER} performance of the \ac{RIS}-assisted \textcolor{black}{intra-plane} \ac{ISL} is depicted in~\FGR{fig:ber_scintillation}. The error performance is inversely proportional to the square of the scintillation index at a low scintillation levels~\cite{tekbiyik2020reconfigurable}. Compared to a weak scintillation effect, it is necessary to increase the output power by approximately $14$ dB in order to remain the same error performance in the transition zone. \textcolor{black}{These results reveal that communication units should adapt their transmit power according to their positions relative to the Sun.}
\subsection{Misalignment Fading}
\begin{figure}[!t]
\centering
\includegraphics[width=0.9\linewidth, page=3]{./figs/utils/ris_nonterrestrial}
\caption{Atmospheric drag results in the satellite moving towards the Earth and the conditions can destabilize a \ac{HAPS} deployment. For this reason, a misalignment between antennas can occur, especially for high-frequency, sharp beams.}
\label{fig:atmospheric_drag}
\end{figure}
Another challenging aspect of \ac{RIS}-assisted \acp{NTN} is the misalignment fading that can occur due to an error in the alignment between the transmitter and receiver antennas. There may be various reasons for this error. For instance, the position and altitude of a HAPS system in the Earth's stratosphere might change suddenly on account of heavy wind and related atmospheric disturbances. Also, the density of gases in the mesosphere is high, so there is likely to be molecular absorption and the refraction of waves. This layer of the atmosphere is warmer than the others, as most of the radiation from the sun is absorbed by the particles in the thermosphere. The difference in density between the layers may cause diffraction of the waves, which may cause misalignment.
\ac{LEO} satellites can also experience misalignment errors due to satellite drag, as illustrated in~\FGR{fig:atmospheric_drag}. This results from a collision of gas molecules. The particles, X-ray, and ultraviolet beams ejected from the Sun heat and expand the atmosphere; thus, they rise the collision probability. The atmospheric drag results in the satellite moving towards the Earth. The drag force on satellites is related to the flow between the higher and lower density layers of atmosphere. As atmospheric warming during active period of the Sun increases this flow, the drag force rises. In order to ensure reliable communications via \acp{NTN}, it is critical to evaluate possible misalignment effects on RIS-assisted communication systems. In our previous work~\cite{tekbiyik2020reconfigurable}, misalignment fading denoted by jitter, $\sigma_{s}$, was considered in \ac{RIS}-aided communications for \acp{ISL}.
Drawing on the findings in~\cite{tekbiyik2020reconfigurable}, we examine the BER performance of RIS-assisted intra-plane ISLs with misalignment fading by employing the same simulation parameters as in \SEC{sec:scintillation}. \FGR{fig:ber_misalignment} demonstrates that significantly high power is needed to maintain the same BER performance level. $250$ dB of additional power is required to keep the \ac{BER} at a level of about $10^{-7}$ when the offset between the centers of the receiver and transmitter antenna beams is around $1$ m. \textcolor{black}{The simulation results regarding misalignment fading in \ac{RIS}-assisted communication systems show the necessity for instantaneous beam alignment methodologies, which are still open in the literature.}
\begin{figure}[!t]
\centering
\includegraphics[width=\linewidth]{./figs/results/ber_vs_misalignment_pt_no.eps}
\caption{\ac{BER} versus $P_{t}/N_{0}$ under the misalignment fading, which can seriously decrease the communication performance in terms of error probability, for \ac{BPSK} signaling. The radius of equivalent beamwidth is chosen as $10$ m and $\sigma_{s}$ denotes the misalignment jitter.}
\label{fig:ber_misalignment}
\end{figure}
\subsection{Channel Estimation}
Because \acp{RIS} can adjust the amplitude and/or phase of the incident electromagnetic wave, they can largely override the randomness of the propagation medium. But doing this certainly requires high-quality \ac{CSI} first. Therefore, channel estimation plays a vital role in terms of \ac{QoS} in non-terrestrial communication as well. For this purpose, \ac{DL} can be used to achieve high performance in channel estimation under challenging channel conditions. Recently, we proposed the use of \acp{GAT} for channel estimation in \ac{RIS}-assisted full-duplex \ac{HAPS} backhauling~\cite{tekbiyik2020channel}. This model can also be used in different network architectures. The \ac{GAT} can reduce the computational complexity and increase the learning capacity by taking unseen nodes within the proposed \ac{RIS}-assisted \ac{NTN} in the process~\cite{tekbiyik2020channel}. The model is trained by assigning the received signal samples and known pilot samples to the nodes and vertices of graphs, respectively. By utilizing the trained model proposed in~\cite{tekbiyik2020channel}, we investigate the channel estimation performance from the point of normalized mean square error under varying solar scintillation by employing the same model introduced in~\cite{tekbiyik2020channel}. \FGR{fig:nmse_scintillation} indicates that the proposed model is robust against the changes in the channel characteristics. Not even strong solar scintillation degrades the performance of the \ac{GAT} estimator.
The proposed model can also estimate channel coefficients with almost the same high performance as at various solar scintillation levels, namely weak, transition zone, and strong. The reason for this performance is the attention mechanism which enables the model to focus on the most relevant information in the input; hence, \ac{GAT} can be effective under changing channel parameters even if the model has been trained under better conditions than those of the test stage. \textcolor{black}{However, further research along this direction is required for a holistic exploration.}
\begin{figure}[!t]
\centering
\includegraphics[width=\linewidth]{./figs/results/nmse_vs_scintillation.eps}
\caption{\ac{GAT} can perform accurate channel estimation under varying solar scintillation effects, which can be grouped into weak ($m<0.3$), transition zone ($0.3\leq m<1$), and strong scintillation ($m=1$). The circular, square and dot markers denote misalignment index $m = 0.1$, $0.5$, and $1$, respectively.}
\label{fig:nmse_scintillation}
\end{figure}
\vspace{-0.2cm}
\section{Open Issues and Research Directions}\label{sec:open_issues}
\ac{RIS}-assisted \acp{NTN} have the potential to significantly improve the capacity and reliability of personal communications and deep space communications. Despite these promises, many open issues remain and must be carefully considered.
\paragraph{Channel Models for \acp{NTN} and \acp{DSN}}
As is well known, an understanding of channel behavior on both a large scale and small scale is essential for developing a link budget analysis and determining the waveform and its parameters. However, the \ac{RIS}-assisted network topologies discussed in this study are challenging in two ways.
First, channel models for stratospheric transceivers do not allow a proper understanding of the channel behavior as the atmospheric effects have not yet been considered in detail~\cite{kurt2020vision}. Also, \ac{HAPS}-to-\ac{HAPS} channels have yet to be investigated for mobility cases. 3-D deployments are now replacing 2-D deployments, channel models need to be handled in accordance with this dimensional expansion. Furthermore, channels between satellites and \ac{HAPS} systems should be examined in terms of both mobility and near-space effects.
Conversely, since channel modeling studies for satellitea and \acp{DSN} are still at an early stage, this area also needs to be developed with appropriate channel models. For example, while the small-scale fading effect created by the solar scintillation in the amplitude of the signal is Rician distributed for low scintillation levels, a full-fledged model has not yet been proposed for strong scintillations~\cite{xu2019effects}. Therefore, the coronal channel modeling should be further studied. A comprehensive study on \acp{DSN} and channel models suitable for the atmospheres of other planets is also essential.
The second challenge is that all these channel modeling studies should also be generalized for \ac{RIS}-assisted communication, but RIS channel models are still at an early stage, even for terrestrial use cases. Although meta-atoms are referred to as theoretically perfect reflectors, their performance may vary depending on the type of material used in practice. Also, the correlation between meta-atoms can affect the behavior of the communication channel~\cite{basar2019wireless}. For this reason, the channel models recommended for \acp{RIS} must first be verified through measurements.
\paragraph{Design and Operation of \acp{RIS} under Harsh Environmental Conditions}
Temperature changes and other environmental factors must be taken into consideration in the design of RISs. RISs must be designed in such a way that they are resistant to temperature differences between day and night. Even though space is presumptively accepted as a vacuum, solar winds fill space with plasma and particles ejected from the Sun. Furthermore, coronal mass ejections can impact wireless channel by changing the radiation and temperature in the operating environment of the \ac{RIS}-assisted nodes. For example, charged particles encountered by spacecraft in Van Allen Belts can damage the electronic components of communication systems. Thus, radiation-tolerant \acp{RIS} and all instrumentation making \acp{RIS} operational must be investigated. Dust storms on other planets and temperature differences occurring over a short time duration are also important considerations.
\paragraph{Super Smart Energy Management}
To be able to manage and maintain a high performance \ac{NTN}, smart and efficient energy management should be investigated. Active use of energy harvesting from hybrid sources, including RF and solar energy, is a necessity in resource-limited deployment scenarios.
Distributed low-power sensor networks in planetary exploration can pave the way towards an Internet of Space Things, which will result in a prolific amount of information for deep space. These sensor networks can be powered by harvesting RF-energy with RISs. Also, wearable biomedical devices can harvest energy via \ac{RIS}, as continually monitoring the health status of astronauts is essential. Moreover, the potential benefits of \acp{RIS} in transmitting power from solar power satellites to a planetary surface can be considered.
Managing this extremely complex multi-connectivity network with multi-layer communication by integrating \acp{NTN} with \acp{DSN} can be supported by machine learning algorithms. Conventional radio resource and energy management methods might be inadequate to provide the desired \ac{QoS}. Reinforcement learning both leverages communication performance and provides the minimum power consumption by solving the complex optimization problems with near-optimal results.
\paragraph{Beyond Connectivity: Computing, Caching, and Sensing}
Since satellite and space networks contain cryogenically-cooled low-noise front-ends, the system noise temperature is quite low; in other words, the \ac{SNR} for the received signal is high. By utilizing \acp{RIS}, the required transmit power can be substantially reduced for energy-limited rovers and \ac{HAPS} systems. It should be also noted that positioning RISs closer to a source improves performance~\cite{yildirim2020modeling}. As a result, \ac{RIS}-assisted space stations and \ac{HAPS} systems, which perform computation and caching, can reduce computation delays in the core network. Moreover, supporting edge computation and caching by \ac{DL} can significantly reduce latency.
\section{Concluding Remarks}\label{sec:conclusion}
In this paper we proposed an RIS-assisted DSN integrated with NTNs to extend the duration of active links between Earth and space stations by using relay stations compatible with SWaP requirements. The challenges and opportunities were detailed by interpreting simulation results.
Open issues and research directions for using \acp{RIS} in \ac{DSN}-integrated \acp{NTN} were elaborated. Perhaps in the near future the sound of ``one giant leap for mankind'' will reach us via \ac{RIS}-assisted \acp{DSN} integrated with \acp{NTN}—but with better quality and reliability, of course.
\balance
\bibliographystyle{IEEEtran}
|
\section{Introduction}
Deep Neural Networks (DNNs) have achieved great success in different computer vision problems, e.g., image classification~\cite{krizhevsky2017imagenet}, detection~\cite{ren2016faster}, and semantic segmentation~\cite{long2015fully}. Such success is demanding for large datasets with clean human-annotated labels. However, it is costly and time-consuming to correctly label massive images for building a large-scale dataset like ImageNet~\cite{imagenet_cvpr09}. Some common and less expensive ways to collect large datasets are through online search engines~\cite{schroff2010harvesting} or crowdsourcing~\cite{yu2018learning}, which would, unfortunately, bring noisy labels to the collected datasets. Besides, an in-depth study~\cite{zhang2016understanding} showed that deep learning with noisy labels can lead to severe performance deterioration. Thus, it is crucial to alleviate the negative effects caused by noisy labels for training DNNs.
A typical strategy is to conduct {\em sample selection} (SS) and to train DNNs with selected samples~\cite{han2018co,jiang2018mentornet,song2019selfie,wei2020combating,yu2019does}. Since DNNs tend to learn simple patterns first before fitting noisy samples~\cite{arpit2017closer}, many studies utilize the small-loss trick, where the samples with smaller losses are taken as clean ones. For example, {\em Co-teaching}~\cite{han2018co} leverages two networks to select small-loss samples within each mini-batch for training each other. Later, \citet{yu2019does} pointed out the importance of the disagreement between two networks and proposed {\em Co-teaching+}, which updates the two networks using the data on which the two networks hold different predictions. By contrast, {\em JoCoR}~\cite{wei2020combating} proposes to reduce the diversity between two networks by training them simultaneously with a joint loss calculated from the selected small-loss samples.
Although these methods have achieved satisfactory performance by training with selected small-loss samples, they simply discard other large-loss samples which may contain potentially useful information for the training process.
To make full use of all given samples, a prominent strategy is to
consider selected samples as labeled ``clean'' data and other samples as unlabeled data, and to perform {\em semi-supervised learning} (SSL)~\cite{arazo2020pseudo,berthelot2019mixmatch,laine2016temporal,tarvainen2017mean}. Following this strategy, {\em SELF}~\cite{nguyen2019self} detects clean samples by progressively removing noisy samples whose self-ensemble predictions of the model do not match the given labels in each iteration. With the selected labeled and unlabeled data, the problem becomes an SSL problem, and a {\em Mean-Teacher} model \cite{tarvainen2017mean} can be trained.
Another recent method, {\em DivideMix}~\cite{li2020DivideMix}, leverages Gaussian Mixture Model (GMM)~\cite{permuter2006study} to distinguish clean (labeled) and noisy (unlabeled) data, and then uses a strong SSL backbone called {\em MixMatch}~\cite{berthelot2019mixmatch}. {\em DivideMix} achieves state-of-the-art results across different benchmark datasets.
As shown above, both methods rely on a specific SS strategy and a specific SSL model. The two components play a vitally important role for combating label noise, and stronger components are expected to achieve better performance.
This motivates us to investigate a general algorithmic framework that can leverage various SS strategies and SSL models.
In this paper, we propose {\em SemiNLL}, which is a versatile framework to bridge the gap between SSL and {\em noisy-label learning} (NLL).
Our framework can absorb various SS strategies and SSL backbones, utilizing their power to achieve promising performance. Guided by our framework, one can easily instantiate a specific learning algorithm for NLL, by specifying a commonly used SSL backbone with an SS strategy.
The key contributions of our paper can be summarized as follows:
\begin{itemize}
\item To avoid reinventing the wheel for NLL using SSL algorithms, we propose a versatile framework that can absorb various SS strategies and SSL algorithms. Our framework is advantageous since better performance would be achieved if stronger components (including the ones proposed in the future) are used.
\item To instantiate our framework, we propose {\em DivideMix+} by replacing the epoch-level selection strategy of {\em DivideMix}~\cite{li2020DivideMix} with a mini-batch level one. We also propose {\em GPL}, another instantiation of our framework that leverages a two-component {\em \textbf{G}aussian mixture model}~\cite{li2020DivideMix,permuter2006study} to select labeled (unlabeled) data and uses {\em \textbf{P}seudo-\textbf{L}abeling}~\cite{arazo2020pseudo} as the SSL backbone.
\item We conduct extensive experiments on benchmark-simulated and real-world datasets with noisy labels.
Empirical results show that the stronger SS strategies and SSL backbones we use, the better performance {\em SemiNLL} could achieve. In addition, our instantiations, {\em DivideMix+} and {\em GPL}, outperform other state-of-the-art noisy-label learning methods.
\end{itemize}
\section{Related work}
In this section, we briefly review several related aspects on which our framework builds.
\subsection{Learning with noisy labels}
For NLL, most of the existing methods could be roughly categorized into the following groups:
\noindent\textbf{Sample selection}.\quad
This family of methods regards samples with small loss as ``clean'' and trains the model only on selected clean samples. For example, {\em self-paced MentorNet}~\cite{jiang2018mentornet}, or equivalently {\em self-teaching}, selects small-loss samples and uses them to train the network by itself.
To alleviate the sample-selection bias in {\em self-teaching}, \citet{han2018co} proposed an algorithm called {\em Co-teaching}~\cite{han2018co}, where two networks choose the next batch of data for each other for training based on the samples with smaller loss values. {\em Co-teaching+}~\cite{yu2019does} bridges the {\em disagreement strategy}~\cite{malach2017decoupling} with {\em Co-teaching}~\cite{han2018co} by updating the networks over data where two networks make different predictions. In contrast, \citet{wei2020combating} leveraged the agreement maximization algorithm~\cite{kumar2010co} by designing a joint loss to train two networks on the same mini-batch data and selected small-loss samples to update the parameters of both networks. The mini-batch SS strategy in our framework belongs to this direction. However, instead of ignoring the large-loss unclean samples, we just discard their labels and exploit the associated images in an SSL setup.
\noindent\textbf{Noise transition estimation}.\quad Another line of NLL is to estimate the noise transition matrix for loss correction~\cite{goldberger2016training,hendrycks2018using,menon2015learning,natarajan2013learning,patrini2017making,wang2020training,xiao2015learning}. \citet{patrini2017making} first estimated the noise transition matrix and trained the network with two different loss corrections. \citet{hendrycks2018using} proposed a loss correction technique that utilizes a small portion of trusted samples to estimate the noise transition matrix. However, the limitation of these methods is that they do not perform well on datasets with a large number of~classes.
\noindent\textbf{Other deep learning methods}.\quad Some other interesting and promising directions for NLL include meta-learning~\cite{finn2017model,snell2017prototypical} based, pseudo-label estimation~\cite{lee2013pseudo} based, and robust loss~\cite{feng2020can,ghosh2017robust,ma2020normalized,wang2019symmetric,xu2019l_dmi,zhang2018generalized} based approaches. For meta-learning based approaches, most studies fall into two main categories: training a model that {\em adapts fast to different learning tasks} without overfitting to corrupted labels~\cite{garcia2016noise,li2019learning}, and {\em learning to reweight} loss of each mini-batch to alleviate the adverse effects of corrupted labels~\cite{ren2018learning,shu2019meta,zhang2020distilling}. Pseudo-label estimation based approaches reassign the labels for noisy samples. For example, {\em Joint-Optim}~\cite{tanaka2018joint} corrects labels during training and updates network parameters simultaneously. {\em PENCIL}~\cite{yi2019probabilistic} proposes a probabilistic model, which can update network parameters and reassign labels as label distributions. The family of pseudo-label estimation has a close relationship with semi-supervised learning~\cite{han2019deep,lee2013pseudo,tanaka2018joint,yi2019probabilistic}. Robust loss based approaches focus on designing loss functions that are robust to noisy labels.
\subsection{Semi-supervised learning}
SSL methods leverage unlabeled data to provide additional information for the training model. A line of work is based on the concept of consistency regularization: if a perturbation is given to an unlabeled sample, the model predictions of the same sample should not be too different. Laine and Aila~\cite{laine2016temporal} applied consistency between the output of the current network and the exponential moving average (EMA) of the output from the past epochs. Instead of averaging the model outputs, Tarvainen and Valpola~\cite{tarvainen2017mean} proposed to update the network on every mini-batch using an EMA of model parameter values. \citet{berthelot2019mixmatch} introduced a holistic approach that well combines {\em MixUp}~\cite{zhang2017MixUp}, entropy minimization, and consistency regularization. Another line of SSL is pseudo-labeling, the objective of which is to generate pseudo-labels for unlabeled samples to enhance the learning process. Recently, \citet{arazo2020pseudo} proposed a method to improve previous pseudo-labeling methods~\cite{iscen2019label,lee2013pseudo} by adding {\em MixUp} augmentation~\cite{zhang2017MixUp} and setting a minimum number of labeled samples per mini-batch to reduce accumulated error of wrong pseudo-labels.
\subsection{Combination of SS and SSL}
Some previous studies that combine a specific SS strategy and a specific SSL backbone could be regarded as special cases in our framework. \citet{ding2018semi} used a pre-trained DNN on the noisy dataset to select labeled samples. In the SSL stage, {\em Temporal Ensembling}~\cite{laine2016temporal} was used to handle labeled and unlabeled data. \citet{nguyen2019self} proposed a progressive noise filtering mechanism based on the {\em Mean-Teacher} model~\cite{tarvainen2017mean} and its self-ensemble prediction. \citet{li2020DivideMix} used a Gaussian Mixture Model (GMM) to divide noisy and clean samples based on their training losses and fitted them into a recent SSL algorithm called {\em MixMatch}~\cite{berthelot2019mixmatch}.
Each specific component used in these methods has its own pros and cons. This motivates us to propose a versatile framework that can build on a variety of SS strategies and SSL backbones. In other words, all the methods mentioned in this subsection could be taken as instantiations of our framework.
\begin{figure*}[ht]
\centering
\includegraphics[width=0.99\textwidth]{pic/framework.pdf}
\caption{The schematic of {\em SemiNLL}. First, each mini-batch of data is forwarded to the network to conduct SS, which divides the original data into the labeled/unlabeled sets. Second, labeled/unlabeled samples are used to train the SSL backbone to produce accurate model output.}
\label{fig:framework}
\end{figure*}
\section{The overview of SemiNLL}
\begin{algorithm}[!t]
\caption{SemiNLL}
\label{alg:seminll}
\begin{algorithmic}[1]
\REQUIRE Network $f_{\theta}$, SS strategy $\textsc{select}$, SSL method $\textsc{semi}$, epoch $T_{\text{max}}$, iteration $I_{\text{max}}$;
\FOR{$t$ = 1,2,\ldots,$T_{\text{max}}$}
\STATE \textbf{Shuffle} training set ${\mathcal D}_{train}$;
\FOR{$n = 1,\ldots,I_{\text{max}}$}
\STATE \textbf{Fetch} mini-batch $D_n$ from ${\mathcal D}_{train}$;
\STATE \textbf{Obtain} ${\mathcal{X}}_m, {\mathcal{U}}_m \leftarrow \textsc{select}(D_n, f_{\theta}) $;
\STATE \textbf{Update} $f_{\theta} \leftarrow \textsc{semi}({\mathcal{X}}_m, {\mathcal{U}}_m, f_{\theta}) $;
\ENDFOR
\ENDFOR
\ENSURE $f_{\theta}$
\end{algorithmic}
\end{algorithm}
In this section, we present {\em SemiNLL}, a versatile framework of learning with noisy labels by SSL. The idea behind our framework is that we effectively take advantage of the whole training set by trusting the labels of undoubtedly correct samples and utilizing only the image content of potentially corrupted samples. Previous sample selection methods~\cite{han2018co,jiang2018mentornet,wei2020combating,yu2019does} train the network only with selected clean samples, and they discard all potentially corrupted samples to avoid the harmful memorization of DNNs caused by the noisy labels of these samples. In this way, the feature information contained in the associated images might be discarded without being exploited. Our framework, alternatively, makes use of those corrupted samples by ignoring their labels while keeping the associated image content, transforming the NLL problem into an SSL setup. The mechanism of SSL that leverages labeled data to guide the learning of unlabeled data naturally fits well in training the model with the clean and noisy samples divided by our SS strategy. The schematic of our framework is shown in~Figure~\ref{fig:framework}. We first discuss the advantages of the mini-batch SS strategy in our framework and then introduce several SSL backbones used in our framework.
\subsection{Mini-batch sample selection}\label{subsec:minibatch-separation}
During the SS process, a hazard called confirmation bias \cite{tarvainen2017mean} is worth noting. Since the model is trained using the selected clean (labeled) and noisy (unlabeled) samples, wrongly selected clean samples in this iteration may keep being considered clean ones in the next iteration due to the model overfitting to their labels. Most existing methods~\cite{li2020DivideMix,nguyen2019self} divide the whole training set into the clean/noisy set on an epoch level. In this case, the learned selecting knowledge is incorporated into the SSL phase and will not be updated till the next epoch. Thus, the confirmation bias induced from those wrongly divided samples will accumulate within the whole epoch. To overcome this problem, our mini-batch SS strategy divides each mini-batch of samples into the clean subset ${\mathcal{X}}_m$ and the noisy subset ${\mathcal{U}}_m$ (Line 5 in Algorithm~\ref{alg:seminll}) right before updating the network using SSL backbones. In the next mini-batch, the network can know better to distinguish clean and noisy samples, alleviating the confirmation bias mini-batch by mini-batch.
\subsection{SSL backbones\label{subsec:backbones}}
The mechanism of SSL that uses labeled data to guide the learning of unlabeled data fits well when dealing with clean/noisy data in NLL. The difference lies in an extra procedure, as introduced in Subsection~\ref{subsec:minibatch-separation}, that divides the whole dataset into clean and noisy data. After the SS process, clean samples are considered labeled data and keep their annotated labels. The others are considered noisy samples, and their labels are discarded to be treated as unlabeled ones in SSL backbones. {\em SemiNLL} can build on a variety of SSL algorithms without any modifications to form an end-to-end training scheme for NLL. Concretely, we consider the following representative SSL backbones ranging from weak to strong according to their performance in SSL~tasks:
\begin{enumerate}[(i)]
\setlength\itemsep{0.005em}
\item {\em Temporal Ensembling}~\cite{laine2016temporal}, where the model uses an exponential moving average (EMA) of label predictions from the past epochs as a target for the unsupervised loss. It enforces consistency of predictions by minimizing the difference between the current outputs and the EMA outputs.
\item {\em MixMatch}~\cite{berthelot2019mixmatch}, which is a holistic method that combines {\em MixUp}~\cite{zhang2017MixUp}, entropy minimization, consistency regularization, and other traditional regularization tricks. It guesses low-entropy labels for augmented unlabeled samples and mixes labeled and unlabeled data using {\em MixUp}~\cite{zhang2017MixUp}.
\item {\em Pseudo-Labeling}~\cite{arazo2020pseudo}, which learns from unlabeled data by combining soft {\em pseudo-label} generation~\cite{tanaka2018joint} and {\em MixUp} augmentation~\cite{zhang2017MixUp} to reduce confirmation bias in training.
\end{enumerate}
In the next section, we will instantiate our framework by applying specific SS strategies and SSL backbones to the \textsc{select} and \textsc{semi} placeholders in Algorithm~\ref{alg:seminll}.
\section{The instantiations of SemiNLL}
\subsection{Instantiation 1: DivideMix+\label{subsec:dividemix+}}
\begin{figure*}[ht]
\centering
\begin{subfigure}{.7\textwidth}
\centering
\includegraphics[width=1\linewidth]{pic/dividemix-.pdf}
\setlength{\belowcaptionskip}{0.20cm}
\caption{DivideMix-}
\label{fig:a}
\end{subfigure}
\begin{subfigure}{.7\textwidth}
\centering
\includegraphics[width=1\linewidth]{pic/dividemix.pdf}
\setlength{\belowcaptionskip}{0.20cm
\caption{DivideMix}
\label{fig:b}
\end{subfigure}
\begin{subfigure}{.7\textwidth}
\centering
\includegraphics[width=1\linewidth]{pic/dividemix+.pdf}
\caption{DivideMix+}
\label{fig:c}
\end{subfigure}
\caption{Comparisons between: (a) {\em DivideMix-}, (b) {\em DivideMix}, and (c) {\em DivideMix+}. Squares represent data. Circles represent SS strategy. Octagons represent SSL backbone.}
\end{figure*}
In Algorithm~\ref{alg:seminll}, if we (i) specify the \textsc{select} placeholder as a GMM~\cite{permuter2006study}, (ii) specify the \textsc{semi} placeholder as {\em MixMatch}~\cite{berthelot2019mixmatch} mentioned in Subsection~\ref{subsec:backbones}, and (iii) train two independent networks wherein each network selects clean/noisy samples in the SS phase and predicts labels in the SSL phase for the other network, then our framework is instantiated into a mini-batch version of {\em DivideMix}~\cite{li2020DivideMix}. Specifically, during the SS process, {\em DivideMix}~\cite{li2020DivideMix} fits a two-component GMM to the loss $\ell_i$ of each sample using the Expectation-Maximization technique and obtains the posterior probability of a sample being clean or noisy:
\begin{equation}
p\!\left(k\mid\ell_{i}\right)=\frac{p\!\left(k\right)p\!\left(\ell_{i}\mid k\right)}{p\!\left(\ell_{i}\right)},
\end{equation}
where $k=0\left(1\right)$ denotes the clean (noisy) set.
During the SSL phase, the clean set ${\mathcal{X}}_e$ and the noisy set ${\mathcal{U}}_e$ are fit into an improved {\em MixMatch}~\cite{berthelot2019mixmatch} strategy with label co-refinement and co-guessing. As shown in Figure~\ref{fig:b}, the SS strategy (GMM) of {\em DivideMix}~\cite{li2020DivideMix} is conducted on an epoch level. Since ${\mathcal{X}}_e$ and ${\mathcal{U}}_e$ are updated only once per epoch, the confirmation bias induced from the wrongly divided samples will be accumulated within the whole epoch. However, our mini-batch version, which is called {\em DivideMix+} (Figure~\ref{fig:c}), divides each mini-batch of data into a clean subset ${\mathcal{X}}_m$ and a noisy subset ${\mathcal{U}}_m$, and updates the networks using the SSL backbone right afterwards. In the next mini-batch, the updated networks could better distinguish clean and noisy samples.
\begin{table*}
\centering
\resizebox{0.71\textwidth}{!}{
\setlength{\tabcolsep}{4.25mm}{
\begin{tabular}{c|c|cccc|c}
\toprule
\multirow{3}{*}{Datasets} & \multirow{3}{*}{Method} & \multicolumn{4}{c|}{Symmetric} \\ \cline{3-6}
& & \multicolumn{4}{c|}{Noise ratio} & \multirow{1}{*}{Mean}\\
& & 20\% & 40\% & 60\% & 80\% \\ \midrule
\multirow{7}{*}{MNIST}
& Cross-Entropy & $ 86.16 \pm 0.34 $ & $ 70.39 \pm 0.59 $ & $ 50.35 \pm 0.51 $ & $23.41 \pm 0.96$ & $57.58$ \\
& Coteaching & $ 91.20 \pm 0.03 $ & $90.02 \pm 0.02 $ & $ 83.21 \pm 0.71 $ & $25.33 \pm 0.84 $ & $72.44 $\\
& F-correction & $ 93.93 \pm 0.10 $ & $ 84.30 \pm 0.43 $ & $ 65.06 \pm 0.64 $ & $ 29.81 \pm 0.63 $ & $ 68.27 $ \\
& GCE & $ 94.36 \pm 0.11 $ & $93.61 \pm 0.17 $ & $ 92.46 \pm 0.20 $ & $85.04 \pm 0.66 $ & $91.37 $ \\
& M-correction & $ \boldsymbol{97.25 \pm 0.03} $ & $ \underline{96.63 \pm 0.04}$ & $ 95.07 \pm 0.08 $ & $ 86.19 \pm 0.42$ & $ 93.79 $\\
& DivideMix & $ 96.80 \pm 0.08 $ & $96.53 \pm 0.06 $ & $ \underline{96.47 \pm 0.04} $ & $\underline{95.15 \pm 0.25} $ & $ \underline{96.24} $\\
\cmidrule{2-7}
& \textbf{GPL (ours)}& $ 96.67 \pm 0.09 $ & $96.27 \pm 0.08 $ & $ 95.82 \pm 0.09 $ & $94.81 \pm 0.15 $ & $ 95.89 $\\
& \textbf{DivideMix+ (ours)}& $ \underline{96.83 \pm 0.06} $ & $\boldsymbol{96.79 \pm 0.06} $ & $ \boldsymbol{96.69 \pm 0.03} $ & $\boldsymbol{95.91 \pm 0.10} $ & $ \boldsymbol{96.56} $\\
\midrule
\multirow{7}{*}{\begin{tabular}[c]{@{}c@{}}FASHION\\MNIST\\ \end{tabular}}
& Cross-Entropy & $ 90.83 \pm 0.26 $ & $ 86.44 \pm 0.11 $ & $ 77.27 \pm 0.56 $ & $ 61.84 \pm 1.27 $ & $ 79.10$\\
& Coteaching & $ 89.18 \pm 0.32 $ & $89.13 \pm 0.05 $ & $ 80.08 \pm 0.25 $ & $ 60.36 \pm 2.15$ & $ 79.69$\\
& F-correction & $ \boldsymbol{ 93.37 \pm 0.17} $ & $ 92.27 \pm 0.06 $ & $ 90.32 \pm 0.30 $ & $85.78 \pm 0.06 $ & $90.43$\\
& GCE & $ \underline{93.35 \pm 0.09} $ & $92.58 \pm 0.11 $ & $ 91.30 \pm 0.20 $ & $88.01 \pm 0.22 $ & $ \underline{91.31} $ \\
& M-correction & $ 93.03 \pm 0.15 $ & $\underline{92.74 \pm 0.42} $ & $ \underline{91.61 \pm 0.02} $ & $85.25 \pm 0.23 $ & $90.66 $ \\
& DivideMix & $ 92.98 \pm 0.17 $ & $92.55 \pm 0.13 $ & $ 91.55 \pm 0.31 $ & $\underline{88.55 \pm 0.24} $ & $ 90.66 $ \\
\cmidrule{2-7}
& \textbf{GPL (ours)} & $ 92.94 \pm 0.20 $ & $91.38 \pm 0.54 $ & $ 89.97 \pm 0.16 $ & $87.14 \pm 0.65 $ & $ 90.36 $\\
& \textbf{DivideMix+ (ours)}& $ 93.20 \pm 0.08 $ & $\boldsymbol{92.89 \pm 0.15} $ & $ \boldsymbol{92.15 \pm 0.16} $ & $ \boldsymbol{88.70 \pm 0.17} $ & $ \boldsymbol{91.74}$ \\
\midrule
\multirow{7}{*}{CIFAR-10}
& Cross-Entropy & $ 83.48 \pm 0.17 $ & $ 68.49 \pm 0.40 $ & $ 48.65 \pm 0.06 $ & $27.56 \pm 0.43 $ & $ 57.05$\\
& Coteaching & $ 67.73 \pm 0.71 $ & $62.83 \pm 0.72 $ & $ 48.81 \pm 0.78 $ & $27.56 \pm 2.71 $ & $ 51.73 $ \\
& F-correction & $ 83.27 \pm 0.04 $ & $ 73.67 \pm 0.30 $ & $ 77.64 \pm 0.11 $ & $63.95 \pm 0.32$ & $74.63$ \\
& GCE & $ 89.72 \pm 0.10 $ & $87.75 \pm 0.05 $ & $ 84.11 \pm 0.26 $ & $ 72.84 \pm 0.30$ & $ 83.61$ \\
& M-correction & $ 92.01 \pm 0.40 $ & $90.09 \pm 0.68 $ & $ 85.90 \pm 0.22 $ & $ 70.57 \pm 0.85$ & $ 84.64$\\
& DivideMix & $ \underline{94.82 \pm 0.09} $ & $93.95\pm 0.14 $ & $ 92.28 \pm 0.08 $ & $89.30 \pm 0.17 $ & $ 92.59 $\\
\cmidrule{2-7}
& \textbf{GPL (ours)}& $ 94.45 \pm 0.20$ & $\underline{94.00 \pm 0.22}$ & $ \boldsymbol{93.32 \pm 0.10}$ & $\underline{91.76 \pm 0.23}$ & $ \underline{93.38}$\\
& \textbf{DivideMix+ (ours)}& $\boldsymbol{94.84 \pm 0.12}$ & $\boldsymbol{94.03 \pm 0.20}$ & $ \underline{93.08 \pm 0.19}$ & $\boldsymbol{91.91 \pm 0.07} $ & $\boldsymbol{93.47} $ \\
\midrule
\multirow{6}{*}{CIFAR-100}
& Cross-Entropy & $ 60.93 \pm 0.40 $ & $ 46.24 \pm 0.74 $ & $ 29.00 \pm 0.38 $ & $ 11.42 \pm 0.19$ & $ 36.90 $\\
& F-correction & $ 60.49 \pm 0.29 $ & $ 48.93 \pm 0.21 $ & $ 48.74 \pm 0.41 $ & $22.93 \pm 0.78$ & $45.27$ \\
& GCE & $ 69.20 \pm 0.10$ & $ 65.90 \pm 0.25$ & $57.33 \pm 0.18 $ & $18.19 \pm 1.15 $ & $ 52.66$ \\
& M-correction & $ 67.96 \pm 0.17$ & $64.48 \pm 0.76$ & $ 55.37 \pm 0.72$ & $ 24.21 \pm 1.06$ & $ 53.01 $\\
& DivideMix & $ \underline{73.17 \pm 0.28}$ & $\underline{71.01 \pm 0.16}$ & $ \underline{66.61 \pm 0.18}$ & $ 43.25 \pm 0.82$ & $63.51 $ \\
\cmidrule{2-7}
& \textbf{GPL (ours)}& $ 71.24 \pm 0.24$ & $68.89 \pm 0.07$ & $ 65.80 \pm 0.63 $ & $\boldsymbol{59.96 \pm 0.15}$ & $\underline{66.47} $ \\
& \textbf{DivideMix+ (ours)}& $ \boldsymbol{73.22 \pm 0.21}$ & $ \boldsymbol{71.03 \pm 0.32}$ & $ \boldsymbol{67.52 \pm 0.19}$ & $ \underline{58.07 \pm 0.71}$ & $ \boldsymbol{67.46} $ \\
\bottomrule
\end{tabular
}
}
\caption{Average test accuracy (\%) and standard deviation (5 runs) of all the methods in various datasets under symmetric label noise. The best accuracy is \textbf{bold-faced}. The second-best accuracy is \underline{underlined}.} \label{table:sym}
\end{table*}
\subsection{Instantiation 2: GPL}
Intuitively, the choice of stronger SS strategies and SSL models would achieve better performance based on our framework. Thus, we still choose GMM to distinguish clean and noisy samples due to its flexibility in the sharpness of distribution~\cite{li2020DivideMix}. As for the SSL backbone, we choose the strongest {\em Pseudo-Labeling}~\cite{arazo2020pseudo} introduced in Subsection~\ref{subsec:backbones}. We call this instantiation {\em GPL} ({\em \textbf{G}MM} + {\em \textbf{P}seudo-\textbf{L}abeling}).
Note that we do not train two networks in {\em GPL} as in {\em DivideMix}~\cite{li2020DivideMix} and {\em DivideMix+}. To our understanding, training two networks simultaneously might provide significant improvements in performance. However, this is outside the scope of this paper, since our goal is to demonstrate the versatility of our framework.
\subsection{Self-prediction divider}
Inspired by {\em SELF}~\cite{nguyen2019self}, we introduce the {\em self-prediction divider}, a simple yet effective SS strategy which leverages the information provided by the network's own prediction to distinguish clean and noisy samples. Based on the phenomenon that DNN's predictions tend to be consistent on clean samples and inconsistent on noisy samples in different training iterations, we select the correctly annotated samples via the consistency between the original label set and the model's own predictions. The {\em self-prediction divider} determines potentially clean samples in a mini-batch if the samples' maximal likelihood predictions of the network match their annotated labels. Specifically, the samples are divided into the labeled set only if the model predicts the annotated label to be the correct class with the highest likelihood. The others are considered noisy samples, and their labels will be discarded to be regarded as unlabeled ones in SSL backbones. Compared to previous small-loss SS methods~\cite{han2018co,wei2020combating,yu2019does}, which depend on a known noise ratio to control how many small-loss samples should be selected in each training iteration, {\em self-prediction divider} does not need any additional information to perform SS strategy where the clean subset and the noisy subset are determined by the network itself. Concretely, we instantiate three learning algorithms by combining our {\em \textbf{s}elf-\textbf{p}rediction \textbf{d}ivider (SPD)} with three SSL backbones introduced in Subsection~\ref{subsec:backbones} and denote them as {\em SPD-Temporal Ensembling}, {\em SPD-MixMatch}, and {\em SPD-Pseudo-Labeling}, respectively.
\begin{table*}
\centering
\resizebox{0.71\textwidth}{!}{
\setlength{\tabcolsep}{4.25mm}{
\begin{tabular}{c|c|cccc|c}
\toprule
\multirow{3}{*}{Datasets} & \multirow{3}{*}{Method} & \multicolumn{4}{c|}{Asymmetric} \\ \cline{3-6}
& & \multicolumn{4}{c|}{Noise ratio} & \multirow{1}{*}{Mean}\\
& & 10\% & 20\% & 30\% & 40\% \\ \midrule
\multirow{7}{*}{MNIST}
& Cross-Entropy & $ 95.78 \pm 0.19 $ & $ 91.15 \pm 0.26 $ & $ 86.01 \pm 0.25 $ & $79.92 \pm 0.32 $ & $ 88.22 $\\
& Coteaching & $ 90.32 \pm 0.02 $ & $89.03 \pm 0.02 $ & $79.80 \pm 0.27 $ & $64.94 \pm 0.02 $ & $ 81.02 $\\
& F-correction & $ 96.39 \pm 0.04 $ & $ 94.27 \pm 0.21 $ & $ 89.33 \pm 0.94 $ & $ 81.61 \pm 0.42 $ & $ 90.40 $ \\
& GCE & $ 94.61 \pm 0.13 $ & $94.43 \pm 0.07 $ & $94.00 \pm 0.12 $ & $93.42 \pm 0.12 $ & $ 94.12 $\\
& M-correction & $ \underline{96.74 \pm 0.03} $ & $ \underline{96.70 \pm 0.10} $ & $ \boldsymbol{96.67 \pm 0.07}$ & $94.85 \pm 0.40 $ & $ 96.24 $ \\
& DivideMix & $96.17 \pm 0.06 $ & $96.11 \pm 0.09 $ & $ 95.88 \pm 0.05$ & $95.83 \pm 0.05 $ & $96.00 $\\
\cmidrule{2-7}
& \textbf{GPL (ours)} & $ \boldsymbol{96.76 \pm 0.04} $ & $\boldsymbol{96.71 \pm 0.03} $ & $96.49 \pm 0.08 $ & $ \underline{96.45 \pm 0.04}$ & $ \boldsymbol{96.60}$ \\
& \textbf{DivideMix+ (ours)} & $ 96.67 \pm 0.04 $ & $96.66 \pm 0.07 $ & $ \underline{96.50 \pm 0.04} $ & $\boldsymbol{96.46 \pm 0.04} $ & $ \underline{96.57} $ \\
\midrule
\multirow{7}{*}{\begin{tabular}[c]{@{}c@{}}FASHION\\MNIST\\ \end{tabular}}
& Cross-Entropy & $ \underline{93.88 \pm 0.16} $ & $ 92.20 \pm 0.33 $ & $ 90.41 \pm 0.67 $ & $84.56 \pm 0.41$ & $ 90.26 $ \\
& Coteaching & $ 88.01 \pm 0.03$ & $ 78.88 \pm 0.20$ & $ 70.07 \pm 0.38$ & $ 61.97 \pm 0.21$ & $74.73 $\\
& F-correction & $ \boldsymbol{94.17} \pm 0.12 $ & $ \boldsymbol{93.88 \pm 0.10} $ & $ \boldsymbol{93.50 \pm 0.10} $ & $ \boldsymbol{93.25 \pm 0.16} $ & $ \boldsymbol{93.7} $ \\
& GCE & $ 93.51 \pm 0.17$ & $ \underline{93.24 \pm 0.14}$ & $\underline{92.21 \pm 0.27}$ & $89.53 \pm 0.53$ & $ 92.12 $\\
& M-correction & $92.11 \pm 0.93 $ & $ 91.26 \pm 1.35 $ & $89.79 \pm 1.28 $ & $89.58 \pm 2.20$ & $ 90.69 $\\
& DivideMix & $ 91.83 \pm 0.24 $ & $91.09 \pm 0.08 $ & $89.90 \pm 0.26 $ & $ 87.58 \pm 0.26 $ & $ 90.10 $\\
\cmidrule{2-7}
& \textbf{GPL (ours)}& $ 92.52 \pm 0.22 $ & $92.23 \pm 0.09 $ & $ 92.15 \pm 0.26$ & $\underline{91.64 \pm 0.31} $& $ \underline{92.14} $ \\
& \textbf{DivideMix+ (ours)}& $ 92.56 \pm 0.39 $ & $ 92.25 \pm 0.21 $ & $91.62 \pm 0.08 $ & $ 89.67 \pm 0.44 $& $ 91.53 $ \\
\midrule
\multirow{7}{*}{CIFAR-10}
& Cross-Entropy & $ 90.85 \pm 0.06 $ & $ 87.23 \pm 0.40$ & $ 81.92 \pm 0.32 $ & $ 76.23 \pm 0.45 $ & $ 84.06 $\\
& Coteaching & $ 62.85 \pm 2.20 $ & $ 61.04 \pm 1.31 $ & $ 54.50 \pm 0.39 $ & $ 51.68 \pm 1.66$& $57.52 $ \\
& F-correction & $ 89.79 \pm 0.33 $ & $ 86.79 \pm 0.67 $ & $ 83.34 \pm 0.30 $ & $ 76.81 \pm 1.08 $ & $ 84.18 $ \\
& GCE & $ 90.40 \pm 0.09$ & $89.30 \pm 0.13$ & $86.89 \pm 0.22$ & $ 82.60 \pm 0.17$& $ 87.30 $ \\
& M-correction & $ 92.28 \pm 0.12$ & $92.13 \pm 0.17$ & $91.38 \pm 0.11$ & $ 90.43 \pm 0.23$& $ 91.56 $ \\
& DivideMix & $ 93.61 \pm 0.15$ & $ 92.99 \pm 0.21$ & $ 91.79 \pm 0.36$ & $ 90.57 \pm 0.31$& $92.24 $ \\
\cmidrule{2-7}
& \textbf{GPL (ours)}& $ \boldsymbol{94.32 \pm 0.01}$ & $ \boldsymbol{94.23 \pm 0.07}$ & $\boldsymbol{93.79 \pm 0.06}$ & $ \boldsymbol{93.02 \pm 0.30}$ & $ \boldsymbol{93.84} $\\
& \textbf{DivideMix+ (ours)}& $ \underline{94.27 \pm 0.23}$ & $ \underline{93.92 \pm 0.20}$ & $ \underline{92.82 \pm 0.28}$ & $ \underline{91.91 \pm 0.24}$& $ \underline{93.23} $ \\
\midrule
\multirow{7}{*}{CIFAR-100}
& Cross-Entropy & $ 68.58 \pm 0.34 $ & $ 68.82 \pm 0.22 $ & $ 53.99 \pm 0.50 $ & $ 44.31 \pm 0.23 $& $ 58.93 $ \\
& F-correction & $ 68.87 \pm 0.06 $ & $ 64.11 \pm 0.37 $ & $ 56.45 \pm 0.59 $ & $ 46.44 \pm 0.50 $ & $58.97 $ \\
& GCE & $70.77 \pm 0.14 $ & $ 69.22 \pm 0.15$ & $64.60 \pm 0.25$ & $51.72 \pm 1.17 $& $ 64.08$ \\
& M-correction & $ 69.44 \pm 0.52$ & $67.25 \pm 0.81$ & $63.16 \pm 1.55 $ & $52.90 \pm 1.79 $ & $ 63.19 $\\
& DivideMix & $ \boldsymbol{74.00 \pm 0.29}$ & $ \underline{73.28 \pm 0.42}$ & $\boldsymbol{72.84 \pm 0.36}$ & $ 54.33 \pm 0.69$& $ 68.61 $ \\
\cmidrule{2-7}
& \textbf{GPL (ours)}& $ 71.94 \pm 0.29$ & $71.22 \pm 0.11$ & $ 70.56 \pm 0.23$ & $\boldsymbol{69.84 \pm 0.41} $& $\boldsymbol{ 70.89} $ \\
& \textbf{DivideMix+ (ours)}& $ \underline{73.49 \pm 0.31}$ & $\boldsymbol{73.30 \pm 0.22}$ & $ \underline{72.36 \pm 0.43}$ & $ \underline{55.63 \pm 0.60}$& $ \underline{68.70} $ \\
\bottomrule
\end{tabular}%
}
}
\caption{Average test accuracy (\%) and standard deviation (5 runs) of all the methods in various datasets under asymmetric label noise. The best accuracy is \textbf{bold-faced}. The second-best accuracy is \underline{underlined}.} \label{table:asy}
\end{table*}
\subsection{Effects of the two components}
This section demonstrates the effects of SS strategies and SSL backbones in our framework. To prove that a more robust SS strategy can boost performance for our framework, we propose {\em DivideMix-} (Figure~\ref{fig:a}) by replacing the GMM in {\em DivideMix}~\cite{li2020DivideMix} with our {\em self-prediction divider} on an epoch level. Since {\em self-prediction divider} is supposed to be weaker than GMM, {\em DivideMix-} is expected to achieve lower performance than {\em DivideMix}~\cite{li2020DivideMix}. To prove the effectiveness of the SSL backbone, we remove it after the SS process and only update the model using the supervised loss calculated from the clean samples. We will give detailed discussions in Subsection~\ref{subsec:effectofstrategy} and Subsection~\ref{subsec:effectofbackbone}.
\section{Experiments}
In this section, we first compare two instantiations of our framework, {\em DivideMix+} and {\em GPL}, with other state-of-the-art methods.
We also analyze the effects of SS strategies by comparing {\em DivideMix-}, {\em DivideMix}~\cite{li2020DivideMix}, and {\em DivideMix+}, then analyze the effects of SSL backbones by combining three representative SSL methods with our {\em self-prediction divider}. More information of our experiments can be found in supplementary~materials.
\subsection{Experiment setup}
\noindent \textbf{Datasets}. We thoroughly evaluate our proposed {\em DivideMix+} and {\em GPL} on five datasets, including MNIST~\cite{lecun1998gradient-based}, FASHION-MNIST~\cite{xiao2017fashion}, CIFAR-10, CIFAR-100~\cite{krizhevsky2009learning}, and~Clothing1M~\cite{xiao2015learning}.
MNIST and FASHION-MNIST contain 60K training images and 10K test images of size $28\times28$. CIFAR-10 and CIFAR-100 contain 50K training images and 10K test images of size $32\times32$ with three channels. According to previous studies~\cite{li2020DivideMix,wei2020combating,zhang2018generalized}, we experiment with two types of label noise: symmetric noise and asymmetric noise. Symmetric label noise is produced by changing the original label to all possible labels randomly and uniformly according to the noise ratio. Asymmetric label noise is similar to real-world noise, where labels are flipped to similar classes.
Clothing1M is a large-scale real-world dataset that consists of one million training images from online shopping websites with labels annotated from surrounding texts. The estimated noise ratio is approximately 40\%~\cite{xiao2015learning}.
\begin{table*}
\centering
\resizebox{1.00\textwidth}{!}{
\setlength{\tabcolsep}{3mm}{
\begin{tabular}{c|c|cccc|cccc}
\toprule
\multirow{3}{*}{Datasets} & \multirow{3}{*}{Method} & \multicolumn{4}{c|}{Symmetric} & \multicolumn{4}{c}{Asymmetric} \\ \cline{3-10}
& & \multicolumn{4}{c|}{Noise ratio} & \multicolumn{4}{c}{Noise ratio} \\
& & 20\% & 40\% & 60\% & 80\% & 10\% & 20\% & 30\% & 40\% \\
\midrule
\multirow{3}{*}{CIFAR-10}
& DivideMix- & $ 94.49 \pm 0.02 $ & $93.64 \pm 0.12 $ & $ 91.65 \pm 0.34 $ & $76.61 \pm 1.26$ & $ 93.58 \pm 0.02 $ & $92.87 \pm 0.14 $ & $91.21 \pm 0.21 $ & $ 90.42 \pm 0.23$ \\
& DivideMix & $ 94.82 \pm 0.09 $ & $93.95\pm 0.14 $ & $ 92.28 \pm 0.08 $ & $89.30 \pm 0.17 $ & $ 93.61 \pm 0.15$ & $ 92.99 \pm 0.21$ & $ 91.79 \pm 0.36$ & $ 90.57 \pm 0.31$ \\
& DivideMix+ (ours) & $\boldsymbol{94.84 \pm 0.12}$ & $\boldsymbol{94.03 \pm 0.20}$ & $ \boldsymbol{93.08 \pm 0.19}$ & $\boldsymbol{91.91 \pm 0.07} $ & $ \boldsymbol{94.27 \pm 0.23}$ & $ \boldsymbol{93.92 \pm 0.20}$ & $ \boldsymbol{92.82 \pm 0.28}$ & $ \boldsymbol{91.91 \pm 0.24}$\\
\midrule
\multirow{3}{*}{CIFAR-100}
& DivideMix- & $72.51 \pm 0.32 $ & $ 69.27 \pm 0.46$ & $ 61.13 \pm 0.60 $ & $25.96 \pm 0.78$ & $ 73.62 \pm 0.12 $ & $ 72.32 \pm 0.24$ & $70.64 \pm 0.20 $ & $\boldsymbol{68.04 \pm 1.24} $ \\
& DivideMix & $ 73.17 \pm 0.28$ & $71.01 \pm 0.16$ & $ 66.61 \pm 0.18$ & $ 43.25 \pm 0.82$ & $ \boldsymbol{74.00 \pm 0.29}$ & $ 73.28 \pm 0.42 $ & $\boldsymbol{72.84 \pm 0.36}$ & $ 54.33 \pm 0.69$ \\
& DivideMix+ (ours) & $ \boldsymbol{73.22 \pm 0.21}$ & $ \boldsymbol{71.03 \pm 0.32}$ & $ \boldsymbol{67.52 \pm 0.19}$ & $ \boldsymbol{58.07 \pm 0.71}$ & $ 73.49 \pm 0.31$ & $\boldsymbol{73.30 \pm 0.22}$ & $ 72.36 \pm 0.43$ & $ 55.63 \pm 0.60$ \\
\bottomrule
\end{tabular}
}
}
\caption{Test accuracy (\%) of {\em DivideMix-}, {\em DivideMix}, and {\em DivideMix+}.}\label{table:dividemix-}
\end{table*}
\noindent\textbf{Network Structure and Optimizer.}\quad Following previous works~\cite{arazo2020pseudo,li2020DivideMix,wei2020combating,zhang2018generalized}, we use a 2-layer MLP for MNIST, a ResNet-18~\cite{he2016deep} for FASHION-MNIST, the well-known ``13-CNN'' architecture~\cite{tarvainen2017mean} for CIFAR-10 and CIFAR-100, and an 18-layer PreAct Resnet~\cite{he2016identity} for Clothing1M. To ensure a fair comparison between the instantiations of our framework and other methods, we keep the training settings for MNIST, CIFAR-10, CIFAR-100, and Clothing1M as close as possible to {\em DivideMix}~\cite{li2020DivideMix} and FASHION-MNIST close to {\em GCE}~\cite{zhang2018generalized}.
For FASHION-MNIST, the network is trained using stochastic gradient descent (SGD) with 0.9 momentum and a weight decay of $1\times10^{-4}$ for 120 epochs. For MNIST, CIFAR-10, and CIFAR-100, all networks are trained using SGD with 0.9 momentum and a weight decay of $5\times10^{-4}$ for 300 epochs. For Clothing1M, the momentum is 0.9, and the weight decay is 0.001.
\noindent\textbf{Baselines}.\quad We compare {\em DivideMix+} and {\em GPL} with the following state-of-the-art algorithms and implement all methods by PyTorch on NVIDIA Tesla V100 GPUs.
\begin{enumerate}[(i)]
\setlength\itemsep{0.005em}
\item {\em Coteaching}~\cite{han2018co}, which trains two networks and cross-updates the parameters of peer networks.
\item {\em GCE}~\cite{zhang2018generalized}, which uses a theoretically grounded and easy-to-use loss function, the $\mathcal{L}_q$ loss, for NLL.
\item {\em F-correction}~\cite{patrini2017making}, which corrects the prediction by the label transition matrix. As suggested by the authors, we first train a standard network using the cross-entropy loss to estimate the transition matrix.
\item {\em M-correction}~\cite{arazo2019unsupervised}, which models clean and noisy samples by fitting a two-component BMM and applies {\em MixUp} data augmentation~\cite{zhang2017MixUp}.
\item {\em DivideMix}~\cite{li2020DivideMix}, which divides clean and noisy samples by using a GMM on an epoch level and leverages {\em MixMatch}~\cite{berthelot2019mixmatch} as the SSL backbone.
\end{enumerate}
\subsection{Performance Comparison}
The results of all the methods under symmetric and asymmetric noise types on MNIST, FASHION-MNIST, CIFAR-10, and CIFAR-100 are shown in Table~\ref{table:sym} and Table~\ref{table:asy}. The results on Clothing1M are shown in Table~\ref{table:clothing}.
\noindent\textbf{Results on MNIST.}\quad {\em DivideMix+} surpasses {\em DivideMix} across symmetric and asymmetric noise at all noise ratios, showing the effectiveness of the mini-batch SS strategy in our framework.
{\em M-correction} performs well under low noise ratios. However, in the hardest symmetric 80\% case, {\em DivideMix+} achieves best test accuracy.
\noindent\textbf{Results on FASHION-MNIST.}\quad FASHION-MNIST is quite similar to MNIST but more complicated. {\em DivideMix+} still outperforms {\em DivideMix} on symmetric and asymmetric noise at all noise ratios.
In the harder asymmetric 40\% noise, {\em DivideMix+} and {\em GPL} outperform the other methods by a large margin.
\noindent\textbf{Results on CIFAR-10.}\quad {\em DivideMix+} constantly outperforms {\em DivideMix}, especially in the cases with higher noise ratios.
We believe the reason is that the mini-batch SS strategy used in our framework can better mitigate the confirmation bias induced from wrongly divided samples in more challenging scenarios.
Overall, {\em GPL} and {\em DivideMix+} surpass the other methods over a large margin, with the latter performing extremely well on asymmetric noise.
\noindent\textbf{Results on CIFAR-100.}\quad
In most cases, {\em DivideMix+} and {\em DivideMix} achieve higher test accuracy than the other approaches, with {\em DivideMix+} performing better. Specifically, {\em DivideMix+} surpasses {\em DivideMix} by 14.82\% in the hardest symmetric 80\% case. An interesting phenomenon is that all the approaches suffer from performance deterioration in the asymmetric 40\% cases except {\em GPL}, which significantly outperforms the second-best algorithm over +14\%.
\begin{table*}[ht]
\centering
\resizebox{0.86\textwidth}{!}{
\setlength{\tabcolsep}{4mm}{
\begin{tabular}{l |c c c|c c c}
\toprule
Dataset & \multicolumn{3}{c|}{CIFAR-10}&\multicolumn{3}{c}{CIFAR-100}\\
\midrule
Method/Noise ratio& 20\% & 50\%& 80\% &20\% & 50\%& 80\% \\
\midrule
SPD-Cross-Entropy & $83.13 \pm 0.16$ & $79.74 \pm 0.10$& $49.14 \pm 0.15$ & $45.07 \pm 0.55$ & $35.02 \pm 0.57$& $10.22 \pm 0.10$\\
SPD-Temporal Ensembling & $83.15 \pm 0.06$ & $80.16 \pm 0.36$ & $49.10 \pm 0.13$ &$46.16 \pm 0.12$ & $ 39.91 \pm 0.60 $ & $ 12.37 \pm 0.67 $\\
SPD-MixMatch & $ 93.53 \pm 0.52$ & $90.22 \pm 0.18$ & $88.77 \pm 0.20$ & $72.89 \pm 0.30$ &$68.57 \pm 0.20$ & $33.92 \pm 0.20$\\
SPD-Pseudo-Labeling & $ \boldsymbol{94.52 \pm 0.06}$ & $\boldsymbol{93.24 \pm 0.36}$ &$ \boldsymbol{90.27 \pm 0.34}$ & $\boldsymbol{73.84 \pm 0.48}$ & $ \boldsymbol{68.61 \pm 0.40}$ &$ \boldsymbol{55.37 \pm 0.34}$ \\
\bottomrule
\end{tabular}
}
}
\setlength{\belowcaptionskip}{-0.43cm}
\caption{Test accuracy (\%) of the baseline and three SSL backbones integrated into our proposed framework.}\label{table:sslbackbones}
\end{table*}
\noindent\textbf{Results on Clothing1M.}\quad
To show the robustness of our framework under real-world noisy labels, we demonstrate the effectiveness of {\em DivideMix+} and {\em GPL} on Clothing1M. As shown in Table ~\ref{table:clothing}, the performance of {\em DivideMix+} is better than that of {\em DivideMix} and other methods.
\subsection{The effects of SS strategies\label{subsec:effectofstrategy}}
To study how SS strategies can affect the performance of our framework, we propose {\em DivideMix-} by replacing the GMM component in {\em DivideMix} with our {\em self-prediction divider} yet maintaining the epoch-level SS strategy for a fair comparison. Due to constraints of space, we only provide the mean value of the results in Table \ref{table:dividemix-}, which can show the overall tendency. Results with mean and standard deviation can be found in supplementary materials. In CIFAR-10, the difference between {\em DivideMix-} and {\em DivideMix} is not obvious in the lower noise ratios. However, in the most difficult symmetric 80\% case, the test accuracy of {\em DivideMix} is +12.69\% higher than {\em DivideMix-}. The difference is even greater in CIFAR-100, showing that GMM is better able to distinguish clean and noisy labels in most cases. An impressive phenomenon to note is that {\em DivideMix-} excels in the asymmetric 40\% case in CIFAR-100, which means the {\em self-prediction divider} performs better in nosier asymmetric cases than GMM.
The reason is explained in the original paper of {\em DivideMix}~\cite{li2020DivideMix}, that GMM cannot effectively distinguish clean and noisy samples under asymmetric noise with high noise ratio in datasets with a large number of classes. At the same time, the fact that {\em DivideMix+} constantly outperforms {\em DivideMix} in most cases shows that the mini-batch SS strategy in our framework is better than the epoch-level one in {\em DivideMix}.
\subsection{The effects of SSL backbones\label{subsec:effectofbackbone}}
We evaluate the effects of SSL backbones in our framework by combining the {\em self-prediction divider (SPD)} with three different SSL methods and a baseline which only updates the model using the cross-entropy loss calculated from clean samples. We denote them as {\em SPD-Temporal Ensembling}, {\em SPD-MixMatch}, {\em SPD-Pseudo-Labeling}, and {\em SPD-Cross-Entropy}, respectively. For a fair comparison, we use the ``13-CNN'' architecture~\cite{tarvainen2017mean} for all methods across different datasets. We keep most hyperparameters introduced by the SSL methods close to their original papers~\cite{arazo2020pseudo,berthelot2019mixmatch,laine2016temporal}, since they can be easily integrated into our framework without massive adjustments.
In Table~\ref{table:sslbackbones}, we list these four algorithms in the left column from weak to strong according to their performance in their original papers. The test accuracies demonstrate their corresponding performance for NLL based on our framework. {\em SPD-MixMatch} and {\em SPD-Pseudo-Labeling} outperform {\em SPD-Temporal Ensembling} by a large domain in both CIFAR-10 and CIFAR-100, especially under 80\% noise ratio (over 40\% in CIFAR-10). This phenomenon is reasonable because {\em Temporal Ensembling}~\cite{laine2016temporal} only uses consistency regularization for unsupervised loss, while {\em MixMatch}~\cite{berthelot2019mixmatch} and {\em Pseudo-Labeling}~\cite{arazo2020pseudo} also leverage entropy regularization as well as {\em MixUp} data augmentation~\cite{zhang2017MixUp}. Moreover, {\em SPD-Pseudo-Labeling} achieves remarkable test accuracy under 80\% noise ratio in CIFAR-100, which is +21.44\% higher than {\em SPD-MixMatch} and +42.66\% higher than {\em SPD-Temporal Ensembling}. We assume that this is due to the additional loss used in {\em SPD-Pseudo-Labeling} that prevents the model from assigning all labels to a single class at the early training stage.
From the results of {\em SPD-Cross-Entropy}, we can see that after the removal of the SSL backbone, the test accuracy drops dramatically compared to {\em SPD-MixMatch} and {\em SPD-Pseudo-Labeling}, especially in high noise ratios and datasets with more classes (e.g., CIFAR-100). This is possibly due to the substantial amount of data that has been removed by the {\em self-prediction divider}, leaving very few samples per class. Thus, instead of discarding noisy samples, transferring them to unlabeled ones in SSL backbones is an effective way to combat noisy labels.
\begin{table}[!t]
\centering
\resizebox{0.40\textwidth}{!}{
\setlength{\tabcolsep}{5mm}{
\begin{tabular}{l|c}
\toprule
Methods & Test Accuracy \\
\midrule
Cross-Entropy & 69.21 \\
F-correction~\cite{patrini2017making} & 69.84 \\
M-correction~\cite{arazo2019unsupervised} & 71.00 \\
Joint-Optim~\cite{tanaka2018joint} & 72.16 \\
Dividemix~\cite{li2020DivideMix} & 73.91 \\
\midrule
GPL(ours) & 73.19 \\
Dividemix+(ours) & $\boldsymbol{74.14}$ \\
\bottomrule
\end{tabular}
}
}
\setlength{\belowcaptionskip}{-0.6cm}
\caption{Test accuracy (\%) on Clothing1M.}\label{table:clothing}
\end{table}
\section{Conclusion}
This paper proposes a versatile framework called {\em SemiNLL} for NLL. This framework consists of two main parts: the mini-batch SS strategy and the SSL backbone. We conduct extensive experiments on benchmark-simulated and real-world datasets to demonstrate that {\em SemiNLL} can absorb a variety of SS strategies and SSL backbones, leveraging their power to achieve state-of-the-art performance in different noise scenarios. Moreover, we throughly analyze the effects of the two components in our framework.
\bibliographystyle{plainnat}
\section{Introduction}
Deep Neural Networks (DNNs) have achieved great success in different computer vision problems, e.g., image classification~\cite{krizhevsky2017imagenet}, detection~\cite{ren2016faster}, and semantic segmentation~\cite{long2015fully}. Such success is demanding for large datasets with clean human-annotated labels. However, it is costly and time-consuming to correctly label massive images for building a large-scale dataset like ImageNet~\cite{imagenet_cvpr09}. Some common and less expensive ways to collect large datasets are through online search engines~\cite{schroff2010harvesting} or crowdsourcing~\cite{yu2018learning}, which would, unfortunately, bring noisy labels to the collected datasets. Besides, an in-depth study~\cite{zhang2016understanding} showed that deep learning with noisy labels can lead to severe performance deterioration. Thus, it is crucial to alleviate the negative effects caused by noisy labels for training DNNs.
A typical strategy is to conduct {\em sample selection} (SS) and to train DNNs with selected samples~\cite{han2018co,jiang2018mentornet,song2019selfie,wei2020combating,yu2019does}. Since DNNs tend to learn simple patterns first before fitting noisy samples~\cite{arpit2017closer}, many studies utilize the small-loss trick, where the samples with smaller losses are taken as clean ones. For example, {\em Co-teaching}~\cite{han2018co} leverages two networks to select small-loss samples within each mini-batch for training each other. Later, Yu \etal~\cite{yu2019does} pointed out the importance of the disagreement between two networks and proposed {\em Co-teaching+}, which updates the two networks using the data on which the two networks hold different predictions. By contrast, {\em JoCoR}~\cite{wei2020combating} proposes to reduce the diversity between two networks by training them simultaneously with a joint loss calculated from the selected small-loss samples.
Although these methods have achieved satisfactory performance by training with selected small-loss samples, they simply discard other large-loss samples which may contain potentially useful information for the training process.
To make full use of all given samples, a prominent strategy is to
consider selected samples as labeled ``clean'' data and other samples as unlabeled data, and to perform {\em semi-supervised learning} (SSL)~\cite{arazo2020pseudo,berthelot2019mixmatch,laine2016temporal,tarvainen2017mean}. Following this strategy, {\em SELF}~\cite{nguyen2019self} detects clean samples by progressively removing noisy samples whose self-ensemble predictions of the model do not match the given labels in each iteration. With the selected labeled and unlabeled data, the problem becomes an SSL problem, and a {\em Mean-Teacher} model \cite{tarvainen2017mean} can be trained.
Another recent method, {\em DivideMix}~\cite{li2020DivideMix}, leverages Gaussian Mixture Model (GMM)~\cite{permuter2006study} to distinguish clean (labeled) and noisy (unlabeled) data, and then uses a strong SSL backbone called {\em MixMatch}~\cite{berthelot2019mixmatch}. {\em DivideMix} achieves state-of-the-art results across different benchmark datasets.
As shown above, both methods rely on a specific SS strategy and a specific SSL model. The two components play a vitally important role for combating label noise, and stronger components are expected to achieve better performance.
This motivates us to investigate a general algorithmic framework that can leverage various SS strategies and SSL models.
In this paper, we propose {\em SemiNLL}, which is a versatile framework to bridge the gap between SSL and {\em noisy-label learning} (NLL).
Our framework can absorb various SS strategies and SSL backbones, utilizing their power to achieve promising performance. Guided by our framework, one can easily instantiate a specific learning algorithm for NLL, by specifying a commonly used SSL backbone with an SS strategy.
The key contributions of our paper can be summarized as follows:
\begin{itemize}
\item To avoid reinventing the wheel for NLL using SSL algorithms, we propose a versatile framework that can absorb various SS strategies and SSL algorithms. Our framework is advantageous since better performance would be achieved if stronger components (including the ones proposed in the future) are used.
\item To instantiate our framework, we propose {\em DivideMix+} by replacing the epoch-level selection strategy of {\em DivideMix}~\cite{li2020DivideMix} with a mini-batch level one. We also propose {\em GPL}, another instantiation of our framework that leverages a two-component {\em \textbf{G}aussian mixture model}~\cite{li2020DivideMix,permuter2006study} to select labeled (unlabeled) data and uses {\em \textbf{P}seudo-\textbf{L}abeling}~\cite{arazo2020pseudo} as the SSL backbone.
\item We conduct extensive experiments on benchmark-simulated and real-world datasets with noisy labels.
Empirical results show that the stronger SS strategies and SSL backbones we use, the better performance {\em SemiNLL} could achieve. In addition, our instantiations, {\em DivideMix+} and {\em GPL}, outperform other state-of-the-art noisy-label learning methods.
\end{itemize}
\section{Related work}
In this section, we briefly review several related aspects on which our framework builds.
\subsection{Learning with noisy labels}
For NLL, most of the existing methods could be roughly categorized into the following groups:
\noindent\textbf{Sample selection}.\quad
This family of methods regards samples with small loss as ``clean'' and trains the model only on selected clean samples. For example, {\em self-paced MentorNet}~\cite{jiang2018mentornet}, or equivalently {\em self-teaching}, selects small-loss samples and uses them to train the network by itself.
To alleviate the sample-selection bias in {\em self-teaching}, Han \etal~\cite{han2018co} proposed an algorithm called {\em Co-teaching}~\cite{han2018co}, where two networks choose the next batch of data for each other for training based on the samples with smaller loss values. {\em Co-teaching+}~\cite{yu2019does} bridges the {\em disagreement strategy}~\cite{malach2017decoupling} with {\em Co-teaching}~\cite{han2018co} by updating the networks over data where two networks make different predictions. In contrast, Wei \etal~\cite{wei2020combating} leveraged the agreement maximization algorithm~\cite{kumar2010co} by designing a joint loss to train two networks on the same mini-batch data and selected small-loss samples to update the parameters of both networks. The mini-batch SS strategy in our framework belongs to this direction. However, instead of ignoring the large-loss unclean samples, we just discard their labels and exploit the associated images in an SSL setup.
\noindent\textbf{Noise transition estimation}.\quad Another line of NLL is to estimate the noise transition matrix for loss correction~\cite{goldberger2016training,hendrycks2018using,menon2015learning,natarajan2013learning,patrini2017making,wang2020training,xiao2015learning}. Patrini \etal~\cite{patrini2017making} first estimated the noise transition matrix and trained the network with two different loss corrections. Hendrycks \etal~\cite{hendrycks2018using} proposed a loss correction technique that utilizes a small portion of trusted samples to estimate the noise transition matrix. However, the limitation of these methods is that they do not perform well on datasets with a large number of~classes.
\noindent\textbf{Other deep learning methods}.\quad Some other interesting and promising directions for NLL include meta-learning~\cite{finn2017model,snell2017prototypical} based, pseudo-label estimation~\cite{lee2013pseudo} based, and robust loss~\cite{feng2020can,ghosh2017robust,ma2020normalized,wang2019symmetric,xu2019l_dmi,zhang2018generalized} based approaches. For meta-learning based approaches, most studies fall into two main categories: training a model that {\em adapts fast to different learning tasks} without overfitting to corrupted labels~\cite{garcia2016noise,li2019learning}, and {\em learning to reweight} loss of each mini-batch to alleviate the adverse effects of corrupted labels~\cite{ren2018learning,shu2019meta,zhang2020distilling}. Pseudo-label estimation based approaches reassign the labels for noisy samples. For example, {\em Joint-Optim}~\cite{tanaka2018joint} corrects labels during training and updates network parameters simultaneously. {\em PENCIL}~\cite{yi2019probabilistic} proposes a probabilistic model, which can update network parameters and reassign labels as label distributions. The family of pseudo-label estimation has a close relationship with semi-supervised learning~\cite{han2019deep,lee2013pseudo,tanaka2018joint,yi2019probabilistic}. Robust loss based approaches focus on designing loss functions that are robust to noisy labels.
\begin{figure*}[ht]
\centering
\includegraphics[width=0.99\textwidth]{pic/framework.pdf}
\caption{The schematic of {\em SemiNLL}. First, each mini-batch of data is forwarded to the network to conduct SS, which divides the original data into the labeled/unlabeled sets. Second, labeled/unlabeled samples are used to train the SSL backbone to produce accurate model output.}
\label{fig:framework}
\end{figure*}
\subsection{Semi-supervised learning}
SSL methods leverage unlabeled data to provide additional information for the training model. A line of work is based on the concept of consistency regularization: if a perturbation is given to an unlabeled sample, the model predictions of the same sample should not be too different. Laine and Aila~\cite{laine2016temporal} applied consistency between the output of the current network and the exponential moving average (EMA) of the output from the past epochs. Instead of averaging the model outputs, Tarvainen and Valpola~\cite{tarvainen2017mean} proposed to update the network on every mini-batch using an EMA of model parameter values. Berthelot \etal~\cite{berthelot2019mixmatch} introduced a holistic approach that well combines {\em MixUp}~\cite{zhang2017MixUp}, entropy minimization, and consistency regularization. Another line of SSL is pseudo-labeling, the objective of which is to generate pseudo-labels for unlabeled samples to enhance the learning process. Recently, Arazo \etal~\cite{arazo2020pseudo} proposed a method to improve previous pseudo-labeling methods~\cite{iscen2019label,lee2013pseudo} by adding {\em MixUp} augmentation~\cite{zhang2017MixUp} and setting a minimum number of labeled samples per mini-batch to reduce accumulated error of wrong pseudo-labels.
\subsection{Combination of SS and SSL}
Some previous studies that combine a specific SS strategy and a specific SSL backbone could be regarded as special cases in our framework. Ding \etal~\cite{ding2018semi} used a pre-trained DNN on the noisy dataset to select labeled samples. In the SSL stage, {\em Temporal Ensembling}~\cite{laine2016temporal} was used to handle labeled and unlabeled data. Nguyen \etal~\cite{nguyen2019self} proposed a progressive noise filtering mechanism based on the {\em Mean-Teacher} model~\cite{tarvainen2017mean} and its self-ensemble prediction. Li \etal~\cite{li2020DivideMix} used a Gaussian Mixture Model (GMM) to divide noisy and clean samples based on their training losses and fitted them into a recent SSL algorithm called {\em MixMatch}~\cite{berthelot2019mixmatch}.
Each specific component used in these methods has its own pros and cons. This motivates us to propose a versatile framework that can build on a variety of SS strategies and SSL backbones. In other words, all the methods mentioned in this subsection could be taken as instantiations of our framework.
\section{The overview of SemiNLL}
\begin{algorithm}[!t]
\caption{SemiNLL}
\label{alg:seminll}
\begin{algorithmic}[1]
\REQUIRE Network $f_{\theta}$, SS strategy $\textsc{select}$, SSL method $\textsc{semi}$, epoch $T_{\text{max}}$, iteration $I_{\text{max}}$;
\FOR{$t$ = 1,2,\ldots,$T_{\text{max}}$}
\STATE \textbf{Shuffle} training set ${\mathcal D}_{train}$;
\FOR{$n = 1,\ldots,I_{\text{max}}$}
\STATE \textbf{Fetch} mini-batch $D_n$ from ${\mathcal D}_{train}$;
\STATE \textbf{Obtain} ${\mathcal{X}}_m, {\mathcal{U}}_m \leftarrow \textsc{select}(D_n, f_{\theta}) $;
\STATE \textbf{Update} $f_{\theta} \leftarrow \textsc{semi}({\mathcal{X}}_m, {\mathcal{U}}_m, f_{\theta}) $;
\ENDFOR
\ENDFOR
\ENSURE $f_{\theta}$
\end{algorithmic}
\end{algorithm}
In this section, we present {\em SemiNLL}, a versatile framework of learning with noisy labels by SSL. The idea behind our framework is that we effectively take advantage of the whole training set by trusting the labels of undoubtedly correct samples and utilizing only the image content of potentially corrupted samples. Previous sample selection methods~\cite{han2018co,jiang2018mentornet,wei2020combating,yu2019does} train the network only with selected clean samples, and they discard all potentially corrupted samples to avoid the harmful memorization of DNNs caused by the noisy labels of these samples. In this way, the feature information contained in the associated images might be discarded without being exploited. Our framework, alternatively, makes use of those corrupted samples by ignoring their labels while keeping the associated image content, transforming the NLL problem into an SSL setup. The mechanism of SSL that leverages labeled data to guide the learning of unlabeled data naturally fits well in training the model with the clean and noisy samples divided by our SS strategy. The schematic of our framework is shown in~Figure~\ref{fig:framework}. We first discuss the advantages of the mini-batch SS strategy in our framework and then introduce several SSL backbones used in our framework.
\subsection{Mini-batch sample selection}\label{subsec:minibatch-separation}
During the SS process, a hazard called confirmation bias \cite{tarvainen2017mean} is worth noting. Since the model is trained using the selected clean (labeled) and noisy (unlabeled) samples, wrongly selected clean samples in this iteration may keep being considered clean ones in the next iteration due to the model overfitting to their labels. Most existing methods~\cite{li2020DivideMix,nguyen2019self} divide the whole training set into the clean/noisy set on an epoch level. In this case, the learned selecting knowledge is incorporated into the SSL phase and will not be updated till the next epoch. Thus, the confirmation bias induced from those wrongly divided samples will accumulate within the whole epoch. To overcome this problem, our mini-batch SS strategy divides each mini-batch of samples into the clean subset ${\mathcal{X}}_m$ and the noisy subset ${\mathcal{U}}_m$ (Line 5 in Algorithm~\ref{alg:seminll}) right before updating the network using SSL backbones. In the next mini-batch, the network can know better to distinguish clean and noisy samples, alleviating the confirmation bias mini-batch by mini-batch.
\subsection{SSL backbones\label{subsec:backbones}}
The mechanism of SSL that uses labeled data to guide the learning of unlabeled data fits well when dealing with clean/noisy data in NLL. The difference lies in an extra procedure, as introduced in Subsection~\ref{subsec:minibatch-separation}, that divides the whole dataset into clean and noisy data. After the SS process, clean samples are considered labeled data and keep their annotated labels. The others are considered noisy samples, and their labels are discarded to be treated as unlabeled ones in SSL backbones. {\em SemiNLL} can build on a variety of SSL algorithms without any modifications to form an end-to-end training scheme for NLL. Concretely, we consider the following representative SSL backbones ranging from weak to strong according to their performance in SSL~tasks:
\begin{enumerate}[(i)]
\setlength\itemsep{0.005em}
\item {\em Temporal Ensembling}~\cite{laine2016temporal}, where the model uses an exponential moving average (EMA) of label predictions from the past epochs as a target for the unsupervised loss. It enforces consistency of predictions by minimizing the difference between the current outputs and the EMA outputs.
\item {\em MixMatch}~\cite{berthelot2019mixmatch}, which is a holistic method that combines {\em MixUp}~\cite{zhang2017MixUp}, entropy minimization, consistency regularization, and other traditional regularization tricks. It guesses low-entropy labels for augmented unlabeled samples and mixes labeled and unlabeled data using {\em MixUp}~\cite{zhang2017MixUp}.
\item {\em Pseudo-Labeling}~\cite{arazo2020pseudo}, which learns from unlabeled data by combining soft {\em pseudo-label} generation~\cite{tanaka2018joint} and {\em MixUp} augmentation~\cite{zhang2017MixUp} to reduce confirmation bias in training.
\end{enumerate}
In the next section, we will instantiate our framework by applying specific SS strategies and SSL backbones to the \textsc{select} and \textsc{semi} placeholders in Algorithm~\ref{alg:seminll}.
\begin{table*}
\centering
\resizebox{1.00\textwidth}{!}{
\setlength{\tabcolsep}{4.25mm}{
\begin{tabular}{c|c|cccc|c}
\toprule
\multirow{3}{*}{Datasets} & \multirow{3}{*}{Method} & \multicolumn{4}{c|}{Symmetric} \\ \cline{3-6}
& & \multicolumn{4}{c|}{Noise ratio} & \multirow{1}{*}{Mean}\\
& & 20\% & 40\% & 60\% & 80\% \\ \midrule
\multirow{7}{*}{MNIST}
& Cross-Entropy & $ 86.16 \pm 0.34 $ & $ 70.39 \pm 0.59 $ & $ 50.35 \pm 0.51 $ & $23.41 \pm 0.96$ & $57.58$ \\
& Coteaching & $ 91.20 \pm 0.03 $ & $90.02 \pm 0.02 $ & $ 83.21 \pm 0.71 $ & $25.33 \pm 0.84 $ & $72.44 $\\
& GCE & $ 94.36 \pm 0.11 $ & $93.61 \pm 0.17 $ & $ 92.46 \pm 0.20 $ & $85.04 \pm 0.66 $ & $91.37 $ \\
& M-correction & $ \boldsymbol{97.25 \pm 0.03} $ & $ \underline{96.63 \pm 0.04}$ & $ 95.07 \pm 0.08 $ & $ 86.19 \pm 0.42$ & $ 93.79 $\\
& DivideMix & $ 96.80 \pm 0.08 $ & $96.53 \pm 0.06 $ & $ \underline{96.47 \pm 0.04} $ & $\underline{95.15 \pm 0.25} $ & $ \underline{96.24} $\\
\cmidrule{2-7}
& \textbf{GPL (ours)}& $ 96.67 \pm 0.09 $ & $96.27 \pm 0.08 $ & $ 95.82 \pm 0.09 $ & $94.81 \pm 0.15 $ & $ 95.89 $\\
& \textbf{DivideMix+ (ours)}& $ \underline{96.83 \pm 0.06} $ & $\boldsymbol{96.79 \pm 0.06} $ & $ \boldsymbol{96.69 \pm 0.03} $ & $\boldsymbol{95.91 \pm 0.10} $ & $ \boldsymbol{96.56} $\\
\midrule
\multirow{7}{*}{\begin{tabular}[c]{@{}c@{}}FASHION\\MNIST\\ \end{tabular}}
& Cross-Entropy & $ 90.83 \pm 0.26 $ & $ 86.44 \pm 0.11 $ & $ 77.27 \pm 0.56 $ & $ 61.84 \pm 1.27 $ & $ 79.10$\\
& Coteaching & $ 89.18 \pm 0.32 $ & $89.13 \pm 0.05 $ & $ 80.08 \pm 0.25 $ & $ 60.36 \pm 2.15$ & $ 79.69$\\
& GCE & $ \boldsymbol{93.35 \pm 0.09} $ & $92.58 \pm 0.11 $ & $ 91.30 \pm 0.20 $ & $88.01 \pm 0.22 $ & $ \underline{91.31} $ \\
& M-correction & $ 93.03 \pm 0.15 $ & $\underline{92.74 \pm 0.42} $ & $ \underline{91.61 \pm 0.02} $ & $85.25 \pm 0.23 $ & $90.66 $ \\
& DivideMix & $ 92.98 \pm 0.17 $ & $92.55 \pm 0.13 $ & $ 91.55 \pm 0.31 $ & $\underline{88.55 \pm 0.24} $ & $ 90.66 $ \\
\cmidrule{2-7}
& \textbf{GPL (ours)} & $ 92.94 \pm 0.20 $ & $91.38 \pm 0.54 $ & $ 89.97 \pm 0.16 $ & $87.14 \pm 0.65 $ & $ 90.36 $\\
& \textbf{DivideMix+ (ours)}& $ \underline{93.20 \pm 0.08} $ & $\boldsymbol{92.89 \pm 0.15} $ & $ \boldsymbol{92.15 \pm 0.16} $ & $ \boldsymbol{88.70 \pm 0.17} $ & $ \boldsymbol{91.74}$ \\
\midrule
\multirow{7}{*}{CIFAR-10}
& Cross-Entropy & $ 83.48 \pm 0.17 $ & $ 68.49 \pm 0.40 $ & $ 48.65 \pm 0.06 $ & $27.56 \pm 0.43 $ & $ 57.05$\\
& Coteaching & $ 67.73 \pm 0.71 $ & $62.83 \pm 0.72 $ & $ 48.81 \pm 0.78 $ & $27.56 \pm 2.71 $ & $ 51.73 $ \\
& GCE & $ 89.72 \pm 0.10 $ & $87.75 \pm 0.05 $ & $ 84.11 \pm 0.26 $ & $ 72.84 \pm 0.30$ & $ 83.61$ \\
& M-correction & $ 92.01 \pm 0.40 $ & $90.09 \pm 0.68 $ & $ 85.90 \pm 0.22 $ & $ 70.57 \pm 0.85$ & $ 84.64$\\
& DivideMix & $ \underline{94.82 \pm 0.09} $ & $93.95\pm 0.14 $ & $ 92.28 \pm 0.08 $ & $89.30 \pm 0.17 $ & $ 92.59 $\\
\cmidrule{2-7}
& \textbf{GPL (ours)}& $ 94.45 \pm 0.20$ & $\underline{94.00 \pm 0.22}$ & $ \boldsymbol{93.32 \pm 0.10}$ & $\underline{91.76 \pm 0.23}$ & $ \underline{93.38}$\\
& \textbf{DivideMix+ (ours)}& $\boldsymbol{94.84 \pm 0.12}$ & $\boldsymbol{94.03 \pm 0.20}$ & $ \underline{93.08 \pm 0.19}$ & $\boldsymbol{91.91 \pm 0.07} $ & $\boldsymbol{93.47} $ \\
\midrule
\multirow{6}{*}{CIFAR-100}
& Cross-Entropy & $ 60.93 \pm 0.40 $ & $ 46.24 \pm 0.74 $ & $ 29.00 \pm 0.38 $ & $ 11.42 \pm 0.19$ & $ 36.90 $\\
& GCE & $ 69.20 \pm 0.10$ & $ 65.90 \pm 0.25$ & $57.33 \pm 0.18 $ & $18.19 \pm 1.15 $ & $ 52.66$ \\
& M-correction & $ 67.96 \pm 0.17$ & $64.48 \pm 0.76$ & $ 55.37 \pm 0.72$ & $ 24.21 \pm 1.06$ & $ 53.01 $\\
& DivideMix & $ \underline{73.17 \pm 0.28}$ & $\underline{71.01 \pm 0.16}$ & $ \underline{66.61 \pm 0.18}$ & $ 43.25 \pm 0.82$ & $63.51 $ \\
\cmidrule{2-7}
& \textbf{GPL (ours)}& $ 71.24 \pm 0.24$ & $68.89 \pm 0.07$ & $ 65.80 \pm 0.63 $ & $\boldsymbol{59.96 \pm 0.15}$ & $\underline{66.47} $ \\
& \textbf{DivideMix+ (ours)}& $ \boldsymbol{73.22 \pm 0.21}$ & $ \boldsymbol{71.03 \pm 0.32}$ & $ \boldsymbol{67.52 \pm 0.19}$ & $ \underline{58.07 \pm 0.71}$ & $ \boldsymbol{67.46} $ \\
\bottomrule
\end{tabular
}
}
\caption{Average test accuracy (\%) and standard deviation (5 runs) in various datasets under symmetric label noise. The best accuracy is \textbf{bold-faced}. The second-best accuracy is \underline{underlined}.} \label{table:sym}
\end{table*}
\section{The instantiations of SemiNLL}
\subsection{Instantiation 1: DivideMix+\label{subsec:dividemix+}}
\begin{figure}[ht]
\begin{subfigure}{.47\textwidth}
\centering
\includegraphics[width=.99\linewidth]{pic/dividemix-.pdf}
\setlength{\belowcaptionskip}{0.5cm}
\caption{DivideMix-}
\label{fig:a}
\end{subfigure}
\begin{subfigure}{.47\textwidth}
\centering
\includegraphics[width=.972\linewidth]{pic/dividemix.pdf}
\setlength{\belowcaptionskip}{0.5cm
\caption{DivideMix}
\label{fig:b}
\end{subfigure}
\begin{subfigure}{.47\textwidth}
\centering
\includegraphics[width=.97\linewidth]{pic/dividemix+.pdf}
\caption{DivideMix+}
\label{fig:c}
\end{subfigure}
\caption{Comparisons between: (a) {\em DivideMix-}, (b) {\em DivideMix}, and (c) {\em DivideMix+}. Squares represent data. Circles represent SS strategy. Octagons represent SSL backbone.}
\end{figure}
In Algorithm~\ref{alg:seminll}, if we (i) specify the \textsc{select} placeholder as a GMM~\cite{permuter2006study}, (ii) specify the \textsc{semi} placeholder as {\em MixMatch}~\cite{berthelot2019mixmatch} mentioned in Subsection~\ref{subsec:backbones}, and (iii) train two independent networks wherein each network selects clean/noisy samples in the SS phase and predicts labels in the SSL phase for the other network, then our framework is instantiated into a mini-batch version of {\em DivideMix}~\cite{li2020DivideMix}. Specifically, during the SS process, {\em DivideMix}~\cite{li2020DivideMix} fits a two-component GMM to the loss $\ell_i$ of each sample using the Expectation-Maximization technique and obtains the posterior probability of a sample being clean or noisy:
\begin{equation}
p\!\left(k\mid\ell_{i}\right)=\frac{p\!\left(k\right)p\!\left(\ell_{i}\mid k\right)}{p\!\left(\ell_{i}\right)},
\end{equation}
where $k=0\left(1\right)$ denotes the clean (noisy) set.
During the SSL phase, the clean set ${\mathcal{X}}_e$ and the noisy set ${\mathcal{U}}_e$ are fit into an improved {\em MixMatch}~\cite{berthelot2019mixmatch} strategy with label co-refinement and co-guessing. As shown in Figure~\ref{fig:b}, the SS strategy (GMM) of {\em DivideMix}~\cite{li2020DivideMix} is conducted on an epoch level. Since ${\mathcal{X}}_e$ and ${\mathcal{U}}_e$ are updated only once per epoch, the confirmation bias induced from the wrongly divided samples will be accumulated within the whole epoch. However, our mini-batch version, which is called {\em DivideMix+} (Figure~\ref{fig:c}), divides each mini-batch of data into a clean subset ${\mathcal{X}}_m$ and a noisy subset ${\mathcal{U}}_m$, and updates the networks using the SSL backbone right afterwards. In the next mini-batch, the updated networks could better distinguish clean and noisy samples.
\subsection{Instantiation 2: GPL}
Intuitively, the choice of stronger SS strategies and SSL models would achieve better performance based on our framework. Thus, we still choose GMM to distinguish clean and noisy samples due to its flexibility in the sharpness of distribution~\cite{li2020DivideMix}. As for the SSL backbone, we choose the strongest {\em Pseudo-Labeling}~\cite{arazo2020pseudo} introduced in Subsection~\ref{subsec:backbones}. We call this instantiation {\em GPL} ({\em \textbf{G}MM} + {\em \textbf{P}seudo-\textbf{L}abeling}).
Note that we do not train two networks in {\em GPL} as in {\em DivideMix}~\cite{li2020DivideMix} and {\em DivideMix+}. To our understanding, training two networks simultaneously might provide significant improvements in performance. However, this is outside the scope of this paper, since our goal is to demonstrate the versatility of our framework.
\begin{table*}
\centering
\resizebox{1.00\textwidth}{!}{
\setlength{\tabcolsep}{4.25mm}{
\begin{tabular}{c|c|cccc|c}
\toprule
\multirow{3}{*}{Datasets} & \multirow{3}{*}{Method} & \multicolumn{4}{c|}{Asymmetric} \\ \cline{3-6}
& & \multicolumn{4}{c|}{Noise ratio} & \multirow{1}{*}{Mean}\\
& & 10\% & 20\% & 30\% & 40\% \\ \midrule
\multirow{7}{*}{MNIST}
& Cross-Entropy & $ 95.78 \pm 0.19 $ & $ 91.15 \pm 0.26 $ & $ 86.01 \pm 0.25 $ & $79.92 \pm 0.32 $ & $ 88.22 $\\
& Coteaching & $ 90.32 \pm 0.02 $ & $89.03 \pm 0.02 $ & $79.80 \pm 0.27 $ & $64.94 \pm 0.02 $ & $ 81.02 $\\
& GCE & $ 94.61 \pm 0.13 $ & $94.43 \pm 0.07 $ & $94.00 \pm 0.12 $ & $93.42 \pm 0.12 $ & $ 94.12 $\\
& M-correction & $ \underline{96.74 \pm 0.03} $ & $ \underline{96.70 \pm 0.10} $ & $ \boldsymbol{96.67 \pm 0.07}$ & $94.85 \pm 0.40 $ & $ 96.24 $ \\
& DivideMix & $96.17 \pm 0.06 $ & $96.11 \pm 0.09 $ & $ 95.88 \pm 0.05$ & $95.83 \pm 0.05 $ & $96.00 $\\
\cmidrule{2-7}
& \textbf{GPL (ours)} & $ \boldsymbol{96.76 \pm 0.04} $ & $\boldsymbol{96.71 \pm 0.03} $ & $96.49 \pm 0.08 $ & $ \underline{96.45 \pm 0.04}$ & $ \boldsymbol{96.60}$ \\
& \textbf{DivideMix+ (ours)} & $ 96.67 \pm 0.04 $ & $96.66 \pm 0.07 $ & $ \underline{96.50 \pm 0.04} $ & $\boldsymbol{96.46 \pm 0.04} $ & $ \underline{96.57} $ \\
\midrule
\multirow{7}{*}{\begin{tabular}[c]{@{}c@{}}FASHION\\MNIST\\ \end{tabular}}
& Cross-Entropy & $ \boldsymbol{93.88 \pm 0.16} $ & $ 92.20 \pm 0.33 $ & $ 90.41 \pm 0.67 $ & $84.56 \pm 0.41$ & $ 90.26 $ \\
& Coteaching & $ 88.01 \pm 0.03$ & $ 78.88 \pm 0.20$ & $ 70.07 \pm 0.38$ & $ 61.97 \pm 0.21$ & $74.73 $\\
& GCE & $ \underline{93.51} \pm 0.17$ & $ \boldsymbol{93.24 \pm 0.14}$ & $\boldsymbol{92.21 \pm 0.27}$ & $89.53 \pm 0.53$ & $ \underline{92.12} $\\
& M-correction & $92.11 \pm 0.93 $ & $ 91.26 \pm 1.35 $ & $89.79 \pm 1.28 $ & $89.58 \pm 2.20$ & $ 90.69 $\\
& DivideMix & $ 91.83 \pm 0.24 $ & $91.09 \pm 0.08 $ & $89.90 \pm 0.26 $ & $ 87.58 \pm 0.26 $ & $ 90.10 $\\
\cmidrule{2-7}
& \textbf{GPL (ours)}& $ 92.52 \pm 0.22 $ & $92.23 \pm 0.09 $ & $ \underline{92.15 \pm 0.26}$ & $\boldsymbol{91.64 \pm 0.31} $& $ \boldsymbol{92.14} $ \\
& \textbf{DivideMix+ (ours)}& $ 92.56 \pm 0.39 $ & $ \underline{92.25 \pm 0.21} $ & $91.62 \pm 0.08 $ & $ \underline{89.67 \pm 0.44} $& $ 91.53 $ \\
\midrule
\multirow{7}{*}{CIFAR-10}
& Cross-Entropy & $ 90.85 \pm 0.06 $ & $ 87.23 \pm 0.40$ & $ 81.92 \pm 0.32 $ & $ 76.23 \pm 0.45 $ & $ 84.06 $\\
& Coteaching & $ 62.85 \pm 2.20 $ & $ 61.04 \pm 1.31 $ & $ 54.50 \pm 0.39 $ & $ 51.68 \pm 1.66$& $57.52 $ \\
& GCE & $ 90.40 \pm 0.09$ & $89.30 \pm 0.13$ & $86.89 \pm 0.22$ & $ 82.60 \pm 0.17$& $ 87.30 $ \\
& M-correction & $ 92.28 \pm 0.12$ & $92.13 \pm 0.17$ & $91.38 \pm 0.11$ & $ 90.43 \pm 0.23$& $ 91.56 $ \\
& DivideMix & $ 93.61 \pm 0.15$ & $ 92.99 \pm 0.21$ & $ 91.79 \pm 0.36$ & $ 90.57 \pm 0.31$& $92.24 $ \\
\cmidrule{2-7}
& \textbf{GPL (ours)}& $ \boldsymbol{94.32 \pm 0.01}$ & $ \boldsymbol{94.23 \pm 0.07}$ & $\boldsymbol{93.79 \pm 0.06}$ & $ \boldsymbol{93.02 \pm 0.30}$ & $ \boldsymbol{93.84} $\\
& \textbf{DivideMix+ (ours)}& $ \underline{94.27 \pm 0.23}$ & $ \underline{93.92 \pm 0.20}$ & $ \underline{92.82 \pm 0.28}$ & $ \underline{91.91 \pm 0.24}$& $ \underline{93.23} $ \\
\midrule
\multirow{7}{*}{CIFAR-100}
& Cross-Entropy & $ 68.58 \pm 0.34 $ & $ 68.82 \pm 0.22 $ & $ 53.99 \pm 0.50 $ & $ 44.31 \pm 0.23 $& $ 58.93 $ \\
& GCE & $70.77 \pm 0.14 $ & $ 69.22 \pm 0.15$ & $64.60 \pm 0.25$ & $51.72 \pm 1.17 $& $ 64.08$ \\
& M-correction & $ 69.44 \pm 0.52$ & $67.25 \pm 0.81$ & $63.16 \pm 1.55 $ & $52.90 \pm 1.79 $ & $ 63.19 $\\
& DivideMix & $ \boldsymbol{74.00 \pm 0.29}$ & $ \underline{73.28 \pm 0.42}$ & $\boldsymbol{72.84 \pm 0.36}$ & $ 54.33 \pm 0.69$& $ 68.61 $ \\
\cmidrule{2-7}
& \textbf{GPL (ours)}& $ 71.94 \pm 0.29$ & $71.22 \pm 0.11$ & $ 70.56 \pm 0.23$ & $\boldsymbol{69.84 \pm 0.41} $& $\boldsymbol{ 70.89} $ \\
& \textbf{DivideMix+ (ours)}& $ \underline{73.49 \pm 0.31}$ & $\boldsymbol{73.30 \pm 0.22}$ & $ \underline{72.36 \pm 0.43}$ & $ \underline{55.63 \pm 0.60}$& $ \underline{68.70} $ \\
\bottomrule
\end{tabular}%
}
}
\caption{Average test accuracy (\%) and standard deviation (5 runs) in various datasets under asymmetric label noise. The best accuracy is \textbf{bold-faced}. The second-best accuracy is \underline{underlined}.} \label{table:asy}
\end{table*}
\subsection{Self-prediction divider}
Inspired by {\em SELF}~\cite{nguyen2019self}, we introduce the {\em self-prediction divider}, a simple yet effective SS strategy which leverages the information provided by the network's own prediction to distinguish clean and noisy samples. Based on the phenomenon that DNN's predictions tend to be consistent on clean samples and inconsistent on noisy samples in different training iterations, we select the correctly annotated samples via the consistency between the original label set and the model's own predictions. The {\em self-prediction divider} determines potentially clean samples in a mini-batch if the samples' maximal likelihood predictions of the network match their annotated labels. Specifically, the samples are divided into the labeled set only if the model predicts the annotated label to be the correct class with the highest likelihood. The others are considered noisy samples, and their labels will be discarded to be regarded as unlabeled ones in SSL backbones. Compared to previous small-loss SS methods~\cite{han2018co,wei2020combating,yu2019does}, which depend on a known noise ratio to control how many small-loss samples should be selected in each training iteration, {\em self-prediction divider} does not need any additional information to perform SS strategy where the clean subset and the noisy subset are determined by the network itself. Concretely, we instantiate three learning algorithms by combining our {\em \textbf{s}elf-\textbf{p}rediction \textbf{d}ivider (SPD)} with three SSL backbones introduced in Subsection~\ref{subsec:backbones} and denote them as {\em SPD-Temporal Ensembling}, {\em SPD-MixMatch}, and {\em SPD-Pseudo-Labeling}, respectively.
\subsection{Effects of the two components}
This section demonstrates the effects of SS strategies and SSL backbones in our framework. To prove that a more robust SS strategy can boost performance for our framework, we propose {\em DivideMix-} (Figure~\ref{fig:a}) by replacing the GMM in {\em DivideMix}~\cite{li2020DivideMix} with our {\em self-prediction divider} on an epoch level. Since {\em self-prediction divider} is supposed to be weaker than GMM, {\em DivideMix-} is expected to achieve lower performance than {\em DivideMix}~\cite{li2020DivideMix}. To prove the effectiveness of the SSL backbone, we remove it after the SS process and only update the model using the supervised loss calculated from the clean samples. We will give detailed discussions in Subsection~\ref{subsec:effectofstrategy} and Subsection~\ref{subsec:effectofbackbone}.
\section{Experiment}
In this section, we first compare two instantiations of our framework, {\em DivideMix+} and {\em GPL}, with other state-of-the-art methods.
We also analyze the effects of SS strategies by comparing {\em DivideMix-}, {\em DivideMix}~\cite{li2020DivideMix}, and {\em DivideMix+}, then analyze the effects of SSL backbones by combining three representative SSL methods with our {\em self-prediction divider}. More information of our experiments can be found in supplementary~materials.
\begin{table*}
\centering
\resizebox{1.00\textwidth}{!}{
\setlength{\tabcolsep}{4.25mm}{
\begin{tabular}{c|c|cccc|cccc}
\toprule
\multirow{3}{*}{Datasets} & \multirow{3}{*}{Method} & \multicolumn{4}{c|}{Symmetric} & \multicolumn{4}{c}{Asymmetric} \\ \cline{3-10}
& & \multicolumn{4}{c|}{Noise ratio} & \multicolumn{4}{c}{Noise ratio} \\
& & 20\% & 40\% & 60\% & 80\% & 10\% & 20\% & 30\% & 40\% \\
\midrule
\multirow{3}{*}{CIFAR-10}
& DivideMix- & $ 94.49 $ & $93.64 $ & $ 91.65 $ & $76.61 $ & $ 93.58 $ & $92.87 $ & $91.21 $ & $ 90.42$ \\
& DivideMix & $ 94.82 $ & $93.95 $ & $ 92.28 $ & $89.30 $ & $ 93.61 $ & $ 92.99$ & $ 91.79$ & $ 90.57$ \\
& DivideMix+ (ours) & $\boldsymbol{94.84} $ & $\boldsymbol{94.03} $ & $ \boldsymbol{93.08} $ & $\boldsymbol{91.91} $ & $\boldsymbol{94.27}$ & $\boldsymbol{93.92} $ & $\boldsymbol{92.82} $ & $ \boldsymbol{91.91} $ \\
\midrule
\multirow{3}{*}{CIFAR-100}
& DivideMix- & $72.51 $ & $ 69.27$ & $ 61.13 $ & $25.96 $ & $ 73.62 $ & $ 72.32$ & $70.64 $ & $\boldsymbol{68.04} $ \\
& DivideMix & $ 73.17 $ & $71.01$ & $ 66.61 $ & $ 43.25$ & $ \boldsymbol{74.00} $ & $73.28 $ & $\boldsymbol{72.84} $ & $ 54.33 $ \\
& DivideMix+ (ours) & $ \boldsymbol{73.22} $ & $ \boldsymbol{71.03}$ & $ \boldsymbol{67.52} $ & $ \boldsymbol{58.07}$ & $ 73.49 $ & $\boldsymbol{73.30} $ & $72.36 $ & $55.63 $ \\
\bottomrule
\end{tabular}
}
}
\caption{Test accuracy (\%) of {\em DivideMix-}, {\em DivideMix}, and {\em DivideMix+}.}\label{table:dividemix-}
\end{table*}
\subsection{Experiment setup}
\noindent \textbf{Datasets}. We thoroughly evaluate our proposed {\em DivideMix+} and {\em GPL} on five datasets, including MNIST~\cite{lecun1998gradient-based}, FASHION-MNIST~\cite{xiao2017fashion}, CIFAR-10, CIFAR-100~\cite{krizhevsky2009learning}, and~Clothing1M~\cite{xiao2015learning}.
MNIST and FASHION-MNIST contain 60K training images and 10K test images of size $28\times28$. CIFAR-10 and CIFAR-100 contain 50K training images and 10K test images of size $32\times32$ with three channels. According to previous studies~\cite{li2020DivideMix,wei2020combating,zhang2018generalized}, we experiment with two types of label noise: symmetric noise and asymmetric noise. Symmetric label noise is produced by changing the original label to all possible labels randomly and uniformly according to the noise ratio. Asymmetric label noise is similar to real-world noise, where labels are flipped to similar classes.
Clothing1M is a large-scale real-world dataset that consists of one million training images from online shopping websites with labels annotated from surrounding texts. The estimated noise ratio is approximately 40\%~\cite{xiao2015learning}.
\noindent\textbf{Network Structure and Optimizer.}\quad Following previous works~\cite{arazo2020pseudo,li2020DivideMix,wei2020combating,zhang2018generalized}, we use a 2-layer MLP for MNIST, a ResNet-18~\cite{he2016deep} for FASHION-MNIST, the well-known ``13-CNN'' architecture~\cite{tarvainen2017mean} for CIFAR-10 and CIFAR-100, and an 18-layer PreAct Resnet~\cite{he2016identity} for Clothing1M. To ensure a fair comparison between the instantiations of our framework and other methods, we keep the training settings for MNIST, CIFAR-10, CIFAR-100, and Clothing1M as close as possible to {\em DivideMix}~\cite{li2020DivideMix} and FASHION-MNIST close to {\em GCE}~\cite{zhang2018generalized}.
For FASHION-MNIST, the network is trained using stochastic gradient descent (SGD) with 0.9 momentum and a weight decay of $1\times10^{-4}$ for 120 epochs. For MNIST, CIFAR-10, and CIFAR-100, all networks are trained using SGD with 0.9 momentum and a weight decay of $5\times10^{-4}$ for 300 epochs. For Clothing1M, the momentum is 0.9, and the weight decay is 0.001.
\noindent\textbf{Baselines}.\quad We compare {\em DivideMix+} and {\em GPL} with the following state-of-the-art algorithms and implement all methods by PyTorch on NVIDIA Tesla V100 GPUs.
\begin{enumerate}[(i)]
\setlength\itemsep{0.005em}
\item {\em Coteaching}~\cite{han2018co}, which trains two networks and cross-updates the parameters of peer networks.
\item {\em GCE}~\cite{zhang2018generalized}, which uses a theoretically grounded and easy-to-use loss function, the $\mathcal{L}_q$ loss, for NLL.
\item {\em M-correction}~\cite{arazo2019unsupervised}, which models clean and noisy samples by fitting a two-component BMM and applies {\em MixUp} data augmentation~\cite{zhang2017MixUp}.
\item {\em DivideMix}~\cite{li2020DivideMix}, which divides clean and noisy samples by using a GMM on an epoch level and leverages {\em MixMatch}~\cite{berthelot2019mixmatch} as the SSL backbone.
\end{enumerate}
\subsection{Performance Comparison}
The results of all the methods under symmetric and asymmetric noise types on MNIST, FASHION-MNIST, CIFAR-10, and CIFAR-100 are shown in Table~\ref{table:sym} and Table~\ref{table:asy}. The results on Clothing1M are shown in Table~\ref{table:clothing}.
\noindent\textbf{Results on MNIST.}\quad {\em DivideMix+} surpasses {\em DivideMix} across symmetric and asymmetric noise at all noise ratios, showing the effectiveness of the mini-batch SS strategy in our framework.
{\em M-correction} performs well under low noise ratios. However, in the hardest symmetric 80\% case, {\em DivideMix+} achieves best test accuracy.
\noindent\textbf{Results on FASHION-MNIST.}\quad FASHION-MNIST is quite similar to MNIST but more complicated. {\em DivideMix+} still outperforms {\em DivideMix} on symmetric and asymmetric noise at all noise ratios.
In the harder asymmetric 40\% noise, {\em DivideMix+} and {\em GPL} outperform the other methods by a large margin.
\noindent\textbf{Results on CIFAR-10.}\quad {\em DivideMix+} constantly outperforms {\em DivideMix}, especially in the cases with higher noise ratios.
We believe the reason is that the mini-batch SS strategy used in our framework can better mitigate the confirmation bias induced from wrongly divided samples in more challenging scenarios.
Overall, {\em GPL} and {\em DivideMix+} surpass the other methods over a large margin, with the latter performing extremely well on asymmetric noise.
\noindent\textbf{Results on CIFAR-100.}\quad
In most cases, {\em DivideMix+} and {\em DivideMix} achieve higher test accuracy than the other approaches, with {\em DivideMix+} performing better. Specifically, {\em DivideMix+} surpasses {\em DivideMix} by 14.82\% in the hardest symmetric 80\% case. An interesting phenomenon is that all the approaches suffer from performance deterioration in the asymmetric 40\% cases except {\em GPL}, which significantly outperforms the second-best algorithm over +14\%.
\begin{table*}[ht]
\centering
\resizebox{1.00\textwidth}{!}{
\setlength{\tabcolsep}{6mm}{
\begin{tabular}{l |c c c|c c c}
\toprule
Dataset & \multicolumn{3}{c|}{CIFAR-10}&\multicolumn{3}{c}{CIFAR-100}\\
\midrule
Method/Noise ratio& 20\% & 50\%& 80\% &20\% & 50\%& 80\% \\
\midrule
SPD-Cross-Entropy & 83.07 & 79.63& 48.97 & 44.43 & 35.63&10.34\\
SPD-Temporal Ensembling & 83.12 & 80.57& 48.96 &46.29& 40.39&12.74\\
SPD-MixMatch & 94.06&90.20 &88.91 & 73.23 & 68.76 &33.96\\
SPD-Pseudo-Labeling & $\boldsymbol{94.54}$ & $\boldsymbol{93.27}$& $\boldsymbol{90.34} $ & $\boldsymbol{73.29}$ & $\boldsymbol{68.97}$ & $\boldsymbol{55.40}$\\
\bottomrule
\end{tabular}
}
}
\setlength{\belowcaptionskip}{-0.43cm}
\caption{Test accuracy (\%) of the baseline and three SSL backbones integrated into our proposed framework.}\label{table:sslbackbones}
\end{table*}
\noindent\textbf{Results on Clothing1M.}\quad
To show the robustness of our framework under real-world noisy labels, we demonstrate the effectiveness of {\em DivideMix+} and {\em GPL} on Clothing1M. As shown in Table ~\ref{table:clothing}, the performance of {\em DivideMix+} is better than that of {\em DivideMix} and other methods.
\subsection{The effects of SS strategies\label{subsec:effectofstrategy}}
To study how SS strategies can affect the performance of our framework, we propose {\em DivideMix-} by replacing the GMM component in {\em DivideMix} with our {\em self-prediction divider} yet maintaining the epoch-level SS strategy for a fair comparison. Due to constraints of space, we only provide the mean value of the results in Table \ref{table:dividemix-}, which can show the overall tendency. Results with mean and standard deviation can be found in supplementary materials. In CIFAR-10, the difference between {\em DivideMix-} and {\em DivideMix} is not obvious in the lower noise ratios. However, in the most difficult symmetric 80\% case, the test accuracy of {\em DivideMix} is +12.69\% higher than {\em DivideMix-}. The difference is even greater in CIFAR-100, showing that GMM is better able to distinguish clean and noisy labels in most cases. An impressive phenomenon to note is that {\em DivideMix-} excels in the asymmetric 40\% case in CIFAR-100, which means the {\em self-prediction divider} performs better in nosier asymmetric cases than GMM.
The reason is explained in the original paper of {\em DivideMix}~\cite{li2020DivideMix}, that GMM cannot effectively distinguish clean and noisy samples under asymmetric noise with high noise ratio in datasets with a large number of classes. At the same time, the fact that {\em DivideMix+} constantly outperforms {\em DivideMix} in most cases shows that the mini-batch SS strategy in our framework is better than the epoch-level one in {\em DivideMix}.
\subsection{The effects of SSL backbones\label{subsec:effectofbackbone}}
We evaluate the effects of SSL backbones in our framework by combining the {\em self-prediction divider (SPD)} with three different SSL methods and a baseline which only updates the model using the cross-entropy loss calculated from clean samples. We denote them as {\em SPD-Temporal Ensembling}, {\em SPD-MixMatch}, {\em SPD-Pseudo-Labeling}, and {\em SPD-Cross-Entropy}, respectively. For a fair comparison, we use the ``13-CNN'' architecture~\cite{tarvainen2017mean} for all methods across different datasets. We keep most hyperparameters introduced by the SSL methods close to their original papers~\cite{arazo2020pseudo,berthelot2019mixmatch,laine2016temporal}, since they can be easily integrated into our framework without massive adjustments.
In Table~\ref{table:sslbackbones}, we list these four algorithms in the left column from weak to strong according to their performance in their original papers. The test accuracies demonstrate their corresponding performance for NLL based on our framework. {\em SPD-MixMatch} and {\em SPD-Pseudo-Labeling} outperform {\em SPD-Temporal Ensembling} by a large domain in both CIFAR-10 and CIFAR-100, especially under 80\% noise ratio (over 40\% in CIFAR-10). This phenomenon is reasonable because {\em Temporal Ensembling}~\cite{laine2016temporal} only uses consistency regularization for unsupervised loss, while {\em MixMatch}~\cite{berthelot2019mixmatch} and {\em Pseudo-Labeling}~\cite{arazo2020pseudo} also leverage entropy regularization as well as {\em MixUp} data augmentation~\cite{zhang2017MixUp}. Moreover, {\em SPD-Pseudo-Labeling} achieves remarkable test accuracy under 80\% noise ratio in CIFAR-100, which is +21.44\% higher than {\em SPD-MixMatch} and +42.66\% higher than {\em SPD-Temporal Ensembling}. We assume that this is due to the additional loss used in {\em SPD-Pseudo-Labeling} that prevents the model from assigning all labels to a single class at the early training stage.
From the results of {\em SPD-Cross-Entropy}, we can see that after the removal of the SSL backbone, the test accuracy drops dramatically compared to {\em SPD-MixMatch} and {\em SPD-Pseudo-Labeling}, especially in high noise ratios and datasets with more classes (e.g., CIFAR-100). This is possibly due to the substantial amount of data that has been removed by the {\em self-prediction divider}, leaving very few samples per class. Thus, instead of discarding noisy samples, transferring them to unlabeled ones in SSL backbones is an effective way to combat noisy labels.
\begin{table}[!t]
\centering
\resizebox{0.40\textwidth}{!}{
\setlength{\tabcolsep}{5mm}{
\begin{tabular}{l|c}
\toprule
Methods & Test Accuracy \\
\midrule
Cross-Entropy & 69.21 \\
F-correction~\cite{patrini2017making} & 69.84 \\
M-correction~\cite{arazo2019unsupervised} & 71.00 \\
Joint-Optim~\cite{tanaka2018joint} & 72.16 \\
Dividemix~\cite{li2020DivideMix} & 73.91 \\
\midrule
GPL(ours) & 73.19 \\
Dividemix+(ours) & $\boldsymbol{74.14}$ \\
\bottomrule
\end{tabular}
}
}
\setlength{\belowcaptionskip}{-0.6cm}
\caption{Test accuracy (\%) on Clothing1M.}\label{table:clothing}
\end{table}
\section{Conclusion}
This paper proposes a versatile framework called {\em SemiNLL} for NLL. This framework consists of two main parts: the mini-batch SS strategy and the SSL backbone. We conduct extensive experiments on benchmark-simulated and real-world datasets to demonstrate that {\em SemiNLL} can absorb a variety of SS strategies and SSL backbones, leveraging their power to achieve state-of-the-art performance in different noise scenarios. Moreover, we throughly analyze the effects of the two components in our framework.
\newpage
{\small
\bibliographystyle{ieee_fullname}
|
\section{Introduction}
\IEEEPARstart{I}{nfertile} couples worldwide use In-Vitro Fertilization (IVF) to treat their infertility. In a typical IVF treatment, clinicians stimulate the woman to produce many eggs, fertilize those eggs, and culture the resulting embryos for 3--5 days. The clinicians then visually inspect the embryos, select the one that appears most likely to form a viable pregnancy, and transfer it back to the mother. To aid in embryo selection, many modern clinics employ sophisticated time-lapse imaging systems~\cite{armstrong2019Timelapse} that record three-dimensional videos of the embryos as they develop.
One feature known to be predictive of an embryo's viability is the shape and symmetry among the cells in the early developing embryo, which are known as blastomeres~\cite{racowsky2011National}. However, current clinical practice is to visually score the symmetry at a few distinct points in time, which is time-consuming, inaccurate, and omits much information about the embryo, especially when time-lapse imaging is used. This makes replacing visual symmetry scoring with automated blastomere segmentation a prime candidate for improving clinical IVF practice.
However, while clinics have collected a lot of embryo images from IVF cycles, most existing blastomere segmentation algorithms~\cite{rad2018hybrid,sidhu2019Automated,moradirad2019CellNet,kheradmand2019Preimplantation,khan2016Segmentation} use hand-crafted features instead of data-driven approaches. Since hand-crafted methods are tailored to a certain dataset, they may not be robust on different datasets that are collected in varying environments. In this work, we propose a convolutional neural network, which performs amodal visual reconstruction for blastomere segmentation.
\input{fig-tex/fig_teaser.tex}
Amodal visual reconstruction, predicting the complete shape of partially-visible objects, is part of human ordinary perception. Two common examples of this are: 1) translucent objects visually overlap within the camera's view, such as when observing biomedical images of cells (Fig.~\ref{fig:motivation}\red{a}), and 2) opaque objects occlude each other and only a portion of the object is visible, such as when looking down a street at a row of parked cars (Fig.~\ref{fig:motivation}\red{b}). Beyond its importance in cognitive psychology, amodal visual perception can greatly benefit computer vision applications in practice. With it, biologists can examine new hypothesis through automatic large-scale cell shape measurement from light microscopy images and robotic agents can better navigate through complex environments with partially visible objects.
Unlike the typical instance segmentation setting, which only requires us to label the visible pixels~\cite{li2017fully,he2017mask,liu2018path,dai2016instance}, our wish to predict the shape for invisible or partially-visible object regions requires us to fit a model of shape to the image. Classic solutions have tried known rigid templates of the target object~\cite{knoll1986recognizing}, statistical models which capture object shape variation~\cite{Cootes1995}, or discriminative parts-based models learned from a dataset~\cite{parizi2014automatic} potentially with explicit occlusion reasoning~\cite{winn2006layout}.
Many recent deep-learning-based models have been proposed for amodal segmentation~\cite{li2016amodal,zhu2017semantic,ehsani2018segan,follmann2019learning,hu2019sail,qi2019amodal}. However, these approaches often do not have prior knowledge of the underlying shape, which makes the shape difficult to predict from instance observations under different amounts of occlusion. Further, unlike normal instance segmentation, images of an object under different amounts of occlusion should have the same amodal mask output. Thus, it will be more robust to classify input features into an intermediate robust representation instead of working on the pixel-level.
To exploit this additional information, we propose to learn discrete supervised learning amodal instance segmentation algorithm for partially-visible objects. From binary masks of our object class, we create a deep shape prior as an embedding space with a vector quantized-variational autoencoder (VQ-VAE;~\cite{kingma2013auto}). Then, we train our segmentation model to predict the latent representation of an object mask in a bounding box.
Segmentation performance of proposal-based instance segmentation methods~\cite{he2017mask,liu2018path} highly depends on the bounding box quality. In amodal segmentation, occlusion makes having accurate bounding boxes even more difficult. To tackle this occlusion problem, we add an occlusion detection module to a backbone network. This allows our network to propose better bounding boxes by integrating the occlusion information with the backbone features.
We experiment with a real embryo cell biomedical dataset. Furthermore, we conduct experiments on a synthetic dataset and natural images of street scenes via the KINS dataset~\cite{alhaija2018augmented} to show generalizability of our method. Our approach of encoding objects outperforms state-of-the-art instance segmentation algorithms~\cite{he2017mask,qi2019amodal} on both the translucent and occluded types of tested partial visibility.
In summary, our contribution is to propose a novel formulation that incorporates a vector quantized shape code into the amodal instance segmentation pipeline. Additionally, we exploit occlusion information when detecting and segmenting amodal objects via occlusion detection, which can be a new direction for amodal segmentation. This method achieves state-of-the-art performance on not only an internal biomedical image dataset but also the KINS natural image dataset.
Finally, to the best of our knowledge, this is the first approach that applies amodal instance segmentation method to blastomere segmentation.
\section{Related Works}
\bfsection{Blastomere Segmentation}
Traditional methods predict semantic blastomere masks using hand-crafted features without the instance-level segmentation.
Khan~\textit{et al.}~\cite{khan2016Segmentation} set seeds inside and outside of cells and optimize Markov random field for segmentation.
Rad~\textit{et al.}~\cite{rad2018hybrid} and Kheradmand~\textit{et al.}~\cite{kheradmand2019Preimplantation} generate blastomere candidates from extracted edges and select the best candidate with in terms of edge coverage.
Sidhu and Mills~\cite{sidhu2019Automated} apply thresholding and morphological operations to find the regions of blastomeres and find centers of each cell by measuring distances from pixels to the closest boundary.
Cell-Net proposed by Rad~\textit{et al.}~\cite{moradirad2019CellNet} is the closest method to ours, training a convolutional neural network for cell localization. However, Cell-Net only predicts blastomere centers, while we perform amodal instance segmentation.
\bfsection{Amodal Instance Segmentation}
Partially-visible object segmentation is typically studied in biomedical image analysis where cells are often translucent. For nuclei segmentation, Molnar~\textit{et al.}~\cite{molnar2016accurate} fit a circular active contour model~\cite{Kass1988} using multiple layered distributions of the number of nuclei per pixel. Plissiti and Nikou~\cite{plissiti2012overlapping} segment overlapping nuclei by combining nuclei boundary features with priori knowledge of nuclei shape. However, the proposed method works only when given two nuclei centers and requires parameter tuning. Lee and Kim~\cite{LeeKim2016} approach translucent cell data as a problem of superpixel segmentation for seed location, and of contour attribution and refinement via graph cuts. B\"{o}hm et al.~\cite{Bohm2018} segment translucent cell data by learning to lift the image into 3D via a UNet architecture. In both cases, the shape of the object is not specifically represented (e.g., implicitly via a prior), which makes handling occlusion-based partial visibility difficult.
Some works exist on more natural images, e.g., Kihara~\textit{et al.}~\cite{kihara2016shadows} exploit occlusion as a signal to recover full masks for object instances via a Shape Boltzmann machine~\cite{eslami2014shape}, but not for translucent objects.
Li and Malik~\cite{li2016amodal} introduce the first amodal segmentation method. They predict bounding-boxes of modal parts of objects using the object detector~\cite{ren2015faster} and extract segmentation masks using a neural network accepting a pair of an image and a bounding-box as the input. The proposed algorithm iteratively updates segmentation masks by recomputing the bounding-boxes from the output of the network. Zhu~\textit{et al.}~\cite{zhu2017semantic} announce datasets for class-independent amodal segmentation. Multiple subjects annotate the BSDS dataset to analyze the consistency between them. For computational model comparison, they evaluate modal and amodal object proposal algorithms on the proposed amodal COCO dataset. Ehsani~\textit{et al.}~\cite{ehsani2018segan} first perform amodal segmentation and then apply a generative adversarial network to have a complete object image by synthesizing the amodal area. Follman~\textit{et al.}~\cite{follmann2019learning} predict amodal masks as well as visible masks for occlusion reasoning. Hu~\textit{et al.}~\cite{hu2019sail} present a synthetic dataset for amodal instance level video object segmentation. Qi~\textit{et al.}~\cite{qi2019amodal} present an amodal segmentation dataset, KINS, by annotating the KITTI detection dataset. They also propose an amodal segmentation network by adding occlusion classification and amodal segmentation branches to the Mask R-CNN framework~\cite{he2017mask}.
Recently, Isack~\textit{et al.}~\cite{isack2018k} introduce the notion of K-convexity, and demonstrated its application in translucent instance segmentation via an energy minimization on an MRF. This allows enforcing a convexity prior on the shape of an instance (such as star~\cite{veksler2008star}, geodesic-star~\cite{gulshan2010geodesic}, hedgehog~\cite{isack2016hedgehog}, or regular~\cite{gorelick2017convexity}). However, K-convexity optimization requires seed annotation for each object instance. In contrast, our method learns pixel-wise shape priors and does not require seed annotations.
\input{fig-tex/fig_vqvae.tex}
\bfsection{Deep Learning for Shape Prior}
These approaches are common in 3D shape completion. Wu~\textit{et al.}~\cite{wu2016single} reconstruct 3D shapes by training shape priors from 3D skeleton parameters; they also later consider the naturalness of reconstructed shapes when training shape priors~\cite{wu2018learning}. Dai~\textit{et al.}~\cite{dai2017shape} transform incomplete 3D scans into complete 3D shapes by learning from template shapes. Stutz and Geiger~\cite{stutz2018learning} adopt a variational autoencoder~\cite{kingma2013auto}. For detection-based instance object segmentation, Kuo~\textit{et al.}~\cite{kuo_shapemask_2019} construct a set of prior masks for each object class and align one of the templates within a bounding box to use it as a shape prior for mask generation. However, deep shape priors are less common for amodal segmentation, where objects overlap each other.
\bfsection{Deep Learning for Vector Quantization}
Vector quantization methods have been widely used for image compression~\cite{agustsson2017soft,theis2017lossy}. Recently, van den Oord \textit{et al.}~\cite{van2017neural} proposed a vector quantized variational autoencoder for image generation. They show that the proposed method generates more realistic images using learned template codewords. Based on the vector quantized variational autoencoder, Razavi \textit{et al.}~\cite{razavi2019generating} developed a hierarchical autoencoder, which encodes an input image in high and low levels. While the high-level codewords contain global information, the low ones have local features. The hierarchical method synthesizes high-quality images by utilizing both global and local information.
\section{Vector Quantized Shape Code}
\label{sec:vq_vae}
Our goal is to learn a discrete representation of amodal shape masks. With it, we can re-formulate the amodal instance segmentation as a classification problem in the low-dimensional latent space. Comparing to previous dense pixel-level mask prediction, the proposed approach can be robust to occlusion changes and regularized in geometry.
To this end, we train a vector quantized variational autoencoder (VQ-VAE) model on the amodal masks to learn the vector quantized (VQ) shape code.
\bfsection{Comparing Latent Variable Models}
To learn a compact representation of the input,
variational autoencoder models (VAE)~\cite{kingma2013auto} are commonly used with the Gaussian prior distribution of the latent variable.
VAEs learn a global continuous code of the input with the mask encoder model $\mathcal{E}$, which can be decoded back for input reconstruction with the mask decoder model $\mathcal{D}$.
To discretize the learned code, VAE-based clustering methods jointly learn a codebook of embedding vectors that serve as clustering centers. However, as the learned embedding is global, it takes a large codebook for the input to find a similar quantized code. It requires an even larger codebook for a larger number of object categories.
VQ-VAEs~\cite{van2017neural} predict embeddings with spatial resolution and jointly learn a global codebook (Fig.~\ref{fig:vqvae}). With it, we can use the quantized embeddings to reconstruct input with a limited codebook size.
\bfsection{VQ-VAE Model}
The key component of VQ-VAE models is the embedding quantizer module. During inference, the mask encoder first transforms the input binary mask $\mathbf{x}$ into a set of latent vectors $\mathbf{e}$. Then, the embedding quantizer assigns each latent vector to the nearest code in the pre-trained codebook $\{\mathbf{q}_1,\ldots, \mathbf{q}_K\}$. Lastly, the mask decoder transforms the quantized embeddings $\mathbf{\hat{e}}$ back into a binary mask.
\bfsection{Learning}
The loss function combines a reconstruction loss, a codebook loss, and a commitment loss. The reconstruction loss is defined as the cross-entropy loss between input mask $\mathbf{x}$ and the reconstructed mask $\mathcal{D}(\mathbf{\hat{e}})$.
The codebook loss, which only applies to the codebook, makes the selected codes $\mathbf{\hat{e}}$ close to the predicted latent vector $\mathbf{e}$. The commitment loss, which only applies to the mask encoder, forces the latent vectors $\mathcal{E}(\mathbf{x})$ to stay close to the matched codes to prevent excessive fluctuations of codes. The full VQ-VAE loss function $\mathcal{L}_\textrm{v}$ is
\begin{align}
\mathcal{L}_\textrm{v}
=\|\mathbf{x}-\mathcal{D}(\mathbf{\hat{e}})\|_{2}^{2}
+\|[\mathbf{e}]-\mathbf{\hat{e}}\|_{2}^{2}
+\beta&\|\mathcal{E}(\mathbf{x})-[\mathbf{\hat{e}}]\|_{2}^{2},
\end{align}
where the operator $[.]$ stands for a stop gradient operation that blocks gradients from flowing into its argument, and $\beta$ is a hyper-parameter, which is set to 0.25.
\bfsection{Implementation Details}
The mask encoder has three convolution layers, two residual modules, and one convolution layer. The stride for each convolution layer is 2, which reduces the spatial resolution by half at each layer.
For the three convolutional layers, we use 32, 64, and 128 $4 \times 4$ sized filters, respectively.
Thus, the mask encoder changes the spatial resolution from $H \times W$ to $H/8 \times W/8$. In the last convolution layer, we set the embedding dimension to 16 empirically. Hence, the mask encoder yields a $H/8 \times W/8 \times 16$ tensor, which is a set of 16-dimensional latent vectors in embedding space.
For the embedding quantizer, we set the number of codewords $K$ to 4 empirically, as binary masks are much easier to model than natural images. Also, $K$ codewords have $K \times H/8 \times W/8$ possible combinations, which is large enough to model binary object masks.
The mask decoder has one convolutional layer, two residual modules, and three deconvolutional layers.
Note that each axis of the input image is reconstructed to its original size via the deconvolutional layers. At the end of the decoder, we add a sigmoid layer to constrain values in the reconstructed masks ranging from 0 to 1. The VQ-VAE model is trained separately, and its parameters are fixed after training.
\section{Amodal Instance Segmentation Pipeline}\label{sec:method}
We propose the VQ-VAE segmentation module to improve amodal instance segmentation. We take the proposal-based instance segmentation approach that contains two modules: object detection and mask prediction (Fig.~\ref{fig:overview}).
We attach an occlusion detection branch to object detection (Sec.~\ref{subsec:objectdetection}) and replace previous fully convolutinal network (FCN) with the proposed module for mask prediction (Sec.~\ref{subsec:method_seg}).
The whole pipeline is trained end-to-end (Sec.~\ref{subsec:method_learn}).
\input{fig-tex/fig_model_overview.tex}
\input{4_1_overview.tex}
\input{fig-tex/fig_model_detail.tex}
\input{4_2_segment.tex}
\input{4_3_refinement.tex}
\input{4_5_learn.tex}
\input{4_6_details.tex}
\subsection{Object Detection Module}
\label{subsec:objectdetection}
\bfsection{Backbone} To extract effective features from the input image, we use ResNet50~\cite{he2016deep} as a backbone network, which is trained using the ImageNet dataset~\cite{deng2009imagenet}. We drop the average pooling layer from the ResNet50 to use spatial features.
\bfsection{Bounding Box Detection}
The region proposal network (RPN)~\cite{ren2015faster} takes features from the backbone network and measures object existence probabilities and regression parameters of bounding boxes. We employ feature pyramidal networks to extract features across five scales, and minimize the sum of the loss functions at all scales. For each region of interests (ROI), we predict regression parameters and classify its object category.
\bfsection{Occlusion Detection}
Unlike Faster-RCNN~\cite{ren2015faster}, our detection module predicts both bounding boxes and a binary occlusion map. Detecting locations of occlusions allows our object detection module to predict accurate bounding boxes for partially visible objects. Using the backbone features, we estimate probabilities of each pixel being occluded $\{d_{i}\}$ via four convolution layers. We adopt the binary cross entropy loss:
\begin{equation}
\mathcal{L}_\textrm{o}=-\sum_{i \in H \times W}{\left\{l_{i} \log d_{i}+\left(1-l_{i}\right) \log \left(1-d_{i}\right)\right\}},
\end{equation}
where $H \times W$ is the spatial resolution of the backbone feature map and $l_{i}$ is the ground-truth occlusion label at pixel $i$. We concatenate the output of the second-to-last convolution layer and the backbone feature map to exploit occlusion information in the detection and segmentation modules.
\subsection{VQ-VAE Segmentation Module}\label{subsec:method_seg}
As shown in Fig~\ref{fig:maskGeneration}, the proposed VQ-VAE segmentation module has two steps: initial mask generation and mask refinement. It first generates an initial mask through decoding the predicted VQ-VAE shape code. Then, the refinement step learns to better align the initial mask with the visible object boundaries.
\bfsection{Initial Mask Generation}
Given the instance-level feature from the object detection module, we first predict the vector quantized shape code and use a pre-trained VQ-VAE decoder model to decode it into object masks with complete shapes.
We first predict a vector quantized shape code instead of a pixel-level binary mask to capture complete shapes using VQ-VAE. We use three convolution layers and one fully connected layer to predict codewords of vector quantized shape code $\mathbf{c}$. We formulate the problem of vector quantized shape code prediction as a classification problem. For the classification target, we use the pre-trained VQ-VAE mask encoder $\mathcal{E}$ to encode the ground truth instance mask $\mathbf{g}$ as shown in the right block in Fig~\ref{fig:maskGeneration}. One hot encoding makes the encoded mask $\mathcal{E}(\mathbf{g})$ as a binary representation $\mathbf{b}$. For the codeword classification at each spatial location, the binary cross entropy loss is defined as
\begin{equation}
\mathcal{L}_\textrm{e}=-\sum_{i \in M \times M \times K }{\left\{b_{i} \log c_{i}+\left(1-b_{i}\right) \log \left(1-c_{i}\right)\right\}},
\end{equation}
where $M \times M$ is a spatial resolution of a vector quantized shape code and $K$ is the number of codewords.
We then feed the predicted VQ shape code $\mathbf{c}$ into the VQ-VAE mask decoder $\mathcal{D}$ to obtain an initial mask.
\subsection{Bounding Box Adjustment for Amodal Segmentation}\label{subsec:method_centerbox}
\wj{if the bounding box adjustment works well.}
While bounding boxes for inmodal segmentation surround visible parts of objects, ones for amodal segmentation can include opaque areas occluded by other objects. For the segmentation module, therefore, it can be ambiguous to determine which object is the target, as exemplified in Fig.~\ref{fig:centerbox}(a). To tackle this ambiguity, we adjust the bounding box by making its center to be located at the center of visible pixels, as shown in Fig.~\ref{fig:centerbox}(b). By rearranging the bounding boxes, the segmentation module conduct segmentation with the center prior (\textit{i.e.}, a target object should be at the center of a bounding box).
\subsection{Learning Strategy}\label{subsec:method_learn}
During training, parameters in the region proposal network, detection, mask generation, and refinement modules are updated together to minimize the sum of the loss functions: $\mathcal{L}=\mathcal{L}_\textrm{p}+\alpha\mathcal{L}_\textrm{d}+\beta\mathcal{L}_\textrm{o}+\gamma\mathcal{L}_\textrm{e}+\delta\mathcal{L}_\textrm{r}$, where $\mathcal{L}_\textrm{p}$ and $\mathcal{L}_\textrm{d}$ indicate the losses for the region proposal network and the detection module, respectively. Hence, we train the proposed network in an end-to-end manner. Empirically, we set the hyper-parameters $\alpha=\gamma=\delta=1$ and $\beta=0.01$.
\subsection{Implementation Details}
We provide implementation details of the proposed algorithm including architectures and learning strategies.
\bfsection{Architecture}
We shrink the spatial resolution of the initial mask using two convolution layers with strides 2. The refinement network consists of four convolution layers and two deconvolution layers. It outputs class-wise masks at each output channel to decouple segmentation and classification.
\bfsection{Learning}
We initialize parameters in the proposed networks with random values except for the backbone network, which uses weights from the ResNet50~\cite{he2016deep}. We train the network via the stochastic gradient descent optimizer. We set the initial learning rate to $0.04$, and reduce it to $0.004$ and $0.0004$ after 10,000 and 11,000 iterations, respectively. We train networks for 12,000 iterations. We use a minibatch size of 16. It takes less than two days to train the proposed networks.
\bfsection{Running time}
We measure the average computational time of the proposed algorithm on a single Titan X GPU. For images whose shorter axis is fixed to 800, the average running time is 1.75 frames per second. Note that we set the number of proposals to 1,000.
\subsection{Ablation Study}
\bfsection{Ablation Study}
We perform two ablation studies on the KINS dataset. We chose KINS over the embryo dataset for more general analysis, since the objects in KINS have more diverse shapes. We use Mask R-CNN in these studies. First, we remove the occlusion detection branch (VQ-VAE + Refinement). To this end, we train the network without the loss function for occlusion detection $\mathcal{L}_\textrm{o}$. Second, we exclude the refinement decoder in the segmentation module (VQ-VAE). To train the network without the refinement decoder, we minimized the embedding loss $\mathcal{L}_\textrm{e}$ only.
We compare these two settings with the full architecture (VQ-VAE + Refinement + Occlusion map) on the KINS dataset. Table~\ref{table:ablation} lists the mAP scores for each ablation setting. Our full architecture performs 0.303 mAP, which is better than the other settings. It indicates that all our components are necessary for accurate amodal segmentation. The inferior performance of the setting without refinement comes from the lack of low-level features.
\subsection{Vector Quantized Shape Code}
\bfsection{Comparing with VAE}
We use the Mask-RCNN pipeline and compare different shape modeling from FCN (no latent code), VAE (continuous latent code) and VQ-VAE (discrete latent code) modules.
Table~\ref{table:synthesized} compares mAP indices of the three methods on the synthesized dataset. The proposed VQ-VAE method is better able to delineate complete shapes versus the other two methods. Especially, there is a considerable margin between the proposed algorithm and FCN.
Fig.~\ref{fig:synthesized} shows segmentation results for partially-visible object segmentation. The proposed algorithm discovers these objects; though sometimes the baseline methods fail to predict full geometric shapes. We provide more segmentation results in the supplementary materials.
\bfsection{Robustness to Occlusion}
We assess the occlusion handling abilities of FCN and VQ-VAE when the perfect bounding boxes are given. To this end, we design the following \emph{in silico} psychophysics experiment.
Given an input shape, we add another shape in front with different degrees of occlusion. We feed these test images into the trained amodal segmentation models with FCN or VQ-VAE modules.
We quantitatively compare the shape completion performance of the proposed algorithm and FCN. Our algorithm (0.968) outperforms FCN (0.935) in terms of IoU. Moreover, as shown in Fig.~\ref{fig:occlusion}, the proposed algorithm reliably complete the full shapes of occluded objects. However, FCN makes soft predictions on unseen regions, thus restored regions are blurry.
\subsection{Cell Segmentation on Embryo Images}
\subsection{Main Results on Embryo Cell Images}
In vitro fertilization clinicians predict embryo transfer success by visually observing cell properties like size, granularity, and cleavage (cell split) timing. Cell segmentation of embryo images would automate this property collection for more efficient prediction. Note that our method is more interpretable by clinicians compared to predicting a single number (cell count) from the input image~\cite{khan_deep_2016}.
\bfsection{Data}
From the IVF clinic in Tel Aviv Medical Center, Israel, we collect 11,671 embryo images, each with a spatial resolution of 500$\times$500 pixels. The numbers of cells in each embryo image varies from 2 to 8. Note that we exclude one cell images to evaluate amodal instance segmentation methods. To obtain ground-truth segments, we annotate cells and then ask experts to proofread the annotations. We use 7,054 images for training and the remaining 4,617 for testing.
Fig.~\ref{fig:embryo_data} show examples of embryos and their ground-truth annotations for blastomere instances.
We observe that cells are highly overlapping and only partially visible. The size of cells varies as cells cleave and shrink.
\input{tab-tex/tab_embryo.tex}
\input{fig-tex/fig_embryo.tex}
\bfsection{Results}
Table~\ref{table:embryo} compares the results of our proposed algorithm with Mask R-CNN~\cite{he2017mask}. We report mean average precision metrics for the evaluation of the cell segmentation methods. The proposed algorithm outperforms the baseline methods. Qualitatively, we observe that the proposed network faithfully detect embryo cells (Fig.~\ref{fig:embryo_result}). Even though partial boundaries of cells are missing, the proposed algorithm generates masks accurately by considering the shape prior of embryo cells.
\subsection{Amodal Segmentation on KINS Dataset}
\input{tab-tex/tab_kins.tex}
To demonstrate the general applicability of our proposed method, we test on an amodal instance segmentation dataset for natural images with a greater diversity of shapes.
\bfsection{Data}
The KINS dataset~\cite{qi2019amodal} is a benchmark for amodal instance segmentation algorithms, which is originally from the KITTI dataset~\cite{geiger2012we}. It consists of 7,474 training and 7,517 test images of driving scenes. The annotated objects belong to one of 7 object classes: pedestrian, cyclist, car, van, tram, truck, and misc-vehicle. The KINS dataset provides both amodal and inmodal ground-truth annotations.
\bfsection{Results}
Table~\ref{table:kins} lists mean average precision metrics of the results of the proposed algorithm with Mask R-CNN~\cite{he2017mask} and Mask R-CNN + ASN~\cite{qi2019amodal}. Our proposed algorithm performs better than the conventional FCN method on the Mask R-CNN pipeline and yields a slightly better result on Mask R-CNN + ASN. Qualitatively, it finds complete masks of occluded cars (Fig.~\ref{fig:kins_result}). In the last two rows in Fig.~\ref{fig:kins_result}, the proposed method fails to segment out cars on the left. This is because the non-maximum suppression removes highly overlapped bounding boxes.
\section{Experiments}
We compare the proposed method with state-of-the-art methods on a microscopy image dataset and a natural image dataset. Then, we perform ablation studies on the natural dataset to better understand each component and to validate our design choices.
\subsection{Experiment Setup}
\bfsection{Comparison methods}
For amodal instance segmentation, we can use different object detection pipelines, \textit{e.g.}, Mask-RCNN~\cite{he2017mask}.
With the same pipeline, the proposed VQ-VAE segmentation module is compared with the fully convolutional network (FCN) on two datasets.
\bfsection{Metrics} We use mean average precision (mAP), which is standard for object instance segmentation~\cite{lin2014microsoft}. Let $\textrm{AP}_{k}$ denotes a predicted segmentation as correct if its mask intersection over union (IoU) is higher than $k$.
mAP score is the average of $\{\textrm{AP}_{k}\}$ where $k$ ranges from 0.5 to 0.95 at 0.05 intervals.
\input{fig-tex/fig_embryodata.tex}
\input{5_exp_embryo.tex}
\input{tab-tex/tab_syn.tex}
\subsection{Additional Results on Synthetic Images}
\input{fig-tex/fig_syn.tex}
\input{fig-tex/fig_occlusion.tex}
\input{fig-tex/fig_kins.tex}
\input{5_exp_design_choice.tex}
\subsection{Additional Results on Natural Images}
\input{5_exp_kins.tex}
\input{tab-tex/tab_ablation.tex}
\input{5_exp_ablation_model.tex}
\section{Conclusion}
We proposed an image segmentation method for blastomere instances, which outputs complete masks of cells automatically. The proposed algorithm predicts bounding boxes first and then generates masks. We show that it is effective to learn a mapping from the bounding box features to a shape prior embedding space from a VQ-VAE. This allows us to cope with translucent cells. We also show the benefits of occlusion detection for amodal object detection and segmentation.
Our method is applicable for any partially visible objects, not only cells but also geometric shapes, cars, or pedestrians.
Experimental results on the embryo, synthesized, and KINS demonstrated that our proposed algorithm outperforms state-of-the-art object instance segmentation methods~\cite{he2017mask,qi2019amodal}.
Our future works include application to other objects in natural scenes and expanding to biomedical problems that suffer occlusions, such as human blood cell segmentation. We also suggest proposal-free amodal segmentation networks with the center prediction to achieve real-time running speed. Lastly, by adopting generative adversarial networks~\cite{goodfellow2014generative}, we might be able to learn shape priors better.
\subsubsection{#1}\vspace*{-0.1in}}
\newcommand{\red}[1]{\textcolor{red}{#1}}
\newcommand{\textit{et al.}}{\textit{et al.}}
\newcommand{\textit{e.g.}}{\textit{e.g.}}
\newcommand{\textit{i.e.}}{\textit{i.e.}}
\def\BibTeX{{\rm B\kern-.05em{\sc i\kern-.025em b}\kern-.08em
T\kern-.1667em\lower.7ex\hbox{E}\kern-.125emX}}
\markboth{\journalname, VOL. XX, NO. XX, XXXX 2020}
{Jang \MakeLowercase{\textit{et al.}}: Learning Vector Quantized Shape Code for Amodal Blastomere Instance Segmentation}
\begin{document}
\title{Learning Vector Quantized Shape Code for Amodal Blastomere Instance Segmentation}
\author{Won-Dong Jang$^{*}$, \and Donglai Wei, \and Xingxuan Zhang, \and Brian Leahy, \and Helen Yang, \and James Tompkin, \and Dalit Ben-Yosef, \and Daniel Needleman, \and and Hanspeter Pfister
\thanks{W. Jang, D. Wei, B. Leahy, H. Yang, H. Pfister, D. Needleman are with the School of Engineering and Applied Sciences, Harvard University, Cambridge, MA, USA.
corresponding author email: <EMAIL>}
\thanks{X. Zhang is with the School of Engineering, Jiaotong University, China.}
\thanks{J. Tompkin is with the Department of Computer Science, Brown University, Cambridge, MA, USA.
\thanks{D. Ben-Yosef is with the Tel Aviv Sourasky Medical Center
}
\maketitle
\input{0_abstract}
\begin{IEEEkeywords}
Blastomere segmentation, Cell segmentation, Amodal segmentation, Shape prior, Vector Quantization, Autoencoder.
\end{IEEEkeywords}
\input{1_intro.tex}
\input{2_related.tex}
\input{3_vqvae.tex}
\input{4_0_method.tex}
\input{5_exp.tex}
\input{6_conclusion.tex}
\bibliographystyle{IEEEtran}
|
\section{Introduction}
A recurring theme across many diverse areas of mathematics is the natural occurrence of structures for which a key aspect can take one of only two values. We mention several examples of this two-valued phenomenon.
In coding theory, a projective two-weight code is a linear code whose codeword weights take one of exactly two distinct values \cite{CK}.
In finite geometry, a projective two-intersection set is a point set in projective space whose intersection with each hyperplane has one of exactly two distinct sizes \cite{CK};
and an $m$-ovoid or a tight set in a polar space is a point set whose intersection with every tangent hyperplane of the polar space has one of exactly two distinct sizes \cite[Chapter 2]{BvM}, \cite[Section 4.5]{MWX}.
In graph theory, a strongly regular graph has exactly two distinct eigenvalues \cite[Chapter 1]{BvM}.
In group theory, a rank $3$ permutation group has exactly two nontrivial orbitals \cite[Section 10]{CK}; and a symmetric Schur ring of dimension~$3$ over a group $G$ is a partition of the nonidentity group elements into exactly two nontrivial subsets which, together with the identity element, form a subalgebra of $\CC[G]$ \cite{Wie}.
In number theory, uniform cyclotomy is a partition of the nonzero elements of a finite field into cyclotomy classes such that exactly two distinct cyclotomic numbers occur~\cite{BMW}.
An underlying connection between many instances of these two-valued phenomena is provided by a partial difference set.
Indeed, let $D$ be a $k$-subset of an additive abelian group $G$ of order~$v$. The subset $D$ is a $(v,k,\la,\mu)$ \emph{partial difference set} in $G$ if the multiset $\{x-y \mid x,y \in D, x \ne y \}$ contains each nonidentity element of $D$ exactly $\la$ times and each nonidentity element of $G \sm D$ exactly $\mu$ times.
Partial difference sets, possibly having some additional properties, give rise to examples of each of the above structures. The construction of partial difference sets is therefore of great interest. We refer to \cite{CK,Ma} for excellent surveys of partial difference sets and equivalent structures, and to \cite{A,ACK,BKLP,BLMX,BDMR,BL,BWX,B,CGT,CM,CTZ,CF13,CP11,CP13,CRX,CHOP,CDMPS,CoM,CoMa,CoMP,CP14,CP17,CP18,CoPe13,D,DX00,DX04,FMX15JCT,FMX15C,FWXY,FX,FKM,HP,H00,H02,H03,HLX,HN,IZZ,LM95,LM96,MS,M14,M18,M21,MO,MX14,MX18,O16,O16DCC,O16JG,OP,P,P08EJC,P09JCD,P09DCC,P10,P19,PDS,P02,P08DCC,RX,SW,TPF,WQWZ} for investigations spanning twenty-five years into which groups $G$ contain a partial difference set with specified parameters $(v,k,\la,\mu)$.
This paper is concerned with the construction of two of the richest classes of partial difference sets, whose parameters are determined by only two integers. A partial difference set whose parameters $(v,k,\la,\mu)$ take the form
$\big(n^2,r(n-1), n+r(r-3),r(r-1)\big)$ has \emph{$(n,r)$ Latin square type},
and one whose parameters take the form
$\big(n^2,r(n+1), -n+r(r+3),r(r+1)\big)$ has \emph{$(n,r)$ negative Latin square type}.
A partial difference set $D$ in $G$ is \emph{regular} if $1_G \notin D$ and $D = \{d^{-1} \mid d \in D\}$.
A great deal of research has been conducted into the existence of a collection of $t > 1$ disjoint regular $(tc,c)$ Latin square type partial difference sets in an abelian group $G$ of order~$t^2c^2$.
The number of elements of $G$ avoided by the partial difference sets of such a collection is $tc$, so the collection has maximum possible size when $c>1$.
The principal motivation for this paper was the discovery of a widespread structural property that does not seem to have been previously recognized: in all previous constructions, possibly after some modification, the $tc$ avoided elements of $G$ form a subgroup~$U$. We therefore refer to such a collection of partial difference sets as a \emph{$(c,t)$ LP-packing in $G$ relative to~$U$}.
The presence of the associated subgroup $U$ is key to the recursive lifting construction of LP-packings of increasing exponent that we present in Section~\ref{sec5}. In general, lifting constructions (that increase the group exponent) have historically been more challenging to find than product constructions (that combine objects into the direct product of the starting groups).
Table~\ref{tab-LPpacking} illustrates how widely the concept of LP-packings has been previously studied, often implicitly, and in each case identifies the subgroup $U$ explicitly.
In some cases, identifying the crucial subgroup $U$ from the original reference requires considerable effort.
We shall unify and extend many results for LP-packings, whose original derivation relied on a variety of sometimes delicate approaches, by means of a common framework that depends only on elementary methods.
We shall show that we have considerable control over the choice of the associated subgroup~$U$.
In addition to the recursive construction of Section~\ref{sec5}, we shall give a product construction for LP-packings in Section~\ref{sec4}.
We present the following result (to be proved as Corollary~\ref{coro-LPpackingeven}) as a showcase for our constructions.
\begin{theorem}\label{thm-preview}
Let $p$ be prime, let $s_1,\dots,s_v$ be nonnegative integers (not all zero), and let $m=\min\{s_i \mid s_i>0\}$. For each $i=1,\dots,v$, let $G_i=\Z_{p^i}^{2s_i}$ and let $U_i$ be a subgroup of $G_i$ of order~$p^{is_i}$.
Let $G = \prod_{i=1}^v G_i$, let $U = \prod_{i=1}^v U_i$, and let $n = \sqrt{|G|}$.
Then for each $j = 0,\dots,m-1$, there exists an
$\big(\frac{n}{p^{m-j}}, p^{m-j}\big)$ LP-packing in $G$ relative to~$U$:
a collection of $p^{m-j}$ disjoint regular $\big(n, \frac{n}{p^{m-j}}\big)$ Latin square type partial difference sets in $G$ avoiding~$U$.
\end{theorem}
We shall show in Section~\ref{sec3} that a collection of $t$ disjoint regular $(tc,c)$ Latin square type partial difference sets with $c=1$ does not have maximum possible size, because there exists a size $t+1$ collection known equivalently as a $(t,t+1)$ partial congruence partition. However, we regard the $(1,t)$ LP-packing as a more natural object than the larger partial congruence partition, because it forms the base case of the recursive construction of Section~\ref{sec5} and so allows a powerful generalization to nonelementary abelian groups.
We similarly refer to a collection of $t-1 > 0$ disjoint regular $(tc,c)$ negative Latin square type partial difference sets in an abelian group $G$ of order~$t^2c^2$,
for which the $(c-1)(tc+1)$ nonidentity avoided elements of $G$ form a regular $(tc,c-1)$ negative Latin square type partial difference set in $G$, as a \emph{$(c,t-1)$ NLP-packing}; such a collection has maximum possible size for all $c \ge 1$.
The NLP-packing structure was already known, and previous constructions are summarized in Table~\ref{tab-NLPpacking}.
We shall use a product construction in Section~\ref{sec6} to combine small examples of NLP-packings with the families of LP-packings constructed recursively here, and so extend previous constructions of families of NLP-packings.
We mention that both LP-packings and NLP-packings give rise to group-based amorphic association schemes~\cite{vM}.
Our principal objective is to determine for which abelian groups $G$ and subgroups $U$ there exists a $(c,t)$ LP-packing in $G$ relative to $U$, and for which abelian groups $G$ there exists a $(c,t-1)$ NLP-packing in~$G$. We propose that the natural framework for studying regular $(tc,c)$ Latin square type and negative Latin square type partial difference sets is as these respective collections, rather than as single examples.
The rest of the paper is organized as follows.
Section~\ref{sec2} gives a historical overview of packings of partial difference sets of Latin square type and negative Latin square type in abelian groups.
Section~\ref{sec3} introduces LP-packings, describes their relationship to partial congruence partitions, establishes constraints on their structure, and presents a product construction.
Section~\ref{sec4} introduces an LP-partition as an auxiliary configuration in the construction of an LP-packing in a larger group from an LP-packing in a smaller group.
Section~\ref{sec5} recursively constructs infinite families of LP-partitions and LP-packings of increasing exponent, inspired by a recursive lifting construction of difference sets using relative difference sets~\cite{DJ}.
Section~\ref{sec6} introduces NLP-packings, and uses a product construction to combine an NLP-packing with an LP-packing to give an NLP-packing in the direct product of the starting groups.
Section~\ref{sec7} examines how LP-packings and NLP-packings are related to hyperbolic and elliptic strongly regular bent functions, and
Section~\ref{sec8} examines how these packings are related to reversible Hadamard difference sets.
Section~\ref{sec9} proposes some open problems for future research.
\begin{table}[ht]
\ra{1.3}
\caption{Known $(c,t)$ LP-packings in $G$ relative to $U$, where $p$ is prime and $\s$ is defined in~\eqref{eqn-s}}
\label{tab-LPpacking}
\begin{center}
\scalebox{0.86}{
\begin{tabular}{|c|c|c|c|c|c|}
\hline
$c$ & $t$ & Group $G$ & Subgroup $U$ & Restrictions & Source \\ \hline
\mr{2}{*}{$p^{w-s}$} & \mr{2}{*}{$p^s$} & \mr{2}{*}{$\Z_{p^2}^w$} & \mr{2}{*}{$(p\Z_{p^2})^w$} & $s$ is proper & \mr{2}{*}{\cite[Thm.~3.1]{CRX}} \\
& & & & divisor of $w>1$ & \\ \hline
$p$ & $p$ & $\Z_{p^2}^2$ & $(p\Z_{p^2})^2$ & & \cite[Thm.~3.1]{D} \\ \hline
$2^{a-1}$ & $2$ & $\Z_{2^a}^2$ & order $2^a$, noncyclic & & \cite[Thm.~3.1]{DP} \\ \hline
$p^{w-\s(p,w)}$ & $p^{\s(p,w)}$ & $\Z_{p^2}^w$ & $(p\Z_{p^2})^w$ & $w > 1$ & \cite[Lem.~3.1]{HLX} \\ \hline
$\frac{n}{p^s}$ & $p^s$ & $\Z_{p^{a+1}}^{2us} \times \Z_{p^{a}}^{2ws}$ & $\Z_{p^{a+1}}^{us} \times \Z_{p^{a}}^{ws}$ & $n=\sqrt{|G|}$ and $a \ge 1$ & \cite[Thm.~3.1]{LM90} \\ \hline
$p^{w-1}$ & $p$ & $\Z_{p^2}^w$ & $(p\Z_{p^2})^w$ & $w > 1$ & \cite[Thm.~2.2]{LM96} \\ \hline
$p^s$ & $p^s$ & $\Z_{p^2}^{2s}$ & $(p\Z_{p^2})^{2s}$ & $s > 1$ & \cite[Prop.~4.1]{P02} \\ \hline
\mr{2}{*}{$p^{(2a-1)s}$} & \mr{2}{*}{$p^s$} & \mr{2}{*}{$\Z_{p^{2a}}^{2s}$} & \mr{2}{*}{$(p^a\Z_{p^{2a}})^{2s}$} & & \cite[Lem.~4.3]{P06} \\
& & & & & \cite[Thm.~3.1]{P08DCC} \\ \hline
\mr{2}{*}{$\frac{n}{p}$} & \mr{2}{*}{$p$} & \mr{2}{*}{$\Z_p^{2u} \times
\prod_{i=1}^v\Z_{p^{2i}}^{2s_{2i}}$} & \mr{2}{*}{$\Z_p^u \times
\prod_{i=1}^v (p^i\Z_{p^{2i}})^{2s_{2i}}$} & $n=\sqrt{|G|}$ and & \mr{2}{*}{\cite[Cor.~6.2]{P08DCC}} \\
& & & & $u, s_{2i} \ge 0$ & \\ \hline
$p^{w-1}$ & $p$ & $\Z_{p^2}^w$ & $(p\Z_{p^2})^w$ & $p$ is odd divisor of $w$ & \cite[Thm.~4.1]{RX} \\ \hline
$2^{\ceiling{w/2}}$ & $2^{\floor{w/2}}$ & $\Z_{4}^w$ & $(2\Z_{4})^w$ & $w > 1$ & \cite[Lem.~4.5]{XD} \\ \hline
\end{tabular}
}
\end{center}
\end{table}
\begin{table}[ht]
\ra{1.3}
\caption{Known $(c,t-1)$ NLP-packings in $G$}
\label{tab-NLPpacking}
\begin{center}
\scalebox{0.84}{
\begin{tabular}{|c|c|c|c|c|}
\hline
$c$ & $t-1$ & Group $G$ & Restrictions & Source \\ \hline
\mr{3}{*}{$\frac{n}{2}$}& \mr{3}{*}{$1$} & \mr{3}{*}{$\Z_2^{2u} \times \Z_4^w \times \prod_{i=3}^v \Z_{2^i}^{2s_i} \times
\Z_3^{2y} \times \prod_{i=1}^{z} \Z_{p_i}^{4t_i}$} & $n=\!\sqrt{|G|}$ and prime $p_i > 3$, & \mr{3}{*}{\cite[Prop.~3.12]{CP13}} \\
& & & $u,w,s_i,y,t_i \ge 0$, & \\
& & & if $u=0$ then $y=t_i=0$ & \\ \hline
\mr{2}{*}{$\frac{n}{3}$}& \mr{2}{*}{$2$} & \mr{2}{*}{$\Z_3^{2+2u} \times \Z_9^w \times \prod_{i=3}^v \Z_{3^i}^{2s_i}$} & $n=\sqrt{|G|}$ and $u,w,s_i \ge 0$, & \mr{2}{*}{\cite[Prop.~3.13]{CP13}} \\
& & & $w \ne 1$ & \\ \hline
\mr{2}{*}{$\frac{n}{4}$}& \mr{2}{*}{$3$} & $\Z_2^{4+2u} \times \Z_4^w \times \prod_{i=3}^v \Z_{2^i}^{4s_i}$ & $n=\sqrt{|G|}$ and $u,w,s_i \ge 0$, & \mr{2}{*}{\cite[Prop.~3.14]{CP13}} \\ \cline{3-3}
& & $\Z_2^{2u} \times \Z_4^{w+2} \times \prod_{i=3}^v \Z_{2^i}^{4s_i}$ & $u \ne 1$ and $w \notin \{1,3\}$ & \\ \hline
\mr{2}{*}{$4^{w-1}$} & \mr{2}{*}{$3$} & \mr{2}{*}{$\Z_{4}^{2u}\times \Z_2^{4w-4u}$} & $0 \le u \le w$ if $w$ odd & \cite[Cor.~2.2]{DX04} \\
& & & $0 \le u < w$ if $w$ even & \cite[Thm.~3.6]{DX06} \\ \hline
\mr{2}{*}{$\frac{n}{4}$}& \mr{2}{*}{$3$} & \mr{2}{*}{$\Z_2^{4u} \times \Z_4^{2w} \times \prod_{i=2}^v \Z_{2^{2i}}^{4s_{2i}}$} & $n=\sqrt{|G|}$ and $u,w,s_{2i} \ge 0$,& \mr{2}{*}{\cite[Thm.~4.2]{P08DCC}} \\
& & & $u+w \ge 1$ & \\ \hline
$\frac{n}{3}$ & $2$ & $\Z_3^{2u+2} \times \prod_{i=1}^v \Z_{3^{2i}}^{2s_{2i}}$ & $n=\sqrt{|G|}$ and $u, s_{2i} \ge 0$ & \cite[Cor.~5.1]{P08DCC} \\ \hline
$3^{w-1}$ & $2$ & $\Z_{3}^{2w}$ & $w \ge 1$ & \cite[Thm.~1]{TPF} \\ \hline
\end{tabular}
}
\end{center}
\end{table}
\section{Historical Overview}\label{sec2}
In this section, we give an overview of previous constructions of regular Latin square type and negative Latin square type partial difference sets in abelian groups. In many cases, these results produce only a single partial difference set, whereas our objective in later sections will be to construct disjoint collections.
Let $G$ be an abelian group.
We firstly present some basic results involving the group ring $\Z[G]$ and character theory.
For $A \in \Z[G]$ and $g \in G$, denote the coefficient of $g$ in $A$ by~$[A]_g$ (so that $A = \sum_{g \in G} [A]_g g$).
For $A \in \Z[G]$, write $A^{(-1)}=\sum_{g \in G} [A]_gg^{-1}$.
For a subset $A$ of~$G$, by a standard abuse of notation we also denote the group ring element $\sum_{g \in A}g$ by $A$.
A \emph{character} $\chi$ of $G$ is a group homomorphism from $G$ to the multiplicative group of the complex field~$\mathbb{C}$. Write $\wh{G}$ for the character group of~$G$.
A character $\chi \in \wh{G}$ is \emph{principal} on a subgroup $H \leqslant G$ if $\chi(h)=1$ for each $h \in H$. For a subgroup $H \leqslant G$, write \[
H^{\perp}=\{ \chi \in \wh{G} \mid \mbox{$\chi$ is principal on $H$} \}
\]
(so that $G^\perp$ consists of only the principal character on~$G$).
For $\chi \in \wh{G}$ and $A \in \Z[G]$, write $\chi(A)$ for the character sum $\sum_{g \in G} [A]_g\chi(g)$. For a more detailed treatment of group rings and character theory, see \cite[Chap.~VI]{BJL} and \cite[Chap.~1]{Pott}.
\begin{definition}\label{def-pds}
Let $D$ be a $k$-subset of an additive group $G$ of order~$v$. The subset $D$ is a $(v,k,\la,\mu)$ \emph{partial difference set} (PDS) in $G$ if the multiset $\{x-y \mid x,y \in D, x \ne y \}$ contains each nonidentity element of $D$ exactly $\la$ times and each nonidentity element of $G-D$ exactly $\mu$ times.
The partial difference set is \emph{regular} if $1_G \notin D$ and $D=D^{(-1)}$.
\end{definition}
\noindent
The condition $D=D^{(-1)}$ in Definition~\ref{def-pds} is guaranteed to hold except in the special case $\la = \mu$ \cite[Prop.~1.2]{Ma}, in which case the partial difference set is a \emph{$(v,k,\la)$-difference set} in~$G$ (see Section~\ref{sec8}).
Provided that $D=D^{(-1)}$, the condition $1_G \notin D$ is not restrictive \cite[p.~222]{Ma}; we shall be concerned only with partial difference sets that are regular.
Let $D$ be a $k$-subset of a group $G$ of order $v$ for which $1_G \notin D$.
Then, in group ring notation, $D$ is a $(v,k,\la,\mu)$ PDS in $G$ if and only if
\begin{equation}\label{eqn-PDSgpring}
DD^{(-1)} = k-\mu + (\la-\mu) D + \mu G \quad \mbox{in $\Z[G]$}.
\end{equation}
The following result is a consequence of the orthogonality properties of characters.
\begin{proposition}[Fourier inversion formula] \label{prop-fourier}
Let $G$ be an abelian group and let $A \in \Z[G]$. Then
\[
[A]_g=\frac{1}{|G|}\sum_{\chi \in \wh{G}} \chi(A)\ol{\chi(g)} \quad \mbox{for each $g \in G$}.
\]
In particular, elements $A, B$ of $\Z[G]$ are equal if and only if $\chi(A)=\chi(B)$ for all $\chi \in \wh{G}$.
\end{proposition}
\noindent
For a regular PDS $D$, we can use the relation $D=D^{(-1)}$ to rewrite \eqref{eqn-PDSgpring} as $D^2 = k-\mu + (\la-\mu) D + \mu G$ in~$\Z[G]$. Applying a nonprincipal character $\chi$ to both sides then shows that the character sum $\chi(D)$ satisfies the quadratic equation $\chi(D)^2-(\la-\mu)\chi(D)-(k-\mu)=0$. Using Proposition~\ref{prop-fourier}, we therefore obtain the following characterization of a regular partial difference set in terms of its character sums.
\begin{lemma}[{\cite[Cor.~3.3]{Ma}}]\label{lem-pdschar}
Let $G$ be an abelian group of order $v$. Let $D$ be a $k$-subset of~$G$ for which $1_G \notin D$.
Let $\la,\mu$ be nonnegative integers satisfying
$k^2=k-\mu+(\la-\mu)k+\mu v$ and $(\la-\mu)^2+4(k-\mu) \ge 0$.
Then $D$ is a regular $(v,k,\la,\mu)$ partial difference set in $G$ if and only if
\[
\chi(D) = \tfrac{1}{2}\Big(\la-\mu \pm \sqrt{(\la-\mu)^2+4(k-\mu)}\Big)
\quad \mbox{for all nonprincipal characters $\chi$ of $G$}.
\]
\end{lemma}
We are interested in the following two types of parameters for partial difference sets.
\begin{definition}\label{def-LatinPDS}
Let $n \ge 1$ and $r \ge 0$ be integers.
An $(n^2,r(n-\ep),\ep n+r^2-3\ep r,r^2-\ep r)$ PDS
is an \emph{$(n,r)$ Latin square type PDS} if $\ep=1$, and is an \emph{$(n,r)$ negative Latin square type PDS} if $\ep=-1$.
\end{definition}
\noindent
The connection between Latin square type partial difference sets and Latin squares is ``rather indirect'' \cite[p.~2]{HN}.
We allow the case $r=0$ in Definition~\ref{def-LatinPDS}, which corresponds to the empty set (see Remark~\ref{rem-NLPpacking}~$(iii)$).
We next give a character-theoretic description of the two parameter sets of interest.
\begin{lemma}\label{lem-LatinPDS}
Let $n \ge 1$ and $r \ge 0$ be integers, and let $\ep \in \{1,-1\}$.
Let $G$ be an abelian group of order $n^2$, and let $D$ be an $r(n-\ep)$-subset of $G$. Then $D$ is a regular $(n,r)$ Latin square type PDS in $G$ (when $\ep=1$), and is a regular $(n,r)$ negative Latin square type PDS in $G$ (when $\ep=-1$), if and only if
\[
\chi(D) \in \{ -\ep r, \ep (n-r) \} \quad \mbox{for all nonprincipal characters $\chi$ of $G$}.
\]
If $D$ is a regular $(n,r)$ Latin square type or negative Latin square type PDS in $G$, then the character set $\wti{D}=\{ \chi \in \wh{G} \mid \chi(D)=\ep(n-r) \}$ is also a regular $(n,r)$ Latin square type or negative Latin square type PDS in~$\wh{G}$, respectively.
\end{lemma}
\begin{proof}
The first statement follows directly from Lemma~\ref{lem-pdschar}.
The second statement is obtained from \cite[Thm.~3.4]{Ma}: let $D^+$ be the dual of $D$ as defined in \cite{Ma}, and note that $\wti{D} = D^+$ in the case $\ep=1$ and that $\wti{D} = \wh{G}-1_{\wh{G}}-D^+$ in the case $\ep=-1$.
\end{proof}
We now briefly survey the numerous previous constructions of Latin square type and negative Latin square type partial difference sets, which we classify into three classes that are restricted to elementary abelian groups and five that are not.
The three construction classes that are restricted to elementary abelian groups arise from the following sources.
\begin{enumerate}[(1)]
\item \emph{Projective two-intersection sets}.
Projective two-intersection sets (sometimes called two-character sets) are classical configurations from finite geometry that provide a series of constructions \cite{A,ACK,BL,CHOP,CDMPS,CoM,CoMa,CoMP,CoPe13,HP,IZZ,P} (see also \cite[Sect.~9]{Ma}).
Recently, projective two-intersection sets have been intensively studied in the context of intriguing sets in polar spaces (see \cite{BKLP,BLMX,BDMR,B,CP14,CP17,CP18,FMX15JCT}, for example). See \cite{C} for a recent survey of intriguing sets, and \cite[Sect.~4.5]{MWX} for a detailed account of the connection between these sets and Latin square type and negative Latin square type PDSs.
\item \emph{Nondegenerate quadratic forms and their generalizations}.
A fundamental construction employs nondegenerate quadratic forms over finite fields \cite[Thm.~2.6]{Ma}. This construction has been greatly extended to allow the quadratic form to be replaced by other functions, including bent and vectorial bent functions \cite{CGT,CM,CTZ,CP13,FWXY,OP,TPF}.
\item \emph{Cyclotomic classes of a finite field}.
A seminal construction due to Baumert, Mills and Ward involves uniform cyclotomy in a finite field \cite{BMW}, \cite[Thm.~10.3]{Ma}. Further cyclotomic constructions have been proposed, based on sophisticated number-theoretic techniques \cite{M14,M18,MO,MX14,MX18,O16,O16JG}.
\end{enumerate}
We now describe the five construction classes that are not restricted to elementary abelian groups.
We shall later use LP-packings and NLP-packings to streamline many of these constructions, using only elementary methods. In some cases, we shall give a unified treatment for previously known parameter sets; in other cases, we shall produce examples with entirely new parameter sets.
\begin{enumerate}[(1)]
\item \emph{Direct constructions via partial congruence partitions}
\begin{definition}\label{def-PCP}
Let $G$ be a group of order~$n^2>1$. An \emph{$(n,r)$ partial congruence partition of degree~$r$ in $G$} is a collection $\{U_1,\dots,U_r\}$ of order $n$ subgroups of $G$ such that $U_i \cap U_j=\{1_G\}$ for all $i \ne j$.
\end{definition}
If $U_1, \dots, U_r$ is an $(n,r)$ partial congruence partition in $G$, then $\sum_{i=1}^r (U_i-1_G)$ is a regular $(n,r)$ Latin square type partial difference set in $G$ \cite[Sect.~4.1]{BJ}.
A central objective in the study of $(n,r)$ partial congruence partitions in groups of order~$n^2$ is to determine the largest possible degree~$r$ \cite{BJ,J89a,J89b}. The following result determines this value for abelian $p$-groups of the form~$H \times H$.
\begin{result}[{\cite[Thm.~2.7]{BJ}}]\label{res-PCPPDS}
Let $s_1,\dots,s_v$ be nonnegative integers and let $m = \min \{s_i \mid s_i > 0\}$. Let $p$ be prime and let $G=\prod_{i=1}^v\Z_{p^i}^{2s_i}$ and let $n = \sqrt{|G|}$.
Then the largest integer~$r$ for which there exists an $(n,r)$ partial congruence partition in $G$ is~$p^m+1$.
Therefore there exists a regular $(n,r)$ Latin square type PDS in $G$ for each positive integer $r \le p^m+1$.
\end{result}
We shall strengthen Result~\ref{res-PCPPDS} using Corollary~\ref{coro-LPpackingeven}: see Remark~\ref{rem-PDSeven}~$(i)$.
\item \emph{Direct constructions via finite local rings}
For positive integers $p$ and $w$, define the function
\begin{equation}\label{eqn-s}
\s(p,w)=\max_{\ell \mid w} \ell \left(\floor{\frac{w/\ell-2}{p}}+1\right),
\end{equation}
which determines the parameters of Latin square type partial difference sets in $\Z_{p^2}^w$ under the construction of \cite[Corollary 3.4]{HLX}. Various direct constructions of Latin square type partial difference sets exploit the structure of finite local rings \cite{CRX,DX04,H02,H03,HLX,HN,LM90,LM96,P02,RX}. We summarize below the constructions from \cite{CRX,H02,H03,HLX,HN,LM90,LM96,RX}, which we shall strengthen using Corollaries~\ref{coro-LPpackingeven} and~\ref{coro-LPpackingodd}:
see Remarks~\ref{rem-PDSeven}~$(ii)$ and~\ref{rem-PDSodd}.
\begin{result}\label{res-ringPDS}
Let $p$ be prime.
There exists an $(n,r)$ Latin square type PDS in a group $G$ of order $n^2$ for the following $G$ and $r$, where $r \le n$.
\begin{enumerate}[(i)]
\item $G=\Z_{p^2}^w$ for $w > 1$, and $r$ is a positive integer multiple of $\frac{n}{p^{\s(p,w)}}$ {\rm (\cite[Thm.~3.2, Cor.~3.4]{HLX})}.
\item $G=\Z_{p^2}^w \times \Z_{p}^{2uw}$ for $w > 1$ and a positive integer~$u$, and $r$ is a positive integer multiple of $\frac{n}{p^{\s(p,w)}}$ {\rm (\cite[Thm.~3.1]{H02})}.
\item $G = \prod_{i=1}^v \Z_{p^i}^{2s_i}$ for nonnegative integers $s_1,\dots,s_v$, and $r$ is a positive integer multiple of $\frac{n}{p^{\,\gcd(s_1,\ldots,s_v)}}$ {\rm (\cite[Thm.~3.4, Cor.~3.5]{H03})}.
\end{enumerate}
\end{result}
\item \emph{Lifting constructions via finite local rings}
Finite local rings also provide constructions for Latin square type partial difference sets that lift examples from a group $G \times G$ to a larger group $G' \times G'$ \cite{H00,H03,HN}.
We shall present in Theorem~\ref{thm-composition} a ring-free lifting construction that applies to a collection of Latin square type partial difference sets, rather than just a single one.
This construction leads to the result stated in Theorem~\ref{thm-preview}, which recovers the parameters of partial difference sets constructed via lifting in each of \cite{H00}, \cite[Sects.~4,~5]{H03}, and \cite{HN}.
\item \emph{Product constructions}
Various delicate product constructions for Latin square type and negative Latin square type partial difference sets have been found \cite{M14,P08EJC,P09JCD,P19,PDS,P08DCC}. The following examples occur in $p$-groups having arbitrarily large exponent.
\begin{result}[{\cite[Cor.~6.2]{P08DCC}}]\label{res-PDSproductLP}
Let $p$ be prime, and let $u$ and $s_2,s_4,\dots,s_{2v}$ be nonnegative integers.
Let $G = \Z_p^{2u} \times \prod_{i=1}^v\Z_{p^{2i}}^{2s_{2i}}$ and let $n = \sqrt{|G|}$.
Then there is a partition of $G-1_G$ into $p$ regular partial difference sets in~$G$, of which $p-1$ are of $(n,\frac{n}{p})$ Latin square type and one is of $(n,\frac{n}{p}+1)$ Latin square type.
\end{result}
We shall strengthen Result~\ref{res-PDSproductLP} using Corollary~\ref{coro-LPpackingeven} (see Remark~\ref{rem-PDSeven}~$(v)$).
\begin{result}[{\cite[Thm.~4.2, Cor.~5.1]{P08DCC}}]\label{res-PDSproductNLP}
\quad
\begin{enumerate}[$(i)$]
\item
Let $u,w$ and $s_4,s_6,\dots,s_{2v}$ be nonnegative integers satisfying $u+w \ge 1$.
Let $G = \Z_2^{4u} \times \Z_4^{2w} \times \prod_{i=2}^v \Z_{2^{2i}}^{4s_{2i}}$ and let $n = \sqrt{|G|}$.
Then there is a partition of $G -1_G$ into $4$ regular partial difference sets in~$G$,
of which three are of $(n,\frac{n}{4})$ negative Latin square type and one is of $(n,\frac{n}{4}-1)$ negative Latin square type.
\item
Let $u$ and $s_2,s_4,\dots,s_{2v}$ be nonnegative integers.
Let $G = \Z_3^{2u+2} \times \prod_{i=1}^v \Z_{3^{2i}}^{2s_{2i}}$ and let $n = \sqrt{|G|}$.
Then there is a partition of $G -1_G$ into $3$ regular partial difference sets in~$G$, of which two are of $(n,\frac{n}{3})$ negative Latin square type and one is of $(n,\frac{n}{3}-1)$ negative Latin square type.
\end{enumerate}
\end{result}
We shall strengthen Result~\ref{res-PDSproductNLP} using Corollary~\ref{coro-NLPpacking} (see Remark~\ref{rem-PDSproduct}).
\item \emph{Constructions in Galois domains}
A final construction class uses cyclotomy \cite{FKM} and character sums \cite{M21,O16DCC} over Galois domains (direct products of finite fields) to produce partial difference sets in the direct product of elementary abelian groups.
\end{enumerate}
\section{LP-Packings}\label{sec3}
In this section, we introduce a $(c,t)$ LP-packing as a collection of $t$ disjoint regular $(tc,c)$ Latin square type PDSs, in an abelian group of order $t^2c^2$, whose union satisfies an additional property. We provide examples of LP-packings, describe their relationship to partial congruence partitions, establish constraints on their structure, and present a product construction.
A principal motivation for this paper was the realization that in many previous constructions of such collections,
possibly after some modification as summarized in Table~\ref{tab-LPpacking}, the $tc$-set $G-\sum_{i=1}^t P_i$ of elements avoided by the partial difference sets forms a subgroup $U$ of~$G$.
Our definition of a $(c,t)$ LP-packing is based on this observation.
\begin{definition}[LP-packing]\label{def-LPpacking}
Let $t>1$ and $c>0$ be integers. Let $G$ be an abelian group of order~$t^2c^2$, and let $U$ be a subgroup of $G$ of order~$tc$. A $(c,t)$ \emph{LP-packing in $G$ relative to $U$} is a collection $\{P_1,\dots,P_t\}$ of $t$ regular $(tc,c)$ Latin square type PDSs in $G$ for which $\sum_{i=1}^t P_i = G-U$.
\end{definition}
\begin{remark}
In Definition~\ref{def-LPpacking}, each $P_i$ is a $c(tc-1)$-subset of $G$, and $|G-U| = tc(tc-1)$, so the relation $\sum_{i=1}^t P_i = G-U$ is equivalent to the statement that the $P_i$ are disjoint and their union is~$G-U$.
\end{remark}
We next give a characterization of a $(c,t)$ LP-packing involving character sums.
\begin{lemma}\label{lem-LPpacking}
Let $P_1,\dots,P_t$ be $c(tc-1)$-subsets of an abelian group $G$ of order~$t^2c^2$ and let $U$ be a subgroup of $G$ of order~$tc$.
Then $\{P_1,\dots,P_t\}$ is a $(c,t)$ LP-packing in $G$ relative to~$U$ if and only if the multiset equality
\begin{equation}
\label{eqn-U}
\{\chi(P_1),\dots,\chi(P_t)\} = \begin{cases}
\{-c,\dots, -c\} & \mbox{if $\chi \in U^\perp$,} \\
\{(t-1)c, -c, \dots, -c\} & \mbox{if $\chi \notin U^\perp$}
\end{cases}
\end{equation}
holds for all nonprincipal characters $\chi$ of~$G$.
\end{lemma}
\begin{proof}
Let $\chi$ be a nonprincipal character of~$G$.
Suppose firstly that $\{P_1,\dots,P_t\}$ is a $(c,t)$ LP-packing in $G$ relative to~$U$. Apply $\chi$ to the equation $\sum_{i=1}^t P_i = G-U$ to give
\[
\sum_{i=1}^t \chi(P_i) = \begin{cases}
-tc & \mbox{if $\chi \in U^\perp$,} \\
0 & \mbox{if $\chi \notin U^\perp$.}
\end{cases}
\]
Since $\chi(P_i) \in \{-c, (t-1)c\}$ for each $i$ by Lemma~\ref{lem-LatinPDS}, this implies that
\[
\{\chi(P_1),\dots,\chi(P_t)\} = \begin{cases}
\{-c,\dots, -c\} & \mbox{if $\chi \in U^\perp$,} \\
\{(t-1)c, -c, \dots, -c\} & \mbox{if $\chi \notin U^\perp$},
\end{cases}
\]
as required.
Now suppose that \eqref{eqn-U} holds. Then each $P_i$ is a $c(tc-1)$-subset of $G$ for which $\chi(P_i) \in \{-c, (t-1)c\}$, and so by Lemma~\ref{lem-LatinPDS} is a regular $(tc,c)$ Latin square type PDS in~$G$. It remains to show that $\sum_{i=1}^t P_i = G-U$. We have from $|P_i|=c(tc-1)$ and \eqref{eqn-U} that, for all $\chi \in \wh{G}$,
\[
\chi \Big(\sum_{i=1}^t P_i\Big) = \begin{cases}
tc(tc-1) & \mbox{if $\chi \in G^\perp$}, \\
-tc & \mbox{if $\chi \in U^\perp \sm G^\perp$,} \\
0 & \mbox{if $\chi \notin U^\perp$.}
\end{cases}
\]
By the Fourier inversion formula (Proposition~\ref{prop-fourier}), this implies that
$\sum_{i=1}^t P_i = G-U$, as required.
\end{proof}
A $(t,t+1)$ partial congruence partition (see Definition~\ref{def-PCP}) lies at the heart of many configurations {\rm\cite{Dillon}, \cite[Sect.~4]{ES}}. We next show that this structure can equivalently be expressed as a $(1,t)$ LP-packing.
\begin{proposition}\label{prop-PCP}
Let $G$ be an abelian group of order $t^2$, let $U_1, \dots, U_t$ be $t$-subsets of~$G$, and let $U_0$ be a subgroup of $G$ of order~$t$. Then $\{ U_0,U_1,\dots,U_t\}$ is a $(t,t+1)$ partial congruence partition in $G$ if and only if
$\{ U_1-1_G,\dots,U_t-1_G\}$ is a $(1,t)$ LP-packing in $G$ relative to~$U_0$.
\end{proposition}
\begin{proof}
By Definition~\ref{def-PCP},
$\{ U_0,U_1,\dots,U_t\}$ is a $(t,t+1)$ partial congruence partition in $G$ if and only if each of $U_1,\dots,U_t$ is a subgroup of $G$ and $\sum_{i=0}^t U_i = G + t 1_G$.
By Definitions~\ref{def-LatinPDS} and~\ref{def-LPpacking},
$\{ U_1-1_G,\dots,U_t-1_G\}$ is a $(1,t)$ LP-packing in $G$ relative to~$U_0$
if and only if each of $U_1-1_G,\dots,U_t-1_G$ is a regular $(t^2,t-1,t-2,0)$ PDS in $G$ and $\sum_{i=1}^t (U_i-1_G) = G-U_0$. It is therefore sufficient to show for $i=1,\dots,t$ that $U_i$ is a subgroup of $G$ if and only if $U_i-1_G$ is a regular $(t^2,t-1,t-2,0)$ partial difference set in~$G$. This follows directly from Definition~\ref{def-pds}.
\end{proof}
The following result greatly constrains the degree of a partial congruence partition.
\begin{result}\label{res-PCPdegree}
Suppose there exists an $(n,r)$ partial congruence partition of degree $r$ in an abelian group $G$ of order~$n^2$. Then $r \le n+1$, and
\begin{enumerate}[(i)]
\item
if $r \ge \floor{\sqrt{n}}+2$ then $G$ is elementary abelian {\rm(\cite[Thms.~3.1,~3.4]{J89a})}.
\item
if $r=\sqrt{n}+1$ then either $G$ is elementary abelian or $G=(\Z_{p^2})^w$ for some prime $p$ and positive integer~$w$ {\rm(\cite[Cor.~5.3]{J89b})}.
\end{enumerate}
\end{result}
By Result~\ref{res-PCPdegree}~$(i)$, a $(t,t+1)$ partial congruence partition in an abelian group~$G$ can exist only when $t = p^s$ for some prime $p$ and positive integer~$s$, and $G \cong \Z_p^{2s}$.
By Proposition~\ref{prop-PCP}, such a structure exists if and only if there is a $(1,p^s)$ LP-packing in $G$ relative to an order $p^s$ subgroup.
We now show, in an example we shall refer to frequently in the rest of the paper, that this structure exists for all primes $p$ and positive integers~$s$ by reference to a spread of a vector space over a finite field.
\begin{definition}\label{def-spread}
Let $p$ be prime and $s$ a positive integer, and let $V$ be a $2s$-dimensional vector space over the field $\F_p$ with identity~$1_V$. A \emph{spread} of $V$ is a $(p^s+1)$-set $\{H_0,H_1,\dots,H_{p^s}\}$ of $s$-dimensional vector subspaces of $V$ such that $H_i \cap H_j = \{1_V\}$ for all $i \ne j$.
\end{definition}
\begin{example}\label{ex-spread}
Let $p$ be prime and $s$ a positive integer, and let $V$ be a $2s$-dimensional vector space over the field $\F_p$ with identity~$1_V$.
Then there exists a spread $\{H_0, H_1, \dots, H_{p^s}\}$ of $V$ \cite[Chap.~VII, Sect.~5]{HP73}.
By Definitions~\ref{def-PCP} and~\ref{def-spread},
$\{ H_0, \ldots, H_{p^s} \}$ forms a $(p^s,p^s+1)$ partial congruence partition in~$V$, because $H_i$ is a subgroup of order $p^s$ in $V \cong \Z_p^{2s}$ if and only if $H_i$ is an $s$-dimensional vector subspace of~$V$.
Then by Proposition~\ref{prop-PCP} with $t=p^s$, we see that
$\{ H_1-1_V, \ldots, H_{p^s}-1_V \}$ is a $(1,p^s)$ LP-packing in $V \cong \Z_{p}^{2s}$ relative to~$H_0 \cong \Z_p^s$,
and by Lemma~\ref{lem-LPpacking}
each nonprincipal character of $V$ is principal on exactly one of $H_0,\dots,H_{p^s}$.
\end{example}
\begin{remark}\label{rem-PCP}
According to Definition~\ref{def-LPpacking}, a $(c,t)$ LP-packing involves a collection of $t$ disjoint regular $(tc,c)$ Latin square type PDSs in an abelian group $G$ of order~$t^2c^2$. Since this collection covers all but $tc$ elements of $G$, and each PDS contains $c(tc-1)$ elements, the collection has maximum possible size when $c>1$. The collection is not necessarily of maximum possible size when $c=1$: for $p$ prime and $t=p^s$, the $(t,t+1)$ partial congruence partition of Example~\ref{ex-spread} is a collection of $t+1$ disjoint regular $(t,1)$ Latin square type PDSs in the elementary abelian group of order $t^2$.
However, as Result~\ref{res-PCPdegree} shows, if we instead attempt to create a partial congruence partition in a non-elementary abelian group of order $t^2$ then the degree must drop from $t+1$ to at most $\floor{\sqrt{t}}+1$, covering a proportion of only about $1/\sqrt{t}$ of the elements of~$G$.
In contrast, in Section~\ref{sec5} we shall provide constructions of $(c,t)$ LP-packings with $c>1$ in various nonelementary abelian groups.
For this reason, we regard a $(c,t)$ LP-packing with $c>1$ as a natural generalization of a $(t,t+1)$ partial congruence partition.
\end{remark}
The characterization in Lemma~\ref{lem-LPpacking}, together with Lemma~\ref{lem-LatinPDS}, shows that we can combine subsets from an LP-packing to obtain Latin square type PDSs with various parameters.
\begin{lemma}\label{lem-LPpackingtoPDS}
Suppose that $\{P_1, \ldots, P_t\}$ is a $(c,t)$ LP-packing in an abelian group $G$ of order $t^2c^2$ relative to a subgroup $U$ of order~$tc$. Let $I$ be a $b$-subset of $\{1,\dots,t\}$. Then
\begin{enumerate}[$(i)$]
\item $\sum_{i \in I} P_i$ is a regular $(tc,bc)$ Latin square type PDS in~$G$.
\item $\sum_{i \in I} P_i+U-1_G$ is a regular $(tc,bc+1)$ Latin square type PDS in~$G$.
\end{enumerate}
\end{lemma}
\begin{remark}\label{rem-LPpacking}
Suppose that $\{P_1, \ldots, P_t\}$ is a $(c,t)$ LP-packing in an abelian group $G$ of order $t^2c^2$ relative to a subgroup $U$ of order~$tc$.
\begin{enumerate}[(i)]
\item
By Definition~\ref{def-LPpacking} and Lemma~\ref{lem-LPpackingtoPDS}, we see that
$\{P_1,\dots,P_{t-1},P_t+U-1_G\}$ is a partition of $G-1_G$ into $t$ regular partial difference sets in~$G$, of which the first $t-1$ are of $(tc,c)$ Latin square type and the last is of $(tc,c+1)$ Latin square type.
\item
For each $i$, let $\wti{{P\mkern 0mu}_i}
= \{\chi \in \wh{G} \mid \chi(P_i) = (t-1)c\}$.
By Lemma~\ref{lem-LatinPDS}, each $\wti{{P\mkern 0mu}_i}$ is a regular $(tc,c)$ Latin square type PDS in~$\wh{G}$, and by Lemma~\ref{lem-LPpacking} each $\chi \in \wh{G} \sm U^\perp$ belongs to exactly one~$\wti{{P\mkern 0mu}_i}$ and each $\chi \in U^\perp$ belongs to no~$\wti{{P\mkern 0mu}_i}$. By Definition~\ref{def-LPpacking}, the collection $\{\wti{P_1}, \ldots, \wti{{P\mkern 0mu}_t}\}$
is therefore a $(c,t)$ LP-packing in $\wh{G}$ relative to~$U^{\perp}$.
\end{enumerate}
\end{remark}
Combining the subsets of an LP-packing into equally-sized collections gives an LP-packing with fewer subsets.
\begin{lemma}\label{lem-LPpackingunion}
Suppose there exists a $(c,t)$ LP-packing in an abelian group $G$ of order $t^2c^2$ relative to a subgroup $U$ of order~$tc$, and suppose $s$ divides~$t$. Then there exists an $(sc,\frac{t}{s})$ LP-packing in $G$ relative to~$U$.
\end{lemma}
\begin{proof}
Let $\{P_1, \dots, P_t\}$ be a $(c,t)$ LP-packing in $G$ relative to~$U$, and let
\[
P'_i = P_{is+1} + P_{is+2} + \dots + P_{is+s} \quad \mbox{for $i = 0, 1,\dots, \frac{t}{s}-1$}.
\]
Then Lemma~\ref{lem-LPpacking} shows that $\{P'_0,\dots,P'_{\frac{t}{s}-1}\}$ is an $(sc,\frac{t}{s})$ LP-packing in $G$ relative to~$U$.
\end{proof}
The following result, which is inspired by \cite[Lemma 3.2]{DP}, allows us to establish some constraints on LP-packings in Proposition~\ref{prop-LPpackingstruc}. These constraints assist in finding examples computationally, as demonstrated in Example~\ref{ex-LPpacking444}.
\begin{lemma}\label{lem-LSPDS}
Let $G$ be a group of order $n^2$ and let $U$ be a subgroup of order~$n$.
Let $\{g_0,g_1,\dots, $ $g_{n-1}\}$ be a complete set of right coset representatives for $U$ in $G$, where $Ug_0=U$. Suppose $P$ is a regular $(n,r)$ Latin square type PDS in $G$ for which $P \cap U = \es$. Then $|P \cap Ug_i|=r$ for each $i = 1, \dots, n-1$.
\end{lemma}
\begin{proof}
Since $P \cap U = \es$,
\begin{equation}\label{eqn-|P|}
r(n-1) = |P| = \sum_{i=1}^{n-1} |P\cap U g_i|.
\end{equation}
For all $x,y \in P$, we have
$x y^{-1} \in U$ if and only if $x,y$ belong to the same right coset of $U$.
Therefore
\begin{equation}\label{eqn-gUPP}
\sum_{i=1}^{n-1} |P \cap U g_i|^2 = \sum_{g \in U}[PP^{(-1)}]_g.
\end{equation}
Now $P$ is an $(n^2,r(n-1),n+r^2-3r,r^2-r)$ PDS in $G$, and so by \eqref{eqn-PDSgpring}
\[
PP^{(-1)} = r(n-r) + (n-2r)P + (r^2-r)G.
\]
Since $P \cap U = \es$, this implies that
\[
\sum_{g \in U} [PP^{(-1)}]_g = r(n-r) + (r^2-r)n = r^2(n-1).
\]
Substitution in \eqref{eqn-gUPP} then gives
\[
\sum_{i=1}^{n-1} |P \cap U g_i|^2 = r^2(n-1),
\]
which, by the Cauchy-Schwarz inequality and \eqref{eqn-|P|}, gives
$|P \cap Ug_i|=r$ for each $i=1,\dots,n-1$ as required.
\end{proof}
\begin{proposition}\label{prop-LPpackingstruc}
Let $G$ be an abelian group of order $t^2c^2$, and let $U$ be a subgroup of $G$ of order~$tc$.
Let $\{g_0,g_1,\dots,g_{tc-1}\}$ be a complete set of right coset representatives for $U$ in $G$, where $Ug_0=U$, and define subsets $S,N$ of $\{1,\dots,tc-1\}$ by
\begin{align*}
& S = \{j \mid Ug_j = Ug_j^{-1}\}, \\
& \mbox{$N$ contains exactly one of $j,k$ when $Ug_j \ne Ug_j^{-1} = Ug_k$}.
\end{align*}
Suppose that $\{P_1,\dots,P_t\}$ is a $(c,t)$ LP-packing in $G$ relative to~$U$.
Then there are $c$-subsets $H_{ij}$ of $U$ for $i=1,\dots,t$ and $j=1,\dots,tc-1$ satisfying
\begin{align*}
P_i &=\sum_{j \in S} H_{ij}g_j + \sum_{j \in N} \big(H_{ij}g_j + H_{ij}^{(-1)}g_j^{-1}\big) \quad \mbox{for each $i$}, \\
H_{ij}^{(-1)}g_j^{-1} &= H_{ij}g_j \quad \mbox{for $j \in S$}, \\
H_{ij}^{(-1)}g_j^{-1} &= H_{ik}g_k \quad \mbox{for $j \in N$ where $Ug_j^{-1} = Ug_k$}, \\
U &= \sum_{i=1}^t H_{ij} \quad \mbox{for each~$j$}.
\end{align*}
\end{proposition}
\begin{proof}
By Definition~\ref{def-LPpacking}, each $P_i$ is a regular $(tc,c)$ Latin square type PDS in $G$, and $G-U$ is the disjoint union of the~$P_i$.
Write
\begin{align}
P_i &= \sum_{j=1}^{tc-1} (P_i \cap Ug_j) \nonumber \\
&= \sum_{j \in S} (P_i \cap Ug_j) +
\sum_{j \in N} \big((P_i \cap Ug_j) + (P_i \cap Ug_j^{-1})\big). \label{eqn-PiUgj}
\end{align}
By Lemma~\ref{lem-LSPDS}, for each $i = 1,\dots, t$ and each $j = 1,\dots,tc-1$ we may write
\[
P_i \cap U g_j = H_{ij} g_j \quad \mbox{for some $c$-subset $H_{ij}$ of~$U$}.
\]
Substitute in \eqref{eqn-PiUgj} and use $P_i = P_i^{(-1)}$ to give
\[
P_i = \sum_{j \in S} H_{ij} g_j + \sum_{j \in N} \big(H_{ij}g_j + H_{ij}^{(-1)}g_j^{-1}\big),
\]
and the constraints
$H_{ij}^{(-1)}g_j^{-1} = H_{ij}g_j$ for $j \in S$, and
$H_{ij}^{(-1)}g_j^{-1} = H_{ik}g_k$ for $j \in N$ where $Ug_j^{-1} = Ug_k$.
For each $j$, we have $Ug_j = \sum_{i=1}^t (P_i \cap Ug_j) = \sum_{i=1}^t H_{ij} g_j$ and therefore $U = \sum_{i=1}^t H_{ij}$.
\end{proof}
\begin{example}\label{ex-LPpacking444}
Using Proposition~\ref{prop-LPpackingstruc}, we obtain the following exhaustive search results
for LP-packings in $\Z_4^3$ relative to two nonisomorphic order~8 subgroups.
\begin{enumerate}[$(i)$]
\item
There are exactly $1536$ distinct $(4,2)$ LP-packings $\{P_1,P_2\}$ in $\Z_4^3=\lan x,y,z \ran$ relative to $(2\Z_4) \times (2\Z_4) \times (2\Z_4)=\lan x^2,y^2,z^2 \ran$, one of which is
\begin{align*}
P_1=&(1+x^2+y^2+x^2y^2)x+(x^2+x^2y^2+z^2+y^2z^2)y+(x^2+x^2z^2+y^2+y^2z^2)z\\
&+(y^2+z^2+x^2+x^2y^2z^2)xy+(1+x^2+z^2+x^2z^2)xz \\
&+(x^2+x^2y^2+x^2z^2+x^2y^2z^2)yz+(y^2+z^2+x^2y^2+x^2z^2)xyz, \\
P_2=&\lan x,y,z \ran-\lan x^2,y^2,z^2 \ran-P_1.
\end{align*}
The presented form of $P_1$ shows that both $P_1$ and $P_2$ intersect the seven nonidentity cosets of the subgroup $\lan x^2,y^2,z^2 \ran$ in $\lan x,y,z \ran$ in exactly 4 elements.
\item
There are exactly $512$ distinct $(4,2)$ LP-packings $\{R_1,R_2\}$ in $\Z_4^3=\lan x,y,z \ran$ relative to $\Z_4 \times (2\Z_4)=\lan x,y^2 \ran$, one of which is
\begin{align*}
R_1=&(y^2+xy^2+x^2y^2+x^3)z+(x+xy^2+x^3+x^3y^2)z^2+(y^2+x+x^2y^2+x^3y^2)z^3\\
&+(x+x^2+x^2y^2+x^3y^2)y+(1+x^2+x^3+x^3y^2)yz \\
&+(x+x^2+x^2y^2+x^3y^2)yz^2+(1+x+xy^2+x^2)y^3z^3, \\
R_2=&\lan x,y,z \ran-\lan x,y^2 \ran-R_1.
\end{align*}
The presented form of $R_1$ shows that both $R_1$ and $R_2$ intersect the seven nonidentity cosets of the subgroup $\lan x,y^2 \ran$ in $\lan x,y,z \ran$ in exactly 4 elements.
\end{enumerate}
\end{example}
We conclude this section with a product construction for LP-packings.
\begin{theorem}\label{thm-LPpackingproduct}
For $j=1,2$, suppose there exists a $(c_j,t)$ LP-packing in an abelian group $G_j$ of order $t^2c_j^2$ relative to a subgroup $U_j$ of order~$tc_j$. Then there exists a $(tc_1c_2,t)$ LP-packing in $G_1 \times G_2$ relative to $U_1 \times U_2$.
\end{theorem}
\begin{proof}
For $j=1,2$, let $\{P_{j,0},\dots,P_{j,t-1}\}$ be a $(c_j,t)$ LP-packing in $G_j$ relative to~$U_j$ and define
\[
K_\ell = P_{1,\ell} U_2 + U_1 P_{2,\ell} + \sum_{i=0}^{t-1} P_{1,i}P_{2,i+\ell} \quad \mbox{for $\ell = 0,\dots,t-1$},
\]
where the subscript $i+\ell$ is reduced modulo~$t$. We shall use Lemma~\ref{lem-LPpacking} to show that $\{K_0,K_1,\dots,$ $K_{t-1}\}$ is a $(tc_1c_2,t)$ LP-packing in $G_1 \times G_2$ relative to $U_1 \times U_2$.
Each $P_{j,i}$ is a $c_j(tc_j-1)$-subset of $G_j-U_j$, so each $K_\ell$ is a subset of $G_1 \times G_2$ of size $c_1(tc_1-1)tc_2+tc_1 c_2(tc_2-1) + \sum_{i=0}^{t-1} c_1(tc_1-1)c_2(tc_2-1) = tc_1c_2(t^2c_1c_2-1)$. Let $\chi$ be a nonprincipal character of $G_1 \times G_2$, and let $\chi_j=\chi|_{G_j}$ for $j=1,2$. Then
\[
\chi(K_\ell) = \chi_1(P_{1,\ell}) \chi_2(U_2) + \chi_1(U_1) \chi_2(P_{2,\ell}) + \sum_{i=0}^{t-1} \chi_1(P_{1,i})\chi_2(P_{2,i+\ell}) \quad \mbox{for $\ell = 0,\dots,t-1$}.
\]
By Lemma~\ref{lem-LPpacking}, it remains to prove that
\begin{equation}\label{eqn-LPprod}
\{\chi(K_0),\dots,\chi(K_{t-1})\} = \begin{cases}
\{-tc_1c_2,\dots, -tc_1c_2\} & \mbox{if $\chi \in (U_1 \times U_2)^\perp$,} \\
\{t(t-1)c_1c_2, -tc_1c_2, \dots, -tc_1c_2\} & \mbox{if $\chi \notin (U_1 \times U_2)^\perp$.}
\end{cases}
\end{equation}
For $j=1,2$, by Lemma~\ref{lem-LPpacking} we are given that
\[
\{\chi_j(P_{j,0}),\dots,\chi_j(P_{j,t-1})\} = \begin{cases}
\{c_j(tc_j-1),\dots, c_j(tc_j-1)\} & \mbox{if $\chi_j \in G_j^\perp$,} \\
\{-c_j,\dots, -c_j\} & \mbox{if $\chi_j \in U_j^\perp \sm G_j^\perp$,} \\
\{(t-1)c_j, -c_j, \dots, -c_j\} & \mbox{if $\chi_j \notin U_j^\perp$.}
\end{cases}
\]
Since $\chi$ is nonprincipal on $G_1 \times G_2$, we may assume by symmetry that $\chi_2$ is nonprincipal on $G_2$. Again by symmetry, we need to consider only the following three cases, which together establish~\eqref{eqn-LPprod}.
\begin{description}
\item[{\bf Case 1:}] $\chi_1$ is principal on $U_1$ and $\chi_2$ is principal on $U_2$. \\
Then $\chi_1(P_{1,i})$ is constant over $i$ (regardless of whether $\chi_1$ is principal or nonprincipal on $G_1$), so
$\chi(K_\ell) = \chi_1(P_{1,\ell}) tc_2 + tc_1 (-c_2) + \chi_1(P_{1,\ell}) \sum_{i=0}^{t-1} (-c_2) = -tc_1c_2$ for each~$\ell$.
\item[{\bf Case 2:}] $\chi_1$ is principal on $U_1$ and $\chi_2$ is nonprincipal on $U_2$. \\
Then $\chi_1(P_{1,i})$ is again constant over $i$ and $\sum_{i=0}^{t-1} \chi_2(P_{2,i}) = 0$, so
$\chi(K_\ell) = tc_1 \chi_2(P_{2,\ell})$ which gives
$\{\chi(K_0),\dots,\chi(K_{t-1})\} = \{t(t-1)c_1c_2,-tc_1c_2,\dots,-tc_1c_2\}$.
\item[{\bf Case 3:}] $\chi_1$ is nonprincipal on $U_1$ and $\chi_2$ is nonprincipal on $U_2$. \\
Then
$\chi(K_\ell) = \sum_{i=0}^{t-1} \chi_1(P_{1,i}) \chi_2(P_{2,i+\ell})$,
so the multiset $\{\chi(K_0),\dots,\chi(K_{t-1})\}$ contains
one occurrence of the value $(t-1)c_1(t-1)c_2 + (t-1)(-c_1)(-c_2) = t(t-1)c_1c_2$, and
$t-1$ occurrences of the value $(t-1)c_1(-c_2) + (-c_1)(t-1)c_2 + (t-2)(-c_1)(-c_2) = -tc_1c_2$.
\end{description}
\end{proof}
\begin{remark}
The construction of Theorem~\ref{thm-LPpackingproduct} is closely modelled on the product construction for collections of Latin square type PDSs presented by Polhill in {\rm \cite[Thm.~2.2]{P08DCC}}.
Indeed, the configuration described in {\rm \cite[Lemma 2.1]{P08DCC}} for $e=1$ can be represented as $\{P_1,\dots,P_{t-1},P_t+U-1_G\}$, where $\{ P_1, \dots, P_t\}$ is a $(c,t)$ LP-packing in $G$ relative to~$U$ (see also Remark~\ref{rem-LPpacking}~$(i)$).
Our construction has the advantage that, by identifying the role of the avoided subgroup $U$ in the definition of an LP-packing, we are able to control the avoided subgroup $U_1 \times U_2$ in Theorem~\ref{thm-LPpackingproduct}.
\end{remark}
We extend the construction of Theorem~\ref{thm-LPpackingproduct} using Lemma~\ref{lem-LPpackingunion}.
\begin{corollary}\label{coro-LPpackingproduct}
For $j=1,2$, suppose there exists a $(c_j,t_j)$ LP-packing in an abelian group $G_j$ of order $t_j^2c_j^2$ relative to a subgroup $U_j$ of order~$t_jc_j$, and suppose that $t_1$ divides~$t_2$. Then there exists a $(t_2c_1c_2,t_1)$ LP-packing in $G_1 \times G_2$ relative to $U_1 \times U_2$.
\end{corollary}
\begin{proof}
By Lemma~\ref{lem-LPpackingunion} with $s=\frac{t_2}{t_1}$, there exists a $(\frac{t_2}{t_1}c_2,t_1)$ LP-packing in $G_2$ relative to $U_2$. Then by Theorem~\ref{thm-LPpackingproduct}, there exists a $(t_2c_1c_2,t_1)$ LP-packing in $G_1 \times G_2$ relative to $U_1 \times U_2$.
\end{proof}
\section{LP-partitions}\label{sec4}
In this section, we introduce a $(c,t)$ LP-partition as an auxiliary configuration in the construction of an LP-packing in a larger group from an LP-packing in a smaller group. We then show how to construct an LP-partition from a collection of LP-partitions in various factor groups. We shall apply these two constructions recursively in Section~\ref{sec5} to produce infinite families of LP-packings.
\begin{definition}[LP-partition]\label{def-LPpartition}
Let $t>1$ and $c>0$ be integers. Let $G$ be an abelian group of order~$t^2c^2$, let $V$ be a subgroup of $G$ of order~$tc^2$, and let $H \leqslant V$.
A \emph{$(c,t)$ LP-partition in $G-V$ relative to $H$} is a collection $\{R_1, \ldots, R_t\}$ of $t$ disjoint $(t-1)c^2$-subsets of $G$ whose union is $G-V$ and for which the multiset equality
\[
\{\chi(R_1), \dots, \chi(R_t)\}= \begin{cases}
\{-c^2,\dots, -c^2\} & \mbox{if $\chi \in V^{\perp}$,} \\
\{0,\dots,0\} & \mbox{if $\chi \in H^{\perp} \sm V^{\perp}$,} \\
\{(t-1)c,-c,\dots,-c\} & \mbox{if $\chi \notin H^{\perp}$}.
\end{cases}
\]
holds for all nonprincipal characters $\chi$ of~$G$.
\end{definition}
\begin{remark}\label{rem-LPpartitiondefn}
\quad
\begin{enumerate}[$(i)$]
\item
In Definition~\ref{def-LPpartition}, we can deduce that the $R_i$ are disjoint and their union is $G-V$ from the other conditions, by applying the Fourier inversion formula to the equation
\[
\chi \Big(\sum_{i=1}^t R_i\Big) = \begin{cases}
t(t-1)c^2 & \mbox{if $\chi \in G^\perp$}, \\
-tc^2 & \mbox{if $\chi \in V^\perp \sm G^\perp$}, \\
0 & \mbox{if $\chi \notin V^\perp$}.
\end{cases}
\]
\item
Each subset of a $(c,t)$ LP-partition in $G-V$ relative to $H$ is a $(G,V,H;c)$ Latin shell, as introduced in {\rm\cite[Defn.~3.1]{HN}} as an auxiliary configuration in the construction of a single Latin square type PDS.
\end{enumerate}
\end{remark}
A simple example of an LP-partition is given by a spread of a vector space (see Definition~\ref{def-spread}).
\begin{example}\label{ex-spread2}
By Lemma~\ref{lem-LPpacking} and Definition~\ref{def-LPpartition}, a $(1,t)$ LP-packing in $G$ relative to~$U$ is identical to a $(1,t)$ LP-partition in $G-U$ relative to~$U$.
Let $t=p^s$ for a prime $p$ and positive integer~$s$, and let $V$ be a $2s$-dimensional vector space over~$\F_p$ with identity~$1_V$.
Let $\{H_0,\dots,H_{p^s}\}$ be a spread of $V \cong \Z_p^{2s}$.
Following Example~\ref{ex-spread}, $\{H_1-1_V,\dots,H_{p^s}-1_V\}$ is therefore a $(1,p^s)$ LP-partition in $V-H_0$ relative to~$H_0$ as well as a $(1,p^s)$ LP-packing in $V$ relative to~$H_0$.
\end{example}
We next combine the subsets of an LP-partition with the subsets obtained by lifting an LP-packing, in order to produce an LP-packing in a larger group.
\begin{theorem}\label{thm-composition}
Let $G$ be an abelian group of order $t^4c^2$ containing subgroups $H \leqslant U \leqslant V \leqslant G$, where $H$, $U$, $V$ have order $t$, $t^2c$, $t^3c^2$, respectively.
Suppose there exists a $(tc,t)$ LP-partition in $G-V$ relative to $H$, and a $(c,t)$ LP-packing in $V/H$ relative to~$U/H$.
Then there exists a $(tc,t)$ LP-packing in $G$ relative to~$U$.
\end{theorem}
\begin{proof}
Let $\{R_1,\dots,R_t\}$ be a $(tc,t)$ LP-partition in $G-V$ relative to $H$, and let $\{P_1,\dots,P_t\}$ be a $(c,t)$ LP-packing in $V/H$ relative to~$U/H$.
For $i=1,\dots,t$, let $P'_i = \{g \in V \mid gH \in P_i\}$ be the pre-image of $P_i$ under the quotient mapping from $V$ to $V/H$. We shall use Lemma~\ref{lem-LPpacking} to show that $\{P'_1+R_1,\dots,P'_t+R_t\}$ is a $(tc,t)$ LP-packing in $G$ relative to~$U$.
Each $P_i$ is a $c(tc-1)$-subset of $V/H-U/H$ by Definition~\ref{def-LPpacking}, so
each $P'_i$ is a $tc(tc-1)$-subset of~$V-U$.
Since each $R_i$ is a $(t-1)(tc)^2$-subset of $G-V$ by Definition~\ref{def-LPpartition}, it follows that each $P'_i+R_i$ is a subset of $G$ of size
$|P'_i| + |R_i| = tc(t^2c-1)$.
By Lemma~\ref{lem-LPpacking}, for all characters $\psi$ of $V/H$ we have
\[
\{\psi(P_1),\dots,\psi(P_t)\}=
\begin{cases}
\{c(tc-1),\dots,c(tc-1)\} & \mbox{if $\psi \in (V/H)^{\perp}$,} \\
\{-c,\dots, -c\} & \mbox{if $\psi \in (U/H)^{\perp} \sm (V/H)^{\perp}$,} \\
\{(t-1)c, -c, \dots, -c\} & \mbox{if $\psi \notin (U/H)^{\perp}$.}
\end{cases}
\]
Let $\chi$ be a nonprincipal character of~$G$. If $\chi$ is principal on $H$ then it induces a nonprincipal character $\psi$ on~$G/H$. Since $P'_i$ is a union of cosets of the order~$t$ subgroup $H$, we obtain
\[
\chi(P'_i) = \begin{cases}
t\psi(P_i) & \mbox{if $\chi \in H^\perp$}, \\
0 & \mbox{if $\chi \notin H^\perp$}.
\end{cases}
\]
Therefore the value of $\{\chi(P'_1),\dots,\chi(P'_t)\}$ is as specified in the second column of the following table:
\[
\ra{1.5}
\scalebox{0.85}{
$
\begin{array}{|l|c|c|c|}
\cline{2-4}
\mc{1}{l|}{} & \{\chi(P'_1),\dots,\chi(P'_t)\} & \{\chi(R_1),\dots,\chi(R_t)\} & \{\chi(P'_1+R_1),\dots,\chi(P'_t+R_t)\} \\ \hline
\chi \in V^\perp \sm G^\perp & \{tc(tc-1),\dots,tc(tc-1)\} & \{-t^2c^2,\dots,-t^2c^2\} & \mr{2}{*}{$\{-tc,\dots,-tc\}$} \\ \cline{1-3}
\chi \in U^\perp \sm V^\perp & \{-tc,\dots,-tc\} & \mr{2}{*}{$\{0,\dots,0\}$} & \\ \cline{1-2} \cline{4-4}
\chi \in H^\perp \sm U^\perp & \{t(t-1)c,-tc,\dots,-tc\} & & \mr{2}{*}{$\{t(t-1)c,-tc,\dots,-tc\}$} \\ \cline{1-3}
\chi \notin H^\perp & \{0,\dots,0\} & \{t(t-1)c,-tc,\dots,-tc\} & \\ \hline
\end{array}
$
}
\]
By Definition~\ref{def-LPpartition}, the value of $\{\chi(R_1),\dots,\chi(R_t)\}$ is as given in the third column of the table, and the fourth column contains the sum of the second and third columns. Lemma~\ref{lem-LPpacking} then implies that
$\{P'_1+R_1,\dots,P'_t+R_t\}$ is a $(tc,t)$ LP-packing in $G$ relative to~$U$, as required.
\end{proof}
\begin{remark}\label{rem-lifting}
Lifting constructions similar to that of Theorem~\ref{thm-composition} were proposed in {\rm \cite{H00}, \cite[Sects.~4,~5]{H03}}, {\rm\cite{HN}}, each producing a single Latin square type partial difference set. These previous constructions make use of the delicate structure of finite local rings. In contrast, the construction of Theorem~\ref{thm-composition} applies simultaneously to a collection of such partial difference sets, and does not require ring theory.
\end{remark}
We now construct an LP-partition by lifting and combining the subsets of a collection of LP-partitions in various factor groups. The construction makes use of two spreads of a vector space over a finite field (as introduced in Definition~\ref{def-spread}).
\begin{theorem}\label{thm-LPpartitionrecursive}
Let $t=p^s$ for a prime $p$ and positive integer~$s$.
Let $G$ be an abelian group of order $t^4c^2$ containing subgroups $Q \leqslant G' \leqslant G$, where $Q \cong \Z_p^{2s}$ and $G/G' \cong \Z_p^{2s}$.
Let $H_0,\dots,H_t$ be subgroups of $G$ forming a spread when viewed as subgroups of~$Q$.
Let $V_0,\dots,V_t$ be subgroups of $G$ for which $\{V_0/G',\dots,V_t/G'\}$ is a spread of~$G/G'$, and for which $H_i \leqslant V_i$ for each $i=1,\dots,t$.
Suppose that for each $i = 1,\dots,t$ there exists a $(c,t)$ LP-partition in $V_i/H_i - G'/H_i$ relative to $Q/H_i$.
Then there exists a $(tc,t)$ LP-partition in $G-V_0$ relative to~$H_0$.
\end{theorem}
\begin{proof}
For each $i = 1, \dots, t$, let $\{ S_{i1},\dots,S_{it}\}$ be a $(c,t)$ LP-partition in $V_i/H_i-G'/H_i$ relative to~$Q/H_i$. By Definition~\ref{def-LPpartition}, for each $i=1,\dots,t$ and for all nonprincipal characters $\psi_i$ of $V_i/H_i$ we have
\[
\{\psi_i(S_{i1}),\dots,\psi_i(S_{it})\}= \begin{cases}
\{-c^2,\dots, -c^2\} & \mbox{if $\psi_i \in (G'/H_i)^{\perp}$}, \\
\{0,\dots,0\} & \mbox{if $\psi_i \in (Q/H_i)^{\perp} \sm (G'/H_i)^{\perp}$,} \\
\{(t-1)c, -c, \dots, -c\} & \mbox{if $\psi_i \notin (Q/H_i)^{\perp}$.}
\end{cases}
\]
For each $i,j$ satisfying $1 \le i, j \le t$, let
$S'_{ij}=\{g \in V_i \mid gH_i \in S_{ij}\}$
be the pre-image of $S_{ij}$ under the quotient mapping from $V_i$ to~$V_i/H_i$, and let $R_j=\sum_{i=1}^t S'_{ij}$ for $j = 1,\dots,t$.
We shall show that $\{R_1,\dots,R_t\}$ is a $(tc,t)$ LP-partition in $G-V_0$ relative to~$H_0$.
By Definition~\ref{def-LPpartition},
each $S_{ij}$ is a $(t-1)c^2$-subset of $V_i/H_i-G'/H_i$, and so
each $S'_{ij}$ is a $t(t-1)c^2$-subset of~$V_i-G'$.
By Example~\ref{ex-spread}, the $V_i/G'-1_{G/G'}$ are disjoint in $G/G'$ and so
the $V_i-G'$ are disjoint in~$G$.
Therefore each $R_j = \sum_{i=1}^t S'_{ij}$ is a $t^2(t-1)c^2$-subset of~$G$.
By Definition~\ref{def-LPpartition} and Remark~\ref{rem-LPpartitiondefn}~$(i)$, it is now sufficient to show that, for all nonprincipal characters $\chi$ of~$G$,
\begin{equation}\label{eqn-Richars}
\{\chi(R_1),\dots,\chi(R_t)\} = \begin{cases}
\{-t^2c^2,\dots, -t^2c^2\} & \mbox{if $\chi \in V_0^\perp$}, \\
\{0,\dots,0\} & \mbox{if $\chi \in H_0^\perp \sm V_0^\perp$},\\
\{t(t-1)c, -tc, \dots, -tc\} & \mbox{if $\chi \notin H_0^\perp$}.
\end{cases}
\end{equation}
Let $\chi$ be a nonprincipal character of~$G$.
Since the $H_i$ correspond to a spread of $Q$, if $\chi$ is nonprincipal on $Q$, then $\chi$ is principal on exactly one of the~$H_i$ (see Example~\ref{ex-spread}). If $\chi$ is nonprincipal on $V_i$ and principal on $H_i$, then $\chi$ induces a nonprincipal character $\psi_i$ on~$V_i/H_i$; since each $S'_{ij}$ is a union of cosets of the order $t$ subgroup $H_i$, we have for each $i,j$ satisfying $1 \le i,j \le t$ that
\[
\chi(S'_{ij}) = \begin{cases}
t \psi_i(S_{ij}) & \mbox{if $\chi \in H_i^\perp \sm V_i^\perp$}, \\
0 & \mbox{if $\chi \notin H_i^\perp$}.
\end{cases}
\]
If $\chi$ is principal on $G'$ then it induces a nonprincipal character~$\tau$ on~$G/G'$;
since the $V_i/G'$ form a spread of $G/G'$, we then have that $\tau$ is principal on exactly one of the~$V_i/G'$ and so $\chi$ is principal on exactly one of the~$V_i$.
We shall use the subgroup inclusions
\begin{equation}\label{eqn-inclusions}
H_i \leqslant Q \leqslant G' \leqslant V_i \quad \mbox{for each $i = 0,\dots,t$}.
\end{equation}
The conclusions of the following five cases collectively establish~\eqref{eqn-Richars}.
\begin{description}
\item[{\bf Case 1:}] $\chi$ is principal on $V_0$. \\
Then by \eqref{eqn-inclusions}, $\chi$ is principal on $G'$ and on each~$H_i$.
For each $i=1,\dots,t$, therefore $\chi$ is nonprincipal on $V_i$ and principal on~$G'/H_i$.
For each $j$, we obtain
$\chi(R_j) = \sum_{i=1}^t \chi(S'_{ij}) = t \sum_{i=1}^t \psi_i(S_{ij}) = t \sum_{i=1}^t (-c^2) = -t^2c^2$.
\item[{\bf Case 2:}] $\chi$ is principal on $G'$ and nonprincipal on $V_0$. \\
Then by \eqref{eqn-inclusions}, $\chi$ is principal on each~$H_i$.
Therefore $\chi$ is principal on each $G'/H_i$ and on exactly one of $V_1,\dots,V_t$, say~$V_I$.
For each $j$, we obtain
$\chi(R_j) = \sum_{i=1}^t \chi(S'_{ij}) = |S'_{Ij}| + t \sum_{1 \le i \le t,\, i\ne I} \psi_i(S_{ij}) = t(t-1)c^2 + t \sum_{1 \le i \le t,\, i \ne I} (-c^2) = 0$.
\item[{\bf Case 3:}] $\chi$ is principal on $Q$ and nonprincipal on $G'$. \\
Then by \eqref{eqn-inclusions}, $\chi$ is principal on each~$H_i$ and nonprincipal on each~$V_i$.
Therefore $\chi$ is principal on each $Q/H_i$ and nonprincipal on each~$G'/H_i$.
For each $j$, we obtain
$\chi(R_j) = t \sum_{i=1}^t \psi_i(S_{ij}) = 0$.
\item[{\bf Case 4:}] $\chi$ is principal on $H_0$ and nonprincipal on $Q$. \\
Then for $i=1,\dots,t$, we have that $\chi$ is nonprincipal on $H_i$ and so $\chi(S'_{ij}) = 0$ for each~$j$. Therefore $\chi(R_j)=0$ for each~$j$.
\item[{\bf Case 5:}] $\chi$ is nonprincipal on $H_0$.\\
By \eqref{eqn-inclusions}, $\chi$ is nonprincipal on $Q$ and on each $V_i$.
Therefore $\chi$ is principal on exactly one of $H_1,\dots,H_t$, say $H_I$, and is nonprincipal on~$Q/H_I$.
For each $j$, we obtain
$\chi(R_j) = t \psi_I(S_{Ij})$ and so
$\{\chi(R_1),\dots,\chi(R_t)\} = \{t(t-1)c, -tc, \ldots, -tc\}$.
\end{description}
\end{proof}
The following example illustrates the use of Theorems~\ref{thm-composition} and~\ref{thm-LPpartitionrecursive} to construct a $(4,4)$ LP-packing in $\Z_4^4$ relative to each of the order~$16$ subgroups $\Z_4\times(2\Z_4)^2$ and $\Z_4^2$ and $(2\Z_4)^4$ in turn, starting from a spread of~$\F_4^2$. In Section~\ref{sec5} we shall show how to apply this procedure recursively to produce infinite families of LP-packings and LP-partitions in groups of increasing exponent and fixed rank.
\begin{example}
Let $(\F_4,\oplus,\cdot) = \{0,1,\alpha,\alpha^2\}$, where $\alpha^2=\alpha \oplus 1$.
Then a spread of $(\F_4^2,\oplus)$ is $\big\{\lan (1,0) \ran$, $\lan (0,1) \ran$, $\lan (1,1) \ran$, $\lan (\alpha,1) \ran$, $\lan (\alpha \oplus 1,1) \ran \big\}$.
Using the isomorphism from $(\F_4^2,\oplus)$ to $\Z_2^4 = \lan y_1,y_2,y_3,y_4 \ran$ given by
$(1,0) \mapsto y_1$, $(\alpha,0) \mapsto y_2$, $(0,1) \mapsto y_3$, $(0,\alpha) \mapsto y_4$,
this spread in multiplicative notation is
\begin{equation}\label{eqn-spreadZ24}
\big\{ \lan y_1, y_2 \ran, \lan y_3, y_4 \ran, \lan y_1y_3, y_2y_4 \ran, \lan y_2y_3,y_1y_2y_4 \ran, \lan y_1y_2y_3, y_1y_4 \ran \big\}.
\end{equation}
Let
\[
\begin{split}
S_1(y_1,y_2,y_3,y_4) &= y_3+y_4+y_3y_4, \\
S_2(y_1,y_2,y_3,y_4) &= y_1y_3+y_2y_4+y_1y_2y_3y_4, \\
S_3(y_1,y_2,y_3,y_4) &= y_2y_3+y_1y_2y_4+y_1y_3y_4, \\
S_4(y_1,y_2,y_3,y_4) &= y_1y_2y_3+y_1y_4+y_2y_3y_4.
\end{split}
\]
Then by Examples~\ref{ex-spread} and~\ref{ex-spread2},
\begin{equation}\label{eqn-packingpartition}
\big\{ S_1(y_1,y_2,y_3,y_4), S_2(y_1,y_2,y_3,y_4), S_3(y_1,y_2,y_3,y_4), S_4(y_1,y_2,y_3,y_4) \big\}
\end{equation}
is both a $(1,4)$ LP-partition in $\lan y_1,y_2,y_3,y_4 \ran -\lan y_1, y_2 \ran \cong \Z_2^4 - \Z_2^2$ relative to $\lan y_1, y_2 \ran \cong \Z_2^2$,
and a $(1,4)$ LP-packing in $\lan y_1,y_2,y_3,y_4 \ran = \Z_2^4$ relative to $\lan y_1, y_2 \ran \cong \Z_2^2$.
Let $G=\Z_4^4 = \lan x_1,x_2,x_3,x_4 \ran$.
We firstly use Theorem~\ref{thm-LPpartitionrecursive} to construct a $(4,4)$ LP-partition in $G-\lan x_1,x_2,x_3^2,x_4^2 \ran = \Z_4^4 - \Z_4^2\times(2\Z_4)^2$
relative to~$\lan x_1^2,x_2^2 \ran \cong \Z_2^2$.
Let $Q = \lan x_1^2,x_2^2,x_3^2,x_4^2 \ran \cong \Z_2^4$,
and let $G' = \lan x_1^2,x_2^2,x_3^2,x_4^2 \ran$ so that $G/G' \cong \Z_2^4$.
With reference to \eqref{eqn-spreadZ24}, the subgroups
\[
\begin{split}
& H_0 = \lan x_1^2, x_2^2 \ran, \quad H_1 = \lan x_3^2, x_4^2 \ran, \quad H_2 = \lan x_1^2x_3^2, x_2^2x_4^2 \ran, \quad \\
& H_3 = \lan x_2^2x_3^2,x_1^2x_2^2x_4^2 \ran, \quad H_4 = \lan x_1^2x_2^2x_3^2, x_1^2x_4^2 \ran
\end{split}
\]
form a spread when viewed as subgroups of~$Q$.
Let
\[
\begin{split}
& V_0 = \lan x_1,x_2,x_3^2,x_4^2 \ran, \quad
V_1 = \lan x_1^2,x_2^2,x_3,x_4 \ran, \quad
V_2 = \lan x_1^2,x_2^2,x_1x_3,x_2x_4 \ran, \quad \\
& V_3 = \lan x_1^2,x_2^2,x_2x_3,x_1x_2x_4 \ran, \quad
V_4 = \lan x_1^2,x_2^2,x_1x_2x_3,x_1x_4 \ran,
\end{split}
\]
so that, with reference to \eqref{eqn-spreadZ24},
$\{V_0/G',V_1/G',V_2/G',V_3/G',V_4/G'\}$ is a spread of $G/G'$ and $H_i \leqslant V_i$ for $i=1,2,3,4$.
Then the factor groups $V_i/H_i, G'/H_i, Q/H_i$ take the following form for $i=1,2,3,4$:
\[
\ra{1.2}
\begin{array}{r|c|c|c}
i & V_i/H_i & G'/H_i & Q/H_i \\ \hline
1 & \lan x_1^2H_1, x_2^2H_1, x_3H_1, x_4H_1 \ran \cong \Z_2^4 & \lan x_1^2H_1, x_2^2H_1 \ran \cong \Z_2^2 & \lan x_1^2H_1, x_2^2H_1 \ran \cong \Z_2^2 \\
2 & \lan x_1^2H_2, x_2^2H_2, x_1x_3H_2, x_2x_4H_2 \ran \cong \Z_2^4 & \lan x_1^2H_2, x_2^2H_2 \ran \cong \Z_2^2 & \lan x_1^2H_2, x_2^2H_2 \ran \cong \Z_2^2 \\
3 & \lan x_1^2H_3, x_2^2H_3, x_2x_3H_3, x_1x_2x_4H_3 \ran \cong \Z_2^4 & \lan x_1^2H_3, x_2^2H_3 \ran \cong \Z_2^2 & \lan x_1^2H_3, x_2^2H_3 \ran \cong \Z_2^2 \\
4 & \lan x_1^2H_4, x_2^2H_4, x_1x_2x_3H_4, x_1x_4H_4 \ran \cong \Z_2^4 & \lan x_1^2H_4, x_2^2H_4 \ran \cong \Z_2^2 & \lan x_1^2H_4, x_2^2H_4 \ran \cong \Z_2^2
\end{array}
\]
Since \eqref{eqn-packingpartition} is a $(1,4)$ LP-partition in $\lan y_1,y_2,y_3,y_4 \ran - \lan y_1,y_2 \ran$ relative to $\lan y_1,y_2 \ran$,
where $\lan y_1,y_2,y_3,y_4\ran = \Z_2^4$, it follows for $i=1,2,3,4$ that the $4$ subsets of $V_i/H_i$
shown in row $i$ of the following table form a $(1,4)$ LP-partition in $V_i/H_i-G'/H_i$ relative to $Q/H_i$:
\begin{small}
\[
\ra{1.2}
\scalemath{0.83}{
\begin{array}{|l|l|l|l|}
\hline
S_1(x_1^2,x_2^2,x_3,x_4)H_1 & S_2(x_1^2,x_2^2,x_3,x_4)H_1 & S_3(x_1^2,x_2^2,x_3,x_4)H_1 & S_4(x_1^2,x_2^2,x_3,x_4)H_1 \\
S_1(x_1^2,x_2^2,x_1x_3,x_2x_4)H_2 & S_2(x_1^2,x_2^2,x_1x_3,x_2x_4)H_2 & S_3(x_1^2,x_2^2,x_1x_3,x_2x_4)H_2 & S_4(x_1^2,x_2^2,x_1x_3,x_2x_4)H_2 \\
S_1(x_1^2,x_2^2,x_2x_3,x_1x_2x_4)H_3 & S_2(x_1^2,x_2^2,x_2x_3,x_1x_2x_4)H_3 & S_3(x_1^2,x_2^2,x_2x_3,x_1x_2x_4)H_3 & S_4(x_1^2,x_2^2,x_2x_3,x_1x_2x_4)H_3 \\
S_1(x_1^2,x_2^2,x_1x_2x_3,x_1x_4)H_4 & S_2(x_1^2,x_2^2,x_1x_2x_3,x_1x_4)H_4 & S_3(x_1^2,x_2^2,x_1x_2x_3,x_1x_4)H_4 & S_4(x_1^2,x_2^2,x_1x_2x_3,x_1x_4)H_4 \\ \hline
\end{array}
}
\]
\end{small}
Now regard each entry of the above table as a $12$-subset of $G$, by interpreting $H_i$ as the sum of four elements in~$G$.
By Theorem~\ref{thm-LPpartitionrecursive}, summing the entries in each column of the table then gives a $(4,4)$ LP-partition $\{R_1,R_2,R_3,R_4\}$ in $G-V_0$ relative to~$H_0$.
We next use Theorem~\ref{thm-composition} to construct a $(4,4)$ LP-packing in $G = \Z_4^4$ relative to $U = \lan x_1,x_2^2,x_3^2 \ran \cong \Z_4 \times (2\Z_4)^2$.
We have $V_0/H_0 = \lan x_1H_0, x_2H_0, x_3^2H_0, x_4^2H_0 \ran \cong \Z_2^4$ and $U/H_0 = \lan x_1H_0, x_3^2H_0 \ran \cong \Z_2^2$.
Since \eqref{eqn-packingpartition} is a $(1,4)$ LP-packing in $\lan y_1,y_2,y_3,y_4 \ran = \Z_2^4$ relative to $\lan y_1,y_2 \ran$,
it follows that
\[
\big\{
S_1(x_1,x_3^2,x_2,x_4^2)H_0,\,\,
S_2(x_1,x_3^2,x_2,x_4^2)H_0,\,\,
S_3(x_1,x_3^2,x_2,x_4^2)H_0,\,\,
S_4(x_1,x_3^2,x_2,x_4^2)H_0
\big\}
\]
is a $(1,4)$ LP-packing in $V_0/H_0$ relative to~$U/H_0$.
Regard each of these subsets as a $12$-subset of $G$, by interpreting $H_0$ as the sum of four elements in~$G$.
Then by Theorem~\ref{thm-composition},
\[
\begin{split}
\big\{
R_1+ S_1(x_1,x_3^2,x_2,x_4^2)H_0,\,\,
R_2+ S_2(x_1,x_3^2,x_2,x_4^2)H_0,\,\, \\
R_3+ S_3(x_1,x_3^2,x_2,x_4^2)H_0,\,\,
R_4+ S_4(x_1,x_3^2,x_2,x_4^2)H_0
\big\}
\end{split}
\]
is a $(4,4)$ LP-packing in $G$ relative to~$U$.
Using the same $(4,4)$ LP-partition $\{R_1,R_2,R_3,R_4\}$, we can similarly use
Theorem~\ref{thm-composition} to construct a $(4,4)$ LP-packing in $G$ relative to the subgroup $U' = \lan x_1,x_2 \ran \cong \Z_4^2$ and to $U'' = \lan x_1^2,x_2^2,x_3^2,x_4^2 \ran \cong (2\Z_4)^4$.
Following the above procedure, and using
$U'/H_0 = \lan x_1H_0, x_2H_0 \ran \cong \Z_2^2$ and
$U''/H_0 = \lan x_3^2H_0, x_4^2H_0 \ran \cong \Z_2^2$, shows that
\[
\begin{split}
\big\{
R_1+ S_1(x_1,x_2,x_3^2,x_4^2)H_0,\,\,
R_2+ S_2(x_1,x_2,x_3^2,x_4^2)H_0,\,\, \\
R_3+ S_3(x_1,x_2,x_3^2,x_4^2)H_0,\,\,
R_4+ S_4(x_1,x_2,x_3^2,x_4^2)H_0
\big\}
\end{split}
\]
is a $(4,4)$ LP-packing in $G$ relative to~$U'$, and that
\[
\begin{split}
\big\{
R_1+ S_1(x_3^2,x_4^2,x_1,x_2)H_0,\,\,
R_2+ S_2(x_3^2,x_4^2,x_1,x_2)H_0,\,\, \\
R_3+ S_3(x_3^2,x_4^2,x_1,x_2)H_0,\,\,
R_4+ S_4(x_3^2,x_4^2,x_1,x_2)H_0
\big\}
\end{split}
\]
is a $(4,4)$ LP-packing in $G$ relative to~$U''$.
\end{example}
\section{Recursive Construction of LP-partitions and LP-packings}\label{sec5}
In this section, we recursively construct infinite families of LP-partitions and LP-packings, as shown schematically in Figure~\ref{fig-ladder}.
On the right side of the figure, we iteratively use Theorem~\ref{thm-LPpartitionrecursive} to produce an LP-partition in the group $\Z_{p^a}^{2s} - \Z_{p^a}^s \times (p\Z_{p^a})^s$ relative to $H_0 \cong \Z_p^s$ by lifting with respect to $H_i \cong \Z_p^s$ an LP-partition in a factor group, for each $i=1,\dots,p^s$ (see Theorem~\ref{thm-LPpartitionuniform}).
On the left side of the figure, we iteratively use Theorem~\ref{thm-composition} to produce an LP-packing in $\Z_{p^a}^{2s}$ relative to an order $p^{as}$ subgroup, by lifting with respect to~$H_0$ an LP-packing in a factor group and taking the union of the resulting subsets with the subsets of the LP-partition at the same height (see Theorem~\ref{thm-LPpackinguniform}).
\begin{figure}[ht]
\centering
\setlength{\unitlength}{1mm}
\begin{picture}(150,130)(10,5)
\put(50,5){\makebox(0,0){$(1,p^s)$ LP-packing in $\Z_p^{2s}$}}
\put(50,0){\makebox(0,0){relative to order $p^s$ subgroup}}
\put(50,10){\vector(0,1){25}}
\put(22,22.5){\makebox(0,0){Lift using $H_0$}}
\put(7,15){\framebox(30,15){}}
\put(125,5){\makebox(0,0){$(1,p^s)$ LP-partition in}}
\put(125,0){\makebox(0,0){$\Z_{p}^{2s}-\Z_p^s$ relative to $H_0 \cong \Z_p^s$}}
\put(120,10){\vector(0,1){25}}
\put(150,25){\makebox(0,0){Lift using}}
\put(150,20){\makebox(0,0){$H_1,\dots,H_{p^s}$}}
\put(135,15){\framebox(30,15){}}
\put(50,45){\makebox(0,0){$(p^s,p^s)$ LP-packing in $\Z_{p^2}^{2s}$}}
\put(50,40){\makebox(0,0){relative to order $p^{2s}$ subgroup}}
\put(50,50){\vector(0,1){25}}
\put(22,62.5){\makebox(0,0){Lift using $H_0$}}
\put(7,55){\framebox(30,15){}}
\put(125,45){\makebox(0,0){$(p^s,p^s)$ LP-partition in}}
\put(125,40){\makebox(0,0){$\Z_{p^2}^{2s}-\Z_{p^2}^s \times (p\Z_{p^2})^s$ relative to $H_0 \cong \Z_p^s$}}
\put(120,50){\vector(0,1){25}}
\put(105,45){\vector(-1,0){30}}
\put(150,65){\makebox(0,0){Lift using}}
\put(150,60){\makebox(0,0){$H_1,\dots,H_{p^s}$}}
\put(135,55){\framebox(30,15){}}
\put(50,85){\makebox(0,0){$(p^{2s},p^s)$ LP-packing in $\Z_{p^3}^{2s}$}}
\put(50,80){\makebox(0,0){relative to order $p^{3s}$ subgroup}}
\put(50,90){\vector(0,1){25}}
\put(22,102.5){\makebox(0,0){Lift using $H_0$}}
\put(7,95){\framebox(30,15){}}
\put(125,85){\makebox(0,0){$(p^{2s},p^s)$ LP-partition in}}
\put(125,80){\makebox(0,0){$\Z_{p^3}^{2s}-\Z_{p^3}^s \times (p\Z_{p^3})^s$ relative to $H_0 \cong \Z_p^s$}}
\put(120,90){\vector(0,1){25}}
\put(105,85){\vector(-1,0){30}}
\put(150,105){\makebox(0,0){Lift using}}
\put(150,100){\makebox(0,0){$H_1,\dots,H_{p^s}$}}
\put(135,95){\framebox(30,15){}}
\put(50,120){\circle*{1}}
\put(50,125){\circle*{1}}
\put(120,120){\circle*{1}}
\put(120,125){\circle*{1}}
\end{picture}
\vspace{1cm}
\caption{Recursive construction of LP-packings and LP-partitions
according to Theorems~\ref{thm-LPpartitionuniform} and~\ref{thm-LPpackinguniform}}
\label{fig-ladder}
\end{figure}
We begin with a technical lemma.
\begin{lemma}\label{lem-quotient}
Let $p$ be prime, and let $s$ and $a_1, \dots, a_{2s}$ be positive integers.
Let $G = \prod_{j=1}^{2s} \Z_{p^{a_j}}$, and let $Q$ be the unique subgroup of $G$ isomorphic to~$\Z_{p}^{2s}$. Let $H_0,\dots,H_{p^s}$ be a spread of~$Q$, where (without loss of generality) $H_0$ is contained in the first $s$ direct factors of~$G$. Then for each $i = 1, \dots, p^s$, we have
$G/H_i \cong \prod_{j=1}^s \Z_{p^{a_j}} \times \prod_{j=s+1}^{2s} \Z_{p^{a_j-1}}$, and the quotient group $Q/H_i$ is contained in the first $s$ direct factors of $G/H_i$.
\end{lemma}
\begin{proof}
Fix $i \in \{1,\dots,p^s\}$.
Since $Q=H_0H_i$, we can choose a set of generators for $Q$ (depending on $i$) so that $H_i$ is contained in the last $s$ direct factors of~$Q$. Therefore $G/H_i \cong \prod_{j=1}^s \Z_{p^{a_j}} \times \prod_{j=s+1}^{2s} \Z_{p^{a_j-1}}$. Let the first $s$ direct factors of~$G$ be $\lan x_1,\dots,x_s \ran$, so that the first $s$ direct factors of $G/H_i$ are $\lan x_1H_i,\dots,x_sH_i \ran$.
Since $H_0 = \lan x_1^{p^{a_1-1}}, \dots, x_s^{p^{a_s-1}} \ran$, we have $Q/H_i=\lan x_1^{p^{a_1-1}}H_i,\dots,x_s^{p^{a_s-1}}H_i \ran$. Therefore $Q/H_i$ is contained in the first $s$ direct factors of~$G/H_i$.
\end{proof}
We now apply Theorem~\ref{thm-LPpartitionrecursive} iteratively to produce an infinite family of LP-partitions.
\begin{theorem}\label{thm-LPpartitionuniform}
Let $p$ be prime, let $a,s$ be positive integers, and let $G=\Z_{p^a}^{2s}$.
Let $H_0$ and $V_0$
be the unique subgroups of $G$ for which $H_0 \cong \Z_p^s$ is contained in the first $s$ direct factors of $G$ and $V_0 \cong \Z_{p^a}^s \times \Z_{p^{a-1}}^s$ satisfies $V_0/H_0 \cong \Z_{p^{a-1}}^{2s}$.
Then there exists a $(p^{(a-1)s},p^s)$ LP-partition in $G-V_0$ relative to~$H_0$.
\end{theorem}
\begin{proof}
The proof is by induction on $a \ge 1$. The case $a=1$ is given by Example~\ref{ex-spread2}. Suppose that all cases up to $a-1 \ge 1$ are true.
Let $Q$ and $G'$ be the unique subgroups of $G$ for which $Q \cong \Z_p^{2s}$ and $G' \cong \Z_{p^{a-1}}^{2s}$.
Since $G/G' \cong \Z_p^{2s}$, by Example~\ref{ex-spread} there exists a spread in $G/G'$ one of whose elements we may take to be~$V_0/G'$. Let the remaining elements of this spread be $V_1/G',\dots,V_{p^s}/G'$, where $V_1,\dots,V_{p^s}$ are subgroups of~$G$ containing $G'$.
Let $\phi$ be a group isomorphism from $G/G'$ to $Q$ for which $\phi(V_0/G') = H_0$, and let $H_i = \phi(V_i/G')$ for $i = 1,\dots,p^s$.
Since $\{V_0/G',\dots,V_{p^s}/G'\}$ is a spread of $G/G'$, it follows that $\{H_0,\dots,H_{p^s}\}$ is a spread of~$Q$.
Furthermore, for each $i$ we have $H_i \leqslant Q \leqslant G' \leqslant V_i$ and therefore $H_i$ is a subgroup of $V_i$, and $V_i/H_i \cong G'$.
Let $i \in \{1,\dots,p^s\}$.
By construction, we have $H_i \leqslant V_i$ and $V_i/H_i \cong G' \cong \Z_{p^{a-1}}^{2s}$.
By Lemma~\ref{lem-quotient}, we have $G'/H_i \cong \Z_{p^{a-1}}^s \times \Z_{p^{a-2}}^s$ and $Q/H_i \cong \Z_p^s$, and $Q/H_i$ is contained in the first $s$ direct factors of~$V_i/H_i$. We also have $(G'/H_i)/(Q/H_i) \cong G'/Q \cong \Z_{p^{a-2}}^{2s}$. Therefore by the inductive hypothesis there exists a $(p^{(a-2)s},p^s)$ LP-partition in $V_i/H_i - G'/H_i$ relative to $Q/H_i$. Then by Theorem~\ref{thm-LPpartitionrecursive} there exists a $(p^s \cdot p^{(a-2)s},p^s)$ LP-partition in $G-V_0$ relative to~$H_0$, so the case $a$ is true.
\end{proof}
We now apply Theorem~\ref{thm-composition} iteratively, making use of the LP-partitions constructed in Theorem~\ref{thm-LPpartitionuniform}, to produce an infinite family of LP-packings in groups of increasing exponent relative to an arbitrary subgroup of the appropriate order. The initial LP-packing in an elementary abelian group is obtained from a spread.
\begin{theorem}\label{thm-LPpackinguniform}
Let $p$ be prime, and let $a,s$ be positive integers. Let $G=\Z_{p^a}^{2s}$ and let $U$ be a subgroup of $G$ of order~$p^{as}$. Then there exists a $(p^{(a-1)s},p^s)$ LP-packing in $G$ relative to~$U$.
\end{theorem}
\begin{proof}
The proof is by induction on $a \ge 1$. The case $a=1$ follows from the construction of a spread in Example~\ref{ex-spread}, because by a suitable choice of generators of $G = \Z_p^{2s}$ we may assume that the element $H_0$ of the spread is~$U$.
Suppose that all cases up to $a-1 \ge 1$ are true.
Since $U$ is a subgroup of $G$ of order $p^{as}$, it has at most $s$ direct factors isomorphic to~$\Z_{p^a}$ and has rank at least~$s$. We may therefore choose $V_0 \cong \Z_{p^{a}}^{s} \times \Z_{p^{a-1}}^{s}$ for which $U \leqslant V_0 \leqslant G$, and also assume that the subgroup of $V_0$ isomorphic to $\Z_{p^a}^s$ is contained in the first $s$ direct factors of~$G$
and has intersection of rank $s$ with~$U$.
Therefore $U$ contains a subgroup $H_0 \cong \Z_p^s$ for which $H_0$ is contained in the first $s$ direct factors of~$G$.
We then have $V_0/H_0 \cong \Z_{p^{a-1}}^{2s}$ and $|U/H_0| = p^{(a-1)s}$.
Then by Theorem~\ref{thm-LPpartitionuniform}, there exists a $(p^{(a-1)s},p^s)$ LP-partition in $G-V_0$ relative to $H_0$. By the inductive hypothesis, there exists a $(p^{(a-2)s},p^s)$ LP-packing in $V_0/H_0$ relative to~$U/H_0$. Therefore by Theorem~\ref{thm-composition}, there exists a $(p^{(a-1)s},p^s)$ LP-packing in $G$ relative to~$U$, so the case $a$ is true.
\end{proof}
The construction process illustrated in Figure~\ref{fig-ladder} shares many features with the recursive construction of difference sets from relative difference sets introduced in~\cite{DJ}, which is illustrated in \cite{DJ-nato} using an analogous representation to Figure~\ref{fig-ladder}. In both settings, an ingredient on the left side is lifted from a factor group and combined with an ingredient on the right side to form a larger example on the left side; and multiple instances of an ingredient on the right side are lifted from factor groups and combined to form a larger example on the right side.
The ingredients in \cite{DJ} corresponding to an LP-packing and an LP-partition are, respectively, a covering extended building set for constructing a difference set, and a building set for constructing a relative difference set.
However, in \cite{DJ} the ingredients are placed into distinct cosets of a subgroup and so can be combined without regard to their intersections as subsets of the subgroup.
In contrast, in the construction of Theorem~\ref{thm-composition}, the lifted subset $P'_j$ of an LP-packing is combined with the subset $R_j$ of an LP-partition by taking their union.
Likewise, in the construction of Theorem~\ref{thm-LPpartitionrecursive}, the subsets of the $i^{\rm th}$ LP-partition are lifted to $S'_{i1},\dots,S'_{it}$, and then
subset $R_j$ of the new LP-partition is formed as the union of $S'_{1j}, \dots, S'_{tj}$ (with one lifted subset derived from each of the $t$ LP-partitions).
This leads to an additional constraint, that the subsets $P'_j$ and $S'_{1j}, \dots, S'_{tj}$ must be disjoint for each~$j$.
We now extend Theorem~\ref{thm-LPpackinguniform} using constructions from Section~\ref{sec3} in order to obtain the result stated in Theorem~\ref{thm-preview}.
\begin{corollary}\label{coro-LPpackingeven}
Let $p$ be prime, let $s_1,\dots,s_v$ be nonnegative integers (not all zero), and let $m=\min\{s_i \mid s_i>0\}$. For each $i=1,\dots,v$, let $G_i=\Z_{p^i}^{2s_i}$ and let $U_i$ be a subgroup of $G_i$ of order~$p^{is_i}$.
Let $G = \prod_{i=1}^v G_i$, let $U = \prod_{i=1}^v U_i$, and let $n = \sqrt{|G|}$.
Then for each $j = 0,\dots,m-1$, there exists an $\big(\frac{n}{p^{m-j}}, p^{m-j}\big)$ LP-packing in $G$ relative to~$U$.
\end{corollary}
\begin{proof}
For each $i=1,\dots,v$ for which $s_i > 0$, by Theorem~\ref{thm-LPpackinguniform} there exists a $(p^{(i-1)s_i},p^{s_i})$ LP-packing in $G_i$ relative to~$U_i$. By a straightforward induction, Corollary~\ref{coro-LPpackingproduct} then gives an
$\big(\frac{n}{p^m}, p^m\big)$ LP-packing in $G$ relative to~$U$.
For $j \in \{0,\dots,m-1\}$, apply Lemma~\ref{lem-LPpackingunion} with $s=p^j$ to give an $\big(\frac{n}{p^{m-j}}, p^{m-j}\big)$ LP-packing in $G$ relative to~$U$.
\end{proof}
\begin{remark}\label{rem-PDSeven}
By Definition~\ref{def-LPpacking}, the constructed LP-packing in Corollary~\ref{coro-LPpackingeven} is a collection of $p^{m-j}$ disjoint regular $\big(n,\frac{n}{p^{m-j}}\big)$ Latin square type PDSs in $G$ avoiding~$U$.
By Lemma~\ref{lem-LPpackingtoPDS}, we also obtain a regular Latin square type PDS in $G$ with parameters $(n,r)$ and $(n,r+1)$, for each positive integer multiple $r$ of $\frac{n}{p^{m-j}}$ satisfying $r \le n$.
This improves on numerous previous constructions of PDSs in three respects: by constructing a collection of disjoint PDSs in $G$ where previously only a single PDS was constructed; by identifying the order~$n$ subgroup~$U$ and showing it can be chosen arbitrarily; and by producing Latin square type PDSs with new parameters~$(n,r)$.
We give an illustrative selection of examples.
\begin{enumerate}[$(i)$]
\item
Result~\ref{res-PCPPDS} constructs a single regular $(n,r)$ Latin square type PDS in $G$ for each positive integer $r \le p^m+1$, whereas the case $j=0$ of Corollary~\ref{coro-LPpackingeven} gives a disjoint collection of $p^m$ such PDSs with $r = \frac{n}{p^m}$.
\item
Result~\ref{res-ringPDS}~$(iii)$ constructs a regular $(n,r)$ Latin square type PDS in $G$ for each positive integer multiple~$r$ of $\frac{n}{p^{\,\gcd(s_1,\dots,s_v)}}$ satisfying $r \le n$, whereas the case $j=0$ of Corollary~\ref{coro-LPpackingeven} gives such a PDS for each positive integer multiple~$r$ of $\frac{n}{p^m}$ satisfying $r \le n$ and whose parameters therefore differ when $m > \gcd(s_1,\dots,s_v)$.
\item
The LP-packing of {\rm \cite[Cor.~6.2]{P08DCC}} described in Table~\ref{tab-LPpacking} is a disjoint collection of $p$ regular $(n,\frac{n}{p})$ PDSs in certain groups $G$ constrained to have at most one direct factor with an exponent that is an odd power of~$p$, whereas the case $j=m-1$ of Corollary~\ref{coro-LPpackingeven} gives such disjoint collections of PDSs without this constraint.
\item
In each group for which the lifting approach in {\rm \cite{H00}, \cite[Sects.~4,~5]{H03}, \cite{HN}} constructs a single PDS, Corollary~\ref{coro-LPpackingeven} provides a collection of disjoint PDSs having the same parameters and avoiding an arbitrary order~$n$ subgroup $U$ of~$G$.
\item
Result~\ref{res-PDSproductLP} constructs a partition of $G-1_G$ into $p$ regular PDSs in $G$, of which $p-1$ are of $(n,\frac{n}{p})$ Latin square type and one is of $\big(n,\frac{n}{p}+1\big)$ Latin square type, where $G$ is constrained to have $s_{2i+1} = 0$ for each~$i > 0$.
In contrast, by applying Remark~\ref{rem-LPpacking}~$(i)$ to the case $j=0$ of Corollary~\ref{coro-LPpackingeven} we obtain a partition of $G-1_G$ into $p^m$ regular PDSs in~$G$, of which $p^m-1$ are of $\big(n,\frac{n}{p^m}\big)$ Latin square type and one is of $\big(n,\frac{n}{p^m}+1\big)$ Latin square type, without constraining~$G$.
\end{enumerate}
\end{remark}
The LP-packings in Corollary~\ref{coro-LPpackingeven} are all contained in $p$-groups of the form $H \times H$, having even rank. We now extend this result to $p$-groups of odd rank, starting from the following infinite family of odd rank examples that was obtained from an elegant construction using Galois rings. (We could likewise make use of the further odd rank example of Example~\ref{ex-LPpacking444}~$(ii)$, relative to a nonelementary abelian group.)
\begin{result}[{\cite[Lem.~3.1]{HLX}}]\label{res-HLX}
Let $p$ be prime and let $w>1$ be an odd integer.
Then there exists a $(p^{w-\s(p,w)},p^{\s(p,w)})$ LP-packing in $\Z_{p^2}^w$ relative to~$(p\Z_{p^2})^w$, where the function $\sigma(p,w)$ is defined in~\eqref{eqn-s}.
\end{result}
\begin{corollary}\label{coro-LPpackingodd}
Let $p$ be prime, let $w>1$ be an odd integer, and let $s_1,\dots, s_v$ be nonnegative integers (possibly all zero).
Let $m = \min\big(\{\s(p,w)\} \cup \{s_i \mid s_i>0\}\big)$.
For each $i=1,\dots,v$, let $G_i=\Z_{p^i}^{2s_i}$ and let $U_i$ be a subgroup of $G_i$ of order~$p^{is_i}$.
Let $G = \Z_{p^2}^{w} \times \prod_{i=1}^v G_i$, let $U = (p\Z_{p^2})^{w} \times \prod_{i=1}^v U_i$, and let $n = \sqrt{|G|}$.
Then for each $j = 0,\dots,m-1$, there exists an $\big(\frac{n}{p^{m-j}}, p^{m-j}\big)$ LP-packing in $G$ relative to~$U$.
\end{corollary}
\begin{proof}
In the case that the $s_i$ are all zero, we have $m = \s(p,w)$ and the result follows by applying Lemma~\ref{lem-LPpackingunion} with $s=p^j$ to Result~\ref{res-HLX}.
Otherwise, at least one of the $s_i$ is positive and we may define \mbox{$m' = \min\{s_i \mid s_i > 0\}$}.
The case $j=0$ of Corollary~\ref{coro-LPpackingeven} then gives an $\big(\frac{n/p^w}{p^{m'}}, p^{m'}\big)$ LP-packing in $\prod_{i=1}^v G_i$ relative to~$\prod_{i=1}^v U_i$.
Combine with the LP-packing of Result~\ref{res-HLX} using Corollary~\ref{coro-LPpackingproduct} to give an $\big(\frac{n}{p^{m}}, p^m\big)$ LP-packing in $G$ relative to~$U$.
Then for $j \in \{0,\dots,m-1\}$, apply Lemma~\ref{lem-LPpackingunion} with $s=p^j$ to give an $\big(\frac{n}{p^{m-j}}, p^{m-j}\big)$ LP-packing in $G$ relative to~$U$.
\end{proof}
\begin{remark}\label{rem-PDSodd}
By Definition~\ref{def-LPpacking}, the constructed LP-packing in Corollary~\ref{coro-LPpackingodd} is a collection of $p^{m-j}$ disjoint regular $\big(n,\frac{n}{p^{m-j}}\big)$ Latin square type PDSs in $G$ avoiding~$U$.
By Lemma~\ref{lem-LPpackingtoPDS}, we also obtain a regular Latin square type PDS in $G$ with parameters $\big(n,r)$ and with parameters $\big(n,r+1)$, for each positive integer multiple $r$ of $\frac{n}{p^{m-j}}$ satisfying $r \le n$.
We can show that this improves on previous constructions along similar lines to Remark~\ref{rem-PDSeven}. For example,
Result~\ref{res-ringPDS}~$(ii)$ constructs a single regular $(n,r)$ Latin square type PDS in $G$ for $v=1$ and $i =1$ and $s_1$ a positive integer multiple of $w$, for each positive integer multiple $r$ of $\frac{n}{p^{\s(p,w)}}$ satisfying $r \le n$.
In contrast, the case $j=0$ of Corollary~\ref{coro-LPpackingodd} gives a disjoint collection of $p^m$ such PDSs in much more general groups $G$ (having arbitrary $v$ and~$s_i$).
\end{remark}
\section{NLP-packings}\label{sec6}
In this section, we introduce a $(c,t-1)$ NLP-packing as a collection of $t-1$ disjoint regular $(tc,c)$ negative Latin square type PDSs, in an abelian group of order $t^2c^2$, whose union satisfies an additional property.
This structure forms a counterpart to a $(c,t)$ LP-packing.
We provide examples of NLP-packings, and show how to combine them with LP-packings via a product construction to form NLP-packings in the direct product of the starting groups.
In conjunction with the LP-packings of Corollaries~\ref{coro-LPpackingeven} and~\ref{coro-LPpackingodd}, this produces infinite families of NLP-packings.
However, we do not have a lifting result for NLP-packings similar to Theorem~\ref{thm-composition}.
\begin{definition}[NLP-packing]\label{def-NLPpacking}
Let $t > 1$ and $c>0$ be integers, and let $G$ be an abelian group of order $t^2c^2$.
A \emph{$(c,t-1)$ NLP-packing in $G$} is a collection $\{P_1,\dots,P_{t-1}\}$ of $t-1$ regular $(tc,c)$ negative Latin square type PDSs in~$G$ for which $G-1_G-\sum_{i=1}^{t-1}P_i$ is a regular $(tc,c-1)$ negative Latin square type PDS in~$G$.
\end{definition}
\begin{remark}\label{rem-NLPpacking}
\quad
\begin{enumerate}[$(i)$]
\item
We can rephrase Definition~\ref{def-NLPpacking} to say that a $(c,t-1)$ NLP-packing in $G$ is a partition of $G-1_G$ into $t$ regular partial difference sets, of which $t-1$ are of $(tc,c)$ negative Latin square type and one is of $(tc,c-1)$ negative Latin square type. Previous results on NLP-packings use this phrasing.
\item
In Definition~\ref{def-NLPpacking}, each $P_i$ is a $c(tc+1)$-subset of $G$, and $G-1_G-\sum_{i=1}^{t-1}P_i$ is a $(c-1)(tc+1)$-subset of $G$, so the subsets $P_i$ of a $(c,t-1)$ NLP-packing are necessarily disjoint.
For all $c \ge 1$, if a $(c,t-1)$ NLP-packing exists in $G$, then it contains the maximum possible number of disjoint regular $(tc,c)$ negative Latin square type PDSs.
\item
The case $r=0$ was allowed in Definition~\ref{def-LatinPDS} so that the definition of a $(c,t-1)$ NLP-packing applies when $c=1$.
\end{enumerate}
\end{remark}
NLP-packings have been studied in a series of previous papers, as summarized in Table~\ref{tab-NLPpacking}. We now characterize a $(c,t-1)$ NLP-packing using character sums.
\begin{lemma}\label{lem-NLPpacking}
Let $P_1,\dots,P_{t-1}$ be disjoint $c(tc+1)$-subsets of an abelian group $G$ of order~$t^2c^2$.
Then $\{P_1, \ldots, P_{t-1}\}$ is a $(c,t-1)$ NLP-packing in $G$ if and only if the multiset equality
\[
\{\chi(P_1),\dots,\chi(P_{t-1})\} = \{c,\dots,c\} \mbox{ or } \{-(t-1)c,c,\dots,c\}
\]
holds for all nonprincipal characters $\chi$ of~$G$.
\end{lemma}
\begin{proof}
Each $P_i$ is a $c(tc+1)$-subset of $G$, and the $P_i$ are disjoint, so $G-1_G-\sum_{i=1}^{t-1}P_i$ is a $(c-1)(tc+1)$-subset of~$G$. Lemma~\ref{lem-LatinPDS} then shows that $\{P_1,\dots,P_{t-1}\}$ is a $(c,t-1)$ NLP-packing in $G$ if and only if
\begin{equation}
\label{eqn-chiPi}
\chi(P_i) \in \{c, -(t-1)c\} \quad \! \mbox{and} \! \quad
\chi\Big(G-1_G-\sum_{i=1}^{t-1}P_i\Big) \in \{c-1, -(tc-c+1)\}
\end{equation}
for all nonprincipal characters $\chi$ of $G$.
For each nonprincipal character $\chi$ of $G$, condition \eqref{eqn-chiPi}
is equivalent to
\[
\chi(P_i) \in \{c, -(t-1)c\} \quad \mbox{ and } \quad
\sum_{i=1}^{t-1}\chi(P_i) \in \{-c, (t-1)c)\},
\]
which in turn is equivalent to
\[
\{\chi(P_1),\dots,\chi(P_{t-1})\} = \{c,\dots,c\} \mbox{ or } \{-(t-1)c,c,\dots,c\},
\]
as required.
\end{proof}
Lemmas~\ref{lem-LatinPDS} and~\ref{lem-NLPpacking} allow us to combine subsets from an NLP-packing to obtain negative Latin square type PDSs with various parameters.
\begin{lemma}\label{lem-NLPpackingtoPDS}
Suppose that $\{P_1, \ldots, P_{t-1}\}$ is a $(c,t-1)$ NLP-packing in an abelian group $G$ of order~$t^2c^2$, and let $I$ be a $b$-subset of $\{1,\dots,t-1\}$. Then $\sum_{i \in I} P_i$ is a regular $(tc,bc)$ negative Latin square type PDS in~$G$.
\end{lemma}
\begin{remark}
Suppose that $\{P_1, \ldots, P_{t-1}\}$ is a $(c,t-1)$ NLP-packing in an abelian group $G$ of order~$t^2c^2$.
\begin{enumerate}[$(i)$]
\item
Let $I$ be a $b$-subset of $\{1,\dots,t-1\}$.
By Lemma~\ref{lem-LatinPDS}, an $r(n+1)$-subset $D$ of an abelian group $G$ of order $n^2$ is a regular $(n,r)$ negative Latin square type PDS in $G$ if and only if
$G-1_G-D$ is a regular $(n,n-r-1)$ negative Latin square type PDS in~$G$.
Lemma~\ref{lem-NLPpackingtoPDS} then shows that
$G-1_G-\sum_{i \in I} P_i$ is a regular $(tc,(t-b)c-1)$ negative Latin square type PDS in~$G$.
\item
For each $i$, let $\wti{{P\mkern 0mu}_i} = \{\chi \in \wh{G} \mid \chi(P_i) = -(t-1)c\}$.
Let $P = G-1_G-\sum_{i=1}^{t-1}P_i$ and let $\wti{P} = \{\chi \in \wh{G} \mid \chi(P) = -(tc-c+1)\}$.
By Lemma~\ref{lem-LatinPDS} and Definition~\ref{def-NLPpacking}, each $\wti{{P\mkern 0mu}_i}$ is a regular $(tc,c)$ negative Latin square type PDS in~$\wh{G}$, and $\wti{P}$ is a
regular $(tc,c-1)$ negative Latin square type PDS in~$\wh{G}$.
By Lemma~\ref{lem-NLPpacking}, each nonprincipal character $\chi$ of $G$ belongs to exactly one of $\wti{P_1},\dots,\wti{P_{t-1}},\wti{P}$.
By Definition~\ref{def-NLPpacking}, the collection $\{\wti{P_1}, \dots, \wti{P_{t-1}}\}$ is therefore a $(c,t-1)$ NLP-packing in~$\wh{G}$.
\end{enumerate}
\end{remark}
We give some small examples of NLP-packings.
\begin{example}\label{ex-NLPpackingsmall}
\quad
\begin{enumerate}[$(i)$]
\item
{\rm\cite[p.~373]{P08DCC}}.
A $(1,3)$ NLP-packing $\{P_1,P_2,P_3\}$ in $\Z_2^4=\lan x,y,z,w \ran$ is given by
\begin{align*}
P_1 &= x+z+yw+xw+yz, \\
P_2 &= y+w+xyzw+yzw+xyw, \\
P_3 &= xy+zw+xz+xyz+xzw.
\end{align*}
\item
{\rm\cite[p.~373]{P08DCC}}.
A $(1,3)$ NLP-packing $\{P_1,P_2,P_3\}$ in $\Z_4^2=\lan x,y \ran$ is given by
\begin{align*}
P_1 &= x^2+y+y^3+xy+x^3y^3, \\
P_2 &= x+x^3+y^2+xy^3+x^3y, \\
P_3 &= xy^2+x^3y^2+x^2y+x^2y^3+x^2y^2.
\end{align*}
\item
{\rm\cite[Ex.~1]{P08DCC}}.
A $(1,2)$ NLP-packing $\{P_1,P_2\}$ in $\Z_3^2=\lan x,y \ran$ is given by
\begin{align*}
P_1 &= x+x^2+y+y^2, \\
P_2 &= xy+x^2y^2+xy^2+x^2y.
\end{align*}
\end{enumerate}
\end{example}
We next show that combining the subsets of an NLP-packing into equally-sized collections gives an NLP-packing with fewer subsets.
\begin{lemma}\label{lem-NLPpackingunion}
Suppose there exists a $(c,t-1)$ NLP-packing in an abelian group $G$ of order $t^2c^2$, and suppose $s$ divides~$t$. Then there exists an $(sc,\frac{t}{s}-1)$ NLP-packing in~$G$.
\end{lemma}
\begin{proof}
Let $\{P_1, \dots, P_{t-1}\}$ be a $(c,t-1)$ NLP-packing in $G$, and let
\[
P'_i = P_{is+1} + P_{is+2} + \dots + P_{is+s} \quad \mbox{for $i = 0, 1,\dots, \frac{t}{s}-2$}.
\]
Then Lemma~\ref{lem-NLPpacking} shows that $\{P'_0,\dots,P'_{\frac{t}{s}-2}\}$ is an $(sc,\frac{t}{s}-1)$ NLP-packing in~$G$.
\end{proof}
The following product construction combines an NLP-packing with an LP-packing to form an NLP-packing in a larger group.
\begin{theorem}[{\cite[Thm.~2.1]{P08DCC}}]\label{thm-NLPLPpackingproduct}
Suppose there exists a
$(c_1,t-1)$ NLP-packing in an abelian group $G_1$ of order $t^2c_1^2$, and a
$(c_2,t)$ LP-packing in an abelian group $G_2$ of order $t^2c_2^2$ relative to a subgroup $U_2$ of order~$tc_2$.
Then there exists a $(tc_1c_2,t-1)$ NLP-packing in $G_1 \times G_2$.
\end{theorem}
\begin{proof}
Let $\{ P_{1,0},\dots,P_{1,t-2} \}$ be a $(c_1,t-1)$ NLP-packing in~$G_1$, and let $\{ P_{2,0},\dots,P_{2,t-1} \}$ be a $(c_2,t)$ LP-packing in~$G_2$ relative to~$U_2$.
Define $P_{1,t-1}=G_1-\sum_{i=0}^{t-2} P_{1,i}$ and
\[
K_{\ell}=P_{1,\ell}U_2+\sum_{i=0}^{t-1} P_{1,i}P_{2,i+\ell} \quad \mbox{for $\ell = 0, \dots, t-2$},
\]
where the subscript $i+\ell$ is reduced modulo~$t$.
We shall use Lemma~\ref{lem-NLPpacking} to show that
$\{ K_0, K_1, \dots,$ $K_{t-2} \}$ is a $(tc_1c_2,t-1)$ NLP-packing in $G_1 \times G_2$.
The $P_{1,i}$ are disjoint $c_1(tc_1+1)$-subsets of $G_1$ for $i=0,\dots,t-2$, and by definition $P_{1,t-1}$ is a $c_1(tc_1-t+1)$-subset of $G_1$ disjoint from the other~$P_{1,i}$. The $P_{2,i}$ are $c_2(tc_2-1)$-subsets of $G_2-U_2$. Therefore the $K_\ell$ are disjoint subsets of $G_1 \times G_2$, each of size
$c_1(tc_1+1)tc_2+c_1(tc_1-t+1)c_2(tc_2-1)+(t-1)c_1(tc_1+1)c_2(tc_2-1)=tc_1c_2(t^2c_1c_2+1)$.
Let $\chi$ be a nonprincipal character of $G_1 \times G_2$, and let $\chi_j=\chi|_{G_j}$ for $j=1,2$. Then
\[
\chi(K_\ell) = \chi_1(P_{1,\ell}) \chi_2(U_2) + \sum_{i=0}^{t-1} \chi_1(P_{1,i})\chi_2(P_{2,i+\ell}) \quad \mbox{for $\ell = 0,\dots,t-2$}.
\]
By Lemma~\ref{lem-NLPpacking}, it remains to prove that
\begin{equation}\label{eqn-NLPprod}
\{\chi(K_0),\dots,\chi(K_{t-2})\} = \{tc_1c_2,\dots,tc_1c_2\} \mbox{ or } \{-t(t-1)c_1c_2,tc_1c_2,\dots,tc_1c_2\}.
\end{equation}
We are given that
\begin{equation}\label{eqn-chi1P10}
\{\chi_1(P_{1,0}),\dots,\chi_1(P_{1,t-2})\} = \begin{cases}
\{c_1(tc_1+1),\dots,c_1(tc_1+1)\} & \mbox{if $\chi_1 \in G_1^\perp$,} \\
\{c_1,\dots,c_1\} \mbox{ or } \{-(t-1)c_1,c_1,\dots,c_1\}
& \mbox{if $\chi_1 \not\in G_1^\perp$}
\end{cases}
\end{equation}
by Lemma~\ref{lem-NLPpacking}, and that
\[
\{\chi_2(P_{2,0}),\dots,\chi_2(P_{2,t-1})\} = \begin{cases}
\{c_2(tc_2-1),\dots, c_2(tc_2-1)\} & \mbox{if $\chi_2 \in G_2^\perp$,} \\
\{-c_2,\dots, -c_2\} & \mbox{if $\chi_2 \in U_2^\perp \sm G_2^\perp$,} \\
\{(t-1)c_2, -c_2, \dots, -c_2\} & \mbox{if $\chi_2 \notin U_2^\perp$}
\end{cases}
\]
by Lemma~\ref{lem-LPpacking}.
The following two cases together establish~\eqref{eqn-NLPprod}.
\begin{description}
\item[{\bf Case 1:}] $\chi_2$ is principal on $U_2$.\\
Then $\chi_2(P_{2,i})$ is constant over $i$ (regardless of whether $\chi_2$ is principal or nonprincipal on $G_2$), so
$\chi(K_\ell) = \chi_1(P_{1,\ell}) tc_2 + \chi_1 \big(\sum_{i=0}^{t-1}P_{1,i}\big) \chi_2(P_{2,1}) = \chi_1(P_{1,\ell}) tc_2 + \chi_1(G_1) \chi_2(P_{2,1})$.
In the case that $\chi_1$ is principal on $G_1$, and therefore $\chi_2$ is nonprincipal on $G_2$, we obtain
$\chi(K_\ell) = c_1(tc_1+1) tc_2 + t^2c_1^2 (-c_2) = tc_1c_2$ for each~$\ell$.
Otherwise $\chi_1$ is nonprincipal on $G_1$, and then we obtain
$\chi(K_\ell) = \chi_1(P_{1,\ell}) tc_2$ so that
$\{\chi(K_0),\dots,\chi(K_{t-2})\} =
\{tc_1c_2,\dots,tc_1c_2\}$ or $\{-(t-1)tc_1c_2,tc_1c_2,\dots,tc_1c_2\}$.
\item[{\bf Case 2:}] $\chi_2$ is nonprincipal on $U_2$.\\
Then
$\chi(K_\ell) = \sum_{i=0}^{t-1} \chi_1(P_{1,i})\chi_2(P_{2,i+\ell})$.
By definition of $P_{1,t-1}$, we can extend \eqref{eqn-chi1P10} to obtain
\[
\begin{split}
&\{\chi_1(P_{1,0}),\dots,\chi_1(P_{1,t-1})\} \\
=& \begin{cases}
\{c_1(tc_1-t+1),c_1(tc_1+1),\dots,
c_1(tc_1+1)\}
& \mbox{if $\chi_1 \in G_1^\perp$,} \\
\{-(t-1)c_1,c_1,\dots,c_1\} & \mbox{if $\chi_1 \not\in G_1^\perp$},
\end{cases}
\end{split}
\]
which can be written as
\[
\{\chi_1(P_{1,0}),\dots,\chi_1(P_{1,t-1})\} = \{d-tc_1,d,\dots,d\} \quad \mbox{for all $\chi_1$}
\]
(where the integer $d$ equals $ c_1(tc_1+1)$ if $\chi_1$ is principal on $G_1^\perp$, and equals $c_1$ otherwise).
Therefore the multiset $\{\chi(K_0),\dots,\chi(K_{t-2})\}$ contains
either one or zero occurrences of the value
$(d-tc_1)(t-1)c_2 + (t-1)d(-c_2) = -t(t-1)c_1c_2$,
and (respectively)
$t-2$ or $t-1$ occurrences of the value $(d-tc_1)(-c_2)+d(t-1)c_2 + (t-2)d(-c_2) = tc_1c_2$.
\end{description}
\end{proof}
\noindent
\begin{remark}
The construction of Theorem~\ref{thm-NLPLPpackingproduct} is a restatement (without restricting $t$ to be a prime power) of the product construction for collections of negative Latin square type PDSs and Latin square type PDSs given in {\rm \cite[Thm.~2.1]{P08DCC}}.
Indeed, the configuration described in {\rm \cite[Lemma 2.1]{P08DCC}} for $e=-1$ can be represented as $\{ P_1, \dots, P_{t-1}, G-1_G-\sum_{i=1}^{t-1} P_i\}$, where $\{P_1,\dots,P_{t-1}\}$ is a $(c,t-1)$ NLP-packing in $G$.
Although our formulation explicitly identifies the subgroup $U_2$ of the starting LP-packing, this does not seem to confer an advantage because the subgroup does not appear in the conclusion of the theorem.
\end{remark}
We now extend the construction of Theorem~\ref{thm-NLPLPpackingproduct} using Lemmas~\ref{lem-LPpackingunion} and~\ref{lem-NLPpackingunion}.
\begin{corollary}\label{coro-NLPLPpackingproduct}
Suppose there exists a $(c_1,t_1-1)$ NLP-packing in an abelian group $G_1$ of order $t_1^2c_1^2$, and there exists a $(c_2,t_2)$ LP-packing in an abelian group $G_2$ of order $t_2^2c_2^2$ relative to a subgroup $U_2$ of order~$t_2c_2$.
\begin{enumerate}[$(i)$]
\item
Suppose that $t_1$ divides~$t_2$. Then there exists a $(t_2c_1c_2,t_1-1)$ NLP-packing in $G_1 \times G_2$.
\item
Suppose that $t_2$ divides~$t_1$. Then there exists a $(t_1c_1c_2,t_2-1)$ NLP-packing in $G_1 \times G_2$.
\end{enumerate}
\end{corollary}
\begin{proof}
\quad
\begin{enumerate}[$(i)$]
\item
By Lemma~\ref{lem-LPpackingunion} with $s=\frac{t_2}{t_1}$, there exists a $(\frac{t_2}{t_1}c_2,t_1)$ LP-packing in $G_2$ relative to $U_2$. Then by Theorem~\ref{thm-NLPLPpackingproduct}, there exists a $(t_2c_1c_2,t_1-1)$ NLP-packing in $G_1 \times G_2$.
\item
By Lemma~\ref{lem-NLPpackingunion} with $s=\frac{t_1}{t_2}$, there exists a $(\frac{t_1}{t_2}c_1,t_2-1)$ NLP-packing in $G_1$. Then by Theorem~\ref{thm-NLPLPpackingproduct}, there exists a $(t_1c_1c_2,t_2-1)$ NLP-packing in $G_1 \times G_2$.
\end{enumerate}
\end{proof}
The product constructions of Theorems~\ref{thm-LPpackingproduct} and~\ref{thm-NLPLPpackingproduct} both combine two packings to form a packing in the direct product of the starting groups. Theorem~\ref{thm-LPpackingproduct} combines two LP-packings to form an LP-packing, whereas Theorem~\ref{thm-NLPLPpackingproduct} combines an NLP-packing and an LP-packing to form an NLP-packing.
We note in passing that we can likewise combine two NLP-packings to form a collection of regular Latin square type PDSs that partitions the nonidentity elements of the product group.
\begin{proposition}[{\cite[Thm.~2.3]{P08DCC}}]\label{prop-NLPpackingproduct}
For $j=1,2$, suppose there exists a $(c_j,t-1)$ NLP-packing in an abelian group $G_j$ of order~$t^2c_j^2$. Then there exists a collection $\{K_1,\dots,K_{t-1}\}$ of $t-1$ disjoint regular $(t^2c_1c_2,tc_1c_2)$ Latin square type PDSs in $G_1 \times G_2$ for which $G_1 \times G_2 -1_{G_1 \times G_2}-\sum_{\ell=1}^{t-1}K_\ell$ is a regular $(t^2c_1c_2,tc_1c_2+1)$ Latin square type PDS in~$G_1 \times G_2$.
\end{proposition}
\begin{proof}[Proof (Outline)]
For $j=1,2$, let $\{P_{j,0},\dots,P_{j,t-2}\}$ be a $(c_j,t-1)$ NLP-packing in~$G_j$ and define $P_{j,t-1} = G_j -\sum_{i=0}^{t-2}P_{j,i}$. Define
\[
K_\ell = \sum_{i=0}^{t-1} P_{1,i}P_{2,i+\ell} \quad \mbox{for $\ell = 1,\dots,t-1$},
\]
where the subscript $i+\ell$ is reduced modulo~$t$.
Model the rest of the proof on that of Theorem~\ref{thm-NLPLPpackingproduct},
distinguishing the case that $\chi|_{G_1}$ is principal on $G_1$ from the case that $\chi|_{G_1}$ is nonprincipal on~$G_1$, and applying Lemma~\ref{lem-LatinPDS}.
\end{proof}
\begin{remark}\label{rem-productconstructions}
In the construction of Proposition~\ref{prop-NLPpackingproduct}, if we could identify an order $t^2c_1c_2$ subgroup $U$ of $G_1 \times G_2$ contained in $G_1 \times G_2 - \sum_{\ell=1}^{t-1}K_\ell$ for which
$G_1 \times G_2 -U - \sum_{\ell=1}^{t-1}K_\ell$ is a regular $(t^2c_1c_2,tc_1c_2)$ Latin square type PDS in $G_1 \times G_2$, then we could conclude that $\{ K_1, \ldots, K_{t-1}, G_1 \times G_2 - U - \sum_{\ell=1}^{t-1}K_\ell \}$ is a $(tc_1c_2,t)$ LP-packing in $G_1 \times G_2$ relative to $U$. However, we do not know how to identify such a subgroup~$U$ (see also Remark~\ref{rem-identifyU}).
\end{remark}
We now apply Theorem~\ref{thm-NLPLPpackingproduct} to the small NLP-packings of Example~\ref{ex-NLPpackingsmall} and LP-packings from Corollaries~\ref{coro-LPpackingeven} and~\ref{coro-LPpackingodd} to produce two infinite families of NLP-packings.
\begin{corollary}\label{coro-NLPpacking}
Let $u,w$ be nonegative integers. Let $s_3,\dots,s_v$ be nonnegative integers (possibly all zero).
\begin{enumerate}[$(i)$]
\item
Let $u \ne 1$, let $w \not\in \{1,2,3\}$ and let each $s_i \ne 1$.
Let $G$ be either $\Z_2^{2u+4} \times \Z_4^w \times \prod_{i=3}^v \Z_{2^i}^{2s_i}$ or $\Z_2^{2u} \times \Z_4^{w+2} \times \prod_{i=3}^v \Z_{2^i}^{2s_i}$, and let $n=\sqrt{|G|}$.
Then there exists an $\big(\frac{n}{4},3\big)$ NLP-packing in~$G$.
\item
Let $u \ne 0$ and let $w \ne 1$. Let $G = \Z_3^{2u} \times \Z_9^w \times \prod_{i=3}^v \Z_{3^i}^{2s_i}$ and let $n = \sqrt{|G|}$.
Then there exists an $\big(\frac{n}{3},2\big)$ NLP-packing in~$G$.
\end{enumerate}
\end{corollary}
\begin{proof}
\quad
\begin{enumerate}[$(i)$]
\item
In the case that $u,w$, and each $s_i$ is zero, the result is given by Example~\ref{ex-NLPpackingsmall}~$(i),(ii)$. Otherwise, let $G' = \Z_2^{2u} \times \Z_4^w \times \prod_{i=3}^v \Z_{2^i}^{2s_i}$ and construct an $\big(\frac{n/4}{4},4\big)$ LP-packing in $G'$ relative to a subgroup of order~$n/4$.
To do so when $w$ is even, take $p=2$ and $(s_1,s_2) = (u,\frac{w}{2})$ and $j = m-2 \ge 0$ in Corollary~\ref{coro-LPpackingeven}.
To do so when $w$ is odd, note that $\s(2,w) = \frac{w-1}{2} \ge 2$ and take $p=2$ and $(s_1,s_2) = (u,0)$ and $j=m-2 \ge 0$ in Corollary~\ref{coro-LPpackingodd}.
Now let $H$ be either $\Z_2^4$ or $\Z_4^2$. Use Theorem~\ref{thm-NLPLPpackingproduct} to combine the $(1,3)$ NLP-packing in~$H$ given by Example~\ref{ex-NLPpackingsmall}~$(i),(ii)$ with the constructed $\big(\frac{n/4}{4},4\big)$ LP-packing in $G'$ to produce the required $\big(\frac{n}{4},3\big)$ NLP-packing in $H \times G'$.
\item
In the case that $u=1$, and $w$ and each of the $s_i$ is zero, the result is given by Example~\ref{ex-NLPpackingsmall}~$(iii)$.
Otherwise, let $G' = \Z_3^{2u-2} \times \Z_9^w \times \prod_{i=3}^v \Z_{3^i}^{2s_i}$ and construct an $\big(\frac{n/3}{3},3\big)$ LP-packing in $G'$ relative to a subgroup of order $n/3$.
To do so when $w$ is even, take $p=3$ and
$(s_1,s_2) = (u-1,\frac{w}{2})$ and $j = m-1 \ge 0$ in Corollary~\ref{coro-LPpackingeven}.
To do so when $w$ is odd, note that $\s(3,w) \ge 1$ and take $p = 3$ and $(s_1,s_2) = (u-1,0)$ and $j=m-1 \ge 0$ in Corollary~\ref{coro-LPpackingodd}.
Now use Theorem~\ref{thm-NLPLPpackingproduct} to combine the $(1,2)$ NLP-packing in~$\Z_3^2$ given by Example~\ref{ex-NLPpackingsmall}~$(iii)$ with the constructed $\big(\frac{n/3}{3},3\big)$ LP-packing in $G'$ to produce the required $\big(\frac{n}{3},2\big)$ NLP-packing in~$G$.
\end{enumerate}
\end{proof}
\begin{remark}\label{rem-PDSproduct}
Using Remark~\ref{rem-NLPpacking}~$(i)$, we can rephrase Result~\ref{res-PDSproductNLP} in terms of $(\frac{n}{t},t-1)$ NLP-packings in groups of order~$n^2$. As we now describe, we then see that Corollary~\ref{coro-NLPpacking} is significantly more general than Result~\ref{res-PDSproductNLP}.
Result~\ref{res-PDSproductNLP}~$(i)$ and Corollary~\ref{coro-NLPpacking}~$(i)$ each construct NLP-packings with parameters $\big(\frac{n}{4},3\big)$. The former result is restricted to groups of the form $\Z_2^{4u} \times \Z_4^{2w} \times \prod_{i=2}^v \Z_{2^{2i}}^{4s_{2i}}$ for $u+w\ge 1$, whereas the latter result applies to the forms
$\Z_2^{2u+4} \times \Z_4^w \times \prod_{i=3}^v \Z_{2^i}^{2s_i}$ and
$\Z_2^{2u} \times \Z_4^{w+2} \times \prod_{i=3}^v \Z_{2^i}^{2s_i}$
for $u \ne 1$ and $w \not\in \{1,2,3\}$ and each $s_i \ne 1$.
Similarly, Result~\ref{res-PDSproductNLP}~$(ii)$ and Corollary~\ref{coro-NLPpacking}~$(ii)$ each construct NLP-packings with parameters $\big(\frac{n}{3},2\big)$. The former result is restricted to groups of the form
$\Z_3^{2u+2} \times \prod_{i=1}^v \Z_{3^{2i}}^{2s_{2i}}$,
whereas the latter result applies to the form
$\Z_3^{2u} \times \Z_9^w \times \prod_{i=3}^v \Z_{3^i}^{2s_i}$ for $u \ne 0$ and $w \ne 1$.
\end{remark}
\section{Relationship with Strongly Regular Bent Functions}\label{sec7}
In this section, we examine how LP-packings and NLP-packings are related to the hyperbolic and elliptic strongly regular bent functions introduced by Chen and Polhill in~\cite{CP13}.
\begin{definition}[{\cite[Thm.~3.7]{CP13}}]\label{def-srbf}
Let $G$ and $H$ be abelian groups. Let $f$ be a surjective function from $G$ to $H$ satisfying $f(1_G)=1_H$, and let $f^{-1}$ be the pre-image of~$f$.
\begin{enumerate}[$(i)$]
\item
$f$ is a \emph{hyperbolic strongly regular bent function} if
$f^{-1}(h)$ is a regular $\big(\sqrt{|G|},\frac{\sqrt{|G|}}{|H|}\big)$ Latin square type PDS for each $h \ne 1_H$, and
$f^{-1}(1_H)-1_G$ is a regular $\big(\sqrt{|G|},\frac{\sqrt{|G|}}{|H|}+1\big)$ Latin square type PDS.
\item
$f$ is an \emph{elliptic strongly regular bent function} if
$f^{-1}(h)$ is a regular $\big(\sqrt{|G|},\frac{\sqrt{|G|}}{|H|}\big)$ negative Latin square type PDS for each $h \ne 1_H$, and
$f^{-1}(1_H) - 1_G$ is a regular $\big(\sqrt{|G|},\frac{\sqrt{|G|}}{|H|}-1\big)$ negative Latin square type PDS.
\end{enumerate}
\end{definition}
We can relate hyperbolic strongly regular bent functions to LP-packings, and elliptic strongly regular bent functions to NLP-packings.
\begin{proposition}\label{prop-bent}
Let $G$ be an abelian group of order $t^2c^2$, let $U$ be a subgroup of $G$ of order~$tc$, and let $H$ be an abelian group of order~$t$.
\begin{enumerate}[$(i)$]
\item
If there exists a $(c,t)$ LP-packing in $G$ relative to~$U$, then there exists a hyperbolic strongly regular bent function from $G$ to~$H$.
\item
The existence of a $(c,t-1)$ NLP-packing in $G$ is equivalent to the existence of an elliptic strongly regular bent function from $G$ to~$H$.
\end{enumerate}
\end{proposition}
\begin{proof}
Write $H=\{1_H,h_1,h_2,\ldots,h_{t-1}\}$.
\begin{enumerate}[$(i)$]
\item
Let $\{P_1,\dots,P_t\}$ be a $(c,t)$ LP-packing in $G$ relative to~$U$, and define the surjective function $f:G \rightarrow H$ by
\[
f(g)=\begin{cases}
h_i & \mbox{if $g \in P_i$ for $i \in \{1,\dots,t-1\}$,} \\
1_H & \mbox{if $g \in P_t + U$}.
\end{cases}
\]
Then by Remark~\ref{rem-LPpacking}~$(i)$, $f$ is a hyperbolic strongly regular bent function.
\item
Suppose firstly that $\{P_1,\dots,P_{t-1}\}$ is a $(c,t-1)$ NLP-packing in $G$, and define the surjective function $f:G \rightarrow H$ by
\[
f(g)=\begin{cases}
h_i & \mbox{if $g \in P_i$,} \\
1_H & \mbox{otherwise}.
\end{cases}
\]
Then by Definition~\ref{def-NLPpacking}, $f$ is an elliptic strongly regular bent function from $G$ to~$H$.
Conversely, suppose that $f:G \rightarrow H$ is an elliptic strongly regular bent function, and define subsets $P_1,\dots,P_{t-1}$ of $G$ by
\[
P_i = f^{-1}(h_i) \quad \mbox{for $i=1,\dots,t-1$}.
\]
Then by Definition~\ref{def-NLPpacking}, $\{P_1,\dots,P_{t-1}\}$ is a $(c,t-1)$ NLP-packing in~$G$ from $G$ to~$H$.
\end{enumerate}
\end{proof}
\begin{remark}\label{rem-identifyU}
Proposition~\ref{prop-bent} shows the equivalence of an NLP-packing and an elliptic strongly regular bent function, but not the equivalence of an LP-packing and a hyperbolic strongly regular bent function~$f$.
In order to establish the converse to Proposition~\ref{prop-bent}~$(i)$, we would need to identify an order $\sqrt{|G|}$ subgroup $U$ of $G$ contained in $f^{-1}(1_H)$ for which $f^{-1}(1_H)-U$ is a regular $\big(\sqrt{|G|},\frac{\sqrt{|G|}}{|H|}\big)$ Latin square type PDS in~$G$.
The additional conditions on $f$ required to guarantee the existence of such a subgroup~$U$ could be substantial, as we discuss in Example~\ref{ex-ternarybent}~$(i)$ below.
\end{remark}
The following example uses strongly regular bent functions to construct collections of PDSs, and provides an alternative derivation of a special case of Corollary~\ref{coro-NLPpacking}~$(ii)$.
\begin{example}[{\rm \cite[Thm.~1, Cor.~3]{TPF}}]\label{ex-ternarybent}
Let $p$ be a prime, and let $\zp$ be a primitive $p^{\rm {th}}$ root of unity.
Let $g$ be a function from $\F_{p^n}$ to~$\F_p$, and let
$\Tr$ be the trace function from $\F_{p^n}$ to~$\F_p$.
The \emph{Walsh transform} $\cW_g : \F_{p^n} \rightarrow \CC$ of $g$ is given by
$$
\cW_g(b)=\sum_{x \in \F_{p^n}} \zp^{g(x)+\Tr(bx)}.
$$
The function $g$ is \emph{bent} if
$|\cW_g(b)|=p^{\frac{n}{2}}$ for all $b \in \F_{p^n}$.
A bent function $g$ is \emph{weakly regular} if there exists a function $g^*: \F_{p^n} \rightarrow \F_p$ satisfying
\[
\cW_g(b)=\mu p^{\frac{n}{2}}\zp^{g^*(b)} \quad \mbox{for some $\mu \in \CC$ with $|\mu|=1$ and for all $b \in \F_{p^n}$},
\]
and is \emph{regular} (a much stronger condition) if there exists a function $g^*: \F_{p^n} \rightarrow \F_p$ such that $\cW_g(b)= p^{\frac{n}{2}}\zp^{g^*(b)}$ for all $b \in \F_{p^n}$.
Suppose $f: \F_{3^{2s}} \rightarrow \F_3$ is a weakly regular
bent function satisfying $f(-x)=f(x)$ and $f(0)=0$.
Then exactly one of the following holds:
\begin{enumerate}[$(i)$]
\item
$|f^{-1}(0)|=3^{2s-1}+2\cdot3^{s-1}$, and $f$ is a hyperbolic strongly regular bent function from $\F_{3^{2s}}$ to~$\F_3$:
each of $f^{-1}(1)$ and $f^{-1}(2)$ is a regular $(3^s,3^{s-1})$ Latin square type PDS in $\F_{3^{2s}}$, and
$f^{-1}(0) \sm \{0\}$ is a regular $(3^s,3^{s-1}+1)$ Latin square type PDS in $\F_{3^{2s}}$.
\item
$|f^{-1}(0)|=3^{2s-1}-2\cdot3^{s-1}$, and $f$ is an elliptic strongly regular bent function from $\F_{3^{2s}}$ to $\F_3$:
each of $f^{-1}(1)$ and $f^{-1}(2)$ is a regular $(3^s,3^{s-1})$ negative Latin square type PDS in $\F_{3^{2s}}$, and
$f^{-1}(0) \sm \{0\}$ is a regular $(3^s,3^{s-1}-1)$ negative Latin square type PDS in $\F_{3^{2s}}$;
equivalently, $\big\{f^{-1}(1), f^{-1}(2)\big\}$ is a
$(3^{s-1},2)$ NLP-packing in~$(\F_{3^{2s}},+) = \Z_3^{2s}$.
\end{enumerate}
Following Remark~\ref{rem-identifyU}, we mention that in order to obtain a $(3^{s-1},3)$ LP-packing in the group $(\F_3^{2s},+) = \Z_3^{2s}$ relative to $\Z_3^s$ from the hyperbolic strongly regular bent function $f$ of $(i)$, we would firstly need to identify an order~$3^s$ subgroup $U$ of $\Z_3^{2s}$ contained in $f^{-1}(0)$, and secondly require $f^{-1}(0)-U$ to be a regular $(3^s,3^{s-1})$ Latin square type PDS in~$\Z_3^{2s}$. The first condition alone requires $f$ to be normal {\rm\cite[p.~106]{CMP}}, which in combination with the property that $f$ is weakly regular requires in turn that $f$ should be regular {\rm\cite[Thm.~6]{CMP}}.
This demonstrates that, in general, an LP-packing is a significantly more constrained structure than a hyperbolic strongly regular bent function.
\end{example}
\begin{remark}\label{rem-stronglyregular}
\quad
\begin{enumerate}[$(i)$]
\item
In {\rm \cite[Lemma 3.15]{CP13}}, an incomplete attempt was made to lift a hyperbolic strongly regular bent function on $\Z_p^{2a}$ to a hyperbolic strongly regular bent function on $\Z_{p^2}^{2a}$. We were able to accomplish this lifting in Theorem~\ref{thm-LPpackinguniform}, using the more constrained structure of an LP-packing. This underscores the importance of identifying the role of the subgroup $U$ in the definition of an LP-packing.
\item
{\rm Corollaries~\ref{coro-LPpackingeven}, ~\ref{coro-LPpackingodd} and~\ref{coro-NLPpacking}} recover the hyperbolic and elliptic strongly regular bent functions of {\rm\cite[Prop.~3.13]{CP13}} and strengthen those of {\rm\cite[Prop.~3.14]{CP13}}. The hyperbolic and elliptic strongly regular bent functions constructed in {\rm\cite[Prop.~3.12]{CP13}} have codomain of size~$2$, and are equivalent to reversible Hadamard difference sets. We examine this connection in more detail in Section~\ref{sec8} below.
\end{enumerate}
\end{remark}
\section{Relationship with reversible Hadamard difference sets}\label{sec8}
In this section, we examine how LP-packings and NLP-packings are closely related to reversible Hadamard difference sets.
\begin{definition}\label{def-HDS}
Let $G$ be a group $G$ of order~$v$.
A \emph{$(v,k,\la)$-difference set in $G$} is a $(v,k,\la,\la)$ partial difference set in~$G$.
A \emph{Hadamard difference set} has parameters
$(v,k,\la) = \big(4c^2,c(2c-1),c(c-1)\big)$ for some integer~$c$.
A difference set $D$ is \emph{reversible} if $D=D^{(-1)}$.
\end{definition}
\noindent
By Definition~\ref{def-pds}, a reversible $(v,k,\la)$-difference set that does not contain~$1_G$ is the same as a regular $(v,k,\la,\la)$ partial difference set.
The following result is a direct consequence of Definition~\ref{def-NLPpacking} and Lemmas~\ref{lem-pdschar} and~\ref{lem-LatinPDS}.
\begin{lemma}\label{lem-NLPrevHDS}
Let $c$ be a positive integer, and let $P$ be a $c(2c-1)$-subset of an abelian group $G$ of order~$4c^2$.
Then the following statements are equivalent.
\begin{enumerate}[$(i)$]
\item
$P$ is a $\big(4c^2,c(2c-1),c(c-1)\big)$ reversible Hadamard difference set in~$G$, and $1_G \notin P$
\item
$\chi(P) \in \{-c,c\}$ for all nonprincipal characters $\chi$ of $G$
\item
$\{G-P\}$ is a $(c,1)$ NLP-packing in~$G$.
\end{enumerate}
\end{lemma}
The next result follows from Definition~\ref{def-LPpacking} and Lemmas~\ref{lem-LatinPDS} and~\ref{lem-NLPrevHDS}.
\begin{lemma}\label{lem-LPpackingHDS}
Let $c$ be a positive integer,
let $G$ be an abelian group of order $4c^2$, and let $U$ be a subgroup of order~$2c$. Then $P_1,P_2$ are disjoint $\big(4c^2,c(2c-1),c(c-1)\big)$ reversible Hadamard difference sets in $G$ whose union is~$G-U$ if and only if $\{P_1,P_2\}$ is a $(c,2)$ LP-packing in $G$ relative to~$U$.
\end{lemma}
We now use LP-packings in $2$-groups to produce infinite families of reversible Hadamard difference sets.
\begin{corollary}\label{cor-revHDS}
Let $u,w$ and $s_3,\dots,s_v$ be nonnegative integers (not all zero).
For each $i =3,\dots,v$, let $G_i=\Z_{2^i}^{2s_i}$ and let $U_i$ be a subgroup of $G_i$ of order~$2^{is_i}$.
Let $G = \Z_2^{2u} \times \Z_4^w \times \prod_{i=3}^v G_i$, let $U = \Z_2^u \times (2\Z_4)^w \times \prod_{i=3}^v U_i$, and let $|G| = 4c^2$. Then
\begin{enumerate}[$(i)$]
\item
for $w \ne 1$, there exist two disjoint
$\big(4c^2,c(2c-1),c(c-1)\big)$ reversible Hadamard difference sets in $G$ whose union is~$G-U$.
\item
there exists a $\big(4c^2,c(2c-1),c(c-1)\big)$ reversible Hadamard difference set in~$G$.
\end{enumerate}
\end{corollary}
\begin{proof}
\quad
\begin{enumerate}[$(i)$]
\item
By Lemma~\ref{lem-LPpackingHDS}, an equivalent statement is that for $w \ne 1$ there exists a $(c,2)$ LP-packing in $G$ relative to~$U$. To show this when $w$ is even, take $p=2$ and $(s_1,s_2)=(u,\frac{w}{2})$ and $j=m-1 \ge 0$ in Corollary~\ref{coro-LPpackingeven}.
To show this when $w$ is odd, note that $\s(2,w) = \frac{w-1}{2} \ge 1$ and take
$p=2$ and $(s_1,s_2) = (u,0)$ and $j=m-1 \ge 0$ in Corollary~\ref{coro-LPpackingodd}.
\item
In view of part $(i)$, we may assume $w=1$.
In the case that $u$ and each of the $s_i$ is zero, there is a trivial reversible Hadamard difference set $D$ in $\Z_4$ comprising just the identity element.
Otherwise, take $p=2$ and $(s_1,s_2)=(u,0)$ and $j=m-1\ge 0$ in Corollary~\ref{coro-LPpackingeven} to produce a $(\frac{c}{2},2)$ LP-packing in
$\Z_2^{2u} \times \prod_{i=3}^v G_i$ relative to some subgroup of order~$c$. Use Theorem~\ref{thm-NLPLPpackingproduct} to combine this with the $(1,1)$ NLP-packing $\{\Z_4-D\}$ in $\Z_4$ (as given by Lemma~\ref{lem-NLPrevHDS}) to produce a $(c,1)$ NLP-packing in~$G$, then apply Lemma~\ref{lem-NLPrevHDS}.
\end{enumerate}
\end{proof}
\begin{remark}\label{rem-revHDS}
\quad
\begin{enumerate}[$(i)$]
\item
Corollary~\ref{cor-revHDS}~$(i)$ significantly extends a previous construction for $G=\Z_{2^a}^2$ and noncyclic $U$ of order~$2^a$ {\rm \cite[Thm.~3.1]{DP}}.
\item
All abelian $2$-groups in which a reversible Hadamard difference set is known to exist are recovered by Corollary~\ref{cor-revHDS}~$(ii)$; several other families of abelian groups are also known to contain reversible Hadamard difference sets (see {\rm\cite[Chap.~VI, Sect.~14]{BJL}}), and all these groups are recovered by {\rm \cite[Prop.~3.12]{CP13}}) using elliptic strongly regular bent functions.
\item
The construction of linking systems of reversible Hadamard difference sets in {\rm \cite[Thm.~5.3]{DMP}} depends on the existence of collections of disjoint partial difference sets in abelian $2$-groups satisfying mutual structural properties. After some modification, we can reinterpret this construction as combining a $(2^{(a-1)s},2^s)$ LP-packing in an abelian group of order $2^{2as}$, relative to some subgroup of order~$2^{as}$, with the complement of a $(2^s-1,2^{s-1}-1,2^{s-2}-1)$-difference set in $\Z_{2^s-1}$ {\rm \cite[Chap.~VI, Thm.~1.10]{BJL}} for $s > 1$ and $a > 0$.
\end{enumerate}
\end{remark}
\section{Open Problems}\label{sec9}
In this section, we propose some open problems for future research.
\begin{enumerate}[P1.]
\item
The recursive constructions of LP-packings described in Section~\ref{sec5} rest on two base cases: a $(1,p^s)$ LP-packing in $\Z_p^{2s}$ relative to $\Z_p^s$ derived from a spread (Example~\ref{ex-spread}), and the LP-packings in $p$-groups of odd rank given by~Result~\ref{res-HLX}. Find more base cases, especially in groups of odd rank (when a spread does not exist).
\item
Section~\ref{sec6} combines small examples of NLP-packings with the constructed families of LP-packings to produce families of NLP-packings. Find new small examples of NLP-packings.
\item
All known $(c,t)$ LP-packings with $t > 2$, and all known $(c,t)$ NLP-packings with $t>1$, occur in $p$-groups (see Section~\ref{sec8} for discussion of $(c,2)$ LP-packings and $(c,1)$ NLP-packings.)
Find examples in groups other than $p$-groups.
\item
Nonexistence results for LP-packings and NLP-packings can be obtained from known nonexistence results for single PDSs (see \cite{W}, for example).
We can also obtain nonexistence results for these packings from known nonexistence results for reversible Hadamard difference sets \cite[Chap.~VI, Sect.~14]{BJL}, by applying Lemmas~\ref{lem-LPpackingunion} and~\ref{lem-LPpackingHDS} to $(c,2t)$ LP-packings, and Lemmas~\ref{lem-NLPpackingunion} and~\ref{lem-NLPrevHDS} to $(c,2t-1)$ NLP-packings. Establish further nonexistence results for LP-packings and NLP-packings.
\item
Construct hyperbolic strongly regular bent functions satisfying additional conditions in order to produce LP-packings in new groups (see Remark~\ref{rem-identifyU}).
\item
Are there collections of partial difference sets with advantageous structure, other than LP-packings and NLP-packings, that can be analyzed using methods similar to those of this paper?
\end{enumerate}
\section*{Acknowledgements}
The authors gratefully acknowledge helpful discussions with Samuel Simon at the outset of this research.
We thank Alexander Pott for his hospitality when the first author visited the second at Otto von Guericke University, Magdeburg in Summer 2018 to pursue this research. We thank the anonymous reviewers for their detailed and helpful comments.
\bibliographystyle{amsplain}
\providecommand{\bysame}{\leavevmode\hbox to3em{\hrulefill}\thinspace}
\providecommand{\MRno}{\relax\ifhmode\unskip\space\fi MR }
\providecommand{\MRhref}[2]
\href{http://www.ams.org/mathscinet-getitem?mr=#1}{#2}
}
\providecommand{\href}[2]{#2}
\providecommand{\MR}[1]{\MRhref{#1}{\MRno{#1}}}
|
\section{Introduction}
In the design of control systems for industrial applications, it is important to achieve a certain level of fault tolerance. There has been a growing interest in the literature recently on developing effective fault tolerant paradigms for reliable control of real-world systems \cite{Frank2004}. This type of control system is particularly useful when the system is subject to unpredictable failures. Recent applications of fault-tolerant control include power systems and aircraft flight control systems \cite{odgaard2014fault,Verhaegen2010}.
It is known that certain class of faults can be modeled as partially observable Markov decision processes (POMDP). Various methods are studied in the literature to find an approximate solution to POMDP. Grid-based methods are used in \cite{lovejoy1991computationally} to compute an approximate value function at a fixed number of points in the belief space, and then interpolate over the entire space. The advantage of such approaches is that their computational complexity remains unchanged at each iteration and does not increase with time. Their drawback, however, is that the fixed points may not be reachable. In point-based methods \cite{Shani2013survey}, the reachability drawback is circumvented by restricting attention to the reachable set. Using the notion of $\alpha$-vectors, an approximate value function is obtained iteratively over a finite number of points in the reachable set. In these methods, the points are not fixed and may change with the value function. In policy-search methods such as finite-state controllers \cite{hansen1998finite}, on the other hand, attention is devoted to a certain class of strategies, and the objective is to find the best strategy in that class using policy iteration and gradient-based techniques. For more details on POMDP solvers, the interested reader is referred to \cite{Cassandra1998survey, Shani2013survey,Murphy2000survey}, and references therein.
In this paper, a fault-tolerant scheme is proposed for a system consisting of a number of homogeneous components, where each component may fail with a certain probability. Three courses of action are defined to troubleshoot the faulty system: (i) let the system operate with faulty components; (ii) inspect the system, and (iii) repair the system. Each course of action has an implementation cost. The problem is formulated as a POMDP but since finding an optimal solution for this problem is intractable, in general, we are interested in seeking a near-optimal solution for it~\cite{madani2003undecidability}. However, identifying an $\epsilon$-optimal solution for this problem is also NP-hard \cite{meuleau1999solving}. To overcome this hurdle, we exploit the structure of the problem to use a different information state (that is smaller than the belief state). The computational complexity of the proposed solution is logarithmic with respect to the desired neighborhood $\epsilon$, and polynomial with respect to the number of components. To derive some of the results, we use some methods developed in \cite{JalalACC2018,arabneydi2016new}.
This paper is organized as follows. The problem is formally stated in Section~\ref{sec:problem} and the main results of the work are presented in the form of three theorems in Section~\ref{sec:main}. Two numerical examples are provided in Section~\ref{sec:examples} and some concluding remarks are given in Section~\ref{sec:conclusions}.
\section{Problem Formulation}\label{sec:problem}
Throughout this paper, $\mathbb{R}$ and $\mathbb{N}$ refer, respectively, to real and natural numbers. For any $n \in \mathbb{N}$, $\mathbb{N}_n$ denotes the finite set $\{1,\ldots,n\}$. Moreover, $\Prob{\boldsymbol \cdot}$ is the probability of a random variable, $\Exp{\boldsymbol \cdot}$ is the expectation of a random variable, and $\ID{\boldsymbol \cdot }$ is the indicator function. The shorthand notation $x_{a:b}$ denotes vector $(x_a,\ldots,x_b)$, $a,b \in \mathbb{N}, a \leq b$. For any finite set $\mathcal{X}$, $\mathcal{P}(\mathcal{X})$ denotes the space of probability measures on $\mathcal{X}$. For any $y,n \in \mathbb{N}$, $y \leq n$, and $p \in [0,1]$, $\Binopdf(y,n,p)$ is the Binomial probability distribution function of $y$ successful outcomes from $n$ trials where the success probability is~$p$.
Consider a stochastic dynamic system consisting of $n \in \mathbb{N}$ internal components. Denote by $x^i_t \in \mathcal{X}:=\{0,1\}$ the state of component $i \in \mathbb{N}_n$ at time $t \in \mathbb{N}$, where $x^i_t=0$ means that the $i$-th component is in the operating mode and $x^i_t=1$ means that it is faulty. If a component is faulty, it remains so until it is repaired. Let $p \in [0,1]$ denote the probability that a component becomes faulty at any time $t \in \mathbb{N}$. It is assumed that the probability of failure of each component is independent of others.
Denote by $m_t=\sum_{i=1}^n x^i_t \in \mathcal{M}:= \{0,1,\ldots,n\}$ the number of faulty components at time $t \in \mathbb{N}$, and note that the state of each component may not be directly available. Let $o_t \in \mathcal{O}= \mathcal{M} \cup \{ \mathtt{blank} \}$ denote the number of faulty components at time $t \in \mathbb{N}$ that are observed. If there is no observation at time $t \in \mathbb{N}$, then $o_t= \mathtt{blank} $. Initially, the system is assumed to have no faulty components, i.e. $o_1=0$.
At any time $t \in \mathbb{N}$, we have three different options (actions) at our disposal, represented by $u_t \in \mathcal{U}$, where $\mathcal{U}=\mathbb{N}_3$ is the action set. The first option is to do nothing and let the system continue operating without disruption at no implementation cost. In this case, no new information on the number of faulty components is collected, i.e.,
\begin{equation}\label{eq:o_1}
\Prob{o_{t+1}= \mathtt{blank} | m_t=m, u_t=1}=1, \quad m \in \mathcal{M},t \in \mathbb{N}.
\end{equation}
The second option is to inspect the system and detect the number of faulty components at some inspection cost, where
\begin{equation}\label{eq:o_2}
\Prob{o_{t+1}=m | m_{t+1}=m, u_t=2}=1, \quad m \in \mathcal{M}, t \in \mathbb{N}.
\end{equation}
The third option is to repair the faulty components at a cost depending on the number of them, i.e. $m_t$. Therefore, at any time $t \in \mathbb{N}$, the following relations hold:
\begin{align}\label{eq:relation_prob}
\Prob{x^i_{t+1}=1 | x^i_t=0, u_{t} \in \{ 1,2\}}&=p, \nonumber \\
\Prob{x^i_{t+1}=1 | x^i_t=1, u_{t} \in \{ 1,2\}}&=1, \nonumber \\
\Prob{x^i_{t+1}=1 | x^i_t=0, u_{t}=3}&=0, \nonumber \\
\Prob{x^i_{t+1}=1 | x^i_t=1, u_{t}=3}&=0.
\end{align}
Let $c(m,u): \mathcal{M}\times \mathcal{U} \rightarrow \mathbb{R}$ denote the cost associated with action $u \in \mathcal{U}$ when the number of faulty components is $m \in \mathcal{M}$. The strategy $g$ is defined as the mapping from the available information by time $t$ to an action in $\mathcal{U}$, i.e.,
\begin{equation}
u_t=g(o_{1:t},u_{1:t-1}).
\end{equation}
The objective is to develop a cost-efficient fault-tolerant strategy in the sense that the system operates with a relatively small number of faulty components, taking the inspection and repair costs into account. To this end, given the discount factor $\beta \in (0,1)$, we define the following cost:
\begin{equation}\label{eq:total_cost}
J(g)=\Exp{\sum_{t=1}^\infty \beta^{t-1} c(m_t,u_t)}.
\end{equation}
\section{Main Result}\label{sec:main}
To present the main result of this paper, we first derive a Bellman equation to identify the optimal solution. Since the corresponding Bellman equation involves an intractable optimization problem, we subsequently present an alternative Bellman equation that is tractable and provides a near-optimal solution. For any $m \in \mathcal{M} \backslash \{0,n\}$, define the following vector-valued function $\phi: \mathcal{M} \backslash \{0,n\} \rightarrow \mathcal{P}(0,1,.\ldots,n)$:
\begin{equation}
\phi(m, \boldsymbol \cdot):=\Binopdf(\boldsymbol \cdot, m,1) \ast \Binopdf(\boldsymbol \cdot, n-m,p),
\end{equation}
where $\ast$ denotes the convolution operator.
\begin{theorem}\label{thm:1}
Given any realization $m_{1:t}$ and $u_{1:t}$, $t \in \mathbb{N}$, the transition probability matrix of the number of faulty components can be computed as follows:
\begin{equation}
\Prob{m_{t+1}|m_{1:t},u_{1:t}}= \Prob{m_{t+1}|m_{t},u_{t}},
\end{equation}
where for any $y \in \{0,1,\ldots,n\}$,
\begin{itemize}
\item if $m_t=0$ and $u_t \in \{1,2\}$, then
\[\Prob{m_{t+1}=y|m_t,u_t}= \Binopdf(y,n,p),\]
\item if $m_t=n$ and $u_t \in \{1,2\}$, then
\[\Prob{m_{t+1}=y|m_t,u_t}= \Binopdf(y,n,1)= \ID{y=n},\]
\item if $m_t \notin \{0,n\}$ and $u_t \in \{1,2\}$, then
\[\Prob{m_{t+1}=y|m_t,u_t }= \phi(m_t, y+1),\]
\item if $u_t=3$, then
\[\Prob{m_{t+1}=y|m_t,u_t }= \ID{y=0}.\]
\end{itemize}
\end{theorem}
\begin{proof}
Define $\hat x^i_{t+1}:= x^i_t x^i_{t+1}$ and $\tilde x^i_{t+1}:= (1-x^i_t) x^i_{t+1}$, $i \in \mathbb{N}_n, t \in \mathbb{N}$. Define also $\mathbf x_t:=(x^1_t,\ldots,x^n_t)$, $t \in \mathbb{N}$. Then, given any realization $\mathbf x_{1:t}$ and $u_{1:t}$, one has
\begin{align}\label{eq:proof_thm1}
\Prob{m_{t+1} | \mathbf x_{1:t}, u_{1:t}}&=\Prob{\sum_{i=1}^n x^i_{t+1} | \mathbf x_{t}, u_{t}} \nonumber \\
&= \Prob{\sum_{i=1}^n \hat x^i_{t+1} + \sum_{i=1}^n \tilde x^i_{t+1}| \mathbf x_{t}, u_{t}}.
\end{align}
On the other hand, one can conclude from the above definitions that $(n-m_t)$ terms of $ \sum_{i=1}^n \hat x^i_{t+1}$ as well as $m_t$ terms of $\sum_{i=1}^n \tilde x^i_{t+1}$ are definitely zero. It is also important to note that $\hat x^i_{t+1}$ and $\tilde x^i_{t+1}$, $t \in \mathbb{N}, i \in \mathbb{N}_n$, are independent Bernoulli random variables with success probability $x^i_t \cdot \Prob{x^i_{t+1}=1|x^i_t=1,u_t}$ and $(1-x^i_t) \cdot \Prob{x^i_{t+1}=1|x^i_t=0,u_t}$, respectively. Therefore, the right-hand side of~\eqref{eq:proof_thm1} is the probability of the sum of $m_t$ i.i.d. Bernoulli random variables with success probability $\Prob{x^i_{t+1}=1|x^i_t=1,u_t}$ and $(n-m_t)$ i.i.d. Bernoulli random variables with success probability $\Prob{x^i_{t+1}=1|x^i_t=0,u_t}$. Since the random variables are independent, the probability of their sum is equal to the convolution of their probabilities. The proof follows from~\eqref{eq:relation_prob}, on noting that $m_{1:t}$ can be represented by $\mathbf{x}_{1:t}$.
\end{proof}
Let $s_t \in \mathcal{M}$ be the last observation before $t \in \mathbb{N}$ that is not \emph{blank} and $z_t \in \mathbb{N} \cup \{0\}$ be the elapsed time associated with it, i.e., the time interval between the observation of $s_t$ and~$t$.
\begin{lemma}\label{lemma:hat-f}
There exists a function $\hat f$ such that
\begin{equation}
(s_{t+1},z_{t+1})=\hat f(s_t,z_t,u_t,o_{t+1}), \quad t \in \mathbb{N},
\end{equation}
where
\begin{equation}
\hat f(s_t,z_t,u_t,o_{t+1}):=\begin{cases}
(s_t, 1+z_t), & u_t=1,\\
(o_{t+1},0), & u_t=2,\\
(0,0), & u_t=3.
\end{cases}
\end{equation}
\end{lemma}
\begin{proof}
The proof follows from the definition of $s_t$ and $z_t$, and equations~\eqref{eq:o_1},~\eqref{eq:o_2} and~\eqref{eq:relation_prob}.
\end{proof}
For the sake of simplicity, denote by $T_{\scriptstyle \mathsf{m}}$ the transition probability matrix of the number of faulty components under actions $\{1,2\}$ given by Theorem~\ref{thm:1}, i.e., for any $m',m \in \mathcal{M}$ and $t \in \mathbb{N}$,
\begin{equation} \label{eq:tm}
T_{\scriptstyle \mathsf{m}}(m',m):=\Prob{m_{t+1}=m' | m_t=m}.
\end{equation}
\begin{lemma}\label{lemma:prob_o}
Given any realization $s_{1:t}$, $z_{1;t}$, $o_{1:t}$ and $u_{1:t}$, $t \in \mathbb{N}$, the following equality holds irrespective of strategy~$g$,
\begin{multline}\label{eq:prob_o}
\Prob{o_{t+1} | s_{1:t}, z_{1:t}, o_{1:t},u_{1:t}}= \ID{u_t=1, o_{t+1}= \mathtt{blank} }\\
\ID{u_t=2} T_{\scriptstyle \mathsf{m}}^{1+z_t}(o_{t+1},s_t)+ \ID{u_t=3,o_{t+1}=0}.
\end{multline}
\end{lemma}
\begin{proof}
The proof follows from equations~\eqref{eq:o_1},~\eqref{eq:o_2},~\eqref{eq:relation_prob} and~\eqref{eq:tm}, and the Chapman–Kolmogorov equation.
\end{proof}
\begin{lemma}
Given any realization $s_{1:t}$, $z_{1;t}$, $o_{1:t}$ and $u_{1:t}$, $t \in \mathbb{N}$, there exists a function $\hat c$ such that
\begin{equation}
\Exp{c(m_t,u_t)| s_{1:t},z_{1:t},o_{1:t}, u_{1:t}}=\hat c(s_t,z_t,u_t),
\end{equation}
where
\begin{equation}\label{eq:hat_c}
\hat c(s_t,z_t,u_t):=\sum_{m \in \mathcal{M}} c(m,u_t) T_{\scriptstyle \mathsf{m}}^{z_t}(m,s_t).
\end{equation}
\end{lemma}
\begin{proof}
The proof follows from the definition of expectation operator, states $(s_t,z_t)$, update function $\hat f$ in Lemma~\ref{lemma:hat-f}, and the Chapman–Kolmogorov equation.
\end{proof}
\begin{theorem}
For any $ s\in \mathcal{M}$ and $z \in \mathbb{N} \cup \{0\}$, define the following Bellman equation
\begin{equation}\label{eq:bellman-optimal}
V(s,z)=\min_{u \in \mathcal{U}} (\hat c(s,z,u) + \beta \Exp{V(\hat f(s,z,u,o))} ),
\end{equation}
where the expectation is taken over observations $ o \in \mathcal{O}$ with respect to the conditional probability function in~\eqref{eq:prob_o}. The optimal strategy for the cost function~\eqref{eq:total_cost} is obtained by solving the above equation.
\end{theorem}
\begin{proof}
The proof follows from the fact that $(s_t,z_t)$, $t \in \mathbb{N}$, is an information state because it evolves in a Markovian manner under control action $u_t$ according to Lemma~\ref{lemma:hat-f}. In addition, the conditional probability~\eqref{eq:prob_o} and expected cost~\eqref{eq:hat_c} do not depend on strategy $g$, and can be represented in terms of state $(s_t,z_t)$ and action $u_t$, $t \in \mathbb{N}$. Thus, the proof is completed by using the standard results from Markov decision theory~\cite{Bertsekas2012book}.
\end{proof}
Since the optimization of the Bellman equation~\eqref{eq:bellman-optimal} is carried out over a countable infinite set, it is computationally difficult to solve it. As a result, we are interested in a strategy which is sufficiently close to the optimal strategy and is tractable. To this end, define the following Bellman equation for any $k \in \mathbb{N}$, $s \in \mathcal{M}$ and $z \in \{0,1,\ldots,k\}$:
\begin{equation}\label{eq:bellman_k}
\hat V_k(s,z)=\min_{u \in \mathcal{U}}(\hat V^u_k(s,z)),
\end{equation}
where
\begin{align}
\hat V^1_k(s,z) \hspace{-.1cm}&:= \hspace{-.1cm} \hat c(s,z,1) \hspace{-.1cm}+\hspace{-.1cm} \beta \hspace{-.05cm}\left( \hspace{-.05cm} \hat V_k(s,1+z) \ID{z \hspace{-.1cm}<\hspace{-.1cm} k} \hspace{-.1cm}+ \hspace{-.1cm} \hat V_k(0,0) \ID{z=k} \hspace{-.1cm} \right) \nonumber \\
\hat V^2_k(s,z)&:= \hat c(s,z,2) + \beta \sum_{m \in \mathcal{M}} T_{\scriptstyle \mathsf{m}}^{1+z}(m,s) \hat V_k(m,0), \nonumber \\
\hat V^3_k(s,z)&:= \hat c(s,z,3) + \beta \hat V_k(0,0).
\end{align}
Let $c_{\text{max}}$ denote an upper bound on the per-step cost and $J^\ast $ denote the cost under the optimal strategy.
\begin{theorem}\label{thm:app}
Given $\varepsilon \in \mathbb{R}_{>0}$, choose a sufficiently large $k \in \mathbb{N}$ such that
$k \geq \log\left( \frac{(1-\beta)\varepsilon}{2c_{\text{max}}}\right) / \log{\beta}.$
An $\varepsilon$-optimal strategy can then be obtained by solving the Bellman equation~\eqref{eq:bellman_k} as follows:
\begin{equation}\label{eq:proposed-strategy}
g^\ast_\varepsilon(s,z):=\begin{cases}
1, & \hat V^1_k(s,z) = \hat V_k(s,z),\\
2, & \hat V^2_k(s,z) =\hat V_k(s,z),\\
3, & \hat V^3_k(s,z) = \hat V_k(s,z),
\end{cases}
\end{equation}
where $|J^\ast - J(g^\ast_\varepsilon)| \leq \varepsilon$.
\end{theorem}
\begin{proof}
Due to space limitations, only a sketch of the proof is provided, which consists of two steps. In the first step, an approximate Markov decision process with state space $\mathcal{M} \times (0,1,\ldots,k)$ and action space $\mathcal U$ is constructed in such a way that it complies with the dynamics and cost of the original model. The optimal solution of the approximate model is obtained from the Bellman equation~\eqref{eq:bellman_k}. In the second step, it is shown that the difference between the optimal cost of the original model and that of the approximate model is upper-bounded by $ \frac{2 \beta^k c_{\text{max}}}{1-\beta}$. If $\varepsilon \leq \frac{2 \beta^k c_{\text{max}}}{1-\beta}$, then the solution of the approximate model is an $\varepsilon$-optimal solution for the original model.
\end{proof}
According to the strategy proposed in Theorem~\ref{thm:app} (described by~\eqref{eq:proposed-strategy}), the near-optimal action at any time $t \in \mathbb{N}$ depends on the latest observation of the number of faulty components by that time ($s_t$) and the elapsed time since then $(z_t)$. Note that the near-optimal action changes sequentially in time based on the dynamics of the state $(s_t,z_t)$, according to~Lemma~\ref{lemma:hat-f}.
\section{Simulations}\label{sec:examples}
In this section, we aim to verify the main result presented in the preceding section by simulations. Consider a computing platform consisting of $n$ processors. Let $m_t$ denote the number of faulty processors at time $t \in \mathbb{N}$ and $p$ be the probability that a processor fails. The per-step cost under action $u_t=1$ is expressed as follows:
$c(m_t, 1)= \ell(m_t),$
where $\ell(m_t) \in \mathbb{R}_{\geq 0}$ is the cost of operating with $m_t$ faulty processors.
The per-step cost under action $u_t=2$ is described as:
$c(m_t, 2)= \ell(m_t) + \alpha(m_t),$
where $\alpha(m_t) \in \mathbb{R}_{\geq 0}$ is the cost of inspecting the system to detect the number of faulty processors. The per-step cost under action $u_t=3$ is formulated as:
$c(m_t, 3)= \ell(m_t) + \gamma(m_t),$
where $\gamma(m_t) \in \mathbb{R}_{\geq 0}$ is the cost of repairing the faulty processors.
\\
\textbf{Example 1. Inspection and repair with fixed price.} Let the cost of inspection and repair be constant, i.e., they do not depend on the number of faulty components. We consider the following numerical parameters:
\begin{multline}
n=10, \quad p=0.02, \quad\beta=0.95, \quad k=100,\\
\quad \ell(m)=m, \quad \alpha (m)=1, \quad \gamma(m)=60, \quad \forall m \in \mathcal{M}.
\end{multline}
Figure~\ref{fig:flat} shows the optimal course of action for the above setting, in different scenarios in terms of the number of faulty processors (based on the most recent observation). In this figure, the black color represents the first option (continue operating without disruption), gray color represents the second option (inspect the system and detect the number of faulty components) and the white color represents the third option (repair the faulty components). It is observed from the figure that that the inspection and repair options become more attractive as the number of faulty processors and/or the elapsed time since the last observation grow.
\begin{figure}
\center
\vspace{-0cm}
\scalebox{0.94}{
\includegraphics[width=\linewidth]{flat.eps}}
\vspace{-0cm}
\caption{The optimal strategy for Example 1, where the options are depicted in three different colors.}\label{fig:flat}
\end{figure}
\\
\textbf{Example 2. Inspection and repair with variable price.} Now, let the cost of inspection and repair be variable. We consider the same parameters as the previous example, except the following ones:
\begin{equation}
\alpha (m)=0.1+0.09m, \gamma(m)=30+3m, \quad \forall m \in \mathcal{M}.
\end{equation}
The results are presented in Figure~\ref{fig:variable}, analogously to Figure~\ref{fig:flat}. The figure shows that the inspection option is less desirable compared to Example 1, where the inspection and repair options prices were independent of the number of faulty processors. The reason is that with the variable rate, the repair option becomes more economical, hence more attractive than the previous case.
\begin{figure}[t!]
\center
\vspace{0cm}
\scalebox{0.94}{
\includegraphics[width=\linewidth]{variable.eps}}
\vspace{0cm}
\caption{The optimal strategy for Example 2, where the options are depicted in three different colors.}\label{fig:variable}
\end{figure}
\section{Conclusions}\label{sec:conclusions}
In this paper, we presented a fault-tolerant scheme for a system consisting of a number of homogeneous components, where each component can fail at any time with a prescribed probability. We proposed a near-optimal strategy to choose sequentially between three options: (1) do nothing and let the system operate with faulty components; (2) inspect to detect the number of faulty components, and (3) repair the faulty components. Each option incurs a cost that is incorporated in the overal cost function in the optimization problem. Two numerical examples are presented to demonstrate the results in the cases of fixed and variable rates. As a future work, one can investigate the case where there are a sufficiently large number of components using the law of large numbers~\cite{JalalCDC2017}.
\bibliographystyle{IEEEtran}
|
\section{Introduction}
\subfile{sections/introduction}
\section{Related Work}
\subfile{sections/related_work}
\section{Materials}
\subfile{sections/materials}
\section{Event Extraction}
\subfile{sections/event_extraction}
\section{COVID-19 Prediction Application}
\subfile{sections/covid_prediction}
\section{Conclusions}
\subfile{sections/conclusions}
\section*{Acknowledgments}
This work was funded by the National Library of Medicine Biomedical and Health Informatics Training Program under Award Number 5T15LM007442-19, the Gordon and Betty Moore Foundation, and the National Center For Advancing Translational Sciences of the National Institutes of Health under Award Number UL1 TR002319. We want to acknowledge Elizabeth Chang, Kylie Kerker, Jolie Shen, and Erica Qiao for their contributions to the gold standard annotations and Nicholas Dobbins for data management and curation. Research and results reported in this publication was partially facilitated by the generous contribution of computational resources from the University of Washington Department of Radiology.
\subsection{Data}
An existing clinical data set from the UW from January 2020 through May 2020 was used to explore the prediction of COVID-19 test results and identify the most prominent predictors of COVID-19. The data set represents 230K patients, including 28K patients with at least one COVID-19 PCR test result. The data set includes telephone encounters, outpatient progress notes, and emergency department (ED) notes, as well as structured data (demographics, vitals, laboratory results, etc.).
For each patient in this data set, all of the COVID-19 tests with either a \textit{positive} or \textit{negative} result \underline{and} at least one note within the seven days preceding the test result were identified. Only COVID-19 tests with a note within the previous seven days are included in experimentation, to improve the robustness of the COVID-19 symptomology exploration. Each of these test results was treated as a sample in this binary classification task (positive or negative). The likelihood of COVID-19 positivity was predicted using structured EHR data and notes within a 7-day window preceding the test result. The pairing of notes and COVID-19 test results was independently performed for each of the note types (ED, outpatient progress, and telephone encounter notes). From this pool of data, we identified the following test counts by note type: 2,226 negative and 148 positive for ED; 7,599 negative and 381 positive for progress; and 7,374 negative and 448 positive for telephone. Within the 7-day window of this subset of COVID-19 test results, there are 5.3K ED, 14.5K progress, and 27.5K telephone notes. This data set has some overlap with the data set used in Section \ref{methods_section} but is treated as a separate data set in this COVID-19 prediction task. The notes in the CACT training set are less than 1\% of the notes used in this secondary use application.
\subsection{Methods}
\textbf{Features:} Symptom information was automatically extracted from the notes using the Span-based Event Extractor trained on CACT.\footnote{Only automatically extracted symptom data were used. No supervised (hand annotated) labels were used.} The extracted symptoms were normalized using the mapping in Table \ref{symptom_normalization} in the Appendix. Each extracted symptom with an \textit{Assertion} value of ``present'' was assigned a feature value of 1. The 24 identified predictors of COVID-19 from existing literature (see Section \ref{lit_review}) were mapped to 32 distinct fields within the UW EHR and used in experimentation. Identified fields are listed in Table \ref{ehr_features} of the Appendix. For the coded data (e.g. structured fields like ``basophils''), experimentation was limited to this subset of literature-supported COVID-19 predictors, given the limited number of positive COVID-19 tests in this data set.
Within the 7-day history, features may occur multiple times (e.g. multiple temperature measurements). For each feature, the series of values was represented as the minimum or maximum of the values depending on the specific feature. For example, temperature was represented as the maximum of the measurements to detect any fever, and oxygen saturation was represented as the minimum of the values to capture any low oxygenation events. Table \ref{ehr_features} in the Appendix includes the aggregating function, $\bm{f}$, used for each field.
Where symptom features were missing, the feature value was set to 0. For features from the structured EHR data, which are predominantly numerical, missing features were assigned the mean feature value in the set used to train the COVID-19 prediction model.
\textbf{Model:} COVID-19 was predicted using the Random Forest framework, because it facilitates nonlinear modeling with interdependent features and interpretability analyses (Scikit-learn Python implementation used \citep{scikit-learn}). Alternative prediction algorithms include Logistic Regression, SVM, and FFNN. Logistic Regression assumes feature independence and linearity, which is not valid for this task. For example, the feature set includes both the symptom ``fever'' and temperature measurements (e.g. ``$38.6^{\circ}C$''). Model interpretability is less clear with SVM, and the number of positive test samples is relatively small for a FFNN.
The relative importance of features in predicting COVID-19 was explored using \citet{lundberg2020local}’s SHAP (SHapley Additive exPlanations) approach, which is implemented in the SHAP Python module.\footnote{\url{https://pypi.org/project/shap/}} SHAP generates interpretable, feature-level explanations for nonlinear model predictions. For each prediction, SHAP feature scores are estimated, where larger absolute scores indicate higher importance, and the absolute values of the scores sum to 1.0 for each prediction.
\textbf{Experimental paradigm:} The available data was split into train/test sets using an 80\%/20\% split by patient, although training and evaluation was performed at the test-level (i.e. each COVID-19 test result is a sample). Performance was evaluated using the receiver operating characteristic (ROC) and the associated area under the curve (AUC). Given the relatively small number of positive samples, the train/test splits were randomly created 1,000 times through repeated hold-out testing \citep{kim2009estimating}. \citet{kim2009estimating} demonstrated that repeated hold-out testing can improve the robustness of the results in low resource settings. For each train/test split, the AUC was calculated, and an average AUC was calculated across all hold-out iterations. The random holdout iterations yield a distribution of AUC values, which facilitate significance testing. The significance of the AUC performance was assessed using a two-sided T-test. The Random Forest models were tuned using 3-fold cross validation on the training set and evaluated on the withheld test set. COVID-19 prediction experimentation included three feature sets: \textit{structured} (32 structured EHR fields), \textit{notes} (automatically extracted symptoms), and \textit{all} (combination of structured fields and automatically extracted symptoms). Separate models were trained and evaluated for each note type (ED, progress, and telephone) and feature set (\textit{structured}, \textit{notes}, and \textit{all}). The selected Random Forest hyperparameters are summarized in Table \ref{covid_hyperparams} in the Appendix.
\subsection{Results}
Figure \ref{roc} presents the ROC for the COVID-19 predictors with the average AUC across repeated hold-out partitions. The AUC evaluates model performance across all operating points, including operating points that are not clinically significant, for example extremely low true positive rate (TPR). To address this AUC limitation and provide an alternative method for comparing feature sets, we selected a fixed operating point on the ROC, comparing the false positive rate (FPR) at a specific TPR. We selected a TPR (sensitivity) of 80\%, as a value that has clinical value for identifying individuals with COVID-19, and we examined the FPR (specificity) at this fixed TPR. In this use case, we are attempting to see how well structured EHR fields and symptoms perform compared to the reference standard of a laboratory PCR test.
Table \ref{fpr_at_tpr} presents the FPR at TPR=0.80, including the FPR mean and standard deviation across the repeated holdout iterations. Lower FPRs (better performance) are achieved for all three note types, when automatically extracted symptoms are added to the structured data. We would not expect a combination of clinical features to have particularly high sensitivity. \citet{smith2020symptom} achieved similar performance in predicting COVID-19 using clinical prediction rules. While detecting COVID-19 in 80\% of patients with the disease, the inclusion of automatically extracted symptoms decreases the FPR (the ``cost'') by 2-7 percentage points. For all note types, the inclusion of the automatically extracted symptom information (\textit{all} feature set) improves performance over structured data only (\textit{structured}-only feature set) for both AUC and FPR$@$TPR=0.80 with significance ($p<0.001$ per two-sided T-test). The \textit{structured} features achieve higher performance in the ED note experimentation, than experimentation with progress and telephone notes, due to the higher prevalence of vital sign measurements and laboratory testing in proximity to ED visits. In ED note experimentation, over 99\% of samples include vital signs and 72\% include blood work. In progress and telephone note experimentation, 23-38\% of samples includes vital signs and 19-26\% include blood work. The automatically extracted symptoms are especially important in clinical contexts, like outpatient and tele-visit, where vital signs, laboratory results, and other structured data are less available.
\begin{figure}[t!]
\centering
\iftif
\includegraphics[width=6.0in]{figures/roc.tif}
\else
\includegraphics[width=6.0in]{figures/roc.png}
\fi
\caption{Receiver operating characteristic by note type and feature set combination for repeated hold-out iterations. The solid line indicates the average ROC, and the shaded region around the solid line indicates one standard deviation.}
\label{roc}
\end{figure}
\begin{table}[h]
\small
\centering
\input{tables/fpr_at_tpr}
\caption{COVID-19 prediction false positive rate at a true positive rate of 80\%}
\label{fpr_at_tpr}
\end{table}
Figure \ref{shap_impact_single} presents a SHAP value plot for the five most predictive features from a single Random Forest model from the ED note experimentation with the \textit{all} feature set. In this SHAP plot, each point represents a single test prediction, and the SHAP value (x-axis) describes the feature importance. Positive SHAP values indicate support for COVID-19 positivity, and negative values indicate support for negative test result. The color coding indicates the feature value, where red indicates higher feature values and blue indicates lower feature values. For example, high and moderate basophils values (coded in red and purple, respectively) have negative SHAP values, indicating support COVID-19 negativity. Low basophils values (coded in blue) have positive SHAP values, indicating support COVID-19 positivity.
\begin{figure}[ht!]
\centering
\iftif
\includegraphics[width=5.0in]{figures/shap_impact.tif}
\else
\includegraphics[width=5.0in]{figures/shap_impact.png}
\fi
\caption{SHAP plot for a single Random Forest model from the ED note experimentation with the \textit{all} feature set, explaining the importance of features in making predictions for the withheld test set. * indicates the feature is an automatically extracted symptom}
\label{shap_impact_single}
\end{figure}
Given the relatively small sample size and low proportion of positive COVID-19 tests, the SHAP impact values presented in Figure \ref{shap_impact_single} were aggregated across repeated hold-out runs. Figure \ref{shap_impact_violin} presents the averaged SHAP values for each repeated hold-out run for the eight most predictive features for the \textit{all} feature set. For each repeated hold-out run, the absolute value of the SHAP values were averaged, yielding a single feature score per repetition. The mean SHAP values (x-axis) represents the importance of the feature in predicting COVID-19, where positive values indicate a positive correlation between the feature values and COVID-19 positivity and negative values indicate a negative correlation. The most predictive features vary by note type, although fever is a prominent indicator of COVID-19 across note types. For each note type, the top five symptoms indicating COVID-19 positivity include: ED - fever, cough, myalgia, fatigue, and flu-like symptoms; progress - fever, myalgia, respiratory symptoms, cough, and ill; and telephone - fever, cough, myalgia, fatigue, and sore throat. The differences in symptom importance by note type reflects differences in documentation in the clinical settings (e.g., emergency department, outpatient, and tele-visit).
\begin{figure}[t!]
\centering
\iftif
\includegraphics[width=6.0in]{figures/shap_violin.tif}
\else
\includegraphics[width=6.0in]{figures/shap_violin.png}
\fi
\caption{Distribution of averaged SHAP values by note type with the \textit{all} feature set. The vertical lines in each violin indicate the quartiles. * indicates the feature is an automatically extracted symptom.}
\label{shap_impact_violin}
\end{figure}
\ifsubfile
\subsection{Methods}
\label{methods_section}
Event extraction tasks, like ACE05 \citep{ace2005}, typically require prediction of the following event phenomena:
\begin{itemize}[nolistsep]
\item trigger span identification
\item trigger type (event type) classification
\item argument span identification
\item argument type/role classification
\end{itemize}
The CACT annotation scheme differs from this configuration in that labeled arguments require the argument type (e.g. \textit{Assertion}) and the subtype (e.g. \textit{present}, \textit{absent}, etc.) to be predicted. Resolving the argument subtypes require a classifier with additional predictive capacity.
We implement a span-based, end-to-end, multi-layer event extraction model that jointly predicts all event phenomena, including the trigger span, event type, and argument spans, types, and subtypes. Figure \ref{extraction_model} presents our Span-based Event Extractor framework, which differs from prior related work in that multiple span classifiers are used to accommodate the argument subtypes.
\begin{figure}[ht!]
\centering
\iftif
\frame{\includegraphics[width=2.91in]{figures/extraction_model.tif}}
\else
\frame{\includegraphics[width=2.91in]{figures/extraction_model.pdf}}
\fi
\caption{Span-based Event Extractor}
\label{extraction_model}
\end{figure}
Each input sentence consists of tokens, $X=\{x_1, x_2, . .. x_n\}$, where $n$ is the number of tokens. For each sentence, the set of all possible spans, $S=\{s_1, s_2, . .. s_m\}$, is enumerated, where $m$ is the number of spans with token length less than or equal to $M$ tokens. The model generates trigger and argument predictions for each span in $S$ and predicts the pairing between arguments and triggers to create events from individual span predictions.
\textbf{Input encoding:} Input sentences are mapped to contextualized word embeddings using \textit{Bio+Clinical BERT} \citep{alsentzer-etal-2019-publicly}. To limit computational cost, the contextualized word embeddings feed into a bi-LSTM without fine tuning BERT (no backpropagation to BERT). The bi-LSTM has hidden size $v_h$. The forward and backward states, $\bm{h}_{t,f}$ and $\bm{h}_{t,b}$, are concatenated to form the $1 \times 2v_h$ dimensional vector $\bm{h}_t=[\bm{h}_{t,f}, \bm{h}_{t,b}]$, where $t$ is the token position.
\textbf{Span representation:} Each span is represented as the attention weighted sum of the bi-LSTM hidden states. Separate attention mechanisms, $c$, are implemented for trigger and each labeled argument, and a single attention mechanism is implemented for all span-only arguments, $c \in \{1,2\ldots6\}$ (1 for trigger, 4 for labeled arguments, and 1 for span-only arguments). The attention score for span representation $c$ at token position $t$ is calculated as
\begin{equation}
\alpha_{c,t} =\bm{w}_{\alpha,c} \bm{h}_t^T
\end{equation}
where $\bm{w}_{\alpha,c}$ is a learned $1 \times 2v_h$ vector. For span representation $c$, span $i$, and token position $t$, the attention weights are calculated by normalizing the attention scores as
\begin{equation}
a_{c,i,t} =\frac{\mbox{exp}(\alpha_{c,t})}{\sum\limits_{k = start(s_i)}^{end(s_i)} \mbox{exp}(\alpha_{c,k})},
\end{equation}
where $start(s_i)$ and $end(s_i)$ denote the start and end token indices of span $s_i$. Span representation $c$ for span $i$ is calculated as the attention-weighted sum of the bi-LSTM hidden state as
\begin{equation}
\bm{g}_{c,i} = \sum\limits_{t = start(s_i)}^{end(s_i)} a_{c,i,t}\bm{h}_t.
\end{equation}
\textbf{Span prediction:} Similar to the span representations, separate span classifiers, $c$, are implemented for trigger and each labeled argument, and a single classifier predicts all span-only arguments, $c \in \{1,2\ldots6\}$ (1 for trigger, 4 for labeled arguments, and 1 for span-only arguments). Label scores for classifier $c$ and span $i$ are calculated as
\begin{equation}
\phi_{c}(s_i) = \bm{w}_{s,c}\mbox{FFNN}_{s,c}(\bm{g}_{c,i}),
\end{equation}
where $\phi_{c}(s_i)$ yields a vector of label scores of size $|L_{c}|$, $\mbox{FFNN}_{s,c}$ is a non-linear projection from size $2v_h$ to $v_s$, and $\bm{w}_{s,c}$ has size $|L_{c}| \times v_s$. The trigger prediction label set is $L_{trigger}=\{null, \textit{COVID}, \textit{Symptom}\}$. Separate classifiers are used for each labeled argument (\textit{Assertion}, \textit{Change}, \textit{Severity}, and \textit{Test Status}) with label set, $L_{c}=\{null \cup L_l\}$, where $L_l$ is defined in Table \ref{annotated_phenomena}.\footnote{The assertion classifier uses the larger label set associateed with \textit{Symptom}.} For example, $L_{Severity}=\{null, mild, moderate, severe\}$. A single classifier predicts all span-only arguments with label set, $L_{span-only}= \allowbreak \{null, \allowbreak Anatomy, \allowbreak Characteristics, \allowbreak Duration, \allowbreak Frequency\}$.
\textbf{Argument role prediction:} The argument role layer predicts the assignment of arguments to triggers using separate binary classifiers, $d$, for each labeled argument and one classifier for all span-only arguments, $d \in \{1,2,\ldots5\}$ (4 for labeled arguments and 1 for span-only arguments). Argument role scores for trigger $j$ and argument $k$ using argument role classifier $d$ are calculated as
\begin{equation}
\bm{\psi}_{d}(s_j,s_k) = \bm{w}_{r,d}\mbox{FFNN}_{r,d}([\bm{g}_j,\bm{g}_k])
\end{equation}
where $\bm{\psi}_{d}(s_j,s_k)$ is a vector of size $2$, $\mbox{FFNN}_{r,d}$ is a non-linear projection from size $4v_h$ to $v_r$, and $\bm{w}_{r,d}$ has size $2 \times v_r$.
\textbf{Span pruning:} To limit the time and space complexity of the pairwise argument role predictions, only the top-$K$ spans for each span classifier, $c$, are considered during argument role prediction. The span score is calculated as the maximum label score in $\phi_{c}$, excluding the $null$ label score.
\subsection{Model Configuration}
The model configuration was selected using 3-fold cross validation (CV) on the training set. Table \ref{event_hyperparameters} in the Appendix summarizes the selected configuration. Training loss was calculated by summing the cross entropy across all span and argument role classifiers. Models were implemented using the Python PyTorch module \citep{NEURIPS2019_9015}.
\subsection{Data Representation}
During initial experimentation, \textit{Symptom} \textit{Assertion} extraction performance was high for the \textit{absent} subtype and lower for \textit{present}. The higher \textit{absent} performance is primarily associated with the consistent presence of negation cues, like ``denies'' or ``no.'' While there are affirming cues, like ``reports'' or ``has,'' the \textit{present} subtype is often implied by a lack of negation cues. For example, an entire sentence could be ``Short of breath.'' To provide the \textit{Symptom} \textit{Assertion} span classifier with a more consistent span representation, we replaced each \textit{Symptom} \textit{Assertion} span (token indices) with the \textit{Symptom} trigger span in each event and found that performance improved. We extended this trigger span substitution approach to all labeled arguments (\textit{Assertion}, \textit{Change}, \textit{Severity}, and \textit{Test Status}) and found performance improved. By substituting the trigger spans for the labeled argument spans, trigger and labeled argument prediction is roughly treated as a multi-label classification problem, although the model does not constrain trigger and labeled argument predictions to be associated with the same spans. As previously discussed, the scoring routine does not consider the span indices of labeled arguments.
\subsection{Results}
\begin{table*}[htb!]
\small
\centering
\input{tables/extraction_performance.tex}
\caption{Extraction performance}
\label{extraction_performance}
\end{table*}
Table \ref{extraction_performance} presents the event extraction performance on the training set using CV and the withheld test set. Extraction performance is similar on the train and test sets, even though the training set has higher rates of COVID-19 positive notes. \textit{COVID} trigger extraction performance is very high (0.97 F1) and comparable to the annotator agreement (0.95 F1). The \textit{COVID} \textit{Assertion} performance (0.73 F1) is higher than \textit{Test Status} performance (0.62 F1), which is likely due to the more consistent \textit{Assertion} annotation. \textit{Symptom} trigger and \textit{Assertion} extraction performance is high (0.83 F1 and 0.79 F1, respectively), approaching the annotator agreement (0.86 F1 and 0.83 F1, respectively). \textit{Anatomy} extraction performance (0.61 F1) is lower than expected, given the high annotator agreement (0.81 F1). \textit{Duration} extraction performance is comparable to annotator agreement, and \textit{Frequency} extraction performance is lower than annotation agreement. \textit{Change}, \textit{Severity}, and \textit{Characteristics} extraction performance is low, again likely related to low annotator agreement for these cases.
Existing symptom extraction systems do not extract all of the phenomena in the CACT annotation scheme; however, we compared the performance of our Span-based Event Extractor to MetaMapLite++ for symptom identification and assertion prediction. MetaMapLite++ is an analysis pipeline that includes a UMLS concept extractor with assertion prediction \citep{yetisgen2016new}. Table \ref{umls_performance} presents the performance of MetaMapLite++ on the CACT test set. The spans associated with medical concepts in MetaMapLite++ differ slightly from our annotation scheme. For example, ``dry cough'' was extracted by MetaMapLite++ as a symptom, whereas our annotation scheme labels ``cough'' as the symptom and ``dry'' as a characteristic. To account for this difference, Table \ref{umls_performance} presents the performance of MetaMapLite++ for two trigger equivalence criteria: 1) \textit{exact match} for triggers is required, as defined in Section 3.3 and 2) \textit{any overlap} for triggers is considered equivalent. The Span-based Event Extractor outperforms MetaMapLite++ for symptom identification precision (0.81 vs. 0.66), recall (0.85 vs. 0.67), and F1 (0.83 vs. 0.66), even when MetaMapLite++ is evaluated with the more relaxed \textit{any overlap} trigger scoring. The lower recall of MetaMapLite++ is partially the result of the UMLS not including symptom acronyms and abbreviations that frequently occur in our data, for example ``N/V/D'' for ``nausea, vomiting, and diarrhea.'' Table \ref{umls_performance} only reports the performance for the UMLS ``Sign or Symptom'' semantic type. When all UMLS semantic types are used, the recall improves; however, the precision is extremely low ($P = 0.02-0.03$) \footnote{In the UMLS, 15\% of the unique gold symptoms in the CACT training set are covered when only the ``Sign or Symptom'' semantic type is used. The UMLS coverage increases to 48\% when all semantic types are used, and the unique gold symptoms occur in 76 different UMLS semantic types.}.
\begin{table*}[htb!]
\small
\centering
\input{tables/umls_extraction_performance}
\caption{MetaMapLite++ extraction performance for \textit{Symptom} \textit{trigger} and \textit{Assertion}}
\label{umls_performance}
\end{table*}
Table \ref{assertion_performance} presents the assertion prediction performance for both systems, only considering the subset of predictions with exact trigger matches (i.e. assertion prediction performance is assessed without incurring penalty for trigger identification errors). The number of gold assertion labels (``\# Gold'') is greater for the Span-based Event Extractor, because more of the symptom triggers predictions are correct. The Span-based Event Extractor outperforms MetaMapLite++ in assertion prediction precision (0.95 vs. 0.81), recall (0.94 vs. 0.81), and F1 (0.94 vs. 0.81). MetaMapLite++'s lower performance is partly the result of differences between the distribution of assertion labels in CACT and the dataset used to train MetaMapLite++'s assertion classifier (2010 i2b2).
\begin{table*}[htb!]
\small
\centering
\input{tables/assertion_comparison}
\caption{\textit{Symptom} \textit{Assertion} comparison for events with equivalent triggers (exact span match)}
\label{assertion_performance}
\end{table*}
\ifsubfile
\subsection{Data}
This work used inpatient and outpatient clinical notes from the UW clinical repository. COVID-19-related notes were identified by searching for variations of the terms \textit{coronavirus}, \textit{covid}, \textit{sars-cov}, and \textit{sars-2} in notes authored between February 20-March 31, 2020, resulting in a pool of 92K notes. Samples were randomly selected for annotation from a subset of 53K notes that include at least five sentences and correspond to the note types: telephone encounters, outpatient progress, emergency department, inpatient nursing, intensive care unit, and general inpatient medicine. Multiple note types were used to improve extraction model generalizability.
Early in the outbreak, the UW EHR did not include COVID-19 specific structured data; however, structured fields indicating COVID-19 test types and results were added as testing expanded. We used these structured fields to assign a \textit{COVID-19 Test} label describing COVID-19 polymerase chain reaction (PCR) testing to each note based on patient test status within the UW system (no data external to UW was used):
\begin{itemize}[nolistsep]
\item \textit{none}: patient testing information is not available
\item \textit{positive}: patient will have at least one future positive test
\item \textit{negative}: patient will only have future negative tests
\end{itemize}
\noindent
More nuanced descriptions of COVID-19 testing (e.g. conditional or unordered tests) or diagnoses (e.g. possible infection or exposure) are not available as structured data. For the 53K note subset, the \textit{COVID-19 Test} label distribution is 90.8\% \textit{none}, 7.9\% \textit{negative}, and 1.3\% \textit{positive}.\footnote{The COVID-19 test positivity rate cannot be inferred from these label distributions, as there can be multiple test results associated with each note-level label.}
Given the sparsity of \textit{positive} and \textit{negative} notes, CACT is intentionally biased to increase the prevalence of these labels. To ensure adequate \textit{positive} training samples, the CACT training partition includes 46\% \textit{none}, 5\% \textit{negative}, and 49\% \textit{positive} notes. Ideally, the test set would be representative of the true distribution; however, the expected number of \textit{positive} labels with random selection is insufficient to evaluate extraction performance. Consequently, the CACT test partition was biased to include 50\% \textit{none}, 46\% \textit{negative}, and 4\% \textit{positive} notes. Notes were randomly selected in equal proportions from the six note types. CACT includes 1,472 annotated notes, including 1,028 train and 444 test notes.
\subsection{Annotation Scheme}
\begin{table*}[htb!]
\small
\centering
\input{tables/annotated_phenomena.tex}
\caption{Annotation guideline summary. \textsuperscript{*} indicates the argument is required. \textsuperscript{$\dagger$} indicates at least one of the arguments, \textit{Test Status} or \textit{Assertion}, is required}
\label{annotated_phenomena}
\end{table*}
We created detailed annotation guidelines for COVID-19 and symptoms, using the event-based annotation scheme in Table \ref{annotated_phenomena}. Each event includes a trigger that identifies and anchors the event and arguments that characterize the event. The annotation scheme includes two types of arguments: \textit{labeled arguments} and \textit{span-only arguments}. \textit{Labeled arguments} (e.g. \textit{Assertion}) include an argument span, type, and subtype (e.g. \textit{present}). The subtype label normalizes the span information to a fixed set of classes and allows the extracted information to be directly used in secondary use applications. \textit{Span-only arguments} (e.g. \textit{Characteristics}) include an argument span and type but do not include a subtype label, because the argument information is not easily mapped to a fixed set of classes.
For \textit{COVID} events, the trigger is generally an explicit COVID-19 reference, like ``COVID-19" or ``coronavirus." \textit{Test Status} characterizes implicit and explicit references to COVID-19 testing, and \textit{Assertion} captures diagnoses and hypothetical references to COVID-19. \textit{Symptom} events capture subjective, often patient reported, indications of disorders and diseases (e.g ``cough''). For \textit{Symptom} events, the trigger identifies the specific symptom, for example ``wheezing" or ``fever," which is characterized through \textit{Assertion}, \textit{Change}, \textit{Severity}, \textit{Anatomy}, \textit{Characteristics}, \textit{Duration}, and \textit{Frequency} arguments. Symptoms were annotated for all conditions/diseases, not just COVID-19. Notes were annotated using the BRAT annotation tool \citep{stenetorp2012brat}. Figure \ref{annotation_examples} presents BRAT annotation examples.
\begin{figure}[ht]
\begin{adjustbox}{varwidth=\textwidth,fbox,center}
\begin{subfigure}{3.0in}
\iftif
\includegraphics[scale=0.20]{figures/brat_example1.tif}
\else
\includegraphics[scale=0.20]{figures/brat_example1.png}
\fi
\end{subfigure}
\par\medskip
\begin{subfigure}{3.0in}
\iftif
\includegraphics[scale=0.20]{figures/brat_example2.tif}
\else
\includegraphics[scale=0.20]{figures/brat_example2.png}
\fi
\end{subfigure}
\end{adjustbox}
\caption{BRAT annotation examples for \textit{COVID} and \textit{Symptom} (\textit{SSx}) event types}
\label{annotation_examples}
\end{figure}
Most prior medical problem extraction work, including symptom extraction, focuses on identifying the specific problem, normalizing the extracted phenomenon, and predicting an assertion value (e.g. present versus absent). This approach omits many of the symptom details that clinicians are taught to document and that form the core of many clinical notes. This symptom detail describes change (e.g. improvement, worsening, lack of change), severity (e.g. intensity and impact on daily activities), particular characteristics (e.g. productive, dry, or barking for cough), and location. We hypothesize that this symptom granularity is needed for many clinical conditions to improve timely diagnosis and validate diagnosis prediction rules.
\subsection{Annotation Scoring and Evaluation}
Annotation and extraction is scored as a slot filling task, focusing on information most relevant to secondary use applications. Figure \ref{annotation_comparison} presents the same sentence annotated by two annotators, along with the populated slots for the \textit{Symptom} event. Both annotations include the same trigger and \textit{Frequency} spans (``cough'' and ``intermittent'', respectively). The \textit{Assertion} spans differ (``presenting with'' vs. ``presenting''), but the assigned subtypes (\textit{present}) are the same, so the annotations are equivalent for purposes of populating a database. Annotator agreement and extraction performance are assessed using scoring criteria that reflects this slot filling interpretation of the labeling task.
\begin{figure}[ht]
\begin{adjustbox}{varwidth=\textwidth,fbox,center, padding=0ex 0ex 0ex 0ex, margin=0ex 0ex 0ex 0ex}
\begin{subfigure}{3.0in}
\centering
\iftif
\includegraphics[scale=0.20]{figures/slot_filling.tif}
\else
\includegraphics[scale=0.20]{figures/slot_filling.png}
\fi
\end{subfigure}
\par\medskip
\begin{subfigure}{3.0in}
\centering
\large
$\bm{\Downarrow}$ \\
\footnotesize
SSx(trigger=``cough'',
\textit{Assertion}=\textit{present},
\textit{Frequency}=``intermittent'')
\end{subfigure}
\end{adjustbox}
\caption{Annotation examples describing event extraction as a slot filling task}
\label{annotation_comparison}
\end{figure}
The \textit{Symptom} trigger span identifies the specific symptom. For \textit{COVID}, the trigger anchors the event, although the span text is not salient to downstream applications. For labeled arguments, the subtype label captures the most salient argument information, and the identified span is less informative. For span-only arguments, the spans are not easily mapped to a fixed label set, so the selected span contains the salient information. Performance is evaluated using precision (P), recall (R), and F1.
\textbf{Trigger:} Triggers, $T_i$, are represented by a pair (event type, $e_i$; token indices, $x_i$). Trigger equivalence is defined as
\[
T_i \equiv T_j \mbox{ if } (e_i \equiv e_j) \land (x_i \equiv x_j).
\]
\textbf{Arguments:} Events are aligned based on trigger equivalence. The arguments of events with equivalent triggers are compared using different criteria for \textit{labeled arguments} and \textit{span-only arguments}. Labeled arguments, $L_i$, are represented as a triple (argument type, $a_i$; token indices, $x_i$; subtype, $l_i$). For labeled arguments, the argument type, $a$, and subtype, $l$, capture the salient information and equivalence is defined as
\[
L_i \equiv L_j \mbox{ if } (T_i \equiv T_j) \land (a_i \equiv a_j) \land (l_i \equiv l_j).
\]
Span-only arguments, $S_i$, are represented as a pair (argument type, $a_i$; token indices, $x_i$). Span-only arguments with equivalent triggers and argument types, $(T_i \equiv T_j) \land (a_i \equiv a_j)$, are compared at the token-level (rather than the span-level) to allow partial matches. Partial match scoring is used as partial matches can still contain useful information.
\subsection{Annotation Statistics}
CACT includes 1,472 notes with a 70\%/30\% train/test split and 29.9K annotated events (5.4K \textit{COVID} and 24.4K \textit{Symptom}). Figure \ref{covid_stats} contains a summary of the \textit{COVID} annotation statistics for the train/test subsets. By design, the training and test sets include high rates of COVID-19 infection (\textit{present} subtype for \textit{Assertion} and \textit{positive} subtype for \textit{Test Status}), with higher rates in the training set. CACT includes high rates of \textit{Assertion} \textit{hypothetical} and \textit{possible} subtypes. The \textit{hypothetical} subtype applies to sentences like, ``She is mildly concerned about the coronavirus'' and ``She cancelled nexplanon replacement due to COVID-19.'' The \textit{possible} subtype applies to sentences like, ``risk of Covid exposure'' and ``Concern for respiratory illness (including COVID-19 and influenza).'' \textit{Test Status} \textit{pending} is also frequent.
\begin{figure}[ht]
\centering
\iftif
\frame{\includegraphics[scale=1.0]{figures/covid_histogram.tif}}
\else
\frame{\includegraphics[scale=1.0]{figures/covid_histogram.png}}
\fi
\caption{COVID annotation summary}
\label{covid_stats}
\end{figure}
There is some variability in the endpoints of the annotated \textit{COVID} trigger spans (e.g. ``COVID'' vs. ``COVID test''); however 98\% of the \textit{COVID} trigger spans in the training set start with the tokens ``COVID,'' ``COVID19,'' or ``coronavirus.'' Since the \textit{COVID} trigger span is only used to anchor and disambiguate events, the \textit{COVID} trigger spans were truncated to the first token of the annotated span in all experimentation and results.
The training set includes 1,756 distinct uncased \textit{Symptom} trigger spans, 1,425 of which occur fewer than five times. Figure \ref{ssx_histogram} presents the frequency of the 20 most common \textit{Symptom} trigger spans in the training set by \textit{Assertion} subtypes \textit{present}, \textit{absent}, and other (\textit{possible}, \textit{conditional}, \textit{hypothetical}, or \textit{not patient}). The extracted symptoms in Figure \ref{ssx_histogram} were manually normalized to aggregate different extracted spans with similar meanings (e.g. ``sob'' and ``short of breath'' $\rightarrow$ ``shortness of breath''; ``febrile'' and ``fevers'' $\rightarrow$ ``fever''). Table \ref{symptom_normalization} in the Appendix presents the the symptom normalization mapping, provided by a medical doctor. These 20 symptoms account for 62\% of the training set \textit{Symptom} events. There is ambiguity in delineating between some symptoms and other clinical phenomena (e.g. exam findings and medical problems), which introduces some annotation noise.
\begin{figure}[ht]
\centering
\iftif
\frame{\includegraphics[scale=1.0]{figures/ssx_histogram.tif}}
\else
\frame{\includegraphics[scale=1.0]{figures/ssx_histogram.png}}
\fi
\caption{Most frequent symptoms in the training set broken down by \textit{Assertion} subtype}
\label{ssx_histogram}
\end{figure}
Given the long tail of the symptom distribution and our desire to understand the more prominent COVID-19 symptoms, we focused annotator agreement assessment and extraction model training/evaluation on the symptoms that occurred at least 10 times in the training set, resulting in 185 distinct, unnormalized symptoms that cover 82\% of the training set \textit{Symptom} events. The set of 185 symptoms was determined only using the training set, to allow unbiased experimentation on the test set. The subsequent annotator agreement and information extraction experimentation only incorporate these 185 most frequent symptoms.
\subsection{Annotator Agreement}
All annotation was performed by four UW medical students in their fourth year. After the first round of annotation, annotator disagreements were carefully reviewed, the annotation guidelines were updated, and annotators received additional training. Additionally, potential \textit{COVID} triggers were pre-annotated using pattern matching (``COVID,'' ``COVID-19,'' ``coronavirus,'' etc.), to improve the recall of \textit{COVID} annotations. Pre-annotated \textit{COVID} triggers were modified as needed by the annotators, including removing, shifting, and adding trigger spans. Figure \ref{agreement} presents the annotator agreement for the second round of annotation, which included 96 doubly annotated notes. For labeled arguments, F1 scores are micro-average across subtypes.
\begin{figure}[ht]
\small
\centering
\iftif
\frame{\includegraphics[scale=1.0]{figures/agreement.tif}}
\else
\frame{\includegraphics[scale=1.0]{figures/agreement.png}}
\fi
\caption{Annotator agreement}
\label{agreement}
\end{figure}
\ifsubfile
\subsection{Annotated Corpora}
Given the recent onset of COVID-19, there are limited COVID-19 corpora for natural language processing (NLP) experimentation. Corpora of scientific papers related to COVID-19 are available \citep{Wang2020CORD19TC, WHO_corpora}, and automatic labels for biomedical entity types are available for some of these research papers \citep{wang2020comprehensive}. However, we are unaware of corpora of clinical text with supervised COVID-19 annotations.
Multiple clinical corpora are annotated for symptoms. As examples, \citet{south2009developing} annotated symptoms and other medical concepts with negation (present/not present), temporality, and other attributes. \citet{koeling2011annotating} annotated a pre-defined set of symptoms related to ovarian cancer. For the i2b2/VA challenge, \citet{uzuner20112010} annotated medical concepts, including symptoms, with assertion values and relations. While some of these corpora may include symptom annotations relevant to COVID-19 (e.g. ``cough'' or ``fever''), the distribution and characterization of symptoms in these corpora may not be consistent with COVID-19 presentation. To fill the gap in clinical COVID-19 annotations and detailed symptom annotation, we introduce CACT to provide a relatively large corpus with COVID-19 diagnosis, testing, and symptom annotations.
\subsection{Medical Concept and Symptom Extraction}
The most commonly used United Medical Language System (UMLS) concept extraction systems are the clinical Text Analysis and Knowledge Extraction System (cTAKES) \cite{savova2010mayo} and MetaMap \cite{aronson2001effective}. The National Library of Medicine (NLM) created a lightweight Java implementation of MetaMap, \textit{MetaMapLite}, which demonstrated real-time speed and extraction performance comparable to or exceeding the performance of MetaMap, cTAKES, and DNorm \cite{demner2017metamap}. In previous work, we built on MetaMapLite, incorporating assertion value predictions (e.g. present versus absent) using classifiers trained on the 2010 i2b2 challenge dataset to create the extraction pipeline referred to here as \textit{MetaMapLite++} \cite{yetisgen2016new}. MetaMapLite++ assigns each extracted UMLS Metathesaurus concept an assertion value with an Support Vector Machine (SVM)-based assertion classifier that utilizes syntactic and semantic knowledge. The SVM assertion classifier achieved state-of-the-art assertion performance (Micro-F1 94.23) on the i2b2 2010 assertion dataset \cite{bejan2013assertion}. Here, we use MetaMapLite++ as a baseline for evaluating extraction performance for a subset of our annotated phenomena, specifically symptoms with assertion values, using the UMLS ``Sign or Symptom'' semantic type. The Mayo Clinic updated its rule-based medical tagging system, MedTagger \citep{wen2019desiderata}, to include a COVID-19 specific module that extracts 18 phenomena related to COVID-19, including 11 common COVID-19 symptoms with assertion values \cite{ohnlp}. We do not use the COVID-19 MedTagger variant as a baseline, because our symptom annotation and extraction is not limited to known COVID-19 symptoms.
\subsection{Relation and Event Extraction}
There is a significant body of information extraction (IE) work related to coreference resolution, relation extraction, and event extraction tasks. In these tasks, spans of interest are identified, and linkages between spans are predicted. Many contemporary IE systems use end-to-end multi-layer neural models that encode an input word sequence using recurrent or transformer layers, classify spans (entities, arguments, etc.), and predict the relationship between spans (coreference, relation, role, etc.) \citep{ZHENG201759, orr2018event, Shi_2019_extracting, Pang2019deep, chen_2019_extracting, christopoulou2020adverse}. Of most relevance to our work is a series of developments starting with \citet{lee-etal-2017-end}, which introduces a span-based coreference resolution model that enumerates all spans in a word sequence, predicts entities using a feed-forward neural network (FFNN) operating on span representations, and resolves coreferences using a FFNN operating on entity span-pairs. \citet{luan-2018-multi} adapts this framework to entity and relation extraction, with a specific focus on scientific literature. \citet{luan2019general} extends the method to take advantage both of co-reference and relation links in a graph-based approach to jointly predict entity spans, co-reference, and relations. By updating span representations in multi-sentence co-reference chains, the graph-based approach achieved state-of-the-art on several IE tasks representing a range of different genres. \citet{wadden2019entity} expands on \citet{luan2019general}'s approach, adapting it to event extraction tasks. We build on \citet{luan-2018-multi} and \citet{wadden2019entity}'s work, augmenting the modeling framework to fit the CACT annotation scheme. In CACT, event arguments are generally close to the associated trigger, and inter-sentence events linked by co-reference are infrequent, so the graph-based extension, which adds complexity, is unlikely to benefit our extraction task.
Many recent NLP systems use pre-trained language models (LMs), such as ELMo, BERT, and XLNet, that leverage unannotated text \cite{peters2018deep, devlin2019bert, yang2019xlnet}. A variety of strategies for incorporating the LM output are used in IE systems, including using the contextualized word embedding sequence: as the input to a Conditional Random Field entity extraction layer \citep{HuangW2019BMSf}, as the basis for building span representations \citep{luan2019general,wadden2019entity}, or by adding an entity-aware attention mechanism and pooled output states to a fully transformer-based model \citep{wang-etal-2019-extracting}. There are many domain-specific LM variants. Here, we use \citet{alsentzer-etal-2019-publicly}'s \textit{Bio+Clinical BERT}, which is trained on PubMed papers and MIMIC-III \citep{johnson2016mimic} clinical notes.
\subsection{COVID-19 Outcome Prediction}
There are many pre-print and published works exploring the prediction of COVID-19 outcomes, including COVID-19 infection, hospitalization, acute respiratory distress syndrome, need for intensive care unit (ICU), need for a ventilator, and mortality \citep{tian2020predictors, figliozzi2020predictors, jain2020predictive, dong2020novel, xu2020risk, Izquierdo2020, bertsimas2020predictions, Wynantsm1328, SIORDIA2020104357, zhang2020risk, Brinati2020, Mei2020, wollenstein2020personalized}. These COVID-19 outcomes are typically predicted using existing structured data within the EHR, including demographics, diagnosis codes, vitals, and lab results, although \citet{Izquierdo2020} incorporates automatically extracted information from the existing EHRead tool. Our literature review identified 24 laboratory, vital sign, and demographic fields that are predictive of COVID-19 (see Table \ref{lit_features} in the Appendix details). While there are some frequently cited fields, there does not appear to be a consensus across the literature regarding the most prominent predictors of COVID-19 infection. These 24 predictive fields informed the development of our COVID-19 prediction work in Section \ref{covid_prediction}. Prediction architectures includes logistic regression, SVM, decision trees, random forest, K-nearest neighbors, Na\"ive Bayes, and multilayer perceptron \citep{Izquierdo2020, bertsimas2020predictions, Brinati2020, Mei2020, wollenstein2020personalized}.
\ifsubfile
|
\section{Introduction} \label{section:introduction}
It is increasingly common in the natural and social sciences to amass large quantities of geo-referenced data. Researchers seek to use these data to understand phenomena and make predictions via interpretable models that quantify uncertainty taking into account the spatial and temporal dimensions. Gaussian processes (GP) are flexible tools that can be used to characterize spatial and temporal variability and quantify uncertainty, and considerable attention has been devoted to developing GP-based methods that overcome their notoriously poor scalability to large data. The literature on scaling GPs to big data is now extensive. We mention low-rank methods \citep{candelarasmussen05, snelsonghahramani07, gp_predictive_process, frk}; their extensions \citep{lowetal2015, ambikasaran16, huang2018, geoga20}; methods that exploit special structure or simplify the representation of multidimensional inputs---for instance, a Toeplitz structure of the covariance matrix scales GPs to big time series data, and tensor products of scalable univariate kernels can be used for multidimensional inputs \citep{gilboaetal2013, moranwheeler2020, loperetal2020}. These methods may be unavailable or perform poorly in geostatistical settings, which focus on small-dimensional inputs, i.e. the spatial coordinates plus time. In these scenarios, low-rank methods oversmooth the spatial surface \citep{gp_pp_biasadj}, Toeplitz-like structures are typically absent, and so-called \textit{separable} covariance functions obtained via tensor products poorly characterize spatial and temporal dependence. To overcome these hurdles, one can use covariance tapering and domain partitioning \citep{taper1, taper2, fsa, stein2014, katzfuss_jasa17} or composite likelihood methods and sparse precison matrix approximations \citep{vecchia88, grmfields, block_composite_likelihood}; refer to \cite{sunligenton}, \cite{sudipto_ba17}, \cite{Heaton2019} for reviews of scalable geostatistical methods.
\begin{figure}
\centering
\includegraphics[width=.95\textwidth]{figures/application/data_plot.png}
\caption{Observed data of Section \ref{section:applications:modisnoaa}. Missing outcomes are in magenta. GHCN data are much more sparsely observed compared to satellite imaging from MODIS.}
\label{fig:data_plot}
\end{figure}
Additional difficulties arise in multivariate (or multi-output) regression settings. Multivariate geostatistical data are commonly misaligned, i.e. observed at non-overlapping spatial locations \citep{spatial_handbook}. Figure \ref{fig:data_plot} shows several variables measured at non-overlapping locations, with one measurement grid considerably sparser than the others. This issue can be solved by modeling cross-dependence of the outputs via latent spatial random effects thought of as a realization of an underlying GP and embedded in a larger hierarchical model.
Unfortunately, GP approximations that do not correspond to a valid stochastic process may inaccurately characterize uncertainty, as the models used for estimation and interpolation may not coincide. Rather than seeking approximations to the full GP, one can develop valid standalone spatial processes by introducing conditional independence across spatial locations as prescribed by a sparse directed acyclic graph (DAG). These models are advantageous because they lead to scalability by construction; in other words, posterior computing algorithms for these methods can be interpreted as approximate algorithms for the full GP, but also as exact algorithms for the standalone process.
This family of method includes nearest-neighbor Gaussian processes, which limit dependence to a small number of neighboring locations (NNGP; \citealt{nngp, nngp_aoas}), and block-NNGPs \citep{prates}. There is a close relation between DAG structure and computational performance of NNGPs: some orderings may be associated to improved approximations \citep{guinness_techno}, and graph coloring algorithms \citep{molloyreed2002, lewis2016} can be used for parallel Gibbs sampling. Inferring ordering or coloring can be problematic when data are in the millions, but these issues can be circumvented by forcing DAGs with known properties onto the data; in meshed GPs \citep[MGPs;][]{meshedgp}, patterned DAGs associated to domain tiling are associated to more efficient sampling of the latent effects. Alternative so-called multiscale or multiresolution methods correspond to DAGs with hierarchical node structures (trees), which are typically coupled with recursive domain partitioning; in this case, too, efficiencies follow from the properties of the chosen DAG. There is a rich literature on Gaussian processes and recursive partitioning, see e.g
\cite{ferreira, treed, foxdunson12}; in geospatial contexts, in addition to the GMRF-based method of \cite{nychka15}, multi-resolution approximations \citep[MRA;][]{katzfuss_jasa17} replace an orthogonal basis decomposition with approximations based on tapering or domain partitioning and also have a DAG interpretation \citep{katzfuss_vecchia}.
Considerably less attention has been devoted to process-based methods that ensure scalability in multivariate contexts, with the goal of modeling the spatial and/or temporal variability of several variables jointly via flexible cross-covariance functions \citep{genton_ccov}. When scalability of GP methods is achieved via reductions in the conditioning sets, including more distant locations is thought to aid in the estimation of unknown covariance parameters \citep{steinetal2004}. However, the size of such sets may need to be reduced excessively when outcomes are not of very small dimension. One could restrict spatial coverage of the conditioning sets, but this works best when data are not misaligned, in which case all conditioning sets will include outcomes from all margins; this cannot be achieved for misaligned data, leading to pathological behavior. Alternatively, one can model the multivariate outcomes themselves as a DAG; however this may only work on a case-by-case basis. Similarly, recursive domain partitioning strategies work best for data that are measured uniformly in space as this guarantees similarly sized conditioning sets; on the contrary, recursive partitioning struggles in predicting the outcomes at large unobserved areas as they tend to be associated to the small conditioning sets making up the coarser scales or resolutions.
In this article, we solve these issues by introducing a Bayesian regression model that encodes spatial dependence as a latent spatial multivariate tree (\modelname); conditional independence relations at the \textit{reference} locations are governed by the branches in a treed DAG, whereas a map is used to assign all \textit{non-reference} locations to leaf nodes of the same DAG. This assignment map controls the nature and the size of the conditioning sets at all locations; when severe restrictions on the reference set of locations become necessary due to data size, this map is used to improve estimation and predictions and overcome common issues in standard nearest-neighbor and recursive partition methods while maintaining the desirable recursive properties of treed DAGs.
Unlike methods based on defining conditioning sets based solely on spatial proximity, \textsc{SpamTrees}\ scale to large data sets without excessive reduction of the conditioning sets. Furthermore, \textsc{SpamTrees}\ are less restrictive than methods based on recursive partitioning and can be built to guarantee similarly-sized conditioning sets at all locations.
The present work adds to the growing literature on spatial processes defined on DAGs by developing a method that targets efficient computations of Bayesian multivariate spatial regression models. \textsc{SpamTrees}\ share similarities with MRAs \citep{katzfuss_jasa17}; however, while MRAs are defined as a basis function expansion, they can be represented by a treed graph of a \modelname\ with full ``depth'' as defined later (the DAG on the right of Figure \ref{figure:spamtree}), in univariate settings, and ``response'' models. All these restrictions are relaxed in this article. In considering spatial proximity to add ``leaves'' to our treed graph, our methodology also borrows from nearest-neighbor methods \citep{nngp}. However, while we use spatial neighbors to populate the conditioning sets for non-reference locations, the same cannot be said about reference locations for which the treed graph is used instead. Our construction of the \modelname\ process also borrows from MGPs on tessellated domains \citep{meshedgp}; however, the treed DAG we consider here induces markedly different properties on the resulting spatial process owing to its recursive nature. Finally, a contribution of this article is in developing self-contained sampling algorithms which, based on the graphical model representation of the model, will not require any external libraries.
The article builds \textsc{SpamTrees}\ as a standalone process based on a DAG representation in Section \ref{section:construction}. A Gaussian base process is considered in Section \ref{section:spamtree_gp} and the resulting properties outlined, along with sampling algorithms. Simulated data and real-world applications are in Section \ref{section:applications}; we conclude with a discussion in Section \ref{section:discussion}. The Appendix provides more in-depth treatment of several topics and additional algorithms.
\section{Spatial Multivariate Trees}\label{section:construction}
\begin{figure}
\centering
\includegraphics[width=.32\textwidth]{figures/spamtree_dag_limited.png}
\includegraphics[width=.32\textwidth]{figures/spamtree_dag_partial.png}
\includegraphics[width=.32\textwidth]{figures/spamtree_dag_full.png}
\caption{Three \textsc{SpamTrees}\ on $M=4$ levels with depths $\delta=1$ (left), $\delta=3$ (center), and $\delta=4$ (right). Nodes are represented by circles, with branches colored in brown and leaves in green. }
\label{figure:spamtree}
\end{figure}
Consider a spatial or spatiotemporal domain ${\cal D}$. With the temporal dimension, we have ${\cal D} \subset \Re^{d} \times [0, \infty)$, otherwise ${\cal D} \subset \Re^{d}$. A $q$-variate spatial process is defined as an uncountable set of random variables $\{\bolds{w}(\bolds{\ell}) : \bolds{\ell}\in{\cal D}\}$, where $\bolds{w}(\bolds{\ell})$ is a $q\times 1$ random vector with elements $w_i(\bolds{\ell})$ for $i=1,2,\ldots,q$, paired with a probability law $P$ defining the joint distribution of any finite sample from that set. Let $\{\bolds{\ell}_1,\bolds{\ell}_2, \ldots, \bolds{\ell}_{n_{{\cal L}}}\} = {\cal L} \subset {\cal D}$ be of size $n_{{\cal L}}$. The $n_{{\cal L}}q\times 1$ random vector $\bolds{w}_{{\cal L}} = (\bolds{w}(\bolds{\ell}_1)^{\top},\bolds{w}(\bolds{\ell}_2)^{\top},\ldots,\bolds{w}(\bolds{\ell}_{n_{{\cal L}}})^{\top})^{\top}$ has joint density $p(\bolds{w}_{{\cal L}})$. After choosing an arbitrary order of the locations, $p(\bolds{w}_{{\cal L}}) = \prod_{i=1}^{n_{{\cal L}}} p( \bolds{w}(\bolds{\ell}_{i}) \,|\, \bolds{w}(\bolds{\ell}_1), \ldots, \bolds{w}(\bolds{\ell}_{i-1})),$ where the conditioning set for each $\bolds{w}(\bolds{\ell}_i)$ can be interpreted as the set of nodes that have a directed edge towards $\bolds{w}(\bolds{\ell}_i)$ in a DAG. Some scalable spatial processes result from reductions in size of the conditioning sets, following one of several proposed strategies \citep{vecchia88, steinetal2004, gramacy_apley14, nngp, katzfuss_vecchia, meshedgp}. Accordingly,
\begin{align}\label{eq:dag_basic_multivar}
p(\bolds{w}_{{\cal L}}) &= \prod_{i=1}^{n_{{\cal L}}} p( \bolds{w}(\bolds{\ell}_{i}) \,|\, \bolds{w}(\pa{\bolds{\ell}_i})),
\end{align}
where $\pa{\bolds{\ell}_i}$ is the set of spatial locations that correspond to directed edges pointing to $\bolds{\ell}_i$ in the DAG. If $\pa{\bolds{\ell}_i}$ is of size $J$ or less for all $i=1, \dots, n_{{\cal L}}$, then $\bolds{w}(\pa{\bolds{\ell}_i}))$ is of size $Jq$. Methods that rely on reducing the size of parent sets are thus negatively impacted by the dimension $q$ of the multivariate outcome; if $q$ is not very small, reducing the number of parent locations $J$ may be insufficient for scalable computations. As an example, an NNGP model has $\pa{\bolds{\ell}_i} = N(\bolds{\ell}_i)$, where $N(\cdot)$ maps a location in the spatial domain to its neighbor set. It is customary in practice to consider $Jq = m \leq 20$ for accurate and scalable estimation and predictions in univariate settings, but this may be restrictive in some multivariate settings as one must reduce $J$ to maintain similar computing times, possibly harming estimation and prediction accuracy.
We represent the $i$th component of the $q\times 1$ vector $\bolds{w}(\bolds{\ell})$ as $w(\bolds{\ell}, \xi_i)$, where $\xi_i = (\xi_{i1}, \dots, \xi_{ik})^\top \in \Xi$ for some $k$ and $\Xi$ serves as the $k$-dimensional latent spatial domain of variables. The $q$-variate process $\bolds{w}(\bolds{\ell})$ is thus recast as $\{ w(\bolds{\ell}, \xi) : (\bolds{\ell}, \xi) \in {\cal D} \times \Xi \}$, with $\xi$ representing the latent location in the domain of variables. We can then write (\ref{eq:dag_basic_multivar}) as
\begin{align}\label{eq:dag_basic_latentrep}
p(\bolds{w}_{{\cal L}^*}) &= \prod_{i=1}^{n_{{\cal L}^*}} p( w(\bolds{\ell}^*_{i}) \,|\, w(\pa{\bolds{\ell}^*_i})),
\end{align}
where ${\cal L}^* = \{ \bolds{\ell}^*_i \}_{i=1}^{n_{{\cal L}^*}}$, $\bolds{\ell}^*_i \in {\cal D} \times \Xi = {\cal D}^*$, and $w(\cdot)$ is a univariate process on the expanded domain ${\cal D}^*$. This representation is useful as it provides a clearer accounting of the assumed conditional independence structure of the process in a multivariate context.
\subsection{Constructing spatial multivariate DAGs} \label{section:construction_parts}
\textsc{SpamTrees}\ are defined by \textit{(i)} a treed DAG ${\cal G}$ with \textit{branches} and \textit{leaves} on $M$ levels and with depth $\delta \leq M$; \textit{(ii)} a reference set of locations ${\cal S}$; \textit{(iii)} a \textit{cherry picking} map. The graph is ${\cal G} = \{ \bolds{V}, \bolds{E} \}$ where the nodes are $\bolds{V}= \{\bolds{v}_1, \dots, \bolds{v}_{m_V}\} = \bolds{A} \cup \bolds{B}$, $\bolds{A} \cap \bolds{B} = \emptyset$. The \textit{reference} or \textit{branch} nodes are $\bolds{A} = \{ \bolds{a}_1, \dots, \bolds{a}_{m_A} \} = \bolds{A}_0 \cup \bolds{A}_1 \cup \dots \cup \bolds{A}_{M-1}$, where $\bolds{A}_i = \{ \bolds{a}_{i, 1}, \dots, \bolds{a}_{i, m_i} \}$ for all $i=0, \dots, M-1$ and with $\bolds{A}_i \cap \bolds{A}_j = \emptyset$ if $i\neq j$. The \textit{non-reference} or \textit{leaf} nodes are $\bolds{B} = \{ \bolds{b}_1, \dots, \bolds{b}_{m_B} \}$, $\bolds{A} \cap \bolds{B} = \emptyset$. We also denote $\bolds{V}_r = \bolds{A}_r$ for $r=0, \dots, M-1$ and $\bolds{V}_{M} = \bolds{B}$. The edges are $\bolds{E} = \{ \pa{\bolds{v}} \subset \bolds{V} : \bolds{v} \in \bolds{V} \}$ and similarly $\ch{\bolds{v}} = \{ \bolds{v}' \in \bolds{V} : \bolds{v} \in \pa{\bolds{v}'} \}$. The reference set ${\cal S}$ is partitioned in $M$ levels starting from zero, and each level is itself partitioned into reference subsets: ${\cal S}=\cup_{r=0}^{M-1} {\cal S}_r = \cup_{r=0}^{M-1} \cup_{i=1}^{m_i} S_{ri}$, where $S_{ri} \cap S_{r'i'} = \emptyset$ if $r\neq r'$ or $i \neq i'$ and its complement set of \textit{non-reference} or \textit{other} locations ${\cal U} = {\cal D}^* \setminus {\cal S}$. The \textit{cherry picking} map is $\eta: {\cal D}^* \rightarrow \bolds{V}$ and assigns a node (and therefore all the edges directed to it in ${\cal G}$) to any location in the domain, following a user-specified criterion.
\subsubsection{Branches and leaves}
For a given $M$ and a depth $\delta \leq M$, we impose a treed structure on ${\cal G}$ by assuming that if $\bolds{v} \in \bolds{A}_i$ and $i>M-\delta = M_{\delta}$ then there exists a sequence of nodes $\{\bolds{v}_{r_{M_{\delta}}}, \dots, \bolds{v}_{r_{i-1}}\}$ such that $\bolds{v}_{r_{j}} \in \bolds{A}_{j}$ for $j = M_{\delta}, \dots, i-1$ and $\pa{\bolds{v}} = \{\bolds{v}_{r_{M_{\delta}}}, \bolds{v}_{r_1}, \dots, \bolds{v}_{r_{j-1}} \}$. If $i\leq M-\delta = M_{\delta}$ then $\pa{\bolds{v}} = \{ \bolds{v}_{i-1} \}$ with $\bolds{v}_{i-1} \in \bolds{A}_{i-1}$.
$\bolds{A}_0$ is the tree \textit{root} and is such that $\pa{\bolds{v}_0} = \emptyset$ for all $\bolds{v}_0 \in \bolds{A}_0$. The depth $\delta$ determines the number of levels of ${\cal G}$ (from the top) across which the parent sets are nested. Choosing $\delta=1$ implies that all nodes have a single parent; choosing $\delta=M$ implies fully nested parent sets (i.e. if $\bolds{v}_i \in \pa{\bolds{v}_j}$ then $\pa{\bolds{v}_i} \subset \pa{\bolds{v}_j}$ for all $\bolds{v}_i, \bolds{v}_j \in \bolds{V}$). The $m_i$ elements of $\bolds{A}_i$ are the branches at level $i$ of ${\cal G}$ and they have $i-M_{\delta}$ parents if the current level $i$ is above the depth level $M_{\delta}$ and 1 parent otherwise. We refer to \textit{terminal branches} as nodes $\bolds{v} \in \bolds{A}$ such that $\ch{\bolds{v}} \subset \bolds{B}$. For all choices of $\delta$, $\bolds{v} \in \bolds{A}_i, \bolds{v}'\in \bolds{A}_j$ and $\bolds{v} \in \pa{\bolds{v}'}$ implies $i<j$; this guarantees acyclicity.
As for the leaves, for all $\bolds{v} \in \bolds{B}$ we assume $\pa{\bolds{v}} = \{\bolds{v}_{r_{M_{\delta}}}, \dots, \bolds{v}_{r_{k}}\}$ for some integer sequence $\{ r_{M_{\delta}}, \dots, r_{k} \}$ and $\bolds{v}_{r_i} \in \bolds{A}_i$ with $i\geq M_{\delta}$. We allow the existence of multiple leaves with the same parent set, i.e. there can be $k$ and $\bolds{b}_{i_1}, \dots, \bolds{b}_{i_k}$ such that for all $i_2, \dots, i_k$, $\pa{\bolds{b}_{i_h}} = \pa{\bolds{b}_{i_1}}$. Acyclicity of ${\cal G}$ is maintained as leaves are assumed to have no children. Figure \ref{figure:spamtree} represents the graph associated to \textsc{SpamTrees}\ with different depths.
\subsubsection{Cherry picking via $\eta(\cdot)$}
The link between ${\cal G}$, ${\cal S}$ and ${\cal U}$ is established via the map $\eta : {\cal D}^* \rightarrow \bolds{V}$ which associates a node in ${\cal G}$ to any location $\bolds{\ell}^*$ in the expanded domain ${\cal D}^*$:
\begin{align}\label{eq: map}
\eta(\bolds{\ell}^*) &= \left\{
\begin{array}{l}
\eta_{A}(\bolds{\ell}^*) = \bolds{a}_{ri} \in \bolds{A}_r\; \mbox{ if }\; \bolds{\ell}^* \in S_{ri},\\
\eta_{B}(\bolds{\ell}^*) = \bolds{b} \in \bolds{B} \; \mbox{ if }\; \bolds{\ell}^* \in {\cal U}.
\end{array}
\right.
\end{align}
This is a many-to-one map; note however that all locations in $S_{ij}$ are mapped to $\bolds{a}_{ij}$: by calling $\eta(X) = \{ \eta(\bolds{\ell}^*) : \bolds{\ell}^* \in X \}$ then for any $i=0, \dots, M-1$ and any $j=1, \dots, m_i$ we have $\eta(S_{ij}) = \eta_A(S_{ij}) = \bolds{a}_{ij}$. \textsc{SpamTrees}\ introduce flexibility by cherry picking the leaves, i.e. using $\eta_B : {\cal U} \rightarrow \bolds{B}$, the restriction of $\eta$ to ${\cal U}$. Since each leaf node $\bolds{b}_j$ determines a unique path in ${\cal G}$ ending in $\bolds{b}_j$, we use $\eta_B$ to assign a convenient parent set to $w(\bolds{u})$, $\bolds{u} \in {\cal U}$, following some criterion.
For example, suppose that $\bolds{u} = (\bolds{\ell}, \xi_s)$ meaning that $w(\bolds{u}) = w(\bolds{\ell}, \xi_s)$ is the realization of the $s$-th variable at the spatial location $\bolds{\ell}$, and we wish to ensure that $\pa{w(\bolds{u})}$ includes realizations of the same variable. Denote $\bolds{T} = \{\bolds{v} \in \bolds{A} : \ch{\bolds{v}} \subset \bolds{B} \}$ as the set of terminal branches of ${\cal G}$. Then we find $(\bolds{\ell}, \xi_s)_{\text{opt}} = \arg\min_{(\bolds{\ell}',\xi'=\xi_s) \in \eta_A^{-1}(\bolds{T})} d( \bolds{\ell}' , \bolds{\ell} )$ where $d(\cdot, \cdot)$ is the Euclidean distance. Since $(\bolds{\ell}, \xi_s)_{\text{opt}} \in S_{ij} $ for some $i,j$ we have $\eta_A((\bolds{\ell}, \xi_s)_{\text{opt}}) = \bolds{a}_{ij}$. We then set $\eta_B(\bolds{u}) = \bolds{b}_{k}$ where $\pa{\bolds{b}_k} = \{\bolds{a}_{ij} \}$. In a sense $\bolds{a}_{ij}$ is the terminal node nearest to $\bolds{u}$; having defined $\eta_B$ in such a way forces the parent set of any location to include at least one realization of the process from the same variable. There is no penalty in using ${\cal D}^* = {\cal D} \times \Xi$ as we can write $p(\bolds{w}(\bolds{u})\,|\, \pa{\bolds{w}(\bolds{u})}) = p(\bolds{w}((\bolds{\ell}, \xi_1), \dots, (\bolds{\ell}, \xi_q))\,|\, \pa{\bolds{w}(\bolds{u})}) = \prod_{s=1}^q p(w(\bolds{\ell},\xi_s) \,|\, w(\bolds{\ell}, \xi_1), \dots, w(\bolds{\ell}, \xi_{s-1}), \pa{\bolds{w}(\bolds{\ell})}))$, which also implies that the size of the parent set may depend on the variable index.
Assumptions of conditional independence across variables can be encoded similarly. Also note that any specific choice of $\eta_B$ induces a partition on ${\cal U}$; let $U_{j} = \{ \bolds{u} \in {\cal U} : \eta_B(\bolds{u}) = \bolds{b}_j \}$, then clearly ${\cal U} = \cup_{j=1}^{m_U} U_j$ with $U_i \cap U_j = \emptyset$ if $i\neq j$. This partition does not necessarily correspond to the partitioning scheme used on ${\cal S}$. $\eta_B$ may by designed to ignore part of the tree and result in $m_U < m_B$. However, we can just drop the unused leaves from ${\cal G}$ and set $\ch{\bolds{a}} = \emptyset$ for terminal nodes whose leaf is inactive, resulting in $m_U = m_B$. We will thus henceforth assume that $m_U = m_B$ without loss of generality.
\subsection{\textsc{SpamTrees}\ as a standalone spatial process}
We define a valid joint density for any finite set of locations in ${\cal D}^*$ satisfying the Kolmogorov consistency conditions in order to define a valid process. We approach this problem analogously to \cite{nngp} and \cite{meshedgp}. Enumerate each of the $m_S$ reference subsets as $S_i = \{ \bolds{s}_{i_1}, \dots, \bolds{s}_{i_{n_i}} \}$ where $\{i_1,\ldots,i_{n_i}\}\subset \{1,\ldots,n_{{\cal S}}\}$, and each of the $m_U$ non-reference subsets as $U_i = \{ \bolds{u}_{i_1}, \dots, \bolds{u}_{i_{n_i}} \}$ where $\{i_1,\ldots,i_{n_i}\}\subset \{1,\ldots,n_{{\cal U}}\}$. Then introduce ${\cal V} = \{ V_1, \dots, V_{m_V} \}$ where $m_V = m_S + m_U$ and $V_i = S_i$ for $i=1, \dots, m_S$, $V_{m_S+i} = U_i$ for $i=1, \dots, m_U$. Then take $\bolds{w}_i = (w(\bolds{\ell}_{i_1}), \ldots, w(\bolds{\ell}_{i_{n_i}}))^{\top}$ as the $n_i\times 1$ random vector with elements of $w(\bolds{\ell})$ for each $\bolds{\ell}\in V_i$. Denote $\bwpa{i} = \bolds{w}( \eta^{-1}(\pa{\bolds{v}_i}))$. Then
\begin{equation}\label{eq:pws}
\begin{aligned}
\tilde{p}(\bolds{w}_{{\cal S}}) = \tilde{p}(\bolds{w}_1, \dots, \bolds{w}_{m_S})
= \prod_{r=0}^{M-1} \prod_{i : \{ \bolds{v}_i \in \bolds{A}_r \} } p(\bolds{w}_i \,|\, \bwpa{i}) \qquad & \quad
\tilde{p}(\bolds{w}_{{\cal U}} \mid \bolds{w}_{{\cal S}}) = \prod_{i : \{ \bolds{v}_i \in \bolds{B} \} } p( \bolds{w}_i \,|\, \bwpa{i}) \\
\tilde{p}(\bolds{w}_{{\cal S}}) \tilde{p}(\bolds{w}_{{\cal U}} \mid \bolds{w}_{{\cal S}}) =
\prod_{r=0}^{M-1} \prod_{i : \{ \bolds{v}_i \in \bolds{A}_r \} } p(\bolds{w}_i \,|\, \bwpa{i})
&\prod_{i : \{ \bolds{v}_i \in \bolds{B} \} } p( \bolds{w}_i \,|\, \bwpa{i})
\end{aligned}
\end{equation}
which is a proper multivariate joint density since ${\cal G}$ is acyclic \citep{lauritzen}.
All locations inside $U_j$ always share the same parent set, but a parent set is not necessarily unique to a single $U_j$. This includes as a special case a scenario in which one can assume
\begin{align}\label{equation:u_locations_independent}
\tilde{p}(\bolds{w}_{{\cal U}} \mid \bolds{w}_{{\cal S}}) = \prod_{j=1}^{m_U} \prod_{i=1}^{|U_j|} p( w(\bolds{u}_i) \,|\, \bolds{w}(\eta^{-1}(\pa{\bolds{b}_j})));
\end{align}
in this case each location corresponds to a leaf node.
To conclude the construction, for any finite subset of spatial locations $\mathcal{L} \subset \mathcal{D}$ we can let ${\cal U} = \mathcal{L} \setminus {\cal S}$ and obtain
\[\tilde{p}(\bolds{w}_\mathcal{L}) = \int \tilde{p}(\bolds{w}_{{\cal U}} \mid \bolds{w}_{{\cal S}}) \tilde{p}(\bolds{w}_{{\cal S}}) \prod_{ \bolds{s}_i \in {\cal S} \setminus \mathcal{L}} d(\bolds{w}(\bolds{s}_i)),\]
leading to a well-defined process satisfying the Kolmogorov conditions (see Appendix \ref{appx:kolmogorov}).
\subsubsection{Positioning of spatial locations in conditioning sets}
In spatial models based on sparse DAGs, larger conditioning sets yield processes that are closer to the base process $p$ in terms of Kullback-Leibler divergence \citep{sudipto_ss20, meshedgp}, denoted as $KL(\cdot \| p)$. The same results cannot be applied directly to \textsc{SpamTrees}\ given the treed structure of the DAG. For a given ${\cal S}$, we consider the distinct but related issues of placing individual locations into reference subsets (1) at different levels of the treed hierarchy; (2) within the same level of the hierarchy.
\begin{proposition}\label{prop:prop1}
Suppose ${\cal S} = {\cal S}_0 \cup {\cal S}_1$ where $S_{0} \cap S_{1} = \emptyset$ and ${\cal S}_1 = S_{11} \cup S_{12}$, $S_{11} \cap S_{12} = \emptyset$. Take $\bolds{s}^* \notin {\cal S}$. Consider the graph ${\cal G} = \{ \bolds{V} = \{ \bolds{v}_0, \bolds{v}_1, \bolds{v}_2 \}, \bolds{E} = \{ \bolds{v}_0\to\bolds{v}_1, \bolds{v}_0\to\bolds{v}_2 \} \}$; denote as $p_0$ the density of a \modelname\ using $\eta({\cal S}_0 \cup \{ \bolds{s}^* \}) = \bolds{v}_0$, $\eta(S_{11}) = \bolds{v}_1$ and $\eta(S_{12}) = \bolds{v}_2$, whereas let $p_1$ be the density of a \modelname\ with $\eta({\cal S}_0) = \bolds{v}_0$, $\eta(S_{11} \cup \{ \bolds{s}^*\}) = \bolds{v}_1$ and $\eta(S_{12}) = \bolds{v}_2$. Then $KL(p_1 \| p) - KL(p_0 \| p) > 0.$
\end{proposition}
The proof proceeds by an ``information never hurts'' argument \citep{coverthomas91}.
Denote ${\cal S}^* = {\cal S} \cup \{ \bolds{s}^* \}$, $\bolds{w}^*= \bolds{w}_{{\cal S}^*}$, $w^* = w(\bolds{s}^*)$ and $\bolds{w}_j^* = (\bolds{w}_j^\top, w^*)^\top$. Then
\begin{align*}
p_0(\bolds{w}^*) &= p(\bolds{w}_0^*) p(\bolds{w}_1 \,|\, \bolds{w}_0^*) p(\bolds{w}_2 \,|\, \bolds{w}_0^*) = p(\bolds{w}_0) p(w^* \,|\, \bolds{w}_0) p(\bolds{w}_1 \,|\, \bolds{w}_0, w^*) p(\bolds{w}_2 \,|\, \bolds{w}_0^*) \\
p_1(\bolds{w}^*) &= p(\bolds{w}_0) p(\bolds{w}_1^* \,|\, \bolds{w}_0) p(\bolds{w}_2 \,|\, \bolds{w}_0) = p(\bolds{w}_0) p(w^* \,|\, \bolds{w}_0) p(\bolds{w}_1 \,|\, \bolds{w}_0, w^*) p(\bolds{w}_2 \,|\, \bolds{w}_0),
\end{align*}
therefore $p_0(\bolds{w}^*) / p_1(\bolds{w}^*) = p(\bolds{w}_2 \,|\, \bolds{w}_0^*) / p(\bolds{w}_2 \,|\, \bolds{w}_0)$; then by Jensen's inequality
\begin{equation}\label{KL_approx}
\begin{aligned}
KL(p_1 \| p) - KL(p_0 &\| p) = \int \left\{ \log\left(\frac{p(\bolds{w}^*)}{p_1(\bolds{w}^*)}\right) - \log\left(\frac{p(\bolds{w}^*)}{p_0(\bolds{w}^*)} \right) \right\} p(\bolds{w}^*) d\bolds{w}^* \\
&= \int \log\left(\frac{p_0(\bolds{w}^*)}{p_1(\bolds{w}^*)}\right) p(\bolds{w}^*) d\bolds{w}^* = \int \log \left(\frac{ p(\bolds{w}_2 \,|\, \bolds{w}_0^*) }{ p(\bolds{w}_2 \,|\, \bolds{w}_0) }\right) p(\bolds{w}^*) d\bolds{w}^*\\
&= \int \log \left(\frac{ p(\bolds{w}_2 \,|\, \bolds{w}_0^*) }{ p(\bolds{w}_2 \,|\, \bolds{w}_0) }\right) p(\bolds{w}_1, \bolds{w}_2, \bolds{w}_0^*) d\bolds{w}_1 d\bolds{w}_2 d\bolds{w}_0^* \\
&= \int \left\{ \int \log \left(\frac{ p(\bolds{w}_2 \,|\, \bolds{w}_0^*) }{ p(\bolds{w}_2 \,|\, \bolds{w}_0) }\right) p(\bolds{w}_1, \bolds{w}_2 \,|\, \bolds{w}_0^*) d\bolds{w}_1 d\bolds{w}_2 \right\} p(\bolds{w}_0^*) d\bolds{w}_0^* \geq 0.
\end{aligned}
\end{equation}
Intuitively, this shows that there is a penalty associated to positioning reference locations at higher levels of the treed hierarchy. Increasing the size of the reference set at the root augments the conditioning sets at all its children; since this is not true when the increase is at a branch level, the density $p_0$ is closer to $p$ than $p_1$. In other words there is a cost of branching in ${\cal G}$ which must be justified by arguments related to computational efficiency. The above proposition also suggests populating near-root branches with locations of sparsely-observed outcomes. Not doing so in highly imbalanced settings may result in possibly too restrictive spatial conditional independence assumptions.
\begin{proposition}
Consider the same setup as Proposition \ref{prop:prop1} and let $p_2$ be the density of a \modelname\ such that $\eta(S_{12} \cup \{ \bolds{s}^* \}) = \bolds{v}_2$. Let $H_p$ be the conditional entropy of base process $p$. Then $H_p(w^* \,|\, \bolds{w}_0, \bolds{w}_2) < H_p(w^* \,|\, \bolds{w}_0, \bolds{w}_1)$ implies $KL(p_2 \| p) < KL(p_1 \| p)$.
\end{proposition}
\noindent The density of the new model is \[p_2(\bolds{w}^*) = p(\bolds{w}_0) p(\bolds{w}_1 \,|\, \bolds{w}_0) p(\bolds{w}_2^* \,|\, \bolds{w}_0) = p(\bolds{w}_0) p(\bolds{w}_1 \,|\, \bolds{w}_0) p(\bolds{w}_2 \,|\, \bolds{w}_0) p(w^* \,|\, \bolds{w}_0, \bolds{w}_2).\] Then, noting that $p(\bolds{w}^*_1 \,|\, \bolds{w}_0) = p(\bolds{w}_1 \,|\, \bolds{w}_0) p(w^* \,|\, \bolds{w}_0, \bolds{w}_1)$, we get $\frac{p_1(\bolds{w}^*)}{p_2(\bolds{w}^*)} = \frac{p(w^* \,|\, \bolds{w}_0, \bolds{w}_1)}{p(w^* \,|\, \bolds{w}_0, \bolds{w}_2)}$ and
\begin{align*}
KL(p_2 \| p) - KL(p_1 &\| p)
= \int \log p(w^* \,|\, \bolds{w}_0, \bolds{w}_1) p(\bolds{w}^*) d\bolds{w}^* - \int \log p(w^* \,|\, \bolds{w}_0, \bolds{w}_2) p(\bolds{w}^*) d\bolds{w}^*\\
&= H_p(w^* \,|\, \bolds{w}_0, \bolds{w}_2) - H_p(w^* \,|\, \bolds{w}_0, \bolds{w}_1).
\end{align*}
This result suggests placing a new reference location $\bolds{s}^*$ in the reference subset \textit{least} uncertain about the realization of the process at $\bolds{s}^*$. We interpret this as justifying recursive domain partitioning on ${\cal S}$ in spatial contexts in which local spatial clusters of locations are likely less uncertain about process realization in the same spatial region.
In the remainder of this article, we will consider a given reference set ${\cal S}$ which typically will be based on a subset of observed locations; the combinatorial problem of selecting an optimal ${\cal S}$ (in some sense) is beyond the scope of this article. If ${\cal S}$ is not partitioned, it can be considered as a set of knots or ``sensors'' and one can refer to a large literature on experimental design and optimal sensor placement \citep[see e.g.][and references therein]{krauseetal2008}. It might be possible to extend previous work on adaptive knot placement \citep{pp_adaptive_knots}, but this will come at a steep cost in terms of computational performance.
\section{Bayesian spatial regressions using \textsc{SpamTrees}} \label{section:spamtree_gp}
Suppose we observe an $l$-variate outcome at spatial locations $\bolds{\ell}\in {\cal D} \subset \Re^{d}$ which we wish to model using a spatially-varying regression model:
\begin{equation}\label{eq:linear_svc}
\begin{aligned}
y_j( \bolds{\ell} ) &= \bolds{x}_j(\bolds{\ell})^\top \bolds{\beta}_j + \sum_k z_{jk}(\bolds{\ell}) w(\bolds{\ell}, \bolds{\xi}_k) + \varepsilon_j(\bolds{\ell}), \quad j=1, \dots, l,
\end{aligned}
\end{equation}
where $y_j(\bolds{\ell})$ is the $j$-th point-referenced outcome at $\bolds{\ell}$, $\bolds{x}_j(\bolds{\ell})$ is a $p_j \times 1$ vector of spatially referenced predictors linked to constant coefficients $\bolds{\beta}_j$, $\varepsilon_j(\bolds{\ell})\iidsim N(0, \tau^2_j)$ is the measurement error for outcome $j$, and $z_{jk}(\bolds{\ell})$ is the $k$-th (of $q$) covariates for the $j$-th outcome modeled with spatially-varying coefficient $w(\bolds{\ell}, \bolds{\xi}_k)$, $\bolds{\ell} \in {\cal D}$, $\bolds{\xi}_k \in \Xi$.
This coefficient $w(\bolds{\ell}, \bolds{\xi}_k)$
corresponds to the $k$-th margin of a $q$-variate Gaussian process $\{\bolds{w}(\bolds{\ell}) : \bolds{\ell} \in {\cal D}\}$ denoted as $\bolds{w}(\bolds{\ell}) \sim GP(\mathbf{0}, \bolds{C}_{\bolds{\theta}}(\cdot, \cdot))$ with cross-covariance $\bolds{C}_{\bolds{\theta}}$ indexed by unknown parameters $\bolds{\theta}$ which we omit in notation for simplicity. A valid cross-covariance function is defined as $\bolds{C}_{\bolds{\theta}}: {\cal D}\times{\cal D} \rightarrow {\cal M}_{q\times q}$, where ${\cal M}_{q\times q}$ is a subset of the space of all $q\times q$ real matrices $\Re^{q\times q}$. It must satisfy $\bolds{C}(\bolds{\ell},\bolds{\ell}') = \bolds{C}(\bolds{\ell}',\bolds{\ell})^{\top}$ for any two locations $\bolds{\ell}, \bolds{\ell}' \in {\cal D}$, and $\sum_{i=1}^n\sum_{j=1}^n \bolds{z}_i^{\top}\bolds{C}(\bolds{\ell}_i,\bolds{\ell}_j)\bolds{z}_j > 0$ for any integer $n$ and finite collection of points $\{\bolds{\ell}_1,\bolds{\ell}_2,\ldots,\bolds{\ell}_n\}$ and for all $\bolds{z}_i \in \Re^{q}\setminus \{\bolds{0}\}$.
We replace the full GP with a Gaussian \modelname\ for scalable computation considering the $q$-variate multivariate Gaussian process $\bolds{w}(\cdot)$ as the base process. Since the $(i,j)$-th entry of $\bolds{C}(\bolds{\ell},\bolds{\ell}')$ is $\bolds{C}(\bolds{\ell}, \bolds{\ell}')_{i,j}=\mbox{Cov}(w_i(\bolds{\ell}),w_j(\bolds{\ell}'))$, i.e. the covariance between the $i$-th and $j$-th elements of $\bolds{w}(\bolds{\ell})$ at $\bolds{\ell}$ and $\bolds{\ell}'$, we can obtain a covariance function on the augmented domain $\bolds{C}^* : {\cal D}^* \times {\cal D}^* \rightarrow \Re$ as $\bolds{C}^*((\bolds{\ell}, \bolds{\xi}), (\bolds{\ell}', \bolds{\xi}')) = \bolds{C}(\bolds{\ell}, \bolds{\ell}')_{i,i'}$ where $\bolds{\xi}$ and $\bolds{\xi}'$ are the locations in $\Xi$ of variables $i$ and $j$, respectively. \cite{apanasovich_genton2010} use a similar representation to build valid cross-covariances based on existing univariate covariance functions; their approach amounts to considering $\bolds{\xi}$ or $\| \bolds{\xi} - \bolds{\xi}' \|$ as a parameter to be estimated. Our approach can be based on any valid cross-covariance as we may just set $\Xi = \{1, \dots, q\}$. Refer to e.g. \cite{genton_ccov} for an extensive review of cross-covariance functions for multivariate processes. Moving forward, we will not distinguish between $\bolds{C}^*$ and $\bolds{C}$. The linear multivariate spatially-varying regression model (\ref{eq:linear_svc}) allows the $l$ outcomes to be observed at different locations; we later consider the case $l=q$ and $\bolds{Z}(\bolds{\ell}) = I_q$ resulting in a multivariate space-varying intercept model.
\subsection{Gaussian \textsc{SpamTrees}}
Enumerate the set of nodes as $\bolds{V} = \{ \bolds{v}_1, \dots, \bolds{v}_{m_V}\}$, $m_V = m_S + m_U$ and denote $\bolds{w}_i = w(\eta^{-1}(\bolds{v}_i))$, $\bolds{C}_{ij}$ as the $n_i \times n_j$ covariance matrix between $\bolds{w}_i$ and $\bolds{w}_j$, $\bolds{C}_{i,[i]}$ the $n_i\times J_i$ covariance matrix between $\bolds{w}_i$ and $\bwpa{i}$, $\bolds{C}_{i}$ the $n_i \times n_i $ covariance matrix between $\bolds{w}_{i}$ and itself, and $\bolds{C}_{[i]}$ the $J_i \times J_i $ covariance matrix between $\bwpa{i}$ and itself.
A base Gaussian process induces $\widetilde{p}(\bolds{w}_{{\cal S}}) = \prod_{j : \{\bolds{v}_j \in \bolds{A} \}} N( \bolds{w}_j \mid \bolds{H}_{j} \bwpa{j}, \bolds{R}_{j})$, where
\begin{equation}\label{equation:h_and_r}
\bolds{H}_{j} = \bolds{C}_{j, [j]} \bolds{C}^{-1}_{[j]} \quad\text{and}\quad \bolds{R}_{j} = \bolds{C}_{j} - \bolds{C}_{j, [j]}\bolds{C}^{-1}_{[j]} \bolds{C}_{[j],j},
\end{equation}
implying that the joint density $\tilde{p}(\bolds{w}_{{\cal S}})$ is multivariate normal with covariance $\tilde{\bolds{C}}_{{\cal S}}$ and precision matrix $\tilde{\bolds{C}}^{-1}_{{\cal S}}$. At ${\cal U}$ we have $\widetilde{p}( \bolds{w}_{{\cal U}} \mid \bolds{w}_{{\cal S}}) = \prod_{j : \{ \bolds{v}_j \in \bolds{B} \}} N( \bolds{w}_{j} \mid \bolds{H}_{j}\bwpa{j}, \bolds{R}_{j} )$, where $\bolds{H}_{j}$ and $\bolds{R}_{j}$ are as in (\ref{equation:h_and_r}). All quantities can be computed using the base cross-covariance function. Given that the $\widetilde{p}$ densities are Gaussian, so will be the finite dimensional distributions.
The treed graph ${\cal G}$ leads to properties which we analyze in more detail in Appendix \ref{appendix:gp_properties} and summarize here.
For two nodes $\bolds{v}_i, \bolds{v}_j \in \bolds{V}$ denote the \textit{common descendants} as $\text{cd}(\bolds{v}_i, \bolds{v}_j) = (\{ \bolds{v}_i \} \cup \ch{\bolds{v}_i}) \cap (\{ \bolds{v}_j \} \cup \ch{\bolds{v}_j})$. If $\bolds{v}_i \in \pa{\bolds{v}_j}$ denote $\bolds{H}_{i\to j}$ and $\bolds{H}_{\setminus i \to j}$ as the matrix obtained by subsetting $\bolds{H}_j$ to columns corresponding to $\bolds{v}_i$, or to $\pa{\bolds{v}_j}\setminus\{\bolds{v}_i \}$, respectively. Similarly define $\bolds{w}_{[i\to j]} = \bolds{w}_i$ and $\bolds{w}_{[\setminus i \to j]}$. As a special case, if the tree depth is $\delta=1$ and $\{ \bolds{v}_j\} = \pa{\bolds{v}_i}$ then $\text{cd}(\bolds{v}_i, \bolds{v}_j) = \{ \bolds{v}_i \}$, $\bolds{H}_{i\to j} = \bolds{H}_j$, and $\bolds{w}_{[i\to j]} = \bwpa{j}$.
Define $\bolds{\mathcal{H}}$ as the matrix whose $(i,j)$ block is $\bolds{\mathcal{H}}_{ij} = \bolds{O}_{n_i \times n_j}$ if $\bolds{v}_j \notin \pa{\bolds{v}_i}$, and otherwise $\bolds{\mathcal{H}}_{ij} = \bolds{H}_{j\to i}$.
\subsubsection{Precision matrix}
The $(i,j)$ block of the precision matrix at both reference and non-reference locations $\tilde{\bolds{C}}^{-1}$ is denoted by $\tilde{\bolds{C}}^{-1}(i, j)$, with $i,j=1, \dots, m_V$ corresponding to nodes $\bolds{v}_i, \bolds{v}_j \in \bolds{V}$ for some $i,j$; it is nonzero if $\text{cd}(\bolds{v}_i, \bolds{v}_j) = \emptyset$, otherwise:
\begin{equation}\label{equation:precision_blocks_main}
\begin{aligned}
\tilde{\bolds{C}}^{-1}(i, j) &= \sum\limits_{\bolds{v}_k \in \text{cd}(\bolds{v}_i, \bolds{v}_j)}
(\bolds{I}_{ki} - \bolds{H}_{i\to k})^\top \bolds{R}_k^{-1} (\bolds{I}_{kj} - \bolds{H}_{j\to k})\\
&= \sum\limits_{\bolds{v}_k \in \text{cd}(\bolds{v}_i, \bolds{v}_j)}
(\bolds{I}_{ki} - \bolds{\mathcal{H}}_{ki})^\top \bolds{R}_k^{-1} (\bolds{I}_{kj} - \bolds{\mathcal{H}}_{kj}),
\end{aligned}
\end{equation}
where $\bolds{I}_{ij}$ is the $(i,j)$ block of an identity matrix with $n_{{\cal S}}+n_{{\cal U}}$ rows and is nonzero if and only if $i=j$. We thus obtain that the number of nonzero elements of $\tilde{\bolds{C}}^{-1}$ is
\begin{align} \label{equation:sparsity}
\text{nnz}(\tilde{\bolds{C}}^{-1}) &= \sum_{i=1}^{m_V} \left( 2n_{i} J_{i} + n_{i}^2\bolds{1}\{\bolds{v}_i \in \bolds{V}\}\right),
\end{align}
where $n_{i} = |\eta^{-1}(\bolds{v}_{i})|$, $J_{i} = | \eta^{-1}(\pa{\bolds{v}_{i}})|$, and by symmetry $(\tilde{\bolds{C}}^{-1}(i,j))^{\top} = \tilde{\bolds{C}}^{-1}(j,i)$.
If $\delta > 1$, the size of $\bolds{C}_{[i]}$ is larger for nodes $\bolds{v}_i$ at levels of the treed hierarchy farther from $\bolds{A}_{M_{\delta}}$. However suppose $\bolds{v}_i, \bolds{v}_j$ are such that $\pa{\bolds{v}_j} = \{ \bolds{v}_i \} \cup \pa{\bolds{v}_i}$. Then computing $\bolds{C}_{[j]}^{-1}$ proceeds more cheaply by recursively applying the following:
\begin{align} \label{equation:nested_inverse_main}
\bolds{C}_{[j]}^{-1} = \begin{bmatrix} \bolds{C}_{[i]}^{-1} + \bolds{H}_i^\top \bolds{R}_i^{-1}\bolds{H}_i & - \bolds{H}_i^\top \bolds{R}_i^{-1} \\
-\bolds{R}_i^{-1}\bolds{H}_i & \bolds{R}_i^{-1}\end{bmatrix}.
\end{align}
\subsubsection{Induced covariance}
Define a path from $\bolds{v}_k$ to $\bolds{v}_j$ as ${\cal P}_{k \to j} = \{\bolds{v}_{i_{1}}, \dots, \bolds{v}_{i_{r}} \}$ where $\bolds{v}_{i_{1}} = \bolds{v}_k$, $\bolds{v}_{i_{r}} = \bolds{v}_j$, and $\bolds{v}_{i_h} \in \pa{\bolds{v}_{i_{h+1}}}$. The longest path $\tilde{{\cal P}}_{k \to j}$ is such that if $\bolds{v}_k \in \bolds{A}_{r_k}$ and $\bolds{v}_j \in \bolds{A}_{r_j}$ then $|\tilde{{\cal P}}_{k \to j}| = r_j - r_k + 1$. The shortest path $\bar{{\cal P}}_{k \to j}$ is the path from $\bolds{v}_k$ to $\bolds{v}_j$ with minimum number of steps. We denote the longest path from the root to $\bolds{v}_j$ as $\tilde{{\cal P}}_{0\to j}$; this corresponds to the full set of ancestors of $\bolds{v}_j$, and $\pa{\bolds{v}_j} \subset \tilde{{\cal P}}_{0\to j}$. For two nodes $\bolds{v}_i$ and $\bolds{v}_j$ we have $(\pa{\bolds{v}_i} \cap \pa{\bolds{v}_j}) \subset (\tilde{{\cal P}}_{0\to i} \cap \tilde{{\cal P}}_{0\to j})$. We define the \textit{concestor} between $\bolds{v}_i$ and $\bolds{v}_j$ as $\text{con}(\bolds{v}_i, \bolds{v}_j) = \arg \max_{\bolds{v}_k \in \bolds{V}} \{ k : {\cal P}_{k\to i} \cap {\cal P}_{k\to j} \neq \emptyset \} $ i.e. the last common ancestor of the two nodes.
Take the path $\tilde{{\cal P}}_{M_{\delta} \to j}$ in ${\cal G}$ from a node at $\bolds{A}_{M_{\delta}}$ leading to $\bolds{v}_j$. After defining the cross-covariance function $\bolds{K}_i(\bolds{\ell}, \bolds{\ell}') = \bolds{C}_{\bolds{\ell}, \bolds{\ell}'} - \bolds{C}_{\bolds{\ell}, [i]} \bolds{C}^{-1}_{[i]} \bolds{C}_{[i], \bolds{\ell}'}$ and denoting $\bolds{K}_i(\bolds{\ell}, s) = \bolds{K}_i(\bolds{\ell}, \eta^{-1}(\bolds{v}_s))$ we can write
\begin{equation}\label{equation:additive_repres_main}
\begin{aligned}
\bolds{w}_j = \sum\limits_{s=i_{M_{\delta}}}^{i_{r-1}} \bolds{K}_{s}(j,s) \bolds{K}_{s}^{-1}(s,s) \bolds{e}_s + \bolds{e}_j,
\end{aligned}
\end{equation}
where for $s>i_{M_{\delta}}$ the $\bolds{e}_s$ are independent zero-mean GPs with covariance $K_s(\bolds{\ell}, \bolds{\ell}')$ and we set $K_{i_{M_{\delta}}}(\bolds{\ell}, \bolds{\ell}') = \bolds{C}(\bolds{\ell}, \bolds{\ell}')$ and $\bolds{e}_{i_{M_{\delta}}} = \bolds{w}_{i_{M_{\delta}}} \sim N(0, \bolds{C}_{i_{M_{\delta}}})$.
Take two locations $\bolds{\ell}, \bolds{\ell}'$ such that $\bolds{v}_i = \eta(\bolds{\ell}), \bolds{v}_j = \eta(\bolds{\ell}')$ and let $\bolds{v}_z = \text{con}(\bolds{v}_i, \bolds{v}_j)$; if $\pa{\bolds{v}_i} \cap \pa{\bolds{v}_j} \neq \emptyset$ then the above leads to
\begin{align}\label{equation:cov_recursive}
Cov_{\widetilde{p}}(\bolds{w}(\bolds{\ell}), \bolds{w}(\bolds{\ell}')) &= \sum\limits_{s \in \pa{\bolds{v}_i} \cap \pa{\bolds{v}_j}} \bolds{K}_{s}(\bolds{\ell},s) \bolds{K}_{s}^{-1}(s,s)\bolds{K}_{s}(s,\bolds{\ell}') + \bolds{1}\{\bolds{\ell} = \bolds{\ell}'\}\bolds{K}_j(\bolds{\ell}, \bolds{\ell}'),
\end{align}
where $\bolds{K}_z(\bolds{\ell}, \bolds{\ell}') = \bolds{C}(\bolds{\ell}, \bolds{\ell}')$. If $\pa{\bolds{v}_i} \cap \pa{\bolds{v}_j} = \emptyset$ take the shortest paths $\bar{{\cal P}}_{z \to i} = \{ i_1, \dots, i_{r_i}\}$ and $\bar{{\cal P}}_{z\to j} = \{ j_1, \dots, j_{r_j} \}$; setting $\bolds{F}_{i_h} = \bolds{C}_{i_h, i_{h-1}} \bolds{C}_{i_{h-1}}^{-1}$ we get
\begin{align}\label{equation:cov_limited}
Cov_{\widetilde{p}}(\bolds{w}(\bolds{\ell}), \bolds{w}(\bolds{\ell}')) &= \bolds{F}_{i_{r_i}} \cdots \bolds{F}_{i_{1}} \bolds{C}_{z} \bolds{F}_{j_{1}}^{\top} \cdots \bolds{F}_{j_{r_j}}^{\top}.
\end{align}
In particular if $\delta=M$ then $\pa{\bolds{v}_i} \cap \pa{\bolds{v}_j} \neq \emptyset$ for all $i,j$ and only (\ref{equation:cov_recursive}) is used, whereas if $\delta=1$ then the only scenario in which (\ref{equation:cov_recursive}) holds is $\{ \bolds{v}_z \} = \pa{\bolds{v}_i} \cap \pa{\bolds{v}_j}$ in which case the two are equivalent.
In univariate settings, the special case in which $\delta = M$, and hence $M_{\delta} = 0$, leads to an interpretation of (\ref{equation:additive_repres_main}) as a basis function decomposition; considering all leaf paths ${\cal P}_j$ for $\bolds{v}_j \in \bolds{B}$, this leads to an MRA \citep{katzfuss_jasa17, katzfussgong2019}. On the other hand, keeping other parameters constant, $\delta<M$ and in particular $\delta=1$ may be associated to savings in computing cost, leading to a trade-off between graph complexity and size of reference subsets; see Appendix \ref{appendix:computing_cost}.
\subsubsection{Block-sparse Cholesky decompositions}
In recent work \cite{jurekkatzfuss2020} consider sparse Cholesky decompositions of covariance and precision matrices for treed graphs corresponding to the case $\delta=M$ above in the context of space-time filtering; their methods involve sparse Cholesky routines on reverse orderings of $\tilde{\bolds{C}}^{-1}$ at the level of individual locations. In doing so, the relationship between Cholesky decompositions and ${\cal G}$, $\tilde{\bolds{C}}^{-1}$ and the block structure in ${\cal S}$ remains somewhat hidden, and sparse Cholesky libraries are typically associated to bottlenecks in MCMC algorithms. However we note that a consequence of (\ref{equation:precision_blocks_main}) is that it leads to a direct algorithm, for any $\delta$, for the block-decomposition of any symmetric positive-definite matrix $\bolds{\Lambda}$ conforming to ${\cal G}$, i.e. with the same block-sparse structure as $\tilde{\bolds{C}}^{-1}$. This allows us to write $\bolds{\Lambda} = (\bolds{I} - \bolds{L})^\top \bolds{D} (\bolds{I} - \bolds{L})$ where $\bolds{I}$ is the identity matrix, $\bolds{L}$ is block lower triangular with the same block-sparsity pattern as $\bolds{\mathcal{H}}$ above, and $\bolds{D}$ is block diagonal symmetric positive-definite. In Appendix \ref{appendix:precision_decomposition} we outline Algorithm \ref{algorithm:precision_decomposition} which \textit{(i)} makes direct use of the structure of ${\cal G}$, \textit{(ii)} computes the decomposition at blocks of reference and non-reference locations, and \textit{(iii)} requires no external sparse matrix library, in particular no sparse Cholesky solvers. Along with Algorithm \ref{algorithm:precision_decomp_invtri} for the block-computation of $(\bolds{I}-\bolds{L})^{-1}$, it can be used to compute $\bolds{\Lambda}^{-1}= (\tilde{\bolds{C}}^{-1} + \bolds{\Sigma})^{-1}$ where $\bolds{\Sigma}$ is a block-diagonal matrix; it is thus useful in computing the Gaussian integrated likelihood.
\subsection{Estimation and prediction} \label{section:general:estimation}
We introduce notation to aid in obtaining the full conditional distributions. Write (\ref{eq:linear_svc}) as
\begin{equation}\label{eq:linear_svc_exp}
\begin{aligned}
\bolds{y}(\bolds{\ell}) &= \bolds{X}(\bolds{\ell}) \bolds{\beta} + \bolds{Z}(\bolds{\ell}) \bolds{w}(\bolds{\ell}) + \bolds{\varepsilon}(\bolds{\ell}),
\end{aligned}
\end{equation}
where $\bolds{y}(\bolds{\ell}) = ( \{y_j(\bolds{\ell})\}_{j=1}^l )^{\top}$, $\bolds{\varepsilon}(\bolds{\ell}) = ( \{\varepsilon_j(\bolds{\ell})\}_{j=1}^l )^{\top} \sim N(\mathbf{0}, \bolds{D}_{\tau})$, $\bolds{D}_{\tau} = \text{diag}(\tau^2_1, \dots, \tau^2_l)$, $\bolds{X}(\bolds{\ell}) = \text{blockdiag}\{ \bolds{x}_j(\bolds{\ell})^{\top}, j=1,\dots,l \}$, $\bolds{\beta} = (\bolds{\beta}_{p_1}^\top, \dots, \bolds{\beta}_{p_j}^{\top})^{\top}$. The $l \times q$ matrix $\bolds{Z}(\bolds{\ell}) = (\bolds{z}_j(\bolds{\ell})^{\top}, j=1,\dots,l)$ with $\bolds{z}_j(\bolds{\ell})^{\top} = (z_{jk}(\bolds{\ell}), k=1, \dots,q)$ acts a design matrix for spatial location $\bolds{\ell}$. Collecting all locations along the $j$-th margin, we build ${\cal T}_j = \{ \bolds{\ell}_1^{(j)}, \dots, \bolds{\ell}_{N_j}^{(j)}\}$ and ${\cal T} = \cup_j {\cal T}_j$. We then call $\bolds{y}^{(j)} = (y_j(\bolds{\ell}_1^{(j)}), \dots, y_j(\bolds{\ell}_{N_j}^{(j)}))^\top$ and $\bolds{\varepsilon}^{(j)}$ similarly, $\bolds{X}^{(j)}= (\bolds{x}_j(\bolds{\ell}_1^{(j)}), \dots, \bolds{x}_j(\bolds{\ell}_{N_j}^{(j)}))^\top$, $\bolds{w}^{(j)}= (\bolds{w}(\bolds{\ell}_1^{(j)}, \bolds{\xi})^{\top}, \dots, \bolds{w}(\bolds{\ell}_{N_j}^{(j)}, \bolds{\xi})^{\top})^\top$ and $\bolds{Z}^{(j)} = \text{blockdiag}\{ \bolds{z}_{j}(\bolds{\ell}_s^{(j)})^\top \}_{s=1}^{N_j}$. The full observed data are $\bolds{y}, \bolds{X}, \bolds{Z}$. Denoting the number of observations as $n = \sum_{j=1}^l N_j$, $\bolds{Z}$ is thus a $n \times q n$ block-diagonal matrix, and similarly $\bolds{w}$ is a $qn \times 1$ vector. We introduce the diagonal matrix $\bolds{D}_n$ such that $\text{diag}(\bolds{D}_n) = (\tau^2_1\bolds{1}_{N_1}^{\top}, \dots, \tau^2_l \bolds{1}_{N_l}^{\top})^{\top}$.
By construction we may have $\eta(S_i) = \bolds{v}_i$ and $\eta(S_j) = \bolds{v}_j$ such that $(\bolds{\ell}, \bolds{\xi}) \in S_i$ and $(\bolds{\ell}', \bolds{\xi}') \in S_j$ where $\bolds{\ell}'=\bolds{\ell}$, $\bolds{\xi}\neq \bolds{\xi}'$ and similarly for non-reference subsets. Suppose ${\cal A} \subset {\cal D} \times \Xi$ is a generic reference or non-reference subset. We denote $\bar{{\cal A}} \subset {\cal D} \times \Xi$ as the set of all combinations of spatial locations of ${\cal A}$ and variables i.e. $\bar{{\cal A}} = {\cal A}\big|_{{\cal D}} \times {\cal A}\big|_{\Xi}$ where ${\cal A}\big|_{{\cal D}} \subset {\cal D}$ is the set of unique spatial locations in ${\cal A}$ and ${\cal A}\big|_{\Xi}$ are the unique latent variable coordinates.
By subtraction we find ${\cal A}_{-} = \bar{{\cal A}} \setminus {\cal A}$ as the set of locations whose spatial location is in ${\cal A}$ but whose variable is not. Let $\bolds{y}(\bar{{\cal A}}) = \bolds{y}({\cal A}) = (\{ \bolds{y}(\bolds{\ell}), \bolds{\ell} \in {\cal A}\big|_{{\cal D}} \} )^\top$, $\bolds{X}(\bar{{\cal A}}) = \bolds{X}({\cal A}) = \text{blockdiag}\{ \bolds{X}(\bolds{\ell})^{\top}, \bolds{\ell} \in {\cal A}\big|_{{\cal D}}\}$; values corresponding to unobserved locations will be dealt with by defining $\tilde{\bolds{D}}_n({\cal A})$ as the diagonal matrix obtained from $\bolds{D}_n$ by replacing unobserved outcomes with zeros.
Denote $\bolds{Z}(\bar{{\cal A}}) = \text{blockdiag}\{ \bolds{Z}(\bolds{\ell}), \bolds{\ell} \in {\cal A}\big|_{{\cal D}} \}$ and $\bolds{w}(\bar{{\cal A}})$ similarly. If ${\cal A}$ includes $L$ unique spatial locations then $\bolds{y}(\bar{{\cal A}})$ is a $L\ l\times 1$ vector and $\bolds{X}({\cal A})$ is a $L\ l \times pl$ matrix. In particular, $\bolds{Z}(\bar{{\cal A}})$ is a $L\ l \times Lql$ matrix; the subset of its columns with locations in ${\cal A}$ is denoted as $\bolds{Z}({\cal A})$ whereas at other locations we get $\bolds{Z}({\cal A}_{-})$. We can then separate the contribution of $\bolds{w}({\cal A})$ to $\bolds{y}({{\cal A}})$ from the contribution of $\bolds{w}({\cal A}_{-})$ by writing $\bolds{y}({{\cal A}}) = \bolds{X}({{\cal A}}) \bolds{\beta} + \bolds{Z}({\cal A}_{-}) \bolds{w}({\cal A}_{-}) + \bolds{Z}({\cal A}) \bolds{w}({\cal A}) + \bolds{\varepsilon}({{\cal A}})$, using which we let $\tilde{\bolds{y}}({{\cal A}}) = \bolds{y}({{\cal A}}) - \bolds{X}({{\cal A}}) \bolds{\beta} - \bolds{Z}({\cal A}_{-}) \bolds{w}({\cal A}_{-})$.
With customary prior distributions $\bolds{\beta} \sim N(\mathbf{0}, \bolds{V}_{\beta})$ and $\tau^2_j \sim Inv.Gamma(a_{\tau}, b_{\tau})$ along with a Gaussian \modelname\ prior on $\bolds{w}$, we obtain the posterior distribution as
\begin{align} \label{equation:posterior_distribution}
p(\bolds{w}, \bolds{\beta}, \{\tau_j^2\}_{j=1}^l, \bolds{\theta} \,|\, \bolds{y}) &\propto p(\bolds{y} \,|\, \bolds{w}, \bolds{\beta}, \{\tau_j^2\}_{j=1}^l) p(\bolds{w} \,|\, \bolds{\theta}) p(\bolds{\theta}) p(\bolds{\beta}) \prod_{j=1}^l p(\tau^2_j).
\end{align}
We compute the full conditional distributions of unknowns in the model, save for $\bolds{\theta}$; iterating sampling from each of these distributions corresponds to a Gibbs sampler which ultimately leads to samples from the posterior distribution above.
\subsubsection{Full conditional distributions}
The full conditional distribution for $\bolds{\beta}$ is Gaussian with covariance $\bolds{\Sigma}^*_{\bolds{\beta}} = (\bolds{V}_{\bolds{\beta}}^{-1} + \bolds{X}^{\top} \bolds{D}_n^{-1} \bolds{X})^{-1}$ and mean $\mu^*_{\bolds{\beta}} = \bolds{\Sigma}_{\bolds{\beta}} \bolds{X}^{\top} \bolds{D}_n^{-1} (\bolds{y} - \bolds{Z} \bolds{w})$. For $j=1, \dots, l$, $p(\tau^2_j \,|\, \bolds{\beta}, \bolds{y}, \bolds{w}) = Inv.Gamma(a_{\tau, j}^*, b^*_{\tau, j})$ where $a_{\tau, j}^* = a_{\tau} + N_j/2$ and $b^*_{\tau, j} = b_{\tau} + \frac{1}{2}\bolds{E}^{(j)\top}\bolds{E}^{(j)}$ with $\bolds{E}^{(j)} = \bolds{y}^{(j)} - \bolds{X}^{(j)}\bolds{\beta}_j - \bolds{Z}^{(j)}\bolds{w}^{(j)}$.
Take a node $\bolds{v}_i \in \bolds{V}$. If $\bolds{v}_i \in \bolds{A}$ then $\eta^{-1}(\bolds{v}_i) = S_i$ and for $\bolds{v}_j \in \ch{\bolds{v}_i}$ denote $\tilde{\bolds{w}}_j = \bolds{w}_j - \bolds{H}_{\setminus i \to j} \bolds{w}_{[\setminus i \to j]}$. The full conditional distribution of $\bolds{w}_i$ is $N(\bolds{\mu}_{i}, \bolds{\Sigma}_{i})$, where
\begin{equation}\label{equation:reference_full_conditional}
\begin{aligned}
\bolds{\Sigma}_{i}^{-1} = \bolds{Z}(S_i)^{\top}\bolds{D}_n(S_i)^{-1} & \bolds{Z}(S_i) + \bolds{R}_i^{-1} + \bolds{F}^{(c)}_i \\
\bolds{\Sigma}_{i}^{-1}\bolds{\mu}_{i} = \bolds{Z}(S_i)^{\top} \bolds{D}_n(S_i)^{-1} & \tilde{\bolds{y}}(S_i) + \bolds{R}_i^{-1}\bolds{H}_i\bolds{w}_{[i]} + \bolds{m}^{(c)}_i\\
\bolds{F}^{(c)}_i = \sum_{j: \{\bolds{v}_j \in \ch{\bolds{v}_i}\}} \bolds{H}_{i\to j}^{\top} \bolds{R}_j^{-1} \bolds{H}_{i\to j} \qquad & \qquad \bolds{m}^{(c)}_i = \sum_{j: \{\bolds{v}_j \in \ch{\bolds{v}_i}\}} \bolds{H}_{i\to j}^{\top} \bolds{R}_j^{-1} \tilde{\bolds{w}}_j
\end{aligned}
\end{equation}
If $\bolds{v}_i \in \bolds{B}$ instead $\bolds{\Sigma}_{i} = (\bolds{Z}(U_i)^{\top}\bolds{D}_n(U_i)^{-1} \bolds{Z}(U_i) + \bolds{R}_i)^{-1}$ and $\bolds{\mu}_{i} = \bolds{\Sigma}_i (\bolds{Z}(U_i)^{\top} \bolds{D}_n(U_i)^{-1}\tilde{\bolds{y}}(U_i) + \bolds{R}_i^{-1}\bolds{H}_i\bolds{w}_{[i]})$. Sampling of $\bolds{w}$ at nodes at the same level $r$ proceeds in parallel given the assumed conditional independence structure in ${\cal G}$. It is thus essential to minimize the computational burden at levels with a small number of nodes to avoid bottlenecks. In particular computing $\bolds{F}^{(c)}_i$ and $\bolds{m}^{(c)}_i$ can become expensive at the root when the number of children is very large. In Algorithm \ref{algorithm:gibbs} we show that one can efficiently sample at a near-root node $\bolds{v}_i$ by updating $\bolds{F}^{(c)}_i$ and $\bolds{m}^{(c)}_i$ via message-passing from the children of $\bolds{v}_i$.
\begin{figure}[tp]
\vspace*{-\baselineskip}
\begin{minipage}{\columnwidth}
\begin{algorithm}[H]
\small
\SetKwInOut{Input}{Input}
\textbf{Initialize:} \( \text{\Large$\ell $} = 0 \)\;
\For{$r \in \{0, \dots, M\}$}{
\For(\tcp*[f]{[parallel for]}){$j : \{ \bolds{v}_{j} \in \bolds{V}_r\}$}{
Compute $\bolds{R}_j^{-1} = (\bolds{C}_j - \bolds{C}_{j,[j]} \bolds{C}_{[j]}^{-1} \bolds{C}_{[j],j})^{-1}$ and $|\bolds{R}_j^{-1}|$\;
$\text{\Large$\ell $} = \text{\Large$\ell $} + \frac{1}{2}\log|\bolds{R}_j^{-1}| - \frac{1}{2} (\bolds{w}_j - \bolds{H}_j \bolds{w}_{[j]})^{\top} \bolds{R}_j^{-1} (\bolds{w}_j - \bolds{H}_j \bolds{w}_{[j]})$\;
\If{$\ch{\bolds{v}_j}\neq \emptyset $}{
Identify $\bolds{v}_i \in \ch{\bolds{v}_j}$ such that $\bolds{v}_i \in \bolds{V}_{r+1}$\;
Compute and store $\bolds{C}_{[i]}^{-1}$ (possibly via (\ref{equation:nested_inverse_main}))\;
}
}
}
\textbf{Result:} \( \exp(\text{\Large$\ell $}) \propto p(\bolds{w} \,|\, \bolds{\theta}) = \prod_i N(\bolds{w}_i \,|\, \bolds{H}_i \bolds{w}_{[i]}, \bolds{R}_i) \).
\caption{Computing $p(\bolds{w} \,|\, \bolds{\theta})$.}\label{algorithm:wpriorcompute}
\normalsize
\end{algorithm}
\end{minipage}
\begin{minipage}{\columnwidth}
\begin{algorithm}[H]
\small
\SetKwInOut{Input}{Input}
\textbf{Input:}$\bolds{C}_{[j]}$ for all $j$ from Algorithm \ref{algorithm:wpriorcompute};\\
$\bolds{W}_e = \bigcup\limits_{r \text{ is even}} \bolds{V}_r$; $\bolds{W}_o = \bigcup\limits_{r \text{ is odd}} \bolds{V}_r$;\\
\For{$i \in \{e, o\}$}{
\For(\tcp*[f]{[parallel for]}){$j : \{ \bolds{v}_{j} \in \bolds{W}_i\}$}{
Sample $\bolds{w}_j \sim N(\bolds{\mu}_j, \bolds{\Sigma}_j)$ using (\ref{equation:reference_full_conditional})\;
Let $\pa{\bolds{v}_{j}}=\{ \bolds{v}_{p} \}$, then $\bolds{m}^{(c)}_p = \bolds{H}^{\top}_{j} \bolds{R}^{-1}_j \bolds{w}_j$ and
$\bolds{F}^{(c)}_p = \bolds{H}^{\top}_{j} \bolds{R}^{-1}_j \bolds{H}_{j}$\;
}
}
\textbf{Result:} sample from $p(\bolds{w}_j \,|\, \bolds{w}_{-j}, \bolds{y}, \bolds{\beta}, \bolds{\theta}, \bolds{\tau})$ for all $\bolds{v}_j \in \bolds{V}$.
\caption{Sampling from the full conditional distribution of $\bolds{w}_i$ when $\delta = 1$.}\label{algorithm:gibbs_limited}
\normalsize
\end{algorithm}
\end{minipage}
\begin{minipage}{\columnwidth}
\begin{algorithm}[H]
\small
\SetKwInOut{Input}{Input}
\Input{$\bolds{C}_{[j]}$ for all $j$ from Algorithm \ref{algorithm:wpriorcompute}}
\textbf{Initialize:} for all $i$, $\bolds{m}^{(c)}_i = \mathbf{0}_{n_i \times 1}$ and $\bolds{F}^{(c)}_i = \bolds{O}_{n_i \times n_i}$\;
\For{$r \in \{M, \dots, 0\}$}{
\For(\tcp*[f]{[parallel for]}){$j : \{ \bolds{v}_{j} \in \bolds{V}_r\}$}{
Sample $\bolds{w}_j \sim N(\bolds{\mu}_j, \bolds{\Sigma}_j)$ using (\ref{equation:reference_full_conditional})\;
\For{$p : \{ \bolds{v}_{p} \in \pa{\bolds{v}_{j}} \} $}{
$\bolds{m}^{(c)}_p = \bolds{m}^{(c)}_p + \bolds{H}^{\top}_{p \to j} \bolds{R}^{-1}_j \bolds{w}_j$\;
$\bolds{F}^{(c)}_p = \bolds{F}^{(c)}_p + \bolds{H}^{\top}_{p \to j} \bolds{R}^{-1}_j \bolds{H}_{p \to j}$\;
}
}
}
\textbf{Result:} sample from $p(\bolds{w}_j \,|\, \bolds{w}_{-j}, \bolds{y}, \bolds{\beta}, \bolds{\theta}, \bolds{\tau})$ for all $\bolds{v}_j \in \bolds{V}$.
\caption{Sampling from the full conditional distribution of $\bolds{w}_j$ when $\delta = M$.}\label{algorithm:gibbs}
\normalsize
\end{algorithm}
\end{minipage}
\end{figure}
\subsubsection{Update of $\bolds{\theta}$}
The full conditional distribution of $\bolds{\theta}$---which may include $\bolds{\xi}_j$ for $j=1,\dots, q$ or equivalently $\delta_{ij} = \| \bolds{\xi}_i - \bolds{\xi}_j\|$ if the chosen cross-covariance function is defined on a latent domain of variables---is not available in closed form and sampling a posteriori can proceed via Metropolis-Hastings steps which involve accept/reject steps with acceptance probability $\alpha = \min \{1, \frac{p(\bolds{w} \,|\, \bolds{\theta}') p(\bolds{\theta}') q(\bolds{\theta} \,|\, \bolds{\theta}')}{p(\bolds{w} \,|\, \bolds{\theta}) p(\bolds{\theta}) q(\bolds{\theta}' \,|\, \bolds{\theta})}\} $. In our implementation, we adaptively tune the standard deviation of the proposal distribution via the robust adaptive Metropolis algorithm \citep[RAM;][]{vihola2012}. In these settings, unlike similar models based on DAG representations such as NNGPs and MGPs, direct computation via $p(\bolds{w} \,|\, \bolds{\theta}) = \prod_i N(\bolds{w}_i \,|\, \bolds{H}_i \bolds{w}_{[i]}, \bolds{R}_i)$ is inefficient as it requires computing $\bolds{C}^{-1}_{[i]}$ whose size grows along the hierarchy in ${\cal G}$. We thus outline Algorithm \ref{algorithm:wpriorcompute} for computing $p(\bolds{w} \,|\, \bolds{\theta})$ via (\ref{equation:nested_inverse_main}). As an alternative we can perform the update using ratios of $p(\bolds{y} \,|\, \bolds{\beta}, \bolds{\theta}, \bolds{\tau}) = \int p(\bolds{y} \,|\, \bolds{w}, \bolds{\beta}, \bolds{\tau}) p(\bolds{w} \,|\, \bolds{\theta}) d\bolds{w} = N(\bolds{y} \,|\, \bolds{X} \bolds{\beta}, \bolds{Z} \tilde{\bolds{C}} \bolds{Z}^{\top} + \bolds{D}_n)$ using Algorithms \ref{algorithm:precision_decomposition} and \ref{algorithm:precision_decomp_invtri} outlined in Appendix \ref{appendix:precision_decomposition} which require no sparse matrix library.
\subsubsection{Graph coloring for parallel sampling}
An advantage of the treed structure of ${\cal G}$ is that it leads to fixed graph coloring associated to parallel Gibbs sampling; no graph coloring algorithms are necessary \citep[see e.g.][]{molloyreed2002, lewis2016}. Specifically, if $\delta=M$ (full depth) then there is a one to one correspondence between the $M+1$ levels of ${\cal G}$ and graph colors, as evidenced by the parallel blocks in Algorithms \ref{algorithm:wpriorcompute} and \ref{algorithm:gibbs}. In the case $\delta=1$, ${\cal G}$ is associated to only two colors alternating the odd levels with the even ones. This is possible because the Markov blanket of each node at level $r$, with $r$ even, only includes nodes at odd levels, and vice-versa.
\subsubsection{Prediction of the outcome at new locations}
The Gibbs sampling algorithm will iterate across the above steps and, upon convergence, will produce samples from $p(\bolds{\beta}, \{ \tau^2_j \}_{j=1}^q, \bolds{w} \mid \bolds{y})$. We obtain posterior predictive inference at arbitrary $\bolds{\ell} \in {\cal D}$ by evaluating $p(\bolds{y}(\bolds{\ell})\,|\, \bolds{y})$. If $\bolds{\ell} \in {\cal S} \cup {\cal U}$, then we draw one sample of $\bolds{y}(\bolds{\ell}) \sim N(\bolds{X}(\bolds{\ell})^\top \bolds{\beta} + \bolds{Z}(\bolds{\ell})^\top \bolds{w}(\bolds{\ell}), \bolds{D}_n(\bolds{\ell}))$ for each draw of the parameters from $p(\bolds{\beta}, \{ \tau^2_j \}_{j=1}^l, \bolds{w} \mid \bolds{y})$. Otherwise, considering that $\eta(\bolds{\ell}) = \bolds{v}_j \in \bolds{B}$ for some $j$, with parent nodes $\pa{\bolds{v}_j}$, we sample $\bolds{w}(\bolds{\ell})$ from the full conditional $N(\bolds{\mu}_{\bolds{\ell}}^*, \bolds{\Sigma}_{\bolds{\ell}}^*)$, where $\bolds{\Sigma}_{\bolds{\ell}}^* = (\bolds{Z}(\bolds{\ell})\bolds{D}_n(\bolds{\ell})^{-1} \bolds{Z}(\bolds{\ell})^{\top} + \bolds{R}_{\bolds{\ell}}^{-1})^{-1}$ and $\bolds{\mu}_{\bolds{\ell}}^* = \bolds{\Sigma}_{\bolds{\ell}}^*(\bolds{Z}(\bolds{\ell}) \bolds{D}^{-1} (\bolds{y}(\bolds{\ell}) - \bolds{X}(\bolds{\ell})^\top \bolds{\beta} ) + \bolds{R}_{\bolds{\ell}}^{-1} \bolds{H}_{\bolds{\ell}} \bwpa{j})$, then draw $\bolds{y}(\bolds{\ell}) \sim N(\bolds{X}(\bolds{\ell})^\top \bolds{\beta} + \bolds{Z}(\bolds{\ell})^\top \bolds{w}(\bolds{\ell}), \bolds{D}_n)$.
\subsubsection{Computing and storage cost}\label{section:compute_cost}
The update of $\tau_j^2$ and $\bolds{\beta}$ can be performed at a minimal cost as typically $p = \sum_{j=1}^l p_j$ is small; almost all the computation budget must be dedicated to computing $p(\bolds{w} \,|\, \bolds{\theta})$ and sampling $p(\bolds{w} \,|\, \bolds{y}, \bolds{\beta}, \bolds{\tau}^2)$. Assume that reference locations are all observed ${\cal S} \subset {\cal T}$ and that all reference subsets have the same size i.e. $|S_i| = N_s$ for all $i$. We show in Appendix \ref{appendix:computing_cost} that the cost of computing \textsc{SpamTrees}\ is $O(n N_s^2)$. As a result, \textsc{SpamTrees}\ compare favorably to other models specifically in not scaling with the cube of the number of samples. $\delta$ does not impact the computational order, however, compared to $\delta=M$, choosing $\delta=1$ lowers the cost by a factor of $M$ or more. For a fixed reference set partition and corresponding nodes, choosing larger $\delta$ will result in stronger dependence between leaf nodes and nodes closer to the root, and this typically corresponds to leaf nodes being assigned conditioning sets that span larger distances in space. The computational speedup corresponding to choosing $\delta=1$ can effectively be traded for a coarser partitioning of ${\cal S}$, resulting in large conditioning sets that are more local to the leaves.
\section{Applications} \label{section:applications}
We consider Gaussian \textsc{SpamTrees}\ for the multivariate regression model (\ref{eq:linear_svc_exp}). Consider the spatial locations $\bolds{\ell}, \bolds{\ell}' \in {\cal D}$ and the locations of variables $i$ and $j$ in the latent domain of variables $\bolds{\xi}_i, \bolds{\xi}_j \in \Xi$, then denote $\bolds{h} = \| \bolds{\ell} - \bolds{\ell}' \|$, $\Delta = \delta_{ij} = \| \bolds{\xi}_i - \bolds{\xi}_j \|$, and \begin{align*}
C(\bolds{h}, \Delta) = \frac{\exp\left\{ - \phi \| \bolds{h} \|/\exp\left\{\frac{1}{2} \beta \log(1+\alpha \Delta )\right\} \right\} }{ \exp\left\{ \beta \log(1+\alpha \Delta ) \right\}}.
\end{align*}
For $j=1, \dots, q$ we also introduce $C_j(\bolds{h}) = \exp\left\{ - \phi_j \| \bolds{h} \|\right\}$. A non-separable cross-covariance function for a multivariate process can be defined as
\begin{equation}\label{eq:apanasovich_genton_covariance2}
\begin{aligned}
\text{Cov}(w(\bolds{\ell}, \bolds{\xi}_i), w(\bolds{\ell}', \bolds{\xi}_j)) = \bolds{C}_{ij}(\bolds{h}) &= \begin{cases}
\sigma^2_{i1} C(\bolds{h}, \delta_{ij}) + \sigma^2_{i2} C_i(\bolds{h}) & \text{if } i=j \\
\sigma_{i1} \sigma_{j1} C(\bolds{h}, \delta_{ij}) & \text{if } i\neq j,
\end{cases}
\end{aligned}
\end{equation}
which is derived from eq. (7) of \cite{apanasovich_genton2010}; locations of variables in the latent domain are unknown, therefore $\bolds{\theta} = \{\sigma_{i1}, \sigma_{i2}, \phi_i\}_{i=1,\dots,q} \cup \{ \delta_{ij} \}_{i=1,\dots, q}^{j<i} \cup \{ \alpha, \beta, \phi\}$ for a total of $3q + q(q-1)/2 + 3$ unknown parameters.
\subsection{Synthetic data} \label{section:simulations}
In this section we focus on bivariate outcomes ($q=2$) using the cross-covariance (\ref{eq:apanasovich_genton_covariance2}). For simplicity we set $\delta_{21} = 1, \alpha=1, \beta=1$. For each combination of $\sigma_{ij} \in \{1, 2\}$ for $i,j \in \{1,2\}$, $\phi_1 = \phi_2 \in \{0.1, 1, 10\}$, and $\phi \in \{ 0.1, 1, 10 \}$ we generate 25 data sets. Considering model (\ref{eq:linear_svc_exp}), we set $\bolds{\beta} = \mathbf{0}$, $\bolds{Z} = I_q$ and take the sampling locations as a regular grid of size $70\times70$ for a total of 4,900 spatial locations.
\begin{figure}
\centering
\includegraphics[width=.95\textwidth]{figures/example_simulated.png}
\caption{Left half: \textit{Full data set} -- a bivariate outcome is generated on 4,900 spatial locations. Right half: \textit{Observed data set} -- the training sample is built via subsampling each outcome at a smaller set of locations.}
\label{figure:synthetic_example}
\end{figure}
We simulate the spatial effects by sampling the full GP; the nuggets for the two outcomes are set to $\tau^2_1 = 0.01$ and $\tau^2_{2}=0.1$. As a result, we obtain 900 data sets, each simulating a realization of a bivariate outcome. We mimick real-world data measured irregularly in space by replacing the outcomes with missing values at $\approx 80\%$ of the spatial locations chosen uniformly at random and independently across the two margins of the outcome. In order to replicate the occurrence of regions with more sparsely observed outcomes, we replace outcomes with missing values at $\approx 99\%$ of spatial locations inside small circular areas whose center is chosen uniformly at random in $[0,1]^2$.
Figure \ref{figure:synthetic_example} shows one of the resulting 900 data sets which we use to evaluate the performance of \textsc{SpamTrees}. We consider multivariate \textsc{SpamTrees}\ with $\delta = M$ and $\delta=1$ and compare them with multivariate cubic meshed GPs \citep[Q-MGPs;][]{meshedgp}, integrated nested Laplace approximations \citep[INLA;][]{inla} implemented via \texttt{R-INLA} using a $15\times 15$ grid, a low-rank multivariate GP method (labeled \textsc{lowrank}) on 25 knots obtained via \textsc{SpamTrees}\ by setting $M=1$ with no domain partitioning, and an independent partitioning GP method (labeled \textsc{ind-part}) implemented by setting $M=1$ and partitioning the domain into 25 regions. Refer e.g. to \cite{Heaton2019} for an overview of low-rank and independent partitioning methods. We also include results from a non-spatial regression using Bayesian additive regression trees \citep[BART;][]{bart}. Each method was setup to target a compute time of 15 seconds for each data set. The total runtime for the 900 data sets thus amounted to about 26 hours.
Tables \ref{table:synthetic_results:predictions} and \ref{table:synthetic_results:estimation} summarise the results across all 900 data sets. All Bayesian methods based on latent GPs exhibit very good coverage; in these simulated scenarios, \textsc{SpamTrees}\ exhibit comparatively lower out-of-sample prediction errors. We highlight that the construction of DAG-based Bayesian methods for spatial regression depends on the underlying covariance function or kernel; comparisons of different covariance specifications across different methods is beyond the scope of this article.
Additional implementation details and figures can be found in Appendix \ref{appendix:synth_comparison}.
\begin{table}
\centering
\resizebox{.60\columnwidth}{!}{%
\begin{tabular}{|l|rrr|}
\hline
Model & Cov. (95\%) & \textsc{RMSE}$(\bolds{y})$ & \textsc{MAE}$(\bolds{y})$ \\
\hline
\textsc{SpamTrees}\ $\delta=M$ & 96.14 & \textbf{0.7171} & \textbf{0.5144} \\
\textsc{inla} & 92.83 & 0.7521 & 0.5742 \\
\textsc{q-mgp} & \textbf{95.87} & 0.7529 & 0.5410 \\
\textsc{SpamTrees}\ $\delta=1$ & 96.31 & 0.8027 & 0.5737 \\
\textsc{bart} & 92.61 & 0.8820 & 0.6895 \\
\textsc{lowrank} & 96.31 & 1.0230 & 0.7676 \\
\textsc{ind-part} & 95.97 & 1.0796 & 0.8198\\
\hline
\end{tabular}
}
\caption{Prediction performance on multivariate synthetic data: average coverage of 95\% prediction intervals, root mean square error (\textsc{RMSE}), and mean absolute error in prediction (\textsc{MAE}), over 900 data sets, sorted by lowest \textsc{RMSE}} \label{table:synthetic_results:predictions}
\end{table}
\begin{table}[ht]
\centering
\resizebox{.50\columnwidth}{!}{%
\begin{tabular}{|l|rr|}
\hline
Model & \textsc{RMSE}$(\bolds{y})$ & \textsc{MAE}$(\bolds{y})$ \\
\hline
\textsc{SpamTrees}\ $\delta=M$ & \textbf{2.2882} & 1.6873\\
\textsc{lowrank} & 2.6953 & \textbf{1.4895} \\
\textsc{q-mgp} & 2.8944 & 1.9630 \\
\textsc{ind-part} & 2.9604 & 1.6079 \\
\textsc{SpamTrees}\ $\delta=1$ & 3.5883 & 2.3322 \\
\hline
\end{tabular}
}
\caption{RMSE and MAE in the estimation of $\bolds{\theta}$ from covariance function (\ref{eq:apanasovich_genton_covariance2}), averaged over 900 data sets.} \label{table:synthetic_results:estimation}
\end{table}
\subsection{Climate data: MODIS-TERRA and GHCN} \label{section:applications:modisnoaa}
Climate data are collected from multiple sources in large quantities; when originating from satellites and remote sensing, they are typically collected at high spatial and relatively low temporal resolution. Atmospheric and land-surface products are obtained via post-processing of satellite imaging, and their quality is negatively impacted by cloud cover and other atmospheric disturbances. On the other hand, data from a relatively small number of land-based stations is of low spatial but high temporal resolution. An advantage of land-based stations is that they measure phenomena related to atmospheric conditions which cannot be easily measured from satellites (e.g. precipitation data, depth of snow cover).
We consider the joint analysis of five spatial outcomes collected from two sources. First, we consider Moderate Resolution Imaging Spectroradiometer (MODIS) data from the Terra satellite which is part of the NASA's Earth Observing System. Specifically, data product \texttt{MOD11C3} v. 6 provides monthly Land Surface Temperature (LST) values in a 0.05 degree latitude/longitude grid (the Climate Modeling Grid or CMG). The monthly data sets cover the whole globe from 2000-02-01 and consist of daytime and nighttime LSTs, quality control assessments, in addition to emissivities and clear-sky observations.
The second source of data is the Global Historical Climatology Network (GHCN) database which includes climate summaries from land surface stations across the globe subjected to common quality assurance reviews. Data are published by the National Centers of Environmental Information (NCEI) of the National Oceanic and Atmospheric Administration (NOAA) at several different temporal resolutions; daily products report five core elements (precipitation, snowfall, snow depth, maximum and minimum temperature) in addition to several other measurements.
We build our data set for analysis by focusing on the continental United States in October, 2018. The MODIS data correspond to 359,822 spatial locations. Of these, 250,874 are collected at the maximum reported quality; we consider all remaining 108,948 spatial locations as missing, and extract (1) daytime LST (\texttt{LST\_Day\_CMG}), (2) nighttime LST (\texttt{LST\_Night\_CMG}), (3) number of days with clear skies (\texttt{Clear\_sky\_days}), (4) number of nights with clear skies (\texttt{Clear\_sky\_nights}). From the GHCN database we use daily data to obtain monthly averages for precipitation (\texttt{PRCP}), which is available at 24,066 spatial locations corresponding to U.S. weather stations; we log-transform \texttt{PRCP}. The two data sources do not share measurement locations as there is no overlap between measurement locations in MODIS and GHCN, with the latter data being collected more sparsely---this is a scenario of complete spatial misalignment. For this reason we build \textsc{SpamTrees}\ favoring placement of GHCN locations at root nodes following Proposition \ref{prop:prop1}. Further implementation details are outlined at Appendix \ref{appendix:implementation}.
From the resulting data set of size $n=$1,027,562 we remove all observations in a large $3\times 3$ degree area in the central U.S. (from -100W to -97W and from 35N to 38N, i.e. the red area of Figure \ref{fig:usamap}) to build a test set on which we calculate coverage, MAE and RMSE of the predictions. We implement \textsc{SpamTrees}\ on the covariance function (\ref{eq:apanasovich_genton_covariance2}).
Figure \ref{fig:results_plot} maps the predictions at all locations and the corresponding posterior uncertainties. Comparisons with other methods are difficult due to data size and complete misalignment. We implemented a tessellated MGP with the same covariance function and targeting similar computing time (predictive performance is reported in Appendix \ref{appendix:modisnoaa}); \textsc{SpamTrees}\ displayed lower prediction errors and better coverage as seen in Table \ref{tab:predict_perf}.
We report posterior summaries of $\bolds{\theta}$ in Appendix \ref{appendix:modisnoaa}. Opposite signs of $\sigma_{i1}$ and $\sigma_{j1}$ for pairs of variables $i,j \in \{1, \dots, q\}$ imply a negative relationship; however, the degree of spatial decay of these correlations is different for each pair as prescribed by the latent distances in the domain of variables $\delta_{ij}$. Figure \ref{fig:covariance_plot} depicts the resulting cross-covariance function for three pairs of variables.
\begin{figure}
\centering
\includegraphics[width=.85\textwidth]{figures/application/usamap.png}
\caption{Prediction area}
\label{fig:usamap}
\end{figure}
\begin{table}
\begin{tabular}{c}
\resizebox{16cm}{!}{
\begin{tabular}{|r|ccccc|}
\hline
Measure & \footnotesize \texttt{Clear\_sky\_days} & \footnotesize\texttt{Clear\_sky\_nights} & \footnotesize\texttt{LST\_Day\_CMG} &\footnotesize \texttt{LST\_Night\_CMG} & \texttt{PRCP} \\
\hline
95\% Coverage & 0.9798 & 0.9894 & 1.0000 & 0.9993 & 0.9717 \\
MAE & 1.2824 & 1.3029 & 0.9686 & 0.8440 & 0.3517 \\
RMSE & 1.6114 & 1.6214 & 1.2547 & 1.0764 & 0.5168 \\
\hline
\end{tabular}}
\\
\resizebox{16cm}{!}{
\begin{tabular}{|c|c|c|c|}
\hline
$n=$1,014,017 & Total iterations: 30,000 & Total time: 16.14h &
Average time/iteration: 1.9s \\
\hline
\end{tabular}}
\end{tabular}
\caption{Prediction results of \textsc{SpamTrees}\ over the $3 \times 3$ degree area shown in Figure \ref{fig:usamap}}\label{tab:predict_perf}
\end{table}
\begin{figure}
\centering
\includegraphics[width=.95\textwidth]{figures/application/results_plot.png}
\caption{Predicted values of the outcomes at all locations (top row) and associated 95\% uncertainty (bottom row), with darker spots corresponding to wider credible intervals.}
\label{fig:results_plot}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=.95\textwidth]{figures/application/covariance_plot.png}
\caption{Given the latent dimensions $\delta_{ij}$, the color-coded lines represent $C(\bolds{h}, \delta_{ij})$ whereas $\bolds{C}_{ij}(\bolds{h}) = \sigma_{1i} \sigma_{1j} C(\bolds{h}, \delta_{ij})$ is shown as a dashed grey line.}
\label{fig:covariance_plot}
\end{figure}
\section{Discussion} \label{section:discussion}
In this article, we introduced \textsc{SpamTrees}\ for Bayesian spatial multivariate regression modeling and provided algorithms for scalable estimation and prediction. \textsc{SpamTrees}\ add significantly to the class of methods for regression in spatially-dependent data settings.
We have demonstrated that \textsc{SpamTrees}\ maintain accurate characterization of spatial dependence and scalability even in challenging settings involving multivariate data that are spatially misaligned. Such complexities create problems for competing
approaches, including recent DAG-based approaches ranging from NNGPs to MGPs.
One potential concern is the need for users to choose a tree, and in particular specify
the number of locations associated to each node and the multivariate composition of locations in each node. Although one can potentially estimate the tree structure based on the data, this would eliminate much of the computational speedup.
We have provided theoretical guidance based on KL divergence from the full GP and computational cost associated to different tree structures. This and our computational experiments lead to practical guidelines that can be used routinely in tree building. Choosing a tree is simpler than the common task of choosing a neural network architecture in deep learning, and provides a useful degree of user-input to refine and improve upon an approach.
We have focused on sampling algorithms for the latent effects because they provide a general blueprint which may be used for posterior computations in non-Gaussian outcome models; efficient algorithms for non-Gaussian big geostatistical data sets are currently lacking and are the focus of ongoing research.
Including time as a dimension is challenging and care must be taken when building a sparse DAG to avoid unreasonable assumptions on temporal dependence. For these reasons, future research may be devoted to building sparse DAG methods combining the advantages of treed structures with e.g. Markov-type assumptions of conditional independence.
\subsection*{Acknowledgements} This project has received funding from the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme (grant agreement No 856506). This project was also partially funded by grant R01ES028804 of the United States National Institutes of Health (NIH).
\newpage
\vspace{2cm}
\begin{center}
{\Huge \textsc{\textbf{Appendix}}}
\end{center}
|
\subsubsection{\@startsection{subsubsection}{3}%
\begin{abstract}
Serverless computing becomes the new trending paradigm in cloud computing, allowing developers to focus on the core application logic and rapidly prototype applications. Due to the great prospects of serverless computing, in recent years, most major cloud vendors have rolled out their commodity serverless computing platforms. However, the characteristics of these platforms have not been systematically studied. To fill this knowledge gap, this paper presents a comprehensive study on characterizing mainstream commodity serverless computing platforms (i.e., AWS Lambda, Azure Functions, Google Cloud Functions, and Alibaba Cloud Function Compute). First, we qualitatively analyze these platforms from development, deployment, and runtime aspects to form the taxonomy of characteristics. Additionally, we quantitatively evaluate the actual performance of different serverless computing platforms through our designed benchmark platform. Our benchmark platform has two types of benchmarks, i.e., microbenchmarks and macrobenchmarks. Based on the results of qualitative and quantitative analyses, we derive a series of findings and provide insightful implications for both developers and cloud vendors.
\end{abstract}
\maketitle
\section{Introduction}\label{intro}
Serverless computing is an emerging trend in cloud computing, allowing developers to deploy applications on the cloud in the auto-scaling and pay-as-you-go manner~\cite{JonasCoRR2019}\cite{cncfnew}. It is predicted that 50\% of global enterprises will employ serverless computing by 2025~\cite{gartner20new}. Serverless computing can make developers solely focus on developing their functions, which are small pieces of programs dedicated to a simple task (Function-as-a-Service, FaaS~\cite{JonasCoRR2019}). Meanwhile, developers can also directly apply existing proprietary serverless services (Backend-as-a-Service, BaaS~\cite{JonasCoRR2019}) provided by many cloud service vendors, e.g., object storage services provided by AWS S3~\cite{awsnew}.
Serverless computing can effectively simplify the application deployment and thus alleviate developers' efforts from tedious and error-prone server management. There have emerged various commodity serverless platforms including AWS Lambda~\cite{awsnew}, Google Cloud Functions~\cite{googlenew}, Azure Functions~\cite{azurenew}, and Alibaba Cloud Function Compute~\cite{alibabanew}. In a sense, most of the current serverless computing platforms often act as a ``black-box''. Without having to pay attention to the underlying details, however, developers may have difficulties to choose the target serverless computing platform that is the most adequate to serve their applications, as they have little knowledge of characteristics, configuration policies, actual running performance, and cost. For example, how can developers evaluate whether and how to transform their existing applications into serverless functions or how to package their tasks with newly developed functions on a specific platform? How many configuration options does a platform provide to developers to quickly meet their requirements? How about the runtime performance for a given task on a specific serverless computing platform? These issues are really non-trivial in practice as they can definitely impact the developer's decision making, and thus potentially impact the quality of service, user experience, and even the revenue of the application.
Unfortunately, to the best of our knowledge, there still exists a gap when developers select the most adequate serverless computing platform. In this paper, we conduct an extensive empirical study to characterize four mainstream serverless computing platforms, including AWS Lambda, Azure Functions, Google Cloud Functions, and Alibaba Cloud Function Compute. Specifically, our study is conducted from two folds in terms of qualitative analysis and quantitative characterization.
First, we seek to qualitatively explore the various characteristics described in the official documentations of these serverless computing platforms. Such characteristics specify the inherent restrictions of different aspects involving developing, deploying, and executing functions, which will result in fatal failures if developers do not comply with it. To this end, we construct a taxonomy with respect to the need-to-consider information from three aspects,i.e., development, deployment, and runtime. Such a taxonomy can help developers better understand the supported characteristics of the serverless computing platforms to facilitate further developers' development practice.
Then, we quantitatively evaluate the actual performance of different serverless computing platforms. We design an automated benchmark platform named \textbf{\textit{TBS}} including a benchmark suite (named \textbf{\textit{SlsBench}}). \textit{TBS} can automatize the package, deployment, execution, log collection, clean-up, and result generation of each round test. Furthermore, \textit{SlsBench} contained in \textit{TBS} has two types of benchmarks, i.e., microbenchmarks and macrobenchmarks. Specifically, microbenchmarks consist of a set of simple workloads focusing on specific resource consumption, such as CPU, memory, network, disk IO, etc. Macrobenchmarks consist of a set of real-world representative applications (e.g., multimedia data process, MapReduce, machine-learning-based serving, etc.), which require to utilize various system resources. Based on \textit{TBS}, we conduct a set of extensive experiments under various configurations to evaluate the actual system performance (i.e., startup latency and resource efficiency) on different serverless computing platforms. Additionally, we make \textit{TBS} and \textit{SlsBench} public and open source\footnote{They will be made public later.}. Our experimental results show (i) how different configurations impact the running performance of functions and (ii) the performance variation when running different kinds of workloads. Finally, we report a series of findings and implications. Our findings cannot only help serverless developers choose the right configurations and platforms to obtain the optimal performance based on their actual workloads, but also guide cloud vendors to improve their serverless computing platforms.
Some highlights of our findings about the fatal restrictions of different platform and best practice to choose the platform that can achieve better performance (e.g., high throughput, low startup latency, and execution latency), including:
\noindent $\bullet$ \textbf{Development.} The availability and support of programming languages in developing functions are yet limited on most serverless computing platforms. Only 9 popular programming languages along with limited versions are supported on at least one platform, 4 out of which (PowerShell, Ruby, PHP, and TypeScript) are supported on only one platform. As a result, developers have to narrow their choice of languages to construct their functions. Except for Azure, the other three platforms can deploy functions with only uncompressed package size that is no more than 500 MB, and can not deploy Python function solely with the latest libraries such as TensorFlow 2.3. Additionally, different languages can result in obviously different cold start time. Languages with ``cumbersome'' runtime (e.g., Java) can account for much longer cold start time (7x) than other script languages (e.g., Python and Node.js).
\noindent $\bullet$ \textbf{Deployment.} When deploying the functions, we find that loading redundant libraries can introduce non-negligible cold start time as well as increase package size. In particular, the cold start time can increase 4.6x if the function loads some unused library. Therefore, developers need to trim their code (e.g., removing the useless code) and apply the lazy loading of libraries to alleviate the overhead of initializing functions. It is better for cloud vendors to provide related tools to help developers optimize their function packages before development instead of directly deploying the zipped function packages submitted by developers.
\noindent $\bullet$ \textbf{Runtime}. Allocating more memory within a certain range can obviously reduce the cold start time. For example, increasing the memory from 128 MB to 1,024 MB in AWS Lambda can save more than 80\% of the cold start time of a Hello-world function developed with Java. For functions with high memory requirements (e.g., \textit{sls-fib}), allocating more memory can also reduce the execution time dramatically. However, the rate of reduction will drop when allocated memory exceeds 1024 MB. The findings can help developers make a trade-off between the cost and performance by allocating memory adequately. We also find that different serverless computing platforms can have discrepant performance for a specific type of task. For example, Alibaba Cloud Function Compute performs the best for CPU-bound and memory-bound benchmarks, but other platforms can beat it with other benchmarks under certain conditions (e.g., Google Functions has higher random-IO throughput at 1,024 MB of memory).
\noindent $\bullet$ \textbf{Implications:} Overall, our findings can reveal the mystery of existing commodity serverless computing platforms, motivating future research, and development. Specifically, the findings show strong implications for both developers and cloud vendors of the serverless computing ecosystem.
\textbf{To Developers.} Our findings show the shortcomings of existing serverless computing platforms, such as limited programming languages, package size restriction, and so on. Developers should carefully choose the appropriate platform based on their preferred languages, application package size, and other features that may be restricted on some platforms. Our findings also show how different serverless computing platforms actually perform under various configurations. Developers should choose a relatively lightweight language (e.g., Python) to construct their functions and tune configurations (e.g., allocating enough memory) and programs (e.g., removing dead code) to achieve better performance. Developers should choose the appropriate platform to obtain better performance based on their resource needs of functions, or even directly get an intuitive result by running their functions with our automated benchmarking platform \textit{TBS}.
\textbf{To Cloud Vendors.} Our findings encourage continuous improvement and optimization for serverless computing platforms. Cloud vendors can provide some personalized guides of configuration based on developers' functions and provide more tools to alleviate the restrictions of some configurations, e.g., provide tools for removing dead code and reducing package size instead of directly deploy zipped packages submitted by developers.
The remainder of this paper is organized as follows. We present our research goal and methodology to conduct our study in section~\ref{sec:methodlogy}. We summarize and compare the key characteristics of different serverless computing platforms in section~\ref{sec:char}. We introduce the details of the benchmark platform \textit{TBS} in section~\ref{sec:testbed}. We describe the evaluation results among different serverless computing platforms in section~\ref{sec:evaluation}. We present discussion about our study in section~\ref{sec:discussion}, survey related work in section~\ref{sec:related_work}, and conclude the paper with future work in section~\ref{sec:conclusion}.
\section{Approach Overview}\label{sec:methodlogy}
This section illustrates the methodology that we adopt to evaluate serverless computing platforms, including Amazon Web Service Lambda (released in November 2014), Microsoft Azure Functions\footnote{Azure Functions offers three different hosting plans~\cite{azurenew}, we only focus on the consumption plan that is the most similar to other serverless computing platforms.} (released in November 2016), Google Cloud Functions (released in July 2018), and Alibaba Cloud Function Compute (released in April 2017). We first present the research goal of our study. Then, driven by this goal, a workflow of our research approach is presented.
\subsection{Research Goal}
To fill the gap when developers select the most adequate serverless computing platform. We compare these serverless computing platforms with both qualitative analysis and quantitative analysis. Qualitative analysis refers to the comparison of different characteristics provided by these serverless computing platforms, whereas quantitative analysis is the comparison of the actual performance of running workloads on different serverless computing platforms.
First, we qualitatively summarize a taxonomy of characteristics on serverless computing platforms from three aspects, i.e., development, deployment, and runtime. Such a taxonomy allows developers to have an intuitive understanding to judge whether developers can develop their functionalities on these serverless computing platforms. Then, we design an open-source benchmark platform to quantitatively explore the actual system performance (i.e., startup latency, and resource efficiency) of serverless computing platforms. This platform can help developers choose the appropriate platform and configurations to deploy functionalities in the optimal performance. We show the corresponding results in Section~\ref{sec:char} and Section~\ref{sec:evaluation}, respectively. A workflow overview of our approach is presented and showed as follows.
\subsection{Workflow Overview}
\begin{figure}[!thb]
\centering
\includegraphics[width=0.45\textwidth]{architecture.pdf}
\caption{A workflow overview of our approach}
\label{fig:workflow}
\end{figure}
In our study, we try to reveal the mystery of the commodity serverless computing platforms to help developers make choices among them based on both the qualitative analysis with official documentations and the quantitative analysis with practical performance evaluation. Figure~\ref{fig:workflow} shows the overall workflow of our approach.
The first step of our approach is to extract and summarize features provided by official documentations of different serverless computing platforms. Such features specify the inherent restrictions of different aspects, e.g., developing, deploying, and runtime.
In our study, we do not focus on features such as security, usability, service level agreement, etc. We double-check the value of each feature to ensure correctness.
In order to evaluate and compare the actual performance of different platforms. We design and implement a benchmarking platform named \textbf{\textit{TBS}} to automatically conduct the measurement analysis for commodity serverless computing platforms, and \textit{TBS} is shown in Figure~\ref{fig:workflow}. On one hand, \textit{TBS} can fully automatize the package, deployment, execution, log collection, clean-up, result generation for each measurement of developers on four serverless computing platforms. On the other hand, \textit{TBS} contains a benchmark suite named \textbf{\textit{SlsBench}}, which is used to explore and compare the actual performance from various aspects (e.g., cold start time, execution time, throughput, etc.) of these serverless computing platforms. With such a benchmark suite, it can help cloud providers design and evaluate their serverless computing platforms and provide useful hints for developers to architect their applications on serverless computing platforms. Specifically, we conduct different workloads on these serverless computing platforms, and these workloads are measured under varied experimental configurations. To make a fair comparison, our experiments are repeated several measurements. Based on experimental results, the key factors that may affect the performance of functions are determined.
When developers use such a \textit{TBS}, structured logs will be stored to further serve developers to extract the key data information, e.g., event timestamp, cold start time of functions, execution time of functions, output result of functions, etc. Based on this data information, the visualized results across different platforms can be shown in the Web-based frontend of \textit{TBS}. In our study, \textit{TBS} adopts a design principle of modularity and extensibility. This kind of principle allows developers to integrate other serverless computing platforms into \textit{TBS} without restraint. Consequently, we report a series of our findings based on our experimental results in our study, and give some insightful implications for both application developers and cloud vendors.
\section{Taxonomy of Characteristics}\label{sec:char}
Generally, developers need to know about characteristics related to the development, deployment, and runtime of serverless computing platforms in the process of application development. With regard to this, we construct the taxonomy of characteristics on serverless computing platforms and it is shown in Figure~\ref{fig:info}.
Such a taxonomy can help developers quickly understand the pros and cons of different serverless computing platforms, and decide whether an existing application or a new functionality should be implemented on serverless computing platforms. The specific information is illustrated as follows.
\begin{figure*}[!thb]
\centering
\includegraphics[width=0.9\textwidth]{sls_info_for_platforms.pdf}
\caption{Taxonomy of characteristics on serverless computing platforms.}
\label{fig:info}
\end{figure*}
\begin{figure*}[!thb]
\centering
\includegraphics[width=0.93\textwidth]{language.pdf}
\caption{Supported languages of four mainstream serverless providers ($\checkmark$: supported;$\times$: unsupported; $^\beta$: beta; $^{1,2,3}$: generation of Azure Functions; \textcircled{$\times$}: deprecated.)}
\label{fig:language}
\end{figure*}
\subsection{Development-related Characteristics}
When developers are developing applications, they should consider whether a certain serverless computing platform can implement their functionalities in terms of development language, trigger type, package size limit, etc. Thus, the detailed information about the development-related characteristics is explained as follows.
Generally, an application executed in serverless computing platforms consists of one or more serverless functions. Each serverless function is a small, stateless, event-driven, and pay-as-you-go unit dedicated to handling a specific task. Such a function is often constructed with a small piece of code written in different languages, e.g., Java, Python, NodeJS, etc. As shown in Figure~\ref{fig:language}, different serverless computing platforms support various \textbf{languages and versions}. We find that the three most popular languages (i.e., Python, Node.js, and Java~\cite{toplanguagenew}) are natively supported by all these four platforms. However, Python is not supported by Azure on Windows .Net Core. Furthermore, Google Cloud Functions is the sole platform that does not support Microsoft, and it has the fewest supported languages. From Figure~\ref{fig:language}, \emph{we also find that certain languages are only natively supported by only one platform}. For instance, PowerShell is only natively supported by Azure Functions, Ruby is only natively supported by AWS Lambda, and PHP is only natively supported by Alibaba Cloud Function Compute. Each platform supports various languages, and each language may have different versions. In our later experiments, we will choose the consistent language and version to compare the dynamic feature of the platform performance fairly. Last but not least, four platforms use the custom manner to allow developers to build their runtime and execute any preferred languages, whereas may bring extra non-negligible efforts and concerns of runtime stability. Thus, \textit{it is better to choose the appropriate platform to deploy an application based on developers' preferred languages.}
In the process of programming, an application or a function is triggered to execute by events~\cite{JonasCoRR2019} (e.g., HTTP requests, timer, storage conditions). Besides, functions with different \textbf{event triggers} may have different main methods with given parameters, which will be invoked by serverless computing platforms to deal with specific incoming events.
In order to reduce the cold start time of serverless functions, serverless computing platforms often consider the \textbf{package size limit} of deployment. At the time of writing of our study, the package size limit of AWS Lambda is 50 MB with a compressed format and 250 MB with an uncompressed format. For Alibaba Cloud Function Computing, it allows from 50 MB with a compressed format to 500 MB with an uncompressed format, whereas Google Cloud Functions lifts the restrictions to 100 MB with a compressed format and 500 MB with an uncompressed format. Differently, Azure Functions supports the higher deployment package as much as several GB in size. Unfortunately, the package size of applications, especially, deep-learning-based tasks with huge libraries, is big enough to exceed the limit of certain platforms. For example, the compressed size of the newest ``Tensorflow'' library (version 2.3.1-2~\cite{tensorflownew}) is 106.4 MB, and the uncompressed installed size is as high as 661.8 MB. It is impossible to deploy such an application with the ``big'' library to serverless computing platforms directly, thus the development enthusiasm of developers may be dampened. In this situation, \textit{developers can apply existing technologies and tools~\cite{RomanoTSE20} to remove the useless code and compress them. Cloud vendors can also provide some tools to automatically reduce the package size to ease the burden of developers.}
\subsection{Deployment-related Characteristics}
After developing an application or a function, developers need to deploy them on appropriate serverless computing platforms. In this situation, which deployment method or tool can be adopted, and different platforms are how to allocate memory for a function instance. This part of the information is worth thinking about to avoid deployment failure and poor performance. Next, we will introduce this part of the information.
As shown in Figure~\ref{fig:info}, we summarize the three most popular \textbf{deployment methods} (i.e., source code, docker container, and external services) of functions on serverless computing platforms. Four serverless computing platforms all support the deployment method with the source code, e.g., deployment with a zipped package that contains function code and dependent third-party libraries. Besides, we find that only Azure Functions (with consumption plan) does not allow developers to deploy functions by building their custom runtime (e.g., building custom Docker images), and other platforms can. Additionally, Azure Functions and Google Cloud Functions both support the deployment with other external services, e.g., source control services (Git), FTP services, whereas AWS Lambda and Alibaba Cloud Function Compute do not. When deploying functions on serverless computing platforms, developers can use \textbf{deployment tools}, e.g., command-line tool, console editor, and API $\&$ SDK. Besides, developers can even deploy functions from specific development tools, e.g., visual studio code for Azure Functions.
In the deployment process, some platforms (e.g., AWS Lambda, and Alibaba Cloud Function Compute) need to specify the memory of functions in advance. For the \textbf{memory capability} of different serverless computing, we find that AWS Lambda allocates memory to a function instance from 128 MB to 3,008 MB in steps of 64 MB, Alibaba Cloud Function Compute is from 128 MB to 3,072 MB in steps of 64 MB, and Google Cloud Functions is from 128 MB to 2,048 MB with assigned values. For Azure Functions, developers cannot specify the memory in the development or deployment, but they can use 1,536 MB at most. Generally, memory and CPU are closely related. The \textbf{CPU capability} of a function instance increases proportionally with allocated memory. Specifically, AWS Lambda and Alibaba Cloud Function Compute get 1 vCPU at 1,792 MB and 1,024 MB, respectively. Regrettably, we find that \textbf{GPU support} for the function instance is not available in these serverless computing platforms. An alternative way is to provide GPU power as serverless services~\cite{Jun18PDPnew, alibabaGPUnew}.
Each serverless computing platform allows a function to run within the \textbf{timeout} limits as shown in Figure~\ref{fig:info}. Thus, \textit{it is not recommended to run long-time tasks with serverless functions}. Specifically, the timeout limit of Google Cloud Functions is only 9 minutes, whereas both Azure Functions and Alibaba Cloud Function Compute allow functions to run within 10 minutes. For AWS Lambda, its timeout specifies the maximum timeout duration as long as 15 minutes. Serverless computing platforms have a \textbf{local disk} to restrict the local storage (roughly 0.5 GB by default). However, Alibaba Cloud Function Compute enables developers to extend the local storage capacity with network-attached storage (NAS).
In addition, we collect the detailed information of the \textbf{supported regions} through deploying functions on each serverless computing platform factually. At the time of writing, AWS Lambda spans 20 geographical regions, whereas Google Cloud Functions is available in 19 regions. For Alibaba Cloud Function Compute, it offers 15 regions to deploy functions, whereas Azure Functions actually supports the deployment of functions in 32 regions. However, \textit{we find an inconsistency compared to its official documentation\footnote{Products available by region. \url{https://azure.microsoft.com/en-us/global-infrastructure/services/?products=functions®ions=all}},} which mentions that 43 regions overall are available. According to supported regions, \textit{developers can choose the appropriate region near to them to speed up their function executions.}
For the \textbf{runtime OS}, most serverless computing platforms run functions on Linux hosts, and only Azure Functions supports both Windows hosts and Linux hosts. As shown in Figure~\ref{fig:language}, Windows hosts and Linux hosts do not support the same languages set. \textit{We find that certain inconsistencies between the official documentation and the actual execution process on our experiments.} For example, AWS Lambda claims usage of \textit{Amazon Linux} as the operating system for Python 3.7 runtime\footnote{https://docs.aws.amazon.com/lambda/latest/dg/lambda-runtimes.html}. However, we find that it actually employs the \textit{Amazon Linux 2} as the operating system when we deploy a function with Python 3.7 runtime. Thus, we infer the reason for this inconsistency may be due to the belated update of AWS Lambda documentation, and this kind of inconsistency makes developers confused.
\subsection{Runtime-related Characteristics}
Applications and functions that have been successfully deployed on the serverless computing platforms are executed by invocations. Four platforms all support synchronous and asynchronous \textbf{invocations}. In addition, developers often need to consider the request \textbf{payload} when calling, but the payload limit for different serverless computing platforms is shown in Figure~\ref{fig:info}.
Additionally, serverless computing platforms can scale-out by creating multiple instances to run the same function currently for burst invocations. An instance often runs on a separate host \textbf{environment}, i.e., a lightweight VM (e.g., Firecraker~\cite{AgacheNSDI2020} used by AWS Lambda, a dedicated container (e.g., Docker~\cite{alibabanew} used by Alibaba Cloud Function Compute and Azure Functions), or a gVisor~\cite{gvisornew} used by Google Cloud Functions). Furthermore, all the resources (i.e., CPU, memory, storage) of this host are dedicated solely to its instance. For the \textbf{instance currency} of AWS Lambda and Google Cloud Functions, each function execution maps to one exclusive instance, whereas each instance of Azure Functions and Alibaba Cloud Function Compute can run multiple executions (i.e., start multiple processes) at the same time through sharing the resources. Depending on the deployment region, the \textbf{instance limit} supported by AWS Lambda is from 500 up to 3,000 per single function. Google Cloud Functions only allows up to 3,000 invocations to be executed concurrently, because it does not have a maximum number of allocated instances per single HTTP-triggered function. For Azure Functions and Alibaba Cloud Function Computing, their instance limits are 200 and 100, respectively.
Serverless computing platforms allow developers to use normal language and operating system features, such as creating additional threads and processes. Resources allocated to serverless functions, including memory, execution time, disk, and network use, must be shared among all the \textbf{processes/threads} it uses. The specific processes/threads limit is shown in Figure~\ref{fig:info}. In Linux, its kernel use \textbf{file descriptors} to efficiently manage the opened files and create indexes for them. The limit of file descriptors is also shown in Figure~\ref{fig:info} for four serverless computing platforms.
Generally, for the \textbf{billing model}, the price generally rises with the increasing of the allocated memory. Indeed, the overall price of executing a function varies depending on the specific function memory (in AWS, Google, and Alibaba) or the actual consumed memory during invocations (in Azure).
\section{The benchmarking platform}\label{sec:testbed}
In order to facilitate the benchmarking on different serverless computing platforms, we have designed and implemented an open-source benchmarking platform named \textbf{\textit{TBS}}, which contains a benchmark suite named \textbf{\textit{SlsBench}}.
\subsection{Platform Design and Implementation}
\textit{TBS} is an open and modular platform that supports extensible benchmarking on any serverless computing platform through a set of well-defined interfaces. \textit{TBS} abstracts necessary components (i.e., packaging component, deploying component, testing component, logging component, cleaning-up component, and results generating component) to benchmark the target serverless computing platform. If developers want to integrate a new serverless computing platform into \textit{TBS}, they just need to provide the platform-specific implementation with the official APIs and SDK, then they can automatically measure the new serverless computing platform.
The specific illustration for components of \textit{TBS} is explained as follows. The packaging component automatically packages a benchmark (discussed in Section~\ref{sec:benchmark}) as a zipped file or a container image for further deployment. The deploying component can specify configurations (shown in Figure~\ref{fig:info}) and deploy functions to the target platform. The testing component is to execute deployed functions based on configured trigger types, e.g., sending an HTTP request. The logging component is used to retrieve and store the execution logs of each measurement. Although different serverless computing platforms have different formats of logs, they often contain basic information (e.g., timestamp, execution time of functions, output result of functions, etc.) about the life cycle of a function. For each benchmark in our study, we save the important contents, as well as the output result in logs. The result generating component can extract valuable information from structured logs, and generate a comparable result about different serverless computing platforms.
\subsection{Benchmarks}\label{sec:benchmark}
\begin{figure}[!htb]
\centering
\includegraphics[width=0.5\textwidth]{benchmark.pdf}
\caption{Benchmarks}
\label{fig:benchmark}
\end{figure}
\textit{TBS} leverages a set of workloads as shown in Figure~\ref{fig:benchmark}, including microbenchmarks and macrobenchmarks, to evaluate the performance aspect of serverless computing platforms. With such benchmarks, developers can have a better understanding of the actual system performance of serverless computing platforms. Specifically, (i) microbenchmarks consist of a set of simple workloads targeting specific system aspects, i.e., CPU-bound workloads, memory-bound workloads, diskIO-bound workloads, and network-bound workloads. (ii) Macrobenchmarks focus on the real-world workloads (i.e., multimedia data process, MapReduce, machine-learning-based serving, etc), and they go far beyond microbenchmarks. Each benchmark is implemented in multiple programming languages (Python by default) to support the variety of the available runtime systems. Although there are subtle differences between code variants to adapt to the peculiarities of each cloud platform, the code of each programming language is basically the same for all. By default, our functions are implemented using the HTTP trigger, which is the most common event trigger and is supported by all serverless computing platforms.
\subsubsection{Microbenchmarks}
Microbenchmarks consist of a set of workloads exploring different system aspects, especially underlying hardware resources (such as CPU, memory, storage, and network). We will briefly introduce each microbenchmark.
\textbf{CPU-bound benchmarks}. Such benchmarks are mainly used to measure the CPU-bound performance. \textit{sls-matrixMul} in Figure~\ref{fig:benchmark} calculates the result of the multiplication of two N-dimensional square matrices, and \textit{sls-linpack} solves the linear equation.
\textbf{Memory-bound benchmarks}. Such benchmarks are mainly used to measure the memory-bound performance. We built a workload to calculate Fibonacci~\cite{fibonaccinew} values recursively, and this workload will consume a lot of memory.
\textbf{DiskIO-bound benchmarks}. Such benchmarks are used to mainly measure the IO-bound performance. We design three kinds of workloads, i.e., \textit{sls-dd}, \textit{sls-randomIO}, and \textit{sls-sequentialIO}. Specifically, \textit{sls-dd} is a workload that creates files in ``/tmp/'' directory of the local disk through using \textit{dd} command of the Linux system. \textit{sls-randomIO} and \textit{sls-sequentialIO} are workloads that measure the throughput and latency of random IO and sequential IO, respectively.
\textbf{Network-bound benchmarks}. Such benchmarks are mainly used to measure the network-bound performance. We design three kinds of workloads. (i) \textit{sls-http} is a network-bound test, which will be returned immediately after the invocations with a smaller payload in JSON-format. \textit{sls-http} can be used to verify the round-trip time of a geographically distributed deployment. (ii) \textit{sls-iPerf} is designed to actively measure the maximum bandwidth achievable on the IP network. (iii) \textit{sls-cloudstorage} is designed to measure throughput and latency between function instances and cloud storage. For example, download an object from an input bucket of the cloud storage and upload the object to an output bucket.
\subsubsection{Macrobenchmarks}
In order to evaluate the performance of real-world applications, we design benchmarks that can complete complex tasks with actual functionality. The specific illustration is as follows.
\textbf{Multimedia data process}. In this kind of application, \textit{sls-image} and \textit{sls-video} workloads are designed. (i) \textit{sls-image} is an image processing workload, which performs image transformation tasks using Python Pillow library~\cite{Pillow}. \textit{sls-image} fetches an input image from the cloud storage, and applies ten different effects (e.g., copy, rotation, cropping, etc.) to this image. The corresponding output is uploaded back to the cloud storage. (ii) \textit{sls-video} applies the grayscale effect from the OpenCV library~\cite{OpenCV} to the video input, and uploads the converted video to the cloud storage.
\textbf{MapReduce}. MapReduce is a popular programming model that allows developers to process or generate large-scale data in parallel. In our study, we add the \textit{sls-mapreduce} workload, which consists of two types of functions. One is the \textit{Map} function to implement filtering and sorting functionalities, and the other is the \textit{Reduce} function to merge and organize the outputs of the \textit{Map} function.
\textbf{ML Training \& Serving}. We also provide workloads that deal with machine learning (ML) tasks using serverless computing. Such workloads mainly involve with the ML model training and ML model serving. Generally, the raw input data of a machine learning task needs pre-processing to prepare the input of training. In the \textit{sls-lr-training} workload, we use the text dataset about \textit{Amazon Fine Food Review}\footnote{\url{https://snap.stanford.edu/data/web-FineFoods.html}} saved in the cloud storage, and input them into the regression model.
Because the \textit{sls-lr-training} workload needs to access large-size datasets from the cloud storage, it often takes up a lot of CPU, memory, and network. After training a model, this model needs to be served for arbitrary inputs to make predictions. In the \textit{sls-lr-serving} workload, we utilize the model built in the ML model training phase, and input users’ review texts into this model to predict the corresponding sentiment score. To further explore the inference related to deep learning models, we add the image classification workload named \textit{sls-cnn} and words generation workload named \textit{sls-rnn}. \textit{sls-cnn} uses a SqueezeNet model~\cite{SqueezeNet} to achieve an impressive accuracy on an ImageNet~\cite{ImageNet} with 50x fewer parameters than the state-of-the-art model. In our experiments, \textit{sls-cnn} is implemented with Python Tensorflow Keras~\cite{TensorflowKeras}. Due to the limited memory size, attempts to import other convolutional neural network models (CNN~\cite{CNN}) are proved to be a failure. \textit{sls-rnn} uses a recurrent neural network model (RNN~\cite{RNN}) to implement words generation through PyTorch~\cite{PyTorch}.
\section{Performance Evaluation}\label{sec:evaluation}
Based on \textit{TBS} and \textit{SlsBench}, we can automatically evaluate four mainstream serverless computing platforms. Since executing serverless functions with low latency is critical for user experience\cite{HellersteinCIDR19}\cite{DuASPLOS2020}, We mainly focus on the following metrics (i.e., startup latency and resource efficiency) that are closely related with user-perceived latency in this paper. By default, we run each round test for 20 executions and use the median value to represent related evaluation results.
\textbf{Startup latency}. Functions are typically small and executed in seconds or even milliseconds, thereby the startup latency can be a considerable overhead. We mainly focus on the startup latency when cold start happens in this paper. Reducing cold start time is a key challenge in serverless computing~\cite{OakesATC18}\cite{JonasCoRR2019}\cite{DuASPLOS2020}. We will investigate how languages, package size, and memory affect the cold start time of function instance.
\textbf{Resource efficiency}. Different kinds of workloads have various demands for resources. For example, computation-intensive workloads can consume more CPU resources. We will investigate the resource efficiency by executing different workloads on different platforms. In our study, we focus on metrics like execution time of function, throughput of storage and CPU.
Note that we do not show all results in this paper due to space limit. Detailed results are available in our GitHub repository.
\subsection{Startup Latency}
The cold start may involve downloading a package, launching a new container, setting up the runtime environment, and initializing function, and it takes more time to handle a request than reusing an existing function instance (\emph{warm} start). Thus, the cold start can significantly affect application responsiveness and in turn affect the user experience.
For each platform, we create functions with the same workload and configuration, and sequentially invoke them twice to derive the cold start time. We use the difference of overall response time (end-to-end duration from the client perspective) as an estimation of its cold start latency.
\subsubsection{How will program languages affect the cold start?}
\begin{figure}[!htb]
\centering
\includegraphics[width=0.4\textwidth]{lan-python.pdf}
\caption{The distribution of the cold start time with Python on different platforms.}
\label{fig:lan-python}
\end{figure}
\begin{figure}[!htb]
\centering
\includegraphics[width=0.4\textwidth]{lan-Nodejs.pdf}
\caption{The distribution of the cold start time with Node.js on different platforms.}
\label{fig:lan-nodejs}
\end{figure}
\begin{figure}[!htb]
\centering
\includegraphics[width=0.4\textwidth]{lan-Java.pdf}
\caption{The distribution of the cold start time with Java on different platforms.}
\label{fig:lan-java}
\end{figure}
As shown in Figure~\ref{fig:language}, languages are not well supported across all platforms, thus we just compare the cold start time of three most well-supported languages~\cite{toplanguagenew}, i.e., Java, Node.js, and Python.
Figure~\ref{fig:lan-python}, Figure~\ref{fig:lan-nodejs}, and Figure~\ref{fig:lan-java} show the distribution of the cold start time for three different languages, respectively. The most obvious trend is that statically typed languages (e.g., Java) have over 3 times higher cold start time than dynamically typed language (e.g., Python and Node.js), especially for the cases with small memory. AWS Lambda is overall the fastest with allocated memory less than 1024 MB, with a median cold start latency of only 228.37 ms for Python, 139.96 ms for Node.js, and 374.79 ms for Java at 512 MB memory. Executing a Java function needs to launch and initiate a ``cumbersome'' JVM, and will incur much more overhead and result in much more cold start time. Launching a Java function on AWS Lambda has the least cold start time than other platforms, and can even speed up by a factor of about 7 compared to Google Functions when allocating 128 MB memory. Interestingly, we also find that the Google Functions can perform better than other platforms to run Java functions with allocated memory more than 1024 MB, and the cold start time can even be less than functions developed by other languages.
\textbf{Implication}: Without regard to the preference, developers should choose those ``light-weight'' languages (e.g., Python, Node.js) to implement their functions with lower cold start time, especially for those time-sensitive tasks. When developing functions with huge memory (e.g., more than 1024 MB), developers can be free to choose any languages with low cold start latency. For cloud vendors, they should further optimize the initialization of cumbersome runtime of languages like Java. For example, cloud vendors can apply new technologies like Unikernel to reduce the overhead of initialization~\cite{JonasCoRR2019}.
\subsubsection{How will memory affect the cold start?}
Since we can not specify the memory of Azure functions (1,536 MB), we can assume its performance to be in the range from 1024 MB to 2048 MB. As shown in Figure~\ref{fig:info}, developers can specify only different memory when deploying their functions, and the serverless computing platform will allocate CPU capability proportional to allocated memory. Therefore, allocating more memory to executing functions will also increase the computing power of function instance.
As shown in Figure~\ref{fig:lan-python}, Figure~\ref{fig:lan-nodejs}, and Figure~\ref{fig:lan-java}, we can find that the more memory allocated to function means the lower cold start time for most cases for Google Functions, on which the memory size improves the cold start time in a roughly linear fashion. However, we also find that the AWS Lambda and Alibaba Cloud Compute do not seem to have significant differences with different allocated memories for Python and Node.js. This probably means that not so much memory and CPU is required to start the Python or Node.js application. We are also surprised to find that 98.5\% of the cold start time can be reduced for Java on Google Cloud Functions when increasing the memory from 128 MB to 2048 MB.
\textbf{Implication}: For functions written with Python or Node.js on AWS Lambda and Alibaba Cloud Function Compute, allocating more memory will not speed up the start-up and only lead to more cost. Developers should not increase memory allocation if the task is not memory-intensive. In other cases, developers can properly add certain memory to reduce the cold start time. For cloud vendors, they should give some practical guides of memory configuration to make a balance between the cost and performance.
\subsubsection{How will package size affect the cold start?}
\begin{figure}[!htb]
\centering
\includegraphics[width=0.4\textwidth]{packagesize.pdf}
\caption{The distribution of cold start time of different package size.}
\label{fig:clod_start_package}
\end{figure}
Figure~\ref{fig:clod_start_package} compares four Python functions with the various number of referenced third-party packages. Functions are executed using 128 MB of memory. The first bar presents the cold start time of a basal function (504.6 KB). The following three bars show the cold start time of functions constructed by solely adding different third-parity library into the basal function without loading (2.8 MB with a library named \textit{Pillow}, 21.9 MB with a library named \textit{Numpy}, and 48.6 MB with a library named \textit{OpenCV-python}, respectively). The last three bars show the distribution of cold start time of functions with loading the useless library (the dependency libraries are redundantly imported in the package).
Interestingly, we find that increasing the size of the deployment package does not obviously affect the cold start time on AWS Lambda and Azure Functions. Instead, loading more libraries in the function will dramatically increase the cold start time except on the Google Cloud Functions, since it will increase the time of initializing functions (e.g., loading code into memory). In particular, the cold start time can even increase 4.6x if the function loads some unused library (21.9 MB-with-import VS 21.9 MB-no-import).
\textbf{Implications}: Loading redundant libraries may introduce non-negligible cold start time, so developers need to trim their code (e.g., removing the useless code) and apply lazy loading of libraries to alleviate the overhead of initializing functions. It is better for cloud vendors to provide related tools to help developers optimize their function packages before development instead of directly deploying the zipped function packages submitted by developers.
\subsection{Resource Efficiency}
We leverage our benchmark tasks (i.e., microbenchmarks and mac-robenchmarks) to explore the resource efficiency.
\subsubsection{CPU-bound microbenchmarks}
\begin{figure}[!htb]
\centering
\includegraphics[width=0.4\textwidth]{linpack.pdf}
\caption{The distribution of the CPU throughput across different platforms}
\label{fig:linpack}
\end{figure}
We use the \textit{sls-linpack} benchmark to evaluate the CPU capacity of different platforms. \textit{sls-linpack} operates on 1,000$\times$1,000 matrix, and will output a performance rating metric in terms of Millions of Floating Point Operations Per Second (MFLOPS). The experimental results are shown in Figure~\ref{fig:linpack}. We find that Alibaba Cloud Function Compute performs much better than other platforms, and AWS Lambda performs worst than other platforms. For AWS Lambda and Google Cloud Functions, the \textit{mflops} can increase linearly by allocating more memory, which conforms to the finding as we summarized in Figure~\ref{fig:info} that serverless computing platform will allocate CPU capability proportional with allocated memory.
\textbf{Implication}: For computation-intensive workloads, developers are recommended to use the Alibaba Cloud Function Compute. For other platforms, developers can obtain more computing power by allocating more memory.
\subsubsection{Memory-bound microbenchmarks}
\begin{figure}[!htb]
\centering
\includegraphics[width=0.4\textwidth]{memory.pdf}
\caption{The distribution of the execution latency for the memory-bound workload}
\label{fig:memory_bench}
\end{figure}
To measure the performance of memory-bound workloads, we use the \textit{sls-fib} to calculate the 25th Fibonacci value recursively, which will result in high memory consumption. The experimental results are shown in Figure~\ref{fig:memory_bench}. Alibaba Cloud Function Computing have the shortest execution time among all cases, and are insensitive to more memory. For other platforms, allocating more memory can linearly reduce the execution time. However, the rate of reduction will decrease when allocated memory increased from 1,024 MB to 2,048 MB.
\textbf{Implication}: For those memory-bound workloads, developers are recommended to deploy functions on Alibaba Cloud Function Compute to alleviate the burden of limited memory. For other platforms, developers can allocate more memory under a certain limit (1,024 MB) to achieve cost-effective a trade-off between performance and cost.
\subsubsection{DiskIO-bound microbenchmarks}
\begin{figure}[!htb]
\centering
\includegraphics[width=0.5\textwidth]{ramdomIO-read.pdf}
\caption{The random read throughput of different platforms}
\label{fig:ramdomIO-read}
\end{figure}
\begin{figure}[!htb]
\centering
\includegraphics[width=0.5\textwidth]{ramdomIO-write.pdf}
\caption{The random write throughput of different platforms}
\label{fig:ramdomIO-write}
\end{figure}
\begin{figure}[!htb]
\centering
\includegraphics[width=0.5\textwidth]{sequentialIO-read.pdf}
\caption{The sequential read throughput of different platforms}
\label{fig:sequentialIO-read}
\end{figure}
\begin{figure}[!htb]
\centering
\includegraphics[width=0.5\textwidth]{sequentialIO-write.pdf}
\caption{The sequential write throughput of different platforms}
\label{fig:sequentialIO-write}
\end{figure}
We leverage \textit{sls-randomIO} and \textit{sls-sequentialIO} to measure the throughput of local storage in terms of different IO patterns, i.e., read and write. The experimental results are shown in Figures~\ref{fig:ramdomIO-read},~\ref{fig:ramdomIO-write},~\ref{fig:sequentialIO-read} and~\ref{fig:sequentialIO-write}. Alibaba Cloud Function Compute has stable performance with different memory, and has outstanding sequential throughput bandwidth than other platforms. For example, Alibaba Cloud Function Compute has 32.96 times the read throughput of Google Cloud Functions at 128 MB of memory. Interestingly, Google Cloud Functions can perform better than Alibaba Cloud Function Compute at when allocated memory is more than 1024 MB.
For other platforms, the IO throughput increases by allocating memory, but the rate of increase will drop when exceeding a threshold. For example, sequential IO throughput of Google Cloud Functions trends to be stable when allocating memory more that 512 MB. Azure’s memory configuration of 1,536MB implies its performance should be set to the range from 1,024 MB to 2,048 MB. Surprisingly, we find that the throughput of Azure is lower than 1,024 MB instances of other platforms.
\textbf{Implication}: If developers are deploying functions with frequent random IO operations, it is better to deploy such functions on Alibaba Cloud Function Compute, especially for the cases with small allocated memory.
\subsubsection{Network-bound microbenchmarks}
Since the network latency heavily depends on network condition and geographical location, we mainly focus on the throughput measurement with \textit{sls-iPerf} in this paper. \textit{sls-iPerf} use \textit{iPerf3}~\cite{iPerf3} with default configurations to run the throughput test for 30 seconds with different same-region iPerf servers, so that iPerf server-side bandwidth was not a bottleneck~\cite{WangATC2018}. Due to the changeable network conditions, we cannot get stable results and just show some key findings in this part. Network throughput has a trend to increase as function memory increases. Actually, we have also measured the latency with the \textit{sls-http} benchmark deployed in different regions, and the latency indeed varies across different regions.
\textbf{Implication}: Developers can deploy their functions on the regions that are close to function users, which can obviously reduce the overall response time.
\subsubsection{Macrobenchmarks Performance}
Different from microbenchmarks that exclusively evaluate different resources, the macrobenchmarks utilize CPU, memory, disk IO, and network resources together at different degrees. We present two representative workloads here.
\begin{figure}[!htb]
\centering
\includegraphics[width=0.5\textwidth]{video_processing.pdf}
\caption{The distribution of the execution latency of \textit{sls-video} on different platforms}
\label{fig:video}
\end{figure}
\begin{figure}[!htb]
\centering
\includegraphics[width=0.5\textwidth]{ML_Serving.pdf}
\caption{The distribution of the execution latency of \textit{sls-lr-serving} on different platforms}
\label{fig:serving}
\end{figure}
Figures~\ref{fig:video} and~\ref{fig:serving} shows the execution latency to complete video processing task (\textit{sls-video}) and model serving task (\textit{sls-lr-serving}) on each platform, respectively. In this experiment, we also vary the allocated memory size of the function to study its impacts on the execution latency of such macrobenchmarks. Missing bars in the figures indicate that the platform could not complete the given workload with the allocated memory size for platforms except Azure Functions. Besides, Google Functions fail to serve the \textit{sls-lr-serving} task with high memory demanding.
Alibaba Cloud Function Compute shows better performance than other platforms overall with the same allocated memory, which confirms our previous findings of evaluations with microbenchmarks. In other words, Alibaba Cloud Function Compute performs well in most cases with microbenchmarks, thereby also has better performance for complicated workloads. Google Functions performs worst with a low memory allocated (with memory no more than 512 MB). Although Azure Functions allocates 1,536MB of memory, we find that the execution latency on Azure Functions is close to 1,024 MB instances of other platforms. For each platform, allocating more memory can linearly reduce the execution time of functions. However, the ratio of improvement varies for different platforms.
We need to address that developers can also automatically and conveniently obtain the intuitive comparison result of other workloads with \textit{TBS}, so that they can make a quick decision to deploy their functions on the target platform to achieve optimal performance.
\textbf{Implication}: For those complicated workloads like macrobenchmarks in this paper, allocating high memory can narrow the performance gap among different platforms.
\section{Discussion}\label{sec:discussion}
In this section, we describe some issues that may potentially affect the generalization of our results.
\textbf{Other features of serverless computing platform}. In this paper, we do not investigate every aspect of serverless computing, such as permissions, security~\cite{DattaWWW20}, monitoring, and so on. We mainly focus on performance-related features and give practical implications for both developers and cloud vendors.
\textbf{Fairness of evaluation}. In order to make a fair comparison, we deploy all functions in the same region (i.e., us-east in this paper) on all tested platforms. However, serverless computing platforms may have different performance in different regions. With \textit{TBS}, others can conveniently repeat any experiment in other regions. We will study how the deploying regions can affect the performance of functions.
\textbf{Rapid evolution of serverless computing platforms}. The results and findings might change over time. These commodity serverless computing platforms are keeping adding new features and improving their performance, and later benchmarks on these platforms may conflict with our current result. Fortunately, anyone can leverage \textit{TBS} to benchmark target platforms continuously to keep tracing the up-to-date characteristics.
\textbf{Breakdown of cold start time}. The cold start may involve many processes, including downloading a package, launching a new container, setting up the runtime environment, and loading function. Although we have investigated many factors that will affect the overall cold start time, we fail to further locate the mainly affected process due to the coarse-grained logs provided by these commodity serverless computing platforms. We plan to conduct a deeper analysis based on some open-source serverless platforms~\cite{openWhisknew}\cite{fissionnew}.
\section{Related Work}\label{sec:related_work}
In this section, we analyze the related work about serverless computing and its benchmark suites. In addition, we introduce the other work about cloud computing and web applications.
Shahrad \textit{et al.}~\cite{ShahradATC20} characterized the entire production FaaS workload of Azure Functions, including trigger types, invocation frequencies and patterns, and resource needs. Spillner \textit{et al.}~\cite{SpillnerCoRR2019} conducted a quantitative study about how developers use FaaS offerings, and gained insights into how functions are implemented and developed. However, these studies cannot offer insights into the running performance seen by a developer. Some previous work~\cite{back2018using, figiela2018performance, lee2018evaluation, WangATC2018} evaluated and compared how serverless applications behave on different platforms. Such work focused more on the workloads and the performance differences between platforms instead of analyzing the underlying implications on serverless computing. Mohanty \textit{et al.}~\cite{MohantyCloudCom18}, Palade \textit{et al.}~\cite{PaladeService2019}, and Li \textit{et al.}~\cite{LiWOSC2019} focused on the performance of serverless computing with several popular open-source serverless platforms. Differently, we mainly focus on current mainstream commodity serverless computing platforms that have standard guidelines and mature practices.
Yu \textit{et al.}~\cite{yu2020characterizing} proposed an open-source benchmark suite named ServerlessBench to characterize serverless computing platforms leveraging customized test cases. Maissen \textit{et al.}~\cite{MaissenDEBSS2020} and Kim \textit{et al.}~\cite{KimLCLOUD19} also designed two benchmark suites, named FaaSDOM and FunctionBench, respectively, to facilitate the performance testing of serverless computing platforms. They both provided microbenchmarks, and FunctionBench provided some more complicated benchmarks that represent real-world applications to evaluate the performance. In our study, we can provide a comprehensive benchmark suite including both microbenchmarks and macrobenchmarks, and enable developers or end-users to gain insights on the running performance of each platform automatically in \textit{TBS}.
Serverless computing is a new paradigm of cloud computing, which is a great step in the history of computer and data governance~\cite{add7}. In general, computation offloading for applications can be accomplished on cloud~\cite{add3,add14}, network edge and client side (mobile device~\cite{add12,add11}) using different hardware, including CPU, GPU. Particularly, edge computing is an emerging and promising technology dedicated for improving use experience of today's interactive applications~\cite{add1}. In this situation, service-oriented situational applications have shown great potential in solving immediate and quick roll-out problems~\cite{add4}. The concept and technique~\cite{add6,add5} of service computing is becoming more and more mature. Thus, Function-as-a-Service becomes a popular trend. However, complex resource management need to rethink about serverless computing. Previous mobile web work has been rethinking resource management~\cite{add2,add9,add10} and performance~\cite{add13,add8}. Currently, major cloud providers present the correspond strategies, which uniformly manage resources to ensure scalability and load balancing. The detailed infrastructure is abstracted and provide useful interface.
\section{Conclusion}\label{sec:conclusion}
In this paper, we have characterized four mainstream commodity serverless computing platforms (i.e., AWS Lambda, Azure Functions, Google Cloud Functions, and Alibaba Cloud Function Compute) via both qualitative analysis and quantitative analysis. In qualitative analysis, we construct a taxonomy of 20 characteristics of serverless computing platforms from three aspects, i.e., development, deployment, and runtime. In quantitative analysis, we design and implement an open-source benchmark platform \textit{TBS} containing a benchmark suite named \textit{SlsBench}.
In \textit{SlsBench}, we construct a set of workloads to evaluate different aspects (e.g., startup latency and resource efficiency) of these serverless computing platforms. In future work, we plan to extend our benchmark platform to cover more various tasks, and maintain \textit{TBS} continuously to facilitate further applications and research on serverless computing for the whole serverless computing ecosystem.
\balance
\bibliographystyle{ACM-Reference-Format}
|
\section{Conclusion}
\vspace{-0.1in}
We have presented {\textsc{DiTucker}}\xspace, a deep multi-fidelity active learning approach for high-dimensional outputs. Our deep neural network based multi-fidelity model is flexibly enough to capture the strong, complex relationships between the outputs and between the fidelities. We proposed a mutual information based acquisition function that accounts for multi-fidelity queries. To calculate and optimize the acquisition function, we developed an efficient and reliable method that successfully overcomes the computational challenges due to the massive outputs.
\section{Experiment}
\subsection{Solving Partial Differential Equations}
\begin{figure*}
\centering
\setlength\tabcolsep{0pt}
\begin{tabular}[c]{ccc}
\setcounter{subfigure}{0}
\begin{subfigure}[t]{0.33\textwidth}
\centering
\includegraphics[width=\textwidth]{./figs/rmse_Burgers.pdf}
\caption{Burgers' equation}
\end{subfigure} &
\begin{subfigure}[t]{0.33\textwidth}
\centering
\includegraphics[width=\textwidth]{./figs/rmse_Heat2.pdf}
\caption{Heat equation}
\end{subfigure}
&
\begin{subfigure}[t]{0.33\textwidth}
\centering
\includegraphics[width=\textwidth]{./figs/rmse_Poisson2.pdf}
\caption{Poisson's equation}
\end{subfigure}
\end{tabular}
\caption{\small Normalized root-mean-square error (nRMSE) for active learning of PDE solution fields with two-fidelity queries. The normalizer is the mean of the test outputs. The results are averaged from five runs. The shaded regions indicate the standard deviations.} \label{fig:solving-pde-2fid}
\end{figure*}
We first evaluated {\textsc{DiTucker}}\xspace in standard computational physics tasks. Specifically, we used {\textsc{DiTucker}}\xspace to predict the solution fields of three commonly used partial differential equations (PDEs): \textit{Burgers}'s , \textit{Poisson}'s and \textit{Heat} equations~\citep{olsen2011numerical}. The training examples are collected by running a numerical solver with different meshes. The more the nodes/steps to create the mesh, the higher the fidelity. The input includes the PDE parameters and/or parameterized initial or boundary conditions. The output consists of the solution values at the mesh used in the solver. For example, a $50 \times 50$ mesh corresponds to a $2,500$ dimensional output vector. To perform active learning, we considered two-fidelity queries for all the three equations, where the sizes of the corresponding output fields are $16 \times 16$ and $32 \times 32$. In addition, we considered a three-fidelity setting for Poisson's equation, denoted by Poisson-3, and the sizes of the output fields of the three fidelities are $16 \times 16$ , $32 \times 32$, and $64 \times 64$, respectively. For the two-fidelity active learning, we uniformly sampled the inputs, and queried $10$ training examples at the first fidelity and $2$ at the second fidelity. We used those examples as the initial training dataset. Similarly, for Poisson-3, we collected $10$, $5$ and $2$ examples in the first, second, and third fidelity as the initial training set. We generated $500$ samples for test, where the test inputs were uniformly sampled from the domain. The outputs are calculated by running the solver with an even denser mesh --- $128 \times 128$ for Burger's and Poisson's equations, and $100 \times 100$ for Heat equation --- and interpolating the solution values at the target grid~\citep{zienkiewicz1977finite} (this is the standard approach in physical simulation and the accuracy does not change). More details are given in the supplementary material. We ran the solvers at each fidelity for many times and calculated the average running time. We then normalized the average running time to obtain $\lambda_1 = 1$, $\lambda_2 = 3$ and $\lambda_3 = 10$.
\noindent\textbf{Competing methods.} We compared {\textsc{DiTucker}}\xspace with the following active learning approaches. (1) \texttt{MF-BALD}, a straightforward extension of BALD~\citep{houlsby2011bayesian} to integrate multi-fidelity queries. The acquisition function is
\begin{align}
&a_{\text{MF-BALD}}({\bf x}, m) = \frac{1}{\lambda_m} \mathbb{I}\big({\bf y}_m({\bf x}), {\mathcal{W}}|\mathcal{D}\big) =\frac{1}{\lambda_m}\left(\mathbb{H}({\bf y}_m({\bf x})|\mathcal{D}) - \mathbb{E}_{p({\mathcal{W}}|\mathcal{D})}\left[\mathbb{H}({\bf y}_m({\bf x})|{\mathcal{W}}, \mathcal{D})\right]\right) \notag \\
&=\frac{1}{\lambda_m}\left(\mathbb{H}({\bf y}_m({\bf x})|\mathcal{D}) - \frac{d_m}{2}\log(2\pi e \sigma_m^2)\right). \label{eq:mf-bald}
\end{align}
Note that conditioned the NN parameters, the entropy of the observed output ${\bf y}_m$ is only determined by the noise variance $\sigma_m^2$.
(2) \texttt{Dropout-latent}, where we use MC dropout~\citep{gal2017deep} for variational inference, each time draw $100$ dropout samples for the low dimensional latent outputs $\{{\bf f}_1({\bf x}), \ldots, {\bf f}_M({\bf x})\}$ to estimate multi-variate Gaussian posteriors for each ${\bf f}_m$ and ${\widehat{\h}}_m = [{\bf f}_m; {\bf f}_M]$ (via empirical means and covariance matrices), and then follow Section \ref{sect:ac-compute} to calculate and optimize the acquisition function \eqref{eq:ac}. Note that we have also used MC dropout to outright sample the final, high-dimensional outputs $\{{\bf y}_m\}$ and estimate multi-variate Gaussian posteriors to calculate \eqref{eq:ac} and \eqref{eq:mf-bald}. While doing this is much more expensive, the performance did not improve. Instead, it was way worse than \texttt{Dropout-latent} and \texttt{MF-BALD}. See the details in the supplementary material. (3) \texttt{MF-Random}, where each time we randomly select a fidelity and then an input to query.
(4) \texttt{Random-F1}, (5) \texttt{Random-F2}, and (6) \texttt{Random-F3}, where we stick to the first, second and third fidelity, respectively, and randomly sample an input to query each time. \zsdc{(7) \texttt{Random-partial}, where we randomly select a fidelity $m$, and then optimize \eqref{eq:ac} with $m$ fixed.}
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{./figs/rmse_Poisson3.pdf}
\caption{\small Normalized root-mean-square error (nRMSE) for active learning of the solution field of Poisson's equation, with three-fidelity queries. The results are averaged from five runs. }
\label{fig:solving-po-3fid}
\end{figure}
\noindent\textbf{Settings and results.} We introduced a two-layer NN for each fidelity and used \texttt{tanh} as the activation function. The layer width was chosen from $\{8, 16, 32, 64, 128\}$. We set the same dimension for the latent output in each fidelity and selected it from $\{5, 10, 15, 20\}$. For \texttt{Dropout-latent}, we tuned the dropout rate from $\{0.1, 0.2, 0.3, 0.4, 0.5\}$. All the methods were implemented with PyTorch~\citep{paszke2019pytorch}. We used ADAM~\citep{kingma2014adam} for stochastic optimization, where the learning rate was tuned from $\{10^{-4}, 5 \times 10^{-4}, 10^{-3}, 5 \times 10^{-3}, 10^{-2}\}$. We set the number of epochs to $2,000$, which is enough for convergence. We conducted five runs for each method, and in each run, we queried $100$ examples. We report the average normalized root-mean-square-error (nRMSE) \textit{vs.} the accumulated cost in Fig. \ref{fig:solving-pde-2fid} and \ref{fig:solving-po-3fid}. The shaded region shows the standard deviation. We can see that at the beginning, all the methods have the same or comparable performance. Along with more queries, {\textsc{DiTucker}}\xspace quickly achieves better prediction accuracy, and continues to outperform all the other methods by a large margin. Therefore, {\textsc{DiTucker}}\xspace can perform much better with the same cost or achieve the same performance with the smallest cost. It is interesting to see that in Fig. \ref{fig:solving-pde-2fid}a, while all the competing approaches have saturated early, {\textsc{DiTucker}}\xspace keeps improving its prediction accuracy.
\texttt{Dropout-latent} and \texttt{MF-BALD} consistently outperform random querying strategies, demonstrating the effectiveness of the information-based acquisition functions. However, \texttt{MF-BALD} only computes the predictive entropy in each single fidelity (plus the entropy of the noise, see \eqref{eq:mf-bald}), and does not consider how the low-fidelity training examples influence the final predictions (\textit{i}.\textit{e}.,\xspace at the highest fidelity). Its inferior performance to {\textsc{DiTucker}}\xspace indicates that a straightforward extension to the original BALD can be suboptimal. The worse performance of \texttt{Dropout-latent} than {\textsc{DiTucker}}\xspace implies that our structural variational inference can give a better posterior estimation for the latent outputs than the stochastic estimations constructed from the dropout samples. Overall, these results have demonstrated the advantage of our deep multi-fidelity active learning approach.
\subsection{Topology Structure Optimization}
Next, we applied {\textsc{DiTucker}}\xspace in topology structure optimization. A topology structure is a layout of materials, \textit{e}.\textit{g}.\xspace alloy and concrete, in some designated spatial domain. Given the input from the outside environment, \textit{e}.\textit{g}.\xspace external force, we want to find an optimal structure that achieves the maximum (or minimum) interested property, \textit{e}.\textit{g}.\xspace stiffness. Topology structure optimization is crucial to many engineering design and manufacturing problems, including 3D printing, and design of air foils, slab bridges, aerodynamic shapes of race cars, {\textit{etc}.}\xspace The conventional approach is to solve a constraint optimization problem that minimizes a compliance objective subject to a total volume constraint~\citep{sigmund1997design}. However, the numerical computation is usually very costly. We aim to use active learning to learn a model that directly predicts the optimal structure, without the need for running numerical optimization every time.
We considered the stress experiment in~\citep{keshavarzzadeh2018parametric} with an L-shape linear elastic structure. The structure is subjected to a load (\textit{i}.\textit{e}.,\xspace input) on the bottom right half and is discretized in a $[0, 1] \times [0, 1]$ domain. The load is represented by two parameters, the location (in $[0.5, 1]$) and angle (in $[0, \frac{\pi}{2}]$). The goal is to find the structure that achieves the maximum stiffness given the load. Optimizing the structure needs to repeatedly call a numerical solver, where the choice of the mesh determines the fidelity. We used two fidelities to query the training examples. One uses a $50 \times 50$ mesh, the other $75 \times 75$. Correspondingly, the output dimensions are $2,500$ and $5,625$. The costs are measured by the average running time: $\lambda_1 = 1, \lambda_2 = 3$. We randomly generated $500$ structures for test, where the internal solver uses a $100 \times 100$ mesh.
\begin{figure}[H]
\centering
\includegraphics[width=0.48\textwidth]{./figs/rmse_Lbracket.pdf}
\caption{\small Prediction accuracy of active learning in topology structure optimization.}
\label{fig:tpo}
\end{figure}
Initially, we randomly queried $10$ examples at the first fidelity and $2$ at the second fidelity. We then ran all the active learning methods to query $100$ examples. We conducted the experiments for five times, and report the average nRMSE along with the cost in Fig. \ref{fig:tpo}. As we can see, {\textsc{DiTucker}}\xspace achieves much better prediction accuracy than the competing approaches (with the same cost). That implies our predicted structures are much closer to the optimal structures. While the performance of the other methods tended to converge early, {\textsc{DiTucker}}\xspace 's performance kept improving and the trend did not stop even when all the queries were finished. It is worth noting that \texttt{MF-BALD} is even worse than random query strategies. This might because the acquisition function \eqref{eq:mf-bald} does not take into account the relationships between the fidelities, which are critical for this task.
\setlength{\columnsep}{10pt}
\begin{wrapfigure}{r}{0.5\textwidth}
\centering
\begin{subfigure}[b]{0.5\textwidth}
\centering
\includegraphics[width=1.0\linewidth]{figs/topt_ground_v1.eps}
\caption{\small Ground Truth}
\label{fig:tpo-gt}
\end{subfigure}
\begin{subfigure}[b]{0.5\textwidth}
\centering
\includegraphics[width=1.0\linewidth]{figs/topt_mutual_v1.eps}
\caption{\small {\textsc{DiTucker}}\xspace}
\label{fig:tpo-ours}
\end{subfigure}
\begin{subfigure}[b]{0.5\textwidth}
\centering
\includegraphics[width=1.0\linewidth]{figs/topt_dropout_v1.eps}
\caption{\small Dropout-latent}
\label{fig:tpo-dropout}
\end{subfigure}
\begin{subfigure}[b]{0.5\textwidth}
\centering
\includegraphics[width=1.0\linewidth]{figs/topt_bald_v1.eps}
\caption{\small MF-BALD}
\label{fig:tpo-bald}
\end{subfigure}
\begin{subfigure}[b]{0.5\textwidth}
\centering
\includegraphics[width=1.0\linewidth]{figs/topt_full_random_v1.eps}
\caption{\small MF-Random}
\label{fig:tpo-random}
\end{subfigure}
\begin{subfigure}[b]{0.5\textwidth}
\centering
\includegraphics[width=1.0\linewidth]{figs/topt_fix_random_f1_v1.eps}
\caption{\small Random-F1}
\label{fig:tpo-random-1}
\end{subfigure}
\begin{subfigure}[b]{0.5\textwidth}
\centering
\includegraphics[width=1.0\linewidth]{figs/topt_fix_random_f2_v1.eps}
\caption{\small Random-F2}
\label{fig:tpo-random-2}
\end{subfigure}
\caption{\small The predicted topology structures for $8$ loads. All the active learning approaches started with the same training set (10 fidelity-1 and 2 fidelity-2 examples), and ran with $100$ queries.}
\label{fig:pred-top-structure}
\end{wrapfigure}
To perform a fine-grained comparison, we visualize eight structures predicted by all the methods, after the active learning is finished. As shown in Fig. \ref{fig:pred-top-structure}, {\textsc{DiTucker}}\xspace predicted much more accurate structures, which capture both the global shapes and local details, and the density of the materials is closer to the ground-truth. Although \texttt{Dropout-latent} captures the global shapes as well, its predictions are more blurred and miss many local details, \textit{e}.\textit{g}.\xspace the second to seventh structure in Fig. \ref{fig:pred-top-structure}c. The other methods often provided wrong structures (\textit{e}.\textit{g}.\xspace the first and last structure in Fig. \ref{fig:pred-top-structure}d, e, and f) and insufficient density (\textit{e}.\textit{g}.\xspace the second, third and sixth structure in \ref{fig:pred-top-structure}f).
\cmt{
\begin{figure}[!htb]
\centering
\begin{subfigure}[b]{0.5\textwidth}
\centering
\includegraphics[width=1.0\linewidth]{figs/topt_ground_v1.eps}
\caption{\small Ground Truth}
\label{fig:tpo-gt}
\end{subfigure}
\begin{subfigure}[b]{0.5\textwidth}
\centering
\includegraphics[width=1.0\linewidth]{figs/topt_mutual_v1.eps}
\caption{\small {\textsc{DiTucker}}\xspace}
\label{fig:tpo-ours}
\end{subfigure}
\begin{subfigure}[b]{0.5\textwidth}
\centering
\includegraphics[width=1.0\linewidth]{figs/topt_dropout_v1.eps}
\caption{\small Dropout-latent}
\label{fig:tpo-dropout}
\end{subfigure}
\begin{subfigure}[b]{0.5\textwidth}
\centering
\includegraphics[width=1.0\linewidth]{figs/topt_bald_v1.eps}
\caption{\small MF-BALD}
\label{fig:tpo-bald}
\end{subfigure}
\begin{subfigure}[b]{0.5\textwidth}
\centering
\includegraphics[width=1.0\linewidth]{figs/topt_full_random_v1.eps}
\caption{\small MF-Random}
\label{fig:tpo-random}
\end{subfigure}
\begin{subfigure}[b]{0.5\textwidth}
\centering
\includegraphics[width=1.0\linewidth]{figs/topt_fix_random_f1_v1.eps}
\caption{\small Random-F1}
\label{fig:tpo-random-1}
\end{subfigure}
\begin{subfigure}[b]{0.5\textwidth}
\centering
\includegraphics[width=1.0\linewidth]{figs/topt_fix_random_f2_v1.eps}
\caption{\small Random-F2}
\label{fig:tpo-random-2}
\end{subfigure}
\caption{\small The predicted topology structures for $8$ loads. All the active learning approaches started with the same training set (10 fidelity-1 and 2 fidelity-2 examples), and ran with $100$ queries.}
\label{fig:pred-top-structure}
\end{figure}
}
\cmt{
\begin{figure*}
\centering
\setlength\tabcolsep{0pt}
\begin{tabular}[c]{ccc}
\setcounter{subfigure}{0}
\begin{subfigure}[t]{0.33\textwidth}
\centering
\includegraphics[width=\textwidth]{./figs/ns_t_1.eps}
\caption{$t=1$}
\end{subfigure} &
\begin{subfigure}[t]{0.33\textwidth}
\centering
\includegraphics[width=\textwidth]{./figs/ns_t_5.eps}
\caption{$t=5$}
\end{subfigure}
&
\begin{subfigure}[t]{0.33\textwidth}
\centering
\includegraphics[width=\textwidth]{./figs/ns_t_9.eps}
\caption{$t=9$}
\end{subfigure}
\end{tabular}
\vspace{-0.05in}
\caption{\small Examples of the first component of the velocity field (with contour lines) at three time points.} \label{fig:cfd-example}
\vspace{-0.2in}
\end{figure*}
}
\subsection{Predicting Fluid Dynamics}
Third, we applied {\textsc{DiTucker}}\xspace in a computational fluid dynamics (CFD) problem. The task is to predict the first component of the velocity field of a flow within a rectangular domain in $[0, 1] \times [0, 1]$. The flow is driven by the boundaries with a prescribed velocity (\textit{i}.\textit{e}.,\xspace input)~\citep{bozeman1973numerical}.
Along with time, the local velocities inside the fluid will vary differently, and eventually result in turbulent flows. Computing these fields along with time requires us to solve the incompressive Navier-Stokes equations~\citep{chorin1968numerical}, which is known to be challenging to solve because of their complex behaviors under big Reynolds numbers. We considered active learning with two-fidelity queries to predict the first component of the velocity field at evenly spaced $20$ time points in $[0, 10]$ (temporal domain). The examples in the first fidelity were generated with a $50 \times 50$ mesh in the spatial domain (\textit{i}.\textit{e}.,\xspace $[0, 1]\times [0, 1]$), and the second fidelity $75 \times 75$. The corresponding output dimensions are $50,000$ and $112,500$. The input is a five dimensional vector that consists of the prescribed boundary velocity and Reynold number. See more details in the supplementary material. To collect the test dataset, we randomly sampled $256$ inputs and computed the solution with a $128 \times 128$ mesh. The test outputs are obtained by the cubic-spline interpolation. At the beginning, we randomly queried $10$ and $2$ training examples in the first and second fidelity, respectively. Then we ran each active learning method with $100$ queries. We repeated the experiments for five times. The average nRMSE along with the accumulated cost is shown in Fig. \ref{fig:ns}. It can be seen that during the training, {\textsc{DiTucker}}\xspace consistently outperforms all the competing methods by a large margin. On the other hand, to achieve the same level of accuracy, {\textsc{DiTucker}}\xspace spends a much smaller cost. That means, our methods requires much less (high-fidelity) simulations to generate the training examples. This is particularly useful for large-scale CFD applications, in which the simulation is known to be very expensive.
\begin{figure}[H]
\centering
\includegraphics[width=0.48\textwidth]{./figs/rmse_Navier.pdf}
\caption{\small Performance of active learning in predicting velocity fields at $20$ time steps in a flow driven by rectangular boundaries. }
\label{fig:ns}
\end{figure}
\zsdc{Finally, we also examined our deep multi-fidelity model in non-active learning. We compared with several state-of-the-art high-dimensional regression models. They all lack effective active learning approaches. Our model consistently outperforms these methods, often by a large margin. This confirms the advantage of our model in surrogate modeling. See the details in the supplementary material.}
\section{Experiment}
\vspace{-0.1in}
\subsection{Solving Partial Differential Equations}
\vspace{-0.1in}
We first evaluated {\textsc{DiTucker}}\xspace in standard computational physics tasks. Specifically, we used {\textsc{DiTucker}}\xspace to predict the solution fields of three commonly used partial differential equations (PDEs): \textit{Burgers}' , \textit{Poisson}'s and \textit{Heat} equations~\citep{olsen2011numerical}. The training examples are collected by running a numerical solver with different meshes. The more the nodes/steps to create the mesh, the higher the fidelity. The input includes the PDE parameters and/or parameterized initial or boundary conditions. The output consists of the solution values at the mesh used in the solver. For example, a $50 \times 50$ mesh corresponds to a $2,500$ dimensional output vector. For active learning, we considered two-fidelity queries for all the three equations, where the sizes of the corresponding output fields are $16 \times 16$ and $32 \times 32$. In addition, we considered a three-fidelity setting for Poisson's equation, denoted by \textit{Poisson-3}, and the sizes of the output fields of the three fidelities are $16 \times 16$ , $32 \times 32$, and $64 \times 64$, respectively. For the two-fidelity active learning, we uniformly sampled the inputs, and queried $10$ training examples at the first fidelity and $2$ at the second fidelity. We used those examples as the initial training dataset. Similarly, for \textit{Poisson-3}, we collected $10$, $5$ and $2$ examples in the first, second, and third fidelity as the initial training set. We generated $500$ samples for test, where the test inputs were uniformly sampled from the domain. The outputs are calculated by running the solver with an even denser mesh --- $128 \times 128$ for Burger's and Poisson's equations, and $100 \times 100$ for Heat equation --- and interpolating the solution values at the target grid~\citep{zienkiewicz1977finite} (this is the standard approach in physical simulation and the accuracy does not change). More details are given in the Appendix. \textit{We ran the solvers at each fidelity for many times and calculated the average running time. We then normalized the average running time to obtain $\lambda_1 = 1$, $\lambda_2 = 3$ and $\lambda_3 = 10$}.
\noindent\textbf{Competing methods.} We compared {\textsc{DiTucker}}\xspace with the following active learning approaches. (1) {MF-BALD}, a straightforward extension of BALD~\citep{houlsby2011bayesian} to integrate multi-fidelity queries. The acquisition function is $a_{\text{MF-BALD}}({\bf x}, m) = \frac{1}{\lambda_m} \mathbb{I}\big({\bf y}_m({\bf x}), {\mathcal{W}}|\mathcal{D}\big) = \frac{1}{\lambda_m}\left(\mathbb{H}({\bf y}_m({\bf x})|\mathcal{D}) - \mathbb{E}_{p({\mathcal{W}}|\mathcal{D})}\left[\mathbb{H}({\bf y}_m({\bf x})|{\mathcal{W}}, \mathcal{D})\right]\right) = \frac{1}{\lambda_m}\left(\mathbb{H}({\bf y}_m({\bf x})|\mathcal{D}) - \frac{d_m}{2}\log(2\pi e \sigma_m^2)\right)$.
Note that conditioned the NN parameters, the entropy of the observed output ${\bf y}_m$ is only determined by the noise variance $\sigma_m^2$.
(2) {MF-PredVar}, a straightforward extension of the popular predictive variance principle, $a_{\text{MF-PredVar}} = \frac{1}{\lambda_m} \frac{1}{d_m}\sum_{j=1}^{d_m} \text{Var}(y_{mj}|\mathcal{D})$. (3) {Dropout-latent}, where we use MC dropout~\citep{gal2017deep} for variational inference, each time draw $100$ dropout samples for the low dimensional latent outputs $\{{\bf h}_1({\bf x}), \ldots, {\bf h}_M({\bf x})\}$ to estimate multi-variate Gaussian posteriors for each ${\bf h}_m$ and ${\widehat{\h}}_m = [{\bf h}_m; {\bf h}_M]$ (via empirical means and covariance matrices), and then follow Section \ref{sect:ac-compute} to calculate and optimize the acquisition function \eqref{eq:ac}. Note that we have also used MC dropout to outright sample the final, high-dimensional outputs $\{{\bf y}_m\}$ and estimate multi-variate Gaussian posteriors to calculate \eqref{eq:ac} and $a_{\text{MF-BALD}}$. While doing this is much more expensive, the performance did not improve. Instead, it was way worse than {Dropout-latent} and {MF-BALD}. See the details in the Appendix. (4) {MF-Random}, where each time we randomly select a fidelity and then an input to query.
(5) {Random-F1}, (6) {Random-F2}, and (7) {Random-F3}, where we stick to the first, second and third fidelity, respectively, and randomly sample an input to query each time.
\begin{figure*}
\centering
\setlength\tabcolsep{0pt}
\begin{tabular}[c]{ccc}
\setcounter{subfigure}{0}
\begin{subfigure}[t]{0.33\textwidth}
\centering
\includegraphics[width=\textwidth]{./figs/new/Burgers_legend.pdf}
\caption{\small \textit{Burgers}}
\end{subfigure} &
\begin{subfigure}[t]{0.33\textwidth}
\centering
\includegraphics[width=\textwidth]{./figs/new/Heat2.pdf}
\caption{\small \textit{Heat}}
\end{subfigure}
&
\begin{subfigure}[t]{0.33\textwidth}
\centering
\includegraphics[width=\textwidth]{./figs/new/Poisson2.pdf}
\caption{\small \textit{Poisson}}
\end{subfigure}\\
\begin{subfigure}[t]{0.33\textwidth}
\centering
\includegraphics[width=\textwidth]{./figs/new/Poisson3.pdf}
\caption{\small \textit{Poisson-3}}\label{fig:solving-po-3fid}
\end{subfigure} &
\begin{subfigure}[t]{0.33\textwidth}
\centering
\includegraphics[width=\textwidth]{./figs/new/Lbracket.pdf}
\caption{\small \textit{Topology structure optimization}} \label{fig:tpo}
\end{subfigure}
&
\begin{subfigure}[t]{0.33\textwidth}
\centering
\includegraphics[width=\textwidth]{./figs/new/Navier.pdf}
\caption{\small \textit{Fluid dynamics}}\label{fig:ns}
\end{subfigure}
\end{tabular}
\vspace{-0.1in}
\caption{\small Normalized root-mean-square error (nRMSE) \textit{vs.} total cost (running time) in active learning. The results are averaged from five runs. The shaded regions indicate the standard deviations.} \label{fig:al-acc}
\vspace{-0.2in}
\end{figure*}
\noindent\textbf{Settings and results.} We introduced a two-layer NN for each fidelity and used \texttt{tanh} as the activation function\footnote{We tried more layers per fidelity, \textit{i}.\textit{e}.,\xspace 3-5, and the overall performance did not improve. In addition, our experience shows that alternative activation functions, like \texttt{ReLU} and \texttt{LeakyReLu}, are inferior to \texttt{tanh}. This is consistent with the choice of typical physics informed neural networks~\citep{raissi2019physics}.}. The layer width was chosen from $\{8, 16, 32, 64, 128\}$. We set the same dimension for the latent output in each fidelity and selected it from $\{5, 10, 15, 20\}$. For {Dropout-latent}, we tuned the dropout rate from $\{0.1, 0.2, 0.3, 0.4, 0.5\}$. All the methods were implemented with PyTorch~\citep{paszke2019pytorch}. We used ADAM~\citep{kingma2014adam} for stochastic optimization, where the learning rate was selected from $\{10^{-4}, 5 \times 10^{-4}, 10^{-3}, 5 \times 10^{-3}, 10^{-2}\}$. We set the number of epochs to $2,000$, which is enough for convergence. We conducted five runs for each method, and in each run, we queried $100$ examples. We report the average normalized root-mean-square-error (nRMSE) \textit{vs.} the accumulated cost in Fig. \ref{fig:al-acc}a-d. The shaded region shows the standard deviation. We can see that at the beginning, all the methods have the same or comparable performance. Along with more queries, {\textsc{DiTucker}}\xspace quickly achieves better prediction accuracy, and continues to outperform all the other methods by a large margin. Therefore, {\textsc{DiTucker}}\xspace can perform much better with the same cost or achieve the same performance with the smallest cost. It is interesting to see that in Fig. \ref{fig:al-acc}a, while all the competing approaches have saturated early, {\textsc{DiTucker}}\xspace keeps improving its prediction accuracy.
\setlength{\columnsep}{5pt}
\begin{figure}[!htb]
\centering
\begin{subfigure}[b]{0.4\textwidth}
\centering
\includegraphics[width=1.0\linewidth]{figs/topt_ground_v1-eps-converted-to.pdf}
\caption{\small Ground Truth}
\label{fig:tpo-gt}
\end{subfigure}
\begin{subfigure}[b]{0.4\textwidth}
\centering
\includegraphics[width=1.0\linewidth]{figs/topt_mutual_v1-eps-converted-to.pdf}
\caption{\small {\textsc{DiTucker}}\xspace}
\label{fig:tpo-ours}
\end{subfigure}
\begin{subfigure}[b]{0.4\textwidth}
\centering
\includegraphics[width=1.0\linewidth]{figs/topt_dropout_v1-eps-converted-to.pdf}
\caption{\small Dropout-latent}
\label{fig:tpo-dropout}
\end{subfigure}
\begin{subfigure}[b]{0.4\textwidth}
\centering
\includegraphics[width=1.0\linewidth]{figs/topt_bald_v1-eps-converted-to.pdf}
\caption{\small MF-BALD}
\label{fig:tpo-bald}
\end{subfigure}
\begin{subfigure}[b]{0.4\textwidth}
\centering
\includegraphics[width=1.0\linewidth]{figs/new/pdv-eps-converted-to.pdf}
\caption{\small MF-PredVar}
\label{fig:tpo-pdv}
\end{subfigure}
\begin{subfigure}[b]{0.4\textwidth}
\centering
\includegraphics[width=1.0\linewidth]{figs/topt_full_random_v1-eps-converted-to.pdf}
\caption{\small MF-Random}
\label{fig:tpo-random}
\end{subfigure}
\begin{subfigure}[b]{0.4\textwidth}
\centering
\includegraphics[width=1.0\linewidth]{figs/topt_fix_random_f1_v1-eps-converted-to.pdf}
\caption{\small Random-F1}
\label{fig:tpo-random-1}
\end{subfigure}
\begin{subfigure}[b]{0.4\textwidth}
\centering
\includegraphics[width=1.0\linewidth]{figs/topt_fix_random_f2_v1-eps-converted-to.pdf}
\caption{\small Random-F2}
\label{fig:tpo-random-2}
\end{subfigure}
\vspace{-0.1in}
\caption{\small The predicted topology structures for $8$ loads. All the active learning approaches started with the same training set (10 fidelity-1 and 2 fidelity-2 examples), and ran with $100$ queries.}
\label{fig:pred-top-structure}
\vspace{-0.2in}
\end{figure}
The inferior performance of {MF-BALD} and {MF-PredVar} to {\textsc{DiTucker}}\xspace indicates that the straightforward extension of BALD and predictive variance can be suboptimal, which might partly because they miss the influence of the low-fidelity training examples on the final predictions (\textit{i}.\textit{e}.,\xspace at the highest fidelity). How to better generalize BALD and PredVar to the multi-fidelity setting remains an open problem.
Overall, these results have demonstrated the advantage of our deep multi-fidelity active learning approach.
\vspace{-0.1in}
\subsection{Topology Structure Optimization}
\vspace{-0.1in}
Next, we applied {\textsc{DiTucker}}\xspace in topology structure optimization. A topology structure is a layout of materials, \textit{e}.\textit{g}.\xspace alloy and concrete, in some designated spatial domain. Given the input from the outside environment, \textit{e}.\textit{g}.\xspace external force, we want to find an optimal structure that achieves the maximum (or minimum) interested property, \textit{e}.\textit{g}.\xspace stiffness. Topology structure optimization is crucial to many engineering design and manufacturing problems, including 3D printing, and design of air foils, slab bridges, aerodynamic shapes of race cars, {\textit{etc}.}\xspace The conventional approach is to solve a constraint optimization problem that minimizes a compliance objective subject to a total volume constraint~\citep{sigmund1997design}. However, the numerical computation is usually very costly. We aim to use active learning to learn a model that directly predicts the optimal structure, without the need for running numerical optimization every time.
We considered the stress experiment in~\citep{keshavarzzadeh2018parametric} with an L-shape linear elastic structure. The structure is subjected to a load (\textit{i}.\textit{e}.,\xspace input) on the bottom right half and is discretized in a $[0, 1] \times [0, 1]$ domain. The load is represented by two parameters, the location (in $[0.5, 1]$) and angle (in $[0, \frac{\pi}{2}]$). The goal is to find the structure that achieves the maximum stiffness given the load. Optimizing the structure needs to repeatedly call a numerical solver, where the choice of the mesh determines the fidelity. We used two fidelities to query the training examples. One uses a $50 \times 50$ mesh, the other $75 \times 75$. Correspondingly, the output dimensions are $2,500$ and $5,625$. The costs are measured by the average running time: $\lambda_1 = 1, \lambda_2 = 3$. We randomly generated $500$ structures for test, where the internal solver uses a $100 \times 100$ mesh. Initially, we randomly queried $10$ examples at the first fidelity and $2$ at the second fidelity. We then ran all the active learning methods to query $100$ examples. We conducted the experiments for five times, and report the average nRMSE along with the cost in Fig. \ref{fig:tpo}. As we can see, {\textsc{DiTucker}}\xspace achieves much better prediction accuracy than the competing approaches (with the same cost). That implies our predicted structures are much closer to the optimal structures. While the performance of the other methods tended to converge early, {\textsc{DiTucker}}\xspace 's performance kept improving and the trend did not stop even when all the queries were finished.
For a fine-grained comparison, we visualize eight structures predicted by all the methods, after the active learning is finished. As shown in Fig. \ref{fig:pred-top-structure}, {\textsc{DiTucker}}\xspace predicted much more accurate structures, which capture both the global shapes and local details, and the density of the materials is closer to the ground-truth. Although {Dropout-latent} captures the global shapes as well, its predictions are more blurred and miss many local details, \textit{e}.\textit{g}.\xspace the second to seventh structure in Fig. \ref{fig:pred-top-structure}c. The other methods often provided wrong structures (\textit{e}.\textit{g}.\xspace the first and last structure in Fig. \ref{fig:pred-top-structure}d, f, and g, the second and fourth structure in Fig. \ref{fig:pred-top-structure}e) and insufficient density (\textit{e}.\textit{g}.\xspace the second, third and sixth structure in \ref{fig:pred-top-structure}g).
\vspace{-0.15in}
\subsection{Predicting Fluid Dynamics}
\vspace{-0.1in}
Third, we applied {\textsc{DiTucker}}\xspace in a computational fluid dynamics (CFD) problem. The task is to predict the first component of the velocity field of a flow within a rectangular domain in $[0, 1] \times [0, 1]$. The flow is driven by the boundaries with a prescribed velocity (\textit{i}.\textit{e}.,\xspace input)~\citep{bozeman1973numerical}.
Along with time, the local velocities inside the fluid will vary differently, and eventually result in turbulent flows. Computing these fields along with time requires us to solve the incompressive Navier-Stokes equations~\citep{chorin1968numerical}, which is known to be challenging to solve because of their complex behaviors under big Reynolds numbers. We considered active learning with two-fidelity queries to predict the first component of the velocity field at evenly spaced $20$ time points in $[0, 10]$ (temporal domain). The examples in the first fidelity were generated with a $50 \times 50$ mesh in the spatial domain (\textit{i}.\textit{e}.,\xspace $[0, 1]\times [0, 1]$), and the second fidelity $75 \times 75$. The corresponding output dimensions are $50,000$ and $112,500$. The input is a five dimensional vector that consists of the prescribed boundary velocity and Reynold number. See more details in the Appendix. To collect the test dataset, we randomly sampled $256$ inputs and computed the solution with a $128 \times 128$ mesh. The test outputs are obtained by the cubic-spline interpolation. At the beginning, we randomly queried $10$ and $2$ training examples in the first and second fidelity, respectively. Then we ran each active learning method with $100$ queries. We repeated the experiments for five times. The average nRMSE along with the accumulated cost is shown in Fig. \ref{fig:ns}. It can be seen that during the training, {\textsc{DiTucker}}\xspace consistently outperforms all the competing methods by a large margin. On the other hand, to achieve the same level of accuracy, {\textsc{DiTucker}}\xspace spends a much smaller cost. That means, our methods requires much less (high-fidelity) simulations to generate the training examples. This is particularly useful for large-scale CFD applications, in which the simulation is known to be very expensive.
To further confirm the gains in computational efficiency, we compared the cost of running the standard numerical methods for the topological optimization and CFD tasks. After active learning, our learned surrogate models give $42$x and $466$x speed-up in computing (predicting) the high-fidelity solution fields against the numerical methods. Even counting the whole active learning cost, with the growth of test cases, the average cost (time) of {\textsc{DiTucker}}\xspace in computing a solution quickly becomes much smaller than the standard numerical methods. See details in Sec. 2 of Appendix.
\textbf{Non-Active Learning.}{ Finally, to confirm the capacity of our deep NN model, we compared with state-of-the-art high-dimensional non-active learning surrogate models. Note that these models all lack effective active learning strategies. Our model consistently outperforms these methods, often by a large margin (Sec. 3 of the Appendix).}
\section{Introduction}
Many applications require us to compute a mapping from low-dimensional inputs to high-dimensional outputs. For example, topology optimization~\citep{rozvany2009critical} aims to find an optimal structure (high-dimensional output) given several design parameters (low-dimensional input). Physical simulation uses numerical solvers to solve partial differential equations (PDEs), which maps the PDE parameters and parameterized initial/boundary conditions (low dimensional input) to the high-dimensional solution field on a mesh. The exact computation of these mappings is often costly. Hence, learning a surrogate model to directly predict the output (rather than computing from scratch every time) is of great interest and importance~\citep{kennedy2000predicting,conti2010bayesian}.
However, collecting training examples is a bottleneck, because to obtain these examples we still have to conduct the original, expensive computation (\textit{e}.\textit{g}.\xspace running numerical solvers). To reduce the cost, we can compute the training examples at different fidelities to enable a trade-off between the cost and quality. Low fidelity examples are cheap to acquire but inaccurate while high-fidelity examples are much more accurate yet expensive. Despite the disparity in accuracy, the outputs at different fidelities are strongly correlated. Hence, examples from different fidelities can all be useful in learning the target mapping.
To reduce the cost while maximizing the learning performance, we develop {\textsc{DiTucker}}\xspace, a deep multi-fidelity active learning approach that can identify both the fidelity and input location to query (or generate) new training examples so as to achieve the best benefit-cost ratio. To our knowledge, this is the first work that incorporates multi-fidelity queries in active learning of high-dimensional outputs (See Fig. 1 in the Appendix for more illustrations). Our work naturally extends the standard active learning, where the query cost is assumed uniform, and minimizing the total cost is equivalent to minimizing the number of queries.
Specifically, we first propose an expressive deep multi-fidelity model. We use a chain of neural networks (NNs) to model the outputs in each fidelity. Each NN generates a low-dimensional latent output first, and then projects it to the high-dimensional observational space. Both the original input and latent output are fed into the NN of the next fidelity so that we can efficiently propagate information throughout the fidelities and flexibly capture their complex relationships. Second, we propose an acquisition function based on the mutual information of the outputs between each fidelity and the highest fidelity (at which we predict the target function). This can be viewed as an extension of the predictive uncertainty principle for the traditional, singe-fidelity active learning. When we seek to query with the highest fidelity, the acquisition function is reduced to the output entropy. We found empirically our acquisition function outperforms the popular BALD~\citep{houlsby2011bayesian} and predictive variance principle~\citep{gal2017deep} adapted to the multi-fidelity setting. Third, we address the challenges of computing and optimizing the acquisition function. Due to the large output dimension, it is very expensive or even infeasible to estimate the required covariance and cross covariance matrices with popular Monte-Carlo (MC) Dropout~\citep{gal2016dropout} samples. To overcome this problem, we consider the NN weights in the latent output layers as random variables and all the other weights as hyper-parameters. We develop a stochastic structural variational learning algorithm to jointly estimate the hyper-parameters and posterior of the random weights. We then use the multi-variate delta method to compute the moments of the hidden outputs in each fidelity, and use moment-matching to estimate a joint Gaussian posterior of the hidden outputs. We use Weinstein-Aronszajn identity to compute the entropy of their projection --- the observed high-dimensional outputs. In this way, we can analytically calculate and optimize the acquisition function in a tractable, reliable and efficient way.
For evaluation, we examined {\textsc{DiTucker}}\xspace in three benchmark tasks of computational physics, a topology structure optimization problem, and a computational dynamic fluids (CFD) application to predict flow velocity fields.\cmt{ of flows. driven by rectangular boundaries.} The output dimensions of these applications vary from hundreds to hundreds of thousands. Our method consistently achieves much better learning performance with the same query cost, as compared with using random query strategies, approximating the acquisition function with dropout samples of the latent outputs, and using other acquisition functions. Our learned surrogates gain 40x and 460x speed-up in solving (predicting) optimal structures and velocity fields than running standard numerical methods. Even counting the total cost of active training, the per-solution cost of {\textsc{DiTucker}}\xspace rapidly becomes far below that of the numerical methods with the increase of acquired solutions.
\subsubsection*{\bibname}}
\usepackage[utf8]{inputenc}
\usepackage[T1]{fontenc}
\usepackage{hyperref}
\usepackage{url}
\usepackage{booktabs}
\usepackage{amsfonts}
\usepackage{nicefrac}
\usepackage{microtype}
\usepackage{xcolor}
\usepackage{microtype}
\usepackage{graphicx}
\usepackage{booktabs}
\usepackage{microtype}
\usepackage{graphicx}
\usepackage{booktabs}
\usepackage{natbib}
\usepackage{amssymb, amsmath,amsthm}
\usepackage{amsfonts}
\usepackage{algorithm}
\usepackage{algorithmic}
\usepackage{paralist}
\usepackage{multirow}
\usepackage{times}
\usepackage{wrapfig}
\usepackage{url,enumerate}
\usepackage{color,xcolor}
\usepackage{makeidx}
\usepackage{amsmath,amssymb}
\usepackage{mathtools}
\usepackage[small, compact]{titlesec}
\usepackage{xspace}
\usepackage{epstopdf}
\usepackage{cite}
\usepackage{mathrsfs}
\usepackage{times}
\usepackage{enumerate}
\usepackage{color}
\usepackage{graphicx,epsfig}
\usepackage{amsmath,amssymb,xspace}
\usepackage{url}
\usepackage{hyperref}
\usepackage{bm}
\usepackage{bbm}
\usepackage{upgreek}
\usepackage{cleveref}
\usepackage{multirow}
\usepackage{ulem}
\usepackage{cancel}
\usepackage{subcaption}
\usepackage{dsfont}
\usepackage{hyperref}
\newcommand{\zsdc}[1]{[\textcolor{blue}{#1}]}
\newtheorem{theorem}{Theorem}[section]
\newtheorem{corollary}{Corollary}[theorem]
\newtheorem{lem}[theorem]{Lemma}
\newcommand{\Wei}[1]{{\color{blue}#1}}
\newcommand{{\textsc{DiTucker}}\xspace}{{DMFAL}\xspace}
\newcommand{{{POST}}\xspace}{{{POST}}\xspace}
\newcommand{\cmt}[1]{}
\newcommand{\textit{e}.\textit{g}.\xspace}{{\textit{e.g.},}\xspace}
\newcommand{\textit{i}.\textit{e}.,\xspace}{{\textit{i.e.},}\xspace}
\newcommand{{\textit{etc}.}\xspace}{{\textit{etc}.}\xspace}
\input{emacscomm.tex}
\begin{document}
\twocolumn[
\aistatstitle{Deep Multi-Fidelity Active Learning of High-Dimensional Outputs}
\aistatsauthor{ Shibo Li \And Robert M. Kirby \And Shandian Zhe }
\aistatsaddress{ School of Computing, University of Utah \\
\texttt{ \{shibo, kirby, zhe\}@cs.utah.edu} }
]
\input{./abstract}
\input{./intro}
\input{./model}
\input{./related}
\input{./exp_zhe2}
\input{./conclusion}
\bibliographystyle{apalike}
\section{Background}
\textbf{Problem setting.} We assume that we can query training examples with $M$ fidelities, which correspond to $M$ mappings, $\{{\bf f}_m({\bf x}) \in \mathds{R}^{d_m}\}_{1 \le m \le M}$ where ${\bf x} \in {\mathcal{X}}$ is an $r$-dimensional input, $r$ is small, $d_m$ is the output dimension, often very large, and $\{d_m\}$ are not necessarily identical. In general, we assume $d_1 \le \ldots \le d_M$. For instance, in structure design, the input can be the design parameters. The high fidelity examples correspond to high-resolution structures while the low fidelity ones low-resolution structures. We denote by $\lambda_m$ the cost to query with fidelity $m$. To perform active learning, we begin with a small set of training examples with mixed fidelities. Each step, we select both the fidelity and input location to query new training examples, so as to best balance the learning improvement and computational cost, namely, maximizing the benefit-cost ratio.
\textbf{Dropout active learning.} A very successful application of active learning with deep neural networks is image classification~\citep{gal2017deep,kirsch2019batchbald}. Typically, a pool of unlabeled input examples, \textit{i}.\textit{e}.,\xspace images, is pre-collected. Each time, we rank the input examples according to an acquisition function computed based on the current model. We then query the labels of the top examples, and add them into the training set. There are two popular acquisition functions. The first one is the predictive entropy of the label,
\begin{align}
\mathbb{H}[y|{\bf x}, \mathcal{D}] = -\sum_{c} p(y=c|{\bf x}, \mathcal{D}) \log p(y=c|{\bf x}, \mathcal{D}), \label{eq:image-ac-1}
\end{align}
where ${\bf x}$ is the input, and $\mathcal{D}$ are the training data. The other one is BALD (Bayesian active learning by disagreement)~\citep{houlsby2011bayesian} --- the mutual information between the label and model parameters,
\begin{align}
\mathbb{I}[y, \boldsymbol{\theta}|{\bf x}, \mathcal{D}] = \mathbb{H}[y|{\bf x}, \mathcal{D}]- \mathbb{E}_{p(\boldsymbol{\theta}|\mathcal{D})}\left[\mathbb{H}({\bf y}|{\bf x}, \boldsymbol{\theta}, \mathcal{D})\right], \label{eq:image-ac-2}
\end{align}
where $\boldsymbol{\theta}$ are the NN weights. The computation of these acquisition functions require us to first estimate the posterior of the NN output and weights. A popular approach is to use Monte-Carlo (MC) Dropout~\citep{gal2016dropout}, which can be viewed as a variational inference method of Bayesian NNs. MC Dropout uses a probability $p_i$ to randomly drop the neurons in each layer $i$ (equivalent to zeroing out the corresponding weights) in the forward pass and performs backward pass over the remaining neurons. After training, the weights and output obtained from one such forward pass can be viewed as a sample of the approximate posterior. We can run Dropout multiple times to collect a set of posterior samples and calculate the acquisition functions by Monte-Carlo approximation.
\vspace{-0.05in}
\section{Deep Multi-Fidelity Modeling for High-Dimensional Outputs}
\vspace{-0.05in}
Despite its success, dropout active learning might be inappropriate for high-dimensional (continuous) outputs. When the output is a continuous vector, it is natural to use a multivariate Gaussian posterior, $p({\bf y}|{\bf x}, \mathcal{D}) \approx {\bf N}({\bf y}|\boldsymbol{\mu}({\bf x}), \boldsymbol{\Sigma}({\bf x}))$, and the entropy $\mathbb{H}({\bf y}|{\bf x}, \mathcal{D})$ in the acquisition functions (see \eqref{eq:image-ac-1} and \eqref{eq:image-ac-2}) require us to compute the log determinant of the covariance matrix, $\log|\boldsymbol{\Sigma}({\bf x})|$. While we can use dropout samples to construct an empirical estimate $\widehat{\boldsymbol{\Sigma}}({\bf x})$, due to the large output dimension $n$, the computation of this $n\times n$ matrix and its log determinant is very expensive or even infeasible.
One might seek to only estimate the variance of each individual output. However, this will ignore the strong, complex output correlations, which is critical to effectively evaluate the uncertainty and calculate the acquisition function. Furthermore, in our active learning task, we want to optimize the acquisition function to find the best input (rather than rank the inputs in a pre-collected pool). The empirical covariance estimation from random samples can lead to numerical instability in optimization and inconsistent results from different runs.
To overcome these problems for multi-fidelity active learning, we propose a highly expressive deep multi-fidelity model, for which we develop a structural variational inference method to capture the posterior dependency of the outputs. We propose a novel acquisition function to allow multi-fidelity queries, and develop an efficient, tractable, and reliable approach to calculate the acquisition function.
Specifically, we use deep neural networks to build a multi-fidelity high-dimensional output model (see Fig. \ref{fig:graphical}) that can flexibly, efficiently capture various complex relationships between the outputs and between the fidelities, taking advantage of these relationships to enhance the predictive performance at the highest fidelity (where we make final predictions). For each fidelity $m$, we introduce a neural network (NN) that first generates a $k_m$ dimensional latent output, and then projects it to the $d_m$ dimensional observed vector, where $k_m \ll d_m$. The NN is parameterized by $\{{\bf W}_m, \boldsymbol{\theta}_m, {\bf A}_m\}$ where ${\bf W}_m$ is a $k_m \times l_m$ weight matrix in the latent output layer, ${\bf A}_m$ is a $d_m \times k_m$ projection matrix, and $\boldsymbol{\theta}_m$ the weights in all the other layers. Denote by ${\bf x}_m$ the NN input, by ${\bf f}_m({\bf x})$ the latent NN output, and by ${\bf y}_m({\bf x})$ the observation. The model is defined by
\begin{align}
{\bf x}_m &= [{\bf x}; {\bf f}_{m-1}({\bf x})], \notag \\
{\bf f}_m({\bf x}) &= {\bf W}_m\boldsymbol{\phi}_{\boldsymbol{\theta}_m}({\bf x}_m),\notag \\
{\bf y}_m({\bf x}) &= {\bf A}_m {\bf f}_m({\bf x}) + \boldsymbol{\epsilon}_m, \label{eq:our-model}
\end{align}
where $\boldsymbol{\phi}_{\boldsymbol{\theta}_m}$ is the $l_m$ dimensional input to the latent output layer (therefore parameterized by $\boldsymbol{\theta}_m$), and $\boldsymbol{\epsilon}_m \sim {\bf N}(\boldsymbol{\epsilon}_m|{\bf 0}, \sigma^2_m{\bf I})$ is an isotropic Gaussian noise. We set $k_0 = 0$ and so ${\bf f}_0({\bf x})$ is empty; when $m=1$, we have ${\bf x}_m = {\bf x}$. Note that $\boldsymbol{\phi}_{\boldsymbol{\theta}_m}$ can be viewed as a set of nonlinear basis functions. Through their combinations via ${\bf W}_m$ and ${\bf A}_m$, we can flexibly capture the complex, strong relationships between the elements of ${\bf y}_m$ to improve the prediction. Furthermore, the input ${\bf x}_{m}$ is constructed by appending to the original input ${\bf x}$ the latent output ${\bf f}_{m-1}$, which can be viewed as a compact (or low-rank) summary of all the information up to fidelity $m-1$. After a series of linear and nonlinear transformations, we obtain the latent output ${\bf f}_m$ --- the summary up to fidelity $m$ --- and then generate the high dimensional observation ${\bf y}_m$. In this way, we efficiently integrate the information from lower fidelities and capture the complex relationship between the current and previous fidelities by learning a nonlinear mapping, ${\bf f}_m({\bf x}) = h({\bf x}, {\bf f}_{m-1}({\bf x}))$, where $h(\cdot)$ is fulfilled by an NN.
\begin{figure}
\centering
\includegraphics[width=0.4\textwidth]{./graphic}
\caption{\small Graphical representation of the deep multi-fidelity model. The low dimensional latent output in each fidelity ${\bf f}_m({\bf x})$ ($1\le m \le M$) is generated by a (deep) neural network.} \label{fig:graphical}
\vspace{-0.15in}
\end{figure}
We place a standard Gaussian prior over the elements in each ${\bf W}_m$. Similar to~\citep{snoek2015scalable}, we consider all the remaining parameters as hyper-parameters to ease the posterior inference and uncertainty reasoning. Given the training dataset $\mathcal{D} = \{\{({\bf x}_{nm}, {\bf y}_{nm})|m\}_{n=1}^{N_m}\}_m$, the joint probability of our model is given by
\begin{align}
&p({\mathcal{W}}, {\mathcal{Y}}|{\mathcal{X}}, \Theta, {\bf s}) = \prod\nolimits_{m=1}^M {\bf N}\left(\vec({\bf W}_m)|{\bf 0}, {\bf I}\right)\notag \\
&\cdot \prod\nolimits_{n=1}^{N_m} {\bf N}\big({\bf y}_{nm}|{\bf A}_m {\bf f}_m({\bf x}_{nm}), \sigma_m^2{\bf I}\big), \label{eq:joint-prob}
\end{align}
where ${\mathcal{W}} = \{{\bf W}_m\}_{1\le m\le M}$, $\Theta=\{\boldsymbol{\theta}_m, {\bf A}_m\}_{1\le m \le M}$, ${\bf s} = [\sigma^2_1, \ldots, \sigma^2_M]^\top$, and $\{{\mathcal{X}}$, ${\mathcal{Y}}\}$ are the inputs and outputs in $\mathcal{D}$.
To estimate the posterior of our model (which is used to calculate the acquisition function and query new examples), we develop a stochastic structural variational learning algorithm. Specifically, for each ${\bf W}_m$, we introduce a multivariate Gaussian posterior, $q({\bf W}_m) = {\bf N}\left(\vec({\bf W}_m)|\boldsymbol{\mu}_m, \boldsymbol{\Sigma}_m\right)$. To ensure the positive definiteness, we parameterize $\boldsymbol{\Sigma}_m$ by its Cholesky decomposition, $\boldsymbol{\Sigma}_m = \L_m \L_m^\top$, where $\L_m$ is a lower triangular matrix. We then assume the posterior $q({\mathcal{W}}) =\prod_{m=1}^M q({\bf W}_m)$, and construct a variational model evidence lower bound (ELBO)~\citep{wainwright2008graphical},
\begin{align}
&{\mathcal{L}}\big(q({\mathcal{W}}), \Theta, {\bf s}\big) = \mathbb{E}_{q}\left[ \log\frac{p({\mathcal{W}}, {\mathcal{Y}}|{\mathcal{X}}, \Theta, {\bf s})}{q({\mathcal{W}})}\right] \notag \\
&=-\text{KL}\big(q({\mathcal{W}})\|p({\mathcal{W}})\big) \notag \\
&+ \sum_{m=1}^M\sum_{n=1}^{N_m} \mathbb{E}_q\left[\log {\bf N}({\bf y}_{nm}|{\bf A}_m {\bf f}_m({\bf x}_{nm}),\sigma_m^2{\bf I})\right], \label{eq:elbo}
\end{align}
where $\text{KL}(\cdot \| \cdot)$ is Kullback Leibler divergence and $p({\mathcal{W}})$ the prior of ${\mathcal{W}}$. We maximize ${\mathcal{L}}$ to jointly estimate $q({\mathcal{W}})$ and the hyper-parameters. While ${\mathcal{L}}$ is intractable, it is straightforward to use the reparameterization trick~\citep{kingma2013auto} to generate parameterized samples for each $\vec({\bf W}_m)$: $\boldsymbol{\mu}_m + \L_m \boldsymbol{\xi}_m$ where $\boldsymbol{\xi}_m \sim {\bf N}({\bf 0}, {\bf I})$, substitute them into \eqref{eq:elbo} to obtain a stochastic estimate of ${\mathcal{L}}$, and perform efficient stochastic optimization.
\vspace{-0.1in}
\section{Multi-Fidelity Active Learning}
\vspace{-0.1in}
\subsection{Mutual Information based Acquisition Function}
\vspace{-0.1in}
We now consider how to perform active learning with multi-fidelity queries. We assume that at each fidelity $m$, the most valuable training example is the one that can best improve our final prediction, namely, the prediction at the highest fidelity $M$.
To this end, we propose our acquisition function based on the mutual information between the outputs at fidelity $m$ and $M$,
\begin{align}
&a({\bf x}, m) = \frac{1}{\lambda_m} \mathbb{I}\left({\bf y}_m({\bf x}), {\bf y}_M({\bf x})|\mathcal{D}\right) \label{eq:ac} \\
&= \frac{1}{\lambda_m} \big(\mathbb{H}({\bf y}_m|\mathcal{D}) +\mathbb{H}({\bf y}_M|\mathcal{D}) - \mathbb{H}({\bf y}_m, {\bf y}_M|\mathcal{D}) \big), \notag
\end{align}
where $\lambda_m>0$ is the cost of querying a training example with fidelity $m$. When $m=M$, we have $a({\bf x}, M) = \frac{1}{\lambda_M} \mathbb{H}\left({\bf y}_M({\bf x})|\mathcal{D}\right)$ --- the output entropy. Therefore, our acquisition function is an extension of the popular predictive entropy principle in conventional active learning. At each step, we maximize our acquisition function to identify a pair of fidelity and input location that give the biggest benefit-cost ratio.
\vspace{-0.1in}
\subsection{Efficient Acquisition Function Calculation}\label{sect:ac-compute}
\vspace{-0.1in}
To maximize the acquisition function \eqref{eq:ac}, a critical challenge is to compute the posterior of the outputs $\{{\bf y}_m\}$ in every fidelity, based on our model estimation results, $p({\mathcal{W}}|\mathcal{D}) \approx q({\mathcal{W}})$. Due to high dimensionality of each ${\bf y}_m$ and the nonlinear coupling of the latent outputs across the fidelities (see \eqref{eq:our-model}), the computation is challenging and analytically intractable. To address this issue, we first consider approximating the posterior of the low dimensional latent output ${\bf f}_m$, which can be viewed a function of the random NN weights $\boldsymbol{\Omega}_m = \{{\bf W}_1, \ldots, {\bf W}_m\}$ (given the input ${\bf x}$). We use multivariate Delta's method~\citep{oehlert1992note,bickel2015mathematical} to compute the moments of ${\bf f}_m$. Specifically, we approximate ${\bf f}_m$ with a first-order Taylor expansion,
\begin{align}
{\bf f}_m(\boldsymbol{\Omega}_m) \approx {\bf f}_m(\mathbb{E}[\boldsymbol{\Omega}_m]) + {\bf J}_m\left(\boldsymbol{\eta}_m - \mathbb{E}[\boldsymbol{\eta}_m]\right), \label{eq:1order}
\end{align}
where the expectation is under $q(\cdot)$, $\boldsymbol{\eta}_m = \vec(\boldsymbol{\Omega}_m)$, ${\bf J}_m = \frac{\partial {\bf f}_m}{\partial \boldsymbol{\eta}_m}|_{\boldsymbol{\eta}_m = \mathbb{E}[\boldsymbol{\eta}_m]}$ is the Jacobian matrix at the mean. The rationale is as follows. First, ${\bf f}_m$ is linear to ${\bf W}_m$, and the second-order derivative is simply ${\bf 0}$. Second, as the NN output, ${\bf f}_m$ is highly nonlinear to the random weights in previous layers, \textit{i}.\textit{e}.,\xspace ${\bf W}_j$ ($j<m$). Hence, we can assume the change rate of ${\bf f}_m$ (\textit{e}.\textit{g}.\xspace gradient or Jacobian) has much greater scales than the posterior covariance of ${\bf W}_j$ in the second-order term of the Taylor expansion. Note that $q({\bf W}_j)$ is much more informative and hence much more concentrated than the prior $ {\bf N}(\vec({\bf W}_j)|{\bf 0}, {\bf I})$. The scale of the posterior covariance should be much less than $1$. Integrating both cases, the first-order term can dominate the Taylor expansion and hence we ignore the higher order terms.
Based on \eqref{eq:1order}, we can easily calculate the first and second moments of ${\bf f}_m$,
\begin{align}
\boldsymbol{\alpha}_m &= \mathbb{E}[{\bf f}_m] \approx {\bf f}_m\left(\mathbb{E}[\boldsymbol{\Omega}_m]\right), \notag \\
{\bf V}_m &= \text{cov}[{\bf f}_m] \approx {\bf J}_m \text{cov}(\boldsymbol{\eta}_m) {\bf J}_m^\top,
\end{align}
where $\text{cov}(\boldsymbol{\eta}_m) = {\rm diag}\left(\left\{\text{cov}\big(\vec({\bf W}_j)\big)\right\}_{1\le j\le m}\right)$.
Then we use moment matching to estimate a joint Gaussian posterior, $q({\bf f}_m) = {\bf N}({\bf f}_m|\boldsymbol{\alpha}_m, {\bf V}_m)$. Next, according to our model definition \eqref{eq:our-model}, we can obtain the posterior of the ouptut ${\bf y}_m$,
\begin{align}
q({\bf y}_m) = {\bf N}({\bf y}_m|{\bf A}_m \boldsymbol{\alpha}_m, {\bf A}_m {\bf V}_m {\bf A}_m^\top + \sigma_m^2{\bf I}).
\end{align}
The output entropy is $\mathbb{H}({\bf y}_m|\mathcal{D}) = \frac{1}{2}\log|{\bf A}_m {\bf V}_m {\bf A}_m^\top + \sigma_m^2{\bf I}| + d_m \log\sqrt{2\pi e}$. However, directly computing the log determinant of a $d_m \times d_m$ matrix is very expensive. We further use the Weinstein-Aronszajn identity~\citep{kato2013perturbation} to derive
\begin{align}
&\mathbb{H}({\bf y}_m|\mathcal{D}) = \frac{1}{2}\log|\sigma_m^{-2}{\bf A}_m {\bf V}_m {\bf A}_m^\top + {\bf I}| + d_m \log \sqrt{2\pi e\sigma_m^2} \notag \\
&=\frac{1}{2}\log|\sigma_m^{-2}{\bf A}_m^\top{\bf A}_m{\bf V}_m + {\bf I}| + d_m \log \sqrt{2\pi e\sigma_m^2}. \label{eq:en_ym}
\end{align}
Now, the log determinant is calculated on a much smaller, $k_m \times k_m$ matrix, which is very cheap and efficient.
Using a similar approach, we can calculate the joint posterior of ${\widehat{\h}}_m = [{\bf f}_m; {\bf f}_M]$ and then ${\widehat{\y}}_m=[{\bf y}_m; {\bf y}_M]$. First, we view ${\widehat{\h}}_m$ as a function of all the random weights, ${\mathcal{W}} = \{{\bf W}_1, \ldots, {\bf W}_M\}$. We use the multivariate Delta's method and moment matching to estimate a joint Gaussian posterior,
$q({\widehat{\h}}_m) = {\bf N}({\widehat{\h}}_m|{\widehat{\boldsymbol{\alpha}}}_m, {\widehat{{\bf V}}}_m)$ where ${\widehat{\boldsymbol{\alpha}}}_m = {\widehat{\h}}_m\left(\mathbb{E}[{\mathcal{W}}]\right)$, ${\widehat{{\bf V}}}_m = {\widehat{\J}}_m \text{cov}(\boldsymbol{\eta}){\widehat{\J}}_m^\top$, $\boldsymbol{\eta} = \vec({\mathcal{W}})$, ${\widehat{\J}}_m = \frac{\partial {\widehat{\h}}_m}{\partial \boldsymbol{\eta}}|_{\boldsymbol{\eta} = \mathbb{E}[\boldsymbol{\eta}]}$, and $\text{cov}(\boldsymbol{\eta}) = {\rm diag}\left(\left\{\text{cov}\big(\vec({\bf W}_j)\big)\right\}_{1\le j\le M}\right)$. According to our model, we can represent
\[
{\widehat{\y}}_m = {\widehat{\A}}_m {\widehat{\h}}_m + {\widehat{\bepsilon}}_m,
\]
where ${\widehat{\A}}_m = {\rm diag}({\bf A}_m, {\bf A}_M)$ and ${\widehat{\bepsilon}}_m = [\boldsymbol{\epsilon}_m; \boldsymbol{\epsilon}_M]$. Therefore, the joint posterior of ${\widehat{\y}}_m$ is
\begin{align}
q({\widehat{\y}}_m) = {\bf N}({\widehat{\y}}_m|{\widehat{\A}}_m {\widehat{\boldsymbol{\alpha}}}_m, {\widehat{\A}}_m {\widehat{{\bf V}}}_m {\widehat{\A}}_m^\top + \S_m),
\end{align}
where $\S_m = {\rm diag}(\sigma^2_m{\bf I}_{d_m}, \sigma^2_M {\bf I}_{d_M})$, ${\bf I}_{d_m}$ and ${\bf I}_{d_M}$ are identity matrices of $d_m \times d_m$ and $d_M \times d_M$, respectively. Again, we use Weinstein-Aronszajn identity to simplify the entropy computation of ${\widehat{\y}}_m$ (\textit{i}.\textit{e}.,\xspace ${\bf y}_m$ and ${\bf y}_M$),
\begin{align}
&\mathbb{H}({\bf y}_m, {\bf y}_M|\mathcal{D}) =\frac{1}{2}\log|\S_m^{-1}{\widehat{\A}}_m {\widehat{{\bf V}}}_m {\widehat{\A}}_m^\top + {\bf I}| + \xi_m \notag \\
&=\frac{1}{2}\log|{\widehat{\A}}_m^\top \S_m^{-1} {\widehat{\A}}_m {\widehat{{\bf V}}}_m + {\bf I}| + \xi_m, \label{eq:en_ymM}
\end{align}
where $\xi_m = d_m \log \sqrt{2\pi e\sigma_m^2} + d_M \log \sqrt{2\pi e\sigma_M^2}$ and the log determinant is computed from a $(k_m + k_M) \times (k_m + k_M)$ matrix, which is cheap and efficient. Note that we can use matrix blocks to compute ${\widehat{\A}}_m^\top \S_m^{-1} {\widehat{\A}}_m = {\rm diag}\left(\sigma_m^{-2}{\bf A}_m^\top {\bf A}_m, \sigma_M^{-2}{\bf A}_M^\top {\bf A}_M\right)$.
Now, based on \eqref{eq:en_ym} and \eqref{eq:en_ymM}, we can calculate our acquisition function \eqref{eq:ac} in an analytic and deterministic way. For each fidelity $m$, we maximize $a({\bf x}, m)$ w.r.t to ${\bf x}$ to find the optimal input. We can use automatic differential libraries to calculate the gradient and feed it to any optimization algorithm, \textit{e}.\textit{g}.\xspace L-BFGS. We then use the optimal input at the fidelity that has the largest acquisition function value to query the next training example. Our deep multi-fidelity active learning is summarized in Algorithm \ref{alg:dmf-al}.
\setlength{\textfloatsep}{0.05cm}
\begin{algorithm}
\small
\caption{{\textsc{DiTucker}}\xspace($\mathcal{D}$, $T$, $\{\lambda_m\}_{m=1}^M$ )}
\label{alg:dmf-al}
\begin{algorithmic}[1]
\STATE Train the deep multi-fidelity model \eqref{eq:joint-prob} on the initial dataset $\mathcal{D}$ with stochastic structural variational learning.
\FOR {$t=1,\ldots, T$}
\STATE Based on \eqref{eq:en_ym} and \eqref{eq:en_ymM}, calculate and optimize the acquisition function \eqref{eq:ac} to find
\begin{align}
({\bf x}_{t}, m_{t}) = \operatornamewithlimits{argmax}_{{\bf x} \in {\mathcal{X}}, 1\le m \le M} a({\bf x}, m).\notag
\end{align}
\STATE Query the output ${\bf y}_{t}$ with fidelity $m_t$.
\STATE $\mathcal{D} \leftarrow \mathcal{D} \cup \{({\bf x}_{t}, {\bf y}_{t})|m_t\}$.
\STATE Re-train the deep multi-fidelity model on $\mathcal{D}$.
\ENDFOR
\end{algorithmic}
\end{algorithm}
\setlength{\floatsep}{0.1cm}
\vspace{-0.05in}
\subsection{Algorithm Complexity}
\vspace{-0.05in}
The time complexity of training our deep multi-fidelity model is ${\mathcal{O}}(N(\sum_{m=1}^M (k_ml_m)^2 + F))$ where $N$ and $F$ are the total number of training examples and NN parameters, respectively. The space complexity is ${\mathcal{O}}(F + \sum_{m=1}^M (k_ml_m)^2)$, which is to store the NN parameters, and posterior mean and covariance of each random weight matrix ${\bf W}_m$. The time complexity of calculating the acquisition function is ${\mathcal{O}}(\sum_{m=1}^M d_mk_m^2 + k_m^3)$. Since $k_m \ll d_m$, the complexity is linear to the output dimensions. Due to the usage of the learned variational posterior $q({\mathcal{W}})$, the space complexity of computing the acquisition function is the same as that of training the multi-fidelity model.
\section{Background}
\textbf{Problem setting.} We assume that we can query training examples with $M$ fidelities, which correspond to $M$ mappings, $\{{\bf f}_m({\bf x}) \in \mathds{R}^{d_m}\}_{1 \le m \le M}$ where ${\bf x} \in {\mathcal{X}}$ is an $r$-dimensional input, $r$ is small, $d_m$ is the output dimension, often very large, and $\{d_m\}$ are not necessarily identical. In general, we assume $d_1 \le \ldots \le d_M$. For instance, in structure design, the input can be the design parameters. The high fidelity examples correspond to high-resolution structures while the low fidelity ones low-resolution structures. We denote by $\lambda_m$ the cost to query with fidelity $m$. To perform active learning, we begin with a small set of training examples with mixed fidelities. Each step, we select both the fidelity and input location to query new training examples, so as to best balance the learning improvement and computational cost, namely, maximizing the benefit-cost ratio.
\textbf{Dropout active learning.} A very successful application of active learning with deep neural networks is image classification~\citep{gal2017deep,kirsch2019batchbald}. Typically, a pool of unlabeled input examples, \textit{i}.\textit{e}.,\xspace images, is pre-collected. Each time, we rank the input examples according to an acquisition function computed based on the current model. We then query the labels of the top examples, and add them into the training set. There are two popular acquisition functions. The first one is the predictive entropy of the label,
\begin{align}
\mathbb{H}[y|{\bf x}, \mathcal{D}] = -\sum_{c} p(y=c|{\bf x}, \mathcal{D}) \log p(y=c|{\bf x}, \mathcal{D}), \label{eq:image-ac-1}
\end{align}
where ${\bf x}$ is the input, and $\mathcal{D}$ are the training data. The other one is BALD (Bayesian active learning by disagreement)~\citep{houlsby2011bayesian} --- the mutual information between the label and model parameters,
\begin{align}
\mathbb{I}[y, \boldsymbol{\theta}|{\bf x}, \mathcal{D}] = \mathbb{H}[y|{\bf x}, \mathcal{D}]- \mathbb{E}_{p(\boldsymbol{\theta}|\mathcal{D})}\left[\mathbb{H}({\bf y}|{\bf x}, \boldsymbol{\theta}, \mathcal{D})\right], \label{eq:image-ac-2}
\end{align}
where $\boldsymbol{\theta}$ are the NN weights. The computation of these acquisition functions require us to first estimate the posterior of the NN output and weights. A popular approach is to use Monte-Carlo (MC) Dropout~\citep{gal2016dropout}, which can be viewed as a variational inference method of Bayesian NNs. MC Dropout uses a probability $p_i$ to randomly drop the neurons in each layer $i$ (equivalent to zeroing out the corresponding weights) in the forward pass and performs backward pass over the remaining neurons. After training, the weights and output obtained from one such forward pass can be viewed as a sample of the approximate posterior. We can run Dropout multiple times to collect a set of posterior samples and calculate the acquisition functions by Monte-Carlo approximation.
\vspace{-0.05in}
\section{Deep Multi-Fidelity Modeling for High-Dimensional Outputs}
\vspace{-0.05in}
Despite its success, dropout active learning might be inappropriate for high-dimensional (continuous) outputs. When the output is a continuous vector, it is natural to use a multivariate Gaussian posterior, $p({\bf y}|{\bf x}, \mathcal{D}) \approx {\bf N}({\bf y}|\boldsymbol{\mu}({\bf x}), \boldsymbol{\Sigma}({\bf x}))$, and the entropy $\mathbb{H}({\bf y}|{\bf x}, \mathcal{D})$ in the acquisition functions (see \eqref{eq:image-ac-1} and \eqref{eq:image-ac-2}) require us to compute the log determinant of the covariance matrix, $\log|\boldsymbol{\Sigma}({\bf x})|$. While we can use dropout samples to construct an empirical estimate $\widehat{\boldsymbol{\Sigma}}({\bf x})$, due to the large output dimension $n$, the computation of this $n\times n$ matrix and its log determinant is very expensive or even infeasible.
One might seek to only estimate the variance of each individual output. However, this will ignore the strong, complex output correlations, which is critical to effectively evaluate the uncertainty and calculate the acquisition function. Furthermore, in our active learning task, we want to optimize the acquisition function to find the best input (rather than rank the inputs in a pre-collected pool). The empirical covariance estimation from random samples can lead to numerical instability in optimization and inconsistent results from different runs.
To overcome these problems for multi-fidelity active learning, we propose a highly expressive deep multi-fidelity model, for which we develop a structural variational inference method to capture the posterior dependency of the outputs. We propose a novel acquisition function to allow multi-fidelity queries, and develop an efficient, tractable, and reliable approach to calculate the acquisition function.
Specifically, we use deep neural networks to build a multi-fidelity high-dimensional output model (see Fig. \ref{fig:graphical}) that can flexibly, efficiently capture various complex relationships between the outputs and between the fidelities, taking advantage of these relationships to enhance the predictive performance at the highest fidelity (where we make final predictions). For each fidelity $m$, we introduce a neural network (NN) that first generates a $k_m$ dimensional latent output, and then projects it to the $d_m$ dimensional observed vector, where $k_m \ll d_m$. The NN is parameterized by $\{{\bf W}_m, \boldsymbol{\theta}_m, {\bf A}_m\}$ where ${\bf W}_m$ is a $k_m \times l_m$ weight matrix in the latent output layer, ${\bf A}_m$ is a $d_m \times k_m$ projection matrix, and $\boldsymbol{\theta}_m$ the weights in all the other layers. Denote by ${\bf x}_m$ the NN input, by ${\bf f}_m({\bf x})$ the latent NN output, and by ${\bf y}_m({\bf x})$ the observation. The model is defined by
\begin{align}
{\bf x}_m &= [{\bf x}; {\bf f}_{m-1}({\bf x})], \notag \\
{\bf f}_m({\bf x}) &= {\bf W}_m\boldsymbol{\phi}_{\boldsymbol{\theta}_m}({\bf x}_m),\notag \\
{\bf y}_m({\bf x}) &= {\bf A}_m {\bf f}_m({\bf x}) + \boldsymbol{\epsilon}_m, \label{eq:our-model}
\end{align}
where $\boldsymbol{\phi}_{\boldsymbol{\theta}_m}$ is the $l_m$ dimensional input to the latent output layer (therefore parameterized by $\boldsymbol{\theta}_m$), and $\boldsymbol{\epsilon}_m \sim {\bf N}(\boldsymbol{\epsilon}_m|{\bf 0}, \sigma^2_m{\bf I})$ is an isotropic Gaussian noise. We set $k_0 = 0$ and so ${\bf f}_0({\bf x})$ is empty; when $m=1$, we have ${\bf x}_m = {\bf x}$. Note that $\boldsymbol{\phi}_{\boldsymbol{\theta}_m}$ can be viewed as a set of nonlinear basis functions. Through their combinations via ${\bf W}_m$ and ${\bf A}_m$, we can flexibly capture the complex, strong relationships between the elements of ${\bf y}_m$ to improve the prediction. Furthermore, the input ${\bf x}_{m}$ is constructed by appending to the original input ${\bf x}$ the latent output ${\bf f}_{m-1}$, which can be viewed as a compact (or low-rank) summary of all the information up to fidelity $m-1$. After a series of linear and nonlinear transformations, we obtain the latent output ${\bf f}_m$ --- the summary up to fidelity $m$ --- and then generate the high dimensional observation ${\bf y}_m$. In this way, we efficiently integrate the information from lower fidelities and capture the complex relationship between the current and previous fidelities by learning a nonlinear mapping, ${\bf f}_m({\bf x}) = h({\bf x}, {\bf f}_{m-1}({\bf x}))$, where $h(\cdot)$ is fulfilled by an NN.
\begin{figure}
\centering
\includegraphics[width=0.4\textwidth]{./graphic}
\caption{\small Graphical representation of the deep multi-fidelity model. The low dimensional latent output in each fidelity ${\bf f}_m({\bf x})$ ($1\le m \le M$) is generated by a (deep) neural network.} \label{fig:graphical}
\vspace{-0.15in}
\end{figure}
We place a standard Gaussian prior over the elements in each ${\bf W}_m$. Similar to~\citep{snoek2015scalable}, we consider all the remaining parameters as hyper-parameters to ease the posterior inference and uncertainty reasoning. Given the training dataset $\mathcal{D} = \{\{({\bf x}_{nm}, {\bf y}_{nm})|m\}_{n=1}^{N_m}\}_m$, the joint probability of our model is given by
\begin{align}
&p({\mathcal{W}}, {\mathcal{Y}}|{\mathcal{X}}, \Theta, {\bf s}) = \prod\nolimits_{m=1}^M {\bf N}\left(\vec({\bf W}_m)|{\bf 0}, {\bf I}\right)\notag \\
&\cdot \prod\nolimits_{n=1}^{N_m} {\bf N}\big({\bf y}_{nm}|{\bf A}_m {\bf f}_m({\bf x}_{nm}), \sigma_m^2{\bf I}\big), \label{eq:joint-prob}
\end{align}
where ${\mathcal{W}} = \{{\bf W}_m\}_{1\le m\le M}$, $\Theta=\{\boldsymbol{\theta}_m, {\bf A}_m\}_{1\le m \le M}$, ${\bf s} = [\sigma^2_1, \ldots, \sigma^2_M]^\top$, and $\{{\mathcal{X}}$, ${\mathcal{Y}}\}$ are the inputs and outputs in $\mathcal{D}$.
To estimate the posterior of our model (which is used to calculate the acquisition function and query new examples), we develop a stochastic structural variational learning algorithm. Specifically, for each ${\bf W}_m$, we introduce a multivariate Gaussian posterior, $q({\bf W}_m) = {\bf N}\left(\vec({\bf W}_m)|\boldsymbol{\mu}_m, \boldsymbol{\Sigma}_m\right)$. To ensure the positive definiteness, we parameterize $\boldsymbol{\Sigma}_m$ by its Cholesky decomposition, $\boldsymbol{\Sigma}_m = \L_m \L_m^\top$, where $\L_m$ is a lower triangular matrix. We then assume the posterior $q({\mathcal{W}}) =\prod_{m=1}^M q({\bf W}_m)$, and construct a variational model evidence lower bound (ELBO)~\citep{wainwright2008graphical},
\begin{align}
&{\mathcal{L}}\big(q({\mathcal{W}}), \Theta, {\bf s}\big) = \mathbb{E}_{q}\left[ \log\frac{p({\mathcal{W}}, {\mathcal{Y}}|{\mathcal{X}}, \Theta, {\bf s})}{q({\mathcal{W}})}\right] \notag \\
&=-\text{KL}\big(q({\mathcal{W}})\|p({\mathcal{W}})\big) \notag \\
&+ \sum_{m=1}^M\sum_{n=1}^{N_m} \mathbb{E}_q\left[\log {\bf N}({\bf y}_{nm}|{\bf A}_m {\bf f}_m({\bf x}_{nm}),\sigma_m^2{\bf I})\right], \label{eq:elbo}
\end{align}
where $\text{KL}(\cdot \| \cdot)$ is Kullback Leibler divergence and $p({\mathcal{W}})$ the prior of ${\mathcal{W}}$. We maximize ${\mathcal{L}}$ to jointly estimate $q({\mathcal{W}})$ and the hyper-parameters. While ${\mathcal{L}}$ is intractable, it is straightforward to use the reparameterization trick~\citep{kingma2013auto} to generate parameterized samples for each $\vec({\bf W}_m)$: $\boldsymbol{\mu}_m + \L_m \boldsymbol{\xi}_m$ where $\boldsymbol{\xi}_m \sim {\bf N}({\bf 0}, {\bf I})$, substitute them into \eqref{eq:elbo} to obtain a stochastic estimate of ${\mathcal{L}}$, and perform efficient stochastic optimization.
\vspace{-0.1in}
\section{Multi-Fidelity Active Learning}
\vspace{-0.1in}
\subsection{Mutual Information based Acquisition Function}
\vspace{-0.1in}
We now consider how to perform active learning with multi-fidelity queries. We assume that at each fidelity $m$, the most valuable training example is the one that can best improve our final prediction, namely, the prediction at the highest fidelity $M$.
To this end, we propose our acquisition function based on the mutual information between the outputs at fidelity $m$ and $M$,
\begin{align}
&a({\bf x}, m) = \frac{1}{\lambda_m} \mathbb{I}\left({\bf y}_m({\bf x}), {\bf y}_M({\bf x})|\mathcal{D}\right) \label{eq:ac} \\
&= \frac{1}{\lambda_m} \big(\mathbb{H}({\bf y}_m|\mathcal{D}) +\mathbb{H}({\bf y}_M|\mathcal{D}) - \mathbb{H}({\bf y}_m, {\bf y}_M|\mathcal{D}) \big), \notag
\end{align}
where $\lambda_m>0$ is the cost of querying a training example with fidelity $m$. When $m=M$, we have $a({\bf x}, M) = \frac{1}{\lambda_M} \mathbb{H}\left({\bf y}_M({\bf x})|\mathcal{D}\right)$ --- the output entropy. Therefore, our acquisition function is an extension of the popular predictive entropy principle in conventional active learning. At each step, we maximize our acquisition function to identify a pair of fidelity and input location that give the biggest benefit-cost ratio.
\vspace{-0.1in}
\subsection{Efficient Acquisition Function Calculation}\label{sect:ac-compute}
\vspace{-0.1in}
To maximize the acquisition function \eqref{eq:ac}, a critical challenge is to compute the posterior of the outputs $\{{\bf y}_m\}$ in every fidelity, based on our model estimation results, $p({\mathcal{W}}|\mathcal{D}) \approx q({\mathcal{W}})$. Due to high dimensionality of each ${\bf y}_m$ and the nonlinear coupling of the latent outputs across the fidelities (see \eqref{eq:our-model}), the computation is challenging and analytically intractable. To address this issue, we first consider approximating the posterior of the low dimensional latent output ${\bf f}_m$, which can be viewed a function of the random NN weights $\boldsymbol{\Omega}_m = \{{\bf W}_1, \ldots, {\bf W}_m\}$ (given the input ${\bf x}$). We use multivariate Delta's method~\citep{oehlert1992note,bickel2015mathematical} to compute the moments of ${\bf f}_m$. Specifically, we approximate ${\bf f}_m$ with a first-order Taylor expansion,
\begin{align}
{\bf f}_m(\boldsymbol{\Omega}_m) \approx {\bf f}_m(\mathbb{E}[\boldsymbol{\Omega}_m]) + {\bf J}_m\left(\boldsymbol{\eta}_m - \mathbb{E}[\boldsymbol{\eta}_m]\right), \label{eq:1order}
\end{align}
where the expectation is under $q(\cdot)$, $\boldsymbol{\eta}_m = \vec(\boldsymbol{\Omega}_m)$, ${\bf J}_m = \frac{\partial {\bf f}_m}{\partial \boldsymbol{\eta}_m}|_{\boldsymbol{\eta}_m = \mathbb{E}[\boldsymbol{\eta}_m]}$ is the Jacobian matrix at the mean. The rationale is as follows. First, ${\bf f}_m$ is linear to ${\bf W}_m$, and the second-order derivative is simply ${\bf 0}$. Second, as the NN output, ${\bf f}_m$ is highly nonlinear to the random weights in previous layers, \textit{i}.\textit{e}.,\xspace ${\bf W}_j$ ($j<m$). Hence, we can assume the change rate of ${\bf f}_m$ (\textit{e}.\textit{g}.\xspace gradient or Jacobian) has much greater scales than the posterior covariance of ${\bf W}_j$ in the second-order term of the Taylor expansion. Note that $q({\bf W}_j)$ is much more informative and hence much more concentrated than the prior $ {\bf N}(\vec({\bf W}_j)|{\bf 0}, {\bf I})$. The scale of the posterior covariance should be much less than $1$. Integrating both cases, the first-order term can dominate the Taylor expansion and hence we ignore the higher order terms.
Based on \eqref{eq:1order}, we can easily calculate the first and second moments of ${\bf f}_m$,
\begin{align}
\boldsymbol{\alpha}_m &= \mathbb{E}[{\bf f}_m] \approx {\bf f}_m\left(\mathbb{E}[\boldsymbol{\Omega}_m]\right), \notag \\
{\bf V}_m &= \text{cov}[{\bf f}_m] \approx {\bf J}_m \text{cov}(\boldsymbol{\eta}_m) {\bf J}_m^\top,
\end{align}
where $\text{cov}(\boldsymbol{\eta}_m) = {\rm diag}\left(\left\{\text{cov}\big(\vec({\bf W}_j)\big)\right\}_{1\le j\le m}\right)$.
Then we use moment matching to estimate a joint Gaussian posterior, $q({\bf f}_m) = {\bf N}({\bf f}_m|\boldsymbol{\alpha}_m, {\bf V}_m)$. Next, according to our model definition \eqref{eq:our-model}, we can obtain the posterior of the ouptut ${\bf y}_m$,
\begin{align}
q({\bf y}_m) = {\bf N}({\bf y}_m|{\bf A}_m \boldsymbol{\alpha}_m, {\bf A}_m {\bf V}_m {\bf A}_m^\top + \sigma_m^2{\bf I}).
\end{align}
The output entropy is $\mathbb{H}({\bf y}_m|\mathcal{D}) = \frac{1}{2}\log|{\bf A}_m {\bf V}_m {\bf A}_m^\top + \sigma_m^2{\bf I}| + d_m \log\sqrt{2\pi e}$. However, directly computing the log determinant of a $d_m \times d_m$ matrix is very expensive. We further use the Weinstein-Aronszajn identity~\citep{kato2013perturbation} to derive
\begin{align}
&\mathbb{H}({\bf y}_m|\mathcal{D}) = \frac{1}{2}\log|\sigma_m^{-2}{\bf A}_m {\bf V}_m {\bf A}_m^\top + {\bf I}| + d_m \log \sqrt{2\pi e\sigma_m^2} \notag \\
&=\frac{1}{2}\log|\sigma_m^{-2}{\bf A}_m^\top{\bf A}_m{\bf V}_m + {\bf I}| + d_m \log \sqrt{2\pi e\sigma_m^2}. \label{eq:en_ym}
\end{align}
Now, the log determinant is calculated on a much smaller, $k_m \times k_m$ matrix, which is very cheap and efficient.
Using a similar approach, we can calculate the joint posterior of ${\widehat{\h}}_m = [{\bf f}_m; {\bf f}_M]$ and then ${\widehat{\y}}_m=[{\bf y}_m; {\bf y}_M]$. First, we view ${\widehat{\h}}_m$ as a function of all the random weights, ${\mathcal{W}} = \{{\bf W}_1, \ldots, {\bf W}_M\}$. We use the multivariate Delta's method and moment matching to estimate a joint Gaussian posterior,
$q({\widehat{\h}}_m) = {\bf N}({\widehat{\h}}_m|{\widehat{\boldsymbol{\alpha}}}_m, {\widehat{{\bf V}}}_m)$ where ${\widehat{\boldsymbol{\alpha}}}_m = {\widehat{\h}}_m\left(\mathbb{E}[{\mathcal{W}}]\right)$, ${\widehat{{\bf V}}}_m = {\widehat{\J}}_m \text{cov}(\boldsymbol{\eta}){\widehat{\J}}_m^\top$, $\boldsymbol{\eta} = \vec({\mathcal{W}})$, ${\widehat{\J}}_m = \frac{\partial {\widehat{\h}}_m}{\partial \boldsymbol{\eta}}|_{\boldsymbol{\eta} = \mathbb{E}[\boldsymbol{\eta}]}$, and $\text{cov}(\boldsymbol{\eta}) = {\rm diag}\left(\left\{\text{cov}\big(\vec({\bf W}_j)\big)\right\}_{1\le j\le M}\right)$. According to our model, we can represent
\[
{\widehat{\y}}_m = {\widehat{\A}}_m {\widehat{\h}}_m + {\widehat{\bepsilon}}_m,
\]
where ${\widehat{\A}}_m = {\rm diag}({\bf A}_m, {\bf A}_M)$ and ${\widehat{\bepsilon}}_m = [\boldsymbol{\epsilon}_m; \boldsymbol{\epsilon}_M]$. Therefore, the joint posterior of ${\widehat{\y}}_m$ is
\begin{align}
q({\widehat{\y}}_m) = {\bf N}({\widehat{\y}}_m|{\widehat{\A}}_m {\widehat{\boldsymbol{\alpha}}}_m, {\widehat{\A}}_m {\widehat{{\bf V}}}_m {\widehat{\A}}_m^\top + \S_m),
\end{align}
where $\S_m = {\rm diag}(\sigma^2_m{\bf I}_{d_m}, \sigma^2_M {\bf I}_{d_M})$, ${\bf I}_{d_m}$ and ${\bf I}_{d_M}$ are identity matrices of $d_m \times d_m$ and $d_M \times d_M$, respectively. Again, we use Weinstein-Aronszajn identity to simplify the entropy computation of ${\widehat{\y}}_m$ (\textit{i}.\textit{e}.,\xspace ${\bf y}_m$ and ${\bf y}_M$),
\begin{align}
&\mathbb{H}({\bf y}_m, {\bf y}_M|\mathcal{D}) =\frac{1}{2}\log|\S_m^{-1}{\widehat{\A}}_m {\widehat{{\bf V}}}_m {\widehat{\A}}_m^\top + {\bf I}| + \xi_m \notag \\
&=\frac{1}{2}\log|{\widehat{\A}}_m^\top \S_m^{-1} {\widehat{\A}}_m {\widehat{{\bf V}}}_m + {\bf I}| + \xi_m, \label{eq:en_ymM}
\end{align}
where $\xi_m = d_m \log \sqrt{2\pi e\sigma_m^2} + d_M \log \sqrt{2\pi e\sigma_M^2}$ and the log determinant is computed from a $(k_m + k_M) \times (k_m + k_M)$ matrix, which is cheap and efficient. Note that we can use matrix blocks to compute ${\widehat{\A}}_m^\top \S_m^{-1} {\widehat{\A}}_m = {\rm diag}\left(\sigma_m^{-2}{\bf A}_m^\top {\bf A}_m, \sigma_M^{-2}{\bf A}_M^\top {\bf A}_M\right)$.
Now, based on \eqref{eq:en_ym} and \eqref{eq:en_ymM}, we can calculate our acquisition function \eqref{eq:ac} in an analytic and deterministic way. For each fidelity $m$, we maximize $a({\bf x}, m)$ w.r.t to ${\bf x}$ to find the optimal input. We can use automatic differential libraries to calculate the gradient and feed it to any optimization algorithm, \textit{e}.\textit{g}.\xspace L-BFGS. We then use the optimal input at the fidelity that has the largest acquisition function value to query the next training example. Our deep multi-fidelity active learning is summarized in Algorithm \ref{alg:dmf-al}.
\setlength{\textfloatsep}{0.05cm}
\begin{algorithm}
\small
\caption{{\textsc{DiTucker}}\xspace($\mathcal{D}$, $T$, $\{\lambda_m\}_{m=1}^M$ )}
\label{alg:dmf-al}
\begin{algorithmic}[1]
\STATE Train the deep multi-fidelity model \eqref{eq:joint-prob} on the initial dataset $\mathcal{D}$ with stochastic structural variational learning.
\FOR {$t=1,\ldots, T$}
\STATE Based on \eqref{eq:en_ym} and \eqref{eq:en_ymM}, calculate and optimize the acquisition function \eqref{eq:ac} to find
\begin{align}
({\bf x}_{t}, m_{t}) = \operatornamewithlimits{argmax}_{{\bf x} \in {\mathcal{X}}, 1\le m \le M} a({\bf x}, m).\notag
\end{align}
\STATE Query the output ${\bf y}_{t}$ with fidelity $m_t$.
\STATE $\mathcal{D} \leftarrow \mathcal{D} \cup \{({\bf x}_{t}, {\bf y}_{t})|m_t\}$.
\STATE Re-train the deep multi-fidelity model on $\mathcal{D}$.
\ENDFOR
\end{algorithmic}
\end{algorithm}
\setlength{\floatsep}{0.1cm}
\vspace{-0.05in}
\subsection{Algorithm Complexity}
\vspace{-0.05in}
The time complexity of training our deep multi-fidelity model is ${\mathcal{O}}(N(\sum_{m=1}^M (k_ml_m)^2 + F))$ where $N$ and $F$ are the total number of training examples and NN parameters, respectively. The space complexity is ${\mathcal{O}}(F + \sum_{m=1}^M (k_ml_m)^2)$, which is to store the NN parameters, and posterior mean and covariance of each random weight matrix ${\bf W}_m$. The time complexity of calculating the acquisition function is ${\mathcal{O}}(\sum_{m=1}^M d_mk_m^2 + k_m^3)$. Since $k_m \ll d_m$, the complexity is linear to the output dimensions. Due to the usage of the learned variational posterior $q({\mathcal{W}})$, the space complexity of computing the acquisition function is the same as that of training the multi-fidelity model.
\section{Background}
\textbf{Problem setting.} We assume that we can query training examples with $M$ fidelities, which correspond to $M$ mappings, $\{{\bf y}_m({\bf x}) \in \mathds{R}^{d_m}\}_{1 \le m \le M}$ where ${\bf x} \in {\mathcal{X}}$ is an $r$-dimensional input, $r$ is small, $d_m$ is the output dimension, often very large, and $\{d_m\}$ are not necessarily identical. In general, we assume $d_1 \le \ldots \le d_M$. For instance, in structure design, the input can be the design parameters. The high fidelity examples correspond to high-resolution structures while the low fidelity ones low-resolution structures. We denote by $\lambda_m$ the cost to query with fidelity $m$. Our goal is to estimate ${\bf y}_M({\bf x})$ (\textit{i}.\textit{e}.,\xspace the target function). Among the examples at different fidelities are strong (but complex) correlations, due to the underlying common bases, say, in physics. Hence, all these examples are valuable to estimate the target function. To perform active learning, we begin with a small set of examples with mixed fidelities. Each step, we select both the fidelity and input location to query (or generate) new examples, so as to best balance the learning improvement (on ${\bf y}_M({\bf x})$) and computational cost, \textit{i}.\textit{e}.,\xspace maximizing the benefit-cost ratio.
\textbf{Dropout active learning.} A successful application of active learning with deep neural networks is image classification~\citep{gal2017deep,kirsch2019batchbald}. Typically, a pool of unlabeled input examples, \textit{i}.\textit{e}.,\xspace images, is pre-collected. Each time, we rank the input examples according to an acquisition function computed based on the current model. We then query the labels of the top examples and add them into the training set. There are two popular acquisition functions. The first one is the predictive entropy of the label,
\begin{align}
\mathbb{H}[y|{\bf x}, \mathcal{D}] = -\sum_{c} p(y=c|{\bf x}, \mathcal{D}) \log p(y=c|{\bf x}, \mathcal{D}), \label{eq:image-ac-1}
\end{align}
where ${\bf x}$ is the input, and $\mathcal{D}$ are the training data. The other one is BALD (Bayesian active learning by disagreement)~\citep{houlsby2011bayesian}, namely the mutual information between the label and model parameters,
\begin{align}
\mathbb{I}[y, \boldsymbol{\theta}|{\bf x}, \mathcal{D}] = \mathbb{H}[y|{\bf x}, \mathcal{D}]- \mathbb{E}_{p(\boldsymbol{\theta}|\mathcal{D})}\left[\mathbb{H}({\bf y}|{\bf x}, \boldsymbol{\theta}, \mathcal{D})\right], \label{eq:image-ac-2}
\end{align}
where $\boldsymbol{\theta}$ are the NN weights. To calculate these acquisition functions, we need to first estimate the posterior of the NN output and weights. A popular approach is to use Monte-Carlo (MC) Dropout~\citep{gal2016dropout}, which is essentially a variational inference method of Bayesian NNs. MC Dropout uses a probability $p_i$ to randomly drop the neurons in each layer $i$ (equivalent to zeroing out the corresponding weights) in the forward pass and performs backward pass over the remaining neurons. After training, the weights and output obtained from one such forward pass can be viewed as a sample of the approximate posterior. We can run Dropout multiple times to collect a set of posterior samples and calculate the acquisition functions by Monte-Carlo approximations.
\vspace{-0.05in}
\section{Deep Multi-Fidelity Modeling for High-Dimensional Outputs}
\vspace{-0.05in}
Despite its success, dropout active learning might be inappropriate for high-dimensional (continuous) outputs. When the output is a continuous vector, it is natural to use a multivariate Gaussian posterior, $p({\bf y}|{\bf x}, \mathcal{D}) \approx {\bf N}({\bf y}|\boldsymbol{\mu}({\bf x}), \boldsymbol{\Sigma}({\bf x}))$\footnote{we can also use a multivariate student $t$ distribution, but the problem remains. }, and the entropy term $\mathbb{H}({\bf y}|{\bf x}, \mathcal{D})$ in the acquisition functions (see \eqref{eq:image-ac-1} and \eqref{eq:image-ac-2}) require us to compute the log determinant of the covariance matrix, $\log|\boldsymbol{\Sigma}({\bf x})|$. While we can use dropout samples to construct an empirical estimate $\widehat{\boldsymbol{\Sigma}}({\bf x})$, due to the large output dimension $n$, the computation of this $n\times n$ matrix and its log determinant is very expensive or even infeasible.
One might seek to only estimate the predictive variance of each individual output. However, this will ignore the strong output correlations, which is critical to effectively evaluate the uncertainty and calculate the acquisition function. Furthermore, in our active learning task, we want to optimize the acquisition function to find the best input (rather than rank the inputs in a pre-collected pool). The empirical covariance estimate from random samples can lead to numerical instability in optimization and inconsistent results from different runs.
To overcome these problems for multi-fidelity active learning, we first propose an expressive deep multi-fidelity model, for which we develop a structural variational inference method to capture the posterior dependency of the outputs. We propose a novel acquisition function to allow multi-fidelity queries, and develop an efficient and reliable approach to calculate the acquisition function.
Specifically, we use deep neural networks to build a multi-fidelity high-dimensional output model (see the graphical representation in Fig. 2 of the Appendix) that can flexibly, efficiently capture complex relationships between the outputs and between the fidelities, taking advantage of these relationships to enhance the learning at the highest fidelity (\textit{i}.\textit{e}.,\xspace target function). For each fidelity $m$, we introduce a neural network (NN) that first generates a $k_m$ dimensional latent output, and then projects it to the $d_m$ dimensional observational space, where $k_m \ll d_m$. The NN is parameterized by $\{{\bf W}_m, \boldsymbol{\theta}_m, {\bf A}_m\}$ where ${\bf W}_m$ is a $k_m \times l_m$ weight matrix in the latent output layer, ${\bf A}_m$ a $d_m \times k_m$ projection matrix, and $\boldsymbol{\theta}_m$ the weights in all the other layers. Denote by $\boldsymbol{\xi}_m$ the NN input, by ${\bf h}_m({\bf x})$ the latent NN output, and by ${\bf y}_m({\bf x})$ the observed output. The model is defined by
\begin{align}
&\boldsymbol{\xi}_m = [{\bf x}; {\bf h}_{m-1}({\bf x})], \;\; {\bf h}_m({\bf x}) = {\bf W}_m\boldsymbol{\rho}_{\boldsymbol{\theta}_m}(\boldsymbol{\xi}_m), \notag \\
&{\bf y}_m({\bf x}) = {\bf A}_m {\bf h}_m({\bf x}) + \boldsymbol{\epsilon}_m, \label{eq:our-model}
\end{align}
where $\boldsymbol{\xi}_1 = {\bf x}$, $\boldsymbol{\rho}_{\boldsymbol{\theta}_m}$ is the second last layer of the NN at fidelity $m$, of dimension $l_m$, and $\boldsymbol{\epsilon}_m \sim {\bf N}(\boldsymbol{\epsilon}_m|{\bf 0}, \sigma^2_m{\bf I})$ is an isotropic Gaussian noise. Note that $\boldsymbol{\rho}_{\boldsymbol{\theta}_m}$ can be considered as $l_m$ nonlinear basis functions parameterized by $\boldsymbol{\theta}_m$. Through their combinations via ${\bf W}_m$ and ${\bf A}_m$, we can flexibly capture the complex relationships between the elements of ${\bf y}_m$ to improve the prediction. Furthermore, the input $\boldsymbol{\xi}_{m}$ is the concatenation of the original input ${\bf x}$ and the latent output ${\bf h}_{m-1}$. The latter can be viewed as a compact (or low-rank) summary of all the information up to fidelity $m-1$. Through an NN mapping, we obtain the latent output ${\bf h}_m = \text{NN}({\bf x}, {\bf h}_{m-1}({\bf x}))$ --- the summary up to fidelity $m$ --- and then generate the high dimensional output ${\bf y}_m$. Thereby, we are able to efficiently integrate the information from lower fidelities and grasp the strong, complex relationship between the current and prior fidelities.
We place a standard Gaussian prior over the elements in each ${\bf W}_m$. Similar to~\citep{snoek2015scalable}, we consider all the remaining parameters as hyper-parameters to ease the posterior inference and uncertainty reasoning. Given the training dataset $\mathcal{D} = \{\{({\bf x}_{nm}, {\bf y}_{nm})\}_{n=1}^{N_m}\}_{m=1}^M$, the joint probability of our model is given by
\begin{align}
&p({\mathcal{W}}, {\mathcal{Y}}|{\mathcal{X}}, \Theta, {\bf s}) = \prod\nolimits_{m=1}^M {\bf N}\left(\vec({\bf W}_m)|{\bf 0}, {\bf I}\right) \notag \\
&\cdot \prod\nolimits_{n=1}^{N_m} {\bf N}\big({\bf y}_{nm}|{\bf A}_m {\bf h}_m({\bf x}_{nm}), \sigma_m^2{\bf I}\big), \label{eq:joint-prob}
\end{align}
where ${\mathcal{W}} = \{{\bf W}_m\}_{m=1}^M$, $\Theta=\{\boldsymbol{\theta}_m, {\bf A}_m\}_{m=1}^M$, ${\bf s} = [\sigma^2_1, \ldots, \sigma^2_M]$, and $\{{\mathcal{X}}$, ${\mathcal{Y}}\}$ are the inputs and outputs in $\mathcal{D}$.
To estimate the posterior of our model (which is used to calculate the acquisition function and query new examples), we develop a stochastic structural variational learning algorithm. Specifically, for each ${\bf W}_m$, we introduce a multivariate Gaussian variational posterior, $q({\bf W}_m) = {\bf N}\left(\vec({\bf W}_m)|\boldsymbol{\mu}_m, \boldsymbol{\Sigma}_m\right)$. To ensure the positive definiteness, we parameterize $\boldsymbol{\Sigma}_m$ by its Cholesky decomposition, $\boldsymbol{\Sigma}_m = \L_m \L_m^\top$, where $\L_m$ is a lower triangular matrix. We then assume the posterior $q({\mathcal{W}}) =\prod_{m=1}^M q({\bf W}_m)$, and construct a variational model evidence lower bound (ELBO)~\citep{wainwright2008graphical}, ${\mathcal{L}}\big(q({\mathcal{W}}), \Theta, {\bf s}\big) = \mathbb{E}_{q}\left[ \log{p({\mathcal{W}}, {\mathcal{Y}}|{\mathcal{X}}, \Theta, {\bf s})}/{q({\mathcal{W}})}\right] =-\text{KL}\big(q({\mathcal{W}})\|p({\mathcal{W}})\big) + \sum_{m=1}^M\sum_{n=1}^{N_m} \mathbb{E}_q\left[\log {\bf N}({\bf y}_{nm}|{\bf A}_m {\bf h}_m({\bf x}_{nm}),\sigma_m^2{\bf I})\right]$, where $\text{KL}(\cdot \| \cdot)$ is Kullback Leibler divergence and $p({\mathcal{W}})$ the prior of ${\mathcal{W}}$. We maximize ${\mathcal{L}}$ to jointly estimate $q({\mathcal{W}})$ and the hyper-parameters. While ${\mathcal{L}}$ is intractable, we use the reparameterization trick~\citep{kingma2013auto} to generate parameterized samples for each $\vec({\bf W}_m)$: $\boldsymbol{\mu}_m + \L_m \boldsymbol{\eta}_m$ where $\boldsymbol{\eta}_m \sim {\bf N}({\bf 0}, {\bf I})$ to obtain a stochastic estimate of ${\mathcal{L}}$, and conduct stochastic optimization.
\section{Multi-Fidelity Active Learning}
\subsection{Mutual Information based Acquisition Function}
We now consider how to perform active learning with multi-fidelity queries. We assume that at each fidelity $m$, the most valuable training example is the one that can best improve our prediction of the target function, namely, the prediction at the highest fidelity $M$.
To this end, we propose our acquisition function based on the mutual information between the outputs at fidelity $m$ and $M$,
\begin{align}
&a({\bf x}, m) = \frac{1}{\lambda_m} \mathbb{I}\left({\bf y}_m({\bf x}), {\bf y}_M({\bf x})|\mathcal{D}\right) \notag \\
&= \frac{1}{\lambda_m} \big(\mathbb{H}({\bf y}_m|\mathcal{D}) +\mathbb{H}({\bf y}_M|\mathcal{D}) - \mathbb{H}({\bf y}_m, {\bf y}_M|\mathcal{D}) \big), \label{eq:ac}
\end{align}
where $\lambda_m>0$ is the cost of querying a training example with fidelity $m$. When $m=M$, we have $a({\bf x}, M) = \frac{1}{\lambda_M} \mathbb{H}\left({\bf y}_M({\bf x})|\mathcal{D}\right)$ --- the output entropy. Therefore, our acquisition function is an extension of the popular predictive entropy principle in conventional active learning. At each step, we maximize our acquisition function to identify a pair of fidelity and input location that give the biggest benefit-cost ratio.
\subsection{Efficient Acquisition Function Calculation}\label{sect:ac-compute}
To maximize the acquisition function \eqref{eq:ac}, a critical challenge is to compute the posterior of the outputs $\{{\bf y}_m\}$ in every fidelity, based on our model estimation results, \textit{i}.\textit{e}.,\xspace $p({\mathcal{W}}|\mathcal{D}) \approx q({\mathcal{W}})$. Due to the high dimensionality of each ${\bf y}_m$ and the nonlinear coupling of the latent outputs across the fidelities (see \eqref{eq:our-model}), the computation is challenging and analytically intractable. To address this issue, we consider approximating the posterior of the low dimensional latent output ${\bf h}_m$ first, which can be viewed as a function of the random NN weights $\boldsymbol{\Omega}_m = \{{\bf W}_1, \ldots, {\bf W}_m\}$ (given the input ${\bf x}$). We then use multivariate delta method~\citep{oehlert1992note,bickel2015mathematical} to compute the moments of ${\bf h}_m$. Specifically, we approximate ${\bf h}_m$ with a first-order Taylor expansion,
\begin{align}
{\bf h}_m(\boldsymbol{\Omega}_m) \approx {\bf h}_m(\mathbb{E}[\boldsymbol{\Omega}_m]) + {\bf J}_m\left(\boldsymbol{\eta}_m - \mathbb{E}[\boldsymbol{\eta}_m]\right), \label{eq:1order}
\end{align}
where the expectation is under $q(\cdot)$, $\boldsymbol{\eta}_m = \vec(\boldsymbol{\Omega}_m)$, ${\bf J}_m = \frac{\partial {\bf h}_m}{\partial \boldsymbol{\eta}_m}|_{\boldsymbol{\eta}_m = \mathbb{E}[\boldsymbol{\eta}_m]}$ is the Jacobian matrix at the mean. The rationale is as follows. First, ${\bf h}_m$ is linear to ${\bf W}_m$ and the second-order derivative is simply ${\bf 0}$. Second, as the NN output, ${\bf h}_m$ is highly nonlinear to the random weights in the previous layers, ${\bf W}_j$ ($j<m$). Hence, we can assume the change rate of ${\bf h}_m$ (\textit{i}.\textit{e}.,\xspace gradient or Jacobian) has much greater scales than the posterior covariance of ${\bf W}_j$ in the second-order term of the Taylor expansion. Note that $q({\bf W}_j)$ is more informative and hence much more concentrated than the prior $ {\bf N}(\vec({\bf W}_j)|{\bf 0}, {\bf I})$. The scale of the posterior covariance should be much smaller than $1$. Considering both cases, the first-order term can dominate the Taylor expansion and hence we ignore the higher order terms. Our ablation study has confirmed our analysis; see Sec. 4 of Appendix for details.
Based on \eqref{eq:1order}, we can easily calculate the first and second moments of ${\bf h}_m$,
\begin{align}
\boldsymbol{\alpha}_m &= \mathbb{E}[{\bf h}_m] \approx {\bf h}_m\left(\mathbb{E}[\boldsymbol{\Omega}_m]\right), \\
{\bf V}_m &= \text{cov}[{\bf h}_m] \approx {\bf J}_m \text{cov}(\boldsymbol{\eta}_m) {\bf J}_m^\top,
\end{align}
where $\text{cov}(\boldsymbol{\eta}_m) = {\rm diag}\left(\left\{\text{cov}\big(\vec({\bf W}_j)\big)\right\}_{1\le j\le m}\right)$.
Then we use moment matching to estimate a joint Gaussian posterior, $q({\bf h}_m) = {\bf N}({\bf h}_m|\boldsymbol{\alpha}_m, {\bf V}_m)$. Next, according to \eqref{eq:our-model}, we can obtain the posterior of the output ${\bf y}_m$,
\begin{align}
q({\bf y}_m) = {\bf N}({\bf y}_m|{\bf A}_m \boldsymbol{\alpha}_m, {\bf A}_m {\bf V}_m {\bf A}_m^\top + \sigma_m^2{\bf I}).
\end{align}
The output entropy is $\mathbb{H}({\bf y}_m|\mathcal{D}) = \frac{1}{2}\log|{\bf A}_m {\bf V}_m {\bf A}_m^\top + \sigma_m^2{\bf I}| + d_m \log\sqrt{2\pi e}$. However, directly computing the log determinant of a $d_m \times d_m$ matrix is very expensive. To bypass this challenge, we use the Weinstein-Aronszajn identity~\citep{kato2013perturbation} to derive
\begin{align}
&\mathbb{H}({\bf y}_m|\mathcal{D}) = \frac{1}{2}\log|\sigma_m^{-2}{\bf A}_m {\bf V}_m {\bf A}_m^\top + {\bf I}| + \log \left({2\pi e\sigma_m^2}\right)^{\frac{d}{2}} \notag \\
&=\frac{1}{2}\log|\sigma_m^{-2}{\bf A}_m^\top{\bf A}_m{\bf V}_m + {\bf I}| + d_m \log \sqrt{2\pi e\sigma_m^2}. \label{eq:en_ym}
\end{align}
Now, the log determinant is calculated on a much smaller, $k_m \times k_m$ matrix, which is very cheap and efficient.
Using a similar approach, we can calculate the joint posterior of ${\widehat{\h}}_m = [{\bf h}_m; {\bf h}_M]$ and then ${\widehat{\y}}_m=[{\bf y}_m; {\bf y}_M]$. First, we view ${\widehat{\h}}_m$ as a function of all the random weights, ${\mathcal{W}} = \{{\bf W}_1, \ldots, {\bf W}_M\}$. We use the multivariate delta method and moment matching to estimate a joint Gaussian posterior,
$q({\widehat{\h}}_m) = {\bf N}({\widehat{\h}}_m|{\widehat{\boldsymbol{\alpha}}}_m, {\widehat{{\bf V}}}_m)$ where ${\widehat{\boldsymbol{\alpha}}}_m = {\widehat{\h}}_m\left(\mathbb{E}[{\mathcal{W}}]\right)$, ${\widehat{{\bf V}}}_m = {\widehat{\J}}_m \text{cov}(\boldsymbol{\eta}){\widehat{\J}}_m^\top$, $\boldsymbol{\eta} = \vec({\mathcal{W}})$, ${\widehat{\J}}_m = \frac{\partial {\widehat{\h}}_m}{\partial \boldsymbol{\eta}}|_{\boldsymbol{\eta} = \mathbb{E}[\boldsymbol{\eta}]}$, and $\text{cov}(\boldsymbol{\eta}) = {\rm diag}\left(\left\{\text{cov}\big(\vec({\bf W}_j)\big)\right\}_{1\le j\le M}\right)$. According to our model \eqref{eq:our-model}, we can represent
\[
{\widehat{\y}}_m = {\widehat{\A}}_m {\widehat{\h}}_m + {\widehat{\bepsilon}}_m,
\]
where ${\widehat{\A}}_m = {\rm diag}({\bf A}_m, {\bf A}_M)$ and ${\widehat{\bepsilon}}_m = [\boldsymbol{\epsilon}_m; \boldsymbol{\epsilon}_M]$. Therefore, the joint posterior of ${\widehat{\y}}_m$ is
\begin{align}
q({\widehat{\y}}_m) = {\bf N}({\widehat{\y}}_m|{\widehat{\A}}_m {\widehat{\boldsymbol{\alpha}}}_m, {\widehat{\A}}_m {\widehat{{\bf V}}}_m {\widehat{\A}}_m^\top + \S_m),
\end{align}
where $\S_m = {\rm diag}(\sigma^2_m{\bf I}_{d_m}, \sigma^2_M {\bf I}_{d_M})$, ${\bf I}_{d_m}$ and ${\bf I}_{d_M}$ are identity matrices of $d_m \times d_m$ and $d_M \times d_M$, respectively. Again, we use Weinstein-Aronszajn identity to simplify the entropy computation of ${\widehat{\y}}_m$ (\textit{i}.\textit{e}.,\xspace ${\bf y}_m$ and ${\bf y}_M$),
\begin{align}
&\mathbb{H}({\bf y}_m, {\bf y}_M|\mathcal{D}) =\frac{1}{2}\log|\S_m^{-1}{\widehat{\A}}_m {\widehat{{\bf V}}}_m {\widehat{\A}}_m^\top + {\bf I}| + \beta_m \notag \\
&=\frac{1}{2}\log|{\widehat{\A}}_m^\top \S_m^{-1} {\widehat{\A}}_m {\widehat{{\bf V}}}_m + {\bf I}| + \beta_m, \label{eq:en_ymM}
\end{align}
where $\beta_m = d_m \log \sqrt{2\pi e\sigma_m^2} + d_M \log \sqrt{2\pi e\sigma_M^2}$ and the log determinant is computed from a $(k_m + k_M) \times (k_m + k_M)$ matrix, which is cheap and efficient. Note that we can use block matrices to compute ${\widehat{\A}}_m^\top \S_m^{-1} {\widehat{\A}}_m = {\rm diag}\left(\sigma_m^{-2}{\bf A}_m^\top {\bf A}_m, \sigma_M^{-2}{\bf A}_M^\top {\bf A}_M\right)$.
Now, based on \eqref{eq:en_ym} and \eqref{eq:en_ymM}, we can calculate our acquisition function \eqref{eq:ac} in an analytic and deterministic way. For each fidelity $m$, we maximize $a({\bf x}, m)$ w.r.t to ${\bf x}$ to find the optimal input. We can use automatic differential libraries to calculate the gradient and feed it to any optimization algorithm, \textit{e}.\textit{g}.\xspace L-BFGS. We then use the optimal input at the fidelity that has the largest acquisition function value to query the next training example. Our deep multi-fidelity active learning is summarized in Algorithm \ref{alg:dmf-al}.
\setlength{\textfloatsep}{20pt}
\begin{algorithm}
\small
\caption{{\textsc{DiTucker}}\xspace($\mathcal{D}$, $T$, $\{\lambda_m\}_{m=1}^M$ )}
\label{alg:dmf-al}
\begin{algorithmic}[1]
\STATE Train the deep multi-fidelity model \eqref{eq:joint-prob} on the initial dataset $\mathcal{D}$ with stochastic structural variational learning.
\FOR {$t=1,\ldots, T$}
\STATE Based on \eqref{eq:en_ym} and \eqref{eq:en_ymM}, calculate and optimize the acquisition function \eqref{eq:ac} to find
\begin{align}
({\bf x}_{t}, m_{t}) = \operatornamewithlimits{argmax}_{{\bf x} \in {\mathcal{X}}, 1\le m \le M} a({\bf x}, m).\notag
\end{align}
\STATE Query the output ${\bf y}_{t}$ at input ${\bf x}_t$ with fidelity $m_t$.
\STATE $\mathcal{D} \leftarrow \mathcal{D} \cup \{({\bf x}_{t}, {\bf y}_{t})|m_t\}$.
\STATE Re-train the deep multi-fidelity model on $\mathcal{D}$.
\ENDFOR
\end{algorithmic}
\end{algorithm}
\vspace{-0.05in}
\subsection{Algorithm Complexity}
\vspace{-0.1in}
The time complexity of training our deep multi-fidelity model is ${\mathcal{O}}(N(\sum_{m=1}^M (k_ml_m)^2 + F))$ where $N$ and $F$ are the total number of training examples and NN parameters, respectively. The space complexity is ${\mathcal{O}}(F + \sum_{m=1}^M (k_ml_m)^2)$, which is to store the NN parameters, and posterior mean and covariance of each random weight matrix ${\bf W}_m$. The time complexity of calculating the acquisition function is ${\mathcal{O}}(\sum_{m=1}^M d_mk_m^2 + k_m^3)$. Since $k_m \ll d_m$, the complexity is linear in the output dimensions. Due to the usage of the learned variational posterior $q({\mathcal{W}})$, the space complexity of computing the acquisition function is the same as that of training the multi-fidelity model.
\section{Background}
\textbf{Problem setting.} We assume that we can query training examples with $M$ fidelities, which correspond to $M$ mappings, $\{{\bf h}_m({\bf x}) \in \mathds{R}^{d_m}\}_{1 \le m \le M}$ where ${\bf x} \in {\mathcal{X}}$ is an $r$-dimensional input, $r$ is small, $d_m$ is the output dimension, often very large, and $\{d_m\}$ are not necessarily identical. In general, we assume $d_1 \le \ldots \le d_M$. For instance, in structure design, the input can be the design parameters. The high fidelity examples correspond to high-resolution structures while the low fidelity ones low-resolution structures. We denote by $\lambda_m$ the cost to query with fidelity $m$. Our goal is to estimate ${\bf h}_M({\bf x})$ (\textit{i}.\textit{e}.,\xspace the target function). Among the examples at different fidelities are strong (but complex) correlations, due to the underlying common bases, say, in physics. Hence, all these examples are valuable to estimate the target function. To perform active learning, we begin with a small set of examples with mixed fidelities. Each step, we select both the fidelity and input location to query (or generate) new examples, so as to best balance the learning improvement (on ${\bf h}_M({\bf x})$) and computational cost, \textit{i}.\textit{e}.,\xspace maximizing the benefit-cost ratio.
\textbf{Dropout active learning.} A successful application of active learning with deep neural networks is image classification~\citep{gal2017deep,kirsch2019batchbald}. Typically, a pool of unlabeled input examples, \textit{i}.\textit{e}.,\xspace images, is pre-collected. Each time, we rank the input examples according to an acquisition function computed based on the current model. We then query the labels of the top examples and add them into the training set. There are two popular acquisition functions. The first one is the predictive entropy of the label,
\begin{align}
\mathbb{H}[y|{\bf x}, \mathcal{D}] = -\sum_{c} p(y=c|{\bf x}, \mathcal{D}) \log p(y=c|{\bf x}, \mathcal{D}), \label{eq:image-ac-1}
\end{align}
where ${\bf x}$ is the input, and $\mathcal{D}$ are the training data. The other one is BALD (Bayesian active learning by disagreement)~\citep{houlsby2011bayesian}, namely the mutual information between the label and model parameters,
\begin{align}
\mathbb{I}[y, \boldsymbol{\theta}|{\bf x}, \mathcal{D}] = \mathbb{H}[y|{\bf x}, \mathcal{D}]- \mathbb{E}_{p(\boldsymbol{\theta}|\mathcal{D})}\left[\mathbb{H}({\bf y}|{\bf x}, \boldsymbol{\theta}, \mathcal{D})\right], \label{eq:image-ac-2}
\end{align}
where $\boldsymbol{\theta}$ are the NN weights. To calculate these acquisition functions, we need to first estimate the posterior of the NN output and weights. A popular approach is to use Monte-Carlo (MC) Dropout~\citep{gal2016dropout}, which can be viewed as a variational inference method of Bayesian NNs. MC Dropout uses a probability $p_i$ to randomly drop the neurons in each layer $i$ (equivalent to zeroing out the corresponding weights) in the forward pass and performs backward pass over the remaining neurons. After training, the weights and output obtained from one such forward pass can be viewed as a sample of the approximate posterior. We can run Dropout multiple times to collect a set of posterior samples and calculate the acquisition functions by the Monte-Carlo approximation.
\vspace{-0.05in}
\section{Deep Multi-Fidelity Modeling for High-Dimensional Outputs}
\vspace{-0.05in}
Despite its success, dropout active learning might be inappropriate for high-dimensional (continuous) outputs. When the output is a continuous vector, it is natural to use a multivariate Gaussian posterior, $p({\bf y}|{\bf x}, \mathcal{D}) \approx {\bf N}({\bf y}|\boldsymbol{\mu}({\bf x}), \boldsymbol{\Sigma}({\bf x}))$\footnote{we can also use a multivariate student $t$ distribution, but the problem remains. }, and the entropy term $\mathbb{H}({\bf y}|{\bf x}, \mathcal{D})$ in the acquisition functions (see \eqref{eq:image-ac-1} and \eqref{eq:image-ac-2}) require us to compute the log determinant of the covariance matrix, $\log|\boldsymbol{\Sigma}({\bf x})|$. While we can use dropout samples to construct an empirical estimate $\widehat{\boldsymbol{\Sigma}}({\bf x})$, due to the large output dimension $n$, the computation of this $n\times n$ matrix and its log determinant is very expensive or even infeasible.
One might seek to only estimate the predictive variance of each individual output. However, this will ignore the strong output correlations, which is critical to effectively evaluate the uncertainty and calculate the acquisition function. Furthermore, in our active learning task, we want to optimize the acquisition function to find the best input (rather than rank the inputs in a pre-collected pool). The empirical covariance estimate from random samples can lead to numerical instability in optimization and inconsistent results from different runs.
To overcome these problems for multi-fidelity active learning, we first propose an expressive deep multi-fidelity model, for which we develop a structural variational inference method to capture the posterior dependency of the outputs. We propose a novel acquisition function to allow multi-fidelity queries, and develop an efficient and reliable approach to calculate the acquisition function.
Specifically, we use deep neural networks to build a multi-fidelity high-dimensional output model (see the graphical representation in Fig. 2 of the Appendix) that can flexibly, efficiently capture complex relationships between the outputs and between the fidelities, taking advantage of these relationships to enhance the learning at the highest fidelity (\textit{i}.\textit{e}.,\xspace target function). For each fidelity $m$, we introduce a neural network (NN) that first generates a $k_m$ dimensional latent output, and then projects it to the $d_m$ dimensional observational space, where $k_m \ll d_m$. The NN is parameterized by $\{{\bf W}_m, \boldsymbol{\theta}_m, {\bf A}_m\}$ where ${\bf W}_m$ is a $k_m \times l_m$ weight matrix in the latent output layer, ${\bf A}_m$ a $d_m \times k_m$ projection matrix, and $\boldsymbol{\theta}_m$ the weights in all the other layers. Denote by ${\bf x}_m$ the NN input, by ${\bf f}_m({\bf x})$ the latent NN output, and by ${\bf y}_m({\bf x})$ the observation. The model is defined by
\begin{align}
&{\bf x}_m = [{\bf x}; {\bf f}_{m-1}({\bf x})], \;\; {\bf f}_m({\bf x}) = {\bf W}_m\boldsymbol{\phi}_{\boldsymbol{\theta}_m}({\bf x}_m), \notag \\
&{\bf y}_m({\bf x}) = {\bf A}_m {\bf f}_m({\bf x}) + \boldsymbol{\epsilon}_m, \label{eq:our-model}
\end{align}
where ${\bf x}_1 = {\bf x}$, $\boldsymbol{\phi}_{\boldsymbol{\theta}_m}$ is the second last layer of the NN at fidelity $m$, of dimension $l_m$,\cmt{$\boldsymbol{\phi}_{\boldsymbol{\theta}_m}$ is the $l_m$ dimensional input to the output layer (hence parameterized by $\boldsymbol{\theta}_m$)} and $\boldsymbol{\epsilon}_m \sim {\bf N}(\boldsymbol{\epsilon}_m|{\bf 0}, \sigma^2_m{\bf I})$ is an isotropic Gaussian noise. Note that $\boldsymbol{\phi}_{\boldsymbol{\theta}_m}$ can be viewed as a set of nonlinear basis functions parameterized by $\boldsymbol{\theta}_m$. Through their combinations via ${\bf W}_m$ and ${\bf A}_m$, we can flexibly capture the complex relationships between the elements of ${\bf y}_m$ to improve the prediction. Furthermore, the input ${\bf x}_{m}$ is constructed by appending to the original input ${\bf x}$ the latent output ${\bf f}_{m-1}$, which can be viewed as a compact (or low-rank) summary of all the information up to fidelity $m-1$. After a series of linear and nonlinear transformations, we obtain the latent output ${\bf f}_m$ --- the summary up to fidelity $m$ --- and then generate the high dimensional observation ${\bf y}_m$. In this way, we efficiently integrate the information from lower fidelities and capture the complex relationship between the current and previous fidelities by learning an NN mapping, ${\bf f}_m({\bf x}) = \psi({\bf x}, {\bf f}_{m-1}({\bf x}))$
We place a standard Gaussian prior over the elements in each ${\bf W}_m$. Similar to~\citep{snoek2015scalable}, we consider all the remaining parameters as hyper-parameters to ease the posterior inference and uncertainty reasoning. Given the training dataset $\mathcal{D} = \{\{({\bf x}_{nm}, {\bf y}_{nm})\}_{n=1}^{N_m}\}_{m=1}^M$, the joint probability of our model is given by
\begin{align}
&p({\mathcal{W}}, {\mathcal{Y}}|{\mathcal{X}}, \Theta, {\bf s}) = \prod\nolimits_{m=1}^M {\bf N}\left(\vec({\bf W}_m)|{\bf 0}, {\bf I}\right) \notag \\
&\cdot \prod\nolimits_{n=1}^{N_m} {\bf N}\big({\bf y}_{nm}|{\bf A}_m {\bf f}_m({\bf x}_{nm}), \sigma_m^2{\bf I}\big), \label{eq:joint-prob}
\end{align}
where ${\mathcal{W}} = \{{\bf W}_m\}_{m=1}^M$, $\Theta=\{\boldsymbol{\theta}_m, {\bf A}_m\}_{m=1}^M$, ${\bf s} = [\sigma^2_1, \ldots, \sigma^2_M]$, and $\{{\mathcal{X}}$, ${\mathcal{Y}}\}$ are the inputs and outputs in $\mathcal{D}$.
To estimate the posterior of our model (which is used to calculate the acquisition function and query new examples), we develop a stochastic structural variational learning algorithm. Specifically, for each ${\bf W}_m$, we introduce a multivariate Gaussian variational posterior, $q({\bf W}_m) = {\bf N}\left(\vec({\bf W}_m)|\boldsymbol{\mu}_m, \boldsymbol{\Sigma}_m\right)$. To ensure the positive definiteness, we parameterize $\boldsymbol{\Sigma}_m$ by its Cholesky decomposition, $\boldsymbol{\Sigma}_m = \L_m \L_m^\top$, where $\L_m$ is a lower triangular matrix. We then assume the posterior $q({\mathcal{W}}) =\prod_{m=1}^M q({\bf W}_m)$, and construct a variational model evidence lower bound (ELBO)~\citep{wainwright2008graphical}, ${\mathcal{L}}\big(q({\mathcal{W}}), \Theta, {\bf s}\big) = \mathbb{E}_{q}\left[ \log{p({\mathcal{W}}, {\mathcal{Y}}|{\mathcal{X}}, \Theta, {\bf s})}/{q({\mathcal{W}})}\right] =-\text{KL}\big(q({\mathcal{W}})\|p({\mathcal{W}})\big) + \sum_{m=1}^M\sum_{n=1}^{N_m} \mathbb{E}_q\left[\log {\bf N}({\bf y}_{nm}|{\bf A}_m {\bf f}_m({\bf x}_{nm}),\sigma_m^2{\bf I})\right]$,
where $\text{KL}(\cdot \| \cdot)$ is Kullback Leibler divergence and $p({\mathcal{W}})$ the prior of ${\mathcal{W}}$. We maximize ${\mathcal{L}}$ to jointly estimate $q({\mathcal{W}})$ and the hyper-parameters. While ${\mathcal{L}}$ is intractable, it is straightforward to use the reparameterization trick~\citep{kingma2013auto} to generate parameterized samples for each $\vec({\bf W}_m)$: $\boldsymbol{\mu}_m + \L_m \boldsymbol{\xi}_m$ where $\boldsymbol{\xi}_m \sim {\bf N}({\bf 0}, {\bf I})$ to obtain a stochastic estimate of ${\mathcal{L}}$, and perform efficient stochastic optimization.
\section{Multi-Fidelity Active Learning}
\subsection{Mutual Information based Acquisition Function}
We now consider how to perform active learning with multi-fidelity queries. We assume that at each fidelity $m$, the most valuable training example is the one that can best improve our prediction of the target function, namely, the prediction at the highest fidelity $M$.
To this end, we propose our acquisition function based on the mutual information between the outputs at fidelity $m$ and $M$,
\begin{align}
&a({\bf x}, m) = \frac{1}{\lambda_m} \mathbb{I}\left({\bf y}_m({\bf x}) {\bf y}_M({\bf x})|\mathcal{D}\right) \notag \\
&= \frac{1}{\lambda_m} \big(\mathbb{H}({\bf y}_m|\mathcal{D}) +\mathbb{H}({\bf y}_M|\mathcal{D}) - \mathbb{H}({\bf y}_m, {\bf y}_M|\mathcal{D}) \big), \label{eq:ac}
\end{align}
where $\lambda_m>0$ is the cost of querying a training example with fidelity $m$. When $m=M$, we have $a({\bf x}, M) = \frac{1}{\lambda_M} \mathbb{H}\left({\bf y}_M({\bf x})|\mathcal{D}\right)$ --- the output entropy. Therefore, our acquisition function is an extension of the popular predictive entropy principle in conventional active learning. At each step, we maximize our acquisition function to identify a pair of fidelity and input location that give the biggest benefit-cost ratio.
\subsection{Efficient Acquisition Function Calculation}\label{sect:ac-compute}
To maximize the acquisition function \eqref{eq:ac}, a critical challenge is to compute the posterior of the outputs $\{{\bf y}_m\}$ in every fidelity, based on our model estimation results, \textit{i}.\textit{e}.,\xspace $p({\mathcal{W}}|\mathcal{D}) \approx q({\mathcal{W}})$. Due to the high dimensionality of each ${\bf y}_m$ and the nonlinear coupling of the latent outputs across the fidelities (see \eqref{eq:our-model}), the computation is challenging and analytically intractable. To address this issue, we consider approximating the posterior of the low dimensional latent output ${\bf f}_m$ first, which can be viewed as a function of the random NN weights $\boldsymbol{\Omega}_m = \{{\bf W}_1, \ldots, {\bf W}_m\}$ (given the input ${\bf x}$). We then use multivariate delta method~\citep{oehlert1992note,bickel2015mathematical} to compute the moments of ${\bf f}_m$. Specifically, we approximate ${\bf f}_m$ with a first-order Taylor expansion,
\begin{align}
{\bf f}_m(\boldsymbol{\Omega}_m) \approx {\bf f}_m(\mathbb{E}[\boldsymbol{\Omega}_m]) + {\bf J}_m\left(\boldsymbol{\eta}_m - \mathbb{E}[\boldsymbol{\eta}_m]\right), \label{eq:1order}
\end{align}
where the expectation is under $q(\cdot)$, $\boldsymbol{\eta}_m = \vec(\boldsymbol{\Omega}_m)$, ${\bf J}_m = \frac{\partial {\bf f}_m}{\partial \boldsymbol{\eta}_m}|_{\boldsymbol{\eta}_m = \mathbb{E}[\boldsymbol{\eta}_m]}$ is the Jacobian matrix at the mean. The rationale is as follows. First, ${\bf f}_m$ is linear to ${\bf W}_m$ and the second-order derivative is simply ${\bf 0}$. Second, as the NN output, ${\bf f}_m$ is highly nonlinear to the random weights in the previous layers, ${\bf W}_j$ ($j<m$). Hence, we can assume the change rate of ${\bf f}_m$ (gradient or Jacobian) has much greater scales than the posterior covariance of ${\bf W}_j$ in the second-order term of the Taylor expansion. Note that $q({\bf W}_j)$ is much more informative and hence much more concentrated than the prior $ {\bf N}(\vec({\bf W}_j)|{\bf 0}, {\bf I})$. The scale of the posterior covariance should be much smaller than $1$. Integrating both cases, the first-order term can dominate the Taylor expansion and hence we ignore the higher order terms.
Based on \eqref{eq:1order}, we can easily calculate the first and second moments of ${\bf f}_m$,
\begin{align}
\boldsymbol{\alpha}_m &= \mathbb{E}[{\bf f}_m] \approx {\bf f}_m\left(\mathbb{E}[\boldsymbol{\Omega}_m]\right), \\
{\bf V}_m &= \text{cov}[{\bf f}_m] \approx {\bf J}_m \text{cov}(\boldsymbol{\eta}_m) {\bf J}_m^\top,
\end{align}
where $\text{cov}(\boldsymbol{\eta}_m) = {\rm diag}\left(\left\{\text{cov}\big(\vec({\bf W}_j)\big)\right\}_{1\le j\le m}\right)$.
Then we use moment matching to estimate a joint Gaussian posterior, $q({\bf f}_m) = {\bf N}({\bf f}_m|\boldsymbol{\alpha}_m, {\bf V}_m)$. Next, according to \eqref{eq:our-model}, we can obtain the posterior of the output ${\bf y}_m$,
\begin{align}
q({\bf y}_m) = {\bf N}({\bf y}_m|{\bf A}_m \boldsymbol{\alpha}_m, {\bf A}_m {\bf V}_m {\bf A}_m^\top + \sigma_m^2{\bf I}).
\end{align}
The output entropy is $\mathbb{H}({\bf y}_m|\mathcal{D}) = \frac{1}{2}\log|{\bf A}_m {\bf V}_m {\bf A}_m^\top + \sigma_m^2{\bf I}| + d_m \log\sqrt{2\pi e}$. However, directly computing the log determinant of a $d_m \times d_m$ matrix is very expensive. We further use the Weinstein-Aronszajn identity~\citep{kato2013perturbation} to derive
\begin{align}
&\mathbb{H}({\bf y}_m|\mathcal{D}) = \frac{1}{2}\log|\sigma_m^{-2}{\bf A}_m {\bf V}_m {\bf A}_m^\top + {\bf I}| + d_m \log \sqrt{2\pi e\sigma_m^2} \notag \\
&=\frac{1}{2}\log|\sigma_m^{-2}{\bf A}_m^\top{\bf A}_m{\bf V}_m + {\bf I}| + d_m \log \sqrt{2\pi e\sigma_m^2}. \label{eq:en_ym}
\end{align}
Now, the log determinant is calculated on a much smaller, $k_m \times k_m$ matrix, which is very cheap and efficient.
Using a similar approach, we can calculate the joint posterior of ${\widehat{\h}}_m = [{\bf f}_m; {\bf f}_M]$ and then ${\widehat{\y}}_m=[{\bf y}_m; {\bf y}_M]$. First, we view ${\widehat{\h}}_m$ as a function of all the random weights, ${\mathcal{W}} = \{{\bf W}_1, \ldots, {\bf W}_M\}$. We use the multivariate delta method and moment matching to estimate a joint Gaussian posterior,
$q({\widehat{\h}}_m) = {\bf N}({\widehat{\h}}_m|{\widehat{\boldsymbol{\alpha}}}_m, {\widehat{{\bf V}}}_m)$ where ${\widehat{\boldsymbol{\alpha}}}_m = {\widehat{\h}}_m\left(\mathbb{E}[{\mathcal{W}}]\right)$, ${\widehat{{\bf V}}}_m = {\widehat{\J}}_m \text{cov}(\boldsymbol{\eta}){\widehat{\J}}_m^\top$, $\boldsymbol{\eta} = \vec({\mathcal{W}})$, ${\widehat{\J}}_m = \frac{\partial {\widehat{\h}}_m}{\partial \boldsymbol{\eta}}|_{\boldsymbol{\eta} = \mathbb{E}[\boldsymbol{\eta}]}$, and $\text{cov}(\boldsymbol{\eta}) = {\rm diag}\left(\left\{\text{cov}\big(\vec({\bf W}_j)\big)\right\}_{1\le j\le M}\right)$. According to our model \eqref{eq:our-model}, we can represent
\[
{\widehat{\y}}_m = {\widehat{\A}}_m {\widehat{\h}}_m + {\widehat{\bepsilon}}_m,
\]
where ${\widehat{\A}}_m = {\rm diag}({\bf A}_m, {\bf A}_M)$ and ${\widehat{\bepsilon}}_m = [\boldsymbol{\epsilon}_m; \boldsymbol{\epsilon}_M]$. Therefore, the joint posterior of ${\widehat{\y}}_m$ is
\begin{align}
q({\widehat{\y}}_m) = {\bf N}({\widehat{\y}}_m|{\widehat{\A}}_m {\widehat{\boldsymbol{\alpha}}}_m, {\widehat{\A}}_m {\widehat{{\bf V}}}_m {\widehat{\A}}_m^\top + \S_m),
\end{align}
where $\S_m = {\rm diag}(\sigma^2_m{\bf I}_{d_m}, \sigma^2_M {\bf I}_{d_M})$, ${\bf I}_{d_m}$ and ${\bf I}_{d_M}$ are identity matrices of $d_m \times d_m$ and $d_M \times d_M$, respectively. Again, we use Weinstein-Aronszajn identity to simplify the entropy computation of ${\widehat{\y}}_m$ (\textit{i}.\textit{e}.,\xspace ${\bf y}_m$ and ${\bf y}_M$),
\begin{align}
&\mathbb{H}({\bf y}_m, {\bf y}_M|\mathcal{D}) =\frac{1}{2}\log|\S_m^{-1}{\widehat{\A}}_m {\widehat{{\bf V}}}_m {\widehat{\A}}_m^\top + {\bf I}| + \xi_m \notag \\
&=\frac{1}{2}\log|{\widehat{\A}}_m^\top \S_m^{-1} {\widehat{\A}}_m {\widehat{{\bf V}}}_m + {\bf I}| + \xi_m, \label{eq:en_ymM}
\end{align}
where $\xi_m = d_m \log \sqrt{2\pi e\sigma_m^2} + d_M \log \sqrt{2\pi e\sigma_M^2}$ and the log determinant is computed from a $(k_m + k_M) \times (k_m + k_M)$ matrix, which is cheap and efficient. Note that we can use matrix blocks to compute ${\widehat{\A}}_m^\top \S_m^{-1} {\widehat{\A}}_m = {\rm diag}\left(\sigma_m^{-2}{\bf A}_m^\top {\bf A}_m, \sigma_M^{-2}{\bf A}_M^\top {\bf A}_M\right)$.
Now, based on \eqref{eq:en_ym} and \eqref{eq:en_ymM}, we can calculate our acquisition function \eqref{eq:ac} in an analytic and deterministic way. For each fidelity $m$, we maximize $a({\bf x}, m)$ w.r.t to ${\bf x}$ to find the optimal input. We can use automatic differential libraries to calculate the gradient and feed it to any optimization algorithm, \textit{e}.\textit{g}.\xspace L-BFGS. We then use the optimal input at the fidelity that has the largest acquisition function value to query the next training example. Our deep multi-fidelity active learning is summarized in Algorithm \ref{alg:dmf-al}.
\setlength{\textfloatsep}{20pt}
\begin{algorithm}
\small
\caption{{\textsc{DiTucker}}\xspace($\mathcal{D}$, $T$, $\{\lambda_m\}_{m=1}^M$ )}
\label{alg:dmf-al}
\begin{algorithmic}[1]
\STATE Train the deep multi-fidelity model \eqref{eq:joint-prob} on the initial dataset $\mathcal{D}$ with stochastic structural variational learning.
\FOR {$t=1,\ldots, T$}
\STATE Based on \eqref{eq:en_ym} and \eqref{eq:en_ymM}, calculate and optimize the acquisition function \eqref{eq:ac} to find
\begin{align}
({\bf x}_{t}, m_{t}) = \operatornamewithlimits{argmax}_{{\bf x} \in {\mathcal{X}}, 1\le m \le M} a({\bf x}, m).\notag
\end{align}
\STATE Query the output ${\bf y}_{t}$ at input ${\bf x}_t$ with fidelity $m_t$.
\STATE $\mathcal{D} \leftarrow \mathcal{D} \cup \{({\bf x}_{t}, {\bf y}_{t})|m_t\}$.
\STATE Re-train the deep multi-fidelity model on $\mathcal{D}$.
\ENDFOR
\end{algorithmic}
\end{algorithm}
\vspace{-0.05in}
\subsection{Algorithm Complexity}
\vspace{-0.1in}
The time complexity of training our deep multi-fidelity model is ${\mathcal{O}}(N(\sum_{m=1}^M (k_ml_m)^2 + F))$ where $N$ and $F$ are the total number of training examples and NN parameters, respectively. The space complexity is ${\mathcal{O}}(F + \sum_{m=1}^M (k_ml_m)^2)$, which is to store the NN parameters, and posterior mean and covariance of each random weight matrix ${\bf W}_m$. The time complexity of calculating the acquisition function is ${\mathcal{O}}(\sum_{m=1}^M d_mk_m^2 + k_m^3)$. Since $k_m \ll d_m$, the complexity is linear in the output dimensions. Due to the usage of the learned variational posterior $q({\mathcal{W}})$, the space complexity of computing the acquisition function is the same as that of training the multi-fidelity model.
\section{Related Work}
\vspace{-0.2in}
Active learning (AL) is a fundamental machine learning topic~\citep{balcan2007margin,settles2009active,balcan2009agnostic,dasgupta2011two,hanneke2014theory}.
Recent research focuses more on deep neural networks. \citet{gal2017deep} used Monte-Carlo (MC) Dropout~\citep{gal2016dropout} to perform variational inference for Bayesian neural networks. The dropout samples can be viewed as the posterior samples of the output. These samples are then used to compute an information measure, such as predictive entropy and BALD~\citep{houlsby2011bayesian}, to select unlabeled examples to query. This method has achieved a great success in image classification. Following this work, \citet{kirsch2019batchbald} developed a greedy approach to select a batch of unlabeled examples each time, so as to improve the efficiency of active learning.
There are also other excellent works along this line. For example, \citep{geifman2017deep,sener2018active} selected representative examples based on core-set search. \citep{gissin2019discriminative} selected maximally indistinguishable samples from the unlabeled pool, and the idea is reminiscent of generative adversarial networks~\citep{goodfellow2014generative}. \citep{ducoffe2018adversarial} used adversarial samples to calculate the distance to the decision boundary and select examples accordingly to label. \citep{ash2019deep} measured the uncertainty in terms of the gradient magnitude and select a disparate batch of inputs in a hallucinated gradient space.
Our work differs from the existing studies in several aspects. First, most methods are pool-based active learning, \textit{i}.\textit{e}.,\xspace a pool of unlabeled examples are collected beforehand, and new examples are only selected from the pool. This is reasonable when the training input is high-dimensional or hard to generate, \textit{e}.\textit{g}.\xspace image classification. By contrast, our work focuses on learning mappings from low-dimensional inputs to high-dimensional outputs, which are common in physical simulation and engineering design. To find the best training example, we \textit{optimize} the acquisition function in the entire domain rather than limit the search in a discrete set. Second, most active learning methods assume a uniform fidelity (or quality) of the training examples. Our work considers the case that the training examples can be queried with multiple fidelities, resulting in different cost/quality trade-offs. Our goal is to maximize the learning improvement while minimizing the query cost, \textit{i}.\textit{e}.,\xspace maximizing the benefit-cost ratio. This naturally extends the standard active learning, where the total cost is proportional to the number of queries due to the uniform fidelity. The pioneer work of \citep{settles2008active}
empirically studies active learning with nonuniform human labeling costs. To our knowledge, our work is the first multi-fidelity active learning approach for high-dimensional outputs.
Many multi-fidelity (MF) models have been proposed, while their active training approaches are lacking. These models are often based on Gaussian processes (GPs). For example, \citet{perdikaris2017nonlinear} successively estimated a set of GPs, where each GP predicts the output for one fidelity, where the input comes from the original input and the output of the previous fidelity. In \citep{cutajar2019deep}, this cascade model is jointly estimated. \citet{hamelijnck2019multi} used Gaussian process regression networks (GPRN) ~\citep{wilson2012gaussian}, deep GP~\citep{damianou2013deep} and mixture of experts~\citep{rasmussen2002infinite} to develop a multi-task, multi-resolution GP model. Despite their success, these models do not scale to high-dimensional outputs. The most recent work by ~\citet{wang2021multi} overcomes this limitation by proposing a nonlinear coregionalization component and then stacking these components for multi-fidelity modeling. How to conduct active learning for this model, however, remains an open problem --- even considering our proposed acquisition function, the computation is challenging due to the complicated couplings of the kernels and weight functions. In our experiments, we have compared the performance of our model with \citep{wang2021multi}, and other state-of-the-art high-dimensional GP regressors (that are single-fidelity) in non-active learning. The results confirmed the advantage of our NN based multi-fidelity model (see Sec. 3 of Appendix). Note that GPs are often used for learning physics and experiment design~\citep{sacks1989design}.
AL also connects to Bayesian optimization (BO)~\citep{mockus2012bayesian,snoek2012practical}, while the latter aims to find the function optimum and so the acquisition function definition, computation, and optimizing techniques/challenges are very different. Interestingly, while multi-fidelity AL methods are scarce, MF BO is common, \textit{e}.\textit{g}.\xspace
\citep{huang2006sequential, lam2015multifidelity,picheny2013quantile,kandasamy2016gaussian,kandasamy2017multi,poloczek2017multi,mcleod2017practical,wu2018continuous,song2019general,takeno2019multi}. Most MFBO approaches are based on GPs --- the classical BO framework. The recent MFBO work of \citet{li2020multi} uses a chain of NNs to estimate a multi-fidelity surrogate of the black-box objective function, queries new examples based on the maximum-value entropy search (MES) principle~\citep{wang2017max}. From the modeling perspective, its key difference (from our model) is that the NN output at each fidelity is directly fed into the NN for the next fidelity. Therefore, their model needs massive weight parameters to connect multiple NNs for high-dimensional outputs, which can be very inefficient. In addition, the massive outputs of the previous fidelity might dominate the original input (low-dimensional) to the model. Furthermore, \citet{li2020multi} developed a recursive one-dimensional quadrature to estimate the posterior of the output at each fidelity. While being suitable for optimizing a single-output black-box function, this technique is difficult to extend to multiple outputs (due to the explosion of the computational cost for multi-dimensional quadrature), not to mention a large number of outputs in our active learning setting.
\subsubsection*{\bibname}}
\onecolumn
\aistatstitle{Appendix}
\section{Experimental Details}
\subsection{Solving Partial Differential Equations}
\noindent \textbf{Burgers' equation} is a canonical nonlinear hyperbolic PDE, widely used to model various physical phenomena, such as nonlinear acoustics~\citep{sugimoto1991burgers}, fluid dynamics~\citep{chung2010computational}, and traffic flows~\citep{nagel1996particle}. Due to its capability of developing discontinuities (\textit{i}.\textit{e}.,\xspace shock waves), Burger's equation is used as a benchmark test example for many numerical solvers and surrogate models~\citep{kutluay1999numerical,shah2017reduced,raissi2017physics}.The viscous version of Burger's equation is given by
\[
\frac{\partial u}{\partial t} + u \frac{\partial u}{\partial x} = v \frac{\partial^2 u}{\partial x^2},
\]
where $u$ is the volume, $x$ is a spatial location, $t$ is the time, and $v$ is the viscosity. We set $x\in[0,1]$, $t \in [0,3]$, and $u(x,0)=\sin(x\pi/2)$ with a homogeneous Dirichlet boundary condition. The input parameter is the viscosity $v \in [0.001, 0.1]$. Given the input, we aim to predict the solution field (\textit{i}.\textit{e}.,\xspace the values of $u$) in the spatial-temporal domain $[0, 1] \times [0, 3]$. To obtain the training and test datasets, we solve the equation using the finite element~\citep{zienkiewicz1977finite} with hat functions in space and backward Euler in time domains on a regular mesh.
\noindent \textbf{Poisson's equation} is an elliptic PDE and commonly used to model potential fields, \textit{e}.\textit{g}.\xspace electrostatic and gravitational fields, in physics and mechanical engineering~\citep{chapra2010numerical}. The equation used in our experiment is given by
\[
\Delta u = \beta \delta({\bf x} - \c),
\]
where $\Delta $ is the Laplace operator~\citep{persides1973laplace}, $\u$ is the volume, $\delta(\cdot)$ is the Dirac-delta function, and $\c$ is the center of the domain. We used a 2D spatial domain, ${\bf x} \in [0, 1] \times [0, 1$], and Dirichlet boundary conditions. We used the constant values of the four boundaries and $\beta$ as the input parameters, each of which ranges from $0.1$ to $0.9$. We solved the equation using the finite difference method with the first order center differencing scheme and regular rectangle meshes.
\noindent \textbf{Heat equation} is a fundamental PDE that models heat conduction over time. It is also widely used in many other areas, such as probability theory~\citep{spitzer1964electrostatic,burdzy2004heat} and financial mathematics~\citep{black1973pricing}.
The equation is defined as
\[
\frac{\partial u}{\partial t} + \alpha \Delta u =0,
\]
where $u$ is the heat, $\alpha$ the thermal conductivity, and $\Delta$ the Laplace operator. In our experiment, we used a 2D spatial-temporal domain $x\in[0,1]$, $t \in [0,5]$ with the Neumann boundary condition at $x=0$ and $x=1$, and $u(x,0)=H(x-0.25)-H(x-0.75)$, where $H(\cdot)$ is the Heaviside step function. We considered three input parameters --- the flux rate $\in [0, 1]$ of the left boundary at $x=0$, the flux rate $\in [-1, 0]$ of the right boundary at $x=1$, and $\alpha \in [0,01, 0.1]$
To generate the training and test examples, we solve the equation with the finite difference in the space domain and backward Euler in the time domain.
\begin{figure*}[htbp!]
\centering
\includegraphics[width=1.0\textwidth]{./figs/illu.pdf}
\caption{\small Illustration of the motivation and goal with physical simulation as an example. Solving every PDE from scratch is expensive. Hence, we learn a multi-fidelity surrogate model that can predict high-fidelity solution outputs outright given the PDE parameters. We develop active learning methods to further reduce the cost of running numerical solvers to generate/collect training examples. }
\label{fig:illustration}
\end{figure*}
\begin{figure*}[htbp!]
\centering
\includegraphics[width=0.6\textwidth]{./figs/graphic-eps-converted-to.pdf}
\caption{\small Graphical representation of the deep multi-fidelity model. The low dimensional latent output in each fidelity ${\bf h}_m({\bf x})$ ($1\le m \le M$) is generated by a (deep) neural network.} \label{fig:graphical}
\end{figure*}
\subsection{Predicting Fluid Dynamics}
We also examined {\textsc{DiTucker}}\xspace in predicting the velocity field of a flow within a rectangular domain with a prescribed velocity along the boundaries~\citep{bozeman1973numerical}. This is a classical computational fluid dynamics (CFD) problem. The simulation of the flow involves solving the incompressible Navier-Stokes (NS) equation~\citep{chorin1968numerical},
\[
\rho (\textbf{u} \cdot \nabla) \textbf{u} = -\nabla p + \mu \nabla^2 \textbf{u},
\]
where $\rho$ is the density, $p$ is the pressure, $\textbf{u}$ is the velocity, and $\mu$ is the dynamic viscosity. The equation is well known to be challenging to solve due to their complicated behaviours under large Reynolds numbers
We set the rectangular domain to $[0, 1] \times [0, 1]$, and time $t \in [0, 10]$.
The input includes the tangential velocities of the four boundaries and the Reynold number $\in [100, 5000]$. The output are the first component of the velocity field at $20$ equally spaced time steps in $[0, 10]$ (see Fig. \ref{fig:cfd-example}). To generate the training and test examples, we used the SIMPLE algorithm~\citep{caretto1973two} with a stagger grid~\citep{versteeg2007introduction}, the up-wind scheme~\citep{versteeg2007introduction} for the spatial difference, and the implicit time scheme with fixed time steps to solve the NS equation.
\begin{figure*}
\centering
\setlength\tabcolsep{0pt}
\begin{tabular}[c]{ccc}
\setcounter{subfigure}{0}
\begin{subfigure}[t]{0.33\textwidth}
\centering
\includegraphics[width=\textwidth]{./figs/ns_t_1-eps-converted-to.pdf}
\caption{$t=1$}
\end{subfigure} &
\begin{subfigure}[t]{0.33\textwidth}
\centering
\includegraphics[width=\textwidth]{./figs/ns_t_5-eps-converted-to.pdf}
\caption{$t=5$}
\end{subfigure}
&
\begin{subfigure}[t]{0.33\textwidth}
\centering
\includegraphics[width=\textwidth]{./figs/ns_t_9-eps-converted-to.pdf}
\caption{$t=9$}
\end{subfigure}
\end{tabular}
\caption{\small Examples of the first component of the velocity field (with contour lines) at three time points.} \label{fig:cfd-example}
\end{figure*}
\begin{figure*}
\centering
\setlength\tabcolsep{0pt}
\begin{tabular}[c]{ccc}
\setcounter{subfigure}{0}
\begin{subfigure}[t]{0.38\textwidth}
\centering
\includegraphics[width=\textwidth]{./figs/new/Burgers_legend_extend.pdf}
\caption{Burgers' equation}
\end{subfigure} &
\begin{subfigure}[t]{0.25\textwidth}
\centering
\includegraphics[width=\textwidth]{./figs/new/Heat2_extend.pdf}
\caption{Heat equation}
\end{subfigure}
&
\begin{subfigure}[t]{0.25\textwidth}
\centering
\includegraphics[width=\textwidth]{./figs/new/Poisson2_extend.pdf}
\caption{Poisson's equation}
\end{subfigure}
\end{tabular}
\caption{\small Normalized root-mean-square error (nRMSE) for active learning of PDE solution fields with two-fidelity queries. The normalizer is the mean of the test outputs. The results are averaged from five runs. The shaded regions indicate the standard deviations.} \label{fig:solving-pde-2fid-all}
\vspace{-0.05in}
\end{figure*}
\subsection{Using Full Dropout}
In the experiment of Section 6.1 of the main paper, we have also applied MC dropout to directly generate the posterior samples of the final output in each fidelity $\{{\bf y}_m\}_{m=1}^M$. We generated $100$ samples. We used these high-dimensional samples to calculate the empirical means and covariance matrices , based on which we estimated a multi-variate Gaussian posterior for each ${\bf y}_m$ and $\widehat{{\bf y}}_m = [{\bf y}_m, {\bf y}_M]$ via moment matching.
These posterior distributions are then used to calculate and optimize our acquisition function and multi-fidelity BALD in the active learning. We denote these methods by {Dropout-full} and {MF-BALD-Dropout}. We report their nRMSE \textit{vs.} cost in Fig. \ref{fig:solving-pde-2fid-all} and \ref{fig:solving-po-3fid-all}, along with all the other methods. As we can see, in all the cases, {Dropout-full} is far worse than {Dropout-latent} and even inferior to random query strategies, except that in Fig. \ref{fig:solving-pde-2fid-all}a, they are quite close.
The prediction accuracy of {MF-BALD} is always much better than {MF-BALD-Dropout}. Those results demonstrate that the posterior distributions of the high-dimensional outputs, if fully estimated by a small number of dropout samples, are far less accurate and reliable than our method. Also, the computation is much more costly --- we need to directly compute an empirical covariance matrix based on these high-dimensional samples, and calculate the log determinant.
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{./figs/new/Poisson3_legend_extend.pdf}
\caption{\small Normalized root-mean-square error (nRMSE) for active learning of the solution field of Poisson's equation, with three-fidelity queries. The results are averaged from five runs. }
\label{fig:solving-po-3fid-all}
\end{figure}
\section{Gains in Computational Efficiency}
To confirm the gains in computational efficiency, we compared the cost of running standard numerical methods for topology structure optimization and fluid dynamics (Sec 6.2 and 6.3 in the main paper). First, after the active learning, we examined the average time of our learned model in computing (or predicting) a high-fidelity solution field, and contrast to that of the standard numerical approaches. As shown in Fig. \ref{fig:cost-ratio}a, our model is mush faster, giving 42x and 466x speed-up for the two tasks. Next, we consider adding the cost (running time) of the active learning, and calculate the average cost of predicting a high-fidelity solution field, $$\text{Cost}_{\text{avg}} = \frac{\text{Active Learning Cost} + \text{Prediction Cost}}{\text{Number of Acquired Solutions}}.$$ Note that for the numerical methods, there is no learning procedures and the active learning cost is zero. We examined how the average cost varies along with more acquired solutions (\textit{i}.\textit{e}.,\xspace test cases). As we can see in Fig. \ref{fig:cost-ratio} b and c, when only a few solutions are needed, directly running numerical methods is more economic because we do not need to conduct an extra learning procedure. However, when we need to compute more and more solutions (test cases), the average cost of using the surrogate model quickly becomes much smaller, and keeps decreasing. That is because after training, the surrogate model can predict solutions way more efficiently than running numerical methods from scratch, and the training cost will be diluted by the number of solutions acquired, finally becoming negligible. All these results have demonstrated the gains of computational efficiency of our method.
\section{Non-Active Learning}
To confirm the surrogate performance of our deep NN model, we also compared with several state-of-the-art high-dimensional non-active learning surrogate models: (1) PCA-GP~\citep{higdon2008computer}, (2) IsoMap-GP~\citep{xing2015reduced}, (3) KPCA-GP~\citep{xing2016manifold}, and (4) HOGP~\citep{zhe2019scalable}. PCA-GP, IsoMap-GP and KPCA-GP are based on the classical linear model of of coregionalization (LMC)~\citep{journel1978mining}, and obtain the bases or low-rank structures from Principal Component Analysis (PCA), IsoMap~\citep{balasubramanian2002isomap} and Kernel PCA~\citep{scholkopf1998nonlinear}, respectively. HOGP tensorizes the outputs, generalizes a multilinear Bayesian model with the kernel trick, and is flexible enough to capture nonlinear output correlations and can efficiently deal with very high-dimensional outputs. These methods are all single-fidelity models. In addition, we also compare with (6) MFHoGP~\citep{wang2021multi}, a state-of-the-art multi-fidelity high-dimensional surrogate model based on matrix GPs. Note that MFHoGP does not support varying output dimensions across the fidelities.
We tested all these methods in the three applications: \textit{Poisson-3}, \textit{Topology structure optimization} and \textit{Fluid dynamics} (see Sec. 6.1, 6.2 and 6.3 of the main paper). For \textit{Poisson-3}, we randomly generated 43, 43, 42 examples for the first, second and third fidelity, respectively. The inputs were uniformly sampled from the input space. For \textit{Topology structure optimization} and \textit{Fluid dynamics}, we randomly generated 64 examples for each fidelity (two fidelities in total). For each application, we randomly generated 512 examples for testing. The settings of those fidelities are the same as in the main paper. Since MFHoGP does not support varying output dimensions in different fidelities, we interpolated the training outputs at low fidelities to ensure all the fidelities have the same output dimension. We varied the number of bases (required by all the competing methods) from \{8, 16, 32, 64\}, which corresponds to the dimension of the latent output in our model (see Sec. 3 of the main paper). We ran the experiments for five times, and report the average nRMSE and its standard deviation in Fig. \ref{fig:na-predict2}. As we can see, our model consistently outperforms all the methods, often by a large margin. MFHoGP is the second best, but in most cases, its prediction accuracy is still significantly worse than our model ($p<0.05$). The remaining single-fidelity models are almost always worse than the multi-fidelity models --- that is reasonable, because the former cannot differentiate examples of different fidelities, blindly mix them in the training, and hence cannot leverage their relationships to improve the prediction. Together these results have shown the advantage of our deep surrogate model, even in the non-active learning setting.
\begin{figure*}
\centering
\setlength\tabcolsep{0pt}
\begin{tabular}[c]{ccc}
\setcounter{subfigure}{0}
\begin{subfigure}[t]{0.333\textwidth}
\centering
\includegraphics[width=\textwidth]{./figs/query_ratio-eps-converted-to.pdf}
\caption{Speed of computing solutions}
\end{subfigure} &
\begin{subfigure}[t]{0.333\textwidth}
\centering
\includegraphics[width=\textwidth]{./figs/cost_ratio_Lbracket-eps-converted-to.pdf}
\caption{Average cost in \textit{topology structure optimization}}
\end{subfigure} &
\begin{subfigure}[t]{0.333\textwidth}
\centering
\includegraphics[width=\textwidth]{./figs/cost_ratio_Navier-eps-converted-to.pdf}
\caption{Average cost in \textit{Fluid dynamics}}
\end{subfigure}
\end{tabular}
\caption{\small Speed of computing solutions after active learning (a) and total average solving (or predicting) cost with active learning included.} \label{fig:cost-ratio}
\end{figure*}
\begin{figure*}
\centering
\setlength\tabcolsep{0pt}
\includegraphics[width=0.7\textwidth]{./figs/legend.pdf}
\begin{tabular}[c]{ccc}
\setcounter{subfigure}{0}
\begin{subfigure}[t]{0.33\textwidth}
\centering
\includegraphics[width=\textwidth]{./figs/NA/s2/NA-nRMSE-Poisson3-eps-converted-to.pdf}
\caption{\textit{Poisson-3}}
\end{subfigure}
&
\begin{subfigure}[t]{0.33\textwidth}
\centering
\includegraphics[width=\textwidth]{./figs/NA/s2/NA-nRMSE-Lbracket-eps-converted-to.pdf}
\caption{\textit{Topology structure optimization}}
\end{subfigure}
&
\begin{subfigure}[t]{0.33\textwidth}
\centering
\includegraphics[width=\textwidth]{./figs/NA/s2/NA-nRMSE-Navier-eps-converted-to.pdf}
\caption{\textit{Fluid dynamics}}
\end{subfigure}
\end{tabular}
\caption{\small Predictive performance of non-active learning. The results were averaged over five runs.} \label{fig:na-predict2}
\end{figure*}
\section{Ablation Study about Multi-Variate Delta Method}
We examined the rationality of the multi-variate delta method used in our work (see (6) in Section 4.2 of the main paper). To this end, we used a Bayesian neural networks (BNN) to learn two nontrivial benchmark functions, \textit{Branin} and \textit{Levy}, which are defined as follows. For \textit{Branin}, the input ${\bf x} \in [-5, 10]\times [0, 15]$, and for \textit{Levy}, ${\bf x} \in [-10, 10]^2$. We used three hidden layers, with 40 neurons per layer, and \texttt{tanh} as the activation function (the same as in our experiment).
\begin{align}
y_{\text{Branin}}({\bf x}) = -\left(\frac{-1.275x_1^2}{\pi^2} + \frac{5x_1}{\pi} + x_2 - 6\right)^2 - \left(10 - \frac{5}{4\pi}\right)\cos(x_1) - 10, \notag \\
y_{\text{Levy}}({\bf x}) = -\sin^2(3\pi x_1) - (x_1-1)^2[1 + \sin^2(3\pi x_2)] - (x_2 - 1)^2[1 + \sin^2(2\pi x_2).
\end{align}
We generated $N$ training examples by randomly sampling from the input domain, and then used the structural variational inference to estimate the BNN. In our work, we used the first-order Taylor expansion of the NN output to approximate the posterior mean and (co-)variance (see (6-8) in the main paper). To confirm the rationality, we examined the ratio between the first-order Taylor expansion and the second-order Taylor expansion. Specifically, given a new input ${\bf x}$, we first sample an instance of the NN weights ${\mathcal{W}}$ from the learned posterior $p({\mathcal{W}}|\mathcal{D})$. Then, we expand the NN output $f_{{\mathcal{W}}}({\bf x})$ at the posterior mean of the weights $\mathbb{E}[{\mathcal{W}}]$, and calculate
\[
\text{Ratio} = \frac{|\text{First-Order Taylor Expansion of } f_{{\mathcal{W}}}({\bf x})|}{|\text{Second-Order Taylor Expansion of } f_{{\mathcal{W}}}({\bf x}) |}.
\]
We varied $N$ from $\{50, 100, 150, 200\}$. To obtain a reliable estimate of the ratio, for each $N$, we randomly sampled $100$ inputs. Given each input, we randomly drew $10$ samples of ${\mathcal{W}}$ to calculate the ratio. We report the average ratio, and its standard deviation in Fig. \ref{fig:approx-ratio}. As we can see, in both cases, the ratio is constantly close to one, and the standard deviation is close to zero. That means, the first-order approximation dominates, and the second-order terms can be ignored. This is consistent with our analysis in the main paper.
\begin{figure*}
\centering
\setlength\tabcolsep{0pt}
\begin{tabular}[c]{cc}
\setcounter{subfigure}{0}
\begin{subfigure}[t]{0.33\textwidth}
\centering
\includegraphics[width=\textwidth]{./figs/new/approx_ratio_Branin-eps-converted-to.pdf}
\caption{\textit{Branin}}
\end{subfigure}
&
\begin{subfigure}[t]{0.33\textwidth}
\centering
\includegraphics[width=\textwidth]{./figs/new/approx_ratio_Levy-eps-converted-to.pdf}
\caption{\textit{Levy}}
\end{subfigure}
\end{tabular}
\caption{\small The ratio between first-order and second-order approximations. \cmt{\textcolor{blue}{Shibo: change y-label to ``Ratio'', add ``0.9'' in the y-axis; x-label ``Number of training examples''}}} \label{fig:approx-ratio}
\end{figure*}
|
\section{Introduction}
Generation of picosecond optical laser pulses with tailored temporal intensity envelopes is a technically challenging task. Their narrow spectral bandwidth restricts methods that attempt to modify temporal intensity through spectral phase and their temporal duration is too short to exploit direct electronic modulation methods. However, picosecond duration pulses and shaping their temporal intensity profile is critically important for photoinjector-based electron instruments. Photoinjectors lie at the heart of a myriad of forefront scientific research areas and technologies in quantum electrodynamics~\cite{yakimenko2019prospect,wang2020coherent}, high-energy physics~\cite{ellis2001new,leemans2006gev}, ultrafast science~\cite{emma2010first,kern2018structures,stankus2019ultrafast,ibrahim2020untangling}, and medical technologies~\cite{neutze2000potential,coquelle2018chromophore,wolf2019photochemical} among many others. Even though the operational requirements for applications dependent on photoinjectors can be vastly different, increased electron control and performance remains a central challenge as these various technologies are brought to new scientific and engineering frontiers. From large scale facilities such as storage rings~\cite{bilderback2005review}, electron-electron colliders~\cite{ellis2001new}, and x-ray free electron lasers (XFELs)~\cite{mcneil2010x}, to laboratory-scale instrumentation such as ultrafast electron diffraction~\cite{hastings2006ultrafast} or transmission electron microscopy~\cite{li2014single}, electron generation in a photoinjector exists as the crucial first step and a significant predictor of instrument performance. One foundational strategy to tailor the electron bunch phase-space for greater uniformity and higher brightness is to control the shape of the optical pulses used for photoemission.
In photoinjectors, electrons are generated via the photoelectric effect with laser pulses comprised of light, typically in the deep ultraviolet (UV), whose photon energy lies above the work function of the cathode material. After photoemission, electrons' phase-space distribution evolves rapidly from the interaction with the external accelerating fields and their own internal coulombic forces~\cite{palmer1998next}. As part of this process, the spatio-temporal distribution of the laser pulses effect the electron distribution, and thereby can be used to influence how the bunches will evolve as they are accelerated in vacuum. A key measure of electron bunch quality is its transverse emittance, $\epsilon_x$, defined as~\cite{Wiedemann2015}
\begin{equation}
\epsilon_x = \sqrt{\langle x_{i}^{2} \rangle \langle x_{i}^{\prime 2} \rangle - \langle x_{i}^{2} x_{i}^{\prime 2} \rangle \label{EQN:Emittance}}
\end{equation}
where $x$ is transverse position and $x^{\prime}$ is the corresponding angle with respect to the ideal trajectory. Generally, emittance reduction is beneficial to the electron beam brightness and any associated secondary emission processes, and can be therefore be used as a metric to optimize in determining spatio-temporal laser distributions~\cite{stephan2020high}.
In XFELs, in addition to low emittance ($<1.5$ $\mu$m), it is particularly crucial to have electron bunches with narrow energy spread ($\Delta \mathrm{E}/\mathrm{E} < 10^{-3}$) and good spatial uniformity, as growth in these parameters can significantly decrease x-ray production in the undulators~\cite{PhysRevSTAB.10.034801}. Beyond conventional Gaussian temporal and spatial distribution of the photoinjector laser pulses, other commonly sought-after laser distributions shown to reduce normalized transverse emittance include flat-top spatio-temporal profiles resembling cylinders~\cite{krasilnikov2012experimentally} or 3D ellipsoids such that the beam size and intensity vary as a function of time~\cite{luiten2004realize}.
Existing shaping techniques rely on modifying or combining femtosecond pulses and can be split into either spectral techniques such as spatial-light modulators~\cite{mironov2016shaping,penco2013optimization} or acousto-optic modulators~\cite{li2009laser,petrarca2007production}, or temporal techniques such as pulse stackers~\cite{krasilnikov2012experimentally,krasilnikov2019studies}. Spectral methods are hindered by tailored phase structure being distorted during pulse amplification and upconversion if shaping in the IR, limited spectral bandwidth if shaping in the UV, and by material damage threshold limitations in the mW-level~\cite{carbajo2018power}. On the other hand, intensity fluctuations inherent to temporal techniques have been shown to induce unwanted microbunching~\cite{PhysRevAccelBeams.23.024401,mitchell2016sensitivity} on the electron bunch, resulting in increased emittance relative to Gaussian distributions. In order to operate at the correct wavelength, these lasers also typically employ a series of nonlinear conversion stages to upconvert infrared (IR) light to UV light below 270 nm~\cite{will2011photoinjector,gilevich2020lcls,alley1999design}. Efficient nonlinear conversion is detrimentally affected by non-flat phase structure and can distort tailored temporal profiles, complicating shaping efforts. As such, these methods are limited in their applicability to high average power, 24/7 facilities, such as LCLS where reliability is paramount.
We present an upconversion method that incorporates temporal shaping into a nonlinear conversion stage exploiting post-amplification phase manipulation, thereby circumventing the pitfalls of existing upconversion and shaping techniques. We utilize non-colinear sum frequency generation (Fig. \ref{FIG:OpChirp_Schematic}) combining two highly dispersed pulses that result in a pulse with tunable temporal profile in duration and shape~\cite{vicario2012deep,kuzmin2021highly}. Here, we build on the elegant approach by Raoult et al~\cite{raoult1998efficient} of efficient narrowband second harmonic generation in thick crystals by adding higher-order dispersion to simultaneously shape the output pulse duration. The output pulse is nearly transform limited and can be directly utilized for application or passed through further nonlinear conversions processes without distortion. As such, this method, which we call dispersion controlled nonlinear shaping (DCNS), can be broadly used to tailor pulses for the reduction of normalized transverse emittance in photoinjector-based instrumentation.
\begin{figure}[htp!]
\centering
\includegraphics[width=1\linewidth]{OpChirp_Model.pdf}
\caption{DCNS method in the spectral domain demonstrating two pulses of equal on opposite chirp mixing in a nonlinear medium to generate chirped second harmonic (SHG) copies of each pulse along with the narrowband, flat phase, sum frequency (SFG) pulse.}
\label{FIG:OpChirp_Schematic}
\end{figure}
\section{Dispersion-controlled Pulses}\label{SEC:optics}
To describe DCNS, we start off by expressing the electric field of a laser in frequency space as $E(\omega)=A(\omega)e^{i\varphi(\omega)}$, where $A(\omega)$ is the spectral amplitude and $\varphi(\omega)$ is the spectral phase as a function of angular frequency $\omega$. The amplitude function is typically Gaussian in shape and the spectral phase is customarily described by a Taylor expansion around the central frequency of the field, $\omega_0$.
\begin{multline}
\varphi(\omega) = \varphi_0 + \varphi_1 (\omega-\omega_0) + \frac{\varphi_2}{2!} (\omega-\omega_0)^2 \\
+ \frac{\varphi_3}{3!} (\omega-\omega_0)^3 + \frac{\varphi_4}{4!} (\omega-\omega_0)^4 + \dots \label{EQN:SpecPhase}
\end{multline}
Here $\varphi_j$ is the $j^{th}$ derivative of $\varphi(\omega)$, evaluated at $\omega_0$.
The first two terms of \eqref{EQN:SpecPhase} correspond to the absolute phase at the central frequency and the group delay, respectively. Since these two terms do not affect the shape of the pulses, our investigation will focus on the next terms, higher order in $\omega$, to control the shape of the pulse envelope. The second order dispersion (SOD) term, $\varphi_2$, is a quadratic phase that gives a linear ordering to the arrival time of the frequency components, stretching the duration of the pulse. Third order dispersion (TOD) results from cubic phase, $\varphi_3$. Here the opposite ends of the spectrum will overlap in time either before or after the main peak, leading to characteristic beating or temporal fringes. Higher orders effect the pulse similarly, but to a lesser degree, with even orders primarily effecting pulse duration and odd orders leading to modulation of the pulse in the time domain. In our description, we will focus on using a combination of SOD and TOD as the two most influential parameters to control the shape of the output pulse in the nonlinear sum frequency mixing process.
\section{Methods \& Results}
We begin by modeling these pulses as a combination of two equal-energy transform-limited Gaussian pulses overlapped in time. From here, the phase of each pulse can be adjusted separately by multiplying with Eq.\ref{EQN:SpecPhase} in frequency space, where $\varphi_2$ and $\varphi_3$ are free parameters and higher-order terms are ignored. These tailored pulses then become our two initial fields, $A_1$ and $A_2$, used in solving the coupled equations (Eq. \eqref{EQN:SFGEqns}) for sum frequency generation~\cite{boyd2019nonlinear}.
\begin{subequations}
\begin{align}
\frac{d A_1}{d z} &= \frac{2 i d_{eff} \omega_1^2}{k_1 c^2}A_2^* A_3 e^{-i \Delta k z} \\
\frac{d A_2}{d z} &= \frac{2 i d_{eff} \omega_2^2}{k_2 c^2}A_1^* A_3 e^{-i \Delta k z} \\
\frac{d A_3}{d z} &= \frac{2 i d_{eff} \omega_3^2}{k_3 c^2}A_1 A_2 e^{i \Delta k z}
\end{align}
\label{EQN:SFGEqns}
\end{subequations}
Propagation and frequency mixing is handled using a symmetrized split-step Fourier method along with a fourth order Runga-Kutta algorithm to solve the coupled nonlinear equations. Nonlinear conversion and nonlinear index effects are handled in the time and position domains while propagation and dispersion through the crystal are handled in the temporal and spatial frequency domain. To ensure accurate results, the resolution of the time, frequency, and spatial grids was increased until further refinement resulted in negligible change to the results.
As stated above, SOD primarily controls duration and TOD controls the sharpness of the leading or trailing edge and ringing in the field on the opposing edge. It is the interplay between SOD and TOD (Fig. \ref{FIG:3x3_Array}) that then determines the pulse duration and the shape.
By defining the ratio between TOD and SOD,
\begin{equation*}
\alpha = \frac{\varphi_3 / ps^3}{\varphi_2/ps^2},
\end{equation*}
we gain a single parameter to describe the general shape of a shaped pulse that is approximately invariant to pulse duration~\footnote{See Supplemental Material at [URL will be inserted by publisher] for reference figure.}.
\begin{figure}[htp!]
\centering
\includegraphics[width=1\linewidth]{3x3_Array_cent_text_w_Words.pdf}
\caption{Pulse envelopes resulting from different combinations of SOD and TOD. Each column has a the same magnitude of the TOD value at the top and each row has the same magnitude of SOD. The blue plot results from combinations with negative SOD and positive TOD while the red plot is positive SOD and negative TOD.}
\label{FIG:3x3_Array}
\end{figure}
Applying these concepts in combination with narrow bandwidth generation suggested by Raoult et al.~\cite{raoult1998efficient} we can search for pulses that fulfill our need for sharp rise time, long pulse duration, and narrow bandwidth.
To best illustrate our conversion scheme and
to explore how it might be implemented in a real-world case, we use the photoinjector system at LCLS-II. This system consists of an Ytterbium-based commercial laser that outputs 50 $\mu$J, 330~fs pulses with an approximately 4~nm full width at half maximum (FWHM) spectral bandwidth. We propose using type I non-colinear SFG in the conversion from 1030~nm to 515~nm. Using Fig.\ref{FIG:3x3_Array} as a reference to which portion of the pulses will mix, and considering that the LCLS-II design assumed a few 10s of picosecond flat-top photoinjector laser pulse~\cite{osti_1029479}, we focus on situations where the magnitude of $\alpha$ is close to $0.125$~ps (center column). This results in the signal and idler pulses having equal and opposite amounts of SOD and $\alpha=-0.128$~ps, for temporal symmetry. The initial value of SOD ($\approx$3.5 ps$^2$) was chosen so that the FWHM of the 515~nm pulse would be 25~ps in time.
This large amount of SOD necessitates a large amount of TOD to maintain $\alpha$. This can be achieved by passing the transform-limited pulses through a matched compressor-stretcher set with gratings detuned from Littrow angle. To maintain grating efficiency one can generate a larger amount of SOD and an appropriate amount of TOD in the set and reduce SOD to desired levels afterwards with highly dispersive optics, such as chirped volume Bragg gratings~\cite{galvanauskas1998use}.
Once the half-pulse fields are constructed from the given phase parameters, they are propagated with the split-step method through BBO with the crystal angle tuned for type I sum mixing and a 2~mm crystal length. The crossing angle was set to 1.5~deg. This angle must be sufficiently large to allow the sum-frequency signal to be separated from the two input beams and suppress intra-beam second harmonic generation, but not so large as to reduce the spatial overlap of the beams in the crystal~\cite{raoult1998efficient}. Additionally, the simulated focusing through the crystal was adjusted to eliminate back-conversion from 515 nm to 1030 nm as this process can distort the temporal and spectral profiles.
\begin{figure}[htp!]
\centering
\includegraphics[width=1\linewidth]{GA_Pulse_w_Words.pdf}
\caption{Genetic algorithm optimized 515 nm laser profile for a bunch length of 1.22 mm resulting in an emittance value of 0.30 $\mu$m. Notably, the genetic algorithm settled on a profile which has slower rise and fall times than an ideal flat top as emittance is optimized rather than profile shape. The two plots shown are, a) the laser profile in time before (grey) and after (blue, dashed) a 0.5 nm spectral filter, and b) the spectrum of the pulse before (grey) and after filtering (blue, dashed) with the super-gaussian spectral filter in black.}
\label{FIG:Fin_Field_Reg_Filt}
\end{figure}
The resultant pulse (Fig. \ref{FIG:Fin_Field_Reg_Filt}a) displays the attractive qualities of a sharp rise time and a flatter profile than the traditionally used Gaussian pulses with upwards of 65\% conversion efficiency in simulation. Experimental efficiency is likely to be less than the 40\% stated by Raoult et al.~\cite{raoult1998efficient} due to added TOD. It is also characterized by large and rapid amplitude fluctuations on the picosecond scale that can be detrimental to e-beam emittance. These fluctuations—the larger oscillations at the edges-—are the result of coherent interaction between the main spectral peak and interference from the higher wavelengths in the asymmetric tail seen in Fig. \ref{FIG:Fin_Field_Reg_Filt}b. However, by applying a spectral amplitude filter after conversion that is significantly wider than the FWHM of the pulse's spectral bandwidth (Fig. \ref{FIG:Fin_Field_Reg_Filt}b), the high frequency components can be attenuated without a major efficiency penalty. In this case, employing a second order super-Gaussian spectral filter with $1/e$ width of 0.5 nm, such as a Bragg grating interference filter, the total power in the field is reduced by less than 10\% and results in a smoother temporal profile (Fig. \ref{FIG:Fin_Field_Reg_Filt}a). As with the mixing, the numerical grids were chosen to reduce possible numerical errors from these filters. Alternatively, the thickness of the nonlinear crystal for the doubling step from 515 nm to 257 nm can be chosen to filter the bandwidth through phase mismatch~\cite{radzewicz2000poor}.
To compare the performance of the DCNS pulses vs. the baseline Gaussian temporal distribution, we simulated the LCLS-II photoinjector performance. To reduce simulation time, the DCNS temporal intensity profiles at 515 nm are squared to directly generate the UV profiles with a root mean squared error of less than 1\%~\footnote{See Supplemental Material at [URL will be inserted by publisher] for quantitative analysis.}. These pulses are supplied as the initial condition to the simulation and used to model the initial electron beam parameters. Electron bunch generation includes variations derived from oscillations in the laser pulse profile. The simulation code used for e-beam dynamics is OPAL~\cite{adelmann2019opal}, and for particle distribution generation, distgen~\cite{distgen}. The FWHM in time and spot size of the laser pulse directly impact the resulting 3D shape of the emitted electron beam. As the beam is emitted, the forces due to external magnetic and electric fields along with internal space charge fields are computed at each time step. In this case, the FWHM is controlled by the magnitude of SOD and shape by $\alpha$. While supplying the DCNS pulses is straightforward, determining the optimal FWHM and spot size of the pulse is not. The strength of the space charge forces are directly impacted by both the FWHM and spot size, which then impacts how strong the external forces need to be to limit emittance growth. To determine optimal laser and machine settings, the parallel simulation code is run in combination with an optimization algorithm (NSGA-II~\cite{dpam:02}) using libEnsemble~\cite{libEnsemble}, a Python library that coordinates ensemble calculations. Standard optimization of the LCLS-II photoinjector includes variables such as laser spot size, laser FWHM, RF cavity phases, RF cavity gradients, and magnet strengths.
To maintain broad applicability of these results, we limit the simulation to only the photoinjector and the first 15 meters of acceleration ($\approx$100 MeV), after which are LCLS-II specific configurations. In this region the laser parameters have the greatest impact on electron bunch evolution as internal space-charge forces are not yet damped by highly relativistic speeds.
\begin{figure}[htp!]
\centering
\includegraphics[width=0.45\textwidth]{compare_paretos_PRL.pdf}
\caption{Pareto front comparison of DCNS and Gaussian performance for the LCLS-II injector. DCNS pulses in combination with a 1.0 nm spectral filter, achieves the lowest emittance values at most bunch lengths. The lowest achieved emittance value is 0.30~$\mu$m at a bunch length of 1.22~mm, using a 0.5~nm filter.}
\label{FIG:optresults}
\end{figure}
The metrics commonly used for determining beam quality for XFELs are emittance \eqref{EQN:Emittance} and bunch length ($\sigma_z$).
If the emittance is extremely small, the beam size, momentum, or both must be small as well.
Note, we do not optimize the orthogonal transverse dimension, $y$, because the simulation is transversely symmetric.
As stated earlier, smaller emittances lead to brighter bunches, and therefore better x-ray production.
The same logic applies to second metric of interest, the bunch length,
which is the rms size of the bunch in the longitudinal dimension.
Shorter bunches require less compression after the injector, which mitigates non-linearities in the bunch compression process and results in better FEL performance.
With these variables and metrics defined, several optimization rounds were performed to compare the performance of DCNS and Gaussian laser pulses in the LCLS-II photoinjector. Final results are shown in Fig.~\ref{FIG:optresults} and Fig.~\ref{FIG:hist}, with the later showing density of simulation points near the Pareto fronts.
\begin{figure}[htp!]
\centering
\includegraphics[width=\linewidth]{pareto_hist_prl.pdf}
\caption{Histogram showing simulation density of all GA solutions, not only Pareto optimal, for Gaussian and DCNS cases. Lighter colored areas indicate a higher number of simulations with valid solutions and thus regions where different methods might be more effective.}
\label{FIG:hist}
\end{figure}
Fig.~\ref{FIG:optresults}, demonstrates that lower emittance values are achieved using DCNS. Further, Fig.~\ref{FIG:hist} indicates the density of low emittance values are shifted for different filters. With a spectral filter of bandwidth 0.5 nm, longer bunch lengths are performing better, whereas with the 1.0 nm filter, there is a higher density of points at shorter bunch lengths. This suggests DNCS could benefit a variety of experimental XFEL configurations with the filter bandwidth and $\alpha$ used as an adjustment knob.
As shown in Fig.~\ref{FIG:optresults}, the best value found is $\epsilon_x=0.30$~$\mu$m with the the DCNS pulse shown in Fig.~\ref{FIG:Fin_Field_Reg_Filt}.
The bunch length in this case, 1.22 mm,
is slightly longer than the typical operating length of 1 mm at LCLS.
For a practical comparison, we choose the minimum emittance values at $\sigma_z=1$ mm.
A 25\% improvement in the emittance value is
obtained from DCNS ($\epsilon_x=0.37$~$\mu$m) vs. Gaussian ($\epsilon_x=0.50$~$\mu$m) pulses.
Note, this Gaussian point is not visible in Fig.~\ref{FIG:optresults},
because it is not Pareto optimal.
For shorter bunch lengths, i.e. 0.5 mm, the difference is slightly larger
reaching about 30\% ($\epsilon_x=0.4$~$\mu$m vs 0.58~$\mu$m). Applying this reduction to both $x$ and $y$ planes, since the simulation is symmetric, the total transverse brightness can be more than doubled. In the case of XFELs, this emittance improvement translates to a twofold increase in undulator peak brightness, 25\% shorter x-ray wavelengths, and an upper bound reduction in undulator lengths by 25\% for similar peak currents, which can substantially reduce cost, complexity, and size.
\section{Conclusion}
The regime of picosecond intensity envelope shaping is a challenging task that traditionally lies outside the capabilities of many methods commonly implemented for optical pulses with more spectral bandwidth or longer duration. Nonetheless, photoinjector-based electron sources, which would benefit from shaping in this regime, driving understanding of sub-atomic interactions, molecular structure, biological processes, and fundamental particle science utilize known sub-optimal techniques. By tailoring the photoinjector drive lasers, electron emittance, and by extension the electron beam brightness, can be improved enabling further exploration of these research areas and the possibility for broad impact across multiple disciplines.
Existing shaping techniques for the up-converted lasers used in the photoinjectors suffer from challenges in maintaining favorable pulse shapes, providing enough photon throughput for excitation, or even increasing electron emittance. Our proposed optical shaping and conversion technique, DCNS, circumvents these issues by directly upconverting optical pulses where the production of favorable temporal distributions is embedded in sum frequency conversion using highly dispersed pulses. In the case of linear accelerators and XFELs such the LCLS-II, this simple solution is expected to improve electron emittance across all investigated bunch lengths over conventional Gaussian pulses with an upwards of 30\% emittance reduction at short bunch lengths (0.25 mm) and 25\% at bunch lengths greater than or equal to 1 mm. With an effective conversion efficiency of upwards of 40\%, we have laid a realistic avenue to substantially extend the brightness of photoinjector systems worldwide without major configuration changes and thus enhance current scientific capabilities on existing accelerators and reduce the cost of future accelerator facilities.
\begin{acknowledgments}
We thank Yuantao Ding and Christopher Mayes at SLAC National Accelerator Laboratory for their helpful discussions.
This work is supported by the U.S. Department of Energy, Office of Science, Office of Basic Energy Sciences under Contract No. DE-AC02-76SF00515.
\end{acknowledgments}
\section{Control of pulse shape and duration}
\begin{figure}[htp!]
\centering
\includegraphics[width=0.7\linewidth]{3x3_Array_sod_alpha_w_Words.pdf}
\caption{Pulse envelopes from different combinations of SOD and $\alpha$ demonstrating invariance of pulse shape to duration (row) instead being determined by the magnitude of $\alpha$ (column). The blue plot results from combinations with negative SOD and positive TOD while the red plot is positive SOD and negative TOD.}
\label{FIG:3x3_Array_Alpha}
\end{figure}
The temporal intensity profile of a broadband laser pulse is heavily affected by the spectral phase of the constituent wavelengths. When representing the spectral phase as a Taylor expansion in frequency around the central frequency of the pulse
\begin{equation}
\varphi(\omega) = \varphi_0 + \varphi_1 (\omega-\omega_0) + \frac{\varphi_2}{2!} (\omega-\omega_0)^2 + \frac{\varphi_3}{3!} (\omega-\omega_0)^3 + \frac{\varphi_4}{4!} (\omega-\omega_0)^4 + \dots,
\end{equation}
the effect on pulse shape is determined by the order of the term along with magnitude and sign. Here $\varphi_j$ is the $j^{th}$ derivative of $\varphi(\omega)$, evaluated at $\omega_0$. Second order dispersion (SOD), $\varphi_2$, is an even function around the central frequency and thus has a symmetric stretching effect to the temporal profile regardless of sign of the coefficient. However, the sign still determines whether the higher or lower frequency components lead the central frequency in time, with higher components leading for negative signs, and thus still has an effect when considering interactions with other dispersion orders. Third order dispersion (TOD), $\varphi_3$, is an odd function and results in a pre or post-pulse ringing depending on the sign of the coefficient with pre-pulses generated by positive sign. Higher orders have a similar even-odd-even pattern of effects but we do not include or cover their effects here.
In addition to the sign of the individual orders, the relative magnitudes also have a significant effect on the pulse shape. The trivial cases are pulses with some SOD and no TOD or visa versa. In these cases the pulse shape is fully representative of the appropriate orders' effect on temporal intensity. However, when varying amounts of SOD and TOD are added to the same pulse, the shape is determined instead by the relative amount of one to the other and thus we define a ratio,
\begin{equation*}
\alpha = \frac{\varphi_3 / ps^3}{\varphi_2/ps^2},
\end{equation*}
to characterize pulse shape regardless of the absolute magnitude of either. $\alpha$ is defined such that value grows as TOD is added so that, at a glance, the magnitude of $\alpha$ represents departure from the typical Gaussian pulse shapes. Fig. \ref{FIG:3x3_Array_Alpha} demonstrates these properties of $\alpha$. Each column highlights the qualitative invariance to the magnitudes of the dispersion by locking the relative amounts. In this case TOD was calculated for a fixed $\alpha$ and changing SOD. Additionally , as $\alpha$ grows (left to right) the temporal intensity becomes less Gaussian due to a relatively larger amount of TOD for a given SOD.
\section{Direct calculation of UV profiles}
A diagram of the ultraviolet (UV) dispersion controlled nonlinear synthesis (DCNS) pulse generation can be seen in Fig. \ref{FIG:OpChirp_Full}. The UV light is generated as a cascade of nonlinear conversion steps with spectral filtering taking place on the 515 nm light before the second conversion stage. Generating a UV profile from the 1030 nm light is involved and requires building the 1030 nm light from desired SOD and $\alpha$, simulating the nonlinear dynamics in the sum frequency generation (SFG) crystal using the methods in section \ref{SEC:optics}, applying an amplitude filter to the SFG light in the spectral domain and re-generating the temporal profile by inverse fourier transform, and finally simulating the nonlinear dynamics in the second crystal. For this process, generating the initial 1030 nm pulses and spectrally filtering the 515 nm light are computationally negligible compared to each of the split-step Fourier crystal simulations. As such, any computational speed increases to aid in the necessarily large amount of trials needed by a genetic algorithm would be found in the crystal simulations. To this end, eliminating one of the crystal simulations would essentially half the computational time of the entire optical pulse simulation.
\begin{figure}[htp!]
\centering
\includegraphics[width=1\linewidth]{OpChirp_Model_Full.pdf}
\caption{UV DCNS pulse generation from highly dispersed infrared pulses. The main stages are opposite chirp sum frequency generation (SFG) from 1030 nm to 515 nm, spectral filtering of SFG pulses, and finally second harmonic generation from 515 nm to 257 nm. The envelope representations shown here are in the time domain}
\label{FIG:OpChirp_Full}
\end{figure}
Second harmonic generation is governed by the coupled equations,
\begin{subequations}
\begin{align}
\frac{d A_1}{d z} &= \frac{2 i d_{eff} \omega_1^2}{k_1 c^2}A_1^* A_2 e^{i \Delta k z} \\
\frac{d A_2}{d z} &= \frac{i d_{eff} \omega_2^2}{k_2 c^2}A_1^2 e^{i \Delta k z}
\end{align}
\label{EQN:SHGEqns}
\end{subequations}
where $A_1$ is the incident field to the crystal, $A_2$, $\Delta k = 2k_1-k_2$, $d_{eff}$ is the nonlinear coefficient of the crystal, and $\omega_i$ and $k_i$ are the central frequency and wave vector of their respective fields~\cite{boyd2019nonlinear}. In the DCNS process, the second harmonic conversion from 515 nm to 257 nm is a remarkably simple situation mathematically. The 515 nm light is spectrally narrow with very little phase variation and as such $\Delta k$ is very small. In this case, the field amplitude of the second harmonic field is well approximated as a scaled version of the incident lights' intensity profile.
\begin{figure}[htp!]
\centering
\includegraphics[width=0.7\linewidth]{square515_v2_w_words.pdf}
\caption{Generation of UV intensity profile for the 0.5 nm filter optimal parameters (SOD: $\pm 3.436$, $\alpha$: -0.128) via full crystal simulation and approximation by squaring the 0.5 nm filtered 515 nm light. The full simulation intensity profile (blue) differs from direct calculation (red) by 0.83\% root mean squared error and while requiring significantly more computation time.}
\label{FIG:Square_515}
\end{figure}
Fig. \ref{FIG:Square_515} displays the UV temporal intensity profile calculated first by running a full simulation of the crystal dynamics (blue) and then approximated by assuming the UV field is given by the intensity profile of the 515 nm light (red). The approximated field in this case displays all of the characteristics of the calculated field including the residual ringing on the sides of the pulse in addition to having a nearly identical full width at half maximum (FWHM). The root mean squared error between the approximated intensity profile and the calculated one is only 0.83\%. This approximation would cease to be valid in the case where the second harmonic field starts converting energy back to the incident field; however, this is easily mitigated in a real world situation by lowering the intensity in the crystal or using decreasing crystal length.
By using this approximation rather than the full simulation we were able to reduce the computational complexity of the problem and save nearly a full day of processing time for less than 1\% of difference.
\section{Pulse shape tolerance to variance of $\alpha$ and filter width}
\begin{figure}[htp!]
\centering
\includegraphics[width=0.7\linewidth]{alpha_var.pdf}
\caption{Pulse shape changes of the 0.5 nm filter 515 nm pulse with the GA optimized parameters (SOD: $\pm 3.436$, $\alpha$: -0.128) to percent variations in $\alpha$ on one input pulse to mimic real world variance and experimental imperfections. As the variance in $\alpha$ is increased (left to right) the side oscillations are asymmetrically changed along with the slope of the filtered pulse.}
\label{FIG:3x6_Alpha_Var}
\end{figure}
\begin{figure}[htp!]
\centering
\includegraphics[width=0.7\linewidth]{filter_var.pdf}
\caption{Pulse shape changes of the 0.5 nm filter 515 nm pulse with the GA optimized parameters (SOD: $\pm 3.436$, $\alpha$: -0.128) to changing filter widths. As the filter widths increase the side oscillations are not as attenuated due to increase high frequency contributions.}
\label{FIG:Filt_Changes}
\end{figure}
As demonstrated in Sec. 1, $\alpha$ is the dominent parameter on controlling the presence of side oscillations in the mixing process with larger $\alpha$ indicating more TOD and more oscillations. When both input pulses experience a uniform variance in $\alpha$ the resulting mixed pulse shape has the side oscillations uniformly changed, based on the sign of the change in $\alpha$. However, when a single input pulse experiences a change in $\alpha$ that is unmatched on the other input pulse, the mixed pulse begins to show a loss of symmetry. Variations of this nature would be nearly unavoidable in a real world implementation of our proposed method.
As seen in Fig.\ref{FIG:3x6_Alpha_Var}, the unfiltered profile of a mixed pulse has a drastically changed profile for even a 5\% decrease in $\alpha$ of one input pulse with respect to the other and this change is further exasurbated when the difference is increased to 10\%. When these profiles are then filtered the difference in pulse shape is reduced to the point that the 5\% difference displays a very similar profile. The filtered 10\% difference case shows a reduction in the effect but still results in an asymmetric shape that is likely to affect an intended application. It is worth noting that this asymmetric degradation is seemly linear and might result in a tunable parameter between square pulse and triangular pulse shapes, should TOD be independently controllable.
Another parameter that has and effect on the resulting pulse shape is the width of the filter used after mixed pulse synthesis. Like $\alpha$, a wider filter will generally result in a greater retention of the side band oscillations.
In Fig \ref{FIG:Filt_Changes} we have generated a mixed pulse using the GA optimal parameters for a 0.5 nm filter (SOD: $\pm 3.436$, $\alpha$: -0.128) and then change the filter width while maintaining SOD and . This results in the intuitive change that larger filters retain more of the high frequency components that interfere with the main spectral lobe and thus retain more side oscillations. While the SOD and $\alpha$ values used are not identical to the GA optimal parameters for 0.7 nm and 1.0 nm filters, they are similar enough to remain a valid qualitative comparison to the changing of filter widths.
|
\section{Introduction}
Even though the idea of a ``canonical'' $1.4 M_{\odot}$ mass for compact stars has been around for years, it became clear in the last decade or so that this single-scale is not tenable \cite{HV}. At least one additional peak, and most likely two \cite{nos1} are present in the mass distribution, irrespective of the large uncertainties in many objects, particularly in those at the highest masses \cite{nos2}. The issue of a maximum compact star mass is also under discussion, since the latest accurate measurement of a compact object using Shapiro delay \cite{Cromartie} yielded $2.14{{+0.1}\atop{-0.09}} M_{\odot}$ for the millisecond pulsar MSP J0740+6620. A few higher measured values exist, although they have been extracted using methods not generally as reliable. Anyhow it has been argued that a handful of compact stars in the group of interacting ``redback/black widow'' systems may achieve masses greater than the MSP J0740+6620 \cite{Cromartie}, and possibly all the way up to the maximum TOV value, due to their evolutionary histories \cite{SciChi}. It is even possible to produce some ``low-mass'' black holes if some of these systems push the mass of the accreting pulsar over the TOV limit.
On the other hand, and as a valuable tool to complement the knowledge of compact stars in binaries, the availability of gravitational wave data has provided exciting news about the systems merging presently. While the celebrated GW170817 event \cite{abbott/2017} inaugurated the Multimessenger Era, being detected by more than 60 instruments around the world, a few other intriguing detections and candidates are available, although the merging rate does not appear to be as high as originally expected \cite{Luciano}.
One particular event denoted as GW190814 has shown a very intriguing component: an individual mass in the range $2.5-2.67 M_{\odot}$ ($90\%$ confidence) \cite{Abbott2020}. In the absence of additional information, it is not clear whether this is a light black hole or rather a very massive compact star (a cumbersome nickname of ``black neutron star'' \cite{BNS} was used in the media to reflect this ambiguity, although its precise meaning is not related to a ``hybrid'' character and may be misleading). Then, if the second possibility holds, it may be necessary to enlarge the maximum mass value to accommodate it. This in turn would be very important for the microphysics of the dense matter above the saturation density. Some works have already appeared discussing this possibility \cite{Ignazio,Xu}.
Motivated by this evidence, we revisit in this article a class of self-bound stellar models entertained in the last decade to provide viable stellar sequences complying with a high maximum mass. We shall present these solutions below, together with a comparison with the recently published data about the radius of a ``lighter'' compact star, namely the pulsar PSR J0030+0451 \cite{Col}, the only case with a reliable determination of the radius at present. It will be shown that there is a quandary with an explanation of both data simultaneously. Finally, we point related additional information related to the maximum mass problem.
In our calculations we shall consider three different
quark matter models with and without anisotropy in the interior pressure. The quark matter models will be described in the next section. For now it is worth mentioning that anisotropy may arise naturally in matter fields at high densities ($\rho>10^{15}$g/cm$^3$) \cite{ruderman/1972,canuto/1974} and play a fundamental role in the interior of compact objects. The natural examples of anisotropy are abundant; here we quote electromagnetic and fermionic fields in neutron stars \cite{sawyer/1973} and superfluidity \cite{carter/1998}. In fact, since the pioneering Reference 16, there has been an considerable number of works \cite{mak/2003,mak/2002,hernandez/2004,maurya/2019,mardan/2019} devoted to the study of anisotropic spherically symmetric static stellar configurations. Let us finally mention that the origin of local anisotropy was investigated in the review of Ref.22 and it was shown that a possible source may be viscosity (see also Ref. 23). Anisotropic models may be important to model very massive compact stars, but we shall see that the class of self-bound versions run into trouble to accommodate radii.
\section{Self-bound matter and stellar models}
The physics of matter at ultra-high densities can be studied in principle with the aid of \textit{Quantum Chromodynamics Theory} (QCD). For large temperatures and densities, QCD matter is \textit{asymptotically free}, the actual region of the phase diagram in which neutron stars or strange stars (SSs) reside present a high uncertainty in the matter behavior \cite{QCD}.
More than 30 years ago, Witten \cite{c} considered the possibility of a self-bound version of quark matter, made stable by the presence of the $s$ quark, elaborating on the previous works of Bodmer \cite{Bod}, Terazawa \cite{Tera} and Itoh \cite{Itoh}. Because of this feature the proposal is known as {\it strange matter}. In this sense, an exotic state of deconfined quarks, could be the true ground state of hadronic matter, not $^{56}Fe$, having a lower energy per baryon than ordinary nuclei. It is generally assumed that in its present version, strange matter is composed of roughly equal numbers of \textit{up}, \textit{down} and \textit{strange} quarks, and a small number of electrons to attain the charge neutrality \cite{farhi/1984,bethe/1987}.
A variety of approaches for a model description of strange matter and, in general, self-bound matter were attempted since Witten's work. The MIT bag model with a quasi-linear equation of state (EoS) has been widely used, but Nambu-Jona-Lasinio \cite{Buballa,Efrain}, density-dependent quark masses \cite{ZhangLi} and a few other variants were also considered. More recently it has been established that {\it paired} quark states should be relevant in the dense deconfined phase. It has been shown that the presence of pairing gaps (quark matter in the Color-Flavor Locked state, neglecting states at intermediate densities for simplicity) actually {\it enhance} the possible stability of the quark matter phase, as discussed by Lugones and Horvath \cite{e,f} because the system's energy is lowered by the negative contribution to the energy from the introduction of a pairing gap. A general parametric study of this possibility has been discussed by Alford and collaborators \cite{Mark}.
A suitable EoS including pairing effects due to the CFL, discussed in Ref. 37, reads
\begin{equation}
\label{eq:t}
P_r = \frac{1}{3}\rho + \frac{2\psi}{\pi}\rho^{1/2} - \left( \frac{3\psi^2}{\pi^2} + \frac{4}{3}B \right),
\end{equation}
where $\psi$ is defined by
\begin{equation}
\label{eq:y}
\psi = -\frac{m_s^2}{6}+\frac{2\Delta^2}{3},
\end{equation}
with $m_s$ being the mass of the strange quark, $\Delta$ is the pairing gap and $B$ is the bag constant. If we assume $m_s \rightarrow 0$ and non-interacting quarks, a \textit{MIT bag model}-like EoS is restored \cite{AFO}.
Other treatments of the quark matter have been presented, and their relevance to the self-boundedness problem addressed. We shall address the model by Franzon et al. \cite{Bruno} applied to the stellar sequences below.
As we have previously mentioned, in our calculations below we shall consider three different quark matter models with and without anisotropy in the interior pressure. Those are NJL Color-flavor locked EoS with vector interactions, Mean Field Theory of QCD (MFTQCD) and Thirukkanesh-Ragel-Malaver {\it ansatz} exact anisotropic models.
\subsection{NJL Color-flavor locked equation of state with vector interactions model}
\qquad
Motivated by the introduction of strange matter, Ferrer \cite{Efrain} discussed a model based on the Nambu-Jona-Lasinio approach, in which vector interactions and gluon components were introduced. This work shows that, even without a gluonic component, NJL EoS can be used to reach high maximum masses in the corresponding sequences provided the ratio of the (repulsive) vector channel to the quark-antiquark $G_{V}/G_{S}$ is high enough. One advantage of this approach is that the gap equation in the CFL phase is calculated in a self-consistent way. In addition, it is well-known that a quantity analogous to a bag constant can also be obtained \cite{Buballa}. The results of Ref.32 for the calculated $B_{0} = 57.3 MeV/fm^{3}$ (note that this quantity numerically coincides with a "MIT bag constant" $B$, but it is determined self-consistently) and $G_{V}/G_{S} = 0.5$ were employed to plot the curve in Fig. 1 below. The introduction of a gluonic component elevates even more the value of the maximum mass for this fixed set $(B_{0}, G_{V}/G_{S})$.
\subsection{MFTQCD (Mean Field Theory of QCD)}
The microphysical model of cold quark matter presented in Ref.39 starts with a separation of ``soft'' and ``hard'' momentum components, and after an analytical calculation one is left with an improved version of a Bag model, termed as MFTQCD. The EoS depends on two quantities: one is $\xi$, the ratio of the coupling to the dynamical gluon mass generated by interactions, and the other is related to the expectation values of the gluon condensates, identified as a ``vacuum energy constant'' and denoted as $B_{QCD}$. Given reasonable values for both quantities, a region in the $\xi - B_{QCD}$ plane in which the self-boundedness condition, namely that the energy per baryon number unit is $\leq m_{n}$ (or, more precisely, the mass of $^{56} Fe/56$) is satisfied. The pairing of quarks was not considered in Franzon et al.\cite{Bruno}, although it can be easily introduced. A stellar sequence generated using the set $\xi = 0.003658 \, MeV^{-1}$ and $B_{QCD} = 62 MeV/fm^{3}$ unpaired quark matter achieved the highest maximum TOV mass, and these are the values shown in Fig.1.
\subsection{Thirukkanesh-Ragel-Malaver {\it ansatz} exact anisotropic models}
\label{subsec:Malaver}
Finally we plot the result for two values of the parametric vacuum energy $B$ that generated anisotropic stellar models based on the Thirukkanesh-Ragel-Malaver {\it ansatz} as proposed by Thirukkanesh and Ragel \cite{k} and employed by Malaver \cite{l} for the spacetime metric. With the EoS of Eq.(1), Rocha et al.\cite{Gardel} were able to integrate analytically the full problem and obtain stellar sequences with very
high mass, as shown in Fig. 1 for these two values of the vacuum energy density and fixed strange quark mass
$m_{s}$ and pairing gap $\Delta$. The exact solutions display a feature which is not uncommon in a variety of anisotropic stellar models, and it is worth to mention that a general procedure to generate all static, spherically symmetric solutions has been recently presented \cite{aniso}. We believe that these kind of solutions will attract much attention in the near future.
\begin{figure}[htbp!]
\centering
\includegraphics[width=4in]{Graph1.jpg}
\caption{\label{fig:k} Mass-radius relation of the theoretical models. The curve generated using the MFTQCD quark EoS, with $\xi = 0.003658 \, MeV^{-1}$ and $B_{QCD} = 62 MeV/fm^{3}$ appears in green. The NJL CFL+vector interactions EoS with $B_{0} = 57.3 MeV/fm^{3}$ and $G_{V}/G_{S} = 0.5$ is the blue curve. The two anisotropic models calculated using the Thirukkanesh-Ragel-Malaver ansatz, both with $\Delta = 100~\mathrm{MeV}$ and $m_s = 150~\mathrm{MeV}$ are the red curve (corresponding to $B = 57.5~\mathrm{MeV/fm^3}$) and the black one ($B = 70~\mathrm{MeV/fm^3}$). The range of masses reported for the lighter object in the merge GW190814 is indicated by the dashed horizontal lines, and the region of the measured values for PSR J0030+0451 with the filled rectangle.}
\end{figure}
\section{Other models for the massive object and implications for the equation of state and stellar physics}
Our considerations of some examples of microphysical models that can be used to model a high-mass
compact star within the class of quark-based models are not exhaustive, other proposals have been published along these lines. There other effects that may allow an effective $M_{max} \geq 2.5 M_{\odot}$. The examples of Most et al. \cite{Most} and Zhang and Li \cite{Nai-Bo}, among others,
show that a rapidly rotating compact star could be involved. As a general feature, the analysis known as I-Love-Q has been often invoked to state that the amount of mass increase due to rotation is $\sim 20 \%$ for almost all equations of state \cite{chin}, although this figure actually depends on the kind of rotating and exact scaling may be broken \cite{Kostas}. Even though there is no clear indication for a rapidly rotation object in the data, this possibility for increasing the mass over the static limit can not be ruled out. Dexheimer et al.\cite{VD} suggest that the rapid rotation does {\it not} preclude an exotic core composition either, while Li, Sedrakian and Weber \cite{Armen} found that $\Delta$-resonance admixed hypernuclear constitution is in difficulty for producing such a massive configuration, even if maximally rotating at the Keplerian rate, and the object is likely to be a black hole. However, they have also shown that if no new degrees of freedom appear to ``hyperonize'' matter, the stiffness would be enough for a compact star interpretation of the object \cite{Sedra}. The black hole interpretation is also favored by Fattoyev et al. \cite{Fatto} after showing that a large stiffening of the equation of state within a covariant density functional theory approach makes difficult to satisfy constraints from heavy ion collisions and deformability of the lower-mass neutron stars obtained from the event GW170817.
Finally, Tsokaros, Ruiz and Shapiro \cite{soco} have argued that the analysis of other systems do not support the rapid rotation idea, and therefore the equation of state is the key element that must be very stiff if the lighter object in the merge GW190814 happens to be a compact star and not a black hole. However, we find very suggestive that a population analysis by the LIGO/Virgo Collaboration found a ``gap'' in the merging population \cite{LIGOpop} analogous to the one suggested in the local environment, namely the absence of masses in the range $\sim 2.5 - 6 M_{\odot}$, which puts the mysterious objects on the ``neutron star'' side. The black hole interpretation, however, would imply a gap among their distribution of masses, suggesting that the ``low-end'' follow a different
formation path (see below), unless the sample is very incomplete indeed.
\section{Discussion}
\qquad
We have presented a few representative examples of self-bound stellar models, two sequences of the isotropic pressure class with different EoS, and two anisotropic sequences of the same model corresponding to different sets of parameters of one model, with the aim of showing a general feature that became likely important for an explanation of a $\geq 2.5 M_{\odot}$ object in the event GW190814. Even though the latter is {\it not} confirmed to be a compact star, and may well be a black hole, there are arguments to believe that we should be prepared to find an explanation for a high mass, as suggested by other measurements in binaries \cite{SciChi}. In fact, this possibility has already been considered by Bombaci et al. \cite{Ignazio} and Wu et al. \cite{Xu} among others.
Versions of paired NJL CFL strange matter with vector interactions, as well as a MFTQCD without pairing were considered as isotropic examples. The first covered wide range of the parameters $B$, $m_s$ and $\Delta$, and the MFTQCD one was selected from the set of self-bound versions, as determined by the parameter $\xi$ to comply with a high maximum mass. While the anisotropic models obtained within the Thirukkanesh-Ragel-Malaver {\it ansatz} are exact within the quasi-linear EoS, the isotropic solutions were calculated numerically. None of these solutions introduced ultra-strong magnetic fields or any other ingredient.
An interesting general feature related to the behavior of the solutions is that up to about $\sim 1 M_{\odot}$, all self-bound stars are essentially Newtonian, and their radius for a given mass is just $R= {\bigl( \frac{3M}{4\pi \rho} \bigr)}^{1/3}$ to a good approximation. General relativistic effects make the radii at higher masses to be {\it smaller} than their Newtonian values (i.e. the $M-R$ curves like the ones in Fig. 1 are above their Newtonian equivalents). Therefore, the agreement with the measured data is worse than it would be
when $M \sim 1.5 M_{\odot}$, due to the General Relativity effects. Of course, this is a relatively minor effect when compared to the essential feature shared by the models: $R \rightarrow 0$ when $M \rightarrow 0$ for all self-bound stellar sequences, in sharp contrast with conventional hadronic models which have the opposite tendency, facilitating the interpretation of a relatively large radius for PSR J0030+0451 \cite{Col}. An exception within quark-based models has been recently presented by Zhang and Mann \cite{ZM}, and Roupas, Panotopulos and Lopes \cite{RPL} achieve both the mass and radius constraints but for extreme values of the parameters, as in Lugones and Horvath \cite{LugH}.
This situation may have different solutions. The first is that any of the observed quantities is accurate
but not precise, in other words, that either the mass of the object in GW190814 is overestimated (although a $\sim 2.5 M_{\odot}$ maximum mass was shown to stem from the observed distribution of binaries \cite{Alsing}), or the radius of the lower-mass PSR J0030+0451 is smaller, and thus the tension would be alleviated. The second solution is that both objects do {\it not} belong to the same class, in the sense that only very high masses contain exotic matter, but not the lighter ones. The third solution is, if both objects do belong to the same class, that self-bound models are not a good explanation of both simultaneously. A fourth solution is very simple: the $2.5 M_{\odot}$ was not a compact star, but rather a low-mass black hole (Lu, Beniamini and Bonnerot \cite{CalTech} favor this interpretation within a triple-system origin). Our discussion in Ref. 5 points out that at least a class of known binaries may be expected to produce such black holes, although the system in the merger GW190814 does not belong in any sense or could have originated in a ``spider'' binary, because the most massive component is surely a massive black hole. Other forms to produce low-mass black holes have been recently discussed (Liu and Lai\cite{LiuLai} 2020, Safarzadeh and Loeb\cite{SLoeb} 2020). Ultimately, the answer to the existence of absence of a ``mass gap'' is an important problem that must be solved gathering and analyzing empirical data.
It is remarkable that anisotropy is known to give a significant contribution to the increase in the maximum mass of compact objects. Recall that Fig. 1 shows that anisotropic models can reach maximum masses as high as $\sim 5 \, M_\odot$.
In fact, one of the earliest works by Heintzmann and Hillebrandt \cite{heintzmann/1974} has shown that anisotropy is a quite promising mechanism to increase the maximum masses of neutron stars, and for an arbitrarily large anisotropy there is no upper limit even when radial stability analysis is taken into account. Anisotropy yields higher maximum masses once the effective pressure inside stars is high enough to make the star to sustain more mass against gravitational collapse. It has also being pointed out in \cite{heintzmann/1974} that this is the case when $P_\perp>P_r$. On the other hand, if $P_r>P_\perp$ the maximum mass is expected to decrease. A very relevant discussion has been also presented by Bowers and Liang \cite{bowers/1974}.
Finally, it is important to state that a confirmation of the compact star character of the puzzling object in GW190814 would automatically establish an absolute upper limit to the central density $\rho_{c}$ of any compact star, as shown by Lattimer and Prakash \cite{Jim}. The Tolman IV and Tolman VII exact solutions provide an envelope for all stellar models, including exotics, which for the former reads
\begin{equation}
\rho_{c} \simeq 1.56 \times 10^{16} {(M_{\odot}/M)}^{2} g \, cm^{-3}
\end{equation}
which is deemed appropriate for microphysical quark models in a broad sense, and yields immediately
$\rho_{c} \simeq 2.18-2.5 \times 10^{15} g \, cm^{-3}$ for the range of the determined mass.
\section{Acknowledgments}
\qquad
J.E. Horvath acknowledge the Fundac\~ao de Amparo
\`a Pesquisa do Estado de S\~ao Paulo and CNPq Federal Agency for partial financial support.
P.H.R.S. Moraes would like to thank CAPES for financial support. An anonymous referee is acknowledged
for providing additional references and observations that helped us to improve the final version.
|
\section{Introduction}
The strong shock wave released from the metal--vacuum/gas interface may eject a great number of metal particles.\cite{Sollier:1,Monfared:1,Asay:1,Speight:1,Ogorodnikov:1} Most of these particles are of micrometer-scale in size. This phenomenon of ejecta, or microjetting, was first observed by Kormer et al. in a plane impact experiment in the 1950s.\cite{Ogorodnikov:1} And the earliest available technical report on ejecta is from research by the Atomic Weapons Research Establishment, Aldermaston(UK).\cite{Bistow:1} The physics of ejecta are understood as a special limiting case of impulse driven Richtmyer–Meshkov.\cite{Richtmyer:1, Meshkov:1} In recent decades, extensive investigations on particle ejection have been performed because of its important role in many scientific and engineering fields, including explosion damage,\cite{Yeager:1} pyrotechnics,\cite{Held:1} and inertial confinement fusion.\cite{Tokheim:1,Masters:1} Many experimental approaches have attempted to measure the ejection production, such as the Asay foil,\cite{Asay:1,Asay:2} foam recovery,\cite{He:1} piezoelectric probes,\cite{Speight:1,Vogan:1} Fraunhofer holography,\cite{Sorenson:1,Sorenson:2} X-ray/proton radiography,\cite{Monfared:1,Hammerberg:1} Mie scattering,\cite{Monfared:2,Hammerberg:1}, and photon Doppler velocimetry (PDV).\cite{La:1, Ogorodnikov:2, Andriyash:1, Franzkowiak:1, Sun:1, Fedorov:1, Arsenii:1} The main quantities of interest are the particles' velocity, diameter, and total area mass. Most approaches can only measure some of these ejecta parameters. To reveal the full particle field of ejecta, multiple measurement approaches must be equipped. However, in real-world conditions, these approaches are hard to apply simultaneously because of limitations on the experimental space or configuration. Recently, PDV has attracted considerable attention\cite{Andriyash:1, Franzkowiak:1, Arsenii:1} owing to its ability to recover the total area mass and the distributions of particle velocity and diameter at the same time. In addition, the light path of PDV is rather concise and its application is convenient. In some complex experimental configurations, PDV may be the only approach that can measure the ejecta particles.
A standard PDV setup is shown in Fig.~\ref{fig1}. The photodetector records a mixture of reference and backscattering light. The reference wave is in the carrier frequency, and the backscattering wave from ejecta particles has a shifted frequency due to the Doppler effect. The interference of the two light waves in the photodetector leads to temporal beats of light intensity. The beat signal consists of a large number of harmonics with different amplitudes and phases. The heterodyne signal may change according to variations in the particles' position and velocity. A discrete Fourier transform is applied to sweep the beats over time, giving a two-dimensional spectrogram on the ``frequency/velocity--time'' plane. In the spectrogram, the brightness of each point represents the corresponding spectral amplitude. The spectrogram is composed of the integral of all particles' scattering effects. Hence, interpreting the spectrogram in detail remains a challenging task.
\begin{figure}
\centering
\includegraphics[width=0.38\textwidth]{fig1.eps}
\caption{ Standard PDV setup. (1) Laser; (2) Reference light; (3) Incident light; (4) Ejecta; (5) Metal plate; (6) Shock; (7) Detonation; (8) Backscattering light; (9) Optical circulator; (10) Photodetector; (11) Photoelectric signal; (12) PDV spectrogram; (13) Instantaneous spectrum.}
\label{fig1}
\end{figure}
There have been several studies on the interpretation of the PDV spectrum. Buttler\cite{Buttler:1,Buttler:2} used the spectrogram boundaries to determine the velocities of the spike and bubble of Richtmyer--Meshkov instability in loaded metal surface. The evolution of the PDV spectrogram in a gas environment was discussed by Sun et al.,\cite{Sun:1} and the upper boundary of the spectrogram was used to obtain the particle size by considering aerodynamic deceleration effects. Fedorov et al.\cite{Fedorov:1} discussed the influence of different particle sizes on the spectrogram boundary in further detail. Recently, Franzkowiak et al.\cite{Franzkowiak:1} and Andriyash et al.\cite{Andriyash:1,Arsenii:1} reconstructed the light field of ejecta and obtained the simulated PDV spectrum using single- and multiple-scattering theory, respectively. They varied the particles' parameters and fitted the simulated PDV spectrum to the experimental data. In this way, the particle velocity profile, diameter, and total area mass were recovered. Andriyash et al. considered the aerodynamic deceleration effects in a gas environment, whereas Franzkowiak et al. only discussed the case of a vacuum.
Franzkowaik et al.\cite{Franzkowiak:1} and Andriyash et al.\cite{Andriyash:1,Arsenii:1} proposed similar approaches for recovering the ejecta parameters from the PDV spectrum through reconstruction and then fitting. However, some assumptions were introduced in the reconstruction of the light field. Franzkowaik et al. assumed that only backscattering light was present, while Andriyash et al. set the light scattering direction to be uniform and random in space. These assumptions affect the accuracy of the spectrum reconstruction, and thus influence the recovery of the ejecta parameters. The fitting model is another factor that affects the interpretation of the PDV spectrum. Different convergence criteria may produce different results. The quantitative relationship between the ejecta parameters and the PDV spectrum remains unclear. Hence, it is difficult to obtain definite ejecta parameters from the PDV spectrum. These issues provide the motivation for the present work.
In this study, we improve the reconstruction method of the ejecta light field, and propose a novel model for extracting the ejecta parameters directly from the PDV spectrum. Mie theory, which gives a rigorous mathematical solution to Maxwell's equations, is applied to calculate the light scattering effects, and a Monte-Carlo (MC) algorithm is used to describe the light transport process realistically. This reconstruction method provides a high-fidelity simulation for the PDV spectrum. The procedure is discussed in detail in Section II. The influence of the ejecta parameters on the PDV parameters is then explored through MC simulations in Section III.A. In Section III.B, we propose an optical model that reveals the relationships between the PDV spectrum characteristics and the ejecta parameters. With this model, the ejecta parameters can be extracted directly from the PDV spectrum, instead of fitting to experimental data. In Section III.C, the estimated ejecta parameters from an experimental PDV spectrum are verified against those measured by a piezoelectric probe. Finally, the conclusions to this study are presented in Section IV.
\section{Reconstruction of PDV spectrum}
\subsection{Theoretical background}
The photodetector records reference and backscattering light waves. The scattering process of incident light is illustrated in Fig.~\ref{fig2}. The scattering light from the ejecta is governed by the superposition of waves propagating in the ejection particles along different light paths \(i\):
\begin{equation}
{E_{bs}}\left( t \right) = \sum\limits_i {{E_i}\left( t \right)}
\label{eq1}
\end{equation}
where \(E_{bs}\) and \(E_{i}\) are the electric vectors of total and partial scattering waves, respectively.
The light intensity measured by the detector can be represented as:
\begin{equation}
\begin{array}{l}
I\left( t \right) = {\left( {{E_r}\left( t \right) + {E_{bs}}\left( t \right)} \right)^2}= {\left( {{E_r}\left( t \right) + \sum\limits_{i} {{E_i}\left( t \right)} } \right)^2}\\
= E_r^2\left( t \right) + \sum\limits_{i} {E_i^2\left( t \right)} + 2\sum\limits_{i} {{E_r}\left( t \right){E_i}\left( t \right) + } \sum\limits_{i \ne j} {{E_i}\left( t \right){E_j}\left( t \right)}
\end{array}
\label{eq2}
\end{equation}
where \(E_r^2\) and \(E_i^2\) denote the intensity of the reference and scattering light signals, respectively. The third term represents the heterodyne beats between the reference and scattering light, and the last term represents the heterodyne beats between the different scattering light paths. The Fourier transform of \(I(t)\) is determined by the relation:
\begin{equation}
\begin{array}{l}
I\left( \omega \right) = \int {dt\exp \left( {i\omega t} \right)} I\left( t \right)\\
\approx \int {dt\exp \left( {i\omega t} \right)} \left( {2\sum\limits_{i} {{E_r}\left( t \right){E_i}\left( t \right)} } \right)\\
= 2{\left| {{E_r}} \right|} \cdot {\left. {\sum\limits_{i} {\left| {{E_i}} \right|} } \right|_{\omega = {\omega _i} - {\omega _r}}}
\end{array}
\label{eq3}
\end{equation}
where \(\left| {{E}} \right|\) is the amplitude of the light wave field. In Eq.~(3), only the third term appearing in Eq.~(2) remains. This is because the frequencies of the first two terms are too high to be measured by the detector and the value of the last term is much smaller than that of the third term. \(\omega_r\) is the carrier frequency, and \(\omega_i\) is the Doppler-shifted frequency, which corresponds to a sequence of scattering events along path \(i\):
\begin{equation}
{\omega _i} = \frac{{{\omega _r}}}{c}\sum\limits_k {\left( {{\bf{n}}_{k,i}^s - {{\bf{n}}_{k,i}}} \right){{\bf{v}}_{k,i}}}
\label{eq4}
\end{equation}
where \(c\) is the speed of light, \({\bf{n}}_{k,i}\) and \({\bf{n}}_{k,i}^s\) are the directions of wave propagation before and after scattering by particle \(k\), and \({\bf{v}}_{k,i}\) is the velocity of this particle.
\begin{figure}
\centering
\includegraphics[width=0.4\textwidth]{fig2.eps}
\caption{Multiple scattering of light waves in ejection particles.}
\label{fig2}
\end{figure}
To reconstruct the PDV spectrogram [Eq.~(\ref{eq3})], the key is to obtain \(\left| {{E_i}} \right|\) and \(\omega_i\), i.e., the detailed scattering process in the particles. For multiple-particle systems, the scattering field can be described by the transport equation:\cite{Ishimaru:1, Reguigui:1, Binzoni:1}
\begin{equation}
\begin{split}
& \left( {\bf{n}\frac{\partial }{{\partial r}} + \sigma_s + \kappa } \right)I\left( {{\bf r,n},t} \right) \\
& = \int {\left\langle {\sigma_s p\left( {\bf n,n'} \right)\exp \left( {i{k_0}\left( {\bf n - n'} \right){\bf v}t} \right)} \right\rangle I\left( {{\bf r,n'},t} \right)d \bf n'}
\end{split}
\label{eq5}
\end{equation}
where \(I\) is the light intensity in the field, which depends on both the detection position \(\bf r\) and the direction \(\bf n\). \(\sigma_s\) and \(\kappa\) are the coefficients of scattering and absorption, respectively. \(p\left( {\bf n,n'} \right)\) is the scattering phase function. The right-hand side of this equation represents the contributions of scattering light from other positions.
At the boundaries of the ejection, the light intensity has the form:
\begin{equation}
\left\{ {\begin{array}{*{20}{l}}
{I\left( {{\bf{r}} = 0,{\bf{n}} = {{\bf{n}}_{\bf{0}}},t} \right) = {I_0}}\\
{I\left( {{\bf{r}} = 0,{\bf{n}} = - {{\bf{n}}_{\bf{0}}},t} \right) = {I_{bs}} = {{(\sum\limits_i {\left| {{E_i}} \right|} } )^2}}
\end{array}} \right.
\label{eq6}
\end{equation}
where \(\bf n_0\) is the direction of incident light, which is usually perpendicular to the free surface. \(I_0\) and \(I_{bs}\) are the intensities of incident light and backscattering light, respectively, which correspond to the input and output of the transport equation.
\subsection{Monte-Carlo algorithm}
Andriyash et al.\cite{Andriyash:1,Arsenii:1} used the discrete ordinate method to solve the transport equation [Eq.~(\ref{eq5})]. In this paper, a more convenient and accurate method of the MC algorithm is applied to calculate the scattering effects.
In the MC algorithm, the incident light is assumed to be a great number of photons. When passing through random granular media, only part of the photons can penetrate. The proportion of permeable photons is approximated by Beer--Lambert's law:\cite{Bohren:1}
\begin{equation}
p_r =\exp \left( { - \tau L} \right)
\label{eq7}
\end{equation}
where \(L\) is the thickness of the medium and \(\tau \) denotes the inverse extinction length, given by:
\begin{equation}
\tau = N{K_{ext}}\bar A = \frac{{\sum\limits_i {\frac{\pi }{4}{d_i}^2 \cdot {K_{ext}}\left( {{d_i}} \right)} }}{{{A_s}L}}
\label{eq8}
\end{equation}
where \(N \) is the number of particles per unit volume, \(K_{ext} \) is the light extinction coefficient (determined by the particle diameter, light wavelength, and metal relative refraction index), \(\bar A\) is the mean cross-section area of the particles, \(d_i\) is the diameter of particle \(i\), and \(A_s\) is the light exposure area.
For particles in the light exposure area, the total mass has the form:
\begin{equation}
{m_0}{A_s} = \sum\limits_i {\frac{1}{6}\pi {d_i^3}{\rho _0}}
\label{eq9}
\end{equation}
where \(m_0\) is the total area mass of ejection and \(\rho_0\) is the density of the particle material.
Combining Eqs.~(\ref{eq8}) and (\ref{eq9}), we can rewrite Eq.~(\ref{eq7}) as:
\begin{equation}
{p_r} = \exp \left( { - \frac{{3{m_0}\sum\limits_i {d_i^2{K_{ext}}\left( {{d_i}} \right)} }}{{2{\rho _0}\sum\limits_i {d_i^3} }}} \right)
\label{eq10}
\end{equation}
The photons staying in the medium are scattered or absorbed by particles. The probabilities of scattering and absorption are calculated by the formula:
\begin{equation}
\left\{ \begin{array}{l}
{p_s} = {K_{sca}}/{K_{ext}}\\
{p_a} = {K_{abs}}/{K_{ext}}
\end{array} \right.
\label{eq11}
\end{equation}
where the scattering coefficient \(K_{sca}\) and the absorption coefficient \(K_{abs}\) are calculated by Mie theory\cite{Mie:1} as:
\begin{equation}
\left\{ \begin{array}{l}
{K_{ext}} = \frac{2}{{{\alpha ^2}}}\sum\limits_{n = 1}^\infty {\left( {2n + 1} \right){\mathop{\rm Re}\nolimits} \left( {{a_n} + {b_n}} \right)} \\
{K_{sca}} = \frac{2}{{{\alpha ^2}}}\sum\limits_{n = 1}^\infty {\left( {2n + 1} \right)\left( {{{\left| {{a_n}} \right|}^2} + {{\left| {{b_n}} \right|}^2}} \right)} \\
{K_{abs}} = {K_{ext}} - {K_{sca}}
\end{array} \right.
\label{eq12}
\end{equation}
where \(\alpha\) is a dimensionless particle diameter parameter, \(\alpha = \pi d/\lambda \), \(\lambda\) is the light wavelength, and \(a_n\), \(b_n\) are Mie coefficients. It is clear that \(p_s+p_a=1\).
If the photon is absorbed by particles, it will completely disappear and be converted into the particle's internal energy. If the photon is scattered, its propagation direction and frequency will change, as shown in Fig.~3. The phase function of the scattering angle \(\theta\) is calculated by the formula:
\begin{equation}
p(\theta ) = \frac{{{\lambda ^2}}}{{2\pi {K_{sca}}}}({\left| {{S_1}(\theta )} \right|^2} + {\left| {{S_2}(\theta )} \right|^2})
\label{eq13}
\end{equation}
where \(S_1\) and \(S_2\) denote the scattering intensity components in the perpendicular and parallel directions:
\begin{equation}
\left\{ \begin{array}{l}
{S_1}\left( \theta \right) = \sum\limits_{n = 1}^\infty {\frac{{2n + 1}}{{n\left( {n + 1} \right)}}\left( {{a_n}\frac{{d{P_n}\left( {\cos \theta } \right)}}{{d\cos \theta }} + {b_n}\frac{{dP_n^{(1)}\left( {\cos \theta } \right)}}{{d\theta }}} \right)} \\
{S_2}\left( \theta \right) = \sum\limits_{n = 1}^\infty {\frac{{2n + 1}}{{n\left( {n + 1} \right)}}\left( {{a_n}\frac{{dP_n^{(1)}\left( {\cos \theta } \right)}}{{d\theta }} + {b_n}\frac{{d{P_n}\left( {\cos \theta } \right)}}{{d\cos \theta }}} \right)}
\end{array} \right.
\label{eq14}
\end{equation}
where \(P_n\) and \(P_n^{(1)}\) are Legendre and first-order associative Legendre functions, respectively.
The incident light is assumed to be non-polarized, so the azimuth angle \(\varphi\) after scattering obeys the uniform random distribution:
\begin{equation}
p(\varphi)=\frac{1}{2\pi}
\label{eq15}
\end{equation}
\begin{figure}
\centering
\includegraphics[width=0.30\textwidth]{fig31.eps}
\caption{Light scattering on a particle.}
\label{fig3}
\end{figure}
After scattering, the scattering angle and azimuth angle are added to the original light direction. The new direction cosine \(\hat u = [\hat u_x, \hat u_y, \hat u_z] \) has the form:
\begin{equation}
\left\{ \begin{array}{l}
{{\hat u}_x} = \frac{1}{{\sqrt {1 - u_z^2} }}\sin \left( \theta \right)\left[ {{u_x}{u_y}\cos \left( \varphi \right) - {u_y}\sin \left( \varphi \right)} \right] + {u_x}\cos \left( \theta \right)\\
{{\hat u}_y} = \frac{1}{{\sqrt {1 - u_z^2} }}\sin \left( \theta \right)\left[ {{u_x}{u_z}\cos \left( \varphi \right) - {u_x}\sin \left( \varphi \right)} \right] + {u_y}\cos \left( \theta \right)\\
{{\hat u}_z} = - \sin \left( \theta \right)\cos (\varphi )\sqrt {1 - u_z^2} + {u_z}\cos \left( \theta \right)
\end{array} \right.
\label{eq16}
\end{equation}
where \(u = [u_x, u_y, u_z] \) is the original direction cosine. When \(\left| {{u_z}} \right| \approx 1 \), the direction cosine \(\hat u\) is calculated by the formula:
\begin{equation}
\left\{ \begin{array}{l}
{{\hat u}_x} = \sin \left( \theta \right)\cos \left( \varphi \right)\\
{{\hat u}_y} = \sin \left( \theta \right)\sin \left( \varphi \right)\\
{{\hat u}_z} = \cos \left( \theta \right){u_z}/\left| {{u_z}} \right|
\end{array} \right.
\label{eq17}
\end{equation}
After scattering, the frequency of the photon will have changed. The new frequency of scattering light has the form:
\begin{equation}
\hat \omega = \omega \left( {1 + \frac{{{\bf{\hat v}} - {\bf{v}}}}{c}{\bf{n}}} \right)
\label{eq18}
\end{equation}
where \(\bf{v}\) is the velocity of the last particle that the photon left, \(\bf{\hat v}\) is the current particle velocity, \(\bf n\) is the photon direction, and \(\omega\) is the photon frequency before scattering.
After passing through the entire ejection layer, few photons reach the free surface. An ideal diffuse reflection is assumed for these photons. After reflection, the space angles \(\theta\) and \(\varphi\) are uniformly random in \([\pi/2, \pi]\) and \([0, 2\pi]\), respectively.
Because the ejection is blocked by the free surface, eventually all the photons are either absorbed by particles or backscattered out from the top of the ejection. The frequency shifts of these ``out'' photons are summarized as the theoretical PDV spectrogram:
\begin{equation}
I\left( \omega \right) \propto \sum\limits_i {\left| {{E_i}} \right|} = \sum\limits_i {\sqrt {{I_i}} } = {\left. {{n_{out}}} \right|_{\omega = \hat \omega - {\omega _r}}} \cdot \sqrt {\frac{{{I_0}}}{{{n_0}}}}
\label{eq19}
\end{equation}
where \(n_0\) is the number of initial photons and \(n_{out}\) is the distribution of out photons in terms of their frequency.
The detailed steps of the calculation procedure are as follows:
(1) First, the initial conditions of the photons and particles are set, such as the number and frequency of photons, and the diameter, velocity, and position of the particles. The photons start at the top of the ejection and then move towards the free surface.
(2) The step sizes of all photons are set to the same and equal to one thousandth of the height of the ejecta.
(3) In one iteration, all photons take one step in the direction of their propagation. Some photons may penetrate the current ejection layer, and the proportion \(p_r\) is determined by Eq.~(\ref{eq10}). For each photon, a random number is generated in (0, 1). If the random number is less then \(p_r\), the corresponding photon travels over the ejection layer boundary successfully. Otherwise, the corresponding photon is absorbed or scattered by particles in the ejection layer.
(4) For the photons that remain in the ejection layer, we use Eq.~(\ref{eq11}) to determine whether they are scattered or absorbed. If the photon is scattered, the direction change is calculated by Eqs.~(\ref{eq13})--(\ref{eq17}). Because the phase function of the scattering angle is very complex, an acceptance--rejection method is applied. The new frequency of the scattered photons is determined by Eq.~(\ref{eq18}).
(5) Overall, if the photon travels across the ejection layer boundary, its position is updated; if the photon is scattered, its direction, frequency, and position are updated; if the photon is absorbed, it is labeled as such and removed from subsequent calculations.
(6) After updating the state of the photons, we check which of them have reached the free surface or left through the top of the ejection. For all photons that have reached the free surface, the ideal diffuse reflection is applied. If any photons have left the ejection, they are labeled accordingly and removed from subsequent calculations.
(7) Steps (3)--(6) are repeated until all photons have been absorbed or have left the ejection. The frequency shifts of outgoing photons are summarized as the spectrogram.
\subsection{Particle models}
The MC algorithm indicates that the PDV spectrum is related to the particle size \(d\), velocity \(v\), position \(z\), and number \(N\) (i.e., total area mass \(m_0\)). This algorithm can be applied in cases where these parameters are completely random. In real situations, however, the particles of the ejecta usually satisfy certain distributions in terms of velocity and diameter.\cite{Monfared:1,Monfared:2,Durand:1,Durand:2,Schauer:1,Sorenson:1,Sorenson:2,He:2} For the sake of discussion, these assumptions are applied in this paper. Previous studies\cite{Monfared:1,Durand:1} indicate that the initial velocities of particles in the ejecta can be approximated by an exponential law:
\begin{equation}
f\left( v \right) = \frac{{m(v)}}{{{m_0}}} = \frac{\beta }{{{v_{fs}}}}\exp \left[ { - \beta \left( {\frac{v}{{{v_{fs}}}} - 1} \right)} \right]
\label{eq20}
\end{equation}
where \(v_{fs}\) is the velocity of the free surface and \(\beta\) is the velocity distribution coefficient. Under this exponential law, most of the particles are located in the low-velocity region, which is near the free surface. \(\beta\) determines the non-uniformity of this distribution.
In this paper, we only consider the ejecta in a vacuum environment. After being ejected, the particles retain an almost constant velocity and the ejecta expands in a self-similar manner over time. The particle position \(z\) is only related to its initial velocity \(v\) and ejection time \(t_e\), \(z=vt_e\). The corresponding PDV spectrum exhibits slight changes over time.\cite{Franzkowiak:1,Bell:1}
The particle size distribution is assumed to obey a log-normal law:\cite{Schauer:1,Sorenson:2}
\begin{equation}
n\left( d \right) = \frac{1}{{\sqrt {2\pi } \sigma d}}\exp \left( { - \frac{{{{\ln }^2}\left( {d/{d_m}} \right)}}{{2{\sigma ^2}}}} \right)
\label{eq21}
\end{equation}
where \(\sigma\) is the width of the distribution and \(d_m\) is the median diameter. These parameters depend on the roughness of the metal surface, shock-induced breakout pressure, and surrounding gas properties. When $\sigma$ = 0, the function becomes a Dirac equation and all of the particles have the same diameter. Obviously, this is the ideal situation. The particle distribution can also be described by a power law,\cite{Durand:2, Sorenson:1, He:2} but this description may be invalid in the range of small particle sizes (less than 10\; \(\rm \mu m\)).\cite{Sorenson:2} In this paper, the particle size is assumed to be independent of its velocity.
With these assumptions, the determining factors of the PDV spectrum change to the velocity profile coefficient \(\beta\), total area mass \(m_0\), median diameter \(d_m\), and size distribution width \(\sigma\). The aim of this paper is to discuss the influence of these parameters on the PDV spectrum, and to explore how they can be extracted from the PDV spectrum most accurately.
\subsection{Convergence and comparison}
The accuracy of the MC algorithm mainly depends on the initial number of photons. Theoretical PDV spectra with \(10^4\), \(10^5\), \(10^6\), and \(10^7\) initial photons are shown in Fig.~\ref{fig4}. The calculation assumes a vacuum environment and the particles distribution assumptions are applied. In this case, the PDV spectra have a single peak. As the initial number of photons increases, the spectrum curves tend to be smooth. The difference between the spectra with \(10^6\) and \(10^7\) initial photons is very slight. Thus, \(10^7\) initial photons are applied in the following calculations.
\begin{figure}
\centering
\includegraphics[height=0.35\textwidth]{fig4.eps}
\caption{Theoretical PDV spectra with different initial numbers of photons. The calculation assumes the ejection of Sn particles in a vacuum environment. The particle velocities obey an exponential distribution \((\beta=10)\) and the particle sizes follow a log-normal distribution \((d_m=1.5\; \rm{\mu m}, \sigma=0.5)\). The total area mass is \(20\; \rm{mg/cm^2}\). \(\omega_{fs}\) is the Doppler frequency shift corresponding to the free surface velocity, \(\omega_{fs}=2\omega_r \cdot v_{fs}/c\). The probing wavelength is \(\lambda=1550\; \rm{nm}\).}
\label{fig4}
\end{figure}
The high initial number of photons leads to considerable computational cost. For the case of \(10^7\) photons, a single-core CPU requires approximately 3 h to determine the spectrum. GPUs can be applied to accelerate the calculation. Although the frequency of GPU processors is much lower than that of CPUs, GPUs contain hundreds or thousands of stream processors that can work simultaneously. The acceleration ratio of a GPU compared to a CPU is shown in Fig.~\ref{fig5}. The Intel Xeon W-2102 CPU (frequency 2.9 GHz) and two GPUs (Quadro P600 and Nvidia GTX960) are applied. As the initial number of photons increases, the acceleration ratio of the GPUs is enhanced. For the case of \(10^7\) photons, the acceleration ratio reaches a factor of 8 for the Quadro P600 and a factor of 20 for the Nvidia GTX960. Because there are many judgment events in the procedure, and the GPUs have few logical units, it is difficult to improve the acceleration ratio with these GPUs. Thus, the Nvidia GTX960, which requires approximately 500 s to compute each case, is used in the following calculations.
\begin{figure}
\centering
\includegraphics[height=0.35\textwidth]{fig5.eps}
\caption{Acceleration ratio of GPU calculation compared with CPU for different initial numbers of photons. The CPU is an Intel Xeon W-2102 and its basic frequency is 2.9 GHz. The Quadro P600 GPU has 384 stream processors; the clock speed of each processor is about 1.3 GHz. The Nvidia GTX960 GPU has 1024 stream processors; the clock speed of each processor is about 1.1 GHz.}
\label{fig5}
\end{figure}
The PDV spectra simulated by the present procedure are compared with those reported by Andriyash et al. and Franzkowiak et al. in Fig.~\ref{fig6}. We use the equivalent ejecta area mass and particle size instead of the transport optical thickness used by Andriyash et al. With the uniform scattering assumption, our simulation (Case 2) is almost the same as that of Andriyash et al. (Case 3), which validates the adequacy of the present numerical method. However, when the Mie scattering theory is applied, there is a remarkable difference between the present procedure (Case 1) and the results of Andriyash et al. (Case 3) and Franzkowiak et al. (Case 4). The difference with Andriyash et al. is mainly in the low-velocity part. This is because the change in the scattering phase function has a great influence on the multiple scattering, which is the main form of scattering in the low-velocity dense part. The difference with Franzkowiak et al. is in the location of the spectrum peak. Franzkowiak et al. applied the single-scattering theory and assumed that all of the light scattered backward. This implies that the optical thickness is overestimated, and so little light would reach the deep region of the ejecta. Thus, the spectrum moves toward high velocities. These differences in spectra indicate that the scattering assumption may introduce some reconstruction inaccuracy that cannot be neglected.
\begin{figure}
\centering
\includegraphics[height=0.37\textwidth]{fig6.eps}
\caption{Comparison of PDV spectra calculated by different reconstruction methods. Case 1: MC + Mie scattering theory (proposed procedure); Case 2: MC + uniform scattering assumption; Case 3: Discrete coordinates + uniform scattering assumption (Andriyash et al.); Case 4: Single scattering theory (Franzkowiak et al.) . The data for Case 3 were extracted directly from the paper of Andriyash et al. The calculations were carried out for a transport scattering thickness of \(\tau_{tr}=10\), which corresponds to \(m_0=10.7\; {\rm {mg/cm^2}}, d_m=1.5\; \rm {\mu m}\), and \(\sigma = 0.5\). The material is Sn and the ejection velocity profile has \(\beta=8\).}
\label{fig6}
\end{figure}
\section{Interpretation of PDV spectrum}
\subsection{Influence of ejecta parameters}
The results of numerical calculations that demonstrate the sensitivity of the PDV spectrum to changes in the ejecta parameters (\(\beta\), \(m_0\), \(d_m\), and \(\sigma\)) are presented in Figs.~\ref{fig7}--\ref{fig10}. The PDV spectra were simulated using the MC algorithm described in the previous section for Sn particles in a vacuum environment. The initial ejecta parameters were set to \(\beta=10\), \(m_0=20\; \rm{mg/cm^2}\), \(d_m=1.5\; \rm{\mu m}\), and \(\sigma=0.5\). In each figure, one of the parameters changes and the others remain constant.
The simulated PDV spectra with different values of the velocity coefficient \(\beta\) are shown in Fig.~\ref{fig7}. With an increase in \(\beta\), the spectrum peak moves towards the low velocities and its magnitude decreases. Furthermore, the spectrum shape becomes sharper and the high-velocity part of the spectrum becomes invisible. The coefficient \(\beta\) determines the distribution of particles in the ejecta. With larger \(\beta\), fewer particles are located at the top of the ejecta and the incident light can penetrate deeper. This results in the movement of the spectrum peak and a decrease in the observability of high-velocity particles.
\begin{figure}
\centering
\includegraphics[height=0.35\textwidth]{fig7.eps}
\caption{Simulated PDV spectra with different velocity coefficients. The calculation was carried out for Sn particles in a vacuum environment. The total area mass was \(20\; \rm mg/cm^2\). The log-normal distribution \((d_m=1.5\; {\rm \mu m}, \sigma = 0.5)\) was applied to the particle sizes.}
\label{fig7}
\end{figure}
The simulated PDV spectra with different values of the total area mass \(m_0\) are shown in Fig.~\ref{fig8}. The changes in the spectra can be divided into two sections. When \(m_0 \ge 10\; \rm mg/cm^2\), the spectrum displays a single peak. With a decrease in \(m_0\), this peak moves to the left, and its magnitude and slope exhibit slight changes. When \(m_0 \le 5\; \rm mg/cm^2\), a new peak appears around the free surface, and the spectrum exhibits a double-peak shape. A smaller area mass produces a more remarkable new peak. For \(m_0=2\; \rm mg/cm^2\), the original peak disappears and the spectrum again exhibits a single peak. The double-peak spectrum has been observed in previous experiments\cite{La:1,Andriyash:2} and simulations,\cite{Arsenii:1,Franzkowiak:2} and is the result of the direct exposure of incident light at the free surface.
\begin{figure}
\centering
\includegraphics[height=0.35\textwidth]{fig8.eps}
\caption{Simulated PDV spectra with different total area mass. The area mass unit is \(\rm {mg/cm^2}\). The calculation was carried out for Sn particles in a vacuum environment. The velocity coefficient \(\beta = 10\) and the size coefficients \(d_m=1.5\; \rm {\mu m}, \sigma = 0.5\).}
\label{fig8}
\end{figure}
There are two parameters that determine the particle size distribution---the median diameter \(d_m\) and the distribution width \(\sigma\). Their influence on the PDV spectrum is illustrated in Figs.~\ref{fig9} and \ref{fig10}, respectively. \(d_m\) and \(\sigma\) exhibit similar effects: as \(d_m\) or \(\sigma\) increases, the original peak of the spectrum moves towards the low velocities and the peak value decreases. A new peak then appears in the position of the free surface and the original peak gradually attenuates. This change in the form of the spectrum peak is similar to that for the area mass.
\begin{figure}
\centering
\includegraphics[height=0.35\textwidth]{fig9.eps}
\caption{Simulated PDV spectra with different particle median diameters. The diameter unit is \(\rm {\mu m}\). The calculation was carried out for Sn particles in a vacuum environment. The total area mass was \(20\; \rm {mg/cm^2}\). The velocity coefficient \(\beta = 10\; \rm{mg/cm^2}\) and the size coefficient \(\sigma = 0.5\).}
\label{fig9}
\end{figure}
\begin{figure}
\centering
\includegraphics[height=0.35\textwidth]{fig10.eps}
\caption{Simulated PDV spectra with different particle size coefficients \(\sigma\). The calculation was carried out for Sn particles in a vacuum environment. The total area mass was \(20\; \rm {mg/cm^2}\). The velocity coefficient \(\beta = 10\) and the median diameter \(d_m = 1.5\; \rm \mu m\).}
\label{fig10}
\end{figure}
\subsection{Theoretical optical model}
The simulations described above using the MC algorithm provide a qualitative understanding of the influence of the ejecta parameters on the PDV spectrum. However, how to solve the reverse problem, i.e., extracting the ejecta parameters from the PDV spectrum, remains unclear. To obtain the quantitative relationships between the ejecta parameters and the characteristics of the PDV spectrum, we introduce the single-scattering theory. In this theory, the light is assumed to be scattered only once, and the scattering direction is always backward. With this assumption, the light direction is always parallel to the motion of the particles. The frequency shift of the light is proportional to the particle velocity, $\omega=2\omega_0 \cdot v/c$. The PDV spectrum can be expressed in terms of velocity, $I (\tilde v)$.
The above calculations have shown that the single scattering leads to an overestimation of the optical thickness. Here, we assume that the extinction process of particles only includes the backscattering and absorption effects, and the forward scattering is ignored. With this assumption, the PDV spectrum is calculated by the formula:
\begin{equation}
\begin{split}
I\left( \omega \right) &= I\left( {\tilde v} \right) \propto \sum\limits_i^{\tilde v < v < \tilde v + dv} {\sqrt {{I_{back,i}}} } \\
&= \sum\limits_i^{\tilde v < v < \tilde v + dv} {\sqrt {\frac{{{I_0}}}{{{A_s}}}{A_{back,i}}\left( v \right) \cdot \exp \left( { - 2\int_{\tilde v}^\infty {\tau \left( v \right)dv} } \right)} } \\
&\propto \sum\limits_i^{\tilde v < v < \tilde v + dv} {\sqrt {{A_{back,i}}\left( v \right)} } \cdot \sqrt {\exp \left( { - 2\int_{\tilde v}^\infty {\tau \left( v \right)dv} } \right)}
\end{split}
\label{eq22}
\end{equation}
where \(\tilde v\) is the velocity corresponding to the frequency shift, \(\tilde v=2\omega c/ \omega_r\), and \(A_{back}\) is the backscattering cross-section area of particles with velocity \(\tilde v\):
\begin{equation}
\begin{split}
&\sum\limits_i^{\tilde v < v < \tilde v + dv} {\sqrt {{A_{back,i}}\left( v \right)} } = \sum\limits_i^{\tilde v < v < \tilde v + dv} {\sqrt {\frac{1}{4}\pi d_i^2{K_{back}}} } \\
&= N\left( {\tilde v} \right)\frac{1}{2}\bar d\sqrt {\pi {K_{back}}} = \frac{{m\left( {\tilde v} \right){A_s}}}{{\frac{1}{6}\pi \overline {{d^3}} {\rho _0}}}\frac{1}{2}\bar d\sqrt {\pi {K_{back}}} \\
&= \frac{{3{A_s}\sqrt {{K_{back}}} }}{{\sqrt \pi {\rho _0}}}\frac{{{m_0}}}{{\overline {{d^3}} /\overline d }}f(\tilde v)
\end{split}
\label{eq23}
\end{equation}
where \(\overline {d}\) is the average particle diameter and \(K_{back}\) is the backscattering coefficient, which is calculated by Mie theory.
The exponent in Eq.~(22) denotes the extinction effects, where the coefficient 2 signifies the back and forth of light in the ejection process. The integral represents the contributions from the extinction of particles above the layer of velocity \(\tilde v\):
\begin{equation}
\begin{split}
\int_{\tilde v}^\infty {\tau \left( v \right)} dv &= \frac{{\sum\limits_i^{v > \tilde v} {\frac{\pi }{4}d_i^2{K^*_{ext}}} }}{{{A_s}}} = \frac{{\int_{\tilde v}^\infty {N\left( v \right)} dv \cdot \frac{\pi }{4}\overline {{d^2}} {K^*_{ext}}}}{{{A_s}}}\\
&= \frac{{3{K^*_{ext}}}}{{2{\rho _0}}}\frac{{\int_{\tilde v}^\infty {m\left( v \right)} dv}}{{\overline {{d^3}} /\overline {{d^2}} }}
\end{split}
\label{eq24}
\end{equation}
where the equivalent extinction coefficient \(K^*_{ext}\) only considers the backscattering and absorption effects:
\begin{equation}
K^*_{ext}=K_{back}+K_{abs}= g_{back}K_{ext}
\label{eq25}
\end{equation}
where the coefficient \(g_{back}\) is 0.5--0.7 for particle diameters of 1--10 \( \rm \mu m\). When \(g_{back}\) = 1, the present model reduces to that of Franzkowiak et al.
Because the velocity profile is exponential, \({\int_{\tilde v}^\infty {m\left( v \right)} dv} = m_0 f(\tilde v) v_{fs}/\beta\). Equation (\ref{eq25}) has the form:
\begin{equation}
\int_{\tilde v}^\infty {\tau \left( v \right)} dv = \frac{{3{K^*_{ext}}}}{{2{\rho _0}}}\frac{{{m_0}}}{{\overline {{d^3}} /\overline {{d^2}} }}\frac{v_{fs} }{{{\beta}}}f(v)
\label{eq26}
\end{equation}
When \(\tilde v\) is equal to the free surface velocity, the integral represents the amount of light that is able to reach the free surface. The optical thickness of the ejecta is defined as:
\begin{equation}
{\tau _0} = \int_{{v_{fs}}}^\infty {\tau \left( v \right)} dv = \frac{{3{K^*_{ext}}}}{{2{\rho _0}}}\frac{{{m_0}}}{{\overline {{d^3}} /\overline {{d^2}} }}
\label{eq27}
\end{equation}
Combining Eqs.~(\ref{eq23}), (\ref{eq26}), and (\ref{eq27}), we can write Eq.~(\ref{eq22}) as:
\begin{equation}
I\left( {\tilde v} \right) \propto \frac{{\sqrt {{K_{back}}} }}{{K_{ext}^*}} \cdot \frac{{\overline d}}{{\overline {{d^2}} }} \cdot {\tau _0}f\left( {\tilde v} \right)\exp \left( { - \frac{{{v_{fs}}}}{\beta }{\tau _0}f\left( {\tilde v} \right)} \right)
\label{eq28}
\end{equation}
This equation provides the theoretical form for determining the PDV spectrum from the ejecta parameters in a vacuum environment. In this formula, the PDV spectrum is proportional to the product of the velocity profile and the extinction term. These two parts are illustrated in Fig.~\ref{fig11}. As the velocity decreases, the corresponding ejecta position moves closer to the free surface, and the incident light becomes weaker because of particle extinction. However, the particles become dense deeper within the ejecta, and this enlarges the cross-section area of scattering. With the contribution of these two parts, the PDV spectrum exhibits a single peak shape.
\begin{figure}
\centering
\includegraphics[height=0.35\textwidth]{fig12.eps}
\caption{Extinction ratio and particle distribution with respect to velocity.}
\label{fig11}
\end{figure}
The simulated PDV spectra of the present model are compared with those of the MC algorithm in Fig.~\ref{fig12}. With the correction of the backscattering coefficient, there is a good agreement between the results, both in the main peak position and the curve shape. However, in the case of a small ejecta mass (\(m_0=5\; \rm mg/cm^2\)), the present model cannot simulate the peak around the free surface. Although the spectrum is nonzero at the position of the free surface in this model, neglecting the multiple-scattering results prevents the second peak from appearing.\cite{Arsenii:1,Franzkowiak:2} In this paper, we mainly consider the information supplied by the original peak of the PDV spectrum. Thus, this defect has only a very slight influence on the accuracy of the present model.
\begin{figure}
\centering
\includegraphics[height=0.35\textwidth]{fig13.eps}
\caption{Simulated PDV spectrum by MC algorithm and single-scattering (SS) model. The area mass unit is \( \rm mg/cm^2\). The particle settings are \(\beta=10\), \(d_m=1.5\; \rm \mu m\), \(\sigma=0.5\). The backscattering coefficient is \(g_{back}=0.67\). }
\label{fig12}
\end{figure}
We now analyze the spectrum function [Eq.~(\ref{eq28})]. First, we take its derivative:
\begin{equation}
\begin{split}
I'\left( {\tilde v} \right) &\propto {\tau _0}f'\left( {\tilde v} \right)\exp \left( { - \frac{{{v_{fs}}}}{\beta }{\tau _0}f\left( {\tilde v} \right)} \right)\left( {1 - \frac{{{v_{fs}}}}{\beta }{\tau _0}f\left( {\tilde v} \right)} \right)\\
&= - \frac{\beta }{{{v_{fs}}}}I\left( {\tilde v} \right)\left( {1 - \frac{{{v_{fs}}}}{\beta }{\tau _0}f\left( {\tilde v} \right)} \right)
\end{split}
\label{eq29}
\end{equation}
where \(f'(\tilde v)=-\beta/v_{fs}f(\tilde v)\).
When \(I'(\tilde v) = 0\), the solution provides the position of the spectrum peak:
\begin{equation}
1 - \frac{{{v_{fs}}}}{\beta }{\tau _0}f\left( {{{\tilde v}_{peak}}} \right) = 0
\label{eq30}
\end{equation}
\begin{equation}
{{\tilde v}_{peak}}/{v_{fs}} = 1 + \ln \left( {{\tau _0}} \right)/\beta
\label{eq31}
\end{equation}
The peak value of the PDV spectrum is:
\begin{equation}
I\left( {{{\tilde v}_{peak}}} \right) \propto \frac{{\sqrt {{K_{back}}} }}{{K_{ext}^*}} \cdot \frac{{\bar d}}{{\overline {{d^2}} }} \cdot \frac{\beta }{{{v_{fs}}}}{e^{ - 1}} \propto \frac{{\beta}}{{\overline {{d^2}} / \overline {{d}} }}
\label{eq32}
\end{equation}
Finally, the relative curvature at the spectrum peak is given by the second derivative of the spectrum function:
\begin{equation}
\begin{split}
I''\left( {\tilde v} \right)/I\left( {{{\tilde v}_{peak}}} \right) = \frac{{{v_{fs}}}}{\beta }e{\left( {{\tau _0}f\left( {\tilde v} \right)\exp \left( { - \frac{{{v_{fs}}}}{\beta }{\tau _0}f\left( {\tilde v} \right)} \right)} \right)^{\prime \prime }}
\end{split}
\label{eq33}
\end{equation}
\begin{equation}
I''\left( {{{\tilde v}_{peak}}} \right)/I\left( {{{\tilde v}_{peak}}} \right) = - {\left( {\frac{\beta }{{{v_{fs}}}}} \right)^2}
\label{eq34}
\end{equation}
In summary, the relationships between the ejecta parameters and the characteristics of the PDV spectrum have the form:
\begin{equation}
\left\{
\begin{aligned}
&{{\tilde v}_{peak}}/{v_{fs}} = 1 + \ln \left( {{\tau _0}} \right)/\beta \\
&I\left( {{{\tilde v}_{peak}}/{v_{fs}}} \right) \propto \frac{\beta }{{\overline {{d^2}} /\overline d }}\\
&I''\left( {{{\tilde v}_{peak}}/{v_{fs}}} \right)/I\left( {{{\tilde v}_{peak}}/{v_{fs}}} \right) = - {\beta ^2}
\end{aligned}
\right.
\label{eq35}
\end{equation}
where the independent variable is normalized by the velocity of the free surface, \(v_{fs}\).
The above relationships allow some of the ejecta parameters to be determined. First, the velocity profile coefficient \(\beta\) is given by the relative curvature at the spectrum peak. In addition, the surface mean diameter \(d^*=\overline d^2/ \overline {d}=d_m e^{0.5\sigma^2}\) can be derived from the value of the spectrum peak. Finally, the optical thickness of the ejecta \(\tau_0\) is obtained from the position of the spectrum peak, where \(\tau_0\) is related to the ejecta area mass \(m_0\) and the Sauter mean diameter \(d_s=\overline {d^3}/ \overline {d^2}=d_m e^{2.5\sigma^2}\). If \(d_s\) is assumed to be approximately \(d^*\), the area mass \(m_0\) can be determined.
Figures 13--15 compare these theoretical relationships with the MC simulation results. Two optical thickness and a large range of velocity profiles are considered. It can be observed that the theoretical relationships largely conform to the MC simulations in these cases, which verifies the present model to some extent.
\begin{figure}
\centering
\includegraphics[height=0.35\textwidth]{fig14.eps}
\caption{Peak positions of PDV spectrum calculated by MC algorithm and theoretical formula. Two optical thickness are considered, where the square points denote \(\tau_0=23.4\) and the circular points denote \(\tau_0=8.6\). The corresponding ejecta parameters are $m_0 =20\;{\rm mg/cm^2}, d_m=1.5\; \mu{\rm m}, \sigma = 0.5$ and $m_0 =10\;{\rm mg/cm^2}, d_m=2.0\; \mu{\rm m}, \sigma = 0.5$, respectively. }
\label{fig13}
\end{figure}
\begin{figure}
\centering
\includegraphics[height=0.35\textwidth]{fig15.eps}
\caption{Peak values of PDV spectrum calculated by MC algorithm and theoretical formula. Two optical thickness are considered, where the square points denote \(\tau_0=23.4\) and the circular points denote \(\tau_0=8.6\). The corresponding ejecta parameters are $m_0 =20\;{\rm mg/cm^2}, d_m=1.5\; \mu{\rm m}, \sigma = 0.5$ and $m_0 =10\;{\rm mg/cm^2}, d_m=2.0\; \mu{\rm m}, \sigma = 0.5$, respectively. }
\label{fig14}
\end{figure}
\begin{figure}
\centering
\includegraphics[height=0.35\textwidth]{fig16.eps}
\caption{Relative curvature of PDV spectrum at the peak calculated by MC algorithm and theoretical formula. Two optical thickness are considered, where the square points denote \(\tau_0=23.4\) and the circular points denote \(\tau_0=8.6\). The corresponding ejecta parameters are $m_0 =20\;{\rm mg/cm^2}, d_m=1.5\; \mu{\rm m}, \sigma = 0.5$ and $m_0 =10\;{\rm mg/cm^2}, d_m=2.0\; \mu{\rm m}, \sigma = 0.5$, respectively.}
\label{fig15}
\end{figure}
\subsection{Experimental verification}
In the above relationships, the peak value of the spectrum is difficult to use in the analysis of PDV experiments. In the experiments, the PDV spectrum is scaled by the reference light intensity, probe reception, photoelectric conversion efficiency, and circuit amplification factor, among other factors. Additional PDV experiments are required to calibrate this scaled factor. For a single PDV vacuum experiment, only the velocity profile \(\beta\) and optical thickness \(\tau_0\) of the ejecta can be extracted. If there is an additional particle granularity measurement, the ejecta area mass \(m_0\) can also be determined.
A set of ejecta PDV experiments performed by Franzkowiak et al.\cite{Franzkowiak:1} was used to verify the present theoretical model. The experiment was carried out in a vacuum environment using Sn material with the surface machined into \(60 \times 8\; \rm \mu m\) grooves. The shock-induced breakout pressure was \(P_{SB}=28\; \rm GPa\). The velocity of the free surface was found to be approximately \(2013\; \rm m/s\). We extracted the PDV spectrum from the experimental spectrogram over the period \(0.2-0.8\; \rm \mu s\), as shown in Fig.~\ref{fig16}(a). The PDV data were then averaged and smoothed using the low-pass filtering of the fast Fourier transform. We converted the spectrum units [dBm] to volts and then took the second derivative to give the smoothed PDV spectrum shown in Fig.~\ref{fig16}(b). The peak of this spectrum is located at \(\tilde v /v_{fs} = 1.32\) and the corresponding relative curvature is approximately $-110$. Combined with Eq.~(\ref{eq35}), this suggests a velocity profile coefficient of \(\beta=10.5\) and an optical thickness of \(\tau_0=28.79\).
Schauer et al.\cite{Schauer:1} conducted a Mie-scattering experiment with similar conditions, where the surface roughness was \(50 \times 8\; \mu m\) and the breakout pressure was about \(30\; \rm GPa\). The particle size distribution was measured to be \(d_m=0.6\;\mu{\rm{m}}, \sigma=0.5\). Using this data, the total area mass was determined to be \(m_0=7.5\;{\rm{mg/cm^2}}\).
In their PDV experiment, Franzkowiak et al. simultaneously measured the area mass with respect to velocity using a piezoelectric probe. The PDV spectrum and area mass given by our estimations and their experiments are compared in Figs.~\ref{fig17} and~\ref{fig18}, respectively. These two results are in good agreement, which verifies the present theoretical model.
\begin{figure}
\centering
\includegraphics[height=0.40\textwidth]{fig17.eps}
\caption{(a) PDV spectrogram extracted from ejecta experiment of Franzkowiak et al.~\cite{Franzkowiak:1} and (b) second derivative of the smooth data.}
\label{fig16}
\end{figure}
\begin{figure}
\centering
\includegraphics[height=0.30\textwidth]{fig19.eps}
\caption{Comparison of the experimental and simulated PDV spectra.}
\label{fig17}
\end{figure}
\begin{figure}
\centering
\includegraphics[height=0.35\textwidth]{fig18.eps}
\caption{Comparison of the area mass and velocity profile between the piezoelectric probe measurement~\cite{Franzkowiak:1} and our estimation.}
\label{fig18}
\end{figure}
\section{Conclusion}
This paper has discussed the PDV spectrum of ejecta particles from shock-loaded samples in a vacuum. A GPU-accelerated MC algorithm that rebuilds the PDV spectrum for the ejecta particles has been proposed, and Mie theory was applied to describe the scattering process. Compared with the reconstruction methods of Andriyash et al. and Franzkowiak et al., a reasonable scattering model is the key to simulating the PDV spectrum accurately. The simulations using the MC algorithm indicate that the particle velocity profile, particle size, and ejecta area mass have a significant influence on the shape and values of the PDV spectrum. As the velocity profile coefficient or particle size increases, the spectrum peak moves to lower velocities. However, this change in the spectrum peak is reversed for the total area mass. In addition, for small values of the optical thickness (few ejecta mass or large particle size), a new spectrum peak appears near the free surface and the original peak gradually decreases or even disappears. For a quantitative analysis, a corrected single-scattering model was proposed for deriving the relationships between the ejecta parameters and the characteristics of the PDV spectrum. It was found that the relative curvature of the spectrum peak is equal to the square of the velocity profile coefficient \(\beta\). The peak value of the spectrum is proportional to the ratio of \(\beta\) to the particle size \(d\), and the peak position of the spectrum is related to \(\beta\) and the total extinction coefficient \(\tau_0\), where \(\tau_0\) is calculated from the total area mass \(m_0\) and the particle size \(d\). Thus, the ejecta parameters \(\beta\), \(d\), and \(m_0\) can be resolved using information about the spectrum peak. However, the spectrum is scaled by multiple experimental parameters, and the relationship with the particle size is difficult to determine. For a single PDV spectrum, only the velocity profile and optical thickness can be determined. Finally, the theoretical interpretation was found to be in good agreement with the MC simulations and PDV experiments of ejecta in a vacuum environment.
The present theoretical model does not consider the multiple scattering near the free surface. When the optical thickness is sufficiently small and the original peak disappears, the present model may be invalid. How to determine the particle size in the PDV experiment is another unsolved issue. In a gas environment, the particles slow down because of aerodynamic deceleration, and this introduces a series of changes to the PDV spectrum over time. The particle deceleration is related to the particle size. In future work, the PDV spectrum in a gas environment will be discussed in an attempt to recover more comprehensive quantities of the ejecta.
\begin{acknowledgments}
This work was supported by a joint fund from the National Natural Science Foundation of China (Grant Nos. 11902043, 11772065) and the Science Challenge Project (Grant No. TZ2016001).
\end{acknowledgments}
\section*{Data Availability}
The data that support the findings of this study are available from the corresponding author
upon reasonable request.
\nocite{*}
|
\section{Introduction}
Deep Neural Networks (DNNs) have achieved human-level performance in many computer vision tasks such as image classification, object detection, and segmentation.
However, the increased compute cost and memory requirement of large DNN models pose a burden on resource-constrained environments such as mobile devices.
To mitigate this problem, various techniques including network quantization~\cite{pact,bnn,abc-net,xnor,dorefa}, network pruning~\cite{deepcompression,structuredpruning}, and efficient architecture design~\cite{mobilenet,efficientnet,shift} were introduced to reduce the compute cost and memory requirement of DNN models.
Among them, the network quantization technique is being actively studied and recent works have shown that a DNN model can even be quantized to a 1-bit model~\cite{bnn,bireal,rtb,xnor}.
When a DNN model is binarized to a Binary Neural Network (BNN) model, the memory requirement of the model is reduced by 32x since 32-bit floating-point weights can be represented by 1-bit weights.
In addition, high precision multiply-and-accumulate operations can be replaced by XNOR-and-popcount logics in BNNs since both activation and weight have 1-bit precision.
Due to the lightweight nature, BNNs are garnering interests as a promising solution for DNN computing on edge devices.
However, BNNs still suffer from the accuracy degradation caused by the aggressive quantization (32-bit to 1-bit).
While recent researches have shown that a DNN model can be quantized to 2-bit precision with marginal accuracy loss~\cite{lsq}, a severe performance gap still exists between a BNN model and its full-precision counterpart DNN model.
In general, it is known that weight quantization is much easier than activation quantization~\cite{hwgq,binaryduo,dorefa}.
In addition, when quantizing the activation, the accuracy loss is marginal until 2-bit quantization, but a significant accuracy drop occurs when quantizing the activation to 1-bit precision~\cite{binaryduo,abc-net,wrpn}.
Previous works tried to explain the sharp accuracy drop from 2-bit activation to 1-bit case based on the gradient mismatch problem caused by the non-differentiable binary activation function~\cite{bnn+,bireal}.
Since the quantization functions are non-differentiable, gradients cannot propagate through the quantization layer in the back-propagation process.
Therefore, previous works used straight-through-estimator (STE) to compute the approximate gradient on non-differentiable layers~\cite{ste,bnn}.
While STE enables back-propagation through quantization layers, the discrepancy between the actual function and the approximated function causes gradient mismatch problem.
Especially in BNNs, sign function is used to binarize activations and is usually approximated as hardtanh function in back-propagation.
Compared to other multi-bit quantization functions, sign function shows more severe gradient mismatch which leads to sharp accuracy degradation.
Hence, several works tried to design better approximation function~\cite{bnn+,bireal} or to reduce the gradient mismatch using neuron coupling~\cite{binaryduo}.
In this work, we argue that there is another reason for the poor performance of BNN in addition to the gradient mismatch.
We speculate that the symmetry of the sign function is also partly responsible for the degradation of BNN performance.
Most DNN models use ReLU activation function instead of sigmoid or Tanh functions.
While the output distributions of sigmoid and Tanh functions are symmetric with respect to zero, ReLU function replaces all the negative values to zero so that the distribution of ReLU output is highly skewed. When quantizing activations to 2-bit or a higher precision, ReLU-based quantization functions are usually used~\cite{pactsawb,pact,lsq,dorefa}.
In other words, multi-bit quantization functions also output unbalanced activation distributions similar to ReLU function.
However, when binarizing activations, sign function is used, and hence the distribution of the binary activation becomes symmetric (or balanced).
We show that a model with unbalanced activation distribution performs better than that with balanced activation distribution.
We first show that the claim is valid even in the full precision activation case by comparing hardtanh and ReLU6 activation functions.
We then show that the performance of BNN can be improved by simply shifting the threshold of the sign function.
We also analyze the effect of training the threshold of the sign function and show that the thresholds cannot be trained efficiently through back-propagation.
Our contributions can be summarized as follows:
\begin{itemize}
\item To the best of our knowledge, we are the first to report that the unbalanced distribution of binary activation actually helps improve the accuracy of BNNs.
\item We propose to shift the thresholds of binary activation functions to make the distribution of binary activation unbalanced.
\item Experimental results show that the accuracy of previous BNN models can be improved at almost no cost by simply shifting the threshold of the binary activation function.
\end{itemize}
\section{Related Work}
\subsection{Network binarization}
There have been several approaches to quantize weights and/or activations into 1-bit precision.
Courbariaux et al.~\cite{binaryconnect} demonstrated binary weight networks which can be successfully trained on small datasets such as CIFAR-10 and SVHN.
Hubara et al.~\cite{bnn} further proposed BNN in which both weights and activations are binarized.
To apply back-propagation through a sign function which is non-differentiable, the straight-through-estimator (STE) concept was used~\cite{ste}.
Rastegari et al.~\cite{xnor} proposed XNOR-Net which uses real-valued scaling factors when binarizing weights and activations and demonstrated acceptable accuracy on large scale ImageNet dataset.
When binarizing ResNet models, real-valued shortcut connections play a critical role in propagating high resolution information.
Liu et al.~\cite{bireal} proposed to use additional shortcut connections so that a shortcut connection exists for every binary convolution layer.
\subsection{Training quantization parameters}
Recent studies suggested to train the quantization intervals and ranges using back-propagation to improve the accuracy of quantized neural networks.
Choi et al.~\cite{pactsawb,pact} proposed the parameterized clipping activation function (PACT) in which the clipping range is trained using back-propagation.
While only the clipping range was trained in PACT, several following works proposed to train both quantization interval and range~\cite{lsq,qil,lqnet}.
Note that these works focused on training multi-bit networks and hence did not report results on BNNs.
For BNNs, several recent works proposed to train the threshold of binary activation function.
Liu et al.~\cite{reactnet} proposed to use trainable activation functions so that the distribution of the activation can be balanced.
Wang et al.~\cite{sibnn} proposed the trainable binarization which learns the threshold as well as the gradient clipping range used in back-propagation.
We, however, observed results contrary to the findings from previous works on trainable threshold.
Our experimental results show that the bias term in Batch Normalization (BN) layer is already learning the threshold and therefore the effect of training the threshold of binary activation is limited. Details will be described in Sec.~\ref{sec:learnable}.
\subsection{Managing activation distribution}
Since there are only two values available for activation in BNNs, the distribution of binary activation plays a critical role in BNNs.
There have been few works that tried to manipulate the distribution of binary activation to improve the accuracy of BNNs.
Ding et al.~\cite{distloss} proposed to regularize the distribution of pre-activation values to tackle the difficulties that occurred during training BNNs.
The work mostly focused on avoiding extreme cases such as the case when all the pre-activation values have the same sign.
Liu et al.~\cite{reactnet} proposed to reshape the distribution of binary activation using trainable thresholds.
Using the trainable activation functions, they made the distribution of binary activation more balanced.
\subsection{Additional activation function}
Another simple yet effective technique to improve the accuracy of BNNs is to use an additional activation function (e.g. PReLU) between the binary convolution layer and the following BN layer.
Rastegari et al.~\cite{xnor} mentioned that inserting a ReLU function after the binary convolution helps training BNNs.
Tang et al.~\cite{aaai} proposed to use PReLU instead of ReLU to absorb the weight scale factors.
Bulat et al.~\cite{bulat2017,bulat2019} explained that increasing the nonlinearity in BNNs helps training, and compared the effects of ReLU and PReLU.
Based on these findings, many recent works utilized additional activation layers in their models~\cite{reactnet,rtb,leastsquares}.
\section{Method}
In this section, we introduce how to improve the accuracy of BNNs using unbalanced activation distribution.
We first discuss the activation functions in conventional full precision models to describe the motivation for using unbalanced activation distribution (Sec.~\ref{sec:motivation}).
Then we show how to make the distribution of the binary activation unbalanced using threshold shifting in BNNs and how much the proposed technique can improve the accuracy on various benchmarks (Sec.~\ref{sec:biasing}).
Experimental results on ImageNet dataset will be given to show that the accuracy of various existing BNN models can be improved by simply shifting the threshold of binary activation functions (Sec.~\ref{sec:imagenet}).
After that, we discuss the ineffectiveness of the methodology to train the threshold of binary activation in BNNs (Sec.~\ref{sec:learnable}).
We also show that the additional activation functions (i.e. ReLU or PReLU) used in recent BNN models make the activation distribution unbalanced thereby helping to improve the accuracy (Sec.~\ref{sec:relu}).
\subsection{Accuracy gap between hardtanh and ReLU6}
\label{sec:motivation}
Before ReLU activation function was proposed, sigmoid or Tanh functions had been used as an activation function.
It is widely known that ReLU works better because it solves the gradient vanishing problem that occurs in sigmoid or Tanh functions~\cite{relu}.
However, the gradient vanishing problem is largely diminished by using BN layers in recent models~\cite{bn}, so there might be other reasons for the success of the ReLU function.
We suspect that another main reason for the good performance of ReLU is that the output distribution of the ReLU function is highly skewed.
While the output distributions of sigmoid and Tanh functions are symmetric with respect to each mean value, ReLU makes all the negative inputs to zero and passes the positive inputs.
Therefore, the output distribution of the ReLU function is positively skewed (mean value is larger than the median value).
We conducted a few experiments to monitor whether the unbalanced activation distribution due to the nature of ReLU function helps to improve accuracy.
To make the problem simple, we compare the performance of hard hyper-tangent (hardtanh) and ReLU6 functions instead of ReLU.
The ReLU6 is a slight variant of the ReLU function where its positive outputs are clipped to 6.
The hardtanh function is shown in Fig.~\ref{fig:hardtanh}a.
It has been reported that the performance of ReLU6 is as good as that of ReLU or sometimes even better~\cite{relu6}.
Also, ReLU6 and hardtanh have very similar shape and ReLU6 can be thought as a shifted-and-scaled form of hardtanh.
\begin{figure}
\centering
\includegraphics[width=\linewidth]{fig1.pdf}
\caption{(a) Original hardtanh function. Hardtanh function shifted along (b) the x-axis by x\_offset and (c) the y-axis by y\_offset. (d) Hardtanh function with the increased range.}
\label{fig:hardtanh}
\end{figure}
As shown in Fig.~\ref{fig:hardtanh}, we modified the hardtanh function in three different ways; (b) shifting along the x-axis, (c) shifting along the y-axis and (d) increasing the range.
Note that the hardtanh function is identical to the ReLU6 function when it is shifted along the x- and y-axis by 3 and its range is increased by 3 times.
We trained vgg-small model~\cite{vgg} on CIFAR-10 dataset with different activation functions.
The vgg-small model has 4 convolution layers with 64, 64, 128, 128 output channels in sequence followed by 3 fully-connected layers with 512 neurons.
We trained the model in the same condition with 10 different seeds and report the mean and the standard deviation of the test accuracy.
Detailed setup for the training is described in the supplementary material.
When ReLU6 function is used as the activation function, 89.21\% of test accuracy was achieved while 88.55\% was achieved when hardtanh function is used without any modification.
When the hardtanh activation function is modified as described in Fig.~\ref{fig:hardtanh}, the test accuracy of the model also changes.
\begin{figure}
\centering
\includegraphics[width=\linewidth]{fig2.pdf}
\caption{Test accuracy of vgg-small model with modified hardtanh activation functions. Red dotted line represents the test accuracy of the model with ReLU6 activation function. The mean (line) and standard deviation (shade) of 10 runs are plotted.
}
\label{fig:vggsmall_hardtanh}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{fig3.pdf}
\caption{Distributions of (a) pre-activation, (b) output of hardtanh activation function, (c) output of hardtanh activation function shifted to the right along the x-axis by 0.5, and (d) output of hardtanh activation function shifted up along the y-axis by 0.5. }
\label{fig:skew_dist}
\vspace{-2mm}
\end{figure}
Fig.~\ref{fig:vggsmall_hardtanh} shows the change in the test accuracy depending on the amount of shift along (a) the x-axis or (b) y-axis or (c) increase in the range of the hardtanh activation function.
We observe that shifting the hardtanh activation function along the x-axis increases the accuracy substantially (Fig.~\ref{fig:vggsmall_hardtanh}a).
The highest test accuracy was obtained when shifting the hardtanh function in a positive direction along the x-axis by 1.2.
On the other hand, shifting the hardtanh activation along the y-axis or increasing the range does not improve the accuracy as much as shifting the hardtanh along the x-axis (Fig.~\ref{fig:vggsmall_hardtanh}(b and c)).
When the hardtanh function is shifted to the right along the x-axis, the number of negative outputs increases and the output distribution becomes positively skewed (Fig.~\ref{fig:skew_dist}c).
Note that shifting the hardtanh function along the y-axis also makes the output activation distribution to have a non-zero mean.
However, it only shifts the activation distribution and the distribution is not skewed in such a case (Fig.~\ref{fig:skew_dist}d).
Therefore, we think that the higher performance of ReLU activation function compared to the hardtanh partly comes from the unbalanced distribution of activation outputs.
Interestingly, the effect of the activation distribution becomes even more noticeable in BNNs when the hardtanh function is replaced by the sign function.
\subsection{BNNs with unbalanced activation distribution}
\label{sec:biasing}
As described in the previous section, breaking the balance of the activation distribution by shifting the activation function along the x-axis helps improve the accuracy of a model.
When quantizing activations, previous multi-bit quantization methods~\cite{hwgq,pactsawb,pact,lsq,syq,qil,wrpn,dorefa} used ReLU-based quantization which outputs unbalanced activation distributions.
However, in BNNs, previous works used sign function for binary activation function thus the distribution of binary activation is balanced~\cite{xnor++,bnn+,distloss,bnn,bireal,xnor}.
Note that in BNNs, there are only two output values (+1 or -1) and the balanced distribution means that the ratio of +1 to -1 is close to 1:1.
We believe that the poor performance of BNNs observed in previous literature~\cite{abc-net,wrpn} is partly due to the use of the sign function.
Since shifting the hardtanh function along the x-axis improved the test accuracy in the full precision model, we also tried shifting the sign function along the x-axis.
Note that shifting the sign function is the same as just changing the threshold value of the activation function.
\begin{figure*}
\centering
\includegraphics[width=\linewidth]{fig4.pdf}
\caption{(a) Test accuracy vs. threshold shift for vgg-small model. (b) Distribution of the binary activation of the first, third, and fifth activation layer when threshold shift = 0. (c) Distribution of the same data with threshold shift = 1.2. (d) Test accuracy vs. threshold shift when Max Pooling layers are replaced by Average Pooling layers.}
\label{fig:vggsmall_sign}
\end{figure*}
As expected, shifting the activation function along the x-axis increases the accuracy of BNNs also (Fig.~\ref{fig:vggsmall_sign}a).
To check the effect of the shape of the distribution of activation outputs, we observed the distributions of the binary activation in different layers.
Fig.~\ref{fig:vggsmall_sign}b shows the distribution of the binary activations of the first, third, and fifth activation layer when the original sign function is used.
For the first and the fifth activation layer, the distribution of pre-activation is Gaussian-like with zero mean and hence the ratio of +1 to -1 in binary activation is close to 1:1.
In the case of the third activation layer, the distribution of pre-activation values does not have zero mean due to the preceding Max Pooling layer even when the original sign function is used.
Note that the convolution results go through the Max Pooling layer first and then a BN layer before the binary activation layer.
The Max Pooling layer makes the distribution positively skewed, thus the mean of the distribution becomes larger than the median of the distribution.
Since the BN layer centers the distribution based on the mean value, there exist more negative values than positive values in pre-activation values.
As a result, the number of -1 is larger than that of +1 in the binary activation after the third activation layer (Fig.~\ref{fig:vggsmall_sign}b).
However, the output distributions of all other layers which do not come after Max Pooling layer are balanced.
When the threshold of binary activation function is shifted by 1.2, which results in the best accuracy, the distribution of binary activation changes.
Fig.~\ref{fig:vggsmall_sign}c shows the distribution of the binary activations when the threshold is shifted by 1.2.
As expected, the output distributions of all the binary activation layers become unbalanced with the shifted threshold.
Note that even though the distribution of binary activation in a layer is unbalanced, the imbalance does not propagate through layers because the binary weights in the following binary convolution layer are zero-centered.
\subsubsection{Effect of Max Pooling}
Another interesting observation from the results of shifting the activation function along the x-axis was that the accuracy improved only when the activation function was shifted in a positive direction.
Since the Max Pooling layer is the only layer that gives asymmetry in the model, we suspect that Max Pooling causes the accuracy difference between shifting of the threshold to the positive direction and the negative direction.
To gauge the effect of the Max Pooling, we replaced the Max Pooling layers with Average Pooling layers which do not have asymmetric effects.
Fig.~\ref{fig:vggsmall_sign}d shows the training results.
Since the model with Average Pooling is symmetric, the effect of shifting the activation function is also symmetric and we can observe that the accuracy is improved when the threshold is shifted away from zero in either direction.
\subsubsection{Effect of other conditions}
In this section, we investigate the effect of threshold shifting with various (1) datasets, (2) models, (3) initialization methods, and (4) optimizers to show that the proposed technique is not limited to the specific condition or benchmark.
\begin{figure}
\centering
\includegraphics[width=\linewidth]{fig5.pdf}
\caption{Test accuracy vs. threshold shift for (a) binary vgg-small model on MNIST dataset, (b) binary 2-layer MLP model on MNIST dataset, (c) binary LeNet-5 on MNIST dataset, (d) full-precision vgg-small model on CIFAR-10 dataset, (e) binary vgg-small model initialized with full-precision pretrained model, and (f) binary vgg-small model optimized using SGD.}
\label{fig:others}
\end{figure}
\noindent\textbf{Dataset.} We first changed the dataset to MNIST.
To solely investigate the dependency on dataset, we used the same vgg-small model which we used for CIFAR-10 in Sec.~\ref{sec:biasing}.
For MNIST dataset, we trained each model with 30 different random seeds and average results are presented.
As shown in Fig.~\ref{fig:others}a, the accuracy of the vgg-small model can also be improved by shifting the threshold when trained on MNIST dataset.
We also evaluate the ImageNet dataset and the results will be described in the Sec.~\ref{sec:imagenet_result}.
\noindent\textbf{Model architecture.}
We also evaluated two additional model architectures (2-layer MLP and LeNet-5) to verify the effect of threshold shifting.
Fig.~\ref{fig:others}b and c show the result of MLP and LeNet-5, respectively.
Although the accuracy improvement with threshold shifting is relatively small ($\sim$0.2\%), a clear trend is observed in the test accuracy as the activation functions are shifted along the x-axis in both models.
Note that the trend in the MLP model is symmetric because the model do not have the Max Pooling layer in it.
We also conducted experiments on much larger and complex models (AlexNet and ResNet), and the results will be described in the Sec.~\ref{sec:imagenet_result}.
\noindent\textbf{Initialization method.}
In all the previous experiments, we trained the models from scratch using Xavier normal initialization~\cite{xavier}.
In recent literature~\cite{bireal,rtb}, pretrained full-precision models are often used to initialize BNN models.
We also demonstrate that the proposed technique is effective with such an initialization method.
As shown in Sec.~\ref{sec:motivation}, the accuracy of a model with hardtanh function can also be improved by shifting the activation function to the right along the x-axis.
Since pretrained full-precision models typically use hardtanh function instead of ReLU function to minimize the mismatch with the sign activation function in the initialization stage, we shifted the hardtanh function in the full-precision model and used the results as the initialization points.
Fig.~\ref{fig:others}d shows the training results of vgg-small model with full-precision weight and activation.
We initialized the BNN version of vgg-small model with the pretrained model in which the hardtanh function is shifted by 1.
While using the same shift amount for binary activation and hardtanh function seems natural, there might be a better shift amount for the binary activation other than the shift amount for the pretrained model.
Therefore, we searched for the optimal shift amounts for the pretrained model and BNN model separately.
Fig.~\ref{fig:others}e shows the trend of test accuracy versus the threshold shifting amount.
While the pretrained model used 1.0 as the shift amount, the best shift amount for the BNN model was 0.6 so using different amounts of shift for the pretrained model and the BNN produces higher accuracy.
\noindent\textbf{Optimizer.}
Although most previous BNN training methods used Adam optimizer~\cite{adam}, we also tested SGD with momentum to see the sensitivity to the optimizer.
Fig.~\ref{fig:others}f shows that a similar trend is maintained with SGD while the accuracy is slightly degraded compared to the Adam case.
\subsection{Experiments on ImageNet}
\label{sec:imagenet}
In this section, we apply the threshold shifting technique to various previous BNN models on ImageNet dataset and show that the technique can be added to existing BNN training techniques without requiring any modification other than the simple threshold shift.
We also introduce a simple method to find the appropriate shifting amount for each model to avoid time-consuming search for optimal threshold shift for large neural networks.
\subsubsection{Finding shift amount}
\begin{figure}
\centering
\includegraphics[width=0.85\linewidth]{fig6.pdf}
\caption{Top-1 train accuracy at the first epoch and Top-1 validation accuracy at the end of training with various threshold shift amount. The two graphs show similar trends.}
\label{fig:train_vs_valid}
\vspace{-2mm}
\end{figure}
As shown in previous results, the best shift amount differs across models and datasets.
To find the optimal shift amount, we propose to use the training accuracy in earlier training steps.
For example, we choose the best shift amount based on the train accuracy at the end of the first epoch for ImageNet dataset.
This approach is valid because the trend in the train accuracy at the first epoch and that in the validation accuracy after training is finished are similar enough.
Fig.~\ref{fig:train_vs_valid} shows the two trends when training XNOR-Net (ResNet-18)~\cite{xnor} on ImageNet dataset.
The shapes of the two trends resemble each other, and both trends show the peak performance when their shift amounts are around 2.0.
Using the proposed method, we can find the optimal value for the threshold shift easily without training the model to the end for every shift amount.
\subsubsection{Results on ImageNet}
\label{sec:imagenet_result}
Using the proposed method to find optimal threshold shift, we trained several previous BNN models for ImageNet dataset with the threshold shifting technique.
\begin{table*}[t]
\centering
\caption{Accuracy of previous BNN models trained on ImageNet dataset with and without the threshold shifting technique. $^\dagger$For Bi-Real-Net, the shift amounts for the pretrained model and the BNN model are shown together.}
\begin{tabular}{c|cc|c|cc|cc}
\toprule
\multirow{2}{*}{Model} & \multicolumn{2}{c|}{Baseline accuracy [\%]} & \multirow{2}{*}{Shift amount} & \multicolumn{2}{c|}{Accuracy w/ shift [\%]} & \multicolumn{2}{c}{Accuracy gain [\%]} \\
&\;\;\;\;Top-1 & Top-5\;\; & & \;\;\;\;Top-1 & Top-5\;\; & \;\;\;Top-1 & Top-5\;\;\\
\hline
BNN (AlexNet) &\;\;\;\;41.5 & 66.1\;\; & 0.6 & \;\;\;\;42.1 & 66.6\;\; & \;\;\;0.6 & 0.5\;\;\\
XNOR-Net (AlexNet) &\;\;\;\;44.4 & 68.6\;\; & 0.8 & \;\;\;\;45.6 & 69.6\;\; & \;\;\;1.2 & 1.0\;\;\\
XNOR-Net (ResNet-18) &\;\;\;\;51.2 & 74.9\;\; & 2.0 & \;\;\;\;54.2 & 77.6\;\; & \;\;\;3.0 & 2.7\;\;\\
Bi-Real-Net (ResNet-18) &\;\;\;\;56.1 & 79.1\;\; & (2.0, 1.6)$^\dagger$ & \;\;\;\;57.2 & 80.2\;\; & \;\;\;1.1 & 1.1\;\;\\
Bi-Real-Net (ResNet-34) &\;\;\;\;61.9 & 83.9\;\; & (2.0, 1.8)$^\dagger$ & \;\;\;\;62.8 & 84.5\;\; & \;\;\;0.9 & 0.6\;\;\\
\bottomrule
\end{tabular}
\label{tab:imagenet}
\end{table*}
Table~\ref{tab:imagenet} shows how much accuracy improvement can be achieved by the proposed threshold shifting technique on different BNN models.
We first reproduced the baseline accuracy of previous works with less than 0.3\% accuracy difference.
With the same hyper-parameter setting, we only changed the threshold of binary activation functions in each model.
The shift amount for each model is shown in the Table~\ref{tab:imagenet}.
The results indicate that the proposed threshold shifting technique also works well for large models on ImageNet dataset.
In addition, the proposed threshold shifting technique can be easily combined with other BNN training methods (i.e. weight scaling factor and double skip connection).
\subsection{Effect of training threshold}
\label{sec:learnable}
In several recent works, it has been proposed to train the interval and range of quantization functions via back-propagation~\cite{pact,lsq,qil,lqnet}.
Training the threshold of binary activation function in BNNs has been also proposed recently~\cite{reactnet,sibnn}.
The threshold training approaches are based on the belief that the best threshold value for each binary activation function can be found using back-propagation.
However, here we show that training the threshold has a limited effect on the performance of BNNs.
\subsubsection{Batch normalization bias}
The most important reason why the threshold training is not effective is that the role of trainable threshold is already covered by the bias values in the BN layer which comes right before the binary activation layer.
The compute process of a BN layer and the following binary activation layer can be represented as
\begin{equation}
\label{eq:bn-sign}
\begin{split}
Y=
\begin{cases}
-1 & \quad \text{if}\;\; \gamma\frac{X-\mu}{\sigma}+\beta \leq th\\
+1 & \quad \text{if}\;\; \gamma\frac{X-\mu}{\sigma}+\beta > th
\end{cases}.
\end{split}
\end{equation}
Here, $X$ and $Y$ are inputs to the BN layer and output from the binary activation layer, respectively.
$\mu, \sigma, \gamma,$ and $\beta$ are mean, standard deviation, weight, and bias of the BN layer, and $th$ is the trainable threshold.
As shown in the Eq.~\ref{eq:bn-sign}, the bias term in BN serves exactly the same role as the trainable threshold.
Both of them are initialized to zero in the beginning and their gradients are values of the same magnitude and the opposite sign.
\begin{figure}
\centering
\includegraphics[width=0.85\linewidth]{fig7.pdf}
\caption{Comparison of the BN bias values in the second BN layer and the threshold values in the following activation layer of the vgg-small model. The thresholds are trained following the approach suggested in \cite{reactnet}.}
\label{fig:bias_vs_threshold}
\vspace{-2mm}
\end{figure}
We trained the vgg-small model with trainable thresholds following \cite{reactnet} and observed the bias values of the second BN layer and the threshold values of the following binary activation layer.
Fig.~\ref{fig:bias_vs_threshold} shows the values of BN biases and the trained thresholds in the vgg-small model.
For each channel, the BN bias and the threshold have the same values with an opposite sign.
Therefore, training the thresholds of binary activation functions means nothing more than doubling the learning rate of BN biases.
Please refer to section S2 in the supplementary material for more discussion.
\subsubsection{Limited learning capability}
\begin{figure}
\centering
\includegraphics[width=0.95\linewidth]{fig8.pdf}
\caption{(a) Test accuracy of vgg-small model with trainable thresholds with various initial threshold values. (b) Distribution of effective thresholds ($th-\beta$) of the second binary activation layer after training is finished. Each distribution from top to bottom represents the case when the thresholds are initialized to -2, -1, 0, 1, and 2, respectively.}
\label{fig:learnable}
\vspace{-3mm}
\end{figure}
While the shifted threshold values in the proposed method can also be absorbed by the bias values in BN after a training is finished, it differs from the trainable threshold method in that the fixed shift amount of the binary activation function serves as an initialization point as well.
If the BN bias and the threshold can be trained to an optimal point via back-propagation from an arbitrary initial point, the threshold shifting technique might not be as effective.
We, however, found that the final BN bias and the trainable threshold values heavily depend on the initial values, and hence initializing them with the proposed threshold shifting technique strongly affects the training performance.
Fig.~\ref{fig:learnable}a shows the training result of vgg-small model with trainable threshold.
To see the effect of initialization values, we varied the amount of the threshold shift and applied the trainable threshold methods.
Similar to the result with a fixed threshold, the test accuracy strongly depends on the initialization point (or threshold shifting amount).
We further analyzed the distribution of effective threshold ($th-\beta$) after training is finished for five different initialization values.
As shown in Fig.~\ref{fig:learnable}b, the final distribution of the effective threshold strongly depends on the initial values and do not change much from the initial points.
For example, the first case (top) in Fig.~\ref{fig:learnable}b is when the threshold of the binary activation function is set to -2 and the BN bias is initialized to 0.
Even though the thresholds of the binary activation functions, which are initially -2, are trainable, back-propagation does not train them to the optimal value which is close to 1.
\subsection{Effect of additional activation function}
\label{sec:relu}
Recently, several works proposed to use additional activation function (i.e. PReLU) after convolution layer in BNNs~\cite{bulat2017,bulat2019,reactnet,rtb,leastsquares,aaai}.
While the performance improvement by the additional activation function was significant, the reason for the accuracy improvement is not clearly understood yet.
Previous works have described that BNNs usually lack nonlinearity in the model due to the simple activation functions and the additional activation functions give more nonlinearity to the model~\cite{bulat2017,xnor}.
In addition to the increased nonlinearity, we observed that the accuracy improvement by additional activation function is also related to the unbalanced activation distribution.
\begin{figure}
\centering
\includegraphics[width=0.95\linewidth]{fig9.pdf}
\caption{(a) Model description of ResNet-20 with double skip connection. ResNet-20 with PReLU model has an additional PReLU layer at the end of every residual block. Test accuracy of ResNet-20 (b) without and (c) with an additional PReLU layer after every binary convolution layer.
}
\label{fig:prelu}
\vspace{-5mm}
\end{figure}
We first evaluate the effect of additional PReLU layers after binary convolution layers in the ResNet-20 model.
For experiments, we used ResNet-20 model with double skip connections~\cite{bireal} as shown in Fig.~\ref{fig:prelu}a.
The model consists of the first convolution layer, 18 residual blocks, and the last fully-connected layer.
In the ResNet-20 model with additional PReLU layers, a PReLU layer is inserted at the end of every residual block as shown in Fig.~\ref{fig:prelu}a.
The effect of threshold shifting on ResNet-20 model with and without the additional PReLU layers is shown in Fig.~\ref{fig:prelu}b and c, respectively.
While the accuracy is improved by shifting the threshold of binary activation function in the ResNet-20 model, the threshold shifting degrades the accuracy when additional PReLU layers are used jointly.
The additional PReLU layers already play a role in distorting the activation distribution, and hence further shifting the threshold of binary activation function is excessive.
To validate the claim, we slightly modified the additional activation function and analyzed the effect of threshold shifting technique.
We replaced the additional PReLU layers with LeakyReLU layers and varied the slope of the negative range of the LeakyReLU layers from 0 to 1.
Note that when the slope is 1, the LeakyReLU layer becomes an identity function and therefore the model becomes the same as the ResNet-20 model without PReLU layers.
\begin{figure}
\centering
\includegraphics[width=0.95\linewidth]{fig10.pdf}
\caption{Effect of the slope of the negative range of LeakyReLU layers on the dependence of the test accuracy on the threshold shift amount. The slope of the negative range of LeakyReLU is (a) 0.0, (b) 0.1, (c) 0.25, (d) 0.5, (e) 0.75, and (f) 0.9.
Inset: distribution of pre-activation when shift amount is 0 in each case.
}
\label{fig:leakyrelu}
\vspace{-3mm}
\end{figure}
Fig.~\ref{fig:leakyrelu} shows the training results of ResNet-20 with LeakyReLU layers with different slopes.
As the slope of the negative range of LeakyReLU decreases, the distortion of the activation distribution becomes more severe.
When the slope is close to 1, the LeakyReLU layer does not skew the distribution of pre-activation much.
Hence, the distribution of pre-activation values is very close to a Gaussian distribution with zero mean.
As a result, the accuracy improvement by shifting the threshold of binary activation function was clearly observed as in the ResNet-20 model without PReLU layers.
However, as the slope decreases, the LeakyReLU layer makes the activation distribution unbalanced as shown in Fig.~\ref{fig:leakyrelu}.
Note that LeakyReLU layer scales down the negative values only so that the mean value of the distribution increases while the median value of the distribution does not change.
Therefore, it has a similar effect to shifting the threshold of binary activation function in that the activation distribution becomes unbalanced.
\section{Conclusion}
In this paper, we analyzed the impact of activation distribution on the accuracy of BNN models.
While previous BNNs used sign function as binary activation function which balances the distribution of binary activation, we claim that the accuracy of BNN models can be improved when the distribution of binary activation is unbalanced.
By simply shifting the thresholds of binary activation functions, we demonstrated that the accuracy of previous BNN models could be further improved.
We also identified that the unbalanced activation distribution partly accounts for the improved accuracy of BNN models which used additional activation functions.
\small{
\textbf{Acknowledgement:} This work was in part supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (NRF-2020R1A2C2004329), Samsung Research Funding Center under Project Number SRFC-TC1603-51, and Institute of Information \& communications Technology Planning \& Evaluation (IITP) grant funded by the Korea government (MSIT) (No.2019-0-01906, Artificial Intelligence Graduate School Program(POSTECH)).
}
{\small
\bibliographystyle{ieee_fullname}
|
\section{Introduction}
During past decades, significant attention has focused on embedded or networked control systems that have limited and time-varying controller's computation capability due to high requirements on multitasking operations.
In particular, assuming constant and limited computational resources, bounds on computational time of specific optimization algorithms for achieving stability were derived in~\cite{McGovern,McGovern1}.
For time-varying computational resources, a dynamic computation task scheduling method was proposed for model predictive controllers~\cite{Henriksson}.
On-demand computation scheduling of control input based on plant states were investigated for periodic, event-triggered and self-triggered policies in~\cite{ondemand1,ondemand2,ondemand3}, respectively.
Another stream of research considers anytime algorithms for robust control and making efficient use of time-varying computational resources.
In general, an anytime algorithm can provide \emph{a} solution even with limited computational resources, and refines the solution when more resources are available.
In the pioneering work~\cite{bhattacharya2004anytime}, an anytime control system was proposed, where the number of updated states varies with the available computation time known to the controller \emph{a priori}.
In~\cite{GuptaAnytime}, an anytime control algorithm for a multi-input linear system was proposed for the scenario when the computation availability is \emph{unknown a priori}. The main idea was to first calculate the most important component of the control vector and then calculate the less important ones as more computational resource becomes available.
In~\cite{anytime5}, a sequence-based anytime control method was proposed, which can calculate a tentative sequence of future control input for as many time steps as allowed by the available computational resources at each time step. The pre-calculated control sequence can compensate for the time steps when no computational resource is available for control calculations.
Following this work, sequence-based anytime control systems with Markovian processor availability, event-triggered sensor updates and multiple control laws were investigated in~\cite{anytime2,anytime3,anytime4}, respectively.
In \cite{bhattacharya2004anytime,GuptaAnytime,anytime5,anytime2}, the sensor-controller and controller-actuator channels were assumed to be perfect and error-free.
In~\cite{anytime3} and~\cite{anytime4}, the sensor-controller channel was assumed to have independent and identically distributed (i.i.d.) packet dropouts\footnote{Note that i.i.d. packet dropout channel is very commonly considered in the literature of networked control~\cite{ZHAN2015214}.}, i.e., only binary-level (on-off) channel states, were considered, while the controller-sensor channel was assumed to be perfect.
{\color{black}\emph{Motivation.}
The existing work of anytime control~\cite{bhattacharya2004anytime,GuptaAnytime,anytime5,anytime2,anytime3,anytime4} cannot effectively handle the fully distributed networked control scenario, where both the sensor-controller and controller-actuator communication links are wireless.
Moreover, the existing research only considered simple wireless channel models, which cannot capture the key features of practical wireless channels that are time-varying and correlated~\cite{Parastoo}.
Therefore, anytime control design and analysis for fully distributed networked control system in practical wireless channels are critical in practice, but also present new challenges.}
{\color{black}\emph{Novelty and contributions.}
In this work, we consider for the first time the sequence-based anytime control of a wireless networked control system (WNCS) in a generalized \textbf{dual imperfect channel}, where the sensor-controller and controller-actuator channel states are \textbf{spatial- and time-correlated} and are modeled as a \emph{multi-state Markov process}.
Different from most of the existing works~\cite{anytime5,anytime2,anytime3,anytime4}, where only the controller has a buffer to keep the calculated sequence of control inputs, we propose to use anytime control with \textbf{buffers at both the controller and the actuator nodes}. The latter is used to compensate for dropouts in the controller-actuator channel.
Moreover, the available computational resource of the controller is allowed to be time-correlated, modeled as a multi-state Markov process.
Such a \textbf{dual-channel-dual-buffer} anytime control system has practical advantages but brings significant challenges to its analysis due to the complex system state updating rule, compared to previous setups.
We propose a novel cycle-cost-based approach to
derive sufficient conditions for stochastic stability of the overall WNCS. Our stability conditions are stated in terms of plant dynamics, network dynamics, buffer properties and computational resource dynamics. We further show that, under suitable assumptions, the conditions derived guarantee robust stability when plant disturbances are taken into account.}
The remainder of the paper is organized as follows:
Section~II presents the system model of anytime control in the dual-channel-dual-buffer WNCS.
Section~III develops the stability condition.
Section~IV provides robust stability analysis.
Section~V draws conclusions.
\emph{Notation:} Sets are denoted by calligraphic capital letters, e.g., $\mathcal{A}$.
$\mathcal{A} \backslash \mathcal{B}$ denotes set subtraction.
Matrices and vectors are denoted by capital and lowercase upright bold letters, e.g., $\mathbf{A}$ and $\mathbf{a}$, respectively.
$\mathsf{E}\left[A\right]$ is the expectation of the random variable $A$.
The conditional probability $\myprob{{A}\vert {B}}=0$ if $\myprob{{B}}=0$.
$(\cdot)^{\top}$ is the matrix transpose operator.
{\color{black}$| \mathbf{v} |$ is the Euclidean norm of vector $\mathbf{v}$. }
$\mathbb{N}$ and $\mathbb{N}_0$ denote the sets of positive and non-negative integers, respectively.
$\mathbb{R}^m$ denotes the $m$-dimensional Euclidean space.
$[\mathbf{A}]_{j,k}$ and $[\mathbf{v}]_{j}$ denote the element at the $j$th row and $k$th column of a matrix $\mathbf{A}$, and the $j$th element of a vector $\mathbf{v}$, respectively.
$\lambda_{\max}(\mathbf{A})$ denotes the spectral radius of $\mathbf{A}$.
$\text{diag}\{\mathbf{v}\}$ denotes the diagonal matrix generated by the vector $\mathbf{v}$.
$\{v\}_{\mathbb{N}_0}$ denotes the semi-infinite sequence $\{v_0,v_1,\cdots\}$.
A function $\phi: \mathbb{R}_{\geq 0} \rightarrow \mathbb{R}_{\geq 0}$ is of class-$\mathcal{K}_{\infty}$ ($\phi \in \mathcal{K}_{\infty}$) if it is continuous, strictly increasing, and zero at zero.
$\mathbf{0}_{i}$ and $\mathbf{0}_{i \times j}$ denotes the all-zero $i \times i$ and $i \times j$ matrices, respectively. $\mathbf{A}=\mathbf{0}$ indicates that $\mathbf{A}$ has all zero elements.
\section{Anytime Control in a Dual-Channel-Dual-Buffer WNCS} \label{sec:sys}
We consider a WNCS consisting of a plant system, a remote controller and a wireless network placed between the plant and the controller.
A sensor sends its measurements of the plant to the controller, and the controller computes and sends control commands to a remote actuator via the wireless network as illustrated in Fig.~\ref{fig:system_model}. The dual wireless channel (i.e., the sensor-controller and controller-actuator channels) setup is different from \cite{anytime2,anytime3,anytime5}, which assumed either perfect transmissions in two channels or a single (imperfect) wireless channel from the sensor to the controller.
\begin{figure}[t]
\centering\includegraphics[scale=0.8]{system_model.pdf}
\vspace{-0.3cm}
\caption{The dual-channel-dual-buffer WNCS.}
\label{fig:system_model}
\vspace{-0.5cm}
\end{figure}
Each sampling period of the plant, denoted by $T_s$, is divided into four phases: sensor-controller (S-C) transmission, command computation, controller-actuator (C-A) transmission and implementation of control as illustrated in Fig.~\ref{fig:process}.
\begin{figure*}[t]
\centering\includegraphics[scale=0.8]{process.pdf}
\vspace{-0.5cm}
\caption{A communications-computing-control process.}
\label{fig:process}
\vspace{-0.5cm}
\end{figure*}
We consider a process-noise-free discrete-time non-linear plant model
\begin{equation}\label{eq:sys}
\mathbf{x}(t+1) = f(\mathbf{x}(t),\mathbf{u}(t)),
\end{equation}
where $\mathbf{x}(t)\in \mathbb{R}^{l_s}$ and $\mathbf{u}(t)\in \mathbb{R}^{l_u}$ are the plant state and the control input at time $t$.
Note that a more practical model with process noise will be investigated in Section~\ref{sec:robust}.
As a consequence of stochastic computational resources at the controller and packet dropouts in the sensing and control channels (which will be described later in more detail), the plant may have to operate in open loop for arbitrarily long time intervals. This may lead to performance degradation and potential loss of stability. Thus, throughout this work we will analyze the stability conditions of the considered system where the stability is defined as follows.
\begin{definition}\label{def:stability}
\normalfont
The process-noise-free dynamical system \eqref{eq:sys} is stochastically stable, if for some $\psi \in \mathcal{K}_\infty$, the expected value $\sum_{k=0}^{\infty} \myexpect{\psi(|\mathbf{x}(k)|)}<\infty$.
\end{definition}
Our standing assumption is that the plant is globally controllable (in the idealized closed loop case):
\begin{assumption}[\!\!\cite{anytime2,anytime3,anytime5}] \label{assum:rho}
\normalfont
There exist functions $V: \mathbb{R}^{l_s} \rightarrow \mathbb{R}_{\geq 0}, \psi_1,\psi_2 \in \mathcal{K}_\infty$, a constant $\rho\in(0,1)$, and a control policy $\kappa: \mathbb{R}^{l_s} \rightarrow \mathbb{R}^{l_u}$, such that for all $x\in\mathbb{R}^{l_s}$
\begin{equation}\notag
\begin{aligned}
&\psi_1(|\mathbf{x}|) \leq V(\mathbf{x}) \leq \psi_2(|\mathbf{x}|)\\
&V(f(\mathbf{x},\kappa(\mathbf{x}))) \leq \rho V(\mathbf{x}).
\end{aligned}
\end{equation}
\end{assumption}
\begin{assumption}[\!\!\cite{anytime2,anytime3,anytime5}]\label{assum:alpha}
\normalfont
There exists $\alpha>0$ such that $$V(f(\mathbf{x},\mathbf{0}))< \alpha V(\mathbf{x}), \forall \mathbf{x} \in \mathbb{R}^{l_s}.$$
The initial plant state satisfies $$\myexpect{\psi_2( \vert \mathbf{x}(0)\vert )}<\infty,$$ where $\psi_2 \in \mathcal{K}_\infty$ as in Assumption~\ref{assum:rho}.
\end{assumption}
\subsection{Dual Markov Fading Channels}
We consider wireless fading channels for the S-C and C-A transmissions~\cite{KangJIoT,Kang2019ICC,KangTWC}, where wireless channel status varies with time due to multi-path propagation and shadowing caused by obstacles affecting the radio-frequency (RF) wave propagation.
The time-varying channel conditions can be modeled as Markov processes~\cite{Parastoo}.
Furthermore, in practice, the two channel conditions can be correlated, named as spatial correlation, which is caused by the same environment obstacles~\cite{tse2005fundamentals}.
For wireless packet transmissions, there exists a fundamental tradeoff between reliability, i.e., the packet drop probability, and data rate, which determines the amount of information bits that a packet can carry~\cite{Polyanskiy}. For a fixed channel condition, increasing the data rate of a packet can lead to a higher packet drop probability.
For a good channel condition, one can reduce the packet drop probability while maintain the fixed data rate. Alternatively, one can increase the data rate while maintaining the packet drop probability~\cite{Goldsmith}.
In this work, we will keep the data rate fixed for the S-C transmissions, since there is nothing gained from aggregating past sensor measurements in the state feedback case. However, it is well known that sending control sequences can be beneficial to compensate for packet dropouts. To accommodate this in a fading channel environment, we allow C-A transmissions to contain packets of varying data rate. The rates depend on the channel condition and provide a guaranteed packet-dropout probability. Thus, a longer control sequence can be transmitted to the actuator under a better channel condition with the same reliability.
\textbf{C-A channel.} At time slot $t$, the controller can at most transmit $B(t) \in \mathcal{B}\triangleq\{0,1,\cdots, \bar{B}\}$ commands to the actuator with a guaranteed packet drop probability $\bar{\gamma}$.
In other words, $B(t)$ denotes the C-A channel quality.
In this sense, $B(t)$ can be treated as the capacity of the channel under the packet drop probability requirement $\bar{\gamma}$.
Let $\gamma(t)=1$ and $\gamma(t)=0$ denote the successful and failed transmissions in time slot $t$.
\textbf{S-C channel.} At time slot $t$,
the S-C channel power gain takes values from $\{h_1,h_2,\cdots,h_{\bar{B}'}\}$.
Let $B'(t) \in \mathcal{B}' \triangleq \{1,2,\cdots,\bar{B}'\}$ denote the index of the channel power gain. Thus, $B'(t)$ denotes the S-C channel quality. Let $\gamma'(t)=1$ and $\gamma'(t)=0$ denote the successful and failed transmissions in time slot $t$.
The packet drop probability at time~$t$ is
\begin{equation}
\bar{\gamma}'(t)=g(h_{B'(t)}) \in \{g(h_1),g(h_2),\cdots,g(h_{\bar{B}'})\},
\end{equation}
where $g(\cdot)$ is the packet drop probability function in terms of the channel power gain.
Then, we assume that the joint C-A and S-C channel condition $\{(B,B')\}_\mathbb{N}$ is a time-homogeneous Markov process, and the state transition probability is given as
\begin{equation}\label{eq:prob_B}
\begin{aligned}
p_{i,j} &\triangleq \myprob{(B(t+1),B'(t+1))=b_j \vert (B(t),B'(t))=b_i}, \\
& \hspace{5.8cm}\forall b_i,b_j\in \mathcal{B} \times \mathcal{B}'.
\end{aligned}
\end{equation}
\begin{remark}
Our current channel model jointly considers both the spatial-correlated S-C and C-A channels, the time-correlated fading channel conditions and variable data rate requirements. To the best of our knowledge, this has never been considered in the literature of WNCSs and is more general than existing models.
For example, independent dual channels with i.i.d.\ packet dropouts were considered in~\cite{schenato2007foundations}, which is a special case of our model when the fading channels degrade to static ones and the channels' spacial correlation is perfectly canceled.
\end{remark}
\subsection{Anytime Control with Dual Buffers}
When considering perfect transmission between the controller and the actuator, as in \cite{anytime2,anytime3,anytime5}, the system only needs one buffer at the controller to store the computed control commands.
If imperfect transmissions are taken into account, it is convenient to include a command buffer at the actuator to provide robustness against packet dropouts, see e.g.~\cite{QUEVEDO20121803} for a general packetized predictive control method\footnote{Markovian communication and computational resources were not considered in~\cite{QUEVEDO20121803}.}.
Clearly, the dual-buffer system introduces a more complex state updating process.
Let $\Lambda_c$ and $\Lambda_a$ denote the length of the controller's and the actuator's buffer, respectively.
Then, the buffer state at the controller after its transmission phase is denoted as
\begin{equation}
\vec{\mathbf{b}}_c(t) \triangleq [\mathbf{b}_{c,1}(t)^\top,\mathbf{b}_{c,2}(t)^\top,\cdots, \mathbf{b}_{c,\Lambda_c}(t)^\top]^\top,
\end{equation}
where $\mathbf{b}_{c,i}(t)\in\mathbb{R}^{l_u},\forall i\in \{1,2,\cdots,\Lambda_c\}$.
The buffer state at the actuator right after the C-A transmission but before the implementation of a control command, i.e., the pre-control buffer, is denoted as
\begin{equation}
\vec{\mathbf{b}}_a(t)\triangleq [\mathbf{b}_{a,1}(t)^\top,\mathbf{b}_{a,2}(t)^\top,\cdots, \mathbf{b}_{a,\Lambda_a}(t)^\top]^\top,
\end{equation}
where $\mathbf{b}_{a,i}(t)\in\mathbb{R}^{l_u},\forall i\in \{1,2,\cdots,\Lambda_a\}$.
In general, the buffers $\vec{\mathbf{b}}_a(t)$ and $\vec{\mathbf{b}}_c(t)$ keep the calculated sequences of control command, and the buffer updating rules will be given in the following part.
The control input is the first element in the actuator's buffer, i.e.,
\begin{equation}
\mathbf{u}(t) = \mathbf{b}_{a,1}(t),
\end{equation}
which can be treated as the previously predicted control command for the current time slot.
The buffer state at the actuator right after the control implementation, i.e., the post-control state, is
\begin{equation}
\vec{\mathbf{b}}'_a(t) \triangleq \mathbf{S}_a \vec{\mathbf{b}}_a(t),
\end{equation}
where the buffer shift matrices are defined as
\begin{equation}
\color{black}
\mathbf{S}_i \triangleq \begin{bmatrix}
\mathbf{0}_{l_u} & \mathbf{I}_{l_u} & \mathbf{0}_{l_u} & \cdots & \mathbf{0}_{l_u}\\
\vdots & \ddots & \ddots & \ddots & \vdots \\
\mathbf{0}_{l_u} & \cdots & \mathbf{0}_{l_u} & \mathbf{I}_{l_u} & \mathbf{0}_{l_u}\\
\mathbf{0}_{l_u} & \cdots & \cdots & \mathbf{0}_{l_u} & \mathbf{I}_{l_u}\\
\mathbf{0}_{l_u} & \cdots & \cdots & \cdots & \mathbf{0}_{l_u}\\
\end{bmatrix} \in \mathbb{R}^{\Lambda_i l_u \times \Lambda_i l_u}, i=c \text{ or } a.
\end{equation}
Let $N(t)$ denote the number of calculated tentative future control commands at time $t$. The process $\{N\}_{\mathbb{N}}$ is a time-homogeneous Markov process with the transition probability
\begin{equation} \label{eq:prob_N}
q_{i,j} \triangleq \myprob{N(t+1)=j \vert N(t)=i}, i,j\in \mathcal{N},
\end{equation}
where $\mathcal{N}\triangleq \{0,1,\cdots,\bar{N}\}$. It is assumed that $\bar{N} \leq \Lambda_c$.
The \textbf{controller}'s operations are described as
\begin{enumerate}
\item If $\gamma'(t)N(t)>0$, the controller has a new update from the sensor and is available for computation.
In this case, it discards all the existing commands in its buffer and generates a sequence of $N(t)$ control commands to control the plant in time slots $t$ to $(t+N(t)-1)$.
The sequence of tentative controls is
\begin{equation}
\vec{\mathbf{u}}(t) = [\mathbf{u}_1(t)^\top,\mathbf{u}_2(t)^\top,\cdots,\mathbf{u}_{N(t)}(t)^\top]^\top.
\end{equation}
The buffer state before transmission is written as
$$[\vec{\mathbf{u}}(t)^\top,\underbrace{\mathbf{0}_{l_u \times 1}^\top,\cdots,\mathbf{0}_{l_u \times 1}^\top}_{\Lambda_c-N(t)}]^\top. $$
Specifically, the controller calculates the control sequence based on the anytime control algorithm proposed in~\cite{anytime5}, which is rewritten as
\begin{equation}
\begin{aligned}
&\mathbf{u}_i(t) = \kappa(\mathbf{x}_i'(t)), i=1,\cdots,N(t)\\
&\mathbf{x}'_i(t) = \begin{cases}
\mathbf{x}(t), &i=1\\
f(\mathbf{x}'_{i-1}(t),\mathbf{u}_{i-1}(t)), &i=2,\cdots,N(t)
\end{cases}\\
\end{aligned}
\end{equation}
where $\kappa(\cdot)$ is defined in Assumption~\ref{assum:rho}.
Considering the C-A channel capacity and the actuator's buffer length, the controller transmits $\min\{B(t),N(t),\Lambda_a\}$ commands to the actuator. If the transmission is successful, the buffer shifts by $\min\{B(t),N(t),\Lambda_a\}$ steps. Otherwise, the controller erases its buffer. This is because the first computed control command $\mathbf{u}_1(t)$ cannot be implemented in the current time slot $t$, and the rest of computed control commands, which are calculated based on the successful implementation of the first control command, become useless. It is clear that at time instances where the actuator has run out of buffer contents, we have $\vec{\mathbf{b}}_a(t) = \mathbf{0}$ and $\mathbf{u}(t) = \mathbf{0} \neq \mathbf{u}_1(t)$. For the case that $\vec{\mathbf{b}}_a(t) \neq \mathbf{0}$, the predicted control input $\mathbf{u}(t)$ is equal to currently calculated control command $\mathbf{u}_1(t)$ only in the perfect process-noise-free scenario, and $\mathbf{u}(t)\neq\mathbf{u}_1(t)$ in general.
\item If $\gamma'(t)=0$, the controller does not have a new update from the sensor. In this case, it does not generate any new control command.
If $N(t)=0$, the controller does not have the computational resource to generate any new control command.\\
In these two cases, if the plant is out of control in the previous time slot, i.e., $\vec{\mathbf{b}}_a(t-1) = \mathbf{0}$, the controller erases its buffer due to the same reason in case 1); otherwise, the controller sends the buffered commands to the actuator as much as it can, subject to the constraints of the C-A channel capacity and the actuator's buffer length.
\end{enumerate}
The \textbf{actuator}'s operations are described as
\begin{enumerate}
\item If $\gamma(t)=0$, the actuator's buffer is shifted by one step, i.e., $\vec{\mathbf{b}}_a(t) = \mathbf{S}_a \vec{\mathbf{b}}_a(t-1)$, since the first command in the buffer of the previous time slot was used for control.
\item If $\gamma(t)=1$ and $\gamma'(t)N(t)>0$, the actuator erases the previous commands and stores the received ones.
{\color{black}This is because the newly calculated control commands are expected to perform better than the previous calculated commands, especially when the process noise has a large variance.}
\footnote{\color{black} Note that when the number of tentative future commands in the newly received packet is less than the previous ones, we still need to erase all the previous commands, not just part of them. The reason is that each buffered control command at the actuator was calculated assuming the successful implementation of the previously calculated control command. Thus, if the first few control commands at the actuator's buffer are removed (i.e., cannot be applied for control), then the rest of the buffer is of little use. }
\item If $\vec{\mathbf{b}}_a(t-1) = \mathbf{0}$ and $\gamma'(t)N(t)=0$, no operation on the actuator's buffer is required as there is no new commands transmitted.
\item If $\vec{\mathbf{b}}_a(t-1) \neq \mathbf{0}$, $\gamma'(t)=1$ and $\gamma'(t)N(t)=0$, the actuator shifts its buffer by one step due to the same reason in 1) and stores the received commands in the buffer right after the existing commands.
\end{enumerate}
Let $\lambda_c(t)$ and $\lambda_a(t)$ denote the effective buffer lengths at the controller and the actuator, respectively.
Intuitively, $\lambda_c(t)$ and $\lambda_a(t)$ jointly determine the closed-loop performance of plant, as a larger $\lambda_c(t)$ and a larger $\lambda_a(t)$ indicate that the plant will be delivered tentative control values for a longer time.
Based on the controller's and the actuator's operations, the actuator is not necessary to have a larger buffer than the controller. Thus, we assume that $\Lambda_a \leq \Lambda_c$.
Let $L(t)$ denote the number of tentative commands to be transmitted. Based on the controller's operations, $L(t)$ can be written as
\begin{equation} \label{eq:V}
L(t) \!=\! \begin{cases}
\!\min\{B(t),N(t),\Lambda_a\}, &\text{\!\!\!\!\!\!if } \gamma'(t)N(t)>0\\
\!\min\{B(t),\lambda_c(t-1),\Lambda_a-\lambda_a(t-1)\}, & \begin{aligned}
&\text{\!\!\!\!\!if } \gamma'(t)N(t)=0, \\
&\!\!\!\!\!\lambda_a(t-1)\neq 0
\end{aligned} \\
\!0, &\text{\!\!\!\!\!\!otherwise.}\\
\end{cases}
\end{equation}
Then, the buffer-updating rules based on the controller's and the actuator's operations are
\begin{equation} \label{eq:b_c}
\vec{\mathbf{b}}_c(t) \!=\! \begin{cases}
\mathbf{S}_c^{L(t)}\begin{bmatrix}
\vec{\mathbf{u}}(t)\\
\mathbf{0}
\end{bmatrix}, &\text{\!\!\!\!\!if } \gamma'(t)N(t)>0, \gamma(t)=1\\
\mathbf{0}, &\text{\!\!\!\!\!if } \gamma'(t)N(t)>0, \gamma(t)=0\\
\mathbf{0}, &\text{\!\!\!\!\!if } \gamma'(t)N(t)=0,\lambda_a(t-1)=0\\
\mathbf{S}_c^{L(t)} \vec{\mathbf{b}}_c(t-1), &\text{\!\!\!\!\!if } \gamma'(t)N(t)=0,\lambda_a(t-1)\!\neq\! 0, \gamma(t)\!=\!1\\
\vec{\mathbf{b}}_c(t-1),&\text{\!\!\!\!\!if } \gamma'(t)N(t)=0,\lambda_a(t-1)\!\neq\! 0, \gamma(t)\!=\!0\\
\end{cases}
\end{equation}
and
\begin{equation} \label{eq:b_a}
\vec{\mathbf{b}}_a(\!t) \!\!=\!\! \begin{cases}
\mathbf{S}_a\vec{\mathbf{b}}_a(t-1), \text{\hspace{2.2cm} if }\gamma(t)=0\\
\!\!\!\begin{bmatrix}
[\mathbf{u}_1(t)^\top,\cdots,\mathbf{u}_{L(t)}(t)^\top]^\top\\
\mathbf{0}
\end{bmatrix}, \text{ if }\gamma'(t)N(t)>0,\gamma(t)=1\\
\mathbf{0}, \begin{aligned}
\text{\hspace{3.6cm} if }\gamma'(t)N(t)=0,
\vec{\mathbf{b}}_a(t) = \mathbf{0}
\end{aligned}\\
\!\!\!\begin{bmatrix}
\![\mathbf{b}_{a,2}(t-1)^\top,\cdots,\mathbf{b}_{a,\lambda_a(t-1)}(t)^\top]^\top\\
[\mathbf{b}_{c,1}(t)^\top,\cdots,\mathbf{b}_{c,L(t)}(t)^\top]^\top\\
\mathbf{0}
\end{bmatrix}\!\!\!,\! \begin{aligned}
\text{if }\gamma'(t)N(t)\!=\!0, \\
\vec{\mathbf{b}}_a(t) \!\neq\! \mathbf{0}, \gamma(t)\!=\!1.
\end{aligned}\\
\end{cases}
\end{equation}
From \eqref{eq:b_c} and \eqref{eq:b_a}, the updating rules of the effective buffer lengths $\lambda_c(t)$ and $\lambda_a(t)$ are
\begin{equation}\label{eq:prob_c}
\lambda_c(t) = \begin{cases}
N(t) - L(t) ,& \text{if } \gamma'(t)N(t)>0,\gamma(t) =1\\
0 ,& \text{if } \gamma'(t)N(t)>0,\gamma(t) =0\\
0 ,& \text{if } \gamma'(t)N(t)=0,\lambda_a(t-1) =0\\
\lambda_c(t+1) - L(t) ,& \text{\!\!\!\!if } \gamma'(t)N(t)=0,\lambda_a(t-1) \! \neq \! 0,\! \gamma(t)\!=\!1\\
\lambda_c(t-1),& \text{\!\!\!\!if } \gamma'(t)N(t)=0,\lambda_a(t-1) \!\neq\! 0,\! \gamma(t)\!=\!0
\end{cases}
\end{equation}
and
\begin{equation}\label{eq:prob_a}
\lambda_a(t) \!=\! \begin{cases}
\!\max\{\lambda_a(t\!-\!1)\!-\!1,\!0\} ,& \text{\!if } \gamma(t) =0\\
\!L(t) ,& \text{\!if } \gamma'(t)N(t)>0,\gamma(t) =1\\
\!0 ,& \text{\!if } \gamma'(t)N(t)\!=\!0,\lambda_a(t-1) \!=\!0\\
\!\lambda_a(t-1) + L(t)-1 ,& \begin{aligned}
&\text{\!if } \gamma'(t)N(t)\!=\!0,\lambda_a(t-1) \!\neq\! 0, \\
&\gamma(t)=1.
\end{aligned}
\end{cases}
\end{equation}
\section{Stability of the Anytime Control System}\label{sec:stability}
Based on the anytime control method described in \eqref{eq:V}, \eqref{eq:b_c} and \eqref{eq:b_a}, and following the established stability analysis framework adopting stochastic Lyapunov functions~\cite{anytime2,anytime3,anytime5}, to investigate the stability condition of the system~\eqref{eq:sys}, we only need to focus on the plant events that the actuator runs out of control commands, i.e., $\lambda_a(t) =0, \forall t\in\mathbb{N}_0$.
However, since the process $\{\lambda_a\}_{\mathbb{N}_0}$ has an infinite memory and is not a Markov process, the methods in~\cite{anytime2,anytime3,anytime5} are not directly applicable. Instead, we shall analyze the control system through the aggregated Markov process $\{Z\}_{\mathbb{N}_0}$ defined as
\begin{equation}\label{eq:Z}
\begin{aligned}
Z(t) &\triangleq \left(\lambda_c(t),\lambda_a(t),B(t+1),B'(t+1),N(t+1)\right)\\
& \hspace{2.6cm } \in \mathcal{X}_c \times \mathcal{X}_a \times \mathcal{B} \times \mathcal{B}' \times\mathcal{N}, t\in \mathbb{N}_0,
\end{aligned}
\end{equation}
where $\mathcal{X}_c \triangleq \{0,1,\cdots,\Lambda_c\}$ and $\mathcal{X}_a \triangleq \{0,1,\cdots,\min\{\Lambda_a,\bar{N}\}\}$.
Assume that $Z(t), \forall t \in\mathbb{N}_0$, belongs to the finite set $\mathcal{S} \triangleq \{s_0,s_1,\cdots,s_{S}\}$ with cardinality $S$.
Different from~\cite{anytime2}, which only needs to analyze an aggregated process of two processes, we need to investigate the aggregation of five processes, where both $\{\lambda_c\}_{\mathbb{N}_0}$ and $\{\lambda_a\}_{\mathbb{N}_0}$ are correlated with $\{B,B',N\}_{\mathbb{N}_0}$.
Since the control process is divided by the open-loop events with $\lambda_a(t)=0$,
we define $\mathcal{K} = \{k_n\}_{n\in \mathbb{N}_0}$ as the sequence of time steps with $\lambda_a(t)=0$.
We name the time sequence between $k_n$ and $k_{n+1}$ as the $(n+1)$th cycle of the process, $\forall n \in \mathbb{N}_0$.
Then, the number of time steps between consecutive elements of $\mathcal{K}$ is
\begin{equation} \label{eq:delta}
\Delta_{n+1} = k_{n+1}-k_n.
\end{equation}
Without loss of generality, let the set $\mathcal{S}_0 \triangleq \{s_0,s_1,\cdots,s_{S_0}\} \in \mathcal{S}$ with cardinality of $S_0$ denote the subset of $\mathcal{S}$ consisting of all the states with $\lambda_a(t)=0$, and hence $Z(k_n) \in \mathcal{S}_0,\forall n\in\mathbb{N}_0$.
In~\cite{anytime2,anytime3,anytime5}, the set $\mathcal{S}_0$ has only one state. In our scenario, $S_0>1$ introduces more challenges in analyzing the process $\{k_n\}_{n\in \mathbb{N}_0}$.
The state transition process of $\{Z\}_{\mathbb{N}_0}$ is illustrated in Fig.~\ref{fig:set}.
\begin{figure}[t]
\centering\includegraphics[scale=0.6]{set.pdf}
\vspace{-0.3cm}
\caption{An illustration of the state transition process of $\{Z\}_{\mathbb{N}_0}$.}
\vspace{-0.5cm}
\label{fig:set}
\end{figure}
In what follows, we study the properties of $\{Z\}_{\mathbb{N}_0}$ and $\{Z(k_n)\}_{n\in\mathbb{N}_0}$ and then analyze the stability condition.
\subsection{Properties of $\{Z\}_{\mathbb{N}_0}$ and $\{Z(k_n)\}_{n\in\mathbb{N}_0}$}
For ease of analysis, we need the following assumption about the aggregated process $\{(B,B',N)\}_{\mathbb{N}_0}$.
\begin{assumption}\label{assump:BBN}
\normalfont
$\{(B,B',N)\}_{\mathbb{N}_0} \in \mathcal{B} \times \mathcal{B}' \times\mathcal{N}$ is an irreducible and aperiodic (IA) Markov process.
\end{assumption}
Note that wireless fading channel conditions are commonly modeled as IA Markov processes~\cite{Parastoo,KangJIoT,liu2020remote,liu2021remote}, thus it is reasonable to consider an IA Markov process $\{(B,B')\}_{\mathbb{N}_0} \in \mathcal{B} \times \mathcal{B}'$.
Also the computation availability $\{N\}_{\mathbb{N}_0} \in \mathcal{N}$ is commonly modeled as an IA Markov process~\cite{anytime2}. Since an aggregation of IA Markov processes is still an IA Markov process~\cite{durrett2019probability}, it is reasonable to consider the IA Markov process $\{(B,B',N)\}_{\mathbb{N}_0}$.
For the initial state of $\{Z\}_{\mathbb{N}_0}$, it is practical to assume that both buffers are empty:
\begin{assumption}\label{assump:initial}
\normalfont
Let $\lambda_c(0)=\lambda_a(0)=0$, $B(0)=B_0 \in \mathcal{B}$, $B'(0) = B'_0 \in \mathcal{B}'$ and $N(0)=N_0\in \mathcal{N}$.
\end{assumption}
\begin{lemma}\label{lem:Z}
\normalfont
Consider Assumptions~\ref{assump:BBN} and \ref{assump:initial}. Then $\{Z\}_{\mathbb{N}_0}$ is an IA Markov process.
\end{lemma}
\begin{proof}
(1) Irreducibility. Considering the initial state $Z(0)$, to prove that it is possible to get to any state from any state in $\{Z\}_{\mathbb{N}_0}$, we only need to show that any state $(\tilde{\lambda}_c,\tilde{\lambda}_a,\tilde{B},\tilde{B}',\tilde{N}) \in \mathcal{S}$ can return to $Z(0)$~\cite{durrett2019probability}.
Since $\{(B,B',N)\}_{\mathbb{N}_0}$ is an IA Markov process, we assume that $(\tilde{B},\tilde{B}',\tilde{N})$ can get to $(B_0,B'_0,N_0)$ in $l$ steps with probability $p>0$, where $l>\Lambda_a$. Since $\bar{\gamma}>0$, when the packet dropout event consecutively occurs for $l$ times with probability $\bar{\gamma}^l>0$, the effective buffer lengths will be zero at both controller and actuator. Thus, the state $(\tilde{\lambda}_c,\tilde{\lambda}_a,\tilde{B},\tilde{B}',\tilde{N})$ can return to $Z(0)$ in $l$ steps with a non-zero probability.
(2) Aperiodicity. To prove the aperiodicity of $\{Z\}_{\mathbb{N}_0}$, we only need to show that a single state state $Z(0)$ is aperiodic due to the irreducebility of $\{Z\}_{\mathbb{N}_0}$~\cite{durrett2019probability}.
Due to the aperiodicity of $\{(B,B',N)\}_{\mathbb{N}_0}$, the period of the state $(B_0,B'_0,N_0)$ is $1$ and is written as~\cite{durrett2019probability}
\begin{equation}\label{eq:gcd}
\begin{aligned}
&\gcd\{t>0: \myprob{(B(t),B'(t),N(t))=(B_0,B'_0,N_0) \vert \right.\\
&\left. \hspace{2.2cm} (B(0),B'(0),N(0))=(B_0,B'_0,N_0)}>0\}=1.
\end{aligned}
\end{equation}
If state $(B_0,B'_0,N_0)$ can reach itself in $t\in \mathbb{N}$ steps, due to the non-zero probability of consecutive packet dropout of $t$ times, the state $(0,0,B_0,B'_0,N_0)$ can reach itself in $t$ steps as well.
Then, from \eqref{eq:gcd}, the period of the state $(0,0,B_0,B'_0,N_0)$ is $1$.
\end{proof}
From Lemma~\ref{lem:Z}, $\{Z\}_{\mathbb{N}_0}$ has a unique stationary distribution.
We shall denote the state transition probability of $\{Z\}_{\mathbb{N}_0}$ as per
\begin{equation}\label{eq:v_ij}
\begin{aligned}
v_{i,j}& \triangleq \myprob{Z(t+1)=s_j \vert Z(t)=s_i}, \forall s_i,s_j \in \mathcal{S}, t\in \mathbb{N}_0.
\end{aligned}
\end{equation}
Note that $v_{i,j}$ can be numerically calculated based on \eqref{eq:prob_B}, \eqref{eq:prob_N}, \eqref{eq:prob_c} and \eqref{eq:prob_a}, though it does not have a closed-from expression due to the complexity introduced by the dual-buffer updating process \eqref{eq:prob_c} and \eqref{eq:prob_a}.
Let $\mathbf{V}\in \mathbb{R}^{S\times S}$ denote the state transition probability matrix, i.e., $[\mathbf{V}]_{i,j}=v_{i,j}$, and
\begin{equation}\label{eq:V_devide}
\mathbf{V} = \begin{bmatrix}
\mathbf{V}_{0,0} &\aug & \mathbf{V}_{0,1} \\ \hline
\mathbf{V}_{1,0} &\aug & \mathbf{V}_{1,1} \\
\end{bmatrix}
\end{equation}
where $\mathbf{V}_{0,0}\in\mathbb{R}^{S_0 \times S_0}$, $\mathbf{V}_{0,1}\in\mathbb{R}^{S_0 \times (S-S_0)}$,
$\mathbf{V}_{1,0}\in\mathbb{R}^{(S-S_0) \times S_0}$ and $\mathbf{V}_{1,1}\in\mathbb{R}^{(S-S_0) \times (S-S_0)}$.
\begin{lemma}\label{lem:transition}
\normalfont
$\{Z(k_n)\}_{n\in\mathbb{N}_0}$ is an IA Markov process under Assumptions~\ref{assump:BBN} and \ref{assump:initial}. The state transition probability matrix is
\begin{equation}\notag
\tilde{\mathbf{V}} \triangleq \sum_{l=1}^{\infty} \mathbf{D}(l) \in \mathbb{R}^{S_0 \times S_0},
\end{equation}
where
\begin{equation}\notag
\mathbf{D}(l)=\begin{cases}
\mathbf{V}_{0,0},& l=1\\
\mathbf{V}_{0,1} \mathbf{V}_{1,1}^{l-2} \mathbf{V}_{1,0} ,& l>1
\end{cases}
\end{equation}
The stationary distribution of $s_i \in \mathcal{S}_0$, $\pi_i$, is the unique solution of
\begin{equation}
\bm{\pi}^\top \tilde{\mathbf{V}} = \bm{\pi}^\top,
\end{equation}
where $\bm{\pi} \triangleq [\pi_1,\pi_2,\cdots,\pi_{S_0}]^\top$.
\end{lemma}
\begin{proof}
The irreducibility of $\{Z(k_n)\}_{n\in\mathbb{N}_0}$ is obvious, because $\{Z\}_{\mathbb{N}_0}$ is irreducible. For the aperiodicity, we only need to prove that there exist one state of $\{Z(k_n)\}_{n\in\mathbb{N}_0}$ with period $1$~\cite{durrett2019probability}.
Due to the aperiodicity of $\{(B,B',N)\}_{\mathbb{N}_0}$, the period of the state $(B,B',N)$ is $1$.
If the state $(B,B',N)$ can reach itself in $n\in \mathbb{N}$ steps, due to the non-zero probability of consecutive packet dropout of $n$ times, the state $(0,0,B,B',N) \in \mathcal{S}_0$ can reach itself in $n\in \mathbb{N}$ steps as well without passing through any state within $\mathcal{S}_1 \triangleq \mathcal{S}\backslash \mathcal{S}_0$.
Therefore, the period of the state $(0,0,B,B',N)$ of the process $\{Z(k_n)\}_{n\in\mathbb{N}_0}$ is $1$. In the following, we derive the state transition probability matrix.
We define the conditional probability
\begin{equation}
\begin{aligned}
& d_{i,j}(l)\\
& \!\triangleq \! \myprob{\Delta_{n+1}\!=\!l, Z(k_{n+1})\!=\! s_j\vert Z(k_n)\!=\!s_i}, s_i,s_j\!\in\! \mathcal{S}_0, l \in\! \mathbb{N},
\end{aligned}
\end{equation}
and note that
\begin{equation}\label{eq:d_ij}
d_{i,j}(1) = v_{i,j}, \forall s_{i}, s_{j} \in \mathcal{S}_0.
\end{equation}
Let $\tilde{\Delta}_{n+1}$ denote the number of steps to go from $Z(k_n) = s_i \in \mathcal{S}_0$ to $s_j \in \mathcal{S}_1$ without passing through any states in $\mathcal{S}_0$. We define the following conditional probability
\begin{equation}\label{eq:d'_ij}
\begin{aligned}
&\!\tilde{d}_{i,j}(l) \\
&\!\!\triangleq\!\! \myprob{\!\tilde{\Delta}_{n+1}\!\!=\!l,\! Z(\!k_n\!+\!\tilde{\Delta}_{n+1}\!)\!\!=\!s_j \!\vert \! Z(\!k_n\!)\!=\!s_i, k_n \!\!+\! \tilde{\Delta}_{n+1}\!<\!k_{n+1}\!}\!,\\
& \hspace{5.3cm} \forall s_i\in \mathcal{S}_0,s_j\in \mathcal{S}_1, l \in \mathbb{N}
\end{aligned}
\end{equation}
and
\begin{equation}
\tilde{d}_{i,j}(1)= v_{i,j}, \forall s_{i} \in \mathcal{S}_0, s_{j} \in \mathcal{S}_1.
\end{equation}
Then, it can be shown that
\begin{align}
\label{eq:d_new}
\tilde{d}_{i,k}(l+1) &= \sum_{k' \in \mathcal{S}_1}^{} \tilde{d}_{i,k'}(l) v_{k',k},\\
\label{eq:d}
d_{i,j}(l+1) &= \sum_{k \in \mathcal{S}_1}^{} \tilde{d}_{i,k}(l) v_{k,j}.
\end{align}
Let's introduce $\mathbf{D}(l) \in \mathbb{R}^{S_0 \times S_0}$ and $\tilde{\mathbf{D}}(l) \in \mathbb{R}^{S_0 \times (S-S_0)}$, where $[\mathbf{D}(l)]_{i,j} = d_{i,j}(l)$ and $[\tilde{\mathbf{D}}(l)]_{i,j} = \tilde{d}_{i,j}(l)$.
From \eqref{eq:d_new}, we have
\begin{equation}
\tilde{\mathbf{D}}(l+1)=\tilde{\mathbf{D}}(l) \mathbf{V}_{1,1},
\end{equation}
and hence
\begin{equation} \label{eq:D_new}
\tilde{\mathbf{D}}(l) = \tilde{\mathbf{D}}(1) \mathbf{V}_{1,1}^{l-1}.
\end{equation}
Similarly, from \eqref{eq:d}, we have
\begin{equation}\label{eq:D}
\mathbf{D}(l+1)=\tilde{\mathbf{D}}(l) \mathbf{V}_{1,0}.
\end{equation}
Using \eqref{eq:d_ij}, \eqref{eq:d'_ij}, \eqref{eq:D_new} and \eqref{eq:D}, we have
\begin{equation}
\mathbf{D}(l)=\begin{cases}
\mathbf{D}(1) = \mathbf{V}_{0,0},& l=1\\
\tilde{\mathbf{D}}(1) \mathbf{V}_{1,1}^{l-2} \mathbf{V}_{1,0}
= \mathbf{V}_{0,1} \mathbf{V}_{1,1}^{l-2} \mathbf{V}_{1,0} ,& l>1.
\end{cases}
\end{equation}
Therefore, the state transition probability matrix satisfies
$
\tilde{\mathbf{V}} \triangleq \sum_{l=1}^{\infty} \mathbf{D}(l).
$
\end{proof}
\subsection{Analysis of the Stability Condition}
Similar to~\cite{anytime2,anytime3,anytime5}, the stability of the WNCS depends on the statistics of $\{\Delta_n\}_{n\in\mathbb{N}}$ in~\eqref{eq:delta}, which denotes the time duration between consecutive open-loop events. Different to~\cite{anytime2,anytime5}, in the considered case,
the process $\{\Delta_n\}_{n\in\mathbb{N}}$ is not i.i.d.
For WNCS with a single channel, in~\cite{anytime3} an event-triggered setup was considered, leading to $\{\Delta_n\}_{n\in\mathbb{N}}$ which is not i.i.d. However, our current setup is different from~\cite{anytime3}. In particular, $\{\Delta_n\}_{n\in\mathbb{N}}$ is formed by the first return time of a set of states in $\mathcal{S}_0$ rather than a single state (as in~\cite{anytime3}) . Therefore, the approach in~\cite{anytime3} cannot be adopted directly.
In the following, we propose a novel cycle-cost-based approach to obtain sufficient stability conditions.
\begin{lemma} \label{lem:Xi}
\normalfont
The plant is stochastically stable if
\begin{equation}\label{eq:Xi}
\sum_{n=1}^{\infty} \myexpect{\Xi(n)} <\infty,
\end{equation}
where
\begin{equation} \label{eq:Xi_def}
\Xi(n) = \alpha^n \rho^{\sum_{i=1}^{n} (\Delta_i-1)}, \forall n \in \mathbb{N},
\end{equation}
and $\rho$ and $\alpha$ were defined in {\color{black}Assumptions \ref{assum:rho} and \ref{assum:alpha}}, respectively.
\end{lemma}
\begin{proof}
From Assumptions~\ref{assum:rho} and \ref{assum:alpha}, and Definition~\ref{def:stability}, the plant is stable if we can prove that
\begin{equation}
\myexpect{\sum_{t=0}^{\infty} V(\mathbf{x}(t))}<\infty.
\end{equation}
By using Assumptions~\ref{assum:rho} and \ref{assum:alpha}, we have
\begin{equation}
\begin{aligned}
\sum_{t=k_n}^{k_{n+1}-1} V(\mathbf{x}(t))
&\leq \left(1+ \alpha \sum_{l=0}^{k_{n+1}-k_{n}-2} \rho^l\right) V(\mathbf{x}(k_n))\\
&< \left(1+ \frac{\alpha}{1-\rho}\right) V(\mathbf{x}(k_n)),
\end{aligned}
\end{equation}
and hence
\begin{equation}
\sum_{t=0}^{\infty} V(\mathbf{x}(t))\leq \left(1+ \frac{\alpha}{1-\rho}\right)
\sum_{n=0}^{\infty} V(\mathbf{x}(k_n)).
\end{equation}
Then, it is easy to obtain that
\begin{equation}
V(\mathbf{x}(k_n)) \leq \alpha^n \rho^{\sum_{i=1}^{n} (\Delta_i-1)} V(\mathbf{x}(0)).
\end{equation}
Since
$
\myexpect{V(\mathbf{x}(k_n))} \leq \alpha^n \myexpect{\rho^{\sum_{i=1}^{n} (\Delta_i-1)} } \myexpect{V(\mathbf{x}(0))}
$ and $\myexpect{V(\mathbf{x}(0))}<\infty$, \eqref{eq:Xi} is proved.
\end{proof}
From Lemma~\ref{lem:Xi}, to find the stability condition of the system, we only need to investigate the process $\{\Xi\}_{\mathbb{N}}$. However, $\{\Xi\}_{\mathbb{N}}$ is not Markovian, whereas the underlying process $\{Z(k_n)\}_{n\in \mathbb{N}_0}$ is.
In the following, we investigate two stability conditions, with and without exploring the state transition properties of the underlying process. The results are stated in Theorems~\ref{theory:loose} and~\ref{theory:tight}, respectively.
Before proceeding, we need the technical lemma below.
\begin{lemma}\label{lem:loose}
\normalfont
The following inequality holds
$$
\myexpect{\Xi(n+1)} < \left(\frac{\alpha}{\rho} \max_{i,j\in\{1,\cdots,S_0\}} r_{i,j}\right) \myexpect{\Xi(n)}, \forall n \in \mathbb{N},
$$
where
\begin{equation}\notag
\begin{aligned}
r_{i,j} & \triangleq \myexpect{\rho^{\Delta_{n+1}} \vert Z(k_n)=s_i,Z(k_{n+1})=s_j} \\
&= \frac{\sum_{l=1}^{\infty} \rho^l [\mathbf{D}(l)]_{i,j}}{[\tilde{\mathbf{V}}]_{i,j}}<\infty, i,j\in\{1,\cdots,S_0\},
\end{aligned}
\end{equation}
and $\mathbf{D}(l)$ and $\tilde{\mathbf{V}}$ are defined in Lemma~\ref{lem:transition}.
\end{lemma}
\begin{proof}
From \eqref{eq:Xi_def}, it can be shown that
\begin{equation}
\Xi(n+1) = \Xi(n) \frac{\alpha}{\rho} \rho^{\Delta_{n+1}},
\end{equation}
and hence
\begin{equation}
\begin{aligned}
&\myexpect{\Xi(n+1)} = \myexpect{\Xi(n) \frac{\alpha}{\rho} \rho^{\Delta_{n+1}}}\\
&= \frac{\alpha}{\rho} \sum_{i=1}^{S_0}\sum_{j=1}^{S_0} \myexpect{\Xi(n) \rho^{\Delta_{n+1}} \vert Z(k_n)=s_i, Z(k_{n+1})=s_j} \\
& \hspace{4cm} \times \myprob{Z(k_n)=s_i, Z(k_{n+1})=s_j}\\
&\leq \frac{\alpha}{\rho} \max_{i,j\in\{1,\cdots,S_0\}} r_{i,j} \sum_{i=1}^{S_0}\sum_{j=1}^{S_0} \myexpect{\Xi(n) \vert Z(k_n)=s_i, Z(k_{n+1})=s_j}\\ &\hspace{4cm} \times \myprob{Z(k_n)=s_i, Z(k_{n+1})=s_j}\\
&= \frac{\alpha}{\rho} \max_{i,j\in\{1,\cdots,S_0\}} r_{i,j} \myexpect{\Xi(n) },
\end{aligned}
\end{equation}
where
\begin{equation}\label{eq:r_ij}
\begin{aligned}
r_{i,j}& \triangleq \myexpect{\rho^{\Delta_{n+1}} \vert Z(k_n)=s_i,Z(k_{n+1})=s_j}\\
&=\sum_{\Delta_{n+1}=1}^{\infty}\rho^{l}\frac{\myprob{\Delta_{n+1}=l, Z(k_{n+1})=s_j \vert Z(k_n)=s_i}}{\myprob{Z(k_{n+1})=s_j \vert Z(k_n)=s_i}}\\
&= \frac{\sum_{l=1}^{\infty} \rho^l [\mathbf{D}(l)]_{i,j}}{[\tilde{\mathbf{V}}]_{i,j}},
\end{aligned}
\end{equation}
and $r_{i,j}$ is bounded due to the fact that $\sum_{l=1}^{\infty} [\mathbf{D}(l)]_{i,j}<1$ and $\rho <1$.
\end{proof}
Using Lemmas~\ref{lem:Xi} and~\ref{lem:loose}, it is straightforward to have the following result.
\begin{theorem}\label{theory:loose}
\normalfont
Suppose that Assumptions~\ref{assum:rho}-\ref{assump:initial} hold. The system~\eqref{eq:sys} is stochastically stable if
$$\Omega' \triangleq \frac{\alpha}{\rho}\max_{i,j\in\{1,\cdots,S_0\}} r_{i,j} < 1.$$
\end{theorem}
\begin{remark}
For the special case with $S_0 = 1$, i.e., there exists only a single state in $\mathcal{S}$ leading to no control being applied, Lemma~\ref{lem:transition} shows that the processes $\{Z(k_n)\}_{n\in\mathbb{N}_0}$ is i.i.d. and so is the process $\{\Delta_n\}_{n\in\mathbb{N}}$. Such a special case is identical to that of \cite{anytime2}, and we see that
the stability condition in Theorem~\ref{theory:loose} reduces to
\begin{equation}
\Omega' = \frac{\alpha}{\rho} \sum_{l=1}^{\infty} \rho^l \myprob{\Delta_n=l} < 1,
\end{equation}
which is identical to that of~\cite{anytime2}.
\end{remark}
\begin{remark}
The sufficient condition in Theorem \ref{theory:loose} is obtained by considering the worst case scenario of the state transitions, i.e., considering the pair of states $s_i$ and $s_j$ such that it has the largest conditional expectation $\myexpect{\rho^{\Delta_{n+1}} \vert Z(k_n)=s_i,Z(k_{n+1})=s_j}$.
However, such a method does not take into account the state transition probabilities in $\mathbf{\tilde{V}}$ defined in Lemma~\ref{lem:transition}. Thus, the sufficient condition in Theorem~\ref{theory:loose} is conservative.
\end{remark}
The following lemma is needed to obtain a less conservative stability condition in Theorem~\ref{theory:tight}.
\begin{lemma} \label{lem:Omega}
\normalfont
Assuming that $\{Z\}_{\mathbb{N}_0}$ evolves in the steady state, for any arbitrarily small $\epsilon>0$, there exist $\mu>0$ and $K>0$ such that
\begin{equation}\label{eq:lem:key}
\myexpect{\Xi(n)} < \mu \Big(\frac{\alpha}{\rho} \left(\lambda_{\max}(\mathbf{U})+\epsilon\right)\Big)^{n-1}, \forall n> K,
\end{equation}
where
$
\mathbf{U} \triangleq
\begin{bmatrix}
\mathbf{U}_1^\top &\aug &
\mathbf{U}_2^\top &\aug &
\cdots &\aug&
\mathbf{U}_{S_0}^\top
\end{bmatrix}^\top\in \mathbb{R}^{S^2_0 \times S^2_0},
$
\begin{equation} \notag
\mathbf{U}_i \triangleq
\begin{bmatrix}
r_{1,i} \mathbf{f}^\top_1& \mathbf{0} &\cdots &\mathbf{0}\\
\mathbf{0} & r_{2,i} \mathbf{f}^\top_2 &\cdots &\mathbf{0}\\
\vdots & \cdots &\ddots & \vdots \\
\mathbf{0} & \cdots &\mathbf{0} & r_{S_0,i} \mathbf{f}^\top_{S_0}
\end{bmatrix} \in \mathbb{R}^{S_0 \times S^2_0},
\end{equation}
and $\begin{bmatrix}
\mathbf{f}_1 &\aug & \mathbf{f}_2&\aug & \cdots &\aug&\mathbf{f}_{S_0}
\end{bmatrix} \triangleq \mathbf{Z} \tilde{\mathbf{V}} \mathbf{Z}^{-1}$ and $\mathbf{Z} = \text{diag}\{\bm{\pi}\}$.
\end{lemma}
\begin{proof}
See Appendix A.
\end{proof}
From Lemmas~\ref{lem:Xi} and~\ref{lem:Omega}, it is straightforward to derive the following result.
\begin{theorem} \label{theory:tight}
\normalfont
Suppose that Assumptions~\ref{assum:rho}-\ref{assump:initial} hold. System~\eqref{eq:sys} is stochastically stable if
$$\Omega \triangleq \frac{\alpha}{\rho}\lambda_{\max}(\mathbf{U}) < 1.$$
\end{theorem}
\begin{remark}
This stability condition has considered the effect of all the conditional expectations $\myexpect{\rho^{\Delta_{n+1}} \vert Z(k_n)=s_i,Z(k_{n+1})=s_j}$ and the state transition probability matrix of $\{Z(k_n)\}_{n\in\mathbb{N}}$, i.e., $\tilde{\mathbf{V}}$.
Using Perron-Frobenius theorem~\cite{Perron}, since $\mathbf{U}$ is non-negative, we have $\lambda_{\max}(\mathbf{U})\leq \max_{i\in\{1,\cdots,S_0\}}\sum_{j=1}^{S_0}[\mathbf{U}]_{i,j} \leq \max_{i,j\in\{1,\cdots,S_0\}}r_{i,j}$. Thus, the sufficient condition Theorem~\ref{theory:tight} is less restrictive than that in Theorem~\ref{theory:loose}.
\end{remark}
\subsection{Numerical Example for Stability Condition Calculation}
{\color{black}
We set the buffer lengths $\Lambda_c = \Lambda_a =2$, the maximum number of calculated control commands per time step $\bar{N} = 2$, the maximum number of commands that can be transmitted via the C-A channel $\bar{B} = 2$, the number of channel states of the S-C channel $\bar{B}'=2$, and the packet drop probabilities in the two states as $0.2$ and $0.01$.
We assume that $\rho = 0.8$.
The Markov state transition probability matrices of the C-A and S-C channel state $(B(t),B'(t)) \in \{0,1,2\}\times \{1,2\}$, and the processor's computational resource $N(t)\in \{0,1,2\}$ are
\begin{equation}
\mathbf{M} = \begin{bmatrix}
0.24& 0.16& 0.06& 0.04& 0.30& 0.20\\
0.04& 0.36& 0.01& 0.09& 0.05& 0.45\\
0.12& 0.08& 0.06& 0.04& 0.42& 0.28\\
0.02& 0.18& 0.01& 0.09& 0.07& 0.63\\
0& 0& 0.30 & 0.20 & 0.30 & 0.20\\
0 & 0& 0.05 & 0.45 & 0.05 & 0.45
\end{bmatrix},
\end{equation}
and
\begin{equation}
\mathbf{N} = \begin{bmatrix}
0.1 & 0.2 & 0.7\\
0 & 0.6 & 0.4\\
0.1 & 0.3 & 0.6
\end{bmatrix}.
\end{equation}
Thus, the process $Z(t)$ defined in~\eqref{eq:Z} has $3 \times 3 \times 3\times 2\times3 =162$ states.
Using the state transition rules of $\lambda_c(t)$ and $\lambda_a(t)$ in \eqref{eq:prob_c} and \eqref{eq:prob_a}, and the channel and computational resource state transition probabilities~\eqref{eq:prob_B} and \eqref{eq:prob_N}, the state transition matrix of $Z(t)$ can be calculated.
Due to the space limitation, the matrix is presented in~\cite{mymatrix} with $72$ transient states (highlighted in yellow), and thus $Z(t)$ has $S= 162-72=90$ recurrent (effective) states including $S_0=54$ states with $\lambda_a(t) =0$ (highlighted in red).
By removing the transient states, the $90 \times 90$ state transition matrix $\mathbf{V}$ of the effective states can be easily obtained.
Due to the space limitation, it is not possible to show the calculation of the stability conditions based on the $90 \times 90$ matrix $\mathbf{V}$. So we only present the comparison result of the stability conditions in Theorems~\ref{theory:loose} and~\ref{theory:tight}:
\begin{equation}
\max_{i,j\in\{1,\cdots,S_0\}} r_{i,j} = 0.8
\gg
\lambda_{\max}(\mathbf{U}) = 0.0016,
\end{equation}
showing that the sufficient stability condition in Theorem~\ref{theory:tight} is less restrictive than Theorem~\ref{theory:loose}.
For ease of illustration, we will show the calculation of the stability conditions based on a randomly generated (small) $\mathbf{V}$ with $S=4$ and assume that $S_0 =2$, where
\begin{equation}
\mathbf{V} =
\begin{bmatrix}
0.10& 0.10& \aug& 0.10& 0.70\\
0.30& 0.20& \aug& 0.10& 0.40\\
\hline
0.60& 0.20& \aug& 0.10& 0.10\\
0.90& 0.05& \aug& 0.02& 0.03
\end{bmatrix}
\in \mathbb{R}^{4\times4},
\end{equation}
and thus $\mathbf{V}_{0,0}$, $\mathbf{V}_{0,1}$, $\mathbf{V}_{1,0}$ and $\mathbf{V}_{1,1}$ are obtained directly from \eqref{eq:V_devide}.
Using Lemma~\ref{lem:transition}, the state transition matrix of $\{Z(k_n)\}_{n\in\mathbb{N}_0}$ and its stationary distribution are obtained as
\begin{equation}\label{eq:V_tilder}
\tilde{\mathbf{V}} = \begin{bmatrix}
0.8378& 0.1622\\
0.7546& 0.2454
\end{bmatrix},
\end{equation}
and
\begin{equation}\label{eq:pi}
\bm{\pi} = [0.8231\ 0.1769]^\top.
\end{equation}
Taking \eqref{eq:V_tilder} and \eqref{eq:pi} into \eqref{eq:r_ij}, we have
$r_{i,j} = [\mathbf{R}]_{i,j}$, where
\begin{equation}\label{eq:R}
\mathbf{R} = \begin{bmatrix}
0.6511& 0.7323\\
0.6971& 0.7673
\end{bmatrix}.
\end{equation}
Taking \eqref{eq:V_tilder}, \eqref{eq:pi} and \eqref{eq:R} into the definition of $\mathbf{U}$ in Lemma~\ref{lem:Omega}, we have
\begin{equation}\label{eq:U}
\mathbf{U} =\begin{bmatrix}
0.3695& 0.0716& 0& 0\\
0& 0& 0.0165& 0.0054\\
0.4156& 0.0805& 0& 0\\
0& 0& 0.0181& 0.0059
\end{bmatrix}.
\end{equation}
By applying \eqref{eq:R} and \eqref{eq:U} onto Theorems~\ref{theory:loose} and~\ref{theory:tight}, respectively, we have
\begin{equation}
\max_{i,j\in\{1,\cdots,S_0\}} r_{i,j} = 0.7673
>
\lambda_{\max}(\mathbf{U}) = 0.3731.
\end{equation}}
\section{Robustness to Process Noise} \label{sec:robust}
In this section, we investigate the stability condition of the plant system below with process noise:
\begin{equation}\label{eq:sys_noise}
\mathbf{x}(t+1) = f(\mathbf{x}(t),\mathbf{u}(t),\mathbf{w}(t)),
\end{equation}
where $\mathbf{w}(t)\in \mathbb{R}^{l_s}$ is a white noise process, which is independent with the other random processes of the system.
For ease of analysis, we consider uniform bounds and continuity as follows.
\begin{assumption}[\!\!\cite{anytime2}]\label{assum:robust}
\normalfont
There exists $\beta_x$, $\beta_u$, $\beta_w$, $\beta_V$, $\beta_{\kappa}$, $\rho$, $\alpha$ and $\eta \in \mathbb{R}_{>0}$, such that, $\forall \mathbf{x}, \mathbf{z},\mathbf{w} \in \mathbb{R}^{l_s}$ and $\forall \mathbf{u},\mathbf{v}\in \mathbb{R}^{l_u}$ the following are satisfied
\begin{equation}
\begin{aligned}
& |f(\mathbf{x,u,w}) - f(\mathbf{z,v,0})| \leq \beta_x |\mathbf{x-z}| + \beta_u |\mathbf{u-v}| + \beta_w |\mathbf{w}|\\
& |V(\mathbf{x}) - V(\mathbf{z})| \leq \beta_V |\mathbf{x-z}|\\
& |\kappa(\mathbf{x}) - \kappa(\mathbf{z})| \leq \beta_{\kappa} |\mathbf{x-z}|\\
& V(f(\mathbf{x},\kappa(\mathbf{x}),\mathbf{w})) \leq \rho V(\mathbf{x}) + \eta \vert \mathbf{w} \vert\\
& V(f(\mathbf{x},\mathbf{0},\mathbf{w})) \leq \alpha V(\mathbf{x}) + \eta \vert \mathbf{w} \vert.
\end{aligned}
\end{equation}
\end{assumption}
When considering unbounded process noise, the stability condition in Definition~\ref{def:stability} cannot be satisfied. Thus, we consider the following stability condition in terms of the average cost~\cite{anytime2}.
\begin{definition}\label{def:stability_rubust}
\normalfont
The dynamical system with process noise \eqref{eq:sys_noise} is stochastically stable, if for some $\psi \in \mathcal{K}_\infty$, the average expected value $\limsup\limits_{T\rightarrow \infty}\frac{1}{T}\sum_{k=0}^{T} \myexpect{\psi(|\mathbf{x}(k)|)}<\infty$.
\end{definition}
\begin{theorem}\label{theory:robust}
\normalfont
Suppose that Assumptions~\ref{assum:rho}-\ref{assum:robust} holds. The system~\eqref{eq:sys_noise} is stochastically stable if
$\Omega < 1$.
\end{theorem}
\begin{proof}
See Appendix B.
\end{proof}
\begin{remark}
Theorem~\ref{theory:robust} shows that the stability condition for the process-noise-free case holds for the process-noise-present one as well under Assumption~\ref{assum:robust}, which is in line with~\cite{anytime2}.
Note that in the case of Markov jump linear systems, stability conditions for noise-free cases are equivalent to conditions for noisy-cases as well, see Theorem 3.33 of Chapter 3 of \cite{costa2006discrete}.
\end{remark}
\section{Simulation Results}
{\color{black}
We present simulation results of the proposed dual-buffer anytime control algorithm and the baseline single-buffer algorithm~\cite{anytime2}, where the S-C and C-A channels are modeled as multi-state Markov chains.
The system parameters are the same as in the numerical example in~Section~\ref{sec:stability}.
We consider an open-loop unstable constrained plant model of~\eqref{eq:sys_noise} in the presence of noise~\cite{anytime3}:
\begin{equation}
\begin{bmatrix}
x_1(t)\\
x_2(t)
\end{bmatrix}
=
\begin{bmatrix}
x_2(t) + u_1(t)\\
-\text{sat}(x_1(t)+x_2(t)) + u_2(t)
\end{bmatrix}
+
\begin{bmatrix}
w_1(t)\\
w_2(t)
\end{bmatrix},
\end{equation}
where
\begin{equation}
\text{sat}(\mu) =
\begin{cases}
-10, \text{ if } \mu<-10\\
\mu, \text{ if } \mu \in [-10,10]\\
10, \text{ if } \mu>10
\end{cases}
\end{equation}
and the process noise $w_1(t)$ and $w_2(t)$ are independent zero-mean i.i.d. Gaussian and $\myexpect{w_1(t)^2}=\myexpect{w_2(t)^2} = 0.1$.
We take the control policy as $\kappa(\mathbf{x}) = [-x_2,\ 0.505 \text{sat}(x_1+x_2)]^\top$.
In Fig.~\ref{fig:simu}, we show the simulation of $|\mathbf{x}(t)|$ with $800$ time steps for both the dual-buffer anytime method and the baseline method. It is clear that the dual-buffer method significantly outperforms the baseline especially when the initial process state $\mathbf{x}(0)$ is far from the origin.
Therefore, adding the command buffer at the actuator can effectively overcome the effect of packet dropouts, leading to significant control performance improvement.
\begin{figure}[t]
\centering\includegraphics[scale=0.6]{simu.pdf}
\vspace{-0.3cm}
\caption{Simulation of $|\mathbf{x}(t)|$ with different initiations.}
\vspace{-0.5cm}
\label{fig:simu}
\end{figure}
}
\section{Conclusions}
We have studied an anytime control algorithm for dual-channel-dual-buffer WNCS with random computational resources over correlated channels.
We have proposed a novel approach to derive sufficient conditions for stochastic stability in the case of nonlinear plant models with disturbances.
The stability conditions are stated in terms of plant dynamics, network dynamics, buffer properties and computational resource dynamics.
Our numerical results have shown that the proposed dual-buffer anytime control system can provide significant control performance improvement when compared to the conventional single-buffer system as it effectively overcomes the effect of packet dropouts.
In future work, we will consider optimal control sequence design.
In addition, we may include an extension to large-scale WNCSs with multiple plants and controllers and that need to share network and communication resources.
|
\section{Introduction}
\setkeys{Gin}{draft=false}
Complex magnetic lattice configurations have gained attention both theoretically and experimentally during the last decade as part of the search for topological and non-collinear magnetic orders, flat-band and Kitaev interactions, and the elusive quantum spin liquid state\cite{kita:06:AoP,knol:19:ARoCMP,gang:08:PRB,chis:15:PRL,hara:09:PRL,hara:09:JPCM,mukh:15:PRL,yang:20:PRR,zhou:17:RMP,sava:16:RPP,bale:10:nature,fish:18:book}. These magnetic systems are so interesting because of their potential in technology for sustainable, energy-efficient memory devices and computational power\cite{dien:11:HMM}.
One such lattice that has increased in popularity is the Kagome lattice, which is described as a two-dimensional triangular three-sublattice structure and the foundational structure for the three-dimensional pyrochlore lattice\cite{essa:17:PRB,ross:09:PRL,gao:18:PRB,bent:20:arx}. This extra sublattice helps distinguish the Kagome from the two-sublattice honeycomb lattice\cite{boyk:18:PRB}. As shown in Fig. \ref{structure}(a), the Kagome lattice consists of a lattice of coupled trimers, which introduces multiple inversion points. Theoretical studies on materials with Kagome lattices have qualified their relevance to these technologies and have further probed for numerous magnetic and electronic properties\cite{wang:20:PCCP,harr:13:PRB,obri:10:PRB,li:18:SA,tran:20:arx,maks:17:PRB,ochi:17:JPSJ,schm:04:PRB,maks:15:PRB,yama:04:IJQC,moes:00:CJP,gov:02:JPCM,ower:16:PRB}. Experimental studies have offered a different perspective on many of the same properties\cite{zhan:20:PRB,shar:20:IEEE,wu:19:IEEE,yin:18:N,yin:19:NP,yazy:19:NP,rigo:07:PRL,zhan:20:PRB,yin:20:NC}.
Previously, Boyko et al. examined the spin-wave dynamics of the Kagome lattice for three different magnetic configurations [out-of-plane ferromagnetic (FM), out-of-plane antiferrimagnetic (AfM), and 120$^{\circ}$ phase] and various isotropic nearest, next-nearest, and next-next-nearest neighbor interactions\cite{boyk:20:AdP}. In this work, it was shown that, to first order, the FM phase produced three modes, wherein two modes were dispersive, long-range order modes, and one was a non-dispersive, cluster-like flat band. These types of cluster modes are not unheard of as they have been observed in structures like the pyrochlore lattice\cite{gao:18:PRB}. Furthermore, Boyko et al. also revealed that, to first order, the 120$^{\circ}$ phase mimics the antiferromagnetic (AFM) honeycomb lattice due to the net in and out spin configurations\cite{boyk:20:AdP}. However, unlike the AFM honeycomb lattice, the Kagome lattice can break this degeneracy with second- and third-order interactions.
\begin{figure}
\includegraphics[width=3.25in]{Fig1-Structures-small.eps}
\caption{The atoms in a Kagome lattice are colored by spin orientation. The various spin configurations are the ferromagnetic (FM - all spin angles in the same direction) (a), antiferrimagnetic (AfM - spin angles are anti-aligned) (b), and the $120^\circ$ phase, where each spin is rotated 120 degrees away from its neighbors' (c). The $120^\circ$ phase makes up triangular ``in" and ``out" atomic spin groups (e), which further illustrate its AFM nature (d).}
\label{structure}
\end{figure}
When considering the collinear Heisenberg model, exchange competition can produce frustration in the system and require an axial anisotropy to stabilize due to underlying non-collinear states\cite{boyk:20:AdP,hara:09:JPCM,fish:18:book}. Frustrated states can also come from the interactions between orbitals that typically result in FM or AFM orders but may become complicated by the competition between interactions as well as any induced crystal-field anisotropies. The frustration in the triangular interactions can lead to the need for more complex interactions to describe the excitations of the system, which is a fairly standard approach as any deviation of a known model indicates the need for new and exciting physics. In the Kagome lattice, this complexity tends to lead researchers into the realm of more exotic interactions like Kitaev model and spin liquids\cite{kita:06:AoP,gang:08:PRB,knol:19:ARoCMP}.
\begin{figure}
\includegraphics[width=3.25in]{Fig2-Brillouin-small.eps}
\caption{(a) The triangular subgeometry on which our calculations are focused. The exchange interactions $(J_{ij})$, which may not all be equal between all atom pairs.
(b) A top-down view of the FM spin waves in a Kagome lattice with symmetry points M, K and $\Gamma$ overlaid.}
\label{brillouin}
\end{figure}
Another avenue that can alleviate these frustrations is the presence of some underlying, possibly non-collinear, magnetic ground state that is not considered\cite{hara:09:PRL}. Given the large number of magnetic configurations for this structure, many groups are forced to examine numerical methods to interpret and understand neutron scattering experiments. While these numerical methods provide critical information for the characterization and identification of magnetic systems, there is little understanding of how the system's interactions compete to produce the given state, which can hinder further development of magnetic materials. By examining analytical solutions and the evolution of complexity for simple models, one can gain insight into the effects of exchange interactions on various configurations of spins in the Kagome lattice. These insights can aid experimentalists and theorists (or modelers) in the identification of different magnetic orders by simple comparison, which can, in turn, allow for the tunability of structures.
In this study, we look to understand many of the Kagome lattice's magnetic configurations by first producing analytical solutions for the spin waves of in-plane and out-of-plane ferromagnetic (FM) and antiferrimagnetic (AfM) arrangements along with the in-plane 120$^{\circ}$ phase, then numerically characterizing various spin states with a distorting nearest neighbor interaction. The term antiferrimagnetic denotes a mixed composition of AFM and FM aligned spins, which result in a structure that still produces a net magnetic moment (shown in Fig. \ref{structure}(b)). Using a Heisenberg spin-spin exchange Hamiltonian with on-site anisotropy, we determine the energy phase diagrams for this distorted system as well as the spin-wave dynamics. Furthermore, we look beyond the 120$^{\circ}$ phase and push the analytical limit by generalizing the in-plane magnetism to a 120$^{\circ}$ + $d\theta$ phase, where $d\theta$ goes from -120$^{\circ}$ (FM phase) to 60$^{\circ}$ (AfM phase).
These calculations allow for a detailed understanding of the changes in the spin-wave dynamics expected for various configurations of the magnetic Kagome lattice, which is useful for experimental identification and interpretation. Furthermore, deviation from these spin-wave dynamics provides evidence for non-standard interactions like those determined by the Kitaev model or other non-Heisenberg models.
\begin{figure}[!]
\includegraphics[width=3.25in]{Fig3-GHandQRLegend.eps}
\caption{The spin angles for the three-atom structure are represented as deviations from a default state both in and out of the xy-plane. (a) The azimuthal spin angles $\theta$, illustrated by the red arrows on each atom, are represented as deviations from the angles $\theta_{FM}$ (gray arrows) that reflect the FM state. As the significance of these angles is their relation to one another rather than absolute orientation, the angle $\theta_A$ was set arbitrarily to 0. Deviation angles $q$ and $r$ are taken as the difference between angle $\theta$ and $\theta_{FM}$. (b) The in-plane spin angles $\phi$, illustrated by the red arrows on each atom, are represented as deviations from the angles $\phi_{120}$ (gray arrows) which make up the 120$^\circ$ phase. The angle $\phi_A$ is set as $\pi/2$. Deviation angles $g$ and $h$ are taken as the difference between angle $\phi$ and $\phi_{120}$}
\label{GHandQRDiagram}
\end{figure}
\begin{figure}[!]
\includegraphics[width=3.5in]{Fig4-PhaseDiagrams.eps}
\caption{(a) Phase diagram for spin waves out of the plane for all values of $J$ and $\mathcal{D}$. (b) and (c) are the in-plane diagrams for all values of $\mathcal{D}$, where (b) is for positive $J$ values and (c) for negative values. The phase borders were determined by setting the simplified classical energies for each pair of phases equal to each other. The configuration for each region was determined by testing which phase offered the lowest energy for the $\alpha$ and $\beta$ values encompassed by that region. There may be other spin configurations to consider in thoroughly characterizing these phases. However, these would require a considerable amount of extra parameterization. To remain within two dimensions and make as few assumptions as possible, we consider only the most basic configurations for this portion of the analysis.}
\label{Phase_Diagrams}
\end{figure}
\section{Spin-exchange Hamiltonian}
To gain a complete understanding of how local interaction changes can affect the spin state of the Kagome lattice, a Heisenberg spin-spin Hamiltonian is used to model first-order interactions with variable exchange values (shown in Fig. \ref{brillouin}(a)). From this Hamiltonian, the energy eigenstates and spin-wave dynamics are examined within the analytical limit and expanded numerically to explore the spin evolution of the magnetic properties, where the spin-spin exchange Hamiltonian with $z$-axis anisotropy is given as
\begin{equation} \label{Energy2}
H = -\frac{1}{2} \sum_{i \neq j} J_{ij} \bar S_i \cdot \bar S_j - D \sum_{i} \bar S_{iz}^2.
\end{equation}
Here, $J_{i,j}$ is the exchange interaction between the spin sites and $D$ is the anisotropy energy\cite{hara:09:JPCM,fish:18:book}. Since the Kagome lattice can produce numerous collinear and non-collinear phases, we must be able to consider the azimuthal and polar angles of the site spins. Therefore, the Hamiltonian must be shifted to study non-collinear spin configurations. As described in Ref. \cite{hara:09:JPCM}, the spin rotation is performed using an Euler rotation matrix $U$ upon the Hamiltonian, which is dependent on spherical coordinates $\theta$ (azimuthal) and $\phi$ (polar) between the two spins \cite{hara:09:JPCM,hara:09:PRL}. Applying this rotation to the Hamiltonian gives
\begin{equation} \label{Energy3}
H = \displaystyle - \frac{1}{2} \sum_{i \neq j} J_{ij} \bar S_i \cdot \underline{U}_i\underline{U}_j^{-1} \bar S_j
- D \sum_{i} \underline{U}_i^{-1} \bar S_{iz}^2.
\end{equation}
\begin{figure}
\includegraphics[width=3.5in]{Fig5-AlphaVsQG.eps}
\caption{$\alpha$ as a function of $q$ and $g$ such that classical energy is minimized for the out-of-plane and in-plane spin configurations, respectively, where $J$ is positive. Negative $J$ values are characterized by a reflection of this graph about the $\alpha$ axis.}
\label{AlphaVsQG}
\end{figure}
Through a $(1/S)$ Holstein-Primakoff expansion of this Hamiltonian, the system breaks into various orders
\begin{equation} \label{hamexpansion}
H = E_0 + H_1 + H_2 + \cdot \cdot \cdot .
\end{equation}
Here, $E_0$ is the classical energy, which can be used to determine the system's overall ground state for a given spin configuration. The $H_1$ term is the vacuum contribution to the spin dynamics, which vanishes in a stable system. $H_2$ produces the first-order contributions to the spin-wave dynamics within the quadratic limit. Higher order terms can also be determined. However, these terms produce quantum fluctuations which can be ignored for large $S$\cite{hara:09:JPCM}.
\begin{figure*}
\includegraphics[width=\textwidth]{Fig6-EquilateralSpinWaves-small.eps}
\caption{The spin-wave dynamics for the simplest spin configurations, where $J = J' = J''$ ($\alpha = \beta = \mathcal{J}$) and no magnetic field is present. (a), (d), (g), and (j) draw out the spin waves through the path in $k$-space which passes through symmetry points $\Gamma$, M, K, then back to $\Gamma$ for the out-of-plane FM, out-of-plane AfM, in-plane FM, and 120$^\circ$ configurations, respectively. (b) and (c) are the out-of-plane FM ($\mathcal{J}= 1$) spin waves with anisotropies $\mathcal{D}$ = 0 and $\mathcal{D}$ = 1, respectively. (e) and (f) show a similar evolution in the AfM configuration ($\mathcal{J} = -1$). However, since the equilateral spin waves are not stable without anisotropy, the anisotropies here are $\mathcal{D}$ = 1 (e) and $\mathcal{D}$ = 2 (f). (h) and (i) are the in-plane FM ($\mathcal{J} = 1$) spin waves with anisotropies $\mathcal{D}$ = 0 and $\mathcal{D}$ = -1, respectively. (k) and (l) are spin waves for the 120$^\circ$ ($\mathcal{J} = -1$) configuration. Although stable with none (k), (l) reflects a large anisotropy value of -3 to illustrate the full transformation that this configuration's spin waves undergo as this parameter is amplified. Color scales applied to the spin waves are grouped by configuration/column. All 3D figures include a projection of the central energy level of the FM configuration spin waves onto the $k$-plane to visualize the symmetry points. The heat map projection's colors are scaled relative to that of that single energy level's minima and maxima, as in figure \ref{brillouin}(b)}
\label{Equilateral_Spinwaves}
\end{figure*}
The Kagome system can produce many spin configurations. The most well-known are collinear FM and AfM systems. The spins in two configurations can be either in-plane or out-of-plane and make distinct changes in the spin dynamics. Outside of the collinear systems is the non-collinear 120$^{\circ}$ phase [shown in Figs. \ref{structure}(c)-(e)], which is typically an in-plane rotation of spin produced through AFM frustration.
First, we examine the classical energy for this system and then move on to the spin-dynamics. Within these sections, we will investigate the out-of-plane and in-plane structures for the FM and AfM phases. With those established, we will then enable a magnetic distortion of the exchange interactions and simulate the evolution of the spin dynamics of both collinear and non-collinear phases with various parameters.
\section{Classical Energy}
Before determining the spin-wave dynamics for specific configurations, the classical energy of the system is used to determine the ground-state spin configuration within the 3 sublattice (3-SL) Kagome system.
\begin{figure*}[!]
\includegraphics[width=6.0in]{Fig7-IsoscelesSpinWavesOut-small.eps}
\caption{Variously distorted isosceles spin waves for the out-of-plane configurations where, in all cases, $\mathcal{J} = 1$. The first row shows the spin-wave energy values through the path in $k$-space that passes through symmetry points $\Gamma, M, K$, and back to $\Gamma$. These 2-D representations exclude half of the $\alpha$ values shown in the 3D graphs due to a large amount of overlap rendering the paths unreadable. The $\alpha$ value for each row of graphs is given by its central-column triangle diagram. The coloring of atoms indicates the AfM configuration. While the FM state would be more accurately by one uniform color, the dichromatic coloring of the AfM phase was chosen to represent either state here. Unstable spin waves were included to illustrate the results of the exchange interaction manipulations that oppose the intuited signs. For both configurations, the evolution of these graphs with anisotropy was included in its second column. All 3D FM graphs (first two columns) are color-grouped together and all 3D AfM graphs are similarly grouped.}
\label{Isosceles_Spinwaves_Out}
\end{figure*}
\begin{figure*}[!]
\includegraphics[width=6.0in]{Fig8-IsoscelesSpinWavesIn-small.eps}
\caption{Variously distorted isosceles spin waves for the in-plane FM and AfM configurations, each of which has $\mathcal{J} = 1$. The first row shows the spin-wave energy values through the path in k-space that passes through symmetry points $\Gamma$, M, K, then back to $\Gamma$. These each exclude half of the $\alpha$ values shown in the 3D graphs, as a large amount of overlap with the other values made them too difficult to read. The $\alpha$ value for each row of graphs is given by the diagram in the central column. For both configurations, the evolution of the spin waves with anisotropy was included in the second column for each. FM graphs are color-grouped together and all AfM graphs are similarly grouped.}
\label{Isosceles_Spinwaves_In}
\end{figure*}
The classical energy of each magnetic configuration is given by
\begin{equation} \label{E_0}
\begin{array}{ll}
E_0 & \displaystyle = -\frac{1}{2}\sum_{i,j}J_{ij}S^2 \Big(\sin(\theta_i)\sin(\theta_j)\cos(\phi_j-\phi_i) \\
& + \cos(\theta_i)\cos(\theta_j)\Big)
\displaystyle - DS^2 \sum_i \cos^2(\theta_i),
\end{array}
\end{equation}
where the various spin angles are illustrated in Fig. \ref{GHandQRDiagram}\cite{hara:09:JPCM}. Within the different spin configurations, we will specialize angles for the purpose of achieving analytical and understandable solutions.
\subsection{Out-of-plane spin configurations}
\begin{figure*}[!]
\includegraphics[width=7.05in]{Fig9-ScaleneSpinwavesFM-small.eps}
\caption{The FM spin waves configuration with scalene exchange interaction proportions. The $\alpha$ and $\beta$ values for each are separated by a comma as $\alpha$, $\beta$ in the corresponding triangle diagrams. All graphs are grouped together in a single color group.}
\label{Scalene_Spinwaves_FM}
\end{figure*}
The out-of-plane spin configurations have collinear spins with $\theta$ = 0 or $\pi$. Here, the classical energy can be written as
\begin{equation} \label{E_0_OUT}
\begin{array}{ll}
\displaystyle \frac{E_0}{| J |S^2} = & -\frac{2}{3}\Big(\alpha\cos(\theta_A)\cos(\theta_B) + \beta\cos(\theta_A)\cos(\theta_C)
\\
&\displaystyle +\mathcal{J} \cos(\theta_B)\cos(\theta_C)\Big) - \mathcal{D},
\end{array}
\end{equation}
where $\mathcal{J} \equiv J/| J | = \pm1$ and determines the general exchange interaction of the system. Additionally, $\mathcal{D} \equiv D/| J |$, $\alpha$ $\equiv$ $\frac{J'}{|J|}$ and $\beta$ $\equiv$ $\frac{J''}{|J|}$. By pulling out an overall $J$ and looking at the ratio of exchange interactions, we reduce the number of variables and produce an overall scaling factor that helps determine the energy scale of a material system.
For ease of calculation, angles $\theta_A$, $\theta_B$, and $\theta_C$ are replaced with the values 0, 0 + $q$, and 0 + $r$, respectively, where $q$ and $r$ are deviation angles from the FM state, as illustrated in Fig. \ref{GHandQRDiagram}(a). The energy becomes
\begin{equation} \label{E_0_OUT_2}
\displaystyle \frac{E_0}{| J |S^2} = -\frac{2}{3}\Big(\alpha\cos(q)+\beta\cos(r) + \mathcal{J}\cos(q)\cos(r)\Big) - \mathcal{D}.
\end{equation}
Assuming no magnetic field nor anisotropy, the energy function was inspected for minima and maxima. The values for $\alpha$ and $\beta$ that minimize $E_0$ out the plane belong to the sinusoids $\alpha = -\mathcal{J}\cos(r)$ and $\beta = -\mathcal{J}\cos(q)$. It is important to remember that this relationship is only accurate for $\theta$ values that are multiples of $\pi$, as the consideration of any other value reinstates the necessity for the first term of Eq. \ref{E_0}, leading these relationships to oscillate discretely between $-1$ and 1 as $(-1)^{n+1}\mathcal{J}$ where $n$ is $q/\pi$ for $\alpha$ and $r/\pi$ for $\beta$.
The case where $\beta$ and $\alpha$ are equal to each other, but not necessarily to $J$, is termed the ``isosceles" case. While $r$ and $q$ are not mathematically required to be equivalent when $\alpha$ and $\beta$ are equal, it is likely that the angles will be equal through symmetry. The minimizing formula is then $\alpha = -\mathcal{J}\cos(q)$, which is shown in Fig. \ref{AlphaVsQG}. In order to minimize $E_0$, $J'$ may never exceed $J$ in magnitude. Assuming a positive $J$ value, the maximum value $J' = J$ requires the AfM configuration ($q$ = $\pi$) to minimize energy and the minimum value ($J' = -J$) requires the FM ($q$ = 0) configuration for minimization.
\begin{figure*}[!]
\includegraphics[width=7.0in]{Fig10-ScaleneSpinwavesAFM.eps}
\caption{The spin waves for the AfM configuration with scalene exchange interaction proportions. The $\alpha$ and $\beta$ values for each case are separated by a comma as $\alpha, \beta$ in the corresponding triangle diagrams. A single color group encompasses all 3D graphs shown.}
\label{Scalene_Spinwaves_AfM}
\end{figure*}
The simplest case, where all exchange interactions are equal ($\beta = \alpha = \mathcal{J}$), is termed the ``equilateral" case. In this state, the value for $g$ that minimizes $E_0/ | J |S^2$ is $\pi$, which represents the AfM configuration. $E_0/ | J |S^2$ is maximized by $g = 0$, indicating the FM configuration. These configurations are therefore the focus of our out-of-plane analyses.
The classical energy for the out-of-plane FM configuration ($\theta = 0$) is reduced to
\begin{equation} \label{E_OUT_FM}
\displaystyle \frac{E_{0,FM}}{| J |S^2} = -\frac{2}{3}(\alpha+\beta + \mathcal{J})- \displaystyle \mathcal{D}
\end{equation}
and the AfM energy to
\begin{equation} \label{E_OUT_AfM}
\displaystyle \frac{E_{0,AfM}}{| J |S^2} = -\frac{2}{3}(-\alpha-\beta + \mathcal{J})-\mathcal{D}.
\end{equation}
These energies were used to generate the phase diagram in Fig. \ref{Phase_Diagrams}(a).
\subsection{In-plane spin configurations}
For in-plane configurations ($\theta = \pi/2$), the classical energy becomes
\begin{equation} \label{E_0_IN}
\begin{array}{ll}
\displaystyle \frac{E_0}{| J |S^2} = & \displaystyle -\frac{2}{3}\Big[\alpha \cos(\phi_A-\phi_B) + \beta \cos(\phi_A-\phi_C)\\
& + \mathcal{J}\cos(\phi_B - \phi_C)\Big].
\end{array}
\end{equation}
To reduce the number of variables, the energy is rearranged in terms of deviation from the $120^{\circ}$ phase such that $\phi_A$, $\phi_B$, and $\phi_C$ become $(1/2)\pi$, $(7/6)\pi + g$, and $(11/6)\pi – h$, respectively, as shown in Fig. \ref{GHandQRDiagram}(b). Assuming all variables are real valued, the energy becomes
\begin{equation}\label{E_0_IN_Isosceles}
\begin{array}{ll}
\displaystyle \frac{E_0}{| J |S^2} = & \displaystyle
\frac{2}{3}\Big[\textstyle\alpha\sin\Big(\frac{\pi}{6} + g\Big) + \beta\sin\Big(\frac{\pi}{6} + h\Big)\\
& + \cos\Big(\textstyle\frac{\pi}{3} + g + h\Big)\Big].
\end{array}
\end{equation}
It is important to note that since we are using the isotropic Heisenberg model, all spins in this configuration can be rotated in the plane by any phase factor of $\phi^{\prime}$ as it is energetically degenerate. We choose this particular orientation to simplify the expressions. Using this system, the relationships between spin angle and exchange interaction strength which minimize the classical energy are determined analytically as:
\begin{equation}\label{alpha_beta_min_E_IN}
\displaystyle \alpha =
\mathcal{J}\frac{\sin\Big(\frac{\pi}{3}+g+h\Big)}{\cos\Big(\frac{\pi}{6}+g\Big)} ~{\rm and}~
\displaystyle \beta =
\mathcal{J}\frac{\sin\Big(\frac{\pi}{3}+g+h\Big)}{\cos\Big(\frac{\pi}{6}+h\Big)} .
\end{equation}
Unlike the out-of-plane case, these relationships hold true for any value of g and h.
\begin{figure*}[!]
\includegraphics[width=\textwidth]{Fig11-IsoscelesSpinWaves120.eps}
\caption{Isosceles distortions of spin waves for the 120$^\circ$ configuration. The first row shows the spin-wave energy values through the path in $k$-space that passes through symmetry points $\Gamma$, M, K, then back to $\Gamma$. The $\alpha$ value for each row of 3D graphs is given by the diagram in the left-hand column. Note that there are no stable spin waves for this configuration without anisotropy, aside from the equilateral ($\alpha = -1$) mode. This being the case, multiple magnitudes of anisotropy were included to better illustrate the stable possibilities for the 120$^{\circ}$ phase. Because no positive-$\alpha$ state could be stabilized with anisotropy lesser in magnitude than $3|J|$, only negative values were included here. A single color group encompasses all 3D graphs in this figure.}
\label{Isosceles_Spinwaves_120}
\end{figure*}
In the isosceles case, the minimizing relationship becomes
\begin{equation} \label{alpha_min_E_IN_Isosceles}
\displaystyle \alpha =
\mathcal{J}\frac{\sin\Big(\frac{\pi}{3}+2g\Big)}{\cos\Big(\frac{\pi}{6}+g\Big)}.
\end{equation}
Equation (\ref{alpha_min_E_IN_Isosceles}), illustrated in Fig. \ref{AlphaVsQG}, shows that to minimize $E_0$, the absolute value of exchange interaction $J'$ may never exceed twice the value of $J$. It may also be observed that when $J$ is positive, the maximum value $J' = J$ requires the AfM configuration ($g$ = $\frac{\pi}{3}$) to minimize classical energy and the minimum value $J'$ = $-J$ requires the FM configuration ($g$ = $-\frac{2\pi}{3}$).
Examination of the equilateral case showed that the minima and maxima of $E_0/ | J |S^2$, regardless of the sign of $J$, are produced by the 120$^{\circ}$ and FM configurations, respectively. The 120$^{\circ}$ phase has the lowest energy of all equilateral structures in and out of the plane and is determined as the ground state.
The AfM configuration features as a local minimum for the classical energy, yet proved too unstable for spin-wave examination in the equilateral case. It is analyzed, however, with further distortions.
The FM, AfM, and 120$^\circ$ phases will therefore be the initial subjects for in-plane analyses. In the plane, the ferromagnetic energy is simplified to
\begin{equation} \label{E_0_IN_FM}
\displaystyle \frac{E_{0,FM}}{| J |S^2} = -\frac{2}{3}(\alpha + \beta + \mathcal{J}),
\end{equation}
the 120$^{\circ}$ configuration energy to
\begin{equation} \label{E_0_IN_120}
\displaystyle \frac{E_{0,120}}{| J |S^2} = \frac{1}{3}(\alpha + \beta + \mathcal{J}),
\end{equation}
and the AfM energy to
\begin{equation} \label{E_0_AfM}
\displaystyle \frac{E_{0,AfM}}{| J |S^2} = \frac{1}{3}(-\alpha - \beta + \mathcal{J}).
\end{equation}
Overall, the classical energies allow for the general understanding of where these few configurations are stable with respect to each other. It is important to note that other magnetic structures, especially canted non-collinear or $>$3 SL magnetic configurations could also exist. However, there are too many to meaningfully characterize all possibilities here. Therefore, it is essential to look at the spin dynamics to gain insight into the spin configurations' stability. If a system is a stable ground state according to the classical considerations but unstable from the standpoint of the spin dynamics, then this is an indication of the presence of a canted non-collinear state. Therefore, the next step is to evaluate these configurations for dynamic stability.
\section{Spin-wave Dynamics}
Exploring the spin-wave dynamics for the five most interesting spin configurations determined by the classical energy, we first generate solutions to the simplest case where all exchange interactions are equal, no magnetic field is applied, the spin angles are held constant, the physical distance between each atom remains static, and the only variation is anisotropy. Beyond this, we produce spin-wave solutions for the sublattice where exchange interactions are no longer equal, first exploring only one distortion (the isosceles case), then examining the case where no two interactions are equal, termed the ``scalene" case. Considering purely in-plane spin configurations, the latter two cases are further probed by varying the spin angles. The varied spin angles considered belong to the classical energy minimizing relationships discussed in the previous section.
There are a few formatting rules applied throughout the figures in this next section. For every 3-D spin-wave graph, a color scale is applied which spans all colors from purple to red, where purple reflects the minimum $\omega/|J|S$ value 0 and red indicates the maximum $\omega/|J|S$ value between all graphs in the group of graphs to which that color scale is applied. For example, in Fig. \ref{Equilateral_Spinwaves}, the $120^\circ$ graphs (k) and (l) are grouped, and the color scale for both spans from purple at 0 to red at 4.5, where 4.5 is the highest $\omega/|J|S$ value between both graphs. The color grouping used between graphs is specified in each figure caption.
To contextualize the spin waves' form with respect to the reciprocal lattice, all 3D spin-wave graphs include a projection of the equilateral FM configuration's central energy level (pictured in Fig. \ref{Equilateral_Spinwaves}(a)) onto the $k$-plane. The projection is a heat map whose colors are scaled relative to that single energy level's minima and maxima, as in Fig. \ref{brillouin}(b).
The 2D spin-wave graphs in the isosceles figures include only three $\alpha$ values, as opposed to the six considered in the 3D graphs, because the inclusion of all six $\alpha$ values created massive overlap in the lines defining the energies, rendering the graphs unreadable. Any destabilizing $\alpha$ values were thus excluded from these representations to offer greater clarity.
All spin-wave figures after the equilateral figure include small triangular diagrams. These are included to visualize the proportions of the exchange interaction strengths between atoms. Red lines indicate negative valued interactions and black positive. The atoms are colored to reflect the spin configurations as in Fig. \ref{structure}, save those in the isosceles graphs: as the FM and AfM configurations share triangle diagrams, the diagrams were colored to reflect the AfM mode.
\begin{figure*}[!]
\includegraphics[width=\textwidth]{Fig12-IsoscelesSpinWavesVariableSpin.eps}
\caption{Isosceles distortions of in-plane spin waves according to the relationship defined in Eq. \ref{alpha_min_E_IN_Isosceles}. Aside from the two graphs whose $\alpha$ value approaches 2, all spin waves pictured have anisotropy $\mathcal{D}$ = 1. Only negative $\alpha$ values are depicted because when both the $\alpha$ and $J$ signs are reversed for any given combination, they produce the same spin waves, with the exception of the addition of anisotropy. Additionally, the spin rotating angles are provided in Table \ref{isosceles_table}, and two groups are defined for clarity in color scaling, where those without anisotropy are grouped together and those with are grouped separately.}
\label{Emin_Isosceles_Spinwaves}
\end{figure*}
\subsection{Out-of-plane Configurations}
\subsubsection{Ferromagnetic Phase}
Since all spins in the FM phase point in the same direction, it is expected this phase is stable when each of the exchange interactions between these aligned sites ($\mathcal{J}$, $\alpha$, and $\beta$) are positive. We examine the effects of interaction competition on the FM phase to understand when it becomes unstable from a spin-wave standpoint.
Figure \ref{Equilateral_Spinwaves}(b) and (c) show how the equilateral ($\alpha$ = $\beta$ = $\mathcal{J}$ = +1) out-of-plane FM spin waves evolve with anisotropy. The anisotropy present in Fig. \ref{Equilateral_Spinwaves}(c) is equal to the exchange interaction ($\mathcal{D}$ = +1). It is observed that, without anisotropy, the spin waves are identical to the FM configuration pointing in the plane Fig. \ref{Equilateral_Spinwaves}(h). As expected, easy-axis anisotropy present in the out-of-plane configuration simply adds uniformly across the $k$-plane to all energy states. The spin-wave energies $\omega_i$ are represented analytically as
\begin{widetext}
\begin{equation}
\begin{array}{ll}
\omega_0/|J|S = & 2\mathcal{D}+6\mathcal{J} \\
\omega_{\pm}/|J|S = & 2\mathcal{D}+3\mathcal{J}\pm\sqrt{8\mathcal{J}^2\cos(k_x)^2\cos(\frac{k_x}{2}-\frac{\sqrt{3}k_y}{2})^2+8\mathcal{J}^2\cos(\frac{k_x}{2}\pm \frac{\sqrt{3}k_y}{2})\cos(k_x)\sin(k_x)\sin(\frac{k_x}{2}-\frac{\sqrt{3}k_y}{2})+\mathcal{J}^2}. \\
\end{array}
\end{equation}
\end{widetext}
Moving to the first distortion of exchange interactions, the spin-wave dynamics of various isosceles states ($\alpha$ = $\beta$) are explored for the FM configuration in Fig. \ref{Isosceles_Spinwaves_Out}, which illustrates the transformation of the spin waves as the value $\alpha$ changes. With an $\alpha$ value of 2, the interaction between atoms B and C has half the magnitude of the other two. This distortion is reflected in the dispersion shown in the $k_y$ direction when inspecting the spin-wave diagram. A similar, perpendicular phenomenon can be seen in the $\alpha = \frac{1}{2}$ case. The inclusion of anisotropy energy, predictably, has a similar additive effect to that observed in the equilateral spin waves. As the spins are uni-directional in a ferromagnet, it is expected that the exchange interactions are all positive. This instability is confirmed visually by inspecting the negative $\alpha$-valued graphs, except for the $\alpha$ = -2 spin waves, which seem to become stable due to a symmetry effect in the interactions. Larger negative $\alpha$ values produce equally stable-looking waves. It becomes apparent, however, that this state is not truly stable when considering its behavior with anisotropy. Notice that the graph in the second column, which reflects $\mathcal{D} = +1$, has lower overall energy than its no-anisotropy counterpart, which is indicative of an unstable system. This is in agreement with the classical energy, which indicates that this state is only metastable. A similar effect in the AfM configuration is realized in the next analysis. While calculable, the analytical representation of the isosceles spin waves is too large to be presented here and cannot be used to provide insight.
Figure \ref{Scalene_Spinwaves_FM} shows the scalene spin waves with all permutations of a 1-2-3 proportionality between $\alpha$, $\beta$, and $J$, given a static FM spin configuration. It is observed that, as in the isosceles case, both the in-plane and out-of-plane FM configurations produce identical spin waves throughout the variation of exchange interaction, so long as anisotropy is not considered. However, their changes with respect to anisotropy differ. For simplicity, anisotropy was excluded from the scalene discussion. These visualizations illustrate the dispersion that occurs along the direction of the most robust exchange interaction.
\subsubsection{Antiferrimagnetic Phase}
In this configuration, sites B and C in Fig. \ref{brillouin}(a) are aligned and site A points in the opposite direction. It is thus expected that this configuration will be stabilized by a positive exchange interaction $J$ and negative interactions $J'$ and $J''$. However, with frustration in this system, the boundaries of stability are of interest, as with the FM phase.
The out-of-plane, equilateral AfM energies are unstable without anisotropy. The anisotropies $\mathcal{D}$ present in Fig. \ref{Equilateral_Spinwaves} (e) and (f) are valued at +1 and +2, respectively. As discussed prior, it is expected that the AfM configuration contains one positive exchange interaction (between the two same-spin atoms) and two negative interactions. However, as all exchange interactions are strictly equal in the equilateral case, all exchange interactions were assigned a negative value to reflect the net negativity of the three interactions together. Therefore, this state has $\mathcal{J} = \alpha = \beta = -1$. The analytical representation of these energies is too large for inclusion.
Figure \ref{Isosceles_Spinwaves_Out} includes the isosceles ($\alpha$ = $\beta$) out-of-plane AfM spin waves. The expected stabilizing state is $\alpha$ = -1, while $J$ is positive, as the exchange interactions $\alpha$ and $\beta$ are equal and belong to the opposite-spin atom pairs. The spin waves at this $\alpha$ value reflect this expectation in their visually apparent stability. Interestingly, it is shown that various $\alpha$ values that stabilize this otherwise unstable configuration. Similar to the FM phase, the AfM system has an unexpected metastable state where all three exchange interactions are positive with an $\alpha$ value of +2. Again, we see an overall lowering of energy when anisotropy is included, revealing the volatile nature of this arrangement.
The scalene spin waves with all permutations of a 1-2-3 proportionality between $\alpha$, $\beta$, and $J$, for an AfM spin configuration, are shown in Fig. \ref{Scalene_Spinwaves_AfM}. As in the isosceles case, the in-plane and out-of-plane configurations, with no anisotropy, produce identical spin waves throughout the exchange interaction variation. For the sake of simplicity, anisotropy-inclusive spin waves are excluded from the narrative. As with the isosceles case, dispersion in the spin waves occurs along the direction of the strongest exchange interaction.
\begin{table}
\begin{tabular}{|p{1cm}|p{1cm}|p{1cm}||p{1cm}|p{1cm}|p{1cm}|}
\hline
\multicolumn{6}{|c|}{Isosceles spins} \\
\hline
$\mathcal{J}$&$\alpha$&$\beta$&$\phi_A (^{\circ}$)&$\phi_B(^{\circ}$)&$\phi_C(^{\circ}$)\\
\hline
-1& -2 & -2 & 90 & 270.0 & 270.0\\
-1& -3/2 & -3/2 & 90 & 228.6 & 311.4\\
-1& -1 & -1 & 90 & 210.0 & 330.0\\
-1& -3/4 & -3/4 & 90 & 202.0 & 338.0\\
-1& -1/2 & -1/2 & 90 & 194.5 & 345.5\\
1& -2 & -2 & 90 & 90.00 & 90.00\\
1& -3/2 & -3/2 & 90 & 131.4 & 48.60\\
1& -1 & -1 & 90 & 150.0 & 30.00\\
1& -3/4 & -3/4 & 90 & 158.0 & 22.00\\
1& -1/2 & -1/2 & 90 & 165.5 & 14.50\\
\hline
\end{tabular}
\caption{\label{tab:table-name}The spin angles calculated for $\alpha$ values by the relationship in Eq. \ref{alpha_min_E_IN_Isosceles}}
\label{isosceles_table}
\end{table}
\subsection{In-plane Configurations}
\subsubsection{Ferromagnetic Phase}
Figure \ref{Equilateral_Spinwaves} (h) and (i) show the evolution of the equilateral in-plane FM spin waves with the presence of anisotropy. Easy-plane anisotropy added to the in-plane configuration stretches the spin waves as interactions strain the easy orientation. For this configuration, the spin-wave energies can be represented analytically and shown to be
\begin{widetext}
\begin{equation}
\begin{array}{ll}
\omega_0/|J|S = & 2\sqrt{9\mathcal{J}^2-3\mathcal{D}\mathcal{J}} \\
\omega_{\pm}/|J|S = & \biggl(8\mathcal{J}^2\sin(k_x)\cos(\frac{k_x}{2}-\frac{\sqrt{3}k_y}{2})\sin(\frac{k_x}{2}-\frac{\sqrt{3}k_y}{2})\cos(k_x)+8\mathcal{J}^2(\cos(\frac{k_x}{2}-\frac{\sqrt{3}k_y}{2}))^2(\cos(k_x))^2-6\mathcal{D}\mathcal{J}+10\mathcal{J}^2 \\
&\pm 2\mathcal{J}\sqrt{8(\mathcal{D}-3\mathcal{J})^2(\sin(k_x)\cos(\frac{k_x}{2}-\frac{\sqrt{3}k_y}{2})\sin(\frac{k_x}{2}-\frac{\sqrt{3}k_y}{2})\cos(k_x)+\cos(\frac{k_x}{2}-\frac{\sqrt{3}k_y}{2})^2(\cos(k_x))^2+1)}\biggr)^\frac{1}{2}. \\
\end{array}
\end{equation}
\end{widetext}
Various isosceles FM spin waves are illustrated in Fig. \ref{Isosceles_Spinwaves_In}. It is shown that, as in the equilateral case, the spin waves are identical to the out-of-plane configuration where anisotropy is not considered. Otherwise, the inclusion of anisotropy energy has the same stretching effect as that in the equilateral spin waves. Like the out-of-plane FM configuration, the analytical representation of the energies that describe an isosceles in-plane FM state were too large to include here.
As no anisotropy was considered for the scalene cases, Fig. \ref{Scalene_Spinwaves_FM} represents both in- and out-of-plane configurations and is not discussed redundantly.
\subsubsection{Antiferrimagnetic Phase}
Even with the presence of anisotropy five times the magnitude of the exchange interaction, the equilateral spin waves for the in-plane AfM configuration were unstable and therefore excluded from figure \ref{Equilateral_Spinwaves}.
Figure \ref{Isosceles_Spinwaves_In} illustrates isosceles in-plane AfM spin waves. As in the out-of-plane case, there are various $\alpha$ values which stabilize this unstable configuration. Without anisotropy, these spin waves are identical to the out-of-plane spin waves. Yet, when present, a stretching of energy is observed as a result of anisotropy.
Scalene spinwaves for the in-plane AfM configuration with $\mathcal{D} = 0$ are shown in figure \ref{Scalene_Spinwaves_AfM} and are identical to those of their out-of-pane counterparts.
\begin{figure*}[!]
\includegraphics[width=\textwidth]{Fig13-ScaleneSpinWavesVariableSpin-small.eps}
\caption{Scalene distortions of in-plane spin waves according to the relationships in Eq. \ref{alpha_beta_min_E_IN}. All spin waves shown have anisotropy $\mathcal{D}$ = 1. The $\alpha$ and $\beta$ values for each graph are presented within the triangle diagrams, respectively. Additionally, the spin rotating angles are provided in Table \ref{scalene_table}. A single color group encompasses all graphs.}
\label{Emin_Scalene_Spinwaves}
\end{figure*}
\subsubsection{120$^{\circ}$ Phase}
The ground state of the Kagome lattice behaves peculiarly in contrast to the behaviors of the FM and AfM configurations. The no-anisotropy spin waves in Fig. \ref{Equilateral_Spinwaves} (k) show a degeneracy of energy levels for the 120$^{\circ}$ phase. This degeneracy is lost, however, as anisotropy increases. As anisotropy energy increases to outweigh the exchange interactions, Dirac nodes form in the system. While lower values of anisotropy maintained stable spin waves, the relatively large amount of $\mathcal{D}$ = -3 is included for the purpose of illustrating this phenomenon. Lesser anisotropies are included in the 2-D spin waves graph (j) to convey its interesting path to this point. Here, the analytical spin-wave energies can be shown as
\begin{widetext}
\begin{equation}
\begin{array}{ll}
\omega_0/|J|S = & \sqrt{6\mathcal{J}\mathcal{D}} \\
\omega_{\pm}/|J|S = & \biggl(-4\mathcal{J}^2\cos(\frac{k_x}{2}-\frac{\sqrt{3}k_y}{2})^2\cos(k_x)^2-4\mathcal{J}^2\cos(\frac{k_x}{2}-\frac{\sqrt{3}k_y}{2})\sin(\frac{k_x}{2}-\frac{\sqrt{3}k_y}{2})\cos(k_x)\sin(k_x)+3\mathcal{D}\mathcal{J}+4\mathcal{J}^2 \\
& \pm \mathcal{J}\sqrt{(8\cos(\frac{k_x}{2}-\frac{\sqrt{3}k_y}{2})^2\cos(k_x)^2+8\cos(\frac{k_x}{2}-\frac{\sqrt{3}k_y}{2})\sin(\frac{k_x}{2}-\frac{\sqrt{3}k_y}{2})\cos(k_x)\sin(k_x)+1)\mathcal{D}^2}\biggr)^\frac{1}{2} . \\
\end{array}
\end{equation}
\end{widetext}
The distortion of exchange energies did not provide any additional stability to the 120$^{\circ}$ configuration. In fact, any distortion destabilized the system and required anisotropy on the order of the exchange interaction for stability. Even anisotropy, however, could not stabilize any state whose $\alpha$ value had a sign opposite to that of $J$. This instability is predicted by the state being a particular case of energy minimization related to the $\alpha$ and $\beta$ values.
Because isosceles distorting exchange interactions for the 120$^{\circ}$ phase offered no stability, scalene spin wave analyses for this configuration were omitted entirely.
\begin{table}
\begin{tabular}{|p{1cm}|p{1cm}|p{1cm}||p{1cm}|p{1cm}|p{1cm}|}
\hline
\multicolumn{6}{|c|}{Scalene spins} \\
\hline
$\mathcal{J}$&$\alpha$&$\beta$&$\phi_A (^{\circ}$)&$\phi_B(^{\circ}$)&$\phi_C(^{\circ}$)\\
\hline
-1& -5/4 & -3/2 & 90 & 217.1 & 311.6\\
-1& -3/4 & -5/4 & 90 & 175.1 & 306.7\\
-1& -1/2 & -3/4 & 90 & 152.7 & 306.3\\
-1& -3/4 & -1/2 & 90 & 233.7 & 27.27\\
-1& -5/4 & -3/4 & 90 & 233.3 & 4.940\\
-1& -3/2 & -5/4 & 90 & 228.4 & 322.9\\
\hline
\end{tabular}
\caption{\label{tab:table-name}The spin angles calculated for various combinations of $\alpha$ and $\beta$ determined by the relationships in Eq. \ref{alpha_beta_min_E_IN}}
\label{scalene_table}
\end{table}
\subsection{Spin Angle Distortions}
The 120$^\circ$ phase showed to be quite a unique configuration. While it is the ground state when all exchange interactions are equal, any distortion of the exchange parameters destabilize the phase, unlike the other configurations which could maintain stability through some distortion. In light of this, we looked for stability in new states whose spin angles and exchange interactions together minimize the classical energy.
Before considering all distorting parameters, we investigate the isosceles ($\alpha$ = $\beta$) states under the assumption of the simplest case, where rotation angles $g$ and $h$ are also equivalent, as described in Eq. \ref{alpha_min_E_IN_Isosceles}. It can be inferred from this minimizing relationship, illustrated in Fig. \ref{AlphaVsQG}, that there are two angle $g$ solutions for any given $\alpha$. However, these two values produce equivalent spin waves as the two $g$ values produce the same two values for spin angles for atoms B and C, but the angles assigned to each site are reversed. Spinwaves for some interesting $\alpha$ values are illustrated in Fig. \ref{Emin_Isosceles_Spinwaves}. As nearly every result was unstable, an anisotropy of $\mathcal{D} = -1$ was applied to all states except those two which have an $\alpha$ value approaching 2. These do not include anisotropy.
Examining further distortion, figure \ref{Emin_Scalene_Spinwaves} shows scalene spin waves with various values of $\alpha$ and $\beta$, given a negative $J$, and with spin angles defined by the classical energy minimizing relationships in Eq. \ref{alpha_beta_min_E_IN}. As the scalene spin waves were also unstable without anisotropy, the solutions pictured reflect a value $\mathcal{D} = -1$. These visualizations illustrate the dispersion that occurs along the direction of the strongest exchange interaction. The fact that these polar spin angle solutions which minimize the classical energy rely on anisotropy for stability may indicate that the ground state configurations for these distorted exchange interaction states are canted out of the lattice plane to some degree.
Overall, these results indicate that purely in-plane magnetic configurations which deviate from the FM, AfM, and 120$^\circ$ states require anisotropy to stabilize within the Kagome lattice. We expect that there will be a reduction in the required anisotropy as the spin angles are allowed to cant in any out-of-plane direction (not necessarily along the $z$-axis) and form distinct non-collinear phases. Further study into the realm of these out-of-plane canted phases needs to be done. However, the results here are important for understanding the evolution of the magnetic Kagome lattice's spin waves with variable exchange interactions.
\section{Discussion and Conclusion}
The quest for an understanding of quantum spin states, especially the quantum spin liquid, has led to a wealth of experimental realizations and studies on the structural, magnetic, and thermodynamics of Kagome systems\cite{han:12:nature,helt:07:PRL,hirs:15:PRL,dun:16:PRL,kass:17:PRB,zork:19:PRB} as well as distorted Kagome systems\cite{wei:10:PRB,wulf:12:JPCM,mata:19:PRB,mata:10:NatPhys}. While many studies focus on the interpretation of either one material or even one magnetic configuration, the ability to discern the various magnetic interactions with and without distortions has been a challenge that leads most to using numerical approaches in modeling experimental data. However, examining these systems within an analytical limit for the symmetric systems and evolving out into the distorted systems allows for one to gain a deeper understanding of the effects of the interactions and how they distort the magnetic systems.
In this paper, we examine the effects of variable first-order magnetic distortions on the spin-wave dynamics of the Kagome lattice. Using an isotropic spin-spin exchange Hamiltonian with $z$-axis anisotropy, we determine the phase diagrams for various out-of-plane and in-plane spin configurations and then examine how the spin waves are affected by exchange interaction distortions.
By analyzing the spin waves with static spin angles and varying exchange interactions, we gain novel insights on the effect this distortion alone has on the magnetic identity of this sublattice. A dispersive effect on the spin waves in accordance with exchange interaction proportionality is illustrated in the isosceles and scalene cases for all five of the most straightforward configurations, and some unexpectedly metastable states fell out of unlikely exchange interaction states for the FM and AfM cases.
Although it is the ground state classically, the special-case nature of the 120$^\circ$ phase is underlined by its inability to retain stability as its exchange interactions are distorted. This led to our employing the energy-minimizing relationships for spin angles in seeking stable spin waves for these changing $\alpha$ values, which pointed to even further spin distortions for stability. The next natural step in this vein would be to numerically minimize the classical energy with full freedom of spin angles, allowing for continuity in both the polar and azimuthal angles. Limiting to purely in-plane or out-of-plane angles offered analytical insights that are invaluable in characterizing magnetic relationships. However, to more deeply describe the most natural behaviors of atoms in this lattice, the allowance of canted configurations is a necessary consideration.
Overall, this paper aims to provide insight into the spin dynamics of the Kagome lattice to help in the characterization of its non-collinear phases. Therefore, we show how the distortion of the exchange parameters and spin angles affect the dynamics, which provides useful information for the characterization of material systems, especially when investigating these phases using inelastic neutron scattering.
\section*{Acknowledgements}
A.A.C. and J.T.H acknowledge support from the Institute for Materials Science at Los Alamos National Laboratory. The work at Los Alamos National Laboratory was carried out under the auspices of the U.S. DOE and NNSA under Contract No. DEAC52-06NA25396 and supported by U.S. DOE (A.S.).
|
\section{Introduction}\label{intro
In the accreting systems, large scale jets are usually steady, while
episodic jets are sometimes related to flares, which are observed on the
smaller scale. One such example is Sgr~A*, a massive black hole in the Galactic
centre, where we observe radio, infrared and X-ray flares several times a
day. It was concluded that delays in peaks in the light curves at different
wavebands and their fast rise and slow decay in the brightness and
polarisation are related to the ejection and expansion of plasmoids from the
accretion flow. Knots in the jets are also observed, e.g. in 3C 120 and
M87, and could be related to episodic emission. There are models, like e.g.
\cite{BZ77} and \cite{BP82} for continuous jets,
but we still do not have a viable model for episodic jets. In \cite{YuanFengetal09},
such a model was proposed, in analogy with Coronal Mass Ejections
(CMEs) in the Sun, with the closed magnetic field lines emerging from the
main body of the accretion flow, expelled to the corona region: The
foot-points of the magnetic loops are positioned in the turbulent accretion
flow, and their twisting results in magnetic reconnection, forming the flux
ropes. Because of the ongoing reconnection below such a flux rope, the
magnetic tension force weakens, and the initial equilibrium between the
magnetic tension and the magnetic pressure is not maintained. The flux ropes
will be accelerated outwards, forming the episodic jet. The flares,
observed from such jets, are from the emission originating from the
electrons accelerated by the reconnection. In \cite{Shende19}, another model was
proposed, in analogy with Toroidal Instability from tokamak research and also used to model the CMEs.
\citet{Nathanail20} present results of two-dimensional (2D) GRMHD simulations with the Black Hole Accretion Code (BHAC, \citet{Porthetal19}), with Adaptive Mesh Refinement (AMR) of both Magnetically Arrested and Standard and Normal Evolution (MAD and SANE) discs. Different initial magnetic field configurations and resolutions are chosen. They find the formation of copious plasmoids and describe their outward motion. Similar simulations based on the same code, but with the physical resistivity included, are presented in \citet{Ripperda20}. They show no difference in results between the ideal and weakly resistive simulations. They conclude that 2D ideal MHD simulations, with only the numerical resistivity dissipating the magnetic field, can capture the physics.
In this work, we perform 3D GRMHD simulations to investigate the formation of magnetic flux ropes, checking the scenario proposed by \cite{YuanFengetal09}.
\section{Numerical simulations setup}\label{numsetup
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.35\linewidth,height=.35\linewidth]{grid1}
\includegraphics[width=0.35\linewidth,height=.35\linewidth]{grid2}
\end{center}
\caption{We use static mesh refinement for the grid, to obtain largest resolution where it is most needed. Resolution is $R\times\theta\times\varphi=(288\times 128\times 64)$ grid cells in spherical coordinates, in a physical domain reaching to 1200 gravitational radii. The different refinements used in this grid are shown.}
\label{grid}
\end{figure}
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.31\linewidth,height=.5\linewidth]{mojinit1b}
\includegraphics[width=0.31\linewidth,height=.4\linewidth]{rez18500dens}
\includegraphics[width=0.31\linewidth,height=.4\linewidth]{rez18500betaplasm}
\end{center}
\caption{Left panel: The initial setup in our SANE simulation. The color denotes the density, solid lines denote the poloidal magnetic field with arrows showing the field direction. Middle panel: A zoomed-in snapshot in the result after $t=18500~r_{\mathrm g}/c$. The positions of two magnetic islands are marked with the cross of two dotted black lines and two dotted red
lines, respectively. Right panel: Same with the middle panel, but the color shows the plasma $\beta=P_{\rm gas}/P_{\rm mag}$. The two magnetic islands are located at the surface with plasma $\beta\sim 1$.}
\label{init}
\end{figure}
We perform numerical simulations using the GRMHD code Athena++ \citep{Whiteetal16} in full 3D, solving the ideal MHD equations in the Kerr metrics, in Kerr-Schild (horizon penetrating) coordinates. Resolution is $R\times\theta\times\varphi=(288\times 128\times 64)$ grid cells in spherical coordinates, in a physical domain reaching to 1200 gravitational radii, $r_g=GM/c^2$. We use different refinements in this grid, as shown in Fig.~\ref{grid}. The staggered mesh Constrained Transport (CT) method is applied to maintain the divergence-free magnetic field. Static mesh refinement is used for the grid, to obtain largest resolution where it is most needed. Initial configuration of density and magnetic field in our SANE simulation is shown in Fig.~\ref{init}. The central object is not rotating.
\section{Formation and motion of the flux rope}\label{fluxrope}
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.4\linewidth,height=.37\linewidth]{spiralkaa}
\includegraphics[width=0.4\linewidth,height=.37\linewidth]{spirallka}
\end{center}
\caption{Left panel: Outward motion of the magnetic island in our simulation, in different colatitudinal planes. Right panel: Spiralling-out of the magnetic island in a schematic plot of the trajectory in 3D. Positions of the blue circles are chosen to approximately represent magnetic islands from the left panel.}
\label{spir}
\end{figure}
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.32\linewidth,height=.25\linewidth]{positionspeedplasm}
\includegraphics[width=0.32\linewidth,height=.25\linewidth]{vplasmint}
\includegraphics[width=0.32\linewidth,height=.26\linewidth]{forc}
\end{center}
\caption{Time evolution of the position of the center of magnetic island (left panel) and velocity of the material near the center of magnetic island (middle panel) for two magnetic islands from Fig.~\ref{init}. Slopes of the least square fits shown in dashed lines in the left panel are 0.03~c and 0.01~c for the black and red lines, respectively. In the middle panel, least square fits of the radial velocity components are shown by the corresponding color thin solid lines. In the right panel are shown forces, along the radial direction, on the material in the flux rope above the disk mid-plane, near the center of the magnetic island at $t=18500~r_{\mathrm g}/c$.}
\label{posvel}
\end{figure}
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.48\linewidth,height=.7\linewidth]{rezc}
\vspace*{-9cm}\includegraphics[width=0.5\linewidth,height=.35\linewidth]{bsinTrecon}\\
\hspace*{6cm}\vspace*{-1cm}\includegraphics[width=0.5\linewidth,height=.35\linewidth]{velsinTforrecon}
\vspace*{5.5cm}
\end{center}
\caption{Two reconnection layers with a magnetic island between them in a snapshot at $t=19500~r_{\mathrm g}/c$ in our simulation are marked with the black and red dotted lines in the left panel. In the right panels are shown the velocity and magnetic field components at the same time, along a part of the black dotted line circle passing through the reconnection layer. A signature of reconnection, change in the direction of poloidal velocity and all three components of magnetic field, is visible in the projections in $\theta$-direction. A similar signature is obtained in the reconnection layer positioned at the intersection of red dotted lines. }
\label{velbs}
\end{figure}
Interchanging directions of the initial magnetic field in the torus prevent the field to grow too large, and magneto-rotational instability can provide the dissipation for successful accretion of material towards the central object. We perform our simulation until $t=40000~r_{\mathrm g}/c$. A snapshot at $t=18500~r_{\mathrm g}/c$ during the evolution of the accretion flow is shown in Fig.~\ref{init}. We find magnetic islands in the colatitudinal $(R, \theta)$ planes, at different azimuths $\varphi$, after the relaxation from the initial conditions and stabilization of the flows. Such magnetic islands start forming after about $t=15000~r_{\mathrm g}/c$. They periodically emerge from the disk surface at similar radii (azimuthal angle $\varphi$ changes with the rotation of the disk), with period of about $t=1000~r_{\mathrm g}/c$. The magnetic islands are extended in the azimuthal direction, forming magnetic flux ropes of various lengths. We trace the extension of the flux ropes in $\varphi$ direction, which is typically about $120^\circ$ or less, and perform slices in the middle of their length at different times--as shown in Fig.~\ref{spir}. In the same Figure we give a sketch in 3D of the counter-clockwise spiral trajectory of the indicated flux rope cross-sections. To understand the launching and motion of the flux rope, we measure positions of the magnetic islands and velocity of the material near their centers in time--see Fig.~\ref{posvel}. Forces acting on the material in the magnetic islands in radial direction are also shown. The pressure gradient and Lorentz forces push the rope radially outwards.
The launching of the flux rope is caused by the reconnection in the disk, near the disk surface, as shown in the left panel Fig.~\ref{velbs}. In the right panels in the same Figure is shown the reconnection signature in the magnetic field and velocity components perpendicular to the reconnection layer: all three magnetic field components and both poloidal velocity components change sign. Reconnection occurs throughout the disk and in the corona, but its effect on the matter depends on the value of plasma $\beta=P_{\rm gas}/P_{\rm mag}$. Only in the locations where it is about unity, material from the disk will be pushed by reconnection. In the rarefied corona, plasma $\beta$ is much smaller, and in the dense disk, it is much larger than unity.
Inside the disk, which is accreting because of magneto-rotational instability (MRI) providing the sufficient dissipation, reconnection layers which are brought close to the disk surface, can result in the formation of flux rope and its further ejection into the corona. Once lifted into the rarefied corona, the flux rope can be expelled outwards or break. In both cases it would be observed as episodic emission from the vicinity of the black hole.
In addition to the reconnection layer below the magnetic flux rope, there is another reconnection layer, above the magnetic flux rope in our simulations--see Fig.~\ref{velbs}. It helps the opening of the magnetic field lines and ejection of the flux rope.
\section{Conclusions}\label{conclus
We have performed 3D ideal GRMHD numerical simulations of a hot accretion flow around a black hole, to study formation and motion of flux ropes. During the time-evolution until $t=40000~r_{\mathrm g}/c$, magnetic flux ropes of the azimuthal extension of about $120^\circ$ or less are formed, which show as magnetic islands in 2D slices in colatitudinal planes at different azimuthal angles. These flux ropes are created by reconnection close to the disk surface, where the plasma $\beta$, defined as the ratio of the gas to magnetic pressure, is close to unity.
Because of the reconnection and disk differential rotation, the flux ropes are twisted and pushed radially outwards and launched into the corona, spiralling-out from the central object. The radial velocity of their outward propagation is of the order of 0.01~c.
Ejection of the flux ropes from the disk surface repeats periodically in our simulation, with the period of about $1000~r_{\mathrm g}/c$. It could cause episodic flaring from the vicinity of the disk around a black hole.
In addition to the reconnection layer near the disk surface, which forms the flux rope, another reconnection layer above the flux rope can form, helping its outward launch.
\ac
M\v{C} was supported by CAS President’s International Fellowship for Visiting Scientists (grant No. 2020VMC0002), and the Polish NCN grant 2019/33/B/ST9/01564. FY and HY are supported in part by the National Key Research and Development Program of China (Grant No. 2016YFA0400704), the Natural Science Foundation of China (grants 11633006), and the Key Research Program of Frontier Sciences of CAS (No. QYZDJSSW-SYS008). This work made use of the High Performance Computing Resource in the Core Facility for Advanced Research Computing at Shanghai Astronomical Observatory. We thank the referee for constructive questions and suggestions.
\bibliographystyle{ragtime}
|
\section{Introduction}
\iftoggle{arxiv}{}{}
Multivariate extreme value theory is complicated by the lack of natural ordering in $\mathbb{R}^d$, and the infinite possibilities for the underlying set of dependence structures between random variables. Some of the earliest characterizations of multivariate extremes were inspired by consideration of the vector of normalized componentwise maxima. Let $\bm{X} =(X_1,\ldots,X_d) \in \mathbb{R}^d$ with $X_{j} \sim F_j$, and consider a sample $\bm{X}_i =(X_{1i},\ldots,X_{di})$, $i=1,\ldots,n$, of independent copies of $\bm{X}$. For a fixed $j$, defining $M_{j,n} = \max_{1 \leq i \leq n}(X_{ji})$, the extremal types theorem tells us that if we can find sequences such that $(M_{j,n} -b_{j,n})/a_{j,n}$ converges to a non-degenerate distribution, then this is the generalized extreme value distribution. Moreover, the sequence $b_{j,n} \sim F_j^{-1}(1-c/n)$, $n \to \infty$, i.e., is of the same order as the $1-1/n$ quantile. A natural multivariate extension is then to examine the distribution of the vector of componentwise maxima, $(\bm{M}_n -\bm{b}_n)/\bm{a}_n$. This is intrinsically tied up with the theory of multivariate regular variation, because it leads to examination of the joint behaviour of the random vector when all components are growing at the rate determined by their $1-1/n$ quantile. If all components were marginally standardized to focus only on the dependence, then all normalizations would be the same.
In normalizing all components by the same amount, we only consider the dependence structure in a single ``direction'' in $\mathbb{R}^d$. In some cases this turns out to provide a rich description of the extremal dependence: if the limiting distribution of componentwise maxima does not have independent components, then an infinite variety of dependence structures are possible, indexed by a moment-constrained measure on a $d-1$ dimensional unit sphere. However, when the limiting dependence structure is independence, or even when some pairs are independent, this representation fails to discriminate between qualitatively different underlying dependence structures. While consideration of componentwise maxima is not necessarily a common applied methodology these days, the legacy of this approach persists: statistical methods that assume multivariate regular variation, such as multivariate generalized Pareto distributions, are still very popular in practice \citep[e.g.,][]{EngelkeHitz20}. A recent theoretical treatment of multivariate regular variation is given in \citet{KulikSoulier20}.
Various other representations for multivariate extremes have emerged that analyze the structure of the dependence when some variables are growing at different rates to others. These include the so-called conditional extreme value model \citep{HeffernanTawn04,HeffernanResnick07}, whereby the components of $\bm{X}$ are normalized according to how they grow with a single component, $X_j$ say. Related work examines behaviour in relation to an arbitrary linear functional of ${\bm X}$ \citep{BE07}. The conditional representation allows consideration of those regions where some or all variables grow at a lesser rate than $X_j$ if this is the region where the observations tend to lie. In other words, the limit theory is suited to giving a more detailed description of a broader range of underlying dependence structures. Another representation that explicitly considers different growth rates is that of \citet{WadsworthTawn13}. They focus particularly on characterizing joint survival probabilities under certain classes of inhomogeneous normalization; this was found to reveal additional structure that is not evident when applying a common scaling. More recently, \citet{Simpsonetal18} have examined certain types of unequal scaling with a view to classifying the strength of dependence in any sub group of variables of $\bm{X}$.
An alternative approach to adding detail to the extremal dependence structure focuses not on different scaling orders, but rather on second order effects when applying a common scaling. This idea was introduced by \citet{LedfordTawn96}, and falls under the broader umbrella of hidden regular variation \citep{Resnick02}. Various manuscripts have focused on analogizing concepts from standard multivariate regular variation to the case of hidden regular variation \citep[e.g.,][]{RamosLedford09}, but this approach still only focuses on a restricted region of the multivariate space where all variables are large simultaneously. For this reason, although higher-dimensional analogues exist, they are often not practically useful for dimension $d>2$.
Another manner of examining the extremal behaviour of $\bm{X}$ is to consider normalizing the variables such that they converge onto a limit set \citep[e.g.,][]{Davisetal88,BalkemaNolde10}, described by a so-called gauge function. This requires light-tailed margins, which may occur naturally or through a transformation. If the margins are standardized to a common light-tailed form, then the shape of the limit set is revealing about the extremal dependence structure of the random variables, exposing in which directions we expect to see more observations.
Although various connections have been made in the literature, many of these representations remain somewhat disjointed. For example, there is no obvious connection between the conditional extremes methodology and the representation of \citet{LedfordTawn96,LedfordTawn97}, and whilst \citet{WadsworthTawn13} provided a modest connection to conditional extremes, many open questions remain. In this paper we reveal several hitherto unknown connections that can be made through the shape of the limit set and its corresponding gauge function, when it exists, and provide a step towards unifying the treatment of multivariate extremes.
We next provide further elaboration and definition of the different representations of extremal dependence. For some definitions, it is convenient to have a standardized marginal form; we focus mainly on standard Pareto or standard exponential margins with notation $\bm{X}_P$ and $\bm{X}_E$, respectively. As mentioned above, working with common margins highlights dependence features. In Section~\ref{sec:def} we recall the formulations of various representations for multivariate extremes, and provide a thorough background to the concepts of limit sets and their gauge functions, proving a useful new result on marginalization.
Section~\ref{sec:connections} details connections linking conditional extremes, the representation of \citet{WadsworthTawn13}, \citet{LedfordTawn96}, and that of \citet{Simpsonetal18}. We provide \iftoggle{arxiv}{several}{}illustrative examples in Section~\ref{sec:examples} and conclude in Section~\ref{sec:Discussion}.
\section{Background and definitions}
\label{sec:def}
\subsection{Multivariate regular variation}
\label{sec:MRVintro}
A measurable function $f: \mathbb{R}_+ \to \mathbb{R}_+$ is regularly varying at infinity (respectively, zero) with index $\rho \in\mathbb{R}$ if, for any $x>0$, $f(tx)/f(t) \to x^\rho$, as $t \to \infty$ (or, respectively, $t \to 0$). We write $f \in \mathrm{RV}_\rho^\infty$ or $f \in \mathrm{RV}_\rho^0$, omitting the superscript in generic cases. If $f \in \mathrm{RV}_0$, then it is called slowly varying.
The random vector $\bm{X}$ is multivariate regularly varying on the cone $\mathbb{E} = [0,\infty]^d \setminus \{\bm{0}\}$, with index $\alpha>0$, if for any relatively compact $B \subset \mathbb{E}$,
\begin{align}
t{\mathbb P}(\bm{X}/b(t) \in B) \to \nu(B), \qquad t \to \infty, \label{eq:mrv}
\end{align}
with $\nu(\partial B) = 0$, $b(t) \in \mathrm{RV}^\infty_{1/\alpha}$, and the limit measure $\nu$ homogeneous of order $-\alpha$; see, e.g., \citet{Resnick2007}, Section~6.1.4. The parts of $\mathbb{E}$ where $\nu$ places mass reveal the broad scale extremal dependence structure of $\bm{X}$. Specifically, note that we have the disjoint union $\mathbb{E} = \bigcup_{C} \mathbb{E}_C$, where
\begin{align}
\mathbb{E}_C = \{\bm{x} \in \mathbb{E}: x_j > 0, j\in C; x_i = 0, i\not\in C\}=:(0,\infty]^C\times\{0\}^{D \setminus C}, \label{eq:EC}
\end{align}
and the union is over all possible subsets $C \subseteq D=\{1,\ldots, d\}$, excluding the empty set. If $\nu(\mathbb{E}_C)>0$ then the variables indexed by $C$ can take their most extreme values simultaneously, whilst those indexed by $D \setminus C$ are non-extreme.
The definition of multivariate regular variation in equation~\eqref{eq:mrv} requires tail equivalence of the margins. In practice, it is rare to find variables that have regularly varying tails with common indices, and multivariate regular variation is a dependence assumption placed on standardized variables. Without loss of generality, therefore, we henceforth consider $\bm{X}=\bm{X}_P$ with standard Pareto(1) margins in which case $\alpha=1$ and $b(t) = t$.
Frequently, the set $B$ in \eqref{eq:mrv} is taken as $[\bm{0},\bm{x}]^c = \mathbb{E}\setminus [\bm{0},\bm{x}]$, leading to the \emph{exponent function},
\begin{align}
V(\bm{x}) = \nu([\bm{0},\bm{x}]^c). \label{eq:V}
\end{align}
Suppose that derivatives of $V$ exist almost everywhere; this is the case for popular parametric models, such as the multivariate logistic \citep{Gumbel60}, H\"{u}sler--Reiss \citep{HueslerReiss89} and asymmetric logistic distributions \citep{Tawn90}. Let $\partial^{|C|}/\partial \bm{x}_C = \prod_{i \in C} \partial/ \partial x_i$. If the quantity $\lim_{x_j \to 0, j \not\in C}\partial^{|C|}V(\bm{x})/\partial \bm{x}_C$ is non-zero, then the group of variables indexed by $C$ places mass on $\mathbb{E}_C$ \citep{ColesTawn91}.
Multivariate regular variation is often phrased in terms of a radial-angular decomposition. If~\eqref{eq:mrv} holds, then for $r\geq1$,
\begin{align*}
{\mathbb P}(\bm{X}_P /\|\bm{X}_P\| \in A, \|\bm{X}_P\|>tr) /{\mathbb P}(\|\bm{X}_P\|>t) \to H(A)r^{-1},\qquad t\to\infty,
\end{align*}
where $A \subset \mathcal{S}=\{\bm{w}\in\mathbb{R}^d_+: \|\bm{w}\| = 1\}$ and $\|\cdot\|$ is any norm. That is, the radial variable $R= \|\bm{X}_P\|$ and the angular variable $\bm{W}=\bm{X}_P /\|\bm{X}_P\|$ are independent in the limit, with $R \sim$ Pareto(1) and $\bm{W} \in \mathcal{S}$ following distribution $H$. The support of the so-called spectral measure~$H$ can also be partitioned in a similar manner to $\mathbb{E}$. Letting
\[
\mathbb{A}_C = \{\bm{w} \in \mathcal{S}: w_j > 0, j\in C; w_i = 0, i\not\in C\},
\]
we have $\mathcal{S} = \bigcup_C \mathbb{A}_C$. The measure $\nu$ places mass on $\mathbb{E}_C$ if and only if $H$ places mass on $\mathbb{A}_C$.
\subsection{Hidden regular variation}
Hidden regular variation arises when: (i) there is multivariate regular variation on a cone (say $\mathbb{E}$), but the mass concentrates on a subcone $\tilde{\mathbb{E}} \subset \mathbb{E}$, and (ii) there is multivariate regular variation on the subcone $\mathbb{E}' \subseteq \mathbb{E} \setminus \tilde{\mathbb{E}}$ with a scaling function of smaller order than on the full cone. Suppose that~\eqref{eq:mrv} holds, and $\nu$ concentrates on $\tilde{\mathbb{E}}$, in the sense that $\nu(\mathbb{E} \setminus \tilde{\mathbb{E}}) = 0$. For measurable $B \subset \mathbb{E}'$, we have hidden regular variation on $\mathbb{E}'$ if
\begin{align}
t{\mathbb P}(\bm{X}_P/c(t) \in B) \to \nu'(B), \qquad t \to \infty, \qquad c(t) = o(t),\quad c(t) \in \mathrm{RV}_{\zeta}^\infty, \quad\zeta \in (0,1], \label{eq:hrv}
\end{align}
with $\nu'(\partial B) = 0$ and the limit measure $\nu'$ homogeneous of order $-1/\zeta$ \citep[][Section 9.4.1]{Resnick2007}.
The most common cone to consider is $\mathbb{E}' = (0,\infty]^d$. This leads to the \emph{residual tail dependence coefficient}, $\eta_D \in (0,1]$ \citep{LedfordTawn96}. That is, suppose that~\eqref{eq:hrv} holds on $(0,\infty]^d$, then the regular variation index $\zeta=\eta_D$. The residual tail dependence coefficient for the subset $C \subset D$ is found through considering cones of the form
\[
\mathbb{E}'_C = \{\bm{x} \in \mathbb{E}: x_j > 0, j\in C; x_i \in[0,\infty], i\not\in C\} =:(0,\infty]^C\times[0,\infty]^{D \setminus C},
\]
for which $\zeta= \eta_C$.
\subsection{Different scaling orders}
\label{sec:dso}
\subsubsection{Coefficients $\tau_C(\delta)$}
\citet{Simpsonetal18} sought to examine the extremal dependence structure of a random vector through determination of the cones $\mathbb{E}_C$ for which $\nu(\mathbb{E}_C)>0$. Direct consideration of (hidden) regular variation conditions on these cones is impeded
by the fact that ${\mathbb P}(\bm{X}_P/b(t) \in B) = 0$ for all $B \subset \mathbb{E}_C$, $C \neq D$, since no components of $\bm{X}_P/b(t)$ are equal to zero for $t< \infty$. \citet{Simpsonetal18} circumvent this issue by assuming that if $\nu(\mathbb{E}_C)>0$, then there exists $\delta<1$ such that
\begin{align}
\lim_{t \to \infty} t{\mathbb P}(\min_{i \in C}X_{P,i}> xt, \max_{j \in D \setminus C}X_{P,j} \leq y t^\delta)= \lim_{t \to \infty} t{\mathbb P}(\min_{i \in C}X_{P,i}/t> x, \max_{j \in D \setminus C}X_{P,j}/t \leq y t^{\delta-1})>0, \qquad x,y>0. \label{eq:SWT1}
\end{align}
As such, under normalization by $t$, components of the random vector indexed by $C$ remain positive, whereas those indexed by $D\setminus C$ concentrate at zero. Note that if assumption~\eqref{eq:SWT1} holds for some $\delta<1$ then it also holds for all $\delta' \in [\delta,1]$. \citet{Simpsonetal18} expanded assumption~\eqref{eq:SWT1} to
\begin{align}
{\mathbb P}(\min_{i \in C}X_{P,i}> xt, \max_{j \in D \setminus C}X_{P,j} \leq y t^\delta) \in \mathrm{RV}_{-1/\tau_C(\delta)}^{\infty}, \qquad \delta \in [0,1] \label{eq:SWT2},
\end{align}
where~\eqref{eq:SWT2} is viewed as a function of $t$, and the regular variation coefficients $\tau_C(\delta) \in (0,1]$. For a fixed $\delta$, $\tau_C(\delta)<1$ implies either that $\nu(\mathbb{E}_C)=0$, or that $\nu(\mathbb{E}_C)>0$, but that $\delta$ is too small for~\eqref{eq:SWT1} to hold; see \citet{Simpsonetal18} for further details. Considering the coefficients $\tau_C(\delta)$ over all $C$ and $\delta \in [0,1]$ provides information about the cones on which $\nu$ concentrates.
\subsubsection{Angular dependence function $\lambda(\bm{\omega})$}
\citet{WadsworthTawn13} detailed a representation for the tail of $\bm{X}_P$ where the scaling functions were of different order in each component. They focussed principally on a sequence of univariate regular variation conditions, characterizing
\begin{align}
{\mathbb P}(\bm{X}_P>t^{\bm{\omega}}) = \ell(t; \bm{\omega})t^{-\lambda(\bm{\omega})}, \qquad \bm{\omega} \in \mathcal{S}_{\Sigma}=\left\{\bm{\omega} \in [0,1]^d : \sum_{j=1}^d \omega_j = 1\right\}, \label{eq:WT}
\end{align}
where $\ell(t; \bm{\omega}) \in \mathrm{RV}_0^\infty$ for each $\bm{\omega}$ and $\lambda:\mathcal{S}_{\Sigma} \to [0,1]$. Equivalently, ${\mathbb P}(\bm{X}_E>\bm{\omega}v) = \ell(e^v; \bm{\omega})e^{-\lambda(\bm{\omega})v}$. When all components of $\bm{\omega}$ are equal to $1/d$, connection with hidden regular variation on the cone $\mathbb{E}_D$ is restored, and we have $\eta_D = d\lambda(1/d,\ldots,1/d)$. When the subcone $\mathbb{E}_D$ of $\mathbb{E}$ is charged with mass in limit~\eqref{eq:mrv}, then $\lambda(\bm{\omega}) = \max_{1\leq j \leq d}\omega_j$. One can equally focus on sub-vectors indexed by $C$ to define $\lambda_{C}(\bm{\omega})$ for $\bm{\omega}$ in a $(|C|-1)$-dimensional simplex; we continue to have $\eta_C = |C|\lambda_C(1/|C|,\ldots,1/|C|)$ and $\nu(\mathbb{E}_C)>0$ implies $\lambda_{C}(\bm{\omega}) = \max_{1 \leq j \leq |C|} \omega_j$.
\subsection{Conditional extremes}
\label{sec:CEintro}
For conditional extreme value theory \citep{HeffernanTawn04,HeffernanResnick07}, we focus on $\bm{X}_E$. Let $\bm{X}_{E,-j}$ represent the vector $\bm{X}_E$ without the $j$th component. The basic assumption is that there exist functions $\bm{a}^j:\mathbb{R} \to \mathbb{R}^{d-1}$, $\bm{b}^j:\mathbb{R} \to \mathbb{R}_+^{d-1}$ and a non-degenerate distribution $K^j$ on $\mathbb{R}^{d-1}$ with no mass at infinity, such that
\begin{align}
{\mathbb P}\left(\frac{\bm{X}_{E,-j} - \bm{a}^j(X_{E,j})}{\bm{b}^j(X_{E,j})} \leq \bm{z}, X_{E,j}-t>x ~\Big|~ X_{E,j}>t\right) \to K^j(\bm{z}) e^{-x}, \qquad t \to \infty. \label{eq:ce}
\end{align}
Typically, such assumptions are made for each $j \in D$. The normalization functions satisfy some regularity conditions detailed in \citet{HeffernanResnick07}, but as \citet{HeffernanResnick07} only standardize the marginal distribution of the conditioning variable (i.e, $X_j$), allowing different margins in other variables, these conditions do not strongly characterize the functions $\bm{a}^j$ and $\bm{b}^j$ as used in~\eqref{eq:ce}.
When joint densities exist, application of L'H\^opital's rule gives that convergence~\eqref{eq:ce} is equivalent to
\begin{align*}
{\mathbb P}\left(\frac{\bm{X}_{E,-j} - \bm{a}^j(t)}{\bm{b}^j(t)} \leq \bm{z} ~\Big|~ X_{E,j}=t\right) \to K^j(\bm{z}), \qquad t \to \infty.
\end{align*}
We will further assume convergence of the full joint density
\begin{align}
\frac{\partial^{d-1}}{\partial \bm{z}}{\mathbb P}\left(\frac{\bm{X}_{E,-j} - \bm{a}^j(t)}{\bm{b}^j(t)} \leq \bm{z} ~\Big|~ X_{E,j}=t\right) \to \frac{\partial^{d-1}}{\partial \bm{z}} K^j(\bm{z}) =: k^j(\bm{z}), \qquad t \to \infty, \label{eq:cedensity}
\end{align}
which is the practical assumption needed for undertaking likelihood-based statistical inference using this model.
Connected to this approach is work in \cite{BE07}, who study asymptotic behaviour of a suitably normalized random vector~$\bm{X}$ conditional on lying in $tH$, where $H$ is a half-space not containing the origin and $t\to\infty$. The distribution of~$\bm{X}$ is assumed to have a light-tailed density whose level sets are homothetic, convex and have a smooth boundary. In this setting, with $H$ taken to be the vertical half-space $\{\bm{x}\in{\mathbb R}^d:\ x_d>1\}$, the limit is the so-called Gauss-exponential distribution with density $\exp\{-\bm{u}^T\bm{u}/2-v\}/(2\pi)^{(d-1)/2}$, $\bm{u}\in{\mathbb R}^{d-1}$, $v>0$.
\subsection{Limit sets}
\label{sec:Gauge}
\subsubsection{Background}
Let ${\bm X}_1,\ldots,{\bm X}_n$ be independent and identically distributed random vectors in ${\mathbb R}^d$. A random set $N_n=\{{\bm X}_1/r_n,\ldots,{\bm X}_n/r_n\}$ represents a scaled $n$-point sample cloud. We consider situations in which there exists a scaling sequence $r_n>0$, $r_n\to\infty$ such that scaled sample clouds $N_n$ converges onto a deterministic set, containing at least two points. Figure~\ref{fig:sc3} illustrates examples of sample clouds for which a limit set exists. Let ${\cal K}_d$ denote the family of non-empty compact subsets of ${\mathbb R}^d$, and $d_H(\cdot,\cdot)$ denote the Hausdorff distance between two sets \citep{Matheron1975}. A sequence of random sets $N_n$ in ${\cal K}_d$ converges in probability onto a \emph{limit set} $G\in{\cal K}_d$ if $d_H(N_n,G)\topr0$ for $n\to\infty$. The following result gives convenient criteria for showing convergence in probability onto a limit set; see \citet{BEN2010}.
\begin{prop}\label{pconvLS} Random samples on ${\mathbb R}^d$ scaled by $r_n$ converge in probability onto a deterministic set $G$ in ${\cal K}_d$ if and only if
\begin{enumerate}[(i)]
\item $n{\mathbb P}({\bm X}/r_n \in U^c)\to0$ for any open set $U$ containing $G$;
\item $n{\mathbb P}({\bm X}/r_n \in \{\bm{x}+\epsilon B\})\to\infty$ for all ${\bm x}\in G$ and any $\epsilon>0$, where $B$ is the Euclidean unit ball.
\end{enumerate}
\end{prop}
\begin{figure}
\centering
\includegraphics[width=1\textwidth]{Fig/plotSC3.png}
\caption{Sample clouds of $n=10^5$ points simulated from meta-Gaussian distributions with standard exponential margins and copula correlation parameter $\rho=0.5$ (left panel), $\rho=0$ (middle panel) and $\rho=-0.5$ (right panel). The samples are scaled by the factor $r_n=\log n$. See Examples~\ref{exGpos} and \ref{exGneg} for details.}
\label{fig:sc3}
\end{figure}
Limit sets under various assumptions on the underlying distribution have been derived in \citet{Geffroy1958,Geffroy1959,Fisher1969,Davisetal88,BEN2010}. \citet{KinoshitaResnick91} give a complete characterization of the possible limit sets as well as describe the class of distribution functions for which sample clouds can be scaled to converge (almost surely) onto a limit set. Furthermore, convergence in probability onto a limit set is implied by the tail large deviation principle studied in \citet{deValk2016a,deValk2016b}.
\citet{KinoshitaResnick91} showed that if sample clouds can be scaled to converge onto a limit set almost surely, then the limit set is compact and star-shaped. A set $G$ in ${\mathbb R}^d$ is star-shaped if ${\bm x}\in G$ implies $t{\bm x} \in G$ for all $t \in [0,1]$. For a set $G\in{\cal K}_d$, if the line segment ${\bf0}+t {\bm x}$, $t\in [0,1)$ is contained in the interior of $G$ for every ${\bm x}\in G$, then $G$ can be characterized by a continuous \emph{gauge function}:
$$g({\bm x})=\inf\{t\ge0:\ {\bm x}\in tG\},\qquad {\bm x}\in{\mathbb R}^d.$$
A gauge function satisfies homogeneity: $g(t{\bm x})=tg({\bm x})$ for all $t>0$, and the set~$G$ can be recovered from its gauge function via $G=\{{\bm x}\in{\mathbb R}^d:\ g({\bm x})\le1\}.$ Examples of a gauge function include a norm $\|\cdot\|$ on ${\mathbb R}^d$, in which case $G=\{{\bm x}\in{\mathbb R}^d:\ \|{\bm x}\|\le1\}$ is the unit ball in that norm.
The shape of the limit set conveys information about extremal dependence properties of the underlying distribution. In particular, \citet{BalkemaNolde10} make a connection between the shape of the limit set and asymptotic independence, whilst \citet{Nolde14} links its shape to the coefficient of residual tail dependence. We emphasize that the shape of the limit set depends on the choice of marginal distributions, as well as dependence structure. For example, if the components of $(X_1,X_2)$ are independent with common marginal distribution, then $G = \{(x,y) \in \mathbb{R}_+^2: x+y \leq 1\}$ if the margins are exponential; $G = \{(x,y) \in \mathbb{R}^2: |x|+|y| \leq 1\}$ if the margins are Laplace; and $G = \{(x,y) \in \mathbb{R}_+^2: (x^\beta +y^\beta)^{1/\beta} \leq 1\}$ if the margins are Weibull with shape $\beta>0$. In contrast, if the margins are exponential but $G$ takes the latter form, this implies some dependence between the components.
\subsubsection{Conditions for convergence onto a limit set}
Proposition~\ref{pconvLS} provides necessary and sufficient conditions for convergence onto the limit set $G$, but these conditions are not particularly helpful for determining the form of $G$ in practice.
In the following proposition, we state a criterion in terms of the joint probability density for convergence of suitably scaled random samples onto a limit set. This result is an adaptation of Proposition~3.7 in \citet{BalkemaNolde10}. The marginal tails of the underlying distribution are assumed to be asymptotically equal to a von Mises function. A function of the form $e^{-\psi}$ is said to be a \emph{von Mises function} if $\psi$ is a $C^2$ function with a positive derivative such that $(1/\psi'(x))'\to 0$ for $x\to\infty$. This condition on the margins says that they are light-tailed and lie in the maximum domain of attraction of the Gumbel distribution, i.e., for a random sample from such a univariate distribution, coordinate-wise maxima can be normalized to converge weakly to the Gumbel distribution \citep[Proposition~1.1]{Resnick1987}.
\begin{prop}
\label{prop:logdens}
Let the random vector ${\bm X}$ on $[0,\infty)^d$ have marginal distribution functions asymptotically equal to a von Mises function: $1-F_j(x) \sim e^{-\psi_j(x)}$ for $\psi_j(x)\sim \psi(x)$, $x\to\infty$ ($j=1,\ldots,d$) and a joint probability density~$f$ satisfying:
\begin{equation}\label{q1}
\dfrac{-\log f(t{\bm x}_t)}{\psi(t)}\to g_*({\bm x}),\qquad t\to\infty,\qquad {\bm x}_t\to{\bm x},\qquad {\bm x}\in[0,\infty)^d
\end{equation}\noindent
for a continuous function $g_*$ on $[0,\infty)^d$, which is positive outside a bounded set. Then a sequence of scaled random samples $N_n=\{{\bm X}_1/r_n,\ldots,{\bm X}_n/r_n\}$ from $f$ converges in probability onto a limit set $G$ with $G=\{{\bm x}\in[0,\infty)^d : g_*({\bm x})\le 1\}$. The scaling sequence $r_n$ can be chosen as $\psi(r_n)\sim \log n$. Moreover, $\max G=(1,\ldots,1)$.
\end{prop}
\par\noindent{\bf Proof } The mean measure of $N_n$ is given by $n{\mathbb P}({\bm X}/r_n\in\cdot)$ with intensity $h_n({\bm x}) = nr_n^df(r_n{\bm x})$. We show the convergence of that mean measure onto $G$, implying convergence of scaled samples $N_n$; see \citet{BEN2010}, Proposition~2.3. By \eqref{q1} and the choice of~$r_n$, we have
\begin{equation}\label{qcc}
-\log f(r_n{\bm x}_n)/\log n\sim -\log f(r_n{\bm x}_n)/\psi(r_n)\to g_*({\bm x}), \quad n\to\infty,\quad {\bm x}_n\to{\bm x}.
\end{equation}
Continuous convergence in~\eqref{qcc} with $g_*$ continuous implies uniform convergence on compact sets. Hence, $g_*$ is bounded on compact sets. For $G=\{g_*({\bm x})\le1\}$, we have $g_*({\bm x})<1$ on the interior of $G$ and $g_*({\bm x})>1$ on the complement of $G$.
Furthermore, applying L'H\^opital's rule and Lemma~1.2(a) in \citet{Resnick1987}, we have $$\log r_n/\psi(r_n)\sim (1/\psi'(r_n))/r_n\to 0,\qquad r_n\to\infty.$$
Combining these results, we see that $-\log h_n({\bm x}_n)\sim (g_*({\bm x}_n)-1)\log n$, which diverges to $-\infty$ on the interior of~$G$ and to $+\infty$ outside of~$G$. This implies that $$h_n({\bm x}_n)\to\begin{cases} \infty, & {\bm x}\in G^o,\\ 0, & {\bm x}\in G^c,\end{cases} $$
giving convergence (in probability) of $N_n$ onto limit set~$G$.
The form of the margins $1-F_j(x) \sim e^{-\psi_j(x)}$ with $\psi_j(x) \sim \psi(x)\to\infty$ gives $-\log(1-F_j(x))\sim\psi(x)$; i.e., $$-\log(1-F_j(r_n))\sim \psi(r_n)\sim\log n,\qquad n\to\infty.$$
This choice of $r_n$ implies that the coordinate-wise maxima scaled by $r_n$ converge in probability to 1 (\citet{G43,deHaan1970}), so that $\max G = (1,\ldots,1).$
\quad\hfill\mbox{\P}
\begin{rmk}
Condition~\eqref{q1} implies that~$-\log f$ is multivariate regularly varying on $[0,\infty)^d$. Such densities are referred to as Weibull-like. The limit function $g_*$ is homogeneous of some positive order~$k$: $g_*(t{\bm x})=t^k g_*({\bm x})$ for all $t>0$. The gauge function $g$ of the limit set $G$ can thus be obtained from $g_*$ by setting $g({\bm x})=g_*^{1/k}({\bm x})$.
\end{rmk}
When the margins are standard exponential, $\psi(t)=t$. Hence, for the random vector $\bm{X}_E$ with a Lebesgue density $f_E$ on $\mathbb{R}^d_+$, condition~\eqref{q1} is equivalent to
\begin{align}\label{ass:gaugedens}
-\log f_E(t\bm{x}_t)/t\to g_*(\bm{x}), \qquad t \to \infty,\qquad {\bm x}_t\to{\bm x},\qquad \bm{x} \in [0,\infty)^d
\end{align}
with the limit function $g_*$ equal to the gauge function $g$.
Whilst the assumption of a Lebesgue density might appear strict, it is a common feature in statistical practice of extreme value analysis. The assumption permits simple elucidation of the connection between different representations for multivariate extremes. Furthermore, many statistical models, including elliptical distributions and vine copulas \citep{Joe96,BedfordCooke01,BedfordCooke02}, are specified most readily in terms of their densities.
Convergence at the density level such as in~\eqref{q1} may not always hold. The condition requires the limit function and hence the gauge function of the limit set to be continuous, excluding limit sets for which rays from the origin cross the boundary in more than one point. We provide an example of such a situation in Section~\ref{sec:examples}; see Example~\ref{exGneg}. A less restrictive set of sufficient conditions for convergence of sample clouds onto a limit set can be obtained using the survival function. The following proposition is Theorem~2.1 in~\cite{Davisetal88}, with a minor reformulation in terms of scaling.
\begin{prop}
\label{prop:DMR}
Suppose that the random vector ${\bm X}$ has support on $[0,\infty)^d$, the margins are asymptotically equal to a von Mises function: $1-F_j(x) \sim e^{-\psi(x)}$ for $x\to\infty$ ($j=1,\ldots,d$), and the joint survival function satisfies
\begin{equation}\label{eq:DMR}
\dfrac{-\log {\mathbb P}({\bm X} \geq t{\bm x})}{\psi(t)}\to g_*({\bm x}),\qquad t\to\infty,\qquad {\bm x}\in [0,\infty)^d \setminus\{0\}.
\end{equation}\noindent
Further assume that $g_*$ is strictly increasing, such that $g_*({\bm x})<g_*({\bm y})$ if ${\bm x} \leq {\bm y}$ and ${\bm x} \neq {\bm y}$. Then for $r_n$ satisfying $\psi(r_n) \sim \log n$, the sample cloud $N_n=\{{\bm X}_1/r_n,\ldots,{\bm X}_n/r_n\}$ converges onto $G=\{{\bm x}\in[0,\infty)^d : g_*({\bm x})\le 1\}$.
\end{prop}
\subsubsection{Marginalization}
\label{sec:marg}
When $d>2$, a key question is the marginalization from dimension $d$ to dimension $m<d$. We prove below that, as long as the minimum over each coordinate of $g$ is well-defined, then the gauge function determining the limit set in $m$ dimensions is found through minimizing over the coordinates to be marginalized.
A continuous map $h$ from the vector space $V$ into the vector space $\widetilde V$ is \emph{positive-homogeneous} if $h(r{\bm x})=rh({\bm x})$ for all ${\bm x}\in V$ and all $r>0$. If $\widetilde V={\mathbb R}^{m}$, the map $h$ is determined by the $m$ coordinate maps $h_j:V\to{\mathbb R}$, $j=1,\ldots,m$ and in this case it suffices that these maps are continuous and positive-homogeneous.
Convergence onto a limit set is preserved under linear transformations (e.g., Lemma~4.1 in \cite{Nolde14}) and more generally under continuous positive-homogeneous maps with the same scaling sequences (Theorem~1.9 in \cite{BN20}). A consequence of the latter result, referred to as the Mapping Theorem, is that projections of sample clouds onto lower-dimensional sub-spaces also converge onto a limit set.
\begin{prop}\label{prop:marg} Let $N_n$ be an $n$-point sample cloud from a distribution of random vector ${\bm X}$ on ${\mathbb R}^d$. Assume $N_n$ converges in probability, as $n\to\infty$, onto a limit set $G=\{{\bm x}\in{\mathbb R}^d : g({\bm x})\le1\}$ for a gauge function $g$. Let $\widetilde{\bm X} = (X_i)_{i\in I_m}$ denote an $m$-dimensional marginal of ${\bm X}$, where $I_m\subset I=\{1,\ldots,d\}$ is an index set with $|I_m|=m$. Sample clouds from $\widetilde{\bm X}$ also converge, with the same scaling, and the limit set $\tilde G = P_m(G) = \{{\bm y}\in{\mathbb R}^m : \tilde g({\bm y})\le1\}$, where $P_m$ is a projection map onto the coordinates of $\widetilde{\bm X}$ and $$\tilde g({\bm y}) = \min_{\{x_i : i\in I\setminus I_m\}} g({\bm x}),\qquad {\bm x}=(x_1,\ldots,x_d),\qquad {\bm y}=(x_i)_{i\in I_m}.$$
\end{prop}
\par\noindent{\bf Proof } Consider the bivariate case first with $\widetilde{\bm X}=X_2$. Sample clouds from $X_2$ converge onto the limit set $\tilde G\subset{\mathbb R}$, which is the projection of $G$ onto the $x_1$-coordinate axis, by the Mapping Theorem. The projection is determined by the tangent to the level curve $\{{\bm x}\in{\mathbb R}^2 : g({\bm x})=1\}$ orthogonal to the $x_1$-coordinate axis. Similarly, level curves of the gauge function $\tilde g$ of the set $\widetilde G$ are determined by tangents to the level curves $\{{\bm x}\in{\mathbb R}^2 : g({\bm x})=c\}$ for $c\in[0,1]$ orthogonal to the $x_1$-coordinate axis. These projections correspond to $x_1$ values which minimize $g(x_1,x_2)$. Sequentially minimizing over each of the coordinates to be marginalized gives the result.
\quad\hfill\mbox{\P}
An illustration of this result is given in Section~\ref{sec:egd3}.
\section{Linking representations for extremes to the limit set}
\label{sec:connections}
For simplicity of presentation, in what follows we standardize to consider exponential margins for the light-tailed case. This choice is convenient when there is positive association in the extremes, but hides structure related to negative dependence. We comment further on this case in Section~\ref{sec:Discussion}. Owing to the standardized marginals, it makes sense to refer to the limit set, rather than a limit set.
Connections between multivariate and hidden regular variation are well established, with the latter requiring the former for proper definition. Some connection between regular variation and conditional extremes was made in \citet{HeffernanResnick07} and \citet{DasResnick11}, although they did not specify to exponential-tailed margins. The shape of the limit set has been linked to the asymptotic (in)dependence structure of a random vector \citep{BalkemaNolde10,BalkemaNolde12}. Asymptotic independence is related to the position of mass from convergence~\eqref{eq:mrv} on $\mathbb{E}$, but regular variation and the existence of a limit set in suitable margins are different conditions and one need not imply the other. \citet{Nolde14} links the limit set $G$ to the coefficient of residual tail dependence, $\eta_D$.
In this section we present some new connections between the shape of the limit set, when it exists, and normalizing functions in conditional extreme value theory, the residual tail dependence coefficient, the function $\lambda(\bm{\omega})$ and the coefficients $\tau_{C}(\delta)$.
\subsection{Conditional extremes}
\label{sec:CE}
For the conditional extreme value model, the form of the normalizing functions $\bm{a}^j,\bm{b}^j$ is determined by the pairwise dependencies between $(X_{E,i},X_{E,j})$, $i \in D \setminus j$. The two-dimensional marginalization of any $d$-dimensional gauge function is given by Proposition~\ref{prop:marg}, and we simply denote this by $g$ here.
\begin{prop}
\label{prop:CEgauge}
Suppose that for $\bm{X}_E = (X_{E,1},X_{E,2})$ convergence~\eqref{eq:cedensity} and assumption~\eqref{ass:gaugedens} hold, where the domain of $K^j$ includes $(0,\infty)$. Define $\alpha_j = \lim_{x \to \infty} a^j(x)/x$, $j= 1,2$. Then
\begin{enumerate}[(i)]
\item $g(1,\alpha_1) = 1$, $g(\alpha_2, 1) = 1$.
\item Suppose that $-\log f_E(t\bm{x}_t)/t = g(\bm{x}_t) + v(t)$, with $v(t) \in \mathrm{RV}_{-1}^{\infty}$ or $v(t)=o(1/t)$, and $a^j(t) = \alpha_j t + B^j(t)$ with either $B^j(t)/b^j(t) \in \mathrm{RV}^\infty_0$, or $B^j(t) = o(b^j(t))$. For $\beta_1,\beta_2 \leq 1$, if $g(1, \alpha_1 + \cdot) -1 \in \mathrm{RV}_{1/(1-\beta_1)}^0$, then $b^1(x) \in \mathrm{RV}_{\beta_1}^\infty$; similarly if $g(\alpha_2 + \cdot,1) -1 \in \mathrm{RV}_{1/(1-\beta_2)}^0$, then $b^2(x) \in \mathrm{RV}_{\beta_2}^\infty$.
\item If there are multiple values $\alpha$ satisfying $g(1,\alpha) = 1$, then $\alpha_1$ is the maximum such $\alpha$, and likewise for $\alpha_2$.
\end{enumerate}
\end{prop}
Before the proof of Proposition~\ref{prop:CEgauge}, we give some geometric intuition. Figure~\ref{fig:2dgauge} presents several examples of the unit level set of possible gauge functions, illustrating the shape of the limit set, for two-dimensional random vectors with exponential margins. On each figure, the slope of the red line indicates the value of $\alpha_1$; i.e., the equation of the red line is $y=\alpha_1 x$. Intuitively, conditional extreme value theory poses the question: ``given that variable $X$ is growing, how does variable $Y$ grow as a function of $X$?''. We can now see that this is neatly described by the shape of the limit set: to first order, the values of $Y$ occurring with large $X$ are determined by the direction for which $X$ is growing at its maximum rate. The necessity of a scale normalization in the conditional extreme value limit depends on the local curvature and particularly the rate at which $g(\alpha_1+u,1)$ approaches 1 as $u \to 0$. For cases (i), (iv), (v) and (vi) of Figure~\ref{fig:2dgauge}, the function approaches zero linearly in $u$: as a consequence $b^j(t) \in \mathrm{RV}_0^\infty$. For case (ii) the order of decay is $u^2$ and so $b^j(t) \in \mathrm{RV}_{1/2}^\infty$, whilst for (iii) the order is $u^{1/\theta}$ so $b^j(t) \in \mathrm{RV}_{1-\theta}^\infty$.
The class of distributions represented by gauge function (vi) (bottom left) can be thought of as those arising from a mixture of distributions with gauge functions (i) and (iii), up to differences in parameter values. In such an example, there are two normalizations that would lead to a non-degenerate limit in~\eqref{eq:ce}, but ruling out mass at infinity produces the unique choice $\alpha_1=\alpha_2=1$, $\beta_1=\beta_2=0$. If instead we chose to rule out mass at $-\infty$, then we would have $\alpha_1=\alpha_2 = 0$ and $\beta_1=\beta_2 = 1-\theta$.
\begin{proof}[Proof of Proposition~\ref{prop:CEgauge}]
In all cases we just prove one statement as the other follows analogously.\\
(i) By assumption~\eqref{ass:gaugedens}, $(X_{E,1},X_{E,2})$ have a joint density $f_E$ and so conditional extremes convergence~\eqref{eq:cedensity} can be expressed as
\[
b^1(t)f_E(t,b^1(t)z+a^1(t))e^t \to k^1(z) =:e^{-h^1(z)}, \qquad t \to \infty,~~~z \in [\lim_{t \to \infty}-a^1(t)/b^1(t), \infty)
\]
with $k^1 = e^{-h^1}$ a density. Taking logs, we have
\begin{align}
-\log f_E(t,b^1(t)z+a^1(t)) - t -\log b^1(t) \to h^1(z),\qquad t\to\infty. \label{eq:logconv}
\end{align}
Now use assumption~\eqref{ass:gaugedens} with $\bm{x}_t =(1,x_t)= (1,a^1(t)/t + zb^1(t)/t)$. That is,
\begin{align}
-\log f_E(t,b^1(t)z+a^1(t)) =t g(1,x)[1+o(1)] = t g(1,x_t)[1+o(1)] , \label{eq:logg}
\end{align}
with $x = \lim_{t\to\infty} a^1(t)/t + zb^1(t)/t$. As the support of $K^1$ includes $(0,\infty)$, $h^1(z)<\infty$ for all $z \in (0,\infty)$, and combining~\eqref{eq:logconv} and~\eqref{eq:logg} we have
\begin{align}
g(1,x_t)[1+o(1)] = 1 + h^1(z)/t +\log b^1(t)/t + o(1/t).\label{eq:prelim}
\end{align}
Suppose that $b^1(t)/t \to \gamma>0$. Then $x_t \to \alpha_1 + \gamma z$, and taking $t\to \infty$ in~\eqref{eq:prelim} leads to $g(1,\alpha_1 + \gamma z) =1$ for any $z$. But since the coordinatewise supremum of $G$ is $(1,1)$, $g(x,y) \geq \max(x,y)$ which would entail $z \leq (1-\alpha_1)/\gamma$. No such upper bound applies, so we conclude $\gamma=0$, i.e., $b^1(t) = o(t)$. Now taking limits in~\eqref{eq:prelim} leads to $g(1,\alpha_1)=1.$
(ii) Let $g(1,\alpha_1+u)-1 =:r(u) \in \mathrm{RV}_\rho^0$, $\rho>0$. We also have from~\eqref{eq:prelim} $g(1,\alpha_1 + b^1(t)/t+B^1(t)/t)-1 = h^1(1)/t + \log b^1(t)/t - v(t) + o(1/t)$, so that the function $b^1(t)$ is a solution to the equation
\begin{align}
r(b^1(t)/t+B^1(t)/t) = h^1(1)/t + \log b^1(t)/t - v(t)+ o(1/t). \label{eq:exprv}
\end{align}
Equation~\eqref{eq:exprv} admits a solution if $b^1$ is regularly varying at infinity. A rearrangement provides that
\begin{align*}
b^1(t) = t r^{-1}(h^1(1)/t + \log b^1(t)/t - v(t)+ o(1/t))[1+ B^1(t)/b^1(t)]^{-1};
\end{align*}
if $b^1$ is regularly varying then $\log b^1(t)/t \in \mathrm{RV}_{-1}^{\infty}$, so that using the fact that $v(t) \in \mathrm{RV}_{-1}^\infty$ or $v(t)=o(1/t)$, combined with $r^{-1} \in \mathrm{RV}_{1/\rho}^0$, yields $b^1(t) \in \mathrm{RV}_{1-1/\rho}^{\infty}$. We now argue that such a solution is unique in this context. We know that the normalization functions $a^1,b^1$ lead to a non-degenerate distribution $K^1$ that places no mass at infinity. By the convergence to types theorem (\citet[][p.7]{Leadbetter83}, see also part (iii) of this proof), any other function $\tilde{b}^1$ leading to a non-degenerate limit with no mass at infinity must satisfy $\tilde{b}^1(t) \sim d b^1(t)$, $t \to \infty$, for some $d>0$, so that $\tilde{b}^1 \in \mathrm{RV}_{1-1/\rho}^\infty$ also. Finally, setting $\beta_1=1-1/\rho$ gives $b^1 \in \mathrm{RV}_{\beta_1}^{\infty}$.
(iii) Suppose that
\[
{\mathbb P}\left(\frac{X_{E,2}-a^1(t)}{b^1(t)} \leq z \big| X_{E,1}>t\right) \to K^1(z), ~~~~ {\mathbb P}\left(\frac{X_{E,2}-\tilde{a}^1(t)}{\tilde{b}^1(t)} \leq z \big| X_{E,1}>t\right) \to \tilde{K}^1(z),
\]
where neither $K^1$ nor $\tilde{K}^1$ has mass at $+ \infty$. Then by the convergence to types theorem, $\tilde{a}^1(t) = a^1(t) + cb^1(t) +o(b^1(t))$ and $\tilde{b}^1(t) = db^1(t) +o(b^1(t))$, for some $d>0$, and $\tilde{K}^1(z) = K^1(z/d+c)$. As such, $\tilde{a}^1(t)/t \sim a^1(t)/t \sim \alpha_1$. We conclude that if there was a non-degenerate $\tilde{K}^1$ limit for which $\tilde{a}^1(t)/t \sim \tilde{\alpha}_1>\alpha_1$ then $K^1$ must place mass at $+\infty$; since by assumption it does not, then $\alpha_1$ is the maximum value satisfying $g(1,\alpha_1) = 1$.
\end{proof}
\begin{figure}
\centering
\includegraphics[width=0.3\textwidth]{Fig/g1all.pdf}
\includegraphics[width=0.3\textwidth]{Fig/g2all.pdf}
\includegraphics[width=0.3\textwidth]{Fig/g3all.pdf}\\
\includegraphics[width=0.3\textwidth]{Fig/g4all.pdf}
\includegraphics[width=0.3\textwidth]{Fig/g8all.pdf}
\includegraphics[width=0.3\textwidth]{Fig/g7all.pdf}
\caption{Unit level sets of six possible gauge functions for bivariate random vectors with exponential margins. The limit set $G$ is the set bounded by these level sets and the axes. In each case the red line is $y=\alpha_1 x$, the blue lines represent $\lambda(\omega,1-\omega)$ and dots $\eta_{1,2}$ (see Section~\ref{sec:lambda}). Dashed lines represent the boundary $\max(x,y)=1$. Clockwise from top left, the gauge functions represented are: (i) $\max(x,y)/\theta+(1-1/\theta)\min(x,y)$; (ii) $(x+y-2\theta\sqrt{xy})/(1-\theta^2)$; (iii) $(x^{1/\theta}+y^{1/\theta})^\theta$; (iv) $\max\{(x-y)/\theta,(y-x)/\theta,\min(x-\mu y,y-\mu x)/(1-\theta-\mu)\}$; (v) $\max((x-y)/\theta,(y-x)/\theta,(x+y)/(2-\theta))$; (vi) $\min\{\max(x,y)/\theta_1+(1-1/\theta_1)\min(x,y), (x^{1/\theta_2}+y^{1/\theta_2})^{\theta_2}\}$. In each case $\theta \in (0,1)$; in some cases the endpoints are permitted as well. For case (iv), $\theta+\mu<1$.}
\label{fig:2dgauge}
\end{figure}
For distributions whose sample clouds converge onto a limit set described by a gauge function with piecewise-continuous partial derivatives possessing finite left and right limits, further detail can be given about $\beta_j$.
\begin{prop}
Let $G$ be a limit set whose gauge function $g$ has piecewise-continuous partial derivatives $g_1(x,y)=\partial g(x,y)/\partial x$, $g_2(x,y)=\partial g(x,y)/\partial y$ possessing finite left and right limits, and for which the conditions of Proposition~\ref{prop:CEgauge} hold. Then (i) $\beta_1\geq0$ if $g_2(1,(\alpha_1)_+)=0$; (ii) $\beta_1=0$ if $0<g_2(1,(\alpha_1)_+)<\infty$. Further, if $\alpha_1>0$ then $0 \leq g_2(1,(\alpha_1)_+)< \infty$, so that $\beta_1 \geq 0$. Analogous statements hold for $\alpha_2,\beta_2$.
\label{prop:betapos}
\end{prop}
\begin{proof}
Consider the partial derivative
\[
g_2(1,(\alpha_1)_+) = \lim_{u\to 0^+} \frac{g(1,\alpha_1+u)-g(1,\alpha_1)}{u} \geq 0,
\]
as $g(1,\alpha_1+u) \geq g(1,\alpha_1)$. We note $g(1,\alpha_1)=1$, such that $g(1,\alpha_1+u)-1 \sim u g_2(1,(\alpha_1)_+)$, $u \to 0^+$. Since this is regularly varying with index $1/(1-\beta_1)$ by assumption, $g_2(1,(\alpha_1)_+)=0$ implies $g(1,\alpha_1+u)-1=o(u)$, hence $1/(1-\beta_1)\geq1$, and $0<g_2(1,(\alpha_1)_+)<\infty$ implies $1/(1-\beta_1)=1$, so (i) and (ii) follow. If $g$ is differentiable at the point $(1,\alpha_1)$, then since $g(1,y)\geq 1$, $g_2(1,(\alpha_1)_+) = g_2(1,(\alpha_1)_-)=0$ and (i) holds. Otherwise, in a neighbourhood of $(1,\alpha_1)$, we can express
\begin{align*}
g(x,y) = \begin{cases}
\hat{g}(x,y),& y \leq \alpha_1 x\\
\tilde{g}(x,y),& y \geq \alpha_1 x,
\end{cases}
\end{align*}
where the homogeneous functions $\tilde{g}$ and $\hat{g}$ have continuous partial derivatives at $(1,\alpha_1)$.
Euler's homogeneous function theorem gives $1 = \tilde{g}_1(1,\alpha_1)+\tilde{g}_2(1,\alpha_1)\alpha_1 = g_1(1_-,\alpha_1)+\alpha_1 g_2(1,(\alpha_1)_+)$ so that for $\alpha_1>0$, $g_2(1,(\alpha_1)_+)<\infty$, and hence (i) or (ii) hold.
\end{proof}
We remark on links with existing work on conditional extreme value limits for variables with a polar-type representation, whereby $(X_1,X_2)=R(W_1,W_2)$ for $R>0$ and $(W_1,W_2)$ constrained by some functional dependence. \citet{AbdousEtAl05}, \citet{FougeresSoulier10} and \citet{Seifert14} consider a type of conditional extremes limit for certain such polar constructions, where in the light-tailed case, the shape of the constraint on $(W_1,W_2)$ feeds into the normalization and limit distribution. However, limit sets are sensitive to marginal choice, and because the above papers do not consider conditional extreme value limits in standardized exponential-tailed margins, further connections are limited.
\subsection{Different scaling orders: $\lambda(\bm{\omega})$}
\label{sec:lambda}
We now focus on the connection with $\lambda(\bm{\omega})$, as defined in Section~\ref{sec:dso}. When $\bm{\omega} = (1/d,\ldots,1/d)$, this yields the link with the residual tail dependence coefficient $\eta_{D}$, which has already been considered in \citet{Nolde14}. Define the region
\[
R_{\bm{\omega}}=\left(\frac{\omega_1}{\max(\omega_1,\ldots,\omega_d)} , \infty\right] \times \cdots \times \left(\frac{\omega_d}{\max(\omega_1,\ldots,\omega_d)}, \infty\right].
\]
\begin{prop}
Suppose that the sample cloud $N_n = \{\bm{X}_E^1/\log n,\ldots,\bm{X}_E^n/\log n\}$ converges onto a limit set $G$, and that for each $\bm{\omega}\in\mathcal{S}_{\Sigma}$, equation~\eqref{eq:WT} holds. Then
\label{prop:lambda}
\[
\lambda(\bm{\omega}) = \max(\bm{\omega}) \times r_{\bm{\omega}}^{-1} ,
\]
where
\[
r_{\bm{\omega}}=\min\left\{r \in [0,1] : r R_{\bm{\omega}} \cap G = \emptyset \right\}.
\]
\end{prop}
\begin{cor}[\citet{Nolde14}]
\[
1/\eta_{D} = d \lambda(1/d,\ldots,1/d) =r_{(1/d,\ldots,1/d)}^{-1} = \left[\min\left\{r \in [0,1] : \left(r , \infty\right]^d \cap G = \emptyset \right\}\right]^{-1}.
\]
\end{cor}
\begin{proof}[Proof of Proposition~\ref{prop:lambda}]
The proof follows very similar lines to Proposition~2.1 of \citet{Nolde14}. Firstly note that $\lambda(\bm{\omega}) = \kappa(\bm{\omega})$, where $\kappa: [0,\infty)^d \setminus \{\bm{0}\} \to (0,\infty)$ is a 1-homogeneous function defined by
\[
{\mathbb P}(\bm{X}_{E}> \bm{\beta} t) = \ell(e^{t}; \bm{\beta})e^{-t \kappa(\bm{\beta})},\qquad \ell(\cdot;\bm{\beta}) \in \mathrm{RV}_{0}^\infty~\text{for every}~ \bm{\beta} \in [0,\infty)^d \setminus \{\bm{0}\} .
\]
As a consequence,
\begin{align}
\frac{\lambda(\bm{\omega})}{\max(\bm{\omega})} = \lim_{t\to\infty}-\log{\mathbb P}\{\bm{X}_{E} > t\bm{\omega}/\max(\bm{\omega}) \}/t. \label{eq:lambdalim}
\end{align}
Without loss of generality, suppose that $\max(\bm{\omega}) = \omega_d$, so that $R_{\bm{\omega}}=(\omega_1/\omega_d,\infty] \times \cdots \times (\omega_{d-1}/\omega_d,\infty] \times (1,\infty]$. Because of the convergence of the sample cloud onto $G$, we have by Proposition~\ref{pconvLS} that for any $\epsilon >0$ and large enough $t$
\[
{\mathbb P}(\bm{X}_E \in t e^\epsilon r_{\bm{\omega}} R_{\bm{\omega}}) \leq {\mathbb P}(\bm{X}_E \in t R_{(0,\ldots,0,1)}) = e^{-t} \leq {\mathbb P}(\bm{X}_E \in t e^{-\epsilon} r_{\bm{\omega}} R_{\bm{\omega}}),
\]
implying $-\log{\mathbb P}(\bm{X}_E \in t r_{\bm{\omega}} R_{\bm{\omega}}) \sim t$. Therefore $-\log{\mathbb P}(\bm{X}_E \in t R_{\bm{\omega}}) \sim t r_{\bm{\omega}}^{-1}$, and combining with equation~\eqref{eq:lambdalim} gives the result.
\end{proof}
Figure~\ref{fig:lambda} illustrates some of the concepts used in the proof of Proposition~\ref{prop:lambda} when $d=2$ and $\bm{\omega} = (\omega,1-\omega)$.
\begin{figure}[h]
\centering
\includegraphics[width=0.4\textwidth]{Fig/LambdaIllustration.pdf}
\caption{Illustration of the concepts used in the proof of Proposition~\ref{prop:lambda}: the green line represents the ray $x=\{\omega/(1-\omega)\}y$, $\omega<1/2$. The region above the purple line represents $R_{(0,1)}$, the region to the north-east of the thick dotted blue lines represents $r_{\omega}R_{\omega}$, whilst the two sets of thin dotted blue lines illustrate the regions $e^{-\epsilon }r_{\omega}R_{\omega}$ and $e^{\epsilon}r_{\omega}R_{\omega}$. The ratio of the distance from the origin to where the green line intersects the boundary of the limit set $G$, and where the green line intersects the boundary $\max(x,y)=1$, is equal to $r_{\omega}$.}
\label{fig:lambda}
\end{figure}
The blue lines in Figure~\ref{fig:2dgauge} represent $\lambda(\bm{\omega})$, depicting the unit level set of $\lambda(\omega,1-\omega)/ \max(\omega,1-\omega)$, and the dots illustrate the value of $r_{1/2} = \eta_{1,2}$. We can now see clearly how, in two dimensions, different dependence features are picked out by the conditional extremes representation and hidden regular variation based on $\eta_{1,2}$. Often, values of $\eta_{1,2}>1/2$ or $\alpha>0$ are associated with positive extremal dependence. From example~(iv) of Figure~\ref{fig:2dgauge} (bottom right), we observe $\eta_{1,2}<1/2$ but $\alpha>0$. We have that $Y$ does grow with $X$ (and vice versa) but only at a specific rate. On the other hand, joint extremes, where $(X,Y)$ take similar values, are rare, occurring less frequently than under independence.
From example~(iv) we can also see that one of the conclusions following Proposition~2.1 in \citet{Nolde14} is not true: the point $(r_{1/2},r_{1/2})$ need not lie on the boundary of $G$, meaning that we do not necessarily have $\eta_{D} = 1/g(\bm{1})$, although we can deduce the bound $\eta_{D} \geq 1/g(\bm{1})$. Similarly, there are occasions when $g(r_{\bm{\omega}}\bm{\omega}/\max(\bm{\omega})) = 1$, implying $\lambda(\bm{\omega})=g(\bm{\omega})$, but clearly this is not always true. In Proposition~\ref{prop:lambdag}, we resolve when this is the case by representing $r_{\bm{\omega}}$ in terms of $g$.
Define $B_{\bm{\omega}}$ to be the boundary of the region $R_{\bm{\omega}}$, i.e.,
\[
B_{\bm{\omega}} = \bigcup_{i=1}^d \{\bm{x} \in \mathbb{R}^d_+: x_i = \omega_i/\max(\bm{\omega}), x_j \geq \omega_j/\max(\bm{\omega}), j \neq i\}.
\]
\begin{prop}
Assume the conditions of Proposition~\ref{prop:lambda}. Then
\label{prop:lambdag}
\[r_{\bm{\omega}} = \left[\min_{\bm{y} \in B_{\bm{\omega}}} g(\bm{y})\right]^{-1},~~~\text{and hence}~~~ \lambda(\bm{\omega}) = \max(\bm{\omega}) \times \min_{\bm{y} \in B_{\bm{\omega}}} g(\bm{y}).
\]
\end{prop}
From Proposition~\ref{prop:lambdag}, we observe that $\lambda(\bm{\omega}) = g(\bm{\omega})$ if $\arg\min_{y \in B_{\bm{\omega}}} g(\bm{y}) = \bm{\omega}/\max(\bm{\omega})$, i.e., the vertex of the set $B_{\bm{\omega}}$. The proof of Proposition~\ref{prop:lambdag} is deferred until after Proposition~\ref{prop:taug}, for which the proof is very similar.
\begin{rmk}
We note that $\min_{\bm{y} \in B_{\bm{\omega}}} g(\bm{y}) = \min_{\bm{y} \in R_{\bm{\omega}}} g(\bm{y})$.
\end{rmk}
\subsection{Coefficients $\tau_C(\delta)$}
\label{sec:tau}
\subsubsection{Connections to limit set $G$}
\label{sec:tau2}
In two dimensions, the coefficients $\tau_{1}(\delta)$ and $\tau_{2}(\delta)$ provide a somewhat complementary concept to the function $\lambda(\bm{\omega})$. Rather than considering the impact of the limit set $G$ on the shape of the function defined by both variables exceeding thresholds growing at different rates, we are considering what is occurring when one variable exceeds a growing threshold and the other is upper bounded by a certain lesser growth rate. The left and centre panels in Figure~\ref{fig:lambdatau} provide an illustration of $\lambda(\bm{\omega})$ and $\tau_j(\delta)$ in two dimensions.
Define the region $R_{C,\delta} = (1,\infty]^C \times [0,\delta]^{D \setminus C} = \{\bm{x}: x_i \in (1,\infty], i \in C, x_j \in [0,\delta], j \in D\setminus C\}$, so that, for example, when $d=3$, $R_{\{1,3\},\delta} = (1,\infty] \times [0,\delta] \times (1,\infty]$.
\begin{prop}
\label{prop:tau}
Suppose that the sample cloud $N_n = \{\bm{X}_E^1/\log n,\ldots,\bm{X}_E^n/\log n\}$ converges onto a limit set $G$, and that the assumption in equation~\eqref{eq:SWT2} holds. For $\delta \in [0,1]$, and $C \subset D$,
\begin{align*}
\tau_C(\delta) &= r_{C,\delta} = \min\left\{r \in [0,1] : r R_{C,\delta} \cap G =\emptyset \right\}.
\end{align*}
The coefficient $\tau_{D} = \eta_{D}$, and does not depend on $\delta$.
\end{prop}
\begin{proof}
The coefficient $\tau_{D}$ describes the order of hidden regular variation on the cone $(0,\infty]^d$, which is precisely the same as $\eta_{D}$. For $\tau_C(\delta)$, $|C|<d$, we consider the function of $t$
\begin{align*}
{\mathbb P}(\min_{i \in C} X_{P,i}>tx, \max_{j \in D \setminus C} X_{P,j} \leq y t^\delta) \in \mathrm{RV}_{-1/\tau_C(\delta)}^\infty, \qquad 0<x,y<\infty.
\end{align*}
Take $x=y=1$. Then
\begin{align}
\tau_C(\delta) &= \lim_{t \to \infty} \frac{-\log{\mathbb P}(X_{P,1}>t)}{-\log {\mathbb P}(\min_{i \in C} X_{P,i}>t, \max_{j \in D \setminus C} X_{P,j} \leq t^\delta)} \notag \\&= \lim_{t \to \infty} \frac{-\log{\mathbb P}(X_{E,1}>t)}{-\log{\mathbb P}(\min_{i \in C} X_{E,i}>t, \max_{j \in D \setminus C} X_{E,j} \leq \delta t)} ,
\label{eq:tau}
\end{align}
where the denominator in~\eqref{eq:tau} can be expressed $-\log{\mathbb P}(\bm{X}_E \in t R_{\delta})$. As in the proof of Proposition~\ref{prop:lambda}, the convergence onto the limit set and exponential margins enables us to conclude that $-\log {\mathbb P}(\bm{X}_E \in t r_{C,\delta} R_{\delta}) \sim t$, and hence $-\log {\mathbb P}(\bm{X}_E \in t R_{\delta}) \sim tr_{C,\delta}^{-1}$. Combining with~\eqref{eq:tau} gives $\tau_C(\delta) = r_{C,\delta}$.
\end{proof}
In the two-dimensional case, it is possible to express $\tau_j(\delta)$ simply in terms of the gauge function. For higher dimensions, we refer to Proposition~\ref{prop:taug}.
\begin{prop}
Assume the conditions of Proposition~\ref{prop:tau}. When $d=2$,
\label{prop:tau2dg}
$\tau_1(\delta) = [\min_{\gamma \in [0,\delta]} g(1,\gamma)]^{-1}$ and $\tau_2(\delta) = [\min_{\gamma \in [0,\delta]} g(\gamma,1)]^{-1}$.
\end{prop}
\begin{proof}
For $\gamma \in [0,1]$, the points $\left(1/g(1,\gamma), \gamma/g(1,\gamma)\right)$ lie on the curve $\{(x,y) \in [0,1]^2:g(x,y)=1, x\geq y\}$. The value $r_{1,\delta}$ is the maximum value of $1/g(1,\gamma)$ for $\gamma \in [0,\delta]$, hence $\tau_1(\delta) = [\min_{\gamma \in [0,\delta]} g(1,\gamma)]^{-1}$. A symmetric argument applies to $\tau_2(\delta)$.
\end{proof}
The right panel of Figure~\ref{fig:lambdatau} provides an illustration: in blue the value of $\delta$ is such that $\tau_1(\delta)<1$; in red the value of $\delta$ is such that $\tau_2(\delta)=1$. Further detail on this example is given in Section~\ref{sec:egg}.
\begin{figure}
\centering
\includegraphics[width=0.3\textwidth]{Fig/lambda.pdf}
\includegraphics[width=0.3\textwidth]{Fig/tau.pdf}
\includegraphics[width=0.3\textwidth]{Fig/taudelta.pdf}
\caption{Illustration of $\lambda(\bm{\omega})$ and $\tau_{j}(\delta)$ for the gauge function $g(x,y) = \max\{(x-y)/\theta,(y-x)/\theta,(x+y)/(2-\theta)\}$. The left panel illustrates $\lambda(\omega,1-\omega)$ in blue. The centre panel illustrates $\tau_j(\delta)$ in purple: $\tau_1(\delta), \delta \in [0,1]$ is represented by the values below the main diagonal, whilst $\tau_2(\delta)$ is represented by those values above. The set $G$ is added on both panels with dashed lines. The right panel illustrates $\tau_1(\delta)$ and $\tau_2(\delta)$ in terms of the gauge function.}
\label{fig:lambdatau}
\end{figure}
The question arises: does Proposition~\ref{prop:tau2dg} still hold for $d>2, |C| = 1$? Let $g_C$ denote the gauge function for the limit set of $(X_{E,j}: j \in C)$. By Proposition~\ref{prop:marg}, we know that $g_{i,j}(x_i,x_j) = \min_{\bm{x}_{-i,-j} \in [0,\infty)^{d-2}}g_D(\bm{x})$. As such, equality will hold if $\arg\min_{\bm{x}_{-i,-j} \in [0,\infty)^{d-2}}g_D(\bm{x}) \in [0,\delta]^{d-2}$. Note that the dimension does indeed play a key role here: when looking at $\tau_j(\delta)$ for a $d$-dimensional problem, we are looking at the situation where $d-1$ coordinates are upper bounded by a growth rate determined by $\delta$. In contrast, when marginalizing and looking at $\tau_j(\delta)$ for a 2-dimensional problem, the $d-2$ coordinates that we have marginalized over are unrestricted and so can represent small or large values. As such, the answer to our question is negative in general.
Proposition~\ref{prop:taug} details the precise value of $\tau_C(\delta)$ in terms of $g$ for any dimension $d$. In a similar spirit to Section~\ref{sec:lambda}, define the boundary of the region $R_{C,\delta}$ as
\[
B_{C,\delta} = B_{C,\delta}^1 \cup B_{C,\delta}^\delta,
\]
where
\begin{align*}
B_{C,\delta}^1 &= \bigcup_{i \in C} \{\bm{x}\in\mathbb{R}^d_+: x_i = 1, x_j \geq 1~ \forall j \in C \setminus i,~ x_k \leq \delta~ \forall k \in D \setminus C \}\\
B_{C,\delta}^\delta &= \bigcup_{i \in D \setminus C} \{\bm{x}\in\mathbb{R}^d_+: x_i = \delta, x_j \leq \delta~ \forall j \in (D\setminus C) \setminus i,~ x_k \geq 1~ \forall k \in C \},
\end{align*}
so, for example, when $d=3$,
\[B_{\{1,3\},\delta} = \{\bm{x}\in \mathbb{R}^3_+: x_1 = 1, x_2 \leq \delta, x_3 \geq 1\} \cup \{\bm{x}\in \mathbb{R}^3_+: x_1 \geq 1, x_2 = \delta, x_3 \geq 1\} \cup \{\bm{x}\in \mathbb{R}^3_+: x_1 \geq 1, x_2 \leq \delta, x_3 = 1\}.\]
For $C=D$, $R_D = (1,\infty]^d$, and $B_D =\{\bm{x}: \min(\bm{x}) = 1\}$.
\begin{prop}
\label{prop:taug}
Assume the conditions of Proposition~\ref{prop:tau}. For any $C \subseteq D$,
\[\tau_C(\delta) = \left[\min_{\bm{y} \in B_{C,\delta}}g(\bm{y})\right]^{-1} =\left[\min_{\bm{y} \in B_{C,\delta}^1}g(\bm{y})\right]^{-1}. \]
\end{prop}
\begin{proof}
The vertex of the region $R_{C,\delta}$, or its boundary $B_{C,\delta}$, which has components 1 on the coordinates indexed by $C$, and $\delta$ in the other coordinates, lies on $\mathcal{S}_{\vee} :=\{\bm{x} \in \mathbb{R}^d_+: \max(\bm{x}) = 1\}$. The region $G \subseteq \mathcal{S}_{\vee}$, and because the coordinatewise supremum of $G$ is $\bm{1}$, the boundary of $G$ intersects with $\mathcal{S}_{\vee}$. Now consider scaling the region $R_{C,\delta}$ by $r_{C,\delta} \in (0,1]$ until it intersects with $G$. The point of intersection must lie on the boundary of the scaled region $r_{C,\delta}R_{C,\delta}$, i.e., on $r_{C,\delta}B_{C,\delta}$, and on the boundary of $G$, $\{\bm{x} \in \mathbb{R}^d_+: g(\bm{x}) = 1\}$. Therefore, there exists $\bm{x}^\star \in B_{C,\delta}$ such that $g(r_{C,\delta}\bm{x}^\star) = 1$, which is rearranged to give $\tau_C(\delta) = r_{C,\delta} = 1/ g(\bm{x}^\star)$. Furthermore, we must have that such a point $\bm{x}^\star = \arg\min_{\bm{y} \in B_{C,\delta}} g(\bm{y})$, otherwise there exists some $\bm{x}' \in B_{C,\delta}$ such that $g(\bm{x}')<g(\bm{x}^\star)$ and so $g(r_{C,\delta}\bm{x}') < 1$, meaning that $r_{C,\delta} \neq \min\{r \in (0,1]: r R_{C,\delta} \cap G = \emptyset\}$. We conclude that $\bm{x}^\star = \arg\min_{\bm{y} \in B_{C,\delta}} g(\bm{y})$, so $\tau_{C}(\delta) = 1/\min_{\bm{y} \in B_{C,\delta}} g(\bm{y})$.
To show that $\arg\min_{\bm{y} \in B_{C,\delta}} g(\bm{y}) \in B_{C,\delta}^1$, let $\bar{\bm{x}} = \arg\min_{\bm{y} \in B_{C,\delta}^1} g(\bm{y})$, $\tilde{\bm{x}} = \arg\min_{\bm{y} \in B_{C,\delta}^\delta} g(\bm{y})$, and let $\tilde{x}_{l} = \min_{k \in C}\tilde{x}_k \geq 1$. Then $g(\tilde{\bm{x}}) = \tilde{x}_l g(\tilde{\bm{x}}/\tilde{x}_l)$, but $\tilde{\bm{x}}/\tilde{x}_l \in B_{C,\delta}^1$, so $g(\tilde{\bm{x}}/\tilde{x}_l) \geq g(\bar{\bm{x}})$ and hence $\tilde{x}_{l}g(\tilde{\bm{x}}/\tilde{x}_l) \geq g(\bar{\bm{x}})$ as $\tilde{x}_l \geq 1$.
\end{proof}
When $d=2$ and $|C|=1$, we note that $B_{\{j\},\delta}^1 = \{\bm{x}:x_j = 1, x_i \leq \delta\}$, which gives the equality in Proposition~\ref{prop:tau2dg}.
\begin{proof}[Proof of Proposition~\ref{prop:lambdag}]
The proof follows exactly as for the first equality in Proposition~\ref{prop:taug}, replacing $R_{C,\delta}, B_{C,\delta}$ and $r_{C,\delta}$ with $R_{\bm{\omega}}, B_{\bm{\omega}}$ and $r_{ \bm{\omega}}$.
\end{proof}
\subsubsection{Estimation of coefficients $\tau_C(\delta)$}
\label{sec:tauest}
When $C=D$, equation~\eqref{eq:SWT2} yields ${\mathbb P}(\min_{i \in D} X_{P,i}> t) \in \mathrm{RV}^{\infty}_{-1/\tau_{D}}$, implying that $\tau_D$ can be estimated as the reciprocal of the tail index of the so-called structure variable $\min_{i \in D} X_{P,i}$. This is identical to estimating the residual tail dependence coefficient $\eta_D$, for which the Hill estimator is commonly employed. However, for $C$ with $|C|<d$, we assume ${\mathbb P}(\min_{i \in C} X_{P,i}> t , \max_{j \in D \setminus C} X_{P,j}< t^{\delta}) \in \mathrm{RV}^{\infty}_{-1/\tau_{C}(\delta)}$, but this representation does not lend itself immediately to an estimation strategy, as there is no longer a simple structure variable for which $1/\tau_{C}(\delta)$ is the tail index.
In order to allow estimation, \citet{Simpsonetal18} considered ${\mathbb P}(\min_{i \in C} X_{P,i}> t , \max_{j \in D \setminus C} X_{P,j}< (\min_{i \in C} X_{P,i})^{\delta})$, but they only offered empirical evidence that the assumed index of regular variation for this probability was the same as in equation~\eqref{eq:SWT2}. We now prove this to be the case.
Define $R_{C,\delta}^x = \{\bm{x}\in \mathbb{R}^d_+: x_i>1, i \in C,~x_j \leq \delta \min_{l \in C} x_l, j \in D\setminus C\}$, and $B_{C,\delta}^x = B_{C,\delta}^{x,1} \cup B_{C,\delta}^{x,\delta}$ to be its boundary, where
\begin{align*}
B_{C,\delta}^{x,1}= B_{C,\delta}^1 &= \bigcup_{i \in C} \{\bm{x}\in\mathbb{R}^d_+: x_i = 1, x_j \geq 1~ \forall j \in C \setminus i,~ x_k \leq \delta \min_{l \in C} x_l = \delta~ \forall k \in D \setminus C \} \\
B_{C,\delta}^{x,\delta} &= \bigcup_{i \in D \setminus C} \{\bm{x}\in\mathbb{R}^d_+: x_i = \delta\min_{l \in C} x_l,~ x_j \leq \delta\min_{l \in C} x_l~ \forall j \in (D\setminus C) \setminus i,~ x_k \geq 1~ \forall k \in C \},
\end{align*}
and we specifically note the equality $ B_{C,\delta}^{x,1} = B_{C,\delta}^{1}$.
\begin{prop} Assume the conditions of Proposition~\ref{prop:tau}. If ${\mathbb P}(\min_{i \in C} X_{P,i}> t , \max_{j \in D \setminus C} X_{P,j}< t^{\delta}) \in \mathrm{RV}^{\infty}_{-1/\tau_{C}(\delta)}$, and ${\mathbb P}(\min_{i \in C} X_{P,i}> t , \max_{j \in D \setminus C} X_{P,j}< (\min_{i \in C} X_{P,i})^{\delta}) \in \mathrm{RV}_{-1/\tilde{\tau}_C(\delta)}^\infty,$ then $\tilde{\tau}_C(\delta) = \tau_C(\delta)$.
\end{prop}
\begin{proof}
Define $r_{C,\delta}^x = \min\left\{r \in [0,1] : r R_{C,\delta}^x \cap G =\emptyset \right\}$, where for $r>0$, $rR_{C,\delta}^x = \{\bm{x}\in \mathbb{R}^d_+: x_i>r, i \in C,~x_j \leq \delta \min_{l \in C} x_l, j \in D\setminus C\}$. Similarly to Propositions~\ref{prop:lambda} and~\ref{prop:tau},
\begin{align*}
-\log {\mathbb P}(\bm{X}_E \in tr_{C,\delta}^x R_{C,\delta}^x) = -\log {\mathbb P}(\min_{i \in C} X_{E,i}>tr_{C,\delta}^x, \max_{j \in D \setminus C} X_{E,j}< \delta \min_{i \in C} X_{E,i}) \sim t,
\end{align*}
and we conclude $\tilde{\tau}_C(\delta) = r_{C,\delta}^{x}$. As in Proposition~\ref{prop:taug}, we have $r_{C,\delta}^{x}=\min_{\bm{y} \in B_{C,\delta}^x} g(\bm{y})$. Noting again that $\arg\min_{\bm{y} \in B_{C,\delta}^x} g(\bm{y}) \in B_{C,\delta}^{x,1}=B_{C,\delta}^1$ shows that $r_{C,\delta}^{x} = r_{C,\delta} = \tau_{C}(\delta)$.
\end{proof}
\section{Examples}
\label{sec:examples}
We illustrate several of the findings of Section~\ref{sec:connections} with some concrete examples. In Section~\ref{sec:egd2} we focus on the intuitive and geometrically simple case $d=2$; in Section~\ref{sec:egd3}, we examine some three-dimensional examples for which visualization is still possible but more intricate. \iftoggle{arxiv}{}{Additional examples are given in the arXiv version of this article.}
Proposition~\ref{prop:logdens} implies that on $\mathbb{R}^d_+$, the same limit set $G$ as in exponential margins will arise for any marginal choice with $\psi_j(x) \sim x$, $x \to \infty$, provided $e^{-\psi_j(x)}$ is a von Mises function. In some of the examples below, it is convenient to establish a limit set and its gauge function using this observation rather than transforming to exactly exponential margins.
Models with convenient dependence properties are often constructed through judicious combinations of random vectors with known dependence structures; see, for example, \citet{Engelkeetal18} for a detailed study of so-called random scale or random location constructions. In Section~\ref{sec:mixing}, we use our results to elucidate the shape of the limit set when independent exponential-tailed variables are mixed additively. The spatial dependence model of \citet{HuserWadsworth18} provides a case study.
\subsection{Examples and illustrations for $d=2$}
\label{sec:egd2}
All of the examples considered in this section are symmetric, so, for the conditional extremes representation and coefficients $\tau_j(\delta)$, we only consider one case, omitting the subscript on the quantities~$\alpha_j$ and~$\beta_j$. Table~\ref{tab:bivariate} summarizes the dependence information from various bivariate distributions described in \iftoggle{arxiv}{Sections~\ref{exGpos}--\ref{se:eghr}.}{Sections~\ref{exGpos}--\ref{sec:egievd}, as well as the arXiv version.}
\afterpage{%
\clearpag
\begin{landscape
\centering
\begin{tabular}[htbp]{llllll}
\hline
Copula & $g(x,y)$ & $\lambda(\omega,1-\omega)$ & $\eta_{1,2}$ & $\tau_{1}(\delta)$ & $\alpha,\beta$\\
\hline
Gaussian $\rho\geq 0$ & $\frac{x + y -2\rho(xy)^{1/2}}{1-\rho^2}$ & $\begin{cases}
\max(\omega,1-\omega), & \frac{\min(\omega,1-\omega)}{\max(\omega,1-\omega)} \leq \rho^2\\
\frac{1 -2\rho(\omega(1-\omega))^{1/2}}{1-\rho^2}, & \frac{\min(\omega,1-\omega)}{\max(\omega,1-\omega)} \geq \rho^2
\end{cases}$ & $\frac{1+\rho}{2}$ & $\begin{cases}1 & \delta \geq \rho^2\\ \frac{1-\rho^2}{1+\delta-2\rho\delta^{1/2}} & \delta < \rho^2 \end{cases}$ & $\begin{array}{l} \alpha=\rho^2 \\ \beta=1/2\end{array}$ \\\hline
Logistic GP & $\frac{1}{\theta}\max(x,y) + (1-\frac{1}{\theta})\min(x,y)$ & $\max(\omega,1-\omega)$ & 1 & $[\theta^{-1}+(1-\theta^{-1}\delta)]^{-1}$ & $\begin{array}{l} \alpha=1 \\ \beta=0\end{array}$\\\hline
\shortstack{Inverted\\ logistic}& $(x^{1/\theta}+y^{1/\theta})^\theta $ & $g(\omega,1-\omega)$ & $2^{-\theta}$ & $1$ & $\begin{array}{l} \alpha=0 \\ \beta=1-\theta\end{array}$\\\hline
\shortstack{H\"{u}sler--Reiss\\ GP} & $\begin{cases}
\infty, & x \neq y\\
x, & x=y
\end{cases}$ & $\max(\omega,1-\omega)$ & 1 & $\begin{cases}
0, & \delta<1\\
1, & \delta=1
\end{cases}$ & $\begin{array}{l} \alpha=1 \\ \beta \mbox{ undetermined}\end{array}$\\\hline
\shortstack{Inverted\\ H\"{u}sler--Reiss}& $ x\Phi\left(\frac{\lambda}{2} + \frac{\log\left(\frac{x}{y}\right)}{\lambda}\right) + y\Phi\left(\frac{\lambda}{2} + \frac{\log\left(\frac{y}{x}\right)}{\lambda}\right) $ & $g(\omega,1-\omega)$ & $\left[2\Phi\left(\frac{\lambda}{2}\right)\right]^{-1} $ & $1$ & $\begin{array}{l} \alpha=0 \\ \beta=1\end{array}$\\\hline
\end{tabular}
\captionof{table}{Summary of dependence measures across a range of bivariate examples}
\label{tab:bivariate}
\end{landscape}
\clearpage
}
\subsubsection{Meta-Gaussian distribution: nonnegative correlation}\label{exGpos}
Starting with a Gaussian bivariate random vector and transforming its margins to standard exponential, we obtain a meta-Gaussian distribution with exponential margins. Such a distribution inherits the copula of the Gaussian distribution. For simplicity, we consider the case where the underlying Gaussian random vector has standard normal components with correlation $\rho$.
Then, for $\rho\ge0$, the joint probability density $f_E$ satisfies:
\begin{align*}
-\log f_E(tx,ty)/t = (x + y -2\rho(xy)^{1/2})/(1-\rho^2) + O(\log t/t),\quad t\to\infty,\quad x,y\ge 0,
\end{align*}
so that $g(x,y) = (x + y -2\rho(xy)^{1/2})/(1-\rho^2)$. The convergence in \eqref{q1} holds on $[0,\infty)^d$ and hence the limit set exists and is given by $\{{\bm x}\in[0,\infty)^d: g({\bm x})\le1\}$. This is example~(ii) in Figure~\ref{fig:2dgauge}.
\paragraph{Conditional extremes:} Setting $g(\alpha,1) = 1$ leads to $(\alpha^{1/2} - \rho)^2 = 0$, i.e., $\alpha = \rho^2$. For $\beta$ we have $g(\rho^2+u, 1) - 1 = u^2/\{2\rho(1-\rho^2)\} + O(u^3) \in \mathrm{RV}_2^0$, hence $\beta = 1/2$.
\paragraph{Function $\lambda(\bm{\omega})$:} By Proposition~\ref{prop:lambdag}, we need to find $1/r_{\omega} = \min_{\bm{y} \in B_{\omega}} g(x,y)$. If $\min(\omega,1-\omega)/\max(\omega,1-\omega) \leq \rho^2$, then $\min_{\bm{y} \in B_{\omega}} g(x,y) = 1$, with the minima occuring at the points $(1,\rho^2)$, $(\rho^2,1)$. Otherwise, if $\min(\omega,1-\omega)/\max(\omega,1-\omega) \geq \rho^2$, then $\min_{\bm{y} \in B_{\omega}} g(x,y) = g(1, \min(\omega,1-\omega)/\max(\omega,1-\omega))$. Putting this together with Proposition~\ref{prop:lambda}, we find
\[
\lambda(\omega,1-\omega) = \begin{cases}
\max(\omega,1-\omega), & \min(\omega,1-\omega)/\max(\omega,1-\omega) \leq \rho^2\\
g(\omega,1-\omega) = \frac{1 -2\rho(\omega(1-\omega))^{1/2}}{1-\rho^2}, & \min(\omega,1-\omega)/\max(\omega,1-\omega) \geq \rho^2.
\end{cases}
\]
This is the same form as given in \citet{WadsworthTawn13}. We therefore have $\eta_{1,2} = [2g(1/2,1/2)]^{-1} = g(1,1)^{-1} = (1+\rho)/2$.
\paragraph{Coefficients $\tau_j(\delta)$:} From Proposition~\ref{prop:tau2dg}, we have $\tau_1(\delta) = [\min_{\gamma \in [0,\delta]} g(1,\gamma)]^{-1} = [g(1,\min(\delta,\rho^2))]^{-1}$. Therefore, $\tau_{1}(\delta) = 1$ if $\delta \geq \rho^2$, else $\tau_{1}(\delta) = (1-\rho^2)/(1+\delta-2\rho\delta^{1/2}) < 1$. Note that these values are very laborious to calculate via Gaussian survival functions, and they were not given in \citet{Simpsonetal18}.
\subsubsection{Meta-Gaussian distribution: negative correlation}\label{exGneg}
When $\rho<0$, Proposition~\ref{prop:logdens} cannot be applied as the continuous convergence condition~\eqref{q1} does not hold along the axes. Hence, we only gain a partial specification, when $x>0, y>0$, through this route. Instead, here we can apply Proposition~\ref{prop:DMR} since the limit function $g$ in~\eqref{eq:DMR} satisfies the monotonicity condition given immediately thereafter. This limit function is given by
\begin{align*}
g(x,y) = \begin{cases}
(x + y -2\rho(xy)^{1/2})/(1-\rho^2), & x>0,~ y>0,\\
x, & y=0,\\
y, & x=0.
\end{cases}
\end{align*}
\begin{figure}[]
\includegraphics[width=1\textwidth]{Fig/plotGauss3.pdf}
\caption{Top row: Limit sets $\{(x,y): g(x,y)\le1\}$ for a bivariate meta-Gaussian distribution with exponential margins. Bottom row: corresponding plots of function $g(x,y)$ for a fixed value of $y$.}
\label{fig:2dgauss}
\end{figure}
Figure~\ref{fig:2dgauss} illustrates the limit sets $G$ for the three cases $\rho>0,\rho=0$ and $\rho<0$. In the latter case, large values of one variable tend to occur with small values of the other, which causes the limit set to include lines along the axes and the function $g$ is not continuous. Such difficulties can be alleviated by consideration of Laplace margins for distributions displaying negative dependence, which is discussed further in Section~\ref{sec:Discussion}.
\subsubsection{Logistic generalized Pareto copula}
\label{sec:eglog}
The logistic generalized Pareto distribution with conditionally exponential margins (${\mathbb P}(X_{\tilde{E}}>x) = {\mathbb P}(X_{\tilde{E}}>0)e^{-x}$, $x>0$) and dependence parameter $\theta \in(0,1)$ satisfies
\begin{align*}
f_{\tilde{E}}(x,y) &= \theta^{-1}2^{-\theta} e^{-(x+y)/\theta}(e^{-x/\theta}+e^{-y/\theta})^{\theta-2}\\
-\log f_{\tilde{E}}(tx,ty)/t &= \theta^{-1}\max(x,y) + (1-\theta^{-1})\min(x,y) + O(1/t),
\end{align*}
so the gauge function is $g(x,y) = \theta^{-1}\max(x,y) + (1-\theta^{-1})\min(x,y)$. This form of gauge function is found throughout several symmetric asymptotically dependent examples, such as those distributions whose spectral measure $H$ places no mass on 0 and 1 and possess densities that are regularly varying at the endpoints 0, 1 such that $\mathrm{d}H(w)/\mathrm{d}w \in \mathrm{RV}_{1/\theta-2}^{0}$, $-\mathrm{d}H(1-w)/\mathrm{d}w \in \mathrm{RV}_{1/\theta-2}^{0}$. This is example~(i) in Figure~\ref{fig:2dgauge}.
\paragraph{Conditional extremes:} Solving for $g(\alpha,1)=1$, we obtain $\alpha=1$, whilst $g(1+u,1) -1 = u/\theta \in \mathrm{RV}_1^0$, hence $\beta=0$.
\paragraph{Function $\lambda(\bm{\omega})$:} We have that $\arg\min_{\bm{y} \in B_\omega}g(\bm{y}) = (1,1)$, so $r_{ \omega} = 1$ and $\lambda(\omega,1-\omega) = \max(\omega,1-\omega)$. Therefore $\eta_{1,2} =1$.
\paragraph{Coefficients $\tau_j(\delta)$:} $\tau_1(\delta) = [\min_{\gamma \in [0,\delta]} g(1,\gamma)]^{-1} = [g(1,\delta)]^{-1} = [\theta^{-1}+(1-\theta^{-1}\delta)]^{-1}$. This matches the value calculated in the Supplementary Material of \citet{Simpsonetal18}.
\subsubsection{Inverted extreme value distribution}
\label{sec:egievd}
The inverted extreme value copula is the joint lower tail of an extreme value copula, translated to be the joint upper tail. That is, if $(U_1,U_2)$ have an extreme value copula with uniform margins, then $(1-U_1,1-U_2)$ have an inverted extreme value copula. In two dimensions, its density in exponential margins may be expressed as
\[
f_E(x,y) = \{l_{1}(x,y)l_2(x,y)-l_{12}(x,y)\}\exp\{-l(x,y)\},
\]
where $l(\bm{x}) = V(1/\bm{x})$, for $V$ the exponent function in~\eqref{eq:V}, is the 1-homogeneous stable tail dependence function \citep[e.g.,][Ch.8]{Beirlantetal04} of the corresponding extreme value distribution, and $l_1(x,y) = \partial l(x,y) /\partial x$, etc. We thus have
\[
-\log f_E(tx,ty)/t = l(x,y) + O(1/t),\qquad t \to\infty,
\]
so $g(x,y) = l(x,y)$.
\paragraph{Conditional extremes:} Stable tail dependence functions always satisfy $l(x,0) = x, l(0,y) = y$ and so $g(1,0) = g(0,1) = 1$. Hence, if $\alpha=0$ is the only solution to $g(\alpha,1) =1$, then $a(x)/x \sim 0$. An example of this is given by the inverted extreme value logistic copula, whereby $l(x,y) = (x^{1/\theta}+y^{1/\theta})^\theta, \theta \in (0,1]$. This is example~(iii) of Figure~\ref{fig:2dgauge}, for which we have $\alpha=0$ and $\beta=1-\theta$.
\iftoggle{arxiv}{\citet{PapastathopoulosTawn16} study conditional extreme value limits for general classes of inverted extreme value distributions. One case they consider is where the spectral measure defining the distribution in two dimensions has support on a sub-interval $[w_l,w_u] \subset [0,1]$. We give a simple example to illustrate their findings in this context. Let $w_l=1-w_u = w^\star \in [0,1/2)$, and $H(w) = (w-w^\star)/(1-2w^\star), w \in [w^\star,1-w^\star]$. Then
\begin{align*}
l(x,y) &= \int_{w^\star}^{1-w^\star} \max(wx,(1-w)y)\mathrm{d}H(w) \\&= \begin{cases}
x, & y< x w^\star/(1-w^\star)\\
\frac{x}{1-2w^\star}\left\{(1-w^\star)^2-\left(\frac{y}{x+y}\right)^2\right\}+\frac{y}{1-2w^\star}\left\{(1-w^\star)^2-\left(\frac{x}{x+y}\right)^2\right\}, & x w^\star/(1-w^\star) \leq y \leq x (1-w^\star)/w^\star\\
y,& y> x (1-w^\star)/w^\star.
\end{cases}
\end{align*}
In this case, $g(\alpha,1) = 1$ for all $\alpha \in [0,w^\star/(1-w^\star)]$, so by Proposition~\ref{prop:CEgauge}~(iii), $\alpha=w^\star/(1-w^\star)$. Following a Taylor expansion in which linear terms in $u$ vanish, we have $g(w^\star/(1-w^\star) + u, 1) - 1 \in \mathrm{RV}^0_{2}$, so $\beta=1/2$. \citet{PapastathopoulosTawn16} show that for this distribution, $a(x) = x w^\star/(1-w^\star)$ and $b(x)=x^{1/2}$.
A further interesting example studied by \citet{PapastathopoulosTawn16} is that of the inverted H\"{u}sler--Reiss distribution, for which
\[
g(x,y) = x\Phi(\lambda/2 + \log(x/y)/\lambda) + y\Phi(\lambda/2 + \log(y/x)/\lambda), \qquad \lambda>0.
\]
The unique solution to $g(\alpha,1) = 1$ is $\alpha=0$ (obtained as a limit) and $g(u,1)-1$ is rapidly varying, i.e., $g(u,1)-1 \in \mathrm{RV}_{\infty}^0$, corresponding to $\beta=1$. The forms of the normalization functions detailed in \citet{PapastathopoulosTawn16} are
\begin{align*}
a(x) = x \exp\left\{-\lambda(2\log x)^{1/2} +\frac{\lambda\log\log x}{(2\log x)^{1/2}}+\lambda^2/2\right\},\qquad b(x) = a(x)/(\log x)^{1/2},
\end{align*}
for which $a(x)/x \to 0$ and $b(x) \in \mathrm{RV}_1^{\infty}$.}{Further examples in this class are given in the arXiv version.}
\paragraph{Function $\lambda(\bm{\omega})$:} Since $g(x,y) = l(x,y)$, and $l$ is a convex function satisfying $l(x,0)=x,l(0,y) = y$, $\arg\min_{\bm{y} \in B_{\omega}} g(\bm{y}) = (\omega,1-\omega)/\max(\omega,1-\omega)$. Hence, $\lambda(\omega,1-\omega) = g(\omega,1-\omega)$ in this case.
\paragraph{Coefficients $\tau_j(\delta)$:} Since $g(1,0)=1$, we have $\tau_1(\delta) = 1$ for all $\delta \in [0,1]$.
\iftoggle{arxiv}{
\subsubsection{H\"{u}sler--Reiss generalized Pareto copula}
\label{se:eghr}
The bivariate H\"{u}sler--Reiss generalized Pareto distribution with conditionally exponential margins has density
\[
f_{\tilde{E}}(x,y) ~~\propto~~ \exp\left\{-\frac{1}{2}\left(\frac{(x-y)^2}{2(1-\rho)}+x+y\right)\right\},
\]
from which it can be seen that
\[
-\log f_{\tilde{E}}(t\bm{x})/t \to \begin{cases}
\infty, & x \neq y,\\
x, & x=y.
\end{cases}
\]
While Proposition 2 cannot be applied here due to the lack of uniform convergence, the form of the limit set is nonetheless $G=\{(x,y):x=y \leq 1\}$, which is the same limit set as arises under perfect dependence. This can be explained by the construction of the H\"{u}sler--Reiss model, which has a dependence structure asymptotically equivalent to that of $E+(Z_1,Z_2)$, where $E \sim \mbox{Exp}(1)$ is independent of $(Z_1,Z_2) \sim N_2(\bm{0},\Sigma)$. The marginal distributions of such a construction satisfy $\psi(x) \sim x$, $x \to \infty$ so that with the sample cloud constructed by taking $r_n = \log n$, the contribution from the Gaussian component converges in probability to zero, leaving only the contribution from the common term $E$.
\paragraph{Conditional extremes:} We have $g(\alpha,1)=1$ for $\alpha=1$, but cannot use Proposition~\ref{prop:CEgauge} to determine $\beta$ since $g(1+u,1) -1 \not\in \mathrm{RV}^0$.
\paragraph{Function $\lambda(\bm{\omega})$:} The quantity $r_{\omega}=1$ and so $\lambda(\omega,1-\omega)=\max(\omega,1-\omega)$, and $\eta_{1,2}=1$.
\paragraph{Coefficients $\tau_j(\delta)$:} $\tau_1(\delta)=[\min_{\gamma \in [0,\delta]} g(1,\gamma)]^{-1} = 0$ for $\delta<1$ and $=1$ for $\delta=1$. This implies that ${\mathbb P}(X_P>t,Y_P \leq t^{\delta}) \in \mathrm{RV}^\infty_{-\infty}$, i.e., is rapidly varying.
}{}
\subsubsection{Density defined by $g$}
\label{sec:egg}
If $g:\mathbb{R}^d_+ \to \mathbb{R}_+$ is a gauge function describing a limit set $G$, then $f(\bm{x}) = e^{-g(\bm{x})}/(d! |G|)$ is a density \citep[see][]{BalkemaNolde10}. In general, except for the case of $g(\bm{x}) = \sum_{i=1}^d x_i$, the margins are not exactly exponential, and may be heavier than exponential, for example in the case $g(\bm{x}) = \max_{1\leq i \leq d}(x_i)$.
We consider the density defined by $g(x,y) = \max\{(x-y)/\theta,(y-x)/\theta,(x+y)/(2-\theta)\}$, $\theta \in (0,1]$: this is example~(vi) in Figure~\ref{fig:2dgauge}, and illustrated in Figure~\ref{fig:lambdatau}. The marginal density is given by
\[
[2 e^{-x} - \theta e^{-x/\theta} -2(1-\theta)e^{-x/(1-\theta)}]/[4\theta - 3\theta^2].
\]
\paragraph{Conditional extremes:} Solving for $g(\alpha,1)=1$, we obtain $\alpha=1-\theta$, whilst $g(1-\theta+u,1) -1 = u/(2-\theta) \in \mathrm{RV}_1^0$, hence $\beta=0$.
\paragraph{Function $\lambda(\bm{\omega})$:} If $\min(\omega,1-\omega)/\max(\omega,1-\omega) \leq 1-\theta$, then $\arg\min_{\bm{y} \in B_\omega} = (1,1-\theta)$, or $(1-\theta,1)$ and $r_{\omega}=1$; otherwise, $\arg\min_{\bm{y} \in B_\omega} = (1,\omega/(1-\omega))$ or $((1-\omega)/\omega, 1)$, and $r_{\omega} =\{1+ \min(\omega,1-\omega)/\max(\omega,1-\omega)/(2-\theta)\}$. As such
\[
\lambda(\omega,1-\omega) = \begin{cases}
\max(\omega,1-\omega), & \min(\omega,1-\omega)/\max(\omega,1-\omega) \leq 1-\theta\\
g(\omega,1-\omega) = \frac{1}{2-\theta}, & \min(\omega,1-\omega)/\max(\omega,1-\omega) \geq 1-\theta,
\end{cases}
\]
and the residual tail dependence coefficient $\eta_{1,2} = 1-\theta/2$.
\paragraph{Coefficients $\tau_j(\delta)$:} $\tau_1(\delta) = [\min_{\gamma \in [0,\delta]} g(1,\gamma)]^{-1} = [g(1,\min(\delta,1-\theta))]^{-1}$. Therefore, $\tau_{1}(\delta) = 1$ if $\delta \geq 1-\theta$, else $\tau_{1}(\delta) = \theta/(1-\delta)< 1$.
\iftoggle{arxiv}{
\subsubsection{Boundary case between asymptotic dependence and independence}
We give two examples of distributions whose limit set is described by the gauge function $g(x,y)=\max(x,y)$. The first of these displays asymptotic dependence, i.e., $\nu(\mathbb{E}_{1,2})>0$, for $\nu$ as in Section~\ref{sec:MRVintro}, while the second displays asymptotic independence. In both cases, $\lambda(\bm{\omega})=\max(\omega,1-\omega), \eta_{1,2}=1$, and $\tau_{1}(\delta)=\tau_2(\delta)=1$ for all $\delta \in (0,1)$. The conditional extremes convergences are however rather different, and need to be derived carefully as some of the hypotheses in Proposition~\ref{prop:CEgauge} fail.
\paragraph{Example 1: asymptotic dependence}
We consider a particular instance of a bivariate extreme value distribution with spectral measure density $\mathrm{d}H(w)/\mathrm{d}w = h(w) = C w^{-1}(1-w)^{-1}\exp\{-(-\log w)^{1/2}-(-\log(1-w))^{1/2}\}$. The density is regularly varying at 0 and 1 with index $-1$, but its integral is finite. For the exponent function $V$ as in~\eqref{eq:V}, $h(w)=-V_{12}(w,1-w)$ with $V_{12}(x,y)=\partial^2 V(x,y)/(\partial x\partial y)$, so
\[
h\left(\frac{x}{x+y}\right) = -V_{12}\left(\frac{x}{x+y}, \frac{y}{x+y}\right) = -(x+y)^3 V_{12}(x,y),
\]
and therefore
\[
-V_{12}(x,y) = \frac{C}{(x+y)xy}\exp\{-(-\log x + \log(x+y))^{1/2}-(-\log y + \log(x+y))^{1/2}\}.
\]
For the corresponding multivariate max-stable or generalized Pareto distribution with Gumbel/exponential margins, the gauge function is determined by
\[
f_{\tilde{E}}(x,y) ~~\propto~~ e^{x+y}V_{12}(e^x,e^y) ~~\propto~~ \frac{1}{e^x+e^y} \exp\{-(- x + \log(e^x+e^y))^{1/2}-(-y + \log(e^x+e^y))^{1/2}\},
\]
and $-\log f_{\tilde{E}}(tx,ty) \sim t\max(x,y)$. That is, $g(x,y) = \max(x,y)$.
We have $g(\alpha,1)=1$ for all $\alpha \in [0,1]$, so $\alpha=1$ is the maximum such value, and $g(1,1+u)-1=u \in \mathrm{RV}^0_1$, giving $\beta=0$. Taking $a^j(t) = t$, $b^j(t) = 1$, we see
\[
f_{\tilde{E}}(t,t+z)e^t ~~\propto~~ (1+e^z)^{-1} \exp\{-[\log(1+e^{-z})]^{1/2}\}, \qquad z \in \mathbb{R}.
\]
The limit is in fact exact for all $t$ here because of the multivariate generalized Pareto form.
\paragraph{Example 2: asymptotic independence}
Consider the density $f(x,y) = e^{-\max(x,y)}/2$, $x,y>0$. The marginal densities are $f(x)=(1+x)e^{-x}/2$, which is a mixture of $\rm{Gamma(1,1)}$ and $\rm{Gamma(2,1)}$ densities, and is heavier tailed than standard exponential. We firstly verify asymptotic independence by examining $\lim_{t \to \infty}{\mathbb P}(X>t|Y>t)$. The joint survival function ${\mathbb P}(X>x,Y>y)=\{2+\max(x,y)-\min(x,y)\}e^{-\max(x,y)}/2$, while ${\mathbb P}(X>x)=e^{-x}(1+x/2)$. Hence $\lim_{t \to \infty} {\mathbb P}(X>t|Y>t) = 0$.
Again, Proposition~\ref{prop:logdens} ensures that the limit set with gauge function $g(x,y)=\max(x,y)$ would also arise in exponential margins. However, we make the transformation explicitly here to study the conditional extremes convergence on the same scale as previously. The change to exponential margins entails $X_{E} = X-\log(1+X/2)$, leading to the density
\[
f_E(x,y) = \exp[-\max\{x + \log (1+x/2+O(\log x)), y + \log (1+y/2+O(\log y))\}][2^{-1}+O(x^{-1})+O(y^{-1})],
\]
from which we can also see $-\log f_E(tx,ty) \sim t \max(x,y)$.
If we were to suppose that a conditional extremes limit with support including $(0,\infty)$ exists, then Proposition~\ref{prop:CEgauge}~(i) and~(iii) would mean $\alpha_1=\alpha_2=1$, while $g(1+u,1)-1 =u \in \mathrm{RV}^0_1$, so that Proposition~\ref{prop:CEgauge}~(ii) would give $\beta_1=\beta_2=0$. However, for positive $z$, consideration of $b(t)f_E(t,a(t)+b(t)z)e^t$ yields no possibilities for a non-degenerate limit. Nonetheless, for $z<0$ we can take $a(t)=b(t)=t$ leading to
\[
tf_E(t,t+tz)e^t \sim (1+t/2)^{-1}t/2 \to 1, \qquad t \to \infty, \qquad z \in (-1,0),
\]
i.e., a uniform limit distribution on $(-1,0)$.
We comment that the results of Proposition~\ref{prop:CEgauge} do indeed focus predominantly on the positive end of the support for limit distributions, but most known examples of conditional limits have support including $(0,\infty)$. A natural next step is to consider the implications relating to negative support. We particularly note the possibility that the order of regular variation of the two functions $g(1,\alpha_1+u)-1 \in \mathrm{RV}^0_{1/(1-\beta_1^+)}$ and $g(1,\alpha_1-u)-1\in \mathrm{RV}^0_{1/(1-\beta_1^-)}$ need not be equal, though for each of our examples where both functions are regularly varying, $\beta^+=\beta^-$. If $\beta^+>\beta^-$, it seems likely that a limit distribution with positive support only would arise, and vice versa when $\beta^+<\beta^-$.}{Further examples presented in the arXiv version include the H\"{u}sler--Reiss generalized Pareto copula, and a two boundary cases with $g(x,y)=\max(x,y)$, one displaying asymptotic dependence and one asymptotic independence. In the latter of these, we find that there is no conditional extreme value limit with positive support, but there is one with negative support. We comment that the results of Proposition~\ref{prop:CEgauge} do indeed focus predominantly on the positive end of the support for limit distributions, but most known examples of conditional limits have support including $(0,\infty)$. A natural next step is to consider the implications relating to negative support. We particularly note the possibility that the order of regular variation of the two functions $g(1,\alpha_1+u)-1 \in \mathrm{RV}^0_{1/(1-\beta_1^+)}$ and $g(1,\alpha_1-u)-1\in \mathrm{RV}^0_{1/(1-\beta_1^-)}$ need not be equal, though for each of our examples where both functions are regularly varying, $\beta^+=\beta^-$. If $\beta^+>\beta^-$, it seems likely that a limit distribution with positive support only would arise, and vice versa when $\beta^+<\beta^-$.}
\subsection{Examples and illustrations for $d=3$}
\label{sec:egd3}
In this section we give two examples, focusing on issues that arise for $d>2$.
\subsubsection{Gaussian copula}
The general form of the gauge function for a meta-Gaussian distribution with standard exponential margins and correlation matrix $\Sigma$ with non-negative entries is
\[
g(\bm{x}) = (\bm{x}^{1/2})^\top \Sigma^{-1} \bm{x}^{1/2}.
\]
Figure~\ref{fig:3dgausstau23} displays the level set $g(\bm{x}) = 1$ when the Gaussian correlations in $\Sigma$ are $\rho_{12}=0.75, \rho_{13}=0.25, \rho_{23} =0.4$. The red dots on the level set are the points $(1,1,\gamma)/g(1,1,\gamma)$, $(1,\gamma,1)/g(1,\gamma,1)$ and $(\gamma,1,1)/g(\gamma,1,1)$ for $\gamma \in [0,1]$. The figure also provides an illustration of $\tau_{2,3}(\delta)$ for $\delta =0.2$ and $\delta=0.8$: in each case the light blue line from the origin is $\gamma \times (\delta,1,1)$, $\gamma \in [0,1]$, whilst the pink lines trace out the boundary $B_{\{2,3\},\delta}$ and $\tau_{2,3}(\delta)B_{\{2,3\},\delta}$. We see that when $\delta = 0.2$ (left panel), $\tau_{2,3}(0.2) = 1/g(0.2,1,1)$, i.e.,\ $\min_{\bm{y} \in B_{\{2,3\},0.2}}g(\bm{y}) = g(0.2,1,1)$. However, when $\delta = 0.8$, $\min_{\bm{y} \in B_{\{2,3\},0.8}}g(\bm{y}) = g(\gamma^\star,1,1)$, for $\gamma^\star \in [0,0.8]$, so $\tau_{2,3}(0.8) = 1/g(\gamma^\star,1,1)$. We note that the same value of $\tau_{2,3}(\delta)$ applies for any $\delta \geq \gamma^\star$: for this example, when $\delta \geq \gamma^\star \approx 0.51$, $\tau_{2,3}(\delta) = 0.7 = \eta_{2,3}$.
The reason that $\tau_{2,3}(\delta) = \eta_{2,3}$ for sufficiently large $\delta$ is because in this case $\arg\min_{x_1} g(x_1,1,1) = \gamma^\star$, meaning that the two-dimensional marginalization $g_{\{2,3\}}(1,1) = g(\gamma^\star,1,1)$, and we further have that $g_{\{2,3\}}(1,1) = \min_{\bm{y} \in B_{2,3}} g_{\{2,3\}}(\bm{y})$, so $\eta_{2,3} = 1/g_{2,3}(1,1)$. In Section~\ref{sec:vine} we will illustrate a gauge function for which $\arg\min_{x_3} g(1,1,x_3)>1$, and consequently $\tau_{1,2}(\delta) < \eta_{1,2}$ for all $\delta\leq 1$.
The right panel of Figure~\ref{fig:3dgausstau23} illustrates $\tau_{1}(\delta)$ for $\delta = 0.2$ and $\delta =0.6$. When $\delta =0.6$, the boundary $B_{1,\delta}$ already touches $G$, and so $\tau_{1}(0.6)=1$. In this example, $\tau_1(\delta) =1$ for any $\delta \geq 0.5625 = \rho_{12}^2$. As such, $\tau_{1}(0.2)<1$ as illustrated in the figure. We comment that if we had marginalized over $X_2$, and were looking at $\tau_1(\delta)$ for the variables $(X_1,X_3)$, then we would have $\tau_1(\delta) = 1$ for any $\delta \geq 0.0625 = \rho_{13}^2$. This provides an illustration of the dimensionality of the problem interacting with $\tau_{C}(\delta)$, and is again related to the point at which the minimum point defining the lower-dimensional gauge function occurs.
\begin{figure}[]
\includegraphics[width=0.33\textwidth]{Fig/GaussTau23_d02_crop.png}
\includegraphics[width=0.33\textwidth]{Fig/GaussTau23_d08_crop.png}
\includegraphics[width=0.33\textwidth]{Fig/GaussTau1_d02_06_crop.png}
\caption{Level set $g(\bm{x})=1$ for a trivariate meta-Gaussian distribution with exponential margins. The left panel illustrates $\tau_{2,3}(0.2)$: the boundary set indicated by the pink lines is scaled along the blue trajectory until it touches $G$, which happens in this case at the corner point $(0.2,1,1)/g(0.2,1,1)$. The centre panel illustrates $\tau_{2,3}(0.8)$: the boundary set is again pulled back along the indicated trajectory until it touches $G$: in this case this does not occur at a corner point. The right panel illustrates $\tau_1(\delta)$ in a similar manner, for $\delta =0.2,0.6$.}
\label{fig:3dgausstau23}
\end{figure}
\subsubsection{Vine copula}
\label{sec:vine}
Three-dimensional vine copulas are specified by three bivariate copulas: two in the ``base layer'', giving the dependence between, e.g.,\ $X_1,X_2$ and $X_2,X_3$ and a further copula specifying the dependence between $X_1|X_2$ and $X_3|X_2$. Here we take the base copulas to be independence for $(X_1,X_2)$, and the inverted Clayton copula with parameter $\beta>0$ for $(X_2,X_3)$. The final copula is taken as inverted Clayton with parameter $\gamma>0$. The gauge function that arises in exponential margins is
\begin{align}
\label{eq:gvine}
g(\bm{x}) = (1+\beta)\max(x_2,x_3)-\beta \min(x_2,x_3)&-\gamma x_1-(\gamma+1)(\beta+1)(\max(x_2,x_3)-x_2) \notag \\&+ (2\gamma+1)\max(x_1,(\beta+1)(\max(x_2,x_3)-x_2)).
\end{align}
Figure~\ref{fig:3dvinetau12} displays the level set $g(\bm{x})=1$. In this figure we also give an illustration of a case where $\tau_C(1)<\eta_C$: in particular, for this example $\tau_{1,2}(1)<\eta_{1,2,3}=\eta_{1,2}=1/2$. The purple lines represent the boundary of the region $\tau_{1,2}(1)R_{\{1,2\},1} = \tau_{1,2}(1,\infty]^2 \times [0,1]$, while the green lines represent the boundary of the region $\eta_{1,2,3}(1,\infty]^3$. Theorem~1 of \citet{Simpsonetal18} tells us that $\eta_{1,2} = \max(\tau_{1,2}(1),\tau_{1,2,3})$, where $\tau_{1,2,3}=\eta_{1,2,3}$. Therefore $\tau_{1,2}(1)<\eta_{1,2}$ guarantees that $\eta_{1,2}=\eta_{1,2,3}$.
We also illustrate Proposition~\ref{prop:marg}, minimizing~\eqref{eq:gvine} over $x_3$. If $x_2>x_3$ then the minimum over $x_3$ occurs by setting $x_3=x_2$ and is equal to $x_2+(1+\gamma)x_1$. If $x_2< x_3$ then owing to the final term we need to consider the cases $x_3 \lessgtr x_1/(1+\beta)+x_2$. In both cases, the mimimum is attained at $x_3 = x_1/(1+\beta)+x_2$, and is equal to $x_1+x_2 < (1+\gamma)x_1+x_2$. As such, $\min_{x_3} g(x_1,x_2,x_3) = x_1+x_2$. This result is as expected since the bivariate margins of vine copulas that are directly specified in the base layer are equal to the specified copula: in this case, independence.
\begin{figure}[]
\includegraphics[width=0.9\textwidth]{Fig/VineTau13_d1.png}
\caption{Level set $g(\bm{x})=1$ with $g$ as in~\eqref{eq:gvine}. The figure illustrates $\tau_{1,2}(1)$ and $\eta_{1,2,3} = \eta_{1,2} = 1/2$.}
\label{fig:3dvinetau12}
\end{figure}
\subsection{Mixing independent vectors}
\label{sec:mixing}
Here we exploit the results from previous sections to consider what happens when independent exponential random vectors are additively mixed such that the resulting vector still has exponential type tails. We consider as a case study the spatial model of \citet{HuserWadsworth18}, which following a reparameterization can be expressed
\begin{align}
\{X_{\tilde{E}}(s) = \gamma S_E + V_E(s): s \in \mathcal{S}\subset \mathbb{R}^2\}, \qquad \gamma \in (0,\infty), \label{eq:HWspat}
\end{align}
where $S_E \sim \mbox{Exp}(1)$ is independent of the spatial process $V_{E}$, which also possesses unit exponential margins and is asymptotically independent at all spatial lags $s_1-s_2 \neq 0$. The process $V_E$ is assumed to possess hidden regular variation, with residual tail dependence coefficient satisfying $\eta^{V}(s_1,s_2)<1$ for all $s_1 \neq s_2$. The resulting process $X_{\tilde{E}}$ is asymptotically independent for $\gamma \in (0,1]$ and asymptotically dependent for $\gamma>1$; see also \citet{Engelkeetal18} for related results.
When $\gamma <1$, ${\mathbb P}(X_{\tilde{E}}(s) > x) \sim e^{-x}/(1-\gamma)$. In this case, \citet{HuserWadsworth18} show that the residual tail dependence coefficient for the process $X_{\tilde{E}}$ is given by
\begin{align}
\eta^{X} = \begin{cases}
\eta^{V}, & \gamma < \eta^V\\
\gamma, & \eta^V \leq \gamma \leq 1.
\end{cases} \label{eq:etaHW}
\end{align}
That is, the strength of the extremal dependence as measured by the residual tail dependence coefficient $\eta^X$ is increasing in $\gamma$ for $\gamma \geq \eta^V$. In contrast, \citet{WadsworthTawn19} showed that under mild conditions, the process~\eqref{eq:HWspat} has the same conditional extremes normalization as the process $V_E(s)$, with identical limit distribution when the scale normalizations $b_{s-s_0}(t) \to \infty$ as $t \to \infty$. Here, the subscript $s-s_0$ alludes to the fact that the conditioning event in~\eqref{eq:ce} is $\{V_{E}(s_0)>t\}$ and we study the normalization at some other arbitrary location $s \in \mathcal{S}$. In combination, we see that the results of \citet{HuserWadsworth18} and \citet{WadsworthTawn19} suggest that the addition of the variable $\gamma S_E$ to $V_E$ affects the extremal dependence of $X_{\tilde{E}}$ differently for different extreme value representations. We elucidate these results further in the context of the limit sets and their gauge functions. \iftoggle{arxiv}{}{A summary is provided here, with full derivations in the arXiv version.}
Let us suppose that $\bm{S}_E \in \mathbb{R}^{d}_+$ has unit exponential margins, density $f_{S_E}$, and gauge function $g_S$, and is independent of $\bm{V}_E \in \mathbb{R}^{d}_+$, which has unit exponential margins, density $f_{V_E}$, and gauge function $g_V$. Let $\bm{Z}_E = (\bm{S}_E,\bm{V}_E) \in \mathbb{R}^{2d}_+$ be the concatenation of these vectors\iftoggle{arxiv}{. Then, since its density $f_{Z_{E}}(\bm{z}) = f_{S_E}(z_1,\ldots,z_{d}) f_{V_E}(z_{d+1},\ldots,z_{2d})$, it is clear that $\bm{Z}_E$ has unit exponential margins and gauge function $g_Z(\bm{z}) = g_{S}(z_1,\ldots,z_{d}) + g_{V}(z_{d+1},\ldots,z_{2d})$.}{; this has exponential margins and gauge function $g_Z(\bm{z}) = g_{S}(z_1,\ldots,z_{d}) + g_{V}(z_{d+1},\ldots,z_{2d})$.}
Now consider the linear transformation of $\bm{Z}_E$ to
\begin{align*}
A\bm{Z}_E = (\gamma Z_{E,1}+ Z_{E,d+1},\ldots, \gamma Z_{E,d}+ Z_{E,2d}, Z_{E,1},\ldots, Z_{E,d}) = (\gamma \bm{S}_E+\bm{V}_{E}, \bm{S}_E) = (\bm{X}_{\tilde{E}},\bm{S}_E),
\end{align*}
where $A \in \mathbb{R}^{2d \times 2d}$ is the matrix describing this transformation\iftoggle{arxiv}{: the first $d$ rows have $\gamma$ and $1$ in the $j$th and $j+d$th positions, respectively, for $j=1,\ldots,d$, while the second $d$ rows have $1$ in the $j$th position for $j=1,\ldots, d$. All other entries are zero. The matrix $A^{-1}$ has the same configuration but with $-\gamma$ in place of $\gamma$.}{.}
By Lemma~4.1 of \citet{Nolde14}, the normalized sample cloud $\{A\bm{Z}_{E,i}/\log n : i=1,\ldots, n\}$ converges onto the set $AG$, where $G=\{\bm{z}\in\mathbb{R}^{2d}_+:g_Z(\bm{z}) \leq 1\}$, so $AG = \{\bm{z}\in\mathbb{R}^{2d}_+:A^{-1}\bm{z} \in G\} = \{\bm{z}\in\mathbb{R}^{2d}_+: g_Z(A^{-1}\bm{z}) \leq 1\}$. Consequently, the gauge function of $A\bm{Z}_E$ is $g_Z(A^{-1}\bm{z})$, i.e., $g_{Z}(\bm{x},\bm{s})=g_{S}(\bm{s}) + g_{V}(\bm{x}-\gamma\bm{s})$, for $\bm{x}>\gamma\bm{s}$.
Next we apply Proposition~\ref{prop:marg} to the vector $A\bm{Z}$, marginalizing over the last $d$ coordinates, which are equal to $\bm{S}_{E}$. This leaves us with the gauge function of $X_{\tilde{E}}$, denoted $g_X$, and given by
\begin{align*}
g_X(\bm{x}) = \min_{\bm{s} \in [\bm{0},\bm{x}/\gamma]} g_S(s_1,\ldots,s_d) + g_{V}(x_1-\gamma s_1,\ldots, x_d - \gamma s_d).
\end{align*}
To illustrate the results of \citet{HuserWadsworth18} and \citet{WadsworthTawn19} concerning model~\eqref{eq:HWspat}, we need to take $\bm{S}_{E} = S_{E}\bm{1}$, i.e., perfect dependence. Although such a vector does not have a $d$-dimensional Lebesgue density, convergence of the sample cloud based on the univariate random variable $S_E$ onto the unit interval $[0,1]$ implies that the limit set is $G_S = \{\bm{x} \in \mathbb{R}^d_+: x_1=x_2=\cdots=x_d=x, x \leq 1\}$. Such a set can be described by the gauge function
\[
g_S(\bm{s}) = \begin{cases}
\infty, & s_i \neq s_j~~\mbox{for any}~~i,j\\
s, & s_1=\cdots=s_d = s.
\end{cases}
\]
As such, in this case, $ g_X(\bm{x}) = \min_{s \in[0,\min(\bm{x})/\gamma]} \{s + g_{V}(\bm{x}-\gamma s)\}.$
\paragraph{Residual tail dependence} To find the residual tail dependence coefficient $\eta^X$, we require
\begin{align*}
\min_{\bm{x}:\min(\bm{x})=1} g_{X}(\bm{x}) &= \min_{\bm{x}:\min(\bm{x})=1}\min_{s \in[0,\min(\bm{x})/\gamma]} \{s + g_{V}(\bm{x}-\gamma s)\}\\
&=\min_{s \in[0,1/\gamma]}\min_{\bm{x}:\min(\bm{x})=1} \{s + g_{V}(\bm{x}-\gamma s)\}.
\end{align*}
For fixed $s$, consider $\min_{\bm{x}:\min(\bm{x})=1} g_{V}(\bm{x}-\gamma s) = \min_{\bm{z}:\min(\bm{z})=1-\gamma s} g_{V}(\bm{z}) = g_{V}(\bm{y}^\star\times(1-\gamma s))$, where $\bm{y}^\star = \arg\min_{\bm{y}:\min(\bm{y})=1} g_V(\bm{y})$. As such
\begin{align*}
\min_{\bm{x}:\min(\bm{x})=1} g_{X}(\bm{x}) &= \min_{s \in[0,1/\gamma]}\{s + g_V(\bm{y}^\star)(1-\gamma s)\} = \begin{cases} g_V(\bm{y}^\star),& \gamma < 1/g_V(\bm{y}^\star),\\ 1/\gamma, & \gamma \geq 1/g_V(\bm{y}^\star). \end{cases}
\end{align*}
Recalling that $\eta^X = [\min_{\bm{x}:\min(\bm{x})=1} g_{X}(\bm{x})]^{-1}$ and $\eta^V=1/g_{V}(\bm{y}^\star)$, this yields~\eqref{eq:etaHW}.
\paragraph{Conditional extremes} For the conditional extremes normalization, we now let $g_V$ and $g_X$ denote two-dimensional gauge functions. Suppose that $\alpha_V, \beta_V$ are such that $g_V(\alpha_V,1) = 1$ and $g_V(\alpha_V + u,1) - 1 \in \mathrm{RV}_{1/(1-\beta_V)}^0$. We have
\begin{align}
1= g_X(\alpha_X,1) = \min_{s} \{s + g_{V}(\alpha_X-\gamma s,1-\gamma s)\}. \label{eq:ceg}
\end{align}
Suppose that the right hand side of~\eqref{eq:ceg} is minimized at $s^\star \geq 0$, i.e., $g_X(\alpha_X,1) = s^\star + g_{V}(\alpha_X-\gamma s^\star,1-\gamma s^\star)$. Because $\alpha_X \leq 1$ and $g_V(v_1,v_2)\geq \max(v_1,v_2)$, this yields 1=$g_X(\alpha_X,1) \geq 1+ (1-\gamma) s^\star$, therefore we must have $s^\star=0$ for $\gamma \in (0,1)$. Consequently, $\alpha_X = \alpha_V = \alpha$.
\iftoggle{arxiv}{For the scale normalization, let $s_u^\star = \arg\min_{s \geq 0} \{s + g_{V}(\alpha_X-\gamma s + u,1-\gamma s)\}$. Then
\begin{align}
0 \leq g_{X}(\alpha + u, 1) -1 = s_u^\star + g_{V}(\alpha-\gamma s_u^\star + u,1-\gamma s_u^\star) - 1 \leq g_{V}(\alpha + u, 1) -1, \label{eq:ineqst1}
\end{align}
and because $g_V(x,y) \geq \max(x,y)$, if $\alpha<1$ then for sufficiently small $u$
\begin{align}
(1-\gamma)s_u^\star \leq s_u^\star + g_{V}(\alpha-\gamma s_u^\star + u,1-\gamma s_u^\star) - 1 \label{eq:ineqst2}.
\end{align}
Combining inequalities~\eqref{eq:ineqst1} and~\eqref{eq:ineqst2} shows that for $\gamma \in (0,1)$
\begin{align}
0 \leq (1-\gamma)s_u^\star \leq g_{X}(\alpha + u, 1) -1 \leq g_{V}(\alpha + u, 1) - 1 \to 0, \qquad u \to 0, \label{eq:ineqst3}
\end{align}
meaning in particular that $s_u^\star \to 0$ as $u \to 0$. To examine the minimizing sequence $s_u^\star$ in further detail, we consider the derivative of $s + g_{V}(\alpha-\gamma s + u,1-\gamma s)$, assuming that $g_V$ has piecewise continuous partial derivatives possessing finite left and right limits, denoted by $g_{V,1}, g_{V,2}$, such that
\begin{align}
\frac{\partial}{\partial s} \left[s + g_{V}(\alpha-\gamma s + u,1-\gamma s) \right]& = 1-\gamma\left[g_{V,1}(\alpha+u -\gamma s,1-\gamma s)+ g_{V,2}(\alpha+u -\gamma s,1-\gamma s)\right] \notag \\
& =
1-\gamma\left[g_{V}(\alpha+u -\gamma s,1-\gamma s) + \{1-(\alpha+u - \gamma s)\} g_{V,1}(\alpha+u -\gamma s,1-\gamma s)\right. \notag \\ & \qquad\qquad \left. -\gamma s g_{V,2}(\alpha+u -\gamma s,1-\gamma s)\right], \label{eq:gXderiv}
\end{align}
using Euler's homogeneous function theorem on the second line. Consider~\eqref{eq:gXderiv} evaluated at $s_u^\star$ and take the limit as $u \to 0$, yielding
\begin{align}
1-\gamma[1+(1-\alpha)g_{V,1}(\alpha_+,1)]. \label{eq:gXderivlim}
\end{align}
For $g_V$ differentiable at $(\alpha,1)$, $g_{V,1}(\alpha_+,1) = g_{V,1}(\alpha_-,1) = 0$, so the limit of~\eqref{eq:gXderiv} is $1-\gamma>0$, and hence there exists $\epsilon>0$ such that \eqref{eq:gXderiv} is positive for all $u< \epsilon$, giving $s_u^\star = 0$ for all $u<\epsilon$. Consequently, $g_{X}(\alpha+u,1)-1 \sim g_{V}(\alpha+u,1)-1$, $u \to 0$.
For $g_{V}$ not differentiable at $(\alpha,1)$,~\eqref{eq:gXderivlim} is positive when $g_{V,1}(\alpha_+,1)<(1/\gamma-1)/(1-\alpha)$, and again in this case $g_{X}(\alpha+u,1)-1 \sim g_{V}(\alpha+u,1)-1$, $u \to 0$. When $g_{V,1}(\alpha_+,1)>(1/\gamma-1)/(1-\alpha)$, then the minimizing sequence $s_u^\star$ should be as large as possible, i.e., equal to its upper bound of $[g_{X}(\alpha+u,1)-1]/(1-\gamma) = [s_u^\star + g_{V}(\alpha-\gamma s_u^\star + u,1-\gamma s_u^\star) - 1]/(1-\gamma)$ from inequality~\eqref{eq:ineqst3}. Further asymptotic detail on this bound is obtained through a Taylor expansion:
\begin{align}
g_{V}(\alpha-\gamma s_u^\star + u,1-\gamma s_u^\star) = g_V(\alpha,1) + (u-\gamma s_u^\star)g_{V,1}(\alpha_+,1) - \gamma s_u^\star g_{V,2}(\alpha,1_-) + O(\max(u^2,s_u^{\star 2},u s_u^\star)), \label{eq:gXTaylor}
\end{align}
giving
\begin{align*}
s_u^\star \leq \frac{u g_{V,1}(\alpha_+,1)}{\gamma[g_{V,1}(\alpha_+,1)+g_{V,2}(\alpha,1_-) - 1]}[1+o(1)] = \frac{u}{\gamma(1-\alpha)} + o(u).
\end{align*}
Taking $s_u^\star$ at this upper bound and using the expansion~\eqref{eq:gXTaylor},
\begin{align*}
g_{X}(\alpha+u,1)-1 = s_u^\star + g_{V}(\alpha-\gamma s_u^\star + u,1-\gamma s_u^\star) -1 = \frac{1-\gamma}{\gamma(1-\alpha)}u + o(u),
\end{align*}
such that $g_{X}(\alpha+u,1)-1 \in \mathrm{RV}^0_{1}$. Since we also have $g_{V}(\alpha+u,1)-1 \sim u g_{V,1}(\alpha_+,1)\in \mathrm{RV}^0_{1}$, the regular variation indices are identical with $\beta_X=\beta_V=0$. This represents a case where the scale normalizations $b_{s-s_0}(t)$ in the conditional extremes representation do not diverge to infinity, meaning a potential difference in the limit distribution.}{Calculations for the scale normalization are more involved and can be found in the arXiv version. We find that for $g_V$ differentiable at $(\alpha,1)$, $g_{X}(\alpha+u,1)-1 \sim g_{V}(\alpha+u,1)-1$, $u \to 0$, whereas in the non-differentiable case we do not necessarily have this link but can deduce that the regular variation indices are $\beta_X=\beta_V = 0$.}
\iftoggle{arxiv}{Figure~\ref{fig:HW} displays examples of gauge functions $g_V$ and $g_X$. We observe from this figure how, when $\gamma$ is sufficiently large, the shape of $g_V$ is modified to produce $g_X$. The modification is focussed around the diagonal, and explains visually why the residual tail dependence coefficient changes while the conditional extremes normalization does not. The left and right panels illustrate differentiable cases, with $s_u^\star \equiv 0$ for sufficiently small $u$; the centre panel depicts an example with $s_u^\star$ linear in $u$ as $u\to 0$.}{Figure~\ref{fig:HW} displays examples of gauge functions $g_V$ and $g_X$. We observe from this figure how, when $\gamma$ is sufficiently large, the shape of $g_V$ is modified to produce $g_X$. The modification is focussed around the diagonal, and explains visually why the residual tail dependence coefficient changes while the conditional extremes normalization does not. The left and right panels illustrate differentiable cases, and the centre panel non-differentiable.}
\begin{figure}
\centering
\includegraphics[width=0.3\textwidth]{Fig/HWGauss.pdf}
\includegraphics[width=0.3\textwidth]{Fig/HWNegDep.pdf}
\includegraphics[width=0.3\textwidth]{Fig/HWLpnorm.pdf}
\caption{Solid red lines depict the level sets $g_V(\bm{x})=1$, where $g_V$ is of the form (ii), (iii) and (iv) (L--R) from Figure~\ref{fig:2dgauge}. Dashed black lines depict the level sets $g_X(\bm{x}) =1$. In each picture the blue solid line segment is from $(0,0)$ to $(\gamma,\gamma)$, and denotes the limit set of the fully dependent random vector $\gamma S_E$. From left to right, $\gamma=0.9,0.5,0.8$.}
\label{fig:HW}
\end{figure}
\section{Discussion}
\label{sec:Discussion}
In this work we have demonstrated how several concepts of extremal dependence can be unified through the shape of the limit set $G$ of the scaled sample cloud $N_n =\{\bm{X}_1/r_n,\ldots,\bm{X}_n/r_n\}$ arising for distributions with light-tailed margins. For concreteness our focus has been on exponential margins, but other choices can be useful. In the case of negative dependence between extremes --- such that large values of one variable are most likely to occur with small values of another --- the double exponential-tailed Laplace margins can be more enlightening. As an example, for the bivariate Gaussian copula with $\rho<0$ we observed that the limit set $G$ is described by a discontinuous gauge function $g$ that cannot be established through the simple mechanism of Proposition~\ref{prop:logdens}. In \citet{Nolde14}, the gauge function for this distribution in Laplace margins is calculated as
\begin{align*}
g(x,y) = \begin{cases}
(|x|+|y|-2\rho|xy|^{1/2})/(1-\rho^2), & x,y \geq 0~~\mbox{or}~~x,y \leq 0, \\
(|x|+|y|+2\rho|xy|^{1/2})/(1-\rho^2), & x\geq 0, y \leq 0~~\mbox{or}~~x\leq 0,y \geq 0.
\end{cases}
\end{align*}
When $\rho<0$, this yields $g(1,-\rho^2) = 1$ , and $g(1,-\rho^2 + u) \in \mathrm{RV}_{2}^0$, so that extending Proposition~\ref{prop:CEgauge}, we would find that the conditional extremes normalizations are $a^j(t) \sim -\rho^2 t$ and $b^j(t) \in \mathrm{RV}^0_{1/2}$, as given in \citet{Keefetal13}.
The study of extremal dependence features through the limit set $G$ is enlightening both for asymptotically dependent and asymptotically independent random vectors, particularly as it can be revealing for mixture structures where mass is placed on a variety of cones $\mathbb{E}_C$ as defined in~\eqref{eq:EC}. However, many traditional measures of dependence within the asymptotically dependent framework, which are typically functions of the exponent function $V$ given in equation~\eqref{eq:V}, or spectral measure $H$, are not revealed by limit set~$G$. For example, it was noted in the example of Section~\ref{sec:eglog} that the limit set described by the gauge function $g(x,y) = \theta^{-1}\max(x,y) + (1-\theta^{-1})\min(x,y)$ can arise for several different spectral measures, although clearly the parameter $\theta$ demonstrates some link between strength of dependence and shape of $G$.
Nonetheless, multivariate regular variation and associated limiting measures have been well-studied in extreme value theory, but representations that allow greater discrimination between asymptotically independent or mixture structures much less so. The limit set elucidates many of these alternative dependence concepts and provides meaningful connections between them. We have not directly considered connections between the various dependence measures without reference to $G$, and we note that the limit set might not always exist. We leave such study to future work.
\subsection*{Acknowledgements}
The authors are grateful to the editor and two reviewers for constructive feedback and valuable comments that helped improve the paper. NN acknowledges financial support of the Natural Sciences and Research Council of Canada. JLW gratefully acknowledges funding from EPSRC grant EP/P002838/1.
\bibliographystyle{apalike}
|
\section{Introduction}
In the past half-decade, there has been a burst of activity surrounding the
application of deep learning methods (particularly neural networks) to a
variety of problems in collider physics. Much of this research has been
focused on tagging jets \cite{Lonnblad:1990qp, Cheng:2017rdo, Luo:2017ncs,
efn, Kasieczka:2018lwf, Qu:2019gqs, calo, jet-images, dl-colour, baldi,
unimelb, Pearkes:2017, Almeida:2015jua, efp, spectra, message-pass,jedi,
qcd-aware, lund, QCDorWhat, autoencoder, Kasieczka:2017nvn,
Macaluso:2018tck, Butter:2017cot,MUST}. Jets are collimated sprays
of particles produced from energetic partons and contribute significantly
to the complexity of a collider event. Optimal methods of jet tagging are
therefore highly desirable in order to maximally exploit the physics
capabilities of the LHC and future colliders. Examples include quark/gluon
discrimination \cite{qgtag, qgdiscrim, FerreiradeLima:2016gcz, Frye:2017yrw,
Davighi:2017hok, Komiske:2018vkc, Larkoski:2019nwj} and the discrimination
of QCD jets from overlapping jets produced by the hadronic decay of a
boosted heavy resonance~\cite{butter, Salam:2009jx, Plehn_2012,
Dasgupta:2015yua, nsub,Ju:2020tbo}, a topic of much interest at the LHC.
Indeed, the ``standard candle" for jet tagging studies is the
identification of boosted W-bosons from their decay products, a process
which we study in this work.
The new ML-based methods leverage the fine resolution of modern detectors
through their ability to automatically construct powerful discrimination
variables from low-level inputs. By choosing a suitable representation for
the inputs, existing deep learning models can be directly applied to the
task of classifying jets. For example, multilayer perceptrons (MLPs) can be
trained on lists of jets' constituent particle momenta~\cite{Pearkes:2017},
and convolutional neural networks (CNNs) can be trained on jet images~
\cite{calo, jet-images, unimelb, dl-colour, baldi}. These methods have led
to substantial improvements on the performances of traditional taggers.
However, the jet representations they use have drawbacks. For example,
jet images tend to be sparsely populated by particles, making a large
number of the input pixels redundant. On the other hand, some methods based
on the treatment of jets as lists (such as recurrent neural networks) fail
to account for the fact that the particles in a jet have no intrinsic
ordering. Accordingly, it is important to continue to thoroughly explore
the landscape of machine learning-based taggers, and better understand
their performance through analytic calculations~\cite{Kasieczka:2020nyd}.
A specific recent example is the Energy Flow Network (EFN), introduced in
Ref.\,\cite{efn}. EFNs are based on the Deep Sets framework~\cite{deep-sets},
which treats data as point clouds -- sets of points in space. Point clouds do
not have a fixed length or notion of ordering, similar to particle jets at
colliders. The Deep Sets framework demonstrates how to incorporate permutation
invariance and equivariance of the data -- jet constituents in our case. The
permutation-invariant case was extensively studied in Ref.\,\cite{efn} which
treats jets as sets of particle four-momenta, and was also able to show that
the resulting observables can be made infrared and collinear (IRC) safe. More
general non-IRC safe networks were also studied, and dubbed Particle Flow
Networks (PFNs).
Along with permutation-invariant networks the Deep Sets framework can
incorporate permutation equivariance. A permutation-equivariant function is
one which commutes with the permutation operation. That is, for input data
$\vec{x}$, a permutation $\pi$ and a function $f$ equivariance requires
$f(\pi(\vec{x}))=\pi(f(\vec{x}))$. Clearly this requires the dimensionality
of the output of $f$ to be the same as the input. For binary classifiers the
output is usually one-dimensional. Accordingly, it is straightforward to have
an entire network which exhibits invariance of its output under permutation of
the input parameters, but equivariance is implemented on individual internal
layers of the network. While equivariance under \textcolor{black}{transformations of inputs } is less obviously motivated than invariance, \textcolor{black}{it has proven to benefit performance on some tasks. Most notably, the convolutions performed within a CNN are translation-equivariant\,\cite{LeCun} and are responsible for such models' suitability to images, which are naturally symmetric to translations. Equivariance to more complex symmetries has also been investigated such as in Refs.\,\cite{group-equi, Thomas2018TensorFN, Bogatskiy:2020tje}. Ref.\,\cite{Bogatskiy:2020tje} constructs networks that are equivariant under
the action of the Lorentz group.}
Since the Lorentz group is a Lie group and hence
continuous, this necessitates a different approach from the finite permutation groups
relevant to this work.
In this paper, we continue the study of Energy Flow Networks by implementing
permutation-equivariant network layers into the EFN architecture. We are able
to find a variant which maintains infrared and collinear safety of the observables
learned by the resulting network. We study the performance of equivariant Energy
Flow and Particle Flow networks on boosted $W$-tagging, including some networks
that use information about the jet constituent ID. While the equivariant EFNs slightly
outperform the invariant EFNs, we find that this can change under decorrelation of the
network output using planing. We also consider some possible future directions.
\section{Energy Flow Networks}
\begin{figure*}[tp!]
\centering
\includegraphics[width=0.92\textwidth]{network.pdf}
\caption{Examples of the two multilayer perceptrons comprising an EFN. The per-particle mapping (left) produces $\ell$ filters and the function $F$ (right) takes the observables $\mathcal{O}_a = \sum_i z_i\Phi_a(y_i, \phi_i)$ to a probability that the input was a signal jet ($F_S$), or a background jet ($F_B$).}
\label{fig:arch}
\end{figure*}
In principle, the output of any neural network that takes jets or jet constituents as
input can be interpreted as a jet observable. This output could be sensitive to the
ordering of the constituents, contrary to the notion that collider events can be viewed
as unordered list of particles. This has motivated the development and application of
neural network architectures which are insensitive to the ordering of the jet constituents,
where jets are viewed as point clouds: a set of data points in space. The representation of
functions on sets from which neural network architectures may be modeled has been much
studied recently in the ML community \cite{deep-sets, pointnet, bruno-perm, janossy-perm,
adversarial-perm}. Collider physics studies in this vein include the use of architectures
based on Graph Neural Networks such as Refs.\,\cite{Qu:2019gqs} and related
Message Passing Neural Networks~\cite{message-pass,Ren:2019xhp,Abdughani:2020xfo}.
We will explore the Energy Flow Network (EFN), an example which is based closely on the Deep
Sets model introduced in Ref.\,\cite{deep-sets}. Among deep learning models for jet tagging
EFNs have the important feature that they are engineered to enforce infrared and collinear
(IRC) safety on the network output. IRC safety is the requirement that a jet observable be
invariant under soft and collinear emissions. We will demonstrate IRC safety of equivariant
EFNs in Section~\ref{sec:augmented}.
\subsection{Deep Sets framework}
\label{deepsets}
Deep Sets~\cite{deep-sets} provides a framework for constructing neural networks
that incorporate permutation invariance of the inputs. An Energy Flow Network is
an implementation of this model which acts on jets (considering jets as sets of
particles with associated momenta), with the additional requirement of IRC safety.
The Deep Sets framework represents permutation-invariant functions $f$ in the
following way. For a function acting on a set \ensuremath{X}, permutation
invariance is ensured by first mapping the individual elements of \ensuremath{X}
to an intermediate \emph{latent space} using a function $\Phi$. The results
$\Phi(x_i)$ are then summed in the latent space. Finally, a different map $F$ takes
the result of the summation in the latent space to the desired range of $f(X)$. Given
that the sum is invariant under permutations of elements in \ensuremath{X}, the same
property is inherited by the total function $f$. The sum also allows such networks
to accept inputs of varying length.
The authors of Ref.\,\cite{deep-sets} prove the following
\begin{center}
\begin{minipage}{\linewidth}
\textbf{Deep Sets Theorem:} \textit{A function $f$ operating on a set \ensuremath{X} having
elements from a countable universe $\mathfrak{X}$ is invariant to the permutation of
instances in \ensuremath{X} if and only if it can be decomposed in the form}
\begin{equation}
f(\ensuremath{X}) = F\left(\,\sum_{x\in \ensuremath{X}}\Phi(x)\right),
\end{equation}
\textit{for suitable transformations $\Phi$ and $F$.}
\end{minipage}
\end{center}
Ref.\,\cite{efn} applies this decomposition to jet observables, which we may consider
as functions on the set of constituent particle features (four-momentum, charge, flavour etc.).
Such observables can therefore be written in the form
\begin{equation}\label{eq:pfn-obs}
\mathcal{O}\left(\left\{p_1,\:\cdots,p_M\right\}\right) = F\left(\sum_{i=1}^M\Phi(p_i)\right),
\end{equation}
where each $p_i\in\mathbb{R}^d$ is a vector containing $d$ features of particles in the jet,
$\Phi:\mathbb{R}^d\to\mathbb{R}^\ell$ is a per-particle map to the latent space, and
$F:\mathbb{R}^\ell\to\mathbb{R}$ is a continuous mapping. The index $i$ runs over the $M$
particles in the jet. In fact, even before applying the function $F$, each of the $\ell$
components of the summed per-particle maps is a jet observable. The final observable
$\mathcal{O}$ is some combination of these, dictated by $F$. A number of familiar examples
can be written in this way. For example, the jet mass corresponds to choices $\Phi(p^\mu)=p^\mu$
and $F(x) = \sqrt{x^\mu x_\mu}$.
Ref.\,\cite{efn} makes use of the Stone-Weierstrass Approximation Theorem\:\cite{weierstrass}
to derive the specialisation of Eq.\,\ref{eq:pfn-obs} to the case where $\mathcal{O}$
is IRC safe. They arrive at
\begin{equation}
\label{irc-obs-decomp}
\mathcal{O}(\left\{p_1,\;\cdots,p_M\right\})=F\left(\sum_{i=1}^Mz_i\Phi(\Hat{p}_i)\right),
\end{equation}
where the per-particle mapping $\Phi:\mathbb{R}^2\to\mathbb{R}^\ell$ now depends only on
the angular coordinates of the jet constituent momenta through $\Hat{p}\equiv(y,\phi)$,
and the sum is weighted by the transverse momentum fractions $z_i=p_{T,i}/\sum_jp_{T,j}$.
Finally, we comment on the dimensionality of the latent space. While any IRC safe observable
can be represented this way, this may require a latent space of high dimensionality. In
Ref.\,\cite{deep-sets}, the Deep Sets Theorem was only proven for the case where instances
of the set \ensuremath{X} are elements of a countable domain. Ref.\,\cite{setfunclimit} argues
that extending consideration to uncountable sets is necessary since continuity on countable set
(such as $\mathbb{Q})$ is a vastly different condition to continuity on the (uncountable) reals.
They show that the Deep Sets theorem can only be extended to \emph{finite subsets} of uncountable
domains, with the condition that the latent dimension should be at least the number of elements
in the set. That is, in order to ensure universal function approximation of a network in the
Deep Sets framework, one should have $\ell\geq M$. This is consistent with the observation in
Ref.\,\cite{efn} that an EFN's performance saturates near $\ell=64$, since the multiplicity
distributions of the jet samples peak around 40--50. We will present results below for
$\ell=128$ and 256.
\subsection{Architecture}
To realise Eq.\,\ref{irc-obs-decomp} as a neural network, Ref.\,\cite{efn} uses two
multi-layer perceptrons (MLPs) to approximate the functions $\Phi$ and $F$. A schematic
of this construction for an EFN with latent dimension $\ell$ is presented in Fig.\,\ref{fig:arch}.
The network on the left represents the per-particle mapping and the number of nodes in its
output is equal to the dimension of the latent space. Each of these output nodes is called
a \emph{filter} and is a function of the jet constituents' angular coordinates $y$ and $\phi$.
Each filter is transformed into a jet observable by performing the $p_T$-weighted sum
$\mathcal{O}_a = \sum_iz_i\Phi_a(y_i,\phi_i)$, where the sum runs over particles in the
input jet. This set of observables is then passed as input to the second network which
represents the map $F$. As is customary for the application of neural networks to classification
tasks, the output layer of $F$ has two nodes and uses the \emph{softmax} activation, which
correlates these nodes such that their outputs take values in the interval $[0,1]$ and
sum to unity. The output of the network may then be interpreted as probabilities, $F_S$ and
$F_B$, that the input is a $W$ (signal) or QCD (background) jet respectively. However, since
the outputs are correlated, the network really only represents a single IRC safe jet observable
for which $W$ jets have a value close to 1, and QCD jets, 0 (or vice-versa).
\section{Augmented Energy Flow Network architecture}
\label{sec:augmented}
\begin{figure*}[t!]
\centering
\includegraphics[width=0.92\textwidth]{equivariant-network.pdf}
\caption{An example implementation of two equivariant layers (blue) into the EFN architecture. The first equivariant layer has $\ell$ input channels and the second has $\ell'$ output channels. A pooling operation is required to merge the vectors $E_i$ into a single vector $\Vec{\mathcal{O}}$ containing jet observables.}
\label{fig:equi-arch}
\end{figure*}
In addition to presenting a permutation-invariant neural network architecture, on which the
EFN is based, Ref.\,\cite{deep-sets} also defines permutation-\emph{equivariant} neural network
layers. Equivariance of a function means that it commutes with permutations, rather than being
invariant to them. Here we detail the way in which an equivariant layer can be incorporated into
the existing structure of an EFN and seek an architecture that is able to maintain IRC safety of
the network output.
\subsection{Equivariant network layers}\label{sec:layer-def}
Permutation equivariance constrains the form of neural network layers. Consider a network layer
$E_\Theta(\bm{x})=\sigma(\Theta\bm{x})$, having activation $\sigma$, weight matrix
$\Theta \in \mathbb{R}^{M\times M}$, and acting on a set $\bm{x}$ with elements $x_i\in\mathbb{R}$.
A result from Ref.\,\cite{deep-sets} states that such a layer is permutation-equivariant if and
only if $\Theta$ can be decomposed as
\begin{equation}
\label{eqvsum}
\Theta = \lambda\bm{I} + \gamma(\bm{1}\bm{1}^T),
\end{equation}
with $\lambda, \gamma\in\mathbb{R}$, $\bm{I}$ = $id\in\mathbb{R}^{M\times M}$ and
$\bm{1}=(1,\:\cdots,1)^T\in\mathbb{R}^M$. That is, the matrix $\Theta$ has all diagonal terms
equal and all off-diagonal terms equal. The function $E$ therefore depends only on a linear
combination of the input $\bm{x}$ itself and $\bm{1}\bm{1}^T\bm{x}$ -- a vector with each
component being the sum of components of $\bm{x}$. The equivariance of the layer arises from
the fact that this summation respects permutation symmetry. One can also consider functions $E$
whose arguments are not of the linear form $\Theta \bm{x}$. For example, the function
\begin{equation}
\label{eqvmax}
E(\bm{x}) = \sigma(\lambda\bm{I}\bm{x} + \gamma\,\text{max}(\bm{x})\bm{1}),
\end{equation}
also defines an equivariant network layer. Eqs~\ref{eqvsum} and \ref {eqvmax} are valid when the
elements of the set $x_i\in\mathbb{R}$. However, the $x_i$ could themselves be vectorial
$x_i\in \mathbb{R}^D$, so that $\bm{x}\in\mathbb{R}^{M\times D}$. In that case the parameters
$\lambda$ and $\gamma$ can be promoted to matrices $\Lambda, \Gamma\in\mathbb{R}^{D\times D'}$
with $D'$ being the dimension of each output node. This leads to
\begin{equation}
\label{eqvmatsum}
E(\bm{x}) = \sigma(\bm{x}\Lambda + \bm{1}\bm{1}^T\bm{x}\Gamma),
\end{equation}
\begin{equation}
\label{eqvmatmax}
E(\bm{x}) = \sigma\big(\bm{x}\Lambda + \bm{1}\text{maxpool}(\bm{x})\Gamma\big),
\end{equation}
where maxpool is an operation that merges feature vectors in the set $\bm{x}\in\mathbb{R}^{M\times D}$
to a single summary vector in $\mathbb{R}^D$ with $i^{\rm{th}}$ component equal to
$\max\left\{{x_j}^i\right\}_{j=1}^M$.
We will see that EFNs using these layers are IRC unsafe in general. In our studies below we
will use layers of the form Eq.\,\ref{eqvmatmax} as an example of such a network, and further
discuss IRC safety below.
\subsection{Layer implementation}
As well as introducing the equivariant layer itself, Ref.\,\cite{deep-sets} also discusses how
such layers can be implemented in a specific model. Two properties of the equivariant layer will
be exploited. The first is that the composition of equivariant layers is also equivariant. This
can be seen by considering equivariant functions $f,g$ acting on a set $\bm{x}$ with $n$ elements
and $\pi\in S_n$ a permutation. Then
{\setlength{\belowdisplayskip}{6pt}
\setlength{\abovedisplayskip}{6pt}
\begin{align}
(f&\circ g)(\pi \bm{x})\notag\\&=f(g(\pi \bm{x})) = f(\pi g(\bm{x})) = \pi(f(g(\bm{x})) = \pi(f\circ g)(\bm{x}) \, ,
\end{align}}\noindent%
so the composition of $f$ and $g$ is equivariant. This implies that equivariant layers can be
stacked to build deep models.
The second property is that following an equivariant layer with any commutative pooling operation
(such as sum, mean or max) across set elements yields a permutation-\textit{invariant} function.
This allows us to replace the sum over per-particle maps (the permutation-invariant function in
the standard EFN architecture) with a sequence of equivariant layers followed by a pooling operation.
The network is then able to learn its own particular permutation-invariant function by which to
combine the per-particle information.
A schematic of the architecture is presented in Fig.\,\ref{fig:equi-arch}. Just as with the
standard permutation-invariant EFN, on the left we have a dense network representing a per-particle
map $\Phi:\mathbb{R}^2\to\mathbb{R}^\ell$ that acts on the $y$-$\phi$ plane. In the centre of the
diagram the vector representation of each particle under this map is then passed as input to a
number of equivariant layers shown with blue shaded nodes, the first of which has $\ell$ input
\emph{channels}. The number of input (output) channels in an equivariant layer is defined as the
dimensionality of each input (output) node.
\textcolor{black}{It is important to note that the size of the equivariant layers must be at least the number of jet constituents in the input, which may vary during training. As such, our new architecture is not naturally able to accept variable-length sets. This is addressed by fixing the layer size, $M$, to be suitably large and zero-padding the inputs. Such treatment is not necessary for an EFN, since its architecture is completely independent of the input jet multiplicity. However, the number of parameters contained in an equivariant layer does not vary with $M$, so zero-padding inputs to this new architecture is not as troublesome as it is in some other examples.}
The final equivariant layer, having $\ell'$ output channels, is followed by a pooling operation
that projects the $M\times \ell'$-dimensional output of the layer to a vector $\Vec{\mathcal{O}}$
of length\;$\ell'$. There is freedom to choose this operation, for example some of the models
described in Sec.\,\ref{sec:models} use a maxpooling to construct the components of the vector
according to $\mathcal{O}^a=\max\left\{{E_i}^a\right\}_{i=1}^M$, where ${E_i}^a$ is a component
of an output node in the final equivariant layer, with $i$ labeling jet constituents and $a$
indexing the latent dimension. These $\mathcal{O}^a$ are permutation-invariant jet observables,
in analogy with the $p_T$-weighted sum over filters in an EFN, and are passed as input to the MLP
representing the function $F:\mathbb{R}^{\ell'}\to\mathbb{R}$.
\subsection{IRC safety}
\label{sec:irc-safety}
The networks generated from the method above are distinct from the IRC safe EFN studied in
Ref.\,\cite{efn}, which corresponds to taking the pooling operation as a $z$-weighted sum and
the function $E$ as the identity. As such, IRC safety of the jet observables from the new
networks is not guaranteed. In this section, we aim to find an architecture that maintains
IRC safety. This allows for a closer comparison with the EFNs of Ref.\,\cite{efn}.
There are a number of potential architectures of the type described in the previous section.
We will consider an equivariant energy flow network denoted by EV-EFN, which for simplicity
we assume has a just a single equivariant layer $E:\mathbb{R}^{M\times\ell}\to\mathbb{R}^{M\times\ell'}$.
We will consider the implications of multiple equivariant layers later in the section. We
may express the action of the network as
\begin{align}\label{evefn-action}
\mathrm{EV\text{-}EFN}\,\left(\{p_i\}_{i=1}^M\right) = \big(F\circ P\circ E\big)(\textcolor{black}{\{\Phi(\Hat{p}_i)\}_{i=1}^M}),
\end{align}
where $P:\mathbb{R}^{M\times\ell'}\to\mathbb{R}^{\ell'}$
is the pooling operation, and $F:\mathbb{R}^\ell\to\mathbb{R}$ is the final function. To specify a network,
we must make choices for the pooling operation $P$ and the form of the equivariant layer
(using Eqs.\,\ref{eqvmatsum},\,\ref{eqvmatmax} above for instance etc.). The maps $F$ and $\Phi$
are given by fully-connected MLPs. We can write the vector of jet observables $\Vec{\mathcal{O}}$
by dropping $F$ from Eq.\,\ref{evefn-action}. This gives
\begin{align}
\Vec{\mathcal{O}} = \big(P\circ E\big)(\textcolor{black}{\{\Phi(\Hat{p}_i)\}_{i=1}^M}).
\label{eq:obs}
\end{align}
Since we know that taking $P$ as a $z$-weighted sum ${P(\bm{x})=\sum_{i}z_ix_i}$ is an IRC safe
choice in an EFN, we will only consider this operation for this section. This may not be the unique
choice for ensuring IRC safety, but we were able to rule out some other possibilities. For example,
a max function is not necessarily invariant to the collinear splitting of a particle.
In an EFN with equivariant layers, this choice of $P$ alone is not enough to ensure that the
observables vector $\Vec{O}$ is IRC safe. The reason for this is that in the new architecture,
the output nodes of the equivariant layer are not strictly per-particle maps. That is, the
vector $E_i$ has dependence on the momentum of particles other than $i$. However, the
fact that this mixing comes through an equivariant layer imposes helpful constraints.
Specifically, permutation equivariance demands that the $i^{\rm{th}}$ output of the layer depends
on inputs other than $i$ only through permutation-invariant functions of the inputs (in linear combination).
For example, in Eq.\,\ref{eqvsum} for $\left(\Theta\bm{x}\right)_i$ the first term depends only
on $x_i$ and the second term is proportional to the permutation-invariant sum $\sum_i x_i$. Thus,
despite the outputs $E_i $ of the equivariant layer not being per-particle maps, they do have a
unique dependence on the $i^{\rm{th}}$ particle.
In addition, the permutation-invariant functions that give the dependence on all particles can be
identified as jet observables. We show below that the IRC safety of the $z$-weighted sum of such an
output, which depends on the linear combination of single particle information and a jet observable,
is tied to the IRC safety of the jet observables in the equivariant layer. We therefore parameterise
an equivariant layer as
\begin{equation}
E(\textcolor{black}{\{\Phi(\Hat{p}_i)\}_{i=1}^M}) = \sigma\left(\Phi(\Hat{p}_i)\cdot\Lambda + \mathcal{Q}(\textcolor{black}{\{\Phi(\Hat{p}_i)\}_{i=1}^M})\cdot\Gamma\right),
\end{equation}
where $\mathcal{Q}\in\mathbb{R}^\ell$ holds jet observables that depend on the set of per-particle
latent vectors, the matrices $\Lambda$ and $\Gamma$ are defined as in Eq.\,\ref{eqvmatsum}, and $\sigma$
is the activation function. \textcolor{black}{However, as it stands, the equivariant layer only takes
directional information $\Hat{p}$ as input through the filters $\Phi$. If $\mathcal{Q}$ is to be IRC
safe, it must also receive the momentum fractions $z$ as input. Such a dependence can be added to the
layer $E$ through $\mathcal{Q}$ while maintaining equivariance if $Q$ is understood to be permutation-invariant
with regard to pairs $\left(\Phi(\Hat{p}_i), z_i\right)$. This yields}
\begin{equation}
E(\textcolor{black}{\{\Phi(\Hat{p}_i)\}_{i=1}^M,z}) = \sigma\left(\Phi(\Hat{p}_i)\cdot\Lambda + \mathcal{Q}(\textcolor{black}{\{(\Phi(\Hat{p}_i),z_i)\}_{i=1}^M})\cdot\Gamma\right),
\end{equation}
Taking $P$ as a $z$-weighted sum and the discussion above into account, Eq.\,\ref{eq:obs} for a
network with a single equivariant layer can be written out in more detail as
\begin{equation}\label{eq:eqv-irc}
\Vec{\mathcal{O}}=\sum_{i=1}^Mz_i\sigma\left(\Phi(\Hat{p}_i)\cdot\Lambda + \mathcal{Q}(\textcolor{black}{\{(\Phi(\Hat{p}_i),z_i)\}_{i=1}^M})\cdot\Gamma\right).
\end{equation}
We wish to show that the observable $\Vec{\mathcal{O}}$ is IRC safe if every component
of $\mathcal{Q}$ is IRC safe. To study infra-red safety, we add an arbitrarily soft particle
labeled by index $s$ that can have momentum in any direction:
\begin{align}
\Vec{\mathcal{O}}_{IR} &= \lim_{z_s\to0}z_s\sigma\left(\Phi(\Hat{p}_s)\cdot\Lambda + \mathcal{Q}_{IR}\cdot\Gamma\right) \notag \\
&\ \ \:+ \sum_{i=1}^Mz_i\sigma\left(\Phi(\Hat{p}_i)\cdot\Lambda + \mathcal{Q}_{IR}\cdot\Gamma\right)\notag\\
&=\sum_{i=1}^Mz_i\sigma\left(\Phi(\Hat{p}_i)\cdot\Lambda + \mathcal{Q}_{IR}\cdot\Gamma\right) \, ,
\label{eqv-ir-safety}
\end{align}
where \textcolor{black}{$\mathcal{Q}_{IR}=\mathcal{Q}(\{(\Phi(\Hat{p}_s),z_s)\}\cup\{(\Phi(\Hat{p}_i),z_i)\}_{i=1}^M)$}.
For collinear safety a particle splits with momentum fractions $\lambda z$ and
$(1-\lambda)z$ with $\lambda\in [0,1]$. Without loss of generality we take this
to be the first particle. Under a collinear splitting the observable becomes
\begin{align}
\Vec{\mathcal{O}}_C &= \lambda z_1\sigma\left(\Phi(\Hat{p}_1)\cdot\Lambda + \mathcal{Q}_C\cdot\Gamma\right)\notag\\
&\quad+ (1-\lambda)z_1\sigma\left(\Phi(\Hat{p}_1)\cdot\Lambda + \mathcal{Q}_C\cdot\Gamma\right)\notag\\
&
\quad
+ \sum_{i=2}^Mz_i\sigma\left(\Phi(\Hat{p}_i)\cdot\Lambda + \mathcal{Q}_C\cdot\Gamma\right)\notag\\
&=\sum_{i=1}^Mz_i\sigma\left(\Phi(\Hat{p}_i)\cdot\Lambda + \mathcal{Q}_{C}\cdot\Gamma\right),
\label{eqv-c-safety}
\end{align}
and \textcolor{black}{$\mathcal{Q}_{C}=\mathcal{Q}(\{(\Phi(\Hat{p}_1),\lambda z_1),(\Phi(\Hat{p}_1),(1-\lambda) z_1)\}\cup\{(\Phi(\Hat{p}_i),z_i)\}_{i=2}^M)$}.
If $\mathcal{Q}_{IR}=\mathcal{Q}_{C}=\mathcal{Q}$, then Eq.\,\ref{eq:eqv-irc} is recovered
from both Eq.\,\ref{eqv-ir-safety} and Eq.\,\ref{eqv-c-safety}. Thus, $\Vec{\mathcal{O}}$ is
IRC safe if $\mathcal{Q}$ is IRC safe. One obvious choice for the components $\mathcal{Q}^a$ is
to again use the $z$-weighted sum over per-particle maps, $\mathcal{Q}^a=\sum_iz_i\Phi^a(\Hat{p}_i)$ \textcolor{black}{leading to}
\begin{equation}\label{eq:irc-equi-def}
E(\textcolor{black}{\{\Phi(\Hat{p}_i)\}_{i=1}^M},z) = \sigma\left(\textcolor{black}{\{\Phi(\Hat{p}_i)\}_{i=1}^M}\cdot\Lambda + \sum_{j=1}^Mz_j\Phi(\Hat{p}_j)\cdot\Gamma\right)
\end{equation}
The IRC safe jet observables are then obtained by composing with the $z$-weighted sum. In full, they read
\begin{align}\label{1equiircobs}
\Vec{\mathcal{O}} &= \sum_{i=1}^Mz_i{E_i}\big(\textcolor{black}{\{\Phi(\Hat{p}_i)\}_{i=1}^M},z\big) \notag\\
&= \sum_{i=1}^Mz_i\sigma\left(\Phi(\Hat{p}_i)\cdot\Lambda + \sum_{j=1}^Mz_j\Phi(\Hat{p}_j)\cdot\Gamma\right).
\end{align}
Interestingly, one finds that \textcolor{black}{in certain cases the network is equivalent to a standard EFN. For example, replacing activations by the identity gives}
\begin{align}\label{eq:id-act}
\Vec{\mathcal{O}} &= \sum_{i=1}^Mz_i\left(\Phi(\Hat{p}_i)\cdot\Lambda + \sum_{j=1}^Mz_j\Phi(\Hat{p}_j)\cdot\Gamma\right)\notag\\
&=\sum_{i=1}^Mz_i\Phi(\Hat{p}_i)\cdot\Lambda + \sum_{j=1}^Mz_j\Phi(\Hat{p}_j)\cdot\Gamma\notag\\
&=\sum_{i=1}^Mz_i\Phi(\Hat{p}_i)\cdot\left(\Lambda + \Gamma\right)\notag\\
&\equiv\sum_{i=1}^Mz_i{\Psi}(\Hat{p}_i)\ ,
\end{align}
where the first step follows from the fact that the $p_T$ fractions sum to unity, and we
have redefined the per-particle map $\Psi^a(\Hat{p}_i) \equiv \sum_{b=1}^\ell\Phi^b(\Hat{p}_i)\left(\Lambda^{ba} + \Gamma^{ba}\right)$.
This is a consequence of the fact that IRC safety imposes a linear structure on the
transformations prior to $\sigma$. Removing the activation allows the action of the
equivariant layer to be absorbed into the definition of the per-particle map. \textcolor{black}{A similar result is reached even when including activations if one restricts the equivariant layer(s) with $\Gamma=0$ or $\Lambda=0$. See the Appendix\,\ref{app:equi-test} for further comment on this.}
Next, we must check that IRC safety is maintained in a network consisting of more
than one such equivariant layer. Composing two layers leads to observables
\begin{align}
\Vec{\mathcal{O}} &= \big(P\circ E\circ E'\big)(\textcolor{black}{\{\Phi(\Hat{p}_i)\}_{i=1}^M}).\notag\\
&= \sum_{i=1}^Mz_i{E_i}\Big(E'\big(\textcolor{black}{\{\Phi(\Hat{p}_i)\}_{i=1}^M},z\big),z\Big)\notag\\
&= \sum_{i=1}^Mz_i\sigma\Bigg(E'_i\big(\textcolor{black}{\{\Phi(\Hat{p}_i)\}_{i=1}^M},z\big)\cdot\Lambda \notag\\
&\phantom{=\sum_{i=1}^Mz_i\sigma\Bigg(}+ \sum_{j=1}^Mz_jE'_j\big(\textcolor{black}{\{\Phi(\Hat{p}_i)\}_{i=1}^M},z\big)\cdot\Gamma\Bigg)\notag\\
&=\sum_{i=1}^Mz_i\sigma\Bigg[\sigma\left(\Phi(\Hat{p}_i)\cdot\Lambda'+ \sum_{k=1}^Mz_k\Phi(\Hat{p}_k)\cdot\Gamma'\right)\cdot\Lambda\notag\\
&\quad\;+\sum_{j=1}^Mz_j\sigma\left(\Phi(\Hat{p}_j)\cdot\Lambda'+ \sum_{k=1}^Mz_k\Phi(\Hat{p}_k)\cdot\Gamma'\right)\cdot\Gamma\Bigg]\ ,
\label{2equiircobs}
\end{align}
where $E'$ denotes the first equivariant layer, which has matrix parameters
$\Lambda'$ and $\Gamma'$. Thus the vector $\Vec{\mathcal{O}}$ is IRC safe. To see this,
notice that the summand for each sum over particles is weighted by $z$ and that these
weights constitute all of the $z$ dependence in the expression. The same treatment can
be applied for additional stacked layers. Although it would be impractical to present
the equations here, one finds that the sums over particles are always $z$-weighted,
leading to IRC safety. This is because adding an additional equivariant layer corresponds
to replacing each $\Phi(\Hat{p}_{\{i,j,k\}})$ by $E_{\{i,j,k\}}$. Since $E_{\{i,j,k\}}$
contains only IRC safe observables and $\Phi(\Hat{p}_{\{i,j,k\}})$, this replacement
leaves the same $z$-weighted structure of the latent vectors.
\section{Training and performance}
Here we evaluate the performance of equivariant EFNs in discriminating jets from boosted
$W$ bosons against the QCD background. We will consider both IRC safe and unsafe equivariant
networks. The IRC unsafe networks will be contrasted with the performance of Particle Flow Networks
(PFNs)~\cite{efn}. PFNs are similar to EFNs, but based around the general observable decomposition
of Eq.\,\ref{eq:pfn-obs} and are thus IRC unsafe. We begin by detailing the data generation
and preparation steps before outlining the construction of each model.
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{feynman-diagrams.pdf}
\caption{\footnotesize{Tree-level Feynman diagrams for the $W$ (blue) and QCD (red) contributions to the processes of interest. Rearrangement of the final diagram gives an additional $s$-channel contribution, $qg\to qZ$}.}
\label{fig:feyn}
\end{figure*}
\subsection{Data generation and pre-processing}
The $W$ and QCD jets used for this work are obtained respectively from $pp\to Z+W$ and $pp\to Z + \text{jet}$
events generated and hadronised with \texttt{PYTHIA~8.2\textcolor{black}{26}}\:\cite{pythia} at centre of mass energy
$\sqrt{s}=13\;\text{TeV}$. For the $W$ events, we use the \texttt{WeakDoubleBoson:ffbar2ZW}
channel with the $W$ decaying to quarks. For the QCD events we use \texttt{WeakBosonAndParton:qg2gmZq}
and \texttt{qqbar2gmZg} with the photon contribution switched off. The Feynman diagrams for the
tree-level contributions to these processes are shown in Fig.\,\ref{fig:feyn}. In both cases, the
$Z$ is required to decay invisibly to neutrinos. Multiple parton interactions as well as initial
and final state radiation are left on.
To study network performance on jets of different transverse momenta we produce two datasets:
the first containing jets with $p_T\in[250,300]$ GeV and the second with $p_T\in[500,550]$ GeV.
The datasets are produced as follows: after each event is generated, particles with $p_T$ less
than $500$ MeV are removed with the intent of mitigating the influence of contamination from the
underlying event. The event is then clustered by the anti-$k_t$ algorithm \cite{anti} using
\texttt{FastJet~\textcolor{black}{3.3.3}}\,\cite{fastjet} with a jet radius of $R=1.0$ for the $p_T\in[250,300]$ GeV dataset
and $R=0.6$ for the $p_T\in[500,550]$ GeV set.
If the hardest jet in the event has mass~$ m\in[65,105]$ GeV, rapidity $\left|y\right|<2.5$, and
transverse momentum within the relevant range, the set of its constituent particle feature vectors
is added to the dataset, with each vector in the format: $(p_T,y,\phi,m,\text{\texttt{pdg\_id}})$.
The field \texttt{pdg\_id} follows the Particle Data Group Monte Carlo particle numbering scheme \cite{pdg}.
We centre the selection window slightly beyond the mass of the $W$ since the final jet will contain
some additional radiation from the underlying event, translating the distribution. The final datasets
contain 1 million $W$ and QCD jets in equal proportion and are split into three subsets for training
(75\%), validation (15\%), and testing (10\%).
We employ four pre-processing steps intended to isolate the differences between the $W$ and QCD jets.
Each jet in the dataset is first re-clustered using the $k_t$ algorithm\:\cite{genkt} with a jet radius
of $R=0.2$ or $R=0.3$ for the high- or low-$p_T$ datasets respectively. The resulting subjets are ordered
by $p_T$ and the following steps applied:
\begin{enumerate}
\item \textbf{Normalise} the $p_T$ of all particles in the jet such that their sum is unity.
\item \textbf{Translate} all particles in the $y$-$\phi$ plane such that the leading subjet lies at the origin.
\item \textbf{Rotate} all particles in the jet such that the secondary subjet lies directly below the origin.
\item \textbf{Reflect} all particles about the axis $\phi=0$ such that the third subjet has $\phi>0$. If there is no third subjet, reflect such that the $p_T$-weighted sum of $\phi$ over all particles in the jet is positive.
\end{enumerate}
The normalisation step serves to partially reduce the dependence on the $p_T$ of the jet.
Although the radial scale of the event still varies with the transverse momentum, we do not
remove this dependence (through zooming~\cite{unimelb} or a similar procedure) since the
angular scale of the radiation pattern is correlated with the jet type whereas the jet $p_T$
itself is not. The remaining steps remove redundancies arising from spatial symmetries, namely
the location and orientation of the event in the detector.
\subsection{Model details and performance}\label{sec:models}
To evaluate the performance of the equivariant architecture, we employ six neural network
models which we call EFN, PFN(-ID), EV-EFN and EV-PFN(-ID). The details of each model are
given below. All nodes in the networks (except for those in an output layer) use the ReLU
activation\:\cite{relu} and all weights and biases are initialised by He-uniform parameter
initialisation\:\cite{he-uni}. The Adam optimisation algorithm\:\cite{adam} is used to minimise
the binary cross-entropy loss function with data split into batches of 250. Equivariant layers
are fixed to a size of $M$=140 (larger than the number of particles in any of jets in the datasets).
The EFN, PFN and PFN-ID models are implemented using the \texttt{Energyflow} package\footnote{Available at \url{https://energyflow.network}.}
for \texttt{Python}\:\cite{efn}. The remaining networks, which involve equivariant layers, were
trained through \texttt{Keras~\textcolor{black}{2.2.4}}\:\cite{keras} with a \texttt{Tensorflow~\textcolor{black}{1.15}} \cite{tensorflow} backend.\footnote{The code used to construct these models can be found at\\ \url{https://github.com/ayo-ore/equivariant-efns}.}
\begin{figure}[!t]
\centering
\includegraphics[width=\linewidth]{paper-equi-roc-withpfn-withbands.pdf}
\caption{ROC curves averaged over 10 networks for each of the EV-EFN (green), EFN (blue) and PFN (orange) models trained on the $p_T\in[500,550]$ GeV dataset. The bands show one standard deviation. The lower plot shows the ratios relative to the EFN ROC curve.}
\label{fig:equi-roc}
\end{figure}
We decide on the number of epochs over which to train a model by monitoring the accuracy and loss
on the validation set during training and stopping when the results begin to diverge from the test
set. The choices of number and size of layers in the EV models are such that the networks contain \textcolor{black}{a comparable number of parameters as their unaugmented partners. We confirm that equalising the parameter counts (either by increasing or decreasing $\ell$ in the EFN or EV-EFN respectivaly) yields no significant changes to the results presented in this paper.}
\begin{table}[t!]
\begin{ruledtabular}
\begin{tabular}{cccc
Model & Jet $p_T$ [GeV] & AUC & $R_{\varepsilon_{W}=0.5}$\\
\colrule
\multirow{2}{*}{EFN} & 500 -- 550 & $0.9339\pm0.0007$ & $87.4\pm5.7$\\
& 250 -- 300 & $0.9092\pm0.0014$ & $49.1\pm1.6$\\ \hline
\multirow{2}{*}{EV-EFN } & 500 -- 550 & $0.9367\pm0.0009$ & $93.4\pm2.5$\\
& 250 -- 300 & $0.9111\pm0.0015$ & $51.5\pm1.9$\\ \hline
\multirow{2}{*}{PFN} & 500 -- 550 & $0.9366\pm0.0012$ & $93.2\pm4.6$\\
& 250 -- 300 & $0.9117\pm0.0011$ & $51.4\pm1.6$\\ \hline
\multirow{2}{*}{EV-PFN} & 500 -- 550 & $0.9333\pm0.0026$ & $80.1\pm3.8$\\
& 250 -- 300 & $0.9083\pm0.0008$ & $46.2\pm1.6$\\ \hline
\multirow{2}{*}{PFN-ID} & 500 -- 550 & $0.9439\pm0.0014$ & $119.0\pm9.0$\\
& 250 -- 300 & $0.9180\pm0.0008$ & $60.0\pm1.8$\\ \hline
\multirow{2}{*}{EV-PFN-ID} & 500 -- 550 & $0.9402\pm0.0010$ & $96.0\pm4.0$\\
& 250 -- 300 & $0.9152\pm0.0012$ & $53.5\pm2.6$\\
\end{tabular}
\end{ruledtabular}
\caption{Performance metrics obtained by the models listed in text on both $p_T$ datasets. The background rejection metric $R$ is defined as $1/\varepsilon_{QCD}$ at 50\% signal efficiency. For both the AUC score and $R$, a greater value corresponds to better classification. The reported uncertainties are interquartile ranges of values from 10 separately trained models.}
\label{tab:equi_performance}
\end{table}
The networks
are defined as follows.
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{paper-evpfn-overfit.pdf}
\caption{Classification accuracies of the PFN (left) and EV-PFN (right) models on the training and validation subsets of the $p_T\in[500,550]$ GeV data over 50 epochs.}
\label{overfit}
\end{figure*}
\begin{itemize}[align=left]
\item[\textbf{EFN:}] An $\ell=256$ energy flow network where the network representing the per-particle
map has two hidden layers of 100 nodes, and the network representing $F$ has three, also with 100 nodes
each. \textcolor{black}{The model contains 82,358 parameters and} is trained for 30 epochs.
\item[\textbf{PFN and PFN-ID:}] These two networks have the same general construction as the EFN network
above, however the constituent particles' $p_T$ fractions are given as input to the per-particle map
instead of weighting the sum over filters. These networks represent the observable decomposition in
Eq.\,\ref{eq:pfn-obs}, which is not IRC safe in general. In the ID variant, the input feature vector
contains both momentum and particle identification. The identities are represented by small floats
beginning at $0$ for an absent particle, then increasing by $0.05$ for each category of constituent:
$\gamma$, $p$, $\Bar{p}$, $n$, $\Bar{n}$, $\pi^+$, $\pi^-$, $K^+$, $K^-$, $K^0_L$, $e^+$, $e^-$, $\mu^+$,
$\mu^-$.
The models \textcolor{black}{contain 82,458 parameters and} are trained for 25 epochs.
\item[\textbf{EV-EFN:}] This is an $\ell=128$ EFN containing two IRC safe equivariant layers as
defined in Eq.\,\ref{eq:irc-equi-def}. We use a smaller latent dimension of $\ell=128$ and reduce
the depth of the MLP representing the function $F$ such that the total number of parameters in the
model is comparable to that of the $\ell=256$ EFN. Specifically, the networks for $\Phi$ and $F$
both contain two hidden layers of 100 nodes. The equivariant layers have size $M=140$ (corresponding
to the maximum number of jet constituents) and 100 output channels to match the size of the layers in
$F$. The model \textcolor{black}{contains 89,330 parameters and} is trained for 30 epochs.
\item[\textbf{EV-PFN and EV-PFN-ID:}]
These are similar to the $\ell=128$ PFN(-ID) networks, but extended with two equivariant layers in
the same fashion as the EV-EFN. \textcolor{black}{
The equivariant layers are of the type in Eq.\,\ref{eqvmatmax} (which is not IRC-safe) and max-pooling is used as the projection operation. Additionally, we use 14 dimensional basis vectors to represent each category of constituent particle in the ID variant. Although these choices for PID encoding and projection are different to those in the PFN(-ID) models, we find that they lead to better stability of training. Models using other choices (such as sum-pooling or the use of equivariant layers of the type in Eq.\,\ref{eqvmatsum}) often failed to converge, having AUC scores of 0.5 after training. In the instances where training did converge, results similar to those achieved by the architecture outlined here were observed. The EV-PFN(-ID) models contain 89,430 (90,830) parameters and, due to poor generalisation, are trained for 20 epochs.
}
\end{itemize}
The AUC scores and background rejections, ${R\equiv1/\varepsilon_{QCD}}$, for the models trained on the
low- and high- $p_T$ datasets are presented in Tab.~\ref{tab:equi_performance}. Comparing the IRC safe models,
we see that the EV-EFN achieves slightly better classification than the EFN across both measures in both $p_T$
ranges. We also show the ROC curves averaged over 10 networks for these models (as well as the PFN) in
Fig.\,\ref{fig:equi-roc}. The curves here again reflect the classification advantage that the EV-EFN achieves
over the EFN. This suggests that the addition of equivariant layers aids the network's ability to learn from
IRC safe information. In fact, the improvement obtained by the EV-EFN is greater than the improvement seen from
the PFN model, which is not IRC safe and uses a latent space of twice the dimensionality. However, the results in
Tab. \ref{tab:equi_performance} show that the EV-PFN does not discriminate better than the PFN model. In fact, on
both datasets it performs worse than even the IRC safe EFN across both metrics. The result for the EV-PFN-ID model
is not as poor -- although it does not match the classification power of the PFN-ID networks, it does at least
outperform the PFN.
A major issue that the EV-PFN(-ID) networks face is \textcolor{black}{poor generalisation}. To illustrate this, in Fig.\,\ref{overfit} we
plot the classification accuracy of the PFN and EV-PFN models on the training and validation datasets over 50
epochs. The test accuracy for the EV-PFN diverges from the validation accuracy after just 15 epochs of training,
whereas this behaviour is not seen to such a strong degree for the PFN despite both models containing \textcolor{black}{approximately} the same
number of parameters.
The plots reveal that the EV-PFN does indeed classify jets from the training set more accurately than the PFN,
however its poor generalisation means that this performance is not transferred to test results. This could possibly be addressed by regularisation techniques. One such option is to apply
\emph{dropout} -- a method wherein nodes in a layer are randomly selected to be switched off during
training\:\cite{dropout}. We investigate application of dropout on layers in the network for $F$, but find
that for a variety of dropout rates, only the performance on the training set is affected. The same result is observed
for the ID variants. In the interest of avoiding potentially unfruitful trial and error, we leave \textcolor{black}{fine-tuning}
of the EV-PFN(-ID) models to future work.
\subsection{Dependence on jet mass}
\begin{figure*}[t!]
\centering
\includegraphics[width=\textwidth]{thesis-network-sculpting-nolog.pdf}
\caption{Jet mass histograms (normalised) for the $p_T\in[500,550]$ GeV QCD jets that are tagged as $W$'s by successively harder cuts on the EFN (left) or EV-EFN (right) output. The distribution for $W$ jets is shown in grey.}
\label{fig:network-sculpting}
\end{figure*}
In this section, we compare the effect of the EV-EFN and EFN on the QCD jet mass distribution. We follow closely
the approach taken in Ref.\,\cite{mass-agnostic}. An unrestricted neural network will make use of all
available information, jet mass included. Consequently, the output of such networks will be correlated
with the jet mass, leading to a sculpting effect of the jet mass distribution. This is not a desirable
feature, since background modeling is often done by extrapolating from a region containing no signal,
relying on an approximately flat background. A background distribution that is smooth and featureless
is associated with less systematic uncertainty. There exists a trade-off between the raw discrimination
power a neural network offers and the degree to which it reshapes the background jet mass distribution.
For this reason, jet observables that are strongly correlated with the jet mass are less valuable for
experimental use compared to those that are uncorrelated. This problem has been studied extensively from
both analytic~\cite{DDT, analytic-decorrelation} and machine learning perspectives~\cite{data-planing, pca,
adversarial, uboost, DisCo, MUST,ATLAS:2018ibz, mass-agnostic, Dorigo:2020ldg}.
Here we compare results first for the standard training procedure described above and then including the
application of a method for decorrelating network output from the jet mass. To compare the two models,
we train 10 EFNs and 10 EV-EFNs on the $p_T\in[500,550]$ GeV dataset, split into training (75\%) and
testing (25\%) subsets. By increasing the size of the testing set, we reduce the impact of statistical
fluctuations on the results. This is important since, at low signal efficiencies, a very small number
of QCD jets are tagged. In Fig.\,\ref{fig:network-sculpting} we visualise the sculpting effect of the
EFN and EV-EFN models on the training set by showing the jet mass distributions of QCD jets that are
tagged as signal by successively tighter cuts the network output. Both models strongly sculpt the QCD
distribution to the $W$ mass peak. In order to quantify the degree of sculpting we make use of the
Hellinger distance, defined as
\begin{equation}
\newcommand*\bigstrut{\vrule height 1.3\baselineskip depth.3\baselineskip width 0pt\relax}
H(p,q) = \sqrt{\bigstrut 1-\sum_i\sqrt{p(i)q(i)}}\ ,
\end{equation}
where $p$ and $q$ are discrete probability distributions. Hellinger distances are normalised between
0 and 1, with a distance of $0$ corresponding to $p$ and $q$ being identical histograms. For each
trained model, the Hellinger distances between the jet mass distributions of tagged QCD jets and the
original are calculated (using histograms of 25 bins) for a range of cuts on network output. We then
average the distances over network instances.
\begin{figure*}[tp!]
\centering
\includegraphics[width=\textwidth]{thesis-hellingers.pdf}
\caption{Hellinger distances between the tagged QCD jet mass distribution and the full distribution as a function of the background rejection achieved by cutting on the output of an EFN or EV-EFN. Results are shown for the standard training procedure (left) and with the inclusion of mass-planing (right). We indicate with markers the positions along the curves corresponding to fixed signal efficiencies of $\varepsilon_W=0.3$ or $\varepsilon_W=0.5$. The ratios to the EFN-line are shown beneath each plot.}
\label{fig:hellinger}
\end{figure*}
A plot of these average distances as a function of the network background rejection for the two models
is shown on the left of Fig.\,\ref{fig:hellinger}. In these terms a better classifier corresponds to a
curve that is closer to the bottom-right corner, reaching greater background rejection at smaller
Hellinger distance. The general behaviour of the two models is very similar, however the EV-EFNs achieve
smaller Hellinger distances across all cuts on the network output. The greatest difference is seen at a
background rejection of $1/\varepsilon_{QCD}\sim90$ (corresponding to a signal efficiency of
$\varepsilon_W=0.5$), where they improve upon the EFNs by approximately 5\%.
While the EV-EFNs have a smaller effect on the background distribution compared to the EFNs, both
models sculpt to such a great degree that this improvement can be considered somewhat inconsequential.
Having said this, a more meaningful result can be obtained by investigating the networks' sculpting after
decorrelating the network output from the jet mass. Ref.\,\cite{mass-agnostic} found that for neural
network taggers, techniques that achieve this through adjustments to the input data provide decorrelation
almost on par with those that alter the training procedure, while having significantly lower computational
cost. For this reason, we compare the response of EFNs and EV-EFNs to \emph{mass-planing} -- a particular
case of \emph{data planing} \cite{data-planing,jet-images}. A dataset can be planed in some observable,
$\rho$, by assigning each instance of the training data $x_i$ with a weight $w$ according to
\begin{equation}
\left[w(x_i)\right]^{-1} = \frac{d\sigma_x}{d\rho}\bigg|_{\rho=\rho(x_i)},
\end{equation}
where $\frac{d\sigma_x}{d\rho}$ is the differential cross-section for jets in $x$. In practice, this
simply corresponds to inverting the histogram for $\rho$ at the bin in which $x_i$ lies. After such a
weighting, the data has a uniform distribution in the planed observable.
To implement mass-planing in the networks' training procedure, we calculate weights from the $W$ and
QCD mass distributions histogrammed into 100 bins. The weights are used to give the relative importance
of a jet's contribution to the loss function during training. The Hellinger distances are then calculated
and averaged in the same fashion as for the un-planed case. On the right side of Fig.\,\ref{fig:hellinger},
we again present a plot of the calculated distances against background rejection for the newly trained
networks. Interestingly, the advantage that the EV-EFN displayed without planing is no longer apparent.
In fact, the standard EFN exhibits less background sculpting across most cuts on the network output.
One should therefore be cautious about interpreting the absolute performance of some given taggers
without considering decorrelation, as this may change the order of the rankings. However, the EV-EFN
still achieves greater background rejection at fixed signal efficiency. This implies that in a bump
hunt scenario, neither model has a clear advantage in terms of achieved significance.
\section{Conclusions}
The development of new and increasingly accurate jet taggers remains a high priority for the collider
physics community. The use of techniques based on machine-learning methods has provided a number of
advances in this area in the past few years. Recent research has explored incorporating permutation
invariance of the constituent particle momenta into the network architecture via the Deep Sets formalism,
leading to Energy Flow Networks (EFNs). However, Deep Sets also allows for the possibility of permutation
equivariance, which is what we have explored in this paper.
We have successfully implemented permutation-equivariant neural network layers into the Energy Flow
Network architecture. Furthermore, we identified a specific construction, the EV-EFN, that maintains
the infrared and collinear safety of the learned jet observable. When trained to discriminate $W$ from
QCD jets in simulated samples in two boosted $p_T$ windows these EV-EFNs exhibited a marginal improvement
in performance over permutation-invariant EFNs.
The performance of the EV-EFN was equal to that achieved by a Particle Flow Network, which is not
restricted by IRC safety. \textcolor{black}{Some issues were observed when extending PFNs by equivariant layers. Specifically, certain choices of equivariant operation or projection prevented the model from converging during training. Further, models that did converge generalised poorly to the testing sample. These issues could possibly be remedied by alternative constructions or regularisation techniques, however we leave that task for future work.}
Finally, we compared the action of the networks on the QCD jet mass distribution before and after
applying mass planing. The planing decorrelates the tagger performance from the jet mass, at some
cost in performance. However, it also changes the relative performances of the EFN and EV-EFN networks.
While the EV-EFN initially had slightly less impact on the background distribution compared to the EFN,
this advantage was lost after mass planing the input jets. Consequently it would be interesting to apply
recent ideas such as Distance Correlation\,\cite{DisCo}, adversarial networks\,\cite{adversarial}, Mass
Unspecific\,\cite{MUST} (or Agnostic\,\cite{mass-agnostic}) Supervised Tagging to see if networks with
equivariant layers continue to outperform invariant ones.
A number of different future directions present themselves. Future work could investigate the use of
other types of equivariant layers and their impact on IRC safety including those which are equivariant
to continuous symmetries of the Lorentz Group as in Ref.\,\cite{Bogatskiy:2020tje} and similar to
Ref.\,\cite{Butter:2017cot}. It may also be interesting to attempt to translate the EV-EFN into a
low-dimensional human-interpretable space using the method outlined in Ref.\,\cite{human-read}. This
could potentially illuminate the effect that equivariant layers have on the representation of IRC
safe jet observables compared to the original EFN.
\section*{Acknowledgements}
We thank John Gargalionis for useful comments and discussions. This work was supported in part by
the Australian Research Council and the Australian Government Research Training Program Scholarship initiative.
|
\section{Introduction}
Let $G$ be a graph with vertices $\{v_1,v_2,\ldots,v_n\}$ that are
totally ordered $v_1<v_2<\cdots <v_n$. In~\cite{St1}, Stanley defined the chromatic symmetric function $\csf{G}$ of $G$ as
\begin{equation}
\csf{G} = \sum_{\kappa: V\to \mathbb{P}, \text{ proper}} {\bf
x}^{\kappa} = \sum_{\kappa: V\to \mathbb{P}, \text{ proper}} x_1^{\#\kappa^{-1}(1)} x_2^{\#\kappa^{-1}(2)}\cdots,
\end{equation}
where $\mathbb{P}=\{1,2,3,\ldots\}$, ${\bf x}=(x_1,x_2,\ldots)$, and
the sum is over the proper colorings of the vertices of $G$.
Stanley and Stembridge~\cite{StSt} conjectured that the chromatic symmetric
functions expand with positive coefficients in the basis $\{e_{\lambda}\}$
of elementary symmetric functions for the following graphs. Given a Dyck path $d$ from $(0,0)$ to $(n,n)$, let $G(d)$ be the graph with vertices $\{1\ldots n\}$ and edges $(i,j)$, $i<j$ iff the cell
$(i,j)$ is below the path $d$. These are also the incomparability graphs of {\em unit interval orders} or graphs obtained from {\em Hessenberg sequences}.
Shareshian--Wachs~\cite{ShW2} introduced a quasisymmetric version of $\csf{G}$
defined by
\begin{align*}
\csft{G}{q} = \sum_{\kappa: V\to \mathbb{P}, \text{ proper}} q^{\asc(\kappa)}{\bf x}^{\kappa},
\end{align*}
where $\asc(\kappa)$ is the number of edges $\{v_i,v_j\}$ of $G$ with
$i<j$ and $\kappa(v_i)<\kappa(v_j)$.
For the graphs $G(d)$ coming from Dyck paths, the quasisymmetric function $\csft{G(d)}{q}$ is
actually symmetric and Shareshian--Wachs gave a refinement of the
Stanley--Stembridge conjecture for this Catalan family of graphs.
\begin{conjecture}[Stanley--Stembridge,
Shareshian--Wachs] \label{conj: StaSteShWa}
Let $d$ be a Dyck path then the coefficients of $\csft{G(d)}{q}$ in
the elementary basis are in $\mathbb{N}[q]$.
\end{conjecture}
The symmetric functions $\csft{G(d)}{q}$ are very actively studied thanks to their connections to \emph{Hessenberg varieties}~\cite{ShW2}, \emph{diagonal harmonics}~\cite{CM}, and {\em Macdonald polynomials}~\cite{AP}.
Conjecture~\ref{conj: StaSteShWa} has been verified independently and by different techniques by Cho--Huh~\cite{ChoHuh}, Hamada--Precup~\cite{HP} , and Abreu--Nigro~\cite{AN} for the case of {\em abelian} Dyck paths: paths $d$ of size $m+n$ of the form $\mathsf{n}^m w(\lambda) \mathsf{e}^n$ where $w(\lambda)$ is the encoding
in North ($\mathsf{n}$) and East ($\mathsf{e}$) steps of the partition $\lambda \subset n\times m$ (see
Figure~\ref{fig:hitex}). We denote the associated graph by $G(\lambda)$ and the chromatic symmetric function by $\csft{\lambda}{q} = \csft{G(\lambda)}{q}$.
The symmetric functions of abelian Dyck paths are deeply related to the {\em $q$-rook theory} of Garsia--Remmel~\cite{GR} as was unveiled in the Abreu--Nigro expansion, itself a $q$-analogue of a result of Stanley--Stembridge \cite{StSt}. The following statements use the standard notation $[n]_k = [n][n-1]\cdots[n-k+1]$, $[n]!=[n]_n$, $\txtqbin{n}{k} =[n]_k/[k]!$, where $[x]=(1-q^x)/(1-q)$. Also $\qhit{j}{n}{\lambda}$ denotes $q$-hit numbers~\cite{D} which are equal to the Garsia--Remmel $q$-hit numbers~\cite{GR} up to a power of $q$. Moreover, the $\qhit{j}{n}{\lambda}$ are symmetric polynomials in $\mathbb{N}[q]$ that at $q=1$ give the number of permutations of size $n$ with permutation matrix having support of size $j$ in the board of$\ \lambda$.
\begin{theorem}[Abreu--Nigro~\cite{AN}]\label{AN:generalLambda}
Let $\lambda$ be partition inside an $n\times m$ board with $\ell(\lambda) = k \leq \lambda_1$. Then
\begin{align*}
\csft{\lambda}{q} &=
\qfactorial{k}\qhit{k}{n+m-k}{\lambda}\cdot e_{m+n-k,k} + \sum_{j=0}^{k-1}
q^j \qfactorial{j}\qnumber{n+m-2j} \qhit{j}{m+n-j-1}{\lambda} \cdot e_{m+n-j,j}.
\end{align*}
\end{theorem}
Central to this paper is a new identity of Guay-Paquet (private communication~\cite{MGP_LR}) that expands $\csft{\lambda}{q}$ in terms of
chromatic symmetric functions for rectangular shape with coefficients given by {\em $q$-hit numbers} of {\em rectangular boards} of size $n\times m$ that we denote by $\qhit{j}{m,n}{\lambda}$ and satisfy $\sum_{j=0}^n \qhit{j}{m,n}{\lambda}=[m]_n$.
\begin{theorem}[Guay-Paquet \cite{MGP_LR}] \label{thm:qhitCSFabelian}
Let $\lambda$ be partition inside an $n\times m$ board ($n\leq m$). Then
\[
\csft{\lambda}{q} =\dfrac{1}{\qfalling{m}{n}} \sum_{j=0}^{n}
\qhit{j}{m,n}{\lambda} \cdot \csft{m^j}{q}.
\]
\end{theorem}
The original proofs of the two statements above use a linear relation satisfied by $\csft{G(d)}{q}$ called the {\em modular relation}~\cite{AN,AS,MGP}. Our {\bf \em first main result} is an elementary proof of Theorem~\ref{thm:qhitCSFabelian} using a desymmetrizing recursive relation and $q$-rook theory (Section~\ref{sec: pf of MGP}) and {\bf\em our second main result} is the equivalence of this result and Theorem~\ref{AN:generalLambda} (Section~\ref{sec: MGP to AN}). As a by-product of our arguments, we obtain a new proof of the Abreu--Nigro expansion, a new recurrence to compute $\csft{\lambda}{q}$ (Lemma~\ref{lem:X_recursion}), and new relations of {\em $q$-rook numbers} and {\em $q$-hit numbers} (Lemma~\ref{prop:q-rook identities} and~\ref{prop:qhit-relations}) that develop further the $q$-rook theory of rectangular boards~\cite{LM17}.
The full version of this paper is available at~\cite{CMP}.
\section{Background on $q$-rook theory}
For the rest of the paper, we assume $m$ and $n$ are non-negative integers with $m\geq n$.
\subsection{$q$-rook numbers}
\begin{definition}[$q$-rook numbers \cite{GR}]\label{def: GR q-rook numbers}
Given a partition
$\lambda=(\lambda_1,\lambda_2,\ldots,\lambda_{n})$ inside an $n\times m$ board, the \emph{Garsia-Remmel $q$-rook numbers}
are defined as $R_k(\lambda) = \sum_p q^{\inv(p)}$,
where the sum is over all placements $p$ of $k$ non-attacking rooks on $\lambda$ and
$\inv(p)$ is the number of cells of $\lambda$ left after each rook cancels its cell, the cells North in its column and the cells West in its row (see Figure~\ref{fig:ex rook placement}).
\end{definition}
\begin{proposition}[Garsia-Remmel~\cite{GR}]\label{prop:GR_rook_gen_fun}
Given a partition $\lambda=(\lambda_1,\ldots,\lambda_{\ell})$ we have that
\begin{equation} \label{eq: def F}
F(x;\lambda):=\sum_{k=0}^{\ell} R_{k}(\lambda)[x]_{\ell-k} = \prod_{i=1}^{\ell} [x+\lambda_{\ell-i+1}-i+1].
\end{equation}
\end{proposition}
\begin{lemma}\label{lem:F_ratio}
Given a partition $\lambda=(\lambda_1,\ldots,\lambda_{\ell})$ we have that
$$q^{\lambda_1} [x] \frac{ F(x-1;\lambda)}{F(x;\lambda)} = [x-\ell+\lambda_1] - \sum_{j=1}^{\lambda_1} q^{\lambda_1-j}\prod_{t=1}^{\lambda_j'} \frac{[x+\lambda_{t}-1-\ell+t] }{[x+\lambda_{t}-\ell+t]}.$$
\end{lemma}
\begin{proof}
We use induction on $\ell(\lambda)$ and apply Proposition~\ref{prop:GR_rook_gen_fun}. For $\ell(\lambda)=1$, we have
\begin{multline*}
[x] \frac{ F(x-1;\lambda)}{F(x;\lambda)} = [x-1] + \sum_{j=1}^{\lambda_1} q^{-j} - q^{-j} \frac{[x+\lambda_{1}-1] }{[x+\lambda_{1}]} = q^{-\lambda_1} \frac{[x+\lambda_1-1]}{[x+\lambda_1]}([x+\lambda_1]-[\lambda_1]).
\end{multline*}
Next, expanding the RHS of the above identity and doing standard manipulations gives
\begin{multline*}
[x-\ell+\lambda_1] - \sum_{j=1}^{\lambda_1} q^{\lambda_1-j}\prod_{t=1}^{\lambda_j'} \frac{[x+\lambda_{t}-1-\ell+t] }{[x+\lambda_{t}-\ell+t]} = \\
= q^{\lambda_1-\lambda_2} \frac{[x+\lambda_1-\ell]}{[x+\lambda_1-\ell+1]} \left( [x+\lambda_2 -(\ell-1)] - \sum_{j=1}^{\lambda_2} q^{\lambda_2-j} \prod_{t=2}^{\lambda_j'} \frac{[x+\lambda_{t}-1-\ell+t] }{[x+\lambda_{t}-\ell+t]} \right).
\end{multline*}
By induction hypothesis the parenthetical on the RHS above is $q^{\lambda_2} [x] F(x-1;\tilde{\lambda})/F(x;\tilde{\lambda})$ where $\tilde{\lambda} = (\lambda_2,\ldots,\lambda_\ell)$. Using $\tilde{\lambda}_t = \lambda_{t+1}$ for the reindexing we obtain the result.
\end{proof}
\subsection{$q$-hit numbers}
The $q$-hit numbers are defined in terms of the $q$-rook numbers by a change of basis. Let $(a;q)_k=\prod_{i=0}^{k-1}(1-aq^i)$ denote the {\em $q$-Pochhammer symbol}.
\begin{definition}[{\cite[Def. 3.1, Prop. 3.5]{LM17}}]\label{def: hit in terms of rs}
For $\lambda$ inside an $n\times m$ board, we define the \emph{$q$-hit polynomial} of $\lambda$ by
\begin{equation} \label{eq: hit rook change of basis}
\sum_{i=0}^n \qhit{i}{m,n}{\lambda} x^i := \frac{q^{-|\lambda|}}{\qfactorial{m-n}}\sum_{i=0}^n \qrook{i}{\lambda} \qfactorial{m-i} (-1)^i q^{mi-\binom{i}{2}} (x;q)_i,
\end{equation}
where the coefficients $\qhit{i}{m,n}{\lambda}$ are the \emph{$q$-hit numbers} associated to $\lambda$.
\end{definition}
\begin{notation}
For square boards with $m=n$, we denote the $q$-hit number by $\qhit{j}{m}{\lambda}$.
\end{notation}
\begin{remark} \label{rem: difference qhits}
For the case $n=m$, Garsia--Remmel defined $q$-hit numbers $\nqhit{k}{n}{\lambda}$ by the relation
\begin{equation} \label{eq: def GR qhit}
\sum_{i=0}^n \nqhit{i}{n}{\lambda} x^i = \sum_{i=0}^n \qrook{i}{\lambda} [n-i]! \prod_{k=n-i+1}^n (x-q^k).
\end{equation}
One can show that the Garsia--Remmel $q$-hit numbers and our $q$-hit numbers differ by a power of $q$, namely $\widetilde{H}^n_{k}(\lambda)=q^{|\lambda|-kn} H^n_k(\lambda)$ (see \cite[Appendix]{CMP}).
\end{remark}
For the case of square boards, Garsia and Remmel showed that $\widetilde{H}^n_k(\lambda)$ are in $\mathbb{N}[q]$. Later, Dworkin~\cite{D} and Haglund~\cite{H} found different Mahonian statistics on rook placements that realize the polynomials $\widetilde{H}_j(\lambda)$. Guay-Paquet~\cite{MGP_LR} defined the rectangular $q$-hit numbers using a statistic similar to Dworkin's statistic in \cite{D} that we define next.
\begin{definition}[Statistic for the $q$-hit numbers]\label{def: Dworking stat qhit}
Let $\lambda$ be a partition inside an $n\times m$ board. Given a placements $p$ of $n$ nonattacking rooks on a $n\times m$ board, with exactly $j$ inside $\lambda$, let $stat(p)$ be the number of cells $c$ in the $n\times m$ board such that (i) there is no rook in $c$,
(ii) there is no rook above $c$ on the same column, and either,
(iii) if $c$ is in $\lambda$ then the rook on the same row of $c$ is in $\lambda$ and to the right of $c$ or (iv) if $c$ is not in $\lambda$ then the rook on same row of $c$ is either in $\lambda$ or to the right of $c$.
\end{definition}
\begin{remark}
Intuitively, Dworkin's statistic $\stat(p)$ is the number of remaining cells in the $n\times m$ board after: wrapping this board on a vertical cylinder and each rook of $p$ cancels the cells South in its column and the cells East in its row until the border of $\lambda$.
\end{remark}
\begin{theorem} \label{thm: qhit statistic rectangular board}
Let $\lambda$ be a partition inside an $n\times m$ board and $j=0,\ldots,n$ then \newline
$\qhit{j}{m,n}{\lambda}= \sum_p q^{stat(p)}$,
where the sum is over all placements $p$ of $n$ non-attacking rooks on a $n\times m$ board, with exactly $j$ rooks inside $\lambda$.
\end{theorem}
The proof of this result is included in \cite[Appendix]{CMP}. Moreover, for each partition $\lambda$, the statistic $\stat(\cdot)$ is \emph{Mahonian}. This results follows readily from Definition~\ref{def: hit in terms of rs}
\begin{corollary}
Let $\lambda$ be a partition inside an $n\times m$ board, then
$\displaystyle{\sum_{j=0}^n \qhit{j}{m,n}{\lambda} = [m]_n}$.
\end{corollary}
\begin{example}
Consider the partition $\lambda=(6,3,3,1)$ inside a $6\times 8$ board. In Figure~\ref{fig:ex rook placement}, we present an example of a placement $p$ of two rooks on $\lambda$ with $\inv(p)=7$ and an example of a placement $p'$ of six rooks on the $6\times 8$ board with two hits on $\lambda$ and $\stat(p')=13$.
\end{example}
\begin{figure}[ht]
\centering
\includegraphics{q-rook-example-NW.pdf}
\quad
\raisebox{5pt}
\includegraphics{q-hit-example-dworkin-easier.pdf}}
\qquad
\includegraphics{q-lam-deletion-contraction.pdf}
\caption{Left: Example of the statistics of a $q$-rook number and a $q$-hit number. Right: Example of the deletion and contraction of the board of a partition $\lambda$.}
\label{fig:ex rook placement}
\end{figure}
We finish this section with some results for $q$-hit numbers.
First, we give a deletion/contraction relation. Given a shape $\lambda$ and a corner cell $e$ in $\lambda$, $\lambda \backslash e$ denotes the shape obtained after deleting the cell $e$ in $\lambda$, and $\lambda /e$ denotes the shape obtained after deleting in $\lambda$ the row and column containing $e$. See Figure~\ref{fig:ex rook placement} for an example.
\begin{lemma} \label{lem: deletion/contration}
We have the following deletion/contraction relation:
\begin{equation}\label{lemma:del con hits}
\qhit{j}{m,n}{\lambda} = \qhit{j}{m,n}{\lambda\backslash e}
+ q^{|\lambda/e|-|\lambda|+j+m-1}\left( \qhit{j-1}{m-1,n-1}{\lambda/e} - q \qhit{j}{m-1,n-1}{\lambda/e}\right).
\end{equation}
\end{lemma}
The next results show the relation between the $q$-hit numbers when we change the dimensions of the board.
\begin{lemma}
Let $\lambda$ be a partition inside an $n\times m$ board. Then,
$\qhit{j}{m,n}{\lambda} = \frac{1}{[m-n]!}\qhit{j}{m,m}{\lambda}$.
\end{lemma}
\begin{lemma} \label{lemma:remove column}
Let $\lambda$ be a partition inside an $(n-1)\times m$ board. Then,
\begin{equation} \label{eq:greta-relation}
\qhit{j}{m,n}{\lambda} = \qnumber{m+1-n} \qhit{j}{m,n-1}{\lambda}.
\end{equation}
\end{lemma}
Finally, we give formulas for $q$-hit numbers for rectangular shapes and omit their standard proofs.
\begin{proposition} \label{prop:qhit_rectangles}
$\qhit{k}{N}{m^j}= q^{ (N-j-m+k)k} \qfalling{m}{k} \qfactorial{N-j} \dfrac{ \qfalling{N-m}{j-k} \qfalling{j}{j-k}}{ \qfactorial{j-k}}.$
\end{proposition}
\begin{proposition} \label{prop: hits small rect in rect}
$\qhit{r}{m,n}{(m-1)^k} = 0$ for $0\leq r\leq k-2$, $\qhit{k-1}{m,n}{(m-1)^k}=[k] [m-1]_{n-1}$, and $\qhit{k}{m,n}{(m-1)^k}=q^k[m-k] [m-1]_{n-1}$.
\end{proposition}
\section{The Guay-Paquet $q$-hit identity} \label{sec: pf of MGP}
In this section we sketch our main result, a proof of Theorem~\ref{thm:qhitCSFabelian} using $q$-rook theory. We start by giving an example of this elegant identity.
\begin{example} \label{ex:tcase}
For $\lambda = (2,1)$ inside a $2\times 3$ board, looking at Figure~\ref{fig:hitex}, we see that
$\qhit{0}{3,2}{\lambda}=q^0=1$, $\qhit{1}{3,2}{\lambda}=2q+2q^2$,
$\qhit{2}{3,2}{\lambda}=q^3$. One can verify that
\[
\csft{21}{q} = \dfrac{1}{[3][2]} \left(
\csft{3^0}{q} + (2q^2+2q)\csft{3^1}{q}+q^3\csft{3^2}{q}\right),
\]
\end{example}
\begin{figure}
\centering
\subfigure{
\raisebox{25pt}{\includegraphics{mnpath_dworkin}}}
\subfigure{
\includegraphics{hitex_dworkin_together}}
\caption{Left: an abelian Dyck path $\lambda$ inside an $n\times m$ board. Top right: the paths for $\lambda=(2,1)$, and for the rectangles $3^0,3^1,3^2$ inside a $2\times 3$ board. Bottom right: the six placements of $2$ rooks in $2\times 3$
divided by how many rooks ``hit'' $(2,1)$ (in gray) and the
associated statistic to each rook placement.}
\label{fig:hitex}
\end{figure}
We consider the chromatic symmetric functions in variables $x_1,\ldots,x_M$ and each monomial appearing as a particular assignment of the variables (i.e. colors) to the vertices. That is, the vertices $1,\ldots,N=m+n$ are colored $\{1,\ldots,M\}$. For simplicity, we denote by $X_\lambda^N(M)$ the chromatic symmetric polynomial $X_{G(\lambda)}(x_1,\ldots,x_M;q)$ where the graph $G(\lambda)$ has $N$ vertices. We will use induction on both $M$ and $n,\ m$ when necessary, driven by the following recursion.
\begin{lemma}\label{lem:X_recursion}
For $\lambda \subset m \times n$ we have the following recursion
\begin{align*
X_{\lambda}^{m+n}(M) =& X_{\lambda}^{m+n-1}(M-1)
+ x_M\sum_{i=1}^{m+n} q^{m+n-i - \lambda_i'} X^{m+n-1}_{\lambda/i}(M-1) \\
&+ x_M^2\sum_{(i,j) \in \lambda} q^{ i-1 + (m+n-j-\lambda_j')} X^{m+n-2}_{\lambda/(i,j)}(M-1),
\end{align*}
where $\lambda/(i,j)$ is the partition obtained by removing row $i$ and column $j$ from $\lambda$, and $\lambda/i$ means we remove from $\lambda$ column $i$, for $i=1,\ldots,m$ or row $m+n-i+1$, for $i=m+1,\ldots,m+n$.
\end{lemma}
\begin{proof}
In the abelian case, the graph $G(\lambda)$ consists of a clique with vertices $\{1,\ldots m\}$, a clique with vertices $\{m+1,\ldots,m+n\}$ and a bipartite graph in between with edges $(i,m+j)$ for each $(i,j)$ in $\overline{\lambda}$. So a coloring of this graph has at most two vertices of the same color. If the colors used are in $\{1,\ldots,M\}$, there are three cases for the color $M$:
\begin{compactitem}
\item[\textbf{1.}] No vertex is colored $M$, this term contributes $X_\lambda^{m+n-1}(M-1)$ to $X_\lambda^{m+n}(M)$.
\item[\textbf{2.}] Only one vertex is colored $M$. Suppose this vertex is in column $j$ (from left) and row $i=N-j$ (from top to bottom). It creates ascents with all vertices above it but not in $\lambda$, giving $N-j-\lambda_j'$ ascents. Deleting this vertex corresponds to deleting its row and column (only one would be a row/column of $\lambda$) and we get a graph on $N-1$ vertices with shape $\lambda/j$ (deleting row $N-j$, column $j$ or row $j$, column $N-j$). These terms contribute
$x_M \sum_{j} q^{N-j-\lambda'_j} X_{\lambda/j}(M-1)$.
\item[\textbf{3.}] Two vertices are colored $M$. Suppose that one is in column $j$ and the other one is in row $i$, necessarily with $(i,j) \in \lambda$. The ascents they contribute are $N-j - \lambda_j'+ i-1$. We can remove these two vertices, by removing row $i$ and column $j$ from $\lambda$ and decreasing $N$ by 2.
These terms contribute
$x_M^2\sum_{(i,j)} q^{N-j-\lambda'_j + i-1 } X^{N-2}_{\lambda/(i,j)}(M-1)$.
\end{compactitem}
\end{proof}
For rectangular shapes $\lambda = (m^k)$, Lemma~\ref{lem:X_recursion} gives the following recursive expansion.
\begin{lemma}\label{lem:X_rect_recursion}
\begin{align*}
X^{m+n}_{m^k}(M) &= X^{m+n}_{m^k}(M-1) + x_M \left(q^{n-k}[m] X^{m+n-1}_{(m-1)^k}(M-1) + [k]X^{m+n-1}_{m^{k-1}}(M-1) \right. \\ &\left. + q^k[n-k] X^{m+n-1}_{m^k}(M-1) \right)+
x_M^2 q^{n-k}[k][m] X^{m+n-2}_{(m-1)^{k-1}}(M-1).
\end{align*}
\end{lemma}
\begin{proof}
This follows by carefully applying Lemma~\ref{lem:X_recursion} to the shape $\lambda = m^k$, since $\lambda/i$ is either $(m-1)^k$ or $m^{k-1}$ and $\lambda/(i,j) = (m-1)^{k-1}$.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{thm:qhitCSFabelian}]
Translating Theorem~\ref{thm:qhitCSFabelian} into chromatic symmetric polynomials, we want to prove that for every $M$ we have
\begin{align}\label{eq:GPpolynomials}
X_\lambda(M) =\dfrac{1}{\qfalling{m}{n}} \sum_{j=0}^{n}
\qhit{j}{m,n}{\lambda} \cdot X_{m^j}(M).
\end{align}
We apply Lemma~\ref{lem:X_rect_recursion} to each term $X_{m^j}(M)$ appearing in the right hand side of the formula in~\eqref{eq:GPpolynomials}. We also apply Lemma~\ref{lem:X_recursion} and the induction hypothesis to the left hand side of the formula in~\eqref{eq:GPpolynomials}, i.e. to $X_{\lambda}(M)$. Then, we obtain an expression where both sides are written in terms of $X_{m^k}^{m+n}(M-1)$, $x_M X^{m+n-1}_{m^k}(M-1)$, $x_M X^{m+n-1}_{(m-1)^k}(M-1)$ and $x_M^2 X^{m+n-2}_{(m-1)^k}(M-1)$.
The term at $x_M^0$ corresponds to the coefficient of $X^{m+n}_{m^k}(M-1)$, which is $\frac{1}{[m]_n} H^{m,n}_k(\lambda)$ on both sides by induction on $M$. For the linear term, matching the coefficients of $x_M X^{m+n-1}_{(m)^k}(M-1)$ and $x_M X^{m+n-1}_{(m-1)^k}(M-1)$ separately, is equivalent to two $q$-hit identities we need to prove:
\begin{align}
\sum_{i=1}^{m} q^{m+n-i - \lambda_i'} H_k^{m-1,n}(\lambda/(m+n-i,i)) &= [m-n] H_k^{m,n}(\lambda) q^{n-k},\label{Eq:linearterm1}\\
[m-n+1]\sum_{i=m+1}^{m+n} q^{m+n-i - \lambda_i'} H_k^{m,n-1}(\lambda/(m+n-i,i)) &= H_k^{m,n}(\lambda)q^k[n-k] + H_{k+1}^{m,n}(\lambda)[k+1]. \label{Eq:linearterm2}
\end{align}
Finally, the quadratic term $x_M^2X^{m+n-2}_{(m-1)^k}(M-1)$ corresponds to the $q$-hit identity:
\begin{align}\label{Eq:quadraticterm}
q^k\sum_{(i,j)\in \lambda} q^{i+(m-j-\lambda'_j)} H^{m-1,n-1}_k(\lambda/(i,j)) = [k+1] H^{m,n}_{k+1}(\lambda).
\end{align}
By Definition~\ref{def: hit in terms of rs}, we translate these three identities involving $q$-hit numbers into three identities involving $q$-rook numbers that are in Lemma~\ref{prop:q-rook identities}. These identities complete the proof of Theorem~\ref{thm:qhitCSFabelian}.
\end{proof}
\begin{lemma}\label{prop:q-rook identities}
The $q$-hit identities~\eqref{Eq:linearterm1}, ~\eqref{Eq:linearterm2}, and~\eqref{Eq:quadraticterm} are equivalent, in that order, to:
\begin{align}
\sum_{j=1}^m q^{m-j} R_k(\lambda/j) &= R_k(\lambda) [m-k] - R_{k+1}(\lambda) (q^m -q^{m-k-1}), \label{eq: rook first linear} \\
\sum_{i=1}^n q^{i-1 +\lambda_i} R_k(\lambda/i)
&= ([n]-[k])R_k(\lambda), \label{eq: rook second linear} \\
\sum_{(i,j) \in \lambda} q^{i-j+\lambda_i} R_k(\lambda/(i,j)) &= q[k+1] R_{k+1}(\lambda). \label{eq: rook third linear}
\end{align}
\end{lemma}
We give the proof of the first relation \eqref{eq: rook first linear}. The arguments for the other two relations have a similar flavor.
\begin{proof}[Proof of \eqref{eq: rook first linear}]
Let $\ell = \ell(\lambda)$.
Multiplying on both sides by $[x]_{\ell-i}$ and summing over $i=0,\ldots,\ell$, the claimed relation is equivalent to the generating function identity:
\begin{align*}
\sum_{j=1}^m q^{m-j} F(x;\lambda/j) &= \sum_i R_i(\lambda) \left( [m-i] [x]_{\ell-i} - (q^m-q^{m-i})[x]_{\ell-i+1} \right) \\
&= [m+x-\ell]F(x;\lambda) - q^m[x] F(x-1;\lambda),
\end{align*}
where $F(x;\lambda)$ is as in \eqref{eq: def F} and we used the observation that
\[
[m-i] [x]_{\ell-i} - (q^m-q^{m-i})[x]_{\ell-i+1}
= [x]_{\ell-i} [x+m-\ell] - q^m[x]_{\ell-i+1}.
\]
We have that
\[
F(x;\lambda/j) = \prod_{i=1}^{\lambda'_j} [x+\lambda_i -1 -\ell+i ] \prod_{i=\lambda'_j+1}^\ell [x+\lambda_i -\ell+i] = F(x;\lambda) \prod_{i=1}^{\lambda'_j} \frac{ [x-1 +\lambda_i -\ell+i ]}{[x+\lambda_i -\ell+i ]}.
\]
Using Lemma~\ref{lem:F_ratio}, we have that since $\lambda/j = \lambda$ if $j>\lambda_1$,
\begin{align*}
\sum_{j=1}^m q^{m-j} F(x;\lambda/j) &= F(x;\lambda)\bigg( [m-\lambda_1] + q^{m-\lambda_1} \sum_{j=1}^m q^{\lambda_1 - j}\prod_{i=1}^{\lambda'_j} \frac{ [x-1 +\lambda_i -\ell+i ]}{[x+\lambda_i -\ell+i ]}\bigg) \\
&= F(x;\lambda)\left( [m-\lambda_1] +q^{m-\lambda_1}[x-\ell+\lambda_1] -q^{m-\lambda_1}q^{\lambda_1}[x] \frac{F(x-1;\lambda)}{F(x;\lambda) }\right) \\
&= F(x;\lambda)[x-\ell+\lambda_1 + m-\lambda_1] -q ^m[x]F(x-1;\lambda),
\end{align*}
as desired.
\end{proof}
\section{The Abreu--Nigro expansion in the elementary basis} \label{sec: MGP to AN}
In this section we show that Guay-Paquet's identity (Theorem~\ref{thm:qhitCSFabelian}) is equivalent to Abreu--Nigro's identity presented (Theorem~\ref{AN:generalLambda}).
We start by giving a proof of Abreu--Nigro's identity for rectangular shapes.
\begin{lemma}[Abreu--Nigro's formula for rectangles]\label{lem:ANrectangle}
\begin{align*}
\csft{m^k}{q} &=\qfactorial{k}\qhit{k}{m+n-k}{m^k}\cdot e_{m+n-k,k}
+ \sum_{r=0}^{k-1}q^r \qfactorial{r}\qnumber{n+m-2r} \qhit{r}{m+n-r-1}{m^k} \cdot e_{m+n-r,r}.
\end{align*}
\end{lemma}
In order to prove this case of the Abreu--Nigro identity we need the following result.
\begin{lemma}[Guay-Paquet formula for rectangles] \label{lem: MGP for rectangles}
\begin{equation} \label{eq: MGP for rectangles}
[m]X_{(m-1)^k}^{m+n-1} =q^k[m-k]X_{m^k} + [k]X_{m^{k-1}}.
\end{equation}
\end{lemma}
\begin{proof}
By Theorem~\ref{thm:qhitCSFabelian} for the shape $\lambda=(m-1)^k \subset n\times m$ and the formula for the $q$-hit numbers $H^{m,n}_r((m-1)^k)$ from Proposition~\ref{prop: hits small rect in rect} we obtain
\begin{align*}
X_{(m-1)^k} &= \frac{1}{[m]_{n}} q^k[m-1]_k [m-k]_{n-k} X_{m^k} + \frac{1}{[m]_{n}}[k][m-1]_{k-1}[m-k]_{n-k} X_{m^{k-1}},\\
[m]X_{(m-1)^k} &= q^k [m-k] X_{m^k} + [k]X_{m^{k-1}}.
\end{align*}
\end{proof}
\begin{proof}[Proof sketch of Lemma~\ref{lem:ANrectangle}]
We use induction on $m$ and $k$. For the base case, note that $X_{m^0} = [m+n]! e_{m+n} = H_0^{m+n}(m^0) e_{m+n}$. By Lemma~\ref{lem: MGP for rectangles} we have that
\[
X_{m^k} = \frac{1}{q^k[m-k]}\left([m] X_{(m-1)^k} - [k]X_{m^{k-1}}\right).
\]
Next, we use the induction hypothesis on $X_{(m-1)^k}$ and $X_{m^{k-1}}$, Proposition~\ref{prop:qhit_rectangles} for the hit numbers of rectangles, and routine simplifications to verify the desired formula for $X_{m^k}$.
\end{proof}
We are now ready to prove that the Guay-Paquet's identity and Abreu--Nigro's follow from each other. As a corollary, we obtain a new proof of the latter.
\begin{proof}[Proof of Theorem~\ref{AN:generalLambda}]
Applying Lemma~\ref{lem:ANrectangle} to the RHS of the formula in Theorem~\ref{thm:qhitCSFabelian}, we obtain that
\begin{align*}
&\dfrac{1}{\qfalling{m}{n}} \sum_{j=0}^{n}
\qhit{j}{m,n}{\lambda} \cdot \csft{m^j}{q}
= \dfrac{1}{\qfalling{m}{n}} \sum_{j=0}^{n}
\qhit{j}{m,n}{\lambda} \left( \qfactorial{j}\qhit{j}{n+m-j}{m^j}\cdot e_{m+n-j,j}\right) \\[0.07in]
&+\dfrac{1}{\qfalling{m}{n}} \sum_{j=0}^{n}
\qhit{j}{m,n}{\lambda} \left( q^r\sum_{r=0}^{j-1}
\qfactorial{r} \qnumber{n+m-2r}\qhit{r}{m+n-r-1}{m^j} \cdot e_{m+n-r,r}\right).
\end{align*}
Now, switching the summation order, we have that
\begin{align*}
&\dfrac{1}{\qfalling{m}{n}} \sum_{j=0}^{n}
\qhit{j}{m,n}{\lambda} \cdot \csft{m^j}{q}
=\sum_{r=0}^n e_{m+n-r,r} \dfrac{1}{\qfalling{m}{n}}\qfactorial{r} \qhit{r}{m+n-r}{m^r}\qhit{r}{m,n}{\lambda} \\
&+ \sum_{r=0}^{n-1} e_{m+n-r,r} \dfrac{1}{\qfalling{m}{n}}\left(q^r\sum_{j=r+1}^n \qfactorial{r} \qnumber{n+m-2r}\qhit{r}{m+n-r-1}{m^j}\qhit{j}{m,n}{\lambda}\right).
\end{align*}
Thus, we need to show that for $r=k=\ell(\lambda)$,
\begin{align*
\qfalling{m}{n} \qhit{k}{m+n-k}{\lambda} &= \qhit{k}{m+n-k}{m^k}\qhit{k}{m,n}{\lambda}
+ q^k\sum_{j=k+1}^n \qnumber{n+m-2k} \qhit{k}{m+n-k-1}{m^j}\qhit{j}{m,n}{\lambda},
\end{align*}
and for $r<k=\ell(\lambda)$,
\begin{align*
\qfalling{m}{n} q^r \qnumber{n+m-2r}\qhit{r}{m+n-r-1}{\lambda}
&= \qhit{r}{m+n-r}{m^r}\qhit{r}{m,n}{\lambda} \\
&+ q^r\sum_{j=r+1}^n \qnumber{n+m-2r}\qhit{r}{m+n-r-1}{m^j}\qhit{j}{m,n}{\lambda}.\nonumber
\end{align*}
After using Proposition~\ref{prop:qhit_rectangles}, these two relations are equivalent to the following identities relating $q$-hit numbers of $\lambda$ in square boards and rectangular boards. Finally, the Abreu-Nigro expansion for $\csft{\lambda}{q}$ follows now from Lemma~\ref{prop:qhit-relations}. \end{proof}
\begin{lemma}\label{prop:qhit-relations}
Let $\lambda$ be a partition inside an $n\times m$ board and $k=\ell(\lambda)$, then
\begin{align}\label{eq:keyrel1}
\qbinom{m-k}{n-k}\qhit{k}{m+n-k}{\lambda} &= q^{k(n-k)} \qfalling{m+n-2k}{m-k} \qhit{k}{m,n}{\lambda}.
\intertext{For $0\leq r< k$, we have}
\label{eq:keyrel2}
\qbinom{m-r}{n-r} \qhit{r}{m+n-r-1}{\lambda} &=
q^{ r(n-r-1)} \qfalling{m+n-2r-1}{m-r-1} \qhit{r}{m,n}{\lambda} \nonumber\\
& + \sum_{j=r+1}^n q^{r(n-1-j)} \qbinom{j}{r} \dfrac{ \qfalling{m+n-r-j-1}{m-r} }{\qnumber{n-r}} \qhit{j}{m,n}{\lambda}.
\end{align}
\end{lemma}
\begin{proof}[Proof sketch of Lemma~\ref{prop:qhit-relations}]
Each identity follows by using Definition~\ref{def: hit in terms of rs} to rewrite both the LHS and RHS in terms of $q$-rook numbers $R_j(\lambda)$ and showing the resulting expressions are equal via routine $q$-factorial manipulations.
We remark that there is a more interesting proof using the deletion/contraction formula in Lemma~\ref{lem: deletion/contration} on both sides of each relation, the rectangle-resizing identity in Lemma~\ref{lemma:remove column}, and induction.
\end{proof}
\section{Open problems}
Since our proof of Theorem~\ref{thm:qhitCSFabelian} uses $q$-rook theory, it would be interesting to find a bijective proof of this result relating colorings with rook placements.
There are other rules for the elementary basis expansion of $\csft{\lambda}{q}$. In particular, Cho--Huh \cite{ChoHuh} give an expansion in terms of {\em $P$-tableaux} of shape $2^j 1^{m+n-2j}$ such that there is no $s \geq j+2$ such that $(a_{i,1},a_{s,1}) \in \lambda$ for all $i \in \{ \ell+1,\ldots,s-1\}$. Let $c_j^{m,n}(q):=\sum_T q^{\inv_{G(\lambda)}(T)}$, where the sum is over such tableaux (see~\cite[Sec. 6]{ShW2}). It would be interesting to find a weight-preserving bijection that shows that
\[
c_j^{m,n}(q) = \begin{cases}
[j]! H_j^{m+n-j}(\lambda) &\text{ if } j=\ell(\lambda),\\
q^j[j]! [m+n-2j]H_j^{m+n-j-1}(\lambda) &\text{ if } j<\ell(\lambda).
\end{cases}
\]
\acknowledgements{We thank Mathieu Guay-Paquet for generously sharing the notes~\cite{MGP_LR} with
Theorem~\ref{thm:qhitCSFabelian} as well as Alex Abreu, Per Alexandersson, and Antonio Nigro for insightful discussions.}
\printbibliography
\end{document}
|
\section{Introduction}
\label{intro}
The study of dense granular packings is an active field of research in materials science. In these packings, the macroscopic and structural properties rely heavily on the characteristics of the individual grains (friction, stiffness, restitution, etc.). For instance, it is known that friction can change the stability criterion of dense packings of frictional particles for contacts from 6 to 4 \cite{R1} and that the jamming transition is changed from a second-order continuous one to a first-order discontinuous one \cite{R22}. Moreover, previous work has shown that when frictional packings are subject to shear, asymmetric, ordered phases emerge within the material \cite{R2}, a process that is important for understanding granular self-assembly. Additionally, a limit-cycle behavior in cyclically sheared granular packings has been observed for a range of friction values, in which the state of the system remains unchanged upon repeated shear cycles \cite{R3}.
The majority of these studies place an emphasis on either macroscopic parameters (volume fraction, pressure, etc.)\ or particle positions and displacements \cite{R4,R5,R6,R7,R8,R9,R10}; however, a key quantity that emerges specifically for frictional systems is particle rotations, which increase the number of degrees of freedom of the system to 6 from 3 in three-dimensional systems. Only a few studies have an explicit focus on grain rotations \cite{R11,R12,R13,R14,R15,R16} especially in the case of spherical particles in 3D, where a direct correlation between displacements and rotations is not evident as it is in the case of irregularly shaped particles or two-dimensional systems where the number of constraints to satisfy is lower. However, this does not mean that rotations are not important. On the contrary: in dense granular assemblies, energy loss is dominated by frictional dissipation that generates rotational motion. At low particle speeds, collisional dissipation is negligible. Understanding rotations is thus essential to comprehend the dynamics of frictional granular systems, even for spherical particles where the relation between translations and rotations is not straightforward.
\begin{figure}[b!]
\centering
\includegraphics[width=1.0\columnwidth]{./schematic.pdf}
\caption{Schematic of the simulation setup. 20,000 spheres are confined in a box of side length 15 cm. The red wall compresses the system from 0 (decompressed) to an amplitude $A$ (compressed), along the $y$ axis, then back to 0. A top weight is used for constant pressure. }
\label{schematic}
\end{figure}
\begin{figure*}[h]
\centering
\includegraphics[width=1.8\columnwidth,clip,trim=0 9.8cm 1cm 10.0cm]{./fig1_final_new.pdf}
\caption{Two-dimensional projection of the amplitude of grain translations (top row) and rotations (bottom row) when fully compressed for varying friction coefficients. All values here were normalized by the compression amplitude $A=0.15$~cm. The scale for the rotations is saturated for clarity. The system is compressed from the left. A weight is free to move on the top}
\label{fig-1}
\end{figure*}
Theoretical studies such as the frictional Cosserat models \cite{R16} have shown the importance of vorticity and rotations in two dimensions to predict the velocity profiles of granular flows. Their predictions have shown a good agreement with experiments and simulations \cite{R18,R23,R24}. However, the model has not yet been linked to three-dimensional particle-scale simulations, and the dependence of the continuum model parameters on the microscopic parameters of the grains, such as interparticle friction, is unknown.
This numerical study is aimed at establishing the connection between translation and rotation of grains in support of related investigations. The first dedicated to the study of how the reversibility of translational motion compares to the reversibility of rotational motion\cite{R12}. The second investigate how friction and rotations affect the memory formation in granular systems \cite{R13}.
In this paper, we present work focused on quantifying granular rotations of a cyclically compressed
system using soft-sphere discete-element-method (DEM) simulations, as a function of inter-particle friction between grains. We study how friction affects the localization of shear zones in the system, which in turn affects the spatial distribution of rotations and their correlation to the translations.
\section{Numerical Methods}
Soft-sphere DEM simulations were performed using in-house software called pkdgrav \cite{R19}. 20,000 spherical grains of radius 0.25~cm and mass 0.0618~g were dropped in a cubic box with side length 15.0~cm. A top weight of mass 1000~g was dropped onto the grains to maintain a constant external pressure, and is free to move along the $z$ axis. The beads settle to a height of approximately 10~cm. A schematic of the simulation setup is given in Figure \ref{schematic}. The system was allowed to settle before starting a quasi-static compression along the $y$ axis. Given the free moving weight on top, the system expands along the $z$ axis when compressed along the $y$ axis, presenting a pure shear geometry. For each cycle, the wall starts off at position 0 (decompressed) then moves to an amplitude $A$ (compressed), then back to position 0. The compression amplitude used in this report is $\sim$1.0\% of the size of the container, corresponding to 0.15~cm. The results presented in this paper are for a system which was subjected to 200 cycles of pure shear. We focus on small shear amplitudes, where the evolution of the system is small \cite{R12,R13}. The linear displacements are calculated as the difference in position of the grains when the wall is at position 0 to an amplitude $A$. Rotations are defined by a rotation matrix between the orientation of each grain when the wall is at position 0 and compressed to $A$. The units are converted to cm by multiplying the angle of rotation by the radius. We have varied the coefficient of inter-particle friction from $\mu_S$ = 0.05 to $\mu_S$ = 0.5. More details on the numerical setup are available in \cite{R12}. Here we expand on this study and describe key simulation results as a function of friction coefficient.
\section{Results and Discussion}
Figure \ref{fig-1} shows the spatial distribution of displacements and rotations at full compression for
static friction values of 0.05, 0.2 and 0.5. We see well-defined shear bands in
the translational displacement for all friction values. The shear band becomes more localized as the friction is increased.
For lower friction, the rotations are uniformly distributed throughout the
container, behaving more fluid-like as contacts have a lower threshold to slide against one another.
In contrast, at high friction, the rotations are predominately localized in the region of the translational shear band. This behavior at high friction is reminiscent of that of non-spherical particles \cite{R20}.
Figure \ref{fig-2} studies the local structure of the shear-zones in Figure
\ref{fig-1}. Specifically, the displacements were binned along a 45-degree
line running from the bottom far right of the container ($y,z$ = 0), to the
top left ($y,z$ = 18~cm) along a line $l=\sqrt{(y-L)^2+z^2}$
; the bins are therefore parallel to the shear band. We crop the extreme
values of $l$ since the boundaries of the box limits the number of grains
within those bins. Figure \ref{fig-2}a shows that at low friction, a linear
increase in the displacements is observed, as expected for a local rheology.
However, for high friction the curve is highly non-linear, reminiscent of
what we could expect for non-local rheology \cite{R21}. As we have seen, the
rotations are uniform at low friction (Figure \ref{fig-2}b). At high
friction, the rotational displacements are highly localized. We can expect
that rotations will be the highest in the shear zones, as was observed for
non-spherical particles \cite{R20}. Accordingly, Figure \ref{fig-3}
calculates the spatial derivative of the curve in Figure \ref{fig-2}a.
\begin{figure}[ht]
\centering
\includegraphics[width=0.8\columnwidth]{./fig3a.pdf}
\caption{Top, the average translational displacement in the binned zones for different values of friction. Bottom, the same for rotations. The binning is performed along an axis parallel to the shear bands. The legend corresponds to the value of friction, $\mu_S$.}
\label{fig-2}
\end{figure}
As for the rotations, we see a peak emerge in the shear zone at high friction. For low values of friction, the function becomes much more uniform, yet is still different from the mostly flat curves of the rotations. This suggests that rotations and translations are mostly decoupled at low values of friction.
\begin{figure}[ht!]
\centering
\includegraphics[width=0.8\columnwidth]{./fig3b.pdf}
\caption{Spatial derivative of the translations in the binned zones for different values of friction. The binning is performed along an axis parallel to the shear bands. The legend corresponds to the value of friction, $\mu_S$.}
\label{fig-3}
\end{figure}
\begin{figure}[ht]
\centering
\includegraphics[width=0.9\columnwidth]{./fig5.pdf}
\caption{Correlation between the spatial derivative of the grain displacements and the rotations for different values of friction. The transparency of symbols corresponds to the distance to the far-right corner as given in Figures \ref{fig-2} and \ref{fig-3}. The dashed line is the theoretical correlation of a sphere rolling between two moving plates, which has a slope of 1/2. The legend corresponds to the value of friction, $\mu_S$.}
\label{fig-4}
\end{figure}
To further probe the correspondence, the derivative of the displacements and the rotations are plotted against each other in Figure \ref{fig-4}.
For reference, the dashed line corresponds to the relationship between rotation of a sphere rolling without slipping while confined between two parallel plates. In this example, the top plate moves an amount $\Delta x$ in a given time interval and the bottom plate remains fixed; this acts as a measure for the gradient in translations given in Figure \ref{fig-3}. If the sphere cannot slide, then its rotational motion comes out to be exactly $ \Delta x/2$, the relative motion of the two plates. Surprisingly, for high friction, the curve collapses to a straight line of slope $\approx 0.70$. This indicates that there appears to be more rotations than expected for rolling without slipping. Thus, the difference between the dashed line and the data points can be thought of as a measure for frictional dissipation. In fact, if one takes the sum of the normalized squared distance of each point from the black dashed line, values of 32.5, 24.2, 11.0, 16.8, and 52.2 are obtained with increasing friction. At low friction, the threshold for a contact to slide is much less, which leads to high frictional loss. Increasing the friction also increases the threshold, however the dissipated energy due to sliding is increased. Thus the values that we have computed first decrease and then increase again with friction.
For the lowest value of friction, $\mu_S = 0.05$, two conclusions can be made. First, rotations do not strongly depend on local shear in this case. Second, there is more rotation than predicted by the amount of local shear, or than observed at higher values of local shear, indicating that rotations are not simply coupled to translational motion. The intermediate friction of $\mu_S = 0.1$ presents an interesting transitional case. While in the region of low translations, the rotations seem to be proportional to the derivative of rotations, as is the case for high friction; in the region of high translations, the two decouple, as is the case for low friction.
As the friction increases, the rotations appear to be more confined, as evident in Figure \ref{fig-1} as well as the peaks in Figure \ref{fig-2}. It appears the material goes through some structural transition as a function of friction. To visualize a transition, we plot the maximum value of the rotations curve in \ref{fig-2} as a function of the friction coefficient. Figure \ref{fig-5} shows the value of maximum rotation as well as the maximum of the derivative. The peak values were normalized by the highest quantity ($\mu_S = 0.5$). The evolution seems to be a derivative function with an increase in the peak value between $\mu_S$ = 0.2 and 0.3. As we have already seen on Figure \ref{fig-4}, for low values of friction, there also appears to be more rotations than expected if there would have been a linear proportionality between rotations and the derivative of translations.
\begin{figure}[t!]
\centering
\includegraphics[width=0.9\columnwidth]{./fig4.pdf}
\caption{Peak value of the derivative of the translations (crosses) and the rotations (circles) as a function of friction. Data points were normalized by the highest peak value for comparison.}
\label{fig-5}
\end{figure}
\section{Conclusion}
In this article, we have analyzed the influence of inter-particle friction on the response of a three-dimensional granular system of spherical particles to cyclic compression. Three main observations can be made from our results. First, that the shear zone is becoming more localized as the inter-particle friction is increased, a result that agrees with the prediction of frictional Cosserat models \cite{R18}. Second, that at low friction, the rotations are mostly decoupled from translations, while at high friction, the rotations are mostly concentrated in the shear band. The latter result is consistent with that observed for non-spherical particles \cite{R20}. Finally, we find that there is always more rotations than for a theoretical rolling-without-sliding system, which can be used to quantify dissipation in a dense granular system.
\fontsize{9}{12}\selectfont
The authors acknowledge the University of Maryland supercomputing resources (http://hpcc.umd.edu) made available for conducting the research reported in this paper. ZB was supported by the National Science Foundation graduate research fellowship program. WL and AP were
supported by National Science Foundation Grant No. DMR-1507964. AP was also supported by DMR-1809318.
|
\section{Introduction}
Let $\Omega\subseteq \mathbb{R}^2$ be a smooth bounded domain and $W_0^{1,2}(\Omega)$ be the completion of $C_0^{\infty}(\Omega)$ under the Sobolev norm
$\|\nabla_{\mathbb{R}^2} u\|_2^2= \int_{\Omega}{|\nabla_{\mathbb{R}^2} u|^2}dx,$
where $\nabla_{\mathbb{R}^2}$ is the gradient operator on ${\mathbb{R}^2}$ and $\|\cdot\|_2$ denotes the standard $L^2$-norm.
The classical Trudinger-Moser inequality \cite{Yudovich, Pohozaev, Peetre, Trudinger1967, Moser1970}, as the limit case of the Sobolev embedding, says
\begin{eqnarray}\label{Trudinger-Moser}
\sup_{u\in W_0^{1, 2}(\Omega), \, \|\nabla_{\mathbb{R}^2} u\|_2\leq 1}
\int_\Omega e^{\ \beta u^2}dx<+\infty, \ \forall \ \beta\leq 4\pi.
\end{eqnarray}
Moreover, $4\pi$ is called the best
constant for this inequality in the sense that when $\beta> 4\pi$, all integrals in (\ref{Trudinger-Moser}) are still finite, but the supremum is infinite.
It is interesting to know whether or not the supremum in (\ref{Trudinger-Moser}) can be attained.
For this topic, we refer the reader to Carleson-Chang \cite{C-C},
Flucher \cite{Flucher}, Lin \cite{Lin}, Struwe \cite{Struwe}, Adimurthi-Struwe \cite{A-Struwe}, Li \cite{Li-Sci}, Yang \cite{Yang-JFA-06},
Zhu \cite{ZhuJY}, Tintarev \cite{Tintarev} and the references therein.
There are many extensions of (\ref{Trudinger-Moser}).
Adimurthi-Druet \cite{AD} generalized (\ref{Trudinger-Moser}) to the following form:
\begin{equation}\label{T-AM}
\sup _ { u \in W _ { 0 } ^ { 1,2 } ( \Omega ) , \| \nabla_{\mathbb{R}^2} u \| _ { 2 } \leq 1 } \int _ { \Omega } e ^ { 4 \pi u ^ { 2 } \left( 1 + \alpha \| u \| _ { 2 } ^ { 2 } \right) } dx <+\infty, \ \forall\ 0 \leq \alpha < \lambda _ { 1 } ( \Omega ),
\end{equation}
where $\lambda _ { 1 } ( \Omega )$ is the first eigenvalue of the Laplacian with Dirichlet boundary condition in $\Omega .$
This inequality is sharp in the sense that if $\alpha \geq \lambda _ { 1 } ( \Omega )$, all integrals in (\ref{T-AM}) are still finite, but the supremum is infinite. Obviously, (\ref{T-AM}) is reduced to (\ref{Trudinger-Moser}) when $\alpha=0$.
Various extensions of the inequality (\ref{T-AM}) were obtained by Yang \cite{Yang-JFA-06,Yang-JDE-15}, Tintarev \cite{Tintarev} and Zhu \cite{ZhuJY} respectively.
It was extended by Lu-Yang \cite{Lu-Yang} to a version, namely
\begin{equation}\label{T-LY}
\sup _ {u \in W^{1,2}(\Omega),
\int_{\Omega} u dx=0, \| \nabla_{\mathbb{R}^2} u \| _ { 2 } \leq 1 } \int _ { \Omega } e ^ { 2 \pi u ^ 2 \left( 1 + \alpha \| u \| _ 2 ^ 2 \right) } d x < + \infty, \ \forall\ 0 \leq \alpha < \overline{\lambda} _ 1 ( \Omega ),
\end{equation}
where
$ \overline{\lambda} _ 1 ( \Omega )$ denotes the first nonzero Neumann eigenvalue of the Laplacian operator.
This inequality is sharp in the sense that all integrals in (\ref{T-LY}) are still finite when $ \alpha \geq \overline{\lambda} _ 1 ( \Omega)$, but the supremum is infinite.
Moreover, for sufficiently small $\alpha> 0$, the supremum is attained.
Trudinger-Moser inequalities were introduced on Riemannian manifolds by Aubin \cite{A}, Cherrier \cite{C} and Fontana \cite{Fontana}. In particular, let $( \Sigma , g )$ be a 2-dimensional compact Riemann surface, $W^ { 1 , 2 } ( \Sigma, g) $ the completion of $C ^ { \infty } ( \Sigma)$ under the norm
$\|u \| ^2_ { W^{1,2}( \Sigma, g) } = \int _ { \Sigma }( u^2+| \nabla_g u | ^2)\, dv_g$, where $\nabla_g$ stands for the gradient operator on $(\Sigma, g)$.
When $( \Sigma , g )$ is closed Riemann surface,
there holds
\begin{equation}\label{T-MM}
\sup _ {u \in W ^ { 1,2 } ( \Sigma, g), \int _\Sigma u dv_g = 0, \| \nabla_g u \| _ { 2 } \leq 1 }\int _ { \Sigma } e ^ {\ \beta u^ 2} dv_g<+\infty, \ \forall \ \beta\leq 4\pi.
\end{equation}
Moreover, $4\pi$ is called the best
constant for this inequality in the sense that when $\beta> 4\pi$, all integrals in (\ref{T-MM}) are still finite, but the supremum is infinite.
Based on the works of Ding-Jost-Li-Wang \cite{DJLW} and Adimurthi-Struwe \cite{A-Struwe}, Li \cite{Li-JPDE,Li-Sci} proved the existence of extremals for the supremum in (\ref{T-MM}).
When $(\Sigma, g)$ is a compact Riemann surface with smooth boundary $\partial\Sigma$,
Yang \cite{Yang2006IJM} obtained the same inequality as (\ref{T-MM}), namely
\begin{equation}\label{2006Y}
\sup _ {u \in W ^ { 1,2 } ( \Sigma, g), \int _\Sigma u dv_g = 0, \| \nabla_g u \| _ { 2 } \leq 1 }\int _ { \Sigma } e ^ {\ \beta u^ 2} dv_g<+\infty, \ \forall \ \beta\leq 2\pi.
\end{equation}
This inequality is sharp in the sense that if $\beta> 2\pi$, all integrals in (\ref{2006Y}) are still finite, but the supremum is infinite. Furthermore, the supremum in (\ref{2006Y}) can be attained.
\\
In view of the inequality (\ref{T-LY}) in the Euclidean space, we strengthen (\ref{2006Y}) on $(\Sigma, g)$ with smooth boundary $\partial\Sigma$. Precisely we have the following:
\begin{theorem}\label{T1}
Let $(\Sigma, g)$ be a compact Riemann surface with smooth boundary $\partial\Sigma$
and
\begin{eqnarray}\label{la}
\lambda_1(\Sigma)= \inf _ { u \in W^{1,2} ( \Sigma, g) , \int _ { \Sigma } u dv_g = 0 , u \not\equiv 0 } \frac { \|\nabla_g u\|^2_2} {\|u\|_2^2}
\end{eqnarray}
be the first eigenvalue of the Laplace-Beltrami operator $\Delta _ { g}$ with respect to the zero mean value condition.
Denote a function space
\begin{eqnarray*}\label{S}
\mathcal { S } = \left\{ u \in W ^ { 1,2 } ( \Sigma, g) : \int _\Sigma {u}\ dv_g = 0, \ \| \nabla_g u \| _ { 2 } \leq 1 \right\}
\end{eqnarray*}
and
\begin{eqnarray*}
F_{\alpha}^{\beta}(u)= \int_{\Sigma} e^{ \beta u^{2}\left(1+\alpha\|u\|_2^{2}\right) }d v_g.
\end{eqnarray*}
Then there hold\\
(i) for any $ \alpha \geq\lambda_1(\Sigma),$ $\sup _{u \in \mathcal{S}}
F_{\alpha}^{2\pi}(u)=+\infty$;\\
(ii) for any $0 \leq \alpha<\lambda_1(\Sigma)$, $\sup _{u \in \mathcal{S}} F_{\alpha}^{2\pi}(u)<+\infty$;\\
(iii) for sufficiently small $\alpha>0$, $\sup _{u \in \mathcal{S}} F_{\alpha}^{2\pi}(u)$ can be attained by some function $u_{\alpha} \in C^\infty\left(\overline{\Sigma}\right)\cap \mathcal {S}$.
\end{theorem}
For the proof,
we employ the method of blow-up analysis, which was originally used by Carleson-Chang\cite{C-C},
Ding-Jost-Li-Wang \cite{DJLW}, Adimurthi-Struwe \cite{A-Struwe}, Li \cite{Li-JPDE},
and Yang \cite{Yang2007, Yang-JDE-15}.
For related works, we refer the reader to
Adimurthi-Druet \cite{AD},
do \'O-de Souza \cite{do-de2014,do-de2016}, Nguyen \cite{N2017,N2018}, Li-Yang \cite{L-Y}, Zhu \cite{Zhu}, Fang-Zhang \cite{F-Z}, Yang-Zhu \cite{Yang-Zhu2018,Yang-Zhu2019} and Csat\'o-Nguyen-Roy \cite{C-N-R}.
We should point out that the blow-up occurs on the boundary $\partial\Sigma$ in our case.
The key ingredient in the proof of our theorem is the isothermal coordinate system on $\partial\Sigma$.
Though such coordinates have been used by many authors (see for example Li-Liu \cite{Li-Liu} and Yang \cite{Yang2006IJM, Yang2006PJM, Yang-2007}), the proof of its existence around has just been provided by Yang-Zhou \cite{Yang-Zhou} via Riemann mapping theorems involving the boundary.
The remaining part of this paper will be organized as follows: In Section 2, we prove (Theorem \ref{T1}, ($i$)) by constructing test functions; in Section 3, we prove (Theorem \ref{T1}, ($ii$)) by using blow-up analysis; in Section 4, we construct a sequence of functions to show (Theorem \ref{T1}, ($iii$)) holds.
Hereafter we do not distinguish the sequence and the subsequence; moreover, we often denote various constants by the same $C$.
\section{The case of $\alpha\geq\lambda_1(\Sigma)$}
In this section, we select test functions to prove Theorem \ref{T1} ($i$).
Let $\lambda_1(\Sigma)$ be defined by (\ref{la}) and $\alpha \geq \lambda_1(\Sigma)$.
From a direct method of variation,
one obtains that
there exists some function $u_0 \in \mathcal { S } $, such that
\begin{eqnarray}\label{s3}
\lambda_1(\Sigma) =\left\|\nabla_g u_0\right\|_{2}^{2}.
\end{eqnarray}
By a direct calculation, we derive that $u_0$ satisfies the Euler-Lagrange equation
\begin{eqnarray}\label{e-ua}
\left\{
\begin{aligned}
&\Delta_g u _ 0 =\lambda_1(\Sigma)\,u _0 \,\,\, \mathrm{in}\,\,\, \Sigma ,\\
&{ \frac { \partial u _0 } { \partial \mathbf{n} } =0 \,\,\, \mathrm{ on } \,\,\, \partial \Sigma },\\
&\int _ { \Sigma } u_0 \,dv_g = 0,\ \int _{\Sigma} {u_0^2} \,dv_g=1,
\end{aligned} \right.
\end{eqnarray}
where $\mathbf{n}$ denotes the outward unit normal vector on $\partial\Sigma$. Applying elliptic estimates to (\ref{e-ua}), we obtain $u_0 \in \mathcal { S } \cap C^{\infty}\left(\overline{\Sigma}\right)$.
Consequently, there exist a point $ x_0\in \partial\Sigma$ with $u _ { 0 } ( x_0 ) > 0 $ and
a neighborhood $U$ of $x_0$ with $u_ { 0 }(x) \geq u_ { 0 } ( x_0 ) / 2$ in $U$ .
Let $\delta = {\left( t _\epsilon \sqrt { -\log \epsilon } \right)}^{-1}$, where $t _ { \epsilon } > 0$ such that $-t _ { \epsilon } ^ { 2 } \log { \epsilon } \rightarrow + \infty$ and $t _ { \epsilon } ^ { 2 } \sqrt { -\log { \epsilon } } \rightarrow 0$ as $\epsilon\rightarrow0$.
We take an isothermal coordinate system $\left(\phi^{-1}\left(\mathbb{B}_\delta^+\right),\phi\right)$ such that $\phi(x_0) = 0$ and $\phi^{-1}\left(\mathbb{B}_\delta^+\right)\subset U $.
In such coordinates, the metric $g$ has the representation $g = e^{2f}\left (dx_1^2 +dx_2^2 \right)$ and $f$ is a smooth function with $f(0) = 0$.
On $\overline{\mathbb{B}_{\delta}^{+}},$ we define a sequence of functions
\begin{eqnarray*}
\tilde{u}_{\epsilon}(x)=\left\{
\linespread{1.5}\selectfont
\begin{aligned}
&\sqrt{\frac{-\log \epsilon}{2 \pi} }, \ \ \ \ \ \ \ \ \ \ \ \ |x| \leq \delta \sqrt{\epsilon}, \\
&\sqrt{\frac{-2}{\pi \log \epsilon}} \log \frac{\delta}{|x|},\ \delta \sqrt{\epsilon}<|x| \leq \delta.
\end{aligned}\right.
\end{eqnarray*}
And we set
\begin{eqnarray}\label{s4}
u_{\epsilon}=\left\{\begin{array}{ll}
\tilde{u}_{\epsilon} \circ \phi & \text { in }\ \ \phi ^{-1}\left(\overline{\mathbb{B}_{\delta}^{+}}\right), \\
s_{\epsilon}\, \varphi & \text { in } \ \ \Sigma \backslash \phi ^{-1}\left(\overline{\mathbb{B}_{\delta}^{+}}\right),
\end{array}\right.
\end{eqnarray}
where $\varphi \in C_{0}^{\infty}\left(\Sigma \backslash \phi^{-1}(\mathbb{B}_\delta)\right)$ and $s_{\epsilon}$ is a real number such that $\int_{\Sigma} u_{\epsilon} \,dv_g=0$.
Set $v_{\epsilon}=u_{\epsilon}+t_{\epsilon} u_{0}$. According to (\ref{s3})-(\ref{s4}), we have
\begin{eqnarray}\label{s5}
\left\|v_{\epsilon}\right\|_{2}^{2}=\left\|u_{\epsilon}\right\|_{2}^{2}+t_{\epsilon}^{2}\left\|u_{0}\right\|_{2}^{2}+2 t_{\epsilon} \int_{\Sigma} u_{\epsilon} u_{0} \,dv_g
= t_{\epsilon}^{2}+2 t_{\epsilon} \int_{\Sigma} u_{\epsilon} u_{0} \,dv_g+O\left(\frac{-1}{ \log\epsilon}\right) \end{eqnarray}
and
\begin{eqnarray}\label{s6}
\left\|\nabla_g v_{\epsilon}\right\|_{2}^{2}
= 1+\lambda_1(\Sigma) t_{\epsilon}^{2}+2 \lambda_1(\Sigma) t_{\epsilon} \int_{\Sigma} u_{\epsilon} u_{0} \,dv_g+O\left(\frac{-1}{ \log\epsilon}\right).
\end{eqnarray}
Take $v_{\epsilon}^{*} =v_{\epsilon} /\lVert \nabla_gv_{\epsilon} \rVert _{2}^{2}\in\mathcal{S}$.
From $\alpha \geq \lambda_1(\Sigma)$ and (\ref{s4})-(\ref{s6}), we have that on $\phi^{-1}\left(\mathbb{B}_{\delta \sqrt{\epsilon}}^{+}\right)$
\begin{eqnarray*}
&\ &{2 \pi v_{\epsilon}^{*2} \left(1+\alpha\left\|v_{\epsilon}^*\right\|_{2}^{2}
\right)}\\
&=&
2 \pi {v_{\epsilon}^{2}}\frac{1}{\left\|\nabla_g v_{\epsilon}\right\|_{2}^{2}} \left(1+\alpha\frac{\left\|v_{\epsilon}\right\|_{2}^{2}}
{\left\|\nabla_g v_{\epsilon}\right\|_{2}^{2}}\right)\\
& \geq &2 \pi\left(t_{\epsilon}^{2} u_{0}^{2}-\frac{\log \epsilon}{2 \pi} +2 t_{\epsilon}\sqrt{\frac{-\log \epsilon}{2 \pi} } u_{0}\right)
\left(1+\left(\alpha-\lambda_1(\Sigma)\right)\left(t_{\epsilon}^{2}+2 t_{\epsilon} \int_\Sigma u_{\epsilon} u_{0}\,dv_g\right)+o\left(\frac{t_{\epsilon}}{\sqrt{-\log \epsilon}}\right)\right)
\\
& \geq &\left(2 \pi t_{\epsilon}^{2} u_{0}^{2}-\log \epsilon +4 \pi t_{\epsilon}\sqrt{\frac{-\log \epsilon}{2 \pi} } u_{0}\right)\left(1+o\left(\frac{t_{\epsilon}}{\sqrt{-\log \epsilon}}\right)\right) \\
& \geq& -\log {\epsilon}+t_{\epsilon}\sqrt{-\log \epsilon}\left(\sqrt{8 \pi} u_{0}+o(1)\right).
\end{eqnarray*}
Hence there holds
\begin{eqnarray*}
\int_{\Sigma} e^{2 \pi v_{\epsilon}^{*2} \left(1+\alpha\left\|v_{\epsilon}^*\right\|_{2}^{2}
\right)} \,dv_g & \geq& \int_{\phi^{-1}\left(\mathbb{B}_{\delta \sqrt{\epsilon}}^{+}\right)} \frac{1}{\epsilon} e^{t_{\epsilon}\sqrt{-\log \epsilon}\left(\sqrt{8 \pi} u_{0}+o(1)\right)} \,dv_g \\
& \geq &C(\delta) e^{t_{\epsilon} \sqrt{-\log {\epsilon}}\left(\sqrt{2 \pi} u_{0}(x_0)+o(1)\right)}
\end{eqnarray*}
for some positive constant $C(\delta) .$ In view of $u_{0}(x_0)>0,$ we get $ \sup _ { u \in \mathcal {S} } F_\alpha^{2\pi}(u) \ge \lim_{\epsilon \rightarrow 0}F_\alpha^{2\pi}\left(v_{\epsilon}^{*}\right)=+\infty$.
This completes the proof of Theorem \ref{T1} ($i$).
\section{The case of\, $0\leq\alpha<\lambda_1(\Sigma)$}
In this section, we will prove Theorem \ref{T1} ($ii$) in three steps:
firstly, we consider the existence of maximizers for subcritical functionals and give the corresponding Euler-Lagrange equation;
secondly, we deal with the asymptotic behavior of the maximizers through blow-up analysis;
finally, we deduce an upper bound of the supremum $\sup_{ u \in \mathcal { S } }F_{\alpha}^{2\pi}(u)$ under the assumption that blow-up occurs.
\subsection*{\textbf{Step 1.} Existence of maximizers for subcritical functionals}
Using the similar proof of (\cite{Lu-Yang}, Step 1), we have the following
\begin{lemma}\label{L1}
For any $\epsilon>0$, there exists some function $ u_{\epsilon} \in \mathcal {S}\cap C^{\infty}\left(\overline{ \Sigma}\right)$ with $\|\nabla_g u_{\epsilon}\|_2^2=1$, such that
\begin{eqnarray*}
\sup_{u \in\mathcal { S}}F_{\alpha}^{2\pi-\epsilon}(u)
=F_{\alpha}^{2\pi-\epsilon}(u_\epsilon).
\end{eqnarray*}
Moreover, $u_{\epsilon}$ satisfies the Euler-Lagrange equation
\begin{eqnarray}\label{e-u}
\left\{ \begin{aligned}
& \frac{\partial u_{\epsilon}}{\partial \mathbf{n}}=0\,\,\mathrm{on\,\,}{\partial \Sigma},\\
&\Delta_g u_{\epsilon}=\frac{\beta _{\epsilon}}{\lambda _{\epsilon}}u_{\epsilon}e^{\alpha _{\epsilon}u_{\epsilon}^{2}}+\gamma _{\epsilon}u_{\epsilon}-\frac{\mu _{\epsilon}}{\lambda _{\epsilon}}\,\,\mathrm{in\,\,}\Sigma,
\\
&\alpha _{\epsilon}=\left( 2\pi -\epsilon \right) \left( 1+\alpha\|u_{\epsilon}\|_2^2 \right),\ \
\beta _{\epsilon}=\frac{1+\alpha \|u_{\epsilon}\|_2^2}{1+2\alpha \|u_{\epsilon}\|_2^2},\\
&\gamma _{\epsilon}=\frac{\alpha}{1+2\alpha \|u_{\epsilon}\|_2^2},\ \
\lambda _{\epsilon}=\int_{\Sigma}{u}_{\epsilon}^{2}e^{\alpha _{\epsilon}u_{\epsilon}^{2}}\,dv_g,\ \ \mu _{\epsilon}=\frac{\beta _{\epsilon}}{{\rm Area}(\Sigma )}\int_{\Sigma}{u}_{\epsilon}e^{\alpha _{\epsilon}u_{\epsilon}^{2}}\,dv_g.
\end{aligned} \right.
\end{eqnarray}
\end{lemma}
It follows from Lebesgue's dominated convergence theorem that
\begin{eqnarray}\label{3}
{\lim_{\epsilon \rightarrow 0}}
F_{\alpha}^{2\pi-\epsilon}(u_\epsilon)=\sup_{u \in\mathcal { S}}F_{\alpha}^{2\pi}(u).
\end{eqnarray}
Seeing the fact of $1+t e^{t}\geq e^{t} $ for any $t \geq 0$, we get
\begin{eqnarray*}
\lambda _\epsilon=\int_{\Sigma}{u}_\epsilon^{2}e^{\alpha _\epsilon u_\epsilon^{2}}\,dv_g
\geq \frac{1}{\alpha_\epsilon}\int_{\Sigma}\left(e^{\alpha_\epsilon u_\epsilon^{2}}-1\right) d v_g,
\end{eqnarray*}
which together with (\ref{3}) leads to
\begin{eqnarray}\label{L2}
\liminf_{\epsilon \rightarrow 0} \ \lambda _{\epsilon}>0.
\end{eqnarray}
In view of (\ref{e-u}), (\ref{L2}) and $\beta _ { \epsilon } \leq 1$, we obtain
\begin{eqnarray}
\left| \frac{\mu _{\epsilon}}{\lambda _{\epsilon}} \right|
\nonumber &\leq& \frac{1}{\lambda _{\epsilon}{\rm Area}(\Sigma )}\left( \int_{\left\{ u\in \Sigma :\left| u_{\epsilon} \right|\ge 1 \right\}}{|u_{\epsilon}}|e^{\alpha _{\epsilon}u_{\epsilon}^{2}}\,dv_g+\int_{\left\{ u\in \Sigma :\left| u_{\epsilon} \right|<1 \right\}}{|u_{\epsilon}}|e^{\alpha _{\epsilon}u_{\epsilon}^{2}}\,dv_g \right) \\
\nonumber &\leq& \frac{1}{\lambda _{\epsilon}{\rm Area}(\Sigma )}\left( \int_{\left\{ u\in \Sigma :\left| u_{\epsilon} \right|\ge 1 \right\}}{u_{\epsilon}^{2}}e^{\alpha _{\epsilon}u_{\epsilon}^{2}}\,dv_g+\int_{\left\{ u\in \Sigma :\left| u_{\epsilon} \right|<1 \right\}}{e^{\alpha _{\epsilon}u_{\epsilon}^{2}}\,dv_g} \right) \\
\nonumber &\leq& \frac{1}{{\rm Area}(\Sigma )}+\frac{e^{\alpha _{\epsilon}}}{\lambda _{\epsilon}}\\
\label{4} &\leq& C.
\end{eqnarray}
\subsection*{\textbf{Step 2.} Blow-up analysis}
Since $u_{\epsilon}$ is bounded in $ W^{1,2}( \Sigma , g)$,
there exists some function $u_0\in W^{1,2}( \Sigma , g)$ such that
\begin{eqnarray}\label{s2}
\left\{ \begin{aligned}
&u_{\epsilon}\rightharpoonup u_0\, \, \mathrm{weakly}\, \, \mathrm{in}\, \, W^{1, 2}\left(\Sigma,g \right),\\
&u_{\epsilon}\rightarrow u_0\, \, \mathrm{strongly}\, \, \mathrm{in}\, \, L^p\left( \Sigma,g\right) , \forall p>1,\\
&u_{\epsilon}\rightarrow u_0\, \, \mathrm{a. e.} \, \, \mathrm{in}\, \, \Sigma.
\end{aligned} \right.
\end{eqnarray}
Then we have $\int _ { \Sigma } u _ { 0 } dv _ { g } = 0$ and $\|\nabla_gu_0\|_2^2\leq1$.
We set $c_{\epsilon}=|u_{\epsilon}( x_\epsilon )| =\max_{\overline{\Sigma}}|u_{\epsilon}|$.
We first assume that $c_{\epsilon}$ is bounded, which together with elliptic estimates completes the proof of Theorem \ref{T1} ($ii$).
Without loss of generality, we assume
\begin{eqnarray}\label{c}
c_{\epsilon}=u_{\epsilon}( x_\epsilon )\rightarrow+ \infty
\end{eqnarray}
and $x_\epsilon\rightarrow x_0$ as $\epsilon\rightarrow0.$
Applying maximum principle to (\ref{e-u}), we have $x_0 \in \partial \Sigma$.
Following (\cite{Yang-Zhou}, Lemma 4),
we can take an isothermal coordinate system $(U,\phi)$ near $x_0$, such that $\phi(x_0) = 0$, $\phi (U)=\mathbb { B }_r^+$ and $\phi (U\cap \partial\Sigma)= \partial\mathbb { R } ^ { 2 }_+\cap\mathbb { B }_r$ for some fixed $r>0$, where $\mathbb { B }_r^+=\{(x_1,\ x_2)\in \mathbb { R } ^ 2: x_1^2+x_2^2\leq r^2,\ x_2>0 \}$ and $\mathbb{R}^2_+=\left\{x=(x_1,x_2)\in \mathbb { R } ^ 2:x_2>0\right\}$. In such coordinates, the metric $g$ has the representation $g = e^{2f} \left(dx_1^2 +dx_2^2 \right)$ and $f$ is a smooth function with $f (0) = 0$.
Denote $\tilde{x}_\epsilon=\phi(x_\epsilon)$ and $\tilde{u}_\epsilon=u_\epsilon\circ\phi^{-1}$.
To proceed, we observe an energy concentration phenomenon of $u_\epsilon$.
\begin{lemma}\label{L3}
There hold $u_0=0$ and $ |\nabla_g u_\epsilon|^2\,dv_g\rightharpoonup \delta _{x_0} $ in sense of measure, where $\delta _{x_0}$ stands for the Dirac measure centered at $x_0$.
\end{lemma}
\begin{proof}
We first prove $u_0 \equiv 0$. Suppose not, we can see that $0< \lVert \nabla_g u_0 \rVert _{2}^{2}\leq1$. Letting $\eta=\lVert \nabla_g u_0 \rVert _{2}^{2}$,
one has $\lVert \nabla_g \left( u_{\epsilon}-u_0 \right) \rVert _{2}^{2}\rightarrow 1-\eta<1$ and $1 + \alpha \left\| u _ { \epsilon } \right\| _ 2 ^ { 2 } \rightarrow 1 + \alpha \left\| u _ { 0 } \right\| _ 2 ^ { 2 } \leq 1+\eta$ as $\epsilon\rightarrow0$.
For sufficiently small $\epsilon$, we obtain
\begin{eqnarray*}
\left(1 + \alpha \left\| u _ { \epsilon } \right\| _ 2 ^ { 2 } \right)\lVert \nabla_g \left( u_{\epsilon}-u_0 \right) \rVert _{2}^{2}\leq \frac{2-\eta^2}{2}<1.
\end{eqnarray*}
From the H\"older inequality, there holds
\begin{eqnarray*}
\int_{\Sigma}{e^{q\alpha _{\epsilon}u_{\epsilon}^{2}}}\,dv_g&\leq& \int_{\Sigma}{e^{q\left( 1+\frac{1}{\delta} \right)\alpha _{\epsilon} u_{0}^{2}+q\left( 1+\delta \right)\alpha _{\epsilon} \left( u_{\epsilon}-u_0 \right) ^2}}\,dv_g\\
&\leq& \left( \int_{\Sigma}{e^{rq\left( 1+\frac{1}{\delta} \right)\alpha _{\epsilon} u_{0}^{2}}}\,dv_g \right) ^{\frac{1}{r}}\left( \int_{\Sigma}{e^{sq \left( 1+\delta \right)\left( 2\pi -\epsilon \right) \left( 1+\alpha \lVert u_{\epsilon}\lVert _2^{2} \right) \lVert \nabla_g \left( u_{\epsilon}-u_0 \right) \rVert _{2}^{2}\frac{\left( u_{\epsilon}-u_0 \right) ^2}{\lVert \nabla_g \left( u_{\epsilon}-u_0 \right) \rVert _{2}^{2}}}}\,dv_g \right) ^{\frac{1}{s}}\\
&\leq& C\left( \int_{\Sigma}{e^{sq \left( 1+\delta \right) \left( 2\pi -\epsilon \right)\frac{2-\eta^2}{2} \frac{\left( u_{\epsilon}-u_0 \right) ^2}{\lVert \nabla_g \left( u_{\epsilon}-u_0 \right) \rVert _{2}^{2}}}}\,dv_g \right) ^{\frac{1}{s}}
\end{eqnarray*}
for sufficiently small $\delta$, some $r,\ s,\ q>1$ satisfying ${1}/{r}+{1}/{s}=1$ and $sq\left( 1+\delta \right) \left( 2-\eta^2 \right)/2 <1 $. In view of the Trudinger-Moser inequality (\ref{2006Y}), we get $e^{\alpha _{\epsilon}u_{\epsilon}^{2}}$ is bounded in $L^q\left( \Sigma, g \right) $. Hence $\Delta_g u_{\epsilon}$ is bounded in some $L^q\left( \Sigma,g \right)$ from (\ref{e-u}) and (\ref{4}). Applying the elliptic estimate to (\ref{e-u}), one gets $u_{\epsilon}$ is uniformly bounded, which contradicts our assumption $c_{\epsilon}\rightarrow +\infty$. That is to say $u_0 \equiv 0$.
Next we prove $|\nabla_g u_{\epsilon}|^2\,dv_g\rightharpoonup \delta_{x_0}$ in sense of measure. Suppose not. There exists some $r>0$ such that \begin{eqnarray*}\lim_{\epsilon \rightarrow 0}\int_{B_{r}(x_0)}{\left| \nabla_g u_{\epsilon} \right|}^2\,dv_g:=\eta <1,\end{eqnarray*}
where ${B_{r}(x_0)}$ is a geodesic ball centered at $x_0$ with radius $r$.
For sufficiently small $\epsilon$, we can see that $\int_{B_{r}(x_0)}{| \nabla _gu_{\epsilon} |}^2\,dv_g\leq (\eta +1)/{2}<1.$
Then we choose a cut-off function $\rho$ in $C_{0}^{1}\left(\phi\left(B_{r_0}(x_0)\right) \right)$, which is equal to $1$ in $\overline{\phi\left(B_{r_0/2}(x_0)\right) }$ such that
\begin{eqnarray*}\int_{\phi \left( B_{r}\left( x_0 \right)\right) }|\nabla_g \left( \rho \tilde{u}_\epsilon \right) |^2dx
\leq \frac{\eta +3}{4}<1.\end{eqnarray*}
Hence we obtain
\begin{eqnarray*}
\int_{ B_{r/2}\left( x_0 \right) }{e^{\alpha _{\epsilon}qu_{\epsilon}^{2}}}\,dv_g&=&\int_{\phi \left( B_{r/2}\left( x_0 \right) \right)}{e^{\alpha _{\epsilon}q\tilde{u}_\epsilon^{2}}}e^{2f}dx
\\
&\leq& C\int_{\phi \left( B_{r}\left( x_0 \right) \right) }{e^{\alpha _{\epsilon}q\left( \rho \tilde{u}_\epsilon \right) ^2}}dx
\\
&\leq& C\int_{\phi \left( B_{r}\left( x_0 \right)\right) }{e^{\alpha _{\epsilon}q \frac{\eta +3}{4} \frac{\left( \rho \tilde{u}_\epsilon \right) ^2}{\int_{\phi \left( B_{r}\left( x_0 \right)\right) }|\nabla_g \left( \rho \tilde{u}_\epsilon \right) |^2dx}}}dx.
\end{eqnarray*}
From the Trudinger-Moser inequality (\ref{2006Y}), we get $e^{\alpha _{\epsilon}u_{\epsilon}^{2}}$ is bounded in $L^q\left( B_{r/2}(x_0), g\right) $ for any $q>1$ satisfying $q(\eta +3)/4\leq1$. Applying the elliptic estimate to (\ref{e-u}), one gets $u_{\epsilon}$ is uniformly bounded in $B_{r/4}\left( x_0 \right) $. This contradicts (\ref{c}) and ends the proof of the lemma.
\end{proof}
\begin{lemma}\label{L4}
Let
\begin{eqnarray}\label{r}
r _ { \epsilon } = \sqrt{\frac { \lambda _ { \epsilon } } { \beta _ { \epsilon } c _ { \epsilon } ^ { 2 } e ^ { \alpha _ { \epsilon } c _ { \epsilon } ^ { 2 } }}}.
\end{eqnarray}
Then there hold $\lim_{\epsilon\rightarrow0}r_\epsilon= 0$ and $\lim_{\epsilon\rightarrow0}r_\epsilon^2c_\epsilon^k= 0$, where $k$ is a positive integer.
\end{lemma}
\begin{proof}
It is easy to know that $\lim_{\epsilon\rightarrow0}r_\epsilon= 0$ from (\ref{e-u}).
Using the H\"older inequality and (\ref{e-u}), we have
\begin{eqnarray*}
r_{\epsilon}^2 c_{\epsilon}^k&=&\frac{\lambda _{\epsilon}}{\beta _{\epsilon}c_{\epsilon}^{2-k}e^{\alpha _{\epsilon}c_{\epsilon}^{2}}}
\leq \frac{\int_{\Sigma}{u_{\epsilon}^{k}e^{\left( 1-\delta \right) \alpha _{\epsilon}u_{\epsilon}^{2}}\,dv_g}}{\beta _{\epsilon}e^{\left( 1-\delta \right) \alpha _{\epsilon}c_{\epsilon}^{2}}}
\\
&\leq& \frac{1}{\beta _{\epsilon}e^{\left( 1-\delta \right) \alpha _{\epsilon}c_{\epsilon}^{2}}}\left( \int_{\Sigma}{u_{\epsilon}^{kr}\,dv_g} \right) ^{\frac{1}{r}}\left( \int_{\Sigma}{e^{\left( 1-\delta \right) s\alpha _{\epsilon}u_{\epsilon}^{2}}\,dv_g} \right) ^{\frac{1}{s}}\\
&\leq &\frac{1}{\beta _{\epsilon}}\left( \int_{\Sigma}{u_{\epsilon}^{kr}\,dv_g} \right) ^{\frac{1}{r}}\left({\rm Area}(\Sigma ) \right) ^{\frac{1}{s}}
\end{eqnarray*}
for any $0<\delta <1$ and $1/r+1/s=1$.
Then the lemma follows from Lemma \ref{L3} immediately.
\end{proof}
Define
\begin{eqnarray*}
\tilde{u}_{\epsilon}(x)=\left\{\begin{aligned}
&u_{\epsilon} \circ \phi^{-1}\left(x_{1}, x_{2}\right), & x_{2} \geq 0, \\
&u_{\epsilon} \circ \phi^{-1}\left(x_{1},-x_{2}\right), &x_{2}<0,
\end{aligned}\right.
\end{eqnarray*}
and
\begin{eqnarray*}
\tilde{f}(x)=\left\{\begin{aligned}
&f \left(x_{1}, x_{2}\right), & x_{2} \geq 0, \\
&f \left(x_{1},-x_{2}\right), &x_{2}<0,
\end{aligned}\right.
\end{eqnarray*}
on $\mathbb{B}_r$.
Let $U_\epsilon=\left\{x\in\mathbb{R}^2:\ \tilde{x}_\epsilon +r_{\epsilon}x\in\mathbb { B }_r\right\}$. Then one has $U_ { \epsilon }\rightarrow \mathbb{R}^2\,\mathrm{as}\, \epsilon \rightarrow 0$ from Lemma \ref{L4}.
Define two blowing up functions on $ U _ { \epsilon }$,
\begin{eqnarray}\label{psi}
\psi _ { \epsilon } ( x ) &=& \frac{\tilde{u}\left(\tilde{x}_\epsilon+ r _ { \epsilon } x \right)} {c _ { \epsilon }},\\
\label{phi}
\varphi _{\epsilon}\left( x \right) &=&c_{\epsilon}\left( \tilde{u}\left(\tilde{x}_\epsilon+r_{\epsilon}x \right) -c_{\epsilon} \right).
\end{eqnarray}
Now we study the convergence behavior of $\psi _{\epsilon}$ and\ $\varphi _{\epsilon}.$
\begin{lemma}
Up to a subsequence, there hold
\begin{eqnarray}
\label{1} \lim_{ \epsilon\rightarrow0} \psi _{\epsilon}(x) &=& 1 \ \ \ \mathrm{in}\ \ \ C_{loc}^{1}( \mathbb{R}^2 ),\\
\label{6}
\lim_{ \epsilon\rightarrow0}\varphi _{\epsilon}(x)&=&\varphi(x) \ \ \ \mathrm{in}\ \ \ C_{loc}^{1}( \mathbb{R}^2 ),
\end{eqnarray}
where \begin{eqnarray}\label{7}
\varphi \left( x \right) =-\frac{1}{2\pi}\log \left( 1+\frac{\pi}{2} \left| x \right|^2 \right)
\end{eqnarray}
and
\begin{eqnarray}\label{8}
\int_{\mathbb{R}^2_+}e^{4\pi \varphi(x)}dx=1.
\end{eqnarray}
\end{lemma}
\begin{proof}
By (\ref{e-u}) and (\ref{r})-(\ref{phi}), a direct computation shows
\begin{eqnarray}\label{e-psi}
-\Delta _{\mathbb{R}^2}\psi _{\epsilon}&=&\left( c_{\epsilon}^{-2}\psi _{\epsilon}e^{\alpha _{\epsilon}\left( \psi _{\epsilon}+1 \right) \varphi _{\epsilon}}+r_{\epsilon}^{2}\gamma _{\epsilon}\psi _{\epsilon}-\frac{r_{\epsilon}^{2}\mu _\epsilon}{c_{\epsilon}\lambda _{\epsilon}} \right) e^{2\tilde{f}\left(\tilde{x}_\epsilon+r_{\epsilon}x \right)},\\
\label{e-phi}
-\Delta_{\mathbb{R}^2}\varphi _{\epsilon}&=&\left( \psi _{\epsilon}e^{\alpha _{\epsilon}\left( \psi _{\epsilon}+1 \right) \varphi _{\epsilon}}+c_{\epsilon}^2r_{\epsilon}^{2}\gamma _{\epsilon}\psi _{\epsilon}-\frac{c_{\epsilon}r_{\epsilon}^{2}\mu _{\epsilon}}{\lambda _{\epsilon}} \right) e^{2\tilde{f}\left( \tilde{x}_\epsilon+r_{\epsilon}x \right)}.
\end{eqnarray}
Since $\left| \psi _{\epsilon} \right|\leq 1$ and $ \lim_{{\epsilon}\rightarrow 0}-\Delta_{\mathbb{R}^2} \psi _{\epsilon}=0$,
we have by the elliptic estimate to (\ref{e-psi}) that $\lim_{{\epsilon}\rightarrow 0}\psi _{\epsilon}=\psi$ in $ C_{loc}^{1}( \mathbb{R}^2 )$, where $\psi$ is a bounded harmonic function in $\mathbb{R}^2$. Note that $\psi \left( 0 \right) =\lim_{\epsilon \rightarrow 0}\,\psi _{\epsilon}\left( 0 \right) =1$. It follows from the Liouville theorem that $\psi \equiv 1$ in $\mathbb{R}^2$. That is to say (\ref{1}) holds.
Note that $\varphi _{\epsilon}\left( x \right) \leq \varphi _{\epsilon}\left( 0 \right) =0$ for any $ x\in U_\epsilon$. Applying Lemma \ref{L4} and the elliptic estimate to (\ref{e-phi}), we obtain (\ref{6}),
where $\varphi$ satisfies
\begin{eqnarray*}
\left\{ \begin{aligned}
&-\Delta_{\mathbb{R}^2} \varphi =e^{4\pi \varphi}\,\,\text{in\,\,}\mathbb{R}^2,\\
&\varphi \left( 0 \right) =0=\sup_{\mathbb{R}^2}\,\varphi,\\
&\int_{\mathbb{R}^2}{e^{4\pi \varphi}dx\leq 2}.
\end{aligned} \right.
\end{eqnarray*}
By the uniqueness theorem in Chen-Li \cite{CL}, we have (\ref{7}).
Moreover, a simple calculation gives
\begin{eqnarray}\label{2}
\int_{\mathbb{R}^2}{e^{4\pi \varphi}dx= 2}.
\end{eqnarray}
Denote $U_{\epsilon}^{+}=\left\{x \in \mathbb{R}^{2}: \tilde{x}_\epsilon +r_{\epsilon}x\in\mathbb { B }_r^{+}\right\}$ and $U_{\epsilon}^{-}=\left\{x \in \mathbb{R}^{2}:\tilde{x}_\epsilon +r_{\epsilon}x\in\mathbb { B }_r^{-}\right\} .$ For any fixed $R>0$,
let $\mathbb{B}_{R}^{\prime}=\left\{x\in \mathbb{B}_{R}: \tilde{x}_\epsilon +r_{\epsilon}x\in\mathbb { B }_r^{+}\right\}$ and $\mathbb{B}_{R}^{\prime \prime}=\left\{x \in \mathbb{B}_{R}:\tilde{x}_\epsilon +r_{\epsilon}x\in\mathbb { B }_r^{-}\right\}$, we have
\begin{eqnarray*}
\int_{\mathbb{B}_{R}} e^{4 \pi \varphi} dx&=&\lim _{\epsilon \rightarrow 0} \int_{\mathbb{B}_{R}} \frac{1}{\beta_{\epsilon}} \psi_{\epsilon}^{2} e^{\alpha_{\epsilon}\left(1+\psi_{\epsilon}\right) \varphi_{\epsilon}} dx \\
&=&\lim _{\epsilon \rightarrow 0} \int_{\mathbb{B}_{R r_{\epsilon}}\left(\tilde{x}_\epsilon\right)} \frac{1}{\lambda_{\epsilon}} \tilde{u}_{\epsilon}^{2} e^{\alpha_{\epsilon} \tilde{u}_{\epsilon}^{2}} dx\\
& \leq& \lim _{\epsilon \rightarrow 0} \int_{\mathbb{B}_{R r_{\epsilon}}^{+}\left(\tilde{x}_\epsilon\right)} \frac{1}{\lambda_{\epsilon}} \tilde{u}_{\epsilon}^{2} e^{\alpha_{\epsilon} \tilde{u}_{\epsilon}^{2}} dx+\lim _{\epsilon \rightarrow 0} \int_{\mathbb{B}_{R r_{\epsilon}}^{-}\left(\tilde{x}_\epsilon\right)} \frac{1}{\lambda_{\epsilon}} \tilde{u}_{\epsilon}^{2} e^{\alpha_{\epsilon} \tilde{u}_{\epsilon}^{2}} dx.
\end{eqnarray*}
This inequality together with $\int_{U_\epsilon} \tilde{u}_{\epsilon}^{2} e^{\alpha_{\epsilon} \tilde{u}_{\epsilon}^{2}}dx \leq \lambda_{\epsilon}$ and (\ref{2}) gives
\begin{eqnarray*}
\lim_{R \rightarrow+\infty} \lim_{\epsilon \rightarrow 0} \int_{\mathbb{B}_{R r_{\epsilon}}^{+}\left(\tilde{x}_\epsilon\right)} \frac{1}{\lambda_{\epsilon}} \tilde{u}_{\epsilon}^{2} e^{\alpha_{\epsilon} \tilde{u}_{\epsilon}^{2}} dx&=&1, \\
\lim_{R \rightarrow+\infty} \lim_{\epsilon \rightarrow 0} \int_{\mathbb{B}_{R r_{\epsilon}}^{-}\left(\tilde{x}_\epsilon\right)} \frac{1}{\lambda_{\epsilon}} \tilde{u}_{\epsilon}^{2} e^{\alpha_{\epsilon} \tilde{u}_{\epsilon}^{2}} dx&=&1.
\end{eqnarray*}
That is to say (\ref{8}) holds. Then we have the lemma.
\end{proof}
Next we discuss the convergence behavior of $u_\epsilon$ away from $x_0$.
Denote $u_{\epsilon,\,\beta }=\min \{\ \beta c_{\epsilon} , u_{\epsilon} \} \in W^{1,2}\left(\Sigma, g \right) $ for any real number $0<\beta<1$. Following (\cite{Yang2006IJM}, Lemma 3.6), we get
\begin{eqnarray}\label{9}
\lim_{\epsilon \rightarrow 0}\left\| \nabla_g u _ { \epsilon ,\ \beta } \right\| _ { 2 } ^ { 2 }=\beta.
\end{eqnarray}
\begin{lemma}\label{L6}
Letting $\lambda _{\epsilon}$ be defined by (\ref{e-u}), we obtain
\begin{linenomath}
\begin{flalign*}
\begin{split}
&(i)\,\limsup_{ \epsilon \rightarrow 0 }F_{\alpha}^{2\pi-\epsilon}(u_\epsilon)
= {\rm Area}( \Sigma)+\lim_{\epsilon \rightarrow 0}\,\frac{\lambda _{\epsilon}}{c_{\epsilon}^{2}}, \\
&(ii)\,\lim_{\epsilon \rightarrow 0}\frac{\lambda _{\epsilon}}{c_{\epsilon}^{2}}
=\lim_{R\rightarrow +\infty} \lim_{\epsilon \rightarrow 0}\int_{\phi ^{-1}\left( \mathbb{B}^+_{Rr_{\epsilon}}( \tilde{x}_{\epsilon} )\ri )}{e^{\alpha_\epsilon u_{\epsilon}^{2}}}\,dv_g.
\end{split}&
\end{flalign*}
\end{linenomath}
\end{lemma}
\begin{proof}
Recalling (\ref{e-u}) and (\ref{9}), for any real number $0<\beta<1$, one gets
\begin{eqnarray*}
F_{\alpha}^{2\pi-\epsilon}(u_\epsilon)-{\rm Area}( \Sigma)
&=&
\int_{\left\{ x\in \Sigma :\, u_{\epsilon}\leq \beta c_{\epsilon} \right\}}{( e^{\alpha _{\epsilon}u_{\epsilon}^{2}}-1 )\, \,dv_g}+\int_{\left\{ x\in \Sigma :\, u_{\epsilon}>\beta c_{\epsilon} \right\}}{( e^{\alpha _{\epsilon}u_{\epsilon}^{2}}-1 )\, \,dv_g}
\\
&\leq& \int_{\Sigma}{( e^{\alpha _{\epsilon}u_{\epsilon ,\,\beta}^{2}}-1 )\, \,dv_g}+\frac{u_{\epsilon}^{2}}{\beta ^2c_{\epsilon}^{2}}\int_{\left\{ x\in \Sigma :\,u_{\epsilon}>\beta c_{\epsilon} \right\}}{e^{\alpha _{\epsilon}u_{\epsilon}^{2}}\,dv_g}
\\
&\leq& \int_{\Sigma}e^{\alpha _{\epsilon}u_{\epsilon ,\,\beta}^{2}}\alpha _{\epsilon}u_{\epsilon}^{2}\,dv_g+\frac{\lambda _{\epsilon}}{\beta ^2c_{\epsilon}^{2}}
\\
&\leq& \left( \int_{\Sigma}e^{r\alpha _{\epsilon}u_{\epsilon ,\,\beta}^{2}}\,dv_g \right) ^{1/r}\left( \int_{\Sigma} \alpha _{\epsilon}^su_{\epsilon}^{2s}\,dv_g \right) ^{1/s}+\frac{\lambda _{\epsilon}}{\beta ^2c_{\epsilon}^{2}}.
\end{eqnarray*}
By (\ref{2006Y}) and (\ref{9}), $e^{\alpha _{\epsilon}u_{\epsilon ,\,\beta}^{2}}$ is bounded in $L^r\left( \Sigma, g \right)$ for some $r>1$.
Then letting $\epsilon \rightarrow 0$ first and then $\beta \rightarrow 1$, we obtain
\begin{eqnarray}\label{s111}
\limsup _ { \epsilon \rightarrow 0 } F_{\alpha}^{2\pi-\epsilon}(u_\epsilon)-{\rm Area}( \Sigma)\leq \limsup_{\epsilon \rightarrow 0}\,\frac{\lambda _{\epsilon}}{c_{\epsilon}^{2}}.
\end{eqnarray}
According to $c_{\epsilon}={\max}_{\overline{\Sigma}}u_{\epsilon}$, (\ref{e-u}) and Lemma \ref{L3}, we have
\begin{eqnarray*}
F_{\alpha}^{2\pi-\epsilon}(u_\epsilon)-{\rm Area}( \Sigma) \geq \frac{{\lambda_{\epsilon}}}{c_{\epsilon}^{2}}-
\frac{1}{c_{\epsilon}^{2}}\int_{\Sigma}{u_{\epsilon}^{2}}\,dv_g,
\end{eqnarray*}
that is to say
\begin{eqnarray}\label{s112}
\limsup _ { \epsilon \rightarrow 0 }F_{\alpha}^{2\pi-\epsilon}(u_\epsilon) -{\rm Area}\left( \Sigma \right) \geq \liminf _ {\epsilon\rightarrow0}\frac{\lambda_{\epsilon}}{c_{\epsilon}^{2}}.
\end{eqnarray}
Combining (\ref{s111}) and (\ref{s112}), one gets ($i$).
Applying (\ref{e-u}) and (\ref{r})-(\ref{phi}), we obtain
\begin{eqnarray*}
\int_{\phi ^{-1}\left( \mathbb{B}^+_{Rr_{\epsilon}}( \tilde{x}_{\epsilon} ) \right)}{e^{\alpha _{\epsilon}u_{\epsilon}^{2}}}\,dv_g
&=&\int_{\mathbb{B}^+_{Rr_{\epsilon}}\left( \tilde{x}_{\epsilon} \right)}{e^{\alpha _{\epsilon}u_{\epsilon}^{2}\left( x \right)}e^{2f\left( x \right)}}dx
\\
&=&\int_{\mathbb{B}^+_R\left( 0 \right)}{r_{\epsilon}^{2}e^{\alpha _{\epsilon}c_{\epsilon}^{2}\left( x \right)}e^{\alpha _{\epsilon}\left( \psi _{\epsilon}\left( x \right) +1 \right) \varphi _{\epsilon}\left( x \right)}e^{2f\left( \tilde{x}_{\epsilon}+r_{\epsilon}x \right)}}dx
\\
&=&\frac{\lambda _{\epsilon}}{c_{\epsilon}^{2}}\int_{\mathbb{B}^+_R\left( 0 \right)}\frac{1}{\beta _{\epsilon}}{e^{\alpha _{\epsilon}\left( \psi _{\epsilon}\left( x \right) +1 \right)\, \varphi _{\epsilon}\left( x \right)}e^{2f\left( \tilde{x}_{\epsilon}+r_{\epsilon}x \right)}}dx.
\end{eqnarray*}
Letting $\epsilon \rightarrow 0$ first and then $R \rightarrow +\infty$, we have ($ii$)
by (\ref{6})-(\ref{8}).
\end{proof}
Next we consider the properties of $c_{\epsilon}u_{\epsilon}$.
Using the similar idea of (\cite{Yang2006IJM}, Lemma 3.9), one gets
\begin{eqnarray}\label{10}
\frac { \beta _ { \epsilon } } { \lambda _ { \epsilon } } c _ { \epsilon } u _ { \epsilon } e ^ { \alpha _ { \epsilon } u _ { \epsilon } ^ { 2 } } \,dv_g \rightharpoonup \delta_{x_0}.
\end{eqnarray}
After a slight modification of (\cite{Yang2007}, Lemma 4.8), we have
\begin{lemma}\label{L5}
Assume $u \in C ^ { \infty } \left( \overline{\Sigma}\right)$ is a solution of $\Delta_g u = f ( x )$ in $(\Sigma, g)$ and satisfies $ { \| u \| _ { 1 } \leq c _ { 0 } \| f \| _ { 1 } }$. Then for any $1 < q < 2 $, there holds $\| \nabla_g u \| _ { q } \leq C \left( q , c _ { 0 } , \Sigma,\ g \right) \| f \| _ { 1 }.$
\end{lemma}
\begin{lemma}\label{LG}
For any $1 < q < 2$, $c _ { \epsilon } u _ { \epsilon }$ is bounded in $W ^ { 1 , q } (\Sigma, g)$.
Moreover, there holds
\begin{eqnarray*}
\left\{ \begin{aligned}
&c_{\epsilon}u_{\epsilon}\rightharpoonup G\, \, \mathrm{weakly\, \, in\,}\, W^{1,q}\left(\Sigma, g \right), \,\forall 1<q<2, \\
&c_{\epsilon}u_{\epsilon}\rightarrow G\, \, \mathrm{strongly\, \, in \,}\, L^s\left(\Sigma, g \right), \,\forall 1<s<\frac{2q}{2-q}, \\
&c_{\epsilon}u_{\epsilon}\rightarrow G\ \mathrm{in\,}\, {C_{loc}^{1}\left( \Sigma\backslash \left\{ x_0 \right\} \right) },
\end{aligned} \right.
\end{eqnarray*}
where $G$ is a Green function satisfying \begin{eqnarray}\label{e-G}
\left\{ \begin{aligned}
&\Delta_g G=\delta _{x_0}+\alpha G-\frac{1}{{\rm Area}\left( \Sigma \right)}\ {\mathrm{in}}\ \Sigma,\\
&\frac{\partial G}{\partial \mathbf{n}}=0\ {\mathrm{on}}\ {\partial \Sigma}\backslash \left\{ x_0 \right\},\\
&\int_{\Sigma}{G\,dv_g}=0.
\end{aligned} \right.
\end{eqnarray}
\end{lemma}
\begin{proof}
It follows from (\ref{e-u}) that
\begin{eqnarray}\label{11}
\Delta_g \left( c_{\epsilon}u_{\epsilon} \right) =\frac{\beta _{\epsilon}}{\lambda _{\epsilon}}c_{\epsilon}u_{\epsilon}e^{\alpha _{\epsilon}u_{\epsilon}^{2}}+\gamma _{\epsilon}c_{\epsilon}u_{\epsilon}-c_{\epsilon}\frac{\mu _{\epsilon}}{\lambda _{\epsilon}}.
\end{eqnarray}
According to (\ref{e-u}) and (\ref{10}), we have
\begin{eqnarray}\label{12}
\left| \frac{c_{\epsilon}\mu _{\epsilon}}{\lambda _{\epsilon}} \right|=\frac{1}{\text{Area}\left( \Sigma \right)}\int_{\Sigma}{\frac{\beta _{\epsilon}}{\lambda _{\epsilon}}}c_{\epsilon}u_{\epsilon}e^{\alpha _{\epsilon}u_{\epsilon}^{2}}\,dv_g
= \frac{1}{{\rm Area}\left( \Sigma \right)}\left( 1+o_{\epsilon}\left( 1 \right) \right).
\end{eqnarray}
In view of Lemma \ref{L3}, (\ref{e-u}), (\ref{4}), (\ref{10}), (\ref{11}) and (\ref{12}), we have $\Delta_g \left( c _ { \epsilon } u _ { \epsilon } \right)$ is bounded in $L^1(\Sigma, g)$.
From Lemma \ref{L5}, there holds $c _ { \epsilon } u _ { \epsilon }$ is bounded in $W ^ { 1 , q } (\Sigma, g)$ for any $1 < q < 2$.
Then $c_{\epsilon}u_{\epsilon}\rightharpoonup G\, \, \mathrm{weakly\, \, in\,}\, W^{1,q}\left(\Sigma, g \right)$ for any $1<q<2$ and $c_{\epsilon}u_{\epsilon}\rightarrow G\, \, \mathrm{strongly\, \, in \,}\, L^s\left(\Sigma, g \right)$ for any $1<s<{2q}/{(2-q)}$.
We choose a cut-off function $\rho$ in $C^{\infty}\left( \overline{\Sigma} \right)$, which is equal to $0$ in $\overline{B_\delta(x_0)}$ and equal to $1$ in $\Sigma\backslash B_{2\delta}(x_0)$ such that
$\lim_{\epsilon\rightarrow0}{\lVert \nabla_g \left( \rho u_{\epsilon} \right) \rVert _{2}^{2}}=0.$
Hence there holds
\begin{eqnarray*}
\int_{\Sigma \backslash B_{2\delta}\left( x_0 \right)}{e^{s\alpha _{\epsilon}u_{\epsilon}^{2}}}dx\leqslant \int_{\Sigma \backslash B_{2\delta}\left( x_0 \right)}{e^{s\alpha _{\epsilon}\lVert \nabla_g \left( \rho u_{\epsilon} \right) \rVert _{2}^{2}\frac{\rho ^2u_{\epsilon}^{2}}{\lVert \nabla_g \left( \rho u_{\epsilon} \right) \rVert _{2}^{2}}}}dx.
\end{eqnarray*}
From the Trudinger-Moser inequality (\ref{2006Y}), ${e^{\alpha _{\epsilon}u_{\epsilon}^{2}}}$ is bounded in $L^s\left( \Sigma, g \right) $ for some $s>1$.
Applying the elliptic estimate and the compact embedding theorem to (\ref{11}), we obtain $c_{\epsilon}u_{\epsilon}\rightarrow G$ in $ C_{loc}^{1}\left( {\Sigma }\backslash \left\{ x_0 \right\} \right).$
Testing (\ref{11}) by $\phi \in C^1\left( \Sigma \right)$, we obtain (\ref{e-G}).
\end{proof}
Applying the elliptic estimate, we can decompose $G$ as the form
\begin{eqnarray}\label{G}
G=-\frac{1}{\pi}\log |x-x_0|+A_{x_0}+\sigma(x),
\end{eqnarray}
where $A_{x_0}$ is a constant only on $x_0$ and $\sigma(x)\in C^{\infty}\left( \overline{\Sigma}\right)$ with $\sigma(x_0)=0$.
\subsection*{\textbf{Step 3.} Upper bound estimate}
To derive an upper bound of $\sup _ { u \in \mathcal { S} } F _\alpha^ { 2 \pi }( u )$, we use the capacity estimate, which was first used by Li \cite{Li-JPDE} in this topic.
\begin{lemma}\label{L8}
There holds
\begin{eqnarray*}\label{15}
\sup _ { u \in \mathcal { S} } F^ { 2 \pi } _ { \alpha } ( u ) \leq {\rm Area}( \Sigma) + \frac{\pi}{2} e ^ { 1 + 2 \pi A _ { x_0} }.
\end{eqnarray*}
\end{lemma}
\begin{proof}
We take an isothermal coordinate system $(U,\phi)$ near $x_0$ such that $\phi(x_0) = 0$, $\phi (U)\subset \mathbb { R } ^ { 2 }_+$ and $\phi (U\cap \partial\Sigma)\subset \partial\mathbb { R } ^ { 2 }_+$. In such coordinates, the metric $g$ has the representation $g = e^{2f} \left(dx_1^2 +dx_2^2\right )$ and $f$ is a smooth function with $f(0) = 0$.
Denote $\tilde{u}_\epsilon=u_\epsilon\circ\phi^{-1}$.
We claim that
\begin{eqnarray}\label{16}
\lim_{\epsilon \rightarrow 0}\ \frac{\lambda _{\epsilon}}{c_{\epsilon}^{2}}\leq \frac{\pi}{2} e ^ { 1 + 2 \pi A _ { x_0} }.
\end{eqnarray}
To confirm this claim, we set
$ a={\sup_ { \partial \mathbb { B }_ { \delta } \cap \mathbb { R } ^ { 2 }_+ } }\tilde{u}_\epsilon$ and $b=\inf _ { \partial \mathbb { B }_{Rr_\epsilon}\cap \mathbb { R } ^ { 2 }_+ }\tilde{u}_\epsilon$ for sufficiently small $\delta>0$ and some fixed $R>0$.
According to (\ref{6}), (\ref{7}), (\ref{G}) and Lemma \ref{LG}, one gets
\begin{eqnarray*}
a=\frac{1}{c_{\epsilon}}\left( \frac{1}{\pi}\log \frac{1}{\delta}+A_{x_0}+o_\delta(1)+o_{\epsilon}(1)\right),\ \ \\
b=c_{\epsilon}+\frac{1}{c_{\epsilon}}\left( -\frac{1}{2\pi}\log \left( 1+\frac{\pi}{2} R^{2} \right) +o_{\epsilon}( 1 ) \right),
\end{eqnarray*}
where $o_\delta(1)\rightarrow0$ as $\delta\rightarrow 0$ and $o_{\epsilon}(1)\rightarrow0$ as $\epsilon\rightarrow 0$.
It follows from a direct computation that
\begin{eqnarray}\label{17}
\pi ( a-b ) ^2=\pi c_{\epsilon}^{2}+2\log \delta -2\pi A_{x_0}-\log \left( 1+\frac{\pi}{2} R^2\ri ) +o_{\delta}( 1 ) +o_{\epsilon}( 1 ).
\end{eqnarray}
Define a function space
\begin{eqnarray*}
W_{a,b}=\left\{ \tilde{u}\in W^{1,2} \left(\mathbb{B}^+_ \delta\setminus \mathbb{B}^+ _{Rr_\epsilon} \right) :\
\left.\tilde{u}\right|_{ \partial \mathbb { B }_ { \delta } \cap \mathbb { R } ^ { 2 }_+ }=a,\,\,
\left.\tilde{u}\right|_{ \partial \mathbb { B }_{Rr_\epsilon}\cap \mathbb { R } ^ { 2 }_+ }=b,\,\,
\left.\frac{\partial \tilde{u}}{\partial \mathbf{v}}\right|_{\partial \mathbb{R}^{2}_+ \cap \left(\mathbb{B}_{\delta} \backslash \mathbb{B}_{R r_{\epsilon}}\right)}=0 \right\},
\end{eqnarray*}
where $\mathbf{v}$ denotes the outward unit normal vector on $\partial{ \mathbb { R } ^ { 2 }_+ }$.
Applying the direct method of variation, we obtain
$\inf_{u\in W_{a,b}} \int_{\mathbb{B}^+_ \delta \setminus \mathbb{B}^+ _{Rr_\epsilon} }{|\nabla_{\mathbb{R}^2} u|^2}dx$ can be attained by some function $m(x)\in W_{a,b}$ with $\Delta_{\mathbb{R}^2} m( x ) =0$.
We can check that
\begin{eqnarray*}
m( x ) =\frac{a\left( \log|x|-\log ( {Rr_\epsilon} ) \right) +b\left( \log \delta -\log|x| \right)}{\log \delta -\log ( {Rr_\epsilon})}
\end{eqnarray*}
and
\begin{eqnarray}\label{19}
\int_{\mathbb{B}^+_ \delta \setminus \mathbb{B}^+ _{Rr_\epsilon} }{|\nabla_{\mathbb{R}^2} m( x ) |^2}dx=\frac{\pi ( a-b) ^2}{\log \delta -\log ( {Rr_\epsilon} )}.
\end{eqnarray}
Recalling (\ref{e-u}) and (\ref{r}), we have
\begin{eqnarray}\label{20}
\log \delta -\log ( {Rr_\epsilon} ) =\log \delta -\log R-\frac{1}{2}\log \frac{\lambda _{\epsilon}}{\beta _{\epsilon}c_{\epsilon}^{2}}+\frac{1}{2}\alpha _{\epsilon}c_{\epsilon}^{2}.
\end{eqnarray}
Letting $u^*_{\epsilon}=\max \left\{ a,\ \min \left\{b,\ \tilde{u}_{\epsilon} \right\} \right\} \in W_{a,b}$, one gets $| \nabla_{\mathbb{R}^2} u^*_{\epsilon} |\leq | \nabla_{\mathbb{R}^2} \tilde{u}_{\epsilon} |$ in $\mathbb{B}^+_ \delta \setminus \mathbb{B}^+ _{Rr_\epsilon} $ for sufficiently small $\epsilon$.
According to this and $\|\nabla_g u_{\epsilon}\|_2^2=1$, we obtain
\begin{eqnarray}
\nonumber\int_{\mathbb{B}^+_ \delta \setminus \mathbb{B}^+ _{Rr_\epsilon} }{|\nabla _{\mathbb{R}^2}m( x ) |^2}dx
&\leq& \int_{\mathbb{B}^+_ \delta \setminus \mathbb{B}^+ _{Rr_\epsilon} }{|\nabla _{\mathbb{R}^2}u_{\epsilon}^{*}( x ) |^2}dx\\
\label{21}&\leq& 1-\int_{\Sigma \backslash \phi ^{-1}\left(\mathbb{B}^+ _ \delta \right)}{| \nabla_g u_{\epsilon} |}^2\,dv_g-\int_{\phi ^{-1}\left( \mathbb{B}^+ _{Rr_\epsilon} \right)}{| \nabla_g u_{\epsilon} |}^2\,dv_g.
\end{eqnarray}
Now we compute $\int_{\Sigma \backslash \phi ^{-1}\left(\mathbb{B}^+ _ \delta \right)}{| \nabla_g u_{\epsilon} |}^2\,dv_g$ and $\int_{\phi ^{-1}\left( \mathbb{B}^+ _{Rr_\epsilon} \right)}{| \nabla_g u_{\epsilon} |}^2\,dv_g$.
In view of (\ref{e-G}) and (\ref{G}), we obtain
\begin{eqnarray*}
\int_{\Sigma \backslash \phi ^{-1}\left(\mathbb{B}^+ _ \delta \right)}{|}\nabla_g G|^2\,dv_g=\frac{1}{\pi}\log \frac{1}{\delta} +A_{x_0}+\alpha \lVert G \rVert _2^{2}+o_{\epsilon}( 1 ) +o_{\delta}( 1 ).
\end{eqnarray*}
Hence we have by Lemma \ref{LG}
\begin{eqnarray}\label{G3}
\int_{\Sigma \backslash \phi ^{-1}\left(\mathbb{B}^+ _ \delta \right)}{| \nabla_g u_{\epsilon} |}^2\,dv_g=\frac{1}{c_{\epsilon}^{2}}\left(\frac{1}{\pi}\log \frac{1}{\delta} +A_{x_0}+\alpha \lVert G \rVert _2^{2}+o_{\epsilon}( 1 ) +o_{\delta}( 1 )\right).
\end{eqnarray}
It follows from (\ref{phi}), (\ref{6}) and (\ref{7}) that
\begin{eqnarray}\label{22}
\int_{\phi ^{-1}\left( \mathbb{B}^+ _{Rr_\epsilon} \right)}{| \nabla_g u_{\epsilon} |}^2\,dv_g=\frac{1}{c_{\epsilon}^{2}}\left( \frac{1}{2\pi}\log \left( 1+\frac{\pi}{2} R^2 \right) -\frac{1}{2\pi}+o_\epsilon\left( 1 \right)+o_{R}\left( 1 \right)\right),
\end{eqnarray}
where $o_R( 1 ) \rightarrow 0$ as $R\rightarrow +\infty $.
Recalling (\ref{17})-(\ref{22}), we obtain
\begin{eqnarray*}
\log\frac{\lambda _{\epsilon}}{c_{\epsilon}^{2}}\le \log \frac{\pi}{2}+1+2\pi A_{x_0} +o(1),
\end{eqnarray*}
where $o( 1 ) \rightarrow 0$ as $\epsilon \rightarrow 0$ first, then $R\rightarrow +\infty $ and $\delta \rightarrow 0$.
Hence (\ref{16}) is followed.
Combining (\ref{3}), (\ref{16}) and Lemma \ref{L6}, we finish the proof of the lemma.
\end{proof}
From Lemma \ref{L8}, the proof of Theorem \ref{T1} ($ii$) follows immediately under the hypothesis of $c_{\epsilon}\rightarrow +\infty$.
\section{Existence of the Extremal Functions}
The content in this section is carried out under the condition $0 \leq \alpha < \lambda_1(\Sigma)$ and $c_{\epsilon}\rightarrow +\infty$.
Set a cut-off function
$\xi \in C _ { 0 } ^ { \infty } \left( B _ { 2 R \epsilon } ( x_0 ) \right)$ with $\xi = 1$ on $\overline{B _ { R \epsilon } ( x_0)}$ and $\| \nabla_g \xi \| _ { L ^ { \infty } } = O ( ( R \epsilon )^{-1} )$. Denote $\tau =G+(\pi)^{-1}\log |x-x_0|-A_{x_0}$, where $G$ is defined as in (\ref{G}). Let $R= -\log \epsilon $, then $R\rightarrow+\infty$ and $R\epsilon\rightarrow0$ as $\epsilon\rightarrow0$.
We construct a blow-up sequence
\begin{eqnarray}\label{23}
v_{\epsilon}=\left\{\begin{aligned}
&\frac{c^2-\frac{1}{2\pi}{\log \left( 1+\frac{\pi}{2} \frac{|x-x_0|^2}{\epsilon ^2} \right)} +b}{\sqrt{c^2+\alpha \lVert G \rVert _2^{2}}},& \,\,& x\in {B_{R\epsilon}\left( x_0 \right) },\\
&\frac{G-\xi \tau}{\sqrt{c^2+\alpha \lVert G \rVert _2^{2}}},& \,\,\ \ &x\in B_{2R\varepsilon}( x_0 )\backslash B_{R\varepsilon}( x_0 ),\\
&\frac{G}{\sqrt{c^2+\alpha \lVert G \rVert _2^{2}}},& \,\,&x\in \Sigma \backslash B_{2R\varepsilon}( x_0 ),\\
\end{aligned} \right.
\end{eqnarray}
where $b$ and $c$ are constants to be determined later.
In order to assure that $v _{\epsilon}\in C^{\infty}\left( \overline{\Sigma}\right)$, we obtain
\begin{eqnarray}\label{c1}
c^2-\frac{1}{2\pi}\log \left( 1+\frac{\pi}{2} R^2 \right) +b=-\frac{1}{\pi}\log \left( R\epsilon\right) +A_{x_0}.\end{eqnarray}
It follows from $\lVert \nabla_g v _{\epsilon} \rVert _2=1$ that
\begin{eqnarray}\label{c2}
{c}^2=A_{x_0}-\frac{1}{\pi}\log \epsilon+\frac{1}{2\pi}\log\frac{ \pi}{2}-\frac{1}{2\pi}+O\left(\frac{1}{R^2}\right)+O\left( R\epsilon \log ( R\epsilon ) \right) +o_{\epsilon}\left( 1 \right).
\end{eqnarray}
In view of (\ref{c1}) and (\ref{c2}), we have
\begin{eqnarray}\label{B}
b = \frac { 1 } { 2 \pi } + O\left(\frac{1}{R^2}\right)+ O \left( R \epsilon \log ( R \epsilon ) \right)+o_{\epsilon}\left( 1 \right).
\end{eqnarray}
A delicate and simple calculation shows
\begin{eqnarray}\label{28}
\lVert v_{\epsilon} \rVert _{2}^{2}=\frac{\lVert G \rVert _{2}^{2}+O\left( R\epsilon \log \left( R\epsilon \right) \right)}{c^2+\alpha \lVert G \rVert _{2}^{2}}\geq \frac{\lVert G \rVert _{2}^{2}+O\left( R\epsilon \log \left( R\epsilon \right) \right)}{c^2}\left( 1-\frac{\alpha \lVert G \rVert _{2}^{2}}{c^2} \right),
\end{eqnarray}
which gives on $\left(B_{R\epsilon}\left( x_0 \right) , g\right)$
\begin{eqnarray*}2\pi v_{\epsilon}^{2}\left( 1+\alpha \lVert v_{\epsilon} \rVert _{2}^{2} \right) \ge 2\pi c^2+4\pi b-\text{2}\log \left( 1+\frac{\pi}{2}\frac{|x-x_0|^2}{\epsilon ^2} \right) -\frac{4\pi \alpha ^2\lVert G \rVert _{2}^{4}}{c^2}+O\left( \frac{\log R}{c^4} \right) .\end{eqnarray*}
Denote $ { v }^* _ { \epsilon } = \int _ { \Sigma } v _ { \epsilon } \,dv_g/{ {\rm Area}( \Sigma) }$.
It is easy to know that $ { v }^* _ { \epsilon } = O \left( ( R \epsilon ) ^ { 2 } \log \epsilon\right)$ and $v_{\epsilon}-v_{\epsilon}^{*} \in \mathcal{S}$.
On the one hand, by (\ref{c1})-(\ref{28}), there holds
\begin{eqnarray}\label{27}
\int_{B_{R\epsilon}\left( x_0 \right)}{e}^{2\pi \left( v_{\epsilon}-v_{\epsilon}^{*} \right) ^2\left( 1+\alpha \lVert v_{\epsilon}-v_{\epsilon}^{*} \rVert _{2}^{2} \right)}\,dv_g\ge \frac{\pi}{2}e^{1+2\pi A_{x_0}}-\frac{2\pi ^2\alpha ^2\lVert G \rVert _{2}^{4}}{c^2}e^{1+2\pi A_{x_0}}+O\left( \frac{\log R}{c^4} \right) +O\left( \frac{\log\log \epsilon}{R^2} \right).
\end{eqnarray}
On the other hand, from the fact of $e^t\geq t+1$ for any $t\geq0$ and (\ref{23}), one gets
\begin{eqnarray}
\nonumber\int_{\Sigma \backslash B_{R\epsilon}( x_0)}{e}^{2\pi \left( v_{\epsilon}-v_{\epsilon}^{*} \right) ^2\left( 1+\alpha \lVert v_{\epsilon}-v_{\epsilon}^{*} \rVert _2^{2}\right)}\,dv_g
&\geq& \int_{\Sigma \backslash B_{2R\varepsilon}( x_0 )}{\left( 1+2\pi ( v_{\epsilon}-v_{\epsilon}^{*} ) ^2\right)}\,dv_g\\
\label{24} &\geq& \text{Area}\left( \Sigma \right) +2\pi \frac{\lVert G \rVert _{2}^{2}}{c^2}+O\left( \frac{\log R}{c^4} \right) +O\left( R^2\epsilon^2 \right) .
\end{eqnarray}
It follows from (\ref{27}) and (\ref{24}) that
\begin{eqnarray*}\label{25}
\int_{\Sigma}{e}^{2\pi \left( v_{\epsilon}-v_{\epsilon}^{*} \right) ^2\left( 1+\alpha \lVert v_{\epsilon}-v_{\epsilon}^{*} \rVert _2^{2} \right)}\,dv_g&\geq
&\text{Area}\left( \Sigma \right) +\frac{\pi}{2}e^{1+2\pi A_{x_0}}+\frac{2\pi \lVert G \rVert _{2}^{2}}{c^2}\left( 1-\pi \alpha ^2\lVert G \rVert _{2}^{2}e^{1+2\pi A_{x_0}} \right) \\
&\ & +O\left( \frac{\log\log \epsilon}{R^2} \right) +O\left( \frac{\log R}{c^4} \right) +O\left( R^2\epsilon^2 \right).
\end{eqnarray*}
According to $R= -\log \epsilon $ and (\ref{c1}), we obtain
\begin{eqnarray}\label{26}
F^ { 2 \pi } _ { \alpha } ( v _ { \epsilon } - v^*_ { \epsilon } ) > {\rm Area}( \Sigma) + \frac{\pi}{2} e ^ { 1 + 2 \pi A _ { x_0} }.
\end{eqnarray}
for sufficiently small $\alpha$ and $\epsilon$.
The contradiction between (\ref{3}) and (\ref{26}) indicates that $c_{\epsilon}$ must be bounded when $\alpha$ is sufficiently small.
When $|c_{\epsilon}|\leq C$, using Lebesgue's dominated convergence, we have
\begin{eqnarray*}
F^{2\pi}_\alpha(u_0)=\sup_{u\in \mathcal{S}}F^{2\pi}_\alpha(u).
\end{eqnarray*}
Moreover, it is easy to see $u_0 \in C^\infty\left(\overline{\Sigma}\right)\cap \mathcal {S}$ from Lemma \ref{L1} and (\ref{s2}).
Therefore, we obtain Theorem \ref{T1} ($iii$).
\nolinenumbers
\section*{References}
|
\section{Introduction}
\label{sec:introduction}
\input{sections/introduction}
\section{Related Work}
\label{sec:related}
\input{sections/related}
\section{Methods}
\label{sec:methods}
\input{sections/methods}
\section{Experiments and Analysis}
\label{sec:experiments}
\input{sections/experiments}
\section{Discussion}
\label{sec:discussion}
\input{sections/discussion}
\myparagraph{Acknowledgements.}
\input{sections/acknowledgements}
{\small
\bibliographystyle{ieee_fullname}
\subsection{Evaluating Image Quality}
\paragraph{Datasets.} We evaluate $\pi$-GAN\xspace on the real-world CelebA~\cite{liu2015faceattributes} and Cats~\cite{cats} datasets, as well as the synthetic CARLA~\cite{DBLP:journals/corr/abs-1711-03938, graf} dataset. CelebA contains 200,000 high-resolution face images of 10,000 different celebrities. We crop the images from the top of the hair to the bottom of the chin. The Cats dataset contains 6,444 $128 \times 128$ images of cat heads. The CARLA dataset contains 10k images of 16 car models with random texture and color properties, rendered with the Carla Driving simulator. We train and evaluate at $128 \times 128$ resolution for all datasets and models. We evaluate all models using a moving average of parameters.
\myparagraph{Baselines.}
We compare against two previous approaches to 3D-aware image synthesis: HoloGAN~\cite{hologan} and Generative Radiance Fields (GRAF)~\cite{graf}. Baseline models were obtained as pre-trained checkpoints directly from the authors or trained until convergence using the recommended hyperparameters.
\myparagraph{Qualitative results.}
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{images/qualitative_comparison.jpg}
\caption{Qualitative comparison on CelebA, Cats, and CARLA.}
\label{fig:qualitative_comparison}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{images/uncurated_images2.jpg}
\caption{Uncurated generated faces, corresponding to the first 30 random seeds.}
\label{fig:uncurated_celeba}
\end{figure*}
Figure~\ref{fig:qualitative_comparison} compares images generated by $\pi$-GAN\xspace, HoloGAN, and GRAF on three datasets.
Qualitatively, HoloGAN achieves good image quality but suffers from multi-view inconsistency. Although it generally produces sharp images, identity shift is visible across rotations, particularly at the edges of the training distribution. HoloGAN struggled on the synthetic CARLA dataset, which featured much larger variations in viewpoint than CelebA or Cats. Previous papers were also unable to obtain consistent HoloGAN baselines on this dataset~\cite{graf}.
GRAF, which allows for explicit camera control, is more capable than HoloGAN at recovering wide viewing angles. Because it utilizes a 3D representation, it renders different views of the same scene with less identity shift than HoloGAN. However, GRAF is less capable than HoloGAN at rendering fine details such as hair and teeth, and generally produces images that are more cartoon-ish and less lifelike than HoloGAN.
Our $\pi$-GAN\xspace combines fine details with the ability to represent a wide range of camera angles. Compared with HoloGAN and GRAF, it better recreates details such as individual teeth (CelebA) and whiskers (Cats). Because we represent each instance with a radiance field,
$\pi$-GAN\xspace generates images that are inherently view consistent, have minimal identity shift, and that recover a wide range of angles.
\myparagraph{Quantitative results.}
We evaluate image quality using Frechet Inception Distance (FID)~\cite{DBLP:journals/corr/HeuselRUNKH17}, Kernel Inception Distance (KID)~\cite{binkowski2018demystifying}, and Inception Score~\cite{DBLP:journals/corr/SalimansGZCRC16}. Tables \ref{tbl:CelebA}, \ref{tbl:Cats}, and \ref{tbl:Carla} show a quantitative comparison on CelebA, Cats, and CARLA, respectively. We show significant improvements in image quality metrics compared with baselines, particularly on real-world datasets with fine details. Additional results, including precision-recall plots \cite{precision_recall_distributions}, are provided in the supplemental material.
Our evaluation was consistently performed across all models for Table~\ref{tbl:ImageQuality}. Note that specific experiment parameters, such as image crop, may differ from those used by other authors.
\begin{table*}[ht]
\centering
\begin{subfigure}[b]{0.3\textwidth}
\caption{CelebA @ 128 $\times$ 128}
\label{tbl:CelebA}
\begin{tabular}{llll}
\toprule
& FID $\downarrow$ & KID $\downarrow$ & IS $\uparrow$ \\
\midrule
HoloGAN & 39.7 & 2.91 & 1.89 \\
GRAF & 41.1 & 2.29 & 2.34 \\
$\pi$-GAN\xspace & \textbf{14.7} & \textbf{0.39} & \textbf{2.62} \\
\bottomrule
\end{tabular}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.3\textwidth}
\caption{Cats @ 128 $\times$ 128}
\label{tbl:Cats}
\begin{tabular}{llll}
\toprule
& FID $\downarrow$ & KID $\downarrow$ & IS $\uparrow$ \\
\midrule
HoloGAN & 40.4 & 3.30 & 2.03 \\
GRAF & 28.9 & 1.43 & 1.66 \\
$\pi$-GAN\xspace & \textbf{16.8} & \textbf{0.92} & \textbf{2.06} \\
\bottomrule
\end{tabular}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.3\textwidth}
\caption{CARLA @ 128 $\times$ 128}
\label{tbl:Carla}
\begin{tabular}{llll}
\toprule
& FID $\downarrow$ & KID $\downarrow$ & IS $\uparrow$ \\
\midrule
HoloGAN & 67.5 & 3.95 & 3.52 \\
GRAF & 41.7 & 2.43 & 3.70 \\
$\pi$-GAN\xspace & \textbf{29.2} &
\textbf{1.36} & \textbf{4.27} \\
\bottomrule
\end{tabular}
\end{subfigure}
\caption{FID, KID mean$\times$100, and IS for CelebA, Cats, and CARLA datasets.}
\label{tbl:ImageQuality}
\end{table*}
\subsection{Generating Approximate 3D Representations}
A key advantage of our approach over previous CNN attempts at 3D representation learning is that by generating an
implicit radiance field, our model learns an underlying 3D-structure-aware representation. This representation allows for explicit camera control, naturally lends itself to rendering poses that were uncommon or unseen at training time, and is interpretable.
\myparagraph{Extrapolation to rare or unseen camera poses.}
$\pi$-GAN\xspace relies on an underlying 3D structural representation and offers explicit camera control. Like previous methods that offer explicit camera control (e.g.,~\cite{graf}), it more readily renders views and poses outside of the training dataset distribution than previous methods that rely on black-box representations or projections (e.g.,~\cite{hologan}).
Figure~\ref{fig:high_angles} shows that the explicit camera control and representation naturally generalizes to rendering views even from steep angles, although visual artifacts are stronger at the edges of the camera distribution. This is a consequence of the distribution of CelebA images being imbalanced towards front-facing images. As shown in Figure \ref{fig:qualitative_comparison}, CARLA, which features uniformly distributed poses, did not suffer from this issue.
Figure~\ref{fig:zoom_out} illustrates that, despite only training on tightly cropped images, the radiance field extrapolates when we zoom out the camera. Because the radiance field may be rendered from any of a wide variety of angles at training time, the generator is encouraged to produce a radiance field that represents the entire scene, even if only a small portion will be visible in any single image.
To demonstrate that the latent space learned by $\pi$-GAN\xspace is semantically meaningful, we show the results of interpolating between two latent codes in Figure~\ref{fig:interpolation}.
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{images/high_angles_2.jpg}
\caption{$\pi$-GAN\xspace is capable of rendering views from steep angles, producing reasonable results even beyond two standard deviations of camera yaw on CelebA. Face yaw on CelebA is approximately zero-centered Gaussian, with a standard deviation of 17º from the centerline.}
\label{fig:high_angles}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{images/zoom_out_3.jpg}
\caption{Explicit camera control at inference enables rendering views completely absent from the training distribution of camera poses. Although $\pi$-GAN\xspace was trained only on close-up images, it extrapolates to zoomed-out poses.}
\label{fig:zoom_out}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{images/interpolation.jpg}
\caption{Linearly interpolating between two latent codes.}
\label{fig:interpolation}
\end{figure}
\myparagraph{Interpreting the 3D representation.}
Although the color output of the implicit representation depends on ray direction to allow for view-dependent effects, such as specularities, the density output $\sigma$ is completely view independent, resulting in a view-consistent 3D structure that represents a proxy shape of the scene. This 3D structure can be extracted and visualized using the marching cubes algorithm~\cite{lorensen1987marching} on the density output of the conditioned radiance field to produce a surface mesh. Figure~\ref{fig:marching_cubes} shows 3D models extracted from the 3D representation.
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{images/marching_cubes.jpg}
\caption{We can extract a proxy 3D representation as a mesh, either by projecting a depth-map (CelebA, Cats), or through marching cubes (CARLA).}
\label{fig:marching_cubes}
\end{figure}
\subsection{Ablations}
\label{sec:experiments:ablations}
\begin{table}
\centering
\begin{tabular}{lcc}
\toprule
\multirow{2}{*}{Conditioning} & \multicolumn{2}{c}{Architecture} \\
\cmidrule{2-3}
& ReLU P.E. & Sine \\
\midrule
Concatenation & 32.0 & 21.6 \\
Mapping Network & 26.8 & \textbf{5.15} \\
\bottomrule
\end{tabular}
\caption{FID scores on CelebA @ $64 \times 64$, when comparing network architectures with different activation functions and conditioning methods.}
\label{tbl:ablations}
\end{table}
We ablate sinusoidal activations and mapping network conditioning to better understand their individual contributions. We compare radiance fields with sinusoidal activations against radiance fields with ReLU activations and positional encodings (P.E.)~\cite{mildenhall2020nerf}. Moreover, we evaluate radiance fields conditioned with a mapping network and FiLM conditioning against radiance fields conditioned via concatenation~\cite{graf}. Table~\ref{tbl:ablations} summarizes the results of these experiments. Ablations were conducted at $64 \times 64$ in order to save computational resources. Sinusoidal activations and mapping network conditioning each yielded improvements against their respective baselines. However, the combined model, with both sinusoidal activations and a mapping network, was more effective than the sum of its parts.
\begin{figure}
\centering
\includegraphics[width=1.0\linewidth]{images/progressive_ablation.pdf}
\caption{Ablation study for training $\pi$-GAN\xspace with and without progressive growing on CelebA @ $128 \times 128$}
\label{fig:progressive_ablation}
\end{figure}
Figure~\ref{fig:progressive_ablation} compares early training steps for a model trained with progressive growing against a model initialized to the full $128 \times 128$ image resolution. Because computational complexity grows quadratically with image size, progressive growing, which begins at low resolutions, allows for the use of much larger batch sizes at the start of training. The large batch sizes are helpful in stabilizing training, while also allowing for a higher throughput in images per iteration. As others have found before us~\cite{karras2018progressive}, progressive growing, and the larger batch sizes it enables, helped ensure quality and diversity for generated images.
\subsection{SIREN-Based Implicit Radiance Field}
We represent 3D objects implicitly with a neural radiance field, which is parameterized as a multilayer perceptron (MLP) that takes as input a 3D coordinate in space $\mathbf{x} = (x,y,z)$ and the viewing direction $\mathbf{d}$. The neural radiance field outputs both the spatially varying density $\sigma(\mathbf{x}) : \mathbb{R}^3 \rightarrow \mathbb{R}$ and the view-dependent color $(r,g,b) = \mathbf{c}(\mathbf{x},\mathbf{d}) : \mathbb{R}^5 \rightarrow \mathbb{R}^3$. Moreover, we leverage a StyleGAN-inspired mapping network to condition the SIREN{} on a noise vector $\mathbf{z}$ through FiLM conditioning~\cite{perez2017film,dumoulin2018feature-wise}.
As shown in Figure~\ref{fig:network_architecture_diagram}, we formalize the FiLM-ed SIREN{} backbone of our representation as
\begin{align}
\Phi \left( \mathbf{x} \right) = & \phi_{n-1} \circ \phi_{n-2} \circ \ldots \circ \phi_0 \left( \mathbf{x} \right), \\
& \phi_i \left( \mathbf{x}_i \right) = \sin \left( \boldsymbol{\gamma}_i \cdot \left( \mathbf{W}_i \mathbf{x}_i + \mathbf{b}_i \right) + \boldsymbol{\beta}_i \right),
\end{align}
where $\phi_i: \mathbb{R}^{M_i} \mapsto \mathbb{R}^{N_i}$ is the $i^{th}$ layer of an MLP. It consists of an affine transform defined by the weight matrix $\mathbf{W}_i \in \mathbb{R}^{N_i \times M_i}$ and the biases $\mathbf{b}_i\in \mathbb{R}^{N_i}$ applied on the input $\mathbf{x}_i\in\mathbb{R}^{M_i}$, followed by the sine nonlinearity applied to each component of the resulting vector (Figure~\ref{fig:filmsiren_diagram}). Our mapping network is a simple ReLU MLP, which takes as input a noise vector $\mathbf{z}$ and outputs the frequencies $\boldsymbol{\gamma}_i$ and phase shifts $\boldsymbol{\beta}_i$, which condition each layer of the SIREN.
We found this mapping network to be more expressive than concatenation-based conditioning. It yielded image-quality improvements, both for conditioning ReLU-based and SIREN-based neural implicit representations. The ablation studies shown in Sec.~\ref{sec:experiments:ablations} give further insight into these conditioning methods.
Both density and color of our implicit volume are then defined as
\begin{align}
\sigma \left( \mathbf{x} \right) & = \mathbf{W}_{\sigma} \Phi \left( \mathbf{x} \right) + \mathbf{b}_{\sigma}, \\
\mathbf{c} \left( \mathbf{x}, \mathbf{d} \right) & = \mathbf{W}_{c} \phi_{c} \left( \left[ \Phi \left( \mathbf{x} \right), \mathbf{d} \right]^T \right) + \mathbf{b}_{c},
\end{align}
where $\mathbf{W}_{\sigma / c}$ and $\mathbf{b}_{\sigma / c}$ are additional weight and bias parameters.
\subsection{Neural Rendering}
\begin{figure}
\centering
\includegraphics[width=0.7\columnwidth]{images/volume_rendering2.pdf}
\caption{A visualization of our neural volume rendering procedure. Given a conditioned radiance field, we cast rays from the camera origin $\mathbf{o}$, sample density \textsigma\ and color $\mathbf{c}$ values along each ray, and calculate pixel color $\mathbf{C}$ using Eq.~\ref{eqn:volumetric_rendering}.}
\label{fig:volume_rendering}
\end{figure}
We render a neural radiance field from arbitrary camera poses $\xi$ using neural volume rendering. For this purpose, we employ a pinhole camera model and cast rays from the camera origin $\mathbf{o}$ to compute the integrals along each ray through the volume. At every sample, our generator predicts the volume density $\sigma$ and color $\mathbf{c}$. The pixel color $\mathbf{C}$ for a camera ray $\mathbf{r}(t) = \mathbf{o} + t \mathbf{d}$ with near and far bounds $t_n$ and $t_f$ is then calculated using the volume rendering equation~\cite{Max:1995}:
\begin{align}
\label{eqn:volumetric_rendering}
\begin{split}
\mathbf{C} (\mathbf{r}) & = \int_{t_n}^{t_f} T(t) \sigma \left(\mathbf{r}(t) \right) \mathbf{c} \left(\mathbf{r}(t),\mathbf{d} \right)dt,
\\
& \text{where} \quad
T(t) = \text{exp} \left(-\int_{t_n}^{t}\sigma(\mathbf{r}(s))ds \right).
\end{split}
\end{align}
Our approach implements a discretized form of this equation using the stratified and hierarchical sampling approach introduced by NeRF~\cite{mildenhall2020nerf} (see Figure~\ref{fig:volume_rendering}).
This neural rendering approach, which is also adopted by GRAF~\cite{graf}, has several advantages over previous 3D-to-2D projections. Neural rendering allows for explicit control over camera pose, focal length, aspect ratio, and other parameters, while simple projections, such as those used by HoloGAN~\cite{hologan}, are restricted to representing poses in the training dataset.
\subsection{Discriminator}
Following ProgressiveGAN~\cite{karras2018progressive}, we use a convolutional discriminator $D_{\discriminatorparams}$ with parameters $\theta_D$ that grows progressively. We begin training at low resolutions and high batch sizes, during which the generator can focus on producing coarse shapes. As training progresses, we increase the image resolution and add new layers to the discriminator to handle the higher resolutions and discriminate fine details. For most experiments, we begin training at $32 \times 32$ and double the resolution twice during training, up to $128 \times 128$. In practice, we found this progressive growing strategy to allow for larger batch sizes at the beginning of training, which helped to stabilize and speed training (see Sec.~\ref{sec:experiments:ablations}). Final results are rendered by sampling $512 \times 512$ pixels.
Unlike ProgressiveGAN~\cite{karras2018progressive}, our generator architecture does not grow; instead, we increase the resolution of the generator by sampling rays more densely from the same implicit representation.
\subsection{Training Details}
At training time, we randomly sample camera poses $\xi$ from a distribution $p_\xi$. The pose distributions for each dataset are known a priori and approximated as either Gaussian, for CelebA and Cats, or uniform, for CARLA (see supplement for details). In our experiments, we constrained camera positions to the surface of a unit sphere and directed the camera to point towards the origin. At training time, pitch and yaw along the sphere were sampled from a distribution that was tuned according to the dataset. Real images $I$ are sampled from the training set with distribution $p_I$. We use the non-saturating GAN loss with R1 regularization~\cite{gan_convergence}:
\begin{align}
\label{eqn:eqlabel}
\begin{split}
\mathcal{L}(\theta, \phi) & = \mathbf{E}_{\mathbf{z}\sim p_z, \xi\sim p_\xi}[f(D_{\discriminatorparams}(G_{\generatorparams}(\mathbf{z}, \xi)))]
\\
& +\mathbf{E}_{I\sim p_\mathcal{D}}[f(-D_{\discriminatorparams}(I)) + \lambda \lvert \nabla D_{\discriminatorparams} (I) \rvert ^ 2],
\\
& \text{where} \quad f(u) = -\log(1 + \exp(-u)).
\end{split}
\end{align}
We train $\pi$-GAN\xspace in a generative adversarial framework in which a generator and discriminator compete in a zero sum game. Our generator tries to minimize Equation~\ref{eqn:eqlabel}, while the discriminator simultaneously tries to maximize Equation~\ref{eqn:eqlabel}. We use the Adam optimizer with $\beta_1 = 0$, $\beta_2 = 0.9$. We initialize learning rates to \SI{5e-5} for the generator and \SI{4e-4} for the discriminator, decayed over training to \SI{1e-5} and \SI{1e-4} respectively. Further training and implementation details can be found in the supplemental materials.
\endinput
\subsection{Implementation Details}
\gw{yes, move this all to the supplement}
[Should this all go to supplemental materials]
We parameterize the neural implicit representation with a SIREN{} network of eight hidden layers of 256 units
Additional layer for view directions
Our mapping network is a fully connected MLP with three hidden layers of 256 units.
Two timescale update rule, with a generator learning rate of 5e-5 decaying to 1e-5 and a discriminator learning rate of 4e-4 decaying to 1e-4
We follow StyleGAN in further lowering the learning rates of the mapping network by a factor of twenty.
We use R1 regularization with a regularization strength of 1
We do not use batch normalization, spectral normalization, attention mechanisms, or further regularization techniques in our models.
\subsection{Generative Adversarial Networks}
\paragraph{Neural representations and rendering.}
Emerging neural implicit scene representations promise 3D-structure-aware, continuous, memory-efficient representations for parts~\cite{genova2019learning,genova2019deep}, objects~\cite{park2019deepsdf,michalkiewicz2019implicit,atzmon2019sal,gropp2020implicit,yariv2020multiview,davies2020overfit,chabra2020deep}, or scenes~\cite{eslami2018neural,sitzmann2019srns,jiang2020local,peng2020convolutional,sitzmann2020siren}. These can be supervised with 3D data, such as point clouds, and optimized as either signed distance functions~\cite{park2019deepsdf,michalkiewicz2019implicit,atzmon2019sal,gropp2020implicit,sitzmann2019srns,jiang2020local,peng2020convolutional,sitzmann2019metasdf,kellnhofer:2021} or occupancy networks~\cite{mescheder2019occupancy,chen2019learning}.
Using neural rendering~\cite{tewari2020state}, implicit neural representations can also be trained using multiview 2D images~\cite{saito2019pifu,sitzmann2019srns,Oechsle2019ICCV,Niemeyer2020CVPR,mildenhall2020nerf,yariv2020multiview,liu2020neural,jiang2020sdfdiff,liu2020dist}.
Temporally aware extensions ~\cite{Niemeyer2019ICCV} and multimodal variants with part-level semantic segmentation~\cite{kohli2020inferring} have also been proposed.
Among these approaches, sinusoidal representation networks (SIREN)~\cite{sitzmann2020siren} and neural radiance fields (NeRF)~\cite{mildenhall2020nerf} are most closely related to our work. Specifically, we use SIREN{} as the representation network architecture of our framework combined with a neural rendering technique inspired by NeRF. Both SIREN{} and NeRF, however, have only been explored in the context of overfitting to individual objects or scenes, whereas we study the combination of aspects of these seminal works for applications in 3D GANs. Exploring the unique challenges of training a neural implicit GAN supervised by natural 2D data is one of the core contributions of our work.
\myparagraph{Generative 3D-aware image synthesis.}
Generative Adversarial Nets (GANs)~\cite{goodfellow2014generative}, or more generally the paradigm of adversarial learning, have led to significant progress in various image synthesis tasks, including image generation~\cite{radford2016unsupervised,karras2018progressive,karras2019style,Karras2020stylegan2}, image-to-image translation \cite{zhu2017unpaired}, interactive image editing~\cite{wang2018high}, and learning from partial and noisy observations~\cite{bora2018ambientgan}. These methods operate on the 2D space of pixels, ignoring the 3D nature of our physical world. This has limited the application of these generative models in tasks such as view synthesis.
Visual Object Networks~\cite{zhu2018visual} and PrGANs~\cite{Gadelha:2017} learn to synthesize 2D images by first generating a voxelized 3D shape using a 3D-GAN~\cite{wu2016learning} and then projecting it into 2D.
HoloGAN~\cite{hologan} and BlockGAN~\cite{nguyen2020blockgan} have extended the system by incorporating a volumetric but implicit 3D representation. While these methods attempt to model the 3D structure of the object in the synthesized image, the use of an explicit volume representation has constrained their resolution~\cite{lunz2020inverse}.
Szab\'o~\etal~\cite{Szabo:2019} and Liao~\etal~\cite{Liao2020CVPR} instead proposed to model 3D shapes as meshes and collections of primitives for image synthesis, respectively. However, these representations lack the expressiveness needed to synthesize high-fidelity pictures.
The work most similar to ours is GRAF~\cite{graf}, which learns a generative model for implicit radiance fields for 3D-aware image synthesis. Although $\pi$-GAN\xspace operates in a similar setting, its network architecture and training strategy differ from GRAF in several ways. First, we use SIREN\ rather than a positionally encoded ReLU MLP as a choice of neural implicit representation. Second, GRAF conditioned its MLP generator on both a shape noise code and an appearance noise code by concatenation; in contrast, we leverage a StyleGAN-inspired mapping network, which conditions the entire MLP on a single input noise vector through FiLM conditioning. Third, we utilize a progressive growing strategy during training. Finally, we did not employ a patch-based discriminator, as used by GRAF, as SIREN{} is prone to local overfitting to the last batch if sufficient coverage of the space is not maintained. Our experiments demonstrate that all of our innovations are critical to high-quality image synthesis results.
Beyond unconditional 3D-aware image generation, there is an orthogonal line of work on conditional reconstruction of 3D shape and texture from partial observations. These reconstructions can later be used for novel view synthesis. Various 3D representations have been considered for the task, including voxels~\cite{henzler2019platonicgan,tulsiani2017learning}, meshes~\cite{kanazawa2018learning,chen2019learning,goel2020shape,henderson2020leveraging,mustikovela2020self}, point clouds~\cite{tatarchenko2016multi}, a depth map~\cite{wu2020unsupervised}, and implicit functions~\cite{rajeswar2020pix2shape,tulsiani2020implicit}. Some of these methods are also grounded in adversarial training. While these methods focus on 3D reconstruction, $\pi$-GAN\xspace aims to learn an unconditional generative model of radiance fields.
\section{Novel View Synthesis Details}
We demonstrate a potential application of $\pi$-GAN\xspace: we can use a trained generator, without modifications, to perform single-view reconstruction. We base our method on the inverse projection procedure outlined by Karras et al.~\cite{Karras2020stylegan2}.
We freeze the parameters of our implicit representation and seek the frequencies $\boldsymbol{\gamma}_i$ and phase shifts $\boldsymbol{\beta}_i$ for each MLP layer $i$ which produce a radiance field that, when rendered, best matches the target image. We initialize $\boldsymbol{\gamma}_i$ and $\boldsymbol{\beta}_i$ to $\bar{\boldsymbol{\gamma}_i}$ and $\bar{\boldsymbol{\beta}_i}$, the center of mass of frequencies and phase shifts for each layer. We calculate $\bar{\boldsymbol{\gamma}_i}$ and $\bar{\boldsymbol{\beta}_i}$ simply by averaging the frequencies and phase shifts of ten thousand random noise vector inputs. We then run gradient descent to minimize the mean-squared-error image reconstruction loss. We additionally introduce an $\mathcal{L}_2$ penalty with a weight of 0.1 during the optimization process to prevent $\boldsymbol{\gamma}_i$ and $\boldsymbol{\beta}_i$ from straying too far from $\bar{\boldsymbol{\gamma}_i}$ and $\bar{\boldsymbol{\beta}_i}$. We optimize the frequencies and phase shifts with the Adam optimizer over 700 iterations. We initialize the learning rate to 0.01, decaying by a factor of 0.5 every 200 iterations.
\section{Model Details}
\paragraph{Mapping Network.}
The mapping network is parameterized as an MLP with three hidden layers of 256 units each. The mapping network uses leaky-ReLU activations with a negative slope of 0.2.
\paragraph{SIREN-based Implicit Radiance Field.}
The FiLMed-SIREN~\cite{sitzmann2020siren} backbone of the generator is parameterized as an MLP with eight FiLMed-SIREN\ hidden layers of 256 units each.
\paragraph{Discriminator.}
\begin{table}
\centering
\caption{Discriminator architecture, showing progressive growing stages.}
\label{tbl:discriminator_layers}
\begin{tabular}{lll}
\toprule
\begin{tabular}[c]{@{}l@{}}\\\end{tabular} & Activation & Output Shape \\
\midrule
\begin{tabular}[c]{@{}l@{}}Input Image\\Adapter Block (1$\times$1)\\Coord Conv 1 (3$\times$3)\\Coord Conv 2 (3$\times$3)\\Avg Pool Downsample\end{tabular} & \begin{tabular}[c]{@{}l@{}}-\\LeakyReLU (0.2)\\LeakyReLU (0.2)\\LeakyReLU (0.2)\\-\\ \end{tabular} & \begin{tabular}[c]{@{}l@{}}3$\times$128$\times$128~\\64$\times$128$\times$128~\\128$\times$128$\times$128\\128$\times$128$\times$128~\\128$\times$64$\times$64\\\end{tabular} \\
\midrule
\begin{tabular}[c]{@{}l@{}}Coord Conv 1 (3$\times$3)\\Coord Conv 2 (3$\times$3)\\Avg Pool Downsample\\\end{tabular} & \begin{tabular}[c]{@{}l@{}}LeakyReLU (0.2)\\LeakyReLU (0.2)\\-\\\end{tabular} & \begin{tabular}[c]{@{}l@{}}256$\times$64$\times$64\\256$\times$64$\times$64\\256$\times$32$\times$32\\\end{tabular} \\
\midrule
\begin{tabular}[c]{@{}l@{}}Coord Conv 1 (3$\times$3)\\Coord Conv 2 (3$\times$3)\\Avg Pool Downsample\\\end{tabular} & \begin{tabular}[c]{@{}l@{}}LeakyReLU (0.2)\\LeakyReLU (0.2)\\-\\\end{tabular} & \begin{tabular}[c]{@{}l@{}}400$\times$32$\times$32\\400$\times$32$\times$32\\400$\times$16$\times$16\\\end{tabular} \\
\midrule
\begin{tabular}[c]{@{}l@{}}Coord Conv 1 (3$\times$3)\\Coord Conv 2 (3x3)\\Avg Pool Downsample\\\end{tabular} & \begin{tabular}[c]{@{}l@{}}LeakyReLU (0.2)\\LeakyReLU (0.2)\\-\\\end{tabular} & \begin{tabular}[c]{@{}l@{}}400$\times$16$\times$16\\400$\times$16$\times$16\\400$\times$8$\times$8\\\end{tabular} \\
\midrule
\begin{tabular}[c]{@{}l@{}}Coord Conv 1 (3$\times$3)\\Coord Conv 2 (3$\times$3)\\Avg Pool Downsample\\\end{tabular} & \begin{tabular}[c]{@{}l@{}}LeakyReLU (0.2)\\LeakyReLU (0.2)\\-\\\end{tabular} & \begin{tabular}[c]{@{}l@{}}400$\times$4$\times$4\\400$\times$4$\times$4\\400$\times$2$\times$2\\\end{tabular} \\
\midrule
Conv 2d (2$\times$2) & & 1$\times$1$\times$1 \\
\bottomrule
\end{tabular}
\end{table}
Table~\ref{tbl:discriminator_layers} shows the architecture of the progressive discriminator. We begin training at low resolutions and progressively add discriminator stages while upsampling image size. In order to smooth transitions between upsamples, we fade in the contributions of new layers over ten-thousand iterations. We utilized CoordConv layers~\cite{liu2018intriguing} and residual connections~\cite{he2016deep} throughout the discriminator.We considered using a patch discriminator similar to GRAF, but found it leads to uneven image quality as SIREN is prone to local overfitting to the last batch if sufficient coverage of the space is not maintained.
\begin{table}
\centering
\caption{FID, KID mean $\times$ 100, and IS for $\pi$-GAN\xspace on CelebA, Cats, and CARLA datasets. }
\label{tbl:piGAN64}
\begin{tabular}{llll}
\toprule
& FID $\downarrow$ & KID $\downarrow$ & IS $\uparrow$ \\
\midrule
CelebA @ $64 \times 64$ & 5.15 & 0.09 & 2.28 \\
Cats @ $64 \times 64$ & 7.36 & 0.23 & 2.07 \\
CARLA @ $64 \times 64$ & 13.59 & 0.34 & 3.85 \\
\bottomrule
\end{tabular}
\end{table}
\begin{figure}
\centering
\includegraphics[width=0.8\linewidth]{images/COLMAP.jpg}
\caption{COLMAP reconstructions for models trained on CelebA, obtained by running COLMAP with default parameters and no known camera poses; GRAF's results were from their supplement.}
\label{fig:COLMAP}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.8\linewidth]{images/precision_recall.jpg}
\caption{Precision-recall plots for $\pi$-GAN\xspace, GRAF, and HoloGAN on CelebA, Cats, and CARLA.}
\label{fig:Precision-Recall}
\end{figure}
\section{Additional Training Details}
We train the majority of our models across two RTX 6000 GPUs or a single RTX 8000 GPU. We begin training at a resolution of $32\times32$, with an initial batch size of 120. At each upsample, we drop the batch size by a factor of four to keep the models and generated images in memory. At higher resolutions, we aggregate across mini-batches to keep an effective batch size at or above 12, given our GPU constraints. To further reduce memory usage, we used PyTorch's Automatic Mixed Precision (AMP). $\pi$-GAN\xspace trained for 10 hours at 32$\times$32, 10 hours at 64$\times$64, and 36 hours at 128$\times$128. Certain rendering and camera parameters were tuned according to the dataset. We use the true pose distribution when it is known, e.g. for synthetic datasets, otherwise we make a guess and tune the distribution as a hyperparameter. We sample camera poses for CelebA from a normal distribution, with a vertical standard deviation of 0.15 radians and a horizontal standard deviation of 0.3 radians. We sample camera poses for Cats from a uniform distribution, with horizontal range $(-0.75, 0.75)$ and vertical range $(-0.4, 0.4)$. We sample poses for CARLA uniformly from the upper hemisphere. We tune the number of samples along each ray to balance memory consumption and depth resolution. We use 24 samples per ray for CelebA and Cats and 64 samples per ray for CARLA. We utilize a pinhole perspective camera with a field of view of 12º for CelebA, 12º for Cats, and 30º for CARLA.
\section{$\pi$-GAN\xspace results @ $64 \times 64$}
Table~\ref{tbl:piGAN64} includes additional quantitative results, evaluated at $64 \times 64$, in order to allow for comparisons of $\pi$-GAN\xspace against models evaluated at lower resolutions.
\section{Additional Visual Results}
We include additional visual results to show the image quality and view consistency of $\pi$-GAN\xspace. Figures~\ref{fig:faces_multi_view} and~\ref{fig:cats_multi_view} demonstrate the wide range of camera poses supported by $\pi$-GAN\xspace for generated faces and cats. Figure~\ref{fig:curated_celeba} shows the fine detail that $\pi$-GAN\xspace renders on larger images. Figure~\ref{fig:cars_multi_views} shows additional cars with varying elevation and rotation. We include several videos of faces and cats with the camera following an elliptical trajectory in our supplementary video.
\section{COLMAP Reconstruction}
In order to demonstrate the images from $\pi$-GAN\xspace are multi-view consistent, we include a COLMAP reconstruction in Figure~\ref{fig:COLMAP}. We observe that proxy shapes extracted from pi-GAN lead to more pleasing novel views when projected to novel camera poses than those from GRAF.
\section{Interpolation and Truncation}
Following the method of StyleGAN~\cite{karras2019style} we can smoothly interpolate between two generated samples by linearly interpolating between the frequencies and phase shifts corresponding to the two latent codes. We include a result in Figure~8 in the paper. Along similar lines, it is also possible to trade off fidelity and diversity at test time following the method proposed in StyleGAN~\cite{karras2019style}. Because truncation reduced the diversity of generated images, we provided all evaluation metrics without truncation.
\section{Precision and Recall}
Recent work in generative models have investigated alternative metrics in order to independently evaluate fidelity and diversity~\cite{precision_recall_distributions, Kynkaanniemi2019}. Figure~\ref{fig:Precision-Recall} provides precision-recall plots on CelebA, Cats, and CARLA, comparing $\pi$-GAN\xspace to GRAF and HoloGAN.
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{images/curated_celeba.jpg}
\caption{Curated examples from our model trained with CelebA~\cite{liu2015faceattributes}.}
\label{fig:curated_celeba}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{images/multi-view-faces.jpg}
\caption{Curated examples from our model trained with CelebA, displayed from multiple viewing angles.}
\label{fig:faces_multi_view}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{images/multi-view-cats.jpg}
\caption{Curated examples from our model trained with Cats~\cite{cats}, displayed from multiple viewing angles.}
\label{fig:cats_multi_view}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{images/cars_multi_view.jpg}
\caption{Curated examples from our model trained with CARLA~\cite{DBLP:journals/corr/abs-1711-03938}, displayed from multiple viewing angles.}
\label{fig:cars_multi_views}
\end{figure*}
|
\section{Introduction}
Full understanding of the radiation generated by accelerating charged particles
is crucial for accelerator physics and electrodynamics in general. The
predictions of classical electrodynamics for pulse-by-pulse average
characteristics of synchrotron radiation, such as the total radiated power, spectral
composition, angular intensity distribution and brightness
\cite{kim2017synchrotron}, are supported by countless observations. In fact,
they are confirmed every day by routine operations of synchrotron radiation
user facilities. On the other hand, the pulse-to-pulse statistical fluctuations
of synchrotron radiation have not been studied at the same level of detail yet,
although substantial progress has been made in the past few decades. The turn-to-turn intensity fluctuations of incoherent spontaneous bending-magnet,
wiggler, and undulator radiation in storage rings have been studied in
Refs.~\cite{teich1990statistical,sajaev2000determination,sajaev2004measurement,lobach2020PRAB,sannibale2009absolute,catravas1999measurement},
both theoretically and experimentally. The statistical properties of the
Free-Electron Laser (FEL) radiation have been studied in
Refs.~\cite{kim1997start,benson1985shot,saldin2013physics,pellegrini2016physics,becker1982photon,becker1983fully,becker1983photon}.
Moreover, Refs.~\cite{chen2001observation,chen1999photon} claimed to observe a non-classical sub-Poissonian photon statistics in the seventh coherent
spontaneous harmonic of an FEL, although it could have been an instrumentation effect
\cite{park2019investigation}. In any case, more experimental studies into the
statistical properties of synchrotron radiation are needed. In this paper, we describe our observation of
turn-to-turn power fluctuations of incoherent spontaneous undulator radiation in
the Integrable Optics Test Accelerator (IOTA) storage ring at Fermilab \cite{antipov2017iota}. Also, we extend the existing theoretical description
\cite{sannibale2009absolute,lobach2020PRAB} of such fluctuations. Namely, in
Refs.~\cite{sannibale2009absolute,lobach2020PRAB}, only the effect of spatial
distribution of the electrons inside the bunch on the turn-to-turn fluctuations
is considered. However, in general, the distribution of electron velocities
affects the fluctuations as well. We present a generalized formula for the
fluctuations in the case of non-negligible beam divergence in this paper.
Fluctuations and noise do not always degrade the results of an experiment. There
are numerous examples when noise is used to measure the parameters of a specific
system, or even fundamental constants. Some examples are the pioneering
determination of the elementary charge $e$ by the shot noise
\cite{hull1925determination}, and the determination of the Boltzmann constant
$k_B$ by the Johnson-Nyquist noise \cite{johnson1928thermal}. Another example,
relevant to the field of accelerator physics, is the use of Schottky noise
pick-ups in storage rings
\cite{boussard1986schottky,van1989diagnostics,caspers20074}
to measure transverse rms emittances, momentum spread, number of particles, etc.
In fact, we will see that the fluctuations in synchrotron radiation are similar
to Schottky noise. Both effects are related to the existence of discrete
point-like charges as opposed to a continuos charge fluid. Therefore,
measurements of electron bunch parameters via synchrotron radiation fluctuations
have been reported too. They were mostly focused on the longitudinal bunch length \cite{sannibale2009absolute,sajaev2004measurement,
sajaev2000determination}. Reference~\cite{catravas1999measurement} reported an
order-of-magnitude measurement of a transverse emittance. In this paper, we
present one example of a measurement of an unknown small vertical emittance of a
flat beam in IOTA, given a known horizontal emittance, a longitudinal bunch
shape, and ring focusing functions, using our new formula for the fluctuations. Taking beam divergence into account is critical in this specific measurement. For
more results regarding beam diagnostics via fluctuations in IOTA, we refer
the reader to our separate publication \cite{lobach2020furjointprl}.
\section{\label{sec:theory}Theoretical description} Let us assume that we have a detector that can measure the number of detected synchrotron radiation photons
$\N$ at each revolution in a storage ring. Then, according to
\cite{lobach2020PRAB,kim2017synchrotron,teich1990statistical}, the variance of this number can be expressed as
\begin{equation}\label{eq:varN_from_book}
\var{\N}=\av{\big(\N-\av{\N}\bigr)^2}=\av{\N}+\frac{1}{M}\av{\N}^2,
\end{equation}
\noindent where the linear term represents the photon shot noise, related to the
quantum discrete nature of light. This effect would exist even if there was only
one electron. Indeed, the electron would radiate photons with Poisson
statistics~\cite{glauber1963quantum,glauber1963coherent,glauber1951some}. The
quadratic term corresponds to the interference of fields radiated by different
electrons. Changes in relative electron positions and velocities, inside the
bunch, result in fluctuations of the radiation power, and, consequently, of the
number of detected photons. In a storage ring, the effect arises because of betatron motion, synchrotron motion, radiation induced diffusion, etc. The
dependence of $\var{\N}$ on the electron bunch parameters is introduced through
the parameter $M$, which will be called the ``number of coherent modes'', following
the nomenclature of \cite{kim2017synchrotron,teich1990statistical,lobach2020PRAB}.
In \cite{lobach2020PRAB}, we derived an equation for $M$ for an electron bunch
with a Gaussian transverse density distribution and an arbitrary longitudinal density
distribution, assuming an rms bunch length much longer than the radiation
wavelength and a negligible electron beam divergence. In this paper, we present
an equation for $M$ extended to an arbitrary beam divergence,
\newcommand{\mathcal{P}_k(\xpv,\phiv_1-\phiv_2)}{\mathcal{P}_k(\xpv,\phiv_1-\phiv_2)}
\begin{widetext}
\begin{equation}\label{eq:Mgeneral}
\frac{1}{M} = (1-1/\Ne)
\frac{\sqrt{\pi}}{\szeff}
\frac{
\int
\dd{k} \dd[2]{\phiv_1} \dd[2]{\phiv_2} \dd[2]{\xpv}
\mathcal{P}_k(\xpv,\phiv_1-\phiv_2)
\mathcal{I}_k(\phiv_1,\xpv)
\mathcal{I}_k^{*}(\phiv_2,\xpv)
}{
\None^2
}
,
\end{equation}
\noindent with
\begin{equation}
\mathcal{P}_k(\xpv,\phiv_1-\phiv_2)=
\frac{1}{4\pi\sxp\syp}
e^{
-\frac{(\xp)^2}{4\sxp^2}
-\frac{(\yp)^2}{4\syp^2}
}
e^{
-ik\dx(\f{1x}-\f{2x})\xp
-ik\dy(\f{1y}-\f{2y})\yp
}
e^{
-k^2\Sx^2(\f{1x}-\f{2x})^2
-k^2\Sy^2(\f{1y}-\f{2y})^2
},
\end{equation}
\end{widetext}
\begin{equation}
\mathcal{I}_k(\phiv,\xpv) =
\sum\limits_{s=1,2}
\etaf{\phiv}
\Ef{\phiv}{}{s}
\Ef{\phiv-\xpv}{*}{s},
\end{equation}
\begin{equation}\label{eq:NoneDef}
\None = \sum\limits_{s=1,2} \int \dd{k}\dd[2]{\phiv}\etaf{\phiv}\bigl|\Ef{\phiv}{}{s}\bigr|^2,
\end{equation}
\noindent where $s=1,2$ indicates the polarization component, $\Ne$ is the
number of electrons in the bunch, $k = 2\pi/\lambda$ is the magnitude of the wave
vector; $\phiv = (\f{x},\f{y})$, $\phiv_1=(\f{1x},\f{1y})$ and $\phiv_2=(\f{2x},\f{2y})$ represent angles of
direction of the radiation in the paraxial approximation. Hereinafter, $x$ and
$y$ refer to the horizontal and the vertical axes, respectively, and
\begin{equation}\label{eq:kappaDef}
\szeff = 1/\bigl(2\sqrt{\pi}\int\rho^2(z)\dd{z}\bigr),
\end{equation}
\noindent where $\rho(z)$ is the electron bunch longitudinal density
distribution function, $\int\rho(z)\dd{z}=1$, and $\szeff$ is equal
to the rms bunch length $\sz$ for a Gaussian bunch; $\xpv=(\xp,\yp)$ represents the
direction of motion of an electron at the radiator center, relative to a reference electron; $\sxp$ and
$\syp$ are the rms beam divergences, $\sxp^2 = \gx \ex + \Dxp^2\sp^2$, $\syp^2 =
\gy\ey$; $\Sx^2 = \ex/\gx+(\gx\Dx+\Dxp\ax)^2\bx\ex\sp^2/\sxp^2$, $\Sy^2 =
\ey/\gy$, $\dx = (\ax\ex-\Dx\Dxp\sp^2)/\sxp^2$, $\dy = \ay/\ey$, where $\ax$,
$\bx$, $\gx$, $\ay$, $\by$, $\gy$ are the Twiss parameters of an uncoupled focusing optics in the synchrotron
radiation source; $\Dx$, $\Dxp$ are the horizontal dispersion and its derivative, and the
vertical dispersion is assumed to be zero; $\ex$, $\ey$ are the unnormalized rms
emittances; $\sp$ is the relative rms momentum spread. The following two useful relations
exist, $\sx^2 = \Sx^2 + \sxp^2\dx^2$, $\sy^2 = \Sy^2 + \syp^2\dy^2$, where $\sx$ and $\sy$ are the transverse rms beam sizes. The complex
radiation field amplitude $\Ef{\phiv}{}{s}$, generated by a reference
electron, is given by the following expression, see
\cite{glauber1963coherent,lobach2020PRAB},\cite[p.~38]{kim2017synchrotron},
\begin{equation}\label{eq:Adef}
\Ef{\phiv}{}{s} = \sqrt{\frac{\alpha k}{2(2\pi)^3}}\int \dd{t} \epolar\cdot\vRef{t}
e^{ick t - i\kv\cdot\Rref{t}}.
\end{equation}
\noindent where $\kv =
k(\f{x},\f{y},1-\f{x}^2/2-\f{y}^2/2)$, $\alpha$ is the fine-structure constant, $\epolar$ is the considered polarization vector ($s=1,2$), $\Rref{t}$ is the trajectory of the
reference electron in the synchrotron radiation source, $\vRef{t}$ is the
velocity of the reference electron as a function of time, $c$ is the speed of light. The electrons are assumed to be ultrarelativistic, $\gamma\gg1$, where $\gamma$ is the Lorentz factor.
The parameter $\None$ in \Cref{eq:Mgeneral} is the average number of detected photons per turn for a single
electron (s.e.) circulating in the ring. We consider the case of an incoherent radiation ($\sz k\gg1$). Therefore, the
average number of detected photons for the entire bunch can be obtained as
\begin{equation}\label{eq:avNeq}
\av{\N} = \Ne \None.
\end{equation}
The integrals in \Cref{eq:Mgeneral,eq:NoneDef} are taken from minus to plus infinity over all integration variables except for $k$, which goes from zero to plus infinity. The spectral sensitivity and the aperture of the detector are assumed to be included in the detection efficiency $\etaf{\phiv}$, which is a
function of polarization, $k$, and $\phiv$ for that reason.
The derivation of
\Cref{eq:Mgeneral} is largely analogous to \cite{lobach2020PRAB} and is outlined
in Appendix~\ref{app:beamdiv}. Appendix~\ref{app:gaussianlightbeam} provides an illustrative closed-form expression for $M$, based on \Cref{eq:Mgeneral} in the
approximation of a Gaussian spectral-angular distribution of the radiation.
In IOTA, we study undulator radiation, because the quadratic term in
\Cref{eq:varN_from_book}, sensitive to bunch parameters, is larger for
undulators and wigglers than it is for dipole magnets \cite{lobach2020PRAB}. The complex field amplitude
$\Ef{\phiv}{}{s}$, generated by a single electron, can be numerically calculated by our computer code \cite{wigradrepo}, based on the equations from
\cite{clarke2004science}, or by using the SRW package
\cite{chubar2013wavefront}. Then, the integrals in \Cref{eq:Mgeneral,eq:NoneDef} can be
calculated by a Monte-Carlo algorithm. Our \CC~code with Python bindings for
calculation of \Cref{eq:Mgeneral,eq:NoneDef} is provided in the repository
\cite{furrepo}.
\section{\label{sec:apparatus}Apparatus}
In our experiment, a single electron bunch circulated in the IOTA ring, see
Fig.~\ref{fig:iota_detector_layout}(a),
\begin{figure}[!h]
\includegraphics[width=0.7\columnwidth]{iota_detector_layout.png}
\caption{\label{fig:iota_detector_layout} (a) Layout of IOTA. (b) Light path
from the undulator to the detector (not to scale).}
\end{figure}
with a revolution period of $\valrevolutionPeriod$ and the beam energy of $\valEbeamDimLess \pm \valEbeamerror$. We studied two transverse focusing
configurations in IOTA: (1) strongly coupled, resulting in approximately equal transverse mode emittances
and (2) uncoupled, resulting in two drastically different emittances.
Henceforth, we will refer to the beams in these configurations as ``round'' and ``flat'' beams, respectively. In both
cases, the bunch length and the emittances depend on the beam current due to
intra-beam scattering \cite{bjorken1982intrabeam, IBS_Nagaitsev}, beam interaction with its environment \cite{haissinski1973exact}, etc.
The longitudinal bunch density distribution $\rho(z)$ was measured and recorded by
a high-bandwidth wall-current monitor \cite{WGM}. It was not exactly Gaussian,
but this fact was properly accounted for by \Cref{eq:kappaDef} for $\szeff$, which
works for any longitudinal bunch shape. The IOTA rf cavity operated at
\SI{30}{MHz} (4th harmonic of the revolution frequency) with a voltage amplitude of about $\valVrf$.
The rms momentum
spread $\sp$ was calculated from the known rf voltage amplitude, the design ring parameters and the measured rms bunch length $\sz$. In our experiments, the relation was
\begin{equation}
\sp \approx \valszTospDimLess\times\sz[\mathrm{cm}].
\end{equation}
\noindent It is an approximate equation, because of the bunch-induced rf voltage (beam loading) and a
small deviation of $\rho(z)$ from the Gaussian shape. However, the effect of
$\sp$ in \Cref{eq:Mgeneral} in IOTA was almost negligible. Therefore, such
estimation was acceptable.
For the round beam, the IOTA transverse focusing
functions (4D Twiss functions) were chosen to produce approximately equal mode emittances at zero
beam current, $\e_1 \approx \e_1 \approx \SI{12}{nm}$ (rms, unnormalized).
It was empirically confirmed that they remained equal at all beam currents
with a few percent precision.
The expected zero-current emittances for a flat beam were $\ex
\approx \SI{50}{nm}$, $\ey \gtrsim \SI{0.33}{p m}$ (set by the quantum excitation in a perfectly uncoupled ring). The expected zero-current rms bunch length and the rms momentum spread for both round and flat beams were $\sz=\valszZeroCur$,
$\sp=\valspZeroCur$. In our experiment, the electron beam sizes were monitored
and recorded by visible synchrotron light image monitors (SLMs) \cite{kuklev2019synchrotron} in
seven dipole bend locations, at M1L-M4L and at M1R-M3R, see
Fig.~\ref{fig:iota_detector_layout}(a).
The smallest reliably resolvable emittance by the SLMs in our experiment configuration was about $\SI{20}{nm}$.
\begin{figure}[!h]
\includegraphics[width=1.0\columnwidth]{ex_ey_round_beam.png}
\caption{\label{fig:ex_ey_round_beam}
Mode emittances ($\e_1=\e_2=\e$) and bunch lengths (rms and
effective) of a round beam in IOTA as functions of beam current. SLMs had a
monitor-to-monitor variation of $\pm\valSLMroundErr$ (not shown) in the measurement of $\e$. }
\end{figure}
Figure~\ref{fig:ex_ey_round_beam} illustrates the bunch parameters of the round
beam as a function of current. Below, we will present measurements with a flat
beam at only one value of beam current, $\valIbeamND$, measured with a direct-current current transformer (DCCT). The parameters of the flat
beam at this current value were $\sz=\valszND$, $\szeff=\valszeffND$,
$\sp=\valdppND$. The horizontal emittance was $\ex=\valexND$, as measured by the
SLMs with a monitor-to-monitor variation of $\pm\valSLMflatErr$. The small vertical
emittance of the flat beam was unresolvable by the SLMs. However, in
Sec.~\ref{sec:discussion}, we will demonstrate how it can be reconstructed using the fluctuations measurements.
At the center of the undulator, in the uncoupled optics, the Twiss parameters were $\bx=\valbetax$,
$\by=\valbetay$, $\ax=\valalphax$, $\ay = \valalphay$, the horizontal dispersion
$\Dx=\valDx$, its derivative $\Dxp=\valdDxND$.
The strongly coupled optics was created
from the uncoupled optics by changing the current in one skew-quad located at a zero dispersion location.
The coupling parameter $u$ \cite{lebedev2010betatron} was about \SI{0.5}{} everywhere in the ring. Therefore, the following is correct for the coupled case 4D Twiss functions, $\beta_{1x}\approx\beta_{2x}$, $\beta_{1y}\approx\beta_{2y}$. Moreover, their sums, $\beta_{1x}+\beta_{2x}$, $\beta_{1y}+\beta_{2y}$, were approximately equal to the Twiss beta functions in the uncoupled case, $\beta_x$, $\beta_y$. Equation~\eqref{eq:Mgeneral} assumes an uncoupled optics. However, this specific strongly coupled optics used in IOTA can be approximated by the uncoupled optics with equal horizontal and vertical emittances $\ex=\ey=\e$. More specifically, what is used in the derivation of \Cref{eq:Mgeneral} (see Appendix~\ref{app:beamdiv}) is the 6D phase-space distribution of the electrons, \Cref{eq:psixdim}. This distribution, for the round beam, when calculated using the approximation of uncoupled optics with equal emittances $\ex=\ey=\e$, and the distribution, calculated using the exact 4D Twiss functions and equal mode emittances, $\e_1=\e_2=\e$, are almost indistinguishable. This property was intentionally included in the initial design of the coupled optics in IOTA.
The undulator strength parameter
(peak) was $\Ku=\valKu$ with the number of periods $\Nu=\valNundPer$ and the period length $\lamu=\SI{5.5}{cm}$, the total length of the undulator was $\Lu=\Nu\lamu=\SI{58}{cm}$. A photodetector was installed in a dark box above the
M4R dipole magnet, see Fig.~\ref{fig:iota_detector_layout}(b). The light produced in the undulator was directed to the dark box by a
system of two mirrors ($\diameter \SI{2}{\inch}$). Then, it was focused by a lens
($\diameter \SI{2}{\inch}$, focal distance $F=\SI{150}{mm}$) into a spot, smaller
than the sensitive area of the detector ($\diameter \SI{1.0}{mm}$). The lens was $\valzobs$ away from the center of the undulator. Because of
the two round mirrors, which are at $\SI{45}{\degree}$ to the direction of
propagation of the radiation, the angular aperture takes an elliptical shape
with the vertical axis smaller than the horizontal by a factor of $\sqrt{2}$.
Namely, the horizontal and the vertical semi-axes were $\valsemiApertureX$ and $\valsemiApertureY$, respectively. The measurements were performed in the vicinity of the fundamental of the undulator
radiation, $\lambda_1=\lamu(1+\Ku^2/2)/(2\gamma^2)=\vallambdaone$.
As a photodetector we used an InGaAs PIN
photodiode \cite{hamamatsu_photodiode}, which has a high quantum efficiency ($\approx\SI{80}{\percent}$) around the fundamental.
Using the elliptical angular aperture mentioned above and the manufacturers'
specifications for the spectral transmission of the vacuum chamber window at the
M4R dipole magnet, the two mirrors, the focusing lens, and the quantum
efficiency of the InGaAs photodiode, we constructed the detection efficiency
function $\etaf{\phiv}$ for our system. The lens's spectral transmission had to
be linearly extrapolated for a small interval outside of the range provided in the
manufacturer's specifications. There were no free adjustable parameters. We
calculated the field amplitude $\Ef{\phiv}{}{s}$, generated by a single
electron, for the parameters of our undulator on a 3D grid ($k$, $\phi_x$,
$\phi_y$) with our code \cite{wigradrepo}.
Figure~\ref{fig:spectral_angular_distribution}(a) shows the simulated spectrum,
where the intensity is integrated over the elliptical aperture,
\begin{equation}
\dv{\None}{k}=\sum\limits_{s=1,2}\int \dd[2]{\phiv}\etaf{\phiv}\abs{\Ef{\phiv}{}{s}}^2.
\end{equation}
\begin{figure}[!h]
\includegraphics[width=1.0\columnwidth]{spectral_angular_distribution.png}
\caption{\label{fig:spectral_angular_distribution}
(a) Spectral distribution of the average number of detected photons per turn for a single electron (s.e.)
assuming no losses (blue) and accounting for the detection efficiency of the
system (red). Also, the detection efficiency (dashed, right vertical scale). (b) Angular distribution of the number of detected photons
accounting for the detection efficiency of our system. Both (a) and (b) are calculated for
an elliptical aperture with the horizontal and the vertical semi-axes
$\valsemiApertureX$ and $\valsemiApertureY$, respectively.}
\end{figure}
\begin{figure*}[t]
\includegraphics[width=\textwidth]{full_circuit.png}
\caption{\label{fig:integrator_and_comb_filter} (a) Photocurrent integrator
circuit. (b) Comb filter. (c) Sample waveforms of hybrid outputs (\D- and
\S-channels).}
\end{figure*}
The blue line is calculated for an ideal detection system, where
$\etaf{\phiv}=1$ inside the elliptical aperture, and zero outside. The red line
is calculated with $\etaf{\phiv}$, constructed using the manufacturers'
specifications of the optical elements in our system. This $\etaf{\phiv}$ is equal to zero outside of the elliptical aperture. Whereas, inside, it is equal to the detection efficiency of our system. In our case, the detection efficiencies for the horizontal and vertical polarizations ($s=1,2$) are practically the same. Only the reflectance of the mirrors is slightly polarization dependent (under \SI{1}{\percent} difference). Moreover, the radiation is dominated by the horizontal polarization (about \SI{96.5}{\percent}). The dashed line in Fig.~\ref{fig:spectral_angular_distribution}(a) is the detection efficiency of our system for the horizontal polarization.
Figure~\ref{fig:spectral_angular_distribution}(b) shows the angular distribution
with $\etaf{\phiv}$ of our system,
\begin{equation}
\frac{\dd[2]{\None}}{\dd\phi_x\dd\phi_y} = \sum\limits_{s=1,2}\int \dd{k}\etaf{\phiv}\abs{\Ef{\phiv}{}{s}}^2.
\end{equation}
With the spectral properties of all optical elements in the system taken into account, the spectral width of the radiation was $\valfwhm$ (FWHM), and the angular size was $\approx\SI{2}{mrad}$, which could be fully transmitted through the $\diameter \SI{2}{\inch}$ optical system.
Figure~\ref{fig:integrator_and_comb_filter} illustrates our full photodetector
circuit. First, the radiation pulse is converted into a photocurrent pulse by
the photodiode, see Fig.~\ref{fig:integrator_and_comb_filter}(a). Then, the
photocurrent pulse is integrated by an op-amp-based RC integrator, which outputs
a longer pulse with a voltage amplitude that can be easily measured. The op-amp \cite{opamp} was capable of driving the 50-$\Omega$ input load
of a fast digitizing scope, located $\approx\SI{100}{m}$ away. The resistor
$R_0=\SI{580}{\kilo\ohm}$ in the circuit in
Fig.~\ref{fig:integrator_and_comb_filter}(a) was used to remove the offset in
the integrator output signal (about $\SI{0.3}{V}$), produced by the op-amp input
bias current and the photodiode leakage current. The output voltage pulse of the
integrator at the $i$th IOTA turn can be represented as $A_i f(t)$, where $A_i$ is
the signal amplitude at the $i$th turn and $f(t)$ is the average signal for one
turn, normalized so that its maximum value is 1, see
Fig.~\ref{fig:integrator_and_comb_filter}(a). The time $t$ in $f(t)$ is in the range $\SIrange{0}{133.3}{ns}$, i.e., within one IOTA revolution. The number of photoelectrons,
generated by the light pulse at the $i$th turn, $\N_i$, can be calculated as the
time integral of the output pulse of the integrator divided by the electron
charge $e$ and the resistance $\Rf$, i.e.,
\begin{equation}\label{eq:NifromAiInt}
\N_i=\int A_if(t)\dd{t}/(e\Rf).
\end{equation}
\noindent The function $f(t)$ is known --- it was measured with a fast
oscilloscope. It was practically the same during all of our measurements,
because $f(t)$ is rather wide (about \SI{30}{ns} FWHM) and the length of the
input light pulses was much smaller (about \SI{2}{ns} FWHM); moreover, the shape of input pulses did
not change significantly. Therefore, during all of our measurements
\begin{equation}\label{eq:AtoN}
\N_i = \chi A_i,
\end{equation}
\noindent where
\begin{equation}
\chi = \int f(t)\dd{t}/(e\Rf) = \valAmpToPhotoel,
\end{equation}
\noindent
with a $\pm\SI{5}{\percent}$ uncertainty, because of the uncertainty on $\Rf$.
We verified \Cref{eq:AtoN} empirically at different voltage amplitudes $A_i$ and
different bunch lengths, which define the lengths of the input light pulses. During our experiments at different beam currents, $A_i$ was in the range between $\SI{0}{V}$ and $\SI{1.2}{V}$.
Since we also knew the empirical linear relation between the beam current and
the integrator voltage amplitude, we could use it in \Cref{eq:AtoN} to find the
average number of detected photons (photoelectrons) per one electron of the
electron bunch. The result of this calculation was $\valphotonFluxMeas$. This value is quite close to the result
obtained in our simulation,
\begin{multline}
\sum\limits_{s=1,2}\int \dd{k}\dd[2]{\phiv}\etaf{\phiv}\abs{\Ef{\phiv}{}{s}}^2\\
=\valphotonFluxTheor.
\end{multline}
In our experiment, the expected relative fluctuation of $A_i$ was \SI{e-4}{}--\SI{e-3}{} (rms), which is considerably lower than the
digitization resolution of our 8-bit broad-band oscilloscope. To overcome
this problem, we used a passive comb (notch) filter \cite{smith2010physical}, shown in
Fig.~\ref{fig:integrator_and_comb_filter}(b). The signal splitter divides the
integrator output into two identical signals. The lengths and the
characteristics of the cables in the two arms were chosen such that one of the
signals was delayed by exactly one IOTA revolution and, at the same time, the
losses and dispersion in both arms were approximately equal. The time delay in the comb filter could
be adjusted with a $\SI{0.1}{ns}$ precision. Therefore, the time delay error was
negligible, because the pulses at the entrance of the comb filter were about $\SI{30}{ns}$ long (FWHM).
Finally, a passive hybrid \cite{hybrid} generated the difference and the sum
of the signals in the two arms --- its output channels \D~and \S, respectively.
For an ideal comb filter,
\begin{align}
\Dit{} = \xi(A_i-A_{i\shortminus 1})\ft{},\label{eq:Deq}\\
\Sit{} = \xi(A_i+A_{i\shortminus 1})\ft{}\label{eq:Seq},
\end{align}
\noindent where we assume that the pulse shape of input and output
signals of the comb filter is the same --- $f(t)$. This means that we assume a negligible
dispersion in the comb filter, which is a very good approximation according to
our comparison of input and output pulses with the oscilloscope. Also, as a
result of this comparison, we determined the parameter $\xi=\valxi$. Of course,
our comb filter was not perfect. There was some cross-talk between \D- and
\S-channels, some noise in the signals, a small undesirable reflection in one of
the arms, resulting in a small satellite pulse about $\valtr$ away from the main
pulse, see Fig.~\ref{fig:integrator_and_comb_filter}(c). In addition, the hybrid was AC-coupled.
With these effects taken into account \Cref{eq:Deq,eq:Seq} take the form
\begin{multline}
\Dit{} = \xi(A_i-A_{i\shortminus 1})\ft{}+\muD \Sit{}
\\
+\delta_rA_if(t-t_r)
+\nuDit
-\DAC\label{eq:Dit},
\end{multline}
\begin{multline}
\Sit{} = \xi(A_i+A_{i\shortminus 1})\ft{}+\muS \Dit{}
\\
+\delta_r A_if(t-t_r)
+\nuSit
-\SAC\label{eq:Sit},
\end{multline}
\noindent where $t$ is within one IOTA turn (\SIrange{0}{133.3}{ns}), $\muD$ and $\muS$ describe the cross-talk between \D- and
\S-channels ($<\SI{1}{\percent}$), $\delta_r A_if(t-t_r)$ describes the reflected pulse in one of the arms (perhaps the short one), $t_r =
\valtr$, $\delta_r\approx\SI{1.5e-3}{}$;
and it is assumed that the noise contributions $\nuDit$ and $\nuSit$ enter the
equations as sum terms, independent of the signal amplitude; the constants
$\DAC$ and $\SAC$ come from the fact that the hybrid is AC-coupled and the
averages of $\Dit{}$ and $\Sit{}$ over a long time have to be zero.
For each measurement, we recorded $\valwfLen$-long waveforms (about
$\nrev=\valNrev$ IOTA revolutions) of \D- and \S-channels with the oscilloscope
at $\SI{20}{GSa/s}$. The beam current decay was negligible during this $\valwfLen$ acquisition period.
In \Cref{eq:Sit}, the noise, the cross-talk term, and the reflection term are
negligible. The \S-channel can be used to measure the photoelectron count mean $\av{\N}$
during the $\valwfLen$. Using \Cref{eq:AtoN} and the non-negligible part of
\Cref{eq:Sit},
\begin{equation}\label{eq:avNfromSigma}
\av{\N} = \chi\frac{\av{\Sigma(\tpeak)}+\SAC}{2\xi},
\end{equation}
\noindent where we introduced $\tpeak$ --- the time within each turn,
corresponding to the peak of the signal, $f(\tpeak)=1$,
\begin{equation}
\av{\Sigma(\tpeak)} = \frac{1}{\nrev}\sum\limits_{i=1}^{\nrev}\Sigma_i(\tpeak).
\end{equation}
The idea of using a comb filter is that, in the ideal case, see
\Cref{eq:Deq}, the \D-channel would provide the exact difference between two
consecutive turns in IOTA. In this case we would be able to look directly at the
turn-to-turn fluctuations. The offset would be removed, and the oscilloscope
could be used with the appropriate scale setting, with negligible digitization
noise. In our non-ideal comb filter, see \Cref{eq:Dit}, the additional terms
have some impact on the \D-signal, see Fig.~\ref{fig:integrator_and_comb_filter}(c).
Nonetheless,
by analyzing the \D-signal in a special way, described below, it is possible to determine $\var{\N}$ with sufficient precision.
Namely, if we take the variance of \Cref{eq:Dit} with respect to $i$ at a fixed time $t$, then we obtain
\begin{equation}\label{eq:varDt}
\var{\Dt} = 2\xi^2\var{A}\ft{2}+\var{\nuD(t)},
\end{equation}
\noindent where the contribution from $\muD \Sit{}$ and $\delta_rA_if(t)$ may
be dropped, because the fluctuations of $\Sit{}$ and $A_i$ are strongly
attenuated by the factors $\muD$ and $\delta_r$, respectively. Also,
$\var{\DAC}=0$ since $\DAC$ is constant during the $\valwfLen$. The left-hand side of
\Cref{eq:varDt}, as a function of $t\in[0,\valrevolutionPeriodDimLess]\SI{}{ns}$, could be
obtained from the collected waveforms of \D-channel as
\begin{equation}\label{eq:varDtfromWF}
\var{\Dt}
= \frac{1}{\nrev}\sum\limits_{i=1}^{\nrev}\Dit{2}
- \bigl[\frac{1}{\nrev}\sum\limits_{i=1}^{\nrev}\Dit{}\bigr]^2.
\end{equation}
The results of such calculation for 2000 moments of time $t$ within an IOTA
revolution are shown in Fig.~\ref{fig:noise_filtering}. These data are for the round beam. The blue, orange, and green lines
correspond to three significantly different values of beam current within the
range studied in our experiment; the red line corresponds to a zero beam current case.
\begin{figure}[!h]
\includegraphics[width=1.0\columnwidth]{noise_filtering.png}
\caption{\label{fig:noise_filtering} The variance of \D-signal as a function of
time (see \Cref{eq:varDt}) within one IOTA revolution (round-beam data).}
\end{figure}
Figure~\ref{fig:noise_filtering} suggests that there is a constant noise level,
independent of time and independent of the signal amplitude. Specifically, it
suggests that the noise term in \Cref{eq:varDt} is
\begin{equation}
\var{\nuD(t)} = \var{\nuD} = \SI{8.8e-8}{V}^2.
\end{equation}
The observed rms noise amplitude ($\approx\SI{0.3}{mV}$) was analyzed by using the
noise model for the detector electrical schematic,
Figs.~\ref{fig:integrator_and_comb_filter}(a) and (b), as well as the typical
electrical characteristics of the photodiode~\cite{hamamatsu_photodiode} and the
operational amplifier~\cite{opamp}. The three main contributions to the rms
noise in the \D-channel are the following: the oscilloscope input amplifier
noise, \SI{0.21}{mV}; the operational amplifier input voltage noise,
\SI{0.18}{mV}; and the operational amplifier input current noise,
\SI{0.037}{mV}. When added in quadrature, these three sources explain the measured noise level.
The peaks rising above the noise level in Fig.~\ref{fig:noise_filtering} can be
fitted well with $f^2(t)$ (fits not shown). Thus, their shape is in agreement
with \Cref{eq:varDt} as well.
Therefore, using \Cref{eq:AtoN,eq:varDt}, the photoelectron count variance
$\var{\N}$ can be determined as
\begin{equation}\label{eq:varNfromDelta}
\var{\N} = \chi^2\var{A}
= \chi^2\frac{\var{\Delta(\tpeak)}-\var{\nuD}}{2\xi^2},
\end{equation}
\noindent see \Cref{eq:varDtfromWF} for the definition of $\var{\Delta(\tpeak)}$.
The value of the noise level term in \Cref{eq:varNfromDelta} is
\begin{equation}
\frac{\chi^2 \var{\nuD}}{2\xi^2} = \valnoiseLevel.
\end{equation}
We employed a dedicated test light source with known fluctuations to verify this method of measurement of $\av{\N}$ and $\var{\N}$ [\Cref{eq:avNfromSigma,eq:varNfromDelta}]. This verification is described in
Appendix~\ref{app:testlightsource}, where we also estimated the statistical
error of the measurement of $\var{\N}$ by our apparatus, namely,
$\pm\valvarNerror$ --- it is approximately constant in the range of $\var{\N}$
observed with the undulator radiation in IOTA.
\section{Measurement results}
\begin{figure*}[t]
\includegraphics[width=\textwidth]{round_beam_compare_with_no_beam_div.png}
\caption{\label{fig:round_beam_compare_with_no_beam_div}
(a) The fluctuations measurement for the round electron beam in IOTA as a
function of beam current, a prediction by \Cref{eq:Mgeneral} (red solid
line) and a prediction by \MnoDiv~(black dashed line), which does not
account for the beam divergence. (b) The fluctuations measurement for the flat
electron beam at a fixed beam current $\valIbeamND$ with 4 different
optical neutral density filters and one point without any filters, as well as a parabolic fit (a prediction
could not be made, because the vertical emittance of the flat beam was unknown). (c) and (d) present the data of (a) and (b) in terms of the number of coherent modes $M$.}
\end{figure*}
The measured fluctuations data for the round beam at different values of beam current are shown in Fig.~\ref{fig:round_beam_compare_with_no_beam_div}(a)
(blue points).
The blue dashed straight line, $\var{\N}=\av{\N}$, represents the
photon shot noise contribution to the fluctuations --- the first sum term in
\Cref{eq:varN_from_book}. The values of $M$ extracted from the fluctuation data
points using the equation,
\begin{equation}\label{eq:MfromvarN}
M = \av{\N}^2/(\var{\N}-\av{\N}),
\end{equation}
\noindent are shown in Fig.~\ref{fig:round_beam_compare_with_no_beam_div}(c)
(blue points). The error bars in Figs.~\ref{fig:round_beam_compare_with_no_beam_div}(a),(c) correspond to the $\pm\valvarNerror$ statistical error of measurement of $\var{\N}$ by our technique.
Further, Fig.~\ref{fig:round_beam_compare_with_no_beam_div}(c) has a
curve for $M$, simulated by \Cref{eq:Mgeneral} (red line), and, for comparison,
a curve for $M$, simulated by \MnoDiv~(dashed black line), which neglects beam
divergence. Corresponding curves for simulated $\var{\N}$ are shown in
Fig.~\ref{fig:round_beam_compare_with_no_beam_div}(a). The shaded light red
areas in Figs.~\ref{fig:round_beam_compare_with_no_beam_div}(a),(c) show the
uncertainty range of our simulation by \Cref{eq:Mgeneral}.
For this simulation, we needed the values of the following four bunch
parameters, entering \Cref{eq:Mgeneral}, $\ex$, $\ey$, $\sp$, $\szeff$, at all
beam currents. Further, we needed the values of Twiss functions in the
undulator, the parameters of the undulator and of the detection system. All
these aspects were described in Sec.~\ref{sec:apparatus}. We had no free
parameters in this simulation. Numerical calculation of the integrals in
\Cref{eq:Mgeneral} and \MnoDiv~was performed by the Monte-Carlo algorithm on the
Midway2 cluster at the University of Chicago Research Computing Center using our
computer code \cite{furrepo,wigradrepo}.
The simulation uncertainty range (shaded light red area) primarily comes from
the uncertainty in the beam energy $\valEbeamDimLess\pm\valEbeamerror$. The next
source of uncertainty by magnitude, which is a factor of two smaller, is the SLMs' $\pm\valSLMroundErr$
monitor-to-monitor variation of $\e$ [for the round beam we use $\ex=\e$ and $\ey=\e$ in \Cref{eq:Mgeneral}]. The uncertainties of other parameters ($\sp$, $\szeff$, Twiss functions
in the undulator, etc.) had negligible effect. The manufacturers' specifications
for the optical elements of our system did not provide any uncertainties.
Therefore, they were not considered.
We also collected fluctuations data for another experiment configuration.
Figure~\ref{fig:round_beam_compare_with_no_beam_div}(b) shows fluctuation data
points for a flat beam (uncoupled focusing) at a fixed beam current $\valIbeamND$.
The corresponding reconstructed values of $M$ are shown in
Fig.~\ref{fig:round_beam_compare_with_no_beam_div}(d).
The error bars in Figs.~\ref{fig:round_beam_compare_with_no_beam_div}(b),(d) correspond to the $\pm\valvarNerror$ statistical error of measured $\var{\N}$.
In this measurement, the
photoelectron count mean (horizontal axis) was varied by using different optical
neutral density filters (one point without a filter and four points with
filters). Neutral density filters are filters that have constant attenuation in
a certain wavelength range; in our case, around the fundamental harmonic of the
undulator radiation. A new bunch was injected into the ring for each
measurement. The oscilloscope waveforms for \D- and \S-channels were recorded
when the beam current decayed to $\valIbeamND$. The red curve in
Fig.~\ref{fig:round_beam_compare_with_no_beam_div}(b) is a fit with a constant
$M$. A corresponding horizontal line is shown in
Fig.~\ref{fig:round_beam_compare_with_no_beam_div}(d). The value of $M$ in this
fit is $\Mfit=(4.38\pm 0.10)\times 10^6$. This value was calculated as the
average of the five values of $M$ in
Fig.~\ref{fig:round_beam_compare_with_no_beam_div}(d), and the error was
calculated as the standard deviation of these five values.
We do not present any simulation results for the fluctuations $\var{\N}$ in the uncoupled focusing, because
the SLMs provided very inconsistent estimates for the small vertical
emittance $\ey$ of the flat beam --- the max-to-min variation for different SLMs reached a factor of eight.
We believe this happened because the beam images were close to the resolution limit, set by a combination of factors, such as the diffraction limit, the point spread function of the cameras, chromatic aberrations, the effective radiator size of the dipole magnet radiation ($\approx\SI{20}{\micro m}$), and the camera pixel size ($\approx\SI{10}{\micro m}$ in terms of beam size).
Therefore, the monitor-to-monitor emittance value variation primarily came from the Twiss beta-function variation ($\beta_y^{(\mathrm{max})}/\beta_y^{(\mathrm{min})} \approx 12$).
The diffraction limit is primarily caused by the irises, used to reduce the radiation intensity to prevent the cameras from saturating at high beam currents. Alternatively, leaving the irises open and using attenuating optical filters may improve the resolution.
Additional negative effects include the errors in the light focusing optics, calibration errors of the SLMs, and possible Twiss beta-function errors.
The SLMs at locations with larger beta-functions (M4L, M1L) provide estimates for $\e_y$ that agree better with the theoretical predictions \cite{lebedev2020ibs} at lower beam currents, and with the emittance estimates presented below in this paper. Nevertheless, we cannot state that the SLMs could provide a reliable estimate for $\e_y$ during our experiment.
Without $\ey$ we could not use \Cref{eq:Mgeneral} to make a
prediction for $M$ and $\var{\N}$.
However, we attempted the reverse of this procedure, namely, the reconstruction of the
unknown $\ey$ via the measured fluctuations for the flat beam shown in
Figs.~\ref{fig:round_beam_compare_with_no_beam_div}(b),(d). Indeed, the measured
value of the number of coherent modes $\Mmeas$ is a function of four bunch
parameters,
\begin{equation}\label{eq:eyViaMmeas}
\Mmeas = M(\ex,\ey,\sp,\szeff).
\end{equation}
\noindent The full form of the right-hand side is given by \Cref{eq:Mgeneral}.
The horizontal emittance $\ex$ of the flat beam at a beam current of
$\valIbeamND$ could still be reliably measured via the SLMs, yielding $\e_x=\valexND$. The
effective bunch length $\szeff$ could be determined from $\rho(z)$
measured by the wall-current monitor, $\szeff=\valszeffND$. The rms momentum
spread $\sp$ was estimated from $\rho(z)$ and the ring parameters,
$\sp=\valdppND$. The only unknown in \Cref{eq:eyViaMmeas} is $\ey$.
Equation~\eqref{eq:eyViaMmeas} can be solved for $\ey$ by a simple bisection
method. The result is $\ey = \valeyWithDivDimLess \pm \valeyWithDivErr$, where
the uncertainty corresponds to the statistical error of $\Mfit$, mentioned
above. For comparison, we also used \MnoDiv~ in \Cref{eq:eyViaMmeas}, which
neglects beam divergence. In this case, we obtained $\ey =
\valeyWithoutDivDimLess \pm \valeyWithoutDivErr$.
In this reconstruction of $\ey$, there is also a
systematic error due to the uncertainty on the beam energy ($\pm\valEbeamerror$)
and due to the systematic error of $\ex$ measurement by the SLMs ($\pm\valSLMflatErr$
monitor-to-monitor). We estimated these two contributions to the systematic
error of $\ey$ from \Cref{eq:Mgeneral}. They are
$\substack{+2.5 \\ -4.5}\;\SI{}{nm}$ and $\substack{+1.6 \\ -1.4}\;\SI{}{nm}$,
respectively. These systematic errors are rather significant. However, they are
not directly linked to our measurement technique. They are related to the fact
that the beam energy and the horizontal emittance of a flat beam in IOTA were not
known with better precision. Further improvements in beam
characterization in IOTA will reduce the systematic error of our
fluctuations-based technique of $\ey$ measurement.
\section{\label{sec:discussion}Discussion and conclusions}
Power fluctuations in undulator radiation were measured in IOTA under two different experimental conditions and compared with our theoretical predictions.
The fluctuations predicted by \Cref{eq:varN_from_book,eq:Mgeneral} in the round beam configuration agree with the measurements within the uncertainties, as shown in Figs.~\ref{fig:round_beam_compare_with_no_beam_div}(a),(c).
In IOTA, the bunch parameters $\ex$, $\ey$, $\sp$,
$\szeff$ depend on the beam current because of various intensity dependent
effects, e.g., intra-beam scattering \cite{reiser1994theory}, beam interaction
with its environment \cite{haissinski1973exact}, etc. Therefore, $M$ is a
function of the beam current as well, as one can see in
Fig.~\ref{fig:round_beam_compare_with_no_beam_div}(c).
In Figs.~\ref{fig:round_beam_compare_with_no_beam_div}(b),(d), all data points
correspond to one value of the beam current, $\valIbeamND$. The photoelectron count
mean is varied by using neutral density filters with different attenuation
factors \newcommand{\etaND}{\eta_{\mathrm{ND}}} $\etaND$. Such filters linearly
scale down the photoelectron count mean, $\av{\N}\rightarrow\etaND\av{\N}$.
However, they do not change $M$, because if $\etaf{\phiv}$ is replaced by
$\etaND\etaf{\phiv}$ in \Cref{eq:Mgeneral}, then $\etaND$ cancels out in the
numerator and the denominator. This is consistent with
Fig.~\ref{fig:round_beam_compare_with_no_beam_div}(d) --- all measured
values of $M$ are equal within the uncertainty range.
We reconstructed the value
of the vertical emittance $\ey$ of the flat beam from $\Mfit$ via \Cref{eq:Mgeneral}
and via \MnoDiv, which neglects beam divergence. We obtained very different
results, $\ey = \valeyWithDivDimLess \pm \valeyWithDivErr$ and $\ey =
\valeyWithoutDivDimLess \pm \valeyWithoutDivErr$, respectively. This shows that,
for the flat beam, accounting for the beam divergence is critical. In this
measurement, the horizontal beam divergence was $\valbeamDivXND$ and comparable with the rms radiation divergence $\sqrt{\lambda_1/(2\Lu)}=\valoneOverGammaNu$ \cite[Eq.~(2.57)]{kim2017synchrotron},
which gives an estimate of the angular size of $\Ef{\phiv}{}{s}$. Clearly, in
this case it has a significant effect on the integral in the numerator of
\Cref{eq:Mgeneral}. This is why we use \Cref{eq:Mgeneral} in our Letter
\cite{lobach2020furjointprl}, focused on emittance measurements via fluctuations,
as opposed to \MnoDiv, which is simpler, but neglects beam divergence.
In addition, we made an independent estimate of the vertical emittance $\ey$ of
the flat beam based on the beam lifetime, see Appendix~\ref{app:touschek} for
details. At the beam current of $\valIbeamND$, the beam lifetime is solely
determined by Touschek scattering \cite{lebedev2020ibs,lebedev2010betatron,touschek}. The Touschek lifetime is a function of beam
emittances and bunch length. Therefore, since we knew the horizontal emittance $\ex$ of the flat
beam, measured by the SLMs, the bunch length, measured by the wall-current monitor, and the measured beam lifetime
$\abs{I/(\dv*{I}{t})}$, we could find $\ey$. The result is $\ey=\valeyLifetimeRecNDDimLess\pm\valeyLifetimeRecNDErr$, to be compared with the fluctuations-based
measurement, $\ey = \valeyWithDivDimLess \pm \valeyWithDivErr$.
The $\pm\valeyLifetimeRecNDErr$ error in the lifetime-based $\ey$ estimate comes from the $\pm\valSLMflatErr$ uncertainty on $\ex$ of the flat beam.
In the round beam case, at the same beam current of $\valIbeamND$, the beam divergence in the
undulator was about $\valbeamDivRoundX$ (both $x$ and $y$). It was noticeably smaller than the
horizontal beam divergence of the flat beam and than
the rms radiation divergence $\valoneOverGammaNu$. Therefore, the effect of
beam divergence on the fluctuations simulation in
Figs.~\ref{fig:round_beam_compare_with_no_beam_div}(a),(c) is not as dramatic.
However, the deviation from the measurement of the simulation based on \MnoDiv, which neglects beam divergence, is certainly noticeable, whereas the simulation by
\Cref{eq:Mgeneral} agrees well with the measurement.
It would be beneficial to repeat these fluctuation measurements with a
longer and brighter undulator.
In this experiment, we had to avoid using a
monochromator or restricting the angular aperture, because we had to collect all available radiation to
achieve a signal with a voltage amplitude that could be easily measured. Therefore, the integrals in \Cref{eq:Mgeneral,eq:NoneDef} had to be
calculated over a broad range of angles and wavelengths. With a
monochromator, a slit or a pinhole detector, these integrals could be significantly simplified. Further, if our undulator had more periods and if we were able to use a monochromator, we could slowly vary the beam energy and find the energy at which the detected power is at maximum, i.e., when we are
centered on the peak of the fundamental harmonic. In this case, the systematic
error of the $\ey$ measurement via the fluctuations, related to the uncertainty on the beam energy, would
be negligible. Moreover, it can be shown \cite{lobach2020furjointprl} that if one places a narrow vertical slit in front of the detector, then the magnitude of the fluctuations would only depend on $\ey$, i.e., the systematic error of
the $\ey$ measurement related to the
uncertainty of $\ex$, measured by the SLMs, would be minimized, too. Finally, for a
brighter undulator, the statistical error on the measured value of the fluctuations
would be lower as well.
We are considering using our fluctuations measurement apparatus, or an improved variation of it, as
a tool for the diagnostics of the Optical Stochastic Cooling (OSC) experiment in IOTA
\cite{lebedev2015optical,jarvis2018optical,andorf2018computation}. In this experiment, the beam emittance will be even smaller and the existing diagnostic tools may not have sufficient resolution. Even though we cannot measure
transverse emittances and bunch lengths individually in this way, the
fluctuations may serve as an indicator of the cooling process. When the
cooling process starts, the electron bunch shrinks, which, in turn, makes the
fluctuations of the number of detected photons increase.
To conclude, we presented a calculation [\Cref{eq:varN_from_book,eq:Mgeneral}], which describes the turn-to-turn fluctuations in the number
of detected synchrotron radiation photons, produced by an electron bunch with a Gaussian transverse density distribution, an arbitrary longitudinal
density distribution, and non-negligible rms beam divergences.
The rms bunch
length is assumed to be significantly larger than the radiation wavelength. Equation~\eqref{eq:Mgeneral} is
presented for the first time, as beam divergence has been neglected in all previous considerations \cite{sannibale2009absolute,lobach2020PRAB}. Beam divergence can
be neglected if it is significantly smaller than the characteristic radiation
angle --- $\sqrt{\lambda_1/(2\Lu)}$ for the fundamental of undulator radiation. We
presented the results of an experiment with a round beam in IOTA, where beam
divergence had an impact on the fluctuations of the undulator radiation. We
showed that our new \Cref{eq:Mgeneral} agrees better with the measurements than
\MnoDiv, which neglects beam divergence. Finally, we proposed a non-invasive
technique to measure the small unknown vertical emittance of a flat beam via the fluctuations. This new technique is described in detail in a separate publication \cite{lobach2020furjointprl}.
\begin{acknowledgments}
We would like to thank the entire FAST/IOTA team at Fermilab for helping us with
building and installing the experimental setup and taking data, especially Mark
Obrycki, James Santucci, and Wayne Johnson. Greg Saewert constructed the detection circuit and provided the test light source.
Brian Fellenz, Daniil Frolov, David Johnson, and Todd Johnson provided equipment and assisted during our
detector tests.
This work was completed in part with resources provided by the University of Chicago Research Computing Center. This research is supported by the University of Chicago and the US Department of Energy under contracts DE-AC02-76SF00515 and
DE-AC02-06CH11357. This manuscript has been authored by Fermi Research Alliance,
LLC under Contract No. DE-AC02-07CH11359 with the U.S. Department of Energy,
Office of Science, Office of High Energy Physics.
\end{acknowledgments}
|
\section{Introduction}
The era of precision observation of the neutron star, and the multi-messenger astronomy and astrophysics has begun.
One of the most fundamental questions in the neutron star physics is what is the correct equation of state (EoS) of
nuclear matter at densities below and above the saturation density ($\rho_0$).
Equation of state is determined by the energy of a particle and pressure which is gradient of energy with respect to density.
Basic properties of the neutron star such as mass, size, density profile, tidal deformability, species of particles that
consist of the neutron star matter, and their fractions are directly correlated to the EoS.
Neutron star cooling is critically dependent on the fraction of particles, so the EoS also plays an essential role in the thermal
evolution of neutron stars.
Energy of a particle in nuclear matter is conventionally expanded around the saturation density in terms of $x = (\rho - \rho_0)/3 \rho_0$ as
\begin{eqnarray}
E(\rho,\, \delta) &=& E(\rho) + S(\rho) \delta^2 + O(\delta^4), \\
E(\rho) &=& E_B + \frac{1}{2} K_0 x^2 + O(x^3), \label{eq:sat} \\
S(\rho) &=& J + L x + \frac{1}{2} K_{\rm sym} x^2 + \frac{1}{6} Q_{\rm sym} x^3 + O(x^4). \label{eq:sym}
\end{eqnarray}
Neutron-proton asymmetry is accounted by $\delta = (\rho_n - \rho_p)/\rho$ where $\rho_n$ and $\rho_p$ are the
density of the neutron and the proton, respectively.
Most of the nuclear structure models fit their model parameters to the nuclear data, and
the constants and coefficients in Eqs.(\ref{eq:sat}, \ref{eq:sym}) which determine the EoS of nuclear matter are
obtained as results of the fitting.
Among the constants and coefficients in Eqs.(\ref{eq:sat}, \ref{eq:sym}),
saturation density $\rho_0$ and the binding energy $E_B$ are determined highly model-independently,
and the values are converged to $\rho_0=0.16$ fm$^{-3}$ and $E_B = 16$ MeV.
On the other hand, compression modulus $K_0$ that characterizes the density dependence of EoS in the symmetric nuclear matter
is uncertain more than $\rho_0$ and $E_B$.
For example, collection of 240 Skyrme force models shows that the models predict the $K_0$ value in the range $200-400$ MeV \cite{dutra}.
Experimentally constrained range is reduced to $200-260$ MeV, but it is still uncertain much more than $\rho_0$ and $E_B$.
Since the average density of a canonical neutron star, a star of mass $1.4 M_\odot$ is more than $2\rho_0$,
role of $K_0$ in the neutron star matter might be non-negligible and therefore it is needed to reduce the uncertainty of $K_0$
in order to have more exact EoS of the neutron star matter.
Situation is no better than $K_0$ for the symmetry energy $S(\rho)$.
Ranges of $J$, and $L$ constrained from experiments are $30-35$ MeV and $40-76$ MeV, respectively \cite{dutra}.
For $K_{\rm sym}$ experimental access is still very limited, and the value is guessed from relations with $L$ or $3J-L$.
Properties of neutron rich systems ($\delta \simeq 1$) are sensitive to the behavior of the symmetry energy.
$J$, $L$ and $K_{\rm sym}$ being siginificantly uncertain, large uncertainty in the EoS of neutron stars is unavoidable.
Goal of this work is to reduce the uncertainty of $K_0$, $J$ and $L$ so that we have more reliable EoS for the neutron star
matter at densities up to $3\rho_0$ or $4\rho_0$ below which nucleons may be the dominant constituents of the matter.
It has been shown that if the energy density of nuclear matter is expanded in powers of the Fermi momentum of the nucleon,
7 terms are sufficient for the correct description of EoS at densities relevant to nuclei and the center of neutron stars
\cite{kidsnm, kidsnuclei1, kidsnuclei2}.
If higher order terms are added to the basic 7 terms, additional terms give vanishing contribution up to $\rho \simeq 3\rho_0$,
and their effects are marginal at densities larger than $3\rho_0$.
Following this observation, we describe the EoS of infinite nuclear matter with 7 density-dependent terms.
Seven parameters in the functional are fixed to the values of $\rho_0$, $E_B$, $K_0$, $J$, $L$, $K_{\rm sym}$, and $Q_{\rm sym}$.
After the 7 nuclear matter parameters determined, the model is applied to nuclei.
Energy per particle ($E/A$) and the charge radii ($R_c$) of spherical magic nuclei $^{40}$Ca, $^{48}$Ca and $^{208}$Pb
are reproduced by adjusting two additional parameters that account for the gradient of density and spin-orbit interactions \cite{polb}.
At this stage the number of model parameters is increased to 9, but we don't need to introduce additional parameters any more.
By varying $K_0$, $J$, $L$ and $K_{\rm sym}$ values, $\chi^2$ values are calculated for $E/A$ and $R_c$ of $^{40}$Ca, $^{48}$Ca and $^{208}$Pb.
We find that nuclear data are most accurately reproduced in the range $K_0 = 230-250$ MeV.
In the application to the neutron star, we select two sets of $(J,\, L,\, K_{\rm sym})$ that give the smallest $\chi^2$ values
for a given $K_0$ value.
Solving TOV equations, we obtain the mass and radius of the neutron star.
Comparison with neutron star data provides us the ranges of $K_0$, $J$ and $L$ that are most consistent with both nuclear data
and astronomical observation.
In the result, we obtain $K_0$, $J$ and $L$ constrained in ranges
$K_0 \sim 230-250$ MeV, $J \sim 31-33$ MeV and $L\sim 55-65$ MeV.
In Section II, we introduce the model and explain the strategy of investigation.
Neutron star properties are considered, and reduced ranges of the EoS parameters are obtained in Section III.
Parameters that are consistent with the neutron star observation are applied to calculating the properties of nuclei in Section IV.
We summarize the work in Section V.
\section{Model and Fitting}
KIDS (Korea:IBS-Daegu-Sungkyunkwan) model provides nuclear energy density functional based on rules for
systematic expansion of the energy of a nucleon in nuclear matter.
In the homogeneous infinite matter, energy per particle is expanded in terms of the Fermi momentum as
\begin{eqnarray}
{\cal E}(\rho,\, \delta) = {\cal T} + \sum_{k=0} (\alpha_k + \beta_k \delta^2) \rho^{1+k/3},
\end{eqnarray}
where ${\cal T}$ is the kinetic energy, and the terms in the summation denote the potential energy originating from the strong interaction.
In the cold nuclear matter $k_F \propto \rho^{1/3}$, so expansion in powers of $\rho^{1/3}$ is equivalent to the expansion
in terms of the Fermi momentum.
It's been shown that 3 $\alpha_k$'s and 4 $\beta_k$'s are suffcient for a correct description of the EoS of nuclear
matter at high densities \cite{kidsnuclei2}.
By transforming the KIDS functional to the Skyrme-type potential, one can easily obtain the single-particle potential of the nucleon in nuclei.
In addition to $\alpha_k$ and $\beta_k$ which determine the EoS of infinite nuclear matter,
we have two more parameters ($\zeta$ and $W_0$) in the single particle potential for the nucleon in nuclei.
We have 9 free parameters altogether.
Determination of the 9 parameters is divided into two steps:
At first, we assume specific values of the 7 EoS parameters $\rho_0$, $E_B$, $K_0$, $J$, $L$, $K_{\rm sym}$ and $Q_{\rm sym}$,
and determine 3 $\alpha_k$'s and 4 $\beta_k$'s to reproduce these values.
In the second step, we fit $\zeta$ and $W_0$ to $E/A$ (energy per nucleon) and $R_c$ (charge radius) of $^{40}$Ca, $^{48}$Ca and $^{208}$Pb.
Agreement to the 6 nuclear data is measured by the $\chi_6^2$ defined by
\begin{equation}
\chi_6^2 \equiv \sum_{n=1}^6 \left(\frac{O^{\rm exp}_n - O^{\rm calc}_n}{O^{\rm exp}_n} \right)^2.
\label{eq:chi6}
\end{equation}
Among the 7 EoS parameters, we fix $\rho_0$, $E_B$, $Q_{\rm sym}$ to 0.16 fm$^{-3}$, 16.0 MeV and 650 MeV, respectively,
and investigate the dependence on the 4 EoS parameters $K_0$, $J$, $L$, and $K_{\rm sym}$.
In order to avoid too large parameter space, 4 EoS parameters are restricted to the ranges allowed by either experiments
or other theories.
For $K_0$, the range is chosen as $220-260$ MeV, and the interval is divided into 10 MeV, i.e. 220, 230, 240, 250 and 260 MeV.
Ranges of $J$ and $L$ are $30-34$ MeV and $40-70$ MeV, respectively, and the interval is assumed equally 1 MeV.
Instead of $K_{\rm sym}$ we consider $K_\tau$
\begin{equation}
K_\tau \equiv K_{\rm sym} - \left(6 + \frac{Q_0}{K_0} \right) L,
\label{eq:ktau}
\end{equation}
and assume three values $-360$, $-420$ and $-480$ MeV for $K_\tau$.
Total number of the combination of ($K_0$, $J$, $L$, $K_\tau$) is $5 \times 5 \times 30 \times3 =2250$,
and the resulting $\chi_6^2$ ranges from $7.5 \times 10^{-6}$ to $3.2 \times 10^{-4}$.
Among the 2250 results, we plot the results of $\chi_6^2$ less than $10^{-4}$ for each $K_0$ value in Fig.~\ref{fig:chi62}.
The results are comparable with the SLy4 and the GSkI models which
give $\chi_6^2$ values $2.24 \times 10^{-4}$ and $6.9 \times 10^{-5}$, respectively.
\begin{figure}
\begin{center}
\includegraphics[width=0.6\textwidth]{figure1}
\caption{$\chi_6^2$ less than $10^{-4}$ for $K_0=$ 220, 230, 240, 250 and 260 MeV.}
\label{fig:chi62}
\end{center}
\end{figure}
There are several notable features in Fig.~\ref{fig:chi62}.
First, minimal $\chi_6^2$ is likely to be located in the range $K_0 = 230-250$ MeV.
This range is consistent with, and narrower than the one from experiment $K_0=200-260$ MeV.
Second, with larger $K_0$ values, smaller $\chi_6^2$ is obtained with larger $K_\tau$ values (smaller in magnitude).
Third, dependence on $K_\tau$ is relatively weak in the range $K_0=230 - 250$ MeV.
Interestingly this range agrees with the range of minimum $\chi_6^2$.
Since the nuclei in the fitting are not so asymmetric i.e. $\delta^2 \ll 1$,
and the densities relevant to the 6 data may be $\rho_0$ and nearby, i.e. $|(\rho-\rho_0)/3\rho_0| \ll 1$,
it would be natural and reasonable to have weak dependence on $K_\tau$.
From the result of fitting to the nuclear data, one can guess that true $K_0$ value may be located in $K_0 = 230-250$ MeV.
However, it is premature to draw a definite conclusion because
density of nuclei is identically close to $\rho_0$, so the information about the density dependence of the EoS
one can obtain from the nuclear data might be limited.
In addition, asymmetry $\delta$ of $^{48}$Ca and $^{208}$Pb are 0.17 and 0.21, respectively,
so they are not large enough to magnify the effect of the symmetry energy.
Neutron stars provide environment with more wider density range and the large neutron-proton asymmetry.
Not abandoning the possibility of small or large $K_0$ value, we include $K_0=220$ and 260 MeV in the consideration of the neutron star.
For each $K_0$ value, ($J$, $L$, $K_\tau$) values are chosen from the two smallest $\chi_6^2$ values.
Equation of state parameters thus determined are summarized in Tab.~\ref{tab:kjl}.
Ranges of $J$ and $L$ in Tab.~\ref{tab:kjl} are $30-34$ MeV and $40-70$ MeV, respectively.
These ranges are similar to the experimental ranges adopted in \cite{dutra}.
\begin{table}
\begin{center}
\begin{tabular} {|ccc|}\hline
$K_0$ & $(J,\, L\, K_\tau)$ & $\chi^2 (\times 10^{-5})$ \\ \hline\hline
220 & (33, 50, $-480$) & 9.45 \\
& (34, 63, $-480$) & 8.61 \\ \hline
230 & (33, 66, $-420$) & 3.04 \\
& (33, 52, $-480$) & 3.01 \\ \hline
240 & (32, 68, $-360$) & 0.75 \\
& (32, 58, $-420$) & 0.89 \\ \hline
250 & (30, 41, $-360$) & 1.50 \\
& (31, 58, $-360$) & 1.43 \\ \hline
260 & (30, 47, $-360$) & 5.55 \\
& (31, 63, $-360$) & 6.03 \\ \hline
\end{tabular}
\end{center}
\caption{EoS parameters ($J$, $L$, $K_\tau$) giving the two smallest $\chi_6^2$ values for each $K_0$ value.}
\label{tab:kjl}
\end{table}
\section{Neutron star}
In 1990's when data of mass and radius were not sufficient, mass observation played a major role in
constraining the nuclear EoS at high densities.
For various reasons, broadly agreed maximum mass was about $1.5 M_\odot$.
However masses close to or larger than $2M_\odot$ have begun to be reported in the 21st century,
and now the general consensus is that the maximum mass of the neutron star may be $2 M_\odot$ or more.
In the core of $2M_\odot$ neutron star, density is large enough that the nucleons are overlapping to each other.
It is very likely that non-nucleonic state will emerge and exist in the core of $2M_\odot$ neutron stars,
so the EoS at the core is hard to determine definitely because of large uncertainties from various sources.
On the other hand, density at the center of a $1.4M_\odot$ neutron star does not exceed $3 \rho_0$ regardless of the siffness of EoS.
At densities smaller than $3\rho_0$, uncertainty due to the non-nucleonic degrees of freedom can still exist, but their effects
are not significant or critical.
For example, mass and radius of neutron stars with $\Lambda$ hyperons were calculated with EoSs over a wide range
of stiffness for the nucleon-nucleon ($NN$), $N\Lambda$ and $\Lambda\Lambda$ interactions \cite{hyperon2015}.
Maximum mass and the mass-radius curves depend on the $N\Lambda$ and $\Lambda\Lambda$ interactions critically,
but the mass-radius relation of the canonical neutron star is hardly affected by the existence of $\Lambda$ hyperons and the stiffness of
their interactions with $N$ and $\Lambda$.
Therefore $1.4M_\odot$ neutron star provides a good laboratory in which we can study the EoS of nucleonic matter at densities
higher than $\rho_0$.
\begin{figure}
\begin{center}
\includegraphics[width=0.65\textwidth]{figure2}
\end{center}
\caption{Neutron star mass and radius relations with $K_0$, $J$, $L$ and $K_\tau$ selected in Tab. \ref{tab:kjl}.}
\label{fig:ns}
\end{figure}
Now more ample and exact data of mass and radius of neutron stars whose mass is close to $1.4M_\odot$ are available.
We collect the state-of-the-art data of the neutron neutron star in Fig. \ref{fig:ns}.
Horizontal bands around $2 M_\odot$ exhibit the data of large masses \cite{largemass1, largemass2, largemass3}.
Gray zone represents the range obtained from the low mass X-ray binary (LMXB) data \cite{lmxb},
error bars denote the analysis of a soft X-ray source observed in the NICER project \cite{nicer1, nicer2},
and the red and the blue regions in the shape of wings correspond to the result of GW170817 (lower small wings) \cite{gw170817}
and GW190425 (upper large wings) \cite{gw190425}.
In the case of NICER two analyses were performed independently for a single object, so the two results are shown in the figure.
A red horizontal line is drawn at mass $1.4M_\odot$ for reference.
Observation data spread broadly on the mass-radius plane, but there is a narrow band in which all the data overalp.
For the mass $1.4M_\odot$, the commonly overlapping region restricts the radius in the range $11.8-12.5$ km.
Ten candidates for the EoS parameters $(K_0,\, J,\, L,\, K_\tau)$ also show widely spread behavior.
At first compared to the LMXB data at the canonical mass (red line), 4 EoSs are incompatible:
2 soft EoSs corresponding to (220, 33, 50, $-480$), (230, 33, 52, $-480$), and 2 stiff EoSs with (240, 32, 68, $-360$), (260, 31, 63, $-360$).
It is notable that $K_0$ and $K_\tau$ are small simultaneously for the 2 soft EoSs, and they are large simultaneously for the 2 stiff EoSs.
On the other hand, we have 4 EoSs that fall into the common range of LMXB, NICER and GW for $M=1.4M_\odot$:
$(K_0,\, J,\, L,\, K_\tau) = (230,\, 33,\, 66,\, -420),\, (240,\, 32,\, 58,\, -420),\, (250,\, 31,\, 58,\, -360),$ and
$(260,\, 30,\, 47,\, -360)$.
There are several interesting feature.
First, the smallest $K_0$ and $K_\tau$ values, 220 MeV and $-480$ MeV are ruled out.
This provides the lower bound of these parameters.
Second, similar to the 2 soft and 2 stiff EoSs incompatible with the LMXB data,
smaller $K_0$ values are grouped with smaller $K_\tau$ values, and vice versa.
On the other hand, $J$ and $L$ values are smaller when $K_0$ and $K_\tau$ are large.
This means $J$ and $L$ contribute to EoS oppositely to $K_0$ and $K_\tau$, i.e.
if $J$ and $L$ soften the EoS, the latter two parameters make it stiff.
As a result, EoS is adjusted to remain in ranges consistent with data.
Most essential aspect of the result is the range of each EoS parameters.
If the consistency with neutron star observation is accounted, acceptable ranges could be specified as
\begin{eqnarray}
K_0 \sim 230-260,\, J \sim 30-33,\, L \sim 45-65
\label{eq:wide}
\end{eqnarray}
all in the unit of MeV.
On the other hand, if we take into account the consistency with the nuclear data in addition to the nuclear star observation,
$K_0 = 260$ MeV could be excluded because $\chi^2_6$ is much larger than the other $K_0$ values. Then we obtain narrower ranges
\begin{eqnarray}
K_0 \sim 230-250,\, J \sim 31-33,\, L \sim 55-65.
\label{eq:narrow}
\end{eqnarray}
The ranges are reduced significantly compared to those constrained from experiment \cite{dutra},
and thus they are expected to give EoSs less uncertain from sub to supra saturation densities.
Compared to a recent work \cite{arxiv2020} in which $K_0$ is fixed to 240 MeV and symmetry energy parameters are
fit to 13 nuclear data, $J$ is slightly larger in this work, but $L$ and $K_\tau$ ranges are almost identical.
With the newly constrained ranges of the EoS parameters,
we can obtain uncertainty range of the extreme properties of nuclei such as the neutron skip thickness or
the position of the neutron drip line. Consideration about the finite nuclei follows in the next section.
\section{Nuclear property}
Ranges of the EoS parameters in Eq. (\ref{eq:wide}) are determined using 6 nuclear data of $^{40, 48}$Ca, $^{208}$Pb,
and the neutron star mass and radius.
Since we are aiming at a unified description of both nuclear matter and nuclei,
we should check the predictive power of our approach for the nuclear properties.
Top prior quantity among numerous nuclear properties might be the binding energy.
\begin{table}
\begin{center}
\begin{tabular}{|c|cccccc|}\hline
& $^{16}$O & $^{40}$Ca & $^{48}$Ca & $^{90}$Zr & $^{132}$Sn & $^{208}$Pb \\ \hline\hline
Exp. & 7.976 & 8.551 & 8.667 & 8.710 & 8.355 & 7.868 \\ \hline
A & 7.946 (0.38) & 8.564 ($-0.16$) & 8.680 ($-0.15$) & 8.681 (0.33) & 8.378 ($-0.28$) & 7.871 ($-0.04$) \\ \hline
B & 7.940 (0.45) & 8.555 ($-0.04$) & 8.673 ($-0.07$) & 8.676 (0.39) & 8.377 ($-0.26$) & 7.869 ($-0.01$) \\ \hline
C & 7.935 (0.51) & 8.546 (0.06) & 8.667 ($-0.01$) & 8.678 (0.37) & 8.375 ($-0.24$) & 7.868 (0) \\ \hline
D & 7.933 (0.54) & 8.538 (0.15) & 8.653 (0.17) & 8.674 (0.42) & 8.374 ($-0.22$) & 7.867 (0.02) \\ \hline
\end{tabular}
\end{center}
\caption{Binding energy per nucleon in MeV for $(K_0,\, J,\, L,\, K_\tau)$ values satsifying the neutron star mass-radius observation.
Set A, B, C and D correspond to $(230, 33, 66, -420)$, $(240, 32, 58, -420)$, $(250, 31, 58, -360)$, and $(260, 30, 47,-360)$, respectively.
Numbers in parenthesis denote the difference from experiment in \%.}
\label{tab:be}
\end{table}
Table \ref{tab:be} displays the binding energy per nucleon for standard spherical magic nuclei.
Numbers in the parentheses denote the difference from experiment in units of \%.
Energies of $^{40, 48}$Ca and $^{208}$Pb are used in fitting the 2 Skyrme force parameters that are coefficients of the
grandient and spin-orbit terms, so it is natural to have good reproduction of these data.
Energies of $^{16}$O, $^{90}$Zr and $^{132}$Sn are, on the other hand, predictions of the model.
The predictions agree well with experiment, giving difference from experiment at the order of 0.5\% or less.
Another positive aspect of the result is that the predicted energy values are independent of the EoS parameters
($K_0$, $J$, $L$, $K_{\rm sym}$).
This indpendence implies that the 4 EoS parameter sets have equal predictive power for the binding energy of magic spherical nuclei.
\begin{figure}
\begin{center}
\includegraphics[width=0.8\textwidth]{figure3}
\end{center}
\caption{Two neutron separation energies for the O, Ca, Sn and Pb isotopic chains
for the four sets of $(K_0, J, L, K_\tau)$ in Tab. \ref{tab:be}.}
\label{fig:s2n}
\end{figure}
One of the main goal of the work is to have reliable predictions about the objects in extreme conditions in which density is
much larger or smaller than the saturation density, or proton-neutron asymmetry approaches to 1.
Neutron stars are good laboratory to explore these effects, but nuclei close to the neutron drip line also provide a nice test ground.
In Fig. \ref{fig:s2n}, we show the result of two neutron separation energy $S_{2n}$ of O, Ca, Zr, Sn and Pb isotopes
calculated with 4 EoS parameter sets in Tab. \ref{tab:be}.
Experimental data are included with red squares.
In order to take into account the effect of continuum state correctly, we use the Hartree-Fock-Bogolyubov theory.
All the isotopes are assumed spherical, and pairing force is incorporated in the contact interaction form
\begin{eqnarray}
V_{\rm pair} = t'_0 \left( 1- \frac{\rho}{2 \rho_0} \right) \delta(\vec{r}_1 - \vec{r}_2).
\label{eq:pairing}
\end{eqnarray}
Parameters in the pairing force are fit to the mean neutron gap of $^{120}$Sn, 1.392 MeV.
Agreement to the experiment is not as good as the energies of the magic nuclei in Tab. \ref{tab:be}.
More investigation is necessary for a better reproduction of the data.
For example, the disagreement can originate from the way to treat the pairing force.
Non-spherical shape of the isotopes could contribute to the discrepancy.
Nevertheless, there are a few points that are worthy to be discussed.
For the oxigen, results of the 4 EoS parameter sets agree to each other fairly well, but there are huge differences from experiment at $N=8$ and $N=10$.
Position of the neutron drip line is predicted identically at $N=22$.
For the calcium, agreement to experiement is reasonable for $N$ less than 30, but it becomes dramatically bad for $N>30$.
For $N$ between 30 and 40, however, the 4 EoS sets predict similar results,
and for $N>40$, predictions are divided into two branches.
Sets with $K_\tau=-420$ MeV give $S_{2n}$ larger than those of $K_\tau=-360$ MeV sets.
Consequently neutron drip lines differ significantly, $N=54$ for $K_\tau=-360$ MeV and $N=62$ for $K_\tau=-420$ MeV.
For the zirconium, agreement to experiment becomes worse for $N>40$.
Four EoS parameter sets agree to each other well for $N$ less than 82, but the prediction splits into two groups for $N>82$.
Similar behavior is observed for the tin isotopes.
For the lead, the same thing happens at another magic number $N=184$.
Neutron drip of the Pb isotopes spread over a huge range of the neutron number, from $N=186$ to $N=220$.
\section{summary}
This work was motivated by the observation that the EoS of nuclear matter which is crucial in understanding the properties of
extremely neutron-rich nuclei and the neutron star is not determined precisely enough yet.
By reducing the uncertainty in the EoS parameters $K_0$, $J$, $L$, and $K_{\rm sym}$, it may be feasible to have more
exact knowledge about the state of matter at extreme conditions.
KIDS functional theory provides a framework adequate for this analysis.
Four EoS parameters are determined to best reproduce the binding energy and the charge radius of $^{40, 48}$Ca and $^{208}$Pb,
and the radius of $1.4M_\odot$ neutron stars determined from various sources and probes.
As a result we obtain the ranges $K_0 \sim 230-250$ MeV, $J \sim 31-33$ MeV, and $L\sim 55-65$ MeV.
As for $K_\tau$, we used three values $-360$, $-420$ and $-480$ MeV, and find that $-480$ MeV makes the EoS too soft
to satisfy the neutron star observation.
Ranges of $J$ and $L$ are consistent with the ranges in a recent work \cite{arxiv2020} in which $K_0$ is fixed to 240 MeV
and the nuclear data used in the fitting are chosen differently from this work.
EoS parameters thus determined are used to calculate the properties of nuclei.
For the binding energies of spherical magic nuclei $^{16}$O, $^{90}$Zr and $^{132}$Sn, all the 4 EoS parameter sets
consistent with the constraint from the neutron star give results agreeing to nuclear data within the errors less than 0.5\%.
Two-neutron separation energy, on the other hand, shows substantial discrepancy with experiment, so it needs more study
to understand the origin of the discrepancy and make the theoretical prediction better.
From the analysis of this work, it has been shown that the uncertainties of $J$ and $L$ are reduced to the ranges
$\pm 1$ MeV and $\pm 5$ MeV, respectively.
The result of two-neutron separation energy shows that the neutron drip line can be sensitive to the value of $K_\tau$.
We can write an approximate form of Eq. (\ref{eq:ktau}) as
\begin{equation}
K_\tau \simeq K_{\rm sym} - 5L.
\end{equation}
If $L$ is determined with errors $\pm 5$ MeV, maximum difference 10 MeV can give about 50 MeV uncertainty in $K_\tau$.
For a better decision of $K_\tau$, it might be necessary to constrain $L$ in regions narrower than $\pm 5$ MeV.
Aside from the uncetainty of the EoS parameters, it is important to have more correct description of the exotic nuclei.
There are many things to consider, e.g. refined treatment of the pairing force, the effect of deformation, and etc.
Some of these subjects are under way.
\section*{Acknowledgments}
This work was supported by the Daegu University Research Grant 2018.
|
\section{Introduction}
Semimartingale and non-semimartingale properties of the standard fractional Brownian motion (FBM) $B^H$ and its mixtures are well understood. These properties are important in modeling stock price \cite{kluppelberg2004fractional,rostek2009option}, constructing arbitrage strategies and hedging policies \cite{rogers1997arbitrage,shiryaev1998arbitrage,salopek1998tolerance,cheridito2003arbitrage},
and modeling rough volatility \cite{gatheral2018volatility,bayer2016pricing,roughvolatility-webpage}.
The standard FBM $B^H$ captures short/long-range dependence, and possesses the self-similar and stationary increment properties, as well as regular path properties.
It may arise as the limit process of scaled random walks with long-range dependence or an integrated shot noise process \cite{PT17}.
A generalized fractional Brownian motion (GFBM) $X$, introduced by Pang and Taqqu \cite{pang-taqqu}, is defined via the following (time-domain) integral representation:
\begin{equation} \label{def-X}
\{X(t)\}_{t\in {\mathbb R}} \stackrel{\rm d}{=} \left\{ c \int_{\mathbb R} \left( (t-u)_+^{\alpha} - (-u)_+^{\alpha} \right) |u|^{-\gamma/2} B({\rm d}u) \right\}_{t\in {\mathbb R}},
\end{equation}
where $(\alpha, \gamma)$ are constants in the region
\begin{equation} \label{def-X-g-c}
\quad \gamma \in [0,1), \quad
\alpha \in \Big(-\frac{1}{2}+ \frac{\gamma}{2}, \ \frac{1}{2}+ \frac{\gamma}{2}\Big)\, ,
\end{equation}
$B({\rm d} u)$ is a Gaussian random measure on ${\mathbb R}$ with the Lebesgue control measure ${\rm d}u$, and $c=c(\alpha,\gamma) \in {\mathbb R}_+$ is a normalization constant (see \eqref{eq: normalizeC}). It is a $H$-self-similar Gaussian process, that is, $\{X(\kappa t): t\in {\mathbb R}\} \stackrel{\rm d}{=} \{\kappa^{H}X(t): t \in {\mathbb R}\}$ for any $\kappa > 0$, with
$$
H:=\alpha - \frac{\gamma}{2} + \frac{1}{2} \in (0,1)\,,$$
but does not have stationary increments.
The Hurst parameter $H$ is determined by two-parameters $(\alpha,\gamma)$ in the range shown in Figure \ref{fig-1}. One may regard $\gamma \in (0,1)$ as a scale/shift parameter.
It appears that the component $ |u|^{-\gamma/2}$ renders the paths rougher.
On the one hand, it is somewhat surprising that the GFBM $X$ has the H{\"o}lder continuity property with the same parameter $H-\epsilon$ for $\epsilon>0$ as the FBM $B^H$, and that the parameter $\gamma$ does affect the differentiability of the paths \cite{ichiba-pang-taqqu}.
The GFBM $X$ defined in \eqref{def-X} can be also written as
\begin{equation} \label{def-X-alternative}
\{X(t)\}_{t\in {\mathbb R}} \stackrel{\rm d}{=} \left\{ c \int_{\mathbb R} \left( (t-u)_+^{H- \frac{1}{2} +\frac{ \gamma}{2} } - (-u)_+^{H- \frac{1}{2} +\frac{ \gamma}{2} } \right) |u|^{-\gamma/2} B( {\mathrm d} u) \right\}_{t\in {\mathbb R}}.
\end{equation}
It is clear that when $\gamma=0$, this becomes the standard FBM $B^H$ with Hurst parameter $H = \alpha +1/2 \in (0,1)$ (equivalently, $\alpha \in ( -1/2, 1/2)$):
\begin{equation} \label{FBM-rep}
\{B^H(t)\}_{t\in {\mathbb R}} \stackrel{\rm d}{=} \left\{ c \int_{\mathbb R} \left( (t-u)_+^{H-\frac{1}{2}} - (-u)_+^{H-\frac{1}{2}} \right) B( {\mathrm d} u) \right\}_{t\in {\mathbb R}}.
\end{equation}
The GFBM $X$ is derived as the limit of integrated power-law shot noise processes in \cite{pang-taqqu} (see a brief review in Section \ref{sec-SNP-limit}).
We have studied in \cite{ichiba-pang-taqqu} some important path properties of the GFBM $X$, including the H{\"o}lder continuity property, the differentiability and non-differentiability properties, and functional/local law of the iterated logarithm (LIL), see Section \ref{sec-GFBM-review} for a summary of its fundamental properties. Some additional path properties such as the exact uniform modulus of continuity and Chung's LIL are recently studied in \cite{wang2021exact,wang2021lower}.
\begin{figure}
\begin{center}
\begin{tabular}{cc}
\begin{tikzpicture}[scale=2]
\draw[thick,->] (-0.3, 0.0) -- (1.3, 0.0) node[anchor=north west] {$\gamma$};
\draw[thick,->] (0,-0.8) -- (0, 1.3) node[anchor=south east] {$\alpha$};
\draw[] (0,0) node[anchor = north east] {$0$};
\draw[] (0,-0.5) -- (1, 0) node[below] {$1$};
\draw[] (0, 0.5) -- (1,1) node[anchor = north west]{};
\draw[] (0, 0.5) node[anchor=south east] {$1/2$};
\draw[] (0, -0.5) node[anchor=north east] {$-1/2$};
\draw[] (1,0) -- (1,1) node[]{};
\draw[dotted] (0, 1/2) -- (1,1/2) node[]{};
\draw[] (0.5, 0.5) node[anchor=south west]{(I)};
\draw[] (0.5, 0.4) node[anchor=north]{(II)};
\draw[fill=gray, opacity =0.2] ( (0,0.5) -- (1, 0.5) -- (1, 1) -- cycle;
\end{tikzpicture}
&
\begin{tikzpicture}[scale=2]
\draw[thick,->] (-0.3, 0.0) -- (1.3, 0.0) node[anchor=north west] {$\gamma$};
\draw[thick,->] (0,-0.8) -- (0, 1.3) node[anchor=south east] {$\alpha$};
\draw[] (0,0) node[anchor = north east] {$0$};
\draw[] (0,-0.5) -- (1, 0) node[below] {$1$};
\draw[] (0, 0.5) -- (1,1) node[anchor = north west]{};
\draw[] (0, 0.5) node[anchor=south east] {$1/2$};
\draw[] (0, -0.5) node[anchor=north east] {$-1/2$};
\draw[] (1,0) -- (1,1) node[]{};
\draw[dotted] (0, 1/2) -- (1,1/2) node[]{};
\draw[] (0.5, 0.5) node[anchor=south west]{(I)};
\draw[] (0.35, 0.45) node[anchor=north]{(II)-1};
\draw[] (0.5, 0.02) node[anchor=north]{(II)-2};
\draw[fill=gray, opacity =0.2] ( (0,0.5) -- (1, 1) -- (1,0.5) -- cycle;
\draw[fill=gray, opacity =0.6] ( (0,0.5) -- (1, 0.5) -- (0, 0) -- cycle;
\end{tikzpicture}
\\
(a) & (b) \\
\end{tabular}
\caption{(a) The GFBM $\,X\,$ in \eqref{def-X} is semimartingale in the region (I): $\,0 < \gamma < 1\,$, $\,1/2 < \alpha < (1+\gamma)/2\,$ and not semimartingale in the region (II): $\,0 < \gamma < 1 \,$, $\, (\gamma-1)/2 < \alpha \le 1/2\,$.
(b) The mixed GFBM $\,Y\,$ in \eqref{eq: Y} is semimartingale in the region (I) and the region (II)-1: $\,0 < \gamma < 1 \,$, $\, \gamma/2 < \alpha < (1+\gamma)/2 \,$.
The fractional Brownian motion corresponds to the line segment: $\, \gamma = 0 \,$, $\, -1/2 < \alpha < 1/2 \,$ with $\, H = \alpha + (1/2)\,$. The standard Brownian motion corresponds to the point $\, \alpha = \gamma = 0\,$.
}
\label{fig-1}
\end{center}
\end{figure}
In this paper we focus on the semimartingale properties associated with the GFBM and its mixture with an independent BM.
For FBM $B^H$, it is shown in \cite{rogers1997arbitrage} that $B^H$ is not a semimartingale for $H \in (0,1/2) \cup (1/2,1)$. When $H=1/2$, $B^H$ is a Brownian motion, and is a martingale (thus, also a semimartingale).
In \cite{cheridito2001mixed}, it is shown that the mixture of independent BM and FBM is a semimartingale for $H \in (3/4,1)$. This important finding is proved using a filtering approach in \cite{cai2016mixed}.
These results have significant implications in financial applications, in particular, arbitrage theory and pricing, see, e.g., \cite{rogers1997arbitrage,cheridito2003arbitrage}.
The proofs for these results rely heavily upon the stationary increments property of the FBM. The lack of stationary increments of the GFBM $X$ requires new approaches to establish similar results.
We identify the regions of the two parameter values $(\alpha, \gamma)$ in which the GFBM $X$ is a semimartingale (see the regions (I) and (II) in Figure \ref{fig-1}(a) and Proposition \ref{prop-sm-X}). To do so, we first establish the necessary and sufficient condition for the square integrability of the derivative of the kernel of this Gaussian process.
We use this to distinguish the parameter regions for the semimartingale property. We can then apply the characterization of the spectral representation of Gaussian semimartingales by Basse \cite{basse2009spectral}.
Another approach for establishing the semimartingale property is to study the so-called $p$-variations as was done for FBM $B^H$ in \cite{MR1785063,cheridito2001regularizing} and for bifractional Brownian motion $B^{H,K}$ in \cite{MR2218338}. We also establish the $p$-variation ($p<1/H, p=1/H, p>1/H$) results for the GFBM $X$ (see Proposition \ref{prop: variation}) and use them to conclude the non-semimartingale property of the GFBM $X$ when $H<1/2$ (see Proposition \ref{prop-sm-X-2}).
Recall that the standard FBM $B^H$ is a semimartingale if and only if $H=1/2$ while the GFBM $X$ can be a semimartingale for $H\in (1/2,1)$ in region (I) of parameters of $(\alpha,\gamma)$, and in the very special case when the GFBM $X$ becomes a BM (see Remark \ref{rem-sm} and Figure \ref{fig-1} for more discussions, noting that there is a
quadrilateral
shape of $(\alpha,\gamma)$ that result in $H \in (1/2,1)$, and there is a line segment of $(\alpha,\gamma)$ over $[0,1]$ resulting in $H=1/2$ but only $\gamma=0$ gives a BM). However, although $X$ is the semimartingale in region (I), it is a process of finite variation and has an explicit expression, thanks to the result of spectral representation of Gaussian semimartingales in Basse \cite{basse2009spectral}.
We then study the semimartingale properties of the mixed GFBM process (sum in \eqref{eq: Y} of an independent BM and GFBM).
It is shown in \cite{cheridito2001mixed,cai2016mixed,bender2011fractional} that the mixed FBM $B^H$ process is a semimartingale with respect to its own filtration if and only if $H\in \{1/2\} \cup (3/4,1]$.
We show in Proposition \ref{prop-mixture} that the mixed GFBM process is a semimartingale in the region of the two parameter values $(\alpha, \gamma)$ that is equivalent to $H\in (1/2,1)$ (see regions (I) and (II)-1 in Figure \ref{fig-1}(b)),
as well as in the very special case when GFBM is also a BM.
It is also worth noting that when $\gamma > 0 $, the wide range of values of the Hurst index $H$ for which the mixed GFBM process is a semimartingale. In the two regions (I) and (II)-1, the mixed GFBM process has different behaviors due to the fact that in region (I) the GFBM $X$ is a process of finite variation.
For both regions (I) and (II)-1,
we can use the characterization of the equivalence of Gaussian measures in Shepp \cite{shepp1966radon}, and show that the absolute continuity of the measure of the mixed GFBM with respect to that of the standard BM, and provide an expression of the Radon--Nikodym density, using the solutions to the associated Wiener--Hopf integral equations.
For that purpose, we establish that $H \in (1/2,1)$ is the necessary and sufficient condition for the second partial derivative function fo the covariance function of the GFBM $X$ to be square integrable (see Lemma \ref{lm:sq-integrability}).
For region (I), thanks to the finite variation property of the GFBM $X$,
the mixed GBFM becomes a Brownian motion plus a random drift of finite variation, and as a consequence, we provide another expression of the Radon--Nikodym density, in terms of the conditional expectation, applying the results in \cite{MR281279}.
We also conjecture that the mixed process is not a semimartingale when $H\in (0,1/2)$ (see further discussions in Remark \ref{rem-mixture}).
We then use the GFBM and its mixtures in the study of financial asset models.
The first application is stock price process with long range dependence in Section \ref{sec-finance-price}.
The price models we introduce in Section \ref{sec-price} generalize those using shot noise process and FBMs in the literature \cite{kluppelberg2004fractional,altmann2008shot,schmidt2017shot,Stute}.
When the mixed BM and GFBM processes are semimartingales, we derive the Radon--Nikodym derivative for the equivalent martingale measure (Proposition
\ref{prop-pricing-RN}), which can then be used for the price dynamics of various options.
This occurs when the parameters $(\alpha, \gamma)$ lie in the regions (I) and (II)-1, in each of which the Hurst index of GFBM is in $(1/2,1)$. This is in contrast with the model using the standard FBM, which only occurs when the Hurst index of the FBM is in $(3/4,1)$, see \cite{cheridito2001mixed,cai2016mixed}.
In our framework,
the larger range of the Hurst index thus provides greater flexibility in modeling and for further theoretical analysis through the use of It{\^o}'s formula and the properties of semimartingales.
We discuss the implications on arbitrage in asset pricing for stock price processes using the GFBM and its mixtures.
With FBM $B^H$,
arbitrage in fractional Bachelier and Black-Scholes models has been well studied in \cite{rogers1997arbitrage,shiryaev1998arbitrage,salopek1998tolerance,cheridito2003arbitrage}.
If one uses the GFBM $X$ as a stock price process, we find that the only non-arbitrage scenario is the Brownian case (with both $\alpha=0$ and $\gamma=0$).
Although the GFBM $\,X\,$ is a semimartingale in parameter region (I), since it is a process of finite variation, arbitrage exists as shown in \cite{HPS84}.
On the other hand, if one uses the mixed GFBM $Y$ as a stock price process, in the parameter region (I), the process becomes a semimartingale as a Brownian motion with a finite-variation drift, and thus, no arbitrage exists. This price process with the mixed GFBM is of particular interest, since it also exhibits the long range dependence, similar to that with the mixed FBM.
Rogers \cite{rogers1997arbitrage} pointed out that it is possible to construct a process similar to the FBM to model long-range dependence of returns while avoiding arbitrage. We provide an example of a price model with these desirable properties.
We identify self-financing arbitrage strategies in the Bachelier and Black-Scholes models with the GFBM $X$ for the region of parameters that results in $H \in (1/2,1)$ in which the process $X$ is not a semimartingale, taking the approach in Shirayaev \cite{shiryaev1998arbitrage}. See more discussions in Section \ref{sec-arbitrage}.
The second application is about rough stochastic volatility models. There have been many activities in the study of rough volatility using FBM since the seminal research initiated by Gatheral et al. \cite{gatheral2018volatility}, see the extensive literature in \cite{roughvolatility-webpage}.
VIX and other volatility derivatives modeled with Brownian motion have been extensively studied, see, e.g., the books \cite{gatheral2011volatility,bergomi2015stochastic}.
Bayer, Friz and Gatheral \cite{bayer2016pricing} has recently proposed a rough Bergomi model using the
Riemann--Liouville (R-L) FBM. In particular, they provide an approximation for the VIX variance swap which is a forward contract on realized variances. See also \cite{jacquier2018vix} on VIX futures using the model.
We particularly focus on a generalization of the rough Bergomi model using the generalized R-L FBM, and provide an approximation for the associated VIX variance swap.
We use numerical examples to illustrate the impact of the ``non-stationarity" parameter $\gamma$ on the estimates of the VIX variance swaps. We observe that for the same Hurst parameter value $H$, a slight positive $\gamma$ value can dramatically affect the VIX variance. See Table \ref{Table:1} and Figures \ref{Figure:5}, \ref{Figure:4} and \ref{fig:3}.
We then propose various models for the volatility process using the GFBM, the mixed GFBM process, and the generalized fractional Ornstein--Ulenbeck (fOU) processes driven by the GFBM or the mixed GFBM process.
These models may be more advantageous since the GFBM and its mixture can be a semimartingale for a wide range of Hurst parameter values while possessing the long range dependence and roughness properties.
The paper is organized as follows. In Section \ref{sec-GFBM-review}, we give the precise definition of the GFBM and summarize some of its properties. In Sections \ref{sec-sm} and \ref{sec-var-sm}, we study the semimartingale properties of the GFBM $X$. In Section \ref{sec-mix}, the semimartingale property of the mixed BM and GFBM process is investigated. We present the applications in financial models in Sections \ref{sec-finance-price} and \ref{sec-roughvol}.
\section{A generalized fractional Brownian motion} \label{sec-GFBM-review}
The GFBM process $X$ defined in \eqref{def-X} has the following properties:
\begin{itemize}
\item[(i)] $X(0) = 0$ and $\mathbb{E}[X(t)]=0$ for all $t\ge 0$;
\item[(ii)] $X$ is a Gaussian process and $\mathbb{E}[X(t)^2] = t^{2H}$ for $t \ge 0$;
\item[(iii)] $X$ has continuous sample paths almost surely;
\item[(iv)] $X$ is self-similar with Hurst parameter $H \in (0,1)$;
\item[(v)] the paths of $X$ are H{\"o}lder continuous with parameter $H-\epsilon$ for $\epsilon>0$;
\item[(vi)] the paths of $X$ is non-differentiable if $\, \alpha \in (0, 1/2] \,$ and differentiable if $\, \alpha > 1/2\,$ almost surely (see non-differentiable region (II) $\,\alpha \in (0, 1/2] \,$ and differentiable region (I) $\, \alpha > 1/2\,$ in Figure \ref{fig-1}).
\end{itemize}
These properties are established in \cite{pang-taqqu,ichiba-pang-taqqu}. See Proposition 5.1 \cite{pang-taqqu} for properties (iii) and (iv), and Theorems 3.1 and 4.1 in \cite{ichiba-pang-taqqu} for (v) and (vi).
Also recall that the normalization constant $c= c(\alpha, \gamma) \in {\mathbb R}_+$ is given by \begin{equation} \label{eq: normalizeC}
\begin{split}
c(\alpha,\gamma) \, :=\, & \Big( \int^{1}_{0} ( 1 - v)^{2 \alpha } v^{-\gamma} {\mathrm d} v + \int^{\infty}_{0} [(1+v)^{\alpha} - v^{\alpha}]^{2}v^{-\gamma} du v \Big)^{-1/2} \\
\, =\, & \Big( \text{Beta} ( 1 - \gamma, 2\alpha + 1) \\
& {} + \Big( \frac{\,\Gamma ( 1 - \gamma) \,}{\, \Gamma ( - 2 \alpha ) \,} - \frac{\,2 \Gamma ( 1 + \alpha - \gamma) \,}{\,\Gamma ( - \alpha)\,}\Big) \Gamma ( -1 - 2 \alpha + \gamma) \Big) \Big)^{-1/2} \, ,
\end{split}
\end{equation}
as shown in Lemma 2.1 of \cite{ichiba-pang-taqqu}.
Here, $\, \Gamma ( a) \, :=\, \int^{\infty}_{0} x^{a-1} e^{-x} {\mathrm d} x \,$, $\,a > 0 \,$ and $\, \text{Beta}(a,b) \, :=\, \int^{1}_{0} x^{a-1} (1-x)^{b-1} {\mathrm d} x \,$, $\,a, b > 0 \,$ are the Gamma and Beta functions, respectively, and $\, \Gamma (-a) := (-a)^{-1} \Gamma(-a+1) \,$ for positive non-integer $\,a\,$.
Recall that in the case $\gamma = 0$, the FBM $B^H$ has stationary increments.
Namely,
the second moment of its increment:
$$
\mathbb{E}\bigl[(B^H(s) - B^H(t))^2\bigr] = c^2 |t-s|^{2H}\, ,
$$
and the covariance function
\begin{equation} \label{eqn-cov-FBM}
\mathbb{E}\bigl[B^H(s) B^H(t)\bigr] = \frac{1}{2}c^2 \big(t^{2H} + s^{2H} - |t-s|^{2H}\big)\,,
\end{equation}
where $c=c(\alpha,0) =c(H-1/2,0)$ in \eqref{eq: normalizeC}.
When $\gamma \in (0,1)$, in comparison with the FBM $B^H$, the process $X$ loses the stationary increment property. In particular,
the second moment of its increment is
\begin{align} \label{eqn-Phi}
\Phi(s,t)&:=\mathbb{E}\bigl[(X(t) - X(s))^2\bigr] \nonumber\\
&\, = c^2 \int_{\mathbb R} \Big( (t-u)_+^{\alpha} - (s-u)_+^{\alpha} \Big)^2 |u|^{-\gamma} {\rm d}u \nonumber\\
&\,= c^2 \int_s^t (t-u)^{2\alpha} u^{-\gamma} {\rm d}u
+ c^2\int_0^s ( (t-u)^{\alpha} - (s-u)^{\alpha} )^2 u^{-\gamma} {\rm d}u \nonumber\\
&\qquad + \int_0^\infty ((t+u)^{\alpha} -(s+u)^\alpha)^2 u^{-\gamma} {\rm d}u\, ,
\end{align}
and the covariance function is
\begin{align} \label{eqn-Psi}
\Psi(s,t) & = \text{\rm Cov}(X(t), X(s)) = \mathbb{E}[X(s)X(t)] \nonumber\\
&= c^2 \int_{\mathbb R} \Big( \left( (t-u)_+^{\alpha} - (-u)_+^{\alpha} \right) \left( (s-u)_+^{\alpha} - (-u)_+^{\alpha} \right) \Big) |u|^{-\gamma} {\rm d}u \nonumber\\
&= c^2 \int_0^s (t-u)^{\alpha} (s-u)^{\alpha} u^{-\gamma} {\rm d}u \nonumber\\
& \qquad + c^2 \int_0^{\infty} ((t+u)^{\alpha} - u^{\alpha}) ((s+u)^{\alpha} - u^{\alpha})u^{-\gamma}{\rm d}u\, ,
\end{align}
for $0\le s \le t$. %
\medskip
We also remark that generalized FBMs are stated in a more general form with the additional terms involving $(t-u)_{-}^\alpha - (-u)_{-}^\alpha$ in the integrands in \cite[Sections 5.1 and 5.2]{pang-taqqu}. In this paper we focus on the representations of $X$ in \eqref{def-X} since the other forms with additional terms can be treated similarly.
\section{When is the GFBM a semimartingale?} \label{sec-sm}
Any centered Gaussian process $X$ with right-continuous sample paths has a spectral representation in distribution \cite{kuelbs1973representation}, that is,
$$
X(t) \stackrel{\rm d}{=} \int_{-\infty}^t K_t(s) {\rm d} N(s), \quad t\ge 0\, ,
$$
where $N$ is an independently scattered centered Gaussian random measure and
$(t,s) \to K_t(s)$ is a square-integrable deterministic function.
Basse \cite[Theorem 4.6]{basse2009spectral} characterizes the spectral representation of Gaussian semimartingales, identifying the family of kernels $K_t(s)$ for which the representation $ \big\{ \int_{-\infty}^t K_t(s) {\rm d} N(s): t\ge 0\big\}$ is a semimartingale with respect to the natural filtration $\{\mathcal{F}^N_t: t \ge 0\}$.
Specifically for one-dimensional case $N(\cdot) \equiv B(\cdot)$ being the Brownian Gaussian random measure, it says that $\{X(t): t\ge 0\}$ is an $\{\mathcal{F}^N_t: t \ge 0\}$ semimartingale if and only if for $t\ge 0$, the kernel can be represented as
\begin{equation} \label{eqn-K-g-Psi}
K_t(s) \;=\; g(s) + \int_0^t \Psi_r(s) \mu({\rm d}r)\,,
\end{equation}
where $g:\mathbb R_{+} \to \mathbb R$ is locally square integrable with respect to the Lebesgue measure, $\mu (\cdot)$ is a Radon measure on $\mathbb R_{+} $ and a measurable mapping $\Psi_{r}(s): (r, s) \to \mathbb R $ is square integrable with respect to the Lebesgue measure, $\int_{-\infty}^{\infty} \lvert \Psi_{r}(s)\rvert^{2} {\mathrm d} s = 1 $, and $\Psi_{r}(s) = 0 $, $r < s$.
In the case of FBM $B^H$, we have the kernel function
$$
K_t(s) = (t-s)_+^{H-1/2} - (-s)_+^{H-1/2}\, ,
$$
and $N(\cdot) = B(\cdot)$.
Since it is a $\{\mathcal{F}^B_t: t \ge 0\}$ semimartingale if and only if $H=1/2$, i.e., a Brownian motion,
applying \cite[Theorem 4.6]{basse2009spectral}, we have $g\equiv 0$, and $ \Psi_r(s) \equiv 1$.
For the GFBM $X$, we have the kernel function
\begin{equation} \label{eqn-Kt}
K_t(s) = [( (t-s)_{+})^{\alpha} - ((-s)_{+})^{\alpha} ] \, \lvert s \rvert^{-\gamma/2}\, ,
\end{equation}
and $N(\cdot) = B(\cdot)$.
Define the function
$\,\Psi_{t}(\cdot) \,$ by
\[
\Psi_{t}(s) \, :=\, C_{t}^{-1} \, \big[ \alpha ( t - s)^{\alpha - 1 } s^{-\gamma/2} \cdot {\bf 1}_{ \{0 < s < t \}} +
\alpha ( t - s)^{\alpha - 1 } (-s)^{-\gamma/2} \cdot {\bf 1}_{ \{s < 0 \}} \big] \, ,
\]
where $C_{t}$ is a time-dependent, normalizing constant defined by
\[
C_{t} := \alpha \, t^{H} \, ( \text{Beta} (1 - \gamma, 2\alpha - 1) + \text{Beta} ( 1 - \gamma, 1 - 2 \alpha + \gamma)){^{1/2}} \, , \quad t > 0 \, .
\]
\begin{lemma} \label{lm: 3.1}
The function $\,\Psi_{t}(\cdot) \,$ is square integrable with respect to the Lebesgue measure if and only if $\, 1/2 < \alpha < (1+\gamma)/2\,$ and $\, \gamma \in (0, 1) \,$ (region (I) in Figure \ref{fig-1}(a)). In this case, $\int_{-\infty}^{\infty} \lvert \Psi_{t}(s)\rvert^{2} {\mathrm d} s = 1 $.
\end{lemma}
\begin{proof}
If $\, 1/2 < \alpha < (1+\gamma)/2\,$ and $\, \gamma \in (0, 1) \,$,
it follows from the definition that
\[
\alpha^{-2}C_{r}^{2} \int^{\infty}_{-\infty} \lvert \Psi_{r}(s)\rvert^{2} {\mathrm d}s = \int^{r}_{0} ( r-s)^{2(\alpha - 1)} s^{-\gamma}{\mathrm d} s +
\int^{0}_{-\infty} ( r-s)^{2(\alpha - 1)} (-s)^{-\gamma}{\mathrm d} s \, ,
\]
where the first and the second terms are rewritten by the change of variables as
\begin{equation} \label{eqn-Psi-intp-1}
\begin{split}
\int^{r}_{0} (r-s)^{2(\alpha - 1)} s^{-\gamma} {\mathrm d} s & \, =\, r^{2H} \int^{1}_{0} (1-u)^{2(\alpha - 1)} u^{-\gamma} {\mathrm d} u \\
& \, =\, r^{2H}\text{Beta} ( 1-\gamma, 2 \alpha - 1) \,, \\
\int^{0}_{-\infty} (r-s)^{2(\alpha - 1)} (-s)^{-\gamma} {\mathrm d} s & \, =\, r^{2H} \int^{\infty}_{0} (1+u)^{2(\alpha - 1)} u^{-\gamma} {\mathrm d} u \\
& \, =\, r^{2H} \text{Beta} ( 1 - \gamma, 1 + \gamma - 2 \alpha ) \, .
\end{split}
\end{equation}
Thus, in this case, $\int^{\infty}_{-\infty} \lvert \Psi_{t}(s)\rvert^{2} {\mathrm d} s \, =\, 1 $. On the other hand, if the conditions on the parameters $\alpha$ and $\gamma$ are not satisfied, then these terms are not integrable. Thus we conclude the proof.
\end{proof}
\begin{proposition} \label{prop-sm-X}
The following properties hold:
\begin{itemize}
\item[(i)] If $\, \gamma \in (0, 1) \,$ and $\, 1/2 < \alpha < (1+\gamma)/2 \,$ (region (I) in Figure~\ref{fig-1}(a)), then $\, \{ X(t), t \ge 0 \} \,$ in \eqref{def-X} is a semimartingale with respect to $\, \mathcal F^{B}(\cdot)\,$, more specifically, a process of finite variation:
\begin{equation} \label{eq: dvofX}
\frac{\,{\mathrm d} X(t)\,}{\,{\mathrm d} t \,} \, =\,
c \int^{t}_{-\infty}\Psi_{t}(s) {\mathrm d} B(s) \, , \quad t \ge 0 \,,
\end{equation}
where $\,\Psi_{t}(\cdot) \,$ is defined by
\begin{equation} \label{eq: Psi_{t}(s)}
\Psi_{t}(s) \, :=\, \alpha ( t - s)^{\alpha - 1 } \lvert s \rvert^{-\gamma/2}
\end{equation}
for $\, s < t \,$, and $\, \Psi_{t}(s) := 0 \,$ for $\, s > t \,$.
\item[(ii)] If $\, \gamma = 0 \,$ and $\, \alpha \in (-1/2, 0) \cup (0, 1/2) \,$, $\, \{X(t) , t \ge 0 \}\,$ in \eqref{def-X} is reduced to a fractional Brownian motion and is not semimartingale. If $\, \gamma \in (0, 1) \,$ and $\, \alpha \le 1/2 \,$ (region (II) in Figure~\ref{fig-1}(a)), then $\, \{X(t) , t \ge 0 \}\,$ in \eqref{def-X} is not a semimartingale with respect to $\, \mathcal F^{B}(\cdot)\,$.
\item[(iii)] Particularly, if $\,\gamma = 0 \,$ and $\alpha = 0$, that is, $H=1/2$, it is a Brownian motion and thus a semimartingale. However, for $\gamma=2 \alpha \in (0,1)$ with $H=1/2$, it is not a semimartingale with respect to $\, \mathcal F^{B}(\cdot)\,$.
\end{itemize}
\end{proposition}
\begin{remark} \label{rem-sm}
{\rm
In the range (region (I) of Figure~\ref{fig-1}(a)) of parameters $(\alpha, \gamma)$: $\, \gamma \in (0, 1) \,$ and $\, 1/2 < \alpha < (1+\gamma)/2 \,$, the Hurst parameter $H = \alpha - \frac{\gamma}{2} + \frac{1}{2} \in \big(\frac{1}{2}, 1 \big)$.
Note that when $\alpha$ is close to $1/2$, and $\gamma$ is close to $1$, the Hurst parameter $H$ is also close to $1/2$, which differs from the standard FBM case with $\gamma =0$ and $\alpha=0$ resulting in $H=1/2$.
The range of values of Hurst parameter $H$ possessing the semimartingale property is expanded from a single value $1/2$ for the standard FBM, to the half interval $[1/2,1)$ for the process $X(t)$, although the process $X(t)$ is a process of finite variation in region (I) with $H\in (1/2,1)$.
Moreover, note that $H=1/2$ here only corresponds to the singular point $\alpha=0, \gamma=0$. The GFBM $X$ can have $H=1/2$ on the line segment $\alpha =\gamma/2$, which is a BM if and only if $\alpha=0$. }
\end{remark}
\begin{remark}[Differentiability]
{\rm It is shown in \cite{ichiba-pang-taqqu} that the regions (I) and (II) of Figure~\ref{fig-1}(a) correspond to the regions of almost sure differentiable and non-differentiable paths, respectively, that is, in the region (I) the sample path of GFBM is differentiable, while in region (II) the sample path of GFBM is not differentiable. This is not just a coincidence but it turns out that when $\,\alpha > 1/2\,$, it is a semimartingale and its (local) martingale part in the semimartingale decomposition is zero, and its finite variation part is the integral of a Gaussian process. When $\,\alpha \, =\, \gamma =0\,$, it is a Brownian motion with non-differentiable sample path. }
\end{remark}
\begin{proof}
(i) Let us consider the case $\, 1/2 < \alpha < (1+\gamma)/2\,$ and $\, \gamma \in (0, 1) \,$. Thanks to the integrability of \eqref{eqn-Psi-intp-1}
in this parameter set, the square integrability of $\, \Psi_{t}(\cdot)\,$ is assured, by Lemma \ref{lm: 3.1}, and hence, by Theorem 4.6 of Basse (2009), we have the representation
\begin{equation} \label{eqn-X-regionI}
X(t) \, =\, c \int^{t}_{0} \Big( \int^{r}_{-\infty} \Psi_{r}(s) {\mathrm d} B(s) \Big) {\mathrm d} r \,, \quad t \ge 0 \, ,
\end{equation}
where $\,\Psi_{t}(\cdot) \,$ is defined in \eqref{eq: Psi_{t}(s)}.
Thus it is a process of finite variation with the first derivative \eqref{eq: dvofX}, in particular, it is a semimartingale. See also section 4 of \cite{ichiba-pang-taqqu}.
\smallskip
(ii) When $\, \gamma \, =\, 0 \,$, $\, X(\cdot) \,$ in \eqref{def-X} is a FBM with Hurst index $\, \alpha + 1/2\,$, and it is not a semimartingale if $\,\alpha \in (-1/2, 0) \cup (0, 1/2)\,$. Thus, let us consider the case $\, \gamma \in (0, 1)\,$, $\, \alpha \le 1/2\,$ and show the claim by contradiction.
Suppose that $\, X(\cdot) \,$ in \eqref{def-X} is a semimartingale with respect to $\, \mathcal F^{B}(\cdot)\,$. We know $\, \mathbb E [ X^{2}(t)] \, =\,
t^{2\alpha - \gamma +1}\,$ for $\, t \ge 0 \,$.
Then by Theorem 4.6 of \cite{basse2009spectral}
with $\,N_{\cdot} \, =\,B(\cdot) \,$, $\,g(s) \, :=\, 0\,$, $\, C_{t}\, =\, (-\infty, t ]\,$,
there is a canonical decomposition
\begin{equation}
\begin{split}
\, X(t) \, =\, &\int^{t}_{-\infty} K_{t}(s) {\mathrm d} B(s) \, =\, \int^{t}_{-\infty} [ ( (t-s)_{+})^{\alpha} - ((-s)_{+})^{\alpha} ] \, \lvert u \rvert^{-\gamma/2} {\mathrm d} B(s)
\end{split}
\end{equation}
for $\, t \ge 0 \,$, where the integrand $\, K_{t}(s) \,$ has the form:
\begin{equation} \label{eq: Kts}
K_{t}(s) \, :=\, g(s) + \int^{t}_{0} \Psi_{r}(s) \mu ({\mathrm d} r ) \,, \quad 0 \le s \le t \, .
\end{equation}
Here $\, g(\cdot) \,$ is square integrable with respect to the Lebesgue measure, $\, \mu (\cdot) \,$ is a Radon measure on $\, \mathbb R_{+}\,$, and $\, \Psi_{t}(s) \,$ is a measurable mapping satisfying
\begin{equation} \label{eq: BasseTHM4.6}
\int^{\infty}_{-\infty} \lvert \Psi_{t}(s) \rvert^{2} {\mathrm d} s \, =\, 1 \, , \quad \text{ and } \quad \Psi_{t}(s) \equiv 0 \, \, \, (s> t) \, .
\end{equation}
Taking derivatives with respect to $\, t\,$ in (\ref{eq: Kts}), we have
\[
\frac{\,{\mathrm d} K_{t}(s)\,}{\,{\mathrm d} t \,} \, =\, \Psi_{t}(s) \cdot \frac{\, \mu ( {\mathrm d} t ) \,}{\, {\mathrm d} t \,} \, ,
\]
while the integrand $\,K_{t}(s)\,$ in \eqref{eqn-Kt}
has the derivative with respect to $\,t \,$:
\[
\frac{\,{\mathrm d} K_{t}(s) \,}{\,{\mathrm d} t \,} \, =\,
\alpha ( t - s)^{\alpha - 1 } \lvert s \rvert^{-\gamma/2}\,.
\]
Thus by comparing these two expressions and by setting $\,\mu ( {\mathrm d} r ) \, =\, {\mathrm d} r \,$
as the Lebesgue measure, we identify $\, \Psi_{t}(s) \,$ as in \eqref{eq: Psi_{t}(s)}.
However, as in Lemma \ref{lm: 3.1}, if $\, \alpha \le 1/2\,$ and $\, \gamma \in (0, 1)\,$, $\, \Psi_{t}(\cdot) \,$ is not square integrable for every $\,t > 0\,$.
This yields a contradiction to (\ref{eq: BasseTHM4.6}). Thus, we claim that $\, X(\cdot) \,$ in \eqref{def-X} is not semimartingale with respect to $\, \mathcal F^{B}(\cdot)\,$, if $\, \gamma \in (0, 1)\,$ and $\, \alpha \le 1/2\,$.
\smallskip
(iii) The standard Brownian motion case $H = 1 / 2 $ is indeed a semimartingale. The second statement on the parameter sets $\, \gamma \, =\, 2 \alpha \in (0, 1) \,$ with $\, H \, =\, 1/ 2\,$ is proved as a special case of (ii).
\end{proof}
\section{Variations of the GFBM $X$}
\label{sec-var-sm}
For the standard FBM $B^H$, it is shown in \cite[Proposition 3.14]{MR1785063} that
$B^H$ has a $1/H$-variation, that is,
$$
\lim_{\varepsilon \downarrow 0} \int_0^t \frac{\,1\,}{\,\varepsilon\,} \lvert B^H(s+\varepsilon) - B^H(s)\rvert^{1/H} {\mathrm d} s \,= \, \varrho_H t\,
$$
in the sense of convergence uniformly in compact sets,
where $\varrho_H = \mathbb{E}[|Z|^{1/H}]$ for $Z\sim N(0,1)$. Here, the limit is in the sense of convergence in probability uniformly on every compact interval (\text{ucp}). It can be also shown that the classical variation
$$
\sum_{i=0}^{n-1} |B^H_{t_{i+1}} - B^H_{t_i}|^{1/H} \xrightarrow[n \to\infty ]{L^{1}}\, \varrho_H t\,,
$$
where $0=t_0 < \dots < t_n =t$ is a partition of $[0,t]$, see Proposition 3.14 in \cite{MR1785063} and Remark 1 in \cite{MR2218338}.
Then by Propositions 1.9 and 1.11 of \cite{cheridito2001regularizing} one can conclude that $B^H$ is not a semimartingale with respect to $\, \mathcal F^{B}(\cdot)\,$, if $H < 1/2$. Note that the results in \cite{cheridito2001regularizing} involve the notion of the weak semimartingales (see Definition 1.5 of \cite{cheridito2001regularizing}).
This approach of evaluating the variations is also used in \cite{MR2218338} to show that the bifractional Brownian motion $B^{H,K}$ with parameters $\,(H, K)\,$, $\,H \in (0, 1)\,$, $\,K \in (0, 1]\,$ is not a semimartingale, if $HK\neq 1/2$, see Propositions 1--3 and Remark 1 there. Recall that the bifractional Brownian motion $\,B^{H,K}\,$ is a centered Gaussian process with $\,B^{H,K}(0) \, =\, 0 \,$ and covariance function
\[
\mathbb E [ B^{H,K}(s) \, B^{H,K}(t) ] \, =\, \frac{\,1\,}{\,2^{K}\,} ( ( t^{2H} + s^{2H})^{K} - \lvert t - s \rvert^{2\, H\, K} ) \,, \quad s, t \ge 0 \, .
\]
The bifractional Brownian motion $\,B^{H,K}\,$ is a FBM with Hurst index $\,H \in (0, 1) \,$ if $\, K \, =\, 1\,$.
We use this approach to establish the following properties of the GFBM $X$ in \eqref{def-X}. The proof of the proposition below is given in Appendix \ref{sec-app-1}.
\begin{proposition} \label{prop: variation} Let
$$\, \rho_{\alpha, \gamma} \, :=\
(c^{2} \, \text{\rm Beta} ( 1 + 2\alpha , 1 - \gamma) )^{1/(2H)} \, \sqrt{2^{1/H}/\pi } \, \Gamma ( (1 + (1/H))/2 )\,.$$ Then we have the convergence in $\,L^{1}\,$:
\begin{equation} \label{eq: prop variation}
\frac{\,1\,}{\,\varepsilon\,} \int^{t}_{0} \lvert X(s+\varepsilon) - X(s)\rvert^{1/H} {\mathrm d} s \, \xrightarrow[\varepsilon \downarrow 0 ]{L^{1}}\, \rho_{\alpha, \gamma} \, t \,, \quad t > 0 \, ,
\end{equation}
and similarly, for every partition $\, \pi: 0 = t_{0} < \cdots < t_{n} \, =\, t \,$ of $\, [0, t] \,$ with size $\, \lvert \pi \rvert \, :=\, \max_{1\le i \le n} \lvert t_{i} -t_{i-1}\rvert\,$, we have
\begin{equation} \label{eq: prop variation 2}
\sum_{i=0}^{n-1} \lvert X(t_{i+1}) - X(t_{i}) \rvert^{1/H}\xrightarrow[n \to \infty]{L^{1}} \rho_{\alpha, \gamma} t \, .
\end{equation}
\end{proposition}
It is immediate from \eqref{eq: prop variation 2} that for $p<1/H$, we have
\begin{equation} \label{eq: prop variation 3}
\sum_{i=0}^{n-1} \lvert X(t_{i+1}) - X(t_{i}) \rvert^{p}\xrightarrow[n \to \infty]{L^{1}} +\infty \,,
\end{equation}
and for $p>1/H$, we have
\begin{equation} \label{eq: prop variation 4}
\sum_{i=0}^{n-1} \lvert X(t_{i+1}) - X(t_{i}) \rvert^{p}\xrightarrow[n \to \infty]{L^{1}} 0 \, .
\end{equation}
\begin{proposition} \label{prop-sm-X-2}
The process $X$ in \eqref{def-X} is not a semimartingale with respect to its own filtration $\, \mathcal F^{X}(\cdot)\,$ if $H < 1/2$, that is, $\gamma \in (0,1)$ and $\alpha \in (-1/2+\gamma/2, \gamma/2)$.
Since $\mathcal F^{X}\subset \mathcal F^{B}$, it is not a
semimartingale with respect to $\, \mathcal F^{B}(\cdot)\,$ if $H<1/2$ (which is part of Proposition \ref{prop-sm-X} (ii)).
\end{proposition}
\begin{proof}When $\, H < 1/2\,$, as a consequence of \eqref{eq: prop variation 3} with $p=2 < 1/H$, by Propositions 1.9 and 1.11 of \cite{cheridito2001regularizing}, $X$ in \eqref{def-X} is not a semimartingale. Note that when $\, H > 1/2\,$, $X$ in \eqref{def-X} is a process of zero quadratic variation by \eqref{eq: prop variation 4}.
\end{proof}
\section{Mixed BM and GFBM} \label{sec-mix}
In this section we consider the semimartingale properties of the following process
\begin{equation} \label{eq: Y}
Y(t) = \widetilde{B}(t)+ X(t), \quad t\ge 0,
\end{equation}
where $\widetilde{B}(t)$ is a standard Brownian motion and $X(t)$ is the GFBM defined in \eqref{def-X}, independent of $\widetilde{B}(t)$. Let us call $\,Y\,$ the {\em mixed GFBM}.
In the case of FBM $B^H$, we denote
$$
Y^H(t) = \widetilde{B}(t) + B^H(t)\, , \quad H \in (0,1)\, , \quad t \ge 0\, .
$$
It is shown in \cite[Theorem~1.7]{cheridito2001mixed} and \cite[Theorem~2.7]{cai2016mixed} (see also \cite{bender2011fractional}) that $Y^H(t)$ is a semimartingale with respect to its own filtration if and only if $H \in \big\{\frac{1}{2}\big\} \cup \big(\frac{3}{4}, 1\big]$. In \cite{cheridito2001mixed}, the concept of weak semimartingale and a theorem on Gaussian processes in \cite{stricker1984quelques} is used. On the other hand,
in \cite{cai2016mixed}, the filtering approach is used.
In particular, the mixed FBM $Y^H$ is innovated by a martingale in its natural filtration for all $H \in (0,1]$. Then the equivalence property with respect to
the Wiener measure is established for $H \in (3/4, 1]$ and the equivalence property with respect to
the Wiener measure is established for $H \in (1/4)$. The associated Radon-Nikodym density formulas are then derived in these ranges of the parameter $H$.
Let $\mu^{Y,H}$ be the probability measure induced by $Y^H$ on the space of its paths in ${\mathbb C}({\mathbb R}_+;{\mathbb R})$, and $\mu^B$ be the Wiener measure.
For $H>1/2$, the covariance function of $B^H(t)$ in \eqref{eqn-cov-FBM} is written as
\begin{align} \label{eqn-Psi-H}
\Psi^H(t,s) =\mathbb{E}\bigl[B^H(t) B^H(s)\bigr] = \int_0^t \int_0^s K^H(u,v){\rm d}u {\rm d}v,
\end{align}
where
\begin{align} \label{eqn-K-H}
K^H(t,s) = \frac{\partial^2}{\partial t \partial s} E\bigl[B^H(s) B^H(t)\bigr] = c_H |t-s|^{2H-2} .
\end{align}
with $c_{H}:= c^2 H(2H-1)$.
If $H> 3/4$, $K^H(\cdot,\cdot) \in L^2([0,T]^2)$, and $\mu^{Y,H}\sim \mu^B$ (equivalence) by the general criterion in Shepp \cite{shepp1966radon}, and in addition, Shepp's Radon--Nikodym derivative can be written in the form
\begin{align*}
\frac{d \mu^{Y,H}}{d \mu^B}(Y^H) = \exp \left( - \int_0^T \varphi_t(Y^H) {\rm d}Y^H(t) - \frac{1}{2} \int_0^T \varphi_t^2(Y^H) {\rm d}t\right) \, ,
\end{align*}
where
\begin{equation} \label{eqn-varphi-H}
\varphi_{t}(Y^{H}) \, :=\, \int^{t}_{0} L^{H}(s,t) {\mathrm d} Y^{H}(s) \,, \quad 0 \le t \le T\,
\end{equation}
and $L^H \in L^2([0,T]^2)$ is the unique solution of the Wiener--Hopf integral equation
\begin{align*}
L^H(s,t) + c_{H} \int_0^t L^H(r,t)|r-s|^{2H-2} {\rm d}r = - c_{H} |s-t|^{2H}, \quad 0 \le s \le t \le T\,.
\end{align*}
The second partial derivative $\,K(u, v ) \,$ of the covariance function $\Psi$ in \eqref{eqn-Psi} is given by
\begin{equation} \label{eq: lm4.1.1}
K(u, v) \, =\, \frac{\partial^{2} \Psi}{\partial u \partial v}(u,v ) =\, c^{2}(f_{1}(u\wedge v,u \vee v) + f_{2}(u \wedge v , u \vee v) ) \,,
\end{equation}
where
\begin{equation*}
\begin{split}
f_{1}(u,v) \, :=&\, \int^{u}_{0} (v-\theta)^{\alpha -1} (u-\theta)^{\alpha-1} \theta^{-\gamma} {\mathrm d} \theta \,, \\
f_{2}(u,v) \, :=& \, \int^{\infty}_{0} (v+\theta)^{\alpha-1}(u+\theta)^{\alpha-1} \theta^{-\gamma} {\mathrm d} \theta \,.
\end{split}
\end{equation*}
We then obtain the following square--integrability property of $K(u,v)$. Its proof is given in Section~\ref{sec-app-2}.
\begin{lemma} \label{lm:sq-integrability}
Assume $\, 0 < \gamma < 1 \,$ and $\, (\gamma -1 )/2 < \alpha < (\gamma +1)/2\,$. The function $\,K(\cdot, \cdot) \,$ in \eqref{eq: lm4.1.1} is square integrable with respect to the Lebesgue measure in $\,(0, T) \times (0, T) \,$ for every $\, T > 0 \,$, if and only if $\, \gamma/2 < \alpha <1/2+ \gamma /2\, $ and $\, 0 <\gamma <1 \,$ (equivalently, $H \in (1/2,1)$, regions (I) and (II)-1 in Figure~\ref{fig-1}(b)).
\end{lemma}
\begin{remark}
{\rm
We remark that both regions (I) and (II)-1 lead to Hurst parameter $H\in (1/2,1)$.
The previous lemma shows the integrability of the function $\,K(\cdot, \cdot) \,$ in \eqref{eq: lm4.1.1}, from which we can conclude the absolute continuity of the $Y$ with respect to the Brownian motion $\, \widetilde{B}(\cdot)\,$ and obtain an expression of the Radon-Nikodym density as a direct consequence of Shepp's result for general Gaussian processes in \cite{shepp1966radon}. That will involve the Wiener--Hopf integral equation.
However, the two regions (I) and (II)-1 also have distinct behaviors, despite the same Hurst parameter range. In particular, in region (I), we have shown in Proposition~\ref{prop-sm-X} that the process $X$ is of finite variation, with the representation in \eqref{eqn-X-regionI}, which makes $Y$ a Brownian with a random drift of finite variation. As a consequence, we are able to provide a more explicit expression of the Radon-Nikodym density using conditional expectations.
We will next state these results in two propositions. It remains open to show that the explicit Radon-Nikodym density in region (I) is equivalent to the density given by the Wiener--Hopf integral equation. }
\end{remark}
Suppose that $\gamma/2 < \alpha <1/2+ \gamma /2 $ and $\, 0 <\gamma <1 \,$. Let $\,L(s, t) \in L^{2}([0, T]^{2}) \,$ be the unique solution to the Wiener--Hopf integral equation
\begin{equation} \label{eqn-WH}
L(s, t) + \int^{t}_{0} L(r, t ) K(r, s ) {\mathrm d} r \, =\, - K ( s, t) \,, \quad 0 \le s \le t \le T \, ,
\end{equation}
and define
\begin{equation} \label{eq: phiY}
\, \varphi_{t} (Y) \, :=\, \int^{t}_{0} L(s, t) {\mathrm d} Y (s) \,, \quad 0 \le t \le T \, .
\end{equation}
Also, let $\, \ell (s, t) \in L^{2}([0, T]^{2}) \,$ be the unique solution to the Volterra equation
\begin{equation} \label{eqn-WH-V}
\ell (s, t) + \int^{t}_{s} \ell (r, t) L(s, r) {\mathrm d} r \, =\, L(s , t) \,, \quad 0 \le s \le t \le T \, .
\end{equation}
Thus, by \cite{shepp1966radon} and Lemma \ref{lm:sq-integrability}, we obtain the following proposition.
\begin{proposition
\label{prop-mixture}
Suppose that $\gamma/2 < \alpha <1/2+ \gamma /2 $ and $\, 0 <\gamma <1 \,$, i.e., regions (I) and (II)-1 in Figure~\ref{fig-1}(b) (both resulting $H\in (1/2,1)$).
The probability measure $\,\mu^{Y}\,$ induced by $\,Y\,$ in \eqref{eq: Y} is absolutely continuous with respect to the Wiener measure $\,\mu^{ \widetilde{B}}\,$ over $\,[0, T ] \,$ with the Radon--Nikodym density
\begin{equation} \label{eq: RND}
\frac{\,{\mathrm d} \mu^{Y}\,}{\,{\mathrm d} \mu^{ \widetilde{B} }\,} (Y) \, =\, \exp \bigg( - \int^{T}_{0} \varphi_{t} (Y) {\mathrm d} Y (t) - \frac{\,1\,}{\,2\,} \int^{T}_{0} [\varphi_{t}(Y)]^{2} {\mathrm d} t \bigg) \, .
\end{equation}
By the Girsanov theorem
\begin{align} \label{eq: Wbar}
\overline{W} (t) &\, :=\, Y(t) + \int^{t}_{0} \varphi_{s}(Y) {\mathrm d} s \nonumber\\
& \, =\, Y({t}) + \int^{t}_{0} \int^{s}_{0} L(r, s) {\mathrm d} Y(r) {\mathrm d} s \,, \quad 0 \le t \le T \,
\end{align}
is a Brownian motion with respect to its own filtration. Moreover, $\, Y (t) \,$ can be written as
\begin{equation*}
Y(t) \, =\, \overline{W}(t) - \int^{t}_{0} \int^{s}_{0} \ell (r, s ) {\mathrm d} \overline{W}(r) {\mathrm d} s \,, \quad 0 \le t \le T \, .
\end{equation*}
Particularly, the filtration $\,\mathcal F^{Y}(\cdot) \,$generated by $\,Y\,$ and the filtration $\,\mathcal F^{ \overline{W}} (\cdot) \,$ satisfy the identities $\, \mathcal F^{ Y} ({t} ) \, =\, \mathcal F^{ \overline{W}} ({t}) \,$ for $\, 0 \le t \le T \,$.
Therefore, $Y(t)$ is a semimartingale for the pair $(\alpha, \gamma)$ values in this region.
\end{proposition}
In Region (I) of Figure~\ref{fig-1}(b), since the process $X$ has a finite variation, as expressed in \eqref{eqn-X-regionI}, then the mixed process $Y$ is written as
\begin{equation} \label{eqn-wt-lambda}
\begin{split}
Y(t) \, =& \, \widetilde{B}(t) + \int^{t}_{0} \Big( \int^{r}_{-\infty} \Psi_{r}(s) {\mathrm d} B(s) \Big) {\mathrm d} r \, =\, \widetilde{B}(t) + \int^{t}_{0} \widetilde{\lambda}(r) {\mathrm d} r \,, \\
& \text{ with } \quad \widetilde\lambda(t) \,:= \, \int^{t}_{-\infty} \Psi_{t}(s) {\mathrm d} B(s)\,, \quad t \ge 0 \, ,
\end{split}
\end{equation}
where $ \Psi_{t}(s)$ given in \eqref{eq: Psi_{t}(s)}.
Thus $Y$ is, in fact, a Brownian motion with a random drift of finite variation, particularly, it is a semimartingale.
By Theorem 2 and Lemma 4 of \cite{MR281279}
we obtain the Radon-Nikodym density $
\frac{\,{\mathrm d} \mu^{Y}\,}{\,{\mathrm d} \mu^{ \widetilde{B} }\,} (Y)$ in \eqref{eq: RND}, which is stated in the following proposition.
\begin{proposition}
\label{prop-mixture-2}
Suppose that $1/2 < \alpha <1/2+ \gamma /2 $ and $\, 0 <\gamma <1 \,$, i.e., region (I) in Figure~\ref{fig-1}(b) (resulting in $H\in (1/2,1)$).
The Radon--Nikodym density \eqref{eq: RND} over the time interval $\,[0, T]\,$ in Proposition \ref{prop-mixture} is given by
\begin{equation*}
\frac{\,{\mathrm d} \mu^{Y}\,}{\,{\mathrm d} \mu^{ \widetilde{B} }\,} (Y) \, =\, \exp \bigg( \int_0^T \mathbb E [ \widetilde\lambda(s) \vert \mathcal F^{Y}(s) ] {\mathrm d} Y(s) - \frac{1}{2} \int_0^T \big(\mathbb E [ \widetilde\lambda(s) \vert \mathcal F^{Y}(s) ] \big)^2 {\mathrm d} s \bigg) \,.
\end{equation*}
Here, $\,\varphi_{t}(Y) \,$ in \eqref{eq: phiY} is identified as $\, \varphi_{t}(Y) \, \equiv \, - \mathbb E [ \widetilde\lambda(t) \vert \mathcal F^{Y}(t) ] \,$ for $\, t \ge 0 \,$, and
\[
\overline{W} (\cdot) \, =\, Y(\cdot) - \int^{\cdot}_{0} \mathbb E [ \widetilde\lambda(s) \vert \mathcal F^{Y}(s) ] {\mathrm d} s \, =\, \widetilde{B}(\cdot) + \int^{\cdot}_{0} \big( \widetilde{\lambda}(s) - \mathbb E [ \widetilde\lambda(s) \vert \mathcal F^{Y}(s) ] \big) {\mathrm d} s
\]
is a Brownian motion with respect to its own filtration.
\end{proposition}
\begin{remark} \label{rem-mixture}
{\rm
We conjecture that the mixture process $Y$ is not a semimartingale with respect to its own filtration in the parameter region $\, \gamma/2-1/2< \alpha\le \gamma/2\, $ and $\gamma \in (0,1)$ (region (II)-2 including the boundary line segment $\,\alpha \, =\, \gamma / 2 \, , \, \gamma \in (0, 1) \,$ in Figure~\ref{fig-1}(b)). The boundary point $\,\gamma \, =\, \alpha = 0 \,$ represents the standard Brownian motion $\,B^{H}\,$ with $\, H \, =\, 1/ 2\,$ (written as $B^{1/2}$ without confusion) and $\,Y \, =\, \widetilde{B} + B^{1/2}\,$ becomes a semimartingale.
For standard FBM $B^H$, in Cai et al. \cite{cai2016mixed}, representations of the FBM with the Riemann-Liouville fractional integrals and derivatives are used to prove the innovation representations in Theorem 2.4 for $H<1/2$, and equivalence of the measures for $\widetilde{B}+ B^H$ and $B^H$ for $H<1/4$. However, for the GFBM $X$, it still remains open to establish the Riemann-Liouville fractional integrals and derivatives. Therefore, we leave it as future work to prove the non-semimartingale property of the mixture process $Y$ for the parameter pair $(\alpha, \gamma)$ in region (II)-2 of Figure~\ref{fig-1}(b). }
\end{remark}
\subsection{Generalized Riemann-Liouville FBM and its Mixture}
In this section, we discuss the semimartingale properties of the generalized Riemann--Liouville (R-L) FBM and its mixtures.
The generalized R-L FBM is introduced in Remark 5.1 in \cite{pang-taqqu}, and further studied in Section 2.2 in \cite{ichiba-pang-taqqu}.
It is defined by
\begin{equation} \label{eqn-X-RL}
X(t) = c \int_0^t (t-u)^\alpha u^{-\gamma/2} B({\rm d}u)\, , \quad t \ge 0\, ,
\end{equation}
where $B({\rm d}u)$ is a Gaussian random measure on ${\mathbb R}$ with the Lebesgue control measure ${\rm d}u$ and
$c \in {\mathbb R}$, $\gamma \in [0,1)$ and $\alpha \in (\gamma/2-1/2, \gamma/2+1/2)$.
It is a continuous self-similar Gaussian process with Hurst parameter $H=\alpha-\gamma/2+1/2 \in (0,1)$. When $\gamma=0$, it reduces to the standard R-L FBM
$$
B^H(t) = c \int_0^t (t-u)^\alpha B({\rm d} u), \quad t\ge 0\, .
$$
It is clear that the semimartingale properties in Proposition \ref{prop-sm-X} hold for the process $X$ in \eqref{eqn-X-RL}.
In particular, by letting the natural kernel $K_t(s) := (t-s)^\alpha s^{-\gamma/2}$, we have the spectral representation
$
X(t) \stackrel{\rm d}{=} \int_0^t K_t(s) N({\rm d}s)\, ,
$
for a Gaussian measure $N(\cdot)$.
Define
$$
\Psi_t(s) = C_t^{-1} \alpha (t-s)^{\alpha-1} s^{-\gamma/2}\, ,
$$
for a time-dependent normalization constant $C_t$. As shown in the proof of Lemma 3.1, the function $\Psi_1(\cdot)$ is square integrable with respect to the Lebesgue measure if and only if $1/2< \alpha< 1/2 + \gamma/2$ and $\gamma \in (0,1)$, and thus by Basse's characterization of the spectral representation of Gaussian semimartingales (\cite[Theorem 4.6]{basse2009spectral}), we can conclude the the semimartingale property in part (ii) of Proposition \ref{prop-sm-X}. The non-semimartingale property in part (i) of Proposition \ref{prop-sm-X} also follows from a similar argument as in the proof of the proposition. In addition, the variation properties in Proposition \ref{prop: variation} also hold for the process $X$ in \eqref{eqn-X-RL}, from which we can also conclude the non-semimartingale property as in the proof of Proposition \ref{prop-sm-X-2}.
We next discuss the mixed process $Y= \widetilde{B}+X$ with $X$ in \eqref{eqn-X-RL} and an independent BM $\widetilde{B}$.
Let $L(s,t) \in L^2([0,T]^2)$ be the unique solution to the Wiener--Hopf integral equation \eqref{eqn-WH} with
$$
K(u,v) = c^2\int^{u}_{0} (v-\theta)^{\alpha -1} (u-\theta)^{\alpha-1} \theta^{-\gamma} {\mathrm d} \theta\,,
$$
and let $\ell(s,t)\in L^2([0,T]^2)$ be the unique solution to the Volterra equation \eqref{eqn-WH-V}. Then the properties in Proposition \ref{prop-mixture} hold, in particular, $Y=\widetilde{B}+X$ with $X$ in \eqref{eqn-X-RL} is a semimartingale with respect to the filtration generated by itself in the parameter range $\gamma/2 < \alpha <1/2+\gamma/2$ and $0< \gamma<1$.
In addition, Proposition \ref{prop-mixture-2} holds for $\gamma/2 < \alpha <1/2+\gamma/2$ and $0< \gamma<1$, where $\widetilde\lambda(t) = \int^{t}_{0} \Psi_{t}(s) {\mathrm d} B(s)$ in \eqref{eqn-wt-lambda}.
\section{Stock Price Processes associated with GFBM} \label{sec-finance-price}
\subsection{Shot noise process and integrated shot noise} \label{sec-SNP-limit}
In modeling of financial markets, Brownian motion and compound Poisson processes or more generally, L\'evy processes are widely utilized to capture effects of various noises.
Recently, Pang and Taqqu \cite{pang-taqqu} studied a non-stationary, power-law shot noise process $\mathcal Z^*= \{\mathcal Z^*(y): y \in {\mathbb R}\}$ on the whole real line defined by
\begin{equation} \label{eq: Zast}
\mathcal Z^*(y) := \sum_{j=-\infty}^\infty g^*(y-\tau_j) R_j\, , \quad y \in {\mathbb R}\, ,
\end{equation}
where $\{\tau_j: j \in {\mathbb Z}\}$ is a sequence of Poisson arrival times of shots with rate $\lambda$ on the whole real line, and each $R_j$ is the noise associated with shot $j$ at time $\tau_j$ for $\, j \in \mathbb Z\, $. It is a generalization of the compound Poisson process on the positive half line. The variables $\{R_j, j \in {\mathbb Z}\}$ are conditionally independent, given $\{\tau_j\}$, and the marginal distribution of $R_j$ depends on the shot arrival time $\tau_j$, that is, $ \mathbb P(R_j \le r \, |\, \tau_j=u) =: F_u(r)$, $\,r \, \in\, \mathbb R\,$, $\, u \, \in \, \mathbb R\,$ for every $\,j \in \mathbb Z\,$. Assume that
\begin{itemize} \item[(i)] (the power-law property) the function $\,g^{\ast} \,$ satisfies $g^*(y) = y^{-(1-\alpha)} L^*(y)$ for $y\ge 0$ and $g^*(y)=0$ for $y <0$ and $\alpha \in (0, 1/2)$, where $L^*$ is a positive slowly varying function at $+\infty$, and
\item[(ii)] (the moment conditions) the common conditional distribution $\,F_{t} \,$ of the noises $R_\cdot$, given $\,\tau_{\cdot} = t\,$, satisfies the zero mean $K_1(t) := \int_{\mathbb R} r {\rm d} F_t(r) =0$ for every $\, u \in \mathbb R\,$ and finite variance $K_2(t) = \int_{\mathbb R} r^2 {\rm d} F_t(r) = t^{-\gamma} \tilde{L}_+(t)$ for $t>0$ and $K_2(t) = \lvert t \rvert ^{-\gamma} \tilde{L}_{-}(t)$ for $t<0$, where $\gamma \in (0,1)$, and $\tilde{L}_{\pm}$ are some positive slowly varying functions.
\end{itemize}
The shot noise process $\, \mathcal Z^{\ast}\,$ in \eqref{eq: Zast} is a generalization of the compound Poisson process, because if $\, g^{\ast}(y) \, :=\, {\bf 1}_{\{y > 0 \}} \,$, $\, y \in \mathbb R\,$, then $\, \mathcal Z^{\ast} \,$ is a compound process.
The integrated shot noise process $\mathcal Z = \{\mathcal Z(t): t \in {\mathbb R}_+\}$ is defined by
\begin{equation} \label{eq: Z}
\mathcal Z(t) := \int^{t}_{0} \mathcal Z^{\ast}(y) {\mathrm d} y \, =\, \sum_{j=-\infty}^\infty (g(t-\tau_j) - g(-\tau_j)) R_j\, , \quad t \ge 0\, ,
\end{equation}
where the shot shape function $\,g(\cdot) \,$ is differentiable with its derivative $\,g^{\ast} \,$, i.e., $g(t) := \int_0^t g^*(y){\rm d}y$ for $\,t \ge 0 \,$.
\subsection{Stock price models driven by the noise with the long-range dependence} \label{sec-price}
Stock pricing models with a shot-noice component have been developed to study credit and insurance risks \cite{altmann2008shot,schmidt2017shot,Stute}.
In particular, the stock price $P(t)$ is modeled as
\begin{equation} \label{eq: P1}
P(t) := P(0) \exp \left( \Big( \mu -\frac{\sigma^2}{2} \Big) t + \sigma \widetilde{B}(t) + \sigma\int_0^t \sum_{\tau_i \le s } \mathfrak f (s-\tau_i, R_i) {\mathrm d} s \right)\,,\end{equation}
for $t\ge 0$,
where $\{(\tau_i, R_i): i \in {\mathbb N}\}$ is a marked point process, independent of the Brownian motion $\widetilde{B}$, with arrival times $\tau_i$ and marks (noises) $U_i \in {\mathbb R}^d$, and the function $\mathfrak f: {\mathbb R}_+\times {\mathbb R}^d \to {\mathbb R}$ is the deterministic shot shape function. $\, \mu \in \mathbb R \,$ and $\,\sigma > 0 \,$ are some real constants. Equivalent martingale measures for this price process are studied in \cite{schmidt2017shot,schmidt2007shot}. In \cite{schmidt2017shot}, it is also discussed when the shot-noise component is Markovian or a semimartingale.
This is usually when the function $\,\mathfrak f\,$ takes a particular form (exponential function for the Markovian property).
In these studies, the noises $\{R_i\}$ are assumed to be i.i.d. with finite variance. Since it is usually more difficult to work with the shot noise process directly, one may use the diffusion approximations. For example, Kl{\"u}ppelberg and K{\"u}hn \cite{kluppelberg2004fractional} showed that under regular variation conditions, a Poisson shot noise process can be approximated by an FBM (under proper scaling and validated by a functional central limit theorem), and then used the limiting FBM as a stock pricing model.
Here, as a pre-limit, we consider the usual random walk noise and the shock noises on the stock price. We evaluate the effects of these noises, when the frequency of arrival of shot noises is very high with appropriate scaling.
Given a scaling parameter $\, \varepsilon > 0 \,$, we model a pre-limit of price process by
\begin{align} \label{eq: Pepsilon}
P_{\varepsilon}(t) \, :=\, & P ( 0) \exp \left( \Big( \mu -\frac{\sigma^2}{2} \Big) t + \sigma \varepsilon^{1/2} \sum_{j=1}^{ \lfloor t/ \varepsilon \rfloor } \xi_{j} + \sigma\int_0^t \frac{\,1\,}{\,\varepsilon ^{1-H}\,}\mathcal Z^{\ast} \Big( \frac{\, u\,}{\,\varepsilon \,}\Big) {\mathrm d} u \right) \nonumber \\
\, =\, & P ( 0) \exp \left( \Big( \mu -\frac{\sigma^2}{2} \Big) t +\sigma \varepsilon^{1/2} \sum_{j=1}^{ \lfloor t/ \varepsilon \rfloor } \xi_{j} + \sigma \varepsilon^{H}\int_0^{t/\varepsilon} \mathcal Z^{\ast} ( {\, u\,}) {\mathrm d} u \right) \nonumber \\
\, =\, &P ( 0) \exp \left( \Big( \mu -\frac{\sigma^2}{2} \Big) t + \sigma\varepsilon^{1/2} \sum_{j=1}^{ \lfloor t/ \varepsilon \rfloor } \xi_{j} + \sigma \varepsilon^{H} \mathcal Z \Big( \frac{\,t\,}{\,\varepsilon\,} \Big) \right ) \,,
\end{align}
for $t\ge 0$,
where $\, \{ \xi_{j}, j \in \mathbb N\}\,$ are i.i.d. random variables with zero mean and unit variance, independent of the shot noise $\,\mathcal Z^{\ast}\,$ in \eqref{eq: Zast}, and $\, \mathcal Z \,$ is the integrated shot noise process in \eqref{eq: Z}. Here, $\, \mu \,$ and $\,\sigma > 0 \,$ are some real constants. (One may also choose a model without the random walk component, in which case the model in \eqref{eq: price with shot noise} will have only the process $X$ instead of the mixed GFBM.)
Under certain regularity conditions and with a proper scaling, Pang and Taqqu \cite{pang-taqqu} have shown that the scaled process $\widehat{\mathcal Z}^\varepsilon(t) := \varepsilon ^{H} \mathcal Z(t/\varepsilon) $ converge weakly to the GFBM $X$, as $\,\varepsilon \to 0 \,$. The random walk term $\, \varepsilon^{1/2} \sum_{j=1}^{ \lfloor t / \varepsilon \rfloor} \xi_{j}\,$, $\, t \ge 0 \,$ converges weakly to the standard BM, independent of $\, X \,$.
Suppose that the parameters $\,(\alpha, \gamma)\,$ are in the semimartingale region: $\gamma/2 < \alpha <1/2+ \gamma /2 $ and $\, 0 <\gamma <1 \,$ (i.e., $H\in (1/2,1)$ for the GFBM $\, X\,$), as in the assumption of Proposition \ref{prop-mixture}. As a scaling limit of \eqref{eq: Pepsilon}, we propose a stock price model using the mixed GFBM as follows:
\begin{equation} \label{eq: price with shot noise}
\begin{split}
P(t) \, = &\, P(0) \exp \left( \Big( \mu -\frac{\sigma^2}{2} \Big) t + \sigma(\widetilde{B}(t) + X(t)) \right) \\
\, = & P(0) \exp \left( \Big( \mu -\frac{\sigma^2}{2} \Big) t + \sigma Y(t) \right) \,
, \quad t\ge 0 \, ,
\end{split}
\end{equation}
where $\, Y (\cdot) \, =\, \widetilde{B} + X \,$ is the mixed GFBM in \eqref{eq: Y} with $\,X\,$ in \eqref{eqn-X-RL}
(For a recent account of weak convergence in financial models, we refer to Kreps~\cite{kreps2019black}.)
Under the above parameter range, $\, Y \,$ is a semimartingale. The price dynamics is determined as the unique strong solution of the linear stochastic differential equation
\begin{equation} \label{eq: price with shot noise2}
{\mathrm d} P(t) \, =\, P (t) ( \mu {\mathrm d} t + \sigma {\mathrm d} Y(t) ) \, ; \quad t \ge 0 \, ,
\end{equation}
driven by the semimartingale $\, Y\,$, where $\, \mu \,$ is a drift and $\, \sigma \,$ is volatility of stock price under a filtered probability space $\, (\Omega, \mathcal F, (\mathcal F_{t}) , \mathbb P )$. Thus, we take the filtration $\, \mathbb F \, :=\, (\mathcal F_{t}, t \ge 0 ) \, =\, (\mathcal F^{Y}(t), t \ge 0 ) \,$.
As an example with this stock price model \eqref{eq: price with shot noise}, we consider an investor who trades this stock with price \eqref{eq: price with shot noise} and money market account with an instantaneous interest rate $r (> 0)$.
We first recall what is known in the case of standard FBM $B^H$, as shown in \cite{rogers1997arbitrage,cheridito2001mixed,cai2016mixed}, the mixed process $ Y^H=\widetilde{B}+ B^H$ is a semimartingale if and only if $H =1/2$ (the Brownian case) and $H \in (3/4,1)$. Of course, with a BM, i.e., $\,H \, =\, 1/2\,$, the standard results of stock pricing and equivalence of martingale measure can be applied. On the other hand, with $H \in (3/4,1)$, one also obtain the Radon--Nikodym derivative in \eqref{eqn-dQdP} where $Y$ is replaced by $Y^H$, and the function $\, \varphi_{t}(Y)\,$ is replaced by $\varphi_{t}(Y^{H})$ in \eqref{eqn-varphi-H}.
Also, recall that for the GFBM $X$ with $H=1/2$, the parameters $(\alpha, \gamma)$ lies on the line segment $\alpha=\gamma/2$ in Figure \ref{fig-1}. It is a semimartingale, only when $\gamma=0$, which becomes the special case of Black-Shocles pricing model; otherwise, there does not exist an equivalent martingale measure.
In the following proposition, we give the expressions of the equivalent martingale measures for the discounted stock price process modeled with the mixed process $Y$.
\begin{proposition}\label{prop-pricing-RN}
Assume $\gamma/2 < \alpha <1/2+ \gamma /2 $ and $\, 0 <\gamma <1 \,$ (regions (I) and (II)-1 in Figure~\ref{fig-1}(b)). Under the stock price process dynamics \eqref{eq: price with shot noise2}, the discounted stock price process $\, e^{-r t} P(t) \,$, $\, 0 \le t \le T\,$ is a martingale under the new measure $\, \mathbb Q\,$ defined by
\begin{equation} \label{eqn-dQdP}
\frac{\,{\mathrm d} \mathbb Q\,}{\, {\mathrm d} \mathbb P \,} \bigg \vert_{\mathcal F_{T}}\, :=\, \exp \bigg( - \int^{T}_{0} (\theta - \varphi_{t}(Y) ) {\mathrm d} Y(t) - \frac{\,1\,}{\,2\,} \int^{T}_{0} ( \theta^{2} - \lvert \varphi_{t}(Y) \rvert^{2}) {\mathrm d} t \bigg) \, ,
\end{equation}
where $\, \theta \, :=\, (\mu - r ) / \sigma \,$ is the market price of risk and $\, \varphi_{\cdot}(Y)\,$ is defined in \eqref{eq: phiY}. Particularly, if $1/2 < \alpha <1/2+ \gamma /2 $ and $\, 0 <\gamma <1 \,$, i.e., region (I) in Figure~\ref{fig-1}(b), then $\, \varphi_{\cdot}(Y) \, =\, - \mathbb E [ \widetilde{\lambda} (\cdot) \, \vert \, \mathcal F^{Y}(\cdot) ] \, =\, - \mathbb E [ \widetilde{\lambda} (\cdot) \, \vert \, \mathcal F_{\cdot} ] \,$,
where
$ \widetilde\lambda(t)= \int^{t}_{0} \Psi_{t}(s) {\mathrm d} B(s)$ for $t\ge 0$.
\end{proposition}
\begin{proof}
It follows from Proposition \ref{prop-mixture} and the Girsanov theorem that the process $\, \overline{W} (\cdot) \, =\, Y(\cdot) + \int^{\cdot}_{0} \varphi_{s}(Y) {\mathrm d} s \,$ in \eqref{eq: Wbar}
is a Brownian motion for $\,0 \le t \le T\,$ under $\, \mathbb P\,$.
By the simple application of the product rule to \eqref{eq: price with shot noise2}, we have the discounted stock price process
\begin{equation*}
\begin{split}
e^{-r t} P (t) \, &=\, P(0) + \int^{t}_{0} \sigma e^{-r s} P(s) {\mathrm d} \Big( Y(s) + \frac{\,\mu - r \,}{\,\sigma\,} s \Big) \\
& \, =\, P(0) + \int^{t}_{0} \sigma e^{-r s} P(s) {\mathrm d} \Big( \overline{W}(s) - \int^{s}_{0} \varphi_{u}(Y) {\mathrm d}u + \theta s \Big) \,,
\end{split}
\end{equation*}
for $t\ge 0$,
with $\, \theta \, :=\, (\mu - r) \, / \, \sigma \,$. By another application of the Girsanov theorem, $\, Y(t) + \theta t \,$, $\, 0 \le t \le T \,$ is a Brownian motion under the measure $\, \mathbb Q \,$. In particular, the discounted price process $\, e^{-r t } P(t) \,$, $\, 0 \le t \le T\,$ is a martingale under $\,\mathbb Q\,$.
If $1/2 < \alpha <1/2+ \gamma /2 $ and $\, 0 <\gamma <1 \,$, i.e., region (I) in Figure~\ref{fig-1}(b), then $\, \varphi_{\cdot}(Y) \, =\, - \mathbb E [ \widetilde{\lambda} (\cdot) \, \vert \, \mathcal F^{Y}(\cdot) ] \, =\, - \mathbb E [ \widetilde{\lambda} (\cdot) \, \vert \, \mathcal F_{\cdot} ] \,$, because $\, \mathbb F^{Y} \equiv \mathbb F^{P} \, =\, \mathbb F\,$.
\hfill $\,\square\,$
\end{proof}
Consequently, the time-$\,t\,$ price of European option on this stock with payoff function $\, \mathfrak g \,$ and with maturity $\, T\,$ is given by
\begin{equation*}
\mathbb E ^{\mathbb Q} [ e^{-r (T-t)}\, \mathfrak g(P(T)) \, \vert \, \mathcal F_{t}] \, =\, \mathbb E ^{\mathbb P } \bigg[ e^{-r (T-t)} \, \mathfrak g(P(T)) \cdot \frac{\, {\mathrm d} \mathbb Q\,}{\, {\mathrm d} \mathbb P \,} \Big \vert _{\mathcal F_{T}} \, \, \bigg \vert \mathcal F_{t} \bigg] \, ,
\end{equation*}
where the conditional expectations $\, \mathbb E^{\mathbb Q}\,$ and $\, \mathbb E^{\mathbb P}\,$ are calculated under $\, \mathbb Q\,$ and $\, \mathbb P\,$, respectively, given the filtration $\, \{\mathcal F_{t}\}_{t\ge 0}\,$.
Under the stock price model \eqref{eq: price with shot noise2} with the mixed GFBM $\,Y\,$ in the regions (I) and (II)-1 of the parameter sets $\{ \, \gamma / 2 < \alpha < 1/2 + \gamma /2 \,, \, 0 < \gamma < 1 \,\}$,
the pricing and hedging problems of various options (such as American and Asian options) are solved under the measure $\mathbb Q$ in the same way as in the standard Black-Scholes model. Thus, the Black-Scholes pricing formula is still valid under the long range dependence property of the driving mixed GFBM $\,Y\,$.
\subsection{Comments on Arbitrage} \label{sec-arbitrage}
In the theory of asset pricing, the ``First Fundamental Asset Pricing Theorem" requires the existence of an equivalent martingale measure for no arbitrage, and works in the framework of semimartingales for pricing models. For stock price models with FBM $B^H$, since $B^H$ is a semimartingale if and only if $H=1/2$ \cite{liptser2012theory,rogers1997arbitrage}, arbitrage strategies have been discussed in both fractional Bachelier and Black-Scholes models \cite{rogers1997arbitrage,shiryaev1998arbitrage,salopek1998tolerance,cheridito2003arbitrage}.
In particular, Rogers \cite{rogers1997arbitrage} constructed arbitrage for the fractional Bachelier model: a market with a money account $\xi_t =1$ with zero interest rate and a risky stock (no dividends or transaction costs) with price $\tilde{P}(t) = \tilde{P}(0) +\nu t+ \sigma B^H(t)$, for $H \in (0,1/2) \cup (1/2, 1)$ for $\, t \ge 0 \,$.
For the GFBM $X$ in \eqref{def-X}, the only parameter set that guarantees no-arbitrage is $\alpha=0$ and $\gamma=0$ (i.e., $H=1/2$), that is, the standard Brownian motion case. Recall Remark \ref{rem-sm}, $H=1/2$ corresponds to the line $\alpha=\gamma/2$ in Figure \ref{fig-1}, but if $\gamma\neq 0$, it is a ``fake" Brownian motion and it may lead to an arbitrage, because of the non-semimartingale property.
Although we have shown that $\,X\,$ is a semimartingale with respect to the filtration $\mathcal{F}^B(\cdot)$ for $\alpha \in (1/2, 1/2+\gamma/2)$ and $\gamma \in (0,1)$ (resulting in $H \in (1/2,1)$ in region (I)), it is a process of finite variation. As it is shown in \cite{HPS84}, in a frictionless market with continuous trading, arbitrage opportunities are present without unbounded variation of the stock price process. Thus, the differentiable sample path in Proposition \ref{prop-sm-X} leads an arbitrage opportunity in this parameter range.
On the other hand, for the mixed BM and GFBM $Y$ in \eqref{eq: Y}, because in the parameter range $\alpha \in (1/2, 1/2+\gamma/2)$ and $\gamma \in (0,1)$ (resulting in $H \in (1/2,1)$ for $\, X\,$ in region (I)), the process $Y$ becomes a Brownian motion with a random drift of finite variation, there exists an equivalent martingale measure and hence, it forbids arbitrage.
We next discuss the construction of arbitrage opportunities.
For the GFBM $X$, with $\alpha \in (\gamma/2, 1/2)$ and $\gamma \in (0,1)$ (resulting in $H\in (1/2,1)$), one can construct arbitrage strategy using an approach similar that of Shiryaev \cite{shiryaev1998arbitrage} since the $p$-variation of the process $X$ is finite for $p>1/H$ for $H \in (1/2,1)$ and well defined, so that pathwise Riemann-Stieltjes stochastic integral with respect to $X$ is well defined \cite{shiryaev1998arbitrage,salopek1998tolerance}. (See further discussions on stochastic integrals with respect to the GFBM $X$ in Section \ref{sec-stochastic integral}.)
Indeed, replacing the FBM $\,B^{H}\,$ in \cite{shiryaev1998arbitrage} by the GFBM $\,X\,$ in \eqref{def-X}, we find that the self-financing, admissible strategy $\, {{\mathbf m} \pi} \, :=\, ( {{\mathbf m} \beta}, {{\mathbf m} \gamma}) \,$ in the Bachelier model $\, \widetilde{P}(t) \, =\, \widetilde{P}_{0} + X(t) \,$, $\,t \ge 0 \,$ with a risk free bond of zero interest rate yields the portfolio value
\begin{equation} \label{eqn-PV-X}
\mathcal V_{t}^{\pi} \, =\, {{\mathbf m} \beta}_{t} + {{\mathbf m} \gamma}_{t} \widetilde{P}(t) \, =\, \mathcal V_{0}^{\pi} + \int^{t}_{0} {{\mathbf m} \gamma}_{u} {\mathrm d} \widetilde{P}(u) \,, \quad t \ge 0 \, ,
\end{equation}
where the stochastic integral is understood as the pathwise Riemann-Stieltjes integral with respect to $\,X\,$,
and then with $\, \widetilde{P}_{0} \, :=\, 1\,$, $\, {{\mathbf m} \gamma}_{t} \, :=\, 2 X(t)\,$, $\, {{\mathbf m} \beta}_{t} \, :=\, - X(t)^{2} - 2 X(t)\,$, $\,t \ge 0 \,$, the resulting portfolio value satisfies $\, \mathbb P ( \mathcal V_{0}^{\pi} \, =\, 0 \, , \, \mathcal V_{t}^{\pi} > 0 ) \, =\, 1 \,$ for every $\, t > 0 \,$, because $\, \mathcal V_{t} ^{\pi} \, =\, {{\mathbf m} \beta}_{t} + {{\mathbf m} \gamma}_{t} \widetilde{P}(t) \, =\, X(t)^{2} \,$, $\, t \ge 0 \,$.
Similarly, in the fractional Black-Scholes model, which is a market with a money market account $\xi_t =e^{rt}$ and a risky stock (no dividends or transaction costs) $\widetilde{P}(t) = \widetilde{P}_0\exp\big( r t+ B^H(t)\big)$, one can also adopt the self-finance arbitrage strategy for $H \in (1/2,1)$ for standard FBM in Shiryaev \cite{shiryaev1998arbitrage}.
For the GFBM $X$, with $\widetilde{P}(t) = \widetilde{P}_0\exp\big( rt+ X(t)\big)$, we can analogously choose the portfolio ${{\mathbf m} \beta}_t = 1- e^{2X(t)}$ and ${{\mathbf m} \gamma}_t = 2 (e^{X(t)}-1)$ for a self-financing arbitrage opportunity with the portfolio value process $\, \mathcal V^{\pi}_{t}\,$, $\, t \ge 0 \,,$ defined by
\begin{equation} \label{eqn-PV-X-mix}
\mathcal V^{\pi}_{t} \, :=\, {{\mathbf m} \beta}_{t} e^{r t} + {{\mathbf m} \gamma}_{t} \widetilde{P}(t) \, =\, e^{r t } ( e^{X(t)} - 1)^{2} \, =\, \int^{t}_{0} {{\mathbf m} \beta}_{s} r {\mathrm d} s + {{\mathbf m} \gamma}_{s} {\mathrm d} \widetilde{P}(s) \,, \quad t \ge 0 \, .
\end{equation}
However, for $\alpha \in (\gamma/2-1/2, \gamma/2]$
and $\gamma \in (0,1)$ (resulting in $H\in (0,1/2)$), we cannot use the approach in Shiryaev \cite{shiryaev1998arbitrage} to construct arbitrage strategies. Potentially, one could try to construct arbitrage strategies by extending the approaches in \cite{cheridito2003arbitrage}. We leave this to future work.
For the process $Y$ in \eqref{eq: Y}, with $\alpha \in (\gamma/2, 1/2)$ and $\gamma \in (0,1)$ (resulting in $H\in (1/2,1)$ for $\, X\,$ in region (II)-1), one can construct arbitrage strategy using an approach similar that of Shiryaev \cite{shiryaev1998arbitrage}.
The construction of an arbitrage opportunity when the parameter is in the region (II)-2 is also an open problem.
\section{Rough Fractional Stochastic Volatility} \label{sec-roughvol}
In a seminal paper Gatheral et al. \cite{gatheral2018volatility} conducted an empirical study on
stochastic volatility and discovered that the log-volatility behaves like a FBM with Hurst exponent $H<1/2$ (mostly between 0.08 and 0.2), and thus proposed a ``rough" volatility model. For the recent developments in the empirical studies of the Hurst parameter of financial data and the microstructure of leverage effects, see also \cite{alos2019estimating,omar2016microstructural,fukasawa2020volatility} and papers listed on the webpage \cite{roughvolatility-webpage}.
For a given asset with log-price taking the form
\begin{equation} \label{eqn-log-price}
\frac{{\rm d} P(t)}{P(t)} = \mu(t) {\rm d} t + \sigma(t){\rm d} W(t)\, , \quad t \ge 0 \, ,
\end{equation}
where $\mu(t)$ is a drift term and $W(t)$ is a one-dimensional BM, the stochastic volatility $\sigma(t)$ is modeled as
$$
\sigma(t) = \sigma(0) \exp(Z(t))\, , \quad t\ge 0\, ,
$$
where the process $Z(t)$ is a stationary fractional Ornsten-Uhlenbeck (fOU) process, given by the stationary solution of the SDE:
$$
dZ(t) = - a (Z(t) - m) {\rm d} t + \nu {\rm d} B^H(t)
$$
with $m \in {\mathbb R}$ and $a$ and $\nu$ being positive constant parameters.
Here the fOU process $Z$ has an explicit solution given by
$$
Z(t) = Z(0) e^{-a t } + m ( 1- e^{-a t }) + \nu \int_{-\infty}^t e^{-a (t-s)} {\rm d} B^H(s) \, ,
$$
where the stochastic integral with respect to the FBM $B^H$ is a pathwise Reimann-Stieltjes integral (see \cite{cheridito2003fractional}). This model has received great attention in the community of stochastic volatility (see, e.g., \cite{bayer2016pricing} and an extensive relevant literature in \cite{roughvolatility-webpage}).
In addition to the estimation of the Hurst parameter being $H<1/2$, there are several important findings in the empirical study of \cite{gatheral2018volatility}. First, the distribution of the increments of the log-volatility is approximately normal. Second, although the log-volatility estimations are smooth over most intervals, there is observable non-smoothness for some stock prices and indices and for certain time windows.
It is also argued that the reason why the above model of stochastic volatility is used, instead of $\sigma(t) = \sigma \exp(\nu B^H(t))$, is because the above model is stationary. Although mathematical tractability is desirable with the stationary model, it is evident that non-stationarity (in terms of increments) is prominent in financial data (see also \cite{bassler2007nonstationary,mccauley2008martingales}).
In this section we first study a generalization of the rough Bergomi model introduced in \cite{bayer2016pricing}, and then discuss other extensions.
\subsection{A generalized rough Bergomi model}
In \cite{bayer2016pricing}, the rough Bergomi (rBergomi) model was introduced as a non-Markovian generalization of Bergomi model with FBM $B^H$. Specifically, the stock price process $P(u)$ and the instantaneous volatility $v(u)$, $\,u \ge t\,$ under the physical measure are defined by
\begin{equation} \label{eq: rBergomi}
\begin{split}
& \frac{{\rm d} P(u)}{P(u)} = \mu (u) {\mathrm d} u + \sqrt{v(u)}{\rm d} \widetilde{B}(u), \\
&v(u) = \xi(t) \exp \left( \eta \sqrt{2H} \int_t^u (u-s)^{H-1/2} {\rm d} B(s) - \frac{1}{2}\eta^2 (u-t)^{2H}\right) \,,
\end{split}
\end{equation}
for $u\ge t$,
where $\,\mu \,$ is an expected log return process, $\,\eta\,$ is a constant, $\xi(t) = \mathbb{E}[v(u) \, \vert \, \mathcal F (t)]$, $\,u \ge t \,$ is the forward variance curve, $H\in (0,1/2)$, and $\widetilde{B} = \rho B + \sqrt{1-\rho^2} B^{\perp}$ for two independent standard Brownian motions $B, B^{\perp}$ with correlation coefficient $\, \rho \in (-1, 1)\,$. Here, the filtration $\,\mathcal F (t) \,$, $\, t \ge 0 \,$ is generated by the price process $\, P(t) \,$, $\,t \ge 0 \,$, and the process $ \int_t^u (u-s)^{H-1/2} {\mathrm d} B(s)$ is the so-called Riemann-Liouville FBM or Volterra fractional Brownian motion. In \cite{bayer2016pricing}, it was discussed as a first approximation that $P(u)$ becomes a true martingale by the deterministic change of measure under a fixed time horizon $\, t \le u \le T\,$ in the rough Bergomi model under the equivalent martingale measure. For the details of the martingale property of the rough Bergomi model, see \cite{gassiat2019martingale}.
Recently, the rough Bergomi model is studied in the limiting case $\,H \to 0 \,$ in \cite{forde2020rough}.
Using the generalized Riemann-Liouville FBM $X(t)$ in \eqref{eqn-X-RL}, we may modify the above model with replacement of $\,v \,$ in \eqref{eq: rBergomi} by
\begin{align}\label{eq: GrBergomi}
v(u) &= \xi(t) \exp \bigg( \eta (X(u)- X(t)) - \frac{1}{2}\eta^2 \mathbb E [ \lvert X(u) - X(t) \rvert^{2} ]
\bigg) \\
&= \xi(t) \exp \bigg( \eta c \int_t^u (u-s)^\alpha s^{-\gamma/2} B({\rm d}s) - \frac{1}{2}\eta^2 c^{2} \int^{u}_{t} (u-s)^{2\alpha} s^{-\gamma} {\mathrm d} s
\bigg) \,, \nonumber
\end{align}
for $u \ge t$, where $\alpha \in (-(1-\gamma)/2,\gamma/2)$, $\gamma \in (0,1)$ and $H= \alpha - \gamma/2+ 1/2 \in (0,1/2)$. Here, $\, c\,$ is the normalizing constant in \eqref{eq: normalizeC}.
We conduct the following analysis on the approximation of variance of VIX futures as in Section 6 of \cite{bayer2016pricing}.
Let $\sqrt{\varsigma(T)}$ be the terminal value of the VIX futures, where
\begin{equation} \label{eqn-varsigmaT}
\varsigma(T) = \frac{1}{\Delta} \int_T^{T+\Delta} \mathbb{E}[v(u) \, \vert \, \mathcal F (T)] du \,,
\end{equation}
where $\{\mathcal F (t)$, $t\ge 0\}$ is the filtration generated by the process $P(t)$ with $v(u)$ in \eqref{eq: GrBergomi} using the process $X(t)$ in \eqref{eqn-X-RL}, and $\Delta$ is the period under consideration (e.g., a month).
Similar to Section 3.1 in \cite{jacquier2018vix}, we obtain the following formula of forward variance curve: for $u \ge T >t$,
\begin{align*}
&\mathbb{E}[v(u) \, \vert \, \mathcal F (T)] = \xi(t) \exp \bigg( \eta c \int_t^T (u-s)^\alpha s^{-\gamma/2} B({\rm d}s) \bigg) \\
& \quad \times\mathbb{E} \left[ \exp\bigg( \eta c \int_T^u (u-s)^\alpha s^{-\gamma/2} B({\rm d}s) - \frac{1}{2}\eta^2 c^{2} \int^{u}_{t} (u-s)^{2\alpha} s^{-\gamma} {\mathrm d} s
\bigg) \,\Big \vert \, \mathcal F (T) \right] \\
& = \xi(t) \exp \bigg( \eta c \int_t^T (u-s)^\alpha s^{-\gamma/2} B({\rm d}s) \bigg) \\
& \quad \times \exp\bigg(\frac{1}{2} \eta^2 c^2 \int_T^u (u-s)^{2\alpha} s^{-\gamma} {\rm d}s - \frac{1}{2}\eta^2 c^{2} \int^{u}_{t} (u-s)^{2\alpha} s^{-\gamma} {\mathrm d} s
\bigg) \\
&= \xi(t) \exp \bigg( \eta c \int_t^T (u-s)^\alpha s^{-\gamma/2} B({\rm d}s) - \frac{1}{2}\eta^2 c^{2} \int^{T}_{t} (u-s)^{2\alpha} s^{-\gamma} {\mathrm d} s
\bigg)\,.
\end{align*}
Using this and plugging into \eqref{eqn-varsigmaT}, we obtain an expression of $\varsigma(T)$. Direct evaluation of $\text{\rm Var}(\varsigma(T))$ would require a good simulation scheme for the generalized R-L FBM, which we leave for future work. (The hybrid scheme for the Brownian semistationary processes in \cite{bennedsen2017hybrid} cannot be adapted for our model.) We next discuss an approximation scheme for $\text{\rm Var}(\varsigma(T))$.
To estimate the conditional variance of $\varsigma(T)$, we use the approximation:
\begin{equation}
\varsigma(T) \approx \exp\left(\frac{1}{\Delta} \int_T^{T+\Delta} \mathbb{E}[\log v(u) \, \vert \, \mathcal F (T)] du \right)\,.
\end{equation}
We then mimic the calculations in Appendix 3 of \cite{bayer2016pricing}.
By the second expression of \eqref{eq: GrBergomi},
we obtain
\begin{align*}
& \int_T^{T+\Delta} \mathbb{E}[\log v(u) \, \vert \, \mathcal F (T)] {\mathrm d} u \\
&= \eta c \int_T^{T+\Delta} \int_t^T (u-s)^\alpha s^{-\gamma/2} B({\rm d}s) {\rm d} u \,-\, \frac{1}{2}\eta^2 c^{2} \int_T^{T+\Delta} \int_t^T(u-s)^{2\alpha} s^{-\gamma} {\rm d} s {\rm d} u \\
&= \eta c \int_t^T \left( \int_T^{T+\Delta} (u-s)^\alpha {\rm d} u \right) s^{-\gamma/2} B({\rm d}s) \\
& \quad - \frac{1}{2}\eta^2 c^{2} \int_t^T\left( \int_T^{T+\Delta} (u-s)^{2\alpha} {\rm d} u \right) s^{-\gamma} {\rm d} s \\
& = \frac{\eta c}{\alpha+1} \int_t^T \left( (T+\Delta-s)^{\alpha+1} - (T-s)^{\alpha+1} \right) s^{-\gamma/2} B({\rm d}s) \\
& \quad - \frac{\eta^2 c^{2}}{2 (2\alpha+1)} \int_t^T \left( (T+\Delta-s)^{2\alpha+1} - (T-s)^{2\alpha+1} \right) s^{-\gamma} {\rm d} s.
\end{align*}
This gives us
\begin{align*}
& \text{\rm Var}\left(\frac{1}{\Delta}\int_T^{T+\Delta} \mathbb{E}[\log v(u) \, \vert \, \mathcal F (T)] {\mathrm d} u \, \bigg\vert \, \mathcal F (t) \right) \\
& = \left(\frac{\eta c}{\alpha+1}\right)^2 \frac{1}{\Delta^2} \int_t^T \left( (T+\Delta-s)^{\alpha+1} - (T-s)^{\alpha+1} \right)^2 s^{-\gamma} {\rm d}s \,.
\end{align*}
By changing variables with $u=\frac{s-t}{T-t}$, we obtain that the integral is equal to
\begin{align*}
&\int_0^1 \Big[ \big((T-t)(1-u)+\Delta\big)^{\alpha+1}- \big( (T-t)(1-u)\big)^{\alpha+1} \Big]^2 (t+(T-t)u)^{-\gamma} (T-t) {\rm d}u \\
&= (T-t)^{2\alpha+3-\gamma} \int_0^1 \Big[ \Big((1-u)+\frac{\Delta}{T-t}\Big)^{\alpha+1}- (1-u)^{\alpha+1} \Big]^2 \Big(\frac{t}{T-t}+u\Big)^{-\gamma} {\rm d}u\,.
\end{align*}
Thus, we have
\begin{align*}
& \text{\rm Var}\left(\frac{1}{\Delta}\int_T^{T+\Delta} \mathbb{E}[\log v(u) \, \vert \, \mathcal F (T)] {\mathrm d} u \, \bigg\vert \, \mathcal F (t) \right) \\
& = \eta^2
(T-t)^{2\alpha+1-\gamma} f_{\alpha, \gamma} \Big(\frac{\Delta}{T-t}, \frac{t}{T-t}\Big) \,,
\end{align*}
where $\, f_{\alpha, \gamma} \,$ is a function defined by
\begin{equation}\label{eq: falphagamma}
\begin{split}
f_{\alpha, \gamma}(a, b) & \, =\, \left(\frac{ c}{\alpha+1}\right)^2 \frac{1}{a^2} \int_0^1 \Big[ \big((1-u)+a\big)^{\alpha+1} \\
& \hspace{4cm} {} - (1-u)^{\alpha+1} \Big]^2 (b+u)^{-\gamma} {\rm d}u\,.
\end{split}
\end{equation}
Recall that $H= \alpha + 1/2 - \gamma/2$; hence, the component $(T-t)^{2\alpha+1-\gamma} = (T-t)^{2H}$, which is the same as that in \cite{bayer2016pricing}.
When $\gamma=0$, this function $f_{\alpha, \gamma}(a,b)$ reduces to the function $f^H(\theta)$ given in \cite{bayer2016pricing}, that is,
\begin{equation*}
f^H(\theta) = \frac{D_H^2}{\theta^2} \int_0^1 \Big( (1+\theta-x)^{1/2+H} - (1-x)^{1/2+H}\Big)^2 {\rm d}u\,.
\end{equation*}
Indeed, note that $c=c(\alpha,\gamma)$ actually depends on $\alpha$ and $\gamma$, and is given by the formula
\begin{equation} \label{eq:constR-L}
c=c(\alpha,\gamma) = \text{Beta}(1-\gamma, 2\alpha+1)^{-1/2}.
\end{equation}
(This is stated in Section 2.2, following from a similar argument in the proof of Lemma 2.1 in \cite{ichiba-pang-taqqu}.)
In the special case $\gamma=0$, we have $c = (2\alpha+1)^{1/2} = (2H)^{1/2}$, and $\alpha+1= H+1/2$, so the constant $c/(\alpha+1)$ matches $D_H = \frac{\sqrt{2H}}{H+1/2}$ in (6.1) of \cite{bayer2016pricing}.
Therefore, with $\,f_{\alpha, \gamma}\,$ in \eqref{eq: falphagamma}, the VIX variance swaps are then approximated by
\begin{align}
\text{VVIX}^2({t,T}) & \approx \text{\rm Var}\left( \log\sqrt{\varsigma(T)} \, \big\vert \, \mathcal{F}(t) \right) \nonumber \\
&\approx \frac{1}{4} \eta^2
(T-t)^{2\alpha+1-\gamma} f_{\alpha, \gamma} \Big(\frac{\Delta}{T-t}, \frac{t}{T-t}\Big) \, =:\, {\mathbf v} \,. \label{eq:59}
\end{align}
Note that unlike \cite{bayer2016pricing}, we cannot simply regard $\text{VVIX}^2({t,T}) $ as a function of $T-t$. Also note that the $\text{VIX}^2$ process measures the variance of the process of interest
whereas $\text{VVIX}^2$ measures the variance of the $\text{VIX}^2$, that is, the variance of the variance.
To illustrate the impact of the parameters $(\alpha,\gamma)$, we
provide Table \ref{Table:1} to show the different values $\,{\mathbf v}\,$ of $\text{VVIX}^2$ approximations for a fixed $H=0.05$, with the above parameter values $\eta=2$, $T=1$, $\Delta=1/12$ and $t=0.5, 0.75\,$.
(Note that the value of $\eta=2$ is chosen according to the estimates from SPX fits in Section 5.2.1 and Section 6.1 of \cite{bayer2016pricing}.)
We observe that for the same value of $H=0.05$, different values of $(\alpha,\gamma)$ can lead to dramatically different values of $\text{VVIX}^2$ approximations in \eqref{eq:59}.
\begin{table}[htp]
\centering
\begin{tabular}{c}
(a) $\,t = 0.5\,$, $\, T = 1\,$, $\Delta = 1/12$, $\eta = 2$, $H = 0.05$\\
\begin{tabular}{|r | c |r |r| }
\hline
$(\alpha,\gamma)$ & $\,H\,$ & $f_{\alpha,\gamma}$ & $\text{VVIX}^2$ $\,\approx {\mathbf v}\,$ in \eqref{eq:59} \\ \hline
$\,(-0.45, 0)\,$ & $\, 0.05\,$ & $\,0.2413\,$ & $\,0.2251\,$\\ \hline
$\,(-0.4, 0.1)\,$ & $\, 0.05\,$ & $\,0.3930\,$ & $\,0.3666\,$\\ \hline
$\,(-0.325, 0.25)\,$ & $\, 0.05\,$ & $\,0.4856\,$ & $\,0.4531\,$\\ \hline
$\,(-0.2, 0.5)\,$ & $\, 0.05\,$ & $\,0.4043\,$ & $\,0.3772\,$\\ \hline
$\,(0, 0.9)\,$ & $\, 0.05\,$ & $\,0.0718\,$ & $\,0.0670\,$\\ \hline
$\,(0.03, 0.96)\,$ & $\, 0.05\,$ & $\,0.0272\,$ & $\,0.0254\,$\\ \hline
\end{tabular} \\
\\
(b) $\,t = 0.75\,$, $\, T = 1\,$, $\Delta = 1/12$, $\eta = 2$, $H = 0.05$\\
\begin{tabular}{|r | c | r |r| }
\hline
$(\alpha,\gamma)$ & $\,H\,$ & $f_{\alpha,\gamma}$ & $\text{VVIX}^2$ $\,\approx {\mathbf v}\,$ in \eqref{eq:59} \\ \hline
$\,(-0.45, 0)\,$ & $\, 0.05\,$ & $\,0.1936\,$ & $\,0.2100\,$\\ \hline
$\,(-0.4, 0.1)\,$ & $\, 0.05\,$ & $\,0.3008\,$ & $\,0.3421\,$\\ \hline
$\,(-0.325, 0.25)\,$ & $\, 0.05\,$ & $\,0.3434\,$ & $\,0.4227\,$\\ \hline
$\,(-0.2, 0.5)\,$ & $\, 0.05\,$ & $\,0.2455\,$ & $\,0.3520$\\ \hline
$\,(0, 0.9)\,$ & $\, 0.05\,$ & $\,0.0326\,$ & $\,0.0625\,$\\ \hline
$\,(0.03, 0.96)\,$ & $\, 0.05\,$ & $\,0.0118\,$ & $\,0.0237\,$\\ \hline
\end{tabular}
\end{tabular}
\caption{Illustration of the impact of $\gamma$ on the approximation $\,{\mathbf v}\,$ of $\text{VVIX}^2$ in \eqref{eq:59} for (a) $\,t = 0.50\,$ and for (b) $\,t = 0.75\,$.} \label{Table:1}
\end{table}
To further illustrate the significance of the impact of the non-stationarity parameter $\gamma$,
in Figure \ref{Figure:5} we show $\,{\mathbf v}\,$ in \eqref{eq:59} for the same $H=0.05$ with two pairs of $(\alpha, \gamma)$ in two perspectives: (a) as a function of $\,t\,$ for a fixed $\,T \, =\, 1\,$ and (b) as a function of $\,T \ge t\,$ for a fixed $\, t = 0.25\,$. Comparing with the case of standard FBM ($\alpha=-0.45, \gamma=0$),
we observe that the values of $\,{\mathbf v}\,$ are larger for the pair $(\alpha,\gamma) =(-0.2,0.5)$ for $t$ less than a certain threshold (as shown in Table \ref{Table:1} for the two times $t=0.5, 0.75$, and then becomes smaller (but relatively close)
for fixed $T=1$, and they are larger over all $T$ for the fixed $t=0.25$.
More importantly,
for a fixed $T=1$, as $t$ is far from $T$, the difference in the values of $\,{\mathbf v}\,$ can be substantial and as $t$ gets close to $T$, the difference shrinks to zero; on the other hand, for a fixed $t=0.25$, as $T$ increases the difference in the values of $\,{\mathbf v}\,$ remain somewhat constant.
\begin{figure}[h]
\begin{tabular}{cc}
\includegraphics[scale=0.42]{PlotV-t.pdf} &
\includegraphics[scale=0.42]{PlotVvT.pdf}\\
(a) $\,0 \le t \le T = 1\,$& (b) $\,t=0.25 \le T \le 3\,$ \\
\end{tabular}
\caption{ (a) The values $\,{\mathbf v}\,$ in \eqref{eq:59} for different values of $\,t\,$ with $\,H = 0.05 \,$ and $\,T \, =\, 1\,$. (b) The values $\,{\mathbf v}\,$ in \eqref{eq:59} for different values of $\,T\ge t = 0.25\,$ with $\,H = 0.05\,$. } \label{Figure:5}
\end{figure}
Figure \ref{Figure:4} shows the values in \eqref{eq:59} with respect to $\,\gamma \in (0, 1) \,$ for each $\,H \, =\, 0.05, 0.1, 0.15 \,$ and $\,t = 0.5, 0.75\,$ with $\, \alpha = \gamma/2 + H - 1/2\,$.
Note that as $\, \gamma \uparrow 1\,$, the normalizing constant $\, c(\alpha, \gamma) \,$ in \eqref{eq:constR-L} converges $\, \lim_{\gamma \uparrow 1} c(\alpha, \gamma) \, =\, 0\,$, and so does $\,f_{\alpha, \gamma }\,$ in \eqref{eq: falphagamma}. Hence, the value $\,{\mathbf v} \,$ in \eqref{eq:59} converges to $0$, as $\, \gamma \uparrow 1\,$.
Moreover, for different values of $H$, we observe that when $\gamma$ is relatively small (close to zero), the differences of the $\text{VVIX}^2$ values in \eqref{eq:59} are relatively large, while they become smaller as the value of $\gamma$ increases to 1.
This implies that if there is some slight non-stationarity in the increments that is ignored in the analysis and model calibration, it can lead to substantial bias in the estimates of the VIX variance swap.
For different times $t$, the shapes of curves in Figures \ref{Figure:4}(a) and (b) look similar with slightly different values for each $\gamma$.
\begin{figure}[htp]
\begin{tabular}{cc}
\includegraphics[scale=0.42]{fig2a.pdf} &
\includegraphics[scale=0.42]{fig2b.pdf}\\
(a) $\,t = 0.5\,$& (b) $\,t \, =\, 0.75\,$ \\
\end{tabular}
\caption{ (a) The values $\,{\mathbf v}\,$ in \eqref{eq:59} for different values of $\,\gamma\,$ for each $\,H = 0.05, 0.1, 0.15\,$ with $\, \alpha = \gamma/2 + H - (1/2)\,$, $t = 0.5$. (b) The values $\,{\mathbf v}\,$ in \eqref{eq:59} for different values of $\,\gamma\,$ for each $\,H = 0.05, 0.1, 0.15\,$ with $\, \alpha = \gamma/2 + H - (1/2)\,$, $t = 0.75$. } \label{Figure:4}
\end{figure}
We also plot
$\,{\mathbf v}\,$ in \eqref{eq:59} as a surface of the parameters $(\alpha, \gamma)$, again fixing the same parameter values $\eta=2$, $T=1, t=0.5, \Delta=1/12$ and $\,0 \le \gamma < 1 \,$, for (a) $\, t \, =\, 0.5 \,$ and (b) $\,t \, =\, 0.75\,$, in Figure \ref{fig:3}(a) and Figure \ref{fig:3}(b), respectively.
\begin{figure}[htp]
\begin{tabular}{cc}
\includegraphics[scale=0.42]{fig3a.pdf} &
\includegraphics[scale=0.42]{fig3b.pdf} \\
(a) & (b) \\
\end{tabular}
\caption{(a) The values $\,{\mathbf v}\,$ of the $\text{VVIX}^2$ approximations in \eqref{eq:59} for different values of $\,(\alpha, \gamma)\,$ with $\, T = 1\,$, $\Delta = 1/12$, $\eta = 2$, (a) $\,t = 0.5\,$ and (b) $\, t = 0.75\,$. }
\label{fig:3}
\end{figure}
\begin{remark}
We remark that in Remark 2.1 \cite{bayer2016pricing}, the authors stated that the Riemann-Liouville FBM is an example of a Brownian semistationary (BSS) process \cite{barndorff2009brownian}, which is of the form $X(t) = \int_{-\infty}^t g(t-s)\sigma(s) {\rm d} B(s) $ for some deterministic function $g$ and an adapted intermittency process $\sigma(s)$. However, our generalized Riemann-Liouville FBM \eqref{eqn-X-RL} does not belong to this class (BSS) of processes, since the process $u^{-\gamma/2}$ in the definition of the $X(t)$ in \eqref{eqn-X-RL} violates all the ``(semi)stationarity" conditions imposed upon $\sigma(t)$ (see, e.g., \cite{barndorff2011multipower}).
\end{remark}
\subsection{On other stochastic volatility models}
In addition to the generalized rough Bergomi model above, we propose the following candidates to model the volatility process $\sigma(t)$ in \eqref{eqn-log-price}.
(i) The GFBM and its mixture models.
The log-volatility $\ln \sigma(t)$ takes the form $\ln (\sigma(t)/\sigma (0)) = X(t)$ or $Y(t) = \widetilde{B}(t) + X(t)$, $t \ge 0 $, where $X(\cdot)$ is the GFBM given in \eqref{def-X}, and $Y(\cdot)$ is the mixture process in \eqref{eq: Y}, of a standard BM
$\widetilde{B}(\cdot)$ and the process $X(\cdot)$ in \eqref{def-X}.
We recall the semimartingale properties of $X$ and $Y$ in Propositions \ref{prop-sm-X} and \ref{prop-sm-X-2}, and Proposition \ref{prop-mixture}, respectively.
Although in the conventional stochastic volatility models, the volatility $\sigma_t$ is often modeled as a continuous Brownian semi-martingale, for example, the Heston model
and the SABR (stochastic alpha-beta-rho) model, one may also model the log-volatility process as a semimartingale in certain scenarios using the GFBM and its mixtures; see, e.g., the recent development of the corresponding rough volatility models using FBM in \cite{jaisson2016rough,el2019roughening,fukasawa2021rough}.
For instance, in the recent work on rough SABR model by Fukasawa and Gatheral \cite{fukasawa2021rough}, the log-volatility is modeled as a Riemann--Liouville (R-L) FBM, whose driving BM is correlated with the BM driving the stock price. There could be potential extensions of that rough SABR model using the generalized R-L FBM in \eqref{eqn-X-RL} and its mixture.
(ii) The generalized fractional Ornstein-Ulenbeck (fOU) processes driven by the GFBM $X$ and its mixture $Y$. The log-volatility $\log \sigma(\cdot)$ takes the form
\begin{equation} \label{eqn-log-sigma-3}
\log (\sigma(t)/\sigma (0)) = Z(t), \quad t \ge 0 ,
\end{equation}
where $Z(\cdot)$ is the solution to the SDE driven by the GFBM $X(\cdot)$ in \eqref{def-X}:
\begin{equation} \label{eqn-log-sigma-Z1}
{\rm d}Z(t) = - a (Z(t) - m) {\rm d} t + \nu {\rm d} X(t), \quad t \ge 0 ,
\end{equation}
with $m \in {\mathbb R}$ and $a$ and $\nu$ being positive constant parameters. Using pathwise Reimann-Stieltjes integral, we can also write the solution as
\begin{equation} \label{eqn-gen-fOU}
Z(t) = Z(0) e^{-a t } + m ( 1- e^{-a t }) + \nu \int_{0}^t e^{-a (t-s)} {\rm d} X(s), \quad t \ge 0 .
\end{equation}
We refer to this as the generalized fOU process.
In \eqref{eqn-log-sigma-Z1}, one can also use the mixed GFBM process $Y$ in \eqref{eq: Y}.
It is clear that in both models, the log-volatility process $\log \sigma(\cdot)$ is not a semimartingale.
\subsection{Comments on stochastic integrals with respect to the GFBM} \label{sec-stochastic integral}
Observe that we have used stochastic integrals with respect with the GFBM $X$ in \eqref{eqn-PV-X} and \eqref{eqn-PV-X-mix} in portfolio optimization and in \eqref{eqn-gen-fOU}
in rough volatility models. In \eqref{eqn-PV-X} and \eqref{eqn-PV-X-mix}, the stochastic integral is of the type $\int f(X){\rm d}X$, while in \eqref{eqn-gen-fOU},
the integrand is a deterministic function $f(t-s)$.
Stochastic integrals with respect to FBM $B^H$ have been extensively studied in the literature (see for example \cite[Chapter 7]{PT17}, \cite[Chapter 3]{nourdin2012selected} and \cite{biagini2008stochastic,nualart2006malliavin}). For $H>1/2$, the stochastic integral $\int f(B^H){\rm d}B^H$ can be defined pathwise using Young integral due to the regularity of the sample paths of $B^H$, see, e.g., \cite[Chapter 3.1]{nourdin2012selected}. Thanks to the variation property of the GFBM $X$ in Proposition \ref{prop: variation} when $H>1/2$, the integral $\int f(X) {\rm d}X$ for the GFBM $X$ can also be well defined pathwsie using Young integral. On the other hand, for the standard FBM $B^H$ with $H<1/2$, the stochastic integral is studied using rough path theory \cite{friz2010multidimensional} and/or Malliavin calculus \cite{biagini2008stochastic,nualart2006malliavin}. This study for the GFBM $X$ with $H<1/2$ is out of the scope of this paper, and will be investigated in the future.
When the integrand is a deterministic function of the form $\int_0^t f(s) {\rm d} B^H(s)$ or $\int_0^t f (t-s) {\rm d}B^H(s)$, the integral can be defined as a pathwise Riemann-Stieltjes integral, see Proposition A.1 in \cite{cheridito2003fractional}.
By a slight modification of the proof of that proposition, using self-similarity and the H{\"o}lder continuity properties of the GFBM $X$, the same conclusions in Proposition A.1 of \cite{cheridito2003fractional} also hold by replacing $B^H$ with the GFBM $X$. Therefore, the generalized fOU process $Z$ in \eqref{eqn-gen-fOU} is well defined, and so is the process
driven by the mixed GFBM process $Y$.
\section{Appendix}
\subsection{Proof of Proposition \ref{prop: variation}} \label{sec-app-1}
In this subsection, we prove Proposition \ref{prop: variation}.
We first establish the following lemma.
\begin{lemma}\label{lm: variation} For every $\, 0 < u < v\,$, the limits of the covariance of increments at $\,u, v \,$ are given by
\begin{equation} \label{eq: variationLM}
\begin{split}
& \lim_{\varepsilon \downarrow 0 } \frac{\,1\,}{\,\varepsilon^{2H}\,} \, \mathbb E \big[ (X(u+\varepsilon) - X(u))^{2} \big] \, =\, c^{2} \, \text{\rm Beta} ( 1 + 2\alpha , 1 - \gamma) > 0 \, , \\
& \lim_{\varepsilon \downarrow 0 } \frac{\,1\,}{\,\varepsilon^{2H}\,} \, \mathbb E \big[ (X(u+\varepsilon) - X(u)) (X(v+\varepsilon) - X(v)) \big] \, =\, 0\, .
\end{split}
\end{equation}
\end{lemma}
\begin{proof} [Proof of Lemma \ref{lm: variation}]
First, we consider the case $\, u \, =\, v\,$. In this case, since the expectation in \eqref{eq: variationLM} becomes the second moment $\, \Phi (u+\varepsilon, u) \,$ of the increment in \eqref{eqn-Phi}, by a straightforward calculation (see, e.g., \cite{ichiba-pang-taqqu}) for $\, s < t \,$,
\begin{align*}
& \mathbb E [ \lvert X(t) - X(s) \rvert^{2}] \\
&\, =\, c^{2} (t-s)^{2H} \bigg( \text{Beta} ( 1 + 2\alpha, 1 - \gamma) \\
& \qquad + \int^{\infty}_{s/(t-s)} [ (1 + v)^{\alpha} - v^{\alpha} ]^{2} \Big( v - \frac{\,s\,}{\,t-s\,} \Big)^{-\gamma } {\mathrm d} v \\
&\qquad + \Big[\int^{1}_{0} (1-w)^{-\gamma} ( (1 + x w)^{\alpha} - (xw)^{\alpha})^{2} x^{1-\gamma} {\mathrm d} w \Big] \Big \vert_{\{x \, =\, s / ( t -s ) \}} \bigg)\,,
\end{align*}
and hence, substituting $\, t \, =\, u+\varepsilon\,$ and $\, s \, =\, u \,$ and dividing by $\, \varepsilon^{2H}\,$ on both sides, we obtain the first claim in \eqref{eq: variationLM}. Indeed, since
$\, \int^{\infty}_{0} [(1+v)^{\alpha} - v^{\alpha}]^{2} v^{-\gamma}{\mathrm d} v < \infty \,$ under \eqref{def-X-g-c} with $\, \gamma \in (0,1) \,$, we have
\begin{align*}
\int^{\infty}_{u/\varepsilon} [ ( 1 + v)^{\alpha} - v^{\alpha}]^{2} \Big ( v - \frac{\,u\,}{\,\varepsilon \,} \Big)^{-\gamma} {\mathrm d} v
& \, = \, \int^{\infty}_{u/\varepsilon} \Big[ \Big( 1 + v - \frac{\,u\,}{\,\varepsilon\,}\Big)^{\alpha}- \Big( v - \frac{\,u\,}{\,\varepsilon\,} \Big)^{\alpha} \Big]^{2} v^{-\gamma} {\mathrm d} v \\
& \xrightarrow[\varepsilon \downarrow 0]{ } 0 \,, \quad u > 0 \,,
\end{align*}
for the second term, and for the third term, we apply the dominated convergence theorem to obtain
\[
\int^{1}_{0} ( 1- w)^{-\gamma} ( ( 1 + xw)^{\alpha} - ( x w)^{\alpha} )^{2} x^{1-\gamma} {\mathrm d}w \xrightarrow[x \to \infty]{} 0 \, ,
\]
because for every $\, x \ge 1 \,$, by the monotonicity of $\, x \mapsto ((1+x w)^{\alpha} - (xw)^{\alpha} ) x^{1-\gamma}\,,$
\[
\int^{1}_{0} ( 1- w)^{-\gamma} ( ( 1 + xw)^{\alpha} - ( x w)^{\alpha} )^{2} x^{1-\gamma} {\mathrm d}w \, \le \, \int^{1}_{0} (1 - w)^{-\gamma} ((1 + w)^{\alpha}- w^{\alpha})^{2}{\mathrm d} w < \infty \,
\]
in the case of $\, \alpha > 0 \,$, and similarly for every $\, x \ge 1 \,,$
\begin{align*}
\int^{1}_{0} ( 1- w)^{-\gamma} ( ( 1 + xw)^{\alpha} - ( x w)^{\alpha} )^{2} x^{1-\gamma} {\mathrm d}w &\, \le \, \int^{1}_{0} (1 - w)^{-\gamma} w^{2\alpha} {\mathrm d} w \\
& \, =\, \text{Beta} ( 1 - \gamma, 1 +2 \alpha) < \infty
\end{align*}
in the case of $\, \alpha < 0 \,$. See the proof of Theorem 3.1 in \cite{ichiba-pang-taqqu} for the derivation of these upper bounds.
Next, we consider the case $\, u < v \,$ by evaluating the sum of four terms
\begin{equation} \label{eq: lm 4.1 C1e2}
\frac{\,1\,}{\,c^{2}\,}\mathbb E \big[ (X(u+\varepsilon) - X(u)) (X(v+\varepsilon) - X(v)) \big] \, =\, C_{1,\varepsilon} + C_{2, \varepsilon} + C_{3, \varepsilon} + C_{4, \varepsilon}\, ,
\end{equation}
where
\begin{align*}
C_{1, \varepsilon} &\, :=\, \int^{u}_{0} ((u+\varepsilon - w)^{\alpha} - ( u - w)^{\alpha} ) ( (v+\varepsilon - w)^{\alpha} - ( v - w)^{\alpha} ) \lvert w\rvert^{-\gamma} {\mathrm d} w \, ,
\\
C_{2, \varepsilon} &\, :=\, \int^{u+\varepsilon}_{u} ( u+\varepsilon - w)^{\alpha} ( ( v+\varepsilon - w)^{\alpha} - (v-w)^{\alpha}) \lvert w\rvert^{-\gamma} {\mathrm d} w \, ,
\\
C_{3, \varepsilon} &\, :=\, \int^{1}_{0} ((u+\varepsilon + w)^{\alpha} - (u+w)^{\alpha}) ( (v+\varepsilon + w)^{\alpha} - (v + w)^{\alpha}) \lvert w \rvert^{-\gamma} {\mathrm d} w \, ,
\\
C_{4, \varepsilon} &\, :=\, \int^{\infty}_{1} ((u+\varepsilon + w)^{\alpha} - (u+w)^{\alpha}) ( (v+\varepsilon + w)^{\alpha} - (v + w)^{\alpha}) \lvert w \rvert^{-\gamma} {\mathrm d} w \, .
\end{align*}
For the first term $\, C_{1, \varepsilon}\,$, we consider the case $\, \alpha > 0 \,$ first. Note that a simple application of the H\"older continuity, i.e., $\,x ^{\alpha } - y ^{\alpha} \le \lvert x - y \rvert^{\alpha }\,$, $\, x, y > 0 \,$ implies $\,
C_{1, \varepsilon} \le \int^{u}_{0} \varepsilon^{2\alpha} \lvert w \rvert^{-\gamma} {\mathrm d} w \, =\, u^{1-\gamma} \varepsilon^{2\alpha} \, $ but then we may not use this inequality to show $\, \lim_{\varepsilon \downarrow 0 } C_{1, \varepsilon} / \varepsilon^{2H} \, =\, 0\,$, because $\, 2 \alpha \le 2 H \, =\, 2 \alpha - \gamma + 1 \,$. We shall use the estimates of the difference $\, x^{\alpha} - y^{\alpha} \,$ for $\,0 < y < x \,$ outside the neighborhood of the origin.
If $\,\alpha > 0 \,$, using the inequality
\begin{equation} \label{eq: lm 4.1 C1e0}
\, (v+\varepsilon - w )^{\alpha} - (v-w)^{\alpha} \, =\, (v-w)^{\alpha} \Big( \Big( 1 + \frac{\,\varepsilon \,}{\, v - w\,}\Big)^{\alpha} - 1 \Big) \, \le \, \alpha (v-u)^{\alpha -1} \varepsilon\,
\end{equation}
for $\, 0 < w < u < v \,$, we have
\begin{equation} \label{eq: lm 4.1 C1e}
C_{1, \varepsilon} \, \le\, \alpha (v-u)^{\alpha -1} \varepsilon \cdot \int^{u}_{0} ((u+\varepsilon - w)^{\alpha} - (u-w)^{\alpha} ) w^{-\gamma} {\mathrm d} w \, ,
\end{equation}
where the integral on the right hand is evaluated by
\begin{align} \label{eq: lm 4.1 C1e1}
& \int^{u}_{0} ((u+\varepsilon - w)^{\alpha} - (u-w)^{\alpha} ) w^{-\gamma} {\mathrm d} w \, \nonumber\\
& = \, \int^{u+\varepsilon}_{0} ( u+ \varepsilon -w)^{\alpha} w^{-\gamma} {\mathrm d} w - \int^{u}_{0} ( u - w)^{\alpha} w^{-\gamma} {\mathrm d} w - \int^{u+\varepsilon}_{u} ( u + \varepsilon - w)^{\alpha} w^{-\gamma} {\mathrm d} w \nonumber\\
& \, =\, ( (u+\varepsilon)^{\alpha - \gamma +1 } - u^{\alpha - \gamma + 1 }) \, \text{Beta} ( 1 + \alpha, 1 - \gamma) - \int^{u+\varepsilon}_{u} ( u + \varepsilon - w)^{\alpha} w^{-\gamma} {\mathrm d} w \nonumber \\
& \, \le (\alpha - \gamma +1) u^{\alpha - \gamma} \text{Beta} ( 1 + \alpha, 1 - \gamma) \varepsilon - \int^{u+\varepsilon}_{u} (u+\varepsilon - w)^{\alpha} w^{-\gamma} {\mathrm d} w \, ,
\end{align}
and
\[
0 \le \int^{u+\varepsilon}_{u} ( u+\varepsilon - w)^{\alpha} w^{-\gamma} {\mathrm d} w \le u^{-\gamma} \int^{u+\varepsilon}_{u} ( u+\varepsilon -w )^{\alpha} {\mathrm d} w \, =\, \frac{\,u^{-\gamma}\,}{\,1+\alpha\,} \cdot \varepsilon^{\alpha + 1} \, .
\]
Here, $\, \alpha - \gamma + 1 > 0 \,$.
Thus, if $\,\alpha > 0 \,$, combining these inequalities together with \eqref{eq: lm 4.1 C1e}, we obtain
\begin{equation*}
\begin{split}
&\frac{\,1\,}{\,\varepsilon^{2H}\,}C_{1, \varepsilon} \\
&\,\le \, \alpha (v-u)^{\alpha-1} \varepsilon^{1-2H} \bigg( (\alpha - \gamma + 1) u^{\alpha - \gamma} \text{Beta} ( 1 + \alpha, 1 - \gamma) \varepsilon + \frac{\,u^{-\gamma}\,}{\,1+\alpha\,} \cdot \varepsilon^{\alpha + 1 } \bigg) \\
& \, = \, \alpha (v-u)^{\alpha-1} (\alpha - \gamma + 1)u^{\alpha - \gamma} \text{Beta} ( 1 + \alpha, 1 - \gamma) \varepsilon ^{2(1-H)} \\
& \qquad + \alpha (v-u)^{\alpha-1}\frac{\,u^{-\gamma}\,}{\,1+\alpha\,} \cdot \varepsilon^{2 (1-H) + \alpha } \\
& \xrightarrow[\varepsilon \downarrow 0 ]{} 0 \,,
\end{split}
\end{equation*}
because $\, 0 < H < 1 \,$ and $\, 2 ( 1 - H) + \alpha \, =\, 1 - \alpha + \gamma > 0 \,$.
Still for the first term $\, C_{1, \varepsilon}\,$, we consider the case $\, \alpha < 0 \,$. Let $\, \widetilde{\alpha} \, :=\, - \alpha > 0 \, $. Then using \eqref{eq: lm 4.1 C1e0} with the power $\, \widetilde{\alpha} > 0 \,$, instead of the power $\, \alpha\,$, we have
\begin{align*}
\big\lvert ( v + \varepsilon - w)^{\alpha} - ( v - w)^{\alpha} \big\rvert
& \, =\, \bigg \lvert \frac{\, ( v + \varepsilon - w)^{ \widetilde{\alpha }} - (v - w)^{ \widetilde{\alpha}} \,}{\, ( v + \varepsilon - w)^{ \widetilde{\alpha}} \, ( v - w)^{ \widetilde{\alpha}} \,} \bigg \rvert \\
& \,\le\, \bigg \lvert \frac{\, \widetilde{\alpha} ( v - u)^{ \widetilde{\alpha} - 1} \,}{\, ( v + \varepsilon - w)^{ \widetilde{\alpha}} \, ( v - w)^{ \widetilde{\alpha}} \,} \bigg \rvert \varepsilon
\end{align*}
for $\, 0 < w < u \,$. Using this inequality, we obtain
\[
\lvert C_{1, \varepsilon}\rvert \le\bigg \lvert \frac{\, \widetilde{\alpha} ( v - u)^{ \widetilde{\alpha} - 1} \,}{\, ( v + \varepsilon - u)^{ \widetilde{\alpha}} \, ( v - u)^{ \widetilde{\alpha}} \,} \bigg \rvert \varepsilon \cdot \int^{u}_{0}\lvert (u+\varepsilon - w)^{\alpha} - (u-w)^{\alpha} \rvert w^{-\gamma} {\mathrm d} w \, ,
\]
where the integral on the right hand side is now evaluated by
\begin{equation*}
\begin{split}
& \int^{u}_{0}\lvert (u+\varepsilon - w)^{\alpha} - (u-w)^{\alpha} \rvert w^{-\gamma} {\mathrm d} w \\
& \, =\, - \int^{u+\varepsilon}_{0} ( u+ \varepsilon -w)^{\alpha} w^{-\gamma} {\mathrm d} w + \int^{u}_{0} ( u - w)^{\alpha} w^{-\gamma} {\mathrm d} w \\
& \qquad \qquad + \int^{u+\varepsilon}_{u} ( u + \varepsilon - w)^{\alpha} w^{-\gamma} {\mathrm d} w \\
& \,\le\, \big\lvert (u+\varepsilon)^{\alpha - \gamma + 1} - u^{\alpha - \gamma + 1}\big\rvert \, \text{Beta} ( 1 + \alpha , 1 - \gamma) + \int^{u+\varepsilon}_{u} (u+\varepsilon - w)^{\alpha} w^{-\gamma} {\mathrm d} w \, .
\end{split}
\end{equation*}
Thus, from here, we may follow the same steps after the inequality \eqref{eq: lm 4.1 C1e1} and we obtain $\,
\lim_{\varepsilon \downarrow 0 } {\,\lvert C_{1, \varepsilon}\rvert\,}/ {\,\varepsilon^{2H}\,} \, =\, 0 \, $ under the case $\, \alpha < 0 \,$ as well.
With similar reasoning, we obtain
$
\, \lim_{\varepsilon \downarrow 0 }\, \lvert C_{i, \varepsilon}\rvert\,/\,\varepsilon^{2H}\, =\, 0\,
$
for $\, i \, =\, 2, 3, 4\,$. Therefore, combining these estimate with \eqref{eq: lm 4.1 C1e2} we conclude the second claim in \eqref{eq: variationLM}.
\end{proof}
We are now ready to prove Proposition \ref{prop: variation}.
\begin{proof}[Proposition \ref{prop: variation}]
Recall that if $\, \tilde{Z} \,$ is a normal random variable with mean $\,0\,$ and variance $\, \sigma^{2} > 0 \,$, then $\, \mathbb E [ \lvert \tilde{Z} \rvert^{p} ] \, =\, \sigma^{p} \sqrt{ 2^{p} / \pi } \Gamma ( (1+p)/2 ) \,$ for $\,p \ge 1\,$ by the Gamma function formula.
Thus, combining this fact with the consequence
\[
\lim_{\varepsilon \downarrow 0 } \frac{\,1 \,}{\,\varepsilon^{2H}\,} \text{Var} ( X(u+\varepsilon) - X(u)) = c^{2} \text{Beta} ( 1 + 2 \alpha , 1 - \gamma)\, ,
\]
from \eqref{eq: variationLM}, we obtain
\begin{align*}
& \lim_{\varepsilon \downarrow 0 } \frac{\,1\,}{\,\varepsilon\,} \mathbb E \big[ \lvert X(u+\varepsilon) - X(u) \rvert^{1/H} \big] \\
& \,= \,\big ( c^{2} \text{Beta} ( 1 + 2 \alpha , 1 - \gamma) \big)^{1/(2H)} \cdot \sqrt{ \frac{2^{1/H}}{\pi} } \Gamma \bigg( \frac{1+(1/H)}{2} \bigg) \\
& \, =\, \rho_{\alpha, \gamma} \, ,
\end{align*}
as the convergence of the moments of normal random variables.
Thus it suffices to show
\begin{equation} \label{eq: prop var 1}
\lim_{\varepsilon \downarrow 0 } \mathbb E \bigg[ \bigg( \frac{\,1\,}{\,\varepsilon \,}\int^{t}_{0} \lvert X(s+\varepsilon) - X(s) \rvert^{1/H} {\mathrm d} s \bigg)^{2}\bigg] \, =\, (\rho_{\alpha, \gamma} t)^{2} \,, \quad t > 0 \, .
\end{equation}
To show \eqref{eq: prop var 1}, we follow the proofs of Proposition 1 of \cite{MR2218338} and Proposition 3.14 of \cite{MR1785063},
and use our Lemma \ref{lm: variation}. By the Fubini theorem, we rewrite
\begin{equation} \label{eq: prop var 2}
\mathbb E \bigg[ \bigg( \frac{\,1\,}{\,\varepsilon \,}\int^{t}_{0} \lvert X(s+\varepsilon) - X(s) \rvert^{1/H} {\mathrm d} s \bigg)^{2}\bigg] \, =\, 2 \int_{[0, t]^{2}} {\bf m}_{\varepsilon} \cdot {\bf 1}_{\{u < v \}} {\mathrm d} u {\mathrm d} v \, ,
\end{equation}
where we define
\begin{equation} \label{eq: m-epsilon}
{\bf m}_{\varepsilon}\, :=\, \frac{\,1\,}{\,\varepsilon^{2}\,}\mathbb E \Big[ {\lvert (X(u+\varepsilon) - X(u)) ( X(v+\varepsilon) - X (v)) \rvert^{1/H} }\Big] \,, \quad \varepsilon > 0 \, .
\end{equation}
The joint distribution of of the increments $\, X(u+\varepsilon) - X(u) \,$ and $\, X(v+\varepsilon) - X(v) \,$ is normal with zero mean and variances defined by
\begin{align*}\, \sigma_{1, \varepsilon}^{2} &\, :=\, \mathbb E \big[ (X(u+\varepsilon) - X(u))^{2}\big] = \Phi (u+\varepsilon, u) \,,\\
\,\sigma_{2, \varepsilon}^{2} &\, :=\, \mathbb E \big[ (X(v+\varepsilon) - X(v))^{2} \big] = \Phi (v + \varepsilon, v ) \,
\end{align*}
and the covariance defined by
\[
{ \vartheta}_{\varepsilon} \, :=\, \mathbb E \big[ (X(u+\varepsilon) - X(u)) (X(v+\varepsilon) - X(v)) \big] \,.
\]
By the the conditional distribution of $\, X(v+\varepsilon) - X(v) \,$, given $\, X(u+\varepsilon) - X(u) \,$, we compute $\,{\bf m}_{\varepsilon}\,$ in \eqref{eq: m-epsilon} as
\begin{align*}
{\bf m}_{\varepsilon}
&\, =\, \frac{\,1\,}{\,\varepsilon^{2}\,}
\mathbb E \Bigg[ \big \lvert \, \sigma_{1,\varepsilon} \cdot \, Z_{1}\big \rvert^{1/H} \cdot \bigg \lvert \frac{\, {\vartheta_{\varepsilon}} \,}{\,\sigma_{1, \varepsilon}\,} \cdot Z_{1} + \sqrt{ 1 - \frac{\, { \vartheta_{\varepsilon}}^{2}\,}{\,\sigma_{1, \varepsilon}^{2} \, \sigma_{2, \varepsilon}^{2}\,} \, \, } \sigma_{2, \varepsilon} {\, Z_{2}\, } \bigg\rvert ^{1/H} \Bigg]
\\
&\,=\, \mathbb E \Bigg[ \big \lvert \, Z_{1} \, \rvert^{1/H} \cdot \bigg \lvert \frac{\, {\vartheta_{\varepsilon}} \,}{\,\varepsilon^{2H}\,} \cdot Z_{1} + \sqrt{ 1 - \frac{\, { \vartheta_{\varepsilon}}^{2}\,}{\,\sigma_{1, \varepsilon}^{2} \, \sigma_{2, \varepsilon}^{2}\,} \, \, } \frac{\, \sigma_{1, \varepsilon} \sigma_{2, \varepsilon}\,}{\,\varepsilon^{2H}\,} {\, Z_{2}\, } \bigg\rvert ^{1/H} \Bigg] \, .
\end{align*}
Here, $\, Z_{i} \, $, $\, i \, =\, 1, 2\,$ are i.i.d. standard normal random variables with mean $\,0\,$ and variance $\,1\,$.
Then it follows from Lemma \ref{lm: variation} that
$$\,\lim_{\varepsilon \downarrow 0 } \sigma_{i, \varepsilon}^{2}\, / \, \varepsilon^{2H} \, =\, c^{2} \text{Beta} ( 1 + 2 \alpha, 1 - \gamma ) \, \quad \text{for} \quad \, i \, =\, 1, 2 \,,$$
and
\[
\lim_{\varepsilon \downarrow 0 } \frac{\,{\vartheta_{\varepsilon}} \,}{\,\varepsilon^{2H}\,} \, =\, 0 \, , \quad \lim_{\varepsilon \downarrow 0 } \frac{\,\vartheta_{\varepsilon}^{2}\,}{\,\sigma^{2}_{1, \varepsilon} \sigma_{2, \varepsilon}^{2}\,} \, =\, 0 \,, \quad \lim_{\varepsilon \downarrow 0} \frac{\, \sigma_{1, \varepsilon} \sigma_{2, \varepsilon}\,}{\,\varepsilon^{2H}\,} \, =\, c^{2} \text{Beta} ( 1 + 2 \alpha, 1 - \gamma ) \, .
\]
Thus substituting them into \eqref{eq: prop var 2}, we obtain \eqref{eq: prop var 1} and conclude \eqref{eq: prop variation}. The proof of \eqref{eq: prop variation 2} is similar.
\end{proof}
\subsection{Proof of Lemma \ref{lm:sq-integrability} } \label{sec-app-2}
\begin{proof}[Proof of Lemma \ref{lm:sq-integrability}]
(i)
Suppose that $\gamma/2 < \alpha <1/2+ \gamma /2 $ and $\, 0 <\gamma <1 \,$.
Using the inequality $\, (x+y)^{2}\le 2 (x^{2} + y^{2}) \,$ for $\, x, y > 0 \,$ and the symmetry of integral region, we obtain
\begin{equation} \label{eq: lm4.1.2}
\int^{T}_{0} \int^{T}_{0} \bigg[ \frac{\partial^{2} \Psi}{\partial u \partial v}(u,v )\bigg]^{2} {\mathrm d} u {\mathrm d} v \le 4 c^{4} \int^{T}_{0} \bigg( \int^{v}_{0} (f_{1}(u,v))^{2} + (f_{2}(u,v))^{2} {\mathrm d} u \bigg) {\mathrm d} v \,.
\end{equation}
Here, the first term $f_{1}(u,v) $ is bounded by
\[
f_{1}(u,v) \, \le \int^{u}_{0} v^{\alpha-1} (u-\theta)^{\alpha - 1} \theta^{-\gamma} {\mathrm d} \theta \, =\, \text{Beta} ( \alpha , 1 - \gamma) \, u^{\alpha - \gamma}v^{\alpha -1} ,
\]
for $ u < v $. Since $ 0 < \alpha < 1$ and $\,(uw)^{\alpha -1 } (vw)^{\alpha -1} \ge ( 1 + uw)^{\alpha -1 }(1 + vw)^{\alpha -1}\,$ for $ u, v>0 $, we have a bound for the second term $f_{2}(u,v) $,
\begin{equation*}
\begin{split}
f_{2}(u,v) \, &=\, \int^{\infty}_{0} (uv)^{\alpha-\gamma+1} (1 + uw)^{\alpha -1} (1 + vw )^{\alpha - 1} w^{-\gamma} {\mathrm d} w
\\
&\le (uv)^{2\alpha - \gamma} \int^{\infty}_{1} w^{2\alpha -2 - \gamma} {\mathrm d} w \\
&\quad + (uv)^{\alpha - \gamma +1} \int^{1}_{0} (1+uw)^{\alpha -1 } (1 + vw)^{\alpha -1 } w^{-\gamma} {\mathrm d} w \\
& \le \frac{(uv)^{2H-1}}{2 - 2 H } + \frac{\,(uv)^{\alpha - \gamma +1}\,}{\,1 - \gamma \,}\,.
\end{split}
\end{equation*}
Substituting these upper bounds of both terms and using again the inequality $\,(x+y)^{2} \le 2 (x^{2} + y^{2})\,$, $x, y > 0 $ for the second term, we obtain the estimates
\begin{equation*}
\begin{split}
\int^{T}_{0} \int^{v}_{0} (f_{1}(u,v))^{2} {\mathrm d} u {\mathrm d} v \le &\, [\text{Beta} ( \alpha , 1 - \gamma) ]^{2}\int^{T}_{0} \bigg( \int^{v}_{0} u^{2\alpha - 2 \gamma}v^{2\alpha - 2} {\mathrm d} u \bigg){\mathrm d} v
\\
\, = & \, \frac{\,[\text{Beta} ( \alpha , 1 - \gamma) ]^{2} \, T^{4 \alpha - 2 \gamma}\,}{\,2(2\alpha - \gamma) (2\alpha - 2 \gamma +1)\,} < \infty \, ,
\end{split}
\end{equation*}
and
\begin{equation*}
\begin{split}
& \int^{T}_{0} \int^{v}_{0} (f_{2}(u,v))^{2} {\mathrm d} u {\mathrm d} v \\
\le & \int^{T}_{0} \bigg[ \int^{v}_{0} \bigg( \frac{(uv)^{2H-1} }{2 -2H } + \frac{\,(uv)^{\alpha - \gamma +1}\,}{\,1 - \gamma \,} \bigg)^{2} {\mathrm d} u \bigg] {\mathrm d} v
\\
\le&\, 2\, \int^{T}_{0} \int^{v}_{0} \bigg( \frac{(uv)^{4 H-2} }{(2-2H)^{2} } + \frac{\,(uv)^{2\alpha - 2\gamma +2}\,}{\,(1 - \gamma )^{2} \,} \bigg) {\mathrm d} u {\mathrm d} v
\\
\, =& \, \frac{\,T^{8H+1}\,}{\,(2 - 2H)^{2} (4H-1)(4H+1)\,} + \frac{\,T^{4H-2 \gamma +4}\,}{\,(1-\gamma)^{2} (4 H - 2 \gamma +3) (2 H- \gamma + 2)\,} \\
\, <& \, \infty \, .
\end{split}
\end{equation*}
The right hand sides are finite when $\, 2 \alpha > \gamma \,$ and $\, 0 < \gamma < 1\,$.
Therefore, combining these estimates with \eqref{eq: lm4.1.2}, we conclude the second derivative $\, K(u,v) \,$ is square integrable in $\,(0, T) \times (0, T) \,$.
\noindent (ii) Suppose that $\, -1/2+ \gamma/2< \alpha \le \gamma/2 \,$ and $\, 0 < \gamma < 1\,$.
Since $\, (x+y)^{2} \ge x^{2} \,$ for $\,x, y > 0 \,$, we shall show
\[
\int^{T}_{0} \int^{T}_{0} [K(u,v)]^{2} {\mathrm d} u {\mathrm d} v \, \ge \, 2 \int^{T}_{0} \int^{v}_{0} [f_{1}(u,v)]^{2} {\mathrm d} u {\mathrm d} v \, =\, \infty \, .
\]
To do so, by the change-of-variable and by Jensen's inequality, we observe that if $\, u \le v \,$,
\begin{equation*}
\begin{split}
f_{1}(u,v) \, =\, & u^{\alpha - \gamma } \int^{1}_{0} (v - u w)^{\alpha -1}w^{-\gamma} ( 1- w)^{\alpha -1} {\mathrm d} w \\
\, =\, & u^{\alpha - \gamma } \text{Beta} (\alpha , 1 - \gamma) \int^{1}_{0} (v - u w)^{\alpha -1}\frac{ w^{-\gamma} ( 1- w)^{\alpha -1}}{ \text{Beta} (\alpha , 1 - \gamma)} {\mathrm d} w\\
\ge \, & \text{Beta} (\alpha , 1 - \gamma)\, u^{\alpha - \gamma}\, \bigg ( v - u \cdot \frac{\,\alpha \,}{\,\alpha + 1 - \gamma \,} \bigg)^{\alpha - 1} \, ,
\end{split}
\end{equation*}
because $\, w \mapsto (v - u w)^{\alpha - 1}\,$, $\, 0 < w < 1\,$ is a convex function and the expectation of Beta distribution with parameters $\,(\alpha, 1 - \gamma) \,$ is $\, \alpha / ( \alpha + 1 - \gamma ) \,$. Thus, we have a lower bound for $\, \int^{v}_{0} [f_{1}(u,v)]^{2} {\mathrm d} u \,$, that is,
\begin{equation*}
\begin{split}
& \int^{v}_{0} [f_{1}(u,v)]^{2} {\mathrm d} u \\
\, \ge \, & \int^{v}_{0}[ \text{Beta} (\alpha , 1 - \gamma)]^{2}\, u^{2 (\alpha - \gamma)}\, \bigg ( v - u \cdot \frac{\,\alpha \,}{\,\alpha + 1 - \gamma \,} \bigg)^{2 (\alpha - 1)} {\mathrm d} u \\
= \, & [ \text{Beta} (\alpha , 1 - \gamma)]^{2} v^{4 \alpha - 2 \gamma -1} \int^{1}_{0} \bigg( 1 - \theta \cdot \frac{\,\alpha \,}{\,\alpha + 1 - \gamma\,} \bigg)^{2 \alpha - 2 } \theta ^{2\alpha - 2 \gamma} {\mathrm d} \theta\\
\, = \, & [ \text{Beta} (\alpha , 1 - \gamma)]^{2} \cdot \bigg( \frac{\,1 - \gamma \,}{\, \alpha + 1 - \gamma \,}\bigg)^{2\alpha - 2} \frac{ v^{4 \alpha - 2 \gamma -1} }{2 \alpha - 2\gamma + 1 } \,
\end{split}
\end{equation*}
for $\, 0 < v < T\,$. However, if $\, 2 \alpha \le \gamma \,$, this lower bound is not integrable over $\, (0, T ) \,$, and hence, $\, K(\cdot, \cdot) \,$ is not square integrable over $\, (0, T) \times (0, T) \,$.
\end{proof}
\bibliographystyle{plain}
|
\section*{acknowledgments}
This work is supported by NSFC under Grant No.11674058 and NCET-13-0734.
|
\section{Introduction}
Seyfert 1 galaxies and quasars vary in the optical and X-ray continua on timescales of hours to decades. They show typical root mean square (rms)
variabilities of 10–20 percent and, on occasion, changes by a factor up to 2 in their optical luminosity \citep{rumbaugh18} on timescales of years.
Some extremely variable active galactic nuclei (AGN) not only show changes in optical luminosity, but also change their optical spectral type. These optical changing-look AGN exhibit transitions from type 1 to type 2 or vice versa on timescales of months to years. Type 1 AGN show both broad and narrow emission lines in their optical spectra, while type 2 AGN only have narrow emission lines. The optical spectral classification can change as a result of a variation in the accretion rate, accretion disk instabilities, or a variation in absorption, amongst others. Some AGN have shown variability amplitudes by a factor of more than 20 in the X-rays \citep[e.g.,][]{grupe01,grupe10}.
AGN that show X-ray spectral variations, that is,\ when a Compton-thick AGN becomes Compton-thin and vice versa, are also called changing-look AGN \citep{guainazzi02}.
To date, a few dozen Seyfert galaxies and quasars have been known to have changed their optical spectral type, for example NGC\,3516 (\citealt{souffrin73}, \citealt{mehdipour22}), NGC\,7603 \citep{tohline76,kollatschny00}, NGC\,4151 \citep{penston84}, Fairall\,9 \citep{kollatschny85}, NGC\,7582 \citep{aretxaga99}, NGC\,2617 \citep{shappee14}, Mrk\,590 \citep{denney14}, HE\,1136-2304 \citep{parker16, zetzl18, kollatschny18}, Mrk\,1018 \citep{mcelroy16,husemann16,kim18,lyu21}, 1ES\,1927+654 (\citealt{trakhtenbrot19}, \citealt{ricci20}, \citealt{laha22}), NGC\,1566 \citep{Oknyansky19, parker19}, and references therein. Further recent findings are based on spectral variations detected in large-scale surveys, such as the Sloan Digital Sky Survey \citep[SDSS; e.g.,][]{komossa08, lamassa15, rumbaugh18, macleod19}, the Catalina Real-time Transient Survey \citep{graham20},
or the Wide-field Infrared Survey Explorer \citep{stern18}. However, in most of these new findings, only a few spectra of individual galaxies have been acquired to confirm their changing-look character. A recent review on changing-look AGN is given by \cite{komossa22}.
Here, we discuss an outburst of {IRAS\,23226-3843}{} that occurred in 2019. {IRAS\,23226-3843}{} is listed as a bright X-ray source in the ROSAT RASS catalog \citep[e.g.,][]{grupe01}. Other designations for {IRAS\,23226-3843}{} are CTS\,B11.01, 2MASX\,J23252420-3826492, 6dF J2325242-382649, and HE\,2322-3843. {IRAS\,23226-3843}{} is listed as a galaxy in the IRAS Faint Source Catalogue \citep{moshir90}. It has been detected at 60$\mu$m and 100$\mu$m with flux density levels of 0.584\,Jy and 1.49\,Jy (corresponding to $3.5\times10^{10}$ L$_\odot$). It is a radio source and was listed in the NRAO VLA Sky Survey (NVSS) with a flux density level of 4.3\,mJy \citep{condon98}. {IRAS\,23226-3843}{} was classified as a Seyfert type 1 object in an investigation of southern IRAS galaxies in 1991 \citep{allen91}. An apparent magnitude of $13.68\pm 0.10$ was measured for this object in the r band in the Las Campanas Redshift Survey \citep[1991-1996;][]{shectman96}. This
corresponds to an absolute magnitude of M$_{r} = -22.33 \pm 0.51$\,mag. The morphology of {IRAS\,23226-3843}{} has been
classified to be of a S0 type based on UKST sky survey images \citep{loveday96}. Figure~\ref{legsurveyIRAS23_r_ds9_3.jpg} displays an r-band image of {IRAS\,23226-3843}{} based on the
Legacy Survey \citep{dey19} indicating a more complex morphology with weak spiral arms.
\begin{figure}
\centering
\hspace*{-15mm}
\includegraphics[width=12.cm,angle=0]{legsurveyIRAS23_r_ds9_3.pdf}
\caption{DESI legacy image (r band) of {IRAS\,23226-3843}{}. North is to the top, and east is to the left. The scale is 0.44 x 0.44 arcmin corresponding to 18.1 x 18.1 kpc.}
\label{legsurveyIRAS23_r_ds9_3.jpg}
\end{figure}\\
We discovered a very strong X-ray decline of {IRAS\,23226-3843}{} within the XMM-\textit{Newton}{} slew survey in 2017. Subsequently, we carried out deep follow-up \textit{Swift}{}, XMM-\textit{Newton}{}, and NuSTAR{} observations in combination with optical spectra of {IRAS\,23226-3843}{} in 2017 to study its variability behavior. In addition, we reinspected optical, UV, and X-ray data that were taken in the past \citep{kollatschny20}. We found that {IRAS\,23226-3843}{} decreased in X-rays by a factor of more than 30
with respect to ROSAT and \textit{Swift}{} data taken 10 to 27 years earlier.
In August 2019, we detected a strong increase in X-ray and UV fluxes in {IRAS\,23226-3843}{} with \textit{Swift}{} \citep{grupe19}. Thereupon, we carried out optical follow-up spectroscopy with SALT and with the SAAO 1.9\,m telescope, as well as XMM-\textit{Newton}\ and NuSTAR{} observations. Here, we describe and discuss optical and X-ray observations taken between 2019 and 2021. We discuss these results together with findings based on earlier epochs.
Throughout this paper, we assume a lambda cold dark matter ($\Lambda$CDM) cosmology with a Hubble constant of H$_0$~=~73~km s$^{-1}$ Mpc$^{-1}$, $\Omega_{\Lambda}=0.72$, and $\Omega_{\rm M}=0.28$. With a redshift of {\bfseries $z=0.0359$}, this results in a luminosity distance of $D_{\rm L} = 144$\,Mpc for {IRAS\,23226-3843}{} ($\alpha_{2000} =$ 23h 25m 24.2s, $\delta = -38^{\circ}$ 26$^{'}$ 49.2$^{''}$) using the Cosmology Calculator developed by \cite{wright06}.
\section{Observations and data reduction}
A dramatic increase in the X-ray and UV fluxes of {IRAS\,23226-3843}{} was discovered with \textit{Swift}{} on August 11, 2019. We carried out follow-up studies of {IRAS\,23226-3843}{} with \textit{Swift}{} as well as with XMM-\textit{Newton}\ and NuSTAR{} as part of a ToO program to study outbursts of radio-quiet AGN. Furthermore, we took additional optical spectra with SALT as well as with the SAAO 1.9\,m telescope to investigate the optical spectral variations.
\subsection{\bf X-ray, UV, and optical continuum observations with \textit{Swift}{}}
\label{sec:swift_observations}
After the discovery of the increasing X-ray flux in 2019 August, we started monitoring {IRAS\,23226-3843}{} with the NASA Neil Gehrels Gamma-Ray Burst Explorer Mission \textit{Swift}{} \citep{gehrels04} in X-rays and the UV/optical. All Swift observing dates and exposure times are listed in Table\,\ref{swiftlog}. All observations in the 0.3-10\,keV band were performed with \textit{Swift}'s X-ray Telescope \citep[XRT;][]{burrows04} in photon counting mode \citep[pc mode;][]{hills05}. Spectra were extracted in a circular region with a radius of 15 pixel (equal to 35$^{\prime\prime}$) for the source region and 100 (equal to 235$^{\prime\prime}$) for the background region using the FTOOL {\it XSELECT}. Using {\it xrtmkarf}, an auxiliary response file (arf) was created and combined with the XRT response file {\it swxpc0to12s6\_20130101v014.rmf}.
Except for some brighter spectra during the earlier times, generally the spectra were not binned, and analyzed using w-statistics \citep{cash79} as indicated in Table\,\ref{swiftdata}. This table also lists the count rate, hardness ratio\footnote{Here we define the hardness ratio as $HR = \frac{H-S}{H+S}$ were S are the counts in the 0.3-1.0 keV and H in the 1.0-10.0\,keV bands. Hardness ratios were determined applying Bayesian statistics with the program {\it BEHR} by \citet{park06}.}, photon index $\Gamma$, the absorption corrected 0.3-10\,keV flux, and the goodness of the fit.
Typically, observations in the UV-Optical Telescope \citep[UVOT]{roming05} were performed in all 6 photometric filters\footnote{ UVW2 (1928 \AA), UVM2(2246 \AA), UVW1 (2600 \AA), U (3465 \AA), B (4392 \AA), and V (5468 \AA)}. While most observations were performed in the single orbit, in case of multiple orbit observations during one segment, all snapshots were merged using the tool {\it uvotimsum}. The magnitudes in the Vega system and flux densities in each photometric filter were then measured with the tool {\it uvotsource} using the count rate conversion and calibration, as described in \citet{poole08} and \citet{breeveld10}. Because we applied a source extraction radius of 3", we set the {\it uvotsource} parameter {\it apercorr=curveofgrowth}. The fluxes and magnitudes listed in Table\,\ref{swiftmag} are corrected for Galactic reddening. We determined the Galactic reddening correction for each filter by applying equation 2 in \citet{roming05} who used the reddening curves published by \citet{cardelli89}. The attenuation in the direction of\,{IRAS\,23226-3843}{} is $E_{\rm B-V}=0.021$ \citep{schlafly11}, resulting in the following magnitude corrections: UVW2$_{\rm corr} = 0.20$, UVM2$_{\rm corr} = 0.244$,
UVW1$_{\rm corr} = 0.177$, U$_{\rm corr} = 0.136$, B$_{\rm corr} = 0.177$, and V$_{\rm
corr} = 0.083$.
\subsection{{\bf X-ray observations with XMM-\textit{Newton}\ and NuSTAR\ }}
The new XMM-\textit{Newton}\ and NuSTAR\ observations were taken as part of a ToO program (Proposal ID 084080, PI Schartel) aimed
at catching radio quiet AGN in maximum states \citep[see, e.g.,][]{parker16, parker19}. The observing program consisted of one long XMM-\textit{Newton}/NuSTAR\ exposure. In addition, we consider archival XMM-\textit{Newton}\ and NuSTAR\ observations
which were taken as part of a ToO program (Proposal ID 76002, PI Schartel) aimed at catching radio quiet AGN in minimum states. The details of these observations are given in Table~\ref{table:xray_observations.}.
\begin{table*}
\centering
\caption{ Details of the XMM-\textit{Newton}\ and NuSTAR\ X-ray observations in 2019.}
\label{table:xray_observations}
\begin{tabular}{l c c c c}
\hline \hline
\noalign{\smallskip}
Mission & Obs. ID & MJD & Start date & Exp. time\\
& & & & [ks] \\
\hline
XMM-\textit{Newton}\ & 0840800501 & 58795 & 2019-11-07 22:51:54 & 88.29 \\
NuSTAR\ & 80502607002 & 58795 & 2019-11-07 19:10:11 & 54.83 \\
\hline
\end{tabular}
\tablefoot{
Exposure times are the final clean exposures after filtering for background flares.
}
\label{table:xray_observations.}
\end{table*}
We reduced the XMM-\textit{Newton}\ EPIC-pn data using the \texttt{epproc} tool from the XMM-\textit{Newton}\ Science Analysis Software (SAS) version 20.0.0. We extracted source photons from a 20$^{\prime\prime}$ circular extraction region centered on the source PSF, and background photons from a larger circular region on the same chip, avoiding contaminating sources, detector edges, and the high copper background region of the chip. We also reduced the RGS data; however, we found no evidence for absorption or emission line features in the spectra so we do not consider it further in this work.
We reduced the NuSTAR\ data using the NuSTAR\ Data Analysis Software (NuSTARDAS) version 2.1.0, included as part of HEASOFT version 6.29. We extracted source photons from a 30$^{\prime\prime}$ circular extraction region centered on the source PSF, and background photons from a larger circular region on the same chip.
We rebinned all spectra using the \texttt{ftgrouppha} tool, following the optimal binning scheme of \citet{Kaastra16} and with an additional requirement of at least 30 counts per bin. We fit the spectra using \textsc{xspec} version 12.12.0.
\subsection{Optical spectroscopy}
\begin{table}
\centering
\tabcolsep+1.5mm
\caption{Log of spectroscopic observations of {IRAS\,23226-3843}{}.}
\begin{tabular}{c c r l l}
\hline \hline
\noalign{\smallskip}
MJD & UT Date & \mcc{Exp. time} & Tel. & Obs. cond.\\
& & \mcc{[s]} & \\
\hline
50724 & 1997-10-03 & 1800 & SAAO 1.9\,m & clear; 1\farcs{}8 \\
51350 & 1999-06-21 & 600 & Cerro-Tol. 4\,m & clouds; -- \\
57883 & 2017-05-10 & 900 & SALT & clear; 1\farcs{}5 \\
57916 & 2017-06-12 & 900 & SALT & clear; 2\farcs{}0 \\
58727 & 2019-09-01 & 2400 & SAAO 1.9\,m & clear; 1\farcs{}8\\
58736 & 2019-09-10 & 1200 & SALT & clouds; 2\farcs{}0 \\
58750 & 2019-09-24 & 900 & SALT & clear; 1\farcs{}5 \\
58794 & 2019-11-07 & 900 & SALT & clear; 1\farcs{}4 \\
58823 & 2019-12-06 & 900 & SALT & clear; 1\farcs{}3 \\
59053 & 2020-07-23 & 2400 & SAAO 1.9\,m & clear; 1\farcs{}8 \\
\hline
\vspace{-.7cm}
\end{tabular}
\tablefoot{
The observations have been taken with SALT in 2019 as well as with the SAAO 1.9\,m telescope in 1997, 2019 and 2020. In addition, we give the observing dates already presented in \cite{kollatschny20} (Cerro-Tololo 4\,m: 1999; SALT: 2017). We list the Modified Julian date, the UT date, the exposure time, the telescope, and observing conditions.
}
\label{saltlog}
\end{table}
\subsubsection{Spectroscopy with SALT}
After the detection of the increasing X-ray flux, we took four optical spectra of {IRAS\,23226-3843}{} with the Southern African Large Telescope (SALT) between 2019 September and December. One optical spectrum was secured simultaneously to the deep X-ray observation with XMM-\textit{Newton}{} on 2019 November 07. The log of our optical spectroscopic observations with SALT is given in Table~\ref{saltlog}. In addition, we present the observing data from 1999 and 2017 \citep{kollatschny20} and three observations taken with the SAAO 1.9\,m telescope (see \ref{sec:SAAO_1.9}).
The spectroscopic observations were taken under identical instrumental conditions with the Robert Stobie Spectrograph using the PG0900 grating. The slit width was fixed to 2\farcs{}0 projected onto the sky at an optimized position angle to minimize differential refraction. Furthermore, all observations were taken at the same air mass thanks to the particular design feature of the SALT. The spectra were taken with exposure times of 15 to 20 minutes (see Table~\ref{saltlog}). Seeing full width at half maximum (FWHM) values ranged between 1.3 to 2\arcsec.
We covered the wavelength range from 4330 to 7376~\AA\ at a spectral resolution of 6.5~\AA\ . The observed wavelength range corresponds to a wavelength range from 4180 to 7120~\AA\ in the rest frame of the galaxy. There are two gaps in the spectrum caused by the gaps between the three CCDs: one between the blue and the central CCD chip, and one between the central and red CCD chip covering the wavelength ranges 5347--5416~\AA\ and 6386--6454~\AA\
(5162--5228~\AA\ and 6165--6230~\AA\ in the rest frame ). All spectra were wavelength corrected to the rest frame of the galaxy ($z=0.0359$).
In addition to the galaxy spectra, we also observed necessary flat-field and ThAr and Xe arc frames, as well as a spectrophotometric standard star for flux calibration (LTT\,1020). The spatial resolution per binned pixel was
0\farcs{}2534. We extracted eight columns from our object spectrum, corresponding to 2\farcs{}03. The reduction of
the spectra (bias subtraction, cosmic-ray correction, flat-field correction, 2D wavelength calibration, night sky subtraction, and flux calibration) was made in a homogeneous way with IRAF
reduction packages \citep[e.g.,][]{kollatschny01}.
Great care was taken to ensure high-quality intensity and wavelength calibrations to keep the intrinsic measurement errors very low, as described in \citet{kollatschny01,kollatschny03,kollatschny10}. The AGN spectra (and our calibration star spectra) were not always taken under photometric conditions. Therefore, we calibrated the spectra
to the same absolute [\ion{O}{iii}]\,$\lambda$5007 flux of $9.27 \times 10^{-15} \rm erg\,s^{-1}\,cm^{-2}$ taken on 2017 May 10, under clear conditions. The flux of the narrow emission line [\ion{O}{iii}]\,$\lambda$5007 is considered to be constant on timescales of years to decades.
\subsubsection{Spectroscopy with the SAAO 1.9\,m telescope} \label{sec:SAAO_1.9}
The spectra taken on 1997 October 03, 2019 September 01 and 2020 July 23 were obtained with the 1.9\,m telescope at the South African Astronomical Observatory in Sutherland (South Africa). The first spectrum was recorded with the CCD spectrograph in use at the time, whereas we utilized the improved SpUpNIC spectrograph \citep{Crause19} for the 2019 and 2020 observations. All observations used a low resolution grating (300 lines/mm, dispersion of 2.7 \AA/pixel) set at an angle to record the full optical range and using a slit width corresponding to 2\farcs{}7 on the sky. The AGN spectrum taken in 1997 was extracted over a spatial region of 5\arcsec. The spectra taken in 2019 and 2020
were extracted over 10\arcsec.
The wavelength was determined by means of observations of the emission spectrum produced by an argon lamp. The flux calibration was achieved through observations on the same night of spectrophotometric standard stars from the compilation of \citet{Hamuy94} (LTT\,9491, LTT\,7379, LTT\,377). We note that due to seeing fluctuations and possible thin cloud presence the flux calibration was later adjusted so that the [\ion{O}{III}] emission line strengths corresponded to those measured in the SALT spectra. Seeing was 2\arcsec or better for all observations, and weather conditions were generally good.
\section{Results}
\subsection{\bf X-ray, UV, and optical continuum band variations based on \textit{Swift}{}}
\label{sec:swift_results}
The \textit{Swift}{} 0.3--10\,keV, UV, and optical continuum light curves for the years 2007 to 2021 are shown in Fig.~\ref{fig_dirk_xrt_uvot_lc07_19.pdf} along with the X-ray and the hardness ratios (see Sect.~\ref{sec:swift_observations}). All measurements are listed in Tables~\ref{swiftdata} and
\ref{swiftmag}.
\begin{figure*}
\centering
\includegraphics[width=17cm,angle=0]{lc_Dirk_IRAS23226_20221026.pdf}
\vspace*{-5mm}
\caption{Combined X-ray, UV, and optical light curves taken with the \textit{Swift}{} satellite for the years 2007 until 2021. The fluxes are given in units of 10$^{-12}$ ergs s$^{-1}$ cm$^{-2}$. HR is the X-ray hardness ratio. The red lines indicate the dates of our deep XMM-\textit{Newton}{} observations. The observations for the year 2019 (shaded area) are presented in greater detail in Figure~\ref{fig_dirk_xrt_uvot_lc19.pdf}.}
\label{fig_dirk_xrt_uvot_lc07_19.pdf}
\end{figure*}
When \textit{Swift}\ observed {IRAS\,23226-3843}{} as part of our GI filling program on 2019 August 11, it found it in an X-ray high state. The AGN had increased its X-ray flux in the 0.3-10\,keV band by a factor of more than 30 compared with the low-state observation in 2017. On 2019 August 11, \textit{Swift}\ measured a flux of $1.18\times 10^{-11} $~erg~s$^{-1}$~cm$^{-2}$. In the UV, the flux increased by about 2 magnitudes as listed in Table~\ref{swiftmag}. Figure~\ref{fig_dirk_xrt_uvot_lc19.pdf} shows the X-ray, UV, and optical \textit{Swift}{} light curves
for the detailed campaign in 2019.
\begin{figure*}
\centering
\includegraphics[width=17.cm,angle=0]{lc_Dirk_zoom_IRAS23226_20221026.pdf}
\vspace*{-5mm}
\caption{Combined X-ray, UV, and optical light curves
taken with the \textit{Swift}{} satellite for the year 2019. The fluxes are given in units of 10$^{-12}$ ergs s$^{-1}$ cm$^{-2}$. Furthermore, the dates of our optical spectral observations are indicated as blue lines. We note that the deep XMM-Newton observation was performed on MJD 58795 (red line). }
\label{fig_dirk_xrt_uvot_lc19.pdf}
\end{figure*}
The UV and optical \textit{Swift}{} bands closely follow the 0.3--10\,keV X-ray light curve, which exhibits the strongest variability amplitudes and varies by a factor of more than 50. However, even though we do find changes in the hardness ratio over the years from 2007 until 2021, there is no systematic trend of HR with flux, as illustrated in Fig.~\ref{HR_X_flux_20221026.pdf}.
\begin{figure}
\centering
\includegraphics[width=\linewidth]{HR_X_flux_20221026.pdf}
\caption{Hardness ratio (HR) of our Swift observations as a function of X-ray flux. There is no systematic trend of HR with flux.}
\label{HR_X_flux_20221026.pdf}
\end{figure}
Table~\ref{swiftvarstatistics} gives the variability statistics based on the \textit{Swift}{} continua (XRT, W2, M2, W1,
U, B, and V).
%
\begin{table}
\centering
\tabcolsep+1.5mm
\caption{Variability statistics of the \textit{Swift}{} continua XRT, W2, M2, W1, U, B, and V.}
\begin{tabular}{lrrrccc}
\hline \hline
\noalign{\smallskip}
Band & F$_\text{max}$ & F$_\text{min}$ & R$_\text{max}$ & <F> & $\sigma_\text{F}$ & F$_\text{var}$ \\
\noalign{\smallskip}
\hline
\noalign{\smallskip}
XRT & $16.86$ & $0.33$ & $51.09$ & $5.48$ & $4.43$ & $0.79\pm0.09$ \\
W2 & $15.04$ & $2.03$ & $7.41$ & $6.38$ & $4.09$ & $0.63\pm0.07$ \\
M2 & $14.48$ & $1.98$ & $7.31$ & $6.27$ & $4.02$ & $0.64\pm0.08$ \\
W1 & $14.22$ & $2.31$ & $6.16$ & $6.51$ & $3.66$ & $0.56\pm0.07$ \\
U & $15.84$ & $4.69$ & $3.38$ & $9.03$ & $3.73$ & $0.41\pm0.05$ \\
B & $19.16$ & $4.68$ & $4.09$ & $14.32$ & $3.3$ & $0.19\pm0.03$ \\
V & $24.38$ & $15.67$ & $1.56$ & $19.57$ & $2.61$ & $0.12\pm0.02$ \\
\noalign{\smallskip}
\hline
\end{tabular}
\label{swiftvarstatistics}
\tablefoot{ In units of $10^{-12}$ erg s$^{-1}$ cm$^{-2}$.}
\end{table}
We give the minimum and maximum fluxes F$_{\rm min}$ and F$_{\rm max}$, peak-to-peak amplitudes R$_{\rm max}$ = F$_{\rm max}$/F$_{\rm min}$, the mean flux $<$F$>$ over the period of observations from MJD 54266.18 to 59313.26, the standard deviation $\sigma_{\rm F}$, and the fractional variation $F_{\rm var}$ as defined by \cite{rodriguez97}. The $F_{\rm var}$ uncertainties are defined in \cite{edelson02}. We observe a clear decrease of the peak-to-peak amplitude and the fractional variation with increasing wavelength.
\subsection{ Optical continuum and spectral line variations}
The optical spectra were taken with different telescopes, different spectrographs and also under different
atmospheric conditions. For an accurate comparison we therefore had to intercalibrate our spectra. The intercalibration of our spectra shown in Fig.~\ref{IRAS23226_all_spectra_20220326.pdf} was carried out with respect to the [\ion{O}{iii}]\,$\lambda$5007 line as well as with respect to the narrow H$\alpha${}, [\ion{N}{ii}], and [\ion{S}{ii}] lines as the flux of narrow emission lines is considered to be constant on timescales of years. The two SAAO 1.9\,m spectra that were integrated over a wider spatial region of 10\arcsec{} (2019-09-01 and 2020-07-23) contain a higher contribution from the host galaxy as well as additional non-nuclear line emission. We subtracted a continuum flux of 0.2 $\times$ 10$^{-15}$\,erg\,s$^{-1}$\,cm$^{-2}$\,\AA$^{-1}$ to correct for this additional host galaxy contribution.
A relative flux accuracy of only about 5\% was achieved for [\ion{O}{iii}]$\lambda$\,5007 in these nuclear emission line spectra because the spectra of {IRAS\,23226-3843}{} show an additional strong contribution from the underlying host galaxy and the narrow emission line region is slightly extended.
We present our reduced optical spectra of {IRAS\,23226-3843}{} for the years 1997 to 2020 in Fig.~\ref{IRAS23226_all_spectra_20220326.pdf}. The spectra show a strong contribution from the host galaxy with a clear stellar absorption imprint. The observed continuum fluxes at $\sim 5500-6000$\,\AA{} do not vary much between the epochs, except for the outburst spectra from 2019 September to 2019 December. However, the continuum gradient shows slight variations with a varying strength of the continuum slope.
The mean and rms spectra of {IRAS\,23226-3843}{} are shown in Fig.~\ref{IRAS23226_AVG_RMS_20221026_legend.pdf}. They are based on all spectra except for the two SAAO 1.9\,m spectra that were integrated over an extended region of 10\arcsec. The rms spectrum that presents the variable part of the spectra, that is to say the variable part of the emission lines \ in particular, was scaled by a factor of 6.5 to allow for a better comparison with the mean spectrum, and to enhance weaker line structures.
\begin{figure*}
\centering
\includegraphics[width=17cm,angle=0]{IRAS23226_all_spectra_20221116.pdf}
\vspace*{-3mm}
\caption{Reduced optical spectra of {IRAS\,23226-3843}{} for the years 1997 until 2020. The spectra were calibrated to the same absolute [\ion{O}{iii}]\,$\lambda$5007 flux of $9.27 \times 10^{-15} \rm erg\,s^{-1}\,cm^{-2}$ taken on 2017 May 10, under clear conditions. The three outburst spectra from 2019 are indicated by their observation date.}
\label{IRAS23226_all_spectra_20220326.pdf}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=17cm,angle=0]{IRAS23226_AVG_RMS_20221116.pdf}
\caption{Integrated mean (black) and rms (red) spectra of {IRAS\,23226-3843}{}. The rms spectrum has been scaled by a factor of 6.5 to enhance weaker line structures. The Balmer lines H$\alpha${} and H$\beta${} as well as the \ion{Fe}{ii} blends are indicated by black lines.}
\vspace*{-3mm}
\label{IRAS23226_AVG_RMS_20221026_legend.pdf}
\end{figure*}
The broad emission line spectrum is dominated by broad double-peaked H$\beta${} and H$\alpha${} emission lines, as well as permitted broad \ion{Fe}{ii} blends of the multiplets 37, 38 and 42, 48, 49. These broad H$\alpha${} and H$\beta${} lines as well as the \ion{Fe}{ii} blends are highly variable as can be seen in the rms spectrum in Fig.~\ref{IRAS23226_AVG_RMS_20221026_legend.pdf}. In addition to the broad lines, the typical narrow AGN emission lines of [\ion{O}{iii}]$\lambda\lambda\, 4595,5007$, [\ion{O}{i}]$\lambda\, 6300$, [\ion{N}{ii}]$\lambda\lambda\, 6548,6584$, [\ion{S}{ii}]$\lambda\lambda\, 6716,6731$, and narrow H$\alpha${} are present. Before determining the emission line intensities, we subtracted the strong contribution of the host galaxy, as well as the underlying power-law continuum by means of a spectral synthesis. The details of this synthesis are outlined in Sect.~\ref{sec:host_galaxy}. The wavelength ranges of the broad emission lines, of the continua as well as of the pseudo-continua we used to subtract a continuum below the broad Balmer lines and the \ion{Fe}{ii} blends are given in Table~\ref{wavelength-ranges}.
\begin{table}
\centering
\caption{Wavelength ranges of the (pseudo-)continua, the Balmer lines, and \ion{Fe}{ii} blends in the rest-frame.}
\label{wavelength-ranges}
\begin{tabular}{lcc}
\hline \hline
Continuum, Line & Wavelength range \\
& [\AA] \\
\hline
Cont 4230 & 4210 -- 4245 \\
Cont 5040 & 5022 -- 5056 \\
Cont 5545 & 5520 -- 5570 \\
Cont 5965 & 5920 -- 6010 \\
Cont 7020 & 6990 -- 7050 \\
\hline
\ion{Fe}{ii}(37,38) & 4390 -- 4630 \\
H$\beta_{\rm broad}$ & 4770 -- 5019 \\
\ion{Fe}{ii}(42,48,49) & 5057 -- 5519 \\
H$\alpha_{\rm broad}$ & 6230 -- 6850 \\
H$\alpha_{\rm inner}$ & 6439 -- 6776 \\
\hline
\end{tabular}
\tablefoot{
H$\alpha_{\rm inner}$ has the same extent in velocity space as H$\beta_{\rm broad}$.
}
\end{table}
We list in Tab.~\ref{BLR_variability} the extinction corrected observed continua fluxes at $4230\pm20$\,\AA{} and $5040\pm20$\,\AA{}, as well as the broad-line fluxes of H$\alpha${} and H$\beta$. For H$\alpha${}, we also give the flux of the inner region (H$\alpha_{\rm inner}$) which has the same extent in the velocity space as H$\beta_{\rm broad}$ ($-5640$\,km\,s$^{-1}${} to $+7800$\,km\,s$^{-1}${}). In addition, we list the flux of the \ion{Fe}{ii} blend (42, 48, and 49). Here, we linearly interpolated the flux for the gap between 6165 and 6230\,\AA{}. The broad H$\beta_{\rm broad}$ line has been corrected for the contribution of the narrow line fluxes of H$\beta_{\rm narrow}$, [\ion{O}{iii}]$\lambda\,$4959, and [\ion{O}{iii}]$\lambda\,$5007.
The broad H$\alpha_{\rm broad}$ line as well as the inner H$\alpha_{\rm inner}$ line have been corrected for the contribution of the narrow line fluxes of H$\alpha_{\rm narrow}$, [\ion{O}{i}]$\lambda$\,6300, [\ion{N}{ii}]$\lambda\lambda\,6548, 6584$, and [\ion{S}{ii}]$\lambda\lambda$\,6716, 6731.
In addition, we present the Balmer decrement H$\alpha${}/H$\beta${} based on the inner
H$\alpha_{\rm inner}$ region ($-5640$\,km\,s$^{-1}$ up to $+7800$\,km\,s$^{-1}${}). The flux of the \ion{Fe}{ii}(37,38) blend exhibits the same trend of showing a maximum at 2019-11-07 similar to the \ion{Fe}{ii}(42,48,49) blend. However, the \ion{Fe}{ii}(37,38) blend holds a major error caused by the extrapolation of the spectral synthesis and is therefore not listed.
\begin{table*}
\caption{Fluxes of the optical continua at $4230\pm20$\AA{} and $5040\pm20$\AA{},
the broad-line intensities of H$\alpha${} and H$\beta${} after subtraction of the narrow line components,
the flux of the \ion{Fe}{ii} blend (42, 48, and 49), and the Balmer decrement H$\alpha$/H$\beta${} for the
inner line region. }
\label{BLR_variability}
\begin{tabular}{l c c c c c c c}
\hline \hline
UT-Date & Cont$_{4230}$ & Cont$_{5040}$ & H$\beta${} & \ion{Fe}{ii}(42,48,49) & H$\alpha${} & H$\alpha_{\rm inner}$ & (H$\alpha${} / H$\beta${})$_{\rm inner}$\\
\hline
1997-10-03 & 1.54 $\pm$ 0.29 & 1.53 $\pm$ 0.29 & 23.7 $\pm$ 4.0 & 17.6 $\pm$ 8.0 & 104.0 $\pm$ 17.0 & 92.9 $\pm$ 16.0 & 3.9 $\pm$ 1.0 \\
1999-06-21 & 1.50 $\pm$ 0.23 & 1.51 $\pm$ 0.23 & 24.2 $\pm$ 3.5 & 28.6 $\pm$ 8.0 & 100.5 $\pm$ 14.0 & 89.4 $\pm$ 13.0 & 3.7 $\pm$ 0.8 \\
2017-05-11 & 1.11 $\pm$ 0.12 & 1.42 $\pm$ 0.15 & 13.2 $\pm$ 3.0 & 10.1 $\pm$ 6.0 & 93.8 $\pm$ 13.0 & 73.9 $\pm$ 12.0 & 5.6 $\pm$ 1.6 \\
2017-06-13 & 1.24 $\pm$ 0.13 & 1.47 $\pm$ 0.15 & 13.2 $\pm$ 3.0 & 7.7 $\pm$ 6.0 & 96.4 $\pm$ 13.0 & 75.0 $\pm$ 12.0 & 5.7 $\pm$ 1.6 \\
2019-09-01 & 1.58 $\pm$ 0.27 & 1.57 $\pm$ 0.27 & 13.1 $\pm$ 4.0 & 12.5 $\pm$ 7.0 & 86.1 $\pm$ 19.0 & 70.9 $\pm$ 17.0 & 3.7 $\pm$ 2.2 \\
2019-09-10 & 1.58 $\pm$ 0.39 & 1.60 $\pm$ 0.40 & 11.8 $\pm$ 6.0 & 31.7 $\pm$ 12.0 & 115.1 $\pm$ 27.0 & 95.8 $\pm$ 24.0 & 8.1 $\pm$ 4.7 \\
2019-09-24 & 2.32 $\pm$ 0.14 & 2.33 $\pm$ 0.14 & 26.3 $\pm$ 2.0 & 44.6 $\pm$ 5.0 & 181.0 $\pm$ 9.0 & 148.0 $\pm$ 8.0 & 5.6 $\pm$ 0.6 \\
2019-11-07 & 2.46 $\pm$ 0.15 & 2.51 $\pm$ 0.16 & 31.0 $\pm$ 2.0 & 50.2 $\pm$ 5.0 & 199.6 $\pm$ 9.0 & 164.0 $\pm$ 8.0 & 5.3 $\pm$ 0.5 \\
2019-12-06 & 1.96 $\pm$ 0.12 & 2.12 $\pm$ 0.13 & 14.7 $\pm$ 2.0 & 35.4 $\pm$ 5.0 & 155.8 $\pm$ 9.0 & 134.0 $\pm$ 8.0 & 9.1 $\pm$ 1.4 \\
2020-07-23 & 1.39 $\pm$ 0.20 & 1.46 $\pm$ 0.20 & 5.6 $\pm$ 6.0 & 2.5 $\pm$ 4.0 & 18.1 $\pm$ 20.0 & 15.1 $\pm$ 19.0 & -- \\
\hline
mean values & 1.66 $\pm$ 0.07 & 1.75 $\pm$ 0.04 & 17.7 $\pm$ 2.7 & 24.1 $\pm$ 3.1 & 115.0 $\pm$ 8. & 95.9 $\pm$ 8. & 5.7 $\pm$ 1.1 \\
\hline
ratio(11-19/06-17) & 1.98 & 1.71 & 2.35 & 6.52 & 2.07 & 2.19 & 0.93 \\
\hline
\end{tabular}
\tablefoot{
Continuum flux densities in units of 10$^{-15}$\,erg\,s$^{-1}$\,cm$^{-2}$\,\AA$^{-1}$. Line fluxes in units 10$^{-15}$\,erg\,s$^{-1}$\,cm$^{-2}$.
}
\end{table*}
\begin{figure*}
\centering
\includegraphics[width=1.0\linewidth,angle=0]{lightcurves_IRAS23226_withline_20220725.pdf}
\caption{Long-term light curves of the continuum flux density at 5040\,\AA{} (in units of 10$^{-15}$\,erg\,s$^{-1}$\,cm$^{-2}$\,\AA$^{-1}$), as well as of the line
fluxes of H$\alpha$, H$\beta$, and \ion{Fe}{ii}(42,48,49) (in units 10$^{-15}$\,erg\,s$^{-1}$\,cm$^{-2}$) for the years 1997 until 2020. The right panel shows, in addition to the observations from 2017 and 2020 the variations in 2019 in more detail. The epochs of the deep XMM-\textit{Newton}{} observations are indicated by a red arrow.}
\vspace*{-3mm}
\label{Ochmlightcurves_IRAS23226_20220329.pdf}
\end{figure*}
We show the long-term light curves of the
continuum flux density at 5040\,\AA{}, as well as of the line fluxes of
H$\alpha${}, H$\beta${}, and \ion{Fe}{ii}(42,48,49) for the years 1997 until 2020 in Fig.~\ref{Ochmlightcurves_IRAS23226_20220329.pdf}. A second plot gives the
variations in 2019 in more detail. The epochs of the deep XMM observations recorded in parallel at 2017 June 11 and 2019 November 07 are indicated by a red arrow.
\subsection{\bf Host galaxy contribution determined by spectral synthesis}\label{sec:host_galaxy}
We carried out a spectral synthesis of the galaxy spectra using the Penalized Pixel-Fitting method (pPXF) \citep{cappellari04,cappellari17}. This software extracts the stellar population from absorption-line spectra of galaxies, using a maximum penalized likelihood approach. The host galaxy spectrum of {IRAS\,23226-3843}{} is highly contaminated by strong emission line blends. Therefore, we restricted our synthesis to those wavelength ranges that were (nearly) free from emission lines. These regions correspond to 4197 -- 4353\,\AA{}, 4373 -- 4440\,\AA{}, 5020 -- 5155\,\AA{}, 5476 -- 6165\,\AA{}, 6230 -- 6275\,\AA{}, and 6854 -- 7300\,\AA{}. We used a set of 53 high-quality stars from the Indo-US library as stellar templates \citep{valdes04, shetty15, guerou17}. This library provides high-enough resolution with regard to our spectral data, and fully covers the wavelength range of interest. We convolved the templates to an estimate of the average instrumental width and then used pPXF to fit the velocity dispersion for every spectrum. The pPXF method is particularly useful for correcting the emission line profiles for the underlying absorption features. In addition to the stellar templates, we added a power-law component $F_{\lambda} \propto \lambda^{-\alpha}$ to the fitting process in order to take into account the underlying non-stellar AGN continuum. For the exponent of the power law, we used a value of $\alpha = 0.76$. This value is based on the slope of the variable continuum in the rms spectrum (Fig.~\ref{IRAS23226_AVG_RMS_20221026_legend.pdf}).
We present the results of the spectral synthesis using pPXF for the observed spectrum of {IRAS\,23226-3843}{} taken on 2019 September 24 (red) in Fig.~\ref{wkPPXF20190924_1.pdf}. The green line is the combined pPXF fit for the stellar component (orange) and the additional power-law component (blue). The difference between the observed spectrum and the combined pPXF fit is the residuum spectrum (purple) that gives the clean emission line contribution.
\begin{figure*}
\centering
\includegraphics[width=1.0\linewidth,angle=0]{IRAS23226_20190924_pPXF_results.pdf}
\caption{Observed spectrum of {IRAS\,23226-3843}{} taken on 2019 September 24 (red) and synthesis fit with pPXF. The green line is the combined pPXF fit for the stellar component (orange) and the additional power-law component (light blue). The residuum spectrum (purple) i.e., the difference between the observed spectrum and the combined pPXF fit, gives the clean emission line contribution.}
\label{wkPPXF20190924_1.pdf}
\end{figure*}
\subsection{\bf Balmer emission-line profiles and their variations}
From the mean and rms spectrum presented in Fig.~\ref{IRAS23226_AVG_RMS_20221026_legend.pdf}, it is evident that the Balmer lines in {IRAS\,23226-3843}{} are very complex and broad. In particular, we see double-peaked profiles of H$\alpha${} and H$\beta${} that are a signature of an accretion disk. In addition, H$\alpha${} shows an underlying broad and blueshifted component. This very broad component is not identifiable in the H$\beta${} line due to the superposition with the strong \ion{Fe}{ii} blends on the blue side of H$\beta$. Notably, {IRAS\,23226-3843}{} switched to an optical Seyfert 2 spectrum in 2020 showing no longer evidence for broad emission lines.
There is an additional broad blue absorption component at $-10\,400\pm1000$\,km\,s$^{-1}${} in H$\alpha${} and H$\beta${} for all epochs except for the low states in 2017. This absorption was strongest in 1999. We show in Figure~\ref{OchmIRAS23226_19990621_Grupe_20220325.pdf} the emission-line residuum
spectrum of {IRAS\,23226-3843}{} from 1999 June 21 (purple), clearly indicating the additional Balmer
absorption components at $-10\,400$\,km\,s$^{-1}$.
\begin{figure}
\centering
\includegraphics[width=1.0\linewidth,angle=0]{IRAS23226_19990621_Grupe_20220808.pdf}
\caption{Residuum spectrum of {IRAS\,23226-3843}{} from 1991 June 21 indicating the Balmer absorption component in addition to the emission lines.}
\label{OchmIRAS23226_19990621_Grupe_20220325.pdf}
\end{figure}
The absorption seems to be stronger in H$\beta${} than in H$\alpha${} due to the strong blue emission wing in H$\alpha${}.
A comparison of the H$\alpha${} and H$\beta${} line profiles in velocity space is presented in Figure~\ref{OchmIRAS23226_veloplots_20220325.pdf}. As mentioned before, we first subtracted the spectrum of the host galaxy by means of a spectral synthesis using pPXF (see Sect.~\ref{sec:host_galaxy} and Fig.~\ref{wkPPXF20190924_1.pdf}) in order to obtain pure broad emission-line profiles.
\begin{figure}
\centering
\includegraphics[width=1.0\linewidth,angle=0]{IRAS23226_veloplots_20221116.pdf}
\caption{Line profiles of H$\alpha${} and H$\beta${} in velocity space
after subtraction of the host galaxy spectrum. Absorption components in the blue wing of the Balmer lines (i.e., flux below zero) are shaded in green.}
\label{OchmIRAS23226_veloplots_20220325.pdf}
\end{figure}
The double peaks are present at all epochs before 2020 -- except for the red component of H$\alpha${} in the years 1997 and 1999. Profiles with double peaks can be produced through Keplerian rotation of accretion disks \citep{chen89a}, and we probe this interpretation later. The relative velocity of these
components remained constant over all the years. The blue peak originates at a velocity of $-3150\pm150$\,km\,s$^{-1}${} and the red peak originates
at a velocity of $3200\pm150$\,km\,s$^{-1}${} in the mean profiles of both lines.
\begin{figure}
\centering
\includegraphics[width=1.0\linewidth,angle=0]{IRAS23226_velo_avgrms_20220329.pdf}
\caption{Normalized mean and rms profiles of H$\alpha${} (red) and H$\beta${} (blue) in velocity space.}
\label{OchmIRAS23226_velo_avgrms_20220329.pdf}
\end{figure}
We give the full width at half maximum (FWHM) and the full width at zero intensity (FWZI) of the mean and rms profiles of H$\alpha${} and H$\beta${} in Table~\ref{Balmerlinewidths}.
\begin{table*}
{\small
\caption{Balmer line widths FWHM and FWZI, as well as the respective left wing boundary and right wing boundary position, and positions of the blue and red peaks of the mean and rms profiles.}
\label{Balmerlinewidths}
\begin{tabular}{l c c c c c c c c}
\hline \hline
Line & FWHM(left) & FWHM(right) & FWHM & FWZI(left) & FWZI(right) & FWZI & blue peak & red peak\\
\hline
H$\alpha${}$_{\rm mean}$\ & $-5470\pm200$ & $5340\pm200$ & $10810\pm280$ & $-13570\pm400$ & $12200\pm400$ & $25770\pm570$ & $-3100\pm500$ & $3100\pm300$\\
H$\beta_{\rm mean}$\ & $-5680\pm300$ & $6500\pm1000$ & $12180\pm1040$ & $<-6130\pm300$ & $7380\pm500$ & $>13510\pm580$ & $-3200\pm500$ & $3340\pm500$\\
\hline
H$\alpha${}$_{\rm rms}$\ & $-3190\pm500$ & $4860\pm300$ & $8050\pm580$ & $-14120\pm400$ & $11770\pm500$ & $25890\pm640$ & $-2500\pm300$ & $3200\pm300$\\
H$\beta_{\rm rms}$\ & $-5170\pm300$ & $5640\pm400$ & $10810\pm500$ & $<-6630\pm400$ & $7800\pm1000$ & $>15630\pm1080$ & $-2900\pm400$ & $3520\pm400$\\
\hline
\end{tabular}}
\tablefoot{In units of {km\,s$^{-1}$}.}
\end{table*}
The FWHM of H$\beta$$_{\rm mean}$ and H$\alpha$$_{\rm mean}$ amounts to $\sim 11\,000 \pm 1000$\,km\,s$^{-1}$. The FWZI of H$\alpha${} amounts to $\sim 26\,000$\,km\,s$^{-1}$. We can only give a lower limit of $13\,500$\,km\,s$^{-1}${} for the FWZI of H$\beta${} due to the strong blue absorption and the blending of the blue wing with the \ion{Fe}{ii} blend.
A comparison of the individual H$\alpha${} and H$\beta${} line profiles shows differences in the profiles for different epochs. The H$\alpha${} profile was asymmetric with respect to $v=0$\,km\,s$^{-1}${} for the early epochs in 1997 and 1999. The red component was much weaker than the blue component at these epochs. In contrast, H$\alpha${} showed almost symmetrical double-peaked profiles for the later epochs in 2017 and 2019 independent of the integrated line flux. The H$\beta${} profile showed double-peaked profiles at all epochs.
\subsection{\bf X-ray spectrum }
The broadband X-ray spectrum is qualitatively similar to that in 2017, but a factor of 10 brighter across the whole bandpass (see Fig.~\ref{fig:xray_comparison}). A weak, smooth soft excess is present, along with a weak Fe K emission feature and a possible absorption feature in the Fe band (see Fig.~\ref{fig:xray_fe_ratio}) at a slightly higher energy (7.6~keV) than the possible feature seen in the 2017 data ($\sim7.3$~keV). If genuine, this would correspond to an outflow with velocity of $\sim0.12c$ (36000~km~s$^{-1}$). Highly ionized absorbers with outflow velocities higher than 10\,000~km~s$^{-1}$ have been defined as ultra-fast outflows \citep[UFOs;][]{tombesi10}. The soft X-ray band appears featureless, so we first test modeling it with a simple Comptonization model with no atomic features \citep[e.g.,][]{Petrucci18}. We find that the soft excess is well modeled with such a Comptonization model. A broadband fit with a power-law continuum, Comptonized soft excess \citep[\textsc{nthcomp}][]{Zdziarski96,Zycki99}, and Galactic absorption \citep[\textsc{tbabs};][]{Wilms00} gives a good fit to the combined EPIC-pn and NuSTAR spectrum ($\chi^2$/dof of $343/286=1.20$, or $219/208=1.05$ if we exclude the 5--8~keV band). While other models have been proposed for the soft excess, most commonly reflection \citep[e.g.,][]{Jiang19}, the lack of any structure in the soft excess means that more complex models will either be disfavored or unconstrained with the data available, so we assume a Comptonized origin.
\begin{figure}
\centering
\includegraphics[width=\linewidth]{xray_comparison.pdf}
\caption{Comparison of the broadband X-ray data in 2017 and 2019. The change in flux is a uniform increase of a factor of $\sim10$ across the bandpass, and both spectra show residuals in the Fe~K band. All spectra are corrected for the instrumental effective area. The gray line shows the 2017 data scaled up by a factor of 10 to match the 2019 data. Above $\sim3$~keV the agreement in spectral shape is very good, below 3~keV the 2019 data is relatively higher, with the difference increasing toward lower energies.}
\label{fig:xray_comparison}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{xray_fe_ratio.pdf}
\caption{Ratio of the XMM-\textit{Newton}\ and NuSTAR\ spectra to a power-law continuum and Comptonized soft excess model, in the Fe~K band. A moderately broadened emission line is visible, and a possible absorption feature is present between 7 and 8 keV in the EPIC-pn spectrum. The NuSTAR data are consistent with both features, but are noisier.}
\label{fig:xray_fe_ratio}
\end{figure}
While this simple model explains the continuum shape very well, it does not account for the high energy complexity. We therefore focus on modeling the residuals in the Fe~K band. We consider two main possibilities for the moderately broadened iron line, where the emission either is produced in reflection from the accretion disk \citep[e.g.,][]{Fabian89} or scattered from a wind \citep[e.g.,][]{Nardini15}. It is worth noting that the emission from these two processes is highly degenerate \citep[][]{Parker22}, and it is entirely possible that some contribution from both is present in the spectrum. For this reason, we also consider a hybrid model, where X-rays are first reflected from the disk and then pass through a wind which produces further Fe~K emission. We use \textsc{tbabs} \citep{Wilms00} with a column density of $1.58\times10^{20}$~cm$^{-2}$ \citep{HI4PI16} to model Galactic absorption, \textsc{nthcomp} \citep[][]{Zdziarski96,Zycki99} for a Comptonized soft excess, \textsc{relxill} and \textsc{xillver} \citep{Garcia13, Garcia14} for relativistic and distant reflection, and a table of disk wind spectra based on the simulations of \citet{Sim08,Sim10} \citep[this is the same version of the model used in][]{Parker22}.
In \textsc{xspec} format, the models are \textsc{tbabs $\times$ (xillver + gabs $\times$ (nthcomp + powerlaw + relxill))} for the reflection model, \textsc{tbabs $\times$ (xillver + diskwind $\times$ (nthcomp + powerlaw))} for the disk wind model, and \textsc{tbabs $\times$ (xillver + diskwind $\times$ (nthcomp + powerlaw + relxill))} for the hybrid model.
All three models give an acceptable fit to the data ($\chi^2$/dof of 287/277, 306/281, and 276/275 for reflection, disk wind, and hybrid models respectively), without leaving obvious residual structures.
Having established baseline fits, we run MCMC on each fit to explore the parameter space. We use a chain length of $5\times10^5$, 200 walkers, and a burn in period of $5\times10^4$ steps. We then calculate the most likely parameter values and one-sigma errors from the posterior distributions of the chains (Table~\ref{tab:xray_pars}). In Fig.~\ref{fig:xray_models} we show the different models used to fit the data, in each case we plot a random sample of 200 model spectra drawn from the MCMC chains, showing the uncertainty in spectral shape of the different model components.
\begin{figure*}
\centering
\includegraphics[width=0.45\linewidth]{MCMC_model_spectra_ref.pdf}
\includegraphics[width=0.45\linewidth]{MCMC_model_spectra_dw.pdf}
\includegraphics[width=0.45\linewidth]{MCMC_model_spectra_hyb.pdf}
\caption{Model spectra for the three different X-ray models, fit to the 2019 XMM-\textit{Newton}\ EPIC-pn spectrum. Top left shows the reflection model, top right the disk wind model, and bottom the hybrid model. In each case, we plot 200 model spectra with parameters randomly selected from the MCMC chains as transparent lines, and the model corresponding to the peak of the posterior distribution for each parameter as solid lines. This effectively plots the posterior distribution of models, showing the uncertainty in each model component. In each case, the points with error bars show the data, and the black lines show the total model spectra. For the disk wind and hybrid models, the disk wind component is included as a multiplicative component applied to all other spectra components apart from the distant reflection. In the reflection model we apply a Gaussian absorption line to all additive components aside from the distant reflection to account for the possible absorption feature.}
\label{fig:xray_models}
\end{figure*}
\begin{table*}
\centering
\caption{X-ray model parameters.}
\begin{tabular}{l l c c c r}
\hline
\hline
Model & Parameter & Reflection & Disk wind & Hybrid & Description (unit) \\
\hline
\textsc{tbabs} & $N_\mathrm{H}$ & $1.58\times10^{20}$ & $1.58\times10^{20}$ & $1.58\times10^{20}$ & Galactic column (cm$^{-2}$) \\
\\
\textsc{xillver} & $N$ & $(4.1_{-0.4}^{+5.2})\times10^-6$ &$(1.5_{-0.5}^{+0.4})\times10^{-5}$ & $(1.2_{-0.4}^{+0.5})\times10^{-5}$ & normalization$^2$\\
\\
\textsc{gabs} & $E$ & $7.6\pm0.1$ & - & - & line energy (keV)\\
& $N$ & $0.05_{-0.02}^{+0.03}$ & - & - & line depth\\
\\
\textsc{nthcomp} & $\Gamma$ & $3.0_{-0.2}^{+0.1}$ & $2.7\pm0.1$ & $3.1_{ - 0.3}^{ + 0.2}$ & photon index\\
& $kT_\mathrm{e}$ & $134_{-110}^{+244}$ &$>95$ & $4_{-2}^{+4}$ & electron temperature (keV)\\
& $N$ & $(3\pm1)\times10^{-4}$ & $(5.8_{-0.6}^{+2.7})\times10^{-4}$ & $(4_{-1}^{+2})\times10^{-4}$ & normalization \\
\\
\textsc{relxill} & $q$ & $<4.5$ & - & $<5$ & emissivity index\\
& $a$ & $<-0.45$ & - & $0.5_{-0.5}^{+0.3}$ & spin\\
& $i$ & $25_{-2}^{+13}$ & - & $80_{-3}^{+2}$ & inclination (degrees)\\
& $\log(\xi)$ & $2.97_{- 0.11}^{+0.05}$ & - & $2.86_{- 0.08}^{+ 0.17}$ & ionization (erg~cm~s$^{-1}$)\\
& $A_\mathrm{Fe}$ & $>8.4$ & - & $8.5_{-2.4}^{+1.4}$ & iron abundance (solar)\\
& $N$ & ${(6\pm1)}\times10^{-6}$ & - & ${(2.6_{-0.6}^{+0.9})}\times10^{-6}$ & normalization$^2$ \\
\\
\textsc{diskwind} & $\dot{M}$ & - & $0.05_{ - 0.03}^{ + 0.12}$ & $0.039_{-0.006}^{+0.012}$ & mass outflow rate ($\dot{M}_\mathrm{Edd}$)\\
& $f_v$ & - & $1.25^1$ & 2.0$^1$ & terminal velocity parameter\\
& $L_\mathrm{X}/L_\mathrm{Edd}$ & - & $< 0.11$& $<0.04$ & X-ray luminosity\\
& $i$ & - & $62^1$ & $74^1$ & inclination (degrees)\\
\\
\textsc{powerlaw} & $\Gamma$ & $1.74_{ - 0.03}^{ + 0.04}$ & $1.71\pm0.05$ & $1.76_{-0.04}^{+0.03}$ & photon index\\
& $N$ & $(11\pm1)\times10^{-4}$ & $(14_{-1}^{+2})\times10^{-4}$ & $(17\pm1)\times10^{-4}$ & norm. (keV$^{-1}$~cm$^{-2}$~s$^{-1}$) \\
\\
\textsc{constant} & $C_\mathrm{FPM}$ & $1.08\pm0.02$ & $1.07_{-0.01}^{+0.02}$ & $1.07\pm0.02$ & FPM/EPIC-pn ratio\\
\\
& $\chi^2$/dof & $287/277=1.03$ & $306/281=1.09$ & $276/275=1.00$ & fit statistic\\
\hline
\end{tabular}
\label{tab:xray_pars}
\begin{flushleft}
$^1$In this version of the disk wind model the $f_v$ and $\mu=cos(i)$ parameters do not interpolate well between grid points, so we fix these parameters to the closest grid point after the intial fit, and do not calculate errors. This issue is fixed in the new version of the model \citep{matzeu22}.
$^2$The normalizations of the \textsc{xillver} and \textsc{relxill models} are defined based on the flux of the incident spectrum, see \citet{Dauser16}.
\end{flushleft}
\end{table*}
We can estimate the significance of the possible absorption line by examining the change in fit statistic when including or excluding the Gaussian line in the reflection model (the disk wind model describes the absorption line, but also the emission and other features throughout the band, so cannot test for a single feature). Removing the Gaussian worsens the fit by $\Delta\chi^2=5$, for 2 fewer degrees of freedom in the model. This is not highly significant ($\sim2\sigma$), even without taking into account the look-elsewhere effect. Based on these results, we can conclude that some Fe emission from either a mildly relativistic wind or the inner accretion disk is required to explain the observed data, but we cannot conclusively determine which process dominates.
\section{Discussion}
\subsection{\bf X-ray, UV, and optical continuum variations based on \textit{Swift}}
The X-ray, UV, and optical continuum observations of {IRAS\,23226-3843}{} with \textit{Swift}{} all show the same general variability trend from 2007 to 2021 (see Figs.~\ref{fig_dirk_xrt_uvot_lc07_19.pdf} and \ref{fig_dirk_xrt_uvot_lc19.pdf}). The variability behavior can be summarized as follows:
We observe a general slow flux decrease by a factor of 5 in all bands from 2007 until July 2016.
Subsequently, {IRAS\,23226-3843}{} showed a decrease in X-ray flux by a factor of about 4
from July 2016 until July 2017.
Afterwards, the X-ray flux increased to even higher
intensity level than before by a factor of $\sim$ 30 until August 2019.
After a short decline, {IRAS\,23226-3843}{} showed an outburst in the continuum lasting from 2019 September until 2019 December (see Fig.~\ref{fig_dirk_xrt_uvot_lc19.pdf}). This outburst was observed spectroscopically by accompanying optical observations with the SAAO\,1.9\,m telescope and SALT (see Fig.~\ref{Ochmlightcurves_IRAS23226_20220329.pdf}) in 2019.
The overall flux increased again in December 2019 and finally decreased in 2021.
The decreasing and increasing variability trends were strongest in X-rays and less pronounced in the optical B and V bands. However, one has to consider that the U, B, and V \textit{Swift}{} bands were not corrected for the strong contribution of the host galaxy. In order to quantify the variability strength, we calculated the fractional variation $F_{\rm var}$ for each \textit{Swift} \ band in Sect.~\ref{sec:swift_results}.
The results for the X-ray, UV, and optical bands are presented in Tab.~\ref{swiftvarstatistics}. The fractional variations deduced from the \textit{Swift}{} observations from 2007 until 2021 follow the same trend, that is to say decrease in fractional variation with increasing wavelength, as those based on the shorter period from 2007 until 2017 only \citep{kollatschny20}. The fractional variations in the UV and optical bands in {IRAS\,23226-3843}{} are stronger than those in, for example, the changing-look AGN HE\,1136-2304 \citep{zetzl18}, and the prototype AGN NGC\,5548 \citep{edelson15, fausnaugh16}. More precisely, the fractional variations in {IRAS\,23226-3843}{} are stronger by a factor of 2 in comparison to the changing-look AGN HE\,1136-2304, and even more so with regard to NGC\,5548.
\subsection{\bf X-ray spectrum}
The 2019 broadband X-ray spectra of {IRAS\,23226-3843}{} are qualitatively very similar to the 2017 spectrum, but a factor of 10 brighter across the bandpass, with no large change in spectra hardness and no dramatic evolution of the soft excess sometimes seen in changing look AGN \citep[e.g.,][]{noda18,parker19}. The RMS spectrum is consistent with being flat, with no significant absorption or emission features visible.
In the previous paper we modeled the 2017 data with distant reflection, a Comptonized continuum, a warm absorber, and a phenomenological black body soft excess. The 2019 data is of higher quality due to the higher flux, and enabled us to rule out the presence of a warm absorber based on the featureless RGS spectrum. While it is possible that this component was present in 2017 and not in 2019, we consider it more likely that no warm absorber has been present throughout and this component was fitting some other aspect of the spectrum in 2017, such as the spectral complexity introduced by a more complex soft excess including Comptonization and reflection.
We have found low significance ($\sim2\sigma$ in each case) evidence for an absorption line from an outflow in both 2017 and 2019 spectra, at 7.3 and 7.6~keV (observed frame).
We see evidence for a complex iron emission line profile, broadened beyond the level that could be produced by distant reflection alone. We consider relativistic reflection and disk wind models for this emission, along with a combination of both. It is not obvious which model should be preferred: the hybrid model gives the best fit, but both wind and reflection components prefer high inclinations (74 and 80 degrees, respectively) which is hard to reconcile with the unabsorbed spectrum. The reflection model gives a lower inclination (25 degrees), which is consistent with the lack of absorption, but also requires retrograde spin and a highly super-solar iron abundance ($>8.4$), although we note that this may be due to the fixed disk density assumed, as high density reflection can result in lower iron abundances \citep[e.g.,][]{Tomsick18,Jiang19}. The pure disk wind model parameters are more reasonable, with an inclination of 62 degrees and no requirement for extreme parameters, but it also gives a worse fit to the data. This may simply be due to the more limited parameter space of this model relative to the reflection model.
\subsection{\bf Optical spectrum: Extremely broad and double-peaked Balmer profiles}
{IRAS\,23226-3843}{} was originally classified as a Seyfert 1 galaxy \citep{allen91} based on an optical spectrum taken between 1985 and 1990. It was of clear Seyfert type 1 in the years 1997 and 1999 (see Figs.~\ref{IRAS23226_all_spectra_20220326.pdf} and \ref{OchmIRAS23226_veloplots_20220325.pdf}), and the broad H$\alpha${} line profile was asymmetric at both observing epochs, showing
a strong blue peak originating at a velocity of $-3100\pm500$\,km\,s$^{-1}${} without a corresponding red peak. Later on, however,
H$\alpha${} showed almost symmetrical double-peaked profiles for all observing epochs in 2017 and 2019, independent of the integrated line flux. The red peak originated at a velocity of $3100\pm300$\,km\,s$^{-1}${}, hence at the mirrored velocity of the blue peak. In contrast to H$\alpha${}, the H$\beta${} profile shows a double-peaked profile for all epochs between 1997 and 2020. Taking into account the measurement uncertainties, the positions of the blue and red peaks in velocity space were identical in both H$\alpha${} and H$\beta${}, and they remained constant in all our spectra. In 2020, the spectral type of {IRAS\,23226-3843}{} changed to that of a Seyfert 2 type
(see Fig.~\ref{OchmIRAS23226_veloplots_20220325.pdf}).
During our campaign, the line profiles in the red and blue wings of H$\alpha${} and H$\beta${} varied with different amplitudes. This can be seen best in the rms line profiles (see Figs.~\ref{IRAS23226_AVG_RMS_20221026_legend.pdf}, \ref{OchmIRAS23226_veloplots_20220325.pdf}, and \ref{OchmIRAS23226_velo_avgrms_20220329.pdf}). H$\alpha${} shows considerably stronger variations in the red wing than in the blue wing from 1997 until 2017. Afterwards, the blue and red wing varied uniformly.
The broad Balmer line components were faint and barely visible in 2017 in comparison to the optical spectra taken in 1997, 1999, as well as in 2019 (see Fig.~\ref{OchmIRAS23226_veloplots_20220325.pdf}). Especially, the broad H$\beta${} component
was barely visible without correction for the host galaxy. It is usually presumed that broad double-peaked emission lines are generated in relativistic Keplerian disks of gas surrounding the central supermassive black hole, as models developed for this physical scenario produce such profiles \citep[e.g.,][]{chen89a,chen89b}. During the minimum state in 2017 as well as during the maximum state in 2019, the two Balmer lines (H$\alpha${} and H$\beta$) show the same symmetric double-peaked profile. The peak separation corresponds to 6200 km\,s$^{-1}${}. These values are consistent with the peak velocities of many other AGN emission line profiles previously fitted with Keplerian disk models \citep{eracleous94,eracleous03}. In cases with such peak separation values, these earlier studies achieved optimal fits by choosing an accretion disk inner radius of $\sim 500$ Schwarzschild radii.
In our present study, we derive higher Balmer decrement values in comparison to the low values in our earlier investigation in \cite{kollatschny18}. This is due to the H$\beta${} emission-line flux being weaker now after correction for the modeled stellar host component. Furthermore, we now take the additional absorption component on the blue side of H$\beta${} into account. The Case B recombination value of the Balmer decrement H$\alpha${}/H$\beta${} is 2.7 \citep{osterbrock06}. Here we derive Balmer decrement values of 3.7 up to 9, indicating dust extinction or optical
depth effects in the broad-line region.
The values for the Balmer decrement in {IRAS\,23226-3843}{} are similar to those in other AGN: \cite{antonucci83} found an anticorrelation of the Balmer decrement with continuum flux in NGC\,4151.
We observed that the Balmer decrement H$\alpha$/H$\beta${} increases with the decreasing H$\beta${} line flux in variable Seyfert galaxies, such as Mrk\,110 (\citealt{bischoff99}), NGC\,7603 (\citealt{kollatschny00}), and HE\,1136-2304 (\citealt{kollatschny18}). This general trend is also present in {IRAS\,23226-3843}{}. However, the associated errors are larger in {IRAS\,23226-3843}{} than in those galaxies.
There is evidence that the radial structure \citep[e.g.,][]{koratkar91,kaspi05} and the vertical structure \citep{kollatschny14} of the line emitting BLR clouds vary as a function of continuum luminosity. This can lead to variations in the Balmer decrement.
A compilation of Balmer decrement data is given \cite{jaffarian20}. A slightly higher H$\alpha${}/H$\beta${} value in AGN with double-peaked Balmer lines can also be a consequence of higher reddening in these AGN which are assumed to be seen at higher inclination (\citealt{gaskell17}).
The asymmetry of H$\alpha${} during the early states in 1997 and 1999 might have been caused by absorption of the red component. Partial dust obscuration of the BLR by outflowing dust clumps, a model introduced by \citet{gaskell18}, can produce asymmetries and velocity-dependent lags. Other models to explain variable profile humps are conical outflows \citep{zheng91}, an asymmetric distribution of the BLR clouds \citep{wanders96}, and/or asymmetric disks \citep{eracleous94}.
There are clear signs for a blue absorption component in H$\beta${} and H$\alpha${} at $-10\,400\pm1000$,km\,s$^{-1}${} (see Figures~\ref{OchmIRAS23226_19990621_Grupe_20220325.pdf} and
\ref{OchmIRAS23226_veloplots_20220325.pdf}). Furthermore, based on the X-ray spectra, there is additional evidence for an outflow component. This absorption feature is present in the Fe band (see Fig.~\ref{fig:xray_fe_ratio}) at a slightly higher energy (7.6~keV)
than the possible feature seen in the 2017 data ($\sim7.3$~keV). It corresponds to an outflow with velocity of $\sim0.12c$ ($-36\,000$\,km\,s$^{-1}${}).
The broad double-peaked Balmer line profiles and their strong variations in {IRAS\,23226-3843}{} (see Fig.~\ref{OchmIRAS23226_veloplots_20220325.pdf}) are
similar to those of other broad double-peaked galaxies such as NGC\,1097 \citep{storchi97}, and NGC\,7213 \citep{schimoia17}, or broad-line radio galaxies such as 3C390.3 \citep{shapovalova10} and Arp102B \citep{shapovalova13}. \cite{strateva03} investigated the properties of double-peaked Balmer lines in a large sample of AGN based on SDSS galaxies. However, in comparison to all these double-peaked AGN, {IRAS\,23226-3843}{} shows significantly stronger \ion{Fe}{ii} line blends, in contrast to what is expected from Eigenvector 1 studies of AGN \citep{boroson92, sulentic00, shen14}.
\subsection{\bf Variability pattern in {IRAS\,23226-3843}{} }
The changing-look character of {IRAS\,23226-3843}{} is deduced from strong long-term (many years) and short-term (a few weeks) continuum variations in the optical and X-ray. In addition, {IRAS\,23226-3843}{} showed strong line profile variations including a transition from a Seyfert 1 type spectrum to a Seyfert 2 type in 2020.
{IRAS\,23226-3843}{} underwent a strong outburst lasting from 2019 September until 2019 December. The optical
continuum flux -- based on SALT spectra -- increased by a factor of 1.6 within two months. This corresponds to a factor of 3 when correcting the continuum flux for the host galaxy contribution
(1.1 $\times$ 10$^{-15}$\,erg\,s$^{-1}$\,cm$^{-2}$\,\AA$^{-1}$ in the V band). {IRAS\,23226-3843}{} varied
in the X-rays (\textit{Swift}{}) by a factor of nearly 5 during these three months. The optical line intensities of H$\alpha$, H$\beta$, and the \ion{Fe}{ii}(42,48,49) blend varied by factors of 2.3, 2.4, and 4. However, the shape of the double-peaked line profiles of H$\alpha${} and H$\beta${} remained more or less unchanged despite the strong intensity variations during these three months.
We took deep broadband XMM-\textit{Newton}/NuSTAR\ spectra during the optical maximum state (2019 November). These spectra can be compared with those taken during the minimum state in 2017 (see Fig.~\ref{Ochmlightcurves_IRAS23226_20220329.pdf} for
the observing epochs). The 2019 X-ray spectra of {IRAS\,23226-3843}{} are qualitatively very similar to
the 2017 spectra, but brighter by a factor of 10. The X-ray spectra were unabsorbed in 2019 as well as
during the minimum state in 2017. The optical line profiles at these two epochs were also qualitatively identical despite intensity variations by a factor of more than two (see Fig.~\ref{OchmIRAS23226_veloplots_20220325.pdf}).
The optical line profiles, however, changed on timescales of several years. These line-profile changes are independent of the continuum flux intensities. In total, we can discern three separate time intervals, namely the years 1997 - 1999, 2017 - 2019, and 2020. The line profiles of H$\alpha${} showed asymmetric blue excess profiles during the years 1997 and 1999. However, the profiles were symmetric and double peaked during the years 2017 and 2019. Finally, {IRAS\,23226-3843}{} changed to a Seyfert 2 type in 2020. A similar behavior has been observed in
the changing-look AGN HE\,1136-2304 \citep{zetzl18}: The line profiles varied independently of the continuum variations on timescales of years. The strong outburst observed in {IRAS\,23226-3843}{} in autumn 2019 is embedded
in additional variations before and after that epoch.
The cause for the violent as well as long-term variations in {IRAS\,23226-3843}{} can be narrowed down based on the observed long-term and short-term light curves in the X-ray and in the optical, the X-ray spectral variations, as well as optical line intensity and line profile variations.
Microlensing or a tidal disruption event (TDE) can be excluded as an explanation for the long-term variability pattern in {IRAS\,23226-3843}. First, lensing would be monochromatic, while the multiband light curves differ in detail in different bands, the flaring repeats, and lensing would not change the Balmer-line profiles. Second, TDEs in gas-rich environments or illuminated streams of stellar debris sometimes do in fact show a temporary emission-line signal in form of broad Balmer lines that can be double-peaked \citep[e.g.,][]{komossa08}. However, TDEs show single sharp outbursts in the X-rays \citep[e.g.,][]{lin11} or the optical--UV \citep[e.g.,][]{vanvelzen21} that then fade away on a timescale of months to years, very different from the long-term light curve of {IRAS\,23226-3843}.
There were no major changes in the X-ray hardness ratio in {IRAS\,23226-3843}{} despite strong flux variations over 14 years. This behavior is similar to that of HE\,1136-2304 \citep[][]{zetzl18, kollatschny18}. However, other changing-look AGN show strong X-ray gradient variations, for example, 1ES\,1927+654 \citep{ricci20,laha22}, Mrk\,1018 \citep{husemann16}, or NGC\,3516 \citep{laha22}. Strong changes in the gradient are often attributed to enhanced X-ray absorption by warm-absorber gas. The X-ray spectrum of IRAS\,23226-2304 was unabsorbed during the minimum state in 2017 as well as during the maximum state in 2019. Therefore, it is unlikely that the changing-look scenario in {IRAS\,23226-3843}{} is caused by obscuring clouds.
The optical luminosity shows the same variability trends as the X-ray luminosity in {IRAS\,23226-3843}{}. This is inconsistent with magnetic flux inversion. If magnetic flux inversion would be the dominant effect one would not expect the observed correlated trends \citep{scepi21}. For example, the changing-look AGN 1ES1927+654 showed no correlation between the optical/UV and X-ray emission variations \citep{laha22} in contrast to what we observe for {IRAS\,23226-3843}{}.
Currently the assumption of a magnetic flux inversion cannot reproduce the behaviour of changing-look AGN in general.
The changing-look character in {IRAS\,23226-3843}{} is most probably caused by changes of the accretion rate. This explains the short-term variations on timescales of weeks to months. In addition, the accretion disk structure and/or the distribution and ionization state of the emission-line clouds might have changed. These changes might explain the line profile variations on timescales of years that are independent of the continuum flux variations. Line profile variations independent of intensity variations have been observed as well in the changing-look AGN HE\,1136-2304 \citep{zetzl18,kollatschny18}.
Timescales to explain the changing-look phenomenon have been discussed by, for example, \cite{stern18}.
{IRAS\,23226-3843}{} shows a low Eddington ratio as discussed in \cite{kollatschny18}. This is consistent with the general trend based on SDSS spectra that the Eddington ratio decreases when the maximum g-band variability increases \citep{rumbaugh18}. Furthermore, {IRAS\,23226-3843}{} shows very broad emission lines. This is consistent with the correlation between the strength of optical continuum variability amplitudes in
AGN and their emission line widths \citep{kollatschny06}.
\section{Summary}
We present results on long-term variations and on an outburst seen in the X-rays, UV, and optical of the changing-look AGN {IRAS\,23226-3843}{}. The outburst in 2019 has been observed with \textit{Swift}{}, XMM-Newton, NuSTAR, SALT, and the SAAO\,1.9\,m telescope. Our findings are summarized below.
\begin{enumerate}[(1)]
\item The optical, UV, and X-ray continuum light curves of {IRAS\,23226-3843}{} showed strong variations with a similar variability pattern from 1997 until 2021.
There was a strong X-ray, UV, as well as optical outburst in 2019. It varied in the X-ray continuum by a factor of 5 and in the optical continuum by a factor of 1.6 within two months. We got a factor of 3 when we corrected for the host galaxy contribution. The Balmer and \ion{Fe}{ii} emission line intensities showed comparable variability amplitudes for the outburst in 2019.
\item The H$\alpha${} emission line profiles of {IRAS\,23226-3843}{} changed from a blue-peaked profile in the years 1997 and 1999 to a double-peaked profile in the years 2017 and 2019. There were no major variations in the double-peaked profile during the strong outburst in 2019. In 2020, after the outburst, {IRAS\,23226-3843}{} changed from a Seyfert type 1 to a Seyfert type 2 object.
\item A strong outflow component at $v=\,-10\,400$\,km\,s$^{-1}${} is to be seen in the Balmer lines. There is additional evidence for an outflow component
based on the deep X-ray spectra. This absorption feature is present in the Fe band with an outflow velocity of $\sim 0.12c$ ($-36\,000$\,km\,s$^{-1}$).
\item We took a deep broadband XMM-\textit{Newton}/NuSTAR\ spectrum during its maximum state in 2019. This spectrum is qualitatively very similar to a spectrum taken in 2017, but by a factor of 10 more luminous. The soft X-ray band appears featureless. The soft excess is well modeled with a Comptonization model. A broadband fit with a power-law continuum, Comptonized soft excess, and Galactic absorption gives a good fit to the combined EPIC-pn and NuSTAR spectrum. There is no indication of a warm absorber in the X-ray spectra.
In addition, we see a complex and broadened Fe K emission line profile.
\item The changing-look character in {IRAS\,23226-3843}{} is most probably caused by changes in the accretion rate -- based on the short-term variations on timescales of weeks to months. In addition, there are line-profile variations on timescales of years that are independent of the continuum flux variations. These changes might be explained by
changes in the accretion disk structure and/or distribution of the emission line clouds.
\item The double-peaked Balmer lines are extremely broad. In addition, {IRAS\,23226-3843}{} exhibits strong \ion{Fe}{ii} blends. This is in contrast to what is expected from Eigenvector 1 studies.
\end{enumerate}
{IRAS\,23226-3843}{} is an interesting changing-look AGN due to its strong variations in the optical and X-ray continuum, as well as its clear transitions from a single-peaked to a double-peaked Seyfert 1 type object, and finally, to an object of Seyfert type 2. Thus, {IRAS\,23226-3843}{} deserves further studies in the future.
\begin{acknowledgements}
This paper is based on observations obtained with XMM-Newton, an ESA science mission with instruments and contributions directly funded by ESA Member States and NASA, as well as observations made with the Southern African Large Telescope (SALT) under programs 2017-1-SCI-030, 2019-1-DDT-003, and 2019-2-DDT-001.\\
This work also made use of data from the NuSTAR mission, a project led by the California Institute of Technology, managed by the Jet Propulsion Laboratory, and funded by the National Aeronautics and Space Administration.
The authors made use of the NuSTAR Data Analysis Software (NuSTARDAS) jointly developed by the ASI Science Data Center (ASDC, Italy) and the California Institute of Technology (USA).\\
This research has made use of the XRT Data Analysis Software (XRTDAS) developed under the responsibility of the ASI Science Data Center (ASDC), Italy, and of the NASA/IPAC Extragalactic Database (NED) which is operated by the Jet Propulsion Laboratory, Caltech, under contract with the National Aeronautics and Space Administration. \\
This paper includes a Legacy Survey image. The DESI Legacy Imaging Surveys consist of three individual and complementary projects: the Dark Energy Camera Legacy Survey (DECaLS), the Beijing-Arizona Sky Survey (BASS), and the Mayall z-band Legacy Survey (MzLS). DECaLS, BASS and MzLS together include data obtained, respectively, at the Blanco telescope, Cerro Tololo Inter-American Observatory, NSF’s NOIRLab; the Bok telescope, Steward Observatory, University of Arizona; and the Mayall telescope, Kitt Peak National Observatory, NOIRLab. NOIRLab is operated by the Association of Universities for Research in Astronomy (AURA) under a cooperative agreement with the National Science Foundation. Pipeline processing and analyses of the data were supported by NOIRLab and the Lawrence Berkeley National Laboratory (LBNL). Legacy Surveys also uses data products from the Near-Earth Object Wide-field Infrared Survey Explorer (NEOWISE), a project of the Jet Propulsion Laboratory/California Institute of Technology, funded by the National Aeronautics and Space Administration. Legacy Surveys was supported by: the Director, Office of Science, Office of High Energy Physics of the U.S. Department of Energy; the National Energy Research Scientific Computing Center, a DOE Office of Science User Facility; the U.S. National Science Foundation, Division of Astronomical Sciences; the National Astronomical Observatories of China, the Chinese Academy of Sciences and the Chinese National Natural Science Foundation. LBNL is managed by the Regents of the University of California under contract to the U.S. Department of Energy. The complete acknowledgments can be found at \url{https://www.legacysurvey.org/acknowledgment/}. \\
The authors thank the \textit{Swift}{} team for performing the ToO observations, and Eileen Herwig for the reduction of the SALT spectra from 2017. ERC acknowledges support from the South African National Research Foundation. This work has been supported by the DFG grant Ko 857/35-1.
\end{acknowledgements}
\bibliographystyle{aa}
|
\section{Introduction}
\label{sec:intro}
\input{sections/intro}
\section{Related Work}
\label{sec:related}
\input{sections/related}
\section{Background}
\label{sec:background}
\input{sections/background}
\section{Method}
\label{sec:method}
\input{sections/method}
\section{Experiments}
\label{sec:experiments}
\input{sections/experiments}
\section{Discussion \& Conclusions}
\label{sec:conclusion}
\input{sections/conclusion}
\vfill
\section*{Supplementary Material}
\subsection{Representational Power}
\label{sec:rep_power}
We demonstrate the representational power of PhoMoH by randomly sampling from the latent embeddings for identity, appearance, head shape, and expression.
In \cref{fig:sample_buh} and \cref{fig:sample_rp} we show results for both datasets.
The ability to directly sample from the latent embeddings is enabled by the Gaussian prior imposed during training.
In contrast, for i3DMM one has to run PCA on the latent spaces to allow sampling.
Thus, sampling is performed without knowledge of the true underlying data distribution.
In contrast to our results, outputs by i3DMM are less detailed and geometry and color are not always consistent.
Additionally, i3DMM is learned in normalized space, thus all samples have ambiguous scale.
In contrast, PhoMoH is defined in metric space.
Geometric identity and appearance are decoupled in our model.
This allows us to sample multiple plausible appearances for a given geometry, as shown in \cref{fig:vary_appearance}.
Finally, we interpolate between two sampled model configurations in \cref{fig:interpolate}.
The intermediate results are detailed and plausible, demonstrating a compact, well-covered latent space.
\subsection{Mesh Reconstruction}
We compare our model with i3DMM and imGHUM, by fitting all models to the test set of our LHS dataset.
Note that for a fair comparison with i3DMM, as mentioned previously, we created a separate train split, LHS-40, which includes only 40 of the original scan identities.
We trained both our model and i3DMM on this smaller train subset.
The test split contains 5 unseen identities with 40 expressions each -- a total of 200 scans. We report Chamfer distance and Normal Consistency for geometry reconstruction, and a metric similar to Chamfer \cite{yenamandra2021i3dmm} for assessing the color reconstruction quality.
For a fair comparison to i3DMM, we used a similar approach to fitting and the same 3D point observation pairs (\ie 3D point with an associated ground truth signed distance value and color) as them. Results are shown in \cref{tbl:Fitting_buhv1}. Both our geometric and color reconstructions are numerically better than i3DMM. We include imGHUM in this comparison demonstrating that most of the detail in the geometry reconstruction is recovered by the geometric identity layer of our method. imGHUM is a good proxy, but it is restricted to a coarse representation of a scan.
\begin{table}
\centering
\small
\resizebox{0.99\linewidth}{!}{%
\begin{tabular}[t]{l||c|c|c}
& Chamfer $\times 10^{-3}$ $\downarrow$ & Normal Consistency $\uparrow$ & Color $\downarrow$ \\
\hline
imGHUM & $0.160$ & $0.93$ & $-$\\
i3DMM & $0.021$ & $0.96$ & $0.035$\\
\hline
Ours& $0.008$ & $0.96$ & $0.031$\\
\end{tabular}
}
\caption{Fitting results on the LHS-40 test set.}
\label{tbl:Fitting_buhv1}
\end{table}
\begin{figure}
\centering
\begin{overpic}[width=1.0\linewidth]{figures/ablate.png} %
\put(7,0){\footnotesize Full Model}
\put(34,0){\footnotesize $-$feature modulation}
\put(79,0){\footnotesize $-\mathcal{L}_p$}
\end{overpic}
\caption{We ablate the introduction of feature modulation and our novel patch loss $\mathcal{L}_p$. In the absence of feature modulation, the geometric features $\vec{g}$ mostly encode relative position and do not model local semantics. Our patch loss $\mathcal{L}_p$ significantly increases the visual fidelity of results. We show the first 3 PCA components of $\vec{g}$ as RGB colors.}
\label{fig:ablate}
\end{figure}
\subsection{Ablations}
We ablate two important design choices: feature modulation and the introduction of the patch loss $\mathcal{L}_p$.
Our proposed patch loss $\mathcal{L}_p$ significantly improves the visual fidelity of results. See \cref{fig:ablate} for a side-by-side comparison.
The differences are especially apparent for important facial features such as eyes or mouth, but also note that facial hair and the cap are better represented in our full model.
The introduction of feature modulation also slightly improves visual fidelity.
However, feature modulation mostly helps to learn meaningful geometric features $\vec{g}$ and to increase the overall training stability.
As apparent from \cref{fig:ablate}, when trained without feature modulation, the geometric features mostly encode relative position and do not capture local semantics.
We hypothesize, when $\vec{\iota}_a$ is an input to the appearance branch $a$, it can choose to ignore most information in $\vec{g}$.
This is not the case for feature modulation, where $\vec{g}$ is the only direct input to $a$.
By feature modulation, we establish a stronger connection between the two branches of our network.
We also noted training instabilities when using classic conditioning: the appearance branch can choose to entirely ignore $\vec{g}$ and instead correlate appearance with pose and expression, despite the randomization strategy explained in \cref{sec:imp_details}.
If so, geometry identity and appearance are no longer decoupled, hence sampling is no longer possible.
\begin{figure}
\centering
\begin{overpic}[width=\linewidth]{figures/app_transfer.png} %
\put(11,0){\footnotesize Source}
\put(40,0){\footnotesize Target}
\put(63,0){\footnotesize Ours}
\put(83.5,0){\footnotesize i3DMM}
\end{overpic}
\caption{The proposed geometric feature enables semantically coherent appearance transfer between samples useful for \eg creative editing or dataset augmentation and diversification. From left to right: original appearance together with its geometric identity, target appearance, the result of appearance transfer, result of i3DMM~\cite{yenamandra2021i3dmm}. Please note semantically coherent results of our method for \eg hair coloring, despite different hair styles between source and target. i3DMM fails to preserve the hair line, as it is especially visible in the 2nd row (ear region).}
\label{fig:app_transfer}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{figures/edit_hair.png}
\caption{By freezing the appearance code for some portions of the reconstructed mesh, and re-querying other parts, we can perform semantically coherent appearance edits. Here we sample new colors for points corresponding to hair in the source mesh (left). Note how the model correctly returns plausible hair colors.}
\label{fig:edit_hair}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{figures/geo_feature.png}
\caption{Our method learns a geometric feature $\vec{g}$ (without explicit supervision) to serve as semantically informed input to the appearance network. The feature is a useful tool beyond its original use case as it defines dense correspondences between all sampled heads. Here we show the first 3 PCA components as RGB colors.}
\label{fig:geo_feature}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{figures/one_shot.png}
\caption{We transfer semantic labels from one head (left) to others (right) by training a classifier based on the geometric feature $\vec{g}$. The classifier correctly predicts semantic labels for diverse heads despite being trained on a single head instance.}
\label{fig:one_shot}
\end{figure}
\subsection{Applications}
Finally, we are demonstrating possible use-cases of PhoMoH.
The per-point geometric feature $\vec{g}$ learnt by our model can be understood as semantic correspondences between different head instances, see \cref{fig:geo_feature}.
This is a useful property, as we now show.
First, $\vec{g}$ enables semantically correct appearance transfer between different head instances, useful for \eg dataset augmentation or creative editing.
We show results in \cref{fig:app_transfer} and compare to appearance transfer results obtained using i3DMM.
While our model places colors semantically correct, i3DMM struggles to maintain the hair line defined by the geometry.
This is especially apparent in row two, where i3DMM textures the hair region with colors from the ear region.
Similar problems can be observed in \cref{fig:sample_rp}.
i3DMM's inability to follow the semantics defined by the geometry is rooted in model design, where learning of color is decoupled from geometry on a canonical reference shape.
Our model also allows local appearance edits, as shown in \cref{fig:edit_hair}.
In this experiment, we re-queried points corresponding to hair with newly sampled appearance codes.
Our model produces plausible hair colors, following the shape defined by the original hairstyle.
The geometric feature is also useful for classification and segmentation tasks.
In our final experiment, we labeled one single head instance with six semantic labels, corresponding to cap, eyes, ears, nose, mouth, and skin.
We then trained a classifier mapping $\vec{g}$ to one of the six labels.
Despite being trained only on one single head, our classifier gracefully generalizes to other heads produced by PhoMoH, see \cref{fig:one_shot}.
This means, once trained, PhoMoH can drastically simplify and speed up the labeling process for a given dataset.
\subsection{Losses}
For training we require pairs of 3D head scans together with imGHUM parameters approximating the scan; please refer to \cref{sec:data} for more information on how to obtain those.
We start by defining losses for the geometry, and continue with losses for the appearance, as well as regularizers.
\begin{figure*}
\centering
\includegraphics[width=\linewidth]{figures/sample_buh.png}
\caption{Random heads sampled from our model trained on the LHS dataset showing varying head shape, expression, geometric identity, and appearance. For each sample, we show 3D geometry and appearance side-by-side. }
\label{fig:sample_buh}
\end{figure*}
\paragraph{Geometry.}
Given a set of points $\set{O}$ on a ground truth mesh $\mathcal{M}$ describing the surface $\mathcal{S}$ together with its ground truth surface normals $\bar{\normal}$, and corresponding imGHUM parameters $(\vec{\theta},\vec{\beta},\vec{\epsilon})$ approximating $\mathcal{M}$, we minimize
\begin{equation}
\mathcal{L}_{g} = \frac{1}{|\set{O}|} \sum_{i \in \set{O}} \lambda_{g} |d_{\vec{x}_i}| + \lambda_{n} \| \vec{n}_{\vec{x}_i} - \bar{\normal}_i \|,
\end{equation}
where $\vec{n}_{\vec{x}} = \nabla_{\vec{x}} d_{\vec{x}}$ is the estimated surface normal defined by the gradient of the estimated distance \wrt $\vec{x}$ and $\lambda_{*}$ weight the different loss components.
Further, we supervise the SDF with the Eikonal loss \cite{gropp2020igr} for points $\set{F}$ around the surface and additionally supervise their sign
\begin{equation}
\mathcal{L}_{e} = \frac{1}{|\set{F}|} \sum_{i \in \set{F}} (\| \vec{n}_{\vec{x}_i} \| - 1)^2
\end{equation}
\begin{equation}
\mathcal{L}_{l} = \frac{1}{|\set{F}|} \sum_{i \in \set{F}} \text{BCE}\big(l_i, \psi(k d_{\vec{x}_i})\big),
\end{equation}
based on binary sign labels $l$, sigmoid function $\psi$, binary cross-entropy $\text{BCE}$, and the learnable parameter $k$ controlling the sharpness of the decision boundary.
\paragraph{Appearance.}
For all points, we supervise the color component of $f$ to match the ground truth color $\bar{\col}$ of the nearest neighbor on the surface of $\mathcal{M}$
\begin{equation}
\mathcal{L}_{c} = \frac{1}{|\set{O} \cup \set{F}|} \sum_{i \in \set{O} \cup \set{F}} | \vec{c}_{\vec{x}_i} - \bar{\col}_i |.
\end{equation}
While $\mathcal{L}_{c}$ is theoretically sufficient to supervise the color field, in practice only training with $\mathcal{L}_{c}$ leads to overly smooth results.
To this end, we propose a novel perceptual patch loss with importance sampling to improve the visual fidelity of our results.
We sample a patch by first sampling a point on the mesh.
Next, we create a virtual camera looking at the sampled point along its surface normal.
Using this camera, we render a small image and additionally compute the 3D surface locations corresponding to each pixel -- please see \cref{sec:data} for a detailed explanation.
For all 3D points in the resulting patch $\set{P}$ and the corresponding rendered image $\bar\mat{I}^\set{P}$, we enforce the previously introduced color and distance losses, and additionally apply a VGG-loss \cite{chen2017photographic} for an image $\mat{I}^\set{P}$ formed using the colors returned by $f$ for all $\vec{x} \in \set{P}$
\begin{equation}
\small
\mathcal{L}_{p} = \frac{1}{|\set{P}|} \sum_{i \in \set{P}} \left( \lambda_{pc} | \vec{c}_{\vec{x}_i} - \bar\mat{I}^\set{P}_i | + \lambda_{pd} |d_{\vec{x}_i}| \right) + \lambda_{pp} \text{VGG}(\mat{I}^\set{P}, \bar\mat{I}^\set{P}),
\end{equation}
where $\bar\mat{I}^\set{P}_i$ is the $i$-th pixel color in the ground truth rendered patch.
In contrast to rendering losses proposed in prior work \cite{alldieck2022phorhum}, our perceptual patch loss does not require memory or computationally expensive differentiable rendering strategies, and all the expensive operations can be performed during data pre-processing.
\paragraph{Regularization.} To enforce a Gaussian prior on $\vec{\iota}_g$ and $\vec{\iota}_a$, we minimize the Kullback-Leibler (KL) divergence loss~\cite{kingma2013auto} for each mean and standard deviation pair $(\vec{\sigma}, \vec{\mu})$ corresponding to each instance of $\vec{\iota}_g$ and $\vec{\iota}_a$
\begin{equation}
\mathcal{L}_{\text{KL}} = -\frac{1}{2} \sum_{j=1}^{J}\left(1 + \log(\sigma_j^2) - \mu_j^2 - \sigma_j^2 \right),
\end{equation}
where $j$ references the $j$-th dimension.
We expect pairs of meshes $\mathcal{M}$ and imGHUM parameters $(\vec{\theta},\vec{\beta},\vec{\epsilon})$ approximating $\mathcal{M}$ during training -- see \cref{sec:data}.
As it is impossible to obtain perfect pairs due to the ill-posed nature of the problem, we allow our model to ``correct'' imGHUM parameters $(\vec{\theta},\vec{\beta},\vec{\epsilon})$ during training.
To this end, we treat $(\vec{\theta},\vec{\beta},\vec{\epsilon})$ as pre-initialized variables and update those together with the model weights and latent representations.
To constrain the solution space we apply L2 regularization.
We further guide the optimization using 3D facial keypoints $\set{K}$ computed to obtain the initial imGHUM parameter set (\cf \cref{sec:data})
\begin{equation}
\mathcal{L}_{k} = \frac{1}{|\set{K}|} \sum_{i \in \set{K}} |d_{\vec{k}_i}| + | \vec{s}_{\vec{k}_i} - \bar\vec{s}_i |,
\end{equation}
where $\vec{s}$ are reference points returned by imGHUM and $\bar\vec{s}$ are ground-truth reference points for the given keypoint set.
\subsection{Training Data Pre-Processing}
\label{sec:data}
Given a mesh $\mathcal{M}$ with a corresponding color texture, we fit GHUM parameters using a standard 3D facial landmark based optimization for global alignment together with a refinement step based on Iterative
Closest Points (ICP) \cite{besl1992method}. As some of the training scans belong to the same person, we further impose them to share the same shape parameters $\vec{\beta}$. We collect the sets of on surface points $\set{O}$ and near surface points $\set{F}$ using the same procedure as in \cite{alldieck2022phorhum}. To gather a set of patches $\set{P}$, we take the following approach. First, we uniformly sample a set of points on the surface of $\mathcal{M}$. Next, we compute their distance to the closest facial landmark $\set{K}$. We transform these distances into probabilities using a radial basis function (RBF) kernel followed by normalization and draw another set of candidate samples from this distribution. By increasing the bandwidth parameter of the RBF kernel one can skew the distribution to sample more points that belong to the face region. For each point in this candidate set, we create a virtual camera looking at the sampled point along its surface normal at a fixed distance of $4$~cm.
We render the patch at a small resolution of $64\times64$~px using orthographic projection and store the color and 3D surface location of each pixel in the patch. Finally, we select from the candidate set the top patches that have the most coverage (\ie as few pixels associated with empty space).
\subsection{Implementation Details}
\label{sec:imp_details}
The two network branches are modeled with eight 512-dimensional fully-connected layers with Swish activation \cite{ramachandran2017searching} and a skip connection to the middle layer.
We apply positional encoding \cite{tancik2020fourfeat} on $\vec{s}$ and $\vec{g}$ before passing them to the networks.
The geometric identity $\vec{\iota}_g$ and appearance $\vec{\iota}_a$ are learned using 128-dimensional codes.
$\vec{g}$ is 8-dimensional and the combined inputs to our feature modulation are reduced back to 128 dimensions.
To prevent overfitting, especially when trained on small datasets, we replace $(\vec{\theta}, \vec{\epsilon})$ pairs passed to the feature modulation with Gaussian noise with $50\%$ probability.
Further, we start training with the $\mathcal{L}_{\text{KL}}$ loss disabled and then linearly increase its influence.
Finally, we do not follow the auto-decoder learning scheme proposed by DeepSDF where every scan is represented in each training step. Instead, we employ the standard way of training neural networks with randomly sampled batches for each update step.
|
\section{Introduction}
\label{sec:intro}
3D object detection from point clouds has received much attention in recent years~\cite{zhou2018voxelnet,shi2019pointrcnn,shi2020pv,yin2021center,fan2022fully} as its wide applications in autonomous driving, robots navigation, etc.
State-of-the-art 3D detectors~\cite{shi2021pv,shi2022pillarnet,fan2022fully,zhou2022centerformer} still adopt dense predictions with post-processing (\emph{e.g.}, NMS~\cite{canny1986computational}) to obtain final sparse detections. This \emph{indirect} pipeline usually involves many hand-crafted components (e.g., anchors, center masks) based on human experience, which involves much effort for tuning, and prevents dense detectors from being optimized end-to-end to achieve optimal performance.
Recently, DETR-based 2D detectors~\cite{carion2020end,zhu2020deformable,sun2021sparse,zhang2022dino} show that transformers with \emph{direct} sparse predictions can greatly simplify the detection pipeline, and lead to better performance.
However, although many efforts~\cite{Misra_2021_ICCV,nguyen2022boxer,bai2022transfusion} have been made towards \emph{direct} sparse predictions for 3D object detection, because of the different characteristics of images and point clouds (\emph{i.e.}, dense and ordered images \emph{v.s.} sparse and irregular points clouds),
performance of sparse 3D object detectors still largely lags behind state-of-the-art dense detectors.
\begin{figure}[t]
\centering
\includegraphics[width=0.95\linewidth]{motivation}
\caption{Comparison of our baseline Voxel-DETR and ConQueR. GTs ({\color{green}{green}}) and predictions ({\color{blue}{blue}}) of an example scene in the WOD is visualized. Sparse predictions of Voxel-DETR still contain many highly overlapped false positives (in the red dashed circle), while ConQueR can generate much sparser predictions.}
\label{fig:motivation}
\vspace{-3mm}
\end{figure}
To achieve direct sparse predictions, DETRs usually adopt a set of object queries~\cite{carion2020end,zhu2020deformable,sun2021sparse,zhang2022dino,nguyen2022boxer,bai2022transfusion}, and resort to the one-to-one Hungarian Matching~\cite{kuhn1955hungarian} to assign ground-truths (GTs) to object queries.
However, to guarantee a high recall rate, those detectors need to impose much more queries than the actual number of objects in a scene. For example, recent works~\cite{nguyen2022boxer,bai2022transfusion} select top-$300$ scored query predictions to cover only $\sim$40 objects in each scene of Waymo Open Dataset (WOD)~\cite{sun2020scalability}, while 2D DETR detectors~\cite{carion2020end,zhu2020deformable,sun2021sparse,zhang2022dino} use 10$\times$ more predictions than the average GT number of MS COCO~\cite{lin2014microsoft}. As shown in Fig.~\ref{fig:motivation}(a), we visualize an example scene by a baseline DETR-based 3D detector, named Voxel-DETR, which shows its top-$300$ scored predictions.
Objects are generally small and densely populated in autonomous driving scenes, while 3D DETRs adopt the same fixed top-$N$ scored predictions as 2D DETRs, and lack a mechanism to handle such small and dense objects. Consequently, they tend to generate densely overlapped false positives (in the red-dashed circle), harming both the accuracy and \emph{sparsity}~\cite{sun2021sparse,roh2021sparse} of final predictions.
We argue the key reason is that the Hungarian Matching in existing 3D DETRs only assigns each GT to its best matched query, while all other unmatched queries near this GT are not effectively suppressed. For each GT, the one-to-one matching loss solely forces all unmatched queries to predict the same ``no-object'' label, and the best matched query are supervised \emph{without} considering its relative ranking to its surrounding unmatched queries. This design causes the detectors to be insufficiently supervised in discriminating similar query predictions for each GT, leading to duplicated false positives for scenes with densely populated objects.
To overcome the limitations of current supervision, we introduce a simple yet novel Query Contrast strategy to explicitly suppress predictions of all unmatched queries for each GT, and simultaneously enhance the best matched query to generate more accurate predictions in a contrastive manner.
The Query Contrast strategy is integrated into our baseline Voxel-DETR, which consists of a sparse 3D convolution backbone to extract features from voxel grids, and a transformer encoder-decoder architecture with a bipartite matching loss to directly generate sparse predictions.
Our Query Contrast mechanism involves the construction of positive and negative GT-query pairs, and the contrastive learning on all GT-query pairs to supervise both matched and unmatched queries with knowledge of the states of their surrounding queries. Such GT-query pairs are directly created by reusing the Hungarian Matching results: each GT and its best matched query form the positive pair, and all other unmatched queries of the same GT then form negative pairs.
To quantitively measure the similarities of the GT-query pairs, we formulate the object queries to be the same as GT boxes (\emph{i.e.}, using only box categories, locations, sizes and orientations), such that GTs and object queries can be processed by the same transformer decoder, and embedded into a unified feature space to properly calculate their similarities.
Given the GT-query similarities, we adopt the contrastive learning loss~\cite{chen2020simple,he2020momentum,zhu2020eqco} to effectively enhance the positive (matched) query's prediction for each GT, and suppress those of all its negative queries at the same time. Moreover, to further improve the contrastive supervision, we construct multiple positive GT-query pairs for each GT by adding small random noises to the original GTs, which greatly boost the training efficiency and effectiveness.
The resulting sparse 3D detector, named \textbf{Que}ry \textbf{Con}trast Voxel-DET\textbf{R} (\textbf{ConQueR}), significantly improves the detection performance and sparsity of final predictions, as shown in Fig.~\ref{fig:motivation}(b). Moreover, ConQueR abandons the fixed top-$N$ prediction scheme and enables to output a vary number of predictions for different scenes. ConQueR reduces up to {\bf $\sim$60\%} false positives and sets new records on the challenging Waymo Open Dataset (WOD)~\cite{sun2020scalability}. Contributions are summarized as bellow:
\begin{enumerate}
\item We introduce a novel Query Contrast strategy into DETR-based 3D detectors to effectively eliminate densely overlapped false positives and achieve more accurate predictions.
\item We propose to construct multi-positive contrastive training, which greatly improve the effectiveness and efficiency of our Query Contrast mechanism.
\item Our proposed sparse 3D detector ConQueR closes the gap between sparse and dense 3D detectors, and sets new records on the challenging WOD benchmark.
\end{enumerate}
\section{Related Works}
\paragraph{End-to-End 2D Object Detection.} End-to-end object detection aims to generate final sparse predictions without non-differentiable components like NMS. RelationNet~\cite{hu2018relation} proposes an object relation module and DETR~\cite{carion2020end} greatly simplifies the detection pipeline by removing many hand-crafted components like anchors, NMS, etc. DETR introduce a set of object queries and resorts to the Hungarian Matching to associate each GT with the query predictions of minimal matching cost, and selects top-$N$ scored predictions for inference. \cite{wang2021end,sun2021sparse} also reveal that one-to-one matching is the key to achieve sparse predictions. Following works~\cite{zhu2020deformable,meng2021conditional,wang2022anchor,li2022dn,li2022dn,jia2022detrs} improves DETR in many aspects including query design, convergence speed, and performance, surpassing CNN-based dense detectors~\cite{zhang2020bridging,zhu2020autoassign,ge2021ota} by a large margin. However, they still need to select a fixed number of predictions as final results, no matter how many objects are there in an image.
Recently, DINO-DETR~\cite{zhang2022dino} introduces a ``contrastive'' denoising training strategy. It creates positive and negative GTs conceptually, and supervise these GTs with different targets separately, which has no relation with contrastive learning.
\begin{figure*}[t]
\centering
\includegraphics[width=0.9\linewidth]{contrastive_voxel_detr}
\caption{Overall pipeline of the proposed ConQueR. It consists of a 3D Sparse ResNet-FPN backbone to extract dense BEV features, and a transformer encoder-decoder architecture with one-to-one matching. Top-$k$ scored object proposals from a class-agnostic FFN form the object queries to input to the transformer decoder. During training, GTs (noised) are concatenated with object queries to input to the transformer decoder to obtain unified embeddings, which are then used for Query Contrast at each decoder layer. During inference, Top-scored predictions from the last decoder layer are kept as final sparse predictions. ``VFE'' denotes the voxel feature extractor in ~\cite{yan2018second,zhu2019class,yin2021center}.
}
\label{fig:contrastive_voxel_detr}
\end{figure*}
\vspace{-2mm}
\paragraph{3D Object Detection from Point Clouds.} State-of-the-art 3D detectors usually adopts voxel-based~\cite{shi2020pv,yin2021center,shi2021pv,shi2022pillarnet}, range-view~\cite{sun2021rsn,tian2022fully} or point-based~\cite{fan2022fully,yang20203dssd} paradigms to convert raw point clouds into dense feature representations, followed by detection heads to generate dense predictions and resort to NMS to filter out low-quality predictions. Many attempts have also been made to incorporate transformer architectures~\cite{sheng2021improving,mao2021voxel,zhou2022centerformer,Sun2022SWFormerSW} into 3D object detection, but they still rely on post-processing. Others~\cite{nguyen2022boxer,bai2022transfusion} make a step further to use the one-to-one matching loss to achieve direct sparse 3D predictions. \cite{nguyen2022boxer} proposes Box-Attention, a variant of deformable attention to better capture local informations and applies it to 3D object detection. \cite{bai2022transfusion} introduce image features into a decoder-only architecture to enhance query features. However, their performance still largely lags behind state-of-the-art dense 3D detectors.
\vspace{-2mm}
\paragraph{Contrastive Learning for Object Detection.} Contrastive learning aims to learn an embedding space such that similar data pairs stay close while dissimilar ones are far apart. ~\cite{hadsell2006dimensionality} proposes to learn representations by contrasting positive pairs against negative ones. The popular InfoNCE loss~\cite{oord2018representation} uses categorical cross-entropy loss to learn such an embedding space. Following works~\cite{chen2020simple,he2020momentum,caron2020unsupervised} demonstrate the superiority of contrastive learning on providing pre-trained weights for downstream tasks (\emph{e.g.}, 2D detection). Few works explore the use of contrastive loss in object detection. ~\cite{lang2021contrastive} introduces semantically structured embeddings from knowledge graphs to alleviate misclassifications. ~\cite{yao2021g} conducts contrastive distillation between different feature regions to better capture teacher's information. As far as we know, we are the first to introduce the contrastive learning process into DETR-based detectors.
\section{Query Contrast Voxel-DETR (ConQueR)}
\label{sec:method}
State-of-the-art 3D detectors usually generate dense object predictions, which require many hand-designed components (\emph{e.g.}, anchors, box masks) based on prior knowledge, and resort to post-processing to filter out low-quality and duplicated boxes. This \emph{indirect} pipeline hinders the detectors from being optimized end-to-end and achieving optimal performance. 3D DETRs aim at streamlining these hand-crafted modules, and \emph{directly} generating sparse predictions via the transformer architecture and one-to-one matching loss, but they still cannot compete with state-of-the-art dense 3D detectors and face the problem of highly overlapped false positives, as shown in Fig.~\ref{fig:motivation}(a). To solve these challenges, we first introduce our competitive DETR-based 3D framework, named Voxel-DETR in Sec.~\ref{sec:voxel_detr}, and present the Query Contrast strategy to tackle with the duplicated false positives and further improve the detection performance in Sec.~\ref{sec:cns}.
\subsection{Voxel-DETR}
\label{sec:voxel_detr}
As illustrated in Fig.~\ref{fig:contrastive_voxel_detr}, Voxel-DETR consists of a 3D backbone, an encoder-decoder transformer architecture, and a set-matching loss to achieve direct sparse predictions.
\noindent {\bf Backbone.} Point cloud is rasterized into sparse voxel grids and fed into a 3D Sparse ResNet~\cite{he2016deep} backbone network to extract sparse 3D features. These features are transformed into dense Bird Eye View (BEV) feature maps, followed by an FPN~\cite{lin2017feature} to extract multi-scale features.
\noindent {\bf Transformer.}
The encoder-decoder transformer is similar to the two-stage Deformable-DETR~\cite{zhu2020deformable}. The $8\times$ downscaled BEV features from the FPN are input to the transformer encoder, which consists of 3 encoder layers.
Considering the characteristics of 3D detection from point clouds (\emph{i.e.}, all objects are relatively small and densely distributed),
we adopt BoxAttention~\cite{nguyen2022boxer}, which applies spatial in-box constraints to Deformable Attention~\cite{zhu2020deformable}, to perform local self-attention.
A class-agnostic feed-forward network (FFN) head is used to generate initial object proposals from encoder features. Top-$k$ scored box proposals are selected as object queries to input to the 3-layer transformer decoder. Decoder layers conduct inter-query self-attention and cross-attention between query and encoder features, followed by prediction heads to perform iterative box refinement~\cite{zhu2020deformable}. Predicted query boxes from the previous decoder layer's FFN head are transformed by a 3-layer MLP and added with the updated query features (initialized as zero) from the previous decoder layer.
\noindent {\bf Losses.} During training, all FFN prediction heads use the Hungarian Matching to assign GTs to object queries. The detection loss $\mathcal{L}_{det}$ consists of a focal loss~\cite{lin2017focal} for classification, a smooth L1 loss and a 3D GIoU loss for box regression:
\begin{equation}
\label{eq:det_loss}
\mathcal{L}_{\rm det} = \alpha \mathcal{L}_{\rm focal} + \beta \mathcal{L}_{\rm l1} + \gamma \mathcal{L}_{\rm GIoU},
\end{equation}
where $\alpha, \beta, \gamma$ are hyper-parameters to balance the loss terms.
During inference, top-$N$ scored predictions from the last decoder layer are kept as the final sparse detections.
\subsection{Query Contrast}
\label{sec:cns}
\begin{figure}[t]
\centering
\includegraphics[width=1.0\linewidth]{contrastive_illustration}
\caption{Illustration of Query Contrast. Given the GT ({\color{green}green}), Hungarian Matching gives its best matched ({\color{blue}blue}) and all other unmatched ({\color{gray}gray}) object queries. Query embeddings are projected by an extra MLP to align with GT embeddings. The contrastive loss is applied to all positive and negative GT-query pairs based on their feature similarities.
}
\label{fig:contrastive_loss}
\vspace{-3mm}
\end{figure}
Although Voxel-DETR already achieves satisfactory performance, its top-$N$ scored predictions still suffer from densely overlapped false positives (as shown in Fig.~\ref{fig:motivation}(a)). To tackle this problem,
we present a novel Query Contrast mechanism (depicted in Fig.~\ref{fig:contrastive_loss}) to explicitly enhance each GT's best matched query over unmatched ones. We first construct positive and negative GT-query pairs for each GT, which are then processed by each decoder layer to generate aligned GT and query embeddings. To promote the positive queries' similarity towards a GT against negative ones, the contrastive loss is applied at each decoder layer.
\vspace{-2mm}
\paragraph{Construction of positive/negative GT-query pairs.}
To determine queries to be enhanced or suppressed for each GT, we first construct positive and negative GT-query pairs by reusing the Hungarian Matching results (used for Eq.(\ref{eq:det_loss})), which is naturally compatible with our Voxel-DETR framework. Given a GT, the query with the minimal matching cost forms a positive pair with the GT, all other queries and this GT then form negative GT-query pairs. These GT-query pairs help to identify the object queries that need to be further enhanced or suppressed in our Voxel-DETR. Motivated by the SwAV~\cite{caron2020unsupervised} that incorporates multiple image crops to form multiple positive pairs to boost the training process, we further add small noises of different magnitudes on each GT to generate multiple noised GT copies. The multiple noised GT copies then form additional GT-query pairs with the same positive/negative query partitions as original GTs.
In practice, if a noised copy deviates too much from its original GT, the noised GT-query pairs would harm the contrastive training process. However, finding proper noise magnitudes is rather laboursome and cannot generalise well across scenarios. We thus add an auxiliary GT de-noising loss similar to that in DN-DETR~\cite{li2022dn} to obligate the detector to recover the original GT from its noised versions, which ensures that the noised GT copies would not diverge. Note that the ``noising-denoising'' step alone only has marginal effects to detection performance, while our multi-positive Query Contrast based on the noised GT copies leads to superior detection performance, as shown in our ablation studies.
\vspace{-4mm}
\paragraph{Contrast positive pairs against negative pairs.}
Before applying supervisions to the positive and negative GT-query pairs, we need to quantitatively measure the similarities of these pairs.
However, simple geometric metrics (\emph{e.g.}, IoU) cannot sufficiently model the similarities between GTs and queries (\emph{i.e.}, category, appearance, location, size, etc.). We thus propose to embed GTs and queries into a latent space for comprehensive similarity measurement. In our Voxel-DETR, object queries are formulated as proposal boxes (\emph{i.e.} object category, box location, size, and orientation). Therefore, the transformer decoder can naturally be used to encode both GTs and queries into feature embeddings at a chosen layer. We simply select the output layer of the FFN prediction head after each decoder layer (as shown in Fig.~\ref{fig:contrastive_voxel_detr}), followed by a shared MLP for similarity estimation.
However, we observe that the distributions of GT objects and query boxes can be quite different: GTs have no overlap with each other and generally distribute following the roadmap layouts, while queries might correspond to densely overlapped boxes and show up at random locations. As the transformer decoder utilizes self-attention to capture inter-box relations, the different distributions of GTs and query boxes would greatly affect estimation of their similarities. To mitigate the distribution gap, we adopt an extra MLP to project query features to align with GTs' latent space (the ``Projector'' in Fig.~\ref{fig:contrastive_loss}). With the aligned GT and query embeddings, we estimate all positive and negative GT-query pairs' similarities with cosine similarity metric, and adopt the InfoNCE loss~\cite{oord2018representation} to encourage the best matched query to generate more accurate predictions towards its assigned GT, and force all other unmatched queries to deviate away. Moreover, to obtain more stable GT representations for supervising queries, we adopt an exponential moving average (EMA) copy for each decoder layer to embed GTs, which is shown to be effective in our ablations.
Assume that for the $i$-th GT in a point cloud scene, we add $T$ different noises and denote the noised GT embeddings as $\{b^1_i, b^2_i, ... , b^T_i\}$, and denote $K$ query embeddings as $\{q_1, q_2, ... , q_K\}$. Suppose that the Hungarian Matching assigns the $i$-th GT to the $j$-th query, then our Query Contrast loss for the $i$-th GT $\mathcal{L}^{\rm QC}_i$ can be formulated as
\begin{equation}
\mathcal{L}^{\rm QC}_i = -\sum_{t=1}^{T} \log \left( \frac{\exp( \cos(b^t_i, g(q_j) ) / \tau )}{\sum_{k=1}^{K}\exp(\cos(b^t_i, g(q_k)) / \tau)} \right),
\label{eq:qc}
\end{equation}
where $\tau$ is the temperature coefficient, and $g(\cdot)$ denotes the extra MLP projector to align query features to GTs'.
As shown in Fig.~\ref{fig:contrastive_voxel_detr}, the Query Contrast loss is adopted at every decoder layer.
During inference, we abandon the widely adopted top-$N$ scored prediction strategy and use a score threshold (\emph{e.g.}, 0.1) to filter out low-quality query predictions. Query Contrast works quite well on suppressing similar query predictions in local neighborhoods, as shown in Fig.~\ref{fig:motivation}(b). ConQueR greatly boosts the detection accuracy, and reduces up to \textbf{$\sim$60\%} false positives.
\vspace{-1mm}
\paragraph{Discussion: Why does Query Contrast improve DETR-based 3D detectors?}
As discussed in Sec.~\ref{sec:intro}, current detection losses (\emph{i.e.}, focal loss for classification, smooth L1 and GIoU loss for regression) supervise each query without considering its surrounding queries, which lack supervision to train detectors to discriminate similar object queries especially in local regions. The proposed Query Contrast strategy tackles this issue by constructing a contrastive objective to supervise \emph{all} queries simultaneously. As suggested in Eq.(\ref{eq:qc}), for each GT object, the detector is instructed to identify the best matched query, and is forced to learn to differentiate it from all other unmatched counterparts, even if some of them are highly overlapping with the best matched query. As a result, all unmatched queries are trained to deviate from the GT, thus the duplicated false positives in our baseline Voxel-DETR can be effectively suppressed.
Another core design of our Query Contrast is to encode the GTs and queries into a unified learnable latent space. GT objects are encoded to provide better forms of supervision for both matched and unmatched queries. Previous works~\cite{hao2020labelenc,zhang2022lgd} in 2D object detection also show that encoding labels into feature embeddings to serve as extra supervision can perform better than the common hand-designed learning targets (\emph{i.e.}, classification logits and regression offsets), but they generally work in a knowledge distillation (KD) manner, which cannot be utilized to supervise negative queries. In contrast, our contrastive loss does not force matched queries to approach GTs directly, but encourages them to be ``closer'' to their corresponding GT embeddings than other close-by duplicated queries. Note that in our Query Contrast mechanism, GT embeddings are processed in an off-line manner and encoded into a unified space as queries', which serve as a type of supervision and force the detector to generate more similar query features as GTs'.
According to our experiments, the proposed Query Contrast strategy can not only suppress those duplicated false positives, but also contribute to better detection performance, which are consistent with the above discussions.
\section{Experiments}
ConQueR is mainly evaluated on the Waymo Open Dataset~\cite{sun2020scalability} (WOD) benchmark using the official detection metrics: mAP and mAPH (mAP weighted by heading) for Vehicle (Veh.), Pedestrian (Ped.), and Cyclist (Cyc.). The metrics are further splitted into two difficulty levels according to the point numbers in GT boxes: LEVEL\_1 ($>$5) and LEVEL\_2 ($\geq$1). We conduct ablation studies on the \textit{validation} set, and compare with state-of-the-art detectors on both \textit{validation} and \textit{test} set.
\begin{table*}[t]
\footnotesize
\centering
\begin{tabular}{l|>{\columncolor[gray]{0.95}}c|c|c|c|c|c|c}
\specialrule{1pt}{0pt}{1pt}
\toprule
\multicolumn{1}{c|}{\multirow{2}{*}{Methods}} & \multirow{2}{*}{\shortstack[1]{mAP/mAPH\\ L2}} & \multicolumn{2}{c|}{\emph{Vehicle} 3D AP/APH} & \multicolumn{2}{c|}{\emph{Pedestrian} 3D AP/APH} & \multicolumn{2}{c}{\emph{Cyclist} 3D AP/APH}\\
& L2 & L2 & L1 & L2 & L1 & L2 & L1\\
\midrule
\multicolumn{8}{l}{\textbf{Dense Detectors}}\\
CenterPoint$_{\rm ts}$~\cite{yin2021center} & -/67.4 & -/67.9 & -/- & -/65.6 & -/- & -/68.6-/- & -/- \\
PV-RCNN~\cite{shi2020pv} &66.8/63.3& 69.0/68.4 & 77.5/76.9 & 66.0/57.6 & 75.0/65.6 & 65.4/64.0 & 67.8/66.4 \\
AFDetV2~\cite{hu2022afdetv2} & 71.0/68.8 & 69.7/69.2 & 77.6/77.1 & 72.2/67.0 & 80.2/74.6 & 71.0/70.1 & 73.7/72.7\\
SST\_TS~\cite{fan2022embracing} & -/- & 68.0/67.6 & 76.2/75.8 & 72.8/65.9 & 81.4/74.1 & -/- & -/- \\
SWFormer~\cite{Sun2022SWFormerSW} & -/- & 69.2/68.8 & 77.8/77.3 & 72.5/64.9 & 80.9/72.7 & -/- & -/- \\
PillarNet-34~\cite{shi2022pillarnet} & 71.0/68.5 & \underline{70.9}/\textbf{70.5} & \underline{79.1}/\underline{78.6} & 72.3/66.2 & 80.6/74.0 & 69.7/68.7 & 72.3/71.2 \\
CenterFormer\cite{zhou2022centerformer} & 71.2/69.0 & 70.2/69.7 & 75.2/74.7 & 73.6/\underline{68.3} & 78.6/73.0 & 69.8/68.8 & 72.3/71.3 \\
PV-RCNN++~\cite{shi2021pv} & 71.7/69.5 & 70.6/\underline{70.2} & \textbf{79.3/78.8} & 73.2/68.0 & \underline{81.3}/\underline{76.3} & 71.2/70.2 & 73.7/72.7\\
\midrule
\multicolumn{8}{l}{\textbf{Sparse Detectors}}\\
BoxeR-3D & -/- & 63.9/63.7 & 70.4/70.0 & 61.5/53.7 & 64.7/53.5 & -/- & 50.2/48.9 \\
TransFusion-L & -/64.9 & -/65.1 & -/- & -/63.7 & -/- & -/65.9 & -/- \\
Voxel-DETR (ours) & 68.8/66.1 & 67.8/67.2 & 75.4/74.9 & 69.7/63.1 & 77.6/70.5 & 69.0/67.9 & 71.7/70.5\\
ConQueR (ours) & 70.3/67.7 & 68.7/68.2 & 76.1/75.6 & 70.9/64.7 & 79.0/72.3 & 71.4/70.1 & 73.9/72.5 \\
ConQueR \dag (ours) & \underline{73.1}/\underline{70.6} & \textbf{71.0}/\textbf{70.5} & 78.4/77.9 & \underline{73.7}/68.1 & 80.9/75.2 & \underline{74.5}/\underline{73.3} & \underline{77.3}/\underline{76.1} \\
ConQueR \ddag (ours) & \textbf{74.0}/\textbf{71.6} & \textbf{71.0}/\textbf{70.5} & 78.4/77.9 & \textbf{75.8}/\textbf{70.1} & \textbf{82.4}/\textbf{76.6} & \textbf{75.2}/\textbf{74.1} & \textbf{77.5}/\textbf{76.4} \\
\bottomrule
\specialrule{1pt}{1pt}{0pt}
\end{tabular}
\vspace{-1mm}
\caption{Performances on the WOD \emph{validation} split. All models take single-frame input with the same range, no pre-training or ensembling is required. $\dag$ denotes using the 2$\times$ wider ResNet~\cite{zagoruyko2016wide} with $1/4$ downscaled BEV feature map in our backbone. $\ddag$ denotes conducting NMS on pedestrians and cyclists. \textbf{Bold} denotes the best entries, and \underline{underline} denotes the second-best entries. $_{\rm ts}$ denotes the two-stage model.
}
\label{tab:sota}
\end{table*}
\subsection{Implementation Details}
\noindent {\bf Training.} We follow common practice as previous voxel-based methods~\cite{shi2020pv,yin2021center,shi2021pv,shi2022pillarnet} to use point cloud range of $[-75.2m, 75.2m]\times[-75.2m, 75.2m]\times[-2.0m, 4.0m]$ with voxel size $[0.1m, 0.1m, 0.15m]$ in $x$, $y$, and $z$-axes respectively. The same set of augmentations (\emph{i.e.}, GT-Aug, flip, rotation, scaling) are adopted following the previous works~\cite{yin2021center}. We follow ~\cite{wang2021pointaugmenting,bai2022transfusion} to use the ``fade-strategy'' to drop GT-Aug at the last epoch to avoid overfitting. Both our baseline Voxel-DETR and ConQueR are trained for 6 epochs unless otherwise specified. We use the OneCycle~\cite{smith2019super} learning rate scheduler and AdamW~\cite{loshchilov2017decoupled} optimizer with maximal learning rate 0.001.
\noindent{\bf Network.} For the 3D backbone in Fig.~\ref{fig:contrastive_voxel_detr}, we use the same architecture as ResNet-18~\cite{he2016deep} but use sparse 3D convolutions~\cite{graham2017submanifold} to replace the 2D ones. No pre-trained weights are used. The same FPN structure as RetinaNet~\cite{lin2017focal} is used to obtain multi-scale BEV features. For simplicity, we only use the 8$\times$ downscaled features as input to the transformer, which adopts 3 encoder layers and 3 decoder layers for computation efficiency. We select top-$1000$ scored query predictions from the encoder's class-agnostic prediction head as object queries. We adopt top-N (\emph{e.g.}, 300) scored predictions, or score threshold (\emph{e.g.}, $\geq$ 0.1) during inference. We set $\alpha=1$, $\beta=4$, $\gamma=2$ in Eq.~(\ref{eq:det_loss}). For the proposed Query Contrast, we use $\tau=0.7$ in Eq.~(\ref{eq:qc}), and adopt $T=3$ noising groups with a maximal box noise ratio of 0.4~\cite{li2022dn}, and label noise ratio of 0.5~\cite{li2022dn}. Category labels are simply encoded as one-hot embeddings rather than the learnable embeddings in DN-DETR~\cite{li2022dn}.
\subsection{Main Results}
For fair comparison, all methods included use the same point cloud input range, do not use any pre-trained weights, test-time augmentation or model ensembling.
\vspace{-4mm}
\paragraph{Performance.}
As shown in Table~\ref{tab:sota}, state-of-the-art 3D detectors are divided into dense and sparse categories according to whether they can \emph{directly} generate sparse detections. Our sparse detector ConQueR sets new records on \emph{all categories} of the WOD \emph{validation} set. ConQueR with direct sparse predictions (the second-last entry) achieves $\sim$1.0 mAPH/L2 higher than the previous best single-frame model PV-RCNN++~\cite{shi2021pv}, and is over 3.0 mAPH/L2 higher than the popular anchor-free CenterPoint~\cite{yin2021center}. Notably, ConQueR demonstrates overwhelming performance on pedestrians and cyclists, outperforming previous best methods by $\sim$\textbf{2.0} APH/L2, which shows the effectiveness of our Query Contrast strategy especially for densely populated categories. The significant performance improvements can also be validated on the WOD \emph{test} set in Table~\ref{tab:sota_test}. Moreover, ConQueR surpasses previous best sparse detectors TransFusion-L by \textbf{$\sim$6.0} mAPH/L2, closing the performance gap between sparse and dense 3D detectors. When compared with our baseline Voxel-DETR, the proposed Query Contrast mechanism brings over \textbf{1.6} mAPH/L2 without any extra inference cost. Besides, our baseline Voxel-DETR with only 6 epochs of training outperforms previous sparse 3D detectors, and achieves comparable performance with CenterPoint (36-epoch training) with only $1/6$ GPU hours. In addition, ConQueR has an inference latency of 70ms (46ms for CenterPoint)\footnote{Latency is measured with batch size 1 on NVIDIA A100 GPU.}.
Although ConQueR with \emph{direct} sparse predictions already achieves state-of-the-art performance, we find that applying NMS onto ConQueR's sparse predictions can further improve small and densely populated categories such as pedestrians, while NMS causes $\sim$1.2 APH/L2 performance drop on the well-trained vehicles (as shown in Appendix.~\ref{appdenx:a}). This is also the case with our baseline Voxel-DETR. We speculate this is caused by the learning difficulties inherent in the data for extremely similar queries (as shown in Fig.~\ref{fig:motivation}(b)) . We thus report ConQueR's performance after conducting NMS on pedestrians and cyclists (the last entry of Table~\ref{tab:sota}).\footnote{FSD~\cite{fan2022fully} adopts larger point cloud ranges and requires a point segmentation pre-train weights. It is $\sim$1.0 mAPH/L2 lower than our ConQueR.}
\begin{table}[t]
\footnotesize
\centering
\begin{tabular}{l|>{\columncolor[gray]{0.95}}c|c|c|c}
\hline
\multicolumn{1}{c|}{Methods} & All & Veh. & Ped. & Cyc. \\ \hline
CenterPoint~\cite{yin2021center} & 69.0 & 71.9 & 67.0 & 68.2 \\ \hline
PV-RCNN++~\cite{shi2021pv} & 70.2 & 73.5 & 69.0 & 68.2 \\ \hline
AFDetv2~\cite{hu2022afdetv2} & 70.0 & 72.6 & 68.6 & 68.7 \\ \hline
PillarNet-34~\cite{shi2022pillarnet} & 69.6 & \bf74.7 & 68.5 & 65.5 \\ \hline
ConQueR (Ours) & \bf72.0 & 73.3 & \bf70.9 & \bf71.9 \\ \hline
\end{tabular}
\caption{Single-frame performance comparisons on the WOD \emph{test} set. APH/L2 results are reported.}
\label{tab:sota_test}
\vspace{-2mm}
\end{table}
\begin{table}[t]
\footnotesize
\centering
\begin{tabular}{l|c|c|c|c}
\hline
\multicolumn{1}{c|}{Methods} & Preds/Scene & Veh. & Ped. & Cyc. \\ \hline
$\text{CenterPoint}_{\rm nms}$ & 192 & 66.4 & 62.9 & 67.9 \\ \hline
$\text{Transfusion}_{\rm topN}$ & 300 & 65.1 & 63.7 & 65.9 \\ \hline \hline
$\text{Voxel-DETR}_{\rm topN}$ & 300 & 67.1 & 63.0 & 67.8 \\ \hline
$\text{Voxel-DETR}_{\rm score}$ & 222 & 67.2 & 63.1 & 67.9 \\ \hline \hline
$\text{ConQueR}_{\rm topN}$ & 300 & 68.0 & 64.6 & 70.0 \\ \hline
$\text{ConQueR}_{\rm score}$ & 131 & 68.2 & 64.7 & 70.1 \\ \hline
$\text{ConQueR}_{\rm score}$ \dag & 122 & \bf70.5 & \bf68.1 & \bf73.3 \\ \hline
\end{tabular}
\caption{Sparsity of final predictions. APH/L2 results are reported on the WOD \textit{validation} set. The subscripts of each entry denotes the way they obtain final predictions. For example, CenterPoint$_{\rm nms}$ uses NMS to filter out duplicated boxes, and Voxel-DETR$_{\rm topN}$ denotes it uses top-$N$ scored proposals as final predictions, while ConQueR$_{\rm score}$ denotes that using score thresholding to generate final sparse predictions. $\dag$ denotes our best model in Table~\ref{tab:sota}.}
\label{tab:sparsity}
\vspace{-4mm}
\end{table}
\vspace{-2mm}
\paragraph{Sparsity.}
Apart from the performance improvements on the WOD official metrics, ConQueR shows great potential in reducing false positives and improving the sparsity of final predictions. We list the average number of predictions per scene for different 3D detectors in Table~\ref{tab:sparsity}. For the baseline Voxel-DETR, thresholding according to scores helps to reduce $\sim$25\% predictions per sample with slightly better performance. With the help of Query Contrast, ConQueR further reduces the number of predictions substantially by $\sim$60\%. Besides, as the performance of ConQueR continually improves (the last two lines), the sparsity of final predictions steadily improve as well.
When we adopt the same top-300 predictions as baseline Voxel-DETR$_{\rm topN}$ for evaluation, ConQueR$_{\rm topN}$ still improves the detection performance significantly. This indicates the Query Contrast mechanism contributes to generating more accurate predictions from best matched queries.
Furthermore, our ConQueR can achieve much sparser predictions even compared with NMS-based dense detectors such as CenterPoint.
\vspace{1mm}
\subsection{Ablation Study}
\paragraph{Components of Query Contrast.}
We deduce the components of ConQueR to baseline Voxel-DETR by gradually removing multi-positive pairs, auxiliary de-noising loss, and contrastive loss in Table~\ref{tab:QC_components}.
Compared to ConQueR (the first row), removing the multiple noised copies of GTs from contrastive learning (the second row) causes over 0.6 mAPH/L2 performance drop. If we further remove the auxiliary denoising loss (the third row), performances of vehicles and pedestrians classes even become slightly better, indicating that the auxiliary denoising loss alone is not the key for performance improvements. Moreover, we can find that Query Contrast with only original GTs (the second last entry) already improves over the baseline (the last entry) dramatically especially on pedestrians and cyclists. Overall, the Query Contrast scheme brings \textbf{1.1, 1.7, 2.3} APH/L2 improvements for vehicles, pedestrians and cyclists respectively.
\begin{table}[t]
\footnotesize
\begin{center}
\begin{tabular}{ccc|ccc}
\toprule
\multirow{2}{*}{\shortstack[1]{InfoNCE\\Loss}} & \multirow{2}{*}{\shortstack[1]{Aux\\DN}} & \multirow{2}{*}{\shortstack[1]{Multi\\Pos}} & \multicolumn{3}{c}{APH/L2} \\
& & & Veh. & Ped. & Cyc.\\
\midrule
\checkmark & \checkmark & \checkmark & \bf68.2 & \bf64.7 & \bf70.1 \\
\checkmark & \checkmark & & 67.4 ({\color{red} -0.8}) & 64.1 ({\color{red} -0.6}) & 69.6 ({\color{red} -0.5}) \\
\checkmark & & & 67.5 ({\color{blue} +0.1}) & 64.2 ({\color{blue} +0.1}) & 69.3 ({\color{red} -0.3}) \\
& & & 67.1 ({\color{red} -0.4}) & 63.0 ({\color{red} -1.2}) & 67.8 ({\color{red} -1.5}) \\
\bottomrule
\end{tabular}
\vspace{-4mm}
\end{center}
\caption{Effects of components in Query Contrast. The numbers in brackets denotes the performance drop ({\color{red}red}) or increase ({\color{blue}blue}) for each component. Both the multi-positive contrastive loss (Multi-Pos) and the InfoNCE loss (Eq.~(\ref{eq:qc})) from only original GTs have deep impact on performance, while the auxiliary denoising loss (Aux-DN) only has marginal effects.}
\label{tab:QC_components}
\end{table}
\vspace{-2mm}
\paragraph{Effects of different supervisions or similarity metrics for GT-query pairs.}
We demonstrate the effects of different type of supervision or similarity metrics applied to GT-query pairs in Table~\ref{tab:sim_measure}. As discussed in Sec.~\ref{sec:cns}, simple geometric relations like GIoU cannot sufficiently measure the similarities between GTs and queries because they cannot take the appearance information into account, thus only have marginal effects compared to our baseline Voxel-DETR. If we replace Query Contrast with the MSE loss in knowledge distillation (KD) to supervise positive GT-query pairs, performance of vehicles is still comparable with our Query Contrast strategy (the last entry), but it cannot handle densely populated categories like pedestrians and cyclists, indicating the importance of suppressing negative GT-query pairs in our Query Contrast strategy.
\begin{table}[t]
\footnotesize
\centering
\begin{tabular}{l|c|c|c}
\hline
\multicolumn{1}{c|}{Methods} & Veh. & Ped. & Cyc. \\ \hline
Voxel-DETR & 67.1 & 63.0 & 67.8 \\ \hline
ConQueR$_{\rm KD-MSE}$ & 68.1 & 63.4 & 68.2 \\ \hline
ConQueR$_{\rm QC-GIoU}$ & 66.6 & 63.6 & 68.4 \\ \hline
ConQueR$_{\rm QC-Cos}$ & \bf68.2 & \bf64.7 & \bf70.1 \\ \hline
\end{tabular}
\caption{Effects of different supervisions or similarity metrics applied to GT-query pairs. APH/L2 results are reported. $_{\rm QC-Cos}$ denotes our default Query Contrast with the cosine similarity metric, while $_{\rm QC-GIoU}$ denotes using GIoU as the similarity measurement of GT-query pairs. $_{\rm KD-MSE}$ indicates replacing Query Contrast with Knowledge Distillation MSE loss to supervise positive GT-query pairs only.}
\label{tab:sim_measure}
\vspace{-2mm}
\end{table}
\vspace{-2mm}
\paragraph{Number of positive pairs.}
We present the results of using different numbers of noised GT copies in Table~\ref{tab:dn_groups}. We observe that using 3 groups of noised copies without original GTs (default setting) achieves the best performance. Moreover, incorporating original GT into the multi-positive contrastive loss harms the performance. The first two entries show that using single noised copies of GTs is better than using the original GTs. We conjecture this is caused by the lack of training for original GT boxes. The detector is only trained to recover from noised GTs, while having no idea how to deal with perfectly located original GTs.
\begin{table}[t]
\footnotesize
\centering
\begin{tabular}{cc|c|c|c}
\hline
\multicolumn{1}{c|}{Original GTs} & $\#$ Noised GT Groups & Veh & Ped & Cyc \\ \hline
\checkmark & 0 & 67.5 & 64.2 & 69.3 \\
\rowcolor[gray]{0.9}
& 1 & 67.9 & 64.4 & 69.6 \\ \cline{1-5}
& 2 & 68.2 & 64.3 & 69.9 \\
\rowcolor[gray]{0.9}
\checkmark & 2 & 67.8 & 64.4 & 68.8 \\ \cline{1-5}
& 3 & \bf68.2 & \bf64.7 & \bf70.1 \\
\rowcolor[gray]{0.9}
\checkmark & 3 & 68.0 & 64.3 & 69.9 \\
& 4 & 67.7 & 64.4 & 70.1 \\ \hline
\end{tabular}
\caption{Number of positive pairs in the contrastive loss. APH/L2 results are reported on the WOD \emph{validation} split. $\checkmark$ denotes including the original GT group into Eq.~(\ref{eq:qc}).}
\label{tab:dn_groups}
\end{table}
\vspace{-2mm}
\paragraph{Query-GT feature alignment.}
We demonstrate the importance of aligning query embeddings to GTs' with an extra MLP in Table~\ref{tab:afp}. Removing the MLP for query embeddings alignment (the first row) or applying the MLP alignment for both GT and query embeddings (the last row) causes $\sim$1 APH/L2 performance drop, indicating the importance of the asymmetric alignment design to mitigate the distribution gap between GT and query embeddings.
\begin{table}[t]
\footnotesize
\centering
\begin{tabular}{c|c|c|c}
\hline
Projection & Veh. & Ped. & Cyc. \\ \hline
& 67.2 & 64.2 & 69.3 \\ \hline
Q & \bf68.2 & \bf64.7 & \bf70.1 \\ \hline
G\&Q & 67.3 & 64.1 & 68.9 \\ \hline
\end{tabular}
\caption{Design choices of the asymmetric feature alignment. APH/L2 results are reported. `G' and `Q' denotes GT and query embeddings respectively from the selected layer in detector or prediction heads.}
\label{tab:afp}
\vspace{-1mm}
\end{table}
\vspace{-2mm}
\paragraph{Neural Layers for conducting Query Contrast.}
We compare 3 layer alternatives to conduct Query Contrast in Table~\ref{tab:layer_source}: the output layer of each decoder layer, the output layer of each FFN prediction head, and the second-last layer of each FFN prediction head. The Query Contrast scheme can bring consistent improvements for all layer choices, and the features from the last layer of FFN prediction head performs the best, indicating that directly regulate the detection outputs via the contrastive loss can achieve the ``enhance-suppress'' effects onto queries to the utmost.
\begin{table}[t]
\footnotesize
\centering
\begin{tabular}{l|c|c|c}
\hline
Layer to Contrast & Veh. & Ped. & Cyc. \\ \hline
Last$_{\rm decoder}$ & 68.1 & 63.9 & 69.7 \\ \hline
Last$_{\rm FFN}$ & \bf68.2 & \bf64.7 & \bf70.1 \\ \hline
SecondLast$_{\rm FFN}$ & 67.4 & 64.6 & 69.6 \\ \hline
\end{tabular}
\caption{Layers to conduct Query Contrast. Results are the APH/L2 reported on the WOD \emph{validation} split. Last$_{\rm decoder}$ and Last$_{\rm FFN}$ denotes the output layer of each decoder layer and FFN prediction head respectively, while SecondLast$_{\rm FFN}$ indicates the second-last layer of each FFN prediction head is chosen to conduct Query Contrast.}
\label{tab:layer_source}
\vspace{-3mm}
\end{table}
\vspace{-3mm}
\paragraph{Generalisation ability w.r.t. query numbers.}
We verify the generalization ability of Query Contrast by varying query numbers in Table~\ref{tab:vary_query_number}. By default we adopt top-1000 scored proposals as initial queries to input to the transformer decoder. The performance gain of Query Contrast is relatively stable when we gradually reduce query numbers to 500 and 300.
\begin{table}[t]
\footnotesize
\centering
\begin{tabular}{l|c|c|c|c}
\hline
\multicolumn{1}{c|}{Methods} & $\#$Query & Veh. & Ped. & Cyc. \\ \hline
Voxel-DETR & 300 & 66.3 & 62.0 & 66.5 \\ \hline
ConQueR & 300 & 67.0 ({\color{blue} +0.7}) & 63.6 ({\color{blue} +1.6}) & 68.9 ({\color{blue} +2.4}) \\ \hline
Voxel-DETR & 500 & 66.9 & 62.8 & 67.3 \\ \hline
ConQueR & 500 & 67.8 ({\color{blue} +0.9}) & 64.4 ({\color{blue} +1.6}) & 69.0 ({\color{blue} +1.7}) \\ \hline
Voxel-DETR & 1000 & 67.1 & 63.0 & 67.8 \\ \hline
ConQueR & 1000 & {\bf68.2} ({\color{blue} +1.1}) & {\bf64.7} ({\color{blue} +1.7}) & {\bf70.1} ({\color{blue} +2.3}) \\ \hline
\end{tabular}
\caption{Improvements of Query Contrast under different query numbers. APH/L2 results are reported. The {\color{blue} blue} numbers in brackets indicates the performance gains.}
\label{tab:vary_query_number}
\end{table}
\vspace{-2mm}
\paragraph{EMA coefficients for generating GT embeddings.}
Here we show results of different momentums of our EMA decoder, which is used to embed GT boxes, in Table~\ref{tab:ema}. The performance of using the same decoder as queries (the first line) already achieves satisfactory results, while introducing a more stable decoder for GT boxes can further improve the performance especially on categories with fewer instances (\emph{i.e.}, cyclists).
\begin{table}[t]
\footnotesize
\centering
\begin{tabular}{c|c|c|c}
\hline
Momentum & Veh. & Ped. & Cyc. \\ \hline
0 & 67.9 & 64.4 & 69.0 \\ \hline
0.9 & 67.6 & 64.3 & 69.1 \\ \hline
0.99 & 68.0 & 64.5 & 69.2 \\ \hline
0.999 & \bf68.2 & \bf64.7 & \bf70.1 \\ \hline
\end{tabular}
\caption{Effects of EMA momentum coefficient.}
\label{tab:ema}
\end{table}
\vspace{-2mm}
\paragraph{Temperature coefficient in Eq.~(\ref{eq:qc}).}
We shown the effects of different $\tau$ in Table~\ref{tab:tau}. $\tau$ controls the contrastive learning difficulty of the GT-query similarities, and we find $\tau=0.7$ leads to the best performance.
\begin{table}[t]
\footnotesize
\centering
\begin{tabular}{c|c|c|c}
\hline
$\tau$ & Veh. & Ped. & Cyc. \\ \hline
1.0 & 67.9 & 64.2 & 69.8 \\ \cline{1-1} \cline{2-4}
0.7 & \bf68.2 & \bf64.7 & \bf70.1 \\ \cline{1-1} \cline{2-4}
0.5 & 67.6 & 64.5 & 69.7 \\ \cline{1-1} \cline{2-4}
\end{tabular}
\caption{Effects of $\tau$. APH/L2 results are reported.}
\label{tab:tau}
\vspace{-2mm}
\end{table}
\section{Conclusion}
DETR-based sparse 3D detectors faces the problem of duplicated false positives, and lags in detection performance. In this paper, we solve these challenges with our simple yet effective Query Contrast Voxel-DETR (ConQueR). The problem of duplicated false positives is mainly caused by the lack of supervisions in handlings dense similar queries. Based on our sparse 3D detection framework Voxel-DETR, we propose a Query Contrast strategy to explicitly suppress densely overlapping false positives, and simultaneously promote the best matched queries towards their assigned GTs in a contrastive manner. ConQueR reduces $\sim$60$\%$ false positives in the final sparse predictions, closes the gap between sparse and dense 3D detectors, and surpasses previous state-of-the-art 3D detectors by a large margin on the challenging WOD benchmark.
{\small
\bibliographystyle{ieee_fullname}
|
\section{Introduction} \label{sec:introduction}
Storing data in deoxyribonucleic acid (DNA) promises unprecedented density and durability and is seen as the new frontier of data storage. Recent experiments have already demonstrated the viability of DNA-based data storage \cite{church_next-generation_2012, grass_robust_2015,yazdi_portable_2017}.
The DNA storage channel suffers from multiple impairments and constraints due to the synthesis and sequencing of DNA and the limitations of current technologies. In particular, errors in the form of insertions, deletions, and substitutions (IDS) occur. This has spurred a great deal of research on devising coding schemes for the DNA storage channel.
While the literature on error-correcting coding for the DNA storage channel is abundant, most works have considered
a simplified and unrealistic channel model with a small and fixed number of insertions and/or deletions, i.e., an \emph{adversarial} channel. Fewer works have considered a more realistic probabilistic channel model in which errors occur with a given probability, e.g., \cite{Maa22,Srinivasavaradhan2021TrellisBMA,davey_reliable_2001,briffa_improved_2010,inoue_adaptive_2012,shibata_design_2019,Koremura2020,SakogawaMAP_IDS_2020,Shibata2020,Pfister2021,mansour_convolutional_2010,buttigieg_improved_2015}.
These works usually assume independent and identically distributed (i.i.d.) errors. In DNA storage, however, IDS errors are not i.i.d., but the channel has memory \cite{Hamoum_IDSnanopore2021,McBain2022FiniteStateSC}. Potentially, the memory is as large as the whole DNA sequence.
In \cite{Hamoum_IDSnanopore2021}, the authors proposed a statistical model of the DNA storage channel with nanopore sequencing based on a Markov chain that models the inherent memory of the channel. In particular, it builds on the fact that in the MinION technology, strands traverse a nanopore nucleotide by nucleotide, and an electrical signal is generated for every group of $k$ successive nucleotides, called $k$mers. Let $\ensuremath{\boldsymbol{x}}$ be the DNA strand to be synthesized and $\ensuremath{\boldsymbol{z}}$ be the sequence of channel events, where $z_i\in\{\tt{insertion}, \tt{deletion}, \tt{substitution}, \tt{no}\,\tt{error}\}$. The key idea in \cite{Hamoum_IDSnanopore2021} is then to assume that $z_i$ depends on the symbols $x_{i-k+1},\ldots,x_i$ and the previous event $z_{i-1}$ and estimate the probabilities $p(z_i|x_{i-k+1},\ldots,x_i,z_{i-1})$ from experimental data. We refer to this channel model as the memory-$k$ nanopore channel model.
The starting point of this paper is the work \cite{Hamoum_IDSnanopore2021}. Our main contributions are as follows.
\begin{itemize}
\item We derive the optimal maximum a posteriori (MAP) decoder for the memory-$k$ nanopore channel model. The complexity of the decoder increases exponentially with $k$. Based on this decoder, we derive achievable information rates (AIRs).
\item The AIR for a memory-$k$ nanopore channel can be seen as an AIR for the true DNA storage channel of a \emph{mismatched} decoder (i.e., a decoder that it is not matched to the true channel) that assumes that the channel is a memory-$k$ nanopore channel. We show that for increasing $k$, the AIR improves\textemdash meaning that the decoder is better matched to the true channel\textemdash and eventually saturates. This allows us to quantify the trade-off between decoding complexity and performance loss incurred by the suboptimal decoder.
\item The derived MAP decoder can be used to design error-correcting coding schemes tailored to the memory-$k$ nanopore sequencing channel. In particular, we consider the concatenated coding scheme proposed in \cite{Maa22} and multiple reads of the DNA strand, and we optimize the inner and outer code. We show that the concatenated coding scheme of \cite{Maa22} yields excellent performance at rates close to the AIRs.
\item We validate the AIRs and simulation results of the nanopore channel model using the dataset of DNA reads in \cite{Srinivasavaradhan2021TrellisBMA} obtained using Oxford Nanopore Technologies (ONT) sequencing with the MinION technology.
\end{itemize}
\section{The Nanopore Sequencing Channel Model}
\label{sec:system}
We consider the memory-$k$ nanopore channel model introduced in \cite{Hamoum_IDSnanopore2021}. The key property of this channel model is that it removes the assumption of i.i.d. IDS errors. In particular, the occurrence of an error event, or a correct transmission, depends on the previous event (an insertion, deletion, substitution, or no error event), and previous channel input symbols. Let $\ensuremath{\boldsymbol{x}} = (x_1,\ldots ,x_\ensuremath{N})$, $x_i \in \ensuremath{\Sigma} _4 = \{0,1,2,3\}$, be the channel input sequence and $\ensuremath{\boldsymbol{y}} = (y_1,\ldots,y_{\ensuremath{N}'})$ the channel output sequence. Note that $\ensuremath{N}'$ is random and depends on the number of insertion and deletion events, i.e., $N'\neq N$ in general. Also, let $k\text{mer}_i = (x_{i -k + 1},\dots,x_i)$ be the $k$mer at time instant $i$. The channel model is depicted in Fig.~\ref{fig:ids:channel}. Here, we denote the event associated with $k\text{mer}_{i+1}$ by $z_i$, where $z_i$ represents an insertion, deletion, substitution, or no error acting on symbol $x_i$ when it is to be transmitted. Hence, $z_i \in \{{\tt{{Ins}}}, {\tt{{Del}}}, {\tt{{Sub}}}, {\tt{NoErr}} \}$. The transition probabilities $p_1$, $p_2$, $p_3$, and $p_4$ in the figure are of the form $p(z_i|k\text{mer}_i, z_{i-1})$, and their value depends on the $k$mer. In the event of a deletion, the symbol $x_i$ in $k\text{mer}_i$ will be deleted and nothing will be appended to $\ensuremath{\boldsymbol{y}}$, while when a no error event occurs, $\ensuremath{\boldsymbol{y}}$ will be appended with $x_i$. Furthermore, in the event of an insertion, i.e., $z_i = {\tt{{Ins}}}$, the channel will insert $L$ symbols $(a_1,\dots,a_{L})$, where the length $L$ is random. Each symbol $a_j$ will be uniformly picked from $\ensuremath{\Sigma}_4$. In this case, the channel appends $\ensuremath{\boldsymbol{y}}$ with $(a_1,\dots,a_{L})$. This also applies to a substitution event where the symbol $a^*$ is uniformly picked from $\Sigma_4 \setminus \{x_i\}$ and appended to $\ensuremath{\boldsymbol{y}}$ instead of $x_i$. Moreover, we define $p(L| k\text{mer}_i, z_i = {\tt{{Ins}}})$ as the probability of inserting $L$ symbols given $k\text{mer}_i$ and an insertion event, and $p(a^*| k\text{mer}_i, z_i = {\tt{{Sub}}})$ as the probability of $x_i$ being substituted by $a^*$ given $k\text{mer}_i$ and a substitution event. In all these scenarios, the channel always moves from state $k\text{mer}_i$ to state $k\text{mer}_{i+1}$. After the last symbol $x_{\ensuremath{N}}$ has been processed by the channel, the channel outputs $\ensuremath{\boldsymbol{y}}$.
\input{Figures/Channel_Model.tex}
\subsection{Channel Transition Probability Estimation} \label{sec:model_build}
The transition probabilities of the channel model can be estimated using a DNA storage dataset $\mathcal{D}$ containing input (before synthesis) and the corresponding multiple output (after sequencing) sequences.
Let $\ensuremath{\boldsymbol{x}}^{(l)}$ denote the $l$-th input sequence and $\ensuremath{\boldsymbol{y}}^{(l)}_1, \ldots, \ensuremath{\boldsymbol{y}}^{(l)}_{\ensuremath{M}_l}$ the corresponding $\ensuremath{M}_l \geq 1$ output sequences from the dataset $\mathcal{D}$. The method we use to estimate the transition probabilities is as follows. For each pair $(\ensuremath{\boldsymbol{x}}^{(l)}, \ensuremath{\boldsymbol{y}}_j^{(l)})$, we compute the Levenshtein distance between the two vectors using a lattice (see \cite[Sec.~III]{Maa22},\cite{bahl_decoding_1975}). By backtracking from the end of the lattice, a corresponding sequence of events $\ensuremath{\boldsymbol{z}}_j^{(l)}$ can be identified. In case of ties, one event is chosen uniformly at random. Finally, we estimate the probabilities $p(z_i=z|k\text{mer}_i, z_{i-1}=z')$ by a simple counting argument. In particular, let $\mathcal{S}_{z,z',k\text{mer}}^{(l,j)}$ denote the set $\left\{u: (x^{(l)}_{u-k+1},\ldots,x^{(l)}_{u})=k\text{mer}, (z_{j,u-1}^{(l)},z_{j,u}^{(l)})=(z',z) \right\}$. Then, $p(z_i=z|k\text{mer}_i, z_{i-1}=z')$ is estimated by
\begin{align*}
\frac{\sum_{l=1}^{|\mathcal{D}|}\sum_{j=1}^{\ensuremath{M}_l} |\mathcal{S}_{z,z',k\text{mer}_i}^{(l,j)}|}{\sum_{z \in \{{\tt{{Ins}}},{\tt{{Del}}},{\tt{{Sub}}},{\tt{NoErr}} \}} \sum_{l=1}^{|\mathcal{D}|}\sum_{j=1}^{\ensuremath{M}_l} |\mathcal{S}_{z,z',k\text{mer}_i}^{(l,j)}|}\,.
\end{align*}
\vspace{-1ex}
\input{Figures/System_Model.tex}
\subsection{Multiple Reads}
We consider the relevant scenario in which the sequencing process outputs multiple reads ($M$) of the DNA strand. In particular, we model this by assuming that a strand is transmitted over $M$ independent channels.
\section{Coding Scheme and System Model}
We consider the concatenated coding scheme and the low-complexity \emph{separate decoding} scheme proposed in \cite{Maa22}, which decodes separately each noisy strand and combines a posteriori probabilities (APPs) from all reads before passing them to the outer decoder. The primary goal of the inner code is to maintain synchronization with the transmitted sequence and provide likelihoods to the outer decoder. The outer code then corrects remaining errors.
The system model is shown in Fig.~\ref{fig:system:model}. We consider a low-density parity-check (LDPC) code for the outer code and a convolutional code for the inner code. The information data $\u = (u_1, \ldots, u_\ensuremath{K})$, $u_i \in \field{\ensuremath{q_\mathsf{o}}}$, of length $\ensuremath{K}$, is first encoded by an $[\ensuremath{N_\mathsf{o}},\ensuremath{K}]_{\ensuremath{q_\mathsf{o}}}$ outer LDPC code over $\field{\ensuremath{q_\mathsf{o}}}$ into the codeword $\ensuremath{\boldsymbol{w}} = (w_1,\dots,w_{\ensuremath{N_\mathsf{o}}})$, $w_i \in \field{{\ensuremath{q_\mathsf{o}}}}$, where $\field{{\ensuremath{q_\mathsf{o}}}}$ is a binary extension field of size $2^{\ensuremath{k}_{\text{CC}}}$ and $\ensuremath{k}_{\text{CC}}$ is the (binary) dimension of the inner convolutional code. Since current sequencing technologies cannot handle long sequences, the codeword $\ensuremath{\boldsymbol{w}}$ is split into $S$ subsequences $\ensuremath{\boldsymbol{w}}_1,\dots, \ensuremath{\boldsymbol{w}}_S$, each of length $\ensuremath{N_\mathsf{o}^\mathsf{b}} = \nicefrac{\ensuremath{N_\mathsf{o}}}{S}$, where $\ensuremath{\boldsymbol{w}}_j = (w_{(j-1)\ensuremath{N_\mathsf{o}^\mathsf{b}} + 1},\dots, w_{j\ensuremath{N_\mathsf{o}^\mathsf{b}}})$. Each $\ensuremath{\boldsymbol{w}}_j$ is encoded, with an optional offset \cite{Maa22}, by an $(\ensuremath{n},\ensuremath{k}_{\text{CC}}, \ensuremath{m})_4$ inner convolutional code of binary dimension $\ensuremath{k}_{\text{CC}}$, memory $m$, and output alphabet $\ensuremath{\Sigma} _4$ into the inner codeword $\ensuremath{\boldsymbol{x}}_j$ of length $N=(\ensuremath{N_\mathsf{o}^\mathsf{b}} + \ensuremath{m})n$, where $\ensuremath{m}$ is due to terminating the convolutional code to the all-zero state. The overall code rate is given as $R = \ensuremath{R_\mathsf{o}} \ensuremath{R_\mathsf{i}} = \nicefrac{\ensuremath{K} \ensuremath{k}_{\text{CC}}}{S\ensuremath{N}}$, where $\ensuremath{R_\mathsf{o}} = \nicefrac{\ensuremath{K}}{\ensuremath{N_\mathsf{o}}}$ is the rate of the outer code and $\ensuremath{R_\mathsf{i}} = \nicefrac{\ensuremath{k}_{\text{CC}} \ensuremath{N_\mathsf{o}^\mathsf{b}}}{\ensuremath{N}}$ the rate of the inner code.
Each $\ensuremath{\boldsymbol{x}}_j$ is transmitted independently over $\ensuremath{M}$ identical nanopore channels in order to model the multiple copies of a DNA strand at the output of the sequencing process. At the receiver, each noisy read $\ensuremath{\boldsymbol{y}}_{j,1},\ldots,\ensuremath{\boldsymbol{y}}_{j,\ensuremath{M}}$ for transmitted sequence $\ensuremath{\boldsymbol{x}}_j$ is decoded separately with a MAP inner decoder. The APPs at their output are then combined to provide approximate APPs for $\ensuremath{\boldsymbol{w}}_j$. The (approximate) APPs for the symbols in $\ensuremath{\boldsymbol{w}}=(\ensuremath{\boldsymbol{w}}_1,\ldots,\ensuremath{\boldsymbol{w}}_S)$ are concatenated and then passed to the outer decoder. The outer decoder uses these APPs to decide on an estimate $\hat{\u}$ of $\u$.
\section{Symbolwise MAP Decoding for Channel/Inner Code (Inner Decoding)} \label{sec:innerdec-one}
In this section, we derive the optimum (MAP) decoder for the combination of the memory-$k$ nanopore channel model and the inner code. To this end, we use the fact that the combination of the channel and the inner code can be seen as a hidden Markov model (HMM) by introducing a \emph{drift} variable as in \cite{davey_reliable_2001}. The drift $d_i$, $0 \leq i < \ensuremath{N_\mathsf{o}^\mathsf{b}} + m$, is defined as the number of insertions minus the number of deletions that occurred before symbol $x_{ni+1}$ is to be acted on by the event $z_{ni}$, while $d_{\ensuremath{N_\mathsf{o}^\mathsf{b}}+m}$ is defined as the number of insertions minus deletions that occurred after the last symbol $x_{N}$ has been processed by the channel. Thus, by definition, $d_0 = 0$ and $d_{\ensuremath{N_\mathsf{o}^\mathsf{b}}+m} = N'-N$, both known to the decoder. Adding the drift to the joint state of the channel and the inner code, i.e., to $(k\text{mer}_i,z_i, s_i)$, where $s_i$ denotes the state variables of the inner convolutional code, gives the state variable $\sigma_{i} = (k\text{mer}_i,z_i, s_i, d_i)$ of the HMM. In this HMM, a transition from time $i-1$ to time $i$ corresponds to a transmission of symbols $\ensuremath{\boldsymbol{x}}_{(i-1)n+1}^{in}$, where $\ensuremath{\boldsymbol{x}}_a^b = (x_a,x_{a+1},\dots, x_b)$. Further, when transitioning from a state with drift $d_{i-1}$ to a state with drift $d_{i}$, the HMM emits $n+d_{i}-d_{i-1}$ output symbols depending on both the previous and new drift.
In the following, to simplify notation, the subsequence index of $\ensuremath{\boldsymbol{w}}$ and $\ensuremath{\boldsymbol{y}}$ is omitted and $w_i$ simply refers to the $i$-th symbol of an arbitrary subsequence $\ensuremath{\boldsymbol{w}}_j$, while $\ensuremath{\boldsymbol{y}}=(\ensuremath{\boldsymbol{y}}_1,\ldots,\ensuremath{\boldsymbol{y}}_{\ensuremath{M}})$ refers to the corresponding received sequences.
\subsection{Decoding for a Single Received Sequence} \label{sec:single_seq}
The APP for outer code symbol $w_i$ can be computed as $p(w_i|\ensuremath{\boldsymbol{y}}) = \nicefrac{p(\ensuremath{\boldsymbol{y}},w_i)}{p(\ensuremath{\boldsymbol{y}})}$, where the joint probability $p(\ensuremath{\boldsymbol{y}},w_i)$ can be computed by marginalizing the trellis states, corresponding to the HMM, of the channel/inner code that correspond to symbol $w_i$. Then, we can write $p(\ensuremath{\boldsymbol{y}},w_i) = \sum_{(\sigma,\sigma'):w_i} p(\ensuremath{\boldsymbol{y}},\sigma,\sigma')$, where $\sigma$ and $\sigma'$ are realizations of the random variables $\sigma_{i-1}$ and $\sigma_i$, respectively. Here, the summation is over all pairs of states that correspond to symbol $w_i$. We can use the Markov property to decompose the probability $p(\ensuremath{\boldsymbol{y}},\sigma,\sigma')$ into three parts as
\begin{align*}
&p(\ensuremath{\boldsymbol{y}},\!\sigma\!,\!\sigma') \!=\!
p\!\left(\!\ensuremath{\boldsymbol{y}}_{1}^{(i\!-\!1)n+d}, \sigma\!\right)\!p\!\left(\!\ensuremath{\boldsymbol{y}}_{(i\!-\!1)n\!+d+\!1}^{in+d'}, \sigma'\big|\sigma\!\right)\!p\!\left(\!\ensuremath{\boldsymbol{y}}_{in\!+\!d'\!+1}^{\ensuremath{N}'}\Big| \sigma'\!\right)\!.
\end{align*}
We abbreviate the first, second, and third term of the above equation with $\alpha_{i-1}(\sigma)$, $\gamma_i(\sigma,\sigma')$, and $\beta_i(\sigma')$, respectively. Then, the first and third term can computed recursively as
\begin{align*}
\alpha_i(\sigma') &= \sum_{\sigma}\alpha_{i-1}(\sigma) \gamma_{i}(\sigma,\sigma'),
\beta_{i-1}(\sigma) = \sum_{\sigma'}\beta_i(\sigma') \gamma_i(\sigma,\sigma'). \notag
\end{align*}
The term $\gamma_i(\sigma,\sigma')$ (the branch metric) can be decomposed as
\begin{align*}
\gamma_i(\sigma,\sigma') &= \notag \\
&p(w_i)p\left(\ensuremath{\boldsymbol{y}}_{(i-1)n+d+1}^{in+d'}, d'\big|d,s,s',z',z,k\text{mer}',k\text{mer}\right),
\end{align*}
where $p(w_i)$ is the a priori probability of symbol $w_i$. For simplicity, define the state variable $\zeta_i = (k\text{mer}_i, z_i)$. Then, to compute $p(w_i)p(\ensuremath{\boldsymbol{y}}_{(i-1)n+d+1}^{in+d'}, d'\big|d,s,s',\zeta',\zeta)$, we need to consider each possible event for $z'$. For simplicity, we limit our derivation to $n = 1$. The case for general $n$ follows in a straightforward manner.
\begin{enumerate}
\item $z' = {\tt{{Ins}}}$.
Then,
\begin{align*}
&p\left(\ensuremath{\boldsymbol{y}}_{(i-1)+d+1}^{i+d'}, d'\big|d,s,s',\zeta',\zeta\right)\\&=p\left(\ensuremath{\boldsymbol{y}}_{(i-1)+d+1}^{i+d'}, L = d'-d\big|d,s,s',\zeta',\zeta\right)\\
&\quad \times\; p\left(L = d'-d\big|d,s,s',\zeta',\zeta\right)\\
&= \left(\frac{1}{4}\right)^L \hspace{-1ex}\cdot p(L|k\text{mer}',z' = {\tt{{Ins}}}) \cdot p\left(z' = {\tt{{Ins}}}|k\text{mer}',z\right).
\end{align*}
\item $z' = {\tt{{Del}}}$.
Then, $d' = d-1$, which means that $p(\ensuremath{\boldsymbol{y}}_{(i-1)+d+1}^{i+d'}, d' \neq d - 1\big|d,s,s',\zeta',\zeta) = 0$, and
\vspace{-1ex}
\begin{align*}
&p\left(\ensuremath{\boldsymbol{y}}_{(i-1)+d+1}^{i+d'}, d'= d - 1\big|d,s,s',\zeta',\zeta\right)\\
&= p\left(z' = {\tt{{Del}}}\big|k\text{mer}',z\right).
\end{align*}
\item $z' = {\tt{{Sub}}}$.
Then, $d' = d$, which means that $p(\ensuremath{\boldsymbol{y}}_{(i-1)+d+1}^{i+d'}, d' \neq d\big|d,s,s',\zeta',\zeta) = 0$, and
\begin{align*}
&p\left(\ensuremath{\boldsymbol{y}}_{(i-1)+d+1}^{i+d'}, d'= d\big|d,s,s',\zeta',\zeta\right)\\
&= \begin{cases}
p\left(z' = {\tt{{Sub}}}\big|k\text{mer}',z\right) \\\times\;p\left(\ensuremath{\boldsymbol{y}}_{i + d}^{i+d}| k\text{mer}', z' = {\tt{{Sub}}} \right)& \text{if } \ensuremath{\boldsymbol{y}}_{i + d}^{i+d} \neq \ensuremath{\boldsymbol{x}}_{i + d}^{i+d}\\
0& \text{otherwise}\,.
\end{cases}
\end{align*}
\item $z' = {\tt{NoErr}}$.
Then, $d' = d$, which means that $p(\ensuremath{\boldsymbol{y}}_{(i-1)+d+1}^{i+d'}, d' \neq d\big|d,s,s',\zeta',\zeta) = 0$, and
\begin{align*}
&p\left(\ensuremath{\boldsymbol{y}}_{(i-1)+d+1}^{i+d'}, d'= d\big|d,s,s',\zeta',\zeta\right)\\
&= \begin{cases}
p\left(z' = {\tt{NoErr}}\big|k\text{mer}',z\right)& \text{if } \ensuremath{\boldsymbol{y}}_{i + d}^{i+d} = \ensuremath{\boldsymbol{x}}_{i + d}^{i+d}\\
0& \text{otherwise}\,.
\end{cases}
\end{align*}
\end{enumerate}
\subsection{Decoding for Multiple Received Sequences}
\label{eq:separate}
We consider the separate decoding strategy for multiple received sequences $\ensuremath{\boldsymbol{y}}_1,\dots,\ensuremath{\boldsymbol{y}}_{\ensuremath{M}}$ proposed in \cite{Maa22}. Following \cite{Maa22}, the APP $p(w_i|\ensuremath{\boldsymbol{y}}_1,\dots,\ensuremath{\boldsymbol{y}}_{\ensuremath{M}})$ can be approximated as
\begin{align*}
p(w_i|\ensuremath{\boldsymbol{y}}_1,\dots,\ensuremath{\boldsymbol{y}}_{\ensuremath{M}}) \approxprop \frac{\prod_{j=1}^{\ensuremath{M}} p(w_i|\ensuremath{\boldsymbol{y}}_j)}{p(w_i)^{\ensuremath{M}-1}}\,,
\end{align*}
where $p(w_i|\ensuremath{\boldsymbol{y}}_j)$ is computed as outlined in Section~\ref{sec:single_seq}.
This decoder, although suboptimal, is efficient and practical for our scenario, as its complexity grows linearly with $\ensuremath{M}$ \cite{Maa22}.\footnote{Alternatively, one may decode all $M$ reads jointly using a single inner decoder \cite{Maa22}. However, the complexity of this decoder grows exponentially with the number of sequences $\ensuremath{M}$, and becomes infeasible for $\ensuremath{M} > 2$.}
\subsection{Decoding Complexity}
\label{sec:Complexity}
Since the overall decoding complexity is dominated by the combination of the inner code and the channel, we will disregard the complexity of the outer decoder in the complexity analysis. In order to limit the inner decoding complexity, we limit the drift $d_i$ to a fixed interval $[d_{\text{min}},d_{\text{max}}]$ and the number of insertions per symbol $L$ to $L_{\text{max}}$. The complexity of the BCJR algorithm on the joint trellis of the inner code and the channel is directly proportional to the number of trellis edges at each trellis section, which is upper bounded by $2^{\nu} 4^{k+1} \Delta 2^\delta$, where $\Delta=d_{\text{max}}-d_{\text{min}}+1$ is the number of drift states, $\delta = L_{\text{max}} + 2$ is the number of possible drift transitions, and $\nu$ is the number of binary memory elements of the convolutional encoder. Hence, the complexity of decoding a single block is $\ensuremath{N_\mathsf{o}^\mathsf{b}} 2^{\nu + \delta} 4^{k+1} \Delta$, and the complexity of separate decoding (for all $S$ blocks) becomes $C_{\text{sep}} = S \ensuremath{N_\mathsf{o}^\mathsf{b}} 2^{\nu + \delta} 4^{k+1} \Delta \ensuremath{M}$. Note that the complexity increases exponentially with the channel memory $k$.
\vspace{-0.375ex}
\section{Achievable Information Rates}\label{sec:air}
The MAP decoder for the nanopore channel (including the inner code) derived in Section~\ref{sec:innerdec-one} allows us to compute AIRs for this channel. In particular, we compute \emph{BCJR-once} rates \cite{Kavcic2003BinaryII,muller_capacity_2004,soriaga_determining_2007}, defined as the symbolwise mutual information between the input of the channel and the log-likelihood ratios (LLRs) produced by a symbolwise MAP (i.e., optimum) detector. For a given inner code, the BCJR-once rate, denoted by $R_{\text{BCJR-once}}$, is a rate achievable by an outer code that does not exploit possible correlations between the LLRs and when no iterations between the inner and outer decoder are performed.
The BCJR-once rate under separate decoding can be estimated as
$$R_{\text{BCJR-once}} \approx \ensuremath{R_\mathsf{i}} \log \ensuremath{q_\mathsf{o}} + \frac{\ensuremath{R_\mathsf{i}}}{\ensuremath{N_\mathsf{o}^\mathsf{b}} + m} \sum_{i=1}^{\ensuremath{N_\mathsf{o}^\mathsf{b}} + m} \log\frac{\ensuremath{\mathrm{e}}^{L^{\text{BCJR-sep}}_i(w_i)}}{\sum_{a \in \field{\ensuremath{q_\mathsf{o}}}}\ensuremath{\mathrm{e}}^{L^{\text{BCJR-sep}}_i(a)}}$$
by sampling an input sequence $\ensuremath{\boldsymbol{w}}$ and corresponding output sequence $\ensuremath{\boldsymbol{y}} = (\ensuremath{\boldsymbol{y}}_1,\dots,\ensuremath{\boldsymbol{y}}_\ensuremath{M})$ and computing the (mismatched) LLRs $L^{\text{BCJR-sep}}_i(a) = \sum_{j=1}^{\ensuremath{M}} \ln \frac{q(w_i=a|\ensuremath{\boldsymbol{y}}_j)}{q(w_i=0|\ensuremath{\boldsymbol{y}}_j)}$, $a \in \field{\ensuremath{q_\mathsf{o}}}$, where $q(w_i|\ensuremath{\boldsymbol{y}}_j)$ is a (mismatched) inner decoding metric.
BCJR-once rates can also be computed for the true DNA storage channel using a dataset of DNA traces by averaging over pairs of input and output sequences. In this case, given an inner code and a decoder that assumes that the channel has memory $k$, the BCJR-once rate is an AIR for the DNA storage channel of a \emph{mismatched} decoder where the inner decoder is matched to the memory-$k$ nanopore channel model.
For both cases, we assume separate decoding and an inner MAP decoder matched to the memory-$k$ nanopore channel model, i.e., using $q(w_i|\ensuremath{\boldsymbol{y}}_j) = p(w_i|\ensuremath{\boldsymbol{y}}_j)$ where $p(w_i|\ensuremath{\boldsymbol{y}}_j)$ is computed as described in Section~\ref{sec:single_seq}.
\begin{table}[t]
\centering
\caption{Optimized Protographs Found by DE for Different $\ensuremath{M}$} \label{tab:DE_protographs}
\vspace{-2ex}
\begin{tabular}{ccc}
\toprule
\ensuremath{M} & $R = \ensuremath{R_\mathsf{o}} \ensuremath{R_\mathsf{i}}$ & Protograph \\
\midrule
$1$ & $\nicefrac{8}{10} \cdot \nicefrac{3}{2} = \nicefrac{6}{5}$ & $\left(\begin{smallmatrix} 1 & 2 & 0 & 1 & 2 & 2 & 1 & 1 & 1 & 3 \\ 2 & 0 & 3 & 2 & 1 & 0 & 1 & 2 & 2 & 0\end{smallmatrix}\right)$ \\[.2cm]
$2$ & $\nicefrac{7}{8} \cdot \nicefrac{3}{2} = \nicefrac{21}{16}$ & $\left(\begin{smallmatrix} 2 & 2 & 3 & 3 & 2 & 3 & 3 & 3\end{smallmatrix}\right)$\\[.2cm]
$5$& $\nicefrac{14}{15} \cdot \nicefrac{3}{2} = \nicefrac{21}{15}$ &$\left(\begin{smallmatrix} 3 & 2 & 3 & 3 & 2 & 2 & 3 & 3 & 2 & 2 &3 &3 &3 &2 &3\end{smallmatrix}\right)$ \\ \bottomrule
\end{tabular}
\vspace{-3ex}
\end{table}
\section{Concatenated Coding Scheme Design} \label{sec:codes}
\subsection{Inner Code} \label{sec:inner_code}
For the inner code we use the $(1,1,2)_4$ convolutional code with generator polynomial $g = [5,7]_{\text{OCT}}$ and punctured in order to have a higher rate. In particular, we use the puncturing matrix $\bm P = \left( \begin{smallmatrix} 1 & 0 & 1 \\ 1 & 1 & 0\end{smallmatrix}\right)$, which gives an inner code rate of $\ensuremath{R_\mathsf{i}} = \nicefrac{3}{2}$ (in bits per DNA symbol). Moreover, we add a pseudo-random sequence to the output of the inner code.
\subsection{Outer Code}
We consider a protograph-based binary LDPC code as the outer code, which we optimize (also in terms of code rate) via density evolution (DE) using the algorithm proposed in \cite{Kavcic2003BinaryII}
for the memory-$5$ channel model for $M=1$, $2$, and $5$. Moreover, we set $\ensuremath{N} = 110$ as the dataset in \cite{Srinivasavaradhan2021TrellisBMA} contains input sequences with this length. The optimized protographs, when limiting the entries to at most $3$, are shown in Table~\ref{tab:DE_protographs}. The outer LDPC codes are constructed by lifting the protographs using circulants that are optimized using the progressive edge-growth algorithm~\cite{PEG}.
\section{Numerical Results and Discussion}\label{sec:simresults}
In this section, we give AIRs and frame error rate (FER) results for our designed optimized concatenated codes for the memory-$k$ nanopore channel model. We further compute AIRs and FER results for a real DNA channel using the experimental dataset in \cite{Srinivasavaradhan2021TrellisBMA}. The dataset consists of $269709$ output sequences taken from the output of an ONT MinION sequencer and also the corresponding input sequences (before synthesis). There are in total $10000$ input sequences of length $N = 110$.
\input{Figures/AIRs_dataset_vs_model.tex}
In all our simulations, we use $L_{\text{max}} = 2$ and $d_{\max} = -d_\mathrm{min} = 5 \sqrt{N \frac{\max\left(p_\mathsf{I}, p_\mathsf{D}\right)}{1-\max\left(p_\mathsf{I}, p_\mathsf{D}\right)}}$. Moreover, we use separate decoding as described in Section~\ref{eq:separate}. The AIRs are computed by averaging over $10000$ sequences of length $N=110$.
In Fig.~\ref{fig:BCJR-Once-Inner_DS_vs_Model}, we plot BCJR-once rates for the memory-$k$ nanopore channel model (see Section~\ref{sec:system}) with no inner code (dashed curves) for different values of $k$. For each $k$, the inner decoder is matched to the channel. Further, for each $k$, we estimated the transition probabilities of the memory-$k$ nanopore channel model as described in Section~\ref{sec:model_build} using the dataset in \cite{Srinivasavaradhan2021TrellisBMA}. We observe that the AIRs decrease with increasing $k$. This is expected, as increasing the memory $k$ makes the channel more complex.
In the figure, we also plot AIRs for the true DNA storage channel using the dataset in \cite{Srinivasavaradhan2021TrellisBMA} (solid curves). In this case, the curve for a given value of $k$ corresponds to the AIR for the true DNA storage channel of a mismatched decoder where the inner decoder is matched to the memory-$k$ nanopore channel model. We observe the opposite effect to the AIRs for the memory-$k$ nanopore channel, i.e., the AIRs increase with increasing $k$. Again, this behavior is expected: If the memory-$k$ nanopore channel model models well the DNA strorage channel (i.e., the assumption of a Markovian model is good), increasing $k$ makes the decoder better matched to the true DNA storage channel, hence the AIR increases. Equivalently, a low value of $k$ corresponds to a decoder that is more \emph{mismatched} with respect to the true DNA storage channel, resulting in a lower AIR. Our results hence support that this channel model is good. Note that the AIRs saturate for $k$ around $7$. Furthermore, interestingly, comparing the dashed and solid curves, we observe a sandwich effect, where the AIR curve for the memory-$k$ nanopore channel model (dashed curve with orange triangles) and the AIR for the true DNA storage channel with a mismatched decoder (matched to the memory-$7$ nanopore channel model) almost coincide. This indicates that increasing $k$ beyond $7$ does not bring any further gains in AIR for the true DNA storage channel.
Finally, in Fig.~\ref{fig:BCJR-Once-Inner_DS_vs_Model}, we also plot the BCJR-once rate for the true DNA storage channel of a decoder that assumes i.i.d. IDS errors (black curve with star markers), as most decoders in the literature. We observe that this results in a significant performance loss.
\input{Figures/AIRs_dataset_vs_model_CC.tex}
In Fig.~\ref{fig:BCJR-Once-Inner_DS_vs_Model_CC}, we plot AIRs for the memory-$k$ nanopore channel and the true DNA storage channel for the case where we use the convolutional code of Section~\ref{sec:inner_code} as a synchronization inner code for $k$ up to $5$. (The random sequence added to the output of the inner code can be used to match the transmitted coded sequences to the input sequences of the dataset. Hence, the dataset can be used for the true channel with an inner convolutional code as well.) A similar behavior of the AIRs to those of Fig.~\ref{fig:BCJR-Once-Inner_DS_vs_Model} is observed.
In the figure, we also plot the rate $R$ obtained via DE for the true DNA storage channel for the optimized protographs in Table~\ref{tab:DE_protographs} (green filled circles), showing that our coding scheme gives excellent performance at code rates close to the BCJR-once rates.
In the inset of Fig.~\ref{fig:BCJR-Once-Inner_DS_vs_Model_CC}, we plot the FER results of our designed optimized concatenated codes with the inner convolutional code of Section~\ref{sec:inner_code} and outer codes of length $\ensuremath{N_\mathsf{o}}=10000$ based on the protographs in Table~\ref{tab:DE_protographs}. The codes are simulated over both the true DNA storage channel with an inner decoder matched to the memory-$5$ (solid curves)
and the memory-$5$ nanopore channel model for a decoder matched to $k=5$ (green dashed curve with circles) and $k=3$ (red dashed-dotted curve with diamonds). The results are in agreement with the DE results.
The decoding complexity increases exponentially with $k$ (see Section~\ref{sec:Complexity}). Thus, the AIRs and FER results in Figs.~\ref{fig:BCJR-Once-Inner_DS_vs_Model} and \ref{fig:BCJR-Once-Inner_DS_vs_Model_CC} allow us to quantify the performance loss incurred by a decoder assuming a given memory $k$ and hence the trade-off between decoding complexity and error rate performance. We observe that considering memory $5$ incurs almost no loss in terms of AIR, indicating that $k=5$ is enough.
\vspace{-0.1ex}
\section{Conclusion}
\vspace{-0.5ex}
We derived the optimum MAP decoder for the memory-$k$ nanopore channel model. Based on the MAP decoder, we derived AIRs for the true DNA storage channel of a mismatched decoder that is matched to the memory-$k$ model and optimized coding schemes for this channel. We showed that, remarkably, the concatenated coding scheme in \cite{Maa22} (properly optimized) achieves excellent performance for the true DNA storage channel: Considering an optimal inner decoder for the memory-$k$ nanopore channel yields significantly higher AIRs\textemdash hence higher storage density\textemdash for the true DNA storage channel than a decoder that assumes i.i.d. IDS errors, as usually assumed in the literature.
\begin{comment}
\end{comment}
\ifCLASSOPTIONcaptionsoff
\newpage
\fi
\balance
|
\section{Introduction}
In mainstream macroeconometrics literature of structural vector autoregressive (SVAR) models, Gaussian behaviour of structural shocks is assumed either explicitly or implicitly, and the consequence is that only second order information can be exploited. This leads to the well-documented problem about the observational equivalence of any orthonormal rotation of structural disturbances. For overcoming this identification issue, several strategies have been proposed in the literature, and all of them resort to using external restrictions or information. The standard solutions involve to impose zero, linear or non-linear restrictions over the matrix of contemporaneous effects matrix (\citet{sims1980macroeconomics,blanchard1989dynamic,bernanke1998measuring,blanchard2002empirical,gali1992well}, among others). Alternative solutions are based on imposing agnostic restrictions over the impulse response functions, also called sign-restrictions (\citet{uhlig2005effects,canova2007price,arias2018inference}, and the references therein); constructing measures of unobserved shocks using extensive institutional information and use them to estimate their dynamic effects over some outcomes of interest, best known as \textit{narrative approach} (\citet{romer1989does,romer2010macroeconomic}); and employing such proxy measures for the latent structural shocks as instrumental variables, e.g. \citet{mertens2013dynamic,mertens2014reconciliation,stock2018identification}.
Despite the usefulness of such identification schemes, they suffer some drawbacks. First, in case of using external restrictions (zero, linear, non-linear or agnostic ones), these cannot be empirically tested unless the model is overidentified. And, in some cases, those restrictions may be quite arguable. Second, under instrumental or proxy variables approach, it is necessary to construct at least one instrument or proxy for each structural shock, i.e. the researcher can only identify shocks for which such measures are possible to find or construct. Besides, as it is documented by \citet{herwartz2014structural}, the identification power of instrumental variable approach depends on the relevance and exogeneity conditions.
On the other hand, statistical identification of SVAR or SVARMA models has appeared as an alternative to previous identification strategies. This approach, which is part of the \textit{data-driven identification strategies}, is supported by three main requirements: (i) structural shocks are an independent, identically distributed (i.i.d.) process; (ii) the structural error vector has mutually independent components; and (iii) each structural disturbance follows a non-Gaussian distribution. Based on these assumptions, the SVAR or SVARMA model is identify (up to signed permutation) (see \citet{chan2006note,lanne2010structural,lanne2017identification,gourieroux2017statistical,gourieroux2020identification,velasco2022identification} and the references therein). One of the main advantages of this identification scheme is that, since it is not based on external restrictions or exploit the existence of instrumental variables or proxy measures, makes any economically motivated external restriction to be subject of empirical assessment.
Additionally, this identification approach allows for possibly non-fundamental representations for the true structural model. This type of representations are ruled out in most literature for avoiding to deal with the dynamic identification problem.\footnote{There exist some efforts on using alternative methods for estimating impulse response functions when invertibility of the moving average component in the underlying SVARMA model does not hold. The solution is based on using instrumental variables. See \citet{stock2018identification,plagborg2021local,li2022local}.}\footnote{A fundamental SVARMA model has two implications: first, observable endogenous variables are linear combination of present and past values of structural shocks; and, second, structural errors can be recovered using linear combinations of current and past values of endogenous observable variables.}\footnote{Mainstream literature discards non-fundamental representation, otherwise, another identification problem arises. As it discussed in detail for SVAR and SVARMA models by \citet{chan2006note,velasco2022identification}, or its univaraite counterparts, \citet{rosenblatt2000gaussian}, if structural shocks were Gaussian distributed, both fundamental and non-fundamental representations are observational equivalent (\textit{dynamic identification problem}).} Also, under this approach, it is possible to perform prior empirical assessment of the identification conditions, e.g. for testing i.i.d assumption \citet{hong1999hypothesis} proposal can be applied. In this article we stress the attention on condition (iii), since even when (i) and (ii) holds, its failure implies the structural model is not identified.
There is a voluminous literature that tackles down the question regarding whether a random variable (vector) follows a Gaussian distribution or not. For instance, some procedures are based on the empirical distribution (\citet{kolmogorov1933sulla,smirnov1948table,massey1951kolmogorov,shapiro1965analysis}); other approaches employ the characteristic function (\citet{hall1983test,epps1983test}); or those proposals using third and fourth centered moments (\citet{d1971omnibus,bera1982model}). Nonetheless, all these procedures only assess the null hypothesis of joint Gaussianity; which in case of no rejection, leads to conclude that evidence supports the random vector is joint Gaussian distributed and, therefore, this would make any estimation procedure under the statistical identification approach misleading and/or inconsistent. Nevertheless, if the null hypothesis of joint Gaussianity is rejected, it can only be inferred that, provided the components of the vector are mutually independent, at least one element of the random vector is non-Gaussian distributed. Unfortunately, such conclusion is not sufficient for assuring identification under our statistical approach.
The literature for estimating the exact number of non-Gaussian (or Gaussian) components in a random vector is less voluminous, and all the proposals relates the number of non-Gaussian components to the number of eigenvalues different from zero or the rank of some matrix constructed from third or fourth order moments or cumulants. For instance, \citet{nordhausen2017asymptotic} proposes both an asymptotic and a bootstrap test to estimate the Gaussian dimension of an observable random vector, which is assumed to be an affine transformation of another unobservable vector.\footnote{Let $\mathbf{u}$ and $\boldsymbol{\varepsilon}$ be an observable and unobservable random vectors, respectively. \citeauthor{nordhausen2017asymptotic} consider $\mathbf{x} = \boldsymbol{b}_{0} + \boldsymbol{B}\boldsymbol{\varepsilon}$ (\textit{affine} transformation) with $\boldsymbol{b}_{0}$ possibly non-zero vector and $\boldsymbol{B}$ time-invariant, full rank, square matrix.} Their procedure is based on analyzing the scatter matrix\footnote{A scatter matrix is a matrix with the fourth order moments of a vector.} of the $d$-dimensional, observable random vector. They show that when the random vector contains $r$ Gaussian components, the scatter matrix has exactly $r$ eigenvalues equal to $0$.\footnote{To be completely precise, there exists $r$ eigenvalues equal to $d+2$, but \citet{nordhausen2017asymptotic} employ a normalized version of the scatter matrix.} Their estimation method is a sequential hypothesis test procedure. In the first step, they start by assuming that all the eigenvalues are equal to $0$, i.e. the vector is joint Gaussian distributed. If the null is rejected, then goes to the second step, in which the null hypothesis is updated and now it becomes $d-1$ eigenvalues are equal to zero, i.e. $1$ component of the random vector is non-Gaussian distributed while the remaining follows a joint Gaussian distribution. In case of rejection, the procedure continues up to consider a null hypothesis where the number of non-zero eigenvalues of the scatter matrix is $1$, i.e. only one Gaussian element in the random vector. If this were also rejected, the procedure ends and one concludes that all the components in the random vector are non-Gaussian distributed.
\citet{maxand2020identification} applies Nordhausen's method to estimate the number of non-Gaussian structural shocks of a fundamental SVAR model. When SVAR model is imposed to be fundamental, i.e. causal representation, the statistical identification only requires that at most one of the structural shocks may follow a Gaussian (\citet{comon1994independent}). Thus, it is enough to reject up to step $d-2$. Maxand's empirical strategy has two stages: in the first stage, she retrieves the estimated structural residuals from fitting an SVAR model; and, at the second step, apply Nordhausen's test to these estimated structural residuals. It must be noted that under \citeauthor{maxand2020identification} approach, statistical identification is not guaranteed, but she proves it is not necessary to have statistical identification for performing the test, because when the structural shocks vector has more than one Gaussian distributed component, the non-Gaussian block of structural shocks remains identified up to signed-permutation.
The main flaw under \citeauthor{maxand2020identification}'s approach is, precisely, the need for estimating structural shocks of a partial identified model. Hence, \citet{guay2021identification} employs reduced-form errors. His proposal is a bootstrap test that uses \citet{robin2000tests} statistic. \citeauthor{guay2021identification} shows that the rank of the rectangular array constructed using third and fourth order cumulants of reduced-form errors is equivalent to the number of structural shocks that are asymmetric and/or non-mesokurtic.\footnote{A non-mesokurtic distribution has a kurtosis coefficient different from $3$, or has a non-zero excess of kurtosis (fourth order cumulant).} Unlike \citeauthor{maxand2020identification}'s approach, none identification (point or partial) over structural shocks is needed. A technical drawback in \citet{guay2021identification} it is that the asymptotic equivalence between the array of third and/or fourth order moments constructed using unobserved reduced-form errors and the one based on sample residuals is taken for granted.
The former approaches are quite useful, but also restrictive by some reasons. First, many structural macroeconomic models can be represented and fitted more accurately by a VARMA model rather than a VAR one. Second, as it is discussed in \citet{velasco2022identification} and the references therein, the fundamentalness assumption is a restriction to avoid the dynamic identification problem. Therefore, under a more general scenario of possibly non-fundamental representation of structural VARMA models, then \citet{maxand2020identification} or \citet{guay2021identification} strategies would be invalid since, first, it is unknown if a possibly non-fundamental SVARMA model is partially identified if at least one structural shock is Gaussian distributed; and second, they are designed for a situation where reduced-form errors are a simply rotation (linear combination) of structural shocks. Nonetheless, if reduced-form errors were a possibly infinite, two-sided, linear combination of structural shocks, the rank of an array constructed from contemporaneous third and fourth order cumulants of reduced-form errors would not identify the number of structural shocks, unless strong unrealistic assumptions are made.
The aim of this paper is to fill this gap in the literature. Like \citet{guay2021identification}, we use reduced-form errors. But, since it is allowed that true data generating process may be a non-fundamental SVARMA model, reduced-form errors are taken from a fundamental representation, whose identification is guaranteed by Wold Decomposition Theorem. Besides, reduced-form errors from fundamental representation, instead of being a rotation (linear combination) of current values of unobserved structural shocks, can be a filtered version of such disturbances. This implies that reduced-form errors\footnote{Along this article we use the term \textit{reduced-form errors} as equivalent to \textit{reduced-form errors from a fundamental representation}.}, although serially uncorrelated, may exhibit high-order dependence. As usual, we characterize non-Gaussian distributions by its asymmetry or excess of kurtosis, thus why we focus on third and fourth order dependence. For assessing these high-order dependence we follow \citet{velasco2022identification} and employ the third and fourth order cumulant spectrum of the reduced-form errors. Based on this, we construct an array whose rank is equivalent to the number of skewed (asymmetric) and/or non-mesokurtic structural shocks.
Another important difference with respect to previous works is that the asymptotic distribution of our test statistic changes depending on the stage of the sequential procedure, specifically for the null hypothesis of joint Gaussianity of structural disturbances, implying that higher-order cumulant spectrums are zero and making infeasible to use standard Delta method for finding the asymptotic distribution of the test statistic. Although we are able to find the asymptotic distribution of the statistics, its variance is difficult to estimate accurately, specially at finite samples. Consequently, we employ a bootstrap procedure for performing the testing sequence.
The remaining of the article is structured as follows: Section 2 describes our time series model and states the main assumptions. Section 3 shows the connection between the number of non-Gaussian structural shocks and the rank of an array constructed from third and/or fourth order spectrums of reduced-form errors. In section 4, the test procedure and estimation are detailed. Section 5 shows simulation results and the empirical application. We conclude at Section 6.
\section{Model and Assumptions}
\label{sec:Basics}
Let $\boldsymbol{y}_{t}$ be a $d$-dimensional stationary, zero-mean, vector-valued random variable that solves equation (\ref{eq:Eq1}), which is best-known as \textit{structural VARMA model} (\cite{gourieroux2020identification}):
\begin{equation}
\label{eq:Eq1}
\boldsymbol{\Phi}(L)\boldsymbol{y}_{t} = \boldsymbol{\Theta}(L)\boldsymbol{B}\boldsymbol{\varepsilon}_{t},
\end{equation}
where $\boldsymbol{\Phi}(L)=\Id_{d}-\sum_{l=1}^{p}{\boldsymbol{\Phi}_{l}L^{l}}$, $\boldsymbol{\Theta}(L)=\Id_{d}+\sum_{l=1}^{q}{\boldsymbol{\Theta}_{l}L^{l}}$. $L$ is the lag (back-shift) operator, i.e. $L^{k}\boldsymbol{y}_{t}=\boldsymbol{y}_{t-k}$. $\boldsymbol{\varepsilon}_{t}$ is a $d$-dimensional vector of structural shocks. $\boldsymbol{B}$ is a time-invariant, full-rank, squared matrix. When $q=0$, we deal with the so-known SVAR model. Besides, we impose that $\boldsymbol{\Phi}(z)$ and $\boldsymbol{\Theta}(z)$ polynomials satisfy
\begin{align}
\label{eq:Eq2}
\det\left(\boldsymbol{\Phi}(z)\right)\det\left(\boldsymbol{\Theta}(z)\right) & \neq{ 0},\quad \forall\; z\in{\mathbb{T}},
\end{align}
where $\mathbb{T}=\left\{ z\in\mathbb{C}\;|\;|z|=1 \right\}$ is the complex unit circumference\footnote{In most textbooks or articles, it is known as complex unit \emph{circle}.} of radius $1$. Intuitively, equation (\ref{eq:Eq2}) only rules out the presence of unit roots in both auto-regressive (AR) or moving-average (MA) polynomials, although it does not impose the location of polynomial roots\footnote{Compares this to alternative works such as \citet{maxand2020identification,guay2021identification}, where the assumption rules out roots on and inside complex unit circle.}, which implies it is allowed for possibly non-fundamental SVARMA models. Regarding $\boldsymbol{\varepsilon}_{t}$, we assume the following structure:
\begin{assumption}\mbox{}
\label{as:Assumption1}
\begin{enumerate}
\item[(i)] $\{\boldsymbol{\varepsilon}_{t}\}_{t\in\mathbb{Z}}$ is an independent, identically distributed (i.i.d.) sequence.
\item[(ii)] The components of $\boldsymbol{\varepsilon}_{t}$ are mutually independent for each $t$.
\item[(iii)] The number of non-Gaussian distributed shocks in $\boldsymbol{\varepsilon}_{t}$ is denoted by $d_{ng}$, with $d_{ng}\in\{0, 1, \dots, d\}$.
\item[(iv)] $\E[\boldsymbol{\varepsilon}_{t}]=0$, $\E[\boldsymbol{\varepsilon}_{t}\boldsymbol{\varepsilon}_{t}^{\prime}]=\Id_{d}$ and $\E\left\|\boldsymbol{\varepsilon}_{t}\right\|^{8}<\infty$ for all $t\in\mathbb{Z}$.
\end{enumerate}
\end{assumption}
Assumption (\ref{as:Assumption1}) it is not sufficient for identifying the structural model described by Equation (\ref{eq:Eq2}) when AR and MA polynomials satisfy condition (\ref{eq:Eq2}), unless $d_{ng}=d$. In case $q=0$ and condition (\ref{eq:Eq2}) holds not only for $\mathbb{T}$, but for $\{ z\in\mathbb{C}\;|\;|z|\leq1 \}$, i.e. when we deal with a fundamental (also called \textit{stable} or \textit{causal}) SVAR model, statistical identification is achieved if $d_{ng}\geq{d-1}$. There is no result of partial identification when $d_{ng}<d$ and location of polynomials roots is unknown.
Assumption (\ref{as:Assumption1}.i) is an standard requirement (see \citet{lanne2010structural,lanne2017identification,gourieroux2020identification,velasco2022identification,guay2021identification}), and implies that structural errors are not only serially uncorrelated, but do not exhibit any higher-order dependence. Assumption (\ref{as:Assumption1}.ii) is the basis of independent component analysis (\citet{comon1994independent}). This is not as restrictive as it may seem, e.g. in macroeconomics one may think that productivity shock is independent from monetary or fiscal policy shocks. Last condition of Assumption (\ref{as:Assumption1}.iv) is required because we are characterizing non-Gaussianity through asymmetry and/or excess of kurtosis.
\begin{assumption}
\label{as:Assumption2}
Let $\boldsymbol{\vartheta}_{0}\in\mathcal{V}\subset\mathbb{R}^{K}$ be a parameter vector, and $f:\mathcal{V}\mapsto[\mathbb{R}]^{d\times{p}}\times[\mathbb{R}]^{d\times{q}}\times\mathbb{M}^{d\times{d}}$ some mapping, where $[\mathbb{R}]^{d\times{p}}$ is the space of matrix polynomials of size $d$ and degree $p$, and $\mathbb{M}^{d\times{d}}$ the space of full rank, square matrices of dimension $d$. We assume the following:
\begin{enumerate}
\item[(i)] $\boldsymbol{\Phi}(L,\boldsymbol{\vartheta}_{0})$ and $\boldsymbol{\Theta}(L,\boldsymbol{\vartheta}_{0})$ are lag polynomials satisfying condition (\ref{eq:Eq2}), and $\boldsymbol{B}(\boldsymbol{\vartheta}_{0})$ is a time-invariant, full rank matrix. They generates $\boldsymbol{y}_{t}$, according to equation (\ref{eq:Eq1}).
\end{enumerate}
\end{assumption}
Assumption (\ref{as:Assumption2}) states the existence of a \emph{true} value for parameter vector, denoted by $\boldsymbol{\vartheta}_{0}$ which generates the stationary process $\{\boldsymbol{y}_{t}\}$. In other words, $\boldsymbol{y}_{t}$ is the solution of $\boldsymbol{\Phi}(L,\boldsymbol{\vartheta}_{0})\boldsymbol{y}_{t} = \boldsymbol{\Theta}(L,\boldsymbol{\vartheta}_{0})\boldsymbol{B}(\boldsymbol{\vartheta}_{0})\boldsymbol{\varepsilon}_{t}$. This is equivalent to state that model residuals using $\boldsymbol{\vartheta}_{0}$ are equal to the structural shocks vector, i.e $\boldsymbol{e}_{t}(\boldsymbol{\vartheta}_{0})=\boldsymbol{\varepsilon}_{t}$ with $\boldsymbol{e}_{t}(\boldsymbol{\vartheta}_{0}) = \boldsymbol{B}^{-1}(\boldsymbol{\vartheta}_{0})\boldsymbol{\Theta}^{-1}(L,\boldsymbol{\vartheta}_{0})\boldsymbol{\Phi}(L,\boldsymbol{\vartheta}_{0})\boldsymbol{y}_{t} = \boldsymbol{B}^{-1}(\boldsymbol{\vartheta}_{0})\boldsymbol{\Psi}^{-1}(L,\boldsymbol{\vartheta}_{0})\boldsymbol{y}_{t}$ and $\boldsymbol{\Psi}(L,\boldsymbol{\vartheta}_{0})=\boldsymbol{\Phi}^{-1}(L,\boldsymbol{\vartheta}_{0})\boldsymbol{\Theta}(L,\boldsymbol{\vartheta}_{0})$.
Nonetheless, since the aim of this work is not discussing identification of the possibly non-fundamental SVARMA model, and given that Wold's Decomposition Theorem\footnote{This states that any stationary random process can be represented as an infinite moving-average of present and past values of white noise residuals.} (WDT) guarantees the existence of a fundamental representation for any stationary process, we focus the analysis over reduced-form residuals from such representation. Following the previous notation, let $\boldsymbol{\vartheta}_{f}$ the parameter vector associated to the fundamental representation of $\boldsymbol{y}_{t}$. Then, reduced-form residuals, $\boldsymbol{u}_{t}(\boldsymbol{\vartheta}_{f})$, are defined as
\begin{equation}
\label{eq:ReducedFormErrors}
\boldsymbol{u}_{t}=\boldsymbol{u}_{t}(\boldsymbol{\vartheta}_{f})=\boldsymbol{\Psi}^{-1}(L,\boldsymbol{\vartheta}_{f})\boldsymbol{y}_{t}\;,
\end{equation}
where $\boldsymbol{\Psi}(L,\boldsymbol{\vartheta}_{f})=\boldsymbol{\Phi}^{-1}(L,\boldsymbol{\vartheta}_{f})\boldsymbol{\Theta}(L,\boldsymbol{\vartheta}_{f})$ and $\boldsymbol{\Phi}(L,\boldsymbol{\vartheta}_{f})$ and $\boldsymbol{\Theta}(L,\boldsymbol{\vartheta}_{f})$ representing stable (or causal) AR and invertible MA polynomials, respectively. WDT states the existence of white-noise process, $\boldsymbol{u}_{t}$; however, with the exception of a fundamental VARMA model, $\boldsymbol{u}_{t}$ might not be an i.i.d. process.
Using equation (\ref{eq:ReducedFormErrors}) and $\boldsymbol{y}_{t}= \boldsymbol{\Psi}(L,\boldsymbol{\vartheta}_{0})\boldsymbol{B}(\boldsymbol{\vartheta}_{0})\boldsymbol{\varepsilon}_{t}$, then
\begin{equation}
\label{eq:Reduced_StructuralErrors}
\boldsymbol{u}_{t} = \boldsymbol{\Psi}^{-1}(L,\boldsymbol{\vartheta}_{f})\boldsymbol{\Psi}(L,\boldsymbol{\vartheta}_{0})\boldsymbol{B}(\boldsymbol{\vartheta}_{0})\boldsymbol{\varepsilon}_{t} = \boldsymbol{\delta}(L,\boldsymbol{\vartheta}_{f})\boldsymbol{\varepsilon}_{t}
\end{equation}
where $\boldsymbol{\delta}(L,\boldsymbol{\vartheta}_{f}) = \boldsymbol{\Psi}^{-1}(L,\boldsymbol{\vartheta}_{f})\boldsymbol{\Psi}(L,\boldsymbol{\vartheta}_{0})\boldsymbol{B}(\boldsymbol{\vartheta}_{0})=\sum_{j=-\infty}^{\infty}{\boldsymbol{\delta}_{j}(\boldsymbol{\vartheta}_{f})L^{j}}$. That is, reduced-form residuals ($\boldsymbol{u}_{t}$), obtained from adjusting a fundamental VARMA model to $\boldsymbol{y}_{t}$, are a possibly infinite, two-sided filter of true structural shocks, $\boldsymbol{\varepsilon}_{t}$. WDT implies that
\begin{equation*}
\boldsymbol{\Gamma}_{k} = \Cov\left( \boldsymbol{u}_{t}, \boldsymbol{u}_{t+k} \right) = \sum_{j=-\infty}^{\infty}{\boldsymbol{\delta}_{j}(\boldsymbol{\vartheta}_{f})\boldsymbol{\delta}_{j+k}^{\prime}(\boldsymbol{\vartheta}_{f})} = \boldsymbol{0},\quad\forall\;t\in\mathbb{Z},\;\abs{k}>0
\end{equation*}
Since true VARMA model and its fundamental approximation are full rank on the complex unit circle, the previous condition is equivalent to
\begin{equation}
\label{eq:All-Pass}
\boldsymbol{\delta}(e^{i\lambda},\boldsymbol{\vartheta}_{f})\boldsymbol{\delta}^{*}(e^{i\lambda},\boldsymbol{\vartheta}_{f}) = \boldsymbol{\Omega} = \boldsymbol{B}(\boldsymbol{\vartheta}_{0})\boldsymbol{B}^{\prime}(\boldsymbol{\vartheta}_{0})
\end{equation}
where $\boldsymbol{\Omega}$ is positive definite, symmetric constant matrix and $\boldsymbol{\delta}^{*}(e^{i\lambda},\boldsymbol{\vartheta}_{f})$ denotes the transposed, conjugate matrix polynomial of $\boldsymbol{\delta}(e^{i\lambda},\boldsymbol{\vartheta}_{f})$. \citet{velasco2022identification}, and the references therein, denominates filters satisfying equation (\ref{eq:All-Pass}) as an \textit{all-pass filter}.\footnote{To be completely precise, an \emph{all-pass} filter generates $\boldsymbol{\Omega}=\boldsymbol{I}_{d}$. However, given that we are working with only the reduced-form residuals, these errors can exhibit contemporaneous cross-correlations.}
\section{Higher-order Spectral Cumulant Density and Number of non-Gaussian components}
\label{sec:HOSA_CumRandVec}
\subsection{Cumulants of Random Vector}
\label{sec:CumRandVector}
The characteristic function of structural shocks, $\boldsymbol{\varepsilon}_{t}$, can be defined as $\phi_{\boldsymbol{\varepsilon}}(\boldsymbol{\tau})=\mathbb{E}(\exp{(i\boldsymbol{\tau}^{\prime}\boldsymbol{\varepsilon})})$ where $\boldsymbol{\tau}=(\tau_{1},\dots,\tau_{d})^{\prime}$. And, the cumulant generating function is $\boldsymbol{\kappa}_{\boldsymbol{\varepsilon}}(\boldsymbol{\tau})=\log\left(\phi_{\boldsymbol{\varepsilon}}(\boldsymbol{\tau})\right)$.
The (cross) cumulant of order $k$ is
\begin{equation}
\label{eq:Cumulant_sOrder}
\kappa_{k}^{\boldsymbol{\varepsilon}} = \left. \frac{\partial^{k}}{(\partial{\tau_{1}})^{k_{1}}\dots(\partial{\tau_{d}})^{k_{d}}}{\boldsymbol{\kappa}_{\boldsymbol{\varepsilon}}(\boldsymbol{\tau})} \right|_{\boldsymbol{\tau}=\boldsymbol{0}}
\end{equation}
where $\sum_{j=1}^{d}{k_{j}}=k\geq1$ and $k_{j}\geq{0}$. If $k_{j}=0$ for all $j$ except one $l\in\{1,\dots,d\}$, we say it is the $l$-th marginal cumulant of order $k$. The number of different cumulants of order $k$ is given by $\binom{d+k-1}{k}=\frac{(d+k-1)!}{k!(d-1)!}$. \citet{jammalamadaka2006higher} show that for a zero-mean, random-vector $\boldsymbol{\varepsilon}$ it holds
\vspace*{-0.5cm}
\begin{align}
\label{eq:CumK_Vector}
\boldsymbol{\kappa}_{1}(\boldsymbol{\varepsilon})&=\E[\boldsymbol{\varepsilon}]=\boldsymbol{0} \nonumber \\
\boldsymbol{\kappa}_{2}(\boldsymbol{\varepsilon})&=\E[\boldsymbol{\varepsilon}^{\otimes{2}}] = \text{vec}( \E[\boldsymbol{\varepsilon}_{t}\boldsymbol{\varepsilon}^{\prime}_{t}] ) \nonumber \\
\boldsymbol{\kappa}_{3}(\boldsymbol{\varepsilon})&=\E[\boldsymbol{\varepsilon}^{\otimes{3}}] \nonumber \\
\boldsymbol{\kappa}_{4}(\boldsymbol{\varepsilon})&=\E[\boldsymbol{\varepsilon}^{\otimes{4}}]-\left(\mathbf{I}_{d^{4}}+\mathbf{K}^{-1}_{\mathfrak{p}_{2\leftrightarrow{3}}}(d_{[4]})+\mathbf{K}^{-1}_{\mathfrak{p}_{4\leftrightarrow{2}}}(d_{[4]})\right)\left[\boldsymbol{\kappa}_{2}(\boldsymbol{\varepsilon})\otimes\boldsymbol{\kappa}_{2}(\boldsymbol{\varepsilon})\right]
\end{align}
where $\mathbf{K}_{\mathfrak{p}_{j\leftrightarrow{k}}}(d_{[4]})$ is a commutation matrix of dimension $d^{4}$ that interchanges the elements $j$ and $k$ (of a permutation $\mathfrak{p}$ of the indexes $(1,2,...,d)$) of a Kronecker product; and $X^{\otimes{j}}=\underbrace{X\otimes{{\dots}\otimes{X}}}_{j\text{ times}}$. In Online Appendix, we show more details about commutation matrices and a particular example for dimension $d=2$.
Since structural shocks vector, $\boldsymbol{\varepsilon}_{t}$, satisfies the mutual independence assumption, then all its non-marginal cumulants are zero. Let $\kappa_{k,j}$ denotes the $k$-th order, $j$-th marginal cumulant, then
\begin{equation*}
\boldsymbol{\kappa}_{k}(\boldsymbol{\varepsilon}) = \vect\left(\mathbf{v}_{k}\left({\boldsymbol{\varepsilon}}\right)\right),
\end{equation*}
where $\mathbf{v}_{k}\left({\boldsymbol{\varepsilon}}\right) = \begin{bmatrix}
\kappa_{k,1}\boldsymbol{e}_{1}^{\otimes{k-1}} & \kappa_{k,2}\boldsymbol{e}_{2}^{\otimes{k-1}} & \dots & \kappa_{k,d}\boldsymbol{e}_{d}^{\otimes{k-1}}
\end{bmatrix} $ and $\boldsymbol{e}_{j}$ is the $j$-th canonical vector of dimension $d$. In this paper, as we explained above, we employ $k\in{3,4}$, the asymmetry and excess of kurtosis coefficients.
\subsection{Higher-order Cumulant Spectral Density of Reduced-form Errors}
Following \citet{brillinger2001time,velasco2022identification}, the $k$-th order cumulant spectrum of reduced-form errors is
\begin{align}
\label{eq:EqHOSA_Spectrum}
g_{k}(\boldsymbol{\lambda}) &= (2\pi)^{1-k}\left(\boldsymbol{\delta}\left(e^{i\sum_{m=1}^{k-1}{\lambda}_{m}};\boldsymbol{\vartheta}_{f}\right)\otimes \bigotimes_{j=1}^{k-1}\left\{ \boldsymbol{\delta}\left(e^{-i\lambda_{k-j}};\boldsymbol{\vartheta}_{f}\right) \right\}\right)\boldsymbol{\kappa}_{k}\left(\boldsymbol{\varepsilon}\right)
\end{align}
where $\bigotimes$ denotes a cumulative Kronecker product.
Let $G_{k}(\boldsymbol{\lambda})$ be a matrix of dimension $d^{k-1}\times{d}$, such that $g_{k}(\boldsymbol{\lambda})=\vect\left( G_{k}(\boldsymbol{\lambda}) \right)$. Thus, we can define $G_{k}(\boldsymbol{\lambda})$ as
\vspace*{-0.5cm}
\begin{align}
\label{eq:EqHOSA_Spectrum_Matrix}
G_{k}(\boldsymbol{\lambda}) &= \left( \vect(I_{d})^{\prime}\otimes I_{d^{k-1}} \right)\left( I_{d}\otimes{g_{k}(\boldsymbol{\lambda})} \right) \\
{} &= \left( \bigotimes_{j=1}^{k-1}\left\{ \boldsymbol{\delta}\left(e^{-i\lambda_{k-j}};\boldsymbol{\vartheta}_{f}\right) \right\} \right)\mathbf{v}_{k}^{\boldsymbol{\varepsilon}}\boldsymbol{\delta}^{\prime}\left(e^{i\sum_{m=1}^{k-1}{{\lambda}_{m}}} ;\boldsymbol{\vartheta}_{f}\right) \nonumber
\end{align}
Using $G_{k}(\lambda_{k})$, equation (\ref{eq:All-Pass}) and $(A\otimes{B})(C\otimes{D})=\left(AC\otimes{BD}\right)$, provided that $AC$ and $BD$ are conformable, we obtain $G_{k}^{2}(\lambda_{k})$
\begin{align}
\label{eq:GMat_Sq}
\hspace{-0.5cm}
G_{k}^{2}(\lambda_{k})=\left[G_{k}(\boldsymbol{\lambda})\right]^{*}G_{k}(\boldsymbol{\lambda}) &= \boldsymbol{\delta}\left(e^{-i\sum_{m=1}^{k-1}{{\lambda}_{m}} };\boldsymbol{\vartheta}_{f}\right)\left(\mathbf{v}_{k}^{\prime\boldsymbol{\varepsilon}}\boldsymbol{\Omega}^{\otimes{k-1}}\mathbf{v}_{k}^{\boldsymbol{\varepsilon}}\right)\boldsymbol{\delta}^{\prime}\left(e^{i\sum_{m=1}^{k-1}{{\lambda}_{m}}};\boldsymbol{\vartheta}_{f}\right)
\end{align}
Notice that $G_{k}^{2}(\lambda_{k})$ only depends on the scalar $\lambda_{k}=\sum_{m=1}^{k-1}{{\lambda}_{m}}$. Besides, $\mathbf{v}_{k}^{\prime\boldsymbol{\varepsilon}}\boldsymbol{\Omega}^{\otimes{k-1}}\mathbf{v}_{k}^{\boldsymbol{\varepsilon}}$ is positive semi-definite, symmetric matrix. When $d_{ng}=d$, i.e. all components in $\boldsymbol{\varepsilon}_{t}$ are non-Gaussian distributed, then $\mathbf{v}_{k}^{\prime\boldsymbol{\varepsilon}}\boldsymbol{\Omega}^{\otimes{k-1}}\mathbf{v}_{k}^{\boldsymbol{\varepsilon}}$ is full-rank and positive definite. When $\lambda_{k}=0$, $G_{k}^{2}(\lambda_{k})$ is real-valued. Besides, it is straightforward to notice that $\text{Re}\left( G_{k}^{2}(\lambda_{k}) \right)$ is the sum of two full-rank, semi-definite quadratic forms.
\subsection{Number of non-Gaussian components}
Joint multivariate Normality tests for some random vector $\boldsymbol{u}$ based on moments use the combined measure of total skewness, defined as $\left\| \boldsymbol{\kappa}_{3}\left( \boldsymbol{u}\right) \right\|^{2}$, and the total kurtosis, $\text{Tr}\left( \E[\boldsymbol{u}^{\prime}\boldsymbol{\Sigma}_{u}^{-1}\boldsymbol{u}]^{2} \right)$. For standardized data, the null hypothesis of joint Gaussianity, i.e all the components in the vector $\boldsymbol{u}$ are Gaussian, implies that the combined measures are equal to $d(d+2)$, where $d$ is the vector dimension (see \citet{jammalamadaka2006higher,mardia1970measures}). When the null hypothesis is rejected, it only implies there may exist at least one non-Gaussian component in vector $\boldsymbol{u}$. This is not sufficient for our purposes.
The general idea is to relate the number of non-Gaussian components (or the non-Gaussian dimension) to the rank (number of non-zero eigenvalues) for some array constructed from higher-order moments (cumulants). \citet{nordhausen2017asymptotic} employ the scatter matrix $\boldsymbol{S}_{2}=\E[r^{2}\boldsymbol{u}_{t}\boldsymbol{u}_{t}^{\prime}]$ for determining the non-Gaussian dimension, where $r^{2}=\boldsymbol{u}_{t}^{\prime}\boldsymbol{S}_{1}^{-1}\boldsymbol{u}_{t}$ and $\boldsymbol{S}_{1}$ is the variance-covariance matrix of random vector $\boldsymbol{u}_{t}$. They showed that the matrix $\boldsymbol{R}=S_{1}^{-1/2}S_{2}S_{1}^{-1/2}$ has exactly $d_{g}$ eigenvalues equal to $d+2$, with $d_{g}$ the number of Gaussian components in the vector $\boldsymbol{u}_{t}$. This approach is is only valid if $\boldsymbol{u}_{t}$ is a time-invariant linear combination (rotation) of an i.i.d random vector with mutually independent components.\footnote{This implies that \citeauthor{nordhausen2017asymptotic} consider $\boldsymbol{u}_{t}=\boldsymbol{B}\boldsymbol{\varepsilon}_{t}$.} Therefore, the procedure would be invalid in case of $\boldsymbol{u}_{t}$ being an \textit{all-pass filter} of i.i.d. structural shocks, $\boldsymbol{\varepsilon}_{t}$. On the other hand, \citet{guay2021identification} uses an array constructed from contemporaneous third and fourth order cumulants of reduced-form errors, $\boldsymbol{u}_{t}$. He is able to prove that the rank of such array is equal to the number of asymmetric and non-mesokurtic components in the vector of structural shocks.
\paragraph{Why do not use contemporaneous third and/or fourth order cumulants?}\mbox{}\newline
Both \citet{nordhausen2017asymptotic} or \citet{guay2021identification} approaches involve (explicit or implicitly) to compute $\boldsymbol{\kappa}_{k}(\boldsymbol{u})$ for $k\in\{3,4\}$, which we call the contemporaneous $k$-th order cumulants of reduced-form errors, $\boldsymbol{u}_{t}$. In our set-up, $\boldsymbol{\kappa}_{k}(\boldsymbol{u})$ are
\begin{align*}
\boldsymbol{\kappa}_{k}(\boldsymbol{u}) &= \sum_{j=-\infty}^{\infty}{\boldsymbol{\delta}_{j}^{\otimes{k}}\boldsymbol{\kappa}_{k}(\boldsymbol{\varepsilon})}
\end{align*}
hence, the rectangular array $\mathbf{v}_{k}(\boldsymbol{u})$ is equivalent to
\begin{align*}
\mathbf{v}_{k}(\boldsymbol{u}) &= \sum_{j=-\infty}^{\infty}{\boldsymbol{\delta}_{j}^{\otimes{k-1}}\mathbf{v}_{k}\left({\boldsymbol{\varepsilon}}\right)\boldsymbol{\delta}_{j}^{\prime} }
\end{align*}
where $\mathbf{v}_{k}\left({\boldsymbol{u}}\right)=\vect\left( \boldsymbol{\kappa}_{k}(\boldsymbol{u}) \right)$.
Notice that $\rank\left(\mathbf{v}_{k}\left({\boldsymbol{u}}\right)\right)\neq\rank\left(\mathbf{v}_{k}\left({\boldsymbol{\varepsilon}}\right)\right)$\footnote{In general, the sum of matrices do not preserve the rank. For instance, let $A=\boldsymbol{I}_{d}$ and $B=-\boldsymbol{I}_{d}$ both full-rank matrices, but $C=A+B=\boldsymbol{0}$ a null-rank matrix.} unless $\mathbf{v}_{k}\left({\boldsymbol{\varepsilon}}\right)$ is a null matrix for $k=3,4$, i.e. the case of joint Gaussian structural shocks, or $\boldsymbol{\delta}_{j}=\boldsymbol{0}$ for all $\abs{j}>0$, i.e. the case of fundamental SVARMA.
\paragraph{Number of non-Gaussian structural shocks and the rank of $G^{2}_{k}(\lambda_{k})$}\mbox{}\newline
Notice that $\rank\left(\mathbf{v}_{k}({\boldsymbol{\varepsilon}})\right)=\rank\left(\mathbf{v}_{k}^{\prime}({\boldsymbol{\varepsilon}})\mathbf{v}_{k}({\boldsymbol{\varepsilon}})\right)$ is equal to the number of structural shocks with non-zero $k$-th order cumulant. We focus on $k=3,4$, which implies we are characterizing non-Gaussian behavior by the asymmetry and excess of kurtosis. From (\ref{eq:GMat_Sq}) and since $\boldsymbol{\Omega}$ is full-rank, positive definite matrix, $\rank\left(\mathbf{v}_{k}^{\prime}(\boldsymbol{\varepsilon})\boldsymbol{\Omega}^{\otimes{k-1}}\mathbf{v}_{k}({\boldsymbol{\varepsilon}})\right)=\rank\left(\mathbf{v}_{k}({\boldsymbol{\varepsilon}})\right)$. The following proposition summarizes our main finding in this paper.
\begin{proposition}\mbox{}\newline
\label{prop:Proposition1}
Let a SVARMA model described by equations (\ref{eq:Eq1}) and (\ref{eq:Eq2}), with structural shocks satisfying Assumption (\ref{as:Assumption1}). Consider the arrays $G_{3}^{2}(\lambda_{3})$, $G_{4}^{2}(\lambda_{4})$ and $G_{34}^{2}(\lambda_{3},\lambda_{4}) = \begin{bmatrix} G_{3}^{2,\prime}(\lambda_{3}), G_{4}^{2,\prime}(\lambda_{4}) \end{bmatrix}^{\prime}$, therefore
\begin{align*}
\rank\left( G_{3}^{2}(\lambda_{3}) \right) &= \rank\left(\mathbf{v}_{3}({\boldsymbol{\varepsilon}})\right) = d_{3},\quad \forall \;\; \lambda_{3}\in[-\pi,\pi] \\
\rank\left( G_{4}^{2}(\lambda_{4}) \right) &= \rank\left(\mathbf{v}_{4}({\boldsymbol{\varepsilon}})\right) = d_{4},\quad \forall \;\;\lambda_{4}\in[-\pi,\pi] \\
\rank\left(G_{34}^{2}(\lambda_{3},\lambda_{4})\right) &= d_{ng}
\end{align*}
\end{proposition}
where $d_{3}$ and $d_{4}$ is the number of skewed and non-mesokurtic structural shocks, respectively; and $d_{ng}$ the number of non-Gaussian structural shocks.
This proposition can be interpreted in the following way: when we have a possibly non-fundamental SVARMA model, the square of the matrix version of $3$-rd and/or $4$-th order cumulant spectral densities of the reduced-form residuals reveals the number of skewed and/or non-mesokurtic structural shocks, respectively.
\begin{corollary}\mbox{}\newline
\label{coro:Corollary1}
Let a SVARMA model, described by equation (\ref{eq:Eq1}), with structural shocks satisfying Assumption (\ref{as:Assumption1}). If equation (\ref{eq:Eq2}) holds for $\mathbb{T}_{+}=\{ z\in\mathbb{C} \; | \; |z|\leq{1} \}$, then $G_{k}^{2}(\lambda_{k})$ is constant and equivalent to
\begin{align*}
G_{k}^{2}(\lambda_{k}) &= \boldsymbol{B}(\boldsymbol{\vartheta}_{0})\left(\mathbf{v}_{k}^{\prime}(\boldsymbol{\varepsilon})\boldsymbol{\Omega}^{\otimes{k-1}}\mathbf{v}_{k}({\boldsymbol{\varepsilon}})\right)\boldsymbol{B}^{\prime}(\boldsymbol{\vartheta}_{0}),\;\; \forall\;\; \lambda_{k}\in[-\pi, \pi],\;k=3,4
\end{align*}
and $$G_{k}^{2}(\lambda_{k})=\mathbf{v}_{k}^{\prime}(\boldsymbol{u})\mathbf{v}_{k}(\boldsymbol{u}),\quad k=3,4.$$
\end{corollary}
Corollary (\ref{coro:Corollary1}) can be understood as follows: when $\boldsymbol{y}_{t}$ is generated by a fundamental SVARMA model, higher-order cumulant spectral densities of reduced-form errors are constant for any frequency and they are equivalent to contemporaneous $k$-th order cumulants of such errors. In other words, when data is generated by a fundamental model, reduced-form errors are a simply rotation of structural shocks. Besides, this result implies that we can employ our proposal under the restricted set-up that \citet{maxand2020identification,guay2021identification} consider.
\section{Estimating the number of non-Gaussian elements}
\label{sec:EstimatingNumber}
The aim of this paper is to establish a method for assessing empirically the identification requirement of having all components non-Gaussian distributed in the vector of structural shocks in a VARMA model without requiring to estimate the structural parameters of the model, just employing reduced form errors. Thus, although the fundamental filter, $\boldsymbol{\Psi}(z,\boldsymbol{\vartheta}_{f})$, is identified and can be estimated parametrically, the all-pass filter $\boldsymbol{\delta}\left(z,\boldsymbol{\vartheta}_{f}\right)$ cannot be.
Nonetheless, the $k$-th order cumulant spectrum, $g_{k}(\boldsymbol{\lambda})$, based on reduced-form residuals can be estimated non-parametrically. A drawback of our approach is the necessity of a smoothing window in the estimation procedure, provoking that convergence rates are slower than standard $\sqrt{T}$. The spectral estimates follows closely \citet{brillinger1967asymptotic,brillinger2001time}. Nevertheless, before detailing the estimation details, we consider important to discuss our approach for estimating the number of non-Gaussian structural shocks.
\subsection{Estimation of $\rank\left[\text{Re}(G_{k}^{2}(\lambda_{k}))\right]$ }
\label{sec:TestRank}
We have showed that the number of skewed and non-mesokurtic shocks is equal to the rank of $G_{3}^{2}(\lambda_{3})$ and $G_{4}^{2}(\lambda_{4})$, respectively. However, this object is an non-negative integer, i.e. it is discrete. Thus, a suitable way to estimate it is through an iterative hypothesis testing procedure. We initiate with the null hypothesis of joint Gaussianity, i.e. $\rank\left(G_{k}^{2}(\lambda_{k})\right)=0$. If this is rejected, we update the null hypothesis to $\rank\left(G_{k}^{2}(\lambda_{k})\right)=1$, i.e only one asymmetric or non-mesokurtic structural shock. In case this is rejected, we update again the null hypothesis. We continue this until null hypothesis is rejected or if we reject the last null hypothesis, which states $\rank\left(G_{k}^{2}(\lambda_{k})\right)=d-1$. Besides, we employ $\text{Re}\left(G_{k}^{2}(\lambda_{k})\right)$, for simplifying the analysis.
The literature about estimating the rank of a matrix started with the seminal work of \citet{anderson1951estimating}, who proposed a likelihood ratio test for the rank of some matrix coefficients, $\boldsymbol{\Pi}$, in the context of multivariate regression. Let $r=\rank{(\boldsymbol{\Pi})}$, the null hypothesis ($H_{0}$) is $r=r_{0}$ against the alternative ($H_{1}$) $r>r_{0}$. The issue of this proposal is that the test statistic may not be valid under heteroskedasticity or serial dependence, since the variance of the unrestricted estimator has not a Kronecker structure. \citet{lewbel1991rank,cragg1996asymptotic} proposes an statistic based on LDU decomposition of the unrestricted estimation of $\boldsymbol{\Pi}$, and to prevent potential pitfalls, they allow for Gaussian permutation of unrestricted estimator, which imposes strong assumptions for ensuring that Gaussian elimination procedure performs properly. \citet{cragg1997inferring} propose a rank statistic based on a minimum distance criterion. The disadvantage of their proposal is that requires numerical optimization for computing the rank statistic, a situation becomes difficult to handle it as dimension of $\boldsymbol{\Pi}$ grows.
\citet{robin2000tests} employ singular values from the matrix of interest, because of the number of non-zero singular values is equal to the rank of the matrix $\boldsymbol{\Pi}$. Their null hypothesis is equal to the number of the $r$ largest singular values. Then, the statistic employed is the sum of all singular values excluding the $r$ largest ones. The asymptotic distribution of his test statistic is a weighted average of several independent $\chi^{2}$ distributions with one degree of freedom. This makes necessary to tabulate critical values for each case.
In this paper we employ the approach stated by \citet{kleibergen2006generalized}. Like \citet{robin2000tests}, their method is based on the singular value decomposition (SVD) of $\boldsymbol{\Pi}$. The SVD of a matrix $\boldsymbol{\Pi}_{m\times{n}}$ consists on finding orthonormal matrices $\boldsymbol{U}_{m\times{m}}$ and $\boldsymbol{V}_{m\times{m}}$ such that\footnote{See \citet{golub2012matrix} for more details on singular value decomposition}
\begin{equation*}
\boldsymbol{U}^{\prime}\boldsymbol{\Pi}\boldsymbol{V}=\boldsymbol{\mathcal{L}}
\end{equation*}
where $\boldsymbol{\mathcal{L}}$ is a rectangular matrix with a block $\boldsymbol{\mathcal{L}}_{\longbar{m}}=\text{diag}(l_{1},\dots,l_{\longbar{m}})\in\mathbb{R}^{m\times{n}}$ with $\longbar{m}=\min\{m,n\}$ and $l_{1}\geq{l_{2}}\geq\dots\geq{l_{\longbar{m}}}\geq{0}$. For instance, when $m>n$, the matrix $\boldsymbol{\mathcal{L}}=\begin{bmatrix}
\boldsymbol{\mathcal{L}}_{\longbar{m}} & \boldsymbol{0}_{n\times{(m-n)}}
\end{bmatrix}^{\prime}$; instead, when $m<n$, $\boldsymbol{\mathcal{L}}=\begin{bmatrix}
\boldsymbol{\mathcal{L}}_{\longbar{m}} & \boldsymbol{0}_{n\times{(m-n)}}
\end{bmatrix}$.
Now, for a given integer $r\in\{0,\dots,\longbar{m}-1\}$, \citet{kleibergen2006generalized} show
\begin{equation*}
\boldsymbol{\Pi} = \boldsymbol{C}_{r}\boldsymbol{D}_{r} + \boldsymbol{C}_{r,\perp}\boldsymbol{\mathcal{L}}_{r}\boldsymbol{D}_{r,\perp}
\end{equation*}
where $\boldsymbol{C}_{r}\boldsymbol{D}_{r}=\underbrace{\begin{bmatrix}\boldsymbol{U}_{11} \\ \boldsymbol{U}_{21}\end{bmatrix}}_{m\times{r}}\underbrace{\boldsymbol{\mathcal{L}}_{11}}_{r\times{r}}\underbrace{\begin{bmatrix}\boldsymbol{V}_{11}^{\prime} & \boldsymbol{V}_{21}^{\prime}\end{bmatrix}}_{r\times{n}}$, $\boldsymbol{C}_{r,\perp}\boldsymbol{\mathcal{L}}_{r}\boldsymbol{D}_{r,\perp}=\underbrace{\begin{bmatrix}\boldsymbol{U}_{12} \\ \boldsymbol{U}_{22}\end{bmatrix}}_{m\times{(m-r)}}\underbrace{\boldsymbol{\mathcal{L}}_{22}}_{(m-r)\times(n-r)}\underbrace{\begin{bmatrix}\boldsymbol{V}_{12}^{\prime} & \boldsymbol{V}_{22}^{\prime}\end{bmatrix}}_{(n-r)\times{n}}$. Besides, $\boldsymbol{C}_{r,\perp}$ is the orthogonal complement of $\boldsymbol{C}_{r}$, i.e. $\boldsymbol{C}_{r}^{\prime}\boldsymbol{C}_{r,\perp}=\boldsymbol{0}$; similarly, $\boldsymbol{D}_{r,\perp}$ is the orthogonal complement of $\boldsymbol{D}_{r}$.
When $\boldsymbol{\mathcal{L}}_{r}=\boldsymbol{0}$, which implies $l_{i}=0$ for all $\longbar{m}\geq i >r$, then $\boldsymbol{\Pi}=\boldsymbol{C}_{r}\boldsymbol{D}_{r}$. Thus, $\rank(\boldsymbol{\Pi})=\rank(\boldsymbol{C}_{r}\boldsymbol{D}_{r})\leq \min\{ \rank(\boldsymbol{C}_{r}),\;\rank(\boldsymbol{D}_{r}) \}$. And, whenever $\boldsymbol{C}_{r}$ and $\boldsymbol{D}_{r}$ are full rank matrices, then $\rank{(\boldsymbol{\Pi})}=r$. Nonetheless, since $\boldsymbol{\Pi}$ is not observable, the singular values decomposition is applied to its sample counterpart $\boldsymbol{\hat{\Pi}}=\boldsymbol{\Hat{C}}_{r}\boldsymbol{\hat{D}}_{r} + \boldsymbol{\hat{C}}_{r,\perp}\boldsymbol{\hat{\mathcal{L}}}_{r}\boldsymbol{\hat{D}}_{r,\perp}$. Assuming $\sqrt{T}(\vect{[\boldsymbol{\hat{\Pi}}]}-\vect[\boldsymbol{\Pi}]) \rightsquigarrow \mathcal{N}_{mn}(\boldsymbol{0},\boldsymbol{\Omega})$ with $\boldsymbol{\Omega}$ full rank. Thus, under the null hypothesis of $\rank(\boldsymbol{\Pi})=r$, it holds $\sqrt{T}\boldsymbol{\hat{\ell}}_{r} \rightsquigarrow \mathcal{N}_{(m-r)(p-r)}\left(\boldsymbol{0}, \tilde{\boldsymbol{\Omega}}_{r}\right)$ where $\boldsymbol{\Tilde{\Omega}}_{r}=(\boldsymbol{D}_{r,\perp}\otimes{\boldsymbol{C}_{r,\perp}^{\prime}})\boldsymbol{\Omega}(\boldsymbol{D}_{r,\perp}\otimes{\boldsymbol{C}_{r,\perp}^{\prime}})$ and $\boldsymbol{\hat{\ell}}_{r}=\vect\left(\boldsymbol{\hat{\mathcal{L}}}_{r}\right)$. The test statistic is similar to a Wald test statistic:
\begin{equation}
\label{eq:KP_Stat} KP_{r}^{(T)}=T\boldsymbol{\hat{\ell}}_{r}^{\prime}\boldsymbol{\Tilde{\Omega}}_{r}^{-1}\boldsymbol{\hat{\ell}}_{r} \rightsquigarrow \chi^{2}_{(m-r)(p-r)}
\end{equation}
\subsubsection{Null Hypotheses and Test Statistic}
We have some differences respect to \citeauthor{kleibergen2006generalized} set-up. First, our matrix of interest $\boldsymbol{\Pi}=\text{Re}({G}_{k}^{2}(\lambda_{k}))$ is symmetric. The number of non-repeated elements is $\frac{d(d+1)}{2}$. Hence, the asymptotic variance of $\vect\left[\hat{\boldsymbol{\Pi}}\right]$, $\boldsymbol{\Omega}$, is singular. Second, the asymptotic distribution changes when the null hypothesis is joint Gaussianity ($r=0$) of structural shocks, $\boldsymbol{\varepsilon}_{t}$. Finally, the asymptotic distribution has a different convergence rate than the standard $\sqrt{T}$, which we call $a_{T}$.
Before continuing, it is important to characterize the null hypothesis we evaluate at each stage. Let $s$ denotes the stage of the iterative testing procedure, our null hypothesis and alternative hypotheses are
\begin{align*}
H_{0,s}:\quad \rank\left(\boldsymbol{\Pi}\right)&=r_{s}=s\;; \\
H_{1,s}:\quad \rank\left(\boldsymbol{\Pi}\right)&>r_{s}
\end{align*}
we start at $s=0$, i.e. the initial stage, and continue, while $H_{0,s}$ is rejected, up to $s=d-1$.
\paragraph{Test Statistic when $r_{0}=0$}\mbox{}\newline
When $r=0$ we are testing the null hypothesis of joint Gaussianity of structural shocks, i.e. $\boldsymbol{\varepsilon}_{t}$ is an i.i.d. Gaussian process. Therefore, the KP-test statistic is
\begin{align}
\hspace{-0.5cm}
KP^{(T)}_{r_{0}} =a_{T}^{2}\hat{\boldsymbol{\ell}}_{0}^{\prime}\text{Var}\left( \left( \overline{\boldsymbol{D}}_{0,\perp}\otimes\overline{\boldsymbol{C}}^{\prime}_{0,\perp} \right)\boldsymbol{Q} \right)^{\dagger}\hat{\boldsymbol{\ell}}_{0},
\end{align}
where $\boldsymbol{A}^{\dagger}$ denotes the generalized Moore-Penrose inverse of $\boldsymbol{A}$, $\boldsymbol{Q}$ is the vectorized form of a multivariate Whishart distribution, which parameters are detailed below, and $ \overline{\boldsymbol{D}}_{0,\perp}$ and $\overline{\boldsymbol{C}}^{\prime}_{0,\perp}$ are constant full-rank matrices.
\paragraph{Test Statistic when $r_{s}\geq{1}$}\mbox{}\newline
Under this type of null hypotheses, higher-order spectral density is not zero. Hence, we can use a Wald statistic like \citet{kleibergen2006generalized}. However, since the rate of convergence is $a_{T}$ and the asymptotic variance of $\vect(\hat{\boldsymbol{\Pi}})$ is singular, the test statistic is
\begin{equation}
\label{eq:KP_Stat2}
KP^{(T)}_{r_{s}}=a_{T}^{2}\boldsymbol{\hat{\ell}}_{r_{s}}^{\prime}\left(\left[(\boldsymbol{D}_{r_{s},\perp}\otimes{\boldsymbol{C}_{r_{s},\perp}^{\prime}})\boldsymbol{\Omega}(\boldsymbol{D}_{r_{s},\perp}\otimes{\boldsymbol{C}_{r_{s},\perp}^{\prime}})^{\prime}\right]\right)^{\dagger}\boldsymbol{\hat{\ell}}_{r_{s}}
\end{equation}
where $\boldsymbol{D}_{r_{s},\perp}$ and ${\boldsymbol{C}_{r_{s},\perp}}$ are $(d-r_{s})\times{d}$ and $d\times{d-r_{s}}$ full-rank, constant matrices. Then, $\rank{(\boldsymbol{D}_{r_{s},\perp}\otimes{\boldsymbol{C}_{r_{s},\perp}^{\prime}})}=(d-r_{s})^{2}$ but $\rank{\left[(\boldsymbol{D}_{r,\perp}\otimes{\boldsymbol{C}_{r,\perp}^{\prime}})\boldsymbol{\Omega}(\boldsymbol{D}_{r,\perp}\otimes{\boldsymbol{C}_{r,\perp}^{\prime}})^{\prime}\right]}=\text{rk}\leq\min\left\{ \rank\left(\boldsymbol{\Omega}\right),(d-r)^{2} \right\}$. Therefore, the degrees of freedom of the asymptotic distribution of $KP^{(T)}_{r_{s}}$ are unknown.
\subsection{Bootstrap Test}
Performing the estimation procedure using asymptotic distribution may be problematic for several reasons. First, it may be difficult to estimate accurately the asymptotic variance $\boldsymbol{\Omega}$, specially with small/medium sample sizes. Second, the asymptotic distribution under the null hypothesis of joint Gaussianity is unknown; although its critical values may be calculated by Monte-Carlo simulations. Third, under the rest of null hypotheses, the degrees of freedom of the limiting distribution ${\text{rk}}$ are unknown. For all these, we employ a bootstrap test procedure.
We follow \citet{portier2014bootstrap} recommendation, and the bootstrap sample must reflect the null hypothesis. Nonetheless, this is challenging. Under the null hypothesis, there are $d-r_{s}$ Gaussian shocks and $r_{s}$ non-Gaussian, and the columns of $\boldsymbol{U}$ or $\boldsymbol{V}$ are associated to its respective singular values. In particular, notice that $\boldsymbol{\Pi}^{\prime}\boldsymbol{\Pi}\mathbf{v}_{j}=l_{j}^{2}\mathbf{v}_{j}$, i.e. the columns of $\boldsymbol{V}$ span the column space of $\boldsymbol{\Pi}$ and each of them is associated to $l_{j}^{2}$. Thus, if $\rank(\boldsymbol{\Pi})=r_{s}$, then the first $r_{s}$ columns of $\boldsymbol{V}$ ($\boldsymbol{V}_{r_{s}}$) are associated to the $r_{s}$ largest singular values, while the last $d-r_{s}$ columns ($\boldsymbol{V}_{d-r_{s}}$) are associated to the $d-r_{s}$ null singular values. Therefore, $\boldsymbol{V}_{r_{s}}$ and $\boldsymbol{V}_{d-r_{s}}$ span the non-Gaussian and Gaussian dimensions, respectively. Then, both matrices can be employed for projecting the residuals under the non-Gaussian dimension.
The bootstrap sample is obtained following the next algorithm:
\begin{enumerate}
\item[(i)] Given the original sample $\{\boldsymbol{y}_{t}\}_{t=\bar{t}-1}^{T}$ with $\bar{t}=\min\{p,q+1\}$, estimate a fundamental SVARMA model, $(\hat{\boldsymbol{\vartheta}}_{f},p,q)$. Obtain the demeaned residuals $\{\hat{u}_{t}\}_{t=1}^{T}$ and its variance-covariance matrix $\hat{\Sigma}_{\hat{u}}$.
\item[(ii)] Compute the matrix of interest $\hat{\boldsymbol{\Pi}}=\text{Re}\left(\hat{G}_{k}^{2,(T)}(\lambda_{k})\right)$ and its singular value decomposition elements, in particular $\hat{\boldsymbol{V}}$ and $\hat{\boldsymbol{\mathcal{L}}}$.
\item[(iii)] We follow \citet{nordhausen2017asymptotic} to obtain the projection matrix to non-Gaussian dimension. Under the $H_{0,r_{s}}$, take $\hat{\boldsymbol{V}}_{r_{s}}=\hat{\boldsymbol{V}}_{\bullet,(r_{s}+1):d}$ and define the projection matrix $\mathbb{M}_{r_{s}}=\boldsymbol{I}_{d}-\hat{\Sigma}_{\hat{u}}^{1/2}\hat{\boldsymbol{V}}_{r_{s}}\hat{\boldsymbol{V}}_{r_{s}}^{\prime}\hat{\Sigma}_{\hat{u}}^{-1/2}$, where $\hat{\Sigma}_{\hat{u}}$ is the variance of reduced-form residuals.
\item[(iv)] Now, obtain unrestricted bootstrap residuals $\{\tilde{\boldsymbol{u}}_{t}\}_{t=\bar{t}-1}^{T}$. Since the reduced-form residuals are only uncorrelated, but no independent, the standard independent bootstrap re-sampling may affect, critically, the all-pass structure of residuals. Hence, a block-bootstrap procedure is employed\footnote{See \citet{kreiss2011bootstrap,kreiss2012bootstrap}, and the references therein, for more details.}.
\item[(v)] The bootstrap reduced-form residuals restricted to $H_{0,r_{s}}$ is given by
$${\boldsymbol{u}}^{\star}_{t} = \mathbb{M}_{r_{s}}\tilde{\boldsymbol{u}}_{t} + \hat{\Sigma}_{\hat{u}}^{1/2}\hat{\boldsymbol{V}}_{r_{s}}\mathcal{N}_{d-r_{s}}(\boldsymbol{0},\boldsymbol{I}_{d-r_{s}}),$$
where $\mathcal{N}_{d-r_{s}}(\boldsymbol{0},\boldsymbol{I}_{d-r_{s}})$ represents the standard, multivariate Gaussian distribution.
\item[(iv)] Finally, the bootstrap sample $\{\boldsymbol{y}^{\star}_{t}\}_{t=1}^{T}$ is constructed using ${\boldsymbol{u}}^{\star}_{t}$ and estimated parameters $(\hat{\boldsymbol{\vartheta}}_{f},p,q)$.
\end{enumerate}
Once the bootstrap sample restricted to $H_{0,r_{s}}$ is obtained, we proceed to the sequential testing procedure.
\begin{enumerate}
\item[1.] Start with $s=0$, then $H_{0,r_{s}}:\; r_{s}=0$.
\item[2.] Compute the test statistic $KP_{r_{s}}^{(T)}$.
\item[3.] Using the $B$ bootstrapped samples, compute the $b$-th test statistic $KP^{(T),\star}_{r_{s},b}$ and the p-value $\widehat{\text{pv}}_{r_{s}}=\frac{\sum_{b=1}^{B}{\mathbf{1}\left( KP^{(T)}_{r_{s}}\leq KP^{(T),\star}_{r_{s},b} \right)}+1}{B+1}$.
\item[3.] If $\widehat{\text{pv}}_{r_{s}}\geq{\alpha}$\footnote{ $\alpha$ represents the significance level.}, $H_{0,r_{s}}$ is not rejected and the sequence has ended, concluding that estimated rank is $\hat{r}=r_{s}$. Otherwise, go back to point ($1$), update the step to $s+1$ and re-do (2).
\item[4.] The procedure finishes when is no possible to reject $H_{0,r_{s}}$. When $s=d-1$, if $H_{0,r_{d-1}}$ is rejected, then $\hat{r}=d$.
\end{enumerate}
\subsection{Estimation of $k$-th order cumulant spectrum}
The sample periodogram for a $k$-tuple of reduced-form errors, $\boldsymbol{u}_{t}$, is
\begin{equation}
\label{eq:Periodogram_k}
I_{\boldsymbol{c},k}^{(T)}(\boldsymbol{\lambda}) = \frac{1}{(2\pi)^{k-1}T}\prod_{j=1}^{k}{z_{c_{j}}^{T}(\lambda_{t_{j}})},
\end{equation}
where $\boldsymbol{c}=\left(c_{1},\dots,c_{k}\right)\in\upsigma_{k}\left(\{1,\dots,d\}\right)$ and $\upsigma_{k}\left(\{1,\dots,d\}\right)$ the set of all permutations with $k$ elements of the index set $\{1,\dots,d\}$. $\lambda_{t_{j}}=2\pi \frac{t_{j}}{T}$ for $t_{j}=1,\dots,T-1$, $j=1,\dots,k-1$ and $\sum_{j=1}^{k}{\lambda_{t_{j}}}=0[\text{mod}(2\pi)]$. Besides, $z_{c_{j}}^{T}(\lambda_{t_{j}})=\sum_{t=0}^{T-1}{u_{c_{j},t}e^{-i\lambda_{t_{j}}{t}}}$ is the discrete Fourier transform (DFT) of $c_{j}$-th reduced-form error, $u_{c_{j},t}$.
A consistent estimator of the $k$-th joint cumulant spectrum for a $\boldsymbol{c}$-tuple of reduced-form residuals is
\begin{equation}
\hat{g}_{\boldsymbol{c},k}^{(T)}(\boldsymbol{\lambda}) = \left(\frac{2\pi}{T}\right)^{k-1} \sum_{t_{1},\dots,t_{k}=0}^{T-1}{W_{T}\left( \lambda_{1}-\lambda_{t_{1}},\dots,\lambda_{k}-\lambda_{t_{k}} \right)I_{\boldsymbol{c},k}^{(T)}\left( \boldsymbol{\lambda} \right) }
\end{equation}
with $W_{T}(\boldsymbol{a})=H_{T}^{-(k-1)}\sum_{\boldsymbol{j}=(j_{1},\dots,j_{k}):\sum_{m=1}^{k}{(j_{m}+a_{m})}=0}{ W\left( \frac{1}{B_{T}}(\boldsymbol{a}+2\pi\boldsymbol{j}) \right) }$ and $H_{T}$ satisfies $\lim\limits_{T\to{\infty}}{H_{T}}=0$ and $\lim\limits_{T\to\infty}{TH_{T}^{k-1}}=\infty$. The weighting function, $W(\boldsymbol{a})$, is symmetric around $\boldsymbol{0}$ and satisfies conditions stated in \citet[Assumption II]{brillinger1967asymptotic}.
Hence, the estimator of $g_{k}(\boldsymbol{\lambda})$ is $\hat{g}_{k}^{(T)}(\boldsymbol{\lambda}) = \left[ \hat{g}_{\boldsymbol{c},k}^{(T)}(\boldsymbol{\lambda}) \right]_{\boldsymbol{c}\in{\upsigma_{k}\left(\{1,\dots,d\}\right)}}$. Then, $\hat{G}_{k}^{(T)}(\boldsymbol{\lambda})$ is the matrisied version of $\hat{g}_{k}^{(T)}(\boldsymbol{\lambda})$, which is of dimension $d^{k-1}\times{d}$. Hence, $\hat{G}_{k}^{2,(T)}(\lambda_{k})=\left[\hat{G}_{k}^{(T)}(\boldsymbol{\lambda})\right]^{*}\hat{G}_{k}^{(T)}(\boldsymbol{\lambda})$ and the estimator of our matrix of interest is $\hat{\boldsymbol{\Pi}}=\text{Re}\left( \hat{G}_{k}^{2,(T)}(\lambda_{k}) \right)$, which is square, symmetric, non-negative definite matrix.
\subsection{Asymptotics of spectral estimators}
\citet[Theorems (1)-(2)]{brillinger1967asymptotic} showed $\E\left[ \hat{g}_{\boldsymbol{c},k}^{(T)}(\boldsymbol{\lambda}) \right] \to {g}_{\boldsymbol{c},k}(\boldsymbol{\lambda})$
and
\begin{align*}
\hspace{-1.0cm}
H_{T}^{k-1}T\Cov\left( \hat{g}_{\boldsymbol{c},k}^{(T)}(\boldsymbol{\lambda}), \hat{g}_{\boldsymbol{b},k}^{(T)}(\boldsymbol{\mu}) \right) &\to \Lambda_{(\boldsymbol{c},\boldsymbol{b})}^{k}(\boldsymbol{\lambda},\boldsymbol{\mu})\\
\hspace{-1.0cm}
{} & \hspace*{-3.5cm} =2\pi \sum_{\upsigma\left(\{1,\dots,k\}\right)}{ \prod_{m=1}^{k}{\eta\left( \lambda_{m} - (\upsigma\mu)_{m} \right)g_{2,(c_{m},(\upsigma{b})_{m})}(\lambda_{m}) } }\int{W(\boldsymbol{\beta})W(\upsigma\boldsymbol{\beta})\delta_{D}\left( \sum_{m=1}^{k}{\beta_{m}} \right)d\boldsymbol{\beta}}
\end{align*}
as $T\to\infty$, where $\upsigma\boldsymbol{b}=(b_{\upsigma_{1}},\dots,b_{\upsigma_{k}})$, $\upsigma=(\upsigma_{1},\dots,\upsigma_{k})\in\upsigma\left(\{1,\dots,k\}\right)$. $\sum_{m=1}^{k}{\lambda_{m}}=0[\text{mod}(2\pi)]$. $\eta\left({x}\right)=\sum_{j=-\infty}^{\infty}{\delta_{D}(x+2\pi{j})}$ and $\delta_{D}(x)$ is the delta-Dirac function.
Based on this:
\begin{align*}
\hspace*{-1.0cm}
\sqrt{H_{T}^{k-1}T}\left( \hat{g}_{k}^{(T)}(\boldsymbol{\lambda}) - {g}_{k}(\boldsymbol{\lambda}) \right) &\rightsquigarrow \mathcal{N}_{d^{k}}^{\mathcal{C}}\left( \boldsymbol{0},\; \boldsymbol{\Lambda}^{k}(\boldsymbol{\lambda},\boldsymbol{\lambda}) \right) \\
\hspace*{-1.0cm}
\sqrt{H_{T}^{k-1}T}\left( \begin{bmatrix}
\text{Re}(\hat{g}_{k}^{(T)}(\boldsymbol{\lambda})) \\
\text{Im}(\hat{g}_{k}^{(T)}(\boldsymbol{\lambda}))
\end{bmatrix} - \begin{bmatrix}
\text{Re}({g}_{k}(\boldsymbol{\lambda})) \\
\text{Im}({g}_{k}(\boldsymbol{\lambda}))
\end{bmatrix} \right) &\rightsquigarrow \mathcal{N}_{2d^{k}}\left(\boldsymbol{0}, \tilde{\boldsymbol{\Sigma}}(\boldsymbol{\lambda})\right)
\end{align*}
where $\boldsymbol{\Lambda}^{k}(\boldsymbol{\lambda},\boldsymbol{\mu}) = \left[ \Lambda_{(\boldsymbol{c},\boldsymbol{b})}^{k}(\boldsymbol{\lambda},\boldsymbol{\mu}) \right]_{(\boldsymbol{c},\boldsymbol{b})\in\upsigma_{k}\left(\{1,\dots,d\}\right)\times{\upsigma_{k}\left(\{1,\dots,d\}\right)} }$, \\ $\tilde{\boldsymbol{\Sigma}}(\boldsymbol{\lambda})=\begin{bmatrix}
\text{Re}(\boldsymbol{\Lambda}^{k}(\boldsymbol{\lambda},\boldsymbol{\lambda})) & - \text{Im}(\boldsymbol{\Lambda}^{k}(\boldsymbol{\lambda},\boldsymbol{\lambda})) \\
\text{Im}(\boldsymbol{\Lambda}^{k}(\boldsymbol{\lambda},\boldsymbol{\lambda})) & \text{Re}(\boldsymbol{\Lambda}^{k}(\boldsymbol{\lambda},\boldsymbol{\lambda}))
\end{bmatrix}$ and $\mathcal{N}_{d^{k}}^{\mathcal{C}}$ denotes a multivariate complex Normal distribution.
Since $\vect\left( {G}_{k}(\boldsymbol{\lambda}) \right) = {g}_{k}(\boldsymbol{\lambda})$, the asymptotic distribution of $\vect\left( \hat{G}_{k}^{(T)}(\boldsymbol{\lambda}) \right) $ is equivalent to the one for $\hat{g}_{k}^{(T)}(\boldsymbol{\lambda})$. Finally, to find the asymptotic distribution of our matrix of interest $\hat{\boldsymbol{\Pi}}$ notice that
\begin{align*}
\hspace{-1.0cm}
\vect\left(\hat{\boldsymbol{\Pi}}\right)=\vect\left(\text{Re}\left\{\hat{G}_{k}^{2,(T)}(\boldsymbol{\lambda})\right\} \right) &= \left[ \left( I_{d}\otimes{\hat{g}_{\text{Re},k}^{(T)}(\boldsymbol{\lambda})} \right)^{\otimes{2}} + \left(I_{d}\otimes{\hat{g}_{\text{Im},k}^{(T)}(\boldsymbol{\lambda})}\right)^{\otimes{2}} \right]\vect(\mathbf{I}(d))\;,
\end{align*}
where $g_{\text{Re},k}(\boldsymbol{\lambda}) = \text{Re}\left[{g}_{k}(\boldsymbol{\lambda})\right]$ and $g_{\text{Im},k}(\boldsymbol{\lambda}) = \text{Im}\left[{g}_{k}(\boldsymbol{\lambda})\right]$.
Thus, the asymptotic distribution of $\vect\left(\hat{\boldsymbol{\Pi}}\right)$ depends on $H_{0,r_{s}}$. Let $\boldsymbol{D}_{\hat{G}^{2}_{k}}(\boldsymbol{\lambda})$ denotes the Jacobian matrix of $\vect\left( \text{Re}\left\{\hat{G}_{k}^{2,(T)}(\boldsymbol{\lambda})\right\} \right)$ evaluated at population values $\left( g_{\text{Re},k}^{\prime}(\boldsymbol{\lambda}), g_{\text{Im},k}^{\prime}(\boldsymbol{\lambda}) \right)^{\prime}$.
\begin{proposition} \label{prop:Proposition40}
If $ \sqrt{H_{T}^{k-1}T}\left( \hat{g}_{k}^{(T)}(\boldsymbol{\lambda}) - {g}_{k}(\boldsymbol{\lambda}) \right) \rightsquigarrow \mathcal{N}_{d^{k}}^{\mathcal{C}}\left( \boldsymbol{0},\; \boldsymbol{\Lambda}^{k}(\boldsymbol{\lambda},\boldsymbol{\lambda}) \right)$, then:
\begin{enumerate}
\item Under $H_{0,r_{0}}$, i.e. structural shocks $\boldsymbol{\varepsilon}_{t}$ are Gaussian distributed, the asymptotic distribution of $\hat{\boldsymbol{\Pi}}$ is
\begin{align}
\label{eq:DeltaMethod_Gaussian_G2Sq}
H_{T}^{k-1}T\left( \vect\left( \hat{\boldsymbol{\Pi}} \right) \right) &\rightsquigarrow \frac{1}{2}\left(I_{d^{2}}\otimes{[\mathcal{N}_{2d^{k}}^{\prime}]}^{\otimes{2}}\right)\boldsymbol{D}^{\otimes{2}}_{\hat{G}^{2}_{k}}(\boldsymbol{\lambda})\;,
\end{align}
where $\boldsymbol{D}^{\otimes{2}}_{\hat{G}^{2}_{k}}(\boldsymbol{\lambda})$ is the vectorized Hessian\footnote{The Hessian of a vector-valued function $\mathbb{R}^{d}\mapsto{\mathbb{R}^{q}}$ is a $3$-tensor. Intuitively, the Hessian matrix is $3$-dimensional array of dimensions $d\times{d}\times{q}$.} of $\vect\left( \text{Re}\left\{\hat{G}_{k}^{2,(T)}(\boldsymbol{\lambda})\right\} \right)$. $\mathcal{N}_{2d^{k}}$ represents a multivariate Gaussian random $2d^{k}$-dimensional column-vector with mean $\boldsymbol{0}$ and covariance $\tilde{\boldsymbol{\Sigma}}(\boldsymbol{\lambda})$. $\mathcal{N}_{2d^{k}}^{\otimes{2}}=\vect(\mathcal{N}_{2d}\mathcal{N}_{2d}^{\prime})$ and $\mathcal{N}_{2d^{k}}\mathcal{N}_{2d^{k}}^{\prime}\sim{\mathcal{W}(\tilde{\boldsymbol{\Sigma}}(\boldsymbol{\lambda}),1)}$, a Wishart distribution with parameters $(\tilde{\boldsymbol{\Sigma}}(\boldsymbol{\lambda}),1)$.
\item Under $H_{0,r_{s}}$ for $s\geq{1}$, the asymptotic distribution of $\hat{\boldsymbol{\Pi}}$ is
\begin{align}
\hspace{-1.0cm}
\label{eq:DeltaMethod_G2Sq}
\sqrt{H_{T}^{k-1}T}\left( \vect\left(\hat{\boldsymbol{\Pi}} \right) - \vect\left(\boldsymbol{\Pi} \right) \right) &\rightsquigarrow \mathcal{N}_{d^{2}}\left( \boldsymbol{0},\; \boldsymbol{D}^{\prime}_{\hat{G}^{2}_{k}}(\boldsymbol{\lambda})\tilde{\boldsymbol{\Sigma}}(\boldsymbol{\lambda})\boldsymbol{D}_{\hat{G}^{2}_{k}}(\boldsymbol{\lambda}) \right)
\end{align}
\end{enumerate}
\end{proposition}
\subsection{Asymptotic Equivalence of the $k$-th higher order spectrum of $\hat{\boldsymbol{u}}_{t}$ and ${\boldsymbol{u}}_{t}$}
The above estimators and their asymptotic results employ population reduced-form residuals, $\boldsymbol{u}_{t}$, which are not observable. Instead, given a sample $\{\boldsymbol{y}_{t}\}_{t=1}^{T}$, we only can compute the estimated reduced-form residuals $\hat{\boldsymbol{u}}_{t}=\boldsymbol{\Psi}^{-1}(L,\hat{\boldsymbol{\vartheta}}_{f})\boldsymbol{y}_{t}\boldsymbol{1}(1\leq t \leq T )$, where $\boldsymbol{\Psi}(L,\hat{\boldsymbol{\vartheta}}_{f})$ is the estimated fundamental filter and $\boldsymbol{1}(1\leq t \leq T )$ represents indicator function that takes value $1$ when time is between $1$ and $T$. Hence, the $k$-th higher order spectrum for the estimated reduced-form residuals is
\begin{align*}
\hat{g}_{\boldsymbol{c},k}^{\hat{\boldsymbol{u}},(T)}(\boldsymbol{\lambda}) &= \left(\frac{2\pi}{T}\right)^{k-1} \sum_{t_{1},\dots,t_{k}=0}^{T-1}{W_{T}\left( \lambda_{1}-\lambda_{t_{1},T},\dots,\lambda_{k}-\lambda_{t_{k},T} \right)I_{\boldsymbol{c},k}^{\hat{\boldsymbol{u}},(T)}\left( \boldsymbol{\lambda}_{T} \right) }.
\end{align*}
Therefore, it is important to establish the asymptotic equivalence between $\hat{g}_{\boldsymbol{c},k}^{\hat{\boldsymbol{u}},(T)}(\boldsymbol{\lambda})$ and $\hat{g}_{\boldsymbol{c},k}^{{\boldsymbol{u}},(T)}(\boldsymbol{\lambda})$, otherwise the asymptotic results stated for $\hat{g}_{\boldsymbol{c},k}^{{\boldsymbol{u}},(T)}(\boldsymbol{\lambda})$ cannot be employed and new asymptotics must be stated for $\hat{g}_{\boldsymbol{c},k}^{{\boldsymbol{u}},(T)}(\boldsymbol{\lambda})$. When estimators are linear, asymptotic equivalence may be straightforward; but, when the estimators do not possess that property, such result is not that direct. This analysis is omitted in \citet{maxand2020identification,guay2021identification} works.
\begin{proposition} \label{prop:Proposition41}
Under assumptions (\ref{as:Assumption1}-\ref{as:Assumption2}), it holds
\begin{equation*}
\mathbb{E}\sqrt{H_{T}^{k-1}T}\left| \hat{g}_{\boldsymbol{c},k}^{\boldsymbol{u},(T)}(\boldsymbol{\lambda}) - \hat{g}_{\boldsymbol{c},k}^{\hat{\boldsymbol{u}},(T)}(\boldsymbol{\lambda})\right| \leq C H_{T}^{\frac{k-1}{2}}
\end{equation*}
\end{proposition}
Proposition (\ref{prop:Proposition41}) states the asymptotic equivalence between estimators of higher-order cumulant spectrum using unobserved (population) reduced-form residuals and its sample estimate. This result implies that the discrepancy generated by using $\hat{g}_{\boldsymbol{c},k}^{\hat{\boldsymbol{u}},(T)}(\boldsymbol{\lambda})$ decreases to zero at the rate of $C B_{T}^{\frac{k-1}{2}}$, which is slower than $\sqrt{T}$.
\begin{corollary}\mbox{}In case equation (\ref{eq:Eq2}) describes a fundamental, structural VAR, then
\begin{enumerate}
\item[(i)] ${g}_{\boldsymbol{c},k}^{\boldsymbol{u},(T)}(\boldsymbol{\lambda}) = {\boldsymbol{\kappa}}^{(T)}_{k}(\boldsymbol{u})$;
\item[(ii)] $\mathbb{E}\sqrt{H_{T}^{k-1}T}\left| \hat{g}_{\boldsymbol{c},k}^{\boldsymbol{u},(T)}(\boldsymbol{\lambda}) - \hat{g}_{\boldsymbol{c},k}^{\hat{\boldsymbol{u}},(T)}(\boldsymbol{\lambda})\right| \leq C H_{T}^{\frac{k-1}{2}}$;
\item[(iii)] $\mathbb{E}\sqrt{T}\left| \hat{\boldsymbol{\kappa}}^{(T)}_{k}(\boldsymbol{u}) - \hat{\boldsymbol{\kappa}}^{(T)}_{k}(\hat{\boldsymbol{u}})\right| \leq C {T}^{-\frac{1}{2}}$
\end{enumerate}
\end{corollary}
where ${\boldsymbol{\kappa}}^{(T)}_{k}(\boldsymbol{u})$ and $\hat{\boldsymbol{\kappa}}^{(T)}_{k}(\boldsymbol{u})$ are the population and sample vectorized cumulant of $k$-th order for reduced-form residuals, $\boldsymbol{u}_{t}$. This corollary states that, when the true model is a fundamental SVARMA model, instead of using spectral density, the contemporaneous cumulants can be employed and its rate of asymptotic equivalence becomes the standard one, $\sqrt{T}$.
\section{Simulation Study}
Here we present the results from different Monte-Carlo exercises. The true SVARMA model is a non-causal SVAR($1$). We employ different distributions for the structural disturbances. Monte-Carlo and bootstrap repetitions are set to $m=250$ and $B=500$, respectively. The nominal significance level is set to $\alpha=5\%$. The number of points for the DFT (called $\texttt{nfft}$) is set to be $2^{n}$ such that it is close to $T/2$. For instance, if $T=250$, $\texttt{nfft}=128$. Besides, the size of the window is $\texttt{nfft}/2$.
\subsection{Simulation Results}
Table (\ref{tab:Table123_SVARMA10_2D}) shows the rejection rates for the sequence of null hypotheses $\{H_{0,r_{s}}\}_{s=0}^{d-1}$ when $\boldsymbol{y}_{t}$ was generated by a non-causal VAR($1$) of dimension $2$. For computational time issues, when null hypothesis is not reject for $H_{0,s}$, we impose that $H_{0,s^{\prime}}$ is not reject as well for $s<s^{\prime}\leq d-1$. At panel (\ref{tab:Table1_SVARMA10_2D}), we employ a singular frequency, $\lambda_{3}=0$. The first row is associated to the case when structural shocks follows a joint Gaussian distribution. Hence, the rejection rate of $H_{0,0}$ is the size of the test in this case, which is $4\%$. If we consider a mixed case for disturbances (second row of panel (\ref{tab:Table1_SVARMA10_2D})), the rejection rate of $H_{0,0}$ represents the power, $32\%$; while the rejection rate of $H_{0,1}$ represents the size, $8\%$. Finally, when structural shocks are fully non-Gaussian distributed (third row of panel (\ref{tab:Table1_SVARMA10_2D})), rejection rates of $H_{0,0}$ and $H_{0,1}$ represents the power at each step. These values are $42\%$ and $26\%$, respectively.
\input{./Tables/Tables_123_SVAR_2D_LongBoot.tex}
Estimation of higher order spectrum at an specific frequency can be a bit imprecise, specially when sample size is relatively small. In order to make the procedure robust, we compute construct our matrix of interest at a finite grid of frequencies. In this case, $\boldsymbol{\Pi}=\sum_{{\lambda}_{k}\in\mathcal{G}}{\text{Re}\left(G_{k}^{2}({\lambda_{k}})\right)}$, where $\mathcal{G}$ denotes a grid of frequencies. Notice that this new matrix is a particular case of the following $\int_{\lambda_{k}\in[-\pi,\pi]}{w(\lambda_{k}){\text{Re}\left(G_{k}^{2}({\lambda_{k}})\right)}d\lambda_{k}}$, which continues being positive semi-definite, real-valued, square, symmetric matrix. Besides, it is worth to mention that if we employed $\text{Re}\left(G_{k}({\lambda_{k}})\right)$, the rank of $\text{Re}\left(G_{k}({\lambda_{k}})\right)$ identifies the number of non-Gaussian structural shocks; however, the rank of $\int_{\lambda_{k}\in[-\pi,\pi]}{w(\lambda_{k}){\text{Re}\left(G_{k}({\lambda_{k}})\right)}d\lambda_{k}}$ not.
In panel (\ref{tab:Table2_SVARMA10_2D}), a grid of size $11$ is selected from the range $[0,\pi/8]$. For this grid, the size of the test when all disturbances are Gaussian distributed is $4\%$; while, when have one Gaussian shock, the size is around $9\%$. The power is quite similar to case when a single frequency is employed.
In panel (\ref{tab:Table3_SVARMA10_2D}), a grid of $26$ frequencies in the range of $[0,3\pi/16]$ was selected. The size when $\boldsymbol{\varepsilon}_{t}$ is a Gaussian process is $5\%$; while when we have the mixed case, the size is around $10\%$. The power when $\boldsymbol{\varepsilon}_{t}$ has one Gaussian component grows in $5$ percentage points in comparison to the situation of a single frequency. Besides, the power when $\boldsymbol{\varepsilon}_{t}$ is fully non-Gaussian distributed increases significantly with respect to the case of a single frequency.
\input{./Tables/Tables_123_SVAR_2D_LongBoot_T500.tex}
\citet{guay2021identification} employs an alternative method for constructing the bootstrap sample under the null hypothesis. His procedure consists on taking the columns of $\boldsymbol{V}$ associated to the first $r_{s}$ singular values, which under $H_{0,r_{s}}$ they are the only ones different from zero, and project the standardized reduced-form residuals on the space generated by such columns. The rest components $d-r_{s}$ are drawn from multivariate normal distribution with mean zero and variance $\boldsymbol{I}_{d-r_{s}}$. In Table (\ref{tab:Table123_SVARMA10_2D_Guay}) we compute the rejection rates using Guay's bootstrap sampling approach. The size of the test when having fully Gaussian shocks is, disregarding the length of the frequency grid, around $5\%$. The size when $\boldsymbol{\varepsilon}_{t}$ has one Gaussian component is around $7\%$ for the different grid lengths. Nonetheless, the power using \citeauthor{nordhausen2017asymptotic} bootstrap sampling is, consistently, higher than using Guay's approach.
\input{./Tables/Tables_123_SVAR_2D_LongBoot_GuayMethod.tex}
For larger dimensional models, we choose $d=3,4$. We perform these exercises with sample size $T=250$ and the grid $\lambda_{3}=0$. The results can be observed at Table (\ref{tab:Table123_SVARMA10_4D}). The distribution $MN_{1}$ is a mixture of two normal distributions ($\mathcal{N}(10,0.75)$ and $\mathcal{N}(-2,4)$) such that the distribution has a positive skewness coefficient. It can be noted that, while the sample size is fixed, it's more difficult to reject null hypothesis for larger $r_{k}$.
\input{./Tables/Tables_123_SVAR_4D_LongBoot.tex}
\vspace*{-0.5cm}
\subsection{Empirical Application}
\label{sec:EmpiricApp}
We select two well-known data-sets in empirical macroeconomics. The first data-set is used in the seminal work of \citeauthor{blanchard1989dynamic}, where they introduced the long-run restrictions as a identification scheme for causal structural VAR models. This data-base consists on two endogenous variables, the growth of US GNP and unemployment rate. The second data-set is taken from the work of \citet{blanchard2002empirical}. Their data includes three endogenous variables such as tax revenues, government spending and GDP (all of them in real terms). In both cases, we use the same SVAR specification of each work, avoiding our conclusions to be affected by a specification bias.
Table (\ref{tab:ResultApplication_BQ}) shows the results of applying our proposal to Blanchard and Quah database. Using only third order cumulant spectral density, we can identify a single asymmetric structural shock. Using the fourth order spectral density, we cannot identify significantly none component with excess of kurtosis. which also may be affected by the difficulty of estimating fourth-order spectrum with such small sample size ($T=148$). Thus, if the true model followed fundamental SVARMA representation, we can apply the statistical identification approach to this data-set.
\begin{table}[!tbh]
\centering
\caption{Number of non-Gaussian components in \citet{blanchard1989dynamic}}
\label{tab:ResultApplication_BQ}
\begin{minipage}[c][0.25\textheight][t]{0.5\textwidth}
\begin{center}
\subcaption{$\rank\left( \hat{G}_{3}^{2,T}(\lambda_{3}) \right)$}
\begin{tabular}{lccc}
\\[-1.8ex]\toprule
\toprule \\[-1.8ex]
Frequency & & \multicolumn{2}{c}{Null Hypotheses} \\
\cmidrule{3-4} Grid & & $H_{0,0}$ & $H_{0,1}$ \\
[0.5ex]\midrule \\[-1.8ex]
& & & \\
\multirow{1}{*}{$\lambda_{3}=0$} & & $0.022$ & $0.409$ \\
\multirow{1}{*}{Grid of $11$ freqs.} & & $0.097$ & $0.386$ \\
\multirow{1}{*}{Grid of $26$ freqs.} & & $0.064$ & $0.607$ \\
\bottomrule
\bottomrule \\[-1.8ex]
\multicolumn{4}{l}{\scriptsize P-values reported, based on $B=1000$ bootstrap samples.}
\end{tabular}
\par\end{center}
\end{minipage}\hfill{}%
\begin{minipage}[c][0.25\textheight][t]{0.5\textwidth}
\begin{center}
\subcaption{$\rank\left( \hat{G}_{4}^{2,T}(\lambda_{4}) \right)$}
\begin{tabular}{lccc}
\\[-1.8ex]\toprule
\toprule \\[-1.8ex]
Frequency & & \multicolumn{2}{c}{Null Hypotheses} \\
\cmidrule{3-4} Grid & & $H_{0,0}$ & $H_{0,1}$ \\
[0.5ex]\midrule \\[-1.8ex]
& & & \\
\multirow{1}{*}{$\lambda_{4}=0$} & & $0.477$ & $--$ \\
\multirow{1}{*}{Grid of $11$ freqs.} & & $0.896$ & $--$ \\
\multirow{1}{*}{Grid of $26$ freqs.} & & $0.792$ & $--$ \\
\bottomrule
\bottomrule \\[-1.8ex]
\multicolumn{4}{l}{\scriptsize P-values reported, based on $B=500$ bootstrap samples.}
\end{tabular}
\par\end{center}
\end{minipage}
\end{table}
Using \citet{blanchard2002empirical} dataset ($T=246$), unlike \citet{guay2021identification}, we find a skewed structural shock and, at least, two non-mesokurtic structural errors in the trivariate system; implying that, in case of fundamentalness, statistical identification strategy can be applied.
\begin{table}[!tbh]
\centering
\caption{Number of non-Gaussian components in \citet{blanchard2002empirical}}
\label{tab:ResultApplication_BP}
\begin{minipage}[c][0.467\textheight][t]{0.5\textwidth}
\begin{center}
\subcaption{$\rank\left( \hat{G}_{3}^{2,T}(\lambda_{3}) \right)$}
\begin{tabular}{lcccc}
\\[-1.8ex]\toprule
\toprule \\[-1.8ex]
Frequency & & \multicolumn{2}{c}{Null Hypotheses} \\
\cmidrule{3-4} Grid & & $H_{0,0}$ & $H_{0,1}$ & $H_{0,2}$ \\
[0.5ex]\midrule \\[-1.8ex]
& & & & \\
\multirow{1}{*}{$\lambda_{3}=0$} & & $0.225$ & $--$ & $--$ \\
\multirow{1}{*}{Grid of $11$ freqs.} & & $0.078$ & $0.741$ & $--$ \\
\multirow{1}{*}{Grid of $26$ freqs.} & & $0.068$ & $0.224$ & $--$ \\
\bottomrule
\bottomrule \\[-1.8ex]
\multicolumn{4}{l}{\scriptsize P-values reported, based on $B=1000$ bootstrap samples.}
\end{tabular}
\par\end{center}
\begin{center}
\subcaption{$\rank\left( \hat{G}_{4}^{2,T}(\lambda_{4}) \right)$}
\begin{tabular}{lcccc}
\\[-1.8ex]\toprule
\toprule \\[-1.8ex]
Frequency & & \multicolumn{3}{c}{Null Hypotheses} \\
\cmidrule{3-5} Grid & & $H_{0,0}$ & $H_{0,1}$ & $H_{0,2}$ \\
[0.5ex]\midrule \\[-1.8ex]
& & & \\
\multirow{1}{*}{$\lambda_{4}=0$} & & $0.004$ & $0.039$ & $0.088$ \\
\multirow{1}{*}{Grid of $11$ freqs.} & & $0.004$ & $0.004$ & $0.143$ \\
\multirow{1}{*}{Grid of $26$ freqs.} & & $0.004$ & $0.004$ & $0.542$ \\
\bottomrule
\bottomrule \\[-1.8ex]
\multicolumn{4}{l}{\scriptsize P-values reported, based on $B=250$ bootstrap samples.}
\end{tabular}
\par\end{center}
\end{minipage}
\end{table}
\section{Conclusion}
The aim of this paper is to design a testing procedure for verifying empirically one of the main conditions employed by statistical identification scheme for structural VARMA models. As discussed by several recent works, SVARMA models are identified, dynamic and statically, when structural shocks are full independent and all of its components follows a non-Gaussian distributed. Our first contribution is that prior knowledge on the fundamentalness of the model is not needed for performing the testing procedure, contrary to what is required by different current approaches. Instead of using contemporaneous third and fourth order cumulants, we employ third and fourth order spectral density of reduced-form errors.
Our simulation results show that power of the test increases with sample size, as expected, but also if a grid of frequencies is employed rather than a single frequency. The two empirical application exercises show that: we detect a non-Gaussian structural error in \citet{blanchard1989dynamic} dataset, and at least two non-Gaussian structural shocks in \citet{blanchard2002empirical}. This may imply that a fundamental SVAR or SVARMA model could be statistical identified without further restrictions. Besides, our results suggest identification of a possibly non-fundamental SVARMA model, and the method developed by \citet{gourieroux2020identification,velasco2022identification} can be applied.
\bibliographystyle{apalike}
|
\section{Introduction}
In this paper, we address multivariate Hermite interpolation of a function that takes values on a complete Riemannian manifold $\mcM$
with tangent bundle $T\mcM$. More precisely, let $D\subset \R^d$ be a parameter domain and consider a differentiable function
\begin{equation*}
f\colon D \to \mcM, \quad \omega\mapsto f(\omega).
\end{equation*}
Consider a sample data set consisting of $k$ parameter locations $\omega_1,\ldots, \omega_k\in D$ with corresponding function values (manifold locations)
and partial derivatives (tangent vectors)
\begin{equation*}
p_j = f(\omega_j)\in\mcM,
\quad
v_j^i\coloneqq \partial_i f(\omega_j) = \deriv\Big\vert_{t=0} f(\omega_j + te_i)
= \diff f(\omega_j)[e_i] \in T_{f(\omega_j)}\mcM,
\end{equation*}
where $e_1,\ldots,e_d$ denotes the orthonormal basis of unit vectors in $\R^d$.
The Hermite manifold interpolation problem is formalized as follows:
\\
Find a differentiable, manifold-valued function $\hat f\colon D\to \mcM$ such that
\begin{subequations}
\begin{align}
\hat f(\omega_j) &= p_j\in \mcM, \qquad j =1,\ldots,k
\label{eq:basicHermiteMnf1}
\\
\partial_i \hat f(\omega_j) &= v_j^i\in T_{p_j}\mcM, \qquad j=1,\ldots,k; \hspace{0.2cm} i=1,\ldots,d.
\label{eq:basicHermiteMnf2}
\end{align}
\end{subequations}
\subsection{Original contributions}
We develop two approaches to tackle the multivariate manifold Hermite interpolation problem \eqref{eq:basicHermiteMnf1}, \eqref{eq:basicHermiteMnf2}.
The first one is via computing weighted Riemannian barycenters.
This can be considered as an extension of the interpolation approach developed in \cite{Grohs:2015} and in \cite{Sander:2012, Sander:2016}
by incorporating derivative data. We refer to this as barycentric Hermite interpolation \textbf{(BHI)}.
The BHI method has the following main features:
\begin{enumerate}[(i)]
\item The approach works on arbitrary Riemannian manifolds, i.e., no special structure
(Lie Group, homogeneous space, symmetric space,...) is required.
In order to conduct practical computations, only an algorithm for evaluating the Riemannian logarithm map must be available.
All occurrences of Riemannian exponentials can be replaced with retractions and analogously any Riemannian logarithms by inverse retractions \cite[Section 4.1]{AbsilMahonySepulchre:2008}.
For clarity of presentation we just write exponential and logarithmic map, respectively, during this paper.
%
%
\item The differentiability order of the interpolant is the same as that of the weight functions. Hence, when working with smooth weight functions, a smooth interpolant is produced.
%
%
\item Although our theoretical derivation involves covariant derivatives and the Hessian forms of scalar fields on manifolds, the practical implementation of BHI does not require computing any such operators.
\item The method works only locally, on a domain, where the Riemannian center of mass exists and is unique. The number of sample points must exceed the dimension of the manifold.
\end{enumerate}
The second approach is a straightforward translation of Hermite interpolation in Euclidean vector spaces to a selected tangent space of the manifold under consideration.
We refer to this approach as tangent space Hermite interpolation \textbf{(THI)}.
The THI method has the following main features:
\begin{enumerate}[(i)]
\item The approach works on arbitrary Riemannian manifolds, i.e., no special structure
(Lie Group, homogeneous space, symmetric space,...) is required.
In order to conduct practical computations, algorithms for evaluating the Riemannian exponential map and the Riemannian logarithm map must be available (or must be consistently replaced with invertible retractions).
The differential of the Riemannian logarithm must be evaluated (or approximated).
\item The differentiablility order of the interpolant is the same as that of the weight functions.
%
\item The method works only locally, on a star-shaped domain around a manifold location $p_c\in \mcM$, where the Riemannian normal coordinates are well-defined.
The interpolant is constructed in the tangent space $T_{p_c}\mcM$ and thus depends on the choice of the center point $p_c$.
\end{enumerate}
We illustrate the validity and the performance of both approaches by means of numerical examples.
\subsection{Related work}
In the research literature, there are two different tracks of manifold interpolation research.
On the one hand, there is the problem of interpolating scalar functions with manifold-valued \emph{inputs} $f\colon\mcM \to \R$.
This line of research is followed, e.g., in \cite{Allasia:2018, Narcowich:1995}, but is not considered here.
On the other hand, interpolating parametric functions with manifold-valued \emph{outputs} $f\colon\R^d \to \mcM$ is investigated.
The work at hand subordinates to this setting.
Confusion may be caused by the fact that for both types of problems, one speaks of “interpolation \emph{on} a manifold”.\\
To the best of our knowledge, \emph{multivariate} Hermite interpolation for output data
on general manifolds has not yet been considered in the research literature.
Univariate Hermite interpolation has been considered explicitly in \cite{Jakubiak:2006}
for data on compact, connected Lie groups with a bi-invariant metric.
A general approach to Hermite curve interpolation is featured in \cite{Zimmermann:2020}.\\
A related line of research is the generalization of Bézier curves and the De Casteljau-algorithm
to Riemannian manifolds, see \cite{BergmannGousenbourger:2018,GouseMassartAbsil:2018, Polthier:2013, Noakes:2007, SAMIR:2019}.
The transition of this technique to manifolds is via replacing the inherent straight lines with geodesics.
The start and end velocities of the resulting spline are proportional to the velocity vectors of
the geodesics that connect the first two and the last two control points, respectively (\cite[Theorem 1]{Noakes:2007}).
Hence, the method can be adapted to work for univariate Hermite interpolation.
More general subdivision schemes based on similar geodesic averages have been investigated in \cite{Dyn:2017, Wallner:2005}.
The preprint \cite{BenZion:2022} addresses univariate Hermite interpolation on the sphere via subdivision schemes.
Two-variate manifold interpolation via Bézier surfaces with $C^1$-connection of local patches has been investigated in \cite{AbsGouseWirth:2016}.\\
A multivariate Hermite-type method that is specifically tailored to interpolation problems on the Grassmann manifold is sketched in \cite[\S 3.7.4]{Amsallem:2010}.\\
Interpolation with Riemannian barycenters has been proposed in the context of geo\-desic finite elements in
\cite{Grohs:2015} and in \cite{Sander:2012, Sander:2016}.\\
An overview over manifold interpolation methods with a special focus on model reduction applications is given in \cite{Zimmermann:2021}.
\subsection{Notational conventions}
Throughout, we assume that $\mcM$ is a complete Riemannian manifold of dimension $\dim(\mcM)= m$.
The tangent space of $\mcM$ at $p\in \mcM$ is $T_p\mcM$ and the tangent bundle is
$T\mcM= \cup_{p\in\mcM}T_p\mcM$.
As a Riemannian manifold, $\mcM$ carries a family of inner products
$\left\{\langle \cdot,\cdot \rangle_{p}\colon T_p\mcM\times T_p\mcM\to \R, (v,w)\mapsto \langle v,w \rangle_{p}\right\}_{p\in\mcM}$
that depends smoothly on the base point $p\in\mcM$.\\
The differential of a function $f\colon \R^d\to \R^m$ at a point $p$ is denoted by
$\Diff{}f(p)\colon \R^d \to \R^m$. For a two-arguments function $L\colon\R^m\times \R^d\to \R^m, (q,\omega) \mapsto L(q,\omega)$,
the linear maps
$\Diff_q L(q,\omega)\colon \R^m\to \R^m$ and $\Diff_\omega L(q,\omega)\colon \R^d\to \R^m$ denote the differentials of
$q\mapsto L(q,\omega)$ (with $\omega$ considered as fixed) and
$\omega \mapsto L(q,\omega)$ (with $q$ considered as fixed), respectively.
The differential at a point $p$ of a function $f\colon\mcM \to \mcN$ between differentiable manifolds $\mcM,\mcN$ is denoted by
$\diff{}f(p)\colon T_p\mcM \to T_{f(p)}\mcN$.
For $L\colon\mcM\times \R^d\to \mcM, (q, \omega) \mapsto L(q,\omega)$, the notations $\diff{}_qL(q,\omega)$ and $\diff{}_\omega L(q,\omega)$ are to be understood in analogy to the above.
\\
A vector field is a smooth function $X\colon\mcM\to T\mcM$ such that $X(p)\in T_p\mcM$ for all $p\in\mcM$.
The set of smooth vector fields on $\mcM$ is denoted by $\mathcal{X}(\mcM)$.
We use the symbol $\nabla$ to denote the (unique) Levi-Civita connection on $\mcM$. For $X,Y\in\mathcal{X}(\mcM)$, the notation $\nabla_XY\in \mathcal{X}(\mcM)$ denotes the covariant derivative of the vector field $Y$ in the direction of the vector field $X$. A vector field along a curve $c\colon I\to \mcM$ is a mapping
$X_c\colon I \to T\mcM$ such that $X_c(t) \in T_{c(t)}\mcM$. The covariant derivative of $X_c$ along $c$ is $\covDeriv[t]X_c$.
If there is an ambient vector field $X\in\mathcal{X}(\mcM)$ such that $X(c(t)) = X_c(t)$, then $\covDeriv[t]X_c = \nabla_{\dot c(t)} X$.
For a scalar function $g\colon\mcM\to \R$, the gradient vector field $\grad g\in \mathcal{X}(\mcM)$ is point-wise defined by
$\langle \grad g(p), v\rangle_p = \diff{}g(p)[v]$ for all $v\in T_p\mcM$.
The covariant derivative of the gradient field yields the Hessian
$\Hess g [X] = \nabla_X \grad g$
and gives rise to the endomorphism
$T_p\mcM \ni v\mapsto \Hess g(p)[v] = (\nabla_v \grad g)(p)\in T_p\mcM$.\\
For a scalar two-parameter function $L\colon\mcM\times \R^d, (q,\omega)\mapsto L(q,\omega)\in\R$, we will write $\grad_q L(p,\omega)$ for the gradient by $q$ evaluated at $(p,\omega)$, which is defined by
$
\langle \grad_q L(p,\omega), v\rangle_p = \diff{}_q L(p,\omega)[v]$, for all $v\in T_p\mcM$.
Likewise, $\Hess_q L(p,\omega)[v]$ denotes the Hessian of $L(q,\omega)$ by $q$ evaluated at $(p,\omega)$ applied to the tangent vector $v\in T_p\mcM$.
\subsection{Organization of the paper}
Section \ref{sec:RiemannBary} recaps the barycentric interpolation methods. In Section \ref{sec:baryHermite}, the approach is extended to the Hermite setting.
The alternative approach of tangent space Hermite interpolation is outlined in Section \ref{sec:THI}.
Numerical experiments are featured in Section \ref{sec:numex}, and Section \ref{sec:conclusions} concludes the paper.
\section{Interpolation with weighted Riemannian barycenters}
\label{sec:RiemannBary}
The Riemannian barycenter or Riemannian center of mass\footnote{Here, we introduce Riemannian barycenter for discrete data sets; for centers w.r.t.\ a general mass distribution, see Karcher's original paper \cite{Karcher:1977}, Section 1.} or Fr{\'e}chet mean of a sample data set
$\{p_1,\ldots,p_k\}\subset \mcM$ on a manifold
is defined as the minimizer of the Riemannian objective function
\begin{equation*}
\mcM\ni q \mapsto L(q) = \frac{1}{2} \sum_{j=1}^k w_j \dist(q, p_j)^2,
\end{equation*}
where $\dist(q,p_j)$ is the Riemannian distance between the manifold locations $q$ and $p_j$ and $w_j\geq 0$ are scalar weights such that $\sum_{j=1}^k w_j = 1$.
Formally, the latter requirements mean that the center of mass is taken with respect to a discrete positive measure of unit weight.
This definition generalizes the notion of the barycentric mean in Euclidean spaces, cf. \Cref{app:EuclideanCase} of the supplements.
However, on curved manifolds, the global center might not be unique. Moreover,
local minimizers may appear. For more details, see \cite{Karcher:1977} and \cite{AfsariTronVidal:2013}, which also give uniqueness criteria.
\subsection{Interpolation via optimization}
Interpolation can be performed by computing weighted Riemannian centers.
More precisely, let $f\colon\R^d\supset D \to \mcM$ and
let $\omega_1,\ldots,\omega_k\subset D$ be a set of parameter locations and let
$p_j=f(\omega_j)\in \mcM$, $j=1,\ldots,k$ be the corresponding sampled manifold locations on $\mcM$.
The interpolant is then defined on the convex hull $\text{conv}\{\omega_1,\ldots,\omega_k\}\subset D\subset \R^d$ of the samples.\\
Let $\{\varphi_j:\omega\mapsto \varphi_j(\omega)\in \R\mid j=1,\ldots,k\}$ be a suitable set of multivariate, scalar-valued interpolation weight functions with $\varphi_l(\omega_j) = \delta_{lj}$ and $ \sum_{j=1}^k \varphi_j(\omega) \equiv 1$.
Such weight functions can be constructed, e.g., as Lagrangians, \cite{Sander:2016}, or radial basis functions, \cite{Buhmann:2003}.
The interpolant $q^* \approx f(\omega^*)\in \mcM$ at an unsampled parameter
location $\omega^*\in \text{conv}\{\omega_1,\ldots,\omega_k\}$ can be taken to be the minimizer
\begin{equation}
\label{eq:baryInterp}
\omega^* \mapsto f(\omega^*) = q^* \coloneqq \argmin_{q\in \mcM} L(q, \omega^*),
\quad \text{where } L(q, \omega) \coloneqq \frac12 \sum_{j=1}^k \varphi_j(\omega) \dist(q,p_j)^2.
\end{equation}
Since the weight functions may attain negative values, this corresponds to taking the center of mass with respect to a discrete {\em signed} measure of unit weight, see the discussion in \cite[Section 3]{Sander:2016}.
Computing $q^*$ thus requires one to solve a Riemannian optimization problem.
At sample location $\omega_l$, one has indeed that
\begin{equation*}
2L(q, \omega_l) =
\sum_{j=1}^k \varphi_j(\omega_l) \dist(q,p_j)^2
=\sum_{j=1}^k \delta_{lj} \dist(q,p_j)^2= \dist(q,p_l)^2,
\end{equation*}
which has the unique global minimum at $q^* = p_l$.
Hence, the function $\hat f\colon \omega \mapsto \argmin_{q\in \mcM} L(q, \omega)$ satisfies the basic interpolation conditions.
\begin{remark}
Under certain conditions that locally ensure the existence and uniqueness of Riemannian barycenters,
the minima of \eqref{eq:baryInterp} are exactly at the zeros of the associated gradient.
Moreover, they depend smoothly on the parameters $(q, \omega)$ if the weight functions $\varphi_j$ are smooth,
see \cite[Theorems 3.19 \& 4.1]{Sander:2016}.
%
As a consequence, the interpolant is smooth under these conditions.
\end{remark}
\subsection{The gradient and Hessian of the Riemannian distance function}
Interpolation via barycenters is an optimization task.
As a rule, numerical optimization requires the computation of the gradient of the objective function.
The next theorem, due to Karcher, provides the gradient of the squared
Riemannian distance function.
\begin{theorem}[{\cite[Thm. 1.2]{Karcher:1977}}]
\label{thm:grad_dist}
Let $\mcM$ be a complete Riemannian manifold and let $p\in \mcM$.
Let $B_\rho(p)$ be a geodesic ball of radius $\rho$ around $p$ such that the geodesics between any two points inside $B_\rho(p)$ are unique and minimizing.
Define
\begin{equation*}
L_p\colon B_\rho(p)\to \R, \quad q \mapsto \frac12 \dist(q,p)^2.
\end{equation*}
Then
\begin{equation*}
\grad L_p(q) = -\Log_q(p),
\end{equation*}
where $\Log_q = (\Exp_q)^{-1}$ is the Riemannian logarithm map.
\end{theorem}
Because Karcher works with barycenters with respect to mass distributions and also to make this exposition self-contained, we recap the proof in \Cref{app:proofs}.\\
Eventually, we also need information on the Hessian of $L_p(q) = \frac12 \dist(q,p)^2$ at $q=p$.
The Riemannian Hesse form of a scalar function $g\colon\mcM\to \R$ at $q$ is
\begin{equation}
\label{eq:Hesse_def}
\Hess g(q)\colon T_q\mcM\to T_q\mcM, \quad v\mapsto \Hess g(q)[v] = (\nabla_v \grad g)(q),
\end{equation}
see \cite[\S 6A]{Kuhnel:2015}.
\begin{remark} Actually, the Riemannian Hesse $(1,1)$-tensor maps vector fields to vector fields,
\begin{equation*}
\Hess g\colon \mathcal{X}(\mcM) \to \mathcal{X}(\mcM),
\quad X \mapsto \nabla_X \grad g.
\end{equation*}
Yet, since a tensor is a point-wise object, $(\nabla_v \grad g)(q) = (\nabla_X \grad g)(q)$ for all vector fields $X\in\mathcal{X}(\mcM)$ with $X(q) =v$.
Hence, it makes sense to consider the Hessian at $q$ as an endomorphism of $T_q\mcM$.
Additional background information on the Riemannian Hessian is given in \Cref{app:proofs} of the supplements.
\end{remark}
\begin{theorem}
\label{thm:Hesse_id}
Consider the setting of Theorem \ref{thm:grad_dist}.
For $p$ fixed, the Hesse form of the function $q\mapsto L_p(q) = \frac12 \dist(q,p)^2$ at $p$ is the identity,
\begin{equation*}
\Hess L_p(p)= \id_{T_p\mcM}\colon T_p\mcM \to T_p\mcM.
\end{equation*}
\end{theorem}
\begin{proof}
We use Karcher's approach of computing the Hessian
via a variation through geodesics, see \Cref{thm:Karcher1977}
and \Cref{fig:GeoVar} in the supplement.
This shows that for a geodesic $t\mapsto \gamma(t)$ with $\gamma(0) = q$, $\dot \gamma(0) = v$
and the associated variation of geodesics
\begin{equation*}
(s,t)\mapsto c_p(s,t) = \Exp_p(s\Log_p(\gamma(t))),
\end{equation*}
it holds
\begin{equation*}
\langle \Hess L_p(q)[v]\ ,\ v\rangle =
\langle\covDeriv[s] \partial_t c_p(1,0)\ ,\ v \rangle.
\end{equation*}
In the special case, where the geodesic $\gamma$ starts from $q=p$ with velocity $v\in T_p\mcM$,
we obtain
\begin{align*}
\partial_t\big\vert_{t=0} c_p(s,t) &=
\diff\left(\Exp_p\right)_{s\Log_p(\gamma(0))}
\left(s \diff\left(\Log_p\right)_{\gamma(0)}(\dot \gamma (0))\right)
\\
&= \diff\left(\Exp_p\right)_{s\cdot 0}
\left(s \diff\left(\Log_p\right)_{p}(v)\right)
= sv \in T_p\mcM,
\end{align*}
because $\diff{}(\Exp_p)_0=\id_{T_p\mcM} = \diff{}(\Log_p)_p$.
Note that $s\mapsto \partial_t c_p(s,0)\in T_p\mcM$ is a vector field along the point curve
$\alpha\colon s\mapsto \alpha(s)\equiv p$.
Therefore, the covariant derivative coincides with the usual derivative and we obtain
\begin{equation*}
\covDeriv[s] \partial_t c_p(s,0) = \deriv[s](sv) = v.
\end{equation*}
As a consequence,
\begin{equation*}
\langle \Hess L_p(p)[v]\hspace{0.1cm},\hspace{0.1cm} v\rangle =
\langle v, v \rangle \quad \text{ for all } v\in T_p\mcM.
\end{equation*}
The Hessian is symmetric. Via polarization, it is uniquely determined by terms of the above form. This yields $\Hess L_p(p)[v] = v$.
\end{proof}
By \Cref{thm:grad_dist}, the gradient of the objective function $L$ in \eqref{eq:baryInterp} by $q$ is
\begin{equation}
\label{eq:baryGrad}
\grad_q L(q, \omega) = \sum_{j=1}^k \varphi_j(\omega)\cdot (-\Log_q(p_j)) \in T_q\mcM.
\end{equation}
At a sample location $p_l$, the gradient (by $q$) of the objective function $L(q,\omega)$ in the barycentric interpolation problem \eqref{eq:baryInterp} vanishes, because of $\Log_{p_l}(p_l) = 0\in T_{p_l}\mcM$.
By \Cref{thm:Hesse_id}, the Hessian (again with respect to the $q$-argument) is
\begin{equation}
\label{eq:Hesse_id_at_p}
\Hess_{q} L(p_l, \omega_l)[v] = \sum_{j=1}^k \varphi_j(\omega_l) \Hess L_{p_j} (p_l)[v]
= \Hess L_{p_l} (p_l)[v] = v \in T_{p_l}\mcM,
\end{equation}
because $\varphi_j(\omega_l)=\delta_{lj}$.
Thus, the Hessian by $q$ at $(p_l, \omega_l)$ is the identity on $T_{p_l}\mcM$. In particular, it has full rank.
A generic gradient descent algorithm\footnote{
There exist several different ways of computing the weighted Riemannian barycenter, see for example
\url{https://juliamanifolds.github.io/Manifolds.jl/latest/features/statistics.html} for an overview.
} to compute the barycentric interpolant
for a function $f\colon\R^d \ni \omega \mapsto f(\omega) \in \mcM$ is given in \Cref{alg:BaryInt}
\begin{algorithm}{Interpolation via the weighted Riemannian barycenter.}
\begin{algorithmic}[1]
\REQUIRE{Sample data set $\{p_1=f(\omega_1),\ldots, p_k=f(\omega_k)\}\subset \mcM$,
unsampled parameter location $\omega^*\in \text{conv}(\omega_1,\ldots, \omega_{k})\subset \R^d$,
initial guess $q_0$, convergence threshold $\tau$}
\STATE{$k\coloneqq0$}
\STATE{Compute $\grad_q L(q_k, \omega^*)$ according to \eqref{eq:baryGrad}}
\WHILE{$\lVert\grad_q L(q_k, \omega^*)\rVert_{q}> \tau$}
\STATE{select a step size $\alpha_k$}
\STATE{$q_{k+1} \coloneqq \Exp^{\mcM}_{q_k}\left(-\alpha_k \grad_q L(q_k, \omega^*)\right)$}
\STATE{$k\coloneqq k+1$}
\ENDWHILE
\ENSURE{$\hat f(\omega^*)\coloneqq q^*\coloneqq q_k\in \mcM$ interpolant of $f(\omega^*)$.}
\end{algorithmic}
\label{alg:BaryInt}
\end{algorithm}
\section{Barycentric Hermite interpolation}
\label{sec:baryHermite}
In this section, we enhance the me\-thod of weighted barycentric interpolation by including derivative information. The task is to construct an interpolant of the form
\begin{equation*}
\hat f\colon \omega \mapsto \hat f(\omega)= \argmin_{q\in \mcM} \frac12 \sum_{j=1}^k \varphi_j(\omega) \dist(q,p_j)^2 = \argmin_{q\in \mcM} L(q,\omega)
\end{equation*}
that satisfies the interpolation conditions \eqref{eq:basicHermiteMnf1}, \eqref{eq:basicHermiteMnf2}.
The requirement to meet the sampled derivatives $v_j^i=\partial_i f(\omega_j)$ entails conditions on the partial derivatives of the weight functions $\varphi_j(\omega)$.
We work under the general assumption that the interpolation procedure takes place on a domain, where the weighted Riemannian barycenters exist and are unique.
For a detailed analysis of interpolation via Riemannian barycenters, we refer to \cite{Sander:2012, Sander:2016} and \cite{Grohs:2015}.
\subsection{Tracking the barycenters via the implicit function theorem}
\label{sec:TechPrep}
Introduce the parametric gradient field
\begin{equation}
\label{eq:objectiveG_VF}
G\colon\mcM \times \R^d \to T\mcM, \quad (q, \omega) \mapsto G(q,\omega) \coloneqq \grad_q L(q,\omega) = -\sum_{j=1}^k \varphi_j(\omega) \Log_q(p_j)
\end{equation}
and note that $G$ is a smooth vector field on the product manifold $\mcM \times \R^d$.
Suppose that $G$ vanishes at $(q^*, \omega^*)$.
Our strategy is to parameterize the zero-sets locally via the implicit function theorem.
By differentiating the corresponding implicit function, we will establish a relation between the derivatives of the interpolation weight functions $\omega\mapsto \varphi_i(\omega)$ of \eqref{eq:baryInterp} and the sampled derivatives \eqref{eq:basicHermiteMnf2}.\\
The implicit function theorem has a close relative, the inverse function theorem,
and both theorems rely on a full-rank condition for a certain differential.
While a manifold version of the inverse function theorem appears in many standard textbooks on differential geometry
(e.g., \cite[Thm 4.5]{Lee:2012}), we were not able to locate a textbook reference for a manifold counterpart to the implicit function theorem.
Yet, because both $\mcM$ and $T\mcM$ are differentiable manifolds, it is straightforward to transfer the classical implicit function theorem,
and often, it is simply taken for granted (as in \cite[Thm. 2.2]{Sander:2012}, \cite[Thm 4.1.]{Sander:2016}). The interesting part is how the rank condition transforms. This is the contents of the following lemma.
\begin{lemma}
\label{lem:Hessefullrank4implicit}
Let $\mcM$ be a Riemannian manifold.
\begin{enumerate}
\item Let $G\colon\mcM \to T\mcM$ be a smooth vector field.
At locations, where $G$ vanishes, the differential of $G$ coincides with the covariant derivative, i.e.,
at $q^*\in \mcM$ with $G(q^*) = 0$, it holds
\begin{equation*}
\diff{}G(q^*)[v] = (\nabla_v G) (q^*), \quad\text{ for all } v\in T_{q^*}\mcM.
\end{equation*}
\item In the special case, where $G$ is the gradient vector field $\grad g\colon\mcM \to T\mcM$ of a scalar function $g\colon\mcM\to \R$,
the Hesse form of $g$ coincides with the differential of $\grad g$ at locations, where the gradient vanishes, i.e.,
at $q^*\in \mcM$ with $\grad g(q^*) = 0$, it holds
\begin{equation*}
\diff{}(\grad g)(q^*)[v] = \Hess g(q^*) [v], \quad\text{ for all }v\in T_{q^*}\mcM.
\end{equation*}
In particular, the differential $\diff{}(\grad g)(q^*)$ has full rank, if the Hesse form $\Hess g(q^*)$ has full rank.
\end{enumerate}
\end{lemma}
\begin{proof}
Let $v\in T_{q^*}\mcM$. We compute $\diff{}G(q^*)[v]$.
Let $\gamma\colon I\to \mcM$ be a smooth curve with $\gamma(0) = q^*, \dot \gamma (0) = v$.
On a suitably small neighborhood $\tilde M$ around $q^*$, construct a local, orthonormal
frame of vector fields $\{E_i\in \mathcal{X}(\mcM)\mid i=1,\ldots, m\}$
as outlined in \cite[Exercise 3.2, p. 24]{Lee:1997}.
This means that we obtain vector fields $E_i :\tilde M \to T\mcM$, $ i=1,\ldots, m$ such that at each $p\in \tilde M$,
$\{E_i(p)\mid i=1,\ldots, m\}$ is an orthonormal basis of $T_p\mcM$, i.e.,
\begin{equation*}
\langle E_i (p), E_j(p) \rangle_{p} = \delta_{ij}\quad \text{ for all } p\in \tilde M.
\end{equation*}
W.l.o.g., assume that the image of $\gamma$ is contained in $\tilde M$.
We express the vector field $G$ along $\gamma$ in terms of the orthonormal frame
\begin{equation*}
(G \circ \gamma) (t) =
\sum_i \langle (G \circ \gamma) (t), (E_i\circ \gamma)(t) \rangle_{\gamma(t)} E_i(\gamma(t)),
\end{equation*}
and differentiate according to the product rule (\cite[Chapter 2, Prop. 3.2]{DoCarmo:1992})
\begin{align*}
\diff{}G(q^*)[v] &= \deriv[t]\big\vert_{t=0} (G \circ \gamma) (t)\\
&= \sum_i \Bigl[ \deriv[t]\big\vert_{t=0}\left( \langle (G \circ \gamma) (t), E_i(\gamma(t)) \rangle_{\gamma(t)}\right) E_i(q^*)\\
& \hspace{0.5cm} + \langle
G (q^*)
, E_i(q^*) \rangle_{q^*} \deriv[t]\big\vert_{t=0} (E_i\circ \gamma)(t)\Bigr]\\
&= \sum_i \left( \langle \frac{\mathrm{D}(G \circ \gamma)}{\mathrm{d}t}(0), E_i(q^*) \rangle_{q^*} +
\langle
G (q^*)
, \frac{\mathrm{D}(E_i\circ \gamma)}{\mathrm{d}t}(0) \rangle_{q^*}
\right) E_i(q^*)\\
&= \sum_i \langle (\nabla_v G) (q^*), E_i(q^*) \rangle_{q^*} E_i(q^*) = (\nabla_v G) (q^*).
\end{align*}
Note that we used twice that $G(q^*) = 0$.
If $G$ is the gradient vector field of a scalar function $G = \grad g$, then the latter becomes the Hessian,
$(\nabla_v \grad g) (q^*) = \Hess g(q^*) [v]$.
\end{proof}
By Lemma \ref{lem:Hessefullrank4implicit}, the condition on the rank of the differential
of a vector field translates to a condition on the covariant derivative (at points, where the vector field vanishes). This makes it straightforward to transfer any of the classical proofs of the implicit function theorem to the manifold setting.
\begin{theorem}[An implicit function theorem for vector fields]
\label{thm:ImplFunVF}
Let $\mcM$ be an $m$-di\-men\-sio\-nal Rie\-mannian manifold and let $G\colon\mcM\times \R^d \to T\mcM, (q,\omega)\mapsto G(q,\omega)\in T_q\mcM$ be a smooth mapping such that both
$q\mapsto G(q, \omega)\in T_q\mcM$ with $\omega$ considered as fixed and
$\omega \mapsto G(q,\omega)\in T_q\mcM$ with $q$ considered as fixed are smooth vector fields.
Let $(q^*, \omega^*)\in \mcM\times \R^d$ be a point such that $G (q^*, \omega^*) = 0\in T_{q^*}\mcM$
and let $q\mapsto G^*(q) \coloneqq G(q,\omega^*)$.
Assume that the linear form $(\nabla_{(\cdot)} G^*)(q^*)\colon T_{q^*}\mcM \to T_{q^*}\mcM, v\mapsto (\nabla_{v} G^*)(q^*) $ has full rank $m$.\\
Then there exists an open domain $\Omega\subset\R^d$ around $\omega^*$ and a differentiable function $\hat f\colon\Omega \to \mcM, \omega \mapsto \hat f(\omega)$
such that
\begin{equation}
\label{eq:ImplFunVF}
\Omega\ni \omega \mapsto G(\hat f(\omega), \omega))\equiv 0\in T_{\hat f(\omega)}\mcM.
\end{equation}
\end{theorem}
\begin{proof}
Use the standard trick to obtain the implicit function theorem from the inverse function theorem
see, e.g., \cite[Thm. C.40]{Lee:2012}. To this end, consider the auxiliary function
\[
\gamma\colon\mcM \times \R^d \to T\mcM \times \R^d. \quad (q,\omega) \mapsto
\begin{pmatrix}
\Gamma_1(q,\omega)\\
\Gamma_2(q, \omega)
\end{pmatrix}=
\begin{pmatrix}
G(q,\omega)\\
\omega
\end{pmatrix}.
\]
Note that $\Gamma$ is a smooth map between differentiable product manifolds.
By Lemma \ref{lem:Hessefullrank4implicit}, the differential of the first component function of $\Gamma$ by $q$ at the point $(q^*,\omega^*)$ is given by
$\diff{}_qG(q^*, \omega^*)[v] = (\nabla_{v} G^*)(q^*)$ and thus has full rank by assumption.
The differential of the second component function by $q$ vanishes, while its differential by $\omega$ is the identity on $\R^d$. As a consequence, the differential of $\Gamma$ at $(q^*,\omega^*)$
has full rank and the inverse function theorem yields the existence of a local inverse $\Gamma^{-1}= H = (H_1, H_2)$.
It holds
\[
\begin{pmatrix}
v\\
\omega
\end{pmatrix}
=
\Gamma(H(v, \omega)))
=
\begin{pmatrix}
G\left((H_1(v,\omega), H_2(v,\omega)\right)\\
H_2(v,\omega)
\end{pmatrix}.
\]
In particular, $H_2(v,\omega) = \omega$ and the function $\hat f\colon\omega\mapsto H_1(0, \omega)$ is the sought-after implicit function with
$G(\hat f(\omega), \omega) = G(H_1(0, \omega), H_2(v,\omega)) \equiv 0$ on a local domain, where $\hat f$ is defined.
\end{proof}
An alternative proof that works with local coordinates is given in \cite[Theorem 3.1]{Seguin:2022}. However, this proof omits the technical detail stated in Lemma \ref{lem:Hessefullrank4implicit}.
In the setting of Theorem \ref{thm:ImplFunVF}, implicit differentiation of \eqref{eq:ImplFunVF} along a curve
$c:I\to \Omega$ with $c(0) = \omega, \dot c(0) = w$ yields
\begin{align*}
0&=\deriv[t]\big\vert_{t=0} G(\hat f( c(t)), c(t))
= \left(\diff_qG((\hat f(\omega),\omega)), \diff{}_\omega G((\hat f(\omega),\omega))\right)
\begin{bmatrix}
\diff{}\hat f_{\omega} [w]\\
w
\end{bmatrix}\\
&= \diff_qG(\hat f(\omega),\omega)\left[\diff{}\hat f(\omega) [w]\right]
+ \diff_\omega G(\hat f(\omega),\omega)[w]
\end{align*}
Since $G(\hat f(\omega),\omega) = 0$, Lemma \ref{lem:Hessefullrank4implicit} applies and yields
\begin{equation}
\label{eq:ImplDiff_eq}
\left(\nabla_{d\hat f(\omega) [w]} G(\cdot,\omega)\right) (\hat f(\omega)) = - \diff{}_\omega G(\hat f(\omega),\omega)[w] \in T_{\hat f(\omega)}\mcM.
\end{equation}
Similar considerations were made in \cite[Section 6]{Sander:2016}, but for a different purpose.
\subsection{A relation between the derivatives of the interpolation weight functions and the sampled derivatives}
Now, we apply the results of Subsection \ref{sec:TechPrep} to the specific gradient field $G$ of \eqref{eq:objectiveG_VF}.
Recall that $G$ is the gradient vector field (with respect to $q$) of the scalar function $L(q,\omega) = \frac12 \sum_{j=1}^k \varphi_j(\omega)\dist(q,p_j)^2$
so that $G(q,\omega) \coloneqq \grad_q L(q,\omega) = -\sum_{j=1}^k \varphi_j(\omega) \Log_q(p_j)$.
By construction, $G$ vanishes at the sample locations $q=p_l$, $\omega = \omega_l$, $j=1,\ldots, k$.
Moreover, by Theorem \ref{thm:Hesse_id} and \eqref{eq:Hesse_id_at_p}, the Hessians at the sample locations have full rank.
Theorem \ref{thm:ImplFunVF} yields locally an implicit parameterization of the zero set of the gradient field around each sample point.
Since the sought-after interpolant is determined via an optimization problem, and thus, in turn, via the zero set of the gradient field,
the interpolant coincides with the implicit function $\hat f$. In particular, $\hat f(\omega_l) = p_l$.\\
Fix a sample point $(\omega_l,p_l)$, $l\in 1,\ldots k$.
Let $\{e_i\in\R^d\mid i=1,\ldots,d\}$ denote the Cartesian unit vectors in $\R^d$.
With $w=e_i$, $\omega=\omega_l$, eq. \eqref{eq:ImplDiff_eq} relates the partial derivatives $v^i_l \coloneqq\partial_i \hat f(\omega_l)=\diff{}\hat f(\omega_l)[e_i]$ of $\hat f$ to the partial derivatives of the interpolation weight functions,
\begin{align*}
\left(\nabla_{v^i_l} \grad_q L(p_l,\omega_l) \right) (p_l)
&= - \diff{}_\omega G(p_l,\omega_l)[e_i] \quad \Leftrightarrow\\
\Hess_{q} L(p_l,\omega_l)[v^i_l]
&= v^i_l = \sum_{j=1}^k \partial_i\varphi_j(\omega_l) \Log_{p_l}(p_j)
= \sum_{j=1, j\neq l}^k \partial_i\varphi_j(\omega_l) \Log_{p_l}(p_j).
\end{align*}
This is in direct correspondence with \eqref{eq:implderivEuc} in the Euclidean setting, see \Cref{app:EuclideanCase} in the supplements.
In order to be able to match any prescribed, sampled partial derivatives
$v^i_l = \partial_i \hat f(\omega_l)$, we require the sets
\begin{equation*}
\mathcal L_l \coloneqq \{\Log_{p_l}(p_j)\mid j=1,\ldots,k, j\neq l\}
\subset T_{p_l}\mcM, \quad l=1,\ldots k
\end{equation*}
to span the full tangent space $T_{p_l}\mcM$.
Then any choice of interpolation weight function $\varphi_j$ with $\varphi_j(\omega_l)=\delta_{jl}$ that satisfies
\begin{equation}
\label{eq:HermiteConditionSystem}
v^i_l = \sum_{j=1, j\neq l}^k \partial_i\varphi_j(\omega_l) \Log_{p_l}(p_j)
\end{equation}
yields an interpolant $\hat f$ that satisfies the interpolation conditions \eqref{eq:basicHermiteMnf1}, \eqref{eq:basicHermiteMnf2}.
\begin{remark}
Recall that $\text{dim}(\mcM)=m$.
For every fixed index $l\in \{1,\ldots,k\}$, let $\{E^l_j\mid j=1,\ldots,m\}\subset T_{p_l}\mcM$ be a local orthonormal frame. Then, each of the tangent vectors
$\Log_{p_l}(p_j)\in T_{p_l}\mcM$ features a representation with $m$ coordinate coefficients
\[
\Log_{p_l}(p_j) = x^j_l(1) E^l_1 + \cdots + x^j_l(m) E^l_m.
\]
Likewise for the sampled derivatives,
\[
v_l^i = \alpha_l^i(1) E^l_1 + \cdots + \alpha_l^i(m) E^l_m.
\]
With respect to the coordinates of the local frame, the unknowns in \eqref{eq:HermiteConditionSystem}, i.e., the scalar coefficients $\partial_i\varphi_j(\omega_l)$, are determined by
\begin{equation}
\label{eq:coeffs_in_local_frame}
\begin{pmatrix}
x^1_l(1) &\ldots &x^{l-1}_l(1) &x^{l+1}_l(1)&\ldots & x^k_l(1)\\
\vdots& &\vdots &\vdots & &\vdots\\
x^1_l(m) &\ldots &x^{l-1}_l(m) &x^{l+1}_l(m)&\ldots & x^k_l(m)
\end{pmatrix}
\begin{pmatrix}
\partial_i\varphi_1(\omega_l)\\
\vdots\\
\partial_i\varphi_{l-1}(\omega_l)\\
\partial_i\varphi_{l+1}(\omega_l)\\
\vdots\\
\partial_i\varphi_k(\omega_l)\\
\end{pmatrix}
=
\begin{pmatrix}
\alpha_l^i(1)\\
\vdots\\
\vdots\\
\alpha_l^i(m)\\
\end{pmatrix}.
\end{equation}
The system matrix in \eqref{eq:coeffs_in_local_frame} is of dimensions
$m\times (k-1)$.
In order to ensure that solutions exist, the number of sample points must be larger than the manifold dimension, $k> m$, and the logs $\Log_{p_l}(p_j)$, $j\neq k$ must span the tangent space $T_{p_l}\mcM$.
If $k> m+1$, we have more coefficients $\partial_i\varphi_j(\omega_l)$ than are needed for a representation of
$v^i_l$ in terms of the logs. The system is underdetermined.
The additional degrees of freedom may be used to impose extra conditions, e.g. that the derivative coefficients sum up to zero.
In fact, this is required for consistency, because $\sum_j \varphi_j(\omega)\equiv 1$, i.e., the weights correspond to a discrete signed measure of unit weight.
\end{remark}
\section{Multivariate Hermite interpolation in the tangent space}
\label{sec:THI}
As an alternative to BHI, multivariate Hermite interpolation on a manifold may also be conducted in a straightforward manner by moving all data, i.e., sample points and derivatives, to a selected tangent space and to perform the interpolation therein.
The following steps detail this approach.
\begin{enumerate}
\item Choose a manifold location $q_0\in \mcM$ that is to act as the `data center'. For example, $q_0$ may be one of the sample points or it may be the Riemannian barycenter of the given sample data set.
\item Map all sampled manifold locations to the tangent space at $q_0$, i.e., compute
$w_j = \Log_{q_0}(pj)\in T_{q_0}\mcM$, $j=1,\ldots,k$.
\item Transport the sampled velocity vectors to the selected tangent space $T_{q_0}\mcM$.
To this end, compute for $i=1,\ldots, d, \hspace{0.1cm} l = 1,\ldots, k$,
\begin{eqnarray}
\nonumber
T_{q_0}\mcM\ni \hat v_l^i
&=& d(\Log_{q_0})_{p_l}[\partial_i f(\omega_l)]\\
\label{eq:THI_FD_prep}
&\approx&\frac{\Log_{q_0}( \Exp_{p_l}(\Delta t \partial_i f(\omega_l))) - \Log_{q_0}(\Exp_{p_l}(-\Delta t \partial_i f(\omega_l)))}{2\Delta t}.
\end{eqnarray}
The latter finite-differences approximation can be used in cases, where the Riemannian log map is not available in closed form,
for details and remarks on the numerical accuracy, see \cite{Zimmermann:2020}.
\item With all data gathered on one and the same tangent space, classical Hermite interpolation in vector spaces can be pursued. The interpolant is constructed as a weighted linear combination of the sampled data
as
\begin{equation}
\label{eq:tanginterp}
\hat f_{\text{tan}}(\omega) = \sum_{j=1}^k \varphi_j(\omega) w_j + \sum_{i=1}^d\sum_{l=1}^k \psi_{i,l}(\omega) \hat v_l^i \in T_{q_0}\mcM.
\end{equation}
\item The sampled data provides Hermite sample values for the weight functions $\varphi_j(\omega), \psi_{i,l}(\omega)$.
Let $\vec{\varphi_j}:= \left(\varphi_j(\omega_1),\ldots,\varphi_j(\omega_k)\right)$ be the vector
of sampled values of the $j$th coefficient function $\omega \mapsto \varphi_j(\omega)$.
Likewise, let
$\vec{\psi}_{i,l}:= \left(\psi_{i,l}(\omega_1),\ldots,\psi_{i,l}(\omega_k)\right)$.
For each of the coefficient functions, one needs to solve a multiple-input scalar-output
Hermite interpolation problem, where the sample locations are $\omega_1,\ldots,\omega_k$
and the sample values for the various coefficient functions are
\begin{equation}
\label{eq:THI_sample_data}
\begin{pmatrix}
\vec{\varphi}_{j}\\
\partial_1 \vec{\varphi}_{j}\\
\vdots\\
\partial_d \vec{\varphi}_{j}
\end{pmatrix}
=
\begin{pmatrix}
e_j\\
\mathbf{0}\\
\vdots\\
\mathbf{0}
\end{pmatrix},
\quad
\begin{pmatrix}
\vec{\psi}_{1,l}\\
\partial_1 \vec{\psi}_{1,l}\\
\vdots\\
\partial_d \vec{\psi}_{1,l}
\end{pmatrix}
=
\begin{pmatrix}
\mathbf{0}\\
e_l\\
\vdots\\
\mathbf{0}
\end{pmatrix},
\ldots,
\begin{pmatrix}
\vec{\psi}_{d,l}\\
\partial_1 \vec{\psi}_{d,l}\\
\vdots\\
\partial_d \vec{\psi}_{d,l}
\end{pmatrix}
=
\begin{pmatrix}
\mathbf{0}\\
\mathbf{0}\\
\vdots\\
e_l
\end{pmatrix} \in \R^{k(d+1)}.
\end{equation}
Here, the data is written in vector packs of size $k$
($k$ values from $\varphi_j(\omega)$ evaluated at the $k$ sample locations,
$k$ values from $\partial_1\varphi_j(\omega)$ evaluated at the $k$ sample locations, etc. )
and $e_j=(0,\ldots,\stackrel{j}{1},\ldots, 0)^{\mathrm{T}}\in\R^k$ denotes the $j$th canonical unit vector.
\item The interpolants of the coefficient functions can now be obtained from any multivariate interpolation scheme that is able to tackle Hermite data, e.g., the gradient-enhanced Kriging method, \cite[Section 7.2]{ForresterSobesterKeane:2008}.
In summary, this requires to solve $k(d+1)$ Euclidean interpolation problems.
The resulting tangent vector of $\eqref{eq:tanginterp}$ is mapped back to the manifold via
\begin{equation}
\label{eq:THI_interpolant}
\hat f(\omega) = \Exp_{q_0}(\hat f_{\text{tan}}(\omega)) \in \mcM.
\end{equation}
\end{enumerate}
\paragraph{Preliminary comparison.}
For simplicity, the above interpolation approach is referred to as the tangent space Hermite interpolation \textbf{(THI)} method.
The BHI method outlined in Section \ref{sec:baryHermite} works with a number of $k$ weight coefficient functions with properly adjusted partial derivatives, which are obtained by solving the linear equation systems \eqref{eq:coeffs_in_local_frame}.
The THI method works with $k(d+1)$ coefficient functions, each of which comes with its own Hermite sample data set.
At the preprocessing stage of the BHI approach, the Riemannian logarithm function needs to be queried. In contrast, prior to conducting THI, one needs to query the Riemannian logarithm as well as its total derivative (or an approximation thereof).
In the online stage of BHI, the values of the coefficient functions at an untried location $\omega^*\in\R^d$ are determined via a Riemannian optimization problem.
In fact, it is the exact same Riemannian optimization problem that needs to be tackled for performing non-Hermite barycentric interpolation, except that the weight functions are properly adjusted a priori. The practitioner may choose any preferred method from the zoo of optimization algorithms, see, e.g., \cite{AbsilMahonySepulchre:2008}, and
retractions could be used to replace the Riemannian exponential.
For THI, the values of the coefficient functions at $\omega^*$ are obtained via solving $k(d+1)$ Euclidean interpolation problems in a selected tangent space.
The back-mapping to the manifold must be via the Riemannian exponential, for a retraction would compromise the interpolation condition, unless the inverse of the same retraction is also used at the preprocessing stage.
Table \ref{tab:CompMethods} summarizes the main features of both approaches.
\begin{table}[tbp]
\begin{tabular}{ccc}
\toprule
& Barycentric & Tangent space\\ \midrule
\#(coefficient functions) & $k$ & $k(d+1)$\\
minimum \#(samples) & $k>\dim(\mcM) +1$ & no restriction \\
base-point dependent & no & yes \\
manifold data processing via& $\Log_q$ &$\Log_q$, $d(\Log_q)_p$\\
model evaluation via &Riemann. optimization & interp. in tangent space\\
\bottomrule
\end{tabular}
\caption{Comparison of the barycentric and the tangent space approach to Hermite manifold interpolation for a $d$-variate interpolation problem with a number of $k$ sample locations $\omega_j\in \R^d$.}
\label{tab:CompMethods}
\end{table}
\section{Numerical examples}
\label{sec:numex}
In this section, we demonstrate multivariate Hermite manifold interpolation by means of academic examples.
For clarity of the exposition, but for no mathematical reasons, we focus in the experiments on the case, where $\mcM$ is an embedded submanifold of $\R^{N}$ and we express data points $p\in\mcM$ and tangent vectors $v\in T_p\mcM$ in \emph{extrinsic coordinates}. This means that we use coordinates of $\R^N$ rather than local coordinates to address locations on $\mcM$. This situation is also to be expected in practical applications.
\begin{example*}
On the unit sphere $\mcM=S^2=\{p\in\R^3\mid p_1^2+p_2^2 +p_3^2 =1\}\subset \R^3$, which is of (manifold) dimension $m=2$,
points are addressed by their three Euclidean coordinates $p=(p_1,p_2,p_3)^{\mathrm{T}}$.
Likewise, tangent vectors $v\in T_pS^2=\langle p\rangle^\bot$ are also given by extrinsic coordinates $v=(v_1,v_2,v_3)^{\mathrm{T}}\in T_pS^2\subset \R^3$.
\end{example*}
Consider a differentiable, $d$-variate, $\mcM$-valued function
\begin{equation*}
f\colon\R^d\supset D\to \mcM\subset \R^N, \omega \mapsto f(\omega)
\end{equation*}
Suppose that $k$ sample locations $\omega^j=(\omega^j_1,\ldots,\omega^j_d)^{\mathrm{T}}\in \R^d$ are selected
and that sample points $p^j =(p^j_1,\ldots,p^j_N)^{\mathrm{T}} = f(\omega^j)\in\mcM\subset \R^N$ and tangent vectors $v^i_j=\partial_i f(\omega^j)\in T_{p_j}\mcM\subset\R^N,i=1,\ldots,d$ are available.\\
For THI, the coefficient weight functions are obtained by Hermite interpolation of the data sets listed in \eqref{eq:THI_sample_data}.
In the BHI approach, we construct interpolants by computing numerical solutions to the Riemannian optimization problem \eqref{eq:baryInterp}.
In this case, the vector of sample values for the $j$th weight function $\varphi_j$ in \eqref{eq:baryInterp} is
\begin{equation}
\label{eq:sample_weight_funs}
(\varphi_j(\omega^1), \ldots, \varphi_j(\omega^k)) = e_j = (0,\ldots,0,\stackrel{j}{1},0,\ldots,0)\in \R^k
\end{equation}
and is independent of the sample points $p_j\in \mcM$.
The vector of the partial derivatives of $\varphi_j$ at the sample locations is obtained from Algorithm \ref{alg:partial_derivs}.
\begin{algorithm}{Practical computation of the derivatives of the weight functions for BHI.}
\begin{algorithmic}[1]
\REQUIRE{Hermite data set consisting of sample pairs $ (p_j=f(\omega_j), \omega_j)\in \mcM \times\R^d$, $ j=1,\ldots,k$, and partial derivatives $v_l^i = \partial_i f(\omega_l)\in T_{p_l}\mcM$, $i=1,\ldots,d$, $l=1,\ldots,k$.
}
\FOR{$i=1,\ldots,d$}
\FOR{$l=1,\ldots,k$}
\STATE{Compute $\Log_{p_l}(p_j)\in T_{p_l}\mcM\subset \R^N$, $j=1,\ldots,k$, $j\neq l$}
\STATE{$X_l\coloneqq \begin{pmatrix}[c|c|c|c|c|c]
\Log_{p_l}(p_1) & \ldots & \Log_{p_l}(p_{l-1})& \Log_{p_l}(p_{l+1}) &\ldots &\Log_{p_l}(p_k)
\end{pmatrix}\in \R^{N\times (k-1)}$}
\STATE{Solve $
\begin{pmatrix}
X_l\\
1,\ldots,1
\end{pmatrix}
\begin{pmatrix}
c_1\\
\vdots\\
c_{k-1}
\end{pmatrix}
=
\begin{pmatrix}
v^i_l\\
0
\end{pmatrix}$}
\STATE{ $\left(
\partial_i\varphi_1(\omega_l),
\ldots,
\partial_i\varphi_{l-1}(\omega_l),
\partial_i\varphi_l(\omega_l),
\partial_i\varphi_{l+1}(\omega_l),
\ldots,
\partial_i\varphi_k(\omega_l)
\right) \coloneqq
\left(
c_1,
\ldots,
c_{l-1},
0,
c_l,
\ldots,
c_{k-1}
\right)$}
\ENDFOR
\ENDFOR
\ENSURE{Partial derivatives of the weight interpolation functions at the sample locations $\partial_i\varphi_{j}(\omega_l)$, $j,l=1,\ldots,k$, $i=1,\ldots,d$.}
\end{algorithmic}
\label{alg:partial_derivs}
\end{algorithm}
\begin{remark}
In the upcoming experiments, we face the setting, that $\text{dim}(\mcM)=m<k-1$, where $k$ is the number of sample points. Hence, the linear system in step 5 of Alg. \ref{alg:partial_derivs} is underdetermined. The bottom row of all ones is added to the system matrix to enforce that $\sum_j \partial_i\varphi_{j}(\omega_l) = 0$.\\
The minimum 2-norm solution to the system is obtained via the pseudo-inverse based on the SVD, see
\cite[\S 5.7, p. 270--273]{GolubVanLoan:1996}.
Let $r = \rank X$
and let $U_r\Sigma_r V^{\mathrm{T}}_r = X$
be the reduced SVD, i.e.,
$U_r\in \R^{N\times r}$,
$\Sigma_r\in\R^{r\times r}$,
$V_r\in \R^{(k-1)\times r}$.
The columns of $U_r$ form a basis for the column space of $X$.
The underdetermined equation $Xc = v_l^i$ corresponds to an equation for the coordinates with respect to the basis given by $U_r$, namely,
$(U_r^{\mathrm{T}} X) c= (\Sigma_r V_r^{\mathrm{T}}) c = \tilde{v}_l^i = U_r^{\mathrm{T}}v_l^i$.
Writing $\bone\in \R^{k-1}$ for the vector with all entries equal to $1$, we enforce that the derivatives of the coefficient functions sum up to zero via
\[
\tilde{X}_r c :=
\begin{pmatrix}
\Sigma_r V_r^{\mathrm{T}} \\
\bone^{\mathrm{T}}
\end{pmatrix} c
= \begin{pmatrix}
\tilde{v}_l^i\\
0
\end{pmatrix}.
\]
The minimum 2-norm solution is obtained via $c = \tilde{X}_r^{\mathrm{T}}(\tilde{X}_r\tilde{X}_r^{\mathrm{T}})^{-1}\tilde{v}_l^i$. Using the SVD data of $X$ and Schur complement inversion, this can be calculated in closed form
\begin{equation}
\label{eq:underdetX_closed_form}
c = x + \frac{\langle\bone, x\rangle}{s} \left(V_r (V_r^{\mathrm{T}}\bone) - \bone \right),
\quad
x = V_r\Sigma_r^{-1} \tilde{v}_l^i \in \R^{k-1},
\quad
s = (k-1) - \lVert V_r^{\mathrm{T}}\bone\rVert_2^2\in\R.
\end{equation}
\end{remark}
For constructing the interpolation weight functions $\varphi_j$ in BHI and $\varphi_j, \psi_{i,l}$ in THI, we use the method of gradient-enhanced Kriging with the cubic correlation model and a
fixed correlation hyperparameter vector $\theta = (\theta_1,\theta_2) = (0.5, 0.5)$.
The cubic correlation model is
$\rho(\theta,\omega^i,\omega^j)= \prod_{l=1}^{2}{\rho_l(\theta_l, (\omega^i_l - \omega^j_l))}$,
where
\[
\rho_l(\theta_l, (\omega^i_l - \omega^j_l)) =
\left\{
\begin{array}{ll}
1-3\left(\theta_l |\omega^i_l - \omega^j_l|\right)^2 + 2 \left(\theta_l |\omega^i_l - \omega^j_l|\right)^3, & \theta_l |\omega^i_l - \omega^j_l|<1\\
0, & \theta_l |\omega^i_l - \omega^j_l|\geq 1
\end{array}
\right. .
\]
For the details, we refer to \cite[Section 7.2]{ForresterSobesterKeane:2008} and \cite{Zimmermann:2013}.
The input sample values are given by \eqref{eq:sample_weight_funs} and Algorithm \ref{alg:partial_derivs} for BHI and by \eqref{eq:THI_sample_data} for THI, respectively.
BHI requires to solve the optimization problem \eqref{eq:baryInterp} and thus to provide the optimizer with an initial guess.
In the first run of the optimizer, we use the first sample location as the initial guess, i.e., $q_0 = p_1$. For every consecutive run, we use the optimized solution $q^*$ from the previous run as the starting point for the next optimization procedure.
\subsection{Interpolation of the Gauß map of the Helicoid}
\label{sec:numex_Helicoid}
In this section we consider an academic example on the unit sphere $\mcM=S^2$.
The Riemannian exponential and logarithmic maps on $S^2$ are
\begin{align*}
\Exp_q(v) &= \cos(\lVert v\rVert_2)q +\sin(\lVert v\rVert_2) \frac{v}{\lVert v\rVert_2}\in S^2,
\\
\quad \Log_q(p) &= \arccos(\langle q,p\rangle)
\frac{p-\langle q,p\rangle q}{\lVert p-\langle q,p\rangle q\rVert_2}\in T_qS^2,
\end{align*}
respectively.
As a test function, we use the Gauß map of the helicoid in $\R^3$,
\begin{equation*}
f\colon [a,b]^2\to S^2, \quad (\omega_1,\omega_2) \mapsto \frac{1}{(e^{2\omega_1} +1)}
\begin{pmatrix}
2e^{\omega_1}\cos(\omega_2)\\
2e^{\omega_1}\sin(\omega_2)\\
e^{2\omega_1} - 1
\end{pmatrix}.
\end{equation*}
This function yields the normal field of the helicoid and is obtained from the stereographic projection of $(\omega_1,\omega_2)\mapsto \exp(\omega_1 +i\omega_2)\in \C$ onto $S^2$.\\
The partial derivatives of $f$ are
\begin{align*}
\partial_1 f(\omega_1,\omega_2) &= \frac{-2e^{2\omega_1}}{(e^{2\omega_1}+1)^2}
\begin{pmatrix}
2e^{\omega_1}\cos(\omega_2) \\ 2e^{\omega_1}\sin(\omega_2) \\ e^{2\omega_1} - 1
\end{pmatrix}
+ \frac{2}{e^{2\omega_1}+1} \begin{pmatrix}
e^{\omega_1}\cos(\omega_2) \\ e^{\omega_1}\sin(\omega_2) \\ e^{2\omega_1}
\end{pmatrix}, \\
\partial_2 f(\omega_1,\omega_2) &= \frac{1}{e^{2\omega_1}+1} \begin{pmatrix}
-2e^{\omega_1}\sin(\omega_2) \\ 2e^{\omega_1}\cos(\omega_2)\\ 0
\end{pmatrix}.
\end{align*}
We sample the function values and partial derivatives of $f$ on $k=9$ uniformly distributed sample locations $(\omega^j_1,\omega^j_2)\in [a,b]^2 = \left[-\frac14\pi, \frac14\pi\right]^2$, $j=1,\ldots,9$ and conduct BHI and THI.
The numerical gradient descent method of Algorithm \ref{alg:BaryInt} is conducted with a fixed step size of $\alpha =1.0$ and a numerical convergence threshold of $\tau = 1.0\cdot 10^{-8}$. For THI, the Riemannian barycenter of the sample data set is used as the tangent space base point.
The respective interpolant is evaluated on a uniform grid of $101\times 101$ points in the same domain $[a,b]^2$.
\begin{table}[tbp]
\begin{small}
\centering
\begin{tabular}{cccccc}\toprule
\multicolumn{6}{c}{\emph{Parameter settings}}\\%\midrule
Manifold & \#{variables} & domain & \#{samples} & threshold & line step\\
$S^2$ & $d=2$ & $[-\frac14\pi, \frac14\pi]^2$ & $k=9$ & $\tau =1.0\cdot 10^{-8}$ & $\alpha = 1.0$\\[\baselineskip]
\midrule
& \\
\midrule
\multicolumn{6}{c}{\emph{Results: barycentric Hermite interpolation (BHI)}}\\
\multicolumn{2}{c}{Wall clock time} & \multicolumn{2}{c}{Interpolation error} & \multicolumn{2}{c}{FD error}\\
\cmidrule(r){1-2}\cmidrule(lr){3-4}\cmidrule(l){5-6}
offline & online & max & avg & avg.~$\partial_1 \hat f(\omega^j) $ & avg.~$\partial_2 \hat f(\omega^j)$\\
$7.2\cdot 10^{-3}$s & $6.42\cdot 10^{-5}$s & $2.14\cdot 10^{-2}$ &$8.23\cdot 10^{-3}$& $6.17\cdot 10^{-6}$ & $6.27\cdot 10^{-6}$\\
%
\midrule
\multicolumn{6}{c}{\emph{Results: tangent space Hermite interpolation (THI)}}\\
\multicolumn{2}{c}{Wall clock time} & \multicolumn{2}{c}{Interpolation error} & \multicolumn{2}{c}{FD error}\\
\cmidrule(r){1-2}\cmidrule(lr){3-4}\cmidrule(l){5-6}
offline & online & max & avg & avg.~$\partial_1 \hat f(\omega^j) $ & avg.~$\partial_2 \hat f(\omega^j)$\\
$2.5\cdot 10^{-2}$s & $1.5\cdot 10^{-4}$s & $2.36\cdot 10^{-3}$ &$1.10\cdot 10^{-3}$& $6.47\cdot 10^{-6}$ & $7.61\cdot 10^{-6}$\\
\bottomrule
\end{tabular}
\caption{Parameter settings and interpolation results for the experiment of Subsection \ref{sec:numex_Helicoid} associated with Figure \ref{fig:Error_gauss_3x3}.
The `offline' time refers to the construction of the interpolant from the sampled data, while the `online' time accounts for querying the interpolant at a trial location.}
\label{tab:numex1}
\end{small}
\end{table}
Table \ref{tab:numex1} summarizes the parameter settings for this experiment and lists the wall clock time and the interpolation errors.
The wall clock time is split into an offline stage and an online stage.
The offline stage consists of the construction of the Hermite interpolants for the weight coefficient functions $\varphi_j$ for BHI and $\varphi_j, \psi_{i,l}$ for THI, respectively. The online stage accounts for querying the interpolant at a given parameter location.
The evaluation times are averaged over the number of $10,201= 101^2$ runs.\\
The sample data set in form of surface normal vectors, their partial derivatives and the interpolation error surfaces are displayed in Figure \ref{fig:Error_gauss_3x3}.
\begin{figure}[tbp]
\centering
\includegraphics[width=1.0\textwidth]{art/Helicoid_Errors_gauss_map_3x3samples_theta_05_05_tang+bary.png}
\caption{Interpolating the Gauß map of the Helicoid from Subsection \ref{sec:numex_Helicoid}. Left: sample data set and partial derivatives.
Middle: error surface obtained with BHI. Right: error surface obtained with THI.
}
\label{fig:Error_gauss_3x3}
\end{figure}
As can be seen from the Table \ref{tab:numex1} and Figure \ref{fig:Error_gauss_3x3}, the averaged accuracy of the THI interpolant is roughly a factor of $6$ times better than the averaged accuracy of the BHI interpolant. A finite difference check shows that both methods provide interpolants that meet the sampled tangent vectors.
The computation time for the offline stage under the THI approach is $3.5$ times higher than that for BHI.
This is expected, since for THI, the data preprocessing involves evaluating the Riemann exp and log maps plus the differential of the latter for moving the sampled derivatives to the same tangent space.\footnote{For the academic case at hand, a closed form calculation of $\diff{}(\Log_q)_p$ is possible. However, for consistency with the general case, in the numerical experiments, we work with the finite difference approximation of \eqref{eq:THI_FD_prep}.} Moreover, a number of
$k(d+1)=27$ Euclidean Hermite interpolation problems have to be solved to obtain the interpolated weight functions. In contrast, for BHI, only $k=9$ coefficient functions have to be fitted. This, however, involves solving an underdetermined linear equation system.\\
In the online stage of BHI, the basic gradient descent converges after an average number of $4.9$ iterations with the worst case taking $7$ iterations.
On average, for the test case at hand, BHI is online $2.3$ times faster than THI. However, too much weight should not be given to the timing results, as they are highly dependent on the chosen optimization algorithm and choosing a `best one' is beyond the scope of these experiments.\\
The THI method produces different interpolants for different choices of the tangent space base point.
To illustrate this issue, Figure \ref{fig:Error_base_popint_3x3} shows the error surfaces corresponding to three different base-point selections.
The left picture Figure \ref{fig:Error_base_popint_3x3} is produced by working in $T_{p_1}S^2$,
the middle picture by working in $T_{p_4}S^2$,
and the right picture corresponds to working in $T_{p_9}S^2$, where $p_j = f(\omega_j)$.
\begin{figure}[tbp]
\centering
\includegraphics[width=1.0\textwidth]{art/Helicoid_Errors_TSHI_BasePointComp_theta_05_05.png}
\caption{Error surfaces under the THI method from Subsection \ref{sec:numex_Helicoid} for different base tangent spaces.
}
\label{fig:Error_base_popint_3x3}
\end{figure}
\subsection{Interpolation on \texorpdfstring{$SO(3)$}{SO(3)}}
\label{sec:numex_SO3}
The next test case considers a Hermite data set on the special orthogonal group $SO(3) = \{Q\in \R^{3\times 3}\mid Q^{\mathrm{T}}Q = I, \det(Q) = 1\}$.
The tangent space of $SO(3)$ at a point $Q\in SO(3)$ is
$T_QSO(3) = \{V\in\R^{3\times 3}\mid Q^{\mathrm{T}}V +V^{\mathrm{T}}Q = 0\}$. The Riemannian $\exp$ and $\log$ maps on $SO(3)$ are
\begin{align*}
\Exp_Q\colon& T_QSO(3) \to SO(3),& V\mapsto Q\exp_m(Q^{\mathrm{T}}V),\\
\Log_Q\colon& \mathcal{D}_Q \to T_QSO(3),& R\mapsto Q\log_m(Q^{\mathrm{T}}R),
\end{align*}
where $\exp_m$ and $\log_m$ denote the classical matrix exponential and logarithm functions, see \cite{Higham:2008}, and
$\mathcal{D}_Q\subset SO(3)$ is a domain around $Q$ such that for all $R\in \mathcal{D}_Q$, the orthogonal matrix $Q^{\mathrm{T}}R$ does not feature $\lambda = -1$ as an eigenvalue.
As a test function, we consider
\begin{align*}
f\colon &[a,b]^2\to SO(3), \quad (\omega_1,\omega_2) \mapsto
\exp_m X(\omega_1, \omega_2), \text{ where}\\
&\quad X(\omega_1, \omega_2) =
\begin{pmatrix}[ccc]
0 & \omega_1^2 + \frac12\omega_2 & \sin\bigl(4\pi(\omega_1^2+\omega_2^2)\bigr)
\\[.66em]
-\omega_1^2 - \frac12\omega_2 & 0 & \omega_1 + \omega_2^2
\\[.66em]
-\sin(4\pi(\omega_1^2+\omega_2^2)) & -\omega_1 - \omega_2^2 & 0
\end{pmatrix}.
\end{align*}
The sample locations $P_j=\exp_m X(\omega^j)$ at $\omega^j=(\omega_1^j, \omega_2^j)$ and the corresponding partial derivatives $V^i_j =\deriv[t]\big\vert_{t=0} \exp_m(X(\omega^j+t e_i))= \diff{}(\exp_m)(X(\omega^j))[\partial_i X(\omega^j)] $, $ i=1,2$ of the test function can be obtained
by Mathias' theorem, see \cite[Thm. 3.6]{Higham:2008}:
\begin{equation*}
\exp_m\begin{pmatrix}[cc]
X(\omega^j) & \partial_i X(\omega^j)
\\[.66em]
0 & X(\omega^j)
\end{pmatrix}
=
\begin{pmatrix}[cc]
\exp_m(X(\omega^j)) & \diff{}(\exp_m)(X(\omega^j))[\partial_i X(\omega^j)]
\\[.66em]
0 & \exp_m(X(\omega^j))
\end{pmatrix}.
\end{equation*}
For BHI, the partial derivatives of the interpolation weight functions $\varphi_j$ are computed with Algorithm \ref{alg:partial_derivs},
where we use the vectorized tangent matrices $\text{vec}(\Log_{P_l}(P_j))\in\R^{3\cdot 3}$ as columns to form the matrices $X_l$. The input $V_l^i$ on the right hand side of the equation system in step 5 of Algorithm \ref{alg:partial_derivs} is vectorized accordingly. For BHI, for each trial location $\omega^*$, the interpolant $\hat f(\omega^*)$ is computed with Algorithm \ref{alg:BaryInt}, while for THI, it is computed according to \eqref{eq:tanginterp}
and \eqref{eq:THI_interpolant}. FOR THI, the Riemannian barycenter of the sample data set is used as the tangent space base point.\\
For quantifying the accuracy of the interpolation, we compute the relative errors
$\frac{\lVert f(\omega) - \hat f(\omega)\rVert_F}{\sqrt{3}}$. Mind that $\lVert Q\rVert_F= \sqrt{3}$ for any matrix $Q\in SO(3)$.
\begin{table}[tbp]
\begin{small}
\centering
\begin{tabular}{cccccc}\toprule
\multicolumn{6}{c}{\emph{Parameter settings}}\\%\midrule
Manifold & \#{variables} & domain & \#{samples} & threshold & line step\\
$SO(3)$ & $d=2$ & $[0.5, 0.5]^2$ & $k=49$ (Cheby.) & $\tau =1.0\cdot 10^{-6}$ & $\alpha = 1.0$\\[\baselineskip]
\midrule
& \\
\midrule
\multicolumn{6}{c}{\emph{Results: barycentric Hermite interpolation (BHI)}}\\
\multicolumn{2}{c}{Wall clock time} & \multicolumn{2}{c}{Interpolation error} & \multicolumn{2}{c}{FD error}\\
\cmidrule(r){1-2}\cmidrule(lr){3-4}\cmidrule(l){5-6}
offline & online & max & avg & avg.~$\partial_1 \hat f(\omega^j) $ & avg.~$\partial_2 \hat f(\omega^j)$\\
$0.41$s & $0.077$s & $0.029$ &$0.0069$ & $4.45\cdot 10^{-4}$ & $4.91\cdot 10^{-4}$\\
%
\midrule
%
\multicolumn{6}{c}{\emph{Results: tangent space Hermite interpolation (THI)}}\\
\multicolumn{2}{c}{Wall clock time} & \multicolumn{2}{c}{Interpolation error} & \multicolumn{2}{c}{FD error}\\
\cmidrule(r){1-2}\cmidrule(lr){3-4}\cmidrule(l){5-6}
offline & online & max & avg & avg.~$\partial_1 \hat f(\omega^j) $ & avg.~$\partial_2 \hat f(\omega^j)$\\
$0.73$s & $0.0023$s & $0.027$ &$0.0065$ & $3.9\cdot 10^{-4}$ & $4.6\cdot 10^{-4}$\\
\bottomrule
\end{tabular}
\caption{Parameter settings and interpolation results for the experiment of Subsection \ref{sec:numex_SO3}
associated with Figure \ref{fig:Error_SO3_both}.
The `offline' time refers to the construction of the interpolant from the sampled data, while the `online' time accounts for querying the interpolant at a trial location.}
\label{tab:numex_SO3}
\end{small}
\end{table}
In this experiment, we rely on a two-dimensional Chebychev sample plan
\begin{equation*}
\left\{\frac12 (b-a)\cos\left(\frac{(2j-1)\pi}{2k}\right) + \frac12(b+a)\mid j=1,\ldots, k\right\}^2\subset \R^2.
\end{equation*}
\begin{figure}[tbp]
\centering
\includegraphics[width=.4\textwidth]{art/SO3_chebysampling.png}
\caption{Sample plan associated with the experiment of Section \ref{sec:numex_SO3}.
Black dots: Chebychev $7\times7$ grid on the domain $[-0.5, 0.5]^2$. The stars indicate the trial locations that are used for visualization purposes in the upcoming Figure \ref{fig:t-pots}.}
\label{fig:chebysampling}
\end{figure}
We sample the function values and partial derivatives of $f$ on $k=49$ Chebychev sample locations $(\omega^j_1,\omega^j_2)\in [a,b]^2 = \left[-0.5, 0.5\right]^2$, $j=1,\ldots,49$, see \Cref{fig:chebysampling}.
The numerical gradient descent method of Algorithm \ref{alg:BaryInt} is conducted with a fixed step size of $\alpha =1.0$ and a numerical convergence threshold of $\tau = 1.0\cdot 10^{-6}$.
The BHI and THI interpolants are evaluated on a uniform grid of $76\times 76$ trial points in the same domain.
The corresponding interpolation error surfaces are displayed in
\Cref{fig:Error_SO3_both}.
\Cref{tab:numex_SO3} summarizes the parameter settings as well as the timing and accuracy results.\\
In the test case at hand, BHI and THI produce interpolants of roughly the same averaged accuracy, with THI performing slightly better than BHI.
To the naked eye, the error surfaces look identical for both approaches.
A finite difference check shows again that both methods provide interpolants that meet the sampled tangent vectors.
The computation time for the offline stage under the THI approach is roughly $2$ times higher than that for BHI.
In the online stage of BHI, the basic gradient descent converges after an average number of $18.0$ iterations with the worst case taking $87$ iterations.
On average, for the test case at hand, THI is online $33$ times faster than BHI.
We emphasize again that the timing results are expected to vary considerably depending on which optimization algorithm is employed.
\begin{figure}[tpb]
\centering
\includegraphics[width=1.0\textwidth]{art/Errors_SO3_map_7x7samples_a0.5_b0.5_4pi_theta_05_05_tau1e-6_BHI+TSHI.png}
\caption{
Error surfaces for $SO(3)$-interpolation on a Chebychev $7\times 7$ grid from Subsection \ref{sec:numex_SO3}. Left: Barycentric Hermite Interpolation (BHI). Right: Tangent Space Hermite Interpolation (THI).}
\label{fig:Error_SO3_both}
\end{figure}
In this test case, each interpolated value $\hat P = \hat f(\omega)$ is a matrix in $SO(3)$. In order to further visualize the interpolation results,
\Cref{fig:Coords_SO3_cheby} displays the interpolated matrix component functions
$\omega\mapsto \hat P_{11}$, $\omega\mapsto \hat P_{22}$, $\omega\mapsto \hat P_{33}$, and $\omega\mapsto \hat P_{31}$.
Because the plots for THI and BHI virtually coincide, only the results for BHI are shown.
\begin{figure}[tbp]
\centering
\includegraphics[width=1.0\textwidth]{art/SO3_map_7x7samples_a0.5_b0.5_4pi_theta_025_025_cheby_coords_11_22_33_31.png}
\caption{Plots of some selected interpolated matrix component functions
$(\omega_1, \omega_2) \to \left(\hat f(\omega_1, \omega_2)\right)_{i,j}\in \R$ of the experiment from Subsection \ref{sec:numex_SO3} . The black dots indicate the Chebychev $7\times 7$ sample grid.}
\label{fig:Coords_SO3_cheby}
\end{figure}
The first diagonal component of the interpolant is juxtaposed with the corresponding component of the reference function in \Cref{fig:Coord11_SO3_cheby}.\\
\begin{figure}[tbp]
\centering
\includegraphics[width=0.7\textwidth]{art/Errors_SO3_map_7x7samples_a0.5_b0.5_4pi_theta_025_025_cheby_coords_11.png}
\caption{Interpolated matrix component function $\hat P_{11}=(\hat f(\omega))_{11}$ (shaded surface)
and the reference matrix component $P_{11}=f(\omega)$ (white surface) together with the sample locations on a Chebychev $7\times 7$ grid from Subsection \ref{sec:numex_SO3}.}
\label{fig:Coord11_SO3_cheby}
\end{figure}
Matrices in $SO(3)$ induce rotations in the three-dimensional Euclidean space.
Such rotations may be visualized by their action on an object.
In \Cref{fig:t-pots}, we compare the Hermite $SO(3)$-interpolant and the reference $SO(3)$-function via showing their actions on a tea pot object.
Both the interpolant and the reference function are queried at the parameter locations
that are marked with a star in the sampling plan displayed in \Cref{fig:chebysampling}.
\begin{figure}[tpb]
\centering
\includegraphics[width=0.75\textwidth]{art/SO3_map_7x7samples_a0.5_b0.5_4pi_theta_05_05_cheby_interpolatedTpots.png}
\caption{(from upper left to lower right): reference rotations (gray) and interpolated $SO(3)$-matrices (blue) at the 6 trial points displayed in Fig. \ref{fig:chebysampling}. The rotation matrices are visualized via their action on the tea pot object.}
\label{fig:t-pots}
\end{figure}
The interpolated rotations visually agree very well with the reference rotations.
\section{Conclusions and discussion}
\label{sec:conclusions}
We have developed two approaches for multivariate Hermite interpolation of manifold-valued data sets.
On the one hand, Hermite interpolation can be achieved by computing correspondingly weighted Riemannian barycenters, referred to as \emph{barycentric Hermite interpolation} (BHI).
The method we presented should be seen as a generic framework for BHI.
For any practical application, the user has to
\begin{itemize}
\item construct a sample plan, i.e.\ employ a method to select the sample points $\omega_j \in \R^d$ (design of experiment).
\item choose a Riemannian optimization method to compute the weighted Riemannian center of mass in \eqref{eq:baryInterp}.
(Gradient descend, Gauss-Newton-type methods, nonlinear conjugate gradients, etc.)
\item choose interpolation weight functions $\omega \to \varphi_j(\omega)$ with the required properties.
Note that the $\varphi_j$'s are scalar functions on a Euclidean domain so that this subtask does not require any Riemannian considerations.
\item specify side constraints to obtain unique, smooth solutions to the underdetermined linear system in step 5 of Algorithm \ref{alg:partial_derivs}.
\end{itemize}
On the other hand, Hermite interpolation can be achieved by translating any classical approach of Hermite interpolation from Euclidean vector spaces to the tangent space of the manifold under consideration, referred to as \emph{tangent space Hermite interpolation} (THI). This requires to single out a manifold location and use it as the center for the tangent space.
Moreover, all data, i.e., locations and tangent vectors have to be moved to the same tangent space.
As a consequence, the interpolant depends on the base-point selection and the preprocessing stage is more involved.
Eventually, the interpolant is written as a weighted linear combination of the tangent space images of the sample points and the transported, sampled tangent space vectors. This entails that the number of weight functions in the THI approach equals the number of sample locations plus the total number of sampled partial derivatives. In a sense, THI can be considered a `brute-force' approach. Data is forced into a `Procrustean bed' (=the selected tangent space) and processed without regard to its nature.
From an aesthetic view point, the BHI method is the more appealing one: It is base-point independent and works with a number of weight coefficients that equals the number of sample points. Plus, it treats sample locations and tangent vectors, which in fact are incompatible entities, differently.
Yet, in the numerical experiments, the THI method proved to produce the more accurate results and the computational effort of querying the interpolant
in the online stage is much lower than for BHI, because the latter involves solving a Riemannian optimization problem at every trial point.
\section*{Acknowledgements}
The original idea for this research paper stems from a short-term visit of the first author to the "Numerical Algorithms and High-Per\-for\-mance Computing"-group at EPF Lausanne in Spring 2022.
The first author would like to thank the head of group Daniel Kressner and the doctoral assistant Axel E. J. S\'eguin
for their hospitality and for the very stimulating discussions during that visit.
\section*{Appendix}
\section{Barycentric Hermite interpolation: The Euclidean case}
\label{app:EuclideanCase}
To ease the transition to the manifold case of barycentric interpolation, in this section, we outline multivariate, gradient-en\-han\-ced barycentric interpolation in Euclidean spaces.
Let $f\colon\R^d\to \R^m$ be differentiable and suppose that we have sampled data
$f(w_j) = p_j\in \R^m$, $j=1,\ldots,k$ and partial derivatives $\partial_i f(w_j) = v_j^i\in T_{p_j}\R^m\simeq \R^m$, $j=1,\ldots,k$, $i=1,\ldots,d$.
The task is to construct an interpolant $\hat f$ such that
\begin{align}
\label{eq:basicHermiteEuc}
\hat f(w_j) &= p_j\in \R^m, \quad j =1,\ldots,k\\
\partial_i \hat f(w_j) &= v_j^i\in T_{p_j}\R^m,\simeq \R^m, \quad j=1,\ldots,k; \hspace{0.2cm} i=1,\ldots,d.
\end{align}
We approach this task by constructing weighted barycenters
\begin{equation}
\label{eq:baryEuc}
\hat f(\omega) = \text{arg}\min_{q\in\R^m} \frac12 \sum_{j=1}^k \varphi_j(\omega) \lVert q-p_j\rVert^2_2 \coloneqq \text{arg}\min_{q\in\R^m} L(q,\omega).
\end{equation}
Here, the $\varphi_j$, $j=1,\ldots,k$ are weight functions with the following properties
\begin{equation*}
\varphi_j(\omega_i) = \delta_{ij}, \quad
\sum_{j=1}^k \varphi(\omega) \equiv 1.
\end{equation*}
The objective function $L(q,\omega)$ in \eqref{eq:baryEuc} is a weighted sum over squared distance terms
$\frac12 \dist(q,p_j)^2 = \frac12\lVert p_j -q\rVert_2^2 =: L_{p_j}(q)$.
The gradient is $\grad L_{p_j}(q) = (q - p_j)$.
We emphasize that this is exactly minus the Riemannian logarithm $\Log_q(p_j) = p_j-q$ on the Euclidean $\R^m$.
The minimizer of \eqref{eq:baryEuc}
is the unique zero of the gradient equation
\begin{equation}
\label{eq:grad0Euc}
0 = \grad_q L(q, \omega) = \sum_{j=1}^k \varphi_j(\omega)\grad L_{p_j}(q) = \sum_{j=1}^k \varphi_j(\omega) (q-p_j).
\end{equation}
Because of $\sum_{j=1}^k \varphi_j(\omega) \equiv 1$, we obtain the interpolant $q=\hat f(\omega)$ as
\begin{equation}
\label{eq:interpolant_linear_in_data}
\hat f(\omega) = \sum_{j=1}^k \varphi_j(\omega) p_j.
\end{equation}
\begin{remark}
The interpolant in \eqref{eq:interpolant_linear_in_data} is a weighted linear combination of the sample data points
and can thus be constructed in the same fashion in arbitrary vector spaces.
Lagrange interpolation, radial basis function interpolation and Gaussian process regression/Kriging eventually boil down to the form of \eqref{eq:interpolant_linear_in_data} and differ only by the choice of the weight functions $\omega \to \varphi_j(\omega)$.
Hence, all these methods may be considered as examples of weighted barycentric interpolation approaches.
\end{remark}
\paragraph{Parameterized solutions via the implicit function theorem}
Introduce
\begin{equation*}
G\colon\R^m \times \R^d \to \R^m, \quad (q, \omega) \mapsto G(q,\omega) = \grad_q L(q,\omega) = \sum_{j=1}^k \varphi_j(\omega) (q-p_j).
\end{equation*}
Suppose that $G$ has a root at $(q^*, \omega^*)$. The differential of $G$ by $q$ at $(q^*, \omega^*)$ is
\begin{equation*}
\Diff_q G(q^*, \omega^*) = \sum_{j=1}^k \varphi_j(\omega^*) I = I,
\end{equation*}
and is invertible. By the implicit function theorem, there exists
$\hat f\colon \R^d\to \R^m, \omega\mapsto \hat f(\omega)$ such that on a suitably small domain around $\omega^*$, it holds $G(\hat f(\omega),\omega) \equiv 0$.
Implicit differentiation yields (locally)
\begin{equation*}
\Diff_q G(\hat f(\omega), \omega) \circ D\hat f(\omega) + \Diff_\omega G(\hat f(\omega), \omega) =0.
\end{equation*}
Note that $\Diff_q G(\hat f(\omega), \omega) = \Diff_q \grad_q L(\hat f(\omega), \omega)=\Hess_q L(\hat f(\omega),\omega)$.
Hence, we obtain the following equation that relates the partial derivatives of $\hat f$, which acts as the interpolant, to those of the weight functions $\varphi_j$.
\begin{equation}
\label{eq:implderivEuc}
\Hess_q L(\hat f(\omega),\omega) [\partial_i \hat f(\omega)]
= -\sum_{j=1}^k {\partial_i \varphi_k(\omega)(\hat f(\omega)-p_j)}
= \sum_{j=1}^k {\partial_i \varphi_k(\omega)\Log_{\hat f(\omega)}(p_j)}.
\end{equation}
It holds $\Hess_q L(\hat f(\omega),\omega) = I$ so that
\begin{equation*}
\partial_i \hat f(\omega) =
\begin{pmatrix}[ccc]
\Log_{\hat f(\omega)}(p_1) & \cdots &\Log_{\hat f(\omega)}(p_k)
\end{pmatrix}
\begin{pmatrix}
\partial_i \varphi_1(\omega)\\
\vdots\\
\partial_i \varphi_k(\omega)
\end{pmatrix}.
\end{equation*}
If the partial derivatives $\partial_i \hat f(\omega_j)$, $i=1,\ldots,d$ are known at the sample sites $\omega_j, j=1,\ldots,k$, then the above equation imposes conditions
on the partial derivatives of the weight functions thereat.
At sample location $\omega_l$, it holds $\hat f(\omega_l) = p_l$ and
\begin{equation}
\label{eq:partialbaryHermiteEuc}
\partial_i \hat f(\omega_l)
= -\sum_{j=1, j\neq l}^k {\partial_i \varphi_j(\omega_l)(p_l-p_j)}
= \sum_{j=1, j\neq l}^k {\partial_i \varphi_j(\omega_l)\Log_{p_l}(p_j)}.
\end{equation}
In order to be able to match any prescribed, sampled partial derivatives,
$\partial_i \hat f(\omega_l)$, we require the sets
\begin{equation*}
\mathcal L_l \coloneqq \{p_j - p_l\mid j=1,\ldots,k, j\neq l\} = \{\Log_{p_l}(p_j)\mid j\neq l\}
\subset T_{p_l}\R^m\simeq \R^m
\end{equation*}
to span the full tangent space $T_{p_l}\R^m\simeq \R^m$.
As a consequence, at least $k\geq m+1$ sample points must be available to make the barycentric Hermite interpolation problem well-defined.
If $k>m+1$, the equation system \eqref{eq:partialbaryHermiteEuc} is underdetermined and infinitely many valid choices of partial derivative values $\partial_i \varphi_j(\omega_l)$ exist that allow to represent the sampled partial derivatives.
\section{Karcher's computation of the gradient and the Hessian of the Riemannian distance function}
\label{app:proofs}
To make this paper self-contained, we recap the proof of Karcher's Theorem, \Cref{thm:grad_dist}.
To this end, we first list two standard facts about covariant derivatives.
\begin{lemma}[Product rule, {\cite[Chapter 2, Prop. 3.2]{DoCarmo:1992}}]
\label{lem:productrule_codiff_along_c}
Let $\covDeriv$ denote the Le\-vi-Ci\-vi\-ta connection along a curve $c:I\to \mcM$. It holds
\begin{equation*}
\deriv[t] \langle X(t), Y(t)\rangle_{c(t)}
= \langle \frac{\mathrm{D}X}{\mathrm{d}t} (t), Y(t)\rangle_{c(t)} +
\langle X(t), \frac{\mathrm{D}Y}{\mathrm{d}t}(t)\rangle_{c(t)}
\end{equation*}
for all smooth vector fields $X,Y$ along $c$.
\end{lemma}
The next statement corresponds to the Theorem of Schwarz in classical multivariate calculus.
\begin{lemma}[{\cite[Chapter 3, Lemma 3.4]{DoCarmo:1992}}]
\label{lem:schwarz4codiff}
Let $\alpha\colon I\times J \rightarrow \mcM, (s,t)\mapsto \alpha(s,t)$
be smooth.
Note that for $s$ fixed, $t\mapsto \alpha(s,t)$ is a smooth manifold curve,
likewise for $t$ fixed.
Hence,
\begin{align*}
\frac{\partial \alpha}{\partial s} (s,t) &= \deriv[\sigma]\big\vert_{\sigma = 0}\alpha(s+\sigma,t) \in T_{\alpha(s,t)}\mcM,\\
\frac{\partial \alpha}{\partial t} (s,t) &= \deriv[\tau]\big\vert_{\tau = 0}\alpha(s,t+\tau) \in T_{\alpha(s,t)}\mcM,\\
\end{align*}
and it holds
\begin{equation*}
\frac{\mathrm{D}}{\partial t} \frac{\partial\alpha}{\partial s}(s,t) =
\frac{\mathrm{D}}{\partial s} \frac{\partial\alpha}{\partial t} (s,t).
\end{equation*}
Here,
\begin{equation*}
\frac{\mathrm{D}}{\partial t} \frac{\partial\alpha}{\partial s}(s,t) \coloneqq
\covDeriv[\tau]\big\vert_{\tau = 0}
\tau \mapsto \alpha(s,t+\tau).
\end{equation*}
Likewise for $\frac{\mathrm{D}}{\partial s} \frac{\partial\alpha}{\partial t} (s,t)$.
\end{lemma}
\begin{theorem}[{\cite[Thm. 1.2]{Karcher:1977}}]
\label{thm:Karcher1977}
Let $\mcM$ be a complete Riemannian manifold and let $p\in \mcM$ be a fixed point.
Let $B_\rho(p)$ be a geodesic ball around $p_*$ such that the geodesics between any two points inside $B_\rho(p)$ are unique and minimizing.
Define
\begin{equation*}
f\colon B_\rho(p)\to \R, \quad q \mapsto \frac12 \dist(q,p)^2.
\end{equation*}
Then
\begin{equation*}
\grad_q f = -\Log_q(p).
\end{equation*}
\end{theorem}
\begin{proof}
Let $\gamma\colon [0,1]\to B_\rho(p)$ be a geodesic with $\gamma(0) = q, \dot \gamma(0) = v\in T_q\mcM$.
Consider a variation of geodesics
\begin{equation*}
c_p(s,t) = \Exp_p(s\Log_p(\gamma(t)))
\end{equation*}
as illustrated in Figure \ref{fig:GeoVar}.
\begin{figure}[tbp]
\centering
\includegraphics[width=0.5\textwidth]{art/GeoVar_tikzpic.pdf}
\caption{Illustration of the variation through geodesics considered in the proof of Theorem \ref{thm:Karcher1977}.
}
\label{fig:GeoVar}
\end{figure}
As in \cite{Karcher:1977}, we write
\begin{equation*}
c_p'(s,t) = \partial_s c_p(s,t), \quad \dot c_p(s,t) = \partial_t c_p(s,t)
\end{equation*}
and observe that
\begin{enumerate}[(i)]
\item $c_p(0,t) = \Exp_p(0) = p$. (All geodesics $s\mapsto c_p(s,t)$ start from $p$).
\item $c_p(1,t) = \gamma(t)$. (All geodesics $s\mapsto c_p(s,t)$ end at $\gamma(t)$).
Moreover, $\dot \gamma(t) = \dot c_p(1,t)$.
\item the terminal velocity of the geodesic $s\mapsto c_p(s,t)$ matches minus the velocity of the geodesic
emanating from $\gamma(t) = c_p(1,t)$ to reach $p$. Hence,
\begin{equation*}
c_p'(1,t) = -\Log_{\gamma(t)}(p).
\end{equation*}
\item all geodesics are constant-speed curves so that
\begin{equation*}
\dist(p,\gamma(t)) =\lVert \Log_p(\gamma(t)\rVert_p = \lVert c_p'(0,t)\rVert_p = \lVert c_p'(s,t)\rVert_{c_p(s,t)} \text{ for all } s\in[0,1].
\end{equation*}
Because of this, we can write
\begin{equation*}
\lVert c'_p(s,t)\rVert^2 = \int_0^1\lVert c'_p(s,t)\rVert^2 \dInt s.
\end{equation*}
\item When taking covariant derivatives, we may change the order of differentiation according to \Cref{lem:schwarz4codiff}
\begin{equation*}
\frac{\mathrm{D}}{\partial_s}\dot c_p(s,t) = \frac{\mathrm{D}}{\partial_t} c_p'(s,t).
\end{equation*}
\end{enumerate}
We calculate for any $v\in T_q\mcM$ the differential as
\begin{align*}
\diff{}f_q(v)
&= \deriv[t]\big\vert_{t=0} f(\gamma(t)) = \frac12 \deriv[t]\big\vert_{t=0} \dist(\gamma(t),p)^2\\
&= \frac12 \deriv[t]\big\vert_{t=0} \int_0^1\lVert c'_p(s,t)\rVert^2 \dInt s\\
&= \frac12 \int_0^1 \deriv[t]\big\vert_{t=0} \langle c'_p(s,t), c'_p(s,t)\rangle\dInt s\\
&= \int_0^1 \langle \frac{\mathrm{D}}{\partial_t} c'_p(s,t), c'_p(s,t)\rangle\big\vert_{t=0}\dInt s\\
&= \int_0^1 \langle \frac{\mathrm{D}}{\partial_s} \dot c_p(s,t), c'_p(s,t)\rangle\big\vert_{t=0}\dInt s\\
&= \int_0^1 \left(\langle \frac{\mathrm{D}}{\partial_s} \dot c_p(s,t), c'_p(s,t)\rangle
+ \langle\dot c_p(s,t), \frac{\mathrm{D}}{\partial_s} c'_p(s,t)\rangle \right)\big\vert_{t=0}\dInt s,\\
\end{align*}
where the second term vanishes, since $c_p(s,t)$ is a geodesic in $s$ for any fixed $t$. We continue
\begin{align}
\nonumber
\diff{}f_q(v)
&= \int_0^1 \deriv[s] \langle \dot c_p(s,t), c'_p(s,t)\rangle\dInt s\big\vert_{t=0} \\
\nonumber
&= \left[ \langle \dot c_p(s,t), c'_p(s,t)\rangle \right]_0^1 \big\vert_{t=0}, \\
&= \langle \dot c_p(1,t), c'_p(1,t)\rangle + \langle \dot c_p(0,t), c'_p(0,t)\rangle \big\vert_{t=0},
\label{eq:var1stOrder_deriv}
\end{align}
where again the second summand is zero since $c_p(0,t) = p$ is constant. We finally obtain
\begin{equation*}
\diff{}f_q(v)
= \langle \dot \gamma(t), -\Log_{\gamma(t)}(p)\rangle \big\vert_{t=0}\\
\nonumber
= \langle v, -\Log_{q}(p)\rangle
\end{equation*}
and hence $\grad_q f = -\Log_{q}(p)\in T_q\mcM$.
\end{proof}
\paragraph{The Hessian of the Riemannian distance function.}
Let $g\colon\mcM \to \R$ be a smooth scalar function.
Following \cite[Section 6A]{Kuhnel:2015}, it may be considered as a $(0,0)$-tensor.
The covariant derivative of $g$ (in the sense of \cite[Def. 6.2]{Kuhnel:2015})
is a $(0,1)$-tensor $(\nabla g) (X)$ that maps a vector field $X$ to a scalar function.
More precisely, we have
\begin{equation*}
\nabla g\colon \mathcal{X}(\mcM) \to C^\infty(\mcM),
X \mapsto \left(p \mapsto (\nabla f(X))(p) = \diff{}g_p(X(p)) = \langle \grad f(p), X(p) \rangle_p\right).
\end{equation*}
The second covariant derivative, $\nabla^2 g$, called the \emph{Hessian} of $g$,
is obtained by taking the covariant derivative of the $(0,1)$-tensor $(\nabla g)$.
This produces a $(0,2)$-tensor.
A calculation shows
\begin{equation}
\nonumber
\nabla^2 f (X,Y) = \langle \nabla_X \grad f, Y \rangle.
\end{equation}
The Hessian $(1,1)$-tensor associated with $g$ is thus
\begin{equation*}
\Hess f\colon \mathcal{X}(\mcM) \to \mathcal{X}(\mcM) , \quad X \mapsto \nabla_X \grad g.
\end{equation*}
The next lemma provides a path to compute the Hessian operator at a point $p$ and tangent vector $v=X(p)\in T_p\mcM$.
\begin{lemma}[cf. {\cite[Prop. 5.5.4]{AbsilMahonySepulchre:2008}}]
Let $g\colon\mcM\to \R$ be a smooth scalar function on a
Riemannian manifold $\mcM$.
Let $\gamma$ be a geodesic with $\gamma(0)= q$, $\dot\gamma(0)=v$.
Then
\begin{equation*}
\langle \Hess g(q)[v]\hspace{0.1cm},\hspace{0.1cm} v\rangle = \frac{\mathrm{d}^2}{\mathrm{d}t^2}\big\vert_{t=0} (g\circ \gamma)(t).
\end{equation*}
\end{lemma}
\begin{proof}
It holds
\begin{equation*}
\deriv[t] (f\circ \gamma)(t) = \diff{}f_{\gamma(t)}(\dot \gamma(t)) = \langle \grad f(\gamma(t)), \dot \gamma(t) \rangle.
\end{equation*}
Taking the second derivative, we obtain by the rules of covariant differentiation along a curve and
the product rule of Lemma \ref{lem:productrule_codiff_along_c} that
\begin{align*}
\frac{\mathrm{d}^2}{\mathrm{d}t^2} (f\circ \gamma)(t)
&= \deriv[t] \langle \grad f(\gamma(t)), \dot \gamma(t) \rangle\\
&= \langle \covDeriv[t] \grad f(\gamma(t)), \dot \gamma(t) \rangle + \langle \grad f(\gamma(t)), \covDeriv[t] \dot \gamma(t)\rangle\\
&= \langle \nabla_{\dot \gamma(t)} \grad f, \dot \gamma(t) \rangle
\\
&= \langle \Hess f(\gamma(t))[\dot\gamma(t)]\hspace{0.1cm},\hspace{0.1cm} \dot\gamma(t)\rangle.
\end{align*}
The third identity holds, because $\covDeriv[t] \dot \gamma(t) = 0$, since $\gamma$ is a geodesic.
At $t=0$, this yields
\begin{equation}
\label{eq:Hess_vv}
\frac{\mathrm{d}^2}{\mathrm{d}t^2} (f\circ \gamma)(0)
= \langle \Hess f(q)[v]\hspace{0.1cm},\hspace{0.1cm} v\rangle,
\end{equation}
which finishes the proof.
\end{proof}
For computing the Hessian of the distance function $q\mapsto \frac12 \dist(p,q)^2$, we utilize the same variation of geodesics
$c_p(s,t)$ as in the above proof of Theorem \ref{thm:Karcher1977}
and continue the calculation from \eqref{eq:var1stOrder_deriv}.
Using that $c_p(1,t)$ is a geodesic, we obtain
\begin{align*}
\frac{\mathrm{d}^2}{\mathrm{d}t^2} \frac12 \dist(\gamma(t), p)^2
&= \deriv[t] \left(\deriv[t]\frac12 \dist(\gamma(t), p)^2\right)
\stackrel{\eqref{eq:var1stOrder_deriv}}{=}\deriv[t] \langle \dot c_p(1,t), c_p'(1,t) \rangle\\
&= \langle \covDeriv[t]\dot c_p(1,t), c_p'(1,t) \rangle + \langle\dot c_p(1,t) , \covDeriv[t]c_p'(1,t) \rangle\\
&= \langle \dot \gamma(t) , \covDeriv[s] \dot c_p(1,t) \rangle.
\end{align*}
At $t=0$, we obtain for $\gamma(0) = q$, $\dot \gamma(0) = v$,
\begin{equation}
\label{eq:Hessian_via_var_of_geos}
\langle \Hess f(q)[v]\hspace{0.1cm},\hspace{0.1cm} v\rangle =
\frac{\mathrm{d}^2}{\mathrm{d}t^2}\big\vert_{t=0} \dist(\gamma(t), p) =
\langle \covDeriv[s] \dot c_p(1,0), v \rangle.
\end{equation}
\textbf{Remark}
Associated with the variation $c_p(s,t) = \Exp_p(s\Log_p(\gamma(t)))$ is the variation field
\begin{equation*}
s\mapsto \dot c_p(s,0) = \partial_t\big\vert_{t=0} c_p(s,t) = \diff{}(\Exp_p)_{(s\Log_p(q)}\left(s\cdot \diff{}(\Log_p)_{q}(v)\right)
\end{equation*}
For notational convenience, define $x = \Log_p(q), y = \diff{}(\Log_p)_{q}(v) \in T_p\mcM$.
The above variation vector field coincides with
\begin{equation*}
s\mapsto \dot c_p(s,0) = \partial_t\big\vert_{t=0} c_p(s,t) =
\partial_\tau\big\vert_{\tau = 0} \Exp_p(s \cdot (x + \tau y)).
\end{equation*}
This is a variation of the geodesic $s\mapsto \hat c_p(s) \coloneqq c_p(s,0)$ from $p$ to $q$ through geodesics, thus a Jacobi field.\footnote{
More precisely, the unique Jacobi field along the geodesic $c_p(s,0)$ that matches the initial values
$J(0) = \partial_\tau\big\vert_{\tau =0} \Exp_p(0\cdot(x+\tau y)) = \partial_\tau\big\vert_{\tau =0} p = 0$
and
\begin{align*}
J'(0)&= \covDeriv[s]\big\vert_{s =0} \partial_\tau\big\vert_{\tau =0} Exp_p(s \cdot (x + \tau y))
= \covDeriv[\tau]\big\vert_{\tau =0} \partial_s\big\vert_{s =0} Exp_p(s \cdot (x + \tau y)) \\
&= \covDeriv[\tau]\big\vert_{\tau =0} (x + \tau y) = y = \diff{}(\Log_p)_{q}(v).
\end{align*}
}
Hence, with $s\mapsto J(s) = \partial_\tau\big\vert_{\tau = 0} \Exp_p(s \cdot (x + \tau y)) $, we have
\begin{equation*}
J(s) = \dot c_p(s,0), \quad J'(s) = \covDeriv[s]J(s) = \covDeriv[s]\dot c_p(s,0).
\end{equation*}
Thus, we may write
\begin{align*}
\langle \Hess f(q)[v]\hspace{0.1cm},\hspace{0.1cm} v\rangle &=
\langle J'(1), v \rangle.
\end{align*}
|
\section{Introduction}
In the past decades, the error Backpropagation algorithm (BP) \cite{error_backpropagation} has developed into an indispensable technique to train Deep Neural Networks (DNNs) with high performance. However, BP has numerous drawbacks. First, it relies on the sequential propagation of errors through layers which hinders the ability to parallelise the backward pass and inherently slows down the gradient computation \cite{dfa_cnn_online_learning_processor}. Secondly, the transposes of the weights of downstream layers are needed to compute errors of hidden neurons. This requires costly data movement and increases the hardware energy consumption during training \cite{sparse_dfa, dfa_scales_to_modern_dl}. Finally, BP requires saving many intermediate results during inference for the backward pass, leading to high memory usage \cite{sparse_dfa, dfa_cnn_online_learning_processor}.
\par
Training DNNs is computationally expensive and requires high amounts of energy. To address this, massively-parallel dedicated neural processors can be designed to take advantage of the parallel aspect of neural circuits \cite{accelerator_for_deep_cnn, memristor_neural_network, mixed_signal_binarry_cnn_processor, dfa_cnn_online_learning_processor, sparse_dfa, energy_efficient_dnn_learning_processor, energy_efficient_ml}. Such hardware relies on data reuse and locality to avoid costly transport of information and to optimize power efficiency. Neural processors offer numerous advantages that can substantially reduce the energy footprint of DNNs and decrease the latency of data processing. This is especially beneficial in edge devices where energy consumption, computation time and privacy are major challenges. \cite{dl_in_edge_computing}. However, the non-local characteristics of BP make it unsuitable for efficient online learning on dedicated neural processors \cite{dfa_cnn_online_learning_processor, sparse_dfa} and alternatives to BP are needed to optimize the speed, energy efficiency and memory usage of training on hardware \cite{dfa_scales_to_modern_dl}.
\par
Motivated by this, several local alternatives to BP have been proposed. These include unsupervised learning methods such as Deep Infomax \cite{deep_infomax} and Greedy InfoMax \cite{greedy_infomax} that optimize the mutual information of each layer in the network. Other techniques like synthetic gradients \cite{synthetic_gradients} sidestep error backpropagation by creating decoupled neural interfaces that predict gradients or local error signals that define local layerwise loss functions \cite{deep_supervised_learning_using_local_errors, local_error_signals} Another alternative to BP is Random Feedback Alignment (FA) \cite{random_fa}. It uses random feedback matrices during the backward pass to create a propagation path dissociated from the forward weights which removes the need for weight transport.
\par
The Random Direct Feedback Alignment (DFA) algorithm takes the idea of FA one step further by directly projecting output errors to hidden layers using fixed linear random feedback connections \cite{random_dfa, efficient_cnn_with_dfa, sparse_dfa, dfa_cnn_online_learning_processor, dfa_scales_to_modern_dl}. Since the backward computation only relies on linear projections of output errors and local gradients, DFA enables the efficient parallelization of the backward pass or \textit{backward unlocking} \cite{random_dfa, dfa_scales_to_modern_dl}. DFA also demonstrates improved training speed and lower energy consumption on dedicated neural processors compared to BP \cite{energy_efficient_dnn_learning_processor}. However, even though DFA easily scales to modern deep learning architectures such as Transformers \cite{efficient_cnn_with_dfa, dfa_cnn_online_learning_processor, dfa_scales_to_modern_dl}, a performance gap exists with BP which increases with the complexity of the task and becomes catastrophic when DFA is applied to Convolutional Neural Networks (CNNs) \cite{dfa_cnn_online_learning_processor, learning_connections_in_dfa, align_then_memorise}.
\par
One possible approach to improve the performance of FA is to learn feedback connections. Examples are methods like weight mirroring \cite{deep_learning_without_weight_transport} and the Kolen-Pollack algorithms \cite{deep_learning_without_weight_transport, learning_connections_in_dfa}. These algorithms adapt feedback connections by either learning the transpose of weights or reciprocally adjusting feedbacks according to the changes of forward weights \cite{deep_learning_without_weight_transport, learning_connections_in_dfa}. Feedback learning provides improved alignment between forward and feedback weights, leading to better descending directions as well as higher performance compared to fixed random connections. However, weight mirroring and the Kolen-Pollack algorithm require fine hyperparameter balance to be stable \cite{two_routes_to_credit_assignment} and suffer from gradient explosion when used on CNNs without Batch Normalization (BN) \cite{learning_connections_in_dfa}. To the best of our knowledge, only the Kolen-Pollack method has been adapted as a DFA algorithm, henceforth referred to as Direct Kolen-Pollack (DKP) \cite{learning_connections_in_dfa}.
\par
To improve DFA, we conjecture that feedback connections are ideal when they approximate the partial derivatives of outputs with respect to hidden neurons. Based on this assumption, we introduce a novel algorithm that uses Forward-Mode Automatic Differentiation (Forward-Mode AD) to estimate backpropagation paths and learn feedback connections in an entirely online manner. Our main contributions are:
\begin{enumerate}
\item We describe how directional derivatives evaluated in random directions can be used to estimate backpropagation paths from outputs to hidden neurons during inference and propose a corresponding feedback learning rule for DFA. The resulting algorithm, \textit{Directional DFA}, benefits from backward unlocking of DFA while learning feedbacks in an online manner without having to perform backpropagation.
\item We show that, unlike DKP, Directional DFA does not suffer from gradient explosion when used on CNNs and is capable of training shallow CNNs even without the help of batch normalization.
\item We scale Directional DFA to deep convolutional architectures (AlexNet) and benchmark on CIFAR100 to determine whether our method can solve the failure of Random DFA to train deep CNNs. We show that our method significantly improves performance compared to Random DFA but that even when feedbacks approximate backpropagation paths, DFA still fails to train deep CNNs as a gap with BP remains. However, we show that transfer learning allows closing this gap and that Directional DFA still performs the best among DFA methods.
\end{enumerate}
\section{Background} \label{section:method}
In this section, we briefly review the backpropagation algorithm to highlight the causes of backward locking and describe several concepts that are used in our Directional DFA method.
\subsection{Deep Neural Network}
Let $L$ be the depth of a Deep Neural Network and $n_l$ the number of neurons in the layer $l$. We denote by $\boldsymbol{y}_0 \in \mathbb{R}^{n_0}$ the inputs of the network, $W^{(l)} \in \mathbb{R}^{n_l \times n_{l-1}}$ the weights matrix defining the synaptic strengths between the layer $l$ and its preceding $l-1$. The outputs $\boldsymbol{y}^{(l)}$ of the layer $l$ are thus defined as follows:
\begin{equation}
\boldsymbol{a}^{(l)} = W^{(l)} \boldsymbol{y}^{(l-1)}, \quad \boldsymbol{y}^{(l)} = \sigma\left(\boldsymbol{a}^{(l)}\right)
\end{equation}
where $\boldsymbol{a}^{(l)}$ are the preactivations of the layer and $\sigma(x)$ is a pointwise differentiable activation function. We will thus denote by $\sigma^{\prime}\left(x\right)$ its first order derivative, $\frac{\partial \sigma\left(x\right)}{\partial x}$.
\subsection{Error Backpropagation}
We now briefly review the error backpropagation algorithm.
\par
Given any differentiable loss function $\mathcal{L}$, the weight gradient of the layer $l$ is defined by all the partial derivatives of $\mathcal{L}$ with respect to the weights, such as:
\begin{equation}
\begin{split}
\frac{\partial \mathcal{L}}{\partial w^{(l)}_{ij}} &= \delta^{(l)}_{i} \frac{\partial y^{(l)}_{i}}{\partial w^{(l)}_{ij}}
\end{split}
\end{equation}
where
\begin{equation}\label{eq:bp_error}
\begin{split}
\delta^{(l)}_{i} &:= \sum_{o=1}^{n_L} \frac{\partial \mathcal{L}}{\partial y^{(L)}_{o}} \frac{\partial y^{(L)}_o}{\partial y^{(l)}_i}
\end{split}
\end{equation}
is the error associated with the neuron, $\frac{\partial \mathcal{L}}{\partial y^{(L)}_{o}}$ is the error signal at the outputs, provided by the loss module, and
\begin{equation}
\frac{\partial y^{(l)}_{i}}{\partial w^{(l)}_{ij}} = \sigma^{\prime}\left(a^{(l)}_i\right) y^{(l-1)}_j
\end{equation}
is the local partial derivative that only depends on information directly available to the neuron, i.e. inputs, pre-activation and activation.
\par
For output neurons, $\delta^{(L)}_{i}$ is simply equal to the error signal as no backpropagation is required. For hidden neurons, errors are sequentially backpropagated from downstream layers using the chain rule, such as:
\begin{equation} \label{eq:bp_chain_rule}
\begin{split}
\delta^{(l)}_{i} &= \begin{cases}
\frac{\partial \mathcal{L}}{\partial y^{(L)}_{o}} & \text{ if } l=L \\
\sum_{k=1}^{n_{l+1}} \frac{\partial y^{(l+1)}_k}{\partial y^{(l)}_i} \delta^{(l+1)}_{k} & \text{ if } l<L
\end{cases}
\end{split}
\end{equation}
\par
Equation \ref{eq:bp_chain_rule} shows the cause of backward locking in BP. To compute errors of specific hidden neurons, the computation of downstream's neurons errors is required.
This sequential propagation during the backward pass locks the computation of layers' gradients and limits the parallelization of BP. To solve this problem, new approaches have been proposed such as the Direct Feedback Alignment algorithm.
\subsection{Direct Feedback Alignement}
\begin{figure}
\centering
\subfigure[Error Backpropagation]{\label{fig:error_backpropagation} \resizebox{0.9\linewidth}{!}{\input{./Figures/backpropagation.tex}}}
\subfigure[Direct Feedback Alignment]{\label{fig:dfa}\resizebox{0.9\linewidth}{!}{\input{./Figures/dfa.tex}}}
\vskip 0.15in
\caption{Illustrations of the error backpropagation (Figure \ref{fig:error_backpropagation}) and Direct Feedback Alignment (Figure \ref{fig:dfa}). Solid arrows represent forward paths and dotted arrows represent backpropagation paths.}
\label{fig:backprop_vs_dfa}
\vskip -0.1in
\end{figure}
In Direct Feedback Alignment (DFA) \cite{random_dfa}, the partial derivatives $\frac{\partial y^{(L)}_o}{\partial y^{(l)}_i}$ in Equation \ref{eq:bp_error} are replaced by fixed linear feedback connections. in this way, DFA sidesteps error backpropagation as the computation of the chain rule in Equation \ref{eq:bp_chain_rule} is avoided by linearly projecting errors from output to hidden neurons. Figure \ref{fig:backprop_vs_dfa} shows a comparative illustration of BP and DFA.
By denoting $B^{(l)} \in \mathbb{R}^{n_l \times n_L}$ as the feedback matrix that replaces the backpropagation paths between the outputs and the hidden layers $l$, the error $\bar{\delta}^{(l)}_i$ received by the neurons $i$ can be written as follows \cite{random_dfa}:
\begin{equation} \label{eq:feedback_error}
\begin{split}
\bar{\delta}^{(l)}_i &= \sum_{o=1}^{n_L} \frac{\partial \mathcal{L}}{\partial y^{(L)}_{o}} b^{(l)}_{io}
\end{split}
\end{equation}
For the output layer, the feedback matrix $B^{(L)}$ is set as the identity matrix. The gradient of the output layer is thus identical to BP as errors are directly available and no error projection or backpropagation is required. For hidden layers, feedback connections are commonly chosen to be fixed random matrices. In this case, the method is referred as \textit{Random DFA} \cite{random_dfa}. Learning with random feedbacks can seem counter-intuitive. However, DNNs can align their weights with the feedbacks, thus aligning their gradients with the true gradients. If the approximate gradient lies within $90^{\circ}$ of the true gradient, it thus provides a descending direction for small enough learning rates \cite{random_fa, random_dfa, align_then_memorise}.
\subsection{Forward-Mode Automatic Differentiation}
\begin{figure*}
\centering
\subfigure[\label{fig:directional_derivative}]{\resizebox{0.3\textwidth}{!}{\input {Figures/directional_derivative.tikz}}}
\subfigure[\label{fig:directional_gradient}]{\resizebox{0.3\textwidth}{!}{\input {Figures/directional_gradient.tikz}}}
\vskip 0.15in
\caption{Figure \ref{fig:directional_derivative}: Projection of the gradient $\boldsymbol{\nabla f(w)}$ at $\boldsymbol{w}$ onto a given direction $\boldsymbol{v}$. The vector $d_{\boldsymbol{v}}f(\boldsymbol{w}) \boldsymbol{v}$ is obtained by scaling the direction $\boldsymbol{v}$ by the directional derivative $d_{\boldsymbol{v}}f(\boldsymbol{w})$ evaluated at $\boldsymbol{w}$ in the direction of $\boldsymbol{v}$. Figure \ref{fig:directional_gradient}: The expected directional derivative (green arrow), computed by averaging directional gradients (red arrows) over many random directions (black arrows), is an unbiased estimate of the true gradient (blue arrow).}
\label{fig:directional_derivative_gradient}
\vskip -0.1in
\end{figure*}
Contrary to Reverse-Mode Automatic Differentiation (AD) which constructs a computational graph and requires backward evaluations to compute gradients, Forward-Mode AD simultaneously computes functions and directional derivatives in a completely forward way \cite{review_ad, learning_by_directional_gd, gradient_without_backprop}.
\par
Let $f(\boldsymbol{w})$ be a function of parameters $\boldsymbol{w}$. Using Forward-Mode AD simultaneously computes $f(\boldsymbol{w})$ and the directional derivative $d_{\boldsymbol{v}}f(\boldsymbol{w})$ evaluated in a given direction $\boldsymbol{v}$, such as:
\begin{equation} \label{eq:directional_derivative}
d_{\boldsymbol{v}}f(\boldsymbol{w}) = \boldsymbol{\nabla} f(\boldsymbol{w}) \; \boldsymbol{v} = \sum_{i=1}^{n} \frac{\partial f(\boldsymbol{w})}{\partial w_i} v_i
\end{equation}
where $\boldsymbol{\nabla} f(\boldsymbol{w})$ is the true gradient of $f$ evaluated at $\boldsymbol{w}$.
\par
The directional derivative $d_{\boldsymbol{v}}f(\boldsymbol{w})$ is a scalar representing the rate of change of $f$ along the direction $\boldsymbol{v}$ when evaluated at $\boldsymbol{w}$. Geometrically, scaling the given direction $\boldsymbol{v}$ by the directional derivative $d_{\boldsymbol{v}}f(\boldsymbol{w})$ gives the projection of the gradient $\boldsymbol{\nabla f(w)}$ at $\boldsymbol{w}$ onto $\boldsymbol{v}$ – see Figure \ref{fig:directional_derivative}.
\par
In Forward Mode AD, $\boldsymbol{v}$ is usually set with a one-hot vector to compute the exact partial derivative with respect to a single parameter. If $\boldsymbol{v}$ is a one hot vectors with all elements set to 0 except for the element $j$ that is set to 1, Equation \ref{eq:directional_derivative} becomes:
\begin{equation}
d_{\boldsymbol{v}}f(\boldsymbol{w}) = \frac{\partial f(\boldsymbol{w})}{\partial w_j}
\end{equation}
To compose a full gradient, $n$ directional derivatives must be computed with $n$ one hot vector corresponding to each element of $\boldsymbol{w}$. Therefore, this requires $n$ forward differentiation to obtain the full gradient which is more computationally expensive than Reverse-Mode AD, especially when the number of parameters is large as in DNNs \cite{gradient_without_backprop}.
\subsection{Directional Gradient}
An alternative to the use of one-hot vectors in Forward Mode AD has been independently proposed in \cite{learning_by_directional_gd} and \cite{gradient_without_backprop}.
Instead of evaluating the directional derivative in the direction of a single parameter at a time using a one-hot vector, the directional derivative $d_{\boldsymbol{v}}f(\boldsymbol{w})$ is evaluated in a random direction $\boldsymbol{v} \sim \mathcal{N}\left(\boldsymbol{0}, \boldsymbol{I}\right)$ following a multivariate standard normal distribution. The directional gradient (also called forward gradient in \cite{gradient_without_backprop}) is therefore defined by the product $d_{\boldsymbol{v}}f(\boldsymbol{w}) \boldsymbol{v}$ between the directional derivative $d_f(\boldsymbol{w}, \boldsymbol{v})$ and the direction vector $\boldsymbol{v}$.
\par
Because of the nature of $\boldsymbol{v}$, we can demonstrate that the expected directional gradient $\mathbb{E}\left[d_{\boldsymbol{v}}f(\boldsymbol{w}) \boldsymbol{v}\right]$ is an unbiased estimation of the true gradient $\boldsymbol{\nabla} f(\boldsymbol{w})$ when averaged over many different random directions \cite{learning_by_directional_gd, gradient_without_backprop}. Each element $i$ of the expected directional gradient can be decomposed as follows:
\begin{equation} \label{eq:expected_directional_gradient}
\begin{split}
\mathbb{E}\left[d_{\boldsymbol{v}}f(\boldsymbol{w}) v_i\right] =& \sum_{j=1}^{n} \frac{\partial f(\boldsymbol{w})}{\partial w_j} v_j v_i \\
=& \frac{\partial f(\boldsymbol{w})}{\partial w_i} \mathbb{E}\left[v_i^2\right] + \sum_{j \neq i} \frac{\partial f(\boldsymbol{w})}{\partial w_j} \mathbb{E}\left[v_i v_j\right]
\end{split}
\end{equation}
As each element $v_i$ follows a normal distribution of mean 0 and variance 1, $\mathbb{E}\left[v_i\right] = 0$ and $\mathbb{E}\left[v_i^2\right] = \mathbb{E}\left[v_i\right] + \text{Var}\left[v_i\right] = 1$. Therefore, Equation \ref{eq:expected_directional_gradient} reduces to the partial derivative with respect to $w_i$:
\begin{equation}
\begin{split}
\mathbb{E}\left[g_i(\boldsymbol{w}, \boldsymbol{v})\right]
=& \frac{\partial f(\boldsymbol{w})}{\partial w_i} \\
\Leftrightarrow \mathbb{E}\left[\boldsymbol{g(w, v)}\right] =& \boldsymbol{\nabla} f(\boldsymbol{w})
\end{split}
\end{equation}
Figure \ref{fig:directional_gradient} illustrates the expected directional gradient.
\par
However, this method does not perform as well as backpropagation and DFA \cite{learning_by_directional_gd, gradient_without_backprop} because the averaging process requires a large number of random directions.
\section{Method}
In this section, we motivate and describe the iterative feedback learning rule that underpins Directional DFA.
\subsection{Feedback Connections Represent Backpropagation Paths}
To improve DFA, we are interested in learning feedback connections instead of using fixed random weights. Equation \ref{eq:feedback_error} shows that the feedback matrix $B$ replaces the backpropagation paths in Equation \ref{eq:bp_error}, represented by the partial derivatives of the outputs with respect to the hidden neurons. Unlike partial derivatives in BP, feedback connections in DFA are not sample-dependent as they are used for every training sample. We are thus working with the heuristics that an ideal feedback connection $b^{(l)}_{io}$ between the output neuron $o$ and the hidden neuron $i$ in layer $l$ would work best if it approximates the partial derivative $\frac{\partial y^{(L)}_o}{\partial y^{(l)}_i}$ for all the samples of the training dataset, such as:
\begin{equation} \label{eq:optimal_feedback}
b_{io} = \mathbb{E}\left[\frac{\partial y^{(L)}_o}{\partial y^{(l)}_i}\right]
\end{equation}
Reverse-Mode AD could be used to compute the expected partial derivatives over the dataset. However, computing Equation \ref{eq:optimal_feedback} using Reverse-Mode AD is impractical when the dataset contains a large number of samples or in online learning where data is not available at all times. Additionally, Reverse-Mode AD is inefficient, non-local and leads to backward locking. We thus propose an alternative method to estimate these expected partial derivatives in a fully online manner by using Forward-Mode AD.
\subsection{Estimating Backpropagation Paths}
\begin{algorithm}[tb]
\caption{Directional Direct Feedback Alignment algorithm on a feedforward DNN. Here, the operator $\odot$ corresponds to the Hadamard product.}
\label{alg:dir_dfa}
\begin{algorithmic}
\STATE {\bfseries Input:} Training data $D=\left\{\boldsymbol{x_s}\right\}^{M}_{s=1}$
\STATE Randomly initialize $w^{(l)}_{ij}$ for all $l$, $i$ and $j$.
\STATE Initialize $b^{(l)}_{io} = 0$ for all $l$, $i$ and $o$.
\REPEAT
\STATE \COMMENT{Inference (sequential)}
\FOR{$s=1$ {\bfseries to} $M$}
\STATE $\boldsymbol{y^{(0)}} \leftarrow \boldsymbol{x_s}$
\STATE $\boldsymbol{d^{(0)}} \leftarrow \boldsymbol{0}$
\FOR{$l=1$ {\bfseries to} $L$}
\STATE Sample $\boldsymbol{v^{(l)}} \sim \mathcal{N}\left(\boldsymbol{0}, \boldsymbol{I}\right)$
\STATE $\boldsymbol{a^{(l)}} \leftarrow W^{(l)} \boldsymbol{y^{(l-1)}}$
\STATE $\boldsymbol{y^{(l)}} \leftarrow f\left(\boldsymbol{a^{(l)}}\right)$
\STATE $\boldsymbol{d^{(l)}} \leftarrow \left(W^{(l)} \boldsymbol{d^{(l-1)}}\right) \odot \sigma^{\prime}\left(\boldsymbol{a^{(l)}}\right) + \boldsymbol{v^{(l)}}$
\STATE $G^{(l)} \leftarrow \sigma^{\prime}\left(\boldsymbol{a^{(l-1)}}\right) \boldsymbol{y^{(l-1)}}^T$
\ENDFOR
\STATE $\boldsymbol{e} \leftarrow \frac{\partial \mathcal{L}}{\partial \boldsymbol{y^{(L)}}}$
\STATE \COMMENT{Weights and feedbacks updates (parallel)}
\FOR{$l=1$ {\bfseries to} $L$}
\STATE $\nabla W^{(l)} \leftarrow G^{(l)} \odot \left(B^{(l)}\boldsymbol{e}\right)$
\STATE $\nabla B^{(l)} \leftarrow B^{(l)} - \left(\boldsymbol{v^{(l)}} {\boldsymbol{d^{(L)}}}^T\right)$
\STATE $W^{(l)} \leftarrow W^{(l)} - \lambda \nabla W^{(l)}$
\STATE $B^{(l)} \leftarrow B^{(l)} - \alpha \nabla B^{(l)}$
\ENDFOR
\ENDFOR
\UNTIL{$\mathcal{L}<\epsilon$}
\end{algorithmic}
\end{algorithm}
One way to compute gradients while avoiding the limitations of Reverse-Mode AD is to use Forward-Mode AD, as described in \cite{learning_by_directional_gd} and \cite{gradient_without_backprop}. In this method, directions are randomly drawn for each weight of the network which are used to estimate the gradient using Forward-Mode AD. In Directional DFA, we take a different approach. Rather than directly estimating the full gradient, we estimate the expected backpropagation paths as described by Equation \ref{eq:optimal_feedback}. By sampling random directions for the output of each hidden neuron in the network, we can use Forward-Mode AD to estimate the partial derivatives of the output of the network with respect to the hidden neurons in an online manner.
In the following, we thus show how this applies to a DNN. We denote by $\boldsymbol{v}^{(l)} \in \mathbb{R}^{n_l}$ the direction vector for each layer of the network. For each hidden layer $l$, we draw $\boldsymbol{v}^{(l)} \sim \mathcal{N}\left(\boldsymbol{0}, \boldsymbol{I}\right)$ from a multivariate normal distribution. For input and output neurons, the direction vectors $\boldsymbol{v}^{(0)}$ and $\boldsymbol{v}^{(L)}$ respectively are set to 0 as they do not require differentiation. Using Forward-Mode AD we then compute the directional derivative $d^{(l)}_i$ associated with the neuron $i$ of layer $l$:
\begin{equation} \label{eq:directional_derivative_neuron}
d^{(l)}_i := \sum_{j=1}^{n_{l-1}} \frac{\partial y^{(l)}_i}{\partial y^{(l-1)}_j} d^{(l-1)}_j + v^{(l)}_i
\end{equation}
The first term on the right-hand side of Equation \ref{eq:directional_derivative_neuron} applies the chain rule to the directional derivatives of upstream neurons to perform first-order differentiation. The second term corresponds to the drawn direction which allows differentiation in downstream neurons. Note that Equation \ref{eq:directional_derivative_neuron} only contains locally available information and can therefore be computed in an online manner along with the forward pass.
\par
After the inference, a directional derivative $d^{(L)}_o$ is produced along with the activation of each output neuron $o$. Each directional derivative represents the rates of change of the output $o$ in the direction:
\begin{equation}
d^{(L)}_o = \sum_{l=1}^{L-1} \sum_{j=1}^{n^{(l)}} \frac{\partial y^{(L)}_o}{\partial y^{(l)}_j} v^{(l)}_j
\end{equation}
Following \cite{learning_by_directional_gd, gradient_without_backprop} we can show that the products between the output directional derivatives and the drawn directions are, on average, unbiased estimates of the expected backpropagation paths described by Equation \ref{eq:optimal_feedback}:
\begin{equation} \label{eq:dir_derivative_dnn}
\begin{split}
\mathbb{E}\left[d^{(L)}_o v^{(l)}_i\right] &= \mathbb{E}\left[ \sum_{l^{\prime}=1}^{L-1} \sum_{j=1}^{n^{(l^{\prime})}} \frac{\partial y^{(L)}_o}{\partial y^{(l^{\prime})}_j} v^{(l^{\prime})}_j v^{(l)}_i \right] \\
&= \sum_{l^{\prime}=1}^{L-1} \sum_{j=1}^{n^{(l^{\prime})}} \mathbb{E}\left[\frac{\partial y^{(L)}_o}{\partial y^{(l^{\prime})}_j}\right] \mathbb{E}\left[v^{(l^{\prime})}_j v^{(l)}_i\right] \\
&= \mathbb{E}\left[\frac{\partial y^{(L)}_o}{\partial y^{(l)}_i}\right]
\end{split}
\end{equation}
as
\begin{equation}
\begin{split}
\mathbb{E}\left[v^{(l^{\prime})}_j v^{(l)}_i\right] &= \begin{cases}
0 & \text{ if } l^{\prime} \neq l \text{ and } i \neq j \\
1 & \text{ if } l^{\prime} = l \text{ and } i = j
\end{cases}
\end{split}
\end{equation}
Therefore, this estimate can be used as a feedback connection to directly project errors from the output neuron $o$ onto the hidden neurons $i$ of layer $l$, such as:
\begin{equation} \label{eq:directional_feedback}
b_{io} = \mathbb{E}\left[d^{(L)}_o v^{(l)}_i\right]
\end{equation}
Computing the expectation in Equation \ref{eq:directional_feedback} is still impractical due to the averaging over the entire dataset as well as many random directions. Moreover, this estimate becomes obsolete as the network weights are updated which induces changes in the partial derivatives. To relax the computational cost of this estimate and to allow adaptation to weight updates, we define a learning rule for the feedback that implements a moving average of $d^{(L)}_o v^{(l)}_i$ over the past sample evaluations, such as:
\begin{equation} \label{eq:feedback_moving_average}
b^{(l)}_{io} \leftarrow b^{(l)}_{io} + \alpha \left(d^{(L)}_o v^{(l)}_i - b^{(l)}_{io}\right)
\end{equation}
where $0 < \alpha < 1$ is the feedback learning rate.
\par
To be able to use more complex optimization techniques such as Adam \cite{adam}, we can define a feedback gradient as follows:
\begin{equation}
\nabla b^{(l)}_{io} := b^{(l)}_{io} - d^{(L)}_o v^{(l)}_i
\end{equation}
and update feedbacks with gradient descent:
\begin{equation}
b^{(l)}_{io} \leftarrow b^{(l)}_{io} - \alpha \nabla b^{(l)}_{io}
\end{equation}
Algorithm \ref{alg:dir_dfa} summarizes the full algorithm applied to a fully-connected DNN.
\par
We have now defined a rule to update the feedback matrices. Based on the results in \cite{learning_by_directional_gd} and \cite{gradient_without_backprop} discussed above, we expect that the feedback matrices will, over time, approach the expected partial derivatives described by Equation \ref{eq:optimal_feedback} and thus lead to better performing learning in DFA. Note that all computation relies only on Forward-Mode AD which is entirely performed in an online manner without backpropagation. Moreover, Equation \ref{eq:feedback_moving_average} only requires locally-available information such as the output directional derivative, the draw direction of the hidden neuron and the feedback itself. Therefore, the proposed feedback update is a local learning rule that does not require any transport of information.
\section{Results}
We conducted several experiments to measure the improvements of the proposed approach over random DFA and DKP. We first trained DNNs with different depth as well as CNNs on MNIST \cite{mnist}, FashionMNIST \cite{fashion_mnist} CIFAR10 and CIFAR100 \cite{cifar10} and report average test performances for comparison.
Secondly, we trained again each network with Batch Normalization (BN) to measure the impact on performance. Finally, we compare the gradient alignment between each DFA method and the true gradient provided by BP to measure the quality of the weight updates.
\subsection{Experimental Settings}
In our experiments, we trained three fully-connected DNNs with an increasing number of hidden layers (1, 2 and 3 hidden layers) on every dataset. Additionally, we trained a CNN with the following architecture: 15C5-P2-40C5-P2-128-10 where 15C5 represents 15 5x5 convolutional layers and P2 represents a 2x2 max pooling layer. For CIFAR100, we used the AlexNet architecture \cite{alexnet} and up-scaled images to 224x224.
Each network was trained 10 times with BP, random DFA, DKP and our Directional DFA algorithms. We used the following set of hyperparameters for training.
\par
All the inputs were normalized to lie between -1 and 1. No data augmentation was applied except for CIFAR100 where a random crop was used. We used a ReLU activation function and a cross-entropy loss for every layer except for the outputs that remained linear. For random DFA, the weights and biases were initialized to zero, as suggested in \cite{random_fa, random_dfa}, while the feedback matrices were initialized with the uniform Kaiming initialization \cite{kaiming}. For DKP and our directional DFA algorithm, the feedback matrices were initialized to zero while forward weights were initialized with uniform Kaiming. We trained the networks with a batch size of 64 for every dataset except for AlexNet on CIFAR100 where a batch size of 128 was used. Training was performed during 100 epochs using Adam \cite{adam} with $\beta_1=0.9$, $\beta_2=0.999$, $\epsilon=10^{-8}$ and a learning rate $\lambda=10^{-4}$. A learning rate of $\lambda=5 \times 10^{-4}$ was used for the shallow CNNs. We also used Adam for feedback learning with a learning rate of $\alpha=10^{-4}$ in both DKP and directional DFA. For DKP, both forward and feedback weights decay factors were set to $10^{-6}$. In every training, we used a learning rate scheduler for the forward learning rate $\lambda$ with a decay factor of $0.95$ applied at the end of each epoch.
\subsection{Performance Comparison}
\begin{table*}[t]
\caption{Performance comparison of BP, random DFA, DKP and directional DFA on the MNIST, Fashion MNIST and CIFAR10 datasets with different network architectures \textbf{without} batch normalization. The performance of BP is given as a reference and only DFA methods are compared. Best performances between random DFA, DKP and directional DFA are given in bold. 3x800 means a network architecture of 3 fully-connected layers of 800 neurons. \textit{Conv} refers to the shallow CNN.}
\label{table:performances_no_bn}
\vskip 0.15in
\begin{center}
\begin{small}
\begin{sc}
\renewcommand*{\arraystretch}{1.2}
\begin{tabular}{l|l|l|lll}
\toprule
Dataset & Arch & BP & Random DFA & DKP & Dir. DFA \\
\midrule
& 1x800 & 98.25 $\pm$ 0.04\% & 98.04 $\pm$ 0.07\% & 98.12 $\pm$ 0.05\% & \textbf{98.21 $\pm$ 0.02\%} \\
MNIST & 2x800 & 98.39 $\pm$ 0.06\% & 98.18 $\pm$ 0.06 & 98.28 $\pm$ 0.06\% & \textbf{98.3 $\pm$ 0.04\%} \\
& 3x800 & 98.46 $\pm$ 0.05\% & 98.15 $\pm$ 0.05\% & 98.27 $\pm$ 0.05\% & \textbf{98.32 $\pm$ 0.05\%} \\
& Conv & 99.32 $\pm$ 0.03\% & 98.8 $\pm$ 0.08\% & 98.62 $\pm$ 0.13\% & \textbf{99.16 $\pm$ 0.03\%} \\
\midrule
& 1x800 & 89.37 $\pm$ 0.09\% & 88.67 $\pm$ 0.12\% & 89.09 $\pm$ 0.11\% & \textbf{89.27 $\pm$ 0.07\%} \\
Fashion & 2x800 & 90.01 $\pm$ 0.08\% & 89.19 $\pm$ 0.13\% & 89.63 $\pm$ 0.12\% & \textbf{89.82 $\pm$ 0.13\%} \\
MNIST & 3x800 & 89.99 $\pm$ 0.15\% & 89.15 $\pm$ 0.09\% & 89.48 $\pm$ 0.09\% & \textbf{89.56 $\pm$ 0.11\%} \\
& Conv & 92.1 $\pm$ 0.16\% & 89.69 $\pm$ 0.22\% & 86.21 $\pm$ 1.28\% & \textbf{91.54 $\pm$ 0.11\%} \\
\midrule
& 1x1000 & 56.2 $\pm$ 0.12\% & 54.8 $\pm$ 0.22\% & 55.41 $\pm$ 0.18\% & \textbf{55.74 $\pm$ 0.14\%} \\
CIFAR10 & 2x1000 & 56.66 $\pm$ 0.18\% & 54.2 $\pm$ 0.11\% & 55.15 $\pm$ 0.24\% & \textbf{55.7 $\pm$ 0.15\%} \\
& 3x1000 & 56.92 $\pm$ 0.19\% & 53.61 $\pm$ 0.16\% & 54.35 $\pm$ 0.21\% & \textbf{54.97 $\pm$ 0.19\%} \\
& Conv & 68.11 $\pm$ 0.57\% & 58.14 $\pm$ 0.94\% & 41.02 $\pm$ 2.63\% & \textbf{66.96 $\pm$ 0.70\%} \\
\bottomrule
\end{tabular}
\end{sc}
\end{small}
\end{center}
\vskip -0.1in
\end{table*}
\begin{table*}[t]
\caption{Performance comparison of BP, random DFA, DKP and directional DFA on the MNIST, Fashion MNIST, CIFAR10 and CIFAR100 datasets with different network architectures \textbf{with} batch normalization. * transfer learning was used by pre-training convolution layers and freezing them during the training of the fully-connected layers.}
\label{table:performances_bn}
\vskip 0.15in
\begin{center}
\begin{small}
\begin{sc}
\renewcommand*{\arraystretch}{1.2}
\begin{tabular}{l|l|l|lll}
\toprule
Dataset & Arch & BP & Random DFA & DKP & Dir. DFA \\
\midrule
& 1x800 & 98.28 $\pm$ 0.05\% & 98.19 $\pm$ 0.05\% & \textbf{98.3 $\pm$ 0.04\%} & 98.26 $\pm$ 0.03\% \\
MNIST & 2x800 & 98.58 $\pm$ 0.06\% & 98.41 $\pm$ 0.06\% & 98.39 $\pm$ 0.05\% & \textbf{98.52 $\pm$ 0.06\%} \\
& 3x800 & 98.7 $\pm$ 0.06\% & 98.46 $\pm$ 0.04\% & 98.5 $\pm$ 0.06\% & \textbf{98.63 $\pm$ 0.05\%} \\
& Conv & 99.4 $\pm$ 0.05\% & 99.22 $\pm$ 0.05\% & 99.23 $\pm$ 0.04\% & \textbf{99.44 $\pm$ 0.03\%} \\
\midrule
& 1x800 & 90.12 $\pm$ 0.12\% & 90.0 $\pm$ 0.11\% & 90.08 $\pm$ 0.1\% & \textbf{90.09 $\pm$ 0.1\%} \\
Fashion & 2x800 & 90.43 $\pm$ 0.08\% & 90.05 $\pm$ 0.14\% & 89.92 $\pm$ 0.13\% & \textbf{90.16 $\pm$ 0.14\%} \\
MNIST & 3x800 & 90.87 $\pm$ 0.14\% & 89.95 $\pm$ 0.1\% & 90.05 $\pm$ 0.11\% & \textbf{90.25 $\pm$ 0.12\%} \\
& Conv & 91.86 $\pm$ 0.19\% & 90.86 $\pm$ 0.11\% & 91.46 $\pm$ 0.19\% & \textbf{92.12 $\pm$ 0.1\%} \\
\midrule
& 1x1000 & 56.72 $\pm$ 0.23\% & 56.11 $\pm$ 0.18\% & 56.51 $\pm$ 0.18\% & \textbf{56.55 $\pm$ 0.17\%} \\
CIFAR10 & 2x1000 & 57.32 $\pm$ 0.19\% & 56.04 $\pm$ 0.22\% & 55.89 $\pm$ 0.23\% & \textbf{56.56 $\pm$ 0.32\%} \\
& 3x1000 & 58.11 $\pm$ 0.29\% & 55.57 $\pm$ 0.34\% & 55.63 $\pm$ 0.21\% & \textbf{56.6 $\pm$ 0.19\%} \\
& Conv & 69.85 $\pm$ 0.36\% & 62.76 $\pm$ 0.57\% & 65.66 $\pm$ 0.97\% & \textbf{69.34 $\pm$ 0.54\%} \\
\midrule
CIFAR100 & AlexNet & 66.72 $\pm$ 0.28\% & 48.92 $\pm$ 0.55\% & 56.56 $\pm$ 0.37\% & \textbf{57.04 $\pm$ 0.22\%} \\
& AlexNet* & 66.94 $\pm$ 0.04\% & 66.09 $\pm$ 0.14\% & 66.45 $\pm$ 0.11\% & \textbf{66.60 $\pm$ 0.21\%} \\
\bottomrule
\end{tabular}
\end{sc}
\end{small}
\end{center}
\vskip -0.1in
\end{table*}
Table \ref{table:performances_bn} and \ref{table:performances_no_bn} summarizes the testing performances of all methods with and without batch normalization respectively. Here, BP performances are given as references and random DFA, DKP and our Directional DFA are compared.
\par
First of all, we observe that random DFA suffers from performance degradation compared to BP which is consistent with previous work \cite{random_dfa, dfa_cnn_online_learning_processor, learning_connections_in_dfa}. This accuracy gap becomes more important when the network contains convolutional layers, even if the network only contains a few convolutional layers.
\par
Compared to random DFA, DKP generally performs better on all tasks and architectures. These improvements are attributed to the greater alignment provided by the Kolen-Pollack feedback learning, as described in \cite{learning_connections_in_dfa}. However, a performance gap still exists with BP.
\par
The accuracy degradation seen in both random DFA and DKP is less important with the proposed Directional DFA method. Performances achieved by our approach are closer to BP than the other benchmarked methods on most of the tasks and architectures. It is thus clear that the proposed Directional DFA algorithm generally outperforms both random DFA and DKP.
\subsection{Impact of Batch Normalization}
\begin{figure*}
\centering
\subfigure[Without Batch Normalization\label{fig:test_acc_no_bn}]{\input {Figures/acc_conv_no_bn.tikz}}
\subfigure[With Batch Normalization\label{fig:test_acc_bn}]{\input {Figures/acc_conv_bn.tikz}}
\vskip 0.15in
\caption{Evolution of the test accuracy of the CNN trained with BP (black), random DFA (red), DKP (orange) and Directional DFA (blue) on CIFAR10 without BN (Figure \ref{fig:test_acc_no_bn}) and with BN (Figure \ref{fig:test_acc_bn}).}
\label{fig:test_accuracy}
\vskip -0.1in
\end{figure*}
Batch Normalization (BN) \cite{batch_norm} is widely used in today's DNNs. BN normalizes values of pre-activations to reduce internal covariate shift and facilitate training \cite{batch_norm}. The use of BN thus is known to improve the convergence speed and performance of DNNs. A drawback of BN is that it is memory and computationally expensive and strongly depends on the batch size which could be an obstacle to on-chip online learning.
\par
It has been reported in \cite{learning_connections_in_dfa} that DKP suffers from gradient explosion when BN is not used before the activation. During our experiments, we found that our method successfully trains CNNs without BN. We thus report the test accuracies with and without BN for all architectures in Table \ref{table:performances_bn} and \ref{table:performances_no_bn} respectively. BN globally improves convergence speed and test performance of all methods. However, even though DKP successfully trains fully-connected networks without BN, it fails to train CNNs and performs worse than random DFA. Figure \ref{fig:test_accuracy} shows the evolution of the test accuracies of CNNs with and without BN trained with the benchmarked methods on CIFAR10. We can observe that the model without BN trained with DKP fails to converge properly while Random and Directional DFA converges smoothly. Therefore, CNNs trained with our method are capable of learning without the help of BN which offers advantages in terms of memory, computation cost and online learning.
\subsection{Deep Convolutional Networks}
DFA is known to fail on Deep CNNs \cite{dfa_cnn_online_learning_processor, principled_training_of_nn_with_dfa, dfa_scales_to_modern_dl, learning_connections_in_dfa}. To evaluate the scalability of each method on deep CNNs, we trained a network with the AlexNet architecture \cite{alexnet} on CIFAR100. Results were reported in Table \ref{table:performances_bn}. BP achieves an average test performance of 66.72\%. However, Random DFA fails to train AlexNet properly as it only reaches 48.92\% of accuracy. Both DKP and our Directional DFA improve performances compared to Random DFA as they achieve 56.56\% and 57.04\% respectively, with a slight advantage for Directional DFA. Despite the improvements made by feedback learning, performances remain far from BP suggesting that the methods also fail to learn useful features in deeper layers. To confirm this assumption, we pre-trained AlexNet using BP and applied transfer learning \cite{transfer_learning} by freezing the convolution layers and training the 3-layers classifier using each method. As shown in Table \ref{table:performances_bn}, transfer learning brings performances of all DFA methods close to BP, confirming that DFA fails to train deep convolutional feature extractors even when feedbacks approximate backpropagation paths. Despite the increase in performance of every method, Directional DFA remains the approach that performs best among all DFA methods.
\subsection{Gradient Alignment}
\begin{figure*}[t]
\centering
\subfigure[Random DFA \label{fig:align_dfa}]{\input {Figures/alignment_DFA.tikz}}
\subfigure[DKP \label{fig:align_dkp}]{\input {Figures/alignment_DKP.tikz}}
\subfigure[Directional DFA \label{fig:align_dir_dfa}]{\input {Figures/alignment_DirDFA.tikz}}
\vskip 0.15in
\caption{Layerwise alignment between the approximate gradient provided by each method and the true gradient provided by BP. The angles are here given in degree and lower values mean better alignment. Directional DFA provides improved descending directions compared to random DFA and DKP as the layerwise alignment angles are significantly lower.}
\label{fig:alignment}
\vskip -0.1in
\end{figure*}
We now quantify the quality of the gradient provided by each DFA method. An approximate gradient that has a direction close to the true gradient can minimize the loss function faster than an approximate gradient that points away from the minima.
Therefore, to better understand how our approach compares to other DFA methods, we measured the gradient alignment of both methods during training of a 6-layer fully-connected DNN with BN on MNIST. More precisely, we measured the layerwise angle between the approximate and the gradient provided by BP using the cosine distance metric:
\begin{equation}
C\left(\boldsymbol{\nabla_{W^{(l)}}}, \boldsymbol{{\widehat{\nabla}_{W^{(l)}}}}\right) = \frac{180}{\pi} \arccos \left(\frac{\boldsymbol{\nabla_{W^{(l)}}} \boldsymbol{\widehat{\nabla}_{W^{(l)}}}}{\|\boldsymbol{\nabla_{W^{(l)}}}\| \; \|\boldsymbol{\widehat{\nabla}_{W^{(l)}}}\|}\right)
\end{equation}
where $\boldsymbol{\nabla_{W^{(l)}}}$ $\boldsymbol{\widehat{\nabla}_{W^{(l)}}}$ are respectively the flattened true and approximate gradients of layer $l$.
\par
Figure \ref{fig:alignment} shows the layerwise gradient alignment comparison of the trained DNN for each method. As expected from feedback learning, the approximate gradient produced by the DKP is more aligned with the true gradient than random DFA. However, the improvement is minor compared to the proposed Directional DFA. Our method produces approximate gradients that are closer to the true gradients than the ones produced by random DFA and DKP. This suggests that the quality of the Directional DFA gradient is greater than other DFA methods and that our method is capable of converging faster towards minimas.
\subsection{Computational Cost and Hardware Efficiency}
We consider here the computational cost and hardware efficiency of each method. First of all, Random DFA decreases the computational cost of training as no backpropagation is performed. This benefits hardware efficiency as only errors have to be transported and allows backward unlocking for faster weight updates and leads to improved energy consumption \cite{energy_efficient_dnn_learning_processor}.
\par
Both DKP and the proposed Directional DFA have a higher computational cost than Random DFA due to the additional feedback learning rule. Therefore, energy consumption would increase when implemented on hardware because of the additional memory writes that would be required to update feedbacks. In Directional DFA, directional derivatives are computed along with forward activations which double the amount of Multiply-Accumulate (MAC) operations required to perform matrix multiplications. Directional DFA is therefore the most computationally expensive DFA method during inference but data reuse would be possible on hardware as the method relies on locally-available information.
\par
Conversely, BP does not take advantage of data reuse. Many intermediate results have to be stored in memory for backpropagation and global data movements are required \cite{sparse_dfa}. As the majority of the energy consumption of dedicated neural processors comes from read and write from off-chip DRAM rather than MAC operations, BP is inherently limited in how energy efficient it can be made \cite{open_source_dl_accelerator, sparse_dfa}. In contrast, Directional DFA benefits from less transport of information. Consequently, we can expect Directional DFA to be more energy efficient than BP on appropriate hardware as the cost of data movement is higher than the cost of MAC operations \cite{open_source_dl_accelerator, sparse_dfa}.
\section{Discussion}
In this work, we proposed a novel algorithm to learn feedback connections in DFA based on Forward-Mode AD. By evaluating directional derivatives in random directions, Directional DFA provides an unbiased estimate of the expected backpropagation paths as feedbacks. As Forward-Mode AD is performed along with the network's forward pass and takes advantage of data reuse, our method would benefit from fewer data movements than BP and thus reduced energy consumption if implemented on suitable hardware. Moreover, because gradients rely on locally-available information as well as linearly projected error signals, the gradient computation does not require sequential propagation as in BP. Therefore, this unlocks the backward pass and allows weight updates to be performed entirely in parallel which reduces the computation time of training and allows increased data processing rate on hardware \cite{dfa_cnn_online_learning_processor}.
\par
In our experiments, we demonstrated that Directional DFA outperforms other DFA methods trained on MNIST, Fashion MNIST, CIFAR10 and CIFAR100 with both feedforward and convolutional neural networks, thus achieving performances that are closer to BP. We showed that, unlike the DKP algorithm, Directional DFA also achieves stable training on convolution layers without the use of batch normalization. As BN requires full batches to compute the mean and variance of its inputs, it can be an obstacle to online learning where data samples are only available one at a time. Therefore, our method offers a better advantage for on-the-edge online learning than DKP.
\par
We also observed that, like every other DFA method, Directional DFA still fails to train deep CNNs even if backpropagation paths are approximated. This empirically confirms previous findings that there is no general choice of feedbacks that allows alignment in convolution layers due to their constrained structure and their lack of flexibility \cite{principled_training_of_nn_with_dfa, align_then_memorise}. To overcome this limitation, it is common to use transfer learning \cite{energy_efficient_dnn_learning_processor, dfa_scales_to_modern_dl}. We showed that our approach also outperforms other DFA methods using transfer learning.
\par
In terms of computational cost, using fixed random feedback matrices is the least computationally expensive method for training DNNs with DFA. It has been shown \cite{dfa_cnn_online_learning_processor, energy_efficient_dnn_learning_processor} that dedicated neural processors that implement random DFA improve training speed (by a factor of 2) as well as energy-efficiency (by more than 35\%) compared to hardware that is optimized for BP.
We expect that the equivalent improvements for DKP and Directional DFA would be more modest. The reason for this is that feedback connections also need to be updated which would require additional circuitry. Moreover, the number of MAC operations required by Directional DFA during training is double compared to other methods as first-order differentiation has to be performed along with neuron activations. Therefore, Directional DFA would be the most expensive DFA method to implement on hardware. However, the amount of external memory access would remain lower than BP as most of the computation in Directional DFA takes advantage of data reuse. Hence the energy consumption of MAC operations is insignificant compared to external memory access \cite{open_source_dl_accelerator, sparse_dfa, energy_efficient_dnn_learning_processor}, we expect our approach to be more energy-efficient than BP while providing better performance than other DFA methods.
\section{Code Availability}
The code produced in this work will be made available at: \href{https://github.com/Florian-BACHO/DirDFA}{https://github.com/Florian-BACHO/DirDFA}
\bibliographystyle{IEEEtran}
|
\section{Introduction}
\vspace{-0.1cm}
Deep learning methods have demonstrated their tremendous potential when it comes to medical image segmentation. However, the success of most existing architectures relies on the availability of pixel-level annotations, which are difficult to produce \cite{challenges}. Furthermore, these methods are known to be inadequately equipped for distribution shifts. One example is for cross-modality generalization, which is useful when one imaging modality has insufficient training data. For instance, it could be required in segmentation tasks where annotations are sufficiently available for T1 and T2 images but lacking in other MRI sequences. This challenge can be addressed through semi-supervised domain-adaptive approaches, which learn to transfer the knowledge available in\newpage \noindent the ``source" modality from pixel-level labels to the ``target" imaging modality lacking annotations \cite{Domain}.
Several strategies can be employed to address the modality adaptation problem. For instance, generative models attempt to generalize to a target modality by performing unsupervised domain adaptation through image-to-image translation and image reconstruction. In \cite{disentangle}, by learning to translate between CT and MR cardiac images, the proposed model would jointly disentangle the domain specific and domain invariant features between each modality and train a segmenter from the domain invariant features, which can then be applied to the target modality. Other methods \cite{attent,SynSeg,Synergistic,self,constrained} also integrate this translation approach, but the segmenter is trained in an end-to-end manner, on the synthetic target images generated from the source modality using a CycleGAN \cite{cyclegan} model. While these methods perform well, they depend on the availability of pixel-level annotations for source modality images.
In this paper, we propose M-GenSeg, a novel training strategy as illustrated in Fig. \ref{diag1}, which achieves domain adaptation from a partially unannotated source modality to an unannotated target modality. This work extends GenSeg \cite{GenSeg}, a training method that uses image-level "diseased" or "healthy" labels for semi-supervised segmentation. Given these labels, the model imposes an image-to-image translation objective between the image domain presenting tumor lesions and the domain corresponding to an absence of lesions. Teaching the model to find and remove a tumor guides the model to segment the tumor. We incorporate cross-modality image segmentation by adding a second unsupervised objective: an image-to-image translation between source and target modalities. We hypothesize both objectives are complementary since GenSeg helps localizing the tumor on target images, while modality translation enables fine-tuning the segmentor on the target modality by displaying synthetic labelled target images. We evaluate M-GenSeg on a modified version of the BraTS 2020 dataset, in which each type of sequence (T1, T2, T1ce and FLAIR) is considered as a distinct modality. We demonstrate that our model can better generalize than other methods to the target modality. Furthermore, M-GenSeg is semi-supervised in the source modality.
\vspace{-0.3cm}
\section{Methods}
\vspace{-0.2cm}
\subsection{M-GenSeg : semi-supervised segmentation}
\vspace{-0.1cm}
\subsubsection{Healthy-diseased translation.}
The objective of this work is learn to localize and segment tumor lesions in a target modality by making efficient use of the image-level labels via a generative model. The latter introduces translation between presence and absence (of tumor lesions) domains, referred as P and A. This requires disentangling the information common to A and P, and the one specific to P. For this purpose, we split the latent representation of each image into two distinct codes $\mathbf{c}$, $\mathbf{u}$ that respectively refer to common and unique codes. Basically, the common code contains information inherent to both domains, which represents organs and other structures, while the unique code stores features like tumor shapes and localization.\newpage
\begin{figure}[h!]\centering
\includegraphics[width=0.97\columnwidth]{M-GenSeg-combined-compact.pdf}
\vspace{-0.2cm}
\caption{M-GenSeg: Simultaneous domain adaptation and semi-supervised segmentation. The common code is used by common decoders for either absence to presence translation or image reconstruction. Both segmentation and tumor disentangling are performed on diseased inputs by residual decoders. The latter also generates synthetic residuals for absence to presence translation, after sampling from a normal distribution. Finally, based on the same latent representation, cross-modality transposition is performed in a CycleGan-like approach.}
\label{diag1}
\end{figure}
\vspace{-1.2cm}
\paragraph{Presence to absence translation.}
Given an image of modality $m$ in the presence domain $\mathbf{x_{P}^m}$, we use an encoder $E_m$ to compute the latent representation $\mathbf{[}\mathbf{c_{P}^m},\mathbf{u_{P}^m}\mathbf{]}$. A common decoder $G_{com}^m$ takes as input the common code $\mathbf{c_{P}^m}$ and generates a healthy version $\mathbf{x_{PA}^m}$ of that image by removing the apparent tumor region. Simultaneously, both common and unique codes are used by a residual decoder $G_{res}^m$ to output a residual image $\mathbf{\Delta_{PP}^m}$, which corresponds to the additive change necessary to shift the generated healthy image back to the presence domain. In other words, the residual is the disentangled tumor that can be added to the generated healthy image to create a reconstruction $\mathbf{x_{PP}^m}$ of the initial diseased image:
\vspace{-0.2cm}
\begin{equation}
\mathbf{x_{PA}^m}=G_{com}^m(\mathbf{c_{P}^m}),
\end{equation}
\vspace{-0.4cm}
\begin{equation}
\mathbf{\Delta_{PP}^m}=G_{res}^m(\mathbf{c_{P}^m},\mathbf{u_{P}^m}),
\end{equation}
\vspace{-0.4cm}
\begin{equation}
\mathbf{x_{PP}^m}=\mathbf{x_{PA}^m}+\mathbf{\Delta_{PP}^m}.
\end{equation}
\vspace{-0.8cm}
\paragraph{Absence to presence translation.}
Concomitantly, a similar path is achieved for images in the healthy domain. Given an image $\mathbf{x_{A}^m}$ of modality $m$ in domain A, we generate a translated version in domain P. To do so, a synthetic tumor $\mathbf{\Delta_{AP}^m}$ is generated by sampling a code from the normal distribution $\mathcal{N}(0,\mathbf{I})$ and replacing the encoded unique code for that image. The reconstruction of the original image in domain A and the synthetic diseased image $x_{AP}^m$ in domain P are computed from the encoded features $\mathbf{[}\mathbf{c_{A}^m},\mathbf{u_{A}^m}\mathbf{]}$ as follows:
\vspace{-0.2cm}
\begin{equation}
\mathbf{x_{AA}^m}=G_{com}^m(\mathbf{c_{A}^m}),
\end{equation}
\vspace{-0.5cm}
\begin{equation}
\mathbf{x_{AP}^m}=\mathbf{x_{AA}^m}+G_{res}^m(\mathbf{c_{A}^m},\mathbf{u}\sim\mathcal{N}(0,\mathbf{I})).
\end{equation}
There are multiple ways a tumor can appear in the brain. Presence $\rightarrow$ Absence translation thus requires a one-to-many mapping, here enabled by sampling the unique code in a normal distribution. It is important to note that translating from Absence $\rightarrow$ Presence domain performs indirect data augmentation, thus improving segmentation performance. Finally, in both translation directions we reconstruct the latent representations of the generated images, to ensure that the translation task holds the information relative to the initial image. In the Absence $\rightarrow$ Presence direction, it notably enforces that the distribution of unique codes matches with the prior.
\vspace{-0.5cm}
\subsubsection{Modality translation.}
While the GenSeg framework handles primarily the segmentation of tumors, it is done only in the source modality. Indeed, for each modality $m$, the segmenter $G_{seg}^m$ shares most of its weights with the residual decoder, but has its own set of normalization parameters and a supplementary classifying layer. Thus, through the Absence and Presence translations, the segmenters have already learned how to disentangle the tumor from the background. However, supervised training on a few example annotations is still required to learn how to transform the resulting residual representation into appropriate segmentation maps. While this is a fairly straightforward task for the source modality using pixel-level annotations, achieving this for the target modality requires further efforts, justifying the second unsupervised translation objective between source and target modalities.
Based on the CycleGan \cite{cyclegan} approach, modality translations are performed via two distinct generators that share their encoder with the GenSeg task. More precisely, two different decoders $G_S$ and $G_T$ respectively perform the decoding towards source and target modalities. Lastly, given an image $\mathbf{x_{b}^m}$ of modality $m\in \{S,T\}$ in domain $b\in \{P,A\}$, and $m'$ the other modality, we perform a cyclic translation by generating $\mathbf{x_{d}^{m'}}$ in domain $m'$ and $\mathbf{x_{d}^{m''}}$ a reconstruction of the original image:
\vspace{-0.2cm}
\begin{equation}
\mathbf{[}\mathbf{c_{b}^m},\mathbf{u_{b}^m}\mathbf{]}=E_m(\mathbf{x_{b}^{m}}),
\end{equation}
\vspace{-0.5cm}
\begin{equation}
\mathbf{x_{b}^{m'}}=G_{m'}\left(\mathbf{c_{b}^m},\mathbf{u_{b}^m}\right)=\left[G_{m'}\circ E_m\right](\mathbf{x_{b}^{m}}),
\end{equation}
\vspace{-0.4cm}
\begin{equation}
\mathbf{[}\mathbf{c_{b}^{m'}},\mathbf{u_{b}^{m'}}\mathbf{]}=E_{m'}(\mathbf{x_{b}^{m'}}),
\end{equation}
\vspace{-0.4cm}
\begin{equation}
\mathbf{x_{b}^{m''}}=G_m(\mathbf{c_{b}^{m'}},\mathbf{u_{b}^{m'}})=\left[G_{m}\circ E_{m'}\right](\mathbf{x_{b}^{m'}})=\left[G_{m}\circ E_{m'}\circ G_{m'}\circ E_m\right](\mathbf{x_{b}^{m}}).
\end{equation}
Modality translation is only required on source images that have pixel-level annotations. However, performing translation on all types of images and in both directions yield additional training examples for modality generators. Ultimately, modality translation enables to fine-tune the segmentor on annotated synthetic target images, generated from the source images that possess pixel-level annotations $\mathbf{y_S}$. Based on \emph{(i)} the encoded features $\mathbf{[}\mathbf{c_{P}^S},\mathbf{u_{P}^S}\mathbf{]}$ of an original source image in domain P, \emph{(ii)} and the latent representation $\mathbf{[}\mathbf{c_{P}^{T}},\mathbf{u_{P}^{T}}\mathbf{]}$ of its translation to target modality, two segmentation masks for that image are generated:
\vspace{-0.1cm}
\begin{equation}
\mathbf{\hat{y}_S}=G_{seg}^S(\mathbf{c_{P}^S},\mathbf{u_{P}^S}),
\end{equation}
\vspace{-0.4cm}
\begin{equation}
\mathbf{\hat{y}_T}=G_{seg}^T(\mathbf{c_{P}^T},\mathbf{u_{P}^T}).
\end{equation}
By sharing the latent space between segmentation and modality translation tasks, M-GenSeg is forced to learn a common representation relevant to both of them, which can then be interpreted separately by the different decoders.
Furthermore, because $\mathbf{u}$ is sampled from a single prior for the healthy to diseased translation, one can expect that: \emph{(i)} modality information cannot be encoded in $\mathbf{u}$, \emph{(ii)} modality must be encoded in $\mathbf{c}$, \emph{(iii)} residual and translation decoders interpret which features in $\mathbf{u}$ to visualize, and how, based on the code in $\mathbf{c}$.
\vspace{-0.3cm}
\subsection{Loss functions}
\subsubsection{Segmentation Loss.}
For the segmentation objective, we compute a soft Dice loss \cite{vor} on the predictions for both labelled source images and their translations:
\vspace{-0cm}
\begin{equation}
\mathcal{L}_{seg}=Dice\left(\mathbf{y_S},\mathbf{\hat{y}_S}\right)+Dice\left(\mathbf{y_S},\mathbf{\hat{y}_T}\right).
\end{equation}
\vspace{-0.7cm}
\subsubsection{Reconstruction Losses.}
\vspace{-0.1cm}
We impose pixel level image reconstruction constraints on both modality translation task:
\begin{equation}
\mathcal{L}_{cyc}^{mod}=\sum_{\substack{m\in\{S,T\}\\b\in\{A,P\}}}\mathcal{L}_{1}\left(\mathbf{x_{b}^{m''}},\mathbf{x_{b}^{m}}\right),
\end{equation}
and GenSeg task:
\begin{equation}
\mathcal{L}_{rec}^{Gen}=\sum_{m\in\{S,T\}}\left[\mathcal{L}_{1}\left(\mathbf{x_{AA}^m},\mathbf{x_{A}^m}\right)+\mathcal{L}_{1}\left(\mathbf{x_{PP}^m},\mathbf{x_{P}^m}\right)\right].
\end{equation}
Likewise, we enforce the images generated in the GenSeg path to have distributions that match across presence and absence domains by reconstructing their latent codes.
\begin{align}
&\mathcal{L}_{lat}^{Gen} =\sum_{m\in\{S,T\}} \mathcal{L}_{1}\left(\left.E_m(\mathbf{x_{AA}^m})\right|_{\mathbf{c}},\left.E_m(\mathbf{x_{A}^m})\right|_{\mathbf{c}}\right)+\mathcal{L}_{1}\left(\left.E_m\left(\mathbf{x_{PA}^m}\right)\right|_{\mathbf{c}},\left.E_m\left(\mathbf{x_{P}^m}\right)\right|_{\mathbf{c}}\right)\nonumber\\&+\mathcal{L}_{1}(\left.E_m\left(\mathbf{x_{AP}^m}\right)\right|_{\mathbf{c},\mathbf{u}},\left[\left.E_m\left(\mathbf{x_{A}^m})\right|_{\mathbf{c}},\mathbf{u}\right]\right)+\mathcal{L}_{1}(\left.E_m\left(\mathbf{x_{PP}^m}\right)\right|_{\mathbf{c},\mathbf{u}},\left.E_m\left(\mathbf{x_{P}^m}\right)\right|_{\mathbf{c},\mathbf{u}}).
\end{align}
\vspace{-0.8cm}
\subsubsection{Adversarial Loss.}
Lastly, we compute a hinge loss on real images $\mathbf{x_b}$ and generated images $\mathbf{\hat{x}_b}$, for domains $b\in\{SA,SP,TA,TP\}$ in the GenSeg path.
\vspace{-0.1cm}
\begin{align}
\mathcal{L}_{adv}^{Gen}=\sum_{b\in\{SA, SP, TA, TP\}}&\min_{G}\max_{D}\big[-\mathbb{E}_{\mathbf{x_b}}\left(\min\left(0,D_b(\mathbf{x_b})-1\right)\right) \nonumber\\&-\mathbb{E}_{\mathbf{\hat{x}_b}}\left(\min\left(0,-D_b(\mathbf{\hat{x}_b})-1\right)\right)-\mathbb{E}_{\mathbf{\hat{x}_b}}D_b(\mathbf{\hat{x}_b})\big]
\end{align}
Likewise, a loss $\mathcal{L}_{adv}^{mod}$ is computed on the modality translation path to generate synthetic images in source and target modalities.
\newpage
\subsubsection{Overall Loss.}
\vspace{-0.4cm}
The overall loss for training M-GenSeg is a weighted sum of all the losses introduced above:
\vspace{-0.2cm}
\begin{align}
\mathcal{L}_{Total} & = \lambda_{seg}\mathcal{L}_{seg}+ \lambda_{adv}^{mod}\mathcal{L}_{adv}^{mod}+\lambda_{cyc}^{mod}\mathcal{L}_{cyc}^{mod}\label{lambda}\\&+\lambda_{adv}^{Gen}\mathcal{L}_{adv}^{Gen}+\lambda_{rec}^{Gen}\mathcal{L}_{rec}^{Gen}+\lambda_{lat}^{Gen}\mathcal{L}_{lat}^{Gen}.\nonumber
\end{align}
To facilitate the hyper-parameter search, weights are normalized so that their sum always amounts to 1.
\vspace{-0.4cm}
\subsection{Implementation Details}
\vspace{-0.1cm}
\subsubsection{Training and hyper-parameters}
Training and testing of the model were done in PyTorch. We used a batch size of 15 with the AMSGrad optimizer using $\beta_1$ and $\beta_2$ of 0.5 and 0.999 respectively, and a learning rate of 0.0001. Our models were trained for 300 epochs and weights of the segmentation model with the highest validation Dice score were saved for evaluation on the test set. The same on-the-fly data augmentation as in \cite{GenSeg} was applied for all runs. Each training experiment was repeated three times for the baselines and M-GenSeg, with a different random seed for weight initialization. The performance reported is the mean of all test Dice mean scores, with standard deviation, across the three runs.
We employed the following strategy to search for relative loss weights ($\lambda$ in Eq.~\ref{lambda}) : \emph{(i)} with $\lambda_{adv}^{Gen}$, $\lambda_{rec}^{Gen}$ and $\lambda_{lat}^{Gen}$ set to zero, raise $\lambda_{cyc}^{mod}$ with respect to $\lambda_{adv}^{mod}$ until consistent modality translation and reconstruction enables adequate target modality segmentation; \emph{(ii)} slowly incorporate GenSeg unsupervised components by increasing $\lambda_{adv}^{Gen}$, $\lambda_{rec}^{Gen}$ and $\lambda_{lat}^{Gen}$ so that consistent translation between presence and absence domains is achieved; \emph{(iii)} and finally, fine tune $\lambda_{seg}$. We found that the following parameters yielded both great modality and absence/presence translations : $\lambda_{adv}^{mod}=3$, $\lambda_{cyc}^{mod}=20$, $\lambda_{adv}^{Gen}=6$, $\lambda_{rec}^{Gen}=20$ and $\lambda_{lat}^{Gen}=2$. $\lambda_{seg}$ varies depending on the fraction of pixel-level annotations provided to the network for training.
\vspace{-0.4cm}
\subsubsection{Architecture.}
Two distinct encoders, common decoders, residual and segmentation decoders, and modality translation decoders are used, one for each modality. In addition, four distinct discriminators are responsible for discriminating between real and synthetic samples for each of the four domains: Source/Presence, Source/Absence, Target/Presence and Target/Absence. Furthermore, modality translation should be independent from the presence and absence domains. Thus, we introduce two additional discriminators that are responsible for discriminating between real and synthetic images in source and target modalities respectively. The architecture used for encoders, decoders and discriminators is the same as in \cite{GenSeg}. However, in order to better choose the semantic information relevant for each objective, we introduced attention gates \cite{attentionunet} in the skip connections.
\vspace{-0.4cm}
\section{Experimental Results}
\vspace{-0.2cm}
\subsection{Datasets}
\vspace{-0.1cm}
Experiments were performed on the BraTS 2020 challenge dataset, adapted for the multi-modal brain tumor segmentation problem where images are known to be diseased (presence of tumors) or healthy (absence tumors). Amongst the 369 brain volumes available in BraTS, 37 were allocated for validation and test steps, while the 295 left were used for training. Based only on brain tissue, each brain volume was mean-centered and divided by five times the standard deviation. We then split the 3D brain volumes into 2 hemispheres and sliced them along the axial plane into 2D images. Then, slices where at least 1\% of the brain surface came out to be tumors were labelled as diseased, while those that didn't show any tumor lesion were labelled as healthy images. Slices that didn't meet these criteria and those where less than 25\% of the image was brain tissue were discarded. Also, to model the scenario of a bi-modal dataset, we picked source and target modalities out of the four MR contrasts available (T1, T2, T1ce and FLAIR) and discarded the slices corresponding to the two types of sequences left. To constitute unpaired training data, we used only one modality (source or target) per training volume. This was repeated for each possible pair of modalities. We could therefore experiment with twelve different combinations of unpaired source/target modalities. Finally, pixel-level ground-truth annotations were provided only for a portion of the diseased cases.
\vspace{-0.4cm}
\subsection{Model Evaluation}
\vspace{-0.1cm}
\subsubsection{Domain adaptation.}
We compared M-GenSeg with AccSegNet \cite{constrained} and AttENT \cite{attent}, two high performance models for domain-adaptative medical image segmentation. To such an extent, we performed domain-adaptation experiments with source and target modalities drawn from T1, T2, FLAIR and T1ce. For each possible source/target pair, the images of source modalities were provided with pixel-level annotations, while the segmentation maps for target images were discarded. While our model outperforms both AccSegNet and AttENT by a slight margin on the source modality, it greatly improves the segmentation performance on target modality for all source/target pairs, as illustrated in Fig. \ref{raw dice}. On average we report an absolute Dice score increase of 0.04 and 0.08 on target domain, respectively compared to AccSegNet and AttENT, with gain values reaching 0.15. We show in Fig. \ref{mbrats} several presence to absence translations and segmentation examples on different target modality images. As shown in the figure, although no pixel-level annotations were provided for the target modality, tumors were well disentangled from the brain, resulting in a successful presence to absence translation, and segmentation. Note that for T1 and T1ce sequences, where lesions are hypo-intense, M-GenSeg still manages to convert complex residuals into consistent segmentation maps.
\vspace{-0.5cm}
\subsubsection{Annotation deficit.}
We also tested M-GenSeg in scenarios with limited pixel-level annotations available in the source and target modalities. We trained models when only 1\%, 10\%, 40\%, or 70\% of the source modality and 0\% of the target modality annotations were available, using T1 and T2 modalities and compared the performance to AccSegNet and AttENT. We also added the performance of two supervised methods without any domain adaptation, TransUnet \cite{transunet} and a Unet with the same backbone architecture as the one used for M-GenSeg. Results are shown in Fig. \ref{semi}. As expected, supervised methods trained on the source modality generalize poorly on the target modality. However, even in the fully supervised setting (with 100\% of the source images annotated), the semi-supervised M-GenSeg model performs better on the source modality than the Unet model with the same architecture. Thus, it appears that, similarly to UAGAN \cite{uagan}, the modality translation stream is helping to extract valuable features for the segmentation task. Also, the fact that TransUnet is outperforming M-GenSeg on the source modality for an annotation fraction above 40\% indicates that the backbone architecture can be improved for our training strategy. Finally, these experiments show the robustness of our model in both the source and target modalities in a setting using a fraction of the annotated data. In both T1 to T2 and T2 to T1 adaptation experiments, while performance is severely dropping at 1\% of annotations for the four baselines, our model shows in comparison only a slight decrease.
\vspace{-0.9cm}
\begin{figure}[h]\centering
\subfloat{\resizebox{0.23\textwidth}{!}{\begin{tikzpicture}
\begin{axis}[
label style = {font=\Large},
ticklabel style = {font=\large},
width = 0.5\textwidth,
height = 6cm,
ymax=105,
major x tick style = transparent,
ybar=2*\pgflinewidth,
bar width=10pt,
ymajorgrids = true,
grid style=dashed,
xlabel={Source : T1},
ytick={0,20,40,60,80,100},
yticklabels={0\%, 20\%, 40\%, 60\%, 80\%, 100\%},
symbolic x coords={T1ce, FLAIR, T2},
xtick = data,
scaled y ticks = false,
enlarge x limits=0.3,
ymin=0,
legend cell align=left,
legend style={ nodes={scale=0.7, transform shape},
at={(0.39,0.99)},
column sep=1ex
},
legend image post style={scale=0.5}
]
\addplot+[style={black,fill=blue!25,mark=none}, error bars/.cd, y dir=both, y explicit, error bar style=black]
coordinates {(T1ce, 27.58) +- (0, 4.1) (FLAIR, 65.56) +- (0, 1.64) (T2,89.8) +- (0, 2.43) };
\addplot[style={black,fill=green!30,mark=none}, error bars/.cd, y dir=both, y explicit, error bar style=black]
coordinates {(T1ce, 47.9) +- (0, 5.4) (FLAIR, 80.84) +- (0, 1.84) (T2,94.55) +- (0, 1.13) };
\addplot[style={black,fill=red!30,mark=none}, error bars/.cd, y dir=both, y explicit, error bar style=black]
coordinates {(T1ce, 60.9) +- (0, 4.03) (FLAIR, 83.9) +- (0, 1.22) (T2,99.75) +- (0, 0.75) };
\legend{AttENT, AccSegNet, M-GenSeg}
\end{axis}
\end{tikzpicture}}}
\hfill
\subfloat{\resizebox{0.23\textwidth}{!}{\begin{tikzpicture}
\begin{axis}[
label style = {font=\Large},
ticklabel style = {font=\large},
width = 0.5\textwidth,
height = 6cm,
ymax=105,
major x tick style = transparent,
ybar=2*\pgflinewidth,
bar width=10pt,
ymajorgrids = true,
grid style=dashed,
xlabel={Source : T1ce},
ytick={0,20,40,60,80,100},
yticklabels={0\%, 20\%, 40\%, 60\%, 80\%, 100\%},
symbolic x coords={T1, FLAIR, T2},
xtick = data,
scaled y ticks = false,
enlarge x limits=0.3,
ymin=0,
legend cell align=left,
legend style={ nodes={scale=0.5, transform shape},
at={(0.99,0.99)},
column sep=1ex
},
legend image post style={scale=0.5}
]
\addplot+[style={black,fill=blue!25,mark=none}, error bars/.cd, y dir=both, y explicit, error bar style=black]
coordinates {(T1, 22) +- (0, 4.1) (FLAIR, 58.6) +- (0, 1.9) (T2,67.8) +- (0, 1.1) };
\addplot[style={black,fill=green!30,mark=none}, error bars/.cd, y dir=both, y explicit, error bar style=black]
coordinates {(T1, 66.5) +- (0, 3.8) (FLAIR, 69.9) +- (0, 2.82) (T2,88.23) +- (0, 2.18) };
\addplot[style={black,fill=red!30,mark=none}, error bars/.cd, y dir=both, y explicit, error bar style=black]
coordinates {(T1, 77.56) +- (0, 4) (FLAIR, 78) +- (0, 2.45) (T2,91.76) +- (0, 1.76) };
\end{axis}
\end{tikzpicture}}}
\hfill
\subfloat{\resizebox{0.23\textwidth}{!}{\begin{tikzpicture}
\begin{axis}[
label style = {font=\Large},
ticklabel style = {font=\large},
width = 0.5\textwidth,
height = 6cm,
ymax=105,
grid style=dashed,
major x tick style = transparent,
ybar=2*\pgflinewidth,
xlabel={Source : FLAIR},
bar width=10pt,
ymajorgrids = true,
ytick={0,20,40,60,80,100},
yticklabels={0\%, 20\%, 40\%, 60\%, 80\%, 100\%},
symbolic x coords={T1, T1ce, T2},
xtick = data,
scaled y ticks = false,
enlarge x limits=0.3,
ymin=0,
legend cell align=left,
legend style={ nodes={scale=0.5, transform shape},
at={(0.99,0.99)},
column sep=1ex
},
legend image post style={scale=0.5}
]
\addplot+[style={black,fill=blue!25,mark=none}, error bars/.cd, y dir=both, y explicit, error bar style=black]
coordinates {(T1, 64.4) +- (0, 1.35) (T1ce, 69.6) +- (0, 0.94) (T2,12.5) +- (0, 5.6) };
\addplot[style={black,fill=green!30,mark=none}, error bars/.cd, y dir=both, y explicit, error bar style=black]
coordinates {(T1, 68) +- (0, 1.99) (T1ce, 52.5) +- (0, 5.75) (T2,50.14) +- (0, 5.3) };
\addplot[style={black,fill=red!30,mark=none}, error bars/.cd, y dir=both, y explicit, error bar style=black]
coordinates {(T1, 83.6) +- (0, 1.15) (T1ce, 77.92) +- (0, 1.99) (T2,84) +- (0, 2.8) };
\end{axis}
\end{tikzpicture}}}
\hfill
\subfloat{\resizebox{0.23\textwidth}{!}{\begin{tikzpicture}
\begin{axis}[
label style = {font=\Large},
ticklabel style = {font=\large},
width = 0.5\textwidth,
height = 6cm,
ymax=105,
grid style=dashed,
major x tick style = transparent,
ybar=2*\pgflinewidth,
bar width=10pt,
ymajorgrids = true,
xlabel={Source : T2},
ytick={0,20,40,60,80,100},
yticklabels={0\%, 20\%, 40\%, 60\%, 80\%, 100\%},
symbolic x coords={T1, T1ce, FLAIR},
xtick = data,
scaled y ticks = false,
enlarge x limits=0.3,
ymin=0,
legend cell align=left,
legend style={ nodes={scale=0.5, transform shape},
at={(0.99,0.99)},
column sep=1ex
},
legend image post style={scale=0.5}
]
\addplot+[style={black,fill=blue!25,mark=none}, error bars/.cd, y dir=both, y explicit, error bar style=black]
coordinates {(T1, 80.3) +- (0, 0.79) (T1ce, 68.6) +- (0, 1.25) (FLAIR,7.6) +- (0, 3.4) };
\addplot[style={black,fill=green!30,mark=none}, error bars/.cd, y dir=both, y explicit, error bar style=black]
coordinates {(T1, 81.5) +- (0, 1.93) (T1ce, 64.3) +- (0, 2.18) (FLAIR,8.6) +- (0, 4.5) };
\addplot[style={black,fill=red!30,mark=none}, error bars/.cd, y dir=both, y explicit, error bar style=black]
coordinates {(T1, 86.1) +- (0, 1.75) (T1ce, 80.85) +- (0, 1.88) (FLAIR,35.5) +- (0, 4.8) };
\end{axis}
\end{tikzpicture}}}
\vspace{-0.2cm}
\caption{Dice performance gap covered by each method on the target modality for each possible modality pair. At 0\%, the model performs as poorly as supervised segmentation trained on source data without any domain adaptation strategy. At 100\% it performs as well as UAGAN, a method designed for bi-modal datasets, trained on all of the source and target data.}
\label{raw dice}
\end{figure}
\vspace{-1.8cm}
\begin{figure}[h]\centering
\subfloat[T1 to T2 adaptation]{\resizebox{0.49\textwidth}{!}{\begin{tikzpicture}
\begin{axis}[
title style = {font=\huge},
label style = {font=\LARGE},
ticklabel style = {font=\Large},
xlabel={T1 annotations},
ylabel={Dice Score},
ymin=0.25, ymax=0.85,
ytick={0.3, 0.4, 0.5, 0.6, 0.7,0.8},
xtick={0.01, 0.2, 0.4, 0.6, 0.8, 1},
xticklabels={1\%, 20\%, 40\%, 60\%, 80\%, 100\%},
legend style={font=\large,at={(0.61,0.04)},anchor=south},
ymajorgrids=true,
grid style=dashed,
title={Source modality}
]
\addplot[style={bblue,mark=diamond*, mark options={scale=2,fill=white}}]
coordinates {(0.01, 0.607) (0.1, 0.711) (0.4,0.733) (0.7,0.755) (1,0.769)};
\addplot[style={ggreen,mark=triangle*, mark options={scale=2,fill=white}}]
coordinates {(0.01, 0.291) (0.1,0.693) (0.4,0.724) (0.7,0.749) (1,0.768)};
\addplot[style={ppurple,mark=*, mark options={scale=1.5,fill=white}}]
coordinates {(0.01, 0.476) (0.1,0.666) (0.4,0.726) (0.7,0.757) (1,0.769)};
\addplot[style={rred,mark=square*, mark options={scale=1.5,fill=white}}]
coordinates {(0.01, 0.48) (0.1,0.717) (0.4,0.786) (0.7,0.798) (1,0.801)};
\addplot[style={bittersweet,mark=star, mark options={scale=1.5,fill=white}}]
coordinates {(0.01, 0.39) (0.1,0.631) (0.4,0.661) (0.7,0.72) (1,0.746)};
\legend{M-GenSeg,AccSegNet,AttENT,TransUnet,Supervised M-GenSeg}
\end{axis}
\end{tikzpicture}\hspace{0.3cm}\begin{tikzpicture}
\begin{axis}[
title style = {font=\huge},
label style = {font=\LARGE},
ticklabel style = {font=\Large},
xlabel={T1 annotations},
ylabel={Dice Score},
ymin=0.1, ymax=0.9,
ytick={0.1,0.2,0.3,0.4, 0.5, 0.6, 0.7, 0.8},
xtick={0.01, 0.2, 0.4, 0.6, 0.8, 1},
xticklabels={1\%, 20\%, 40\%, 60\%, 80\%, 100\%},
legend style={at={(1.03,0.5)},anchor=west},
ymajorgrids=true,
grid style=dashed,
title={Target modality}
]
\addplot[style={bblue,mark=diamond*, mark options={scale=2,fill=white}}]
coordinates {(0.01, 0.734) (0.1, 0.778) (0.4,0.823) (0.7,0.829) (1,0.833)};
\addplot[style={ggreen,mark=triangle*, mark options={scale=2,fill=white}}]
coordinates {(0.01, 0.423) (0.1,0.756) (0.4,0.802) (0.7,0.813) (1,0.812)};
\addplot[style={ppurple,mark=*, mark options={scale=1.5,fill=white}}]
coordinates {(0.01, 0.478) (0.1,0.674) (0.4,0.721) (0.7,0.747) (1,0.781)};
\addplot[style={rred,mark=square*, mark options={scale=1.5,fill=white}}]
coordinates {(0.01, 0.356) (0.1,0.367) (0.4,0.41) (0.7,0.405) (1,0.413)};
\addplot[style={bittersweet,mark=star, mark options={scale=1.5,fill=white}}]
coordinates {(0.01, 0.353) (0.1,0.251) (0.4,0.218) (0.7,0.198) (1,0.21)};
\end{axis}
\end{tikzpicture}}}
\hfill
\subfloat[T2 to T1 adaptation]{\resizebox{0.49\textwidth}{!}{\begin{tikzpicture}
\begin{axis}[
title style = {font=\huge},
label style = {font=\LARGE},
ticklabel style = {font=\Large},
xlabel={T2 annotations},
ylabel={Dice Score},
xtick={0.01, 0.2, 0.4, 0.6, 0.8, 1},
xticklabels={1\%, 20\%, 40\%, 60\%, 80\%, 100\%},
ytick={0.6,0.65,0.7,0.75, 0.8, 0.85, 0.9},
legend style={at={(1.03,0.5)},anchor=west},
ymajorgrids=true,
grid style=dashed,
title={Source modality}
]
\addplot[style={bblue,mark=diamond*, mark options={scale=2,fill=white}}]
coordinates {(0.01, 0.766) (0.1, 0.818) (0.4,0.848) (0.7,0.856) (1,0.860)};
\addplot[style={ggreen,mark=triangle*, mark options={scale=2,fill=white}}]
coordinates {(0.01, 0.604) (0.1,0.734) (0.4,0.8) (0.7,0.826) (1,0.862)};
\addplot[style={ppurple,mark=*, mark options={scale=1.5,fill=white}}]
coordinates {(0.01, 0.652) (0.1,0.807) (0.4,0.84) (0.7,0.848) (1,0.861)};
\addplot[style={rred,mark=square*, mark options={scale=1.5,fill=white}}]
coordinates {(0.01, 0.72) (0.1,0.821) (0.4,0.854) (0.7,0.866) (1,0.882)};
\addplot[style={bittersweet,mark=star, mark options={scale=1.5,fill=white}}]
coordinates {(0.01, 0.653) (0.1,0.747) (0.4,0.793) (0.7,0.811) (1,0.836)};
\end{axis}
\end{tikzpicture}\hspace{0.3cm}\begin{tikzpicture}
\begin{axis}[
title style = {font=\huge},
label style = {font=\LARGE},
ticklabel style = {font=\Large},
xlabel={T2 annotations},
ylabel={Dice Score},
xtick={0.01, 0.2, 0.4, 0.6, 0.8, 1},
xticklabels={1\%, 20\%, 40\%, 60\%, 80\%, 100\%},
ytick={0.1,0.2,0.3,0.4,0.5,0.6,0.7},
legend style={at={(1.03,0.5)},anchor=west},
ymajorgrids=true,
grid style=dashed,
title={Target modality}
]
\addplot[style={bblue,mark=diamond*, mark options={scale=2,fill=white}}]
coordinates {(0.01, 0.582) (0.1, 0.671) (0.4,0.677) (0.7,0.685) (1,0.698)};
\addplot[style={ggreen,mark=triangle*, mark options={scale=2,fill=white}}]
coordinates {(0.01, 0.43) (0.1,0.651) (0.4,0.66) (0.7,0.661) (1,0.665)};
\addplot[style={ppurple,mark=*, mark options={scale=1.5,fill=white}}]
coordinates {(0.01, 0.522) (0.1,0.648) (0.4,0.655) (0.7,0.65) (1,0.656)};
\addplot[style={rred,mark=square*, mark options={scale=1.5,fill=white}}]
coordinates {(0.01, 0.331) (0.1,0.297) (0.4,0.321) (0.7,0.294) (1,0.328)};
\addplot[style={bittersweet,mark=star, mark options={scale=1.5,fill=white}}]
coordinates {(0.01, 0.247) (0.1,0.177) (0.4,0.151) (0.7,0.135) (1,0.131)};
\end{axis}
\end{tikzpicture}}}
\vspace{-0.3cm}
\caption{Dice scores with T1/T2 pair when using reference annotations for 0\% of target data and various fractions (1\%, 10\%, 40\%, 70\% and 100\%) of source data during training. For readability, standard deviations across the runs are not shown. }
\label{semi}
\end{figure}
\vspace{-1.4cm}
\subsubsection{Reaching supervised performance.}
M-GenSeg provides encouraging improvements on the target modality as compared to a supervised method trained on source data without domain adaptation, as shown in Fig. \ref{semi}. However, when no reference annotations are provided for the target modality, the performance does not reach the one from fully supervised methods trained on the fully annotated target modality. It is thus beneficial to determine the fraction of target modality annotations required to match fully supervised performance. Fig.\ref{increasing} shows the Dice difference between UAGAN, and AttENT, AccSegNet, M-GenSeg with 100\% source annotations and increasing fractions of target annotations (0\%, 10\% and 20\%). We report in average an absolute Dice difference on target modality of 0.021 from T2 to T1 and 0.02 from T1 to T2 between UAGAN trained with all labels and our method trained with all source images annotated and 20\% of the target images annotated. Thus, the annotation burden could be reduced with the M-GenSeg method.
\vspace{-1cm}
\begin{figure}[h]\centering
\subfloat{\resizebox{0.31\textwidth}{!}{\begin{tikzpicture}
\begin{axis}[
label style = {font=\large},
width = 0.5\textwidth,
height = 4.5cm,
ymax=0,
ymin=-15,
ylabel={Dice},
major x tick style = transparent,
ybar=2*\pgflinewidth,
bar width=14pt,
ymajorgrids=true,
grid style=dashed,
xlabel={0\% target annotations},
ytick={-15,-12.5,-10,-7.5,-5,-2.5,0},
yticklabels={-0.15,-0.125,-0.10,-0.075,-0.05,-0.025,0},
symbolic x coords={T1 to T2, T2 to T1},
xtick = data,
scaled y ticks = false,
enlarge x limits=0.5,
legend cell align=left,
legend style={ nodes={scale=0.6, transform shape},
at={(0.35,0.99)},
column sep=1ex
},
legend image post style={scale=0.5}
]
\addplot+[style={black,fill=blue!25,mark=none}, error bars/.cd, y dir=both, y explicit, error bar style=black]
coordinates {(T1 to T2, -9.9) +- (0, 1.2) (T2 to T1,-13.1) +- (0, 0.85)};
\addplot+[style={black,fill=green!30,mark=none}, error bars/.cd, y dir=both, y explicit, error bar style=black]
coordinates {(T1 to T2, -6.8) +- (0, 0.63) (T2 to T1,-12.6) +- (0, 0.95)};
\addplot+[style={black,fill=red!30,mark=none}, error bars/.cd, y dir=both, y explicit, error bar style=black]
coordinates {(T1 to T2, -3.9) +- (0, 0.92) (T2 to T1,-8.9) +- (0, 0.58)};
\end{axis}
\end{tikzpicture}}}
\hfill
\subfloat{\resizebox{0.31\textwidth}{!}{\begin{tikzpicture}
\begin{axis}[
label style = {font=\large},
width = 0.5\textwidth,
height = 4.5cm,
ymax=0,
ymin=-15,
ylabel={Dice},
major x tick style = transparent,
ybar=2*\pgflinewidth,
bar width=14pt,
ymajorgrids=true,
grid style=dashed,
xlabel={10\% target annotations},
ytick={-15,-12.5,-10,-7.5,-5,-2.5,0},
yticklabels={-0.15,-0.125,-0.10,-0.075,-0.05,-0.025,0},
symbolic x coords={T1 to T2, T2 to T1},
xtick = data,
scaled y ticks = false,
enlarge x limits=0.5,
legend cell align=left,
legend style={ nodes={scale=0.5, transform shape},
at={(0.3333,0.99)},
column sep=1ex
},
legend image post style={scale=0.5}
]
\addplot+[style={black,fill=blue!25,mark=none}, error bars/.cd, y dir=both, y explicit, error bar style=black]
coordinates {(T1 to T2, -8.9) +- (0, 0.85) (T2 to T1,-11) +- (0, 0.57)};
\addplot+[style={black,fill=green!30,mark=none}, error bars/.cd, y dir=both, y explicit, error bar style=black]
coordinates {(T1 to T2, -5) +- (0, 0.68) (T2 to T1,-8.6) +- (0, 1.3)};
\addplot+[style={black,fill=red!30,mark=none}, error bars/.cd, y dir=both, y explicit, error bar style=black]
coordinates {(T1 to T2, -3) +- (0, 0.62) (T2 to T1,-5.1) +- (0, 1.1)};
\end{axis}
\end{tikzpicture}}}
\hfill
\subfloat{\resizebox{0.31\textwidth}{!}{\begin{tikzpicture}
\begin{axis}[
label style = {font=\large},
width = 0.5\textwidth,
height = 4.5cm,
ymax=0,
ymin=-15,
major x tick style = transparent,
ybar=2*\pgflinewidth,
bar width=14pt,
ymajorgrids=true,
grid style=dashed,
ylabel={Dice},
xlabel={20\% target annotations},
ytick={-15,-12.5,-10,-7.5,-5,-2.5,0},
yticklabels={-0.15,-0.125,-0.10,-0.075,-0.05,-0.025,0},
symbolic x coords={T1 to T2, T2 to T1},
xtick = data,
scaled y ticks = false,
enlarge x limits=0.5,
legend cell align=left,
legend style={font=\scriptsize,
at={(0.463,0.43), nodes={scale=1, transform shape}},
column sep=1ex
},
legend image post style={scale=0.7}
]
\addplot+[style={black,fill=blue!25,mark=none}, error bars/.cd, y dir=both, y explicit, error bar style=black]
coordinates {(T1 to T2, -7.5) +- (0, 0.6) (T2 to T1,-9.1) +- (0, 1.1)};
\addplot+[style={black,fill=green!30,mark=none}, error bars/.cd, y dir=both, y explicit, error bar style=black]
coordinates {(T1 to T2, -3.8) +- (0, 0.3) (T2 to T1,-5.1) +- (0, 0.7)};
\addplot+[style={black,fill=red!30,mark=none}, error bars/.cd, y dir=both, y explicit, error bar style=black]
coordinates {(T1 to T2, -2) +- (0, 0.75) (T2 to T1,-2.1) +- (0, 0.9)};
\legend{AttENT, AccSegNet, M-GenSeg}
\end{axis}
\end{tikzpicture}}}
\vspace{-0.3cm}
\caption{Dice difference on target modality, as compared to the fully supervised method UAGAN, when using T1/T2 reference annotations for 100\% of source data and increasing fractions of target data (0\%, 10\% and 20\%). A gap of 0 means that fully supervised method performance has been reached.}
\label{increasing}
\end{figure}
\vspace{-1.2cm}
\subsection{Ablation Experiments}
\vspace{-0.1cm}
The following ablation tests, conducted to validate the conception choices and presented in Fig. \ref{ablation:train} and Fig. \ref{ablation:skip}, were run on six different domain adaptation experiments (T1 to T2, T1 to FLAIR, T1 to T1ce and vice-versa) where all the target modality images are unannotated. The variability across the different modality pairs explain the high standard deviations reported in Fig. \ref{ablation:train} and \ref{ablation:skip}.
\vspace{-0.4cm}
\subsubsection{Attention-aware generators.}
To validate the use of the attention gates in the skip connections, we evaluated the performance of M-GenSeg with different shortcut variants. Fig. \ref{ablation:skip} shows the performance with concatenation, skinny cat \cite{GenSeg} and summation skip connections relatively to the performance of the proposed variant with attention gates. Though they provide only a slight performance gain, attention gates have the benefit to give insight on what happens in the model. Fig. \ref{attention} shows the attention maps generated for each type of decoder. As expected, residual decoders focus towards tumor areas. More interestingly, in order not to disturb the process of healthy image generation, common decoders avoid lesion locations. Finally, modality translators tend to focus on salient details of the brain tissue, which facilitates contrast redefinition needed for accurate translation.
\vspace{-0.4cm}
\subsubsection{Modality invariant features.} Some unsupervised cross-modality translation models \cite{attent,constrained,invar,uagan} rely on modality invariant features to perform their image-to-image translation. To do so, additionally to the $m\rightarrow m'\rightarrow m$ CycleGan translation objective, they introduce an identity loss such that $m\rightarrow m$ self-reconstruction can be achieved after encoding and decoding. In our case, this loss could be written $\mathcal{L}_{idt}^{mod}=\sum_{\substack{m\in\{S,T\}\\b\in\{A,P\}}}\mathcal{L}_{1}\left(\left[G_{m}\circ E_m\right](\mathbf{x_{b}^{m}}),\mathbf{x_{b}^{m}}\right)$ and imposes that an image $\mathbf{x_{b}^{m}}$ remains in the same modality $m$ when sequentially going through the encoder and modality decoder specific to this modality, $E_m$ and $G_{m}$. We tested this loss with M-GenSeg and different choices of $\lambda_{idt}^{mod}$, but did not observe performance gains. We thus believe such a loss is unnecessary here.
\newpage
\begin{figure}
\centering
\hspace{0.5cm}\begin{minipage}[t]{0.5\textwidth}\centering
\includegraphics[width=0.96\linewidth]{Mbrats_annote.png}
\vspace{-0.3cm}
\captionof{figure}{Examples of image translation from Presence to Absence domains and resulting segmentation in different adaptation experiments. Each column represents a domain adaptation scenario. For each instance, the input image considered belongs to target modality data, which had no pixel-level annotations provided.
}
\label{mbrats}
\end{minipage}
\hfill
\begin{minipage}[t]{0.37\textwidth}\centering
\includegraphics[width=0.96\linewidth]{Attention_fig.pdf}
\vspace{-0.2cm}
\captionof{figure}{Attention maps for Presence $\rightarrow$ Absence and modality translations. Red color indicates areas of focus while dark blue correspond to locations ignored by the network.}
\label{attention}
\end{minipage}\hspace{0.5cm}
\end{figure}
\vspace{-1.5cm}
\subsubsection{Image-level supervision.} To assess the value of the image-level supervision in this domain adaptation context, we tried M-GenSeg without any GenSeg unsupervised objective. In other words, all $\lambda^{Gen}$ loss components were set to 0. As expected, the source modality segmentation performance wasn't impacted, however segmentation accuracy on the target modality dropped, implying that the use of image-level labels is helpful.
\vspace{-0.6cm}
\subsubsection{Translations for healthy data.} We conjecture that training modality translation only on diseased data is sufficient. However, doing it for healthy data as well provides additional training examples for this task, which is reflected in the performance. Likewise, performing translation from absence to presence domain is not necessary, but yields better Dice scores by making more efficient use of the data. Finally, training M-GenSeg without any healthy path at all (both previous experiments combined) gave even worse Dice scores. We thus claim that healthy tissues, if adequately incorporated to the training process of neural networks like in M-GenSeg, can help to better delineate tumor lesions in segmentation tasks. Note that feeding healthy tissues to supervised methods (along with empty masks) does not suffice, as it only helps to reduce false positives. Indeed, we trained a TransUnet model with either only diseased or both diseased and healthy data. The latter yielded a higher dice score ($0.867 \pm 0.04$ vs $0.836 \pm 0.014$) and precision ($0.878 \pm 0.005$ vs $0.803 \pm 0.017$) in comparison to using only diseased examples. However, the trade-off is a lower recall ($0.856 \pm 0.008$ vs $0.873 \pm 0.009$) which makes it more prone to miss some lesions.
\vspace{-0.6cm}
\subsubsection{Unshared latent space.} We contend that image-level supervision and modality translation benefit from sharing their latent representations. We thus evaluated M-GenSeg with separate latent spaces for these two tasks, which simply requires duplicating the encoders. The drop in performance for this variant indicates that M-GenSeg efficiently combines both tasks when the latent representations share model updates.
\vspace{-0.6cm}
\subsubsection{Residual/Segmentation separation.} Finally, we tried using a separate segmentation decoder instead of sharing the residual and segmentation weights. As for GenSeg \cite{GenSeg}, the separate version under-performed on the brain dataset. This further highlights that disentangling tumors from the background to perform diseased to healthy translation is helpful for tumor segmentation, particularly on the target modality.
\vspace{-0.7cm}
\begin{figure}[h]\centering
\subfloat[Training procedure\label{ablation:train}]{\resizebox{0.59\textwidth}{!}{\begin{tikzpicture}
\begin{axis}[
ybar,
label style = {font=\scriptsize},
ylabel={Relative Dice},
height = 4.9cm,
width = 7cm,
ybar=0.5*\pgflinewidth,
ticklabel style = {font=\scriptsize},
bar width=9pt,
legend style={font=\LARGE,nodes={scale=0.25, transform shape},at={(0.265,0.645)}, anchor=north},
ymajorgrids=true,
ytick = {2.5,0,-2.5,-5,-7.5,-10,-12.5,-15},
yticklabels={2.5\%,0\%,-2.5\%,-5\%,-7.5\%,-10\%,-12.5\%,-15\%},
enlarge x limits=0.5,
grid style=dashed,
symbolic x coords={Source,Target},
xtick=data
]
\addplot+[black,fill=purple!70,
error bars/.cd,
y dir=both,
y explicit, error bar style=black
] coordinates {
(Source, -2.74) +- (0, 0.6)
(Target, -5.43) +- (0, 3.27)
};
\addplot+[black,fill=blue!40,
error bars/.cd,
y dir=both,
y explicit, error bar style=black
] coordinates {
(Source, -1.31) +- (0, 1.9)
(Target, -3.84) +- (0, 1.71)
};
\addplot+[black,fill=cyan!40,
error bars/.cd,
y dir=both,
y explicit, error bar style=black
] coordinates {
(Source, -1.41) +- (0, 1.03)
(Target, -2.41) +- (0, 1.29)
};
\addplot+[black,fill=green!40,
error bars/.cd,
y dir=both,
y explicit, error bar style=black
] coordinates {
(Source, -1.47) +- (0, 1.82)
(Target, -6.61) +- (0, 3.11)
};
\addplot+[black,fill=yellow!40,
error bars/.cd,
y dir=both,
y explicit, error bar style=black
] coordinates {
(Source, -0.56) +- (0, 1.34)
(Target, -8.22) +- (0, 4.71)
};
\addplot+[black,fill=orange!40,
error bars/.cd,
y dir=both,
y explicit, error bar style=black
] coordinates {
(Source, -0.80) +- (0, 1.47)
(Target, -2.66) +- (0, 2.41)
};
\addplot+[black,fill=red!40,
error bars/.cd,
y dir=both,
y explicit, error bar style=black
] coordinates {
(Source, -2) +- (0, 1.2)
(Target, -4.39) +- (0, 1.91)
};
\legend{Modality invariant features,\hspace{0.2cm} No absence to presence translation, No healthy modality translation, No healthy path, No image-level supervision, Residual/Segmentation separation, Unshared latent space}
\end{axis}
\end{tikzpicture}}}
\hfill
\subfloat[Skip connections\label{ablation:skip}]{\resizebox{0.4\textwidth}{!}{\begin{tikzpicture}
\begin{axis}[
ybar,
ylabel={Relative Dice},
ticklabel style = {font=\scriptsize},
height = 5.35cm,
width=6cm,
ymax=0,
ybar=0.5*\pgflinewidth,
bar width=10pt,
legend style={font=\LARGE,at={(0.5,0.29)},nodes={scale=0.3, transform shape},
anchor=north},
ymajorgrids=true,
ytick = {-1,-2,-3,-4},
yticklabels={-1\%,-2\%,-3\%,-4\%},
enlarge x limits=0.5,
grid style=dashed,
symbolic x coords={Source,Target},
xtick=data
]
\addplot+[black,fill=cyan!40,
error bars/.cd,
y dir=both,
y explicit, error bar style=black
] coordinates {
(Source, -2.73) +- (0, 0.33)
(Target, -1.75) +- (0, 0.82)
};
\addplot+[black,fill=green!40,
error bars/.cd,
y dir=both,
y explicit, error bar style=black
] coordinates {
(Source, -0.65) +- (0, 0.12)
(Target, -2.14) +- (0, 1.72)
};
\addplot+[black,fill=red!40,
error bars/.cd,
y dir=both,
y explicit, error bar style=black
] coordinates {
(Source, -0.35) +- (0, 0.14)
(Target, -2.04) +- (0, 0.41)
};
\legend{Skinny-cat, \hspace{0.2cm} Concatenation, Sum}
\end{axis}
\end{tikzpicture}}}
\vspace{-0.2cm}
\caption{Ablation studies : relative Dice change on source and target modalities, as compared to the full version of the proposed model. X\% relative value means that the ablation reached 100\%+X\% of complete M-GenSeg Dice performance. Therefore negative values indicate that the ablation led to decrease in performance.}
\end{figure}
\vspace{-1.2cm}
\section{Conclusion}
\vspace{-0.2cm}
In this study, we propose a M-GenSeg, a new framework for unpaired cross-modality medical image segmentation when the target modality is unannotated or partially annotated. As this is a semi-supervised method, we also support scenarios where only a fraction of the source modality is annotated. Specifically, we combined two unsupervised image-to-image translations to perform both modality adaptation and leverage image-level labels, the latter being much easier to obtain than pixel-level annotations. Experiments showed that in domain-adaptive segmentation tasks, our method outperforms state-of-the-art models, while yielding robust performance even when a small fraction of the source images is annotated. We therefore contend that M-GenSeg is an annotation-efficient framework that greatly reduces the performance gap due to domain shift in cross-modality tumor segmentation. Pending increased computing resources, future work will explore the use of full 3D images rather than 2D slices, along with larger and more optimal architectures.
\bibliographystyle{IEEEtran}
|
\section{\textsc{Cifar} supervised transfer learning}
\label{apx:cifar10_100_sl}
\textsc{Cifar10} supervised transfer learning experiments train a \textsc{resnet18} network on the \textsc{Cifar10} dataset with/without L2 weight decay (4e-5) for $200$ epochs. During training, we use a SGD optimizer \citep{bottou2018optimization} with momentum=0.9, initial learning rate=0.1, cosine learning rate decay, and batch size=128. As to data augmentation, we use \textsc{RandomResizedCrop} (crop scale in $[0.8, 1.0]$), aspect ratio in $[3/4, 4/3]$) and \textsc{RandomHorizontalFlip}. During testing, the input images are resized to $36\times36$ by bicubic interpolation and \textsc{CenterCroped} to $32\times32$. All input images are normalized by $mean=(0.4914, 0.4822, 0.4465), std=(0.2023, 0.1994, 0.2010)$ at the end.
Then transfer the learned representation to \textsc{Cifar100} dataset by training a last-layer linear classifier (linear probing). The linear layer weights are initialized by Gaussian distribution $\mathcal{N}(0, 0.01)$. The linear probing process shares the same training hyper-parameters as the supervised training part except for a zero L2 weight decay in all cases.
The \textsc{Cifar100} supervised transfer learning experiments swap the order of \textsc{Cifar100} and \textsc{Cifar10}.
\section{\textsc{ImageNet} supervised transfer learning}
\label{apx:imagenet_sl}
\subsection{Experiment settings}
\label{apx:imagenet_sl_settings}
{\@startsection{paragraph}{4}{\z@}{.1ex plus 0ex minus .1ex}{-1em}{\normalsize\bf}{Image Preprocessing:} Following \citet{he2016deep}, we use \textsc{RandomHorizontalFlip} and \textsc{RandomResizedCrop} augmentations for all training tasks. For \textsc{ImageNet} and \textsc{Inat18}, the input images are normalized by $mean=(0.485, 0.456, 0.406), std=(0.229, 0.224, 0.225)$. For \textsc{Cifar}, we use the same setting as Appendix \ref{apx:cifar10_100_sl}.}
\@startsection{paragraph}{4}{\z@}{.1ex plus 0ex minus .1ex}{-1em}{\normalsize\bf}{\textsc{Imagenet} Pretraining:} The \textsc{resnet}s are pre-trained on \textsc{ImageNet} with the popular protocol of \citet{goyal2017accurate}: a SGD optimizer with momentum=0.9, initial learning rate=0.1, batch size=256, L2 weight decay=1e-4, and 90 training epochs. The learning rate is multiplied by 0.1 every 30 epochs. By default, the optimizer in all experiments is SGD with momentum=0.9.
\@startsection{paragraph}{4}{\z@}{.1ex plus 0ex minus .1ex}{-1em}{\normalsize\bf}{\textsc{Distill}:} To distill the {\synthcat}$n$ representations $[\phi_1, \dots \phi_n]$ ($n\times$\textsc{resnet50}) into a smaller representation $\Phi$ (\textsc{resnet50}), we use the multi-head architecture as Figure \ref{fig:distill}. Inspired by \citet{hinton2015distilling}, we use the Kullback–Leibler divergence loss to learn $\Phi$ as:
\begin{equation}
\label{eq:distill}
\min_{\Phi, w_0, \dots, w_n}\sum_{i=0}^n\sum_x\bigg[ \tau^2\mathcal{L}_{kl}\Big(s_{\tau}\big({v_i} \circ {\phi_i}(x)\big) ~||~ w_i \circ \Phi(x) \Big)\bigg],
\end{equation}
where $s_{\tau}(v)_i = \frac{e^{v_i/\tau}}{\sum_k{e^{v_k/\tau}}}$ is a softmax function with temperature $\tau$, $v_i$ is the learned last-layer classifier of $i^{th}$ sub-network of \synthcat$n$.
In the {\synthdistill} experiments, we distill five separately trained \textsc{resnet50} into one \textsc{resnet50} according to Eq \ref{eq:distill} with $\tau=10$. We use a SGD optimizer with momentum=0.9, batch size=2048, and weight decay=0. The initial learning rate is 0.1 and warms up to 0.8 within the first 5 epochs. Then learning rate decays to 0.16 and 0.032 at $210^{th}$ and $240^{th}$ epochs, respectively. The total training epochs is 270.
\@startsection{paragraph}{4}{\z@}{.1ex plus 0ex minus .1ex}{-1em}{\normalsize\bf}{Linear probing:}
\begin{itemize}
\item \textbf{\textsc{ImageNet}:} The \textsc{ImageNet} linear probing experiments train a linear classifier with the same hyper-parameters as \textsc{ImageNet} pretraining. By default, the last linear classifier in all linear probing experiments is initialized by $\mathcal{N}(0, 0.01)$.
\item \textbf{\textsc{Inat18}, \textsc{Cifar100}, \textsc{Cifar10}:} Following the settings of \citet{goyal2022vision}, the linear probing experiments (on \textsc{Inat18}, \textsc{Cifar100}, \textsc{Cifar10}) adds a \textsc{BatchNorm} layer before the linear classifier to reduce the hyper-parameter tuning difficulty. The learning rate is initialized to 0.01 and multiplied by 0.1 every 8 epochs. Then train these linear probing tasks for 28 epochs by SGD Nesterov optimizer with momentum=0.9, batch size 256. Note that \textsc{BatchNorm} + a linear classifier is still a linear classifier during inference. We tune L2 weight decay from \{1e-4, 5e-4, 1e-3, 5e-3, 1e-2, 5e-2\} for \textsc{Cifar100} and \textsc{Cifar10}, \{1e-6, 1e-5, 1e-4\} for \textsc{Inat18}.
\end{itemize}
\@startsection{paragraph}{4}{\z@}{.1ex plus 0ex minus .1ex}{-1em}{\normalsize\bf}{Fine-tuning:} As to the fine-tuning experiments (on \textsc{Cifar100}, \textsc{Cifar10}, and \textsc{Inat18}), we tune the initial learning rate from \{0.005, 0.01, 0.05\}, training epochs from \{50, 100\}. We further tune L2 weight decay from \{0, 1e-5, 1e-4, 5e-4\} for \textsc{Cifar100} and \textsc{Cifar10}, \{1e-6, 1e-5, 1e-4\} for \textsc{Inat18}. A cosine learning rate scheduler is used in fine-tuning experiments. A 0.01 learning rate and 100 training epochs usually provide the best performance for these three datasets. So we fix these two hyperparameters in the following supervised learning two-stage fine-tuning experiments and self-supervised learning experiments.
\@startsection{paragraph}{4}{\z@}{.1ex plus 0ex minus .1ex}{-1em}{\normalsize\bf}{Two-stage fine-tuning:} For the two-stage fine-tuning experiments, we separately fine-tune each sub-network (i.e. \textsc{resnet50}) of the \textsc{Cat}$n$ architecture by the same protocol as the normal fine-tuning above. Then train a last-layer linear classifier on top of the concatenated fine-tuned representation. The last-layer linear classifier training can be very efficient with a proper weights initialization strategy. In this work, we initialize the last-layer classifier $w$ (including the bias term) by concatenating the last-layer classifier of each fine-tuned sub-network $w_i$, $w \leftarrow {[w_0^\top, \dots, w_n^\top]^\top}/{n}$. Then we only need to train the last-layer classifier $w$ for 1 epoch with a learning rate = $1e-3$ for \textsc{Cifar} and $1e-5$ for \textsc{Inat18}.
\subsection{Experiments on a deeper architecture: \textsc{resnet152}}
\label{apx:imagenet_sl_resnet152}
Similar to table \ref{tab:imagenet_sl_lineareval} in section \ref{sec:supervisedtransfer}, table \ref{tab:imagenet_sl_rn152_lineareval} provides similar experiments on a deeper architecture \textsc{resnet152}. {\synthcat}$n$ exceeds ERM on \textsc{ImageNet}, \textsc{Cifar10}, \textsc{Cifar100} , and \textsc{Inat18} linear probing tasks.
\begin{table}[ht]
\caption{Imagenet supervised transfer learning performance on a deep architecture \textsc{resnet152}.}
\bigskip
\label{tab:imagenet_sl_rn152_lineareval}
\centering
\begin{tabular}{cc |c| ccc }
\toprule
& & ID & \multicolumn{3}{c}{Linear Probing (OOD)} \\
method & architecture & \textsc{ImageNet} & \textsc{Cifar10} & \textsc{Cifar100} & \textsc{Inat18} \\
\midrule
ERM & \textsc{resnet152} & 77.89 & 92.50 & 76.23 & 39.70 \\
\midrule
{\synthcat}2& $2\times$\textsc{resnet152} & 79.34 & 94.26 & 79.15 & 45.42\\
{\synthcat}5& $5\times$\textsc{resnet152} & 80.14 & 94.91 & 81.35 & 50.32\\
{\synthcat}10& $10\times$\textsc{resnet152} & 80.18 & 95.38 & 82.39 & 52.73 \\
\bottomrule
\end{tabular}
\end{table}
\subsection{Fine-tuning experiments}
For reference, table \ref{tab:imagenet_sl_ft_2ft} provides numerical results for the fine-tuning experiments of Figure \ref{fig:imagenet_sl_ft_2ft}.
\begin{table}[ht]
\caption{Supervised transfer learning by either normal fine-tuning or proposed two-stage fine-tuning. Various representations are pre-trained on \textsc{ImageNet} and then fine-tuned or two-stage fine-tuned on \textsc{Cifar10}, \textsc{Cifar100}, \textsc{Inat18} tasks. }
\bigskip
\label{tab:imagenet_sl_ft_2ft}
\centering
\begin{tabular}{cc c | cc c | ccc}
\toprule
& & & \multicolumn{3}{c|}{fine-tuning} & \multicolumn{3}{c}{two-stage fine-tuning} \\
method & architecture ¶ms & \textsc{Cifar10} & \textsc{Cifar100} & \textsc{Inat18} & \textsc{Cifar10} & \textsc{Cifar100} & \textsc{Inat18} \\
\midrule
ERM & \textsc{resnet50} & 23.5M & 97.54 & 85.58 & 64.19 & - & - & - \\
\midrule
ERM & \textsc{resnet50w2} & 93.9M & 97.76 & 87.13 & 66.72 & - & - & - \\
ERM & \textsc{resnet50w4} & 375M & 97.88 & 87.95 & 66.99 & - & - & -\\
ERM & \textsc{2$\times$resnet50} & 47M & 97.39 & 85.77 & 62.57 & - & - & - \\
ERM & \textsc{4$\times$resnet50} & 94M & 97.38 & 85.56 & 61.58 & - & - & - \\
\midrule
{\synthcat}2& \textsc{2$\times$resnet50} & 47M & 97.56 & 86.04 & 64.49 & 97.87 & 87.07 & 66.96\\
{\synthcat}4& \textsc{4$\times$resnet50} & 94M & 97.53 & 86.54 & 64.54 & 98.14 & 88.00 & 68.42\\
{\synthcat}5& \textsc{5$\times$resnet50} & 118M & 97.57 & 86.46 & 64.86 & 98.19 & 88.11 & 68.48\\
{\synthcat}10& \textsc{10$\times$resnet50} & 235M & 97.19 & 86.65 & 64.39 & 98.17 & 88.50 & 69.07\\
\midrule
{\synthdistill}5& \textsc{resnet50} & 23.5M & 97.07 & 85.31 & 64.17 & - & - & -\\
\bottomrule
\end{tabular}
\end{table}
\subsection{Vision transformer Experiment settings}
{For all vision transformer experiments, we keep the input image resolution at 384 $\times$ 384 and follow a similar protocol as appendix \ref{apx:imagenet_sl_settings}. Specifically, we use a weight decay=5e-4 and a batch size=256 for linear probing, a weight decay=0 and a batch size=512 (following the \citet{dosovitskiy2020image} settings) for fine-tuning and two-stage fine-tuning. Following \citet{dosovitskiy2020image}, all input images are normalized by $mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5)$.}
\section{Self-supervised transfer learning}
\label{apx:ssl}
\subsection{SWAV on \textsc{ImageNet}}
SWAV is a contrastive self-supervised learning algorithm proposed by \citet{caron2020unsupervised}. We train \textsc{resnet50} on \textsc{ImageNet}\footnote{\url{https://github.com/facebookresearch/swav/blob/main/scripts/swav_400ep_pretrain.sh}} by the SWAV algorithm four times, which gives us four pretrained \textsc{resnet50} models. As to the rest four SWAV pre-trained models in this work, we use the public available \textsc{resnet50}\footnote{\url{https://dl.fbaipublicfiles.com/deepcluster/swav_400ep_pretrain.pth.tar}}, \textsc{resnet50w2}\footnote{\url{https://dl.fbaipublicfiles.com/deepcluster/swav_RN50w2_400ep_pretrain.pth.tar}}, \textsc{resnet50w4}\footnote{\url{https://dl.fbaipublicfiles.com/deepcluster/swav_RN50w4_400ep_pretrain.pth.tar}}, and \textsc{resnet50w5}\footnote{\url{https://dl.fbaipublicfiles.com/deepcluster/swav_RN50w5_400ep_pretrain.pth.tar}} checkpoints.
\@startsection{paragraph}{4}{\z@}{.1ex plus 0ex minus .1ex}{-1em}{\normalsize\bf}{Linear probing:} Following the settings in \citet{goyal2022vision}, the linear probing experiments (on \textsc{ImageNet}, \textsc{Inat18}, \textsc{Cifar100}, \textsc{Cifar10}) add a \textsc{BatchNorm} layer before the last-layer linear classifier to reduce the hyper-parameter tuning difficulty. The learning rate is initialized to 0.01 and multiplied by 0.1 every 8 epochs. Then train these linear probing tasks for 28 epochs by SGD Nesterov optimizer with momentum=0.9. We search L2 weight decay from $\{5e-4\}$, $\{5e-4, 1e-3, 5e-3, 1e-2\}$, and $\{1e-6, 1e-5, 1e-4\}$ for \textsc{ImageNet}, \textsc{Cifar}, and \textsc{Inat18} tasks, respectively.
\@startsection{paragraph}{4}{\z@}{.1ex plus 0ex minus .1ex}{-1em}{\normalsize\bf}{Fine-tuning:}
\begin{itemize}
\item \textbf{\textsc{ImageNet}}: Inspired by the semi-supervised \textsc{ImageNet} fine-tuning settings in \citet{caron2020unsupervised}, we attach a randomly initialized last-layer classifier on top of the SSL learned representation. Then fine-tune all parameters, using a SGD optimizer with momentum=0.9 and L2 weight decay=0. Low-layers representation and last-layer classifier use different initial learning rates of 0.01 and 0.2, respectively. The learning rate is multiplied by 0.2 at $12^{th}$ and $16^{th}$ epochs. We train 20 epochs for networks: \textsc{resnet50}, \textsc{resnet50w2}, \textsc{resnet50w4}. We further search training epochs from $\{10, 20\}$ for the wide network (due to overfitting), \textsc{resnet50w5} and then select the best one with 10 training epochs.
\item \textbf{\textsc{Cifar10, Cifar100, Inat18}}: Same as the fine-tuning settings in supervised transfer learning in Appendix \ref{apx:imagenet_sl_settings}.
\end{itemize}
\@startsection{paragraph}{4}{\z@}{.1ex plus 0ex minus .1ex}{-1em}{\normalsize\bf}{Two-stage fine-tuning:}
\begin{itemize}
\item \textbf{\textsc{ImageNet}}: Similar to the two-stage fine-tuning settings in supervised transfer learning, we initialize the last-layer classifier $w$ by concatenation and then train 1 epoch with learning rate=0.001, L2 weight decay=0.
\item \textbf{\textsc{Cifar10, Cifar100, Inat18}}: For \textsc{Cifar10, Cifar100}, we use same two-stage fine-tuning settings as in supervised transfer learning in Appendix \ref{apx:imagenet_sl_settings}. For \textsc{Inat18}, we attach a \textsc{BatchNorm} layer before the last-layer linear classifier to reduce the training difficulty. Note that \textsc{BatchNorm} + a linear classifier is still a linear classifier during inference. Following the linear probing protocol, we train the \textsc{BatchNorm} and linear layers by a SGD optimizer with momentum=0.9, initial learning rate=0.01, and a 0.2 learning rate decay at $12^{th}$ and $16^{th}$ epochs. As to L2 weight decay, we use the same searching space as in the fine-tuning.
\end{itemize}
\subsubsection{Additional results}
\label{apx:swav_additional_exp}
Beside the SWAV \textsc{ImageNet} fine-tuning experiments in Figure \ref{fig:ssl_tf}, Figure \ref{fig:swav_ft_full} provides additional SWAV fine-tuning / two-stage fine-tuning results on \textsc{naturalist18}, \textsc{Cifar100}, and \textsc{Cifar10} tasks. We give a {\plotlabel{[init]cat}} curve on the \textsc{ImageNet} task, but omit the curves on other tasks (\textsc{naturalist18}, \textsc{Cifar100}, and \textsc{Cifar10}) because they are computationally costly.
\begin{figure}[ht]
\centering
\includegraphics[width=\textwidth]{swav_ft_v2.pdf}
\caption{Fine-tuning performance of SWAV on \textsc{ImageNet}, \textsc{naturalist18}, \textsc{Cifar100}, and \textsc{Cifar10} tasks. SWAV is trained on unlabeled \textsc{ImageNet}. {\plotlabel{[2ft]cat}} and {\plotlabel{[init]cat} indicate our two-stage fine-tuning strategy and the normal fine-tuning strategy on $n$ concatenated networks. {\plotlabel{[init]wide}} refers to the normal fine-tuning strategy on wide networks, i.e. \textsc{resnet50}, \textsc{resnet50w2}, \textsc{resnet50w4}, and \textsc{resnet50w5}. }}
\label{fig:swav_ft_full}
\end{figure}
\subsection{SEER on \textsc{Instagram1B}}
SEER \citep{goyal2022vision} trains large \textsc{regnet\{32gf, 64gf, 128gf, 256gf, 10B\}} architectures on the \textsc{Instagram1B} dataset with 1 billion Instagram images, using the SWAV contrastive self-supervised learning algorithm.
\@startsection{paragraph}{4}{\z@}{.1ex plus 0ex minus .1ex}{-1em}{\normalsize\bf}{Linear Probing:} Same as the linear probing settings in SWAV.
\@startsection{paragraph}{4}{\z@}{.1ex plus 0ex minus .1ex}{-1em}{\normalsize\bf}{Fine-tuning:} We use SEER checkpoints\footnote{\url{https://github.com/facebookresearch/vissl/tree/main/projects/SEER}} fine-tuned on \textsc{ImageNet} with $384 \times 384$ resolutions. It is fine-tuned on \textsc{ImageNet} for 15 epochs using SGD momentum 0.9, weight decay 1e-4, learning rate 0.04 and batch size 256. The learning rate is multiplied by 0.1 at $8^{th}$ and $12^{th}$ epochs.
\@startsection{paragraph}{4}{\z@}{.1ex plus 0ex minus .1ex}{-1em}{\normalsize\bf}{Two-stage Fine-tuning:} We keep L2 weight decay 1e-4 the same as fine-tuning. Then keep the other settings the same as in SWAV.
\subsection{Additional experiment: \textsc{SimSiam} on \textsc{Cifar}}
\label{apx:simsiam_cifar}
\textsc{SimSiam} \cite{chen2020simsiam} is a non-contrastive self-supervised learning algorithm. In this section, we pre-train the networks using \textsc{SimSiam} on \textsc{Cifar10}, then transfer the learned representation by linear probing to \textsc{Cifar10}, \textsc{Cifar100}, \textsc{Cifar10} with 1\% training examples, and \textsc{Cifar100} with 10\% training examples.
\@startsection{paragraph}{4}{\z@}{.1ex plus 0ex minus .1ex}{-1em}{\normalsize\bf}{\textsc{SimSiam} pre-training} Following \citet{chen2020simsiam} we pre-train \textsc{resnet18}, \textsc{resnet18w2}, \textsc{resnet18w4}, \textsc{2resnet18}, and \textsc{4resnet18} on \textsc{Cifar10} ($32 \times 32$ resolution) by \textsc{SimSiam} for $800$ epochs, using a SGD optimizer with momentum = $0.9$, initial learning rate = $0.06$, batch size = $512$, L2 weight decay = $5e-4$, and cosine learning rate scheduler. The data augmentations include \textsc{RandomResizedCrop} (crop scale in $[0.2, 1]$), \textsc{RandomHorizontalFlip}, \textsc{RandomGrayScale} ($p=0.2$), and a random applied \textsc{ColorJitter} ($0.4, 0.4, 0.4, 0.1$) with probability $0.8$. All images are normalized by $mean=(0.4914, 0.4822, 0.4465), std=(0.2023, 0.1994, 0.2010)$ before training.
\@startsection{paragraph}{4}{\z@}{.1ex plus 0ex minus .1ex}{-1em}{\normalsize\bf}{\textsc{\synthdistill}} Since self-supervised learning tasks don't contain target labels as supervised learning, we apply knowledge distillation on representation directly. Specifically, we set $v_1, \dots v_n$ in Figure \ref{fig:distill} as Identity matrices. Then we distill $[\phi_1, \dots, \phi_n]$ into $\Phi$ by use a cosine loss:
\begin{equation}
\min_{\Phi, w_0, \dots, w_n}\sum_{i=0}^n\sum_x\bigg[ 1-\cos\Big({\phi_i}(x)~, ~ w_i \circ \Phi(x) \Big)\bigg]
\end{equation}
\@startsection{paragraph}{4}{\z@}{.1ex plus 0ex minus .1ex}{-1em}{\normalsize\bf}{Linear Probing:} Following again the settings of \citet{goyal2022vision}, the linear probing experiments
(on \textsc{Cifar100}, \textsc{Cifar10}, \textsc{Cifar100(1\%)} with 10\% training data, and \textsc{Cifar10(1\%)} with 1\% training data) adds a \textsc{BatchNorm} layer before the last-layer
linear classifier to reduce the hyper-parameter tuning difficulty. We use batch size = 256 for \textsc{Cifar100} and \textsc{Cifar10}, use batch size = 32 for corresponding sampled (10\%/1\%) version. Then we search initial learning rate from $\{0.1, 0.01\}$, L2 weight decay from \{1e-4, 5e-4, 1e-3, 5e-3\}. The learning rate is multiplied by 0.1 every 8 epochs during the total 28 training epochs. As to the optimizer, all experiments use a SGD Nesterov optimizer with momentum=0.9.
\@startsection{paragraph}{4}{\z@}{.1ex plus 0ex minus .1ex}{-1em}{\normalsize\bf}{Results}
Table \ref{tab:simsiam} shows the linear probing accuracy of \textsc{SimSiam} learned representation on various datasets and architectures. When linear probing on the same \textsc{Cifar10} dataset as training, the {\synthcat}$n$ method performs slightly better than width architectures (e.g. \textsc{resnet18w2} and \textsc{resnet18w4}). When comparing them on the \textsc{Cifar100} dataset (OOD), however, {\synthcat}$n$ exceeds width architectures.
\begin{table}[ht]
\centering
\caption{Linear probing accuracy on \textsc{Cifar100}, \textsc{Cifar10}, \textsc{Cifar100(1\%)}, and \textsc{Cifar10(10\%)} tasks. The representation is learned on \textsc{Cifar10} by \textsc{SimSiam} algorithm. {\synthcat}$n$ concatenates $n$ learned representation before linear probing. {\synthdistill}$n$ distills $n$ learned representation into $\textsc{resnet18}$ before linear probing. \textsc{resnet18w}$n$ contains around $n^2$ parameters as {\textsc{resnet18}}.}
\bigskip
\begin{tabular}{cc|cc|cc}
\toprule
& &\multicolumn{2}{c|}{Linear Probing (ID)} & \multicolumn{2}{c}{Linear Probing (OOD) } \\
method & architecture & \textsc{Cifar10} & \textsc{Cifar10(1\%)} & \textsc{Cifar100} & \textsc{Cifar100(10\%)} \\
\midrule
\textsc{SimSiam} &\textsc{resnet18} & 91.88 &87.60 & 55.29 & 42.93 \\
\midrule
\textsc{SimSiam} &\textsc{resnet18w2} & 92.88 & 88.95 & 59.41 & 45.39\\
\textsc{SimSiam} &\textsc{resnet18w4} & 93.50 &90.45 & 59.28 & 44.98\\
\textsc{SimSiam} &\textsc{2resnet18} & 91.62 &87.14 & 55.67 & 43.07\\
\textsc{SimSiam} &\textsc{4resnet18} & 92.54 &85.65 & 64.42 & 49.65\\
\midrule
{\synthcat}2 & 2$\times$\textsc{resnet18} & 92.94 & 88.32 & 59.40 & 46.06\\
{\synthcat}4 & 4$\times$\textsc{resnet18} & 93.42 & 88.81 & 63.06 & 47.48\\
{\synthcat}5 & 5$\times$\textsc{resnet18} & 93.67 & 88.78 & 63.71 & 48.31\\
{\synthcat}10 & 10$\times$\textsc{resnet18} & 93.75 & 88.65 & 66.19 & 49.90\\
\midrule
{\synthdistill}2 & 2$\times$\textsc{resnet18} &93.04 &88.59 &59.65 &45.10 \\
{\synthdistill}5 & 5$\times$\textsc{resnet18} &93.02 &88.56 &60.79 &46.41 \\
{\synthdistill}10 & 10$\times$\textsc{resnet18} &93.11 &88.72 &61.35 &46.75 \\
\bottomrule
\end{tabular}
\label{tab:simsiam}
\end{table}
\subsection{Numerical results}
For reference, Tables \ref{tab:swav} and \ref{tab:seer} provide the numerical results for the linear probing, fine-tuning, and two-stage fine-tuning plots of Figure \ref{fig:ssl_tf}.
\begin{table}[ht]
\centering
\caption{Linear probing, fine-tuning, and two-stage fine-tuning performance of SWAV pre-trained representation and corresponding {\synthcat}$n$ representations. }
\label{tab:swav}
\bigskip
\begin{tabular}{ccc|cccc|c|c}
\toprule
& & & \multicolumn{4}{c|}{ \footnotesize linear-probing} & \footnotesize fine-tuning & \footnotesize two-stage ft \\
\footnotesize method & \footnotesize architecture & \footnotesize params & \footnotesize \textsc{ ImageNet} & \footnotesize \textsc{Cifar10} & \footnotesize \textsc{Cifar100} & \footnotesize \textsc{Inat18} & \footnotesize \textsc{Imagenet} & \footnotesize \textsc{Imagenet} \\
\midrule
SWAV & \textsc{resnet50} & 23.5M & 74.30 & 91.83 & 76.85 & 42.35 & 77.62 & - \\
SWAV & \textsc{resnet50w2} & 93.9M & 77.31 & 93.97 & 79.49 & 47.55 & 80.12 & - \\
SWAV & \textsc{resnet50w4} & 375M & 77.48 & 94.29 & 80.51 & 44.13 & 80.98 & - \\
SWAV & \textsc{resnet50w5} & 586M & 78.23 & 94.84 & 81.54 & 48.11 & 80.40 & - \\
\midrule
{\synthcat}2 & - & 47M & 76.01 & 93.48 & 78.91 & 45.57 & 78.14 & 79.00 \\
{\synthcat}5 & - & 118M & 77.43 & 94.62 & 81.11 & 49.12 & 77.69 & 80.04 \\
{\synthcat}7 & - & 587M & 78.72 & 95.59 & 82.71 & 49.68 & 80.05 & 81.25 \\
{\synthcat}9 & - & 1170M& 78.89 & 95.76 & 83.16 & 50.61 & 80.46 & 81.55 \\
\bottomrule
\end{tabular}
\end{table}
\begin{table}[ht]
\centering
\caption{Linear probing, fine-tuning, and two-stage fine-tuning performance of SEER pre-trained representation and corresponding {\synthcat}$n$ representations. }
\bigskip
\label{tab:seer}
\begin{tabular}{ccc|cccc|c|c}
\toprule
& & & \multicolumn{4}{c|}{ \footnotesize linear-probing} & \footnotesize fine-tuning & \footnotesize two-stage ft \\
\footnotesize method & \footnotesize architecture & \footnotesize params & \footnotesize \textsc{ ImageNet} & \footnotesize \textsc{Cifar10} & \footnotesize \textsc{Cifar100} & \footnotesize \textsc{Inat18} & \footnotesize \makecell{\textsc{Imagenet} \\ (384px)} & \footnotesize \makecell{\textsc{Imagenet} \\ (384px)} \\
\midrule
SEER & \textsc{regnet32gf} & 141M & 73.4 & 89.94 & 71.53 & 39.10 & 83.4 & - \\
SEER & \textsc{regnet64gf} & 276M & 74.9 & 90.90 & 73.78 & 42.69 & 84.0 & - \\
SEER & \textsc{regnet128gf} & 637M & 75.9 & 91.37 & 74.75 & 43.51 & 84.5 & - \\
SEER & \textsc{regnet256gf} & 1270M & 77.5 & 92.16 & 74.93 & 46.91 & 85.2 & - \\
\midrule
{\synthcat}2 & - & 418M & 76.0 & 92.16 & 75.65 & 45.36 & - & 84.5 \\
{\synthcat}3 & - & 1060M & 77.3 & 93.15 & 77.26 & 47.18 & - & 85.1 \\
{\synthcat}4 & - & 2330M & 78.3 & 93.59 & 78.80 & 48.68 & - & 85.5 \\
\bottomrule
\end{tabular}
\end{table}
\section{meta-learning / few-shots learning}
\label{apx:meta-learning}
\subsection{Datasets}
\textbf{\textsc{Cub}} \citep{wah2011caltech} dataset contains $11,788$ images of 200 birds classes, 100 classes ($5,994$ images) for training and 100 classes ($5,794$ images) for testing.
\textbf{\textsc{MiniImageNet}} \citep{matchingnet} dataset contains $60,000$ images of 100 classes with 600 images per class, 64 classes for training, 36 classes for testing.
\subsection{\textsc{Baseline} and \textsc{Baseline++} experiment Settings}
\label{apx:meta_baseline_pretrain}
For \textsc{Baseline} and \textsc{Baseline++} experiments, following \citet{closelookatfewshot}, we use \textsc{RandomSizedCrop}, \textsc{ImageJitter}(0.4, 0.4, 0.4), and \textsc{HorizontalFlip} augmentations, as well as a image normalization $mean=(0.485, 0.456, 0.406)$, $std=(0.229, 0.224, 0.225)$. Then use an \textsc{Adam} optimizer with learning rate = 0.001, batch size = 16, input image size = $224\times224$. Finally, train \textsc{resnet18} on \textsc{Cub} and \textsc{MiniImageNet} datasets for 200 and 400 epochs, respectively. We further tune L2 weight decay from \{0, 1e-5, 1e-4, 1e-3, 1e-2\} and choose 1e-4 for \textsc{Cub}, 1e-5 for \textsc{MiniImageNet} experiments. Compared with the \textsc{Baseline} and \textsc{Baseline++} performance reported by \citet{closelookatfewshot} (table A5), this L2 weight decay tuning process provides $\sim5\%$ and $\sim1\%$ improvement on \textsc{miniImagenet} 5way-1shot and 5way-5shot, respectively. In this work, we use this stronger setting in baseline methods.
As to the few-shots learning evaluation, following \citet{closelookatfewshot}, we scale images by a factor of 1.15, \textsc{CenterCrop}, and normalization. Then randomly sample 1 or 5 images from 5 random classes from the test set (5way-1shot and 5way-5shot). Finally, train a linear classifier on top of the learned representation with a SGD optimizer, momentum = 0.9, dampening = 0.9, learning rate = 0.1, L2 weight decay = 1e-3, batch size = 4, and epochs = 100. We take the average of 600 such evaluation processes as the test score.
The \textsc{Baseline} and \textsc{Baseline++} results in Figure \ref{fig:few_shots} report the mean of five runs with different training and evaluating seeds.
\@startsection{paragraph}{4}{\z@}{.1ex plus 0ex minus .1ex}{-1em}{\normalsize\bf}{Implementation details of the cosine classifier}
Here we summarize the technical details of the cosine classifier implementation used in this work which follows \citet{closelookatfewshot}\footnote{%
\url{https://github.com/wyharveychen/CloserLookFewShot/blob/master/backbone.py##L22}}.
Denote the representation vector as $z$. The cosine classifier calculates the $i^{th}$ element of logits by:
\begin{equation}
h_i = g_i \frac{\left<u_i,z\right>}{||u_i||||z||}
\end{equation}
where $u_i$ is a vector with the same dimension of $z$, $g_i$ is a scalar, $h_i$ is $i^{th}$ element of logits $h$.
Then minimize the cross entropy loss between the target label $y$ and softmax output $s(h)$ by updating $w$ and $g$: $\min_{w,g} \mathcal{L}_{ce}(y, s(h))$.
\subsection{{\synthcat} and {\synthdistill} experiment settings}
For {\synthcat}, we concatenate $n$ representation separately trained by either \textsc{Baseline} or \textsc{Baseline++} as the settings above. For {\synthdistill}, we use the same multi-head architecture as figure \ref{fig:distill} together with a cross-entropy loss function:
\begin{equation}
\label{eq:ce_kl_distill}
\min_{\Phi, w_0, \dots, w_n}\sum_{i=0}^n\sum_x\bigg[ (1-\alpha) \mathcal{L}_{ce}\Big(s\big(w_i \circ \Phi(x)\big), y \Big) + \alpha\tau^2\mathcal{L}_{kl}\Big(s_{\tau}\big({v_i} \circ {\phi_i}(x)\big) ~||~ w_i \circ \Phi(x) \Big)\bigg]
\end{equation},
where $\mathcal{L}_{ce}$ indicates a cross-entropy loss, $\alpha$ is a trade-off parameter between cross-entropy loss and kl-divergence loss. We set L2 weight decay = 0, $\tau=10$, search $\alpha \in \{0.8, 0.9, 1\}$, and keep the other hyper-parameters as Appendix \ref{apx:meta_baseline_pretrain}. We find the impact of $\alpha$ is limited in both \textsc{cub} ($\leq 1\%$) and \textsc{MiniImageNet} ($ \leq 0.5\%$) tasks.
\subsection{Snapshots experiment settings}
In this section, we apply {\synthcat} and {\synthdistill} on 5 snapshots sampled from one training episode (called \textsc{cat5-s} and \textsc{distill5-s}, respectively). We train $\textsc{cub}$ and $\textsc{MiniImageNet}$ respectively for 1000 and 1200 epochs by naive SGD optimizer with a relevant large learning rate 0.8. Then we sample 5 snapshots, $\{200^{th}, 400^{th}, 600^{th}, 800^{th}, 1000^{th}\}$ and $\{400^{th}, 600^{th}, 800^{th}, 1000^{th}, 1200^{th}\}$, for $\textsc{cub}$ and $\textsc{MiniImageNet}$, respectively. The other hyper-parameters are the same as Appendix \ref{apx:meta_baseline_pretrain}.
\subsection{More experimental results}
Table \ref{tab:few_shot_synt_cat} provides the exact number in Figure \ref{fig:few_shots}, as well as additional {\synthcat}$n$ and {\synthdistill}$n$ few-shots learning results with a linear classifier (The orange and gray bars in figure \ref{fig:few_shots} report the few-shots learning performance with a cosine classifier).
{Table \ref{tab:few_shot_synt_cat_snapshot} provides more \synthcat5\textsc{-s} and \synthdistill5\textsc{-s} results with either a linear classifier or a cosine-based classifier.}
\begin{table}[t]
\caption{Few-shots learning performance on \textsc{cub} and \textsc{miniImagenet}. The \textsc{cat5-s} and \textsc{distill5-s} results were obtained using five snapshots taken during a single training episode with a relatively high step size (0.8, SGD). The best snapshot performances are also reported. Standard deviations over five repeats are reported.
}
\label{tab:few_shot_synt_cat_snapshot}
\par\bigskip\centering
\begin{tabular}{ccc |cc |cc}
\toprule
& & & \multicolumn{2}{c|}{\textsc{cub}} & \multicolumn{2}{c}{\textsc{miniImagenet} } \\
& architecture & classifier & 5way 1shot & 5way 5shot & 5way 1shot & 5way 5shot \\
\midrule
best snapshot & \textsc{resnet18} & linear & 59.70$\pm$1.38 & 81.35$\pm$0.79 & 52.79$\pm$0.92 & 75.18$\pm$0.57 \\
\synthcat5\textsc{-s} & \textsc{5$\times$resnet18} & linear& 72.62$\pm$0.98 & 86.56$\pm$0.82 & 61.91$\pm$0.37 & 81.06$\pm$0.14 \\
\synthdistill5\textsc{-s} & \textsc{resnet18} & linear& 68.4$\pm$0.5 & 87.2$\pm$0.4 &
59.9$\pm$0.5 & 80.8$\pm$0.4 \\
\midrule
best snapshot & \textsc{resnet18} & cosine & 65.59$\pm$0.87 & 81.81$\pm$0.50 & 55.67$\pm$0.48 & 75.48$\pm$0.46 \\
\synthcat5\textsc{-s} & \textsc{5$\times$resnet18} & cosine & 73.66$\pm$0.82 & 87.25$\pm$0.77 & 62.94$\pm$0.51 & 81.05$\pm$0.16 \\
\synthdistill5\textsc{-s} & \textsc{resnet18} & cosine &
75.2$\pm$0.8 & 88.6$\pm$0.4 & 62.0$\pm$0.5 & 81.0$\pm$0.3 \\
\bottomrule
\end{tabular}
\end{table}
\begin{table}[ht]
\centering
\caption{Few-shot learning performance on \textsc{cub} and \textsc{miniImagenet} dataset with either a linear classifier or cosine-distance based classifier. Standard deviations over five repeats are reported. }
\bigskip
\label{tab:few_shot_synt_cat}
\begin{tabular}{ccc |cc |cc}
\toprule
& & & \multicolumn{2}{c|}{\textsc{cub}} & \multicolumn{2}{c}{\textsc{miniImagenet} } \\
& architecture & classifier & 5way 1shot & 5way 5shot & 5way 1shot & 5way 5shot \\
\midrule
supervised & \textsc{resnet18} & linear & 63.37$\pm$1.66 & 83.47$\pm$1.23 & 55.20$\pm$0.68 & 76.52$\pm$0.42 \\
{\synthcat}2 & $2\times$\textsc{ResNet18} & linear &66.25$\pm$0.85 & 85.50$\pm$0.34 & 57.30$\pm$0.31 & 78.42$\pm$0.17 \\
{\synthcat}5 & $5\times$\textsc{resnet18} & linear &67.00$\pm$0.18 & 86.80$\pm$0.10 & 58.40$\pm$0.25 & 79.59$\pm$0.17 \\
{\synthdistill}2& \textsc{resnet18} & linear&69.93$\pm$0.74 & 87.72$\pm$0.31 & 58.99$\pm$0.32 & 79.73$\pm$0.21 \\
{\synthdistill}5& \textsc{resnet18} & linear&70.99$\pm$0.31 & 88.52$\pm$0.14 & 59.66$\pm$0.59 & 80.53$\pm$0.27 \\
\midrule
supervised & \textsc{resnet18} & cosine&69.19$\pm$0.88 & 84.41$\pm$0.49 & 57.47$\pm$0.45 & 76.47$\pm$0.27 \\
{\synthcat}2 & $2\times$\textsc{resnet18} & cosine&72.87$\pm$0.43 & 86.82$\pm$0.17 & 60.69$\pm$0.24 & 79.29$\pm$0.23 \\
{\synthcat}5 & $5\times$\textsc{resnet18} & cosine&76.23$\pm$0.55 & 88.87$\pm$0.40 & 63.63$\pm$0.23 & 81.22$\pm$0.17 \\
{\synthdistill}2& \textsc{resnet18} & cosine & 74.81$\pm$0.45 & 88.14$\pm$0.40 & 61.95$\pm$0.11 & 80.79$\pm$0.26 \\
{\synthdistill}5& \textsc{resnet18} & cosine & 76.20$\pm$0.39 & 89.18$\pm$0.24 & 62.89$\pm$0.38 & 81.49$\pm$0.26 \\
\bottomrule
\end{tabular}
\end{table}
\section{Out-of-distribution learning}
\label{apx:ood}
Following \citet{zhang2022rich}, we use the \textsc{Camelyon17} \citep{koh2021wilds} task to showcase the $\synthcat$ and $\synthdistill$ constructed (rich) representation in out-of-distribution learning scenario. The first row of table \ref{tab:camelyon17_synt_cat} is copied from \citet{zhang2022rich}. The rest results use a frozen pre-trained representation, either by concatenating $n$ ERM pre-trained representations ({\synthcat}$n$), distilling of $n$ ERM pre-trained representations ({\synthdistill}$n$), or RFC constructed representations (RFC2). Then train a linear classifier on top of the representation by vREx or ERM algorithms.
For the vREx algorithm, we search the penalty weights from \{0.5, 1, 5, 10, 50, 100\}. For {\synthdistill}$n$ representations in the \textsc{Camelyon17} task, we follow Algorithm 2 in \citet{zhang2022rich}, but use a slightly different dataset balance trick in the loss function (\citet{zhang2022rich} Algorithm 2 line 13-14). We instead balance two kinds of examples: one shares the same predictions on all ERM pre-trained models, and one doesn't. We keep other settings to be the same as \citet{zhang2022rich}\footnote{\url{https://github.com/TjuJianyu/RFC}}.
\section{Replicability}
Code will of course be provided on github to replicate these experiments.
As explained in the paper, many of our experiments are based the setups defined by other authors and rely on the code and the data files they provide. We are therefore very conscious of the value of making well organized replication resources available to the community.
\section{Conclusion}
Using a simple theoretical framework and a broad range of experiments, we show that deep learning scenarios that involve changing tasks or distributions are \emph{better served by representations that are richer than those obtained with a single optimization episode.} In a time where many organizations
deploy considerable resources training huge foundational
models, this conclusion should be sobering.
The simple multiple-training-episode approach $\synthcat$ constructs such richer representation with excellent performances in various scenarios. The \emph{two-stage fine tuning} method works around the poor performance of normal fine-tuning in various transfer scenarios.
More importantly, this work provides a lot of room for new representation learning algorithms that move away from relying solely on a single optimization episode.
\section{Features and representations}
\label{sec:features}
This section provides a conceptual framework for talking about richness and diversity of representations. Although it seems natural to compare representations using information theory concepts such as mutual information, this approach is fraught with problems. For instance, the simplest way to maximize the mutual information $M(\Phi(x),y)$ between the representation $\Phi(x)$ and the desired output $y$ consists of making $\Phi$ equal to the identity. The information theoretic approach overlooks the main role of a feature extraction function, which is not filtering the information present in the inputs $x$, but formatting it in a manner exploitable by a simple learning system such as a linear classifier or a linear regression. The following framework relies on the linear probing error instead.
\@startsection{paragraph}{4}{\z@}{.1ex plus 0ex minus .1ex}{-1em}{\normalsize\bf}{Framework}
We call \emph{feature} a function $x\,{\mapsto}\,\varphi(x)\,{\in}\,\mathbb{R}$, and we call \emph{representation} a set $\Phi$ of features. We use the notation $\mathbf{w}^\top\Phi(x)$ to denote the dot product $\sum_{\varphi\in\Phi} w_\varphi\,\varphi(x)$ where the coefficients $w_\varphi$ of vector $\mathbf{w}$ are indexed by the corresponding feature $\varphi$ and are assumed zero if $\varphi\notin\Phi$.
We assume for simplicity that our representations are exploited with a linear classifier trained with a convex loss~$\ell$. The expected loss of classifier $f$ is
\[ C_P(f) = \mathbb{E}_{(x,y)\sim P}\:\big[\,\ell(f(x),y)\,\big]\]
and the optimal cost achievable with representation $\Phi$
\begin{equation}
\label{eq:ecost}
C^*_P(\Phi) = \min_\mathbf{w} ~ C_P(f) ~~\text{with}~~ f: x\mapsto \mathbf{w}^\top \Phi(x)~.
\end{equation}
This construction ensures:
\begin{proposition}
\label{prop:min}
$C^*_P(\Phi_1 \cup \Phi_2) \leq C^*_P(\Phi_2)$ for all $\Phi_1$, $\Phi_2$.
\end{proposition}
Intuitively, if the combined representation $\Phi_1 \cup \Phi_2$ performs better than $\Phi_2$, then $\Phi_1$ must contain something useful that $\Phi_2$ does not. We formalize this using the word \emph{information} to actually mean \emph{linearly exploitable information about $y$}.
\begin{definition}
\label{def:more}
$\Phi_1$ contains information not present in $\Phi_2$\\
iff~
$C^*_P(\Phi_1\cup\Phi_2) < C^*_P(\Phi_2)$.
\end{definition}
Thanks to proposition~\ref{prop:min}, the opposite property becomes\,:
\begin{definition}
\label{def:notmore}
$\Phi_2$ contains all the information present in $\Phi_1$\\
iff~ $C^*_P(\Phi_1\cup\Phi_2) = C^*_P(\Phi_2)$.
\end{definition}
Finally we say that $\Phi_1$ and $\Phi_2$ carry equivalent information when $\Phi_2$ contains all the information present in $\Phi_1$, and $\Phi_1$ contains all the information present in $\Phi_2$\,:
\begin{definition}
\label{def:equiv}
$\Phi_1$ and $\Phi_2$ carry equivalent information\\
iff~
$C^*_P(\Phi_1) = C^*_P(\Phi_1 \cup \Phi_2) = C^*_P(\Phi_2)$.
\end{definition}
This definition is stronger\footnote{This is also weaker than using the quantity of information~$H$\,: writing $H(\Phi_1){=}H(\Phi_1{\cup}\Phi_2){=}H(\Phi_2)$ would imply that $\Phi_1$ and $\Phi_2$ are equal up to a bijection. Theorems~\ref{prop:ensemble} and~\ref{prop:opt} are important because this is not the case here.} than merely requiring equality $C^*_P(\Phi_1)\,{=}\,C^*_P(\Phi_2)$. In particular, we cannot improve the expected cost by constructing an ensemble\,:
\begin{theorem}
\label{prop:ensemble}
Let representations $\Phi_1$ and $\Phi_2$ carry equivalent information. Let $f_i(x){=}\mathbf{w}_i^{*\top}\Phi_i(x)$, for $i{\in}\{1,2\}$, be corresponding optimal classifiers. Then, for all $0{\leq}\lambda{\leq}1$\,,
\[ C^*_P(\lambda f_1 + (1-\lambda) f_2) = C^*_P(f_1) = C^*_P(f_2). \]
\end{theorem}
\begin{proof}
Let $\Phi=\Phi_1\cup\Phi_2$. Because the loss $\ell$ is assumed convex, the solutions of optimization problem \eqref{eq:ecost} form a convex set $S$. Since $C^*_P(\Phi_1){=}C^*_P(\Phi_1 \cup \Phi_2){=}C^*_P(\Phi_2)$, set $S$ contains $w^*_1$ and $w^*_2$, as well as any mixture thereof.
\end{proof}
We now turn our attention to representations constructed by optimizing
both the representation $\Phi$ and the weights $\mathbf{w}$:
\begin{equation}
\label{eq:optrepr}
\min_\Phi C^*_P(\Phi) = \min_{\Phi} \min_\mathbf{w} \mathbb{E}_{(x,y)\sim P} [ \ell( \mathbf{w}^\top \Phi(x), y) ]\,.
\end{equation}
This idealized formulation optimizes the expected error without constraints on the nature and number of features. All its solutions problem carry equivalent information\,:
\begin{theorem}
\label{prop:opt}
Let $\Phi_1$ and $\Phi_2$ be two solutions of problem~\eqref{eq:optrepr}.
Then $\Phi_1$ and $\Phi_2$ carry equivalent information.
\end{theorem}
\begin{proof}
Proposition~\ref{prop:min} implies $C^*_P(\Phi_1\cup\Phi_2) \leq C^*_P(\Phi_1)$.
Since $\Phi_1$ and $\Phi_2$ are both solutions of problem~\ref{eq:optrepr}, $C^*_P(\Phi_1) = C^*_P(\Phi_2) \leq C^*_P(\Phi_1\cup\Phi_2) \leq C^*_P(\Phi_1)$.
\end{proof}
\@startsection{paragraph}{4}{\z@}{.1ex plus 0ex minus .1ex}{-1em}{\normalsize\bf}{In-distribution viewpoint}
Consider a deep network that is sufficiently overparameterized to accommodate any useful representation in its penultimate layer. Assume that we are able to optimize its expected cost on the training distribution, that is, optimize its in-distribution generalization error. Although repeated optimization episodes need not return exactly the same representations, Theorem~\ref{prop:opt} tells us that these representations \emph{carry equivalent information}; Definition~\ref{def:equiv} tells us that we cannot either improve the in-distribution test error by linear probing, that is, by training a linear layer on top of the concatenated representations; and Theorem~\ref{prop:ensemble} tells us that we cannot improve the test error with an ensemble of such networks. The performance of ensembles depends on the diversity of their components~\cite{dietterich2000ensemble,ganaie2021ensemble}, and nothing has been done here to obtain diverse networks.
In practice, we cannot truly optimize the expected error of an overparameterized network. The representations obtained with separate training episodes tend to carry equivalent information but will not do so exactly.\footnote{Experience shows however that repeated trainings on large tasks, such as \textsc{ImageNet}, yields networks with remarkably consistent training and testing performances.} Although an ensemble of such identically trained networks can still improve both the training and testing errors, using such similarly trained networks remains a poor way to construct ensembles when one can instead vary the training data, the hyper-parameters, or vary the model structure \cite{ganaie2021ensemble}. Engineering diversity escapes the setup of Theorem~\ref{prop:opt} because each component of the ensemble then solves a different problem. This is obviously better than relying on how the real world deviates from the asymptotic setup.
\@startsection{paragraph}{4}{\z@}{.1ex plus 0ex minus .1ex}{-1em}{\normalsize\bf}{Out-of-distribution viewpoint}
Assume now that we train our network on a first data distribution $P(x,y)$, but plan to use these networks, or their representations, or their inner layers, with data that follow a different distribution $Q(x,y)$. Doing so also escapes the assumptions of our framework because the definition of representation carrying similar information (Definition~\ref{def:equiv}) critically depends on the data distribution. Representations that carry equivalent information for the training distribution $P$ need not carry equivalent information for a new distribution $Q$ at all.\footnote{Information theoretical concepts are also tied to the assumed data distribution. For instance, whether two features have mutual information critically depends on the assumed data distribution.}
Consider again representations obtained by performing multiple training episodes of the same network that only differ by their random seed.\footnote{The random seed here may determine the initial weights, the composition of the mini-batches, or the data augmentations. It does not affect the data distribution, the model structure, or even the training algorithm hyper-parameters.} These representations roughly carry equivalent information with respect to the training distribution, but, at the same time, may be very far from carrying equivalent information with respect to a new distribution.
If this is indeed the case, \emph{constructing an ensemble of such similarly trained networks can have a far greater effect on out-of-distribution data than in-distribution.}
Experimental results reported in the following sections will demonstrate this effect. In fact, since we cannot know which of these representations or features might prove more informative on the new distribution, it seems wise to keep them all. \emph{Premature feature selection is not a smart way to prepare for distribution changes.}
\@startsection{paragraph}{4}{\z@}{.1ex plus 0ex minus .1ex}{-1em}{\normalsize\bf}{Optimization dynamics}
There is growing evidence that implicit regularization in deep learning networks is related to various flavors of sparsity (\eg. \citealp{andriushchenko2022sgd,blanc2020implicit}). In an oversimplified account of this complex literature, the learning process explores the feature space more or less randomly; features that carry incrementally useful information stick more than those who do not. Consider for instance a network with representation~$\Phi_t$ at iteration $t$ and a feature $\varphi\in\Phi_t$ whose information is already present in $\Phi_t\!{\smallsetminus}\{\varphi\}$ in the sense of Definition~\ref{def:notmore}. This feature does not incrementally improve the training distribution performance and therefore may not stick. Yet this feature might contain useful information when compared to a different representation, or when compared to $\Phi_t\!{\smallsetminus}\{\varphi\}$ under a different distribution.
Explicit regularization in deep networks, such as the ubiquitous slight weight decay, also tends to destroy features that appear redundant. \citet{papyan-2020} describes how representations collapse when one trains a network for a very long time. \citet{shwartz2017opening} describe competing processes that create representations and prune representations in all layers at once.
\begin{table}[t]
\par\vspace*{-2ex}
\caption{Impact of L2 weight decay on supervised transfer learning between \textsc{Cifar10} and \textsc{Cifar100}.}
\par\vspace*{-1ex}
\label{tab:supervise_transfer_cifar}
\bigskip\centering
\resizebox{0.39\textwidth}{!}{
\begin{tabular}{c|cc}
\toprule
L2 weight decay & $0$ & $5e-4$ \\
\midrule
\textsc{Cifar10} &
91.41$\pm$0.81 & \textbf{94.89$\pm$0.23} \\
\textsc{Cifar10$\rightarrow$Cifar100} &
\textbf{49.68$\pm$0.72} & 29.17$\pm$0.50 \\
\midrule
\textsc{Cifar100} &
70.37$\pm$1.49 &\textbf{76.78$\pm$0.36} \\
\textsc{Cifar100$\rightarrow$Cifar10} &
\textbf{78.87$\pm$0.98} & 75.92$\pm$0.54 \\
\bottomrule
\end{tabular}
}
\par\vspace*{-2ex}
\end{table}
Table~\ref{tab:supervise_transfer_cifar} reports on a simple experiment to illustrate how capacity control with regularization can help in-distribution performance but hurt when the distribution changes. We pre-train a \textsc{resnet18} on the \textsc{Cifar10} task and transfer its learned representation to a \textsc{Cifar100} task by linear probing (see setups in appendix \ref{apx:cifar10_100_sl}). %
Although the best in-distribution performance, 94.9\%, is achieved using a slight weight decay, the representation learned \emph{without weight decay} transfers far better (49.7\% versus 29.2\%). The same observation holds when one reverses the role of the \textsc{Cifar10} and \textsc{Cifar100} datasets.
\@startsection{paragraph}{4}{\z@}{.1ex plus 0ex minus .1ex}{-1em}{\normalsize\bf}{Next steps}
The remaining sections of this paper describe experiments that investigate the effect of concatenating representations obtained by multiple training episodes that only differ by their random seed.
Despite the \emph{intentional lack of diversity} of these ensembles, the performance improvements observed on tasks involving distribution changes are far greater than the in-distribution performance improvements. This proves that representations constructed by multiple training episodes are indeed different. Even though their concatenation carries little additional information for in-distribution, these experiments show how they become substantially more informative when tasks
or distributions change.
Meanwhile, we obtain worse performance (a) when we train a network whose size matches that of the ensemble from scratch, or (b) when we fine-tune the concatenated representations in a single additional run. We contend that this happens because optimization inherently impoverishes the representations in a manner that makes sense in-distribution but hurts out-of-distribution, and we propose \emph{two-stage fine-tuning} (Figure~\ref{fig:twostageft}) to correct this behavior.
\section{Introduction}
Although the importance of features in machine learning systems was already clear when the Perceptron was invented \citep{rosenblatt-1957}, learning features from examples was often considered a hopeless task \citep{minsky-papert-1969}. Some researchers hoped that random features were good enough, as illustrated by the Perceptron. Other researchers preferred to manually design features using substantive knowledge of the problem \citep{simon-1989}. This changed when \citet{rumelhart-1986} showed the possibility of feature learning as a side effect of the risk optimization. Despite reasonable concerns about the optimization of nonconvex cost functions, feature discovery through optimization has driven the success of deep learning methods.
There are however many cues suggesting that learning features no longer can be solely understood through the optimization of the expected error for a single data distribution. First, adversarial examples \citep{szegedy-2014} and shortcut learning \citep{geirhos-2020} illustrate the need to make learning systems that are robust to certain changes of the underlying data distribution and therefore involve multiple expected errors. Second, the practice of transferring features across related tasks \citep{tr-bottou-2011,collobert-2011,oquab-2014} is now viewed as foundational \citep{bommasani-2021} and intrinsically involves multiple data distributions and cost functions. It is therefore timely to question whether the optimization of a single cost function creates and accumulates features in ways that make the most sense in this broader context.
This contribution reports on experiments showing how the out-of-distribution performance of a deep learning model benefits from internal representations that are richer and more diverse than those computed with a single optimization episode. More precisely, \emph{although optimization can produce diverse features, a single run is unable to collect them all into a rich representation that performs better when tasks or distributions change}. In a time where many organizations deploy considerable resources training huge foundational models, this observation should be sobering.
\@startsection{paragraph}{4}{\z@}{.1ex plus 0ex minus .1ex}{-1em}{\normalsize\bf}{Organization of the manuscript}
Section~\ref{sec:features} provides simple theoretical tools to discuss the value of features, discusses their consequences in-distribution and out-of-distribution, and approaches the notion of feature redundancy and feature richness. Sections~\ref{sec:supervisedtransfer}, \ref{sec:ssltransfer}, \ref{sec:metalearning}, and \ref{sec:oodlearning}, present experimental results pertaining respectively to supervised transfer learning, self-supervised transfer learning, meta-learning, and out-of-distribution learning.
\clearpage
\section{Meta-learning \& few-shots learning}
\label{sec:metalearning}
\begin{figure}[ht]
\centering
\includegraphics[width=0.48\textwidth]{plots/meta_learning_full_v4.pdf}
\par\vspace*{-2ex}
\caption{Few-shot learning performance on \textsc{MiniImageNet} and \textsc{Cub}. Four common few-shot learning algorithms are shown in red (results from \citet{closelookatfewshot}). Two supervised transfer methods, with either a linear classifier (\textsc{Baseline}) or cosine-based classifier (\textsc{Baseline++}) are shown in blue. The $\synthdistill$ and $\synthcat$ results, with a cosine-base classifier, are respectively shown in orange and gray. The \synthcat5\textsc{-s} and \synthdistill5\textsc{-s} results were obtained using five snapshots taken during a single training episode with a relatively high step size. The dark blue line shows the best individual snapshot. Standard deviations over five repeats are reported.}
\label{fig:few_shots}
\vspace*{-2ex}
\end{figure}
Each target task in the few-shots learning scenario comes with only a few training examples. One must then consider a large collection of target tasks to obtain statistically meaningful results.
We follow the setup of \citet{closelookatfewshot}\footnote{We are aware of various existing few-shot benchmarks, such as MetaDataset \citep{triantafillou2019meta}, that contain more datasets than \citet{chen2020simple}. We choose \citet{chen2020simple}, because it is enough to validate our ideas in section \ref{sec:features}.
} in which the base task is an image classification task with a substantial number of classes and examples per class, and the target tasks are five-way classification problems involving novel classes that are distinct from the base classes and come with only a few examples. Such a problem is often cast as a \emph{meta learning} problem in which the base data is used to learn how to solve a classification problem with only a few examples. \citet{closelookatfewshot} find that excellent performance can be achieved using simple baseline algorithms such as supervised transfer learning with linear probing (\textsc{Baseline}) or with a cosine-based final classifier (\textsc{Baseline++}). These baselines match and sometimes exceed the performance of common few-shots algorithms such as \textsc{maml} \citep{finn2017model}, \textsc{RelationNet} \citep{sung2018learning}, \textsc{MatchingNet} \citep{matchingnet}, and \textsc{ProtoNet} \citep{protonet}.
Figure~\ref{fig:few_shots} reports results obtained with a \textsc{resnet18} architecture on both the \textsc{MiniImageNet} \citep{matchingnet} and \textsc{Cub} \citep{wah2011caltech} five ways classification tasks with either one or five examples per class as set up by \citet{closelookatfewshot}. The \textsc{maml}, \textsc{RelationNet}, \textsc{MatchingNet}, and \textsc{ProtoNet} results (red bars) are copied verbatim from \citep[table A5]{closelookatfewshot}. The \textsc{Baseline} and \textsc{Baseline++} results were further improved by a systematic L2 weight decay search procedure (see appendix \ref{apx:meta_baseline_pretrain}). All these results show substantial variations across runs, about $4\%$ for \textsc{Cub} and $2\%$ for \textsc{MiniImageNet}.
The \synthcat$n$ and \synthdistill$n$ results were then obtained by first training $n$ \textsc{resnet18} on the base data with different seeds, constructing a combined (rich) representation by either concatenation or distillation (as Figure~\ref{fig:distill}), then, for each task, training a cosine distance classifier using the representation as input. Despite the high replication variance of the competing results, both $\synthdistill$ and $\synthcat$ show very strong performance. Note that naively increasing model architecture, e.g. from \textsc{resnet18} to \textsc{resnet34}, can only gain limited improvements ($\leq 1\%$, \citet{chen2020simple}, table A5) and is still lagging behind $\synthcat$ and $\synthdistill$.
\begin{figure}
\centering
\includegraphics[width=0.37\textwidth]{art/synthdistill.pdf}
\par\vspace*{-1.5ex}
\caption{(\synthdistill$n$) A multiple head network (red) trained to predict the outputs of the pre-trained networks $\Phi_1,\Phi_2,\cdots$ (blue) must develop a representation $\Phi$ that subsumes those of all the blue networks. The same distillation process is used by the \textsc{RFC} algorithm \citep{zhang2022rich} but after training the networks with adversarially re-weighted data. }
\label{fig:distill}
\vspace*{-3ex}
\end{figure}
The pink bars (\synthcat5\textsc{-s} and \synthdistill5\textsc{-s}) in Figure~\ref{fig:few_shots}, concatenate or distill five snapshots taken at regular intervals during a single training episode with a relatively high step size (0.8), achieve a similar few-shots learning performance as \synthcat5 and \synthdistill5, perform substantially better than the best individual snapshot (dark blue line). \emph{It implies that diverse features are discovered and then abandoned but not accumulated during the optimization process.} More results and details are shown in appendix \ref{apx:meta-learning}.
\section{Out-of-distribution generalization}
\label{sec:oodlearning}
In the out-of-distribution generalization scenario, we seek a model that performs well on a family of data distributions, also called environments, on the basis of a finite number of training sets distributed according to some of these distributions. \citet{irm} propose an invariance principle to solve such problems and propose the \textsc{IRMv1} algorithm which searches for a good predictor whose final linear layer is simultaneously optimal for all training distributions.
Since then, a number of algorithms exploiting similar ideas have been proposed, such as \textsc{vREx} \citep{vrex}, \textsc{Fishr} \citep{rame2021fishr}, or \textsc{CLOvE} \citep{wald2021calibration}. Theoretical connections have been made with multi-calibration \citep{hebertjohnson18a,wah2011caltech}. Alas, the performance of these algorithms remains wanting \citep{gulrajani2021in}.
\citet{zhang2022rich} attribute this poor performance to the numerical difficulty of optimizing the complicated objective associated with these algorithms. They propose to work around these optimization problems by providing initial weights that already extract a rich palette of potentially interesting features constructed using the \textsc{RFC} algorithm.
Following \citet{zhang2022rich}, we use the \textsc{Camelyon17} tumor classification dataset \citep{bandi2018detection} which contains medical images collected from five hospitals with potentially different devices and procedures. As suggested in \citet{koh2021wilds}, we use the first three hospitals as training environments and the fifth hospital for testing. \textsc{ood}-tuned results are obtained by using the fourth hospital to tune the various hyper-parameters. \textsc{iid}-tuned results only use the training distributions (see details in appendix \ref{apx:ood}). The purpose of our experiments is to investigate whether initializing with the {\synthdistill} or {\synthcat} algorithm provides a computationally attractive alternative to \textsc{RFC}.
\begin{table}[t]
\par\vspace*{-2ex}
\caption{Test accuracy on the \textsc{Camelyon17} dataset with \textsc{DenseNet121}. We compare various initialization (\textsc{ERM}, \synthcat$n$, \synthdistill$n$, and \textsc{RFC}) for two algorithms \textsc{vREx} and \textsc{ERM} using either the \textsc{iid} or \textsc{ood} hyperparameter tuning method. The standard deviations over 5 runs are reported. }
\par\vspace*{-2ex}
\label{tab:camelyon17_synt_cat}
\centering
\bigskip
\resizebox{0.485\textwidth}{!}{
\begin{tabular}{c | cc |cc}
\toprule
& \multicolumn{2}{c|}{\small IID-Tune} & \multicolumn{2}{c}{\small OOD-Tune} \\
& {\small \textsc{vREx}} & {\small \textsc{ERM}} & {\small \textsc{vREx}} & {\small \textsc{ERM}} \\
\midrule
{\small \textsc{ERM}} & 69.6$\pm$10.5 & 66.6$\pm$9.8 & 70.6$\pm$10.0 & 70.2$\pm$8.7 \\
\midrule
{\synthcat}2 & 74.3$\pm$8.0 & 74.3$\pm$8.0 & 73.7$\pm$8.1 & 74.2$\pm$8.1 \\
{\synthcat}5 & 75.2$\pm$2.9 & 75.0$\pm$2.7 & 74.9$\pm$3.3 & 75.1$\pm$2.8 \\
{\synthcat}20 & 76.4$\pm$0.5 & 76.5$\pm$0.5 & 76.8$\pm$0.9 & 76.4$\pm$0.9 \\
\midrule
{\synthdistill}2 & 67.1$\pm$4.7 & 66.9$\pm$4.8 & 67.4$\pm$4.3 & 66.7$\pm$4.2 \\
{\synthdistill}5 & 69.9$\pm$7.4 & 69.9$\pm$6.9 & 71.8$\pm$5.0 & 69.9$\pm$6.3 \\
{\synthdistill}20 & 73.3$\pm$2.5 & 73.2$\pm$2.3 & 74.8$\pm$3.2 & 73.1$\pm$2.7 \\
\midrule
\textsc{RFC}2 & 77.9$\pm$2.7 & 78.2$\pm$2.6 & 79.5$\pm$2.7 & 78.6$\pm$2.6 \\
\bottomrule
\end{tabular}
}
\par\vspace*{-2ex}
\end{table}
Table~\ref{tab:camelyon17_synt_cat} compares the test performance achieved by two algorithms, \textsc{vREx} and \textsc{ERM}, after initializing with \textsc{ERM}, \synthcat$n$, \synthdistill$n$, and \textsc{RFC2}, in both the \textsc{iid}-tune and \textsc{ood}-tune scenarios. The {\synthcat} and {\synthdistill} initialization perform better than \textsc{ERM} but not as well as \textsc{RFC}. \emph{This result clearly shows the need to research better ways to train networks in a manner that yields diverse representations.} Although this contribution shows that simply changing the seed (as in {\synthcat} and {\synthdistill}) can achieve good results, the experience of deep ensembles \citep{gontijo-lopes2022no} suggests that more refined diversification methods might yield substantially better representations.
\section{Related work}
\@startsection{paragraph}{4}{\z@}{.1ex plus 0ex minus .1ex}{-1em}{\normalsize\bf}{Representations and optimization}
\citet{papyan-2020} shows how deep network representations collapse to a ``simplex equiangular tight frame'' when one trains for a very long time. \citet{shwartz2017opening} argue that the training process first develops representations in unsupervised mode, then prunes away the unnecessary features. Both papers associate this representation impoverishment with better generalization (in-distribution). We argue that it hurts performance when distributions change. Closer to our concerns, \citet{pezeshki2021gradient} describe the gradient starvation phenomenon which makes it hard to find the right features when a network already has spurious features. They do not however consider how to produce rich representations with multiple training episodes.
\@startsection{paragraph}{4}{\z@}{.1ex plus 0ex minus .1ex}{-1em}{\normalsize\bf}{Ensembles}
\cite{dietterich2000ensemble} argues that model diversity is essential for the generalization performance of ensembles. \citet{ganaie2021ensemble} reviews deep ensembles with the same conclusion. Our work focuses instead on scenarios involving multiple tasks or data distributions. We purposely refrain from engineering diversity, still observe substantial improvements, and draw conclusions about the undesirable properties of optimization.
\@startsection{paragraph}{4}{\z@}{.1ex plus 0ex minus .1ex}{-1em}{\normalsize\bf}{Universal representations}
Several authors \citep{wang2022cross,dvornik2020selecting,bilen2017universal,gontijo2021no,li2021universal,li2022universal, chowdhury2021few} have recently proposed to collect features obtained with different tasks, datasets, network architectures, or hyper-parameters. The resulting so-``universal'' representations can be helpful for a variety of tasks. This approach is certainly interesting for practical problems but would not have allowed us to draw our conclusions.
\@startsection{paragraph}{4}{\z@}{.1ex plus 0ex minus .1ex}{-1em}{\normalsize\bf}{Model soups}
Another line of work uses weight averaging to combine the properties
of diverse networks~\citep{wortsman2022model,rame2022diverse}, with an increasing focus on leveraging models trained on multiple tasks to achieve a high performance on a task of interest~\citep{ilharco2022editing,rame2022recycling}. This engineered diversity provides for high performance, but does not allow the authors to draw conclusions about the optimization process itself.
\@startsection{paragraph}{4}{\z@}{.1ex plus 0ex minus .1ex}{-1em}{\normalsize\bf}{Shortcut learning and mitigation}
Several authors \citep[e.g.][]{huang2020self, teney2022evading} propose to work around the shortcut learning problem \citep{geirhos-2020} by shaping the last-layer classifier or introducing penalty terms in a manner that favors richer representations. \citet{zhang2022rich} argue that such additions make the optimization very challenging, but can be managed by initializing the networks with a rich representation constructed with an elaborate multi-step process. We show that rich representations can also be built by merely training the same network multiple times and combining their representations.
\section{Self-supervised transfer learning}
\label{sec:ssltransfer}
In self-supervised transfer learning (SSL), transferable representations are no longer constructed using a supervised auxiliary task, but using a training criterion that does not involve tedious manual labeling. We focus on schemes that rely on the knowledge of a set of acceptable pattern transformations. The training architecture then resembles a siamese network whose branches process different transformations of the same pattern. The SSL training objective must then balance two terms: on the one hand, the representations computed by each branch must be close or, at least, related; on the other hand, they should be prevented from collapsing partially \citep{jing2021understanding} or catastrophically \citep{chen2020simsiam}. Although this second term tends to fill the representation with useful features, what is necessary to balance the SSL training objective might still exclude potentially useful features for the target tasks.
This section presents results obtained using \textsc{swav} pre-training using 1.2 million \textsc{ImageNet} images \citep{caron2020unsupervised} and using \textsc{seer} pre-training using 1 billion \textsc{Instagram1B} images \citep{goyal2022vision}. These experiments leverage the pre-trained models made available by the authors: five \textsc{resnet50} (four from our reproduction), one \textsc{resnet50w2}, one \textsc{resnet50w4} and one \textsc{resnet50w5} for the \textsc{swav} experiments;\footnote{\url{https://github.com/facebookresearch/swav}} one \textsc{regnet32gf}, one \textsc{regnet64gf}, one \textsc{regnet128gf}, and one \textsc{regnet256gf} (1.3B parameters) for the \textsc{seer} experiments.\footnote{\url{https://github.com/facebookresearch/vissl/tree/main/projects/SEER}}
The first four columns of Figure~\ref{fig:ssl_tf} present linear probing results for four target object recognition tasks: supervised \textsc{ImageNet}, \textsc{Inaturalist18}, \textsc{Cifar100}, and \textsc{Cifar10}. The baseline curves (red, labeled \plotlabel{wide} or \plotlabel{wide\&deep}) plot the performance of linear classifiers trained on top of the pre-trained SSL representations. The solid \synthcat$n$ curves were obtained by training a linear classifier on top of the concatenated representations of the $n$ smallest SSL pre-trained representations (solid blue, \plotlabel{cat}). The dash \synthcat$n$ curves train an ensemble of $n$ small classifiers on subsets of the concatenated representation (dash blue, \plotlabel{cat$^{\sf sub}$}).\footnote{Likewise the supervised transfer learning experiments, each small classifier learns on the representation of a sub-network (e.g. \textsc{regnet32gf}, \textsc{regnet64gf}). Now the representation subset cannot be treated as random subsets of the concatenated representation anymore, because the model architectures are not always the same. So we omit the ensemble classifiers for red curves.} Overall, the \synthcat$n$ approach offers the best performance.
The last column of Figure~\ref{fig:ssl_tf} presents results with fine-tuning for the supervised \textsc{ImageNet} task. Our \emph{two-stage fine-tuning} approach (as Figure~\ref{fig:twostageft}) matches the performance of equivalently sized baseline networks. In particular, the largest \synthcat4 model using \textsc{seer} pre-training, with 2.3B parameters, achieves 85.5\% correct classification rate, approaching the 85.8\% rate of the largest baseline network in \textsc{seer} \citep{goyal2022vision}, \textsc{regnet10B} with 10B parameters. Of course, separately training and fine-tuning the components of the $\synthcat$4 network is far easier than training a single \textsc{regnet10B} network.
Additional results using \textsc{SimSiam} \citep{chen2020simple} and with distillation are provided in appendix \ref{apx:simsiam_cifar}. Other experiment details are provided in appendix \ref{apx:ssl}.
\section{Supervised transfer learning}
\label{sec:supervisedtransfer}
This section focuses on supervised transfer learning scenarios in which the representation learned using an auxiliary supervised task, such as the \textsc{ImageNet} object recognition task \citep{imagenet_cvpr09}, is then used for the target tasks, such as, for our purposes, the \textsc{Cifar10}, \textsc{Cifar100}, and \textsc{Inaturalist18} (\textsc{Inat18}) object recognition tasks \citep{krizhevsky2009learning,van2018inaturalist}. We distinguish the \emph{linear probing} scenario where the penultimate layer features of the pre-trained network are used as inputs for linear classifiers trained on the target tasks, and the \emph{fine tuning} scenario which uses back-propagation to further update the transferred features using the target task training data.
\begin{table}[t]
\par\vspace*{-2ex}
\caption{Supervised transfer learning from \textsc{ImageNet} to \textsc{Inat18}, \textsc{Cifar100}, and \textsc{Cifar10} using linear probing. The \textsc{erm} {(empirical risk minimization)} rows provide baseline results. The \synthcat$n$ rows use the concatenated representations of $n$ separately trained networks.
}
\par\vspace*{-1ex}
\label{tab:imagenet_sl_lineareval}
\bigskip
\centering
\setlength\tabcolsep{1.2pt}
\resizebox{0.488\textwidth}{!}{
\begin{tabular}{c cc |c| ccc }
\toprule
& & & {\small ID} & \multicolumn{3}{c}{\small Linear Probing (OOD)} \\
method & architecture & params & \textsc{\small imagenet} & \textsc{\small inat18} & \textsc{\small cifar100} & \textsc{\small cifar10} \\
\midrule
\textsc{erm} & \textsc{resnet50} & 23.5\textsc{m} & 75.58 & 37.91 & 73.23 & 90.57 \\
\textsc{erm} & \textsc{resnet50w2} & 93.9\textsc{m} & 77.58 & 37.34 & 72.65 & 90.86 \\
\textsc{erm} & \textsc{resnet50w4} & 375\textsc{m} & 78.46 & 38.71 & 74.81 & 92.13 \\
\midrule
\textsc{erm} & 2$\times$\textsc{resnet50} & 47\textsc{m} & 75.03 & 39.34 & 74.36 & 90.94 \\
\textsc{erm} & 4$\times$\textsc{resnet50} & 94\textsc{m} & 75.62 & 41.89 & 74.06 & 90.61 \\
\midrule
{\synthcat}2& 2$\times$\textsc{resnet50} & 47\textsc{m} & 77.57 & 43.26 & 76.10 & 91.86 \\
{\synthcat}4& 4$\times$\textsc{resnet50} & 94\textsc{m} & 78.15 & 46.55 & 78.19 & 93.09 \\
{\synthcat}10& 10$\times$\textsc{resnet50} & 235\textsc{m} & 78.36 & 49.65 & 79.61 & 93.75 \\
\bottomrule
\end{tabular}
}
\par\vspace*{-2ex}
\end{table}
\begin{figure*}[t]
\centering
\includegraphics[width=\textwidth]{plots/imagenet_sl_tf_v4.pdf}
\par\vspace*{-2ex}
\caption{Supervised transfer learning from \textsc{ImageNet} to \textsc{Inat18}, \textsc{Cifar100}, and \textsc{Cifar10}. The top row shows the superior linear probing performance of the \synthcat$n$ networks (blue, \plotlabel{cat}). The bottom row shows the performance of fine-tuned \synthcat$n$, which is poor with normal fine-tuning (gray, \plotlabel{[init]cat}) and excellent for two-stage fine tuning (blue, \plotlabel{[2ft]cat}). \synthdistill$n$ (pink, \plotlabel{distill}) representation is obtained by distilling \synthcat$n$ into one \textsc{resnet50} (we omit $\synthdistill$ in this section due to the space limit. see details in the appendix \ref{apx:imagenet_sl}).
}
\label{fig:imagenet_sl_ft_2ft}
\par\vspace*{-1ex}
\end{figure*}
\@startsection{paragraph}{4}{\z@}{.1ex plus 0ex minus .1ex}{-1em}{\normalsize\bf}{Linear probing}
The first three rows of Table~\ref{tab:imagenet_sl_lineareval}, labeled \textsc{erm}, provide baselines for the linear probing scenario, using respectively a \textsc{resnet50} network \citep{he-2016}, as well as larger variants \textsc{resnet50w}$n$ with $n$ times wider internal representations and roughly $n^2$ times more parameters. The following two rows provide additional baseline results using networks \textsc{$n\times$resnet50} composed of respectively $n$ separate \textsc{resnet50} networks joined by concatenating their penultimate layers. Although these networks perform relatively poorly on the pre-training task \textsc{ImageNet}, their linear probing performance is substantially better than that of the ordinary \textsc{resnet}s.
The final three rows of Table~\ref{tab:imagenet_sl_lineareval}, labeled \synthcat$n$, are obtained by training $n$ separate \textsc{resnet50} networks on \textsc{ImageNet} with different random seeds, and using their concatenated representations as inputs for a linear classifier trained on the target tasks. This approach yields linear probing performances that substantially exceed that of comparably sized baseline networks. Remarkably, \synthcat$n$, with separately trained components, outperforms the architecturally similar \textsc{$n\times$resnet50} trained as a single network. See appendix \ref{apx:imagenet_sl} for experimental details.
These results are succinctly\footnote{In order to save space, all further results in the main text of this contribution are presented with such plots, with result tables provided in the appendix.} represented in the top row of Figure~\ref{fig:imagenet_sl_ft_2ft}. For each target task \textsc{Inat18}, \textsc{Cifar100}, and \textsc{Cifar10}, the solid curves show the linear probing performance of the baseline \textsc{resnet50w}$n$ (red, labeled \plotlabel{wide}) and of the \synthcat$n$ networks (blue, \plotlabel{cat}) as a function of the number of parameters of their inference architecture.
The left plot (double height) of Figure~\ref{fig:imagenet_sl_ft_2ft} provides the same information in-distribution, that is, using the pre-training task as target task. In-distribution, the advantage of \synthcat$n$ vanishes when the networks become larger, possibly large enough to approach the conditions of Theorem~\ref{prop:opt}. The out-of-distribution curves (top row) are qualitatively different because they show improved performance all along.
An ensemble of $n$ \textsc{resnet50} networks is architecturally similar to the \synthcat$n$ models. Instead of training a linear classifier on the concatenated features, the ensemble averages $n$ classifiers
independently trained on top of each network. Whether this is beneficial depends on the nature of the target task and its training data (dashed blue, labeled \plotlabel{cat$^{\sf sub}$}). For completeness, we also present an ensemble baseline (dashed red plot, labeled \plotlabel{wide$^{\sf sub}$}) averaging $n$ linear classifiers trained on top of a random partition of the corresponding wide network features.
\@startsection{paragraph}{4}{\z@}{.1ex plus 0ex minus .1ex}{-1em}{\normalsize\bf}{Fine-tuning}
Having established, in the linear probing case, that transferring concatenated representations \synthcat$n$ outperforms transferring the representation of an equivalently sized network, we turn our attention to fine-tuning.
Fine-tuning is usually achieved by setting up a linear classifier on top of the transferred feature and training it on the target task data while allowing back-propagation to update the transferred features as well. The bottom row of Figure~\ref{fig:imagenet_sl_ft_2ft} shows the performance of this approach using the baseline network representations (red curve,
labeled \plotlabel{[init]wide}) and the concatenated representations (gray curve, labeled \plotlabel{[init]cat}), The latter perform very poorly.\footnote{The poor performance of plain fine-tuning had already been pointed out by \citet{kumar2022finetuning} and \citet{kirichenko2022last}.}
\begin{figure}[t]
\centering
\includegraphics[height=0.36\linewidth]{art/2stagefinetune.pdf}
\par\vspace{-2ex}
\caption{\emph{Two-stage fine-tuning} consists of fine-tuning each network separately, then concatenating their feature extractors, now frozen, and training a final classifier. }
\par\vspace*{-2ex}
\label{fig:twostageft}
\end{figure}
\begin{figure*}[ht]
\centering
\includegraphics[height=0.395\textwidth]{plots/ssl_tf_v3.pdf}~~~~~
\includegraphics[height=0.395\textwidth]{plots/ssl_tf_ft_v3.pdf}
\par\vspace*{-2ex}
\caption{Self-supervised transfer learning with \textsc{swav} trained on unlabeled \textsc{ImageNet(1k)} (\emph{top row}) and with \textsc{seer} on \textsc{Instagram1B} (\emph{bottom row}). The constructed rich representation, \synthcat$n$, yields the best linear probing performance (\plotlabel{cat} and \plotlabel{cat$^{\sf sub}$}) for supervised \textsc{ImageNet}, \textsc{Inat18}, \textsc{Cifar100}, and \textsc{Cifar10} target tasks. The two-stage fine-tuning (\plotlabel{[2ft]cat}) matches equivalently sized baseline models (\plotlabel{[init]wide} and \plotlabel{[init]wide\&deep}), but with much easier training. The sub-networks of \synthcat5 (and \synthcat2) in \textsc{swav} hold the same architecture. Due to the space limitation, we put other fine-tuning curves in appendix \ref{apx:swav_additional_exp}.}
\par\vspace*{-1.5ex}
\label{fig:ssl_tf}
\end{figure*}
\begin{figure}
\centering
\includegraphics[width=0.46\textwidth]{plots/vit_tf_v4.pdf}
\par\vspace*{-2ex}
\caption{{Supervised transfer learning from \textsc{ImageNet21k} to \textsc{ImageNet} on vision transformers.}}
\par\vspace*{-2ex}
\label{fig:imagenet_vit_sl_tf}
\end{figure}
We posit that fine-tuning with a single training episode impoverishes the initially rich representation. Instead, we propose \emph{two-stage fine tuning} which consists
of separately training $n$ networks on the pre-training task, separately fine-tuning them on the target task, and finally training a linear classifier on top of the concatenation of the $n$ separately fine-tuned representations (Figure~\ref{fig:twostageft}). The superior performance of two-stage fine-tuning is clear in the bottom row of Figure~\ref{fig:imagenet_sl_ft_2ft} (blue solid curve, labeled \plotlabel{{[2ft]}cat}). Ensembles of fine-tuned networks perform almost as well (blue dashed curve, labeled \plotlabel{[2ft]cat$^{\sf sub}$}).
\@startsection{paragraph}{4}{\z@}{.1ex plus 0ex minus .1ex}{-1em}{\normalsize\bf}{Vision transformers}
Figure~\ref{fig:imagenet_vit_sl_tf} shows that transformer networks behave similarly. We carried out supervised transfer experiments using the original vision transformer, \textsc{ViT}, \citep{dosovitskiy2020image}, and using a more advanced version using carefully crafted data augmentations and regularization, \textsc{ViT(augreg)}, \citep{steiner2021train}.
We use two transformers of two different sizes, {ViT-B/16} and {ViT-L/16}, pre-trained on \textsc{ImageNet21k}.\footnote{Checkpoints provided at \url{https://github.com/google-research/vision_transformer}.} Supervised transfer baselines (red, \plotlabel{wide\&deep} or \plotlabel{[init]wide\&deep}) are obtained by linear-probing and by fine-tuning on \textsc{ImageNet(1k)}. These baselines are outperformed by respectively linear-probing and \emph{two-stage fine tuning} on top of the concatenation of their final representations (\synthcat2).
An even larger transformer architecture, {ViT-H/14}, yields about the same \textsc{ImageNet1k} fine-tuning performance as {ViT-L/16}, but lags 1\% behind \synthcat2, despite having twice as many parameters \cite{dosovitskiy2020image}.
Experiments with two-stage fine-tuned \synthcat2 in \textsc{ViT(augreg)} show even better results, possibly because changing the random seed does not just changes the initial weights and the mini-batch composition, but also affects the data augmentations of the \textsc{ViT(augreg)} networks.
|
\section{Introduction}
Increasing levels of sophistication are required to build large-scale, data-driven models of complicated, nonlinear dynamic processes. Koopman operator theory provides an alluring, though nuanced, approach to building such models \cite{rowley2009spectral, schmid2010dynamic, mezic2013analysis,proctor2018generalizing,hasnaindata,hasnain2022learning,mauroy2020Koopman,budivsic2012applied, williams2014kernel,johnson2022SILL}.
To build universally applicable models, as the scale of systems increase, model classes need to capture as many dynamics as possible in the smallest number of parameters and model dimensions. This requirement has restricted the viability of dictionary-based Koopman models such as the extended dynamic mode decomposition (EDMD \cite{williams2015data}) algorithm. Such models become unwieldy because human defined dictionaries with nonlinearities placed in an evenly spaced grid scale, in dimension, exponentially with the dimension of the measurement space. However, Koopman models, in theory, do not need their parameters to scale exponentially with the dimension of the measurements.
Model dictionaries that satisfy {\em uniform finite approximate closure}, as defined in \cite{johnson2022SILL}, have been show to scale more favorably \cite{johnson2018class}. We have also seen empirical results that suggest deep-learning based models scale much better with measurement dimension as well\cite{yeung2019learning, wehmeyer2018time, lusch2018deep, otto2019linearly,takeishi2017learning}.
Traditionally, dictionary based methods for building Koopman models focus on homogeneous dictionaries. In such a dictionary, each of the nonlinear candidate dictionary functions comes from a specific function basis, such as a polynomial or Fourier basis \cite{williams2016extending}. The SILL dictionary from Part 1 of this paper is one such homogeneous dictionary. Each nonlinear function in the basis is a conjunctive logistic function.
Part 1 of this paper demonstrates the {\em uniform finite approximate closure} of the SILL dictionary, which was inspired by some of the dictionary functions learned by deepDMD, the deep-learning based Koopman model in \cite{yeung2019learning}. To an extent, this analysis explains the success of deepDMD in learning. But, deepDMD learns more than this one type of function (see Fig. 3 of \cite{yeung2019learning}). Indeed, it can learn dictionary functions from multiple function classes simultaneously, and we have observed it to do so (see Fig. \ref{fig:DDMD_0bsTogg}).
Inspired by these observations, we contribute the first analysis of a heterogeneous, or mixed, basis dictionary function for Koopman learning. While mixing elements from different functional basis does increase the representational efficiency of the dictionary, the subspace invariance, or closure, properties of such a dictionary are not straightforward to understand.
Introducing heterogeneous basis leads to added complexity of the resulting dictionaries. This complexity is transferred to the representation of the vector field over which we take Lie-derivatives of each dictionary function (see the proof of Lemma 1 in Section \ref{sec:AppendixProofs} of the appendix). We evaluate the vector field of the new class of Koopman models by computing the Lie-derivatives of their heterogeneous mixture of dictionary functions. We then work from these Lie-derivatives to show that these Koopman models are subspace invariant.
We also follow the pattern we see from deep learning by mixing classes of dictionary functions to build low-dimensional models. The mixed function bases we study give further insight into the strategy adopted by deepDMD to solve the data-driven Koopman learning problem. Combining our new function dictionaries with an effective learning algorithm yields a training loss comparable to deepDMD, but with an order of magnitude reduction in model complexity.
\section{The Koopman Generator Learning Problem}\label{sec:KLP}
The learning problem that we address is explained in full detail in Section 2 of \cite{johnson2022SILL}. In this problem, $r$ $m-$dimensional measurements, $y$, of an $n-$dimensional state, $x$, evolving in time according to an analytic dynamic system, $\dot x=f(x)$, are used to solve the following optimization problem,
\begin{equation}\label{eq:objective}
\min_{K, \psi} \sum_{i=1}^{r} \left\Vert \frac{d\psi(y(x_i))}{dt} - K \psi(y(x_i)) \right\Vert,
\end{equation}
where $K\in\mathbb{R}^{N\times N}$ and $\psi$ is from the set of all differentiable functions from $\mathbb{R}^m$ to $\mathbb{R}^N$, where $N$, the dimension of our approximate Koopman operator, is a natural number greater than $m$, the dimension of our measurements.
In EDMD the dictionary functions $\psi(y)$ are predefined, drawn from a {\it homogeneous} class of nonlinear functions, and assumed to be known. Under these assumptions, Eq. (\ref{eq:objective}) is a convex optimization problem with a closed-form solution.
By contrast, in deepDMD, the dictionary functions and $K$ are learned simultaneously during iterative training. This is, of course, a nonlinear, non-convex optimization problem, for which we employ variants of the stochastic gradient descent algorithm from Tensorflow or Pytorch, such as adaptive gradient descent (AdaGrad, \cite{duchi2011adaptive}) or adaptive momentum (ADAM, \cite{kingma2014adam}). These numerical approaches provide no guarantee that the learned set of dictionary functions $\psi_1(y), \psi_2(y), ...,\psi_{N}(y)$ are homogeneous, or drawn from the same function class. The outcomes of deepDMD often produce heterogeneous dictionaries $\psi(y)$.
\begin{figure*}[ht]
\centering
\includegraphics[width=\linewidth]{DeepDMD2Dict.png}
\caption{Process for the discovery of a novel mixed function dictionary with approximate subspace invariance. First, the deepDMD algorithm takes data from a nonlinear system to build approximate Koopman observables. Then, projections of these observables are analyzed for these functional properties. Finally, high-dimensional functions, whose projections satisfy these properties, are developed and their closure properties are verified.}
\label{fig:DDMD_0bsTogg}
\end{figure*}
\begin{remark}
In this paper, we use the definitions of {\em subspace invariance}, {\em finite closure} and {\em uniform finite approximate closure} in Section 2 of \cite{johnson2022SILL}.
\end{remark}
\begin{remark}
We also follow their notation conventions explained in Section 3 of \cite{johnson2022SILL}. These are restated in Section \ref{sec:AppendixNotation} of the Appendix.
\end{remark}
\section{AugSILL: A Heterogeneous Dictionary Model of DeepDMD's Learned Dictionary}\label{sec:augSILL}
In this section, we define a heterogeneous dictionary for Koopman learning. This dictionary is, like the SILL dictionary in \cite{johnson2022SILL}, inspired from the dictionary functions learned by deepDMD. The nonlinear SILL functions are of the form:
\begin{equation}\begin{aligned}\Lambda(y;\theta_k)\triangleq\Lambda(y;\mu_k, \alpha_k)\triangleq \prod_{i=1}^{m} \lambda(y_i;\theta_{ki})\end{aligned}\end{equation}
where,
\begin{equation}\begin{aligned}\lambda(y_i;\theta_{ki}) \triangleq \lambda(y_i;\mu_{ki}, \alpha_{ki}) \triangleq \frac{1}{1+e^{-\alpha_{ki}(y_i-\mu_{ki})}},\end{aligned}\end{equation}
where $\mu_{ki}$ is a steepness parameter in the $i^{th}$ measurement dimension and $\alpha_{ki}$ is a center parameter in the $i^{th}$ measurement dimension.
In our experience, projections of deepDMD's dictionary functions tended to match the profiles of logistic functions, however there were some punctuated irregularities, divots and bumps not easy to represent as a logistic function, in these projections (See Fig. \ref{fig:DDMD_0bsTogg}). We choose to model these irregularities with radial basis functions (RBFs). We therefore augment our SILL dictionary with conjunctive multivariate RBFs (conjunctive RBFs). We define an RBF with a steepness of $\alpha_{ki}$ and a center of $\mu_{ki}$ as follows: \begin{equation}\begin{aligned}\rho(y_i; \theta_{ki}) \triangleq \rho(y_i; \mu_{ki}, \alpha_{ki}) \triangleq \frac{e^{-\alpha_{ki} (y_i-\mu_{ki})}}{(1+e^{-\alpha_{ki}(y_i-\mu_{ki})})^2}.\end{aligned}\end{equation}
This RBF takes on its global maximum value of $\frac{1}{4}$ when the measurement $y_i=\mu_{ki}$ the center parameter. It radially approaches zero at a rate determined by the value of the steepness parameter $\alpha_{ki}$.
Note the following relationship between our RBF and logistic functions. The RBF $\rho(y_i;\theta_{ki}) = \lambda(y_i;\theta_{ki}) - \lambda(y_i;\theta_{ki})^2$. So, even in one dimension, exactly approximating an RBF would require an infinite linear combination of SILL functions (a piecewise linear spline on infinitesimally small intervals). Thus, conjunctive RBFs contribute distinct nonlinear features that are outside the span of the SILL function space. This mathematical observation is the rationale for referring to this mixture of dictionary functions as heterogeneous.
We define a conjunctive RBF to be $P:\mathbb{R}^m\rightarrow\mathbb{R}$ so that: \begin{equation}\begin{aligned} P(y;\theta_k)\triangleq P(y;\mu_k, \alpha_k)\triangleq\prod_{i=1}^m \rho(y_i;\theta_{ki}).\end{aligned}\end{equation}
Conjunctive RBFs map their center to a value of $4^{-m}$ and radially around that center drop off to zero. The steepness parameters determine how quickly the drop off to zero occurs along each coordinate axis.
When we set the vector $y$ to be constant in all but the $l^{th}$ dimension, $P(y;\mu_j,\alpha_j) = (\prod_{i\neq l}c_i)\rho(y_l;\mu_{jl}, \alpha_{jl})$. This is a constant times the RBF in the $l^{th}$ dimension. When we project dictionary functions learned by deepDMD to a single dimension, we observe that many contours approximate logistic functions, others approximate RBFs and scaled sums of RBFs and logistic functions. Fig. 3 of \cite{yeung2019learning} demonstrates some of these projected functions.
We propose the \textit{augSILL} (augmented SILL) dictionary as the stacked vector of a mixture of dictionary functions of $y$ \begin{equation}\begin{aligned} \psi(y) \triangleq \begin{bmatrix}
1&y^T&\bar \Lambda^T&\bar P^T
\end{bmatrix}^T, \end{aligned}\end{equation} where the vector \begin{equation}\begin{aligned}\bar\Lambda \triangleq [\Lambda(y;\theta_1), ..., \Lambda(y;\theta_{N_L})]^T\end{aligned}\end{equation} contains all conjunctive logistic functions, and \begin{equation}\begin{aligned}\bar P \triangleq [P(y;\theta_{N_L+1}), ..., P(y;\theta_{N_L+N_R})]^T\end{aligned}\end{equation} contains all conjunctive radial basis functions. In this article, the augSILL dictionary includes $N_L$ conjunctive logistic functions and $N_R$ conjunctive RBFs. The augSILL dictionary is of dimension $N=1+m+N_L+N_R$. To our knowledge this is the first time that anyone has analyzed the behavior of mixed dictionary functions for the numerical approximation of a Koopman operator or Koopman generator.
Can a mixture of function basis form a coherent basis that preserves subspace invariance, or at least uniform finite approximate closure? Certainly the more varied basis facilitates approximating the vector field, $F(y)$, due to the increased degrees of freedom in the dictionary. But the subspace invariance properties are just as important to Koopman models as approximating the vector field (see Example 1 of \cite{johnson2022SILL}). Using a mixed basis dictionary complicates the computation of the Lie derivatives of dictionary elements. This is because the representation of the vector field itself (a component in Lie derivative computation) has more degrees of freedom.
\begin{figure*}[ht]
\centering
\includegraphics[width=450pt]{augSILLProds.png}
\caption{ Visual representation of Theorem 1 of \cite{johnson2022SILL}, and Theorems \ref{thm:augSILLconv1}, \ref{thm:augSILLconv2} and \ref{thm:augSILLconv4}.
\textbf{(a)} Theorem 1 of \cite{johnson2022SILL}: The product of two conjunctive logistic functions approximates the conjunctive logistic function whose centers are greater in each dimension.
\textbf{(b)} Theorem \ref{thm:augSILLconv4}: The product of two conjunctive RBFs is nearly zero unless the norm of the difference between their centers is very small.
\textbf{(c1)} Theorem \ref{thm:augSILLconv1}: The product of a conjunctive logistic function and a conjunctive RBF approximates the RBF if at least one dimension of the center in the logistic function is greater than its partner center in the RBF. \textbf{(c2)} Theorem \ref{thm:augSILLconv2}: When this is not the case the product is nearly zero.
}
\label{fig:rbftsig}
\end{figure*}
\section{Uniform Finite Approximate Closure of the AugSILL Dictionary}\label{sec:AugSILLclosure}
In this section we analyze the subspace invariance properties of a mixed dictionary, the augSILL dictionary. To do so, we characterize the error term in Eq. (18) in \cite{johnson2022SILL}. In this section we show that the augSILL dictionary satisfies uniform finite approximate closure.
In Sections \ref{sec:convThms1} and \ref{sec:convThms3} we establish theorems analogous to Theorem 1 in \cite{johnson2022SILL} and its corollaries for the augSILL basis. In Section \ref{sec:augSILL_error} we demonstrate probabilistic results that combine with the corollaries in Section \ref{sec:convThms3} to uniformly bind average error of Eq. (18) in \cite{johnson2022SILL} with a constant, $B$, that can be arbitrarily small. This approach applies to the SILL and augSILL dictionaries to show uniform finite approximate closure.
\subsection{Convergence in Steepness of Bilinear AugSILL Terms to the Span of the AugSILL Dictionary}\label{sec:convThms1}
When we construct a mixed dictionary and compute its Lie derivatives we find mixed bilinear terms involving both types of nonlinear dictionary functions. On the path to demonstrating uniform finite approximate closure, we need to approximate these bilinear terms with a single element of our dictionary.
In Section 5 of \cite{johnson2022SILL} we demonstrated that the product of two conjunctive logistic functions can be well approximated by a single conjunctive logistic function. Here we approximate the product of a conjunctive logistic and RBF. To make this approximation, we first define a decision function, $H$, as follows:
\begin{equation}
H(y;\theta_l, \theta_k) \triangleq \begin{cases}
P(y;\theta_k)\mbox{ if } \mu_{li} < \mu_{ki}, \mbox{ for } i=1,2,...,m \\
0 \mbox{ otherwise,}
\end{cases}
\end{equation} where $l\in\{1,2,...,N_L\}$ and $k\in\{N_L+1, N_L+2,...,N_L+N_R\}$.
The decision function, $H$, uses the relative centers of $\Lambda$ and $P$, $\mu_l$ and $\mu_k$, to choose to take on the value of the conjunctive RBF ($P$) or zero. Both functions are in the augSILL dictionary as the zero function is trivially available.
To approximate the product of a conjunctive logistic and RBF, (see Fig. \ref{fig:rbftsig}\textbf{c1-c2}) we intend to show that: \begin{equation}\label{eq:approx2}
\Lambda(y;\theta_l)P(y;\theta_k) \approx H(y;\theta_l, \theta_k).
\end{equation}
Eq. (\ref{eq:approx2}) is provably quite a poor approximation under the wrong conditions, for example, when elements of $\alpha_l$ are negative. However, the following theorems will demonstrate, that when each element of $\alpha$ is positive, this is a reasonable approximation. This is argued by showing that when the elements of $\alpha$ approach infinity, the error of the approximation approaches zero exponentially for all but a set of points that lie on $m (N_L+N_R)$, $m-1$ dimensional hyperplanes in $\mathbb{R}^m$. These are the points where the measurements exactly match up with the function centers. As we noted in Part 1, this collection of hyperplanes has Lebesgue measure zero in $\mathbb{R}^m$.
In the case where the center of $\Lambda(y;\theta_l)$ is less than the center of $P(y;\theta_k)$ in all dimensions we have that:
\begin{equation}\label{eq:P_errorterm1}
\begin{aligned}
\varepsilon_{\Lambda_lP_k}(y)\triangleq\Lambda(y;\theta_l) & P(y;\theta_k) - H(y;\theta_l,\theta_k) \\
&= (\Lambda(y;\theta_l) - 1)P(y;\theta_k). \\
\end{aligned}
\end{equation}
Theorem \ref{thm:augSILLconv1} demonstrates that Eq. (\ref{eq:P_errorterm1}) goes to zero exponentially in the limit of the steepness parameters, $\alpha$. This theorem's purpose is to argue that the decision function, $H$, approximates that product of a conjunctive logistic and RBF when the center parameters of the RBF are more positive than those of the logistic in each dimension. Since $H$, under these circumstances, decides to approximate this product with the conjunctive RBF to begin with, the approximation is already present in the augSILL dictionary. This fact will allow us to argue the subspace invariance of this dictionary further down the line.
Given an augSILL dictionary, $\psi(y)$, we define a set of $m(N_L+N_R)$, $m-1$ dimensional hyperplanes in $R^m$ corresponding to the centers of each
$M_{\bar\Lambda,\bar P} \triangleq \{ y:\mbox{there exists } i\in\{1,2,...,m\} \mbox{ and } j\in \{1,2,...,N_L+N_R\} \mbox{ so that } y_i=\mu_{ji} \}.$ This defines specific measurements where the approximations in Eq. (\ref{eq:approx2}) and (\ref{eq:approx3}) have maximal error.
\begin{theorem}\label{thm:augSILLconv1}
When the measurements do not exactly align with the centers of the dictionary functions in any dimension and the center of the conjunctive RBF is more positive than the center of the conjunctive logistic function in each dimension, then their product exponentially converges to the conjunctive RBF as their steepness parameters go to infinity. Specifically, if $y\not\in M_{\bar\Lambda,\bar P}$ and for all $i=1, 2, ..., m$, $\mu_{ki} \geq \mu_{li}$, then as $\alpha\rightarrow\infty$, $\Lambda(y;\theta_l) P(y;\theta_k) - H(y;\theta_l,\theta_k)\rightarrow 0$ exponentially.
\end{theorem}
\begin{proof}
As a preliminary, we note that $-1\leq (\Lambda(y;\theta_l) - 1)\leq 0$ for any $y\in\mathbb{R}^m$ and for $\alpha>0$.
We first show that when $\mu_{ki} \geq \mu_{li}$ $P(y;\theta_k)\rightarrow 0$ as $\alpha\rightarrow \infty$. We do so by considering the two possible cases for any $i$ where $\mu_{ki} \geq \mu_{li}$.
\textbf{Case 1}, $y_i>\mu_{ki}$. In this case as $\alpha\rightarrow\infty$ we have that $e^{-\alpha_{ki}(y_i-\mu_{ki})}\rightarrow 0$ exponentially, and that $(1+e^{-\alpha_{ki}(y_i-\mu_{ki})})^2 \rightarrow 1^2=1$. Thus the $i^{th}$ term of $P(y;\theta_k)$ will go to $0$ exponentially.
\textbf{Case 2}, $y_i<\mu_{ki}$. In this case as $\alpha\rightarrow\infty$ we have that $\frac{e^{-\alpha_{ki}(y_i-\mu_{ki})}}{(1+e^{-\alpha_{ki}(y_i-\mu_{ki})})^2}\rightarrow 0$. Thus the $i^{th}$ term of $P(y;\theta_k)$ will go to $0$ exponentially.
Each of these cases implies that $P(y;\theta_k)$ and therefore Eq. (\ref{eq:P_errorterm1}) goes to $0$ exponentially.
$\blacksquare$\end{proof}
Theorem \ref{thm:augSILLconv1} implies that the bilinear augSILL term, $\Lambda P$, a product of two dictionary functions, is well approximated by $P$, a single dictionary function when $P$'s center parameters are more positive than $\Lambda$'s. This fact will be key to showing that bilinear terms in the Lie derivatives of augSILL dictionary functions are approximated as linear terms in Section \ref{sec:convThms3}.
Theorem \ref{thm:augSILLconv2} implies that $\Lambda P$ is well approximated by $0$, when at least one of $P$'s center parameters are more negative than $\Lambda$'s. This, of course, is assuming that all the steepness parameters of these dictionary functions are positive, since the approximation is evaluated in the limit of high steepness.
When the center of $\Lambda(y;\theta_l)$ is greater than the center of $P(y;\theta_k)$ in one or more dimensions we have that:
\begin{equation}\label{eq:P_errorterm2}
\begin{aligned}
\varepsilon_{\Lambda_lP_k}(y)\triangleq\Lambda(y;\theta_l) & P(y;\theta_k) - H(y;\theta_l, \theta_k) \\
&= \Lambda(y;\theta_l)P(y;\theta_k) \\
\end{aligned}
\end{equation}
Theorem \ref{thm:augSILLconv2} demonstrates that the approximation in Eq. (\ref{eq:P_errorterm2}) is a good approximation in the limit of $\alpha$.
\begin{theorem}\label{thm:augSILLconv2}
When the measurements do not exactly align with the centers of the dictionary functions in any dimension and the center of the conjunctive logistic function is more positive than the center of the RBF in all dimensions, then their product exponentially converges to zero as their steepness parameters go to infinity. Specifically, if $y\not\in M_{\bar\Lambda,\bar P}$ and $y_i\neq\mu_{li} $ for all $ i\in \{1, 2, ..., m\}$ so that $\mu_{ki} < \mu_{li}$, then as $\alpha\rightarrow\infty$, \[\Lambda(y;\theta_l) P(y;\theta_k) - H(y;\theta_l, \theta_k) \rightarrow 0\] exponentially.
\end{theorem}
\begin{proof}
We assume that $\mu_{ki} < \mu_{li}$ and note that, in each case as $\alpha\rightarrow \infty$ the $i^{th}$ term in Eq. (\ref{eq:P_errorterm2}) goes to zero as shown in Case 2 of the proof of Theorem \ref{thm:augSILLconv1}. Thus the product of these terms will go to zero as $\alpha\rightarrow\infty$. For the sake of brevity we forgo the explicit computation of the limits which follow and speak in more general terms of rate of growth. We note that when we use ``$\infty_c$'' we mean to say that the term grows to infinity with $\alpha$ with a rate of $e^{c\alpha}$, for some positive constant $c$. Furthermore, we use ``$0_c$'' to mean that the term goes to zero with a rate of $e^{-c\alpha}$, for some positive constant $c$.
\textbf{Case 1}, $y_i > \mu_{ki} $ \textit{Sub-Case 1.1}, $y_i > \mu_{li}$ so as $\alpha\rightarrow\infty$ our $i^{th}$ term goes to $\frac{0_{c_k}}{1}=0$.
\textit{Sub-Case 1.2}, $y_i = \mu_{li}$ so as $\alpha\rightarrow\infty$ our $i^{th}$ term goes to $\frac{0_{c_k}}{2}=0$.
\textit{Sub-Case 1.3}, $y_i < \mu_{li}$ so as $\alpha\rightarrow\infty$ our $i^{th}$ term goes to $\frac{0_{c_k}}{\infty_{c_l}}=0$.
\textbf{Case 2}, $y_i < \mu_{ki}$, thus $x_i<\mu_{li}$ and so as $\alpha\rightarrow\infty$ our $i^{th}$ term goes to $\frac{\infty_{c_k}}{\infty_{c_k}^2\infty_{c_l}}=0$.
\textit{Sub-Case 2.1}, $y_j > \mu_{li}$ so as $\alpha\rightarrow\infty$ our $i^{th}$ term goes to $\frac{\infty_{c_k}}{\infty_{c_k}^2}=0$.
\textit{Sub-Case 2.2}, $y_i = \mu_{li}$ so as $\alpha\rightarrow\infty$ our $i^{th}$ term goes to $\frac{\infty_{c_k}}{2\infty_{c_k}^2}=0$.
\textit{Sub-Case 2.3}, $y_i < \mu_{li}$ so as $\alpha\rightarrow\infty$ our $i^{th}$ term goes to $\frac{\infty_{c_k}}{\infty_{c_k}^2\infty_{c_l}}=0$.
Now we note that, in each case as $\alpha\rightarrow \infty$ the $i^{th}$ term in Eq. (\ref{eq:P_errorterm2}) goes to zero exponentially. Thus the product of these terms will go to zero exponentially as $\alpha\rightarrow\infty$.
$\blacksquare$\end{proof}
Between Theorems \ref{thm:augSILLconv1} and \ref{thm:augSILLconv2} we have that $\Lambda P \approx H$. Specifically, $\Lambda(y;\theta_l) P(y;\theta_k) \approx P(y;\theta_k)$ when $\theta_l\lesssim\theta_k$ and $0$ otherwise. The only pathological case excluded from these theorems is when the conjunctive logistic and conjunctive RBF centers exactly match in some dimension. Distance from the pathology becomes relevant when steepness parameters, $\alpha$, are too small.
Now, we approximate the product of two conjunctive RBFs, completing all the possible combinations of pairwise products between elements of our mixed basis. The product of two conjunctive logistic functions, $\Lambda_l \Lambda_k$, is covered in Section 5 of \cite{johnson2022SILL}. To approximate $P_l P_k$ (see Fig. \ref{fig:rbftsig}\textbf{b}), Theorem \ref{thm:augSILLconv4} describes conditions under which
\begin{equation}\label{eq:approx3}
\varepsilon_{P_lP_k}(y)\triangleq P(y;\theta_l)P(y;\theta_k) \approx 0.
\end{equation}
Because Theorem \ref{thm:augSILLconv4} shows that the approximation in Eq. (\ref{eq:approx3}) is a good approximation in the limit of $\alpha$. It takes the bilinear term, $P_l P_k$, and approximates it with the zero function which is trivially in all augSILL dictionaries.
\begin{theorem}\label{thm:augSILLconv4}
When the measurements do not exactly align with the centers of the dictionary functions in any dimension, then the product of two conjunctive RBFs converges exponentially to zero as their steepness increases.
Specifically, if $y\not\in M_{\bar\Lambda,\bar P}$, then as $\alpha\rightarrow\infty$, $P(y;\theta_l)P(y;\theta_k)\rightarrow 0$ exponentially.
\end{theorem}
The proof of Theorem \ref{thm:augSILLconv4}, as well as a supporting lemma is presented in Section \ref{sec:AppendixProofs} of the appendix.
We now have four approximation theorems for the bilinear mathematical terms which arise when computing the Lie derivatives of augSILL basis functions. We now apply them to these Lie derivatives with a set of approximation corollaries. These corollaries, together with the argument in Section \ref{sec:augSILL_error}, will demonstrate the uniform finite approximate closure of the augSILL dictionary.
\subsection{Showing bilinear Lie derivatives can be approximated linearly to satisfy the Koopman generator equation}\label{sec:convThms3}
Corollary \ref{cor:logApprox} approximates the Lie derivative of a conjunctive logistic function in the context of our mixed basis. The approximation that this corollary suggests is not a Koopman model itself (see Fig. 1 in \cite{johnson2022SILL}). In Section \ref{sec:augSILL_error} we demonstrate uniform finite approximate closure by approximating this intermediate approximation with a full Koopman model.
\begin{corollary}\label{cor:logApprox}
Under the assumptions of Theorem 1 in \cite{johnson2022SILL} and the assumption that the augSILL observables span
the vector field we seek to model, the Lie derivative of a conjunctive logistic function exponentially approaches a nonlinear combination of augSILL functions, specifically,
\begin{equation}\begin{aligned}\label{eq:augSILL_limApproxLog}
\dot{\Lambda}(y&;\theta_l) \rightarrow \sum_{i=1}^{m}\sum_{j=1}^{N_L} \alpha_{li}w_{ij}(1 - \lambda(y_i;\theta_{li})) \Lambda(y;\theta^*) \\& +\sum_{i=1}^{m}\sum_{k=N_L+1}^{N_L+N_R} \alpha_{li}w_{ik}(1 - \lambda(y_i;\theta_{li})) H(y;\theta_{l},\theta_{k})
\end{aligned}\end{equation} exponentially as $\alpha\rightarrow\infty$.
\end{corollary}
Now that we have an intermediate approximation for the Lie derivative of a conjunctive logistic function we need a similar result for conjunctive RBFs. Once we have these results the analysis in Section \ref{sec:augSILL_error} can show a Koopman approximation of the derivatives of these two augSILL functions. This final approximation will be a linear combination of augSILL dictionary functions.
Corollary \ref{cor:rbfApprox} approximates the Lie derivative of a conjunctive RBF in the context of the augSILL basis.
\begin{corollary}\label{cor:rbfApprox}
Under the assumptions of Theorem 1 in \cite{johnson2022SILL} and the assumption that the augSILL observables span
the vector field we seek to model, the Lie derivative of a conjunctive RBF exponentially approaches a nonlinear combination of conjunctive RBFs, specifically,
\begin{equation}\begin{aligned} \label{eq:augSILL_limApproxRbf}
\dot{P}(y;\theta_l)& \rightarrow \sum_{i=1}^{m}\sum_{j=1}^{N_L} \alpha_{li}w_{ij}(1 - 2\lambda(y_i;\theta_{li})) H(y;\theta_{j},\theta_{l})
\end{aligned}\end{equation} exponentially as $\alpha\rightarrow \infty$.
\end{corollary}
Note that Equations (\ref{eq:augSILL_limApproxLog}) and (\ref{eq:augSILL_limApproxRbf}) are not compatible with the Koopman model we seek to learn: $K\psi(y),$ where the dictionary functions, $\psi$, are augSILL functions and $K$ is a real-valued matrix. In Section \ref{sec:augSILL_error} we approximate Equations (\ref{eq:augSILL_limApproxLog}) and (\ref{eq:augSILL_limApproxRbf}) with a mathematical form that is consistent with the approximated Koopman generator, Eq. (7) in \cite{johnson2022SILL}.
One can approximate the Lie derivatives of an augSILL dictionary function as a linear combination of {\em products of pairs} of augSILL functions. The details on approximating these Lie derivatives as a weighted sum of these bilinear terms is detailed in Section \ref{sec:augSILL_error}.
Below, we approximate this weighted sum of bilinear terms (one of the intermediate approximations in Fig. 1 in \cite{johnson2022SILL}) with a weighted sum of augSILL functions. The final weighted sum is of the form of Eq. (7) in \cite{johnson2022SILL}, and therefore admits a Koopman operator model.
\begin{corollary}\label{cor:logH}
Under the assumptions of Theorem 1 in \cite{johnson2022SILL}, the sum of products
\begin{equation}\begin{aligned}\label{eq:cor3pre}
&\sum_{i=1}^{m}\sum_{j=1}^{N_L} \alpha_{li}w_{ij} \Lambda(y;\theta_{l})\Lambda(y;\theta_{j}) \\& +\sum_{i=1}^{m}\sum_{k=N_L+1}^{N_L+N_R} \alpha_{li}w_{ik} \Lambda(y;\theta_{l})P(y;\theta_{k}) \\
\end{aligned}\end{equation}
approaches
\begin{equation}\begin{aligned}\label{eq:cor3post}
& \sum_{i=1}^{m}\sum_{j=1}^{N_L} \alpha_{li}w_{ij} \Lambda(y;\theta^*) \\& +\sum_{i=1}^{m}\sum_{k=N_L+1}^{N_L+N_R} \alpha_{li}w_{ik} H(y;\theta_{l},\theta_{k}),
\end{aligned}\end{equation} a weighted sum of augSILL functions, exponentially as $\alpha\rightarrow\infty$.
\end{corollary}
\begin{proof}
This result is a direct consequence of Theorem 1 of \cite{johnson2022SILL}, Theorem \ref{thm:augSILLconv1} and Theorem \ref{thm:augSILLconv2}. These approximations are outlined in Fig. \ref{fig:rbftsig} \textbf{(a)}, \textbf{(c1)} and \textbf{(c2)}.
$\blacksquare$\end{proof}
\begin{corollary}\label{cor:Hrbf}
Under the assumptions of Theorem 1 in \cite{johnson2022SILL}, the sum of products
\begin{equation}\begin{aligned}\label{eq:cor4pre}
&\sum_{i=1}^{m}\sum_{j=1}^{N_L} \alpha_{li}w_{ij} P(y;\theta_{l})\Lambda(y;\theta_{j}) \\& +\sum_{i=1}^{m}\sum_{k=N_L+1}^{N_L+N_R} \alpha_{li}w_{ik} P(y;\theta_{l})P(y;\theta_{k}) \\
\end{aligned}\end{equation}
approaches
\begin{equation}\begin{aligned}\label{eq:cor4post}
& \sum_{i=1}^{m}\sum_{j=1}^{N_L} \alpha_{li}w_{ij} H(y;\theta_{j},\theta_{l}),
\end{aligned}\end{equation} a weighted sum of conjunctive RBFs, exponentially as $\alpha\rightarrow\infty$.
\end{corollary}
\begin{proof}
This result is a direct consequence of Theorems \ref{thm:augSILLconv1}, \ref{thm:augSILLconv2} and \ref{thm:augSILLconv4}. These approximations are outlined in Fig. \ref{fig:rbftsig} \textbf{(c1)}, \textbf{(c2)} and \textbf{(b)}.
$\blacksquare$\end{proof}
The resulting linear combinations from Corollaries \ref{cor:logH} and \ref{cor:Hrbf} can be stacked and combined into the matrix $K$ (whose entries would be the products of $\alpha_{**}$ and $w_{**})$. Thus, given an augSILL dictionary that admits a weighted bilinear approximation to the Lie derivatives of its functions, Corollaries \ref{cor:logH} and \ref{cor:Hrbf} show the uniform finite approximate closure of that dictionary by showing that the error of approximating the weighted bilinear representation with a linear combination of dictionary functions goes to zero exponentially in steepness.
Corollaries \ref{cor:logApprox}, \ref{cor:rbfApprox}, \ref{cor:logH} and \ref{cor:Hrbf} each imply an error bound for approximating the Lie derivative of an augSILL dictionary function with an intermediate approximation. These error bounds are listed explicitly in Table \ref{tab:ovarallSummary}. They combine with other error bounds in the same table to demonstrate uniform finite approximate closure following the argument given in Section 6.2 of \cite{johnson2022SILL}.
\begin{table*}[ht]
\centering
\begin{tabular}{|p{6mm}|p{8mm}|p{50mm}|p{62mm}|p{28.5mm}|}
\hline
\textbf{Fun.} & \textbf{Ref.} & \textbf{Approximation} & \textbf{Difference (Error)} & \textbf{Error Bound} \\
\hline
Log. & (\ref{eq:augSILL_limApproxLog}) & \(\begin{aligned} &\sum_{i=1}^{m}\sum_{j=1}^{N_L} \alpha_{li}w_{ij} \Lambda(y;\theta^*) \\ &+\sum_{i=1}^{m}\sum_{k=N_L+1}^{N_L+N_R} \alpha_{li}w_{ik} H(y;\theta_{l},\theta_{k}) \end{aligned}\) & \(\begin{aligned} &\sum_{i=1}^{m}\sum_{j=1}^{N_L} \alpha_{li}w_{ij} \lambda(y_i;\theta_{li}) \Lambda(y;\theta^*) \\& +\sum_{i=1}^{m}\sum_{k=N_L+1}^{N_L+N_R} \alpha_{li}w_{ik} \lambda(y_i;\theta_{li}) H(y;\theta_{l},\theta_{k}) \end{aligned}\) & \(\begin{aligned}&\sum_{i=1}^{m}\sum_{j=1}^{N_L}\frac{\nu_{ij}}{2^{m+1}} \\&+ \sum_{i=1}^{m} \sum_{k=N_L+1}^{N_L+N_R} \frac{\nu_{ik}}{2^{3m+1}}\end{aligned}\) \\
\hline
Log. & (\ref{eq:augSILL_logPrime}) & \(\begin{aligned} &\sum_{i=1}^{m}\sum_{j=1}^{N_L} \alpha_{li}w_{ij} \Lambda(y;\theta_l)\Lambda(y;\theta_j) \\ &+\sum_{i=1}^{m}\sum_{k=N_L+1}^{N_L+N_R} \alpha_{li}w_{ik} \Lambda(y;\theta_{l})P(y;\theta_{k})
\end{aligned}\) & \(\begin{aligned}
&\sum_{i=1}^{m}\sum_{j=1}^{N_L} \alpha_{li}w_{ij} \lambda(y_i;\theta_{li})\Lambda(y;\theta_l)\Lambda(y;\theta_j) \\ &+\sum_{i=1}^{m}\sum_{k=N_L+1}^{N_L+N_R} \alpha_{li}w_{ik} \lambda(y_i;\theta_{li})\Lambda(y;\theta_{l})P(y;\theta_{k})
\end{aligned}\) & \(\begin{aligned}&\sum_{i=1}^{m}\sum_{j=1}^{N_L}\frac{\nu_{ij}}{2^{2m+1}} \\&+ \sum_{i=1}^{m}\sum_{k=N_L+1}^{N_L+N_R}\frac{\nu_{ik}}{2^{3m+1}}\end{aligned}\) \\
\hline
RBF & (\ref{eq:augSILL_limApproxRbf}) & \(\begin{aligned} &\sum_{i=1}^{m}\sum_{j=1}^{N_L} \alpha_{li}w_{ij} H(y;\theta_{j},\theta_{l})\end{aligned}\) & \(\begin{aligned} \sum_{i=1}^{m}&\sum_{j=1}^{N_L} 2\alpha_{li}w_{ij} \lambda(y_i;\theta_{li}) H(y;\theta_{j},\theta_{l}) \end{aligned}\) & \(\begin{aligned}\sum_{i=1}^{m}\sum_{j=1}^{N_L}\frac{\nu_{ij}}{2^{3m+1}}\end{aligned}\) \\
\hline
RBF & (\ref{eq:augSILL_rbfPrime}) & \(\begin{aligned} &\sum_{i=1}^{m}\sum_{j=1}^{N_L} \alpha_{li}w_{ij} \Lambda(y;\theta_{j})P(y;\theta_{l}) \\&+\sum_{i=1}^{m}\sum_{k=N_L+1}^{N_L+N_R} \alpha_{li}w_{ik} P(y;\theta_i)P(y;\theta_k)\end{aligned}\) & \(\begin{aligned}
&\sum_{i=1}^{m}\sum_{j=1}^{N_L} \alpha_{li}w_{ij} \lambda(y_i;\theta_{li})\Lambda(y;\theta_{j})P(y;\theta_{l}) \\&+\sum_{i=1}^{m}\sum_{k=N_L+1}^{N_L+N_R} \alpha_{li}w_{ik} \lambda(y_i;\theta_{li})P(y;\theta_i)P(y;\theta_k)
\end{aligned}\) & \(\begin{aligned}&\sum_{i=1}^{m}\sum_{j=1}^{N_L}\frac{\nu_{ij}}{2^{3m+1}} \\&+ \sum_{i=1}^{m}\sum_{k=N_L+1}^{N_L+N_R}\frac{\nu_{ik}}{2^{4m+1}}\end{aligned}\) \\
\hline
\end{tabular}
\caption{Approximations to and properties of error bounds for the four equations referred to in the \textbf{Ref.} column. The reference equation is approximated as the corresponding equation in the \textbf{Approximation} column. We give the error of this approximation in the \textbf{Difference (Error)} column. The \textbf{Error Bound} column gives a bound on this error. The \textbf{Description} column refers to the type of dictionary function approximated in the row. The right side of Fig. 1 in part 1 of this paper \cite{johnson2022SILL} shows where these approximations fit into showing uniform finite approximate closure. The integer $m$ is the dimension of the measurements, $y$.}
\label{tab:linearityErrorAugSILL}
\end{table*}
\subsection{Expectation of Approximation Error Vanishes}\label{sec:augSILL_error}
This section simultaneously addresses the approximation of two related mathematical objects.
\begin{enumerate}
\item It addresses the approximation of Equation (\ref{eq:augSILL_limApproxLog}) with (\ref{eq:cor3post}) and Equation (\ref{eq:augSILL_limApproxRbf}) with Equation (\ref{eq:cor4post}). This is the lower left step in the right side of Fig. 1 in \cite{johnson2022SILL}.
\item It also addresses the approximation of Equations (\ref{eq:cor3pre}) and (\ref{eq:cor4pre}) with the Lie derivatives of conjunctive logistic and RBFs respectively. This is the upper right step in the right side of Fig. 1 in \cite{johnson2022SILL}.
\end{enumerate}
In total, there are four distinct approximations, see Table \ref{tab:linearityErrorAugSILL}. Each case, as illustrated in Fig. 1 in \cite{johnson2022SILL}, our approximation is a step closer to the Koopman model. We either go from: \begin{enumerate}
\item an intermediate approximation of the Lie derivative that is a nonlinear combination of dictionary functions to a linear combination of the same dictionary functions, or
\item the Lie derivative itself to a bilinear intermediate approximation.
\end{enumerate}
To understand our approximation error we compute the expected values of a single dimensional logistic and RBF. We do so with parameters and measurement values sampled from uniform distributions defined on the interval $[-a,a]$. We choose this statistical model for how our data and parameters are sampled, because 1) the data and parameters are assumed to belong to a bounded continuum, and 2) the uniform distribution is the maximum entropy distribution for a continuous random variable on a finite interval. Since our error terms are weighted sums of products of these functions we, under the assumption of independence, estimate the expected value of our error terms via the linearity and product rule of expectation.
We cannot explicitly compute the probability density function (PDF) of our logistic and RBFs, so, we compute the values of these integrals numerically. Intermediate steps and details of this approximation are in Section B of the Appendix in \cite{johnson2022SILL}. The only difference is that a RBF is substituted for a logistic function. In Fig. \ref{fig:expectedVals} we show their calculated expected values and variances for symmetric uniform distributions with different values of $a$.
We find that the expected value of a logistic function will be $1/2$ (see Fig. \ref{fig:expectedVals}). Its variance, as we sample in a wider interval, tend to the functional extremes of zero and one. This is favorable for the linearity of our approximation since, for all $\varepsilon\in (0, 0.5]$, $(0.5-\varepsilon)(0.5+\varepsilon) = 0.25 - \varepsilon^2 < 0.25 = (0.5)^2$. So, products of more extreme samples are lower in value than products of samples near the expected value. The expected value of an RBF will be no greater than $1/4$, and it decreases as $a$ increases
\begin{figure}[ht]
\includegraphics[width=\linewidth ]{expectedVals.png}
\caption{Expected values and variances of logistic and RBFs with parameters and measurement values sampled from symmetric uniform distributions of various interval radii. Note that the expected value of a logistic function is always $1/2$, and that of the RBF is bounded above by $1/4$.}
\label{fig:expectedVals}
\end{figure}
We approximate a single term in the sum of the error function and extrapolate via the sum and product rule of expectation under the assumption of independence to see how nearly linear our approximation is (see Section B in \cite{johnson2022SILL} of the Appendix). We can conservatively bound the expectation of approximation error as a product that decreases exponentially with the number of measurements. The error bounds for a conjunctive logistic and RBF are \begin{equation}\begin{aligned}\label{eq:CEBlogRBF} E[\Lambda]<\frac{1}{2^m}\mbox{ and } E[P]<\frac{1}{4^m}=\frac{1}{2^{2m}}.\end{aligned}\end{equation} Since, $H$ will be a conjunctive RBF in $\frac{1}{2^m}^{th}$ of the measurement-parameter space, its weight of decrease can be bounded by \begin{equation}\begin{aligned}\label{eq:CEBH} E[H]<\left(\frac{1}{2^{2m}}\right)\left(\frac{1}{2^m}\right)=\frac{1}{2^{3m}}.\end{aligned}\end{equation}
We record the full error terms and bounds in Table \ref{tab:linearityErrorAugSILL}.
These bounds are combined with the bounds that arise from Corollaries 1, 2, 3 and 4 to bound the approximation error of a Koopman model using the augSILL dictionary. A summary of all these bounds is given explicitly in Table \ref{tab:ovarallSummary} of the Appendix.
In summary, the error, $\epsilon_l(y)$, is as well or better behaved for augSILL than SILL dictionaries. So, there is a uniform bound on $\epsilon_l$, $B>0$ for augSILL dictionaries, much like there is for a standard SILL dictionary. This means that augSILL models must satisfy uniform finite approximate closure.
It is reasonable to assume that (1) the measurements do not belong to the set $M_{\bar\Lambda,\bar P}$, a subset of $\mathbb{R}^m$ of measure zero, and (2) the dictionary parameters and Koopman approximation matrix are bounded. Under these assumptions, augSILL dictionaries, in the limit of an increasing number of measurements and increasing steepness of their dictionary functions, have that their bounding constant, $B>0$, approaches zero in expectation, $B\rightarrow 0$. Since the uniform bound on the error of this model can be arbitrarily small, the augSILL dictionaries can be used to build accurate Koopman models for any dynamic system of the from of Eq. (1) in \cite{johnson2022SILL}.
\section{Numerical Examples}\label{sec:numerical}
To come full circle, we need to reexamine deepDMD and compare it to heterogeneous dictionary models. DeepDMD uses a feedforward neural network to simultaneously parameterize the matrix $K$ as well as the dictionary functions $\psi(y)$. The deepDMD algorithm has built accurate predictive models and its dimension ($N\in\mathbb{Z}^+$) scales well with that of the modeled system \cite{yeung2019learning}. We build a novel comparison of deepDMD and a much lower-parameter model built from the augSILL basis. The lower parameter augSILL model learns as quickly and accurately as the deep-learning-based model.
To explain why algorithms like EDMD have variable success, we also contribute a head to head comparison of five dictionaries for Koopman learning. We see the deep-learning-inspired dictionaries vastly outperform orthogonal polynomial dictionaries. This suggests that issues with algorithms like EDMD may be resolved by selecting a dictionary proven to satisfy uniform finite approximate closure.
Unless specified otherwise we use simulated data generated from uniformly distributed initial states run with SciPy's ODE integration software. Each of these results is concerned with the discrete-time, data-driven problem statement. The system's state is directly measured at even time intervals.
\begin{figure*}[ht]
\centering
\includegraphics[width=0.7\linewidth ]{algComparison20dim.png}
\caption{A comparison of the full matching pursuit algorithm to SGD on four nonlinear systems. We plot 5-step prediction error. The SGD algorithm (in red) shows consistently better temporal scaling as well as comparable 5-step prediction error.}
\label{fig:mp_vs_sgd}
\end{figure*}
\subsection{Choosing center parameters}
Even when a dictionary class and model dimension are selected, each individual problem will warrant a unique parameterization of the dictionary. Given a dictionary, such as the augSILL dictionary, how do we choose the parameters of each dictionary function? We consider two algorithms, matching pursuit and stochastic gradient descent (SGD).
Matching pursuit \cite{mallat1993matching} considers an expansive list of potential dictionary functions and greedily adds the function that lowers the value of the objective function (Eq. (\ref{eq:objective})) most. Matching pursuit adds one function at a time to the model.
We use SGD to attack a host of non-convex optimization problems. It is famous, in part, because of its use in training artificial neural networks. SGD can be directly applied to parameterize a fixed number of dictionary functions from data, much like it learns the parameters of a neural network.
\subsubsection{Which algorithm do we choose?}\label{sec:numerical-comp}
We compared two variations of matching pursuit, as well as SGD for learning augSILL models of four dynamic systems, the Van der Pol oscillator, the Duffing oscillator, the Lokta-Volterra model and the Gardner-Collins toggle switch. The specific parametrization of these systems is given in Section \ref{sec:AppendixSystems} of the Appendix. The measurements for each system are the state variables themselves.
We found that the full matching pursuit algorithm and SGD had similar performance for a 20 dimensional Koopman operator (see Fig. \ref{fig:mp_vs_sgd}).
We focus on accuracy and performance for 20 dimensional models as a step-in for modeling higher dimensional systems. Also, we note that SDG was about 1.5 times faster when building a 20 dimensional model. Since SGD was the better choice for building larger Koopman models in terms of time to execute, and performed comparably in 5-step prediction error, we compare this algorithm to deepDMD. Since deepDMD utilizes SGD, we can compare model accuracy at each training epoch.
\begin{figure}[ht]
\centering
\includegraphics[width=\linewidth ]{GlycoldDMD_vsSGD_Metric=Test_Error5.jpg}
\caption{A comparison deepDMD to SGD with the augSILL basis on the seven dimensional model of glycolysis given in \cite{daniels2015efficient}. The solid horizontal bar is the mean performance of the DMD algorithm, the dotted bars are one standard deviation above and below this mean. The spines are error bars for the measured epochs.
\label{fig:DeepDMDvsSGD}
\end{figure}
\subsection{AugSILL basis vs deepDMD}
The SILL and augSILL dictionaries are a targeted study of the dictionary functions generated from deepDMD. Our visualization of deepDMD observables showed a convergence to sums of SILL and augSILL dictionary observables. Can we use these dictionaries to build a model on par with deep learning?
To challenge ourselves, we used a seven dimensional glycolysis model to generate our testing and training data \cite{daniels2015efficient}. This data was all generated from a single initial condition $x_0=[1,0.19,0.2,0.1,0.3,0.14,0.05]$. The augSILL model reaches a comparable 5-step prediction error to deepDMD in under 1000 training epochs and neither significantly changes over the next 4000 epochs, see Fig. \ref{fig:DeepDMDvsSGD}. Note that the model we learn using the augSILL basis has $995$ parameters. All in all, the deepDMD model has a total of $3,949$ parameters. All of the augSILL parameters are easily interpreted as center, steepness and weight parameters of a logistic or RBF.
\begin{figure*}[ht]
\centering
\includegraphics[width=.85\linewidth ]{altAlgComparison.png}
\caption{Five-step prediction accuracy vs training epoch of Koopman learning with SGD using various dictionaries. Plotted on a log scale. Note that the Hermite and Legendre polynomial basis (in blue) have greater error throughout the training process. The spines are error bars at each epoch where 5-step error was measured.}
\label{fig:algComp}
\end{figure*}
\subsection{Comparison of dictionaries for Koopman learning}
Now we address the relationship between the choice of dictionary and the success of a Koopman model. What properties do successful Koopman dictionaries have in common? The augSILL basis performed on par with deepDMD. Would we have gotten similar results using other dictionaries
We learn the systems in Eq. (\ref{eq:vanDerPol}), (\ref{eq:duff}), (\ref{eq:pred}), and (\ref{eq:togg}), parameterized with the SGD algorithm. We do so with the SILL, augSILL and summed one-dimensional RBFs, as well as two different orthogonal polynomial dictionaries (Legendre and Hermite). We compare their 5-step average prediction error for a 5, 10 and 20 dimensional Koopman model of each system.
For the 5 dimensional models the choice of dictionary seemed mostly irrelevant. However, for a 5 dimensional model, SGD only outperformed standard DMD for the toggle switch. So, we don't have enough system dimensions with these dictionaries to want to use SGD in the first place.
Building a 10 and 20 dimensional Koopman model of each system, we find that the basis inspired from the outputs of deepDMD (SILL, augSILL and summed RBFs) have lower 5-step prediction errors (see Fig. \ref{fig:algComp}). Each of these three basis had nearly identical errors, though the SILL basis can take more training iterations to perform comparably. The Legendre polynomial basis outperformed the Hermite polynomial basis in each case. For the 20 dimensional model of the Van der Pol system the augSILL basis was over 372 times more accurate than the Legendre polynomial basis, for the Duffing Oscillator it was over 221 times more accurate, for the Predator-Prey System it was over 18 times as accurate and for the Toggle Switch it was over 328 times as accurate.
\section{Conclusion}
In this paper, we demonstrate that, even in a data-driven setting, one can glean real value from mixing basis functions and forming a heterogeneous dictionary for Koopman learning. This is an important insight as current dictionary based Koopman learning typically works with homogeneous dictionaries. The success of heterogeneous dictionaries helps to explain the success of the deepDMD algorithm, and may extend to other deep-learning-based Koopman learning algorithms \cite{wehmeyer2018time, lusch2018deep, otto2019linearly}.
The mixed-basis models can scale efficiently both in model parameters and model dimension as the number of measurements used to construct the model grows. In this paper we demonstrated this point by constructing a heterogeneous dictionary, the augSILL dictionary, as a general solution to the data-driven Koopman learning problem. We then demonstrated the {\em uniform finite approximate closure} of this dictionary and showed it, in numerical simulations, on a number of distinct learning problems.
In under 1000 training epochs augSILL models matched the accuracy of trained deepDMD models. These models were anywhere from 18 to 372 times as accurate as models made using the polynomial dictionaries, which have no guarantees of uniform finite approximate closure. The augSILL dictionary performs like deepDMD to learn a dynamic system from data using an order of magnitude fewer parameters. Further, the augSILL dictionary is fully specified with closed-form analytical expressions (unlike deepDMD) and as a consequence of the theoretical results in this paper, satisfies a unique numerical property of uniform approximate finite closure.
Our methodology provides a template for understanding how deep neural networks successfully approximate governing equations \cite{brunton2016discovering}, the action of operators and their spectra \cite{mezic2005spectral}, and dynamical systems \cite{mezic2013analysis}. Further, these results provide a pattern for improving scalability and interpretability of dictionary-based learning models for dynamical system identification. In studying the dictionary functions learned by deepDMD we found that the algorithm tended to learn redundant observables when the Koopman model was too high dimensional. A future line of work should explore this model-reduction property of DeepDMD, and see if augSILL models do the same.
\begin{ack}
Any opinions, findings and conclusions or recommendations expressed in this material are
those of the author(s) and do not necessarily reflect the
views of the Defense Advanced Research Projects Agency
(DARPA), the Department of Defense, or the United States
Government. This work was supported partially by a Defense Advanced Research Projects Agency (DARPA) Grant
No. FA8750-19-2-0502, PNNL Grant No. 528678, ICB Grant No. W911NF-19-D-0001 and No. W911NF-19-F-0037, and ARO Young Investigator Grant No. W911NF-20-1-0165.
\end{ack}
\bibliographystyle{plain}
|
\section{Introduction}
More than a century after its proposal, General Relativity continues to score successes by accurately describing an ever-increasing range of phenomena, all the way from cosmology to the spectra of gravitational waves from binary mergers and the shadows of black holes. Aside from its truly impressive predictive power, Einstein gravity is a remarkably rigid theoretical structure. In some sense, this is an inevitable consequence of the combined constraints of Lorentz invariance and quantum mechanics. Indeed, General Relativity is the unique theory of a massless spin-2 particle that mediates a $1/r^2$ force at long distances between matter sources~\cite{Weinberg:1965nx}.
\vskip4pt
Despite its myriad empirical confirmations, there is a somewhat disquieting feature of Einstein gravity. Since space and time themselves fluctuate, rigorously defining observable quantities is notoriously slippery. {\it Where} do we measure something, and {\it when}? One recourse is to turn to asymptotic observables, where we pin the spacetime at infinity and define quantities relative to this asymptotic metric. This approach has been very successful in both asymptotically flat spaces (where we study the $S$-matrix) and in asymptotically anti-de Sitter spaces (where we can study boundary correlators). These asymptotic observables provide another viewpoint on the theoretical inevitability of Einstein gravity and reveal highly non-obvious connections to gauge theories~\cite{Bern:2019prr}.
Essentially, these remarkable features are a consequence of the fact that
amplitudes involving massless particles are extremely constrained~\cite{Benincasa:2007xk,Schuster:2008nh,McGady:2013sga}.
For example, there is a finite number of possible three-particle amplitudes involving massless spin-2 fields that are consistent with Lorentz invariance. General Relativity is built from the most relevant one at long distances, and the constraints of
consistent factorization of the four-graviton $S$-matrix then lead uniquely to General Relativity.\footnote{Similar results can also be obtained with the assumption of invariance under Lorentz boosts relaxed~\cite{Pajer:2020wnj,Hertzberg:2020yzl}. In the cosmological context, extensions of the graviton sector were studied for example in~\cite{Creminelli:2014wna,Bordin:2017hal,Bartolo:2017szm,Bordin:2020eui,Bartolo:2020gsh,Cabass:2021fnw}.}
\vskip4pt
In this paper, we study the dynamics of pure gravity in de Sitter space, which is the other natural choice of asymptotics where we could contemplate defining observables. Much as in flat space (or anti-de Sitter space), the symmetries of de Sitter space are very constraining, completely fixing the form of the graviton three-point function up to three overall constants \cite{Maldacena:2011nz}. The four-point function is therefore the first true probe of the dynamics of Einstein gravity, and our goal is to compute this object in exact de Sitter space.\footnote{An earlier computation appears in~\cite{Fu:2015vja}, but their result appears to be incomplete. See also \cite{Li:2018wkt} for a discussion of the exchange part of the correlator using spinor-helicity variables.} The anti-de Sitter space version of this result appears in \cite{Raju:2012zs} (see also \cite{Albayrak:2019yve} for a 5-point graviton exchange correlator in anti-de Sitter space).
In the context of flat space scattering amplitudes, the analogous object---the four graviton $S$-matrix---has served as a font of wisdom, illuminating hidden structures in amplitudes, and elucidating the constraints of theoretical consistency. Our hope is that similar insights will flow from the graviton four-point function in de Sitter space. One motivation is to better understand the
theoretical rigidity of General Relativity in cosmological spacetimes. Another is to gain a glimpse of some hidden simplicity of cosmological correlators from this highly constrained object.\footnote{Indeed, there has been much recent progress in the study of cosmological correlators, e.g.,~\cite{Raju:2012zr,Raju:2012zs,Raju:2011mp,Mata:2012bx,Bzowski:2011ab,Bzowski:2012ih,Bzowski:2013sza,Bzowski:2019kwd,Kundu:2014gxa,Kundu:2015xta,Arkani-Hamed:2015bza,Shukla:2016bnu,Arkani-Hamed:2017fdk,Arkani-Hamed:2018kmz,Baumann:2019oyu,Albayrak:2018tam,Albayrak:2019yve,Sleight:2019mgd,Sleight:2019hfp,Baumgart:2019clc,Gorbenko:2019rza,Mirbabayi:2020vyt,Cohen:2020php,Green:2020whw,Green:2020ebl,BBBB,Baumann:2020dch,Sleight:2020obc,Sleight:2021iix,Baumann:2021fxj,Albayrak:2020fyp,Albayrak:2020isk,Armstrong:2020woi,Meltzer:2021zin,Bonifacio:2021azc,Hogervorst:2021uvp,DiPietro:2021sjt,Sleight:2021plv,Goodhew:2021oqg,Premkumar:2021mlz,Bzowski:2022rlz,Pethybridge:2021rwf, Green:2022slj,Pimentel:2022fsc,Jazayeri:2022kjy,Armstrong:2022csc,Cabass:2022rhr,Armstrong:2022jsa,Armstrong:2022vgl,Mirbabayi:2022gnl}, and the four-graviton correlator provides additional theoretical data in the search for deeper underlying structures.}
\vskip4pt
Given the complexity of this four-point function, we compute it in several complementary ways.
The first, presented in Section~\ref{ss:lagrangian}, is a brute force calculation in the in-in formalism, working in the same gauge used in~\cite{Maldacena:2002vr} to compute the three-point function. After solving the constraint equations of General Relativity, we compute both the graviton trispectrum and the associated quartic graviton wavefunction coefficient. In Section~\ref{ss:dSgggglift} we show how this result could have been derived from the study of exchange correlators involving spinning particles with the help of a few heuristic derivations and explicit calculations. The idea here is to take the flat-space wavefunction coefficient and lift it to de Sitter space by acting with various differential operators, along the lines of~\cite{Benincasa:2019vqr,Baumann:2021fxj,Hillman:2021bnk}. Finally, in Section~\ref{s:dS} we present a third derivation using the boostless cosmological bootstrap~\cite{BBBB,MLT} (see \cite{Baumann:2022jpr,Benincasa:2022gtd} for reviews). In this three-step approach we employ three general results. In the first step, we fix all terms that are singular at vanishing partial energies in terms of three-point functions, as done in~\cite{MLT}, using the cosmological optical theorem~\cite{COT}. In the second step, we employ the flat-space amplitude to fix the residue of the leading total-energy pole~\cite{Maldacena:2011nz,Raju:2012zr}. Finally, all the remaining subleading total-energy poles are fixed by the manifestly local test~\cite{MLT}. All three methods produce the same four-point function, providing a useful cross-check.
\vskip4pt
To explain our approach with as few technical complications as possible, in Section~\ref{s:warmup} we present a warm-up computation of the four-point function of a massless scalar induced by the minimal coupling to gravity, using the same three methods outlined above. For both the scalar and graviton trispectrum, our different methods yield results that are identical up to terms that can be generated by field redefinitions. In Section~\ref{sec:conclusions} we conclude and speculate on future directions and lessons.
\paragraph{Notation and conventions:}\label{Conventions}
We work with the mostly positive metric signature $(-+++)$ and our Fourier convention is
\begin{equation}
f(\textbf{x})=\int \dfrac{{\rm d}^3{\bm k}}{(2\pi)^3}{f}({\bm k})e^{i{\bm k}\cdot\textbf{x}}\equiv\int_{{\bm k}}{f}({\bm k})e^{i{\bm k}\cdot\textbf{x}} \label{FT} \,,
\end{equation}
where bolded letters indicate three-dimensional spatial vectors.
To account for factors of $(2\pi)^3$, it is convenient to define the normalized Dirac delta function
\begin{equation}
\hat \delta({\bm k})\equiv (2\pi)^3\delta^{(3)}({\bm k})\,.
\end{equation}
We will typically denote derivatives with respect to conformal time by a prime, e.g., $\phi'=\partial_\eta \phi$.
The wavefunction of the universe $\Psi$ at conformal time $\eta_0$ is parameterized as
\begin{equation}
\label{eq:WFdef}
\Psi[\eta_{0},\gamma({\bf k})] = \text{exp}\left[-\sum_{n=2}^{\infty} \frac{1}{n!} \sum_{h_i=\pm} \int_{{{\bm k}}_{1}, \ldots, {{\bf{k}}}_{n}} \psi_{n}({\bf k_{1}}, \ldots, {\bf k_{n}})\, \hat \delta \left(\raisebox{.75pt}{$\sum$}\, {\bm k}_a \right) \gamma^{h_1}({\bf k}_{1}) \ldots \gamma^{h_n}({\bf k}_{n}) \right],
\end{equation}
where the functions $\psi_{n}$ are called wavefunction coefficients, and
$\gamma^h({\bm v}{k})$ is the spin-$h$ Fourier mode of the graviton with $h=\pm 2$ (the dependence of $\psi_n$ on the polarizations $h_i$ is left implicit).
In Fourier space, the graviton is parameterized in terms of $\gamma^\pm$ with helicity $\pm 2$ according to
\begin{equation}\label{eq:gamma}
\gamma_{ij}(\eta,{\bf x})=\int_{{\bm v}{k}} e^{i{\bm v}{k}\cdot {\bm v}{x}} \sum_{h=\pm} \epsilon_{ij}^{h}({\bm v}{k}) \gamma_{\textbf{k}}^h(\eta) \,,
\end{equation}
where the polarization tensors $\epsilon_{ij}^h$ satisfy
\begin{subequations}
\label{eq:pol-conditions}
\begin{align}\label{pol1}
\epsilon_{ii}^{h}({\bm v}{k})&=k^{i}\epsilon^{h}_{ij}({\bm v}{k})=0 & \text{(transverse and traceless)}\,,\\
\epsilon_{ij}^{h}({\bm v}{k})&=\epsilon_{ji}^{h}({\bm v}{k})& \text{(symmetric)} \,, \\
\epsilon_{ij}^{h}({\bm v}{k})\epsilon_{jk}^{h}({\bm v}{k})&=0 & \text{(lightlike)} \,, \\
\epsilon^{h}_{ij}({\bm v}{k})\epsilon^{h'}_{ij}({\bm v}{k})^{\ast}&=2\delta_{hh'} & \text{(normalization)}\,, \\
\epsilon_{ij}^h({\bm v}{k})^{\ast}&=\epsilon_{ij}^h(-{\bm v}{k}) &\text{($ \gamma_{ij}(x) $ is real)} \,.\label{poln}
\end{align}
\end{subequations}
Throughout we denote three-dimensional polarisation vectors, used for writing wavefunction coefficients and correlators, by $\epsilon_{i}$, and four-dimensional polarisation vectors, used for writing scattering amplitudes, by $\varepsilon_{\mu}$.
\vskip4pt
We also define primed correlators with the momentum-conserving delta function removed
\begin{equation}
\langle {\cal{O}} {({\bm k}_1)\dots {\cal{O}}({\bm k}_n)} \rangle \equiv \langle {\cal{O}} {({\bm k}_1)\dots {\cal{O} }({{\bm k}_n}) }\rangle' \hat \delta \left(\raisebox{.75pt}{$\sum$}\, {\bm k}_a \right)\,.
\end{equation}
The three-momenta ${\bf k}_a$ have components $k_{a}^i$ and norms $k_a =|{\bf k}_a| $. The corresponding massless four-momenta have components $k_a^{\mu} = (k_a, k_a^i)$. We define $k_{ab} = k_a+k_b$. The elementary symmetric polynomials in three variables are written as
\begin{equation} \label{ESP}
e_1=k_1+k_2+k_3, \qquad e_2=k_1 k_2 +k_1 k_3+k_2k_3, \qquad e_3=k_1 k_2 k_3,
\end{equation}
while those in four variables are written as
\begin{align}
\mathbb{e}_1&=k_1+k_2+k_3+k_{4}\,, & \mathbb{e}_2&=k_1 k_2 +k_1 k_3+k_1k_4 + k_{2}k_{3}+k_{2}k_{4}+k_{3}k_{4}\,, \nonumber \\ \mathbb{e}_3&=k_1 k_2 k_3+k_1 k_2 k_4+k_1 k_3 k_4+k_2 k_3 k_4\,, & \mathbb{e}_4&=k_1 k_2 k_3 k_{4}.
\end{align}
We will often use $k_{T}$ in place of $e_{1}$ and $\mathbb{e}_{1}$ where no confusion arises. We also define
\begin{equation}
\quad s = | {\bf k}_1+{ \bf k}_2|, \qquad\quad t = | {\bf k}_1+{ \bf k}_3|, \qquad\quad u = | {\bf k}_1+{\bf k}_4|.
\end{equation}
Using three-momentum conservation ($\sum_{a=1}^4 {\bf k}_{a}=0$), these satisfy
\begin{equation} \label{NL_energy_relation}
s^2+t^2+u^2=k_1^2+k_2^2+k_3^2+k_4^2.
\end{equation}
The Mandelstam variables for $k_{a}^{\mu}$ are defined such that
\begin{align}
S &=-(k_1^{\mu} +k_2^{\mu})(k_{1, \mu} +k_{2, \mu})=2 k_1 k_2-2 {\bf k}_1 \cdot {\bf k}_2 =k_{12}^2 - s^2, \label{Stos} \\
T &=-(k_1^{\mu} +k_3^{\mu})(k_{1 \mu} +k_{3 \mu})=2 k_1 k_3-2 {\bf k}_1 \cdot {\bf k}_3 = k_{13}^2 - t^2, \label{Ttot} \\
U &=-(k_1^{\mu} +k_4^{\mu})(k_{1 \mu} +k_{4 \mu})=2 k_1 k_4-2 {\bf k}_1 \cdot {\bf k}_4 = k_{14}^2 - u^2. \label{Utou}
\end{align}
Four-momentum conservation ($\sum_{a=1}^4k_{a}^{\mu}=0$) implies that these satisfy $S+T+U=0$.
\newpage
\section{Scalar trispectrum with graviton exchange}\label{s:warmup}
In this section, we discuss the tree-level four-point function for a massless spectator scalar induced by minimal coupling to pure gravity in exact de Sitter space. This calculation already contains most of the features that we will encounter in the graviton four-point function case while being algebraically simpler.
First, in Section~\ref{ss:ssgss}, we review the explicit bulk perturbation theory calculation, which was first done in Ref.~\cite{Seery:2006vu, Seery:2008ax}. Then, in Section~\ref{ss:dSs4}, we show how the same result can be derived from a more ``on-shell" perspective by lifting the corresponding flat-space wavefunction coefficient to de Sitter space. Finally, in Section~\ref{ss:blesss4}, we derive this result in a third way, using the boostless cosmological bootstrap. Namely, from the cosmological optical theorem and energy shifts, the amplitude limit, and the manifestly local test.
\subsection{Lagrangian calculation}\label{ss:ssgss}
We start with the action of a scalar field minimally coupled to gravity,
\begin{equation} \label{eq:minimal-coupling-action}
S = \int {\rm d}^4 x \sqrt{-g} \left[ \frac{M_{\rm Pl}^2}{2} \left(R^{(4)} -6 H^2\right)- \frac{1}{2} g^{\mu \nu} \partial_{\mu} \phi \partial_{\nu} \phi \right].
\end{equation}
Our goal is to compute the four-point correlation function of the massless scalar $\phi$ that arises from the exchange of a graviton.
We first perform a $3+1$ decomposition in the ADM formalism, where the line element is written as
\begin{equation}
{\rm d} s^2 = -N^2 {\rm d} t^2+h_{ij}({\rm d} x^i +N^i {\rm d} t)({\rm d} x^j +N^j {\rm d} t).
\end{equation}
The extrinsic curvature is $K_{ij} = N^{-1} E_{ij}$, where
\begin{equation}
E_{ij} = \frac{1}{2} ( \dot{h}_{ij} - \nabla_i N_j - \nabla_j N_i).
\end{equation}
We also define $E = h^{ij} E_{ij}$.
In terms of these variables, the action is given by
\begin{equation}
\begin{aligned}
S =~&\frac{M_{\rm Pl}^2}{2}\int {\rm d} t {\rm d}^3 x\sqrt{h}\left(N R^{(3)}-6NH^2+N^{-1}\left(E_{ij}E^{ij}-E^2\right)\right) \\
&- \frac{1}{2}\int {\rm d} t {\rm d}^3 x \frac{\sqrt{h}}{N}\left( - \dot{\phi}^2+2 N^i \dot{\phi} \partial_i \phi +(h^{ij} N^2 -N^i N^j) \partial_i \phi \partial_j \phi \right). \label{eq:3+1action}
\end{aligned}
\end{equation}
In this formalism, the lapse, $N$, and shift, $N_i$, are Lagrange multipliers that enforce constraints. Conceptually, we solve these constraints order-by-order in perturbation theory and substitute their solution back into the action~\cite{Maldacena:2002vr}. The resulting action for the dynamical variables serves as the starting point for the field theory calculation of correlation functions.
\subsubsection{Constraint equations}
Varying the action~\eqref{eq:3+1action} with respect to the lapse and shift gives the constraint equations,
\begin{align}
&R^{(3)}-6H^2-\frac{1}{N^2}\left(E_{ij}E^{ij}-E^2\right) +\frac{1}{M_{\rm Pl}^{2} N^{2}}\left(- \dot{\phi}^2+2 N^i \dot{\phi} \partial_i \phi -(h^{ij}N^2+N^i N^j) \partial_i \phi \partial_j \phi \right)=0,\nonumber\\
&\nabla_j\left[N^{-1}\left(E^j_i-E\delta^j_i\right)\right]-M_{\rm Pl}^{-2} N^{-1} \left(\dot{\phi} \partial_i \phi-N^j \partial_i \phi \partial_j \phi \right)=0.
\label{eq:constraints}
\end{align}
We want to solve these constraints perturbatively for the fluctuations around the de Sitter background solution,
\begin{equation}
{\rm d} s^2= - {\rm d} t^2 + a^2 \delta_{ij} {\rm d} x^i {\rm d} x^j, \qquad\qquad \phi = 0, \label{vanish}
\end{equation}
where the scale factor is $a(t)=e^{ Ht}$. Since the scalar has a vanishing background, we will sometimes refer to it as a ``spectator" field to distinguish it from an inflaton, which would have a non-trivial time-dependent background.
\vskip4pt
We fix the gauge such that the spatial metric takes the form
\begin{equation} \label{eq:gamma-def}
h_{ij} = a(t)^2\, (e^\gamma)_{ij}\,,
\end{equation}
where $\gamma_{ij}$ is both transverse and traceless: $ \gamma_{ii}=\partial_i\gamma_{ij}=0$.
To facilitate finding the perturbative solution of the constraints, we expand the lapse and shift order by order in powers of the perturbations,
\begin{align}
N & = 1 + \alpha^{(1)} +\alpha^{(2)} + \dots , \\
N_i & = N_i^{(1)} + N_i^{(2)} + \dots ,
\end{align}
where the superscript labels the order in perturbations of the fields.
We also further decompose $N_i^{(n)}$ into a longitudinal piece and a transverse piece,\footnote{The notation $\psi^{(n)}$ for the longitudinal component of the shift is conventional in the literature. It should not be confused for a wavefunction coefficient---we hope that the distinction is clear from the context.}
\begin{equation}
N_i^{(n)} = \beta_{i}^{(n)} + \partial_i \psi^{(n)},
\end{equation}
where $\beta_{i}^{(n)}$ is transverse: $\partial^i \beta_{i}^{(n)}=0$. (Note that indices on $\gamma_{ij},\ \partial_i$ and $\beta_i$ are raised and lowered using the flat metric $\delta_{ij}$.) We now substitute the decomposition of the lapse and shift into the equations~\eqref{eq:constraints} and solve them order-by-order.
\vskip4pt
To find the interactions to quartic order, it is sufficient to solve the constraint equations to second order (see, e.g.,~\cite{Chen:2006nt} or Appendix A of~\cite{Pajer:2016ieg}). At first order, the constraint equations are
\begin{align}
12H^2 \alpha^{(1)} +4H e^{-2Ht} \partial^2 \psi^{(1)} & = 0, \\
4 H \partial_i \alpha^{(1)} - e^{-2 H t} \partial^2 \beta_{i}^{(1)} & =0.
\end{align}
These can be solved to find $\alpha^{(1)} =\psi^{(1)}=\beta_i^{(1)}=0$ \cite{Maldacena:2002vr}. The solutions are particularly simple because we are considering a spectator field with a vanishing background~\eqref{vanish}.
The constraint equations at second order read
\begin{align}
12H^2 \alpha^{(2)} +4H e^{-2Ht} \partial^2 \psi^{(2)} + \frac{1}{M_{\rm Pl}^2} \left( \dot{\phi}^2 + e^{-2 Ht} \partial_i \phi \partial_i \phi \right)+ \dots & = 0\,, \\
4 H \partial_i \alpha^{(2)} - e^{-2 H t} \partial^2 \beta_{i}^{(2)} - \frac{2}{M_{\rm Pl}^2} \dot{\phi} \partial_i \phi + \dots& =0\,,
\end{align}
where we have only shown the terms needed to compute the scalar trispectrum (the terms involving $\gamma_{ij}$ are written in Section~\ref{ss:constr}). We can solve these equations to find
\begin{align}
\alpha^{(2)} & = \frac{1}{2M_{\rm Pl}^2 H} \frac{1}{\partial^2} \partial_i( \dot{\phi} \partial_i \phi) + \dots, \\
\beta_i^{(2)} & = \frac{2}{M_{\rm Pl}^2} \frac{e^{2Ht}}{\partial^2} \left( \frac{\partial_i \partial_j}{\partial^2} -\delta_{ij} \right) ( \dot{\phi} \partial_j \phi) + \dots, \\
\psi^{(2)} & =-\frac{e^{2 Ht}}{4 M_{\rm Pl}^2 H \partial^2} \left[ \frac{6 H}{\partial^2} \partial_i( \dot{\phi} \partial_i \phi) +\dot{\phi}^2 +e^{-2 Ht} \partial_i \phi \partial_i \phi \right] + \dots,
\end{align}
where again we have only shown the terms needed for the scalar trispectrum.
\subsubsection{Lagrangian interactions to quartic order}
We now want to substitute the solution to the constraints back into the action to obtain the action governing fluctuations around the background up to quartic order.
Expanding~\eqref{eq:3+1action} gives the following interactions relevant for the computation of the scalar four-point function:
\begin{equation}
\begin{aligned}
S \supset & \int {\rm d} t {\rm d}^3 x \frac{e^{3 Ht} }{2} \left[ \dot{\phi}^2 - e^{-2 Ht} \partial_i \phi \partial_i\phi+ e^{-2 Ht} \gamma_{ij} \partial_i \phi \partial_j \phi \right] \\
&+ \int {\rm d} t {\rm d}^3 x \left[ -\frac{e^{ Ht} }{2} \alpha^{(2)} \partial_i \phi \partial_i \phi - e^{ Ht} \beta^{(2)}_i \dot{\phi} \partial_i\phi - e^{ Ht} \partial_i \psi^{(2)} \dot{\phi} \partial_i \phi-\frac{e^{ 3Ht} }{2} \alpha^{(2)} \dot{\phi}^2 \right] \\
&+ \int {\rm d} t {\rm d}^3 x M_{\rm Pl}^2\left[ -3 e^{3 Ht} H^2 (\alpha^{(2)})^2-2 H e^{H t} \alpha^{(2)} \partial^2 \psi^{(2)} +\frac{1}{4} e^{-H t} (\partial_i \beta^{(2)}_j)^2 \right],
\end{aligned}
\label{eq:quartactionscalar}
\end{equation}
where the last line contains the interactions of the constrained fields coming from the Einstein--Hilbert term.
\vskip4pt
The scalar-graviton interaction that contributes to the scalar trispectrum at tree level is the cubic term $\gamma_{ij} \partial_i \phi \partial_j \phi$, which generates a graviton-exchange diagram, as computed in~\cite{Seery:2008ax}. The quartic terms in $\phi$ contribute to the contact term, as computed in~\cite{Seery:2006vu}. The second-order solutions to the constraints make their first appearance in these quartic terms. Substituting the solutions for the constrained fields, we obtain the following quartic scalar interactions:
\begin{equation} \label{ScalarInteractions}
S \supset \int {\rm d} t {\rm d}^3 x \frac{e^{3 Ht} }{4M_{\rm Pl}^2} \left[ \frac{\partial_i}{\partial^2}( \dot{\phi} \partial_i \phi) \frac{\partial_j}{\partial^2}( \dot{\phi} \partial_j \phi) - \frac{1}{H} \frac{\partial_j}{\partial^2}( \dot{\phi} \partial_j \phi) \left( \dot{\phi}^2+e^{-2 Ht} \partial_i \phi \partial_i \phi \right) + 4\dot{\phi}\partial_i \phi \frac{1}{\partial^2}( \dot{\phi} \partial_i \phi)\right].
\end{equation}
Notice that at quartic order in perturbations, inverse Laplacians in interactions make their first appearance. Such interactions appear because we have integrated out auxiliary fields. The appearance at fourth order is a consequence of the fact that we are working with a spectator field. If the field had a nontrivial background, inverse Laplacians would have appeared already at cubic order~\cite{Maldacena:2002vr}. The presence of inverse Laplacians violates the assumptions underlying a useful constraint on wavefunction coefficients called the manifestly local test (MLT) \cite{MLT}. For a quartic wavefunction coefficient, the MLT says that
\begin{equation} \label{MLT-0}
\frac{\partial }{\partial k_{a}} \psi_{4}\Big|_{k_{a}=0}=0\, , \quad a=1, \dots, 4 .
\end{equation}
One might expect that the wavefunction coefficients considered in this paper do not satisfy the MLT because of the presence of inverse Laplacians in the interactions. However, as we argue in Section \ref{ss:MLT}, the MLT does still hold for these wavefunction coefficients.
\subsubsection{Calculating wavefunction coefficients and correlators}
Once we have the action expanded to the appropriate order, we can calculate wavefunction coefficients and correlators following the standard approaches. As a cross-check of our computations, we calculate the relevant correlation function using both the in-in formalism and by computing the relevant wavefunction coefficients.
\paragraph{The wavefunction:}
In order to compute tree-level wavefunction coefficients, we note that given a collection of fields, $\phi$, the path integral
\begin{equation}
\Psi[\varphi] \ = \hspace{-0.4cm} \int\limits_{\substack{\phi(t) \,=\,\varphi\\ \hspace{-0.45cm}\phi(-\infty)\,=\,0}}
\hspace{-0.5cm} \raisebox{-.05cm}{ ${\cal D} \phi\, e^{iS[\phi]}\,,$ }
\end{equation}
formally solves the Schr\"odinger equation (where $\varphi$ is the field profile at time $t$ and the boundary condition at the initial time should be interpreted with the suitable $i\epsilon$ prescription). At tree-level, this path integral is approximated by the on-shell action evaluated on the classical solution with vacuum initial conditions and Dirichlet late-time boundary conditions. The on-shell action is then a functional of this boundary data (see, e.g.,~\cite{Maldacena:2002vr,Anninos:2014lwa,Goon:2018fyu,Baumann:2020dch,COT} for reviews of the formalism).
\vskip4pt
As a matter of practice, we can efficiently compute the wavefunction coefficients that appear in~\eqref{eq:WFdef} by employing Feynman rules analogous to those that compute flat space scattering amplitudes. The main difference is that we have to work in the time domain and integrate over all possible interaction times. Aside from this, there are two different propagators: external lines are associated to the so-called bulk-to-boundary propagator, $K_k(\eta)$, and internal lines are associated to the bulk-to-bulk propagator, $G_k(\eta,\eta')$. In this particular case, the bulk-to-boundary propagator for a massless scalar/graviton is given by
\begin{equation}
K_k(\eta)=(1- ik\eta)e^{ik\eta}\,,
\label{eq:bbdyprop}
\end{equation}
while the bulk-to-bulk propagator for a massless graviton is
\begin{equation}
G_k^{hh'}(\eta,\eta')=i\frac{H^2}{2k^3}\delta^{hh'}\theta(\eta-\eta')\left(K_{k}(\eta')K^{\ast}_{k}(\eta)-K_{k}(\eta')K_{k}(\eta)\right)+(\eta\leftrightarrow \eta')\,,
\label{eq:bbulkprop}
\end{equation}
where $h,h'$ are the helicities of the incoming/outgoing gravitons. In addition to these propagators, we can derive the relevant interaction vertices from the action~\eqref{eq:quartactionscalar} or~\eqref{ScalarInteractions} in the usual way.
\vskip4pt
Once one has the wavefunction, it can be used to compute equal-time correlation functions as in quantum mechanics:
\begin{equation}
\langle\phi_1 \cdots \phi_N\rangle = \int{\cal D}\phi\, \phi_1\cdots\phi_N\left\lvert\Psi[\phi]\right\rvert^2\,.
\end{equation}
This formula can be expanded perturbatively to give a relation between correlation functions and their corresponding wavefunction coefficients, as we will describe.
\paragraph{In-in correlators:}
We can also directly calculate in-in correlation functions in canonical quantization. In order to proceed with this formalism,
we promote the fields to operators by expanding in terms of creation and annihilation operators,
\begin{align}
\phi_{{\bf k}}( \tau) & = f_k(\tau) a_{{\bf k}} + f^*_k(\tau) a_{-{\bf k}}^{ \dagger}, \\
\gamma^h_{{\bf k}}( \tau) & = f_k(\tau) a_{{\bf k}}^h + f^*_k(\tau) a_{-{\bf k}}^{h\, \dagger},
\end{align}
which have the commutation relations
\begin{align}
\big[ a_{{\bf k}}, a_{{\bf k}'}^{ \dagger} \big] & = \hat{\delta}({\bf k}-{\bf k}'), \\
\big[ a_{{\bf k}}^h, a_{{\bf k}'}^{h' \dagger} \big] & = \frac{2}{ M_{\rm Pl}^2}\hat{\delta}({\bf k}-{\bf k}') \delta_{h h'}.
\end{align}
The massless mode functions for both the scalar and graviton with Bunch--Davies vacuum conditions are
\begin{equation}
f_k(\tau) = \frac{H}{\sqrt{2 k^3}} \left( 1+ i k \tau \right) e^{-ik \tau}.
\end{equation}
In-in correlators are then given by the analogue of Dyson's formula, written in terms of the interaction Hamiltonian $ \mathcal{H}_{\rm int}$,
\begin{equation}
\langle \phi_1\cdots \phi_N\rangle = \langle 0\rvert \bar T\,e^{i\int_{-\infty}^\eta{\rm d}\eta'{\cal H}_{\rm int}}\,\phi_1\cdots \phi_N\,T\,e^{i\int_{-\infty}^\eta{\rm d}\eta'{\cal H}_{\rm int}}\lvert 0\rangle\,,
\end{equation}
where $\lvert 0\rangle$ is the ``in" Fock vacuum. It is often convenient to expand out the exponentials and rearrange them into the commutator formula~\cite{Weinberg:2005vy}
\begin{equation}
\begin{aligned}
\langle \phi_{{\bf k}_1}(\eta) \dots \phi_{{\bf k}_n}(\eta) \rangle&= \sum_{N=0}^{\infty} i^N \int^{\eta}_{- \infty} {\rm d}\eta_1 \int^{\eta_1}_{- \infty} {\rm d}\eta_{2} \dots \int^{\eta_{N-1}}_{- \infty} {\rm d} \eta_N \\
& \hspace{.75cm}\times \left\langle \left[ \mathcal{H}_{\rm int} (\eta_N), \left[ \mathcal{H}_{\rm int} (\eta_{N-1}), \dots [ \mathcal{H}_{\rm int} (\eta_{1}), \phi_{{\bf k}_1}( \eta) \dots \phi_{{\bf k}_n}( \eta) ] \dots \right] \right] \right\rangle.
\end{aligned}
\end{equation}
The late-time correlator is obtained by taking the $\eta \rightarrow 0$ limit,
\begin{equation}
\langle \phi_{{\bf k}_1} \dots \phi_{{\bf k}_n} \rangle \equiv \lim_{\eta \rightarrow 0} \langle \phi_{{\bf k}_1}(\eta) \dots \phi_{{\bf k}_n}(\eta) \rangle,
\end{equation}
and all trispectra of interest in this paper are finite in this limit.
\subsubsection{The power spectrum and bispectrum}
We start by computing the two and three-point objects that are needed to relate the four-point wavefunction coefficient to the analogous correlator.
The two-point correlator, or power spectrum, of a massless scalar is
\begin{equation}
P(k) = \langle \phi_{{\bf k}} \phi_{-{\bf k}}\rangle'= \frac{H^2}{2 k^3},
\end{equation}
while the power spectrum of the graviton with our normalizations is given by\footnote{In terms of $\gamma_{ij}$, the two-point function is
\begin{equation}
\lim_{\eta \rightarrow 0} \left\langle \gamma_{ij} ({\bf k}_1, \eta) \gamma_{kl} ({\bf k}_2, \eta) \right\rangle = \frac{2 H^2}{M_{\rm Pl}^2 k_1^3} (2 \pi)^3 \delta^{(3)}\left( {\bf k}_1 + {\bf k}_2\right) \Pi_{ijkl}^{{\bf k}_1},
\end{equation}
where $\Pi_{ijkl}^{{\bf k}_1}$ is the transverse-traceless spin-2 projector defined in~\eqref{eq:pol-sum}.}
\begin{equation} \label{GravitonPS}
\langle \gamma^{h_1}_{{\bf k}} \gamma^{h_2}_{-{\bf k}}\rangle'=P_{\gamma}(k) \delta_{h_1 h_2}, \qquad\quad P_{\gamma}(k)= \frac{H^2}{M_{\rm Pl}^2 k^3}.
\end{equation}
The real part of the cubic wavefunction coefficient for two scalars and one graviton is
\begin{equation} \label{eq:psi3-scalar}
{\rm Re} \, \psi_{\phi \phi \gamma}({\bf k}_1, {\bf k}_2, {\bf k}_3) = \frac{1}{H^2 k_T^2} \left(e_3+e_2 k_T-k_T^3 \right) ( \epsilon_3 \! \cdot \! {\bf k}_1)^2,
\end{equation}
where the elementary symmetric polynomials $e_2$ and $e_3$ are defined in~\eqref{ESP} and we have written the symmetric polarization tensor as $\epsilon_{3, ij} = \epsilon_{3, i} \epsilon_{3, j}$.\footnote{This is just a convenient way to track the index contractions of a symmetric tensor.} From the boundary point of view, this wavefunction coefficient can be fixed by using the symmetries of de Sitter space, which act like conformal transformations on the boundary~\cite{Mata:2012bx}, or as the lowest order solution to the MLT~\cite{MLT}.
The cubic correlator, or bispectrum, is related to this by
\begin{equation}
\langle \phi_{{\bf k}_1} \phi_{{\bf k}_2} \gamma_{{\bf k}_3} \rangle'= -2 P(k_1) P(k_2) P_{\gamma}(k_3)
{\rm Re} \, \psi_{\phi \phi \gamma}({\bf k}_1, {\bf k}_2, {\bf k}_3).
\end{equation}
We can also obtain the bispectrum from an in-in calculation using the cubic interaction Hamiltonian and the formula
\begin{equation}
\langle \phi_{{\bf k}_1} \phi_{{\bf k}_2} \gamma_{{\bf k}_3} \rangle= \lim_{\eta \rightarrow 0} i \int^{\eta}_{- \infty} {\rm d} \eta_1 \left\langle \left[ \mathcal{H}^{(3)}_{\rm int} (\eta_1), \phi_{{\bf k}_1} ( \eta) \phi_{{\bf k}_2}( \eta) \gamma_{{\bf k}_3} ( \eta) \right] \right\rangle.
\end{equation}
We now turn to the computation of the four-point function.
\subsubsection{The quartic wavefunction coefficient}
There are two conceptually different contributions to the quartic wavefunction coefficient. The first arises from the exchange of a massless graviton between two pairs of scalars. The other contribution comes from the quartic vertices~\eqref{ScalarInteractions}. We obtain the full wavefunction coefficient by adding these together. For example, the contribution coming from graviton exchange in the $s$-channel along with the contact diagram with the same permutation symmetries is
\begin{equation}
\psi_{\phi^4}^{(s)}(\{{\bm k}\}) =\frac{1}{6 H^2 M_{\rm Pl}^2} \left[ f^{(s)}_{(2,2)} s^4 \Pi^{(s)}_{2,2} + f^{(s)}_{(2,1)}s^2 \Pi^{(s)}_{2,1}+ f^{(s)}_{(2,0)} (E_L E_R - s k_T )\Pi^{(s)}_{2,0} + f_{c} \right].
\label{eq:schannelWFmasslessscalar}
\end{equation}
The full permutation-invariant answer is then given by summing over channels:
\begin{equation}
\psi_{\phi^4}(\{{\bm k}\}) = \psi_{\phi^4}^{(s)}(\{{\bm k}\})+\psi_{\phi^4}^{(t)}(\{{\bm k}\})+\psi_{\phi^4}^{(u)}(\{{\bm k}\})\,.
\label{eq:fullmassless4pt}
\end{equation}
In~\eqref{eq:schannelWFmasslessscalar},
$\{{\bm k}\}$ is shorthand for $\{{\bm k}_1,{\bm k}_2,{\bm k}_3,{\bm k}_4\}$ and $\Pi^{(s)}_{m,n}$ are contractions of projectors with external momenta whose explicit form is given in Appendix~\ref{app:PolSums}. The $f$s are functions defined by
\begin{align}\label{f22}
f^{(s)}_{(2,2)} = &\,\frac{1}{k_{T}}-\frac{s^2}{k_{T}E_{L}E_{R}}+\frac{s k_{1}k_{2}}{k_{T}E_{L}^2E_{R}}+\frac{s k_{3}k_{4}}{k_{T}E_{R}^2E_{L}}+\frac{2s k_{1}k_{2}k_{3}k_{4}}{k_{T}^2 E_{L}^2 E_{R}^2} -\frac{s(k_{1}k_{2}+k_{3}k_{4})}{k_{T}^2 E_{L}E_{R}}
\nonumber \\ & + \frac{k_{1}k_{2}}{k_{T}^2E_{L}} +\frac{k_{3}k_{4}}{k_{T}^2E_{R}} + \frac{2k_{1}k_{2}k_{3}k_{4}}{k_{T}^3E_{L}E_{R}}, \\
\label{eq:f21}
f_{(2,1)}^{(s)} =& -\frac{2 k_{1}k_{2}k_{3}k_{4}}{k_{T}^3} - \frac{k_{12}k_{3}k_{4}+k_{34}k_{1}k_{2}}{k_{T}^2}-\frac{k_{12}k_{34}}{k_{T}},
\end{align}
\begin{align}
\label{eq:f20}
f^{(s)}_{(2,0)} = & -f^{(s)}_{(2,1)}, \\
f_c =~& -\frac{2k_1k_2k_3k_4(k_{12}k_{34}+s^2)}{k_T^3}-\frac{1}{k_T^2}(k_1k_2k_3+k_1k_2k_4+k_1k_3k_4+k_2k_3k_4)(k_{12}k_{34}+s^2)\nonumber\\
&-\frac{1}{4k_T}\bigg[
24k_1k_2k_3k_4+8(k_1k_3+k_2k_3+k_1k_4+k_2k_4)^2 \nonumber\\
&\hspace{1.3cm}+(k_1k_2+k_3k_4)\left[9(k_1k_3+k_2k_3+k_1k_4+k_2k_4)-12(k_1k_2+k_3k_4)\right] \nonumber\\
&\hspace{1.3cm}+\frac{1}{2}(k_{12}^2+k_{34}^2+2s^2)\Big(3(k_1k_2+k_3k_4)+4(k_1k_3+k_2k_3+k_1k_4+k_2k_4)\Big)
\bigg] \nonumber\\
&-\frac{3(k_1-k_2)^2(k_3-k_4)^2(k_1k_2+k_3k_4)}{4s^2k_T}-3(k_1k_2k_3+k_1k_2k_4+k_3k_4k_1+k_3k_4k_2)\nonumber\\
& +\frac{k_T}{8}\bigg[
21(k_1k_2+k_3k_4)+34(k_1k_3+k_2k_3+k_1k_4+k_2k_4) \nonumber\\
&\hspace{1.3cm}+3(k_{12}^2+k_{34}^2+2s^2)
-\frac{6}{s^2}(k_1-k_2)^2(k_3-k_4)^2\bigg]-\frac{9}{8}k_T^3\,,
\label{eq:fcdSmassless}
\end{align}
where we recall our notation
\begin{align}
k_{ab} &= k_a+k_b\,, & s&=|{\bf k}_1+{\bf k}_2| \,, \\
k_T&=k_1+k_2+k_3+k_4\,, & E_L &=k_1+k_2+s\,, E_R=k_3+k_4+s\,.
\end{align}
The expressions for $\psi_{\phi^4}^{(t)}(\{{\bm k}\})$ and $\psi_{\phi^4}^{(u)}(\{{\bm k}\})$ are obtained from $\psi_{\phi^4}^{(s)}(\{{\bm k}\})$ by permuting energies and momenta.
This trispectrum was first computed in~\cite{Seery:2006vu,Seery:2008ax}, and the result~\eqref{eq:fullmassless4pt} matches their computation.
\vskip4pt
Note that the leading total-energy pole of~\eqref{eq:schannelWFmasslessscalar} is of degree $3$, which is what we expect if the interaction vertices have at most two derivatives~\cite{BBBB}. There are, however, three-derivative quartic self-interactions in the middle terms in~\eqref{ScalarInteractions}, so some cancellations have occurred. In fact, it is possible to integrate by parts and use the scalar's equation of motion to write these three derivative terms as two derivatives ones (up to boundary terms that do not contribute to the wavefunction or correlator), which makes it clear that there should be no total-energy poles of a higher degree.
\vskip4pt
The trispectrum is related to the wavefunction coefficients by
\begin{equation}
\frac{\langle \prod_{a=1}^4\phi_{{\bf k}_a} \rangle'}{4\prod_{a=1}^4 P(k_a)} =-\frac{1}{2} {\rm Re \,} \psi_{\phi^4}(\{{\bm k}\})+\sum_{3 \,\, {\rm perms}} \sum_h P_{\gamma}(s)
{\rm Re} \, \psi_{\phi \phi \gamma^h}({\bf k}_1, {\bf k}_2, -{\bf s}){\rm Re} \, \psi_{\phi \phi \gamma^h}({\bf k}_3, {\bf k}_4, {\bf s})\,,
\end{equation}
where $\bf{s}=\bf{k}_1+\bf{k}_2$.
This expression can be reproduced directly from an in-in calculation, where
the contact contribution is given by
\begin{equation}
\left\langle \prod_{a=1}^4\phi_{{\bf k}_a} \right\rangle_{\rm c} = \lim_{\eta \rightarrow 0} i \int^{\eta}_{- \infty} {\rm d} \eta_1 \left\langle \left[ \mathcal{H}^{(4)}_{\rm int} (\eta_1), \prod_{a=1}^4\phi_{{\bf k}_a} (\eta) \right] \right\rangle,
\end{equation}
and the exchange contribution is given by
\begin{equation}
\left\langle \prod_{a=1}^4\phi_{{\bf k}_a} \right\rangle_{\rm exc.}= \lim_{\eta \rightarrow 0} i^2 \int^{\eta}_{- \infty} {\rm d} \eta_1 \int^{\eta_1}_{- \infty} {\rm d} \eta_2 \left\langle \left[ \mathcal{H}^{(3)}_{\rm int} (\eta_2), \left[ \mathcal{H}^{(3)}_{\rm int} (\eta_1), \prod_{a=1}^4\phi_{{\bf k}_a}(\eta) \right] \right]\right\rangle.
\end{equation}
In these particular cases, the
interaction Hamiltonians are just minus the corresponding Fourier-transformed interaction Lagrangians.
\subsection{Lifting from flat space to de Sitter space}\label{ss:dSs4}
The direct calculation of the wavefunction coefficient~\eqref{eq:schannelWFmasslessscalar} has the benefit of being completely systematic, but the trade-off is that we were forced to introduce a redundant and somewhat complicated description of the bulk physics. In the end, few of these complications propagate all the way to the final answer, which is somewhat simpler than the procedure that gave rise to it. We are therefore motivated to find a more direct construction of~\eqref{eq:schannelWFmasslessscalar} by working at late times. In Section~\ref{ss:blesss4}, we will give a totally systematic (boostless) bootstrap construction of this wavefunction.\footnote{The wavefunction~\eqref{eq:schannelWFmasslessscalar} was bootstrapped in~\cite{Baumann:2020dch} utilizing the de Sitter symmetries of the answer to construct it via weight-shifting~\cite{Arkani-Hamed:2018kmz,Baumann:2019oyu,Karateev:2017jgd}. The difference in Section~\ref{ss:blesss4} is that we do not directly require invariance under de Sitter boosts (it is instead an output). This turns out to be algebraically simpler, which makes the generalization to the graviton four-point function easier.} Here instead we want to provide a somewhat more artisanal approach. What is lost in rigor is made up for by its simplicity. Essentially, the strategy is to take the corresponding flat space wavefunction as an input and transform it to a de Sitter wavefunction, in the spirit of~\cite{Benincasa:2019vqr,Baumann:2021fxj,Hillman:2021bnk}.
\vskip4pt
The starting point is the four-point wavefunction coefficient in flat space arising from four massless scalars exchanging a graviton in the $s$ channel:
\begin{equation}
\label{eq:flatspacespin2exc}
\psi_{\phi^4}^{(s)\, \rm flat}(\{ {\bf k}\}) =\frac{1}{6}\left(\frac{1}{k_TE_LE_R}s^4\Pi^{(s)}_{2,2}-\frac{1}{k_T}s^2\Pi^{(s)}_{2,1}+\frac{1}{k_T}(E_LE_R-sk_T)\Pi^{(s)}_{2,0} +f_c^{\rm flat}\right)\,,
\end{equation}
where we have defined the function
\begin{equation}
f_c^{\rm flat}=-\frac{k_{12}k_{34}+s^2}{k_T} -\frac{3}{2}\left(\frac{k_{12}(k_3-k_4)^2}{s^2}+\frac{k_{34} (k_1-k_2)^2}{s^2}\right)+ \frac{3}{2}k_T\,.
\end{equation}
For this discussion, we will just take~\eqref{eq:flatspacespin2exc} as given, but it can be systematically constructed by requiring that it have only singularities at physical locations, that these singularities have the correct residues, and that
it vanishes when any of the external momenta are taken to be soft~\cite{Baumann:2021fxj}. (Or, alternatively, it can just be computed directly.) Notice that the wavefunction~\eqref{eq:flatspacespin2exc} is given by energy-dependent form factors multiplying polarization sums, which in turn come from the exchanges of the different helicity modes of the graviton. We can understand the presence of these particular polarization sums from the fact that they reduce to Legendre polynomials in the limit $k_T\to 0$~\cite{Baumann:2020dch} (see Appendix \ref{app:PolSums}). It is conceptually useful to separate the helicity-0 contribution into the term multiplying the $\Pi_{2,0}$ polarization sum and the $f_c^{\rm flat}$ piece (which ensures that the soft limit vanishes), because they will behave somewhat differently when we lift them to de Sitter space.
\vskip4pt
We can write the $s$-channel de Sitter wavefunction in the same schematic form as~\eqref{eq:flatspacespin2exc}:
\begin{equation}
\label{eq:dsspin2exc}
\psi_{\phi^4}^{(s)}(\{ {\bf k}\}) =\frac{1}{6}\left( f^{(s)}_{(2,2)} s^4 \Pi^{(s)}_{2,2} + f^{(s)}_{(2,1)}s^2 \Pi^{(s)}_{2,1}+ f^{(s)}_{(2,0)} (E_L E_R - s k_T )\Pi^{(s)}_{2,0} + f_{c}\right)\,.
\end{equation}
Our goal is now to
obtain the various form factors from their flat space counterparts. The strategy is to compare the structure of the cut obtained from the cosmological optical theorem (COT) \cite{COT} between the flat space wavefunction and the corresponding de Sitter wavefunction (which only requires knowledge of lower-point functions). We then want to find a differential operator that transforms the flat space cut into that of the de Sitter wavefunction. Using this operator, we can transmute the highest helicity form factor to de Sitter space. We can then utilize the simple relation between the highest helicity form factor and those corresponding to lower helicity exchanges.
\vskip4pt
Late-time wavefunction coefficients that arise from unitary time evolution in the bulk spacetime must satisfy the COT. This implies that the contribution to the quartic wavefunction that comes from the $s$-channel exchange of gravitons satisfies~\cite{COT,Goodhew:2021oqg}
\begin{align}
\label{COT4Point}
\psi_{\phi^4}^{(s)}(\{k\}, s)+\psi^{\ast\,(s)}_{\phi^4}(\{-k\}, s) &= \sum_h P_{\gamma}(s)\Big[ \psi_{\phi\phi\gamma^h}(k_{1},k_{2},s)+\psi^{\ast}_{\phi\phi\gamma^h}(-k_{1},-k_{2},s) \Big] \nonumber \\ &\hspace{1cm}\times \Big[ \psi_{\phi\phi\gamma^h}(k_{3},k_{4},s)+\psi^{\ast}_{\phi\phi\gamma^h}(-k_{3},-k_{4},s) \Big] \,,
\end{align}
where $P_{\gamma}(s)$ is the power spectrum of the exchanged field.\footnote{This expression assumes that the mode functions of external and internal fields obey Bunch--Davies vacuum conditions, but generalisations have also been derived~\cite{Cespedes:2020xqq}. Loop level cuts have also been derived in~\cite{Melville:2021lst}, and similar relations hold for higher-point diagrams and those with additional internal lines~\cite{Goodhew:2021oqg}. If the exchanged field has spin, only the highest-helicity components of the exchanged field contribute to this cutting rule \cite{Baumann:2021fxj}. This will play an important role for us since we are interested in wavefunction coefficients due to graviton exchange.} The COT shows that the cut of a four-point function is completely fixed by the three-point functions appearing in the corresponding Feynman diagram.
\vskip4pt
In order to construct the four-point function of massless scalars, we begin by taking the cut of~\eqref{eq:flatspacespin2exc}:
\begin{equation}
\psi^{(s)\,\rm flat}_{\phi^4}(\{ k\}, \{ {\bm k} \}, s)+\psi^{\ast\,(s)\,{\rm flat}}_{\phi^4}(\{- k\}, \{- {\bm k} \}, s) = -\frac{s}{3(k_{12}^2-s^2)(k_{34}^2-s^2)}s^4\Pi_{2,2}\,.
\label{eq:masslessflatcut}
\end{equation}
We then want to compare it to that
of the corresponding de Sitter wavefunction, which can easily be computed from~\eqref{eq:psi3-scalar} by contracting with the $\Pi$ tensor defined in~\eqref{eq:pol-sum}:
\begin{equation}
\psi_{\phi^4}^{(s)}(\{ k\}, \{ {\bm k} \}, s)+\psi^{\ast\,(s)}_{\phi^4}(\{- k\}, \{- {\bm k} \}, s)=\frac{s^3(k_{12}^2+2k_1k_2-s^2)(k_{34}^2+2k_3k_4-s^2)}{3(k_{12}^2-s^2)^2(k_{34}^2-s^2)^2}s^4\Pi_{2,2}\,.
\label{eq:masslessdscut}
\end{equation}
Notice that both~\eqref{eq:masslessflatcut} and~\eqref{eq:masslessdscut} involve the same polarization sum (as expected).
We now want to
find a differential operator that acts on the flat space energy structure multiplying $s^4\Pi_{2,2}$ and produces the de Sitter energy structure. An operator that accomplishes this task is\footnote{The derivation of this operator is one of the slightly artisanal parts of this construction. Some general arguments motivating this form are given in~\cite{Baumann:2021fxj}. Alternatively, this operator can be obtained from inspection of the bulk time integrals that compute the wavefunction.}
\begin{equation}
\label{eq:masslessschanlift}
{\cal D}_s \equiv -\,{\cal S}_{k_1}{\cal S}_{k_2}{\cal S}_{k_3}{\cal S}_{k_4}\left(\int{\rm d} k_{12}\, {\rm d} k_{34}\right)^2\,\left(\frac{2}{k_T}(\partial_{k_{12}}+\partial_{k_{34}})+\partial_{k_{12}}\partial_{k_{34}}\right)\,,
\end{equation}
where all operations (including the integrals) should be understood as acting on everything to their right, and the square on the integral indicates that we should integrate twice with respect to each of $k_{12}, k_{34}$.\footnote{We are being somewhat implicit, but the integrals are definite integrals where the integration contour runs from the variable being integrated to $\infty$, as in~(5.74) of~\cite{Arkani-Hamed:2015bza}.} The
differential operator ${\cal S}_k$ is given by
\begin{equation}
{\cal S}_k = 1-k\partial_k\,.
\end{equation}
With these definitions, we can verify that
\begin{equation}
{\cal D}_s \frac{s}{3(k_{12}^2-s^2)(k_{34}^2-s^2)} = -\frac{s^3(k_{12}^2+2k_1k_2-s^2)(k_{34}^2+2k_3k_4-s^2)}{3(k_{12}^2-s^2)^2(k_{34}^2-s^2)^2}\,,
\end{equation}
as desired.
Now, the idea is that we act with the {\it same} operator on the energy structure multiplying the full flat space wavefunction. In doing so, we obtain the energy structure in de Sitter space that multiplies $s^4\Pi_{2,2}$:
\begin{equation}
\begin{aligned}
f_{(2,2)}^{(s)} ={\cal D}_s \frac{1}{k_TE_LE_R} = \frac{1}{k_T}-\frac{s^2}{k_TE_LE_R}&+\frac{sk_1k_2}{k_TE_L^2 E_R}+\frac{sk_3k_4}{k_TE_LE_R^2}+\frac{2s k_1 k_2k_3k_4}{k_T^2 E_L^2E_R^2}\\
&+\frac{k_1k_2}{k_T^2E_L}+\frac{k_3k_4}{k_T^2 E_R} -\frac{s(k_1k_2+k_3k_4)}{k_T^2 E_L E_R}+\frac{2k_1k_2k_3k_4}{k_T^3 E_LE_R}\,,
\end{aligned}
\end{equation}
which agrees with~\eqref{f22}.
\vskip4pt
We can obtain the form factors multiplying the lower-helicity polarization sums by noting that in flat space these energy structures are just the $s\to 0$ limit of the highest-helicity form factor (with an alternating minus sign). We therefore have
\begin{equation}
f_{(2,1)}^{(s)}=-f_{(2,0)}^{(s)}= \lim_{s\to 0}f_{(2,2)}^{(s)}\,,
\end{equation}
which indeed match~\eqref{eq:f21} and~\eqref{eq:f20}.
\vskip4pt
Finally, we want to consider the uplift of $f_c^{\rm flat}$. This is slightly more subtle than the other terms. The most natural thing to do is to act with the analogue of~\eqref{eq:masslessschanlift} appropriate for contact diagrams:
\begin{equation}
\label{eq:contactlift}
{\cal D}_{k_T} \equiv -{\cal S}_{k_1}{\cal S}_{k_2}{\cal S}_{k_3}{\cal S}_{k_4}\left(\int{\rm d} k_T\right)^2\,.
\end{equation}
Doing this, we find the expression
\begin{equation}
{\cal D}_{k_T}f_c =\frac{2k_1k_2k_3k_4(k_{12}k_{34}+s^2)}{k_T^3}+\frac{(k_1k_2k_3+k_1k_2k_4+k_1k_3k_4+k_2k_3k_4)(k_{12}k_{34}+s^2)}{k_T^2}+\cdots\,,
\label{eq:liftedcontact}
\end{equation}
where we have only written the leading two terms in the $k_T\to 0$ limit. These match precisely those in~\eqref{eq:fcdSmassless}, but the more subleading terms we have suppressed do not. However, it is easy to understand the origin of this defect and correct it.\footnote{Another way to understand this is to note that there was ambiguity in our definition of $f_c^{\rm flat}$---we could have just as well grouped some part of it into the helicity-0 form factor, and this ambiguity affects the lifted wavefunction at ${\cal O}(k_T^{-1})$.} We expect that the full wavefunction coefficient will satisfy the MLT in~\eqref{MLT-0}, for reasons that we explain below. It is relatively straightforward to check that
\begin{equation}
\partial_{k_1}\left(f^{(s)}_{(2,2)} s^4 \Pi^{(s)}_{2,2}\right)\Big\rvert_{k_1=0}=\partial_{k_1}\left(f^{(s)}_{(2,1)}s^2 \Pi^{(s)}_{2,1}\right)\Big\rvert_{k_1=0} = 0\,.
\end{equation}
However, the helicity-$0$ piece satisfies
\begin{equation}
\partial_{k_1}\left(f^{(s)}_{(2,0)} (E_L E_R - s k_T )\Pi^{(s)}_{2,0}\right) \hspace{-1pt}\xrightarrow{k_1=0} \frac{(k_{34}^2+k_3k_4+k_2k_{34})(3(k_3-k_4)^2-s^2)(k_2(3k_2+2k_{34})+s^2)}{4s^2(k_2+k_3+k_4)^2}.
\label{eq:leftovermassless4pt}
\end{equation}
This is certainly not zero and must be compensated by the $f_c$ term. However, the full expression~\eqref{eq:liftedcontact} satisfies the MLT by itself because of the way that we have lifted it from flat space using the ${\cal S}$ operators. We therefore need to correct it in order to cancel off~\eqref{eq:leftovermassless4pt}. The fact that~\eqref{eq:leftovermassless4pt} only contains terms with at most a $(k_2+k_3+k_4)^{-2}$ singularity indicates that we only need to correct~\eqref{eq:liftedcontact} by terms that go like $k_T^{-1}$, or are less singular in the limit that the total energy vanishes.
\vskip4pt
In practice, one can make an ansatz including all possible terms with mass dimension three and which are no more singular than $k_{T}^{-1}$ and add them to~\eqref{eq:liftedcontact}. The answer is uniquely fixed by two requirements: the full wavefunction coefficient satisfies the MLT, and the full wavefunction coefficient vanishes in the limit that one of its external momenta is taken to be soft, as it should thanks to the shift symmetry of the scalar. The combination of these two requirements reproduces~\eqref{eq:fcdSmassless}, as we show more explicitly in Section \ref{ss:blesss4}.
\subsubsection{On the validity of the manifestly local test} \label{ss:MLT}
We now give some justification for using the MLT in this paper. The MLT is a condition that must be satisfied by all wavefunction coefficients arising from theories with manifestly local interactions involving fields whose mode functions are those of a massless scalar or graviton in de Sitter space with Bunch--Davies vacuum conditions \cite{MLT}. As written in \eqref{MLT-0}, the MLT for quartic wavefunction coefficients states that
\begin{equation} \label{MLT}
\frac{\partial }{\partial k_{a}} \psi_{4}\Big|_{k_{a}=0}=0\, , \quad a=1, \dots, 4 .
\end{equation}
In this expression we consider $\psi_{4}$ as a function of $k_a$, $s$, and $t$, and hold $s$ and $t$ fixed when we differentiate with respect to $k_a$. This condition is valid away from the physical configuration in which both $k_a$ and ${\bf k}_a$ are taken to zero, so it is distinct from cosmological soft theorems.\footnote{The MLT is also valid for light fields with $m^2<2H^2$ and generalizes to higher even spacetime dimensions \cite{Goodhew:2022ayb}.}
\vskip4pt
The MLT holds for both contact and exchange diagrams and can be derived in two complementary ways. The first comes from the fact that the bulk-to-boundary propagator for massless scalars and gravitons does not contain a term linear in the energy when expanded around $k=0$. This is the case for all conformal time, $\eta$, and so is inherited by the wavefunction coefficient as long as we keep all other variables fixed. The second method comes from demanding that $(2n-2)$-point functions that arise from a single exchange process between two $n$-point functions do not have spurious poles. The cosmological optical theorem (COT) \cite{COT} for such an exchange process constrains the energy dependence of the constituent $n$-point functions to satisfy~\eqref{MLT} in order for the $(2n-2)$-point function to be regular as the energy of the internal line goes to zero, as it should be in a manifestly local theory. The assumption of {\it manifest locality} is that the interactions involve products of fields and positive powers of their derivatives at the same spacetime point~\cite{BBBB}. This in particular assumes that the Lagrangian does not contain any inverse Laplacians. Such inverse Laplacians may arise in local theories when non-dynamical fields are integrated out~\cite{BBBB}, as we have seen above.
\vskip4pt
Here we wish to point out and prove that the MLT is valid more generally. In particular, it applies to theories whose interaction vertices are finite in Fourier space when one of the momenta is taken to zero. To see this, consider the general interaction
\begin{equation}
\mathcal{L}\supset \int_{{\bm k}_1,\ldots, {\bm k}_n} \hat\delta\left(\raisebox{.75pt}{$\sum$}\, {\bm k}_a \right) F({\bm k}_1,\dots,{\bm k}_n)\prod_{a=1}^n \phi_a({\bm k}_a)\,,
\end{equation}
where $\phi_a$ are some (possibly distinct) fields of arbitrary spin (with spin indices omitted) and $F$ is some kernel that depends on the momenta. The MLT applies to external lines corresponding to the field $\phi_a$ in diagrams involving any number of such interactions as long as the interaction is soft,
\begin{equation} \label{FiniteInteractions}
\lim_{k_a\to 0} F({\bm k}_1,\dots,{\bm k}_n)<\infty\,,
\end{equation}
where again we assume that these fields have the usual de Sitter mode functions with Bunch--Davies vacuum conditions. The proof is the same as in Section 3.2 of \cite{MLT} and relies on the fact that the bulk-to-boundary propagator for such fields does not contain a term linear in the energy of such fields when expanded.
\vskip4pt
This result implies the validity of the MLT for manifestly local theories, since in that case $F$ is a polynomial in the momenta ${\bm k}_a$, and so the above condition is satisfied. But the above argument shows that the MLT also applies in the presence of inverse Laplacians as long as they act on a product of fields, since such an interaction vertex still satisfies~\eqref{FiniteInteractions} in Fourier space. The MLT is therefore satisfied by the four-point functions of massless scalars and gravitons (as can easily be checked for the explicit expressions we write below), and we will use it as part of our bootstrap strategy in Sections~\ref{ss:blesss4} and~\ref{s:dS}.\footnote{Another argument for the validity of the MLT is that in axial gauge all of the non-localities can be shuffled into the graviton bulk-to-bulk propagator~\cite{Raju:2011mp,Albayrak:2019yve}, with all the interactions manifestly local.}
\subsection{The boostless bootstrap}\label{ss:blesss4}
Let us now introduce our third method for constructing four-point functions \cite{BBBB,MLT}. The strategy is again centered on the boundary---the advantage compared to the discussion in Section~\ref{ss:dSs4} is that it is more systematic and rigorous, providing a further independent check of our bulk results.
Again, we first concentrate on massless scalars exchanging a graviton, but we will use the same method to derive the pure gravity four-point function in Section~\ref{s:dS}.
\subsubsection{Overview of approach}
The overall strategy contains three separate steps from which we can bootstrap quartic wavefunction coefficients.\footnote{We focus on the real part of the wavefunction coefficients, since for parity-even theories only the real part contributes to expectation values of field operators. See~\cite{Cabass:2021fnw} for details. Similar methods to the ones we employ here can be used to fix the imaginary parts too~\cite{MLT}.}
\begin{itemize}
\item \textbf{Step 1: Unitarity and partial-energy poles:} ~ Late-time wavefunction coefficients that arise from unitary time evolution in the bulk spacetime must satisfy the COT,~\eqref{COT4Point}. Throughout this work we will make a distinction between the parts of wavefunction coefficients that come from time evolution and the kinematic pieces that consist of contractions between polarisation tensors and spatial momentum vectors.
We refer to the former as the~\textit{trimmed} contribution, which is what is constrained by the COT.
Trimmed wavefunction coefficients arising from $s$-channel exchanges can have at most three singularities: they can diverge when the total energy ($k_{T} = \sum_{a=1}^4 k_{a}$) goes to zero, or when either of the partial-energy sums ($E_{L} = k_{1}+k_{2}+s$ and $E_{R} = k_{3}+k_{4}+s$) go to zero.\footnote{This restriction follows from the choice of Bunch--Davies vacuum.}
It was noticed in~\cite{MLT} that the COT in the form~\eqref{COT4Point} fixes the residues of $\textit{all}$ partial-energy poles. Indeed, only the first term on the LHS of~\eqref{COT4Point} has poles when $E_{L}$ or $E_R$ is taken to zero. (In the second term, the sign of the energies $k_1$ and $k_2$ have been flipped, so it is singular at different loci in energy space, where either $k_{12}-s$ or $k_{34}-s$ vanish, but regular when $E_L,E_R$ vanish.)
Since only $\psi_4$ on the LHS of~\eqref{COT4Point} has partial-energy singularities, we can use the (known) RHS of the same equation to fix the Laurent expansion of these singularities.
Since only the highest-helicity components of the exchanged field contribute to the cut, partial-energy poles can only arise due to the exchange of these components. From a Lagrangian point of view, this makes sense since the non-dynamical potential modes can be integrated out in favour of additional contact diagrams, as we saw in Section~\ref{ss:ssgss}, and these diagrams only give rise to total-energy poles.
The residues of partial-energy poles can be efficiently fixed using the de Sitter energy shifts derived in~\cite{MLT,Baumann:2021fxj}, which resemble the flat-space shifts employed in~\cite{Arkani-Hamed:2017fdk}, with the constituent three-point functions given as an input. We will introduce these shifts below.
\item \textbf{Step 2: The amplitude limit and leading total-energy pole:} ~ It is now well-known that cosmological wavefunction coefficients contain scattering amplitudes for the same process on their leading total-energy pole. This was noticed in~\cite{Raju:2012zr,Maldacena:2011nz}. This fact has played an important role in many approaches to bootstrapping cosmological correlators and we will employ it here to fix the residue of the leading total-energy pole of four-point functions.
For the quartic wavefunction coefficients of interest in this paper, the relationship is \cite{COT}\footnote{This differs from the result of~\cite{COT} by an overall minus sign due to different conventions for the amplitude. In the conventions of this paper, the four-point amplitude for $\mathcal{L} = \lambda \phi^4/4!$ is $\mathcal{A} =\lambda$.}
\begin{align} \label{AmplitudeLimit}
\lim_{k_{T} \rightarrow 0} \psi_{\phi^4}(\{ {\bf k}\}) =\frac{2}{H^2} \frac{\mathbb{e}_{4}\mathcal{A}_{4}}{k_{T}^3},
\end{align}
which contains both contact and exchange contributions. The theories of interest in this paper can be written to involve only interactions with net two derivatives, and therefore the order of the leading total-energy pole scales as $k_T^{-3}$.\footnote{In general, for massless fields the degree of the leading pole is given by~\cite{BBBB}
\begin{equation} \label{eq:pole-order}
p =1 + \sum_{V} ([V ]-4),
\end{equation}
where the sum is over the vertices $V$ in a given diagram and $[V]$ is the mass dimension of $V$. For partial-energy singularities, one should instead sum over the vertices associated to the subgraph whose energy is conserved.}
To use~\eqref{AmplitudeLimit}, we write down an ansatz for the quartic wavefunction that has an overall factor of $\mathbb{e}_{4} / k_{T}^3$ and has homogeneity degree 3 under rescaling momenta.
This is the correct scaling for massless scalars and gravitons required by scale invariance.
We then demand that this ansatz satisfies~\eqref{AmplitudeLimit} for some specific amplitude $\mathcal{A}_{4}$.
\item \textbf{Step 3: Manifestly local test and subleading total-energy poles:} ~ The previous two steps fix all partial-energy poles along with the leading total-energy pole of a quartic wavefunction coefficient, given the three-point function and four-point amplitude as inputs. The final step is to fix the subleading total-energy poles. To do this, we demand that the full quartic wavefunction coefficient satisfies the MLT~\eqref{MLT}, which, as we discussed in Section~\ref{ss:MLT}, applies to spectator scalars and gravitons. Like the COT, the MLT constrains the trimmed part of the wavefunction only.
To employ the MLT, we write down an ansatz for subleading total energy poles that is consistent with scale invariance and Bose symmetry and then fix the free parameters by enforcing the MLT.
\end{itemize}
We now apply these three steps to bootstrap the wavefunction coefficient with $s$-channel symmetries, $\psi_{\phi^4}^{(s)}$, which captures the $s$-channel exchange diagram and some contributions from contact diagrams. The other channels are given by permuting the energies and momenta. Throughout we will be working at tree-level, which implies that the quartic wavefunction coefficients we consider are rational functions \cite{Anninos:2014lwa,Goodhew:2022ayb}.
\subsubsection{Unitarity and partial-energy poles}
Since we are considering massless scalars exchanging a graviton, the relevant three-point function is that of two massless scalars and the transverse, traceless graviton $\psi_{\phi \phi \gamma}$. This was written above in~\eqref{eq:psi3-scalar} and is given by
\begin{equation} \label{TwoScalarsOneGraviton}
\psi_{\phi \phi \gamma}({\bf k}_1, {\bf k}_2,{\bf k}_3) = \frac{e_{ij}({\bm k}_{3}) k_{1}^{i} k_{2}^{j}}{H^2} \tilde{\psi}_{3}(k_1, k_2, k_3), \qquad \tilde{\psi}_{3}(k_1, k_2,k_3) = \frac{1}{k_{T}^2}(k_{T}^3 - k_{T}e_{2} - e_{3}).
\end{equation}
We refer to $\tilde{\psi}_{3}$ as the trimmed part, which comes from the time evolution. This wavefunction coefficient can also be fixed using the MLT, where it arises as the solution with the lowest degree total-energy pole \cite{MLT}.
\vskip4pt
We can use this three-point function to fix all the partial-energy poles of $\psi_{\phi^4}^{(s)}$, as explained above. This calculation was done in~\cite{MLT}, but we repeat it here. By gluing together two copies of~\eqref{TwoScalarsOneGraviton}, the contribution to the four-point function will be a product of a tensor structure, coming from the two copies of $e_{ij}k^{i} k^{j}$, and a trimmed part which only depends on the energies $\{k \}$ and $s$. The tensor structure is
\begin{equation}
\sum_{h = \pm} e_{ij}^{h}({\bf s})k_{1}^{i}k_{2}^{j} e_{lk}^{h}(-{\bf s})k_{3}^{l}k_{4}^{k},
\end{equation}
where we sum over the two possible helicities of the exchanged graviton. We can write this in terms of the polarisation sum $ \Pi_{2,2}^{(s)}$, given in Appendix~\ref{app:PolSums}.
If we write the trimmed part as $f_{\text{cut}}$---which we will compute below---then we have
\begin{equation} \label{WFfromCut}
\psi_{\phi^4}^{(s)}( \{ {\bf k} \} ) = \frac{f_{\text{cut}}}{12 H^2 M_{\rm Pl}^2} s^4 \Pi^{(s)}_{2,2} + \ldots,
\end{equation}
where $\ldots$ represents terms without partial-energy poles (which will be fixed at a later stage) and we have included a factor of $H^2 / M_{\rm Pl}^2$ from the graviton power spectrum~\eqref{GravitonPS}.
\vskip4pt
We write the trimmed part as $f_{\text{cut}}$ in~\eqref{WFfromCut} because it is fixed by cosmological cutting rules. This $s$-channel wavefunction coefficient is fundamentally a function of the independent variables $k_{a},s,t$, but in order to write it in a way where the permutation symmetries are manifest, we write the polarisation sum in terms of $u$ as well, even though $u$ can always be eliminated by~\eqref{NL_energy_relation}.
\vskip4pt
To compute this trimmed part, we use de Sitter partial-energy shifts \cite{MLT,Baumann:2021fxj}. The trimmed wavefunction coefficient is a function of $k_{a},s$ only, and to fix it we change to a different set of independent variables.
We consider $f_{\rm cut}$ as a function of $\{E_{L},E_{R},k_{1}k_{2},k_{3}k_{4},s\}$ so that
$f_{\text{cut}} = f_{\text{cut}}(E_{L},E_{R},k_{1}k_{2},k_{3}k_{4},s)$,
and we define
\begin{align}
\tilde{\psi}_{3,L}(E_{L},k_{1}k_{2},s) &= \tilde{\psi}_{3}(k_1, k_2, s), \\
\tilde{\psi}_{3,R}(E_{R},k_{3}k_{4},s) &= \tilde{\psi}_{3}(k_3, k_4, s).
\end{align}
With these definitions, we can write the COT as
\begin{equation}\label{NewCot}
f_{\text{cut}}(E_{L},E_{R},k_{1}k_{2},k_{3}k_{4},s) + f_{\text{cut}}^{\ast}(-E_{L}+2s,-E_{R}+2s,k_{1}k_{2},k_{3}k_{4},s) = \Xi,
\end{equation}
where we have defined
\begin{equation}
\begin{aligned}
\Xi =&\, \frac{1}{s^3}\left[\tilde{\psi}_{3,L}(E_L,k_1 k_2,s)-\tilde{\psi}_{3,L}(E_L-2s,k_1k_2,-s)\right] \\
&~\times \left[\tilde{\psi}_{3,R}(E_R,k_3 k_4,s)-\tilde{\psi}_{3,R}(E_R-2s,k_3k_4,-s)\right].
\end{aligned}
\end{equation}
Here we have used the COT for contact diagrams to eliminate the complex conjugate as~\cite{COT}
\begin{equation}
\tilde{\psi}_{3, L}^{\ast}(-E_{L}+2s,k_{1}k_{2},s) = -\tilde{\psi}_{3,L}(E_{L}-2s,k_{1}k_{2},-s).
\end{equation}
The second term on the LHS of~\eqref{NewCot} is analytic around $E_{L},E_{R}=0$ and therefore does not contribute to the residues of partial-energy poles, so we can fix the poles of the first term from knowledge of $\Xi$. To isolate the partial-energy poles, we perform a complex shift of the partial energies that keeps the total energy fixed:
\begin{equation}
f_{\text{cut}}(E_{L},E_{R},k_{1}k_{2},k_{3}k_{4},s) \mapsto \tilde{f}_{\text{cut}}(z) = f_{\text{cut}}(E_{L}+z, E_{R}-z, k_{1}k_{2},k_{3}k_{4},s).
\end{equation}
This shifted wavefunction coefficient trivially reproduces the unshifted one at $z=0$, is an analytic function in the complex $z$ plane, except for isolated poles in two locations, $z = -E_{L}$ and $z = E_{R}$, and has a Laurent expansion near these poles given by~\cite{MLT}
\begin{align}
\label{Laurentz}
\tilde{f}_{\text{cut}}(z)&=\sum_{0<n\leq m}\dfrac{A_n(E_R,E_L,k_1k_2,k_3k_4,s)}{(z+E_L)^n}+{\cal O}(z+E_L),\\ \label{An}
A_n &=\dfrac{1}{(m-n)!} \left[\partial^{m-n}_z (z+E_L)^m\, \Xi(E_L+z,E_R-z,k_1k_2,k_3k_4,s)\right]_{z=-E_L},
\end{align}
where $m$ is the order of the leading partial energy pole, and the expressions for the $A_{n}$ follow from~\eqref{NewCot} and are completely fixed by the three-point function. In $f_{\text{cut}}$, the expansions around $E_{L}=0$ and $E_{R}=0$ are equivalent since we are using the same three-point function for each sub-diagram. We can use the residue theorem to write
\begin{equation}
f_{\text{cut}}(E_L,E_R,k_1k_2,k_3k_4,s)=\dfrac{1}{2\pi i}\oint\limits_{{\cal C}_0}{\rm d} z\, \dfrac{\tilde{f}_{\text{cut}}(z)}{z}\,,
\end{equation}
where ${\cal C}_0$ is a contour that encircles the origin, oriented clockwise. We can then deform the contour and write the trimmed wavefunction coefficient as a sum over residues plus a boundary term as~\cite{MLT}
\begin{align} \label{Psi4Residue}
f_{\text{cut}}(E_L,E_R,k_1k_2,k_3k_4,s) &=-\text{Res}\left[\dfrac{\tilde{f}_{\text{cut}}(z)}{z}\right]_{z=-E_L}-\text{Res}\left[\dfrac{\tilde{f}_{\text{cut}}(z)}{z}\right]_{z=E_R}+B \\
& = f_{\text{Res}} + B .
\end{align}
The boundary contribution at infinity, $B$, is given by
\begin{align}\label{bound}
B=\dfrac{1}{2\pi i}\oint\limits_{{\cal C}_\infty}{\rm d} z\, \dfrac{\tilde{f}_{\text{cut}}(z)}{z}\,,
\end{align}
and we can use the Laurent expansion~\eqref{Laurentz} to write
\begin{align} \label{psireshere}
f_{\text{Res}}&=\sum_{0<n\leq m}\dfrac{A_n(E_L,E_R,k_1k_2,k_3k_4,s)}{E_L^n}+\sum_{0<n\leq m}\dfrac{A_n(E_R,E_L, k_3k_4,k_1k_2,s)}{E_R^n}\,.
\end{align}
We see that the partial-energy poles are now completely fixed by taking derivatives of the three-point function, which fixes the trimmed four-point function up to a boundary term that can only be singular when the total energy goes to zero: $E_{L}+E_{R}-2s \rightarrow 0$.\footnote{\label{ft:mlts}As explained in~\cite{MLT}, this boundary term might need to contain a simple $s^3$ term in order to satisfy the COT. This can be easily checked by taking~\eqref{Psi4Residue} and comparing the LHS and RHS of~\eqref{NewCot}. We refer the reader to \cite{MLT,Baumann:2021fxj} for further details on this procedure.}
\vskip4pt
We now turn back to the case of interest here and use~\eqref{TwoScalarsOneGraviton} to compute the $A_{n}$. The degree of the leading partial-energy poles is equivalent to the degree of the leading total-energy pole of the corresponding three-point function, which for us is two, so that $m=2$ in~\eqref{Laurentz}. We then have \cite{MLT}
\begin{align}
A_{2}(E_{L},E_{R},k_{1}k_{2},k_{3}k_{4},s) &= \frac{2 k_{1}k_{2}s}{k_{T}^2(E_{L}+E_{R})^2}\big[2k_{3}k_{4} + (E_{L}+E_{R})k_{T}\big], \\
A_{1}(E_{L},E_{R},k_{1}k_{2},k_{3}k_{4},s)&= \frac{1}{k_{T}^3(E_{L}+E_{R})^3}\sum_{n=0}^4 a_{n}s^{n},
\label{eq:A1exp}
\end{align}
where the coefficients appearing in the expansion~\eqref{eq:A1exp} are
\begin{align}
a_{0} &= 2k_{1}k_{2}(E_{L}+E_{R})^2[(E_{L}+E_{R})^2 + 2k_{3}k_{4}], \\
a_{1} &= -4k_{1}k_{2}(E_{L}+E_{R})[(E_{L}+E_{R})^2 - 2k_{3}k_{4}], \\
a_{2} &= -2(E_{L}+E_{R})^4 -4 (E_{L}+E_{R})^2(k_{1}k_{2}+k_{3}k_{4})-16k_{1}k_{2}k_{3}k_{4}, \\
a_{3} &= 8(E_{L}+E_{R})[(E_{L}+E_{R})^2 + k_{1}k_{2}+k_{3}k_{4}], \\
a_{4} &= -8(E_{L}+E_{R})^2,
\end{align}
and $k_{T} = E_{L}+E_{R}-2s$. If we plug these expressions into~\eqref{psireshere}, the resulting $f_{\text{Res}}$ satisfies the COT so there is no need to include any additional $s^3$ pieces (see Footnote~\ref{ft:mlts}). It follows that
\begin{equation}
\begin{aligned} \label{fCut}
f_{\text{cut}} = &-\frac{2s^2}{k_{T}E_{L}E_{R}}+\frac{2s k_{1}k_{2}}{k_{T}E_{L}^2E_{R}}+\frac{2s k_{3}k_{4}}{k_{T}E_{R}^2E_{L}}+\frac{4s k_{1}k_{2}k_{3}k_{4}}{k_{T}^2 E_{L}^2 E_{R}^2} \\ &-\frac{2s(k_{1}k_{2}+k_{3}k_{4})}{k_{T}^2 E_{L}E_{R}}
+ \frac{2k_{1}k_{2}}{k_{T}^2E_{L}}+\frac{2k_{3}k_{4}}{k_{T}^2E_{R}} + \frac{4 k_{1}k_{2}k_{3}k_{4}}{k_{T}^3E_{L}E_{R}}.
\end{aligned}
\end{equation}
We note that in the final expression for $f_{\text{cut}}$, all spurious poles have cancelled out, and consistency dictates the need for total-energy poles even though no information about these residues was input directly. We see that these partial-energy poles match those found by computing nested time integrals in Section \ref{ss:ssgss}, given by $f_{(2,2)}^{(s)}$.
\subsubsection{The amplitude limit and leading total-energy pole}
Having fixed all the partial-energy poles, we now ensure that the leading total-energy pole reproduces the correct scattering amplitude.
The four-point amplitude for minimally coupled scalars is
\begin{equation} \label{ScalarsAmplitude}
\mathcal{A}_4= \frac{1}{4 M_{\rm Pl}^2} \frac{(S^2+T^2+U^2)^2}{STU}.
\end{equation}
Our strategy for constructing the full wavefunction coefficient is to work
``channel-by-channel", meaning that we initially concentrate on the part of the wavefunction coefficient that has the symmetries of an $s$-channel exchange diagram. The relevant symmetry group is $\mathbb{Z}_2 \times \mathbb{Z}_2$, which acts on the momenta as follows: ${\bf k}_{1} \leftrightarrow {\bf k}_{2}$, ${\bf k}_{3} \leftrightarrow {\bf k}_{4}$, and $({\bf k}_{1},{\bf k}_{2}) \leftrightarrow ({\bf k}_{3},{\bf k}_{4})$. We then sum over different channels at the end of the calculation. We have written~\eqref{WFfromCut} in a way that makes these symmetries manifest, using a redundant set of variables since the energies are related by~\eqref{NL_energy_relation}. We therefore need to decompose~\eqref{ScalarsAmplitude} into a sum over channels where each channel has the relevant symmetries when written in terms of $S,T,U$ (which satisfy $S+T+U=0$). When we match the $s$-channel wavefunction coefficient to the $S$-channel amplitude according to~\eqref{AmplitudeLimit}, we will respectively eliminate $u$ and $U$ such that we work with independent variables.
We define the $S$-channel amplitude as
\begin{equation} \label{MasslessScalarsAmp}
\mathcal{A}_{4}^{(S)}
= \frac{1}{M_{\rm Pl}^2}\frac{TU}{S},
\end{equation}
such that $\mathcal{A}_{4} = \mathcal{A}_{4}^{(S)}+\mathcal{A}_{4}^{(T)}+\mathcal{A}_{4}^{(U)}$. Such a definition is ambiguous (the invariant quantity is the residue of the $S=0$ pole), but we must make some choice and our final wavefunction coefficient will not depend on this choice.\footnote{For example, we could add to $\mathcal{A}_{4}^{(S)}$ a term proportional to $S$ without affecting the total amplitude. This extra term would then end up being degenerate with a local term that can always be added to the wavefunction.
}
\vskip4pt
We choose to access the total-energy pole by sending $k_{4} \rightarrow -(k_{1}+k_{2}+k_{3})$ and therefore according to~\eqref{AmplitudeLimit} we must demand that the wavefunction coefficient satisfies
\begin{equation} \label{AmpLimit}
\lim_{k_{T} \rightarrow 0} k_T^3 \psi_{\phi^4}^{(s)} = \frac{2}{H^2 M_{\rm Pl}^2} k_{1}k_{2}k_{3}(k_{1}+k_{2}+k_{3})\frac{(k_{13}^2-t^2)(k_{12}^2+k_{13}^2-s^2-t^2)}{(k_{12}^2-s^2)}.
\end{equation}
We initially take the result from step 1 of the previous section (written as a function of $k_{a},s,t$), ~\eqref{WFfromCut}, and compare its amplitude limit to~\eqref{AmpLimit}. We find that we do not recover the correct expression and that the difference between the two expressions can be written as a series in $s$ with even powers ranging from $s^{-4}$ to $s^{2}$. The $s^{-4}$, $s^{0}$ and $s^{2}$ terms have coefficients that are functions of $k_{1}$, $k_{2}$ and $k_{3}$, while the coefficient of the $s^{-2}$ term also depends linearly on $t^2$. We now need to correct~\eqref{WFfromCut} such that we recover the correct amplitude limit and we can do so by writing down an ansatz order-by-order in $s$.
\vskip4pt
Recall that our ansatz should be a function of \textit{all} internal and external energies, such that all symmetries are manifest, then we eliminate $u$ when we go on the total-energy pole. Our ansatz needs to include terms linear in $t^2$, which means that by symmetry there is also a linear dependence on $u^2$. Our ansatz should scale as $\sim k^3$ by scale invariance and should contain an overall factor of $\mathbb{e}_{4} / k_{T}^3$ given~\eqref{AmpLimit}. We therefore add to~\eqref{WFfromCut}:
\begin{equation} \label{WFfromAmp}
\Delta_{1}\psi_{\phi^4}^{(s)} = \frac{\mathbb{e}_{4}}{s^4 k_{T}^3} \text{Poly}^{(6)}(k_{1},k_{2},k_{3},k_{4},s^2,t^2,u^2),
\end{equation}
where $\text{Poly}^{(6)}$ denotes a general polynomial of its arguments that scales as $\lambda^6$ under ${\bf k}_a \mapsto \lambda {\bf k}_a$.
The wavefunction coefficient should not be singular as $s \rightarrow 0$, as this would correspond to a level of non-locality. We need to use factors of $(k_{1}-k_{2})$ and $(k_{3}-k_{4})$ to cancel such poles. To see this we take ${\bf k}_{2} \rightarrow - {\bf k}_{1} + {\bm \epsilon}$ and ${\bf k}_{4} \rightarrow - {\bf k}_{3} - {\bm \epsilon}$ which fixes ${\bf s} \rightarrow {\bm \epsilon}$. In terms of the energies we have
\begin{align}
k_{2} \rightarrow \sqrt{k_{1}^2 - 2 {\bm \epsilon} \cdot {\bf k}_{1} + \epsilon^2}, \qquad
k_{4} \rightarrow \sqrt{k_{3}^2 + 2 {\bm \epsilon} \cdot {\bf k}_{3} + \epsilon^2}, \qquad
s \rightarrow \epsilon.
\end{align}
Due to the ${\bm \epsilon} \cdot {\bf k}_{1}$ and ${\bm \epsilon} \cdot {\bf k}_{3}$ terms, the only way to guarantee that a polynomial in the external energies cancels a pole in $s$ of order $c$, is for the polynomial to contain overall factors of the form $(k_{1}-k_{2})^{a} (k_{3}-k_{4})^{b}$ subject to $a+b=c$. With these conditions, our ansatz takes the form
\begin{align}
\text{Poly}^{(6)}(k_{1},k_{2},k_{3},k_{4},s^2,t^2,u^2) = \bigg[ & (k_{1}-k_{2})^4 \, \text{Poly}^{(2)}_{1}(k_{12},k_{1}k_{2},k_{34},k_{3}k_{4}) \nonumber \\
& + (k_{1}-k_{2})^2 (k_{3}-k_{4})^2 \,\text{Poly}^{(2)}_{2}(k_{12},k_{1}k_{2},k_{34},k_{3}k_{4}) \nonumber \\ & + (k_{1}-k_{2})^2 s^2 \,\text{Poly}^{(2)}_{3}(k_{12},k_{1}k_{2},k_{34},k_{3}k_{4}) \nonumber \\
& + b_{1} (k_{1}-k_{2})(k_{3}-k_{4}) s^2 (t^2-u^2) \nonumber \\
& + s^4 \, \text{Poly}^{(2)}_{4}(k_{12},k_{1}k_{2},k_{34},k_{3}k_{4}) \nonumber \\ & + b_{2} s^6 \bigg]+ \left( {\bf k}_{1} \leftrightarrow {\bf k}_{3}, {\bf k}_{2} \leftrightarrow {\bf k}_{4} \right) .
\end{align}
The notation $\text{Poly}^{(n)}_{m}(x_1, x_2, \dots)$ denotes a general polynomial of $x_i$ with homogeneity degree $n$ under rescaling momenta, where $m$ is an index distinguishing different polynomials of the same homogeneity degree, and $b_i$ are constants.
We have dropped any terms that depend on $(t^2+u^2)$ since they are degenerate with terms already contained in the ansatz due to the relation in~\eqref{NL_energy_relation}.
\vskip4pt
Demanding that we recover the correct amplitude in the limit $k_T\to 0$ leads to a family of solutions which only differ at $\mathcal{O}(1 / k_{T}^2)$. At this stage we can pick any of these solutions without loss of generality since we will ultimately fix the subleading total-energy poles in the next section. We choose a solution such that our wavefunction coefficient takes the form
\begin{align} \label{AfterStep2}
\psi_{\phi^4}^{(s)} = &\frac{1}{6 H^2 M_{\rm Pl}^2} \left[ \frac{f_{\text{cut}}}{2} s^4 \Pi^{(s)}_{2,2} - \frac{2 \mathbb{e}_{4}}{k_{T}^3}(s^2 \Pi^{(s)}_{2,1} - (k_{12}k_{34}+s^2)\Pi^{(s)}_{2,0}) - \frac{2 \mathbb{e}_{4}(k_{12}k_{34}+s^2)}{k_{T}^3} \right] \\ &+ \ldots \nonumber,
\end{align}
where we see the appearance of familiar polarisation sums, which are written in Appendix~\ref{app:PolSums}. Note that we have not assumed that these sums must appear, rather they are required to realise the correct amplitude limit. This expression now satisfies the COT, has the correct amplitude limit, and the $\ldots$ represents terms that both do not contribute to partial-energy poles and are subleading as $k_{T} \rightarrow 0$. These will be fixed in the next section.
\subsubsection{Manifestly local test and subleading total-energy poles} \label{sec:scalarsMLT}
Finally, we have to ensure that the wavefunction coefficient satisfies the MLT in the form~\eqref{MLT}.
Recall that we take the derivative with respect to an external energy and then set that energy to zero while holding all other variables fixed. We can therefore work order-by-order in $s,t$ (we eliminate $u$ before taking the derivative). We find that~\eqref{AfterStep2} does not satisfy the MLT. Instead, the derivative of this object with respect to $k_{4}$ at $k_{4}=0$ can be written as a series in $s$ with even powers ranging from $s^{-4}$ to $s^{4}$. Each series coefficient depends on $k_{1},k_{2}$ and $k_{3}$, while the $s^{-2}$ and $s^{2}$ coefficients also depend linearly on $t^2$ and the $s^{0}$ coefficient also depends linearly and quadratically on $t^2$. This allows us to write down an ansatz which we add to~\eqref{AfterStep2}. We take the $t^2$ and $u^2$ terms to appear in the combination $(t^2-u^2)$ since this is how they appear in~\eqref{AfterStep2} (regardless of which solution we choose in matching to the amplitude).
\vskip4pt
Our ansatz cannot have partial-energy poles and cannot alter the leading total-energy pole since these have already been fixed. We therefore add to~\eqref{AfterStep2}:
\begin{equation} \label{Step3Ansatz}
\Delta_{2} \psi^{(s)}_{\phi^4} = \frac{1}{s^4 k_{T}^2} \text{Poly}^{(9)}(k_{1},k_{2},k_{3},k_{4},s^2,t^2,u^2),
\end{equation}
where the homogeneity degree of the polynomial is fixed by scale invariance. As before, we need to maintain the symmetries of an $s$-channel exchange diagram and ensure that there are no poles as $s \rightarrow 0$ for physical momenta. This allows us to write
\begin{align}
\text{Poly}^{(9)}(k_{1},k_{2},k_{3},k_{4},s^2,t^2,u^2) = \bigg[ & (k_{1}-k_{2})^4 \text{Poly}^{(5)}_{1}(k_{12},k_{1}k_{2},k_{34},k_{3}k_{4}) \nonumber \\
& + (k_{1}-k_{2})^2 (k_{3}-k_{4})^2 \text{Poly}^{(5)}_{2}(k_{12},k_{1}k_{2},k_{34},k_{3}k_{4}) \nonumber \\ & + (k_{1}-k_{2})^2 s^2 \text{Poly}^{(5)}_{3}(k_{12},k_{1}k_{2},k_{34},k_{3}k_{4}) \nonumber \\
& +(k_{1}-k_{2})(k_{3}-k_{4}) s^2 (t^2-u^2) \text{Poly}^{(3)}_{1}(k_{12},k_{1}k_{2},k_{34},k_{3}k_{4}) \nonumber \\
& + s^4 \text{Poly}^{(5)}_{4}(k_{12},k_{1}k_{2},k_{34},k_{3}k_{4}) \nonumber \\
& +b_{3}(k_{1}-k_{2})(k_{3}-k_{4})s^4(t^2-u^2)k_{T}\nonumber \\
& +b_{4}s^4(t^2-u^2)^2k_{T} \nonumber \\
& + s^6 \text{Poly}^{(3)}_{2}(k_{12},k_{1}k_{2},k_{34},k_{3}k_{4}) \nonumber \\
& + b_{5} s^8 k_{T} \bigg] +\left( {\bf k}_{1} \leftrightarrow {\bf k}_{3}, {\bf k}_{2} \leftrightarrow {\bf k}_{4} \right),
\end{align}
where again the superscripts on the polynomials indicate their homogeneity degree under rescaling the momenta and $b_i$ are constants.
\vskip4pt
In the previous section we found a family of solutions that yields the correct amplitude limit (since we did not use $k_{T}$ as an independent variable), and all of these solutions differ by functions that are captured by~\eqref{Step3Ansatz}. After imposing the MLT we find three free parameters (in addition to $H$ and $M_{\rm Pl}$) and we can write the part of the wavefunction coefficient with $s$-channel symmetries as
\begin{equation}
\begin{aligned} \label{AfterStep3}
\psi_{\phi^4}^{(s)}(\{ {\bm k} \}) & =\frac{1}{6 H^2 M_{\rm Pl}^2} \left[ f^{(s)}_{(2,2)} s^4 \Pi^{(s)}_{2,2} + f^{(s)}_{(2,1)}s^2 \Pi^{(s)}_{2,1}+ f^{(s)}_{(2,0)} (E_L E_R - s k_T )\Pi^{(s)}_{2,0} + f_{c} \right] \\
& + q_{1} (k_{1}^3+k_{2}^3+k_{3}^3+k_{4}^3) \\ & + \frac{q_{2}}{s^2}\big[(k_{1}-k_{2})^2 k_{12}(k_{12}^2-k_{1}k_{2})+(k_{3}-k_{4})^2 k_{34}(k_{34}^2-k_{3}k_{4})\big] \\
& + \frac{q_{3}}{s^4}\big[(k_{1}-k_{2})^4 k_{12}(k_{12}^2+k_{1}k_{2})+(k_{3}-k_{4})^4 k_{34}(k_{34}^2+k_{3}k_{4})\big],
\end{aligned}
\end{equation}
where the constants $q_{i}$ are linear combinations of the constants contained in the polynomial ansatz, and the $f$s are as defined in Section~\ref{ss:ssgss}.
We therefore see that up to these additional three terms, we match the bulk calculation. It is easy to see that these three extra terms satisfy all the conditions we have imposed: they are regular as $E_{L,R} \rightarrow 0$, they do not contribute to the leading total-energy pole, and one can straightforwardly check that they satisfy the MLT.
\vskip4pt
The fact that these additional solutions do not have any
total-energy poles suggests that they arise from field redefinitions~\cite{BBBB}. The $q_{1}$ term is well known, and arises from taking the free theory for the massless scalar and performing the field redefinition $\phi \mapsto \phi + \frac{H^2}{4}q_{1}\phi^3$ (this field redefinition generates a wavefunction coefficient $\psi_{4} = 3 q_{1} \sum_{a=1}^{4}k_{a}^3$, which is what we get when we add the $t$ and $u$ permutations to~\eqref{AfterStep3}). The other two require more complicated field redefinitions which must violate manifest locality given that the resulting wavefunction coefficient contains inverse powers of $s$. They are given by\footnote{We have checked that there are no other field redefinitions that could yield $1/s^2$ or $1/s^4$ terms in $\psi_{\phi^4}^{(s)}$, while also ensuring the absence of $s \rightarrow 0$ poles for physical momenta. This is consistent with what we have found using the MLT in this section.}
\begin{align}
\phi \mapsto \phi & + \frac{ H^2 q_{2} }{3} \left[3 \phi \frac{1}{\partial^2}(\phi \partial^2 \phi) + \phi \frac{1}{\partial^2} (\phi'^2) + \phi' \frac{1}{\partial^2}(\phi \phi') \right], \label{nonlocal1} \\
\phi \mapsto \phi &+
\frac{H^2 q_3}{3}\left[-27\partial^2\phi\frac{1}{\partial^4}(\phi\partial^2\phi)-12\phi\frac{1}{\partial^4}(\phi\partial^4\phi)+6\partial^2\phi\frac{1}{\partial^4}({\phi'}^2)+2\phi'\frac{1}{\partial^4}(\phi\partial^2\phi')\right. \nonumber \\ &\left.-4\phi\frac{1}{\partial^4}(\phi'\partial^2\phi') - 10\phi'\frac{1}{\partial^4}(\phi'\partial^2\phi)-30\partial^2\phi\frac{1}{\partial^4}(\partial_i\phi\partial_i\phi)-60\partial_i\phi\frac{1}{\partial^4}(\phi\partial^2\partial_i\phi)\right]. \label{nonlocal2}
\end{align}
Since any choice of $q_{1,2,3}$ satisfies all our requirements, we can simply set them to any value we like. However, we can also fix these additional terms by imposing a shift symmetry for $\phi$ which is present for minimally coupled massless scalars. A shift symmetry ensures that the wavefunction coefficient has a vanishing soft limit~\cite{Ghosh:2014kba,Bittermann:2022nfh}:
\begin{equation} \label{SoftLimit}
\lim_{{\bm k}_{4} \rightarrow 0} \psi_{\phi^4} = 0.
\end{equation}
We take ${\bm k}_{4} \rightarrow 0$ by sending $k_{4} \rightarrow 0$, $s \rightarrow k_{3}$, $t \rightarrow k_{2}$ (after eliminating $u$), and find that~\eqref{SoftLimit} fixes $q_{1} = q_{2} = q_{3} = 0$. The wavefunction coefficient now exactly matches the bulk expression.
\section{The graviton trispectrum} \label{s:brute}
We now turn to the case of principal interest: the graviton trispectrum. In the previous section, we employed several different methods to derive the analogous wavefunction and correlator for massless minimally coupled scalars interacting via gravity. The consistency of these results gives us confidence in them, and so we will again derive the result in these different ways as a cross-check. We again find agreement, up to terms that can be generated by field redefinitions of the graviton (which are unavoidable ambiguities).
\subsection{Lagrangian calculation} \label{ss:lagrangian}
In this section, we derive the tree-level graviton trispectrum in de Sitter spacetime for pure gravity, starting from the Einstein--Hilbert action. First, in Section~\ref{ss:constr}, after choosing a convenient gauge, we derive the constraint equations for the lapse and shift and then solve them to second order. Second, in Section~\ref{ss:L4}, we obtain the full Lagrangian to quartic order, which we then use to compute the graviton quartic wavefunction coefficient and the corresponding correlator in Section~\ref{ss:psi4}.
\subsubsection{Constraint equations}\label{ss:constr}
The constraint equations~\eqref{eq:constraints} expanded out to second order, including the $\gamma_{ij}$ contributions but not the $\phi$ contributions, can be written as
\begin{equation}
\begin{aligned}
\frac{1}{4a^2}\gamma_{ij,k}\gamma_{ij,k}+\frac{1}{4}\dot{\gamma}_{ij}\dot{\gamma}_{ij}+\frac{4}{a^2}H\partial_i\partial_i\psi^{(2)}+12H^2\alpha^{(2)} &=0,\\
2H\partial_j\alpha^{(2)}+\frac{1}{4}\left(\dot{\gamma}_{ik}\gamma_{ij,k}-\dot{\gamma}_{ik}\gamma_{ik,j}-\dot{\gamma}_{kj,i}\gamma_{ki}-\frac{2}{a^2}\partial_i\partial_i\beta_{j}^{(2)}\right)&=0.
\end{aligned}
\end{equation}
In order to simplify some later equations, it is convenient to define $V_j$ by
\begin{equation}
V_j=\frac{1}{2}\left(-\dot{\gamma}_{ik}\gamma_{ij,k}+\dot{\gamma}_{ik}\gamma_{ik,j}+\dot{\gamma}_{kj,i}\gamma_{ki}\right),
\end{equation}
which has the divergence
\begin{equation}
\partial_iV_i=\frac{1}{2}\partial_i\left(\dot{\gamma}_{jk}\partial_i\gamma_{jk}\right),
\end{equation}
where we have used the fact that the graviton is transverse. The solutions to the constraint equations can then be written in terms of $V_i$ as\footnote{This is the point of disagreement with~\cite{Fu:2015vja}, who have a different solution to the constraints.}
\begin{align}
\alpha^{(2)}&=\frac{1}{4H}\partial^{-2}\partial^jV_j,\\
\psi^{(2)}&=-\frac{a^2}{16H}\partial^{-2}\left(\frac{1}{a^2}\gamma_{ij,k}\gamma_{ij,k}+\dot{\gamma}_{ij}\dot{\gamma}_{ij}+12H\partial^{-2}\partial^jV_j\right),\\
\beta^{(2)}_{j}&=a^2\partial^{-2}\left(\partial^{-2}\partial_j\partial_iV_i-V_j\right).
\end{align}
The presence of inverse Laplacians requires us to specify some boundary conditions for the fields, which---as all three of these terms are perturbations---are that they must vanish on the boundary of our volume. This allows us to integrate spatial derivatives by parts.
\subsubsection{The Lagrangian to quartic order}\label{ss:L4}
Before substituting the solutions of the constraint equations, the Lagrangian to quartic order can be written as
\begin{equation}
\begin{aligned}
\mathcal{L}&=\frac{M_{\rm Pl}^2}{2}a^3\left(-12H^2+\frac{1}{4}\left(\dot{\gamma}_{ij}\dot{\gamma}_{ij}-\frac{1}{a^2}\partial_k\gamma_{ij}\partial_k\gamma_{ij}\right)+\frac{1}{4a^2}\gamma_{ij}\partial_j\gamma_{km}\left(\partial_i\gamma_{km}-2\partial_m\gamma_{ik}\right)\right. \\&\left.-\frac{1}{24}\dot{\gamma}_{ij}\dot{\gamma}_{km}\gamma_{ik}\gamma_{jm}+\frac{1}{24}\dot{\gamma}_{ik}\dot{\gamma}_{ij}\gamma_{jm}\gamma_{km}-\frac{1}{4}\alpha^{(2)}\left(\dot{\gamma}_{ij}\dot{\gamma}_{ij}+\frac{1}{a^2}\partial_k\gamma_{ij}\partial_k\gamma_{ij}\right)-6H^2(\alpha^{(2)})^2\right. \\&\left.+\frac{1}{8a^2}\gamma_{ij}\left(\gamma_{km}\partial_j\gamma_{mn}\partial_k\gamma_{in}-\gamma_{ik}\partial_j\gamma_{mn}\partial_k\gamma_{mn}-\frac{2}{3}\gamma_{km}\partial_m\gamma_{jn}\partial_n\gamma_{ik}\right.\right. \\&\left.\left.+\frac{1}{3}\gamma_{km}\partial_n\gamma_{jm}\partial_n\gamma_{ik}+\frac{4}{3}\gamma_{ik}\partial_k\gamma_{mn}\partial_n\gamma_{jm}+\frac{1}{3}\gamma_{ik}\partial_m\gamma_{kn}\partial_n\gamma_{jm}-\frac{1}{3}\gamma_{ik}\partial_n\gamma_{km}\partial_n\gamma_{jm}\right)\right. \\&\left.+\frac{1}{2a^4}\partial_j\beta^{(2)}_i\partial_j\beta^{(2)}_i-\frac{4H}{a^2}\alpha^{(2)}\partial_i\partial_i\psi^{(2)}-\frac{1}{a^2}V_i(\beta^{(2)}_i+\partial_i\psi^{(2)})\right).
\end{aligned}
\end{equation}
We can then eliminate the lapse and shift and integrate the expression by parts in space (using the fact that $\alpha,\ \beta$ and $\psi$ vanish on the boundary) to obtain the action up to quartic order in graviton fluctuations,
\begin{align}
\mathcal{L}&=\frac{M_{\rm Pl}^2}{2}\left(-12a^4H^2+\frac{a^2}{4}\left(\gamma'_{ij}\gamma'_{ij}-\partial_k\gamma_{ij}\partial_k\gamma_{ij}\right)+\frac{a^2}{4}\gamma_{ij}\partial_j\gamma_{km}\left(\partial_i\gamma_{km}-2\partial_m\gamma_{ik}\right)\right. \nonumber\\&\left.+\frac{a^2}{4}\left(\frac{1}{6}\gamma'_{ik}\gamma'_{ij}\gamma_{jm}\gamma_{km}-\frac{1}{6}\gamma'_{ij}\gamma'_{km}\gamma_{ik}\gamma_{jm}+\frac{1}{2}\gamma_{ij}\gamma_{km}\partial_j\gamma_{mn}\partial_k\gamma_{in}-\frac{1}{2}\gamma_{ij}\gamma_{ik}\partial_j\gamma_{mn}\partial_k\gamma_{mn}\right.\right. \nonumber\\&\left.\left.+\frac{1}{2}\gamma_{ik}\gamma_{ij}\partial_m\gamma_{kn}\partial_n\gamma_{jm}+\gamma_{ik}\gamma_{ij}\partial_k\gamma_{mn}\partial_n\gamma_{jm}+\frac{1}{6}\gamma_{ij}\gamma_{km}\partial_n\gamma_{jm}\partial_n\gamma_{ik}-\frac{1}{6}\gamma_{ik}\gamma_{ij}\partial_n\gamma_{km}\partial_n\gamma_{jm}\right)\right. \nonumber \\&\left.+\frac{a^4}{2}V_j\partial^{-2}V_j-\frac{a}{32H}\partial^{-2}\partial_l\left(\gamma'_{mn}\partial_l\gamma_{mn}\right)\left(\gamma'_{ij}\gamma'_{ij}+\partial_k\gamma_{ij}\partial_k\gamma_{ij}\right) \right. \nonumber\\&\left. +\frac{a^2}{32}\partial^{-2}\partial_i\left(\gamma'_{jk}\partial_i\gamma_{jk}\right)\partial^{-2}\partial_l\left(\gamma'_{mn}\partial_l\gamma_{mn}\right)\right), \label{toApp}
\end{align}
where we have changed variables from cosmic time to conformal time for later convenience. As with the scalar case above, all inverse Laplacians act on a product of fields, so the MLT will be satisfied by our final wavefunction coefficient (see Section~\ref{sec:scalarsMLT}). We can then Fourier transform this expression and rewrite $\gamma_{ij}$ in terms of polarisation tensors using~\eqref{eq:gamma}, where the polarization tensors satisfy the conditions in~\eqref{eq:pol-conditions}.
The Lagrangian is then given by
\begin{align}
\mathcal{L}&=M_{\rm Pl}^2\int_{\bf{k}}\sum_{h} \frac{a^2}{8}\left({\gamma'}_k^h{\gamma'}_k^{h}-k^2\gamma_k^h\gamma_k^{h}\right) \nonumber \\ &+M_{\rm Pl}^2\int_{\bf{k}_1,\bf{k}_2,\bf{k}_3}\sum_{h_1,h_2,h_3}\frac{a^2}{8}\epsilon^{ij}_{1}k^{j}_1\epsilon^{lm}_{2}\left(k_{3}^i\epsilon^{lm}_{3}-2k_{3}^m\epsilon^{il}_{3}\right)\gamma_{k_1}^{h_1}\gamma_{k_2}^{h_2}\gamma_{k_3}^{h_3} \nonumber \\
&+M_{\rm Pl}^2\int_{\textbf{k}_1, \dots, \bf{k}_4}\sum_{h_{1,2,3,4}}-\frac{a^2}{48}\gamma_{k_1}^{h_1}\gamma_{k_2}^{h_2}\gamma_{k_3}^{h_3}\gamma_{k_4}^{h_4}\left(3\epsilon^{ij}_{1}\epsilon^{lm}_{2}k_{3}^j\epsilon^{mn}_{3}k^l_{4}\epsilon^{in}_{4}-3\epsilon^{ij}_{1}\epsilon^{il}_{2}k_{3}^j\epsilon^{mn}_{3}k_{4}^l\epsilon^{mn}_{4} \right. \nonumber \\
&\left.+3\epsilon^{il}_{1}\epsilon^{ij}_{2}k_{3}^m\epsilon^{ln}_{3}k_{4}^n\epsilon^{jm}_{4}+6\epsilon^{il}_{1}\epsilon^{ij}_{2}k_{3}^l\epsilon^{mn}_{3}k_{4}^n\epsilon^{jm}_{4}+k_{3}^nk_{4}^n\epsilon^{ij}_{1}\epsilon^{lm}_{2}\epsilon^{jm}_{3}\epsilon^{il}_{4}-k_{3}^nk_{4}^n\epsilon^{il}_{1}\epsilon^{ij}_{2}\epsilon^{lm}_{3}\epsilon^{jm}_{4}\right)\nonumber\\
&+\frac{a^2}{48}{\gamma_{k_1}^{h_1}}'\gamma_{k_2}^{h_2}{\gamma_{k_3}^{h_3}}'\gamma_{k_4}^{h_4}\left[\epsilon^{ij}_{1}\epsilon^{jm}_{2}\epsilon^{ik}_{3}\epsilon^{km}_4-\epsilon^{ij}_1\epsilon^{ik}_2\epsilon^{km}_3\epsilon^{jm}_4+\frac{3(k_{1}^ik_{2}^i+k_2^2)(k_{3}^jk_{4}^j+k_4^2)}{4s^4}\epsilon^{pq}_1\epsilon^{pq}_2\epsilon^{mn}_3\epsilon^{mn}_4\right. \nonumber \\
&\left.-\frac{3}{s^2}\left(\epsilon^{in}_1k_{2}^n\epsilon^{ij}_2-\epsilon^{in}_1k_{2}^j\epsilon^{in}_2-k_{1}^i\epsilon^{kj}_1\epsilon^{ki}_2\right)\left(\epsilon^{lm}_3k_{4}^m\epsilon^{lj}_4-\epsilon^{lm}_3k_{4}^j\epsilon^{lm}_4-k_{3}^l\epsilon^{mj}_3\epsilon^{lm}_4\right)\right] \nonumber \\
&-\frac{a}{64H}{\gamma'}_{k_1}^{h_1}\gamma_{k_2}^{h_2}\left({\gamma'}_{k_3}^{h_3}{\gamma'}_{k_4}^{h_4}-k_{3}^pk_{4}^p\gamma_{k_3}^{h_3}\gamma_{k_4}^{h_4}\right)\frac{(k_{1}^lk_{2}^l+k_2^2)}{s^2}\epsilon^{mn}_1\epsilon^{mn}_2\epsilon^{ij}_3\epsilon^{ij}_4,
\end{align}
where, for brevity, we have introduced the notation
\begin{equation}
\epsilon^{ij}_a=\epsilon^{h_a}_{ij}(\textbf{k}_a).
\end{equation}
Given the Lagrangian, we can now calculate the cubic and quartic wavefunction coefficients. It turns out to be most convenient to just directly compute the on-shell action, evaluated on the classical solution with the correct boundary conditions. To get the action to fourth order, we need the following perturbative solution to the equations of motion
\begin{equation}
\gamma_k^h=K_k^s\bar{\gamma}^h_k+\sum_{h'}\int {\rm d}\eta' \left.\frac{\delta \mathcal{L}_{3}}{\delta \gamma_k^h}\right\rvert_{\gamma_k^h=K_k^{h'}\bar{\gamma}^{h'}_k}G_k^{hh'}(\eta,\eta')={\gamma_k^h}^{(0)}+\delta \gamma_k^h,
\label{eq:1stordersoln}
\end{equation}
where the bulk-to-boundary and bulk-to-bulk propagators are respectively given by the expressions~\eqref{eq:bbdyprop} and~\eqref{eq:bbulkprop}.
After some redefinition of dummy variables, the variation of the cubic action $\mathcal{L}_{3}$ is given by
\begin{equation}
\begin{aligned}
\frac{\delta \mathcal{L}_{3}}{\delta \gamma_k^h}=M_{\rm Pl}^2\int_{\textbf{k}^3}\sum_{h_4,h_5,h_6}\delta^{hh_4}\delta(\textbf{k}_4-\textbf{k})\frac{a^2}{8}\gamma_{k_5}^{h_5}\gamma_{k_6}^{h_6}\epsilon^{ij}_4&\left(2\epsilon^{il}_5\epsilon^{jm}_6 k_{5}^mk_{6}^l+4\epsilon^{lm}_5\epsilon^{il}_6k_{6}^jk_{6}^m \right.\\&\left.-2\epsilon^{lm}_5\epsilon^{ij}_6k_{6}^lk_{6}^m-\epsilon^{lm}_5\epsilon^{lm}_6k_{6}^i k_{6}^j\right).
\end{aligned}
\end{equation}
Given this expression, we can substitute~\eqref{eq:1stordersoln} into the action and read off the relevant wavefunction coefficients.
\subsubsection{The bispectrum}
We can extract the cubic wavefunction coefficient, which is given by
\cite{Maldacena:2002vr,Maldacena:2011nz}
\begin{equation} \label{GravitonCubicWF}
{\rm Re} \, \psi_{\gamma^3}({\bf k}_{i}; \epsilon_{i}) =- \frac{M_{\rm Pl}^2 }{4 H^2} \left(k_T-\frac{e_2}{k_T} -\frac{e_3}{k_T^2}\right) \big[ \epsilon_1 \! \cdot \! {\bf k}_2 \,\epsilon_2 \! \cdot \! \epsilon_3 + \epsilon_2 \! \cdot \! {\bf k}_3 \,\epsilon_1 \! \cdot \! \epsilon_3 + \epsilon_3 \! \cdot \! {\bf k}_1 \,\epsilon_1 \! \cdot \! \epsilon_2 \big]^2,
\end{equation}
and the bispectrum is related to this by
\begin{equation}
\big\langle \gamma^{s_1}_{{\bf k}_1} \gamma^{s_2}_{ {\bf k}_2} \gamma^{s_3}_{{\bf k}_3} \big\rangle' =-2 {\rm Re} \, \psi_{\gamma^3}({\bf k}_{i}; \epsilon_{i}) \prod_{a=1}^3 P_{\gamma}(k_a).
\end{equation}
We can also compute this correlator from the Hamiltonian using the in-in formalism. These correlators are completely fixed as well by de Sitter symmetries~\cite{Maldacena:2011nz}. Compared to the three-point function for two massless scalars and a graviton, cf.,~\eqref{eq:psi3-scalar}, the only difference lies in the tensor structure. The energy dependence is the same since the bulk-to-boundary propagators for scalars and gravitons are identical.
\subsubsection{The quartic wavefunction coefficient}\label{ss:psi4}
We now turn to compute the quartic wavefunction coefficient for gravitons from the above Lagrangian. There are essentially two complicated aspects of the calculation: the first is simply managing the contractions of polarizations and momenta, while the second is evaluating the time integrals that capture the time evolution of the system.
\vskip4pt
There are five distinct time integrals that we must evaluate to find the quartic wavefunction coefficient. Four are relevant for contact contributions and are given by
\begin{align}
\int{\rm d}\eta\, a^2 K_{k_1}(\eta)K_{k_2}(\eta)K_{k_3}(\eta)K_{k_4}(\eta) &=-i\frac{-2\mathbb{e}_4- \mathbb{e}_3 k_T-\mathbb{e}_2 k_T^2+k_T^4}{H^2k_T^3}-\frac{1}{H^2\eta_0},\\
\int{\rm d}\eta\, a^2 K_{k_1}'(\eta)K_{k_2}(\eta)K_{k_3}'(\eta)K_{k_4}(\eta) &=-i\frac{k_1^2k_3^2\left(k_T^2+k_T(k_2+k_4)+2k_2k_4\right)}{H^2k_T^3},\\
\int{\rm d}\eta\, a K_{k_1}'(\eta)K_{k_2}(\eta)K_{k_3}'(\eta)K_{k_4}'(\eta) &=-2i\frac{k_1^2k_3^2k_4^2k_T+3k_1^2k_2k_3^2k_4^2}{Hk_T^4},\\
\int{\rm d}\eta\, a K_{k_1}'(\eta)K_{k_2}(\eta)K_{k_3}(\eta)K_{k_4}(\eta) &=i\frac{6k_1 \mathbb{e}_4+2k_1^2k_T^3-3k_1^3k_T^2+2k_1^2k_T \mathbb{e}_2+2k_1^4k_T}{Hk_T^4},
\end{align}
where we have only provided expressions for one ordering of momenta since the others can be obtained by permutation.
The final integral is relevant for exchange contributions and is identical to that for massless scalars~\eqref{f22},
\begin{equation}
-iH^2\int{\rm d}\eta\,{\rm d}\eta' a(\eta)^2 a(\eta')^2 K_{k_1}(\eta)K_{k_2}(\eta)K_{k_3}(\eta')K_{k_4}(\eta') G_{s}(\eta,\eta') =
f^{(s)}_{(2,2)}.
\end{equation}
\vskip4pt
We now have all the ingredients we need to compute the wavefunction coefficient: we read off the tensor structures from the Lagrangian and use the above results for the time evolution. We add together all diagrams (as always, the exchange diagram gets an additional factor of one-half from the 4th order correction to the quadratic action~\cite{COT,Goon:2018fyu}). The final quartic wavefunction coefficient can then be written as
\begin{equation}
\label{eq:4gWFdS}
\psi_{\gamma^4}(\{ {\bf k} \} )=\frac{M_{\rm Pl}^2}{16 H^2} \Big[ \psi_{\gamma^4}^{(s)}(\{ {\bf k} \} )+\psi_{\gamma^4}^{(t)}(\{ {\bf k} \} )+\psi_{\gamma^4}^{(u)}(\{ {\bf k} \} )\Big],
\end{equation}
where we have again split the answer into contributions from different channels for convenience. The $s$-channel contribution is
\begin{equation}
\begin{aligned}\label{final}
& \psi_{\gamma^4}^{(s)}(\{ {\bf k} \} ) = M_{\rm Pl}^2 H^2 \psi_{\phi^4}^{(s)} (\epsilon_1 \! \cdot \! \epsilon_2)^2(\epsilon_3 \! \cdot \! \epsilon_4)^2+2f_{(2,2)}^{(s)} \left[ s^2 \Pi^{(s)}_{1,1} (\epsilon_1 \! \cdot \! \epsilon_2)(\epsilon_3 \! \cdot \! \epsilon_4) W_s +2 W_s^2 \right] \\
&\quad -2 f_{2,0}^{(s)} \Pi^{(s)}_{1,0}(\epsilon_1 \! \cdot \! \epsilon_2)(\epsilon_3 \! \cdot \! \epsilon_4) W_s + g_{\rm cont.}^{(s)}(\epsilon_1 \! \cdot \! \epsilon_3)(\epsilon_1 \! \cdot \! \epsilon_4)(\epsilon_2 \! \cdot \! \epsilon_3)(\epsilon_2 \! \cdot \! \epsilon_4)-2 f_{\rm cont.} W_s W_s^{(c)},
\end{aligned}
\end{equation}
where $f_{(2,2)}^{(s)}$ and $f^{(s)}_{(2,0)}$ are defined as in~\eqref{f22} and~\eqref{eq:f20}, and we have defined the following polarization structures:
\begin{align}
W_s & = (\epsilon_1 \! \cdot \! \epsilon_2)\big[ (\epsilon_3 \! \cdot \! k_1)(\epsilon_4 \! \cdot \! k_2)-(\epsilon_3 \! \cdot \! k_2)(\epsilon_4 \! \cdot \! k_1) \big]+(\epsilon_3 \! \cdot \! \epsilon_4)\big[ (\epsilon_1 \! \cdot \! k_3)(\epsilon_2 \! \cdot \! k_4)-(\epsilon_1 \! \cdot \! k_4)(\epsilon_2 \! \cdot \! k_3) \big] \nonumber \\
\label{eq:Wsdef}
& \hspace{.5cm}+ \big[(\epsilon_1 \! \cdot \! k_2) \epsilon_2-(\epsilon_2 \! \cdot \! k_1) \epsilon_1\big]\cdot \left[(\epsilon_3 \! \cdot \! k_4) \epsilon_4-(\epsilon_4 \! \cdot \! k_3) \epsilon_3\right], \\
W_s^{(c)} & =(\epsilon_1 \! \cdot \! \epsilon_4)(\epsilon_2 \! \cdot \! \epsilon_3)-(\epsilon_1 \! \cdot \! \epsilon_3)(\epsilon_2 \! \cdot \! \epsilon_4),
\label{eq:Wcdef}
\end{align}
along with the contact interaction form factors
\begin{align}
f_{\rm cont.} & = \prod_{i=1}^4 (1-k_i \partial_{k_i})k_T (\log k_T -1) =\frac{2 \mathbb{e}_{4}}{k_T^3}+\frac{\mathbb{e}_{3}}{k_T^2} +\frac{\mathbb{e}_{2}}{k_T}-k_T, \\
g_{\rm cont}^{(s)} & = -(k_{12} k_{34}+s^2)\left[\frac{4 \mathbb{e}_{4}}{k_T^3}-\frac{(k_{12}-k_{34})(k_{12}^2+2k_1 k_2-k_{34}^2-2 k_3 k_4 )}{2 k_T^2}+3 \frac{k_1 k_2 +k_3 k_4}{k_T}-\frac{3 k_T}{2} \right] \nonumber \\
& \hspace{13pt}-\frac{4 \mathbb{e}_{4}}{k_T} +\frac{8 k_{12}^2 k_{34}^2 -4 k_{12}^4-4k_{34}^4-2 k_1 k_2 k_{34}^2-2 k_3 k_4 k_{12}^2+8 k_1 k_2 k_{12}^2+8 k_3 k_4 k_{34}^2 }{3 k_T} \nonumber \\
&\hspace{13pt} + \frac{14}{3} k_T (k_1 k_2 + k_3 k_4) - \frac{10 k_T^3}{9}+\frac{16}{9} \sum_{a=1}^4k_a^3.
\end{align}
Notice that the
scalar quartic wavefunction coefficient, $\psi_{\phi^4}^{(s)}$---which we computed in Section~\ref{ss:ssgss}---appears,
and we remind the reader that the various polarisation sums are given in Appendix~\ref{app:PolSums}. Equation~\eqref{eq:4gWFdS} is the main result of this paper. A decompressed version of this expression is included in an ancillary text file.
\vskip4pt
Note that we have only written the real part of the wavefunction coefficient, as this is what contributes to the trispectrum via
\begin{align}
\big\langle \gamma^{h_1}_{{\bf k}_1} \gamma^{h_2}_{{\bf k}_2} \gamma^{h_3}_{{\bf k}_3} \gamma^{h_4}_{{\bf k}_4} \big\rangle' =\,& \Bigg[4 \sum_{3 \, \, {\rm perms}} \sum_h P_{\gamma}(s) \, {\rm Re} \, \psi_{\gamma^{h_1}\gamma^{h_2}\gamma^{h} }({\bf k}_{1},{\bf k}_{2},-{\bf s}) \, {\rm Re} \, \psi_{\gamma^{h_3}\gamma^{h_4}\gamma^{h} }({\bf k}_{3},{\bf k}_{4},{\bf s})\nonumber \\
& -2 \, {\rm Re} \, \psi_{\gamma^4} (\{ {\bf k} \} ) \Bigg] \prod_{a=1}^4 P_{\gamma}(k_a) .
\end{align}
We have also calculated this trispectrum directly from the Hamiltonian using the in-in formalism and obtained the same result.\footnote{Since no time derivatives appear at cubic order, the quartic-order Hamiltonian can be written as minus the quartic-order Lagrangian.} We note that the IR-divergent contributions that appear in the above expressions for contact diagram time integrals drop out when we take the real part of the wavefunction.\footnote{Indeed in those expressions the divergent parts have real coefficients which are then multiplied by a factor of $-i$ coming from the Feynman rules.}
\vskip4pt
We can perform several consistency checks on the result~\eqref{final}.
On general grounds, we expect that tree-level wavefunction coefficients of massless scalars and gravitons in de Sitter space are rational functions of the external energies $ \{k_{1},k_{2},k_{3},k_{4}\}$ and internal energies $ \{s,t,u\}$ \cite{Goodhew:2022ayb} (see also \cite{Anninos:2014lwa} and the connection to holographic renormalization \cite{Skenderis:2002wp}) with poles at vanishing partial and total energy configurations. Therefore, $ \psi_{4} $ should only have poles in the variables $k_{T}$ and $ E_{L,R}^{(s,t,u)}$, which is indeed the case for our result. Moreover, adapting~\eqref{eq:pole-order} to this case, we expect that the wavefunction should scale as
\begin{equation}
\lim_{k_{T}\to0}\psi_{4}= \mathcal{O}(k_T^{-3}), \qquad\quad \lim_{E_{L,R}\to 0}\psi_{4}= \mathcal{O}(E_{L,R}^{-2}) ,
\end{equation}
in the vicinity of its singularities,
which indeed agrees with~\eqref{final}. It also has the correct amplitude limit on the leading $k_T$ pole.
\vskip4pt
Given the result~\eqref{final}, it is natural to try to cast it in a simpler form. For example,
by an appropriate use of partial fractions, one should always be able to write the result in such a way that every term has at most two poles, e.g., $ \{k_{T},E_{L}\} $ or $ \{k_{T},E_{R}\} $ or $ \{E_{L},E_{R}\}$. This re-writing comes at the cost of introducing spurious poles in the denominator and so, without further input, the resulting expressions are not particularly illuminating. It is also natural to eliminate polarisation vectors in favour of cosmological spinor helicity variables~\cite{Maldacena:2011nz}, but this also does not provide a drastic simplification (in contrast to the analogous story in flat space), essentially because helicity scaling does not completely fix the structure of spinor brackets.
\subsection{Lifting from flat space to de Sitter space}
\label{ss:dSgggglift}
We would now like to reproduce the de Sitter four graviton wavefunction~\eqref{eq:4gWFdS} using the same heuristic method employed in Section~\ref{ss:dSs4} for the massless scalar four-point function. We take as our starting point the four-graviton wavefunction in flat space and then transform it to its de Sitter counterpart.
The four-point graviton wavefunction coefficient in flat space is
\begin{equation}
\label{eq:flatggggcorr2}
\begin{aligned}
\psi_{\gamma^4}^{{\rm flat}} = &~\psi^{(s)\,\rm flat}_{\phi^4}(\epsilon_1\cdot \epsilon_2)^2(\epsilon_3\cdot\epsilon_4)^2+\frac{1}{k_TE_LE_R}\bigg[2{\cal P}^{(s)}_1(\epsilon_1\cdot \epsilon_2)(\epsilon_3\cdot\epsilon_4) W_s+4W_s^2\bigg]\\
&+\psi^{(t)\,\rm flat}_{\phi^4}(\epsilon_1\cdot \epsilon_3)^2(\epsilon_2\cdot\epsilon_4)^2+\frac{1}{k_TE^{(t)}_LE^{(t)}_R}\bigg[2{\cal P}^{(t)}_1(\epsilon_1\cdot \epsilon_3)(\epsilon_2\cdot\epsilon_4) W_t+4W_t^2\bigg]
\\
&+\psi^{(u)\,\rm flat}_{\phi^4}(\epsilon_1\cdot \epsilon_4)^2(\epsilon_3\cdot\epsilon_2)^2+\frac{1}{k_TE^{(u)}_LE^{(u)}_R}\bigg[2{\cal P}^{(u)}_1(\epsilon_1\cdot \epsilon_4)(\epsilon_3\cdot\epsilon_2) W_u+4W_u^2\bigg]
\\
&+\frac{1}{k_T}\bigg[\Big(-2(k_{12}k_{34}+s^2)+\lambda k_T^2\Big)(\epsilon_1\cdot \epsilon_3)(\epsilon_1\cdot \epsilon_4)(\epsilon_2\cdot \epsilon_3)(\epsilon_2\cdot \epsilon_4)\\
&~~~~~~~~
+\Big(-2(k_{13}k_{24}+t^2)+\lambda k_T^2\Big)(\epsilon_1\cdot \epsilon_2)(\epsilon_1\cdot \epsilon_4)(\epsilon_3\cdot \epsilon_2)(\epsilon_3\cdot \epsilon_4)\\
&~~~~~~~~
+\Big(-2(k_{14}k_{23}+u^2)+\lambda k_T^2\Big)(\epsilon_1\cdot \epsilon_3)(\epsilon_1\cdot \epsilon_2)(\epsilon_4\cdot \epsilon_3)(\epsilon_4\cdot \epsilon_2)\\
&~~~~~~~~
-2 W_s W_s^{(c)}-2 W_t W_t^{(c)}-2 W_u W_u^{(c)}\bigg]
\,,
\end{aligned}
\end{equation}
where $W_s$ and $W_s^{(c)}$ are defined as in~\eqref{eq:Wsdef} and~\eqref{eq:Wcdef}, and we have defined
\begin{equation}
{\cal P}_1^{(s)} = s^2 \Pi_{1,1}^{(s)} - E_{L}E_{R} \Pi_{1,0}^{(s)}\,,
\end{equation}
where the $t$ and $u$-channel versions can be obtained by permutation.
Additionally, we note that the flat space scalar wavefunction of massless scalars exchanging a graviton $\psi^{(s)\,\rm flat}_{\phi^4}$~\eqref{eq:flatspacespin2exc} enters into this formula, along with its $t$ and $u$-channel permutations.
\vskip4pt
As in the scalar case, the wavefunction~\eqref{eq:flatggggcorr2} can itself be bootstrapped by requiring that it has the right singularities and residues, along with requiring that, when we suitably average over polarizations, it reduces to the massless scalar result (which is why $\psi^{{\rm flat}}_{\phi^4}$ appears multiplying some polarization structures).
Note that there is an inherent ambiguity in that these requirements do not uniquely specify~\eqref{eq:flatggggcorr2}. Indeed, there is a free parameter, $\lambda$.\footnote{We could also introduce another free parameter $\tilde{\lambda}$ by multiplying $(\epsilon_1\cdot \epsilon_2)^2(\epsilon_3\cdot\epsilon_4)^2$ (and the other two permutations) by $\tilde{\lambda} k_{T}$. This would cause the scalar correlator multiplying this tensor structure to not vanish in the soft limit, so we do not include this term here, but we will encounter this possibility in Section~\ref{s:dS}.} In position space, the contribution proportional to this parameter is a contact term (localized at coincident points), and the precise contact terms required to appear in the wavefunction are fixed by our choice of the Ward identity that we require~\eqref{eq:flatggggcorr2} to satisfy~\cite{Baumann:2020dch}.\footnote{From the bulk point of view, different choices of contact terms appearing in the Ward identity correspond to different choices of field variables. For example, using the canonically normalized variable $g_{\mu\nu} = \eta_{\mu\nu}+2\gamma_{\mu\nu}/M_{\rm Pl}$, one finds $\lambda =10/3$.}
\vskip4pt
We now want to contemplate lifting~\eqref{eq:flatggggcorr2} to de Sitter space. To start, we compute the cut of the flat space graviton four-point function,
\begin{align}
\nonumber
\psi_{\gamma^4}^{{\rm flat}}(\{ k\})+\psi_{\gamma^4}^{\ast\,{\rm flat}}(\{- k\}) = &-\frac{2s}{(k_{12}^2-s)(k_{34}^2-s^2)} \bigg[\frac{1}{6}s^4\Pi_{2,2}^{(s)}(\epsilon_1\cdot \epsilon_2)^2(\epsilon_3\cdot\epsilon_4)^2\\
&+2s^2\Pi_{1,1}^{(s)}(\epsilon_1\cdot \epsilon_2)(\epsilon_3\cdot\epsilon_4) W_s+4W_s^2\bigg]+t+u~{\rm channels}\,.
\label{eq:g4flatcut}
\end{align}
We can then compare it with the corresponding de Sitter formula, which we can compute using the graviton three-point function in de Sitter space~\eqref{GravitonCubicWF}.
With this, we compute the cut of the graviton four-point function to be
\begin{align}
\nonumber
\psi_{\gamma^4}(\{ k\})+\psi^*_{\gamma^4}(\{- k\})=\,&\frac{2s^3(k_{12}^2+2k_1k_2-s^2)(k_{34}^2+2k_3k_4-s^2)}{(k_{12}^2-s^2)^2(k_{34}^2-s^2)^2}\bigg[\frac{1}{6}s^4\Pi_{2,2}^{(s)}(\epsilon_1\cdot \epsilon_2)^2(\epsilon_3\cdot\epsilon_4)^2\\
&+2s^2\Pi_{1,1}^{(s)}(\epsilon_1\cdot \epsilon_2)(\epsilon_3\cdot\epsilon_4) W_s+4W_s^2\bigg]+t+u~{\rm channels}\,.
\label{eq:g4dscut}
\end{align}
In both~\eqref{eq:g4flatcut} and~\eqref{eq:g4dscut}, we have suppressed the contributions from the $t$ and $u$ exchange channels, but it will be straightforward to reintroduce them via permutation once we have constructed the $s$-channel contribution. It is simplest to start with the function multiplying the $(\epsilon_1\cdot \epsilon_2)^2(\epsilon_3\cdot\epsilon_4)^2$ tensor structure. In flat space, this factor is the four-point function of massless scalars exchanging a graviton, $\psi_{\phi^4}^{(s)\, \rm flat}$. The natural thing to do is to promote this to its de Sitter space version, which we constructed in Section~\ref{ss:dSs4}. Next, we consider the second line of~\eqref{eq:g4flatcut}; we want to transform the cut of the form factor into its de Sitter counterpart. These are the same energy factors that appeared in the scalar four-point function, so if we act with the operator ${\cal D}_s$, we generate
\begin{equation}
{\cal D}_s\frac{1}{k_T E_L E_R} = f_{(2,2)}^{(s)}\,.
\end{equation}
Finally, in order for the total energy singularity to be Lorentz invariant, we have to include the helicity-0 polarization sum multiplying the tensor structure $(\epsilon_1\cdot \epsilon_2)(\epsilon_3\cdot\epsilon_4) W_s$. In flat space, this is part of ${\cal P}_1^{(s)}$ and the form factor for this piece is (minus) the $s\to0$ limit of the helicity-$1$ form factor, so we replace this form factor in de Sitter by $ -f_{(2,2)}^{(s)}\Big\rvert_{s=0} = -f_{(2,0)}^{(s)}$. Putting this all together, we obtain the $s$-channel contribution to the graviton four-point function:
\begin{equation}
\begin{aligned}
\psi_{\gamma^4}^{(s)} =
~\psi_{\phi^4}^{(s)}(\epsilon_1\cdot \epsilon_2)^2(\epsilon_3\cdot\epsilon_4)^2&+f^{(s)}_{(2,2)}\bigg[2s^2\Pi_{1,1}(\epsilon_1\cdot \epsilon_2)(\epsilon_3\cdot\epsilon_4) W_s+4W_s^2\bigg]\\
&-f^{(s)}_{(2,0)}\bigg[2\Pi_{1,0}(\epsilon_1\cdot \epsilon_2)(\epsilon_3\cdot\epsilon_4) W_s\bigg]\,.
\end{aligned}
\label{eq:schangggglift}
\end{equation}
We can then check that this reproduces the exchange part of~\eqref{final}. We can obtain the $t$ and $u$-channel exchange contributions by permuting momenta.
\vskip4pt
Finally, we turn to the contribution from contact interactions. We notice that there are two different form factors appearing in~\eqref{eq:flatggggcorr2}:
\begin{align}
g_c^{(s)\,{\rm flat}} &= -\frac{2(k_{12}k_{34}+s^2)}{k_T}+\lambda k_T\,,\\
f_c^{\rm flat} &= \frac{1}{k_T}\,.
\end{align}
We can transmute these to their de Sitter counterparts by acting with~\eqref{eq:contactlift}:
\begin{align}
-{\cal D}_{k_T}g_c^{(s)\,{\rm flat}} &= g_{\rm cont}^{(s)}+\frac{2+\lambda}{3}(k_1^3+k_2^3+k_3^3+k_4^3)\,,\\
{\cal D}_{k_T}f_c^{\rm flat} &= f_{\rm cont}\,.
\end{align}
Note that in this case we do not have to correct the result of the lifting procedure to satisfy the MLT, because~\eqref{eq:schangggglift} already satisfies it by itself (this is a consequence of the fact that we have directly used the de Sitter massless scalar four-point function as a building block, which itself satisfies the MLT). We can then construct the $s$-channel contribution to the graviton four-point function
\begin{equation}
\label{eq:liftedgggg}
\begin{aligned}
\psi_{\gamma^4}^{(s)} =
~&\psi_{\phi^4}^{(s)}(\epsilon_1\cdot \epsilon_2)^2(\epsilon_3\cdot\epsilon_4)^2+f^{(s)}_{(2,2)}\bigg[2s^2\Pi_{1,1}(\epsilon_1\cdot \epsilon_2)(\epsilon_3\cdot\epsilon_4) W_s+4W_s^2\bigg]\\
&-f^{(s)}_{(2,0)}\bigg[2\Pi_{1,0}(\epsilon_1\cdot \epsilon_2)(\epsilon_3\cdot\epsilon_4) W_s\bigg]+g_{\rm cont}^{(s)}(\epsilon_1\cdot \epsilon_4)(\epsilon_2\cdot \epsilon_3)(\epsilon_2\cdot \epsilon_4)
-2 f_{\rm cont}W_s W_s^{(c)}\\
&+\frac{2+\lambda}{3}(k_1^3+k_2^3+k_3^3+k_4^3)(\epsilon_1\cdot \epsilon_4)(\epsilon_2\cdot \epsilon_3)(\epsilon_2\cdot \epsilon_4)\,.
\end{aligned}
\end{equation}
We see that this agrees precisely with~\eqref{final} if we fix $\lambda=-2$. It is not surprising that there is a free parameter, because it corresponds to the ambiguity in our choice of bulk field variables. By summing this along with the $t$ and $u$ channels, we produce the graviton four-point wavefunction.
\vskip4pt
This derivation has the virtue of relative simplicity. Nevertheless, one could question whether the final result is unique. In the following section, we systematically construct the most general wavefunction that has the correct singularities and which satisfies the MLT, reproducing both~\eqref{eq:liftedgggg} and the result of the bulk calculation.
\subsection{The boostless bootstrap} \label{s:dS}
We now return to our third method and follow the three-step procedure introduced in Section~\ref{ss:blesss4} to bootstrap the graviton four-point function. Recall that the strategy is as follows: in step 1 we use the Cosmological Optical Theorem (COT)~\cite{COT} and energy shifts~\cite{MLT,Baumann:2021fxj} to fix all partial-energy poles, in step 2 we fix the leading total-energy pole by demanding that we recover the correct scattering amplitude, and in step 3 we use the MLT~\cite{MLT} to fix the subleading total-energy poles. Throughout we use scale invariance, by demanding that the wavefunction coefficient scales as $k^3$, and Bose symmetry. As in Section~\ref{ss:blesss4}, we directly construct the part of the wavefunction coefficient with the symmetries of an $s$-channel diagram, from which we can get the full wavefunction coefficient by adding the other two permutations of momenta.
\subsubsection{Unitarity and partial-energy poles}
We start by fixing the partial-energy poles, for which we need the graviton three-point function~\eqref{GravitonCubicWF}. The trimmed part of this wavefunction coefficient coincides with that of two massless scalars and a graviton. As a consequence, the structure of the partial-energy poles for the graviton four-point function is the same as for massless scalars exchanging a graviton---we already computed this using partial-energy shifts in Section~\ref{ss:blesss4}, with the result given by~\eqref{fCut}. To complete step 1 we need to compute the tensor structure that arises from an $s$-channel exchange diagram. We can find this from the cubic wavefunction coefficient by using the COT, as in \eqref{eq:g4dscut}.
\vskip4pt
After step 1 our wavefunction coefficient is therefore
\begin{align}\label{4GravitonsStep1}
\psi_{\gamma^4}^{(s)} = \frac{M_{\rm Pl}^2}{16 H^{2}} \left[\frac{s^4}{12} \Pi^{(s)}_{2,2} (\epsilon_1 \! \cdot \! \epsilon_2)^2(\epsilon_3 \! \cdot \! \epsilon_4)^2 + s^2 \Pi^{(s)}_{1,1} (\epsilon_1 \! \cdot \! \epsilon_2)(\epsilon_3 \! \cdot \! \epsilon_4) W_{s} + 2 W_{s}^2 \right] f_{\text{cut}} + \ldots,
\end{align}
where here $\ldots$ denotes terms that must be regular in the limits $E_{L,R} \rightarrow 0$. The partial-energy poles match what we found in Section \ref{ss:lagrangian}.
\subsubsection{The amplitude limit and leading total-energy pole}
We now fix the leading total-energy pole by demanding that we recover the correct scattering amplitude. The amplitude is a sum of the $S,T$, and $U$ channel contributions along with a contact contribution:\footnote{This amplitude has an extra overall factor of $(M_{\rm Pl}/2)^4$ compared to the amplitude for canonically normalised graviton perturbations. This is to match the normalization used for $\gamma_{ij}$.}
\begin{equation}
\mathcal{A}_{4} = \frac{M_{\rm Pl}^2}{16}\left( \mathcal{A}_{4}^{(S)}+\mathcal{A}_{4}^{(T)}+\mathcal{A}_{4}^{(T)}+\mathcal{A}_{4}^{(c)}\right).
\end{equation}
The $S$-channel contribution is given by\footnote{As before, the split into different channels is somewhat arbitrary, but the choice of split does not affect the final answer.}
\begin{equation} \label{SchannAmp}
\mathcal{A}_{4}^{(S)} = \frac{T U}{S} (\varepsilon_{1} \cdot \varepsilon_{2})^2(\varepsilon_{3} \cdot \varepsilon_{4})^2 - \frac{2}{S}(U-T) (\varepsilon_{1} \cdot \varepsilon_{2})(\epsilon_{3} \cdot \varepsilon_{4})\mathcal{W}_{S} - \frac{4}{S}\mathcal{W}_{S}^2 ,
\end{equation}
with the $T$ and $U$ channels obtained from this expression, while the contact contribution is
\begin{equation} \label{ContactAmp}
\mathcal{A}_{4}^{(c)} = \left[2S(\varepsilon_{1} \cdot \varepsilon_{3})(\varepsilon_{1} \cdot \varepsilon_{4})(\varepsilon_{2} \cdot \varepsilon_{3})(\varepsilon_{2} \cdot \varepsilon_{4})-2\mathcal{W}_{S}\mathcal{W}_{S}^{(c)}\right] + 2 ~ \text{perms}.
\end{equation}
We have defined the following combinations of polarization vectors and momenta
\begin{align}
\mathcal{W}_S & = (\varepsilon_1 \! \cdot \! \varepsilon_2)\big[ (\varepsilon_3 \! \cdot \! k_1)(\varepsilon_4 \! \cdot \! k_2)-(\varepsilon_3 \! \cdot \! k_2)(\varepsilon_4 \! \cdot \! k_1) \big]+(\varepsilon_3 \! \cdot \! \varepsilon_4)\big[ (\varepsilon_1 \! \cdot \! k_3)(\varepsilon_2 \! \cdot \! k_4)-(\varepsilon_1 \! \cdot \! k_4)(\varepsilon_2 \! \cdot \! k_3) \big] \nonumber \\
& \hspace{13pt}+ \big[(\varepsilon_1 \! \cdot \! k_2)\, \varepsilon_2-(\varepsilon_2 \! \cdot \! k_1)\, \varepsilon_1\big]\cdot \big[(\varepsilon_3 \! \cdot \! k_4)\, \varepsilon_4-(\varepsilon_4 \! \cdot \! k_3)\, \varepsilon_3\big], \\
\mathcal{W}_S^{(c)} & =(\varepsilon_1 \! \cdot \! \varepsilon_4)(\varepsilon_2 \! \cdot \! \varepsilon_3)-(\varepsilon_1 \! \cdot \! \varepsilon_3)(\varepsilon_2 \! \cdot \! \varepsilon_4).
\end{align}
Note that these are conceptually slightly different from $W_{s}$ and $W_{s}^{(c)}$, since we write the amplitude in a manifestly Lorentz invariant form using the four-vector $\varepsilon_{\mu}$. We could write each contribution to this amplitude in axial gauge (with $h_{0 \mu} = 0$) using the three-dimensional polarisation vector $\epsilon_{i}$ rather than $\varepsilon_{\mu}$, and indeed when we match to the amplitude we do so channel-by-channel, so in practice we are matching a de Sitter Feynman diagram for the wavefunction to a flat-space one for the amplitude with both computed in axial gauge. Of course the final answer in both cases is gauge invariant, once we include all diagrams, but thinking in terms of axial gauge gives us a transparent way to track tensor structures (since they are the same in both spacetimes).
\vskip4pt
As in Section~\ref{ss:blesss4}, we have written the $S$-channel amplitude using $S,\ T$, and $U$ such that it has the desired symmetries. When we match our wavefunction coefficient to the amplitude on the leading total-energy pole, we will eliminate $U$ from the amplitude and $u$ from the wavefunction coefficient so that the variables we are using are independent.
We now demand that our wavefunction coefficient satisfies~\eqref{AmplitudeLimit} and we treat each different tensor structure independently.
\vskip4pt
The three tensor structures of~\eqref{SchannAmp} are contained in~\eqref{4GravitonsStep1} so we first make sure each energy-dependent coefficient yields the correct amplitude. For the $(\epsilon_{1} \cdot \epsilon_{2})^2(\epsilon_{3} \cdot \epsilon_{4})^2$ structure we notice that, up to an overall factor, this part of the amplitude matches that of minimally coupled massless scalars, cf.,~\eqref{MasslessScalarsAmp}, while the corresponding part of~\eqref{4GravitonsStep1} also matches---up to the same factor---what we found after step 1 for massless scalars. We can therefore use our result from Section~\ref{ss:blesss4}. Up to subleading total-energy poles, which we will fix in the next section, we therefore take
\begin{equation}
\frac{f_{\text{cut}}}{12} s^4 \Pi^{(s)}_{2,2} \rightarrow \frac{1}{6} \left[ \frac{f_{\text{cut}}}{2} s^4 \Pi^{(s)}_{2,2} - \frac{2 \mathbb{e}_{4}}{k_{T}^3}(s^2 \Pi^{(s)}_{2,1} - (k_{12}k_{34}+s^2)\Pi^{(s)}_{2,0}) - \frac{2 \mathbb{e}_{4}(k_{12}k_{34}+s^2)}{k_{T}^3} \right]
\end{equation}
in~\eqref{4GravitonsStep1}.
\vskip4pt
Next we consider the $(\epsilon_{1} \cdot \epsilon_{2})(\epsilon_{3} \cdot \epsilon_{4}) W_{s}$ structure. If we compare the amplitude limit from~\eqref{4GravitonsStep1} to the correct amplitude we find that they differ by a single term which has an overall factor of $s^{-2}$, and which is independent of $t^2$. We therefore need to correct this part of~\eqref{4GravitonsStep1} and our ansatz is completely fixed up to an overall factor: first, we need to include an overall factor of $\mathbb{e}_{4}/(s^2 k_{T}^3)$. The remaining dependence must scale as $\sim k^2$, by scale invariance, but it also needs to be anti-symmetric in ${\bf k}_{1} \leftrightarrow {\bf k}_{2}$ and ${\bf k}_{3} \leftrightarrow {\bf k}_{4}$, since the tensor structure in this case is anti-symmetric. We also need to cancel the $s \rightarrow 0$ pole for physical momenta. All of these considerations together imply that we need to include an overall factor of $(k_{1}-k_{2})(k_{3}-k_{4})$. If we fix the overall coefficient by matching to the amplitude limit we need to take
\begin{equation}
f_{\text{cut}} s^2 \Pi_{1,1}^{(s)} \rightarrow f_{\text{cut}} s^2 \Pi_{1,1}^{(s)} - \frac{4 \mathbb{e}_{4}}{ k_{T}^3} \Pi_{1,0}^{(s)}
\end{equation}
in~\eqref{4GravitonsStep1}. Lastly, for the $S$-channel exchange we have the $W_{s}^2$ tensor structure and in this case we see that the amplitude limit from~\eqref{4GravitonsStep1} is already the correct one.
\vskip4pt
We now move to the contact interaction terms and we again concentrate on the terms with $S$-channel symmetries. These are the two written in~\eqref{ContactAmp}, and we see that these two tensor structures do not appear in~\eqref{4GravitonsStep1}. First consider the structure $(\epsilon_{1} \cdot \epsilon_{3})(\epsilon_{1} \cdot \epsilon_{4})(\epsilon_{2} \cdot \epsilon_{3})(\epsilon_{2} \cdot \epsilon_{4})$. Our energy-dependent ansatz that needs to multiply this structure should contain an overall factor of $\mathbb{e}_{4}/k_{T}^3$ along with another factor, scaling as $\sim k^2$, which on the total energy pole reduces to $S$. This means that we need to include both $s^{0}$ and $s^{2}$ terms in the ansatz. By fixing the ansatz up to subleading total-energy poles, we make the following addition to~\eqref{4GravitonsStep1}:
\begin{equation}
\Delta_{1} \psi_{\gamma^4}^{(s)} = -\frac{M_{\rm Pl}^2}{16 H^2}\frac{4 \mathbb{e}_{4}}{k_{T}^3}(k_{12}k_{34}+s^2) (\epsilon_{1} \cdot \epsilon_{3})(\epsilon_{1} \cdot \epsilon_{4})(\epsilon_{2} \cdot \epsilon_{3})(\epsilon_{2} \cdot \epsilon_{4}).
\end{equation}
Finally, we need to add the $W_{s} W_{s}^{(c)}$ tensor structure to the wavefunction to match the second tensor structure in~\eqref{ContactAmp}. This structure already scales as $\sim k^2$ and once we include the overall factor of $\mathbb{e}_{4}/k_{T}^3$ the only freedom left is in the overall coefficient. Fixing this by matching to the amplitude means that we add to~\eqref{4GravitonsStep1}:
\begin{equation}
\Delta_{2} \psi_{\gamma^4}^{(s)} = - \frac{M_{\rm Pl}^2}{16 H^2} \frac{4 \mathbb{e}_{4}}{k_{T}^3} W_s W_s^{(c)}.
\end{equation}
The full expression after fixing the total-energy poles and matching to the amplitude is therefore
\begin{align}\label{4GravitonsStep2}
\frac{16 H^2}{M_{\rm Pl}^2}\psi_{\gamma^4}^{(s)} &= \left[ \frac{f_{\text{cut}}}{12} s^4 \Pi^{(s)}_{2,2} - \frac{ \mathbb{e}_{4}}{3 k_{T}^3}(s^2 \Pi^{(s)}_{2,1} - (k_{12}k_{34}+s^2)\Pi^{(s)}_{2,0}) - \frac{ \mathbb{e}_{4}(k_{12}k_{34}+s^2)}{3 k_{T}^3} \right] (\epsilon_1 \! \cdot \! \epsilon_2)^2(\epsilon_3 \! \cdot \! \epsilon_4)^2 \nonumber \\ &+ \left[ f_{\text{cut}} s^2 \Pi_{1,1}^{(s)} - \frac{4 \mathbb{e}_{4}}{ k_{T}^3} \Pi_{1,0}^{(s)} \right] (\epsilon_1 \! \cdot \! \epsilon_2)(\epsilon_3 \! \cdot \! \epsilon_4) W_{s}+ 2 f_{\text{cut}} W_{s}^2 \nonumber \\
& - \frac{4 \mathbb{e}_{4}}{k_{T}^3}(k_{12}k_{34}+s^2) (\epsilon_{1} \cdot \epsilon_{3})(\epsilon_{1} \cdot \epsilon_{4})(\epsilon_{2} \cdot \epsilon_{3})(\epsilon_{2} \cdot \epsilon_{4}) - \frac{4 \mathbb{e}_{4}}{k_{T}^3} W_s W_s^{(c)} + \ldots.
\end{align}
Here $\ldots$ indicates additional terms that are regular as $E_{L,R} \rightarrow 0$ and subleading in the total-energy limit. We note that the amplitude limit has done some heavy lifting for us by tying together the contact and exchange contributions, as required by diffeomorphism invariance.
\subsubsection{Manifestly local test and subleading total-energy poles}
We now turn to the third and final step, where we use the MLT to fix the subleading total-energy poles. Recall that the MLT requires the wavefunction coefficient to satisfy~\eqref{MLT}, where we take the derivative with respect to one external energy and set that energy to zero while holding all other variables fixed. The MLT follows from the form of the bulk-to-boundary propagator for the massless graviton, so it should hold for each tensor structure separately. We therefore work with each tensor structure in~\eqref{4GravitonsStep2} separately.
\vskip4pt
We start with the $(\epsilon_1 \cdot \epsilon_2)^2(\epsilon_3 \cdot \epsilon_4)^2$ structure. The coefficient after the first two steps is equivalent to what we found for massless scalars after the first two steps, so the corrections we need to add to satisfy the MLT are the same. In this case we therefore simply promote the energy-dependent coefficient to~\eqref{AfterStep3}, which we recall has three free parameters.
Now, consider the $(\epsilon_1 \cdot \epsilon_2)(\epsilon_3 \cdot \epsilon_4) W_{s}$ structure. We find that the coefficient does not satisfy the MLT. Taking the derivative with respect to $k_{4}$ followed by setting $k_{4} = 0$ yields a function with $s^{-2}$, $s^{0}$ and $s^2$ terms. The coefficient of $s^{-2}$ is a function of $k_{1},k_{2}$ and $k_{3}$ only, the $s$-independent part depends on $k_{1},k_{2},k_{3}$ and $t^{2}$, while the $s^2$ term again depends only on $k_{1},k_{2}$ and $k_{3}$. We therefore need to write down an ansatz which we add to~\eqref{4GravitonsStep2}. Recall that we need to ensure the absence of poles as $s \rightarrow 0$, so whenever we have a $t^{2}$ dependence we also need to include $u^2$, and we take them to appear in the combination $(t^{2}-u^{2})$ since this is how they appear in~\eqref{4GravitonsStep2}. Finally, this tensor structure's energy-dependent coefficient needs to be anti-symmetric under the permutations ${\bf k}_{1} \leftrightarrow {\bf k}_{2}$ and ${\bf k}_{3} \leftrightarrow {\bf k}_{4}$. We therefore add to~\eqref{4GravitonsStep2}:
\begin{align}
\frac{16 H^2}{M_{\rm Pl}^2}\Delta_{3} \psi_{\gamma^4}^{(s)} = \frac{1}{s^2 k_{T}^2}\big[(k_{1}-k_{2})(k_{3}-k_{4}) &\text{Poly}^{(3)}_{3}(k_{12},k_{1}k_{2},k_{34},k_{3}k_{4}) + \left( {\bf k}_{1} \leftrightarrow {\bf k}_{3}, {\bf k}_{2} \leftrightarrow {\bf k}_{4} \right) \nonumber \\ &+ b_{6} s^{2}(t^2-u^2)k_{T}\big] (\epsilon_1 \! \cdot \! \epsilon_2)(\epsilon_3 \! \cdot \! \epsilon_4) W_{s}.
\end{align}
Note that the symmetries of our ansatz do not allow for an independent $s^2$ term.
\vskip4pt
The tensor structure $W_{s}^2 $ is relatively simple. We find that its coefficient does not satisfy the MLT, but the correction that we need to add only needs to depend on the four external energies. Given that the tensor structure scales as $\sim k^4$, the energy-dependence of our ansatz needs to scale as $\sim k^{-1}$ by scale invariance. There is then a unique choice, up to an overall coefficient, that adheres to the pole structure of the full wavefunction coefficient. We therefore add to~\eqref{4GravitonsStep2}:
\begin{equation}
\frac{16 H^2}{M_{\rm Pl}^2}\Delta_{4} \psi_{\gamma^4}^{(s)} = \frac{b_{7}}{k_{T}}W_{s}^2.
\end{equation}
The story for the final two structures proceeds in the same way and we need to add the following two corrections:
\begin{align}
\frac{16 H^2}{M_{\rm Pl}^2}\Delta_{5} \psi_{\gamma^4}^{(s)} = & \frac{1}{k_{T}^2}\big[(\text{Poly}^{(5)}_{5}(k_{12},k_{1}k_{2},k_{34},k_{3}k_{4}) + s^2 \text{Poly}^{(3)}_{4}(k_{12},k_{1}k_{2},k_{34},k_{3}k_{4})) \nonumber \\ & + \left( {\bf k}_{1} \leftrightarrow {\bf k}_{3}, {\bf k}_{2} \leftrightarrow {\bf k}_{4} \right) \big] (\epsilon_{1} \cdot \epsilon_{3})(\epsilon_{1} \cdot \epsilon_{4})(\epsilon_{2} \cdot \epsilon_{3})(\epsilon_{2} \cdot \epsilon_{4}), \\
\frac{16 H^2}{M_{\rm Pl}^2}\Delta_{6} \psi_{\gamma^4}^{(s)} & = \frac{1}{k_{T}^2}[\text{Poly}^{(3)}_{5}(k_{12},k_{1}k_{2},k_{34},k_{3}k_{4}) + \left( {\bf k}_{1} \leftrightarrow {\bf k}_{3}, {\bf k}_{2} \leftrightarrow {\bf k}_{4} \right)] W_s W_s^{(c)}.
\end{align}
Imposing the MLT for each tensor structure, and adding the other two permutations, then yields
\begin{equation}
\psi_{\gamma^4}=\frac{M_{\rm Pl}^2}{16 H^2} \left[ \psi_{\gamma^4}^{(s)}+\psi_{\gamma^4}^{(t)}+\psi_{\gamma^4}^{(u)}\right],
\end{equation}
where
\begin{equation}
\begin{aligned}\label{finalBoot}
\psi_{\gamma^4}^{(s)} & = M_{\rm Pl}^2 H^2 \psi_{\phi^4}^{(s)}(\epsilon_1 \! \cdot \! \epsilon_2)^2(\epsilon_3 \! \cdot \! \epsilon_4)^2+ 2[f_{(2,2)}^{(s)}s^2 \Pi^{(s)}_{1,1} - f_{2,0}^{(s)} \Pi^{(s)}_{1,0}] (\epsilon_1 \! \cdot \! \epsilon_2)(\epsilon_3 \! \cdot \! \epsilon_4) W_s \\ &
\hspace{12pt}+4 f_{(2,2)}^{(s)} W_s^2
+ \left(g_{\rm cont.}^{(s)}+r_{1} \sum_{a=1}^{4} k_{a}^3 \right)(\epsilon_1 \! \cdot \! \epsilon_3)(\epsilon_1 \! \cdot \! \epsilon_4)(\epsilon_2 \! \cdot \! \epsilon_3)(\epsilon_2 \! \cdot \! \epsilon_4)-2 f_{\rm cont.} W_s W_s^{(c)}.
\end{aligned}
\end{equation}
In this expression $\psi_{\phi^4}^{(s)}$ is given by~\eqref{AfterStep3}. We see that this result matches the one derived in Section~\ref{ss:lagrangian} and given in~\eqref{final} up to four free parameters which are not tied to $M_{\rm Pl}$, and which do not have any poles. These are the constants $q_{1},q_{2},q_{3}$ in~\eqref{AfterStep3} and the new constant $r_{1}$. The $q_{1}$ term can be generated by the field redefinition $\gamma_{ij} \mapsto \gamma_{ij} + \frac{H^2 q_{1}}{4}\gamma_{kl}^2 \gamma_{ij}$, while the $r_{1}$ term can be generated by $\gamma_{ij} \mapsto \gamma_{ij} + \frac{H^2 r_{1}}{4} \gamma_{il} \gamma_{lk} \gamma_{kj}$. The other two terms can be shifted by using the same field redefinitions we used in~\eqref{nonlocal1} and~\eqref{nonlocal2}, with the addition of indices for the graviton:
\begin{align}
\gamma_{ij} \mapsto \gamma_{ij} &+ \frac{ H^2 q_{2} }{3} \left[3 \gamma_{ij} \frac{1}{\partial^2}(\gamma_{kl} \partial^2 \gamma_{kl}) + \gamma_{ij} \frac{1}{\partial^2} (\gamma_{kl}'^2) + \gamma'_{ij} \frac{1}{\partial^2}(\gamma_{kl} \gamma'_{kl}) \right], \\
\gamma_{ij} \mapsto \gamma_{ij} &+
\frac{H^2 q_3}{3}\left[-27\partial^2\gamma_{ij}\frac{1}{\partial^4}(\gamma_{kl}\partial^2\gamma_{kl})-12\gamma_{ij}\frac{1}{\partial^4}(\gamma_{kl}\partial^4\gamma_{kl}) \right. \nonumber \\&\hspace{44pt}\left.+6\partial^2\gamma_{ij}\frac{1}{\partial^4}({\gamma'_{kl}}^2) +2\gamma'_{ij}\frac{1}{\partial^4}(\gamma_{kl}\partial^2\gamma'_{kl})-4\gamma_{ij}\frac{1}{\partial^4}(\gamma'_{kl}\partial^2\gamma'_{kl}) \right. \\\nonumber
&\left.
\hspace{44pt}- 10\gamma'_{ij}\frac{1}{\partial^4}(\gamma'_{kl}\partial^2\gamma_{kl}) -30\partial^2\gamma_{ij}\frac{1}{\partial^4}(\partial_m \gamma_{kl}\partial_m\gamma_{kl})-60\partial_m\gamma_{ij}\frac{1}{\partial^4}(\gamma_{kl}\partial^2\partial_m \gamma_{kl})\right].
\end{align}
We could also construct similar non-local field redefinitions that generate energy-dependent coefficients of the $(\epsilon_1 \cdot \epsilon_3)(\epsilon_1 \cdot \epsilon_4)(\epsilon_2 \cdot \epsilon_3)(\epsilon_2 \cdot \epsilon_4)$ tensor structure involving inverse powers of $s$. This would simply require different contractions between the fields in the above field redefinitions. We did not include terms of this form in our ansatz when we imposed the MLT condition since for this tensor structure no inverse powers of $s$ arose after steps 1 and 2. They can however be dealt with similarly and the resulting free parameters can be removed by field redefinitions. If we were to make a particular choice of field variables, e.g., by enforcing a particular representation of the Ward identity or soft theorem, then we could use this to fix some of these free parameters, as in the quartic scalar calculation.
\section{Conclusions}
\label{sec:conclusions}
In this work, we have computed the graviton four-point correlator and quartic wavefunction coefficient at tree level in de Sitter space for pure Einstein gravity. We derived our result in several different ways and found agreement among them. A few natural directions to pursue present themselves:
\begin{itemize}
\item A consistency check of our result is to ensure that it respects the appropriate soft-graviton theorem, e.g., from \cite{Hinterbichler:2013dpa}. We have found it challenging to prove this, for a few reasons. First, the relevant soft theorem in \cite{Hinterbichler:2013dpa} for a correlator with soft and hard tensors is written in components. To compare to our result one needs to make a specific choice of polarizations, such as the helicity basis, and take derivatives of polarization tensors with respect to their spatial momentum argument, which is subtle because of the arbitrariness in defining this dependence and the necessary discontinuities. This problem appears first for the trispectrum and it is not present in the soft theorem for the bispectrum, where all contractions of polarization tensors can be written in terms of the external energies (see, e.g., \cite{Cabass:2021fnw}). Finally, a naive matching of the coefficients of the polarization tensor structures does not seem to satisfy the trispectrum soft theorem and we could not exclude that different structures cancel each other due to dimension-dependent identities. We postpone further investigations to the future.
\item The graviton four-point function should display appropriate invariance under the isometries of de Sitter spacetime. For correlators at the future conformal boundary, this invariance should reduce to the set of conformal (and current conservation) Ward identities. It would be nice to check explicitly that our result indeed satisfies these constraints.
\item In general we expect that in simple and highly constrained theories---such as pure gravity---higher-order observables should be fixed in terms of the lowest-order interaction. For example, in flat space all $n$-particle graviton amplitudes can in principle be derived from the three-graviton amplitude via BCFW recursion relations~\cite{BCFW,Benincasa:2007qj}. Because of this, one might hope to find variables in which the graviton four-point function we compute is not much more complicated than the graviton three-point function. Our final result~\eqref{final} does not yet realize this hope. It would be nice to find a more natural set of variables, analogous to spinor-helicity variables for maximally helicity-violating amplitudes~\cite{Parke:1986gb}. Cosmological spinor-helicity variables~\cite{Maldacena:2011nz}, as used in \cite{Li:2018wkt}, do not seem to achieve this simplification. One possibility is that simplifications can be achieved by using a Mellin space representation \cite{Sleight:2019mgd, Sleight:2019hfp, Sleight:2021iix}; this would also facilitate a comparison to the anti-de Sitter space result \cite{Raju:2012zs}, as reviewed in \cite{ Sleight:2021plv}. It would be interesting to further pursue these directions.
\end{itemize}
We close with a small historical parable. In a truly heroic computation, Parke and Taylor computed the six-gluon scattering amplitude in 1985, and presented it in~\cite{Parke:1985ax}, ending their discussion with the phrase ``...we hope to obtain a simple analytic form for the answer, making our result not only an experimentalist's, but also a theorist's delight." Less than a year later, they found a spectacular such formula~\cite{Parke:1986gb}, which catalyzed the modern amplitudes revolution. We share a similar hope to find a drastic simplification of the formulas presented in this paper. We should be so lucky.
\paragraph{Acknowledgements} We thank Soner Albayrak, Daniel Baumann, Giovanni Cabass, Zongzhe Du, Carlos Duaso Pueyo, Aaron Hillman, Kurt Hinterbichler, Sadra Jazayeri, Savan Kharel, Hayden Lee, Guilherme Pimentel, and Akhil Premkumar for useful discussions. HG is supported jointly by the Science and Technology Facilities Council through a postgraduate studentship and the Cambridge Trust Vice Chancellor's Award. The work of AJ is supported in part by DOE (HEP) Award DE-SC0009924. EP has been supported in part by the research program VIDI with Project No. 680-47-535, which is (partly) financed by the Netherlands Organisation for Scientific Research (NWO). DS is supported by a UKRI Stephen Hawking Fellowship [grant number EP/W005441/1] and a Nottingham Research Fellowship from the University of Nottingham. For the purpose of open access, the authors have applied a CC BY public copyright licence to any Author Accepted Manuscript version arising.
|
\section{Introduction}
The quantum groups provide the key to $q$-deforming the basic structures of physics from the point of view of non-commutative geometry. The theory plays an important role in the conformal field theory, exact soluble models in statistical physics \cite{Baxte1982} and in a wide range of applications, from cosmic strings and black holes to the solid state physics problems \cite{Anyons1992}, \cite{BD1999}, \cite{LSN2006}, \cite{Wilczek1990}. For instance, the $q$-deformed Lie algebras are extended versions of the usual Lie algebras which are beyond the scope of usual Lie algebras \cite{BD1999}. Moreover, it has recently led to the application of the $q$-calculus in the construction of generalized statistical mechanics where non-extensivity properties can arise from the $q$-deformed theory \cite{GT2004}. For some analytical constructions on quantum groups, we refer to \cite{AMR2018}.
Recently, mathematicians have paid much attention to the $q$-calculus \cite{CK2000}, \cite{E2012}, \cite{E2002} and fractional $q$-differential equations \cite{A2014}, \cite{A1969}, \cite{A1966}, \cite{PMS2007}, \cite{SPT2020}, \cite{PMS2009}, \cite{Shaimardan16}. The $q$-calculus (or quantum calculus) can be dated back to the beginning of the twentieth century, to F. H. Jackson and R. D. Carmichael's work \cite{C1912}, \cite{J1908}, \cite{J1910}. There have been some papers dealing with the existence and uniqueness or multiplicity of solutions for nonlinear fractional $q$-difference equations by the use of some well-known fixed point theorems \cite{KF1968}. For some recent developments on the subject, see e.g. \cite{AM2012}, \cite{F2011}, \cite{F2010}, \cite{ZCZ2013} and the references therein.
The notations used in this introduction are explained in Section 2 below. In this paper, we also focus more on the $q$-analogue of Hilfer’s bi-ordinal fractional derivative or composite fractional derivative operator (see Definition \ref{definition3.1}) and we derive sufficient conditions for the existence of a unique solution of the Cauchy-type $q$-fractional problem \eqref{additive3.3}-\eqref{additive3.4}. Moreover, for the proof of this theorem, we prove an equivalence theorem (Theorem \ref{thm3.4}) of independent interest. In the case when $\mu = 0$ this is the generalized Riemann–Liouville fractional $q$-derivative (see \eqref{additive2.1}) and in case when $\mu = 1$ it corresponds to the Caputo fractional $q$-derivative (see \cite{PMS2009}).
The paper is organized as follows: The main results are presented and proved in Section \ref{section3} and the announced examples are given in Section \ref{section4}. In the final section, we present the explicit solution to the Cauchy problem for the linear case, applying the successive-iteration method. In order to facilitate presentations, we include in Section \ref{section2} some necessary preliminaries.
\section{Preliminaries}\label{section2}
First, we recall several concepts of the $q$-calculus that we need in the paper. We assume that $0<q<1$.
Let $\alpha \in \mathbb{R}$. Then a $q$-real number $[\alpha ]_q$ is defined by (\cite{CK2000})
$$
[\alpha ]_{q}:=\frac{1-q^{\alpha }}{1-q},
$$
and the $q$-binomial coefficient $[n]_{q}!$ is defined by (\cite{CK2000})
\begin{equation*}
[n]_{q}!:=\left\{
\begin{array}{l}
{1,\mathrm{\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;
\;if\;}n={0}}, \\
{[1]_{q}\times [2]_{q}\times \cdots \times [n]_{q},
\mathrm{\;if\;}\,n\in \mathbb{N}}.
\end{array}
\right.
\end{equation*}
The $q$-shifted
operations for $n\in\mathbb{N}$ are defined by
\begin{eqnarray*}
(a;q)_0=1, \; (a;q)_n=\prod\limits_{k=0}^{n-1}\left(1-q^ka\right), \; (a;q)_\infty = \lim\limits_{n\rightarrow\infty}(a;q)_n.
\end{eqnarray*}
Moreover,
\begin{eqnarray}\label{additive2.1C}
(a;q)_\alpha=\frac{(a;q)_\infty}{(q^\alpha a;q)_\infty}.
\end{eqnarray}
The $q$-gamma function $\Gamma_q(x)$ is defined by (\cite{CK2000})
\begin{eqnarray}\label{additive2.1D}
\Gamma_q(x)=\frac{(q; q)_\infty }{(q^x; q)_\infty }(1-q)^{1-x}
\end{eqnarray}
for any $x > 0$. This $q$-Gamma function has the following property (\cite{CK2000}):
\begin{eqnarray*}
\Gamma_q(x)[x]_q=\Gamma_q(x+1).
\end{eqnarray*}
The $q$-derivative of a function $f$ with respect to $x$ is defined by (\cite{J1908})
\begin{eqnarray*}
D_{q}f(x)=\frac{f(x)-f(qx)}{x(1-q)},
\end{eqnarray*}
and the $q$-derivatives $D^{n}_q(f(x))$ of higher order are defined inductively as follows (\cite{PMS2007}):
\begin{eqnarray*}
D_q^0(f(x)):=f(x), \;\;\; D^{n}_q(f(x)):=D_q\left(D_q^{n-1}f(x)\right), (n=1,2,3, \dots).
\end{eqnarray*}
The general $q$-integral (or Jackson integral) is defined (\cite{J1910}) as
\begin{eqnarray*}
\int\limits_a^b f(x)d_{q}x=\int\limits_0^b f(x)d_{q}x-
\int\limits_0^a f(x)d_{q}x,
\end{eqnarray*}
for $0<a<b$, where
\begin{eqnarray*}
\int\limits_0^a f(x)d_{q}x=(1-q)a\sum\limits_{m=0}^\infty q^{m}f(aq^{m}).
\end{eqnarray*}
In the middle of the last century, W. A. Al-Salam \cite{A1969} and R. P. Agarwal \cite{A1966} presented a $q$-analog of the Riemann–Liouville fractional integral with only zero as a lower limit of the $q$-integration and a fractional $q$-derivative. After that, P. M. Rajkovic', S. D. Marinkovic' and M. S. Stankovic' \cite{PMS2007} (see also \cite{PMS2009}) allowed the lower limit of the $q$-integration to be nonzero and reintroduced the Riemann-Liouville $q$-fractional integrals $I^\alpha_{q,a+}f$ of order $\alpha>0$ by
\begin{eqnarray}\label{additive2.1}
\left(I^\alpha_{q,a+}f\right)(x)=\frac{x^{\alpha-1}}{\Gamma_q(\alpha)}\int\limits_a^x(qt/x;q)_{\alpha-1}f(t)d_qt.
\end{eqnarray}
The Riemann-Liouville fractional $q$-derivative $D^\alpha_{q,a+}f$ of order $\alpha>0$ is defined by
\begin{eqnarray*}
\left(D^\alpha_{q,a+}f\right)(x)=\left( D^{[\alpha]}_{q,a+}I^{[\alpha]-\alpha}_{q,a+}f\right)(x),
\end{eqnarray*}
where $[\alpha]$ denotes the smallest integer greater or equal to $\alpha$.
For $\alpha>0$ and $\lambda>-1$ the following identities hold \cite[Lemma 12]{PMS2009}:
\begin{eqnarray}\label{additive2.2}
\left(I^\alpha_{q,a+}t^\lambda(t/a;q)_\lambda\right)(x)=
\frac{\Gamma_q(\lambda+1)}{\Gamma_q(\alpha+\lambda+1)} x^{\alpha+\lambda}(x/a;q)_{\alpha+\lambda}
\end{eqnarray}
and
\begin{eqnarray}\label{additive2.3}
\left(D^\alpha_{q,a+}t^\lambda(t/a;q)_\lambda\right)(x)=
\frac{\Gamma_q(\lambda+1)}{\Gamma_q(\lambda+1-\alpha)} x^{\lambda-\alpha}(x/a;q)_{\lambda-\alpha},\;\;\;\alpha\leq\lambda,
\end{eqnarray}
and
\begin{eqnarray}\label{additive2.3a}
\left(D^\alpha_{q,a+}t^\lambda(t/a;q)_\lambda\right)(x)=0,\;\;\;\alpha>\lambda.
\end{eqnarray}
For all further discussions, we assume that the functions are defined on an interval $[0, b]$ and $b>0$. Moreover, $a \in[0, b]$ is an arbitrary fixed point. In \cite[Subsection 4.3]{AM2012}, M. H. Annaby and Z. S. Mansour presented definitions of spaces of $q$-integrable, $q$-absolutely continuous on $[0, a]$ functions. We also give characterizations of those modified spaces $[a, b]$ which will be used later. For $1\leq p<\infty$ we define the space $L^p_q=L^p_q [a, b]$ by
\begin{eqnarray*}
L^p_q [a, b] :=\left\{f:[a,b]\rightarrow \mathbb{ R}:\left(\int\limits_a^b |f(x)|^pd_qx\right)^\frac{1}{p}<\infty\right\}.
\end{eqnarray*}
Let $\alpha>0$, $\beta>0$ and $1\leq p<\infty$. Then the $q$–fractional integrals have the following
semi-group property (see \cite[Lemma 2.3]{SPT2020}):
\begin{eqnarray}\label{additive2.4}
\left(I^\alpha_{q,a+}I^\beta_{q,a+}f\right)(x) =\left(I^{\alpha+\beta}_{q,a+}f\right)(x),
\end{eqnarray}
for all $x \in[a, b]$ and $f \in L^p_q[a,b]$.
Let $\alpha>\beta > 0$, $1 \leq p<\infty$ and $f \in L^p_q[a,b]$. Then the following (see \cite[Lemma 2.3]{SPT2020})
equalities
\begin{eqnarray}\label{additive2.5}
\left(D^\alpha_{q,a+}I^\alpha_{q,a+}\right)(x)=f(x),\;\;\;\left(D^\beta_{q,a+}I^{\alpha }_{q,a+}f\right)(x)=\left(I^{\alpha-\beta}_{q,a+}f\right)(x),
\end{eqnarray}
hold for all $x\in[a,b]$.
A function $f:[a,b]\rightarrow \mathbb{ R}$ is called $q$-absolutely continuous
if\, $\exists \varphi\in L^1_q[a,b]$ such that
\begin{eqnarray*}
f(x)=f(a)+\int\limits_a^x\varphi(t)d_qt
\end{eqnarray*}
for all $x\in[a,b]$ (see also, \cite[Definition 2.4]{SPT2020}).
The collection of all $q$-absolutely continuous functions on $[a,b]$ is denoted by $AC_q[a,b]$. For $n= {1,2,3,\dots}$ we denote by $AC^n_q[a, b]$ the space of real-valued functions $f(x)$ which have $q$-derivatives up to order $n-1$ on $[a, b]$ such
that $D_q^{n-1}f \in AC_q[a,b]$:
$$
AC^n_q[a, b]:=\left\{f:[a,b]\rightarrow \mathbb{ R}; D_q^{n-1}f \in AC_q[a,b] \right\}.
$$
The function $f:[a,b]\rightarrow \mathbb{ R}$ belongs to $AC^n_q[a, b]$ if the following equality holds (see \cite[Lemma 2.5]{SPT2020}):
\begin{eqnarray*}
f(x) &=& \frac{x^{n-1}}{\Gamma_q\left(n\right)}\int\limits_a^x(qt/x;q)_{n-1}\varphi(t)d_qt+\sum\limits_{k=0}^{n-1}c_kx^k(a/x;q)_k,
\end{eqnarray*}
where $\varphi(x):=D^{n}_qf(x)$ and $c_k=\frac{D^{k}_qf(a)}{\Gamma_q(k)}, k=0,1,2,\dots, n-1$, are arbitrary constants.
Let $f \in L_q^1[a,b]$ and $ \left(I^{n-\alpha}_{q,a+}f\right) \in AC^n_q[a,b]$ with $n=[\alpha], \alpha>0$. Then the following equality holds (see \cite[Lemma 2.5]{SPT2020}):
\begin{eqnarray}\label{additive2.6}
\left(I^\alpha_{q,a+}D_{q,a+}^\alpha f\right)(x)=f(x)-\sum\limits_{k=1}^n\frac{\left(D^{\alpha-k}_{q,a+}f\right)(a)}{\Gamma_q\left(\alpha-k+1\right)}x^{\alpha-k}(a/x;q)_{\alpha-k}.
\end{eqnarray}
Let $\alpha>0$ and $1\leq p<\infty$. Then the fractional integration operator $I^\alpha_{q,a+} $ is bounded in
$L^p_q [a, b]$ (see \cite[Lemma 2.6]{SPT2020}):
\begin{eqnarray}\label{additive2.7}
\|I^\alpha_{q,a+}f\|_{L^p_q [a, b]} &\leq& K\|f\|_{L^p_q [a, b]},
\end{eqnarray}
where $K=\frac{(b-qa)_q^\alpha}{ \Gamma_q(\alpha+1)}$ and $(b-qa)_q^\alpha=b^\alpha(qa/b;q)_\alpha$.
\section{Main results}\label{section3}
Now, we start to present the concept of the bi-ordinal fractional $q$-derivative generalizing the well-known Hilfer’s $q$-derivative (see \cite{H2000}, \cite{H2002} and \cite{T2012}). Bi-ordinal Hilfer derivative was introduced for the first time by Bulavatsky in \cite{B2014} in the case when the fractional order lies between 0 and 1. Later, in \cite{K2021} it was generalized for higher fractional orders. Boundary-value problems for PDEs involving this fractional derivative were considered in recent works \cite{KTR2022} and \cite{T2022}.
Below for the first time, we introduce $q$-analogue of this integral-differential operator together with the key property which allows us to solve the Cauchy problem for the $q$-fractional differential equation.
\begin{definition}\label{definition3.1} We define the bi-ordinal Hilfer fractional $q$-derivative $\mathcal{D}^{(\alpha,\beta)\mu}_{q,a+}$
of orders $\alpha$ and $\beta$, and type $\mu$ with respect to $x$ by
\begin{eqnarray}\label{additive3.1}
\left(\mathcal{D}^{(\alpha,\beta)\mu}_{q,a+}f\right)(x) &:=& \left(I_{q,a+}^{\mu(1-\alpha)}D_q\left(I_{q,a+}^{(1-\mu)(1-\beta)}f\right)\right)(x),
\end{eqnarray}
for $0<\alpha , \beta\leq1$ and $0\leq\mu\leq{1}$.
\end{definition}
In the case when $\mu=0$, the generalized fractional $q$-derivative (\ref{additive3.1}) would correspond to the the Riemann-Liouville fractional $q$-derivative of order $\beta$ (see (\ref{additive2.1})) and in the case when $\mu= 1$ it corresponds to the Caputo fractional $q$-derivative $\left(_cD^\alpha _{q,a+}f\right)(x)$ defined by (see \cite{PMS2009}):
$$
\left(_cD^\alpha _{q,a+}f\right)(x) := \left(I_{q,a+}^{1-\alpha}D_qf\right)(x)=\frac{x^{-\alpha}}{\Gamma_q(1-\alpha)}\int\limits_a^x
\left(qt/x;q\right)_{-\alpha}f(t)d_qt.
$$
\begin{rem}\label{rem3.2} Let $\gamma=\beta+\mu(n-\beta)$ and $\nu=\beta+\mu(\alpha-\beta)$ for $n-1<\alpha,\beta \leq n, n\in \mathbb{N}$ and $0\leq \mu\leq 1$. The generalized fractional $q$-derivative $D_{q,a+}^{(\alpha,\beta)\mu} f$ can be represented as follows:
\begin{eqnarray}\label{additive3.2}
\left(D^{(\alpha,\beta)\mu}_{q,a+}f\right)(x) &:=& \left(I_{q,a+}^{\mu(n-\alpha)}D^n_q\left(I_{q,a+}^{(1-\mu)(n-\beta)}f\right)\right)(x)\nonumber\\
&=&\left(I_{q,a+}^{\gamma-\nu}D^n_q\left(I_{q,a+}^{n-\gamma}f\right)\right)(x)\nonumber\\
&=&\left(I_{q,a+}^{\gamma-\nu}D^{\gamma}_{q,a+} f\right)(x).
\end{eqnarray}
\end{rem}
Next, we study the question of the equivalence between the following Cauchy-type $q$-fractional problem
\begin{eqnarray}\label{additive3.3}
\left(D^{(\alpha,\beta)\mu}_{q,a+}y\right)(x) &=& f\left(x,y(x)\right),\;\;\;n-1<\alpha\leq n; n\in\mathbb{N}, 0\leq \beta\leq 1,
\end{eqnarray}
\begin{eqnarray}\label{additive3.4}
\lim\limits_{x\rightarrow a+}\left(D^k_qI^{(1-\mu)(n-\beta)}_{q,a+}y \right)(x)=\xi_k, \;\;\;\xi_k\in\mathbb{R}, k=1,2,\dots n,
\end{eqnarray}
and the Volterra $q$-integral equation of the second kind:
\begin{eqnarray}\label{additive3.5}
y(x)=\sum\limits_{k=1}^n\frac{\xi_k}{\Gamma_q\left(\gamma-k+1\right)}x^{\gamma-k}(a/x;q)_{\gamma-k}+
\left(I^ \nu_{q,a+}f(\cdot,y(\cdot))\right)(x).
\end{eqnarray}
\begin{thm}\label{thm3.3} We assume $n-1<\alpha,\beta\leq n, n\in\mathbb{N}$, $0\leq \mu\leq 1$, $\gamma=(n-\alpha)(1-\beta)$, $\nu=\beta+\mu(\alpha-\beta)$. Let $f(\cdot, \cdot): [a , b] \times \mathbb{R}\rightarrow \mathbb{R}$ be a function such that $f(\cdot, y(\cdot)) \in L_q^1[a, b]$ for all $y \in L_q^1[a, b]$.Then $y$ satisfies the relations (\ref{additive3.3})-(\ref{additive3.4}) if and only if\, $y$ satisfies the integral equation (\ref{additive3.5}).
\end{thm}
\begin{proof} We assume that $y\in L^1_q[a, b]$ satisfies the relations (\ref{additive3.3})-(\ref{additive3.4}). Since $f(x,y(x)) \in L_q^1[a, b]$ we get that $ \left(D^{(\alpha,\beta)\mu}_{q,a+}y\right)\in L^1_q[a, b]$ for $\gamma=\beta+\mu(n-\beta)$, $\nu=\beta+\mu(\alpha-\beta)$ together with $n-1 < \alpha,\beta \leq n, n \in \mathbb{N}$, $0 \leq \mu \leq 1$. According to (\ref{additive2.7}) we obtain that $\left(I^\nu_{q,a+}f\right)\in L^1_q[a,b]$. Applying the integral operator $I^\nu_{q,a+}$ to both sides of (\ref{additive3.3}) we obtain
\begin{eqnarray}\label{additive3.6}
I^ \nu_{q,a+}D^{(\alpha,\beta)\mu}_{q,a+}y (x)=\left(I^ \nu_{q,a+}f(\cdot,y(\cdot))\right)(x).
\end{eqnarray}
From Remark \ref{rem3.2} and (\ref{additive2.4}), (\ref{additive2.6}) it follows that
\begin{eqnarray}\label{additive3.7}
\left(I^ \nu_{q,a+}D^{(\alpha,\beta)\mu}_{q,a+}y \right)(x)&=&\left(I^ \nu_{q,a+}I^{\gamma-\nu}_{q,a+}
D^{\gamma}_{q,a+}y \right)(x)\nonumber\\
&=&\left( I^{\gamma}_{q,a+}
D^{\gamma}_{q,a+}y \right)(x)\nonumber\\
&=& y(x) -\sum\limits_{k=1}^n\frac{\left(D^{\gamma-k}_{q,a+}f\right)(a)}{\Gamma_q\left(\gamma-k+1\right)}x^{\gamma-k}(a/x;q)_{\gamma-k}\nonumber\\
&=&y(x) -\sum\limits_{k=1}^n\frac{\left(D^k_qI^{(1-\mu)(n-\beta)}_{q,a+}f\right)(a)}{\Gamma_q\left(\gamma-k+1\right)}x^{\gamma-k} (a/x;q)_{\gamma-k}.
\end{eqnarray}
According to (\ref{additive3.4}), (\ref{additive3.6}) and (\ref{additive3.7}) we get
the equation (\ref{additive3.5}). The necessity is proved.
Now we prove the sufficiency. Let $y\in L^1_q[a, b]$ satisfy equation (\ref{additive3.5}). Applying the operator $D^{(\alpha,\beta)\mu}_{q,a+}$ to both sides of (\ref{additive3.5}), we have that
\begin{eqnarray}\label{additive3.8}
\left(D^{(\alpha,\beta)\mu}_{q,a+}y\right)(x)&=&\sum\limits_{k=1}^n\frac{\xi_k}{\Gamma_q\left(\gamma-k+1\right)}
\left(D^{(\alpha,\beta)\mu}_{q,a+}t^{\gamma-k}(a/t;q)_{\gamma-k}\right)(x)\nonumber\\
&+& D^{(\alpha,\beta)\mu}_{q,a+}\left(I^ \nu_{q,a+}f(\cdot,y(\cdot))\right)(x).
\end{eqnarray}
From (\ref{additive2.3a}) it follows that
\begin{eqnarray}\label{additive3.9}
\left(D^{\gamma}_{q,a+}t^{\gamma-k}(a/t;q)_{\gamma-k}\right)(x)=0,\;\;\;1\leq k\leq n.
\end{eqnarray}
By Remark \ref{rem3.2} and (\ref{additive2.4}), (\ref{additive2.5}) and (\ref{additive3.9}), we find that
\begin{eqnarray*}
\left(D^{(\alpha,\beta)\mu}_{q,a+}t^{k-\gamma}(a/t;q)_{k-\gamma}\right)(x)&=&\left(I^{\gamma-\nu}_{q,a+}
D^{\gamma}_{q,a+}t^{\gamma-k}(a/t;q)_{\gamma-k}\right)(x)=0
\end{eqnarray*}
and using (\ref{additive3.2}) we get that
\begin{eqnarray*}
\left(D^{(\alpha,\beta)\mu}_{q,a+}I^\nu_{q,a+}f\right)(x)&=&\left( I_{q,a+}^{\gamma-\nu}D^{\gamma}_qI^\nu_{q,a+}f(\cdot,y(\cdot))\right)(x)\\
&=&\left(I_{q,a+}^{\gamma-\nu}I^{\nu-\gamma}_{q,a+}f(\cdot,y(\cdot))\right)(x)\\
&=&f(x,y(x)).
\end{eqnarray*}
Therefore, we can rewrite (\ref{additive3.8}) in the form
\begin{eqnarray*}
\left(D^{(\alpha,\beta)\mu}_{q,a+}y\right)(x)&=& f(x,y(x)).
\end{eqnarray*}
Finally, we will show that the initial condition of (\ref{additive3.4}) also holds. For this aim, we apply the operator $I^{(1-\mu)(n-\beta)}_{q,a+}$ to both sides of (\ref{additive3.5}) and using (\ref{additive2.2}), (\ref{additive2.4}), we conclude that
\begin{eqnarray}\label{additive3.10}
\left(I^{(1-\mu)(n-\beta)}_{q,a+}y\right)(x)&=&\sum\limits_{k=1}^n\frac{\xi_k}{\Gamma_q\left(\gamma-k+1\right)}
\left(I^{(1-\mu)(n-\beta)}_{q,a+}t^{\gamma-k}(a/t;q)_{\gamma-k}\right)(x)\nonumber\\
&+& \left[I^{(1-\mu)(n-\beta)}_{q,a+}I^\nu_{q,a+}f(\cdot,y(\cdot))\right](x)\nonumber\\
&=&\sum\limits_{k=1}^n\frac{\xi_k}{[n-k]_q!} x^{n-k}(a/x;q)_{n-k}+ \left(I^{\mu(\alpha-n)-n}_{q,a+}f(\cdot,y(\cdot))\right)(x) .
\end{eqnarray}
Let $0\leq m\leq n-1$. Then, we apply the operator $D^m_q$ to both sides of (\ref{additive3.10}) by using (\ref{additive2.3}) and (\ref{additive2.6}), to obtain that
\begin{eqnarray}\label{additive3.11}
\left(D^m_qI^{(1-\mu)(n-\beta)}_{q,a+}y\right)(x)&=&\sum\limits_{k=1}^n\frac{\xi_k}{[n-k]_q!} D^m_q\left[x^{n-k}(a/x;q)_{n-k}\right]+\left( D^m_qI^{\mu(\alpha-n)-n}_{q,a+}f(\cdot,y(\cdot))\right)(x)\nonumber\\
&=&\sum\limits_{k=1}^n\frac{\xi_k}{[n-k-m]_q!} x^{n-k-m}(a/x;q)_{n-k-m}\nonumber\\
&+&\left(I^{\mu(\alpha-n)-n-m}_{q,a+}f(\cdot,y(\cdot))\right)(x).
\end{eqnarray}
Taking in (\ref{additive3.11}) the limit $ x \rightarrow a+$, we get the relations in (\ref{additive3.4}). Thus also the sufficiency is proved, which completes the
proof of Theorem \ref{thm3.3}.
\end{proof}
\vspace{0,3cm}
In the next theorem, we give conditions for a unique global solution to the Cauchy-type
problem (\ref{additive3.3})-(\ref{additive3.4}) in the space $L^1_{\alpha,\beta,\mu, q}[a,b]$ defined for $\alpha>0$ by
\begin{eqnarray*}
L^1_{\alpha,\beta,\mu, q}[a,b]:=\left\{y\in L_q^1[a, b]: D^{(\alpha,\beta)\mu}_{q,a+}y \in L_q^1[a, b]\right\}.
\end{eqnarray*}
The proof of the following existence and uniqueness theorem depends heavily on Theorem \ref{thm3.3} and the
Banach's fixed point Theorem (see \cite[Subsection 8.1]{KF1968}).
\begin{thm}\label{thm3.4} We assume $\gamma=\beta+\mu(n-\beta)$ and $\nu=\beta+\mu(\alpha-\beta)$ for $n-1<\alpha,\beta \leq n, n\in \mathbb{N}$ and $0\leq \mu\leq 1$. Let $n-1<\alpha\leq n; n\in\mathbb{N}$, and let $f:
[a,b]\times \mathbb{R} \rightarrow \mathbb{R}$ be a function such that $f(\cdot,v(\cdot))\in L_q^1[a,b]$ for all $v\in L_q^1[a,b]$, and it
satisfies the Lipschitz condition in the following form:
\begin{eqnarray}\label{additive3.12}
\left|f(x,v_1(x))-f(x,v_2(x))\right|\leq A\left|v_1(x)-v_2(x)\right|,
\end{eqnarray}
where $A>0$ does not depend on $x \in[a,b]$ and $v_1,v_2\in L_q^1[a,b]$. Then there exists a unique solution $y\in L^1_{\alpha,\beta,\mu, q}[a,b]$ to the Cauchy-type problem (\ref{additive3.3})-(\ref{additive3.4}).
\end{thm}
\begin{proof} First we show that there exists a unique solution $y \in L_q^1[a, b]$. In view of Theorem \ref{thm3.3} it is sufficient to prove the existence of a unique solution $y \in L_q^1[a, b]$ of the nonlinear Volterra $q$-integral equation (\ref{additive3.5}). We assume that $[a,a_1] \subset [a,b]$ are such that
\begin{eqnarray}\label{additive3.13}
\omega_1:=AK\frac{\left(a_1-qa\right)_q^\alpha}{\Gamma_q\left(\alpha+1\right)}<1,
\end{eqnarray}
where $K$ is as in (\ref{additive2.7}). Obviously, the equation (\ref{additive3.5}) holds on the interval $[a,a_1]$. Consequently, we rewrite this equation in the form
$y = Ty$ with
\begin{eqnarray}\label{additive3.14}
\left(Ty\right)(x):=y_0+\left(I^\nu_{q,a+}f(\cdot,y(\cdot))\right)(x),
\end{eqnarray}
where $y_0:=\sum\limits_{k=1}^n\frac{\xi_k}{\Gamma_q\left(\gamma-k+1\right)}x^{\gamma-k}(a/x;q)_{\gamma-k}$.
If $y\in L^1_q[a,a_1]$, then $Ty\in L^1_q[a,a_1]$. Moreover, for any $y_1,y_2\in L^1_q[a,a_1]$, $f \left(\cdot, y_j(\cdot)\right) \in L^1_q[a,a_1], \;j=1,2$, and using (\ref{additive2.7}), (\ref{additive3.12}) and (\ref{additive3.13}) we have that
\begin{eqnarray*}
\| Ty_1-Ty_2\|_{L_q^1[a,a_1]}&=& \|I^\nu_{q,a+}\left[f(t,y_1(t))-f(t,y_2(t))\right]\|_{L_q^1[a,a_1]}\\
&\leq& A\|I^\nu_{q,a+} \left[y_1-y_2\right|\|_{L_q^1[a,a_1]}\\
&\leq& \omega_1\| y_1-y_2 \|_{L_q^1[a,a_1]},
\end{eqnarray*}
which completes the proof of our first step. The space $L_q^1[a,a_1]$ is the Banach space and according to the Banach's fixed point Theorem (see \cite[Subsection 8.1]{KF1968}), there exists a unique solution $\widetilde{y}\in L_q^1[a,a_1]$ such that $T\widetilde{y}=\widetilde{y}'$ for $t\in[a,a_1]$.
Therefore, the solution $v\in L_q^1[a,a_1]$ is obtained from a limit of the convergent sequence $\left\{\left(T^i\widetilde{y}_0\right)\right\}$:
\begin{eqnarray}
\lim\limits_{i\rightarrow0}\|T^i\widetilde{y}_0-\widetilde{y}\|_{L_q^1[a,a_1]}=0,
\end{eqnarray}
where $ y_0$ is any function in $ L_q^1[a,a_1]$.
If at least one $\xi_k\neq0$ in the initial condition (\ref{additive3.4}), then we deduce that $\widetilde{y}_0=y_0$. From (\ref{additive3.14}), the sequence $\left\{\left(T^i\widetilde{y}_0\right)(x)\right\}$ can be written as
$$
\left(T^i\widetilde{y}_0\right)(x)=y_0+\left[I^\alpha_{q,a+}f(\cdot,T^{i-1}\widetilde{y}_0(\cdot))\right](x),
$$
for $i\in\mathbb{N}$. Let $y_i =T^i\widetilde{y}_0$. If we write
$y_i(x)=y_0+\left(I^\alpha_{q,a+}f(\cdot,y_{i-1} (\cdot))\right)(x)$, then it is clear that
\begin{eqnarray*}
\lim\limits_{i\rightarrow0}\|y_i-\widetilde{y}\|_{L_q^1[a,a_1]}=0.
\end{eqnarray*}
Hence, we actually used the method of successive approximations to get a unique solution $y'$ to the integral
equation (\ref{additive3.5}) on $[a,a_1]$.
Next, we consider the interval $[a_1, a_2]$ and $a_2= a_1 + h_1$, $a_2<b$, such that
\begin{eqnarray*}
\omega_2:=AK\frac{\left(a_2-qa_1\right)_q^\alpha}{\Gamma_q\left(\alpha+1\right)}<1,
\end{eqnarray*}
for $ h_1>0$. Then (\ref{additive3.5}) can be rewritten as
\begin{eqnarray*}
y(x)=y_0(x)+ \left(I^\alpha_{q,a+}f(\cdot,y(\cdot))\right)(a_1)+\left(I^\alpha_{q,a_1+}f(\cdot,y(\cdot))\right)(x).
\end{eqnarray*}
Since the function $y$ is uniquely given on $[a, a_1]$, the first integral can be considered as the known function,
and we can rewrite the last equation as
\begin{eqnarray*}
\left(Ty\right)(x):=y_{10}+\left(I^\alpha_{q,a_1+}f(\cdot,y(\cdot))\right)(x),
\end{eqnarray*}
where $y_{10}:=y_0(x)+ \left(I^\alpha_{q,a+}f(\cdot,y(\cdot))\right)(a_1)$. We note that $y_{10}\in L^1_q[a_1, a_2]$.
By using the same arguments as above, we derive that there exists a unique solution $y' \in L^1_q[a_1, a_2]$ to the equation (\ref{additive3.5}) on the interval $[a_1, a_2]$. We get the next interval $[a_2, a_3]$, where $a_3 = a_2 + h_2$ such that $ a_3 <b$ ($h_2 > 0$) and repeating the above process, we obtain a unique solution $y' \in L^1_q[a, b]$ to the equation (\ref{additive3.5}) and hence to the Cauchy-type problem (\ref{additive3.3})-(\ref{additive3.4}).
Finally, we must show that such a solution $y\in L^1_{\alpha,\beta,\mu, q}[a,b]$ is unique. According to the above proof, the solution $y\in L^1_q[a, b]$ is a limit of the sequence $\left\{y_i\right\}\in L^1_q[a, b]$:
\begin{eqnarray}\label{additive3.18}
\lim\limits_{i\rightarrow\infty} \|y_i-y\|_{L^1_q[a, b]}=0,
\end{eqnarray}
with the choice of certain $y_i$ on each $[a, a_1], [a_1, a_2], \dots , [a_{L-1}, b]$, for $L\in\mathbb{N}$. Since
\begin{eqnarray*}
\|D^{\alpha,\beta}_{q,a+}y_i-D^{(\alpha,\beta)\mu}_{q,a+}y\|_{L^1_q[a, b]}&=&\|f\left(\cdot,y_i(\cdot)\right)-f\left(\cdot,y(\cdot)\right)\|_{L^1_q[a, b]}\\
&\leq& A
\| y_i- y\|_{L^1_q[a, b]},
\end{eqnarray*}
by (\ref{additive3.18}), we get that
\begin{eqnarray*}
\lim\limits_{i\rightarrow\infty}\|D^{(\alpha,\beta)\mu}_{q,a+}y_i-D^{(\alpha,\beta)\mu}_{q,a+}y\|_{L^1_q[a, b]}=0,
\end{eqnarray*}
and hence $D^{\alpha,\beta}_{q,a+}\in L^1_q[a, b]$. This completes the proof.
\end{proof}
\section{Miscellaneous Examples}\label{section4}
In Section \ref{section3}, we have introduced sufficient conditions for the Cauchy-type problem to have a unique solution in some subspaces of the space of $q$-integrable functions. In fact, in the classical case, these conditions show that the $q$-difference equations of fractional order also can have integrable solutions, provided that their right-hand sides are $q$-integrable. However, these conditions are not sufficient. Below we present two examples and discuss these examples in connection with the results obtained in Section \ref{section3} \cite[See examples 3.1-3.2 for fractional cases]{KST2006}.
Let $\gamma=\beta+\mu(n-\beta)$ and $\nu=\beta+\mu(\alpha-\beta)$ for $n-1<\alpha,\beta \leq n, n\in \mathbb{N}$ and $0\leq \mu\leq 1$.
\begin{example} We assume that $I^\gamma_{q,qa+}y \in AC_q^n[qa, b]$ and $\lambda, \delta \in \mathbb{R}^+$. Then the $q$-differential equation
\begin{eqnarray}\label{additive4.1}
\left(D^{(\alpha,\beta)\mu}_{q,qa+}\right)(x)&=& \lambda \frac{(x-q^{-\nu-\delta+1}a)_q^{\nu+\delta}}{(x-q^{-2\nu-\delta+1}a)_q^\nu} y^2(x) ,
\end{eqnarray}
with the initial condition
\begin{eqnarray}\label{additive4.2}
\lim\limits_{x\rightarrow qa+}\left(D^k_qI^{(1-\mu)(n-\beta)}_{q,qa+}y \right)(x)=0
\end{eqnarray}
has a unique solution in the following form:
\begin{eqnarray}\label{additive4.3}
y(x)=\frac{1}{\lambda} \frac{\Gamma_q\left(1-\alpha-\gamma\right)}{\Gamma_q\left(1-2\alpha-\gamma\right)}
\left(x- qa\right)^{-\nu-\delta}_q,
\end{eqnarray}
for $-2\nu-\delta+1>0$, where $\left(x- a\right)^\alpha_q=x^\alpha(a/x;q)_\alpha$.
Indeed, we consider
$$
f(x,y(x)):=\lambda \frac{(x-q^{-\alpha-\delta+1}a)_q^{\alpha+\delta}}{(x-q^{-2\alpha-\delta+1}a)_q^\alpha} y^2(x),
$$
and assume that $I^\gamma_{q,a+}y \in AC_q^n[qa, b]$ and $G:=\left\{qa\leq x\leq b; \left|y(x)\right|<M<\infty\right\}$. For $-2\nu-\delta+1>0$,
$x\in[qa,b]$ and $y_1,y_2 \in G$,
\begin{eqnarray*}
\left|f(x,y_1(x))-f(x,y_2(x))\right|&=& \lambda \left|\frac{(x-q^{-\nu-\delta}a)_q^{\nu+\delta}}{(x-q^{-2\nu-\delta+1}a)_q^\nu} y^2_1(x)\right.\\
&-&\left.\frac{(x-q^{-\nu-\delta})_q^{\nu+\delta}}{(x-q^{-2\nu-\delta+1}a)_q^\nu} y^2_2(x) \right|\\
&\leq&\lambda\frac{(b-q^{-\nu-\delta+1}a)_q^{\nu+\delta}}{(a-q^{-2\nu-\delta+1}a)_q^\nu} \left|y_1 (x)+y_2(x)\right|\left|y_1 (x)- y_2(x)\right|\\
&\leq&\lambda\frac{(b-q^{-\nu-\delta+1}a)_q^{\nu+\delta}}{(a-q^{-2\nu-\delta+1}a)_q^\nu}2M\left|y_1 (x)- y_2(x)\right|,
\end{eqnarray*}
which proves that $f$ satisfies the Lipschitz condition (\ref{additive3.12}) in Theorem \ref{thm3.4}.
Therefore,
\begin{eqnarray*}
\int\limits_{qa}^b\left|f(x,y (x))\right|d_qx&=& \lambda \int\limits_{qa }^b\frac{(x-q^{-\nu-\delta+1}a)_q^{\nu+\delta}}{(x-q^{-2\nu-\delta+1}a)_q^\nu}\left|y(x)\right|^2 d_qx\\
&\leq& \lambda \frac{(b-q^{-\nu-\delta+1}a)_q^{\nu+\delta}}{(a-q^{-2\nu-\delta+1}a)_q^\nu} M^2,
\end{eqnarray*}
which means that $f(x,y (x))\in L_q^1[qa,b]$. We can conclude that, according to Theorem \ref{thm3.3}, the equation has a unique solution in $L_q^1[qa,b]$. Moreover, by applying the operator $D^{(\alpha,\beta)\mu}_{q,qa+}$ (see \ref{additive3.1})and using the Remark \ref{rem3.2} and (\ref{additive2.2}), (\ref{additive2.3}), we get that
\begin{eqnarray*}
D^{(\alpha,\beta)\mu}_{q,qa+} \left(x- qa\right)^{-\nu-\delta}_q &=& I_{q,qa+}^{\gamma-\nu}D^{\gamma}_{q,qa+}(x-qa)_q^{-\nu-\delta} \\
&=&\frac{\Gamma_q(1-\nu-\delta)}{\Gamma_q{(1-\nu-\delta-\gamma)}} \left[I_{q,qa+}^{\gamma-\nu}
(x-qa)_q^{-\nu-\delta-\gamma} \right]\\
&=& \frac{\Gamma_q(1-\nu-\delta )}{\Gamma_q(1-\delta-2\nu)}
(x-qa)_q^{-2\nu-\delta}
\end{eqnarray*}
for $2\nu+\delta<1$. Hence,
\begin{eqnarray}\label{additive4.4}
D^{(\alpha,\beta)\mu}_{q,qa+} \left[y(x)\right]
=\frac{1}{\lambda}\left[\frac{\Gamma_q(1-\nu-\delta )}{\Gamma_q(1-\delta-2\nu)}\right]^2
\left(x- qa\right)^{-2\nu-\delta}_q.
\end{eqnarray}
By using the well-known formulas (see \cite[formula (3.7)]{CK2000}):
$$
(x-q^{-\nu-\delta+1}a)_q^{\nu+\delta}(x-qa)_q^{-\nu-\delta}=1
$$
and
$$
(x-qa)_q^{-\nu-\delta}=(x-qa)_q^{-2\nu-\delta}(x-q^{-2\nu-\delta}a)_q^\nu,
$$
we have that
\begin{eqnarray}\label{additive4.5}
f(x,y(x))&=&\lambda \frac{(x-q^{-\nu-\delta+1}a)_q^{\nu+\delta}}{(x-q^{-2\nu-\delta+1}a)_q^\nu} y^2(x)\nonumber\\
&=&\frac{1}{\lambda}\left[\frac{\Gamma_q\left(1-\nu-\delta\right)}{\Gamma_q\left(1-2\nu-\delta\right)}\right]^2
\times\frac{(x-q^{-\nu-\delta+1}a)_q^{\nu+\delta}}{(x-q^{-2\alpha-\gamma+1}a)_q^\alpha}
\left[(x-qa)_q^{-\nu-\delta}\right]^2\nonumber\\
&=&\frac{1}{\lambda}\left[\frac{\Gamma_q\left(1-\nu-\delta\right)}{\Gamma_q\left(1-2\nu-\delta\right)}\right]^2
(x-qa)_q^{-2\nu-\delta}=D^{(\alpha,\beta)\mu}_{q,qa+}\left[y(x)\right].
\end{eqnarray}
Now, by combining (\ref{additive4.4}) and using (\ref{additive4.5}) and $\lim\limits_{x\rightarrow qa}\left(x- qa\right)^{-\nu-\delta}_q=0$ for $x\in[qa,b]$, we see that $y(x)$ defined by (\ref{additive4.3}) satisfies (\ref{additive4.1})-(\ref{additive4.2}). The proof of our claim is complete.
\end{example}
\begin{example} Let $\delta, \lambda\in\mathbb{R}^+$. Then the $q$-differential equation
\begin{eqnarray}\label{additive4.6}
\left(D^{(\alpha,\beta)\mu}_{q,a+}y\right)(x) &=& \lambda \left[\frac{(x-qa )_q^{2\nu+2\delta}}
{(x-q^{\nu+2\delta+1}a )_q^\nu}\right]^\frac{1}{2}\left[y(x)\right]^\frac{1}{2},
\end{eqnarray}
with the initial condition
\begin{eqnarray}\label{additive4.7}
\lim\limits_{x\rightarrow qa+}\left(D^k_qI^{(1-\mu)(n-\beta)}_{q,qa+}y \right)(x)&=&0
\end{eqnarray}
has a unique solution in the following form:
\begin{eqnarray}\label{additive4.8}
y(x)=\left[\lambda \frac{\Gamma_q\left(\nu+2\delta+1\right)}{\Gamma_q\left(2\nu+2\delta+1\right)}
\right]^2\left(x-qa\right)^{2\nu+2\delta}_q.
\end{eqnarray}
Since $\left(x-qa\right)^{2\nu+2\delta}_q=\left(x-qa\right)^{\nu+2\delta}_q\left(x-q^{\nu+2\delta}a\right)^\nu_q$ we get
\begin{eqnarray}\label{additive4.9}
f(x,y(x))&:=& \lambda \frac{\left[(x-qa )_q^{2\nu+2\delta}\right]^\frac{1}{2}}
{(x-q^{\nu+2\delta+1}a )_q^\nu} \left[y(x)\right]^\frac{1}{2}\nonumber\\
&=&\lambda^2 \frac{\Gamma_q\left(\nu+2\delta+1\right)}{\Gamma_q\left(2\nu+2\delta+1\right)}\left[\frac{(x-qa )_q^{2\nu+2\delta}}
{(x-q^{\nu+2\delta+1}a )_q^\nu}\right]^\frac{1}{2}
\left[\left(x-qa\right)^{2\nu+2\delta}_q\right]^\frac{1}{2}\nonumber\\
&=&\lambda^2 \frac{\Gamma_q\left(\nu+2\delta+1\right)}{\Gamma_q\left(2\nu+2\delta+1\right)} \left(x-qa\right)^{\nu+2\delta}_q.
\end{eqnarray}
We assume that $I^\gamma_{q,a+}y \in AC_q^n[a, b]$ and
$ G:=\left\{a\leq x\leq b; M_1< y(x) <M_2, 0<M_1,M_2\right\}$. Then, for $x\in[a, b]$ and $y_1, y_2\in G$ we obtain
\begin{eqnarray*}
|f(x,y_1(x))-f(x,y_2(x))|&\leq&\lambda\frac{\left[(b-qa )_q^{2\nu+2\delta}\right]^\frac{1}{2}}
{(a-q^{\nu+2\delta+1}a )_q^\nu }\left|y^\frac{1}{2}_1(x)-y^\frac{1}{2}_2(x)\right|\\
&=&\lambda\frac{\left[(b-qa )_q^{2\nu+2\delta}\right]^\frac{1}{2}}
{(a-q^{\nu+2\delta+1}a )_q^\nu }\frac{\left|y_1(x)-y_2(x)\right|}{y^\frac{1}{2}_1(x)+y^\frac{1}{2}_2(x)}\\
&\leq&\frac{\lambda}{2M_1}\frac{\left[(b-qa )_q^{2\nu+2\delta}\right]^\frac{1}{2}}
{(a-q^{\nu+2\delta+1}a )_q^\nu}\left|y_1(x)-y_2(x)\right|,
\end{eqnarray*}
which proves that $f$ satisfies the Lipschitz condition (\ref{additive3.6}) in Theorem \ref{thm3.4}.
Moreover,
\begin{eqnarray*}
\int\limits_{qa}^b\left|f(x,y (x))\right|d_qx
\leq \lambda \frac{\left[(b-qa )_q^{2\nu+2\delta}\right]^\frac{1}{2}}
{(a-q^{\nu+2\delta+1}a )_q^\delta} M^\frac{1}{2},
\end{eqnarray*}
which means that $f(x,y (x))\in L_q^1[qa,b]$. We can conclude that according to the Theorem \ref{thm3.3}, the equation has a unique solution in $L_{q}^1[qa,b]$. Moreover, by applying the operator $D^{(\alpha,\beta)\mu}_{q,qa+}$ (see \ref{additive3.1}) to (\ref{additive4.8}) and using Remark \ref{rem3.2} and (\ref{additive2.2}), (\ref{additive2.3}), we get that
\begin{eqnarray*}
D^{(\alpha,\beta)\mu}_{q,qa+} \left(x- qa\right)^{2\nu+2\delta}_q &=& I_{q,qa+}^{\gamma-\nu}D^{\gamma}_{q,qa+}(x-qa)_q^{2\nu+2\delta} \\
&=&\frac{\Gamma_q(1+2\nu+2\delta)}{\Gamma_q{(1+2\nu+2\delta-\gamma)}} \left[I_{q,qa+}^{\gamma-\nu}
(x-qa)_q^{2\nu+2\delta-\gamma} \right]\\
&=&\frac{\Gamma_q(1+2\nu+2\delta)}{\Gamma_q{(1+\nu+2\delta)}}
(x-qa)_q^{\nu+2\delta }
\end{eqnarray*}
for $2\nu+\delta<1$. Hence, by \ref{additive4.9}
\begin{eqnarray*}
D^{(\alpha,\beta)\mu}_{q,qa+} \left[y(x)\right]
= \lambda^2 \frac{\Gamma_q(1+2\nu+2\delta)}{\Gamma_q{(1+\nu+2\delta)}}
(x-qa)_q^{\nu+2\delta }=f(x,y(x)).
\end{eqnarray*}
Now, by combining (\ref{additive4.4}) and using that (\ref{additive4.5}) and $\lim\limits_{x\rightarrow qa}\left(x- qa\right)^{-\nu-\delta}_q=0$ for $x\in[qa,b]$, we see that $y(x)$ defined by (\ref{additive4.3}) satisfies (\ref{additive4.6})-(\ref{additive4.7}). The proof of our statement is complete.
\end{example}
\section{A linear Cauchy-type $q$-fractional problem}
Let $\gamma=\beta+\mu(n-\beta)$ and $\nu=\beta+\mu(\alpha-\beta)$ for $n-1<\alpha,\beta \leq n, n\in \mathbb{N}$ and $0\leq \mu\leq 1$.
The (Mittag-Leffler) $q$-function $E_{\alpha,\beta}\left(z;q\right)$ is defined by (\cite{SPT2020})
\begin{eqnarray}\label{additive5.1}
E_{\alpha,\beta}\left[\lambda x^{\gamma}(a/x;q)_{\gamma};q\right]=
\sum\limits_{k=0}^{\infty}\frac{\lambda^{k}x^{k\gamma}(a/x;q)_{k\gamma}}{\Gamma_q(\alpha k+\beta)}.
\end{eqnarray}
We study the following Cauchy type $q$-fractional problem
\begin{eqnarray}\label{additive5.2}
\left(D^{(\alpha,\beta)\mu}_{q,a+}y\right)(x) -\lambda y(x)&=&f(x),
\end{eqnarray}
\begin{eqnarray}\label{additive5.3}
\lim\limits_{x\rightarrow a+}\left(D^k_qI^{(1-\mu)(n-\beta)}_{q,a+}y \right)(x)=\xi_k, \;\;\;\xi_k\in\mathbb{R}, k=0,1,2,\dots n-1.
\end{eqnarray}
\begin{thm}\label{thm5.1}
Let $\gamma=\beta+\mu(n-\beta)$ and $\nu=\beta+\mu(\alpha-\beta)$ for $n-1<\alpha,\beta \leq n, n\in \mathbb{N}$ and $0\leq \mu\leq 1$. We assume that $\lambda \in \mathbb{R}$ is such that
\begin{eqnarray*}
|\lambda|b^{\nu}(1-q)^{\nu}<1,
\end{eqnarray*}
and $f\in L^1_q[a,b]$. Then the Cauchy problem (\ref{additive5.2})-(\ref{additive5.3}) has a unique solution
$y \in L^1_{\alpha,\beta,\mu, q}[a,b]$ and this solution is given by
\begin{eqnarray}\label{additive5.4}
y(x)&=&\sum\limits_{k=1}^{n} b_k x^{\nu-k}E_{\nu,\nu-k+1}\left[\lambda x^\nu;q\right] \nonumber\\
&+&\int\limits_0^{x} x^{\nu-1}\left(qt/x;q\right)_{\nu-1} E_{\nu,\nu}\left[\lambda{x^\nu}(q^\nu{t}/x;q)_\nu;q\right] f(t)d_qt.
\end{eqnarray}
\end{thm}
\begin{proof} First, we solve the Volterra q-integral equation (\ref{additive3.5}) for
\begin{eqnarray*}
f(x,y(x)):=\lambda{y(x)}+f(x).
\end{eqnarray*}
For this aim, we apply the method of successive approximations by setting
$$
y_{0}(x)=\sum\limits_{k=1}^n\frac{\xi_k}{\Gamma_q\left(\gamma-k+1\right)}x^{\gamma-k}(a/x;q)_{\gamma-k}
$$
and
\begin{eqnarray}\label{additive5.5}
y_i(x)&=&y_0(x)+
\frac{\lambda{x^{\nu-1}}}{\Gamma_q\left(\nu\right)}\int_{0}^{x}(qt/x;q)_{\nu-1}y_{i-1}(t)d_qt\nonumber\\ &+&\frac{x^{\nu-1}}{\Gamma_q\left(\nu\right)}\int_{0}^{x}(qt/x;q)^{\nu-1}_{q}f(t)d_qt.
\end{eqnarray}
Using (\ref{additive2.1}) and (\ref{additive5.5}) we find $y_1(x)$:
$$
y_1(x)=y_0(x)+\lambda\left(I^{\nu}_{q,0+}y_0\right)(x)+\left(I^{\nu}_{q,0+}f\right)(x),
$$
that is,
\begin{eqnarray}\label{additive5.6}
y_{1}(x)&=&\sum\limits_{k=1}^{n}\frac{b_k}{\Gamma_q\left(\nu-k+1\right)}x^{\nu-k}+
\lambda\sum\limits_{k=1}^{n}\frac{b_k}{\Gamma_q\left(\nu-k+1\right)}\left(I^\nu_{q,0+}t^{\nu-k}\right)(x)+
\left(I^{\nu}_{q,0+}f\right)(x)\nonumber\\
&=&\sum\limits_{k=1}^{n}\frac{b_k}{\Gamma_q\left(\nu-k+1\right)}x^{\nu-k}+
\lambda\sum\limits_{k=1}^{n}\frac{b_k x^{2\nu-k}}{\Gamma_q\left(2\nu-k+1\right)}+\left(I^{\nu}_{q,0+}f\right)(x) \nonumber\\
&=&\sum\limits_{k=1}^{n} b_k \sum\limits_{m=1}^2\frac{\lambda^{m-1}x^{m\nu-k}}{\Gamma_q\left(\nu m-k+1\right)}+\left(I^{\nu}_{q,0+}f\right)(x).
\end{eqnarray}
Similarly, using (\ref{additive2.1}), (\ref{additive2.2}), (\ref{additive2.4}) and (\ref{additive5.6}) we have for $y_2(x)$ that
\begin{eqnarray*}
y_2(x)&=&y_0(x)+\lambda \left(I^\nu_{q,0+}y_1\right)(x)+\left(I^\nu_{q,0+}f\right)(x) \\
&=&\sum\limits_{k=1}^{n}\frac{b_k}{\Gamma_q\left(\nu-k+1\right)}x^{\nu-k}\\
&+&\frac{\lambda}{\Gamma_q\left(\nu\right)} \sum\limits_{k=1}^{n} b_k \sum\limits_{m=1}^2\frac{\lambda^{m-1}}{\Gamma_q\left(\nu m-k+1\right)}\left(I^\nu_{q,0+}t^{m\nu-k}\right)(x)\\
&+&\lambda\left(I^{\nu}_{q,0+}I^{\nu}_{q,0+}f(t)\right)(x)+\left(I^{\nu}_{q,0+}f\right)(x)\\
&=&\sum\limits_{k=1}^{n}\frac{b_k}{\Gamma_q\left(\nu-k+1\right)}x^{\nu-k}+\lambda\sum\limits_{k=1}^{n} b_k \sum\limits_{m=1}^2\frac{\lambda^{m-1}}{\Gamma_q\left(\nu(m+1)-k+1\right)}x^{\nu(m+1)-k} \\
&+&\lambda\left(I^{2\nu}_{q,0+}f(t)\right)(x)+\left(I^{\nu}_{q,0+}f\right)(x)
\end{eqnarray*}
\begin{eqnarray*}
&=&\sum\limits_{k=1}^{n}\frac{b_k}{\Gamma_q\left(\nu-k+1\right)}x^{\nu-k}+\lambda\sum\limits_{k=1}^{n} b_k \sum\limits_{m=1}^2\frac{\lambda^{m-1}}{\Gamma_q\left(\nu(m+1)-k+1\right)}x^{\nu(m+1)-k} \\
&+&\frac{\lambda{x^{2\nu-1}}}{\Gamma\left(2\nu\right)}\int\limits_0^xf(t)\left(qt/x;q\right)_{2\nu-1}d_qt+\left(I^{\nu}_{q,0+}f\right)(x).
\end{eqnarray*}
Thus,
\begin{eqnarray*}
y_2(x)&=&\sum\limits_{k=1}^{n}b_k\sum\limits_{m=1}^3
\frac{\lambda^{m-1}x^{\nu{m}-k}}{\Gamma_q\left(\nu{m}-k+1\right)}\nonumber\\
&+&\int_0^x\left[\sum\limits_{m=1}^{2}\frac{\lambda^{m-1}x^{\nu{m}-1}(qt/x;q)_{\nu{m}-1}}{\Gamma_q(\nu{m})}\right] f(t)d_qt.
\end{eqnarray*}
Continuing this process, we derive the following relation for $y_i(x)$:
\begin{eqnarray*}
y_i(x)&=&\sum\limits_{k=1}^{n}b_k\sum\limits_{m=1}^{i+1}
\frac{\lambda^{m-1}x^{\nu{m}-k}}{\Gamma_q\left(\nu{m}-k+1\right)}\nonumber\\
&+&\int_{0}^{x}\left[\sum\limits_{m=1}^i\frac{\lambda^{m-1}x^{\nu{m}-1}(qt/x;q)_{\nu{m}-1}}{\Gamma_q(\nu{m})}\right] f(t)d_qt\nonumber\\
&=&\sum\limits_{k=1}^{n}b_k\sum\limits_{m=0}^{i}
\frac{\lambda^{m}x^{\nu{(m+1)}-k}}{\Gamma_q\left(\nu{(m+1)}-k+1\right)}\nonumber\\
&+&\int_{0}^{x}\left[\sum\limits_{m=0}^{i-1}\frac{\lambda^{m}x^{\nu{(m+1)}-1}(qt/x;q)_{\nu{(m+1)}-1}}{\Gamma_q(\nu(m+1))}\right] f(t)d_qt.
\end{eqnarray*}
From (\ref{additive2.1C}) and (\ref{additive2.1D}) its follows that
\begin{eqnarray*}
\frac{\lambda^{m}x^{\nu{m}}({qt}/{x};q)_{\nu{m}}}{\Gamma_q(\nu{m}+\nu)} &=&\left[\lambda{x}^\nu(1-q)^\nu\right]^m\frac{(q^{\nu(m+1)}; q)_\infty }{(q^{\nu{m}+1}t/x; q)_\infty }(1-q)^{\nu-1} \\
&\leq&\left[\lambda{x}^\nu(1-q)^\nu\right]^m\frac{(q^{\nu(m+1)}; q)_\infty }{(q^{\nu{m}+1} ; q)_\infty }(1-q)^{-1}\\
&\leq&\left[\lambda{b}^\nu(1-q)^\nu\right]^m (1-q)^{-1}.
\end{eqnarray*}
Thus,
\begin{eqnarray*}
\left|\sum\limits_{m=0}^\infty\frac{\lambda^{m}x^{\nu{m}}({qt}/{x};q)_{\nu{m}}}{\Gamma_q(\nu{m}+\nu)}\right|&\leq&\frac{1}{1-q}\sum\limits_{m=0}^\infty\left[|\lambda|{b}^\nu(1-q)^\nu\right]^m\leq\frac{1}{|\lambda|{b}^\nu(1-q)^{\nu+1}},
\end{eqnarray*}
which means that this is absolutely and uniformly convergent for $\lambda{b}^\nu(1-q)^\nu<1$. Taking the limit as $i\rightarrow\infty$ and using the well-known formulas (see \cite[formula (3.7)]{CK2000}), we obtain the following explicit solution $y(x)$ to the $q$-integral equation (\ref{additive5.4}):
\begin{eqnarray*}
y(x)&=&\sum\limits_{k=1}^{n}b_kx^{\nu-k}\sum\limits_{m=0}^\infty
\frac{\lambda^{m}x^{\nu{m}}}{\Gamma_q\left(\nu{m}+\nu-k+1\right)}\nonumber\\
&+&\int_{0}^{x}x^{\nu-1}({qt}/{x};q )_{\nu-1}\left[\sum\limits_{m=0}^\infty\frac{\lambda^{m}x^{\nu{m}}({qt}/{x};q)_{\nu{m}}}{\Gamma_q(\nu{m}+\nu)}\right] f(t)d_qt.
\end{eqnarray*}
On the basis of Theorem \ref{thm3.4} and (\ref{additive5.1}), the last expression represents the explicit solution to the Volterra $q$-integral equation (\ref{additive5.4}) and hence to the Cauchy-type problem (\ref{additive5.2})-(\ref{additive5.3}).
\end{proof}
\section{Conflict of Interests}
The authors declare that they have no conflict of interest.
|
\section{Introduction}
\label{sec_intro}
Data-driven approaches are becoming a competitive and viable means for solving some challenging problems in scientific computing. In particular, machine learning and data science techniques are being developed specifically for scientific tasks such as predicting solutions from complicated spatiotemporal data, approximating trajectories from unknown dynamical systems, or constructing surrogate functions for complex nonlinear operations. One of the underlying principles in these approaches is to either augment or approximate the system using observed data. In particular, these approaches are geared toward solving parametric partial differential equations (PDE). Although one should expect that a rigorous and tailored numerical method for solving a specific PDE will typically outperform a general machine learning (ML) algorithm, one of the key contributions of these new ML approaches is the ability to use less prior information. For example, the ML approaches do not necessarily need to have knowledge of the underlying governing system. However, their ability to correctly capture and model indirect physically relevant terms (e.g. energy or conserved quantities) is still an open problem.
In this work, we will focus on the task of computing a family of solutions to parametric PDE, i.e. a PDE whose model/equations, boundary, and/or initial data are parametrized. For example, the conductivity of the heat equation relies on the properties of the underlying material, which can be characterized by some set of parameters. Thus given a set of parameters, one has an equation or algorithm that computes the conductivity and thereby generates a model for the spatiotemporal dynamics. The goal is to find the solution as a function of the parameters, bypassing the need to completely solve the nonlinear system. Recently, deep learning strategies have shown success in this direction \cite{lu2021learning, li2020fourier, zhang2020learning}.
The standard approaches in ML focus on learning functions, i.e. methods that train a mapping from the parameters to the solution of the equation (at a finite set of locations) based on some pre-generated data. We will refer to these approaches as \textit{function learning}, which includes standard multilayer perceptrons, feedforward neural networks, recurrent neural networks, etc. During the testing phase, one obtains a new set of parameters and the trained network predicts the solution in one forward pass, which can be computationally efficient. However, function learning approaches are restrictive. First, the neural network may not be easy to learn due to curse-of-dimensionality. Specifically, when training a function that maps the parameters to the solution of a given PDE, the input (i.e. the parameters) and output (i.e. the value of the solution on a discrete set of time and/or space points) will have a very large dimension. Returning to the example, let us consider the 2-dimensional heat equation whose domain is $[0, 1]^2$. If the solution is defined on a spatial mesh with resolution $100\times 100$, then the dimension of the solution vector (i.e. the output) is $10^4$. As a result, the neural network often requires a large set of parameters which can result in a loss of efficiency and training accuracy. Secondly, for function learning, since the output dimension is fixed, all of the output samples must share the same structure. That is, the outputs must be represented on the same mesh or sequence of points. For example, if the majority of observed outputs are defined on a $100\times 100$ mesh, but there are some samples whose solutions are defined on $99\times 99$ mesh or on a different set of $100\times 100$ points, then the standard approaches will fail. This can be very limiting since one does not necessarily have control over the sampling process and may require the system to adapt to different meshes. Lastly, the predicted values must be on the same mesh in order to be consistent with the trained model.
To overcome these issues, \textit{operator learning} \cite{chen1995universal,chen1995approximation} was introduced and has seen success in various scientific computing tasks \cite{lu2021learning, li2020fourier, kovachki2021neural, o2022derivative, lanthaler2022nonlinear, patel2022variationally, ong2022integral, qian2022reduced, de2022cost}. The operator learning approach trains a map between functions rather than fixed-sized vectors, which allows for an accurate representation of the underlying dependencies between the input and output. In particular, operator learning can handle the training of parametric PDEs since the outputs are represented (correctly) by functions rather than fixed dimensional vectors. Formally, let $V$ and $U$ be two function spaces whose domains are $K_1$ and $K_2\subset\mathbb{R}^d$. Let $G$ be the target operator where $G(u)(\cdot)\in U$ for $u\in V$. We want to learn the operator $G$ by the network $G_{\theta}$, i.e.:
\begin{align}
G(u)(x) \approx G_{\theta}(u)(x),
\label{intro_operator}
\end{align}
for $x\in K_2$, where $\theta$ denotes all network parameters. From the structural point of view, the training and expressibility of function learning and operator learning differ. For function learning, the input is the set of parameters that characterize the input function $u$, while in operator learning, the inputs contain both the parameters and the independent variables $x$ used in evaluating the solution. Additionally, in function learning, the output is the solution given a particular discrete representation, but in operator learning, the output is the solution at the point prescribed by the independent variables $x$. Thus with operator learning, one can also predict the solution at other locations within the domain, providing a more flexible representation.
In \cite{chen1995universal}, the authors provided a universal approximation theorem for nonlinear continuous operators. The original application was for modeling solutions to dynamical systems using a particular shallow architecture. The Deep Operator Neural Network (DON) \cite{lu2021learning}, extended the results of \cite{chen1995universal} (in particular Theorem 5) to account for deep architectures and proposed several new applications. DON has had many extensions and variants. When the governing equations are known, one can incorporate this information using the so-called physics-informed structure \cite{wang2021learning}. In that setting, the user can use fewer samples since the equations provide additional information to the training. Optimization algorithms which exploit DON structure have also been proposed to handle noisy data and train DON \cite{lin2023b, psaros2022uncertainty}, which may be more practical for real data. DON has also been generalized to a wider class of nonlinear approximation problems using shifts \cite{lanthaler2022nonlinear}. Also, detailed numerical experiments on the cost-accuracy trade-off for neural operators are given in \cite{de2022cost}.
The Fourier neural operator (FNO) \cite{li2020fourier, kovachki2021universal} was designed based on a nonlinear generalization of the kernel integral representation for some operators and makes use of the convolutional or Fourier network structure. It was shown to be a universal approximation for operators in \cite{kovachki2021universal}, with respect to the Hilbert space norm (i.e. $H^s$).
FNO has had several extensions with applications to different areas. Some examples include but are not limited to: global weather prediction \cite{pathak2022fourcastnet}, multiphase flow \cite{wen2022u, zhang2022fourier}, and solving PDE with complex geometry \cite{li2022fourier}.
It is worth noting that these neural operator formulations are connected to other approaches in scientific machine learning. FNO can also be seen as a deep generalization to the random Fourier feature approaches \cite{rahimi2007random, rahimi2008weighted} based on approximating kernel integral operators via randomized Fourier networks, see also \cite{bach2015equivalence, rudi2017generalization, hashemi2023generalization, chen2021conditioning, chen2022concentration}. In addition, the FNO is connected to other Fourier learning approaches for PDE, including the compressive spectral methods \cite{schaeffer2013sparse, mackey2014compressive}, which trains a sparse Fourier basis representation for the solution of a given PDE; and the sparse optimization framework for PDE discovery \cite{schaeffer2017learning}, which trains a sparse representation of the PDE operator using a mixture of Fourier and polynomial features.
There are several important open problems with operator learning. First, one often needs to discretize the input function $u$ for each sample, i.e. we observe the inputs at some locations or with respect to some basis. However, we do not want the construction of the operator to depend on a fixed choice of locations or basis elements used in the discretization of the inputs since the locations may differ. If a neural operator allows for different input locations in its training, then it is called \textit{discretization-invariant}. Note that the vanilla DON is not discretization-invariant. Although some methods have been proposed to make DON discretization-invariant \cite{li2022learning}, the universal approximation theorem for DON may not support it. Second, we call a neural operator approach \textit{prediction-free} if the discretization mesh for the input function $u$ and the output function $G(u)(\cdot)$ can differ, i.e. we are not restricted to mapping functions on the same input and output grids. Lastly, a method is \textit{domain-independent} if the input function $u$ and its image $G(u)(\cdot)$ can be defined on different domains. For FNO, because of the use of the fast Fourier transform, it is not prediction-free or domain-independent. If a method has all three properties, we call it \textit{mesh-free}. This work aims to introduce a mesh-free operator learning approach, provide numerical evidence, and show how it can be applied to challenging multiscale problems.
\subsection{Our Contributions}
We propose a mesh-free operator learning approach that builds from the results of \cite{chen1995universal} to incorporate more information on the input and output meshes. Our contributions are as follows.
\begin{enumerate}
\item We construct a {mesh-free} operator learning framework called the basis enhanced learning network (BelNet).
\item We show that BelNet outperforms classical ML approaches and recent neural operators for challenging multiscale systems, i.e. those with high-contrast coefficients or dependencies.
\item Through several numerical experiments, we show that BelNet achieves consistent results between using a fixed grid or free grid, providing evidence that it generalizes the previous operator learning approaches.
\end{enumerate}
The paper is organized as follows. In Section \ref{sec_overview}, we review the popular {operator learning} frameworks and introduce the formal definition of the {mesh-free}. We then discuss the details of the basis enhanced learning (BelNet) in Section \ref{sec_bel}. Lastly, Section \ref{sec_numerical} presents some numerical results.
\section{Overview of Operator Networks}
\label{sec_overview}
We review the universal approximation theory and the two recent learning architectures based on approximating operators. We will then introduce the formal definition of {mesh-free}.
\subsection{Theory and Architectures}
The classical universal approximation theorems (UAP) for shallow \cite{cybenko1989approximation} and multilayer networks \cite{hornik1989multilayer,pinkus1999approximation} show that for any continuous function, there exists a feedforward neural network with a non-polynomial activation of a certain width (and depth) that can approximate the function with arbitrary accuracy in the sup-norm. The hidden layers' width and depth depend on the accuracy level. Numerical methods for scientific computing and other applications that use the classical UAP are developed for training functions and thus can scale poorly (in terms of trainable parameters or nodes) when the inputs are of very high dimension or have complex structures.
For applications in which the inputs to the model are functions themselves, the classical universal approximation results are insufficient. In \cite{chen1995universal}, it was shown that with a sole hidden layer, one could construct a network that can approximate any operator. These \textit{neural operators} are useful for learning families of solutions to a given PDE since they can approximate mappings that have functional parametric dependencies. We will refer to the universal approximation theorem for operators \cite{chen1995universal} as UAP-O in this work and state it below for completeness.
\begin{theorem}[UAP-O, \cite{chen1995universal}]
Suppose $g$ is a Tauber-Wiener (TW) function\footnote{If a function $g:\mathbb{R}\rightarrow \mathbb{R}$ (continuous or discontinuous) satisfies that all linear combinations $\Sigma_{i = 1}^Nc_ig(\lambda_i x+\theta_i)$ are dense in $C[a, b]$, where $c_i, \lambda_i, \theta_i\in\mathbb{R}$, then $g$ is called a Tauber-Wiener (TW) function.}, $X$ is Banach, $K_1\subset X$ and $K_2\subset \mathbb{R}^d$ are compact. Let $V\subset C(K_1)$ be compact and $G: V\rightarrow C(K_2)$ be a nonlinear continuous operator.
Then for any $\epsilon>0$, there are positive integers $M, N, K$, constants $c_i^k, \zeta_k, \theta_i^k, \varepsilon_{ij}^k\in\mathbb{R}$, points $\omega_k\in\mathbb{R}^d$, $y_j\in K_1$, $i = 1, ..., M$, $k = 1, ..., K$, $j = 1, ...., N$ such that
\begin{align*}
\bigg| G(u)(x) - \sum_{k = 1}^K \sum_{i = 1}^M c_i^k\, g\left(\sum_{j = 1}^N\varepsilon_{ij}^ku(y_j)+\theta_i^k\right)\, g(\omega_k\cdot x+\zeta_k)\bigg|<\epsilon
\end{align*}
holds for all $u\in V$ and $x\in K_2$.
\end{theorem}
The DON framework \cite{lu2021learning, lu2022comprehensive, lanthaler2022error} extends the UAP-O to continuous functions and then builds the network based on the following general UAP theorem stated below.
\begin{theorem}[DON Approximation \cite{lu2021learning, lu2022comprehensive, lanthaler2022error}]
Suppose $X$ is Banach, $K_1\subset X$ and $K_2\subset \mathbb{R}^d$ are compact. Let $V\subset C(K_1)$ be compact and $G: V\rightarrow C(K_2)$ be a nonlinear continuous operator. Then for any $\epsilon>0$, there exists positive integers $N, K$, continuous function $f:\mathbb{R}^d\rightarrow\mathbb{R}^K$, $g:\mathbb{R}^N\rightarrow\mathbb{R}^K$ and $y_1, ..., y_N\in K_1 $ such that
\begin{align*}
\left| G(u)(x) - \langle\underbrace{g\big(u(y_1), ..., u(y_N)\big)}_{branch}, \underbrace{f\big(x\big)}_{trunk}\rangle\right|<\epsilon,
\end{align*}
for any $u\in V$ and $x\in K_2$. Here $\langle\cdot, \cdot\rangle$ denotes the inner product in $\mathbb{R}^K$.
\label{deeponet_uap}
\end{theorem}
Theorem~\ref{deeponet_uap} has been used as the basis for many recent developments in neural network approximations based on the DON framework \cite{lu2022comprehensive, lanthaler2022error}. Note that, from the structure of the network used in Theorem~\ref{deeponet_uap}, we see that the collection of training sensors $\{y_j\}_{j=1}^N$ are ``fixed'' for all the sample $u\in V$, i.e. the locations of the observations are the same for each sample. For example, to train the solution operator for nonlinear PDE, the grid (i.e. the choice of discretization) must be the same for all input functions in order of Theorem~\ref{deeponet_uap} to hold.
Figure \ref{deepo_structure} presents an illustration of the architecture of DON based on Theorem \ref{deeponet_uap}. In particular, note that DON depends on the independent variable $x\in\mathbb{R}^d$ within the separate trunk subnetwork (highlighted in blue in Figure \ref{deepo_structure}). As a result, the input and output domains can differ, and the choice of discretization for the two domains do not need to be the same. However, the set of sensor locations $\{y_j\}_{j=1}^N$ is fixed.
\begin{figure}[H]
\centering
\scalebox{.9}{\input{deepo.tex}}
\caption{Stacked version DON. $\bigotimes$ denotes the inner product in $\mathbb{R}^K$.}
\label{deepo_structure}
\end{figure}
Another parallel neural operator introduced in \cite{li2020fourier} is the Fourier neural operator (FNO), which was originally designed as an approximation to the kernel integral operator \cite{li2020fourier, kovachki2021universal} for training a family of solutions to PDE. FNO takes the following form \cite{kovachki2021neural}:
\begin{align}
\mathcal{FNO} (u) = Q\circ \mathcal{L}_L\circ...\circ\mathcal{L}_1\circ R(u)
\label{intro_fno}
\end{align}
where $Q$ and $R$ are the linear projection and lifting operators, $\mathcal{L}_l$ is the neural (sub)network of the form:
\begin{align*}
\mathcal{L}_l(u)(x) = \sigma\bigg(W_lu(x)+b_l(x)+\mathcal{F}^{-1}\big(P_l(k)\cdot \mathcal{F}(u)(k)\big)(x) \bigg),
\end{align*}
where $W_l$ and $b_l$ are the weight and bias.
In this formulation, $\mathcal{F}$ and $\mathcal{F}^{-1}$ are the Fourier transformation and its inverse, and thus the majority of hidden parameters are trained in the Fourier domain. In particular, $P_l$ defines the coefficients of a non-local, linear mapping in the Fourier domain. Although graph-based formulations are also possible \cite{li2020neural}, writing the network as the composition of nonlinear operations in the Fourier domain provides several computational benefits, including lowering the computational cost and simplifying the layer structure. In addition, \textcolor{black}{FNO is discretization invariant in the sense that the network is invariant to the different meshes (see Theorem 5 of \cite{kovachki2021universal}). However, the input and output functions must be defined in the same domain on the same grid.}
Note that in both neural operator frameworks, the sensor locations are fixed. In applications where the sensors are moving (like in particle tracking) or in experiments where not all sensors provide useful information, we would like to be able to use different samples $\{y_j\}_{j=1}^N$. In particular, we construct a neural operator based on DON that is mesh-free, i.e. independent of the locations used in the discretization, prediction, or domains.
\subsection{Discretization-Invariant}
The goal is to construct an operator $G(u)(x): V\rightarrow U$, for $u\in V$, $G(u)(x)\in U$, where $V$ and $U$ are two function spaces.
The function $u$ (which can be the controls, boundary conditions, initial states, etc.) is an input to the network and will only be observed on a set of sensor locations. This is represented by a particular choice of discretization. In practice, the discretization of $u$ could be determined by the data or the sensor locations and thus may not be fixed between different experiments. However, based on the UAP-O, $u$ must be observed at some sensor points to obtain the approximation guarantees. We call a neural operator {discretization-invariant} if the approximation does not require the locations of the sensors $\{y_j\}_{j=1}^N$ to be fixed. As an illustrative example, Figure \ref{intro_dis_free} considers three cases between two different samples, where the sample refers to the different input functions $u$. The first plot of Figure \ref{intro_dis_free} shows the uniform grid used in the construction of finite differences or spectral methods, which require each sampled input, e.g. sample 1 $\{u_1(y_j)\}_{j=1}^N$, and sample 2, $\{u_2(y_j)\}_{j=1}^N$, share the same grid. This is also the case in the middle plot, which uses the same three sensor locations. Our approach is discretization-invariant in the sense that the sensors can be distributed independently for different samples displayed in the plot on the right in Figure \ref{intro_dis_free}. Based on the construction, the FNO approach is discretization-invariant since \textcolor{black}{the layers are trained in the Fourier domain and are thus evaluated in the spatial domain via a projection using the Fourier basis elements}. The vanilla DON formulation is not discretization-invariant. Some recent extensions of DON \cite{li2022learning} could potentially make it discretization-invariant but showing this holds remains to be proven.
\begin{figure}[H]
\centering
\input disc_free.tex
\caption{Illustration of discretization-invariant. Discretization-invariant (right two images) does not require different samples to share the identical discretization of the parameter space, i.e. parameters can be evaluated at different sets of points (dots). }
\label{intro_dis_free}
\end{figure}
A benefit to discretization-invariance is the potential for avoiding aliasing between samples, which is a particular issue for the sampled parametric dependencies in multi-scale problems. This can impact both the network training and the overall accuracy. We illustrate this idea in Figure \ref{intro_dis_free2}.
\begin{figure}[h]
\centering
\begin{tikzpicture}[scale = 0.4]
\begin{axis}[domain=0:2*pi, legend pos= north east, ylabel = $sin(y)$, xlabel = $y$, height = 12cm]
\addplot[blue, samples = 200] {sin( deg(x) ) };
\addplot[only marks, mark=*, mark size=8pt, mark options={fill=red}, samples = 3] {sin( deg(x) )};
\legend{Parameter function of sample 1, Sampled parameter values}
\node[text width = 6cm] at (3, 6) {Sample 1 parameter function};
\end{axis}
\end{tikzpicture}
\begin{tikzpicture}[scale = 0.4]
\begin{axis}[domain=0:2*pi, legend pos= north east, legend style={fill=none}, ylabel = $sin(4y)$, xlabel = $y$, , height = 12cm]
\addplot[blue, samples = 200] {sin( deg(4*x) ) };
\addplot[only marks, mark=*, mark size=8pt, mark options={fill=red}, samples = 3] {sin( deg(4*x) )};
\legend{Parameter function of sample 2, Sampled parameter values};
\end{axis}
\end{tikzpicture}
\begin{tikzpicture}[scale = 0.4]
\begin{axis}[domain=0:2*pi, legend pos= north east, ylabel = $sin(y)$, xlabel = $y$, , height = 12cm]
\addplot[blue, samples = 200] {sin( deg(x) ) };
\addplot[only marks, mark=*, mark size=8pt, mark options={fill=red}, samples = 3] {sin( deg(x) )};
\legend{ Parameter function of sample 1, Sampled parameter values}
\node[text width = 6cm] at (3, 6) {Sample 1 parameter function};
\end{axis}
\end{tikzpicture}
\begin{tikzpicture}[scale = 0.4]
\begin{axis}[domain=0:2*pi, legend pos= north east, legend style={fill=none}, ylabel = $sin(4y)$, xlabel = $y$, height = 12cm]
\addplot[blue, samples = 200] {sin( deg(4*x) ) };
\addplot [only marks, mark=*, mark size=8pt, mark options = {fill = red}] coordinates { (0,0) (2*pi-pi/8, -1 ) (pi-pi/8-pi/4, 1 ) };
\legend{ Parameter function of sample 2, Sampled parameter values};
\end{axis}
\end{tikzpicture}
\caption{Illustration of discretization-invariant. Discretization-invariant (lower two images) can allow sampling at different sensors in the domain for two input functions, such that the sampled function values are different. However, not being discretization-invariant (top two images) may result in the sampled function values being the same for two different samples. }
\label{intro_dis_free2}
\end{figure}
\subsection{{Predication-Free} and {Domain-Independent}}
The goal is to build the neural operator $G(u)$ such that one can evaluate the function values $G(u)(x)$ for $x\in K_2$, i.e. for interpolation or prediction. We say that the neural operator is {prediction-free} if the evaluation point $x$ can be any point in $K_2$ (output function is free of discretization), independent of the choice of the discretization of the input function used in the training of the neural operator. Figure \ref{intro_pred_free} provides an example of {prediction-free} (the mesh on the right) as opposed to the first two plots, which are not {prediction-free} in the sense that the output mesh depends on the choice of the input mesh.
\begin{figure}[H]
\centering
\input pred_free.tex
\caption{Illustration of {prediction-free}. }
\label{intro_pred_free}
\end{figure}
By construction, the DON is a {prediction-free} approximation based on UAP-O. Vanilla FNO is not {prediction-free} since the output mesh must match the discretization of the input mesh \cite{lu2022comprehensive}. While requiring a neural operator to be prediction-free may be challenging, it can be shown to have several benefits. One advantage is that it leverages the operator learning principle and overcomes the limitations of function learning, e.g. lower network complexity, less dependence on data sampling process, and testing flexibility (see also Section \ref{sec_intro}).
In addition, FNO requires that the input and output function spaces be defined on the same domain, so it is {domain-dependent}. We call an operator learning framework {domain-independent} if the output function domain is independent of the input function domain. As an example of domain-independent, see Figure \ref{intro_domain_free}. Domain-independent methods provide the flexibility to solve problems from broader areas. One example is the parametric time-dependent PDE. While the parametrized input function may only have the spatial dependence, the solution operator relies on both space and time.
\begin{figure}[H]
\centering
\input domain_free.tex
\caption{Illustration of domain-independence. }
\label{intro_domain_free}
\end{figure}
\vspace{0.5em}
\subsection{Mesh-Free}\label{goals}
A flexible neural operator tool is one that is {mesh-free}, that is, a technique that is {discretization-invariant}, {domain-independent} and {prediction-free}. As a comparison, we summarize the comparison between the neural operators in Table \ref{intro_table}, see also \cite{lu2022comprehensive}. We will show that the Basis Enhanced Learning Network (BelNet) is a {mesh-free} extension of the current state-of-the-art neural operators.
\begin{table}[h]
\centering
\begin{tabular}{| c | c | c | c|}
\hline
& Discretization-invariant & Prediction-free & Domain-independent\\
\hline
DON & \textcolor{red}{\xmark} & \textcolor{green}{\checkmark} & \textcolor{green}{\checkmark} \\
\hline
FNO & \textcolor{green}{\checkmark} & \textcolor{red}{\xmark} & \textcolor{red}{\xmark}\\
\hline
BelNet & \textcolor{green}{\checkmark} & \textcolor{green}{\checkmark} & \textcolor{green}{\checkmark} \\
\hline
\end{tabular}
\caption{Comparison of DON, FNO and our BelNet.}
\label{intro_table}
\end{table}
\section{BelNet Formulation}
\label{sec_bel}
Similar to FNO, BelNet is derived as a nonlinear generalization of a linear kernel operator.
\textbf{Linear Setting}: We want to approximate an operator $G: V\rightarrow U$ as discussed in Equation (\ref{intro_operator}). As a motivating example, assume that the underlying kernel $\kappa(x, y)$ can be expanded in the following way: $\kappa(x, y) = \sum_{k = 1}^Kp_k(y) q_k(x)$, then
\begin{align*}
G(u)(x) = \int \kappa(x, y)u(y)dy = \sum_{k = 1}^K q_k(x)\int p_k(y)u(y)dy.
\end{align*}
Note that this expansion of the kernel can hold in a variety of useful settings. For example, if the kernel is positive definite then there exists a feature map $\phi$ such that $\kappa(x, y)=\langle \phi(y), \phi(x) \rangle$ by Mercer's Theorem. If the kernel is low-rank (or, more generally, if we approximate it by a low-rank expansion) then we can write the kernel in the form $\kappa(x, y) = \sum_{k = 1}^Kp_k(y) q_k(x)$ where $p_k$ and $q_k$ are the right and left singular functions rescaled by the singular values. To approximate the integral, we introduce a quadrature rule:
\begin{align}
G(u)(x) = \sum_{k = 1}^K q_k(x)\int p_k(y)u(y)dy \approx \sum_{k = 1}^K q_k(x)\sum_{j = 1}^N p_{kj}u_{j}, \label{intro_formulation2}
\end{align}
where $u_j = u(y_j)$, for $y:= [y_i, ..., y_N]^\intercal\subset K_1$, and $p_{kj}\in\mathbb{R}$ for all $j, k$ are the weights defined by $p_k(y_j)$ rescaled by the quadrature coefficients. Thus the natural neural operator in the linear setting becomes
\begin{align}
G_\theta(u)(x) = \sum_{k = 1}^K q_k(x) \, \sum_{j = 1}^N p_{kj}u_{j}= \sum_{k = 1}^K q_k(x) \, p_{k}^\intercal u,
\label{form_linear}
\end{align}
where we can train each function $q_k(x)$ by a standard MLP and collect all of the weights used in the MLP's along with $p_{kj}$'s into a vector of trainable parameters $\theta$.
\textbf{Nonlinear Setting}: To generalize Equation~\eqref{form_linear} to a nonlinear operator, we introduce a nonlinear function $g$:
\begin{align*}
G(u)(x) \approx \sum_{k = 1}^K q_k(x)\, g\left(\sum_{j = 1}^N p_{kj}u_{j}\right),
\end{align*}
which incorporates nonlinear dependencies on the points $x$ through the basis functions $q_k$ and nonlinear dependencies in the function value $u_j$. We construct a neural operator $G_{\theta}$ to approximate by training the functions $q_k$ and $g$ and the values $p_{kj}$.
In particular, we introduce weights and biases, $q^{k}\in\mathbb{R}^{d}$, $W_y^{1, k}\in\mathbb{R}^{N_1\times N}$, $W_y^{2, k}\in\mathbb{R}^{N\times N_1}$,
$b_x^k\in \mathbb{R}$, $b_y^k\in\mathbb{R}^{N_1}$, where $k = 1, ..., K$, and activation functions $a_x$, $a_y$ and $a_u$
such that,
\begin{align}
G(u)(x) \approx G_{\theta}(u)(x) = \sum_{k = 1}^K a_x \left((q^{k})^\intercal x+ b_x^k \right) \, a_u \left(\hat{u}^\intercal W_{y}^{2, k}\left(a_y(W_{y}^{1, k} y + b_y^k ) \right) \right),
\label{intro_formulation3}
\end{align}
for $x\in K_2\subset \mathbb{R}^d$, $u\in V$, and where $\hat{u} = [u(y_1), ..., u(y_N)]^\intercal\in\mathbb{R}^N$.
The motivation for this formulation is its {discretization-invariant} nature. To see this, let us assume $V \subset \mathbb{P}_N$, $p_k(y)>0$ and $p_k(y)\in L^1(\Omega)$ where $\Omega$ is the domain.
For any set of distinct points $\{y_j\}_{j = 0}^N\subset I$, we can define the associated Lagrange basis as,
\begin{align*}
h_j(y) = \prod_{l=0, k\neq j}^N\frac{y-y_j}{y_j-y_l},
\end{align*}
for $j\in[0, N]$.
It is known that \cite{shen2011spectral},
\begin{align*}
\int_I p_k(y)u(y)dy = \sum_{j = 0}^N u(y_j)w_{kj}
\end{align*}
where,
$w_{kj} = \int_I h_j(y)p_k(y)dy, j\in[0, N].$
That is, the approximation of the integral in Equation \ref{intro_formulation2} is exact. However,
it should be noted that $\{y_j\}$ are the sensors of the input function and arbitrary, indicating that the method is independent of the discretization.
\textbf{Full Formulation}: BelNet trains both functions $p$ and $q$ directly. Intuitively, we project $u\in V$ into space spanned by trainable basis $p$, thus the name \textit{basis enhanced learning}. The network also constructs the trainable basis $q$ for the $x$-dependence, which we denote as \textit{construction net}. The entire architecture is displayed in Figure \ref{fig_network_s1}.
\begin{figure}[H]
\centering
\scalebox{.85}{\input{net_structure.tex}}
\caption{Illustration of the network structure. Projection nets are $K$ independent fully connected neural network with weights and bias $W_y^{2, k}\in\mathbb{R}^{N\times N_1}$, $W_y^{1, k}\in\mathbb{R}^{N_1\times N}$ and $b_y^k\in\mathbb{R}^{N_1}$. Construction net is a fully connected neural network with weights and bias $Q\in\mathbb{R}^{K\times d}$ and $b_x\in\mathbb{R}^d$. Here $Q = [q^1, q^2, ..., q^K]$, where $q^i$ are defined in Equation (\ref{intro_formulation3}). In addition, $a_x, a_y, a_u$ are activation functions. }
\label{fig_network_s1}
\end{figure}
\subsection{Connection to FNO}
\label{sec_intuition}
The FNO form is motivated by iterating a block of linear layers in the Fourier domain, which is essentially a convolution representation for linear operators. To explain the connection between BelNet and FNO, let's consider the Fourier integral operator (Equation 4 in \cite{li2020fourier}). Essentially, the basis elements are trained in the full FNO form. In particular, consider the discrete Fourier transformation (DFT) of two functions $u$ and $h$, denoted as:
\begin{align}
&\mathcal{F}(u)(k) = \sum_n u(n)b(n, k), \nonumber\\
&\mathcal{F}(h)(k) = \sum_n h(n)b(n, k),
\label{DFT}
\end{align}
where $b(n, k) = e^{-i2\pi nk}$ is the Fourier basis where $k$ denotes the wave number. By the discrete convolution theorem, we have:
\begin{align}
u*h(m) = \sum_n u(n)h(m-n) = \mathcal{F}^{-1}\big(\mathcal{F}(u*h)\big) = \mathcal{F}^{-1}\left(\mathcal{F}(u)(k)\,\mathcal{F}(h)(k)\right).
\label{convolution}
\end{align}
Substituting Equation~\eqref{DFT} into Equation~\eqref{convolution}, it follows that,
\begin{align*}
u*h(m) & = \sum_k\left( \sum_n u(n)b(n, k)\right) \left( \sum_n h(n)b(n, k) \right)w(k, m) = \sum_k w(k, m) \sum_n r(n, k) u(n)
\end{align*}
where $w(k, m) = e^{i2\pi km}$ is the inverse DFT basis, and:
\begin{align*}
r(n, k) = b(n, k)\sum_lh(l)b(l,k).
\end{align*}
Thus the linear FNO trains the following neural operator:
\begin{align*}
G(u)(x) \approx \mathcal{FNO}(u)(m) = \sum_n u(n)h(m-n) = \sum_k w(k, m) \sum_n r(n, k) u(n)
\end{align*}
where $h$ is the kernel function. As a result, the linear Fourier integral operator is a special case of Equation~\eqref{form_linear}, where $b$ and $w$ play the role of $q$ and $p$ in Equation~\eqref{form_linear}. Note that the full FNO form includes an additional skip-connection and activation layer, which makes FNO a nonlinear approximation. Note that even in the full form, FNO uses the FFT and thus relies on the Cartesian domain with a lattice
grid mesh \cite{lu2022comprehensive}. BelNet does not have this restriction.
\section{Numerical Examples}
\label{sec_numerical}
We focus on four test PDEs: Burgers' equation, often used to compare neural operators, and three other challenging multiscale problems. While of great interest in the applied community \cite{hou1997multiscale, efendiev2013generalized}, multiscale and high-contrast problems pose several difficulties for neural operators. More precisely, one often needs a fine enough resolution to capture the small-scale features, which leads to intractable training problems. In addition, observed data is often mesoscale or macroscale, and thus one needs to infer the homogenized solution operator. Our numerical experiments focus on predicting the solution based on data scarce observations.
For each example, we test BelNet, stacked DON, and unstacked DON. In each case, we present the best results over the tests. We use two variants of BelNet; the first adopts the discretization used in DON experiments, which is referred to as fix-BelNet (since the mesh is the same for all samples).
The second variant uses a randomized discretization for each sample; that is, each input function has a randomized mesh. This is referred to as free-BelNet. As a benchmark, we also refer to established methods \cite{efendiev2013generalized,hou1997multiscale,chetverushkin2021computational,chung2022computational,chung2018constraint} to solve multiscale parametric PDEs.
\subsection{Viscous Burgers' Equation}
\label{sec_vburgers}
Consider the viscous Burgers' equation with periodic boundary conditions:
\begin{align*}
&\frac{\partial u_s}{\partial t} + \frac{1}{2}\frac{\partial (u^2_s)}{\partial x} = \alpha \frac{\partial^2 u_s}{\partial x^2},\hspace{0.2em} x\in[0, 2\pi], \hspace{0.2em} t\in[0, 0.3]\\
&u_s(x, 0) = u^0_s(x),\\
&u_s(0, t) = u_s(2\pi, t),
\end{align*}
where $u^0_s(x)$ is the initial condition that depends on the parameter $s$ and the viscosity is set to $\alpha = 0.1$.
We will consider the operator that maps from the initial condition to the terminal solution at $t = 0.3$.
\textbf{Training Data:} To generate the initial conditions, we first compute the short-time solution ($t=0.1$) to Burgers' equation. We use the periodic boundary conditions, set the viscosity to zero, and use the initial condition $s\sin(x)$ where $s\in [0, 4]$. The solution of the system at $t = 0.1$ is then used as the initial condition $u^0_s$; see the yellow and blue curves in Figure \ref{pic_vburgers_info}. This is used to generate more variability between initial samples for the training phase since different values of $s$ will lead to different levels of sharpness in the slopes for the initial data, even in the short time interval used.
\begin{figure}[H]
\centering
\includegraphics[scale = 0.55]{graphs/vb_ic1.png}
\includegraphics[scale = 0.55]{graphs/vb_ic2.png}
\caption{Plots of two solutions to the viscous Burgers' equation with our initialization procedure. Note that each example's sampling points (black dots) for the initial condition differ. The yellow curves are used to generate the initial conditions for the model problem (viscous Burgers' equation). The initial conditions for the viscous Burgers' equation are displayed in blue.}
\label{pic_vburgers_info}
\end{figure}
The mesh for the input data is as follows. Each initial condition (input function) has $25$ sensors, and we used a total of $200$ initial conditions for training. For each initial condition, the true system is evolved up to time $t=0.3$, and a total of $5$ time-stamps are collected (the terminal time is not included). Therefore the space-time mesh contains $25$-by-$5$ total sample locations for each initial condition.
\textbf{Testing:} To test the neural operators' ability to extrapolate future states, we do not include the solution at the terminal time $t=0.3$ in the training dataset. For testing, we use solutions from $500$ initial conditions and test each neural operator on the solution at the terminal time with a finer mesh of $151$ grid points.
We present the relative errors in Table~\ref{vburgers_results}. We compare BelNet with the (vanilla) DON, where BelNet uses either fixed sensor points matching the discretization in DON or ``free'' sensors that are independent of each sample. With a similar amount of trainable parameters listed in Table~\ref{vburgers_results}, BelNet performs similarly or better in terms of the relative prediction error over the 500 testing sequences.
\begin{table}[H]
\centering
\begin{tabular}{||c c c c c||}
\hline
& DON & fix-BelNet & free-BelNet & FNO\\ [0.5ex]
\hline
Relative error (\%) & $1.17$ & $0.59$ & $1.07$ & $1.61$ \\ [0.5ex]
\hline
Total parameters (in K) & $103.7$ & $102$ & $102$ & $287.425$\\ [0.5ex]
\hline
\end{tabular}
\caption{Viscous Burger's equation example mean relative error of $500$ testing solutions of four experiments. The FNO structure is from \cite{li2020fourier}.}
\label{vburgers_results}
\end{table}
\textbf{BelNet Hyperparameters}: The number of sensors (for the initial condition) is set to $N = 25$. For the projection net, ``$u$'' is connected with $K = 10$ independent subnetworks with identical structures.
Specifically, each sub-network is fully connected and has the structure $[25, 100, 25] \quad (i.e.\ N = 25, N_1 = 100)$. Here $25$, $100$, and $25$ are the input dimension, the width of the first hidden layer (the number of neurons in the first layer), and the output dimension, respectively.
This network representation routine will be used in the rest of the manuscript. In addition, bias and activation (Tanh) are added to the first layer according to the formulation (\ref{intro_formulation3}).
\textcolor{black}{To achieve better accuracy, we replace $a_u$ in Figure \ref{fig_network_s1} with a ReLU-activated trainable layer of size $10\times 10$.} More precisely, we use a ReLU-activated trainable matrix of size $10\times 10$ to replace the activation function $a_u$ in Figure \ref{fig_network_s1}.
For the construction net, the input dimension equals $2$ ($1$ in space and $1$ in time). Hence it is connected with a fully connected network, which has the structure $[2, 100, 100, 100, 10] \quad (i.e.\ K = 10)$. Each layer except the last layer has a bias vector and is activated by the Tanh function.
\subsection{Multiscale Elliptic Equation}
We consider the following multiscale elliptic equation:
\begin{align*}
& -\nabla \cdot (\kappa \nabla u) = f(x, s),\ \ x\in s = [0, 1]^2, \ \ s\in D\\
& u(x) = 0,\ \ x\in\partial \Omega,
\end{align*}
where $s$ is the parameter. The multiscale permeability is denoted by $\kappa$ and an example is shown in Figure \ref{pic_elliptic_info}, specifically:
\begin{align*}
\kappa(x_1, x_2) = 1 + \frac{\sin(2\pi\frac{x_1}{\epsilon_1})\cos(2\pi\frac{x_2}{\epsilon_2}) }{2+\cos(2\pi\frac{x_1}{\epsilon_1})\sin(2\pi\frac{x_2}{\epsilon_2})}
+ \frac{\sin(2\pi\frac{x_1}{\epsilon_1})\cos(2\pi\frac{x_2}{\epsilon_3}) }{2+\cos(2\pi\frac{x_1}{\epsilon_1})\sin(2\pi\frac{x_2}{\epsilon_3})},
\end{align*}
where the parameters are set to $\epsilon_1 = 1/4$, $\epsilon_2 = 1/8$ and $\epsilon_3 = 1/16$.
The multiscale nature of this problem comes from the permeability $\kappa(x_1, x_2)$, which has multiple frequency scales. General machine learning-based approaches have difficulty approximating such spatial systems due to the multi-frequency behavior of the solution \cite{leung2022nh}.
The source function, denoted by $f$, is parameterized by taking the weighted sum of $9$ Gaussian distributions whose centers are uniformly distributed in the domain. A demonstration is presented in Figure \ref{pic_elliptic_info}. We sample the weights for the Gaussians in the source function such that they are uniformly from $[-1, 1]$. The training problem is to learn the map from the source $f$ to the solution of the elliptic equation, $u$.
\begin{figure}[t]
\centering
\includegraphics[scale = 0.5]{graphs/elliptic_kappa.png}
\includegraphics[scale = 0.4]{graphs/elliptic_distribution_centers.png}\\
\includegraphics[scale = 0.5]{graphs/elliptic_source_1.png}\hspace{0.1em}
\includegraphics[scale = 0.5]{graphs/elliptic_source_2.png}
\caption{Plots (from top left to bottom right): the multiscale permeability, centers of all normal distributions, one realization (sample) of the source, a second realization of the source. }
\label{pic_elliptic_info}
\end{figure}
\textbf{Training/Testing Data:} Using the parametric form for $f$, we generate $80$ different source functions for training. For each solution $u$ (corresponding to one source $f$), we choose $19\times 19$ grid points in $\Omega$ uniformly, leading to a total of $80\times 19 \times 19$ training ``samples''. For testing, we generate $100$ testing permeability coefficients and compute the predicted solution on a fine mesh using $100\times 100$ grid points. \textcolor{black}{Each parameter $s$ will determine an input function $f(x, s)$. To measure the input function, we uniform (randomly) placed $10\times 10$ sensors in the input function domain $\Omega$, i.e. $N = 100$. More precisely, we create a uniform coarse mesh of size $10\times 10$ and draw one sensor uniformly from each coarse element. }
\textbf{Results and Comparison:} We compare the results with provably convergent numerical methods in order to better evaluate our approach. This system is a classical multiscale problem, a standard finite element solver would be too slow since one would need a fine discretization level (to capture the multiscale features and obtain sufficient accuracy). Therefore, we compare our approach to the multiscale finite element method (MsFEM) \cite{hou1997multiscale, efendiev2013generalized}, which is an efficient and accurate method for this problem. The fine-scale mesh size is set to
$100\times 100$, and the coarse element size is set to $5\times 5$. Thus, we have $19\times 19$ degrees of freedom, which is the same as the number of solution points per training source used in the operator learning approaches in this example.
The results are summarized in Table \ref{elliptic_results}.
\begin{table}[H]
\centering
\begin{tabular}{||c c c c c||}
\hline
& DON (stack) & fix-BelNet & free-BelNet & MsFEM \\ [0.5ex]
\hline
Euclidean error (\%) & $16.44$ & $14.43$ & $9.66$ & $2.28$\\ [0.5ex]
\hline
Parameters count (in K) & $319.4$ & $312.4$ & $312.4$ & NA\\ [0.5ex]
\hline
\end{tabular}
\caption{Elliptic example mean relative error of $100$ testing solutions of three experiments.
MsFEM adopts $19\times 19$ DOFs while the learning methods are trained on the same $19\times 19$ mesh.
}
\label{elliptic_results}
\end{table}
\textbf{BelNet Hyperparameters:}
As before, we set $N = 100$. For the projection net, $u$ is connected with $K = 10$ independent networks with the structure $[200, 100, 100] \quad (i.e.\ N = 100, N_1 = 100)$. In addition, the bias is added to the first layer, and the first layer is activated by the ReLU function. Since the problem is linear in $f$, to accelerate the training, we removed the nonlinear activation $a_u$. For the construction net, the input dimension is equal to $2$ and is a fully connected network of a structure $[2, 100, 10] \quad (i.e.\ K = 10)$. Each layer except the last layer has a bias vector and is activated by the ReLU function.
\subsection{Radiative Transfer Equation (RTE) with High Contrast Scatter}
We consider the radiative transfer equation (RTE) \cite{newton2020diffusive, li2019diffusion, chung2020generalized} with a high contrast scatter coefficient $\sigma(x, \omega)$ (see Figure \ref{exp_demon_sampling}).
Specifically, the `high contrast' aspect comes from strong scattering in the channels. We assume that the channels are moving and the channels are characterized implicitly by the parameter $\omega$. The goal is to learn the mapping from $\sigma(x, \omega)$ with moving channels to the solution, $I(x,s)$, of the RTE problem:
\begin{align*}
s \cdot \nabla I(x,s) = \frac{\sigma(x, \omega)}{\epsilon} \bigg( \int_{\mathcal{S}^{n-1}} I(x,s') ds' - I(x, s)\bigg) \quad \forall x \in D, s \in \mathcal{S}^{n-1}.
\end{align*}
Here $s$ is a vector on the unit sphere, and $n$ is the dimension of the problem. In our experiments, we considered $n=2$ and thus $\mathcal{S}^{n-1}=\mathcal{S}^{1}$ is the unit circle.
In addition, we set $\epsilon = 1$ and $D = [0, 1]^2$.
To provide closure to the model, we introduce the Dirichlet boundary conditions $I(x,s) = I_{\text{in}}$ used for entrant directions $s \cdot \textbf{n} <0$, i.e. on $\Gamma ^{-} := \{ (x,s) \in \partial D \times \mathcal{S}^{n-1}: s \cdot \textbf{n} <0 \}$. Here, $\textbf{n}$ is the unit outward normal vector field at $x \in \partial D$. The condition is written as
\begin{align*}
I = I_{\text{in}} (x,s) \quad \text{ for all } (x,s) \in \Gamma^-.
\end{align*}
In our examples, the top, bottom, and right boundaries have zero incoming boundary conditions, while we assume the left boundary has non-zero flow injected into the domain.
We chose the multiscale RTE with high contrast channels since it is complicated to solve numerically \cite{chung2020generalized} and a challenging problem for learning-based approaches. In practice, it is impossible to measure $\sigma(x, \omega)$ at all $x\in D$. Instead, one samples $\sigma(x)$ on a mesh $D'$ computes a reduced order model $\hat{\sigma}(x, \omega)$. However, due to the nature of the scattering coefficient, i.e. sparse localized regions, the ROM for two different scatter coefficients may appear to be similar even when the original structures differ. This is illustrated in Figure \ref{exp_demon_sampling} and is often the case when one uses the same grids to construct the ROM. Consequently, a mesh-free version of operator learning should be used.
\begin{figure}[H]
\centering
\includegraphics[scale = 0.35]{graphs/s1_v28of_rte_data_kernel.png}
\includegraphics[scale = 0.35]{graphs/s2_v28of_rte_data_kernel.png}
\includegraphics[scale = 0.35]{graphs/rom1_v28of_rte_data_kernel.png}
\includegraphics[scale = 0.35]{graphs/rom2_v28of_rte_data_kernel.png}
\caption{Left two images: two scatter coefficients. Right two images: the corresponding ROMs of two scatter coefficients. The ROM is obtained by sampling $\sigma(x, \omega)$ in a uniform mesh $5\times 5$ in the domain, and two scatter coefficients are sampled by the same mesh. }
\label{exp_demon_sampling}
\end{figure}
\textbf{Training/Testing Data:} We generate $60$ scattering coefficients and compute the corresponding solutions $I(x,s)$. For each solution, we use $25\times 25$ points in $D$ uniformly sampled with $8$ velocities, which gives a total of $60\times 625 \times 8$ training samples. For testing, we used $100$ scatter coefficients and computed the corresponding solutions on a finer mesh of $50\times50\times 8$, where the last dimension is the velocity. \textcolor{black}{Each $\omega$ determines one scatter coefficient $\sigma(x, \omega)$. To measure each input function $\sigma(x, \omega)$, we draw $25$ sensors randomly uniformly from $D$, so $N = 25$. More precisely, we create a uniform discretization of size $5\times 5$ and draw one sensor randomly from each element in the mesh.}
\begin{table}[H]
\centering
\begin{tabular}{||c | c| c| c | c||}
\hline
\backslashbox{Model}{CR} & $2$ & $10$ & $50$ & $100$ \\ [0.5ex]
\hline
fix-BelNet (279K) & $3.68 $ & $8.36 $ & $13.70$ & $16.40$\\ [0.5ex]
\hline
free-BelNet (279K)& $3.56$ & $7.28$ & $12.30$ & $13.31$\\ [0.5ex]
\hline
DON (294K)& $4.39$ & $15.12$ & $18.30$ & $19.18$\\ [0.5ex]
\hline
\end{tabular}
\caption{RTE mean relative error of $100$ testing solutions for different contrast ratios (CR). Fix-BelNet and free-BelNet share the structure and have 279K trainable parameters, DON has $294K$ parameters. }
\label{rte_cr_table}
\end{table}
\textbf{Results and Comparison:} Table \ref{rte_cr_table} summarizes the results. As expected, in all three models the performance deteriorates as the contrast ratio (CR) increases. However, it is worth noting that the grid-free version of BelNet seems to produce a small relative error with a smaller jump between various CR values.
This may be due to underlying sampling of the grid points, namely since the grids are randomly generated free-BelNet sees more variations between inputs.
Since fix-BelNet has a lower relative error than the DON, this indicates that incorporating mesh coordinate information indeed improves the model's accuracy.
\textbf{BelNet Hyperparameters:} We set $N = 25$, the projection net's parameters are $K = 30$ with fully connected structure of size $[50, 100, 25] \quad (i.e.\ N = 25, N_1 = 100)$. As before, the bias is added to the first layer and the first layer is activated by the ReLU function. For the construction net, the input dimension is equal to $4$ ($2$ in space and $2$ in velocity), hence it has the structure $[4, 100, 100, 30] \quad (i.e. \ K = 30)$. Each layer except the last layer of the construction net has a bias vector and is activated by the ReLU function.
\subsection{Richard's Equation}
The last example is the nonlinear Richard's equation \cite{efendiev2022efficient, efendiev2022hybrid}:
\begin{align*}
& u_t(x, t) - \nabla \cdot e^{\kappa(x) u(x, t)} \nabla u(x, t) = f(x), \ \ x\in\Omega = [0, 1]^2, \ \ t\in[0, T],\\
& \frac{\partial u(x, t)}{\partial n} = 0, \ \ x\in\partial \Omega,
\end{align*}
with a high-contrast multiscale contrast permeability $\kappa (x)$\cite{chung2022contrast, efendiev2022efficient, efendiev2022hybrid} (see Figure \ref{kappa_richard}). The source is defined as $$f(x, y) = 10^4 \, \exp\left(\frac{-(x-0.5)^2-(y-0.5)^2}{0.1}\right).$$
\begin{figure}[H]
\centering
\includegraphics[scale = 0.5]{graphs/kappa_richard_update.png}
\includegraphics[scale = 0.5]{graphs/source_richard.png}
\caption{Left: the permeability field $\kappa(x)$ used in Richard's equation. The permeability in the yellow channels is equal to $1$ while the permeability in the background is equal to $10^6$. Right: the source.}
\label{kappa_richard}
\end{figure}
\textbf{Training/Testing Data:} We map the initial condition to the terminal solution at time $t = 2\times 10^{-5}$.
The initial condition is defined as the weighted sum of $9$ Gaussian distributions, where the weights are uniformly random from $[-2, 2]$. The training set is generated from $60$ initial conditions, where for each initial condition, we observe the solution on a $40\times 40\times 7$ mesh, i.e. $40\times 40$ spatial points and $7$ time-snapshots. For testing, we test the model on $100$ solutions only at the terminal time. For the test solutions, we use a (finer) spatial mesh of size $50\times 50$. To measure the initial condition, free-BelNet randomly draws $25$ points for each sample, while the DON chooses $25$ fixed points for all training and testing samples; in addition, we also test fix-BelNet with the same fixed points used in DON.
\textbf{Results and Comparison:} We compare our results to DON and the solution generated by the generalized multiscale finite element method (GMsFEM) \cite{chetverushkin2021computational, chung2022computational}. For GMsFEM, the fine mesh size is set to $50\times 50$, the coarse element size is $2\times 2$, and we use the first $3$ GMsFEM basis. As a result, GMsFEM has $1728$ degrees of freedom, which is close to the $40\times 40$ training mesh points used in the operator learning models. The results are summarized in Table \ref{40nl_results}.
\begin{table}[H]
\centering
\begin{tabular}{||c c c c c c||}
\hline
& DON (unstack) & DON (stack) & fix-BelNet & free-BelNet & GMsFEM \\ [0.5ex]
\hline
Euclidean error (\%) & $11.98$ & $21.18$ & $7.87$ & $7.64$ & $32.43$\\ [0.5ex]
\hline
Parameters count (in K) & $103$ & $134$ & $127$ & $127$ & NA \\ [0.5ex]
\hline
\end{tabular}
\caption{Richard's equation mean relative error of $100$ testing solutions of three experiments. GMsFEM adopts $1,728$ DOFs while the learning methods are trained on $40\times 40$ mesh.}
\label{40nl_results}
\end{table}
\textbf{BelNet Hyperparameters:} We set $N = 25$. For the projection net, we use $K = 10$ and the fully connected network has the structure $[50, 100, 25] \quad (i.e.\ N = 25, N_1 = 100)$, with bias in the first layer and with the Tanh activation function.
In addition, we replace $a_u$ in Figure \ref{fig_network_s1} with a Tanh-activated trainable layer of size $10\times 10$, which improves accuracy. For the construction net, we use a $[3, 100, 100, 10] \quad (i.e.\ K = 10)$ fully connected neural network with bias and Tahn activation on each layer except the last layer.
\section{Conclusions and Future Works}
We proposed a mesh-free operator learning technique called BelNet. In particular, a method that does not depend on the grid locations for the input function and allows for the domain and discretization of the inputs and outputs to differ. Mesh-free approaches make it possible to solve challenging PDE problems, such as multiscale systems where functions can be aliased more often for fixed grids.
While the form of BelNet is motivated by the convolution representation for linear operators related to PDE, we provide a nonlinear generalization using additional activations. Through several detailed examples, we show that BelNet can outperform standard and recent approaches on important parametric PDEs.
There are several future directions worth investigating. First, the universal approximation theorem for BelNet should follow from a generalization of \cite{chen1995universal}, but one needs to show that the approximation does not require fixed or similar grids. Secondly, in order to improve the computational and storage cost, we plan to optimize the parameter count in the projection net. In particular, leveraging the success of random feature projections, we have seen some preliminary improvements to the efficiency of BelNet. Lastly, while we focused on the applications to multiscale PDE, BelNet can also be applied to complicated input controls for dynamical systems with controls.
\section{Acknowledgement}
Z. Zhang was supported in part by AFOSR MURI FA9550-21-1-0084. H. Schaeffer was supported in part by AFOSR MURI FA9550-21-1-0084.
\bibliographystyle{abbrv}
|
\section{Introduction}
Currently available quantum hardware do not have the capability to perform fully fault tolerant quantum computation due to the non-availability of high fidelity qubits. However, due to the inherent decoherence in quantum mechanics from interactions of the system with the environment, errors are inevitable and any quantum computation is affected by noise. Various error mitigation techniques have been developed that do not require the large quantum overhead necessary in fault tolerant quantum error correction. Some examples are: zero noise extrapolation \cite{Temme2017_ErrorMitigForShorDepthQuantCirc, 1805.04492, tudor2020_digitalZeroNoiseExtrapForQuantErrMitig}, probabilistic error cancellation \cite{Temme2017_ErrorMitigForShorDepthQuantCirc}, readout error mitigation \cite{Nachman_2020ReadoutNoise}, Pauli check sandwiching (PCS) \cite{Debroy_2020ExtFlagGadgetsForLowOverCircVer, gonzales2022quantum}, and symmetry verification \cite{Bonet-Monroig2018_LowCostErrMitigBySymmVerif, McArdle2019_ErrMitigDigitalQuantSimulation, shaydulin2021error, Cai2021quantumErrorMitigSymmExpan}.
\begin{figure}[h]
\centering
\includegraphics[width= \linewidth]{figures/general_cutting2.pdf}
\caption{Circuit cutting framework for quantum error mitigation}
\label{fig:framework}
\end{figure}
Many of these quantum error mitigation techniques insert additional gates and ancillas that need to be executed with the main circuit. One of the drawbacks of these techniques is that these error mitigation subcircuits introduce noise associated with the additional gates and ancillas that they require to run alongside the main circuit on the quantum hardware. In this work, we introduce the Simulated Quantum Error Mitigation (SQEM) framework, where the error mitigation subcircuits are run on a classical simulator whereas the main circuit runs on a quantum device. Essentially, our framework converts error mitigation techniques that require additional gates and ancillas almost entirely into classical pre and post processing procedures and thus makes the execution of the error mitigation protocol almost noiseless.
The SQEM framework relies on the quantum circuit cutting also known as ``quantum knitting". The quantum circuit cutting techniques allow us to divide large quantum circuits into smaller fragments~\cite{peng2020simulating, perlin2021quantum, ayral2021quantum,ayral2020quantum, tang2021cutqc}. These fragments can then be run on separate quantum devices and then the output distributions from each of these fragments can be combined to generate the output of the full quantum circuit. In our error mitigation framework, we fragment the quantum circuit into two parts: the main/original quantum circuit which is executed on a quantum device and the quantum error mitigation subcircuits which are sent to a classical simulator. This is demonstrated in Fig.~\ref{fig:framework}. This allows us to error mitigate almost noise free. The state preparation and measurement (SPAM) errors due to circuit cutting is also mitigated by the mitigating circuit. In our experiments, the error mitigation method that we use is Pauli check sandwiching but other error mitigation schemes can also be used.
One of the objections often raised with the circuit cutting technique is that the computational cost is exponential in the number of cuts. Therefore, if we were to error mitigate all the qubits at once we will have to incur an exponential computational cost. Our framework avoids this scalability problem because we error mitigate one qubit at a time. We provide examples of quantum circuits where this is possible. Each of the instances where we error mitigate a qubit generates a probability distribution.
We introduce a method for recombining these different probability distributions such that the re-combined probability distribution is error mitigated from the noise across all qubits. We demonstrate the efficacy of our framework by hardware experiments. Specifically, we perform three variational quantum eigensolver (VQE) experiments to estimate the ground state energies for the molecules LiH, BeH$_2$, and HF. Our results match the noiseless ansatz result to a very high accuracy as shown in Figs.~\ref{subfig:results_lih} and \ref{subfig:results_HF}.
\section{SQEM framework}
Simulated quantum error mitigation (SQEM) transforms the error mitigation components of an input circuit into classical pre and post processing procedures by applying circuit cutting. The fragments that perform error mitigation are sent to a classical computer to be simulated and the main circuit is sent to the quantum hardware. This procedure enables the gates that comprise the error mitigation scheme to be noiselessly executed, while the original unmitigated quantum computation is run on quantum hardware. Fig~\ref{fig:framework} provides a schematic view of this protocol.
\subsection{Quantum Circuit Cutting}
A quantum circuit can be represented as a graph. In Ref.~\cite{peng2020simulating} it was shown how one can fragment such a graph into two or more subgraphs or subcircuits. These subgraphs can be cut by performing a complete set of measurements on the output of one subcircuit at a cut, and then preparing a corresponding set of states on the input of the other subcircuit at that cut. In practice, one can independently execute and measure the outputs of these subcircuits and classically post-process these results to obtain the output of the original uncut quantum circuit \cite{perlin2021quantum}.
Circuit cutting works by resolving an $n$-qubit quantum state $\rho$ and inserting a set of complete measurement basis,
\begin{equation}
\rho \simeq \f12 \sum_{M\in\mathcal{B}} M \otimes \tr_n\p{M_{(n)}\rho},
\label{eq:cut_identity}
\end{equation}
where $\simeq$ means equality up to a permutation of the qubit order; $\mathcal{B}$ is a basis of Hermitian $2\times 2$ matrices with normalization $\tr\sp{M^{(i)} M^{(j)}} = 2 \delta_{ij}$ for $M^{(i)},M^{(j)}\in\mathcal{B}$; $\tr_n$ is the partial trace of qubit $n$; and $M_{(n)}$ is an operator that acts with $M$ on qubit $n$ and with identity on all the other qubits.
We use the set of Pauli operators plus the identity operator, $\mathcal{B}\equiv\{I,X,Y,Z\}$, as our basis.
\subsection{Pauli Check Sandwiching}
Pauli Check Sandwiching~\cite{gonzales2022quantum} (PCS) uses pairs of checks to error mitigate the input circuit $U$ \cite{Debroy_2020ExtFlagGadgetsForLowOverCircVer,gonzales2022quantum}. Each pair of check sandwiches $U$ and leaves it invariant.
We use a set of $n$ pairs of unitaries $\{\tilde{C}_1^{(i)}, \tilde{C}^{(i)}_2\}$, where $i$ represents the $i^\text{th}$ pair and each pair satisfies the condition
\begin{align}\label{eq:conditionOnChecks1}
\tilde{C}_2^{(i)}U\tilde{C}^{(i)}_1=U.
\end{align}
Then the checks for the circuit are given by the set of pairs of controlled unitaries $\{C_1^{(i)}, C_2^{(i)}\}$
\begin{align}
C_1^{(i)}&=\tilde{C}^{(i)}_1\otimes\op{-}_{i}+\mathbb{I}\otimes\op{+}_{i}\\
C_2^{(i)}&=\tilde{C}^{(i)}_2\otimes\op{-}_{i}+\mathbb{I}\otimes\op{+}_{i},
\end{align}
where the control is on the $i^{th}$ ancilla and the target are the compute qubits. The error mitigation scheme uses the circuit $C_2^{(n)}\cdots C_2^{(2)}C_2^{(1)}UC_1^{(1)}C_1^{(2)}\cdots C_1^{(n)}$ and post selects on the zero outcome for a $Z$ basis measurement on the ancillas. In the ideal case that the error mitigation part of the circuit is noiseless and given arbitrary noise on the compute qubits, there always exists a finite number of checks such that the post selected state is noiseless.x
For this research, we apply the PCS scheme on a hardware-efficient VQE ansatz, which consists of two layers of SU(2) gates and a cascade of controlled-Z gates. The check pairs we use are Z-gates with extra single-qubit gates and we check only a single qubit at a time as shown in Fig.~\ref{fig:SQEM_overview}. The single qubit check Pauli-Z checks are possible because they commute with the ansatz along with the SU(2) gates. The Pauli-Z checks remove all the Pauli-X and Pauli-Y errors.
In PCS, the post selection probability exponentially decreases as the number of pairs of checks increases~\cite{gonzales2022quantum}. Our strategy resolves this problem by error mitigating one qubit at a time. With our strategy the total overhead in the number of runs required is only a polynomial function of the number of pairs of checks used.
\begin{figure*}[htbp]
\centering
\includegraphics[width=\textwidth]{figures/overall_demo3.pdf}
\caption{Overview of the SQEM framework. The red blocks and the red blocks with a dagger in the PCS circuit are complex conjugates of each other. The framework involves three steps: 1) apply the error mitigation circuit for each individual qubit or a subset of qubits 2) cut out the error mitigation circuit fragment and run the original fragment and the mitigation fragment separately on a quantum device and a classical device. 3) recombine a probability distribution $P_R$ based on individual error mitigated distributions $P_0$ ... $P_n$ and a Bayesian probability recombination algorithm. }
\label{fig:SQEM_overview}
\end{figure*}
\subsection{Probabilities Recombination}
Applying the checks for each individual qubit or a subset of qubits makes our approach scalable. However, the mitigated probabilities from each check need to be recombined to mitigate the noise of the system. Here we use a Bayesian recombination model to recombine the probability distributions. The Bayesian recombination model is analogous to Bayesian updating~\cite{berger1994bayesianupdate} in statistics. It is developed based on the assumption that the protected/checked qubits' probability distribution has less noise than unprotected/unchecked ones.
Considering an n-qubit circuit, we use PCS with SQEM to check each qubit individually. For the n experiments we will get n distributions $P_{M\_1}$, ... , $P_{M\_n}$. $P_{UM}$ and $P_{M\_k}$ are the output probability distributions of the unmitigated experiment and that of the mitigated experiment which protects qubit $k$ respectively. Each classical state $\ket{x_1x_2...x_n}$ in the computational basis is represented as a binary string $s = [x_1, x_2, ..., x_n]$, $s \in \mathbb{B}_c$, where $\mathbb{B}_c$ is the computational basis. $s[k] = x_k$, $k \in [0, n]$ indicates the measurement result of the $k$th qubit in the string $s$. $P_{UM}(s)$ is the probability of getting the output $s$. For a distribution $P$, the probability of qubit $k$ having a measured result $\ket{0}$ over all the states is denoted as
\begin{equation}
w_0(P, k) = \sum_{s\in \mathbb{B}_c} P(s)\text{, if } s[k] = 0.
\label{eq:uw0}
\end{equation}
Similarly, we have the probability $w_1(P, k)$ for outcome $\ket{1}$. $w_0(P_{M\_k}, k)$ is considered as the ideal bitwise probability for qubit $k$ with outcome $\ket{0}$ since we use the probability from the experiment that protects qubit $k$. We initialize the recombined probability $P_R$ to $P_{UM}$ and update $P_R$ with the update probabiltities $P_{update}$
\begin{equation}
P_{R}'(s) = P_{R}(s) + P_{update}(s), \forall s\in \mathbb{B}_c.
\label{eq:p_r}
\end{equation}
The probability $P_{update}(s)$ updates the recombined probability distribution $P_R$ such that it has bitwise probabilities close to the protected bitwise probabilities in each mitigated result. For each string $s$, the update probability is calculated based on the gradient of the protected distribution $P_{M\_k}$ for each qubit $k$ and the current distribution $P_R$
\begin{equation}
P_{update}(s) = \sum_{k\in [0,...,n]} P_{R}(s)\times \frac{w_j(P_{M\_k}, k)}{w_j(P_R, k)}, j = s[k].
\label{eq:p_update}
\end{equation}
\begin{figure*}[htbp]
\centering
\subfloat[LiH interatomic distance $l = 1.5$\AA\label{subfig:results_lih}]{\includegraphics[width=0.49\linewidth]{figures/vqe_liH.PNG}}
\subfloat[HF interatomic distance $l = 0.9$\AA\label{subfig:results_HF}]{\includegraphics[width=0.49\linewidth]{figures/vqe_HF.PNG}}
\caption{Experimental results for solving LiH and HF ground state energy with VQE. The red bar represents the unmitigated result. The orange bars (PCS\_k) are the results from the experiments with bitwise PCS on real device. The green bars (SQEM\_k) represent the results that apply SQEM to qubit $k$. The orange and green bar with horizontal stripes represent the result by Bayesian recombining the distributions from the orange and the green bars.}
\label{fig:results}
\end{figure*}
The recombined probability $P_R$ is iteratively updated and renormalized. The algorithm stops when the hellinger distance~\cite{hellinger1909hellinger} between $P_{R'}$ and $P_{R}$ is smaller than a threshold. After the process, the recombined probability distribution $P_{R}$ has bitwise probabilities close to the mitigated bitwise probabilities:
\begin{equation}
w_j(P_R, k) \simeq w_j(P_{M\_k}, k), \forall j\in[0,1], k\in[0,n].
\label{eq:p_eq}
\end{equation}
\section{Experiment}
We apply SQEM to the Variational Quantum Eigensolver (VQE) task of ground state energy estimation for three molecules: LiH, BeH$_2$, and HF. We test our error mitigation framework with the two-local SU(2) ansatz~\cite{hardware-efficient} on the 27-qubit quantum device \texttt{ibm\_auckland}.
The Hamiltonians are constructed in the STO-3G basis with frozen orbitals. The ansatz is initialized with Hartree-Fock~\cite{hartreefork} initialization. We fixed the parameters to the optimal parameters found with the `COBYLA' optimizer on a noiseless simulator. The Bayesian recombination threshold is set to 0.0001. Each experiment is executed with 10000 shots. The code used for experiments in this work is available online at \url{https://github.com/revilooliver/cut4mitigation}.
The blue dashed lines in Figure~\ref{fig:results} represent the noiseless results of the ansatz. They are derived from a noiseless simulator with the same number of shots as our experiments. The black dashed lines represent the exact ground state energy calculated from PySCF~\cite{sun2018pyscf} and Qiskit Nature~\cite{Qiskit}. Figure~\ref{subfig:results_lih} shows the results for LiH. The unmitigated result is -7.8334 hartree which is far from the noiseless ansatz result -7.8634 and the exact gound state energy -7.8810. Then, we run the ansatz with bitwise PCS on a real device and the results are marked with orange bars in the figure. Since the error mitigation circuit introduces extra noise, the results may get worse.
We apply SQEM to simulate the error mitigation fragment on a classical device. As shown in the figure, the error mitigated results (greed bars) with SQEM are closer to the ideal results than the orange bars. At last we use the Bayesian recombination algorithm to recombine the probability distributions from the green bars. The recombined result has energy of -7.8625. It is within the chemical accuracy (an energy error of approximately 0.0016 hartree) of the noiseless ansatz result -7.8634. The results show that SQEM can significantly mitigate the noise on quantum devices and generate results close enough to the results on a noise-free simulator.
We also include the results for HF in Figure~\ref{subfig:results_HF}. The exact groud state energy is -98.5929 hartree. The unmitigated result energy is -98.1218. The recombined result energy is -98.5653, which is very close to the noiseless ansatz result -98.5684. For the BeH$_2$ molecule, the unmitigated, recombined, and noiseless results are -15.3770, -15.4678, and -15.4706, respectively.
\section{Conclusions and Future Directions}
We introduce the SQEM framework which removes the added gates and ancillas required by many error mitigation techniques from the quantum hardware execution by applying circuit cutting. The error mitigation subcircuit is executed on a classical device whereas the main circuit is executed on a quantum device. Furthermore, we demonstrate that this can be performed in a polynomially scalable way for a variety of problems by error mitigating individual qubits with PCS and performing a Bayesian recombination of the probability distributions at the end.
While we used Pauli check sandwiching, our framework is compatible with many error mitigation techniques. We tested our ideas with three hardware VQE experiments: estimating the ground state energies of LiH, BeH$_2$, and HF. In both experiments, the SQEM framework matched the results of a full classical simulation of the VQE circuits up to a very high accuracy.
For future directions we can look at applying SQEM for other problems such as the quantum approximate optimization algorithm \cite{farhi2014QAOA}, investigating the effects of protecting multiple qubits at a time, improving the probability recombination algorithm, testing larger molecules, and trying other error mitigation schemes.
\section*{Acknowledgments}
We thank Mark Byrd, Paul Hovland, and Gokul Subramanian Ravi for helpful discussions.
This material is based upon work supported by the U.S. Department of Energy Office of Science National Quantum Information Science Research Centers. This research was also supported in part by an appointment to the Intelligence Community Postdoctoral Research Fellowship Program at Argonne National Laboratory, administered by Oak Ridge Institute for Science and Education through an interagency agreement between the U.S. This research used resources of the Oak Ridge Leadership Computing Facility, which is a DOE Office of Science User Facility supported under Contract DE-AC05-00OR22725."
The submitted manuscript has been created by UChicago Argonne, LLC, Operator of Argonne National Laboratory (``Argonne”). Argonne, a U.S. Department of Energy Office of Science laboratory, is operated under Contract No. DE-AC02-06CH11357. The U.S. Government retains for itself, and others acting on its behalf, a paid-up nonexclusive, irrevocable worldwide license in said article to reproduce, prepare derivative works, distribute copies to the public, and perform publicly and display publicly, by or on behalf of the Government. The Department of Energy will provide public access to these results of federally sponsored research in accordance with the DOE Public Access Plan (\url{http://energy.gov/downloads/doe-public-access-plan}).
|
\section{Introduction}
Accreting active galactic nuclei (AGN) can be identified by their optical emission line features such as broad and/or narrow high ionization lines.
Doppler broadened Broad Emission Lines (BELs) originate from the gravitationally-bound broad-line region (BLR). In contrast, the narrow-line regions (NLRs) can extend beyond a few hundred parsecs, although the ionizing AGN continuum considerably influences the NLR. AGN that show both the broad- and the narrow-lines are classified as Type-I AGN, while those that do not show BELs are called as Type-II AGN. The AGN unification model \citep{1993ARA&A..31..473A, 1995PASP..107..803U} attributes differences in type to the viewing angle towards an axisymmetric, parsec scale toroidal structure (the so-called ``torus''). This structure can obstruct our direct view of the AGN central engine along our line-of sight, and block our view of the BLR in Type-II AGN.
\par A small subset of AGN, referred to as changing-look (CL) AGN have been reported in recent years \citep[e.g.,][]{2015ApJ...800..144L, 2021MNRAS.508..144G, 2022ApJ...939L..16Z}. These show strong spectral variability in their BELs and continuum, and can therefore appear to transition between the AGN classes over months to several years timescale. CL-AGN transitions have been explained mainly by variable obscuration or a strong variation in the accretion rate \citep[see][for a recent review]{2022arXiv221105132R}.
\par AT2021fxu/ZTF21aalxxzn/Gaia22dgm is an optical transient at a redshift of 0.1097 identified by the Zwicky Transient Facility (ZTF) on 13 February 2021 \citep{2021TNSTR.792....1D}. Prior to the optical outburst, it was cataloged as an AGN (2MASX J13471874+0210579/SWIFT J1347.7+0212) by the BAT AGN Spectroscopic Survey \cite[BASS,][]{2017ApJ...850...74K}. Here we present analysis of AT2021fxu's multiwavelength data taken over a temporal baseline of over two decades.
\section{Data and Analysis}
\subsection{Properties of optical spectra}
In Fig.\ref{fig:panel}a, we show AT2021fxu's optical spectra before and after the optical transition. The archival optical spectrum from Sloan Digital Sky Survey (SDSS) in 2001 shows a continuum dominated by stellar population with no clear BELs. Although a specific AGN class is hard to pin down due to contamination of sky lines in the H$\alpha$ region, the general features are consistent with either a low-ionization nuclear emission-line region \citep[LINER,][]{1980A&A....87..152H} or a Type-II AGN. In the optical spectrum captured by the Faulkes Telescope North in 2021 (Fig.\ref{fig:panel}c), a bright blue continuum and BELs typical for a Type-I AGN are evident. This indicates a change in the spectral class of the AGN from Type-II to Type-I.
\pagebreak
\subsection{Optical and UV photometry}
{\it Swift}'s Ultra-Violet/Optical
Telescope (UVOT) performed observations of the source before and after the optical outburst. The data were processed with the {\tt uvotsource} command
within the HEASoft v6.29c package. We also performed point spread function (PSF) photometry on all publicly available ZTF data at the location
of AT2021fxu using the ZTF forced-photometry service
\citep{2019PASP..131a8003M} in the g- and r-bands.
The measured magnitudes have been corrected for Galactic
extinction using E(B\scalebox{0.75}[1.0]{$-$} V) = 0.021 from \citet{2011ApJ...737..103S}.
\par Fig.\ref{fig:panel}c shows the optical lightcurves since Decmeber 2020 ($\sim$MJD 59200), which is nearly flat until July 2021 ($\sim$MJD 59420). Additionally, the optical magnitudes recorded by SDSS in May 2000 (18.24$\pm$0.01 in the optical-g and 17.33$\pm$0.01 in the optical-r, not shown here) are consistent with the average optical magnitudes between March 2018 to July 2021 (18.22$\pm$0.08 (g-band) and 17.41$\pm$0.05 (r-band)). This suggests that AT2021fxu's optical spectrum did not change between 2000 and 2021. Further, the optical brightness of the source increased by $\approx$0.6 and $\approx$0.34 magnitudes in the g- and r-bands, respectively, during the optical outburst between 1 April 2022 and 27 May 2022. This was followed by an optical decline of $\approx$0.34 magnitudes in the g-band and $\approx$0.2 magnitudes in the r-band over the next two months, until 19 July 2022.
\par Comparing the average UV magnitudes from about a decade prior to the outburst with the values recorded during the optical decline, the magnitudes have changed from $\approx$18.3 to $\approx$17.7 in the U-band, from $\approx$19.3 to $\approx$18.2 in the UVW1 band, and from $\approx$19.6 to $\approx$18.4 in the UVW2 band, indicating that the brightness in all the UV bands increased post-outburst.
\par There is a shift to a bluer spectral energy distribution (SED), with the optical (g\scalebox{0.75}[1.0]{$-$} r) color decreasing from 0.80$\pm$0.06 to 0.46$\pm$0.04, before and after the optical outburst, respectively. At the optical peak, the (g-r) difference is minimum, and is $\approx$0.38. Similarly, the average UV (W1\scalebox{0.75}[1.0]{$-$} U) magnitude difference drops from 1.02$\pm$0.07 to 0.5$\pm$0.01, and the average UV (W2\scalebox{0.75}[1.0]{$-$} U) magnitude decreases from 1.28$\pm$0.05 to 0.74$\pm$0.02, again indicative of a bluer SED.
\subsection{X-ray spectral analysis}
The archival {\it Swift} X-Ray Telescope (XRT) observations between 4 December 2011 to 13 August 2012 suggest that the source was highly variable in X-rays, with the 0.3-10 keV count rate increasing by a factor of 10 in less than 2 months.
We fit the combined 0.3-10 keV XRT spectrum using the \texttt{TBabs*zTBabs*zashift(powerlaw)} model in XSPEC \citep{1996ASPC..101...17A} with the galactic column density ($N_H$) fixed at 1.5$\times$10$^{20}$ cm$^{-2}$ \citep{2016A&A...594A.116H}, and the host galaxy $N_H$ left as a free parameter. The best-fit powerlaw index and the neutral column density were $1.71_{-0.19}^{+0.42}$ and $<$$8\times10^{20}$ cm$^{-2}$, respectively. The observed luminosity and flux at these times were $2.72_{-0.64}^{+0.48}\times10^{43}$ ergs s\textsuperscript{-1} and $0.91\pm 0.20\times10^{-12}$ ergs cm\textsuperscript{-2} s\textsuperscript{-1}, respectively. All the errorbars reported in this work represent 90\% confidence values.
\par {\it Swift}/XRT observations between August 2021 and August 2022 coincided with the recent optical outburst. We fit the combined XRT spectrum during the optical decline (mean countrate of $0.038\pm0.002$ cps) with the same powerlaw model as before and obtained a best-fit photon index and column denstiy values of $1.83_{-0.16}^{+0.32}$, and $<7.5\times10^{20}$ cm$^{-2}$, respectively. The observed 0.3-10 keV luminosity and flux of the post-outburst XRT spectrum were found to be $6.54_{-1.37}^{+0.76}\times10^{43}$ ergs s\textsuperscript{-1} and $2.17_{-0.35}^{+0.31}\times10^{-12}$ ergs cm\textsuperscript{-2} s\textsuperscript{-1}, respectively.
\par Following XRT’s detection, the Neutron Star Interior Composition Explorer{\it (NICER)} performed multiple exposures between 28 June 2022 and 4 July 2022 during the optical decline. The source was variable during this time (Fig.\ref{fig:panel}d) with a mean background-subtracted count rate of $1.33\pm0.02$ cps in the 0.3-2 keV band. We fit a combined {\it NICER} 0.3-2.0 keV spectrum using the powerlaw model described above. The best-fit power law index and the host $N_H$ column upper limits are $2.01^{+0.13}_{-0.06}$ and $1.5\times10^{20}$ cm$^{-2}$, respectively, consistent with the post-outburst values from {\it Swift}/XRT. The observed average luminosity and flux in the 0.3-10 keV energies from {\it NICER} are $7.98_{-0.82}^{+0.73}\times10^{43}$ ergs s\textsuperscript{-1} and $2.60_{-0.30}^{+0.25}\times10^{-12}$ ergs cm\textsuperscript{-2} s\textsuperscript{-1}, respectively.
\par The X-ray lightcurves from {\it Swift}/XRT and {\it NICER} (Fig.\ref{fig:panel}d) show a doubling in the X-ray flux post-outburst and high flux variability with almost no change in the X-ray spectral shape over time. Furthermore, the powerlaw index and the host column density values do not show considerable variation between 2011-12 and 2021-22.
\begin{figure*}[ht!]
\centering
\includegraphics[scale=0.6]{panel_plot_FINALpublish.pdf}
\caption{(a) Optical spectrum of the host galaxy in 2001 (black), and the optical spectrum post-outburst in May 2022 (blue) showing the appearance of BELs and a strong blue continuum; (b) Unfolded spectra and models for {\it Swift}/XRT data between 2011-2012 (before the optical outburst), and {\it Swift}/XRT and {\it NICER} between June-August 2022 (post-outburst); (c) Optical and UV light curves from {\it Swift}/UVOT; (d) {\it Swift}/XRT and {\it NICER} X-ray light curves, zoomed in plot shows short-term X-ray variability.}
\label{fig:panel}
\end{figure*}
\section{Discussion and Conclusion}
We have analyzed AT2021fxu's optical, UV and X-ray data over approximately two decades. The recent optical spectrum clearly indicates the appearance of BELs, suggesting that AT2021fxu underwent a spectral transition from a Type-II to a Type-I AGN. Since the optical g- and r- magnitudes have remained constant between May 2000 to July 2021, it is clear that optical spectrum remained the same during these years, and the recent optical outburst in April-May 2022 marks the onset of the CL transition in this AGN.
\par X-ray spectral analysis shows that there is minimal variation in the column density over the observed time. This eliminates the possibility of variable obscuration driving the spectral transition, implying instead that the CL transition is due to a variable accretion rate. This mechanism of transition due to increased accretion is further confirmed by an increase in the flux in all energy bands and a shift of the SED toward bluer wavelengths.
\par In summary, we show that AT2021fxu is an accretion-drive CL transition in the AGN 2MASX J13471874+021057/SWIFT J1347.7+0212, where the increase in the ionization flux is able to produce strong broad Balmer lines that drives the change in spectral classification.
\par Interestingly, during the writing of this note, the source showed further optical brightening by $\approx$0.12 magnitudes. We have initiated follow-up observations on {\it Swift} and {\it NICER} to further analyze this new outburst, and we encourage observations in other wavelengths to study the ongoing CL transition.
\newpage
|
\section{Introduction}
\label{intro}
Flow of immiscible fluids in porous media \cite{b88,s11,b17,ffh22}
is a problem that has been in the hands of engineers for a long time.
This has resulted in a schism between the physics at the pore scale
and the description of the flow at scales where the porous medium
may be seen as a continuum, also known as the Darcy scale.
This fact has led to the phenomenological {\it relative permeability\/}
equations proposed by Wycoff and Botset in 1936 \cite{wb36}
with the inclusion by Leverett of the concept of capillary pressure in
1940 \cite{l40}, still today being the unchallenged approach to
calculate flow in porous media at large scales.
The basic ideas of the theory are easy to grasp.
Seen from the viewpoint of one of the immiscible fluids,
the solid matrix and the other fluid together reduce the pore space
in which that fluid can move. Hence,
the effective permeability as seen by the fluid is reduced, and the permeability
reduction factor of each fluid is their relative
permeability. The capillary pressure models the interfacial
tension at the interfaces between the immiscible fluids by
assuming that there is a pressure difference between the pressure field
in each fluid. The central variables of the
theory, the relative permeabilities and the capillary pressure curves,
are determined routinely in the laboratory under the
name \textit{special core analysis}, and then used as input in reservoir
simulators \cite{b17,d13}. A key assumption in the
theory is that the relative permeabilities and the capillary
pressure are functions of the saturation alone. This simplifies
the theory tremendously, but it also distances the theory from realism.
Some progress has been made in order to improve on relative permeability
theory. Barenblatt et al.\ \cite{bps02} recognized that a key assumption
in relative permeability theory is that the flow is {\it locally\/} in a
steady state, even if the flow as a whole is developing. This has as an implication
that the central variables of that theory, the relative permeabilities
and the capillary pressure, are functions of the saturation alone. They then go on to
generalize the theory to flow which is locally out of equilibrium, exploring how the
central variables change. Wang et al.\ \cite{waa19} go further by introducing
dynamic length scales due to the mixing zone variations, over which the spatial averaging
is done.
The relative permeability approach is phenomenological. Going beyond relative permeability
theory means making a connection between the physics at the pore scale with
the Darcy scale description of the flow. The attempts at constructing
a connection between these two scales --- which may stretch from micrometers to
kilometers --- have mainly been focused on homogenization. That, is
replacing the original porous medium by an equivalent spatially structure-less one.
The most famous approach to the scale-up problem along these lines is
\textit{Thermodynamically Constrained Averaging Theory} (TCAT)
\cite{hg90,hg93,hg93b,nbh11,gm14}, based on thermodynamically consistent definitions
made at the continuum scale based on volume averages of pore-scale thermodynamic
quantities, combined with closure relations based on homogenization \cite{w86}.
All variables in TCAT are defined in terms of pore-scale variables. However,
this results in many variables and complicated assumptions are
needed to derive useful results.
Another homogenization-approach based on non-equilibrium thermodynamics
uses Euler homogeneity to define the up-scaled pressure. From this, Kjelstrup et al. derive
constitutive equations for the flow while keeping the number of variables down
\cite{kbhhg19,kbhhg19b,bk22}.
There is also an ongoing effort in constructing a scaled-up theory based on geometrical
properties of pore space by using Hadwiger theorem \cite{cacbsbgm18,acbrlasb19,cba19}.
The thorem implies that we can express the properties of the spatial geometry of
the three-dimensional porous medium as a
linear combination of four Minkowski functionals: volume, surface area, mean curvature, and Gaussian curvature.
These are the only four numbers required to characterize the geometric state of
the porous medium. The connectivity of the fluids is described by the Euler
characteristic, which by the Gauss-Bonnet theorem can be computed from the total curvature.
A different class of theories are based on detailed and specific assumptions
concerning the physics involved. An example is \textit{Local Porosity Theory}
\cite{hb00,h06a,h06b,h06c,hd10,dhh12}. Another example is the \textit{Decomposition in
Prototype Flow} (DeProf) \textit{theory} which is a fluid mechanical model combined with
non-equilibrium statistical mechanics based on a classification scheme of
fluid configurations at the pore level \cite{vcp98,v12,v18}. A third example is
that of Xu and Louge \cite{xl15} who introduce a simple model based on Ising-like statistical
mechanics to mimic the motion of the immiscible fluids at the
pore scale, and then scale up by calculating the mean field behavior of the model.
In this way they obtain the wetting fluid retention curve.
The approach we take in this paper is build on the approach to the upscaling problem
found in References \cite{hsbkgv18,rsh20,rpsh22,hfss22,fsh22}. The underlying idea is to
map the flow of immiscible fluids in porous media, a dissipative and
therefore out-of-equilibrium system, onto a system which is in equilibrium.
The \textit{Jaynes principle of maximum entropy} \cite{j57} may then be invoked and the
scale-up problem is transformed into that of calculating a partition function
\cite{hfss22}.
In order to sketch this approach, we need to establish the concept of {\it steady-state flow.\/}
In the field of porous media, ``steady-state flow" has two meanings. The traditional one
is to define it as flow where all fluid interfaces remain static \cite{ea00}. The other one,
which we adopt here, is to state that it is flow where the macroscopic variables remain
fixed or fluctuate around well-defined averages. The fluid interfaces will move and on
larger scales than the pore scale one will see fluid clusters breaking up and merging.
If the porous medium is statistically homogeneous, we will see the same local statistics
describing the fluids everywhere in the porous medium \cite{fsh22}.
Imagine a porous plug as shown in Figure \ref{fig1}. There is a mixture of two immiscible fluids
flowing through it in the direction of the cylinder axis under steady-state conditions.
\begin{figure}
\includegraphics[width=10truecm]{fig1.eps}
\caption{A porous plug in the form of a cylinder. There is a mixture of two immiscible fluids flowing
through it in the direction of the cylinder axis under steady-state conditions. We furthermore imagine two
disks orthogonal to the cylinder axis. These are REAs. }
\label{fig1}
\end{figure}
A central concept in the following is the \textit{Representative Elementary Area} (REA) \cite{bb12,rpsh22,hfss22}.
We pick a point in the porous plug. In a neighbourhood of this point, there will a set of streamlines associated
with the velocity field $\vec v_p$. The overall flow direction
is then defined by the tangent vectors to the stramlines. We place an imaginary disk of area $A$ at
the point, with orientation such that $A$ is orthogonal to the overall flow direction. We illustrate this as
the lower disk in Figure \ref{fig1}. We assume that the disk is small enough for
the porous medium to be homogeneous over the size of the plane with respect to porosity and
permeability. This disk is an REA.
The REA will contain different fluid configurations. By ``fluid
configuration'', we mean the spatial distributions of the two fluids in the disk and their scalar velocity fields. We also show a second disk in Figure
\ref{fig1}, which represents another REA placed further along the average flow.
Since the flow is incompressible, we can imagine the $z$-axis as corresponding
to a pseudo-time axis. We may state that
the fluid configuration in the lower disk of Figure \ref{fig1} evolves into the
fluid configuration in the upper disk under pseudo-time-translation. We now associate an entropy to the fluid configurations in the sense of Shannon \cite{s48}. Even though the system is producing molecular
entropy through dissipation, it is \textit{not} producing Shannon entropy along the $z$-axis. This allows us to
use the Jaynes maximum entropy principle, and a statistical mechanics based on Shannon entropy ensues
\cite{hfss22}.
This statistical mechanics scales up the pore-scale physics, represented by the fluid configurations to
the Darcy scale, which then is represented by a thermodynamics-like formalism involving averaged
velocities.
Rudiments of this thermodynamics-like formalism was first studied by
Hansen et al.\ \cite{hsbkgv18}, who used extensivity to derive a set of equations that relate the
seepage velocity of each of two immiscible fluids flowing under steady state
conditions through a representative volume element (REV). They introduced a
{\it co-moving velocity\/} that together with the average seepage velocity $v$ contained
the necessary information to determine the seepage velocities of the more wetting and the
less wetting fluids, $v_w$ and $v_n$ respectively. Stated in a different way, these equations
made it possible to make the transformation
\begin{equation}
\label{eq:main-map}
(v,v_m) \leftrightarrows (v_w,v_n) \ \;,
\end{equation}
where $S_w$ and $S_n$ are the wetting and non-wetting saturations respectively.
It is the aim of this paper to present a {\it geometrical interpretation\/} of the transformation
in equation \ref{eq:main-map}. We wish to provide
an {\it intuitive\/} understanding of precisely what the transformation is doing and
determine the role of the co-moving velocity. Our aim is not to develop the theory presented in
\cite{hsbkgv18} further by including new results, but rather consolidate and amend the existing
theory with a deeper understanding.
Our geometrical interpretation will rest upon the definition of a space spanned by two central
variables: {\it the wetting and non-wetting transversal pore areas.\/} These areas are defined as the area
of REAs covered by the wetting and non-wetting fluids respectively.
Instead of using the two extensive areas directly, one often work with coordinates where one is the (wetting)
saturation. If the pore area of the REA is fixed, it is often convenient to let
the other variable be this area. A third way --- which is new --- is to use
{\it polar coordinates.\/} This, as we shall see, simplify the theory considerably.
In Section \ref{rea} we describe the REA and the relevant variables. We define
intensive and extensive variables, tying them to how they scale under scaling of the size of
the REV. We also review the central results of Hansen et al.\
\cite{hsbkgv18} here: the introduction of the auxiliary {\it thermodynamic velocities\/} and the
{\it co-moving velocity.\/}
The central concept in our approach to the immiscible two-fluid flow problem is
that of the pore areas. In Section \ref{area} we introduce {\it area space\/}, a
space that simplifies the analysis considerably. We showcase different
coordinate systems on this space, focusing on polar coordinates.
Section \ref{euler} expresses the central results of Hansen et al.\ \cite{hsbkgv18} in polar
coordinates. Expressed in this coordinate system, the geometrical structure implied by the
extensivity versus intensivity conditions impose restrictions that simpliefies
the equations. The section goes on to express
the total pore velocitiy, seepage velocities, thermodynamic velocities and co-moving velocity
in terms of polar coordinates. This provides a clear geometrical interpretation of the
co-moving velocity.
Section \ref{coordinate} focus on relations between the different velocities
that can be expressed without referring to any coordinate system. We do this by invoking
{\it differential forms\/} and {\it exterior algebra\/} \cite{f63,mtw17}. Differential
forms constitute the generalization of the infinitesimal line, area or volume elements used in
integrals and exterior calculus is the algebra that makes it possible to handle them. Flanders
predicted in the early sixties \cite{f63} that within short they would become important tools
in engineering. This did not happen, and they have remained primarily within
mathematics and theoretical physics. Such objects are also especially prevalent in
thermodynamics, and we will here show that the computational
rules of differential forms provide a simple way of expressing the relations
between the velocities in the system at hand.
Section \ref{comov} focuses the discussion on the co-moving velocity by
constructing a coordinate-independent expression
that defines it.
In the last Section \ref{conclusion}, we summarize our main results which may be stated through three equations,
(\ref{eq18}), (\ref{eq29-1}) and (\ref{eqn53}). These equations formulate in a coordinate-free way
the three relations that exist between the average seepage velocity and the co-moving velocity.
\section{Representative Elementary Area}
\label{rea}
We will now elaborate on the definition of the Representative Elementary Area
(REA) presented in Section \ref{intro}.
We define the {\it transversal pore area\/} $A_p$ as previously, which defines a porosity
\begin{equation}
\label{e5}
\phi=\frac{A_p}{A}\;.
\end{equation}
The pore volume contains two immiscible fluids; the more wetting fluid (to be referred to as the
\textit{wetting fluid}) or the less wetting fluid (to be referred to as the \textit{non-wetting} fluid). The transversal
pore area $A_p$ may therefore be split into the area of the REA disk cutting through wetting
fluid $A_w$, or cutting through the non-wetting fluid $A_n$, so that we have
\begin{equation}
\label{eq1}
A_p=A_w+A_n\;.
\end{equation}
We also define the saturations
\begin{eqnarray}
S_w=\frac{A_w}{A_p}\;,\label{eq2}\\
S_n=\frac{A_n}{A_p}\;,\label{e13}\\
\nonumber
\end{eqnarray}
so that
\begin{equation}
\label{e10}
S_w+S_n=1\;.
\end{equation}
We note that the transversal pore areas $A_p$, $A_w$ and $A_n$ are {\it extensive\/}
in the REA area $A$. That is
$A_p\to\lambda A_p$, $A_w\to\lambda A_w$ and $A_n\to\lambda A_n$ when $A\to\lambda A$.
The porosity $\phi$ and the two saturations $S_w$ and $S_n$ are
{\it intensive\/} in the REA area $A$: $\phi\to\lambda^0\phi$, $S_w\to\lambda^0 S_w$
and $S_n\to\lambda^0 S_n$ when $A\to\lambda A$.
There is a time averaged volumetric flow rate $Q$ through the REA. The volumetric
flow rate consists of two components, $Q_{w}$ and $Q_{n}$, which are the
volumetric flow rates of the wetting and the non-wetting fluids. We have
\begin{equation}
\label{e6}
Q=Q_{w}+Q_{n}\;.
\end{equation}
We define the total, wetting and non-wetting {\it seepage velocities\/} respectively as
\begin{eqnarray}
v=\frac{Q}{A_p}\;,\label{e7}\\
v_w=\frac{Q_w}{A_w}\;,\label{e8}\\
v_n=\frac{Q_n}{A_n}\;.\label{e9}
\end{eqnarray}
Using equations (\ref{e6}) to (\ref{e9}), we find
\begin{eqnarray}
v&=&\frac{Q}{A_p}=\frac{A_w}{A_p}\ \frac{Q_w}{A_w}+\frac{A_n}{A_p}\ \frac{Q_n}{A_n}\nonumber\\
&=&S_w v_w+S_n v_n\;.\label{e14}
\end{eqnarray}
The volumetric flow rates $Q$, $Q_w$ and $Q_n$ are extensive in the REA area $A$
and the velocities $v$, $v_w$ and $v_n$ are intensive in the area $A$.
\subsection{Euler homogeneity}
\label{deform}
As the areas $A_w$ and $A_n$ and the volumetric flow rate $Q$ are
extensive in the REA area $A$, we have that
\begin{equation}
\label{eq7}
Q(\lambda A_w,\lambda A_n)=\lambda Q(A_w,A_n)\;.
\end{equation}
This scaling was the basis for the theory presented in \cite{hsbkgv18},
which we now summarize.
We now apply the \textit{Euler homogeneous function theorem} to $Q$. We take the derivative with respect to $\lambda $ on both sides of
(\ref{eq7}) and set $\lambda =1$. This gives
\begin{equation}
\label{eqn11}
Q(A_{w},A_{n})=A_{w}\left( \frac{\partial Q}{\partial A_{w}}\right)
_{A_{n}}+A_{n}\left( \frac{\partial Q}{\partial A_{n}}\right)
_{A_{w}}\;.
\end{equation}
We note here that we assume that $A_w$ and $A_n$ are our independent
control variables. This makes $A_p$, $S_w$ and $S_n$ as dependent variables.
More precisely, equation (\ref{eqn11}) is
the Euler theorem for homogeneous functions applied to a degree-$1$ homogeneous
function $Q$. By dividing this
equation by $A_p$, we have
\begin{equation}
\label{eqn10-1}
v=S_{w}\left( \frac{\partial Q}{\partial A_w}\right)
_{A_n}+S_{n}\left( \frac{\partial Q}{\partial A_n}\right)_{A_w}\;,
\end{equation}
where we have used equations (\ref{eq2}) and (\ref{e13}).
We define the two {\it thermodynamic velocities\/} $\hat{v}_w$ and $\hat{v}_n$ as
\begin{equation}
\label{eqn10-2}
\hat{v}_{w}=\left( \frac{\partial Q}{\partial A_{w}}\right)_{A_{n}}\;,
\end{equation}
and
\begin{equation}
\label{eqn10-3}
\hat{v}_{n}=\left( \frac{\partial Q}{\partial A_{n}}\right)_{A_{w}}\;,
\end{equation}
so that we may write (\ref{eqn10-1}) as
\begin{equation}
\label{eqn10-5}
v=S_w \hat{v}_w+S_n\hat{v}_n\;.
\end{equation}
The thermodynamic velocities $\hat{v}_w$ and $\hat{v}_n$ are {\it not\/} the physical
velocities $v_w$ and $v_n$. Rather, the {\it most general relation\/}
between $v_w$ and $\hat{v}_w$, and $v_n$ and $\hat{v}_n$, that fulfills both
equations (\ref{e14}) and (\ref{eqn10-5}),
\begin{equation}
v=S_w\hat{v}_w+S_n\hat{v}_n=S_wv_w+S_nv_n\;,
\end{equation}
can be expressed as
\begin{align}
\hat{v}_{w} \ =& \ v_{w}+S_n v_m \label{eqn10-5.5} \\
\hat{v}_{n}\ =& \ v_{n}-S_w v_m \label{eqn10-5.6} \;.
\end{align}
This \textit{defines} the {\it co-moving velocity,\/} $v_m$, which
relate the thermodynamic and the physical velocities.
We have up to now used $(A_w,A_n)$ as our control variables. If we now consider
the coordinates $(S_w,A_p)$ instead, we can write
\begin{eqnarray}
\label{eq1000}
\hat{v}_{w}=\left( \frac{\partial Q}{\partial A_{w}}\right)_{A_{n}}
&=&\left(\frac{\partial Q}{\partial S_w}\right)_{A_p}\left(\frac{\partial S_w}{\partial A_w}\right)_{A_n}
+\left(\frac{\partial Q}{\partial A_p}\right)_{S_w}\nonumber\\
&=& \left( \frac{d v}{d S_w} \right)S_n+v\;,
\end{eqnarray}
where we have used that
\begin{equation}
\label{eq1001}
\left(\frac{\partial S_w}{\partial A_w}\right)_{A_n}=\left(\frac{\partial}{\partial A_w}\right)_{A_n}\left(\frac{A_w}{A_w+A_n}\right)
=\frac{S_n}{A_p}\;.
\end{equation}
Likewise, we find that
\begin{equation}
\label{eq1002}
\hat{v}_{n}=\left( \frac{\partial Q}{\partial A_{n}}\right)_{A_{w}}=- \left( \frac{d v}{ d S_w}\ \right) S_w+v\;.
\end{equation}
We combine equations (\ref{eq1000}) and (\ref{eq1002}) with equations (\ref{eqn10-5.5}) and (\ref{eqn10-5.6}) to find
\begin{align}
v_w \ =& \ v+S_n\left(\frac{dv}{dS_w}-v_m\right)\;,\label{e1} \\
v_n \ =& \ v-S_w\left(\frac{dv}{dS_w}-v_m\right)\;,\label{e2}
\end{align}
We see that these two equations gives the map $(v,v_m) \rightarrow
(v_w,v_n)$.
We now subtract equation (\ref{e2}) from (\ref{e1}), finding
\begin{equation}
\label{eq1003}
\frac{dv}{dS_w}=v_w-v_n+v_m\;.
\end{equation}
We now differentiate equation (\ref{e14}) with respect to $S_w$,
\begin{equation}
\label{eq1004}
\frac{dv}{dS_w}=v_w-v_n+S_w\frac{dv_w}{dS_w}+S_n\frac{dv_n}{dS_w}\;.
\end{equation}
We compare equations (\ref{eq1003}) and (\ref{eq1004}), finding
\begin{equation}
\label{eq1005}
v_m=S_w\frac{dv_w}{dS_w}+S_n\frac{dv_n}{dS_w}\;.
\end{equation}
Equations (\ref{e14}) and (\ref{eq1005}) constitute the inverse mapping $(v_w,v_n)\to (v,v_m)$.
The mapping $(v,v_m)\to (v_w,v_n)$ tells us that given the constitutive equations for
$v$ and $v_m$, we also have the constitutive equations for $v_w$ and $v_n$. It turns out that
the constitutive equation for $v_m$ is surprisingly simple \cite{rpsh22}
\begin{equation}
\label{eq1006}
v_m=a+b\frac{dv}{dS_w}\;,
\end{equation}
where $a$ and $b$ are coefficients depending on the entropy associated with the fluid configurations
\cite{hfss22}.
The equations we have used in Section \ref{rea} hint at an underlying mathematical
structure which may seem complex. As we will now show, it is in fact quite the opposite.
\section{Coordinate Systems in Area Space}
\label{area}
The transversal pore areas $A_w \ge 0$ and $A_n \ge 0$ parametrize the first
quadrant of $\mathbb{R}^{2}$. This serves as the ``area space'' mentioned
earlier,which we will continue to denote as such. It is important to realize that this space is {\it not\/}
physical space. A value of the transversal pore area $A_{p}$ corresponds to a point
$\left( A_w, A_n \right)$ in this space. If we consider the entire plane $\mathbb{R}^{2}$ as a whole
and simply restrict our attention to the first quadrant, we may treat the area
space as a vector space. The point $\left( A_w, A_n \right)$ may then
equivalently be described by a vector,
\begin{equation}
\label{e100}
\vec A=A_w \vec e_w+A_n\vec e_n\;,
\end{equation}
where $\vec e_w$ and $\vec e_n$ form an orthonormal basis set,
see the upper figure in Figure \ref{fig2}. Note that in this picture, the bases
are shown as attached to the point $\left(A_w, A_n \right)$. We reiterate that both $A_w$ and $A_n$ are extensive.
\begin{figure}
\includegraphics[width=6truecm]{fig2.eps}
\caption{We illustrate the three coordinate systems we use to parametrize the
transversal pore area. The {\it top figure\/} shows the Cartesian coordinate system $(A_w,A_n)$.
A curve of constant $A_w$ and $A_n$ is indicated as $A_w$ and $A_n$ respectively. Both $A_w$ and $A_n$ are extensive.
The orthonormal basis set $(\vec e_w,\vec e_n)$ is shown as attached to the
point $\left( A_w , A_n \right)$. We show in the {\it middle figure\/}
the saturation coordinate system $\left( A_p, S_w \right)$. We indicate a
curve of constant $A_p$ and $S_w$ as $A_p$ and
$S_w$ in the figure. $A_p$ is extensive and $S_w$ is intensive.
We show the basis set $(\vec e_p,\vec e_s)$, which is not orthonormal. In the {\it lower figure\/}
we show the polar coordinate system, with curves of constant $A_r$ and $\phi$ denoted $A_r$ and $\phi$ respectively. $A_r$ is extensive and $\phi$ is intensive. The
basis set $(\vec e_r,\vec e_\phi)$ is orthonormal. In all three figures, we show the transversal
pore area as a vector $\vec A$.}
\label{fig2}
\end{figure}
Every point in the transversal area space corresponds to a a given saturation
$S_w$ and a transversal pore area $A_p$ and we may view the map
$(A_w,A_n)\to(A_p,S_w)$ given by equation (\ref{eq1}) and (\ref{eq2}) as a coordinate transformation.
This is a more natural coordinate system to work with since $S_w$ is an intensive variable and $A_p$ is
in practice kept constant.
We name this system the {\it saturation coordinate system.\/} We show the normalized basis vector
set in the middle figure in Figure \ref{fig2}. We calculate
\begin{eqnarray}
\vec u_p&=&\left(\frac{\partial \vec A}{\partial A_p}\right)_{S_w}= S_w \vec e_w+(1-S_w)\vec e_n\;,\label{e101}\\
\vec u_s&=&\left(\frac{\partial \vec A}{\partial S_w}\right)_{A_p}= A_p \vec e_w-A_p\vec e_n\;.\label{e102}
\end{eqnarray}
We normalize the two vectors $\vec u_p$ and $\vec u_s$ to find
\begin{eqnarray}
\vec e_p&=&\frac{S_w\vec e_w+(1-S_w)\vec e_n}{[1-2S_w(1-S_w)]^{1/2}}\;,\label{e103}\\
\vec e_s&=& \frac{\vec e_w-\vec e_n}{\sqrt{2}}\;.\label{e104}
\end{eqnarray}
This basis set is not orthonormal, as illustrated in Figure \ref{fig2}. By expressing $\vec A$ in
this coordinate system, we find
\begin{equation}
\label{e104-1}
\vec A=A_p \left[1-2S_w(1-S_w)\right]^{1/2} \vec e_p\;.
\end{equation}
As will become apparent, the most convenient coordinate system to work with from
a theoretical point of view is {\it polar coordinates\/} $\left(A_r, \phi
\right)$, given by
\begin{align}
A_r \ =& \ \sqrt{A_w^2+A_n^2}\;, \label{eq105} \\
\phi \ =& \ \arctan\left(\frac{A_n}{A_w}\right)\;, \label{eq106}
\end{align}
or {\it vice versa\/}
\begin{align}
A_w \ = \ A_r \cos\phi\;, \label{eq107} \\
A_n \ = \ A_r \sin\phi\;. \label{eq108}
\end{align}
We note that $\phi$ is intensive and $A_r$ is extensive. We show the polar
coordinate system in the lower figure in Figure\ \ref{fig2}.
The basis vector set $(\vec e_r,\vec e_\phi)$ is orthonormal, where
\begin{align}
{\vec e}_r \ =& \ \cos\phi\ {\vec e}_w+\sin\phi\ {\vec e}_n=\vec e_p\;, \label{e108-1} \\
{\vec e}_\phi \ =& \ -\sin\phi\ {\vec e}_w+\cos\phi\ {\vec e}_n\;. \label{e108-2}
\end{align}
As for the saturation coordinate system, there is one
intensive variable, $\phi$, and one extensive variable, $A_r$.
However, in contrast to the saturation coordinate system, both variables
are varied in practical situations.
We have that
\begin{equation}
\label{e106}
\vec A=A_r\vec e_r\;.
\end{equation}
We see that this is consistent with equation (\ref{e104-1}) since
\begin{equation}
\label{e106-1}
A_r=A_p\left[1-2S_w(1-S_w)\right]^{1/2}\;
\end{equation}
and
$\vec e_p=\vec e_r$.
\section{Euler Homogeneity in Polar Coordinates}
\label{euler}
We now focus on equation (\ref{eq7}) which we repeat here,
\begin{equation}
Q(\lambda A_w, \lambda A_n)=\lambda Q(A_w,A_n)\;.\nonumber
\end{equation}
We may interpret this equation geometrically. If we follow the value
of $Q$ along a ray passing through the origin
of the two-dimensional space spanned by $(A_w,A_n)$
keeping the ratio $A_n/A_w$ constant, it grows linearly
with the distance from the origin.
In {\it polar coordinates,\/} this means
\begin{equation}
\label{eq8}
Q(A_r,\phi)=\hat v_r A_r\;,
\end{equation}
where
\begin{equation}
\label{eq9}
\hat v_r=\left(\frac{\partial Q}{\partial A_r}\right)_{\phi} =\hat v_r(\phi)\;.
\end{equation}
The important point here is that $\hat v_r$ is {\it not\/} a function of $A_r$.
\begin{figure}
\includegraphics[width=8truecm]{fig3.eps}
\caption{We illustrate here the geometrical meaning of $Q$ being Euler homogeneous,
i.e., fulfilling equation (\ref{eq7}):
Moving along a ray in the space spanned by the $Q$, $A_w$ and $A_n$ directions,
$Q$ increases linearly with the
distance from the origin $(Q,A_w,A_n)=(0,0,0)$. This demonstrates that $Q$ appears as a
``crumpled" cone originating at the origin.
Hence, polar coordinates $(A_r,\phi)$ are the most convenient to use
compared to either the $(A_w,A_n)$ or $(A_p,S_w)$ coordinate
systems.}
\label{fig3}
\end{figure}
We may derive equation (\ref{eq8}) from equation (\ref{eq7}) as follows. First, write
$A_w$ and $A_n$ in terms of $A_r$ and $\phi$ in equation (\ref{eq7}) so that
\begin{eqnarray}
\label{eq8-1}
&&\frac{1}{\lambda}\ Q(\lambda A_r\cos\phi,\lambda A_r\sin\phi)\nonumber\\
&=&Q(A_r\cos\phi,A_r\sin\phi)=Q(A_r,\phi)\;.
\end{eqnarray}
Next, set $\lambda=1/A_r$ to find
\begin{equation}
\label{eq8-2}
A_r\ Q(\cos\phi,\sin\phi)=A_r\hat v_r(\phi)=Q(A_r,\phi)\;,
\end{equation}
where
\begin{equation}
\label{eq8-3-1}
\hat v_r(\phi) \equiv Q(\cos\phi,\sin\phi)
\end{equation}
--- and we are done. We illustrate in Figure \ref{fig3} the geometrical
meaning of equation (\ref{eq7}), i.e., Euler homogeneity. The graph of the volumetric flow
rate $Q$ takes the shape of a ``crumpled" cone in the space spanned by $(Q,A_w,A_n)$.
We now express the different fluid velocities, namely the pore velocity, the seepage velocities,
the thermodynamic velocities and the co-moving velocity, in terms of polar coordinates.
\subsection{Seepage velocities}
\label{velocities}
The area space spanned by $\left( A_w, A_n \right)$ is simply (a subset of) the
real plane. We can therefore identify it
with the \textit{tangent space} of velocities, which is the same Cartesian plane
as that of the areas $\left( A_w, A_n \right)$. We can therefore regard the seepage velocity as a vector in the area space,
\begin{equation}
\label{eq10}
\vec v=v_w \vec e_w+v_n\vec e_n\;.
\end{equation}
The volumetric flow rate $Q$ is then given by the scalar product
\begin{equation}
\label{eq11}
Q=\vec v \cdot \vec A=v_wA_w+v_nA_n\;.
\end{equation}
In polar coordinates, the seepage velocity is given by
\begin{equation}
\label{eq14-1}
\vec v=v_r \vec e_r+v_\phi \vec e_\phi\;.
\end{equation}
Hence, the volumetric flow rate in polar coordinates is given by
\begin{equation}
\label{eq12}
Q=\vec v \cdot \vec A=(v_r\vec e_r+v_\phi\vec e_\phi)\cdot(A_r\vec e_r)=v_r A_r\;.
\end{equation}
Comparing with equation (\ref{eq8}), we find
\begin{equation}
\label{eq13}
v_r=\hat v_r=\hat v_r(\phi)\;.
\end{equation}
By using equations (\ref{eq10}) and (\ref{eq14-1}) combined with equations (\ref{e108-1}) and (\ref{e108-2}), we find
\begin{eqnarray}
\label{eq14}
v_r&=&v_w\cos\phi+v_n\sin\phi\nonumber\\
&=&\frac{S_w v_w+(1-S_w)v_n}{[1-2S_w(1-S_w)]^{1/2}}\;,\\
v_\phi&=&-v_w\sin\phi+v_n\cos\phi\nonumber\\
&=&\frac{-(1-S_w)v_w+S_wv_n}{[1-2S_w(1-S_w)]^{1/2}}\;.
\end{eqnarray}
We have here expressed $\cos\phi$ and $\sin\phi$ in terms of $S_w$.
We now compare equation (\ref{eq8}) with equation (\ref{e7}), rewritten as
\begin{equation}
\label{eq8-3}
Q=vA_p\;.
\end{equation}
Using equation (\ref{e106-1}), we find that their equality demands
\begin{equation}
\label{eq8-4}
v=\frac{\hat v_r}{\left[1-2S_w(1-S_w)\right]^{1/2}}\;.
\end{equation}
Hence, $v$ is \textit{not} the norm of $\vec v$.
\subsection{Thermodynamic velocities}
\label{thermodynamic}
We may write equation (\ref{eqn11}) as
\begin{equation}
\label{eq18}
Q=\hat v_w A_w+\hat v_n A_n = \hat{\vec v} \cdot \vec A\;,
\end{equation}
where we have defined, in the same manner as for equation (\ref{eq10}),
\begin{equation}
\label{eq19}
\hat{\vec v}=\hat v_w \vec e_w+\hat v_n\vec e_n\;.
\end{equation}
We use equations (\ref{e108-1}) and (\ref{e108-2}) to express this equation in
polar coordinates,
\begin{equation}
\label{eq20-1}
\hat{\vec v}=\hat v_r \vec e_r+\hat v_\phi\vec e_\phi\;,
\end{equation}
where
\begin{eqnarray}
\hat v_r&=&\hat v_w\cos\phi+\hat v_n\sin\phi\;,\label{eq20-2}\\
\hat v_\phi&=&-\hat v_w\sin\phi+\hat v_n\cos\phi\;.\label{eq20-3}
\end{eqnarray}
\subsection{Co-moving velocity}
\label{co-moving}
We defined the co-moving velocity $v_m$ as the velocity function that relates the
physical seepage velocities and the thermodynamic velocities, see equations
(\ref{eqn10-5.5}) and (\ref{eqn10-5.6}). We use these two equations together with
equations (\ref{eq10}) and (\ref{eq19}) to form the vector
\begin{eqnarray}
\label{eq23}
\hat{\vec v}-\vec v&=&(\hat v_w-v_w)\vec e_w+(\hat v_n-v_n)\vec e_n\nonumber\\
&=&\frac{v_m}{A_p}\left(A_n\vec e_w-A_w\vec e_n\right)\nonumber\\
&=&\frac{v_m}{A_p}\ \left(A_r\sin\phi\ \vec e_w-A_r\cos\phi\ \vec e_n\right)\;,
\end{eqnarray}
where we have used that $S_w=A_w/A_p$ and equations (\ref{eq107}) and (\ref{eq108}).
We rewrite this equation in terms of $\vec e_\phi$, equation (\ref{e108-2}), finding
\begin{equation}
\label{eq23-2}
\hat{\vec v}-\vec v=-\frac{v_m}{\cos\phi+\sin\phi}\ \vec e_\phi\;.
\end{equation}
We define
\begin{equation}
\label{eq26}
\hat v_m=\frac{v_m}{\cos\phi+\sin\phi}
\end{equation}
so that equation (\ref{eq23}) may be written
\begin{equation}
\label{eq27}
\vec v=\hat{\vec v}+\hat{\vec v}_m\;,
\end{equation}
where
\begin{equation}
\label{eq28}
\hat{\vec v}_m=\hat v_m\vec e_\phi\;.
\end{equation}
\begin{figure}
\includegraphics[width=8truecm]{fig4.eps}
\caption{We illustrate here equations (\ref{eq12}) and (\ref{eq18}),
$Q=\vec v \cdot \vec A=\hat{\vec v}\cdot \vec A =\hat v_r A_r$, where $A_r=|\vec A|$.
The difference in the vectors $\hat{\vec v}$ and $\vec v$, $\hat{v}_m$, is orthogonal to $\vec A$,
and cannot be detected by properties of $\vec A$ alone.}
\label{fig4}
\end{figure}
In polar coordinates, the relation in equation (\ref{eq27}) is isolated to the
$\phi$-coordinate. Using equations (\ref{eq14-1}) and (\ref{eq20-3}), we can
express the relation between the vector components $\hat{v}_{\phi}$ and $v_{\phi}$ as
\begin{equation}
\label{eq29}
v_\phi=\hat v_\phi+\hat v_m\;.
\end{equation}
We illustrate this relation in Figure \ref{fig4}. This figure also demonstrates why
$Q=\hat{\vec v}\cdot\vec A={\vec v}\cdot \vec A$; it is due to $({\vec v}-\hat{\vec v} )=\hat{\vec v}_m\perp
\vec A$. Hence, we have that
\begin{equation}
\label{eq29-1}
\hat {\vec v}_m\cdot \vec A=0\;.
\end{equation}
\section{Coordinate-Free Representation}
\label{coordinate}
Since the difference between the thermodynamic- and seepage velocities in the
previous section was shown to sit in the \textit{tangent vector components}, we can
benefit from a examining the problem from a position where these components are
easier to work with. In this section, we therefore introduce {\it differential
forms\/} and {\it exterior algebra\/} \cite{f63,mtw17} to make this treatment
more economic. In addition, this makes it easier to formulate the equations met
so far in a way that does not depend on the coordinate system used.
Equation (\ref{eq8}), $Q=\hat v_r A_r$, may be differentiated to give
\begin{equation}
\label{eqn1000}
dQ=\hat v_r dA_r+d\hat v_r A_r=\hat v_r dA_r+ \hat v_r'A_r d\phi\;,
\end{equation}
where $\hat v_r'=d\hat v_r/d\phi$. Here, $dQ$, $dA_r$ and $d\phi$ are
one-forms. We furthermore have that
\begin{eqnarray}
\label{eqn1002}
dQ&=&\left(\frac{\partial Q}{\partial A_r}\right)_{\phi}dA_r+
\left(\frac{\partial Q}{\partial \phi}\right)_{A_r}d\phi\nonumber\\
&=&\hat v_rdA_r+\hat v_\phi A_rd\phi\;.
\end{eqnarray}
Comparing equations (\ref{eqn1000}) and (\ref{eqn1002}) gives
\begin{equation}
\label{eqn1003}
\hat v_\phi=\hat v_r'\;.
\end{equation}
We do the same using the coordinate system $(A_w,A_n)$. From equation
(\ref{eq11}), we find
\begin{eqnarray}
\label{eqn1004}
dQ&=&\hat v_w dA_w+\hat v_n dA_n+A_w d\hat v_w+A_nd \hat v_n\nonumber\\
&=&\hat v_w dA_w+\hat v_n dA_n\;.
\end{eqnarray}
This implies the relation
\begin{equation}
\label{eqn1005}
A_w d\hat v_w+A_n d\hat v_n=0\;.
\end{equation}
This equation corresponds to the {\it Gibbs-Duhem\/} equation in thermodynamics,
which is a statement of the dependency amongst the intensive variables.
We express $\hat v_w$ and $\hat v_n$ in terms of $\hat v_r$ and $\hat v_\phi$ by
inverting equations (\ref{eq20-2}) and (\ref{eq20-3}) to find
\begin{align}
\hat v_w \ = \ \hat v_r\cos\phi-\hat v_\phi\sin\phi\;, \label{eq31} \\
\hat v_n \ = \ \hat v_r\sin\phi+\hat v_\phi\cos\phi\;. \label{eq32}
\end{align}
By applying the exterior derivative, we get the relations
\begin{align}
d\hat v_w \ = \ \cos\phi\ d\hat v_r-\sin\phi\ d\hat v_\phi-\hat v_n d\phi\;, \label{eq33} \\
d\hat v_n \ = \ \sin\phi\ d\hat v_r+\cos\phi\ d\hat v_\phi+\hat v_w d\phi\;. \label{eq34}
\end{align}
Combining equations (\ref{eq31}) to (\ref{eq34}) with the Gibbs-Duhem equation
(\ref{eqn1005}) gives
\begin{equation}
\label{eqn1006}
\frac{d\hat v_r}{d\phi}=\hat v_\phi\;,
\end{equation}
which is identical to equation (\ref{eqn1003}). Hence, this is the
Gibbs-Duhem equation in polar coordinates.
The thermodynamic velocity, equation (\ref{eq19}), is therefore given by
\begin{equation}
\label{eq37}
\hat{\vec v}=\hat v_r\vec e_r + \hat v_r^{\prime}\vec e_\phi\;.
\end{equation}
We now have one of the main results of Hansen et al.\ \cite{hsbkgv18} in a very compact form,
\begin{equation}
\label{eq38}
\vec v=\hat v_r\vec e_r+(\hat v_r'+\hat v_m)\vec e_\phi\;.
\end{equation}
We note that since $\hat v_r$ only depends on $\phi$, the same must be true for $\hat v_m$. Hence,
$\hat v_m=\hat v_m(\phi)$. We here see that the relation between the thermodynamic-
and seepage velocities are not captured in the extensive structure tied to the
radial coordinate, but instead in the intensive quantities which only has a
dependency on $\phi$. \\
We now differentiate equation (\ref{eq8-3}), $Q={\vec v}\cdot{\vec A}$ written in the saturation coordinate
system, to find
\begin{equation}
\label{eq39}
dQ=v\ dA_p+\frac{dv}{dS_w} A_pdS_w\;.
\end{equation}
We note that
\begin{equation}
\label{eq40}
dA_p=dA_w+dA_n\;,
\end{equation}
from differentiating equation (\ref{eq1}). Likewise, we note that
\begin{eqnarray}
\label{eq41}
dS_w&=&\left(\frac{\partial S_w}{\partial A_w}\right)_{A_n}dA_w
+\left(\frac{\partial S_w}{\partial A_n}\right)_{A_w}dA_n\nonumber\\
&=&\frac{1}{A_p^2}\left(A_n dA_w-A_w dA_n\right)\;,
\end{eqnarray}
using equation (\ref{eq2}). We combine equations (\ref{eq39}) to
(\ref{eq41}) and find
\begin{equation}
\label{eq42}
dQ=\left(v+(1-S_w)\frac{dv}{dS_w}\right)dA_w+\left(v-S_w\frac{dv}{dS_w}\right)dA_n\;.
\end{equation}
By using equation (\ref{eqn11}), definitions (\ref{eqn10-2}) and
(\ref{eqn10-3}), and the
transformations (\ref{eqn10-5.5}) and (\ref{eqn10-5.6}), we may write
\begin{eqnarray}
\label{eq43}
dQ&=&\hat v_w dA_w+\hat v_n dA_n=\left(v_w-(1-S_w)v_m\right)dA_w\nonumber\\
&+&\left(v_n+S_wv_m\right)dA_n\;.
\end{eqnarray}
Equating (\ref{eq42}) and (\ref{eq43}) gives equations (\ref{e1}) and (\ref{e2}).
Let us now rewrite the differential $dQ$ as follows,
\begin{eqnarray}
\label{eqn43-1}
dQ&=&d[A_pv]=d[A_p(S_w\hat v_w+(1-S_w)\hat v_n)]\nonumber\\
&=&A_p(\hat v_w -\hat v_n)dS_w+v\ dA_p\;,
\end{eqnarray}
where we have used equation (\ref{eqn1005}). We now combine this equation with equation (\ref{eq39})
to get
\begin{equation}
\label{eqn43-2}
\frac{dv}{dS_w}=\hat v_w-\hat v_n\;.
\end{equation}
By using equations (\ref{eqn10-5.5}) and (\ref{eqn10-5.6}), we may rewrite this equation in terms of the
seepage velocities, resulting in equation (\ref{eq1003}).
We differentiate $Q$ a second time, which is zero according to the Poincar{\'e} lemma \cite{f63,mtw17}.
We see that this is indeed so by using equation (\ref{eqn1000}),
\begin{eqnarray}
\label{eqn44}
d^2Q&=&d(\hat v_r dA_r+\hat v_r'A_rd \phi)\nonumber\\
&=&\hat v_r' \left( d \phi\wedge dA_r \right)+\hat v_r' \left( dA_r\wedge d\phi \right)\nonumber\\
&=&(\hat v_r'-\hat v_r')d\phi\wedge dA_r=0\;,
\end{eqnarray}
where the wedge $\wedge$ signifies the antisymmetric exterior product.
The same equation expressed in the coordinate system $(A_w,A_n)$ gives
\begin{eqnarray}
\label{eq45}
d^2Q&=&d(\hat v_wdA_w+\hat v_ndA_n)\nonumber\\
&=&\left[\left(\frac{\partial \hat v_n}{\partial A_w}\right)_{A_n}
-\left(\frac{\partial \hat v_w}{\partial A_n}\right)_{A_n}\right]
dA_w\wedge dA_n\nonumber\\
&=&0\;.
\end{eqnarray}
We rewrite the coefficient in terms of the saturation coordinate system,
\begin{eqnarray}
\label{eqn46}
\left(\frac{\partial \hat v_n}{\partial A_w}\right)_{A_n}
-\left(\frac{\partial \hat v_w}{\partial A_n}\right)_{A_n}&=&\nonumber\\
\frac{d\hat v_n}{dS_w}\left(\frac{\partial S_w}{\partial A_w}\right)_{A_n}-
\frac{d\hat v_w}{dS_w}\left(\frac{\partial S_w}{\partial A_n}\right)_{A_w}&=&\nonumber\\
S_w\frac{d\hat v_w}{dS_w}+(1-S_w)\frac{d\hat v_n}{dS_w}&=&0\;,
\end{eqnarray}
which is nothing but the Gibbs-Duhem equation yet again. Writing this equation in
terms of the seepage velocities gives us equation (\ref{eq1005}).
We see that the different equations relating the velocities can all be found
from the differential geometric structure of the space spanned by the
areas, with the addition to Euler homogeneity, see Figure \ref{fig3}. All
relations are either a
consequence of writing $dQ$ using different coordinate systems, or a consequence
of the Poincar{\'e} lemma, expressed as $d^2Q=0$.
\section{The co-moving velocity in coordinate-free representation}
\label{comov}
The co-moving velocity appears in several equations above.
Common to all of them is that they all are written out in terms of a given coordinate system.
We may write the representation of the co-moving velocity in the saturation coordinate system, equation (\ref{eq1005}), as
\begin{equation}
\label{eqn48}
A_pv_m=\vec A\cdot \frac{d\vec v}{dS_w}\;.
\end{equation}
We may write equation (\ref{eq26}) in polar coordinates as
\begin{equation}
\label{eqn49}
A_r\hat v_m=-\vec A\cdot \frac{d\vec v}{d\phi}\;.
\end{equation}
Expressed in terms of the seepage velocities, we have that
\begin{equation}
\label{eq:dQ-seepage}
dQ \ = \ v_w dA_{w} + v_{n }d A_n + A_w dv_w + A_n dv_n\;.
\end{equation}
In order to obtain an expression for the co-moving velocity which is independent of
coordinate system, we construct
\begin{eqnarray}
\label{eqn50}
&&A_wdv_w+A_ndv_n=v_m\ \frac{A_ndA_w-A_wdA_n}{A_w+A_n}\nonumber\\
&=&v_m A_pdS_w=-\hat v_m A_r d\phi\;,
\end{eqnarray}
where the second line only reflects equations (\ref{eqn48}) and (\ref{eqn49}).
We write $dQ$ as
\begin{align}
\label{eqn51}
dQ=d[\vec A\cdot \hat{\vec v}] \ =& \ \hat{v}^i dA_i \;,
\end{align}
using equation (\ref{eqn1005}). Here, Einstein summation convention has been
applied, with the index $i$ running over
$\left\{ w,n \right\}$ in the basis $\left(A_w, A_n \right)$.
In terms of the seepage velocities, this becomes
\begin{align}
\label{eqn52}
dQ=d[\vec A\cdot \vec v]= v^{i} dA_{i} + A^{i} d v_{i} \;.
\end{align}
Combining equation
(\ref{eqn51}), (\ref{eqn52}) and (\ref{eq27}) gives us
\begin{equation}
\label{eqn53}
\hat{v}_{m}^i dA_i + A^i dv_{i} = 0
\end{equation}
where $ \hat{v}_m^i$ denotes the components of
$\hat{\vec{v}}_{m}$. This holds in any coordinate system, since it is just
obtained by differentiating a function, namely $\vec{A} \cdot \vec{v}$. We have
therefore obtained an expression for the co-moving velocity which is independent
of the chosen coordinate system.
We can introduce \textit{vector-valued differential forms} \cite{mtw17} to make
a connection between the interpretation in terms of (tangent) vectors used up
until now and the language of differential forms.
Let $\left\{ \vec{e}_i \right\}$ be an arbitrary basis for vectors on the area space as
we have discussed until now, be it area-, saturation- or polar coordinates.
We now take $\vec{A}$ as an example. We write
\begin{equation}
\label{eq:A-vector-valued-form}
A \ = \ \beta^i \vec {e}_i \ \;,
\end{equation}
where $A$ is now regarded as a vector-valued differential form (thus seeing $\vec e_i$ as components
written out in the $\beta^i$ basis, i.e., a reversal of view point),
the $\beta_i$ are $0$-forms, just functions, and the index $i$
runs over the coordinate indices. We can still treat $A$ as a vector, but the
difference in viewpoints is apparent when we apply the exterior
derivative $d$ to $A$, which gives
\begin{equation}
\label{eq:derivative-vector-valued-A}
dA = \sum\displaylimits_{i} (\vec{e}_{i} \otimes d\beta_i) = (d\beta^{i})\vec{e}_i \ \;.
\end{equation}
where we in the second equality apply the summation convention and have
suppressed the tensor product of the forms and basis vectors since we are
working over a vector space, as is standard notation in e.g.\ \cite{mtw17}.
Strictly speaking, the exterior derivative $d$ should be replaced by the
\textit{exterior covariant derivative} in this setting, a generalization of $d$
which is defined on both tangent vectors and forms (there is currently no term
in equation~(\ref{eq:derivative-vector-valued-A}) that reflect the changes in
the basis
$\vec{e}_i$). The terms $A^i dv_{i}$ are then possible to interpret in terms of
\textit{connection forms}. However, this is outside the scope of the current
discussion, and we leave this topic for future work.
If we write out equation (\ref{eq:derivative-vector-valued-A}) in the coordinate
system $\left( A_w, A_n \right)$, we get
\begin{equation}
\label{eq:dA-area}
dA \ = \ dA_w \vec{e}_w + dA_n \vec{e}_n \ \;.
\end{equation}
With this formalism in mind, we now return to polar coordinates. Equation
(\ref{eqn53}) may be written in these coordinates by using equation
(\ref{e108-1}) and (\ref{e108-2}) to define the vector valued $0$-forms
\begin{align}
\omega_r \ =& \ \cos\phi\ \vec{e}_w+\sin\phi\ \vec{e}_n \;,\label{eq:wr-form}\\
\omega_\phi \ =& \ -\sin\phi\ \vec{e}_w+\cos\phi\ \vec{e}_n \;,\label{eq:wphi-form}
\end{align}
which are equivalent to the corresponding basis vectors in equation
(\ref{e108-1}) and (\ref{e108-2}), but interpreteted as $0$-forms. We then compute
\begin{align}
d\omega_r \ =& \ \vec{e}_\phi d\phi\;,\label{eqn54}\\
d\omega_\phi \ =& \ - \omega_rd\phi \;,\label{eqn55}
\end{align}
so that
\begin{align}
d A = d( A_r \omega_{r }) \ =& \ dA_r \omega_r+A_rd \omega_r \nonumber \\
=& \ dA_r \vec{e}_r + A_r \vec{e}_{\phi} d\phi \;, \label{eqn56}
\end{align}
where we in the second equality has used that $\omega_r$ is just another way of
writing the basis vector $\vec{e}_r$ that emphasizes its role as a form.
We can similarly define a vector valued $1$-forms $d\hat{v}$ and $dv$ from
equation~(\ref{eq19}) and (\ref{eq10}) respectively. We then straightforwardly obtain
\begin{align}
d\hat{v} \ =& \ d\hat{v}_w \vec{e}_w + d\hat{v}_n \vec{e}_n \label{eq:hat-v-vector-valued-form} \;, \\
dv \ =& \ dv_w \vec{e}_w + dv_n \vec{e}_n \label{eq:v-vector-valued-form} \;.
\end{align}
We express $dv$ in polar coordinates as
\begin{eqnarray}
\label{eqn57}
dv&=&v_r' \vec{e}_rd\phi+v_r\vec{e}_\phi d\phi+v_\phi' \vec{e}_\phi d\phi-v_\phi \vec{e}_r d\phi\nonumber\\
&=&\left[(v_r'-v_\phi) \vec{e}_r +(v_r+v_\phi') \vec{e}_\phi\right]d\phi\;,
\end{eqnarray}
giving
\begin{equation}
\label{eqn58}
A_r {\hat v}_m d\phi= A_r(v_\phi-v_r')d\phi\;,
\end{equation}
or
\begin{equation}
\label{eqn59}
\hat v_m = v_\phi-v_r'\;.
\end{equation}
By combining (\ref{eq14}), (\ref{eq29}) and (\ref{eqn1003}), we obtain the same equation.
Equation (\ref{eqn53}) written in the saturation coordinate system yields equation (\ref{eq1005}).
\section{Conclusion and discussion}
\label{conclusion}
The aim of this work has been to formulate the immiscible two-phase flow in porous media problem
as a {\it geometrical problem\/} in area space. We did this by
\begin{itemize}
\item defining the two pore area variables $A_w$ and $A_n$ and consider the space they span,
\item endow this space with different coordinate systems, $(A_w,A_n)$, $(A_p,S_w)$ and
$(A_r,\phi)$, pointing out that expressions that using the polar coordinate system simplifies
the discussion considerably,
\item recognizing the meaning of the volumetric flow rate being a
degree-$1$ homogeneous function when expressing it in polar coordinates,
and deriving a number of properties related to the different velocities
in the problem; the seepage velocities, the thermodynamic velocities and
the co-moving velocity,
\item using differential forms to derive relations between the different velocities, and lastly,
\item formulate an expression for the co-moving velocity which is independent of the coordinates
on the underlying space.
\end{itemize}
We remind the reader of the following equations.
The difference between the thermodynamic and seepage velocities may be expressed by combining equations (\ref{eqn51})
and (\ref{eqn52}),
\begin{equation}
\label{eqn150}
dQ=\hat{v}^{i} dA_{i}= A^{i} d v_{i} + v^{i} dA_{i} \;.
\end{equation}
The co-moving velocity is given by equation (\ref{eq27}),
\begin{equation}
\vec v=\hat{\vec v}+\hat{\vec v}_m\;.\nonumber
\end{equation}
These equations lead us to the three central equations summarizing the central results of this paper.
They are equation (\ref{eq18}),
\begin{equation}
\vec v\cdot \vec A=Q\;,
\nonumber
\end{equation}
equation (\ref{eq29-1}),
\begin{equation}
\hat{\vec v}_m\cdot\vec A=0\;,
\nonumber
\end{equation}
and equation (\ref{eqn53}),
\begin{equation}
\hat{v}_{m}^i dA_i + A^i dv_{i} = 0
\nonumber
\end{equation}
It is clear from this discussion that the co-moving velocity, $v_m$, which together with the average seepage velocity
$v$, determines $v_w$ and $v_n$ through the transformations (\ref{e1}) and (\ref{e2}), cannot be found by any
measurement of the volumetric flow rate $Q$ or any expression gotten from it and it alone such as $dQ$. The
fundamental question, which remains open is the following: is it possible to measure the co-moving velocity $v_m$
{\it without\/} explicitly measuring $v_w$ and $v_n$?
\bigskip
The authors thank Dick Bedeaux, Carl Fredrik Berg, Eirik Grude Flekk{\o}y, Magnus Aa.\ Gjennestad,
Signe Kjelstrup, Marcel Moura,
Knut J{\o}rgen M{\aa}l{\o}y, Santanu Sinha, Per Arne Slotte and Ole Tors{\ae}ter for
interesting discussions. This work was partly supported by the Research Council of Norway
through its Centres of Excellence funding scheme, project number 262644.
|
\section{Introduction}
\label{sec:intro}
Video Super-Resolution (VSR) aims to recover high-resolution (HR) frame sequences from their low-resolution (LR) counterparts, where the utilization of complementary information across adjacent frames is the key factor to improve performance.
VSR can be roughly classified into two categories -- sliding-window-based methods~\cite{caballero2017real,huang2017video,wang2019edvr,tian2020tdan,li2020mucan,isobe2020video} and recurrent methods~\cite{chan2021basicvsr,chan2022basicvsr++,huang2015bidirectional,sajjadi2018frame,haris2019recurrent,xiang2020zooming,isobe2020video,isobe2020revisiting,lin2021fdan}.
Recurrent methods are the prevalent approaches since they can model long-term dependencies with hidden states, and have won several championships in video restoration competitions~\cite{yang2021ntire,yang2022ntire}.
\begin{figure}[t]
\centering
\includegraphics[width=1.0\columnwidth]{figs/teasor.pdf}
\caption{Our proposed FastRealVSR can ahieve the best trade-off between NIQE, speed and parameters in the real-world VideoLQ dataset.}
\label{fig:speed_cmp}
\end{figure}
When applied to real-world low-quality videos, existing recurrent methods (\eg, BasicVSR~\cite{chan2021basicvsr}) tend to produce unpleasant artifacts, which will be amplified through propagation~\cite{chan2022investigating}. This problem becomes more severe when the degradations are more complicated with unknown blur, noise and compression.
Recent work Real-BasicVSR~\cite{chan2022investigating} makes the first attempt to alleviate those artifacts. It adopts a separate image pre-cleaning module before the video recurrent restoration model to explicitly remove degradations for each video frame.
However, while mitigating unpleasant artifacts, such a strategy can also lead to detail loss, due to the explicit denoising operation in the image level.
Besides, the separate module brings extra computation cost, which slows down the inference speed and is not friendly to practical usages.
In this work, we carefully investigate the cause of the unpleasant artifacts in VSR. We draw two observations.
1) Once the output of the previous frame produces artifacts, the hidden state will also contain such artifacts and further propagate to consequent frames. When we discard the hidden state and only use the current input frame, those artifacts disappear, but at the cost of detail loss.
2) When we replace the ``problematic'' hidden state with a cleaner one (\eg, the corresponding one from the MSE model\footnote{Training with MSE loss tends to produce over-smooth results and thus the hidden state is also cleaner and more smooth.} or a Gaussian-blurred one), the artifacts are largely mitigated and do not further propagate to consequent frames.
Such observations motivate us to directly manipulate the hidden state instead of introducing extra cleaning modules.
We propose a Hidden State Attention (HSA) module to mitigate artifacts in real-world video super-resolution.
Specifically,
we first generate a hidden state pool, where each new hidden state is produced by filtering the original hidden state with different filters, including Gaussian blur filters, box blur filters, sharpening filters, \etc.
Those operations will either blur hidden state to alleviate artifacts or sharpen edges to enhance details.
We choose those classical filters as they only bring a little computation cost.
After that, selective cross attention will calculate the similarity between the features of the current frame and each hidden state in the pool.
As ``problematic'' hidden state usually has a large deviation from input features, such a design can effectively mitigate artifacts.
A cleaned hidden state is then aggregated from the hidden state pool based on the calculated attention.
As a result, this new hidden state not only contains fewer artifacts, but also maintains the details.
\begin{figure}[t]
\centering
\includegraphics[width=.95\columnwidth]{figs/hidden_state.pdf}
\caption{Negative impact of hidden state in real-world scenario. The hidden state could contain artifacts that appears in the previous frames, and further propagate to consequent frames. The regions marked with the ellipse dotted box are the artifacts. (\textbf{Zoom-in for best view})}
\label{fig:w_wo_hidden_state}
\end{figure}
To make our algorithm more practical in the real world, we also take inference speed into consideration. We adopt a compact unidirectional recurrent framework as the backbone of our method. Together with the efficient Hidden State Attention (HSA) module, our proposed method, namely FastRealVSR, achieves $2\times$ speed-up over Real-BasicVSR, while keeping comparable performance.
Our contributions can be summarized as follows.
\textbf{1)} We analyze the hidden state in recurrent VSR methods and show that the hidden state is the main reason for unpleasant artifacts.
\textbf{2)} An effective hidden state attention with cheap hidden states and selective cross attention is proposed to mitigate artifacts in real-world VSR models.
\textbf{3)} FastRealVSR is capable of achieving a better speed-performance trade-off, surpassing previous methods.
\section{Related Work}
\label{sec:related_work}
\begin{figure}[t]
\centering
\includegraphics[width=1\columnwidth]{figs/inference_arch.pdf}
\caption{The diagram of general unidirectional recurrent network. $\mathrm{Model_{combine}}$ is derived by replacing the GAN hidden state in $\mathrm{Model_{gan}}$ with the MSE hidden state in $\mathrm{Model_{mse}}$. We use a single color to represent a complete pre-trained model. (\textbf{Zoom-in for best view})}
\label{fig:inference}
\end{figure}
\textbf{Video Super Resolution.} Most existing VSR methods can be divided into two categories: sliding-window based methods~\cite{caballero2017real,huang2017video,wang2019edvr,tian2020tdan,li2020mucan,isobe2020video} and recurrent methods~\cite{chan2021basicvsr,chan2022basicvsr++,huang2015bidirectional,sajjadi2018frame,haris2019recurrent,xiang2020zooming,isobe2020video,isobe2020revisiting,lin2021fdan}. Sliding-window methods often take several adjacent frames as input and predict the center frame. In contrast to sliding-window based methods, recurrent methods adopt recurrent neural network~\cite{mikolov2010recurrent} to process the input frame and generate the output progressively. Some recent methods~\cite{chan2021basicvsr,chan2022basicvsr++} have achieved excellent performance on synthetic degradations (\eg, bicubic).
Recently, several transformer-based~\cite{cao2021video,liang2022vrt} methods are proposed. However, when these methods~\cite{chan2021basicvsr} are applied to real-world scenarios, their performance deteriorates severely and obvious artifacts will appear since the degradations in real-world scene are unknown and complex.
\noindent \textbf{Real-World Video Super-Resolution.} Recent work RealBasicVSR~\cite{chan2022investigating} is proposed to investigate tradeoffs in real-world VSR. Similar to Real-ESRGAN~\cite{wang2021real}, a second-order degradation model is adopted to simulate the degradations in the real-world. To alleviate the artifacts that are amplified in recurrent
methods (\eg BasicVSR), it uses a pre-cleaning module and a cleaning loss to ``clean'' the input sequence, which is then fed into a general VSR network. In this paper, to mitigate the artifacts, we do not introduce any cleaning modules. Instead, we manipulate the hidden state directly to mitigate the artifacts. Compared to RealBasicVSR, our proposed FastRealVSR achieves a better speed-performance trade-off.
\noindent \textbf{Hidden State Manipulation.} Hidden state is the key component in recurrent methods that adopted in video restoration task~\cite{nah2019recurrent,zhong2020efficient}. To handle the misalignment between the hidden state and the referred current feature, some works~\cite{nah2019recurrent,zhou2019spatio} in video deblurring area are proposed to adapt the hidden states to the current frame by referring to the current feature. Different from them, we aggregate a new hidden state from a hidden state pool and aim to alleviate the mitigate contained in ``problematic'' hidden state.
\section{Methodology}
\label{sec:method}
\subsection{Analysis of Hidden State}
\begin{figure}[t]
\centering
\includegraphics[width=1\columnwidth]{figs/replaced_gan_with_mse.pdf}
\caption{Qualitative comparison between $\mathrm{Model_{mse}}$, $\mathrm{Model_{gan}}$ and $\mathrm{Model_{combine}}$. As shown on the left side of the figure, after replacing GAN hidden state with MSE hidden state, the artifacts in the wall are largely mitigated. As shown on the right side of the figure, the output of $\mathrm{Model_{combine}}$ also loses many necessary details in the railing. (\textbf{Zoom-in for best view})}
\label{fig:replace_gan_with_mse}
\end{figure}
\textbf{Observation 1:} \textit{Hidden state in recurrent methods propagates the artifacts.}
In a unidirectional recurrent method, information in the hidden state serves two purposes. The first is to reconstruct the current frame with a reconstruction module, and the second is to propagate the previous information to the next frame. Therefore, once the output of the previous frame produces artifacts, the hidden state also contains such artifacts and further propagates to consequent frames. As shown in the $1^{st}$ and $3^{rd}$ rows in Fig.~\ref{fig:w_wo_hidden_state}, the contents of ellipse dotted box in $T_{t-1}$, $T_{t}$ retain and amplify the artifacts that appears in $T_{t-2}$. If the hidden state is not used (setting to $0$) and only the information of input image is used, the artifacts will be clearly alleviated. This phenomenon can be observed in the $2^{nd}$ and $4^{th}$ rows in Fig.~\ref{fig:w_wo_hidden_state}.
Directly discarding the hidden state, however, also lead to detail loss and unsharp edges ($2^{nd}$ and $4^{th}$ rows in Fig.~\ref{fig:w_wo_hidden_state}).
In a word, the hidden state not only enhances details but also propagates the artifacts, which is consistent with the observations in RealBasicVSR~\cite{chan2022investigating}.
\begin{figure}[t]
\centering
\includegraphics[width=1.0\columnwidth]{figs/filter.pdf}
\caption{\textbf{Effects of different kernels.} For regions with artifacts, blur filters can mitigate artifacts. For regions with details, sharpening filters can further sharpen details.}
\label{fig:filter}
\end{figure}
\noindent \textbf{Observation 2:} \textit{Replacing the hidden state in the GAN-based model with its corresponding hidden state in the MSE-based model can mitigate the artifacts.}
When the hidden state contains artifacts, although directly discarding the hidden state can alleviate the artifacts, it also leads to obvious details loss in other regions. Therefore, it is better to find a cleaner variant of the ``problematic'' hidden state which contains fewer artifacts and maintains its details. Typically, a GAN-based model is fine-tuned from an MSE-based model with a combination of MSE loss, perceptual loss~\cite{johnson2016perceptual} and GAN loss~\cite{goodfellow2014generative}. As shown in the $1^{st}$ row of Fig.~\ref{fig:replace_gan_with_mse}, the model trained with MSE loss tends to produce over-smooth results, thus the hidden state in the MSE-based model is cleaner and smooth.
Consequently, we wonder whether the artifacts of restored outputs can be mitigated if we replace the 'problematic' hidden state with the corresponding one in the MSE model. For better clarification, the hidden state in the MSE-based model and the hidden state in GAN-based model are referred to as MSE hidden state and GAN hidden state, respectively.
\begin{figure*}[t]
\centering
\includegraphics[width=1.9\columnwidth]{figs/framewrok.pdf}
\caption{\textbf{Framework overview of FastRealVSR.} The left flowchart is a general unidirectional recurrent network. $f_{s}(t)$ and $f_{d}(t)$ represent the shallow and sharp feature of $\bm{X}_{t}$, respectively. $h_{t-1}$ is the hidden state that is propagated to restore $\bm{X}_{t}$. By adopting hidden state attention (HSA), $h_{t-1}$ is transformed into $\hat h_{t-1}$. HSA is composed of a hidden state pool and a selective cross attention module, as shown on the right side of the figure.}
\label{fig:framework}
\end{figure*}
Fig.~\ref{fig:inference} illustrates the simplified diagram of a unidirectional recurrent network that uses both hidden state and input frame as inputs.
In Fig.~\ref{fig:inference}, we use a single color to represent a complete pre-trained inference model. To replace the GAN hidden state with MSE hidden state, firstly, we need to use $ \mathrm{Model_{mse}}$ (pre-trained MSE model) to infer the input sequence and store all generated hidden states. Then, we will infer the input sequence with $ \mathrm{Model_{combine}}$. $ \mathrm{Model_{combine}}$ is modified from $\mathrm{Model_{gan}}$ and the difference is that the GAN hidden state is replaced with the previously stored MSE hidden state. The qualitative comparison between these three models is shown in Fig.~\ref{fig:replace_gan_with_mse}. We can observe that \textbf{1)} from the comparison of the $2^{nd}$ and $3^{rd}$ rows (Fig.~\ref{fig:replace_gan_with_mse} (a)), the artifacts in $ \mathrm{Model_{gan}} $ are reduced in $ \mathrm{Model_{combine}} $. \textbf{2)} directly replacing the GAN hidden state with the MSE hidden state could sometimes lose the necessary details (Fig.~\ref{fig:replace_gan_with_mse} (b)), compared against $\mathrm{Model_{gan}}$.
Although replacing the GAN hidden state with the MSE hidden state can mitigate artifacts, it is unpractical with two shortcomings. \textbf{1)} To obtain final results, it takes the double time since we need to run $\mathrm{Model_{mse}}$ and $\mathrm{Model_{combine}}$ in sequence. \textbf{2)} The output of $\mathrm{Model_{combine}}$ also loses many necessary details compared against $\mathrm{Model_{gan}}$. Therefore, it is preferable to directly manipulate the GAN hidden state, and construct a new hidden state that not only contains the smooth counterpart of the artifact region in the GAN hidden state, but also maintains the details of other regions.
\subsection{Framework}
Motivated by the above observations, we propose a hidden state attention (HSA) module to directly manipulate the hidden state and mitigate artifacts. An overview is illustrated in Fig.~\ref{fig:framework}.
\noindent \textbf{Formulation.} We first briefly introduce the pipeline of the general unidirectional recurrent VSR method. Let $X_{t}$ be the $t$-th image of the input sequence and $Y_{t}$ be the $t$-th frame of the restored sequence. Several residual blocks $\mathrm{RB_{1}}$ are used to extract the shallow feature of $X_{t}$:
\begin{equation}%
f_{s}(t) = \mathrm{RB_{1}}(X_{t}).
\end{equation}
The hidden state from previous frame is first warpped with the estimated optical flow. The warpped hidden state is denoted as $h_{t-1}$. We use a pre-trained SPyNet~\cite{ranjan2017optical} to generate the optical flow. Then $h_{t-1}$ is concatenated with $f_{s}(t)$ and input to another set of residual blocks $\mathrm{RB_{2}}$ to get the deep feature $f_{d}(t)$:
\begin{equation}%
f_{d}(t) = \mathrm{RB_{2}}(\mathrm{Concat}(h_{t-1}, f_{s}(t))),
\end{equation}
\noindent where $h_{t-1} \in \mathbb{R}^{H \times W \times C}$. $f_{d}(t)$ is then passed to an upsampling block $\mathrm{UP}$, and added with the bilinear upsampling of $X_{t}$ to generate $Y_{t}$:
\begin{equation}%
Y_{t} = \mathrm{UP}(f_{d}(t)) + \mathrm{Bilinear}(X_{t}).
\end{equation}
For our method, we replace $h_{t-1}$ with $\hat h_{t-1}$ with
HSA:
\begin{equation}%
\hat h_{t-1} = \mathrm{HSA}(h_{t-1}).
\end{equation}
\noindent where HSA is composed of a hidden state pool module and a selective cross attention module. The specific details of these two modules are described below.
\noindent \textbf{Hidden State Pool.} Hidden state pool is a container that gathers the blurry and sharp versions of $h_{t-1}$. The blurry version is obtained by a blurring operation, while the sharp version is obtained by the simplified sharpening algorithm (\ie unsharp masking (USM)). Specially, the sharp version is obtained by first subtracting the blurry version from the input, and then adding the residual to the input.
The above design is inspired by feature denoising~\cite{xie2019feature} that is adopted to clean feature maps for improving the adversarial robustness of convolutional networks. The effects of different blur operations are shown in Fig.~\ref{fig:filter}. Here, we use images to better illustrate the effects of blurry and sharpening filters. For the region with artifacts, we prefer its blurry version, while for the region with correct details, we prefer its sharp version.
In practical implementation, the operation of generating different versions of $h_{t-1}$ can be easily achieved by a convolution layer with a fixed kernel. Such computational operations are very cheap, and the computational cost is negligible. The process of generating the $i$-th blurry version of $h_{t-1}$ and the $j$-th sharp version of $h_{t-1}$ are formulated as:
\begin{equation}
h_{t-1}^{i}= h_{t-1} \otimes \mathbf{k}^{i},
\end{equation}
\begin{equation}
h_{t-1}^{j}= h_{t-1} + (h_{t-1} - h_{t-1} \otimes \mathbf{k}^{j}),
\end{equation}
\noindent where $\otimes$ denotes the convolving operation and $\mathbf{k}^{i}$, $\mathbf{k}^{j}$ are the blur kernels. The degree of blurring and sharpening can be controlled by adjusting the kernel size and variance (only specifically for the Gaussian filter). In our experiment, two mean filters and one Gaussian filter are adopted to generate blurry versions. We use two sharpening filters to generate the sharp versions. Detailed configurations of these filters are provided in the appendix.
\noindent \textbf{Selective Cross Attention.} After obtaining the hidden state pool $h_{t-1}^{i} \in \mathbb{R}^{H \times W \times C}$ ($1 \leq i \leq N$) ($N$ denotes the number of hidden state included in the hidden state pool), it is vital to aggregate a new hidden state that contains fewer artifacts. From the $2^{nd}$ and $4^{th}$ rows of Fig.~\ref{fig:w_wo_hidden_state}, we can observe that the artifacts in the output are largely mitigated if only the information of the input frame is utilized, indicating that the input feature contains fewer artifacts. Therefore, for the region with artifacts, the ``problematic'' hidden state has a large deviation from input features. Inspired by this, we propose a selective cross attention (SCA) module to aggregate a new hidden state $\hat h_{t-1}$ based on the attention between the hidden state pool $h_{t-1}^{i} \in \mathbb{R}^{H \times W \times C}$ ($1 \leq i \leq N$) and feature of input frame.
As shown in Fig.~\ref{fig:framework}, we first process $f_{s}(t)$ with a convolutional layer $H_{\mathrm{conv}}$, and get the query $Q \in \mathbb{R}^{H \times W \times C}$:
\begin{equation}
Q=H_{\mathrm{conv}}(f_{s}(t)).
\end{equation}
For all $h_{t-1}^{i}$ ($1 \leq i \leq N$), we adopt two $3 \times 3$ convolutional layers to generate the key $K_{t-1}^{i}$ ($1 \leq i \leq N$) and value $V_{t-1}^{i}$ ($1 \leq i \leq N$), respectively. Then we calculate the similarity between each $K_{t-1}^{i}$ and $Q$ by the attention mechanism~\cite{vaswani2017attention,liu2021swin}. Based on the calculated attention maps, we aggregate $\hat h_{t-1}^{i}$ from $V_{t-1}^{i}$ ($1 \leq i \leq N$). The attention matrix is computed as
\begin{equation}
\mathrm{SoftMax}([Q(K^{1})^{T};...;Q(K^{N})^{T}]) \otimes (V^{1};...;V^{N}).
\end{equation}
\section{Experiments}
\label{sec:experiments}
\subsection{Setup}
\noindent \textbf{Training Settings.} We train our FastRealVSR on the REDS~\cite{reds} dataset. The following degradation model is adopted to synthesize training data:
\begin{equation}%
\bm{X} = [(\bm{Y} \circledast \bm{k}_{\sigma} + \bm{n}_{\delta})\downarrow_{r}]_{\mathrm{FFMPEG}}\ ,
\end{equation}
\noindent where $\bm{X}$ and $\bm{Y}$ are paired low-resolution and high-resolution sequences. The $\bm{k}$, $\bm{n}$, and $r$ are blur kernel, additive noise, and downsampling factor, respectively. We use a constant rate factor (\textit{crf}) to control the degree of FFMPEG compression. \textit{crf} uses a specific quality by adjusting the bitrates automatically. The sampling range of $\sigma$, $\delta$, and $\textit{crf}$ are $\{0.2 : 3\}, \{1 : 5\}$, and $\{18 : 35\}$, respectively.
\begin{figure}[t]
\centering
\includegraphics[width=1.0\columnwidth]{figs/training_hidden_state_cmp.pdf}
\caption{Results of URN-GAN, FastRealVSR-Blurry, FastRealVSR-Sharp and FastRealVSR. The outputs of URN-GAN and FastRealVSR-Sharp contain unpleasant artifacts. (\textbf{Zoom-in for best view})}
\label{fig:training_hidden_state_cmp}
\end{figure}
Following RealBasicVSR, we load 15 frames as a sequence and flip them temporally in each iteration. The patch size of input LR frames is $64 \times 64$. The pre-trained SPyNet~\cite{ranjan2017optical} is used as the flow network and its weights are fixed in training. We employ Adam optimizer~\cite{kingma2014adam}. The training process is divided into two stages.
\begin{figure*}[t]
\centering
\includegraphics[width=2.0\columnwidth]{figs/filter_replacement_effect.pdf}
\vspace{-0.1cm}
\caption{Analysis of hidden state pool and SCA. The outputs of URN-GAN contain obvious artifacts (\ding{172}). The artifacts produced in URN-GAN are largely alleviated in FastRealVSR (\ding{173}). When all hidden states in the hidden state pool are replaced with a single blurry hidden state, the artifacts are mitigated (\ding{174}). While the artifacts are retained when all hidden states are replaced with a single sharp hidden state (\ding{175}). The summed blurry/sharp attention maps are visualized in \ding{176} and \ding{177}, respectively. The binary map (\ding{178}) indicates the region where the value of \ding{176} is larger than \ding{177}. The regions marked with ellipse dotted boxes are artifacts. (\textbf{Zoom-in for best view})}
\label{fig:filter_replacement_effect}
\vspace{-0.2cm}
\end{figure*}
In the first stage, we adopt the Unidirectional Recurrent Network (URN) shown in Fig.~\ref{fig:framework}, and train it for $300$K iterations with the
$\textit{L}_{1}$ loss. The batch size and learning rate are set to $16$ and $10^{-4}$. In the second stage, we equip URN with the proposed HSA module to get the network FastRealVSR. We employ the pre-trained MSE model for initialization. Then we train FastRealVSR for $70$K iterations with a combination of $\textit{L}_{1}$ loss, perceptual loss~\cite{johnson2016perceptual} and GAN loss~\cite{goodfellow2014generative}, whose loss weights are set to ${1, 1, 5{\times}10^{-2}}$, respectively.
The batch size is set to $8$. The learning rate of the generator and discriminator is set to $5{\times}10^{-5}$ and $5{\times}10^{-4}$, respectively.
The exponential moving average (EMA)~\cite{yaz2018unusual} is adopted in training for more stable training and better performance. All experiments are implemented on 8 NVIDIA A100 GPUs with PyTorch~\cite{paszke2019pytorch}.
\noindent \textbf{Network Configuration.} In URN, the number of residual blocks that are used to extract the shallow and deep features is set to $2$ and $28$, respectively. The number of feature channels is $64$. In adversarial training, we adopt the discriminator of Real-ESRGAN~\cite{wang2021real}. For the hidden state pool, $3$ blur filters and $2$ sharpening filters are chosen to generate different hidden state variants. Detailed descriptions are provided in the appendix.
\noindent \textbf{Evaluation Dataset and Metrics.} We use VideoLQ~\cite{chan2022investigating} as test set and employ NIQE~\cite{niqe} and BRISQUE~\cite{brisque} to evaluate visual quality.
\subsection{Ablation Studies}
\noindent \textbf{Hidden State Pool.}
We use the GAN version of URN (URN-GAN) as the baseline model. Then we train FastRealVSR, FastRealVSR-Blurry and FastRealVSR-Sharp with the same settings. The only difference among them is the representation type of the hidden states in the hidden state pool. To be specific, FastRealVSR-Blur contains only blurry variants of hidden states, while FastRealVSR-Sharp contains only sharp ones. The quantitative comparison is listed in Tab.~\ref{tab:train_hidden_state}. The performance of FastRealVSR is significantly better than FastRealVSR-Blurry and FastRealVSR-Sharp, indicating that the information in the blurry and sharp representations of hidden states is complementary to each other.
Fig.~\ref{fig:training_hidden_state_cmp} shows the restored results of these four models. It can be observed that FastRealVSR and FastRealVSR-Blurry both produce outputs with fewer artifacts, whereas the outputs of URN-GAN and FastRealVSR-Blurry contain obvious artifacts. Besides, the edges of FastRealVSR are sharper than FastRealVSR-Blurry.
\noindent \textbf{Selective Cross Attention.} In SCA, each hidden state interacts with input features by attention. If the input feature is not used to guide the aggregation of hidden states and only a convolution layer is adopted to aggregate these hidden states, a performance drop is observed in Tab.~\ref{tab:sca}. It indicates that the guidance of input features is necessary when aggregating a new hidden state from a hidden state pool.
\begin{table}[th]
\centering
\caption{Ablation studies of different representation types in hidden state pool. FastRealVSR achieves the best performance.}
\vspace{-0.2cm}
\label{tab:train_hidden_state}\resizebox{0.5\linewidth}{!}{
\begin{tabular}{c|c} \hline
\textit{Network} & \textit{NIQE}$\downarrow$ \\ \hline
URN-GAN & 3.8435 \\ \hline
FastRealVSR-Blurry & 3.8544 \\ \hline
FastRealVSR-Sharp & 3.8210 \\ \hline
FastRealVSR & 3.7658 \\ \hline
\end{tabular}
\vspace{-0.5cm}
}
\end{table}
\begin{table}[th]
\centering
\vspace{-0.7cm}
\caption{Ablation study of SCA. With SCA, FastRealVSR achieves better performance.}
\label{tab:sca}\resizebox{0.6\linewidth}{!}{
\begin{tabular}{c|c|c} \hline
\textit{Variant} & w/o SCA & FastRealVSR (w/ SCA)\\ \hline
\textit{NIQE}$\downarrow$ & 3.8446 & 3.7658 \\ \hline
\end{tabular}
}
\vspace{-0.5cm}
\end{table}
\begin{figure*}[t]
\centering
\vspace{-0.4cm}
\includegraphics[width=2.0\columnwidth]{figs/comparison.pdf}
\caption{Qualitative comparison on the VideoLQ dataset for $\times 4$ video SR. (\textbf{Zoom-in for best view})}
\vspace{-0.2cm}
\label{fig:cmp}
\end{figure*}
\begin{table*}[t]
\centering
\caption{Quantitative comparison on VideoLQ dataset. FastRealVSR is capable of achieving a better speed-performance trade-off. Runtime is computed with an output size of $720 \times 1280$, with an NVIDIA V100 GPU. \redbf{Red} and \blueud{blue} indicates the best and second best performance. Following RealBasicVSR, metrics are computed on the Y-channel of the first, middle, and last frames of each sequence.}
\vspace{-0.3cm}
\label{tab:cmp}\resizebox{\linewidth}{!}{
\begin{tabular}{l|c|c|c|c|c|c|c|c|cl} \hline
& Bicubic & DAN & RealSR & BSRGAN & Real-ESRGAN & RealVSR & DBVSR & RealBasicVSR & \textbf{FastRealVSR} & \\ \hline
Params (M) & - & 4.3 & 16.7 & 16.7 & 16.7 & \redbf{2.7} & 25.5 & 6.3 & \blueud{3.8} & \\
Runtime (ms) & - & 185 & 149 & 149 & 149 & 1082 & 239 & \blueud{63} & \redbf{30} & \\ \hline
NIQE~\cite{niqe} $\boldsymbol{\downarrow}$ & 8.0049 & 7.1230 & 4.1482 & 4.2460 & 4.2091 & 8.0606 & 6.7866 & \blueud{3.7662} & \redbf{3.7658} & \\
BRISQUE~\cite{brisque} $\boldsymbol{\downarrow}$ & 54.899 & 51.563 & 30.542 & 30.213 & 32.103 & 54.988 & 50.936 & \redbf{29.030} & \blueud{29.374} & \\ \hline
\end{tabular}
}
\vspace{-0.2cm}
\end{table*}
\subsection{Analysis of Hidden State Attention}
The proposed HSA module aims to alleviate artifacts in the hidden state. Hence, we further validate the effectiveness of HSA for the ``problematic'' hidden state. When the previous output contains artifacts, the hidden state generated in URN-GAN contains artifacts and is a ``problematic'' hidden state (see Fig.~\ref{fig:training_hidden_state_cmp} and the $1^{st}$ row in Fig.~\ref{fig:filter_replacement_effect}). To get the ``problematic'' hidden state, we run URN-GAN and store all generated hidden states. We use $h_{t}$ ($1 \leq t \leq L - 1$) to refer to the ``problematic'' hidden state that produced by URN-GAN during the process of restoring low-resolution frame $\bm{X}_{t}$.
To test whether HSA in RealFastVSR has the capacity to mitigate artifacts, we replace the original hidden state in RealFastVSR with the ``problematic'' hidden state $h_{t-1}$ from URN-GAN when restoring $\bm{X}_{t}$.
From the $2^{nd}$ row in Fig.~\ref{fig:filter_replacement_effect}, we can observe that HSA has the ability to mitigate artifacts contained in the ``problematic'' hidden state.
\noindent \textbf{Investigation of Different Filters in Hidden State Pool.} We then investigate the respective effectiveness of blurry and sharp filters utilized in the hidden state pool. Specifically, we replace all hidden states in the hidden state pool with a single variant of hidden state, \ie, the hidden state pool contains identical hidden states after same filters. Note that the number of hidden states in the hidden state pool remains the same.
The results of adopting only a single blurry or a single sharp filter in hidden state pool are shown in the $3^{rd}$ and $4^{th}$ rows of Fig.~\ref{fig:filter_replacement_effect}, respectively. We can observe that blurry filters can mitigate artifacts but with the cost of detail loss, whereas sharp filters enhance both artifacts and details. Therefore, \textit{the combination of blurry filters and sharp filters can make a better trade-off between enhancing details and suppressing artifacts.}
\noindent \textbf{Visualization of Attention Map in SCA.} We also visualize the attention map in the selective cross attention (SCA) module to understand how the information in different representations of hidden state is selected based on the input feature. For better visualization, blurry and sharp attention maps are summed together, respectively, resulting in two maps. The visualization of these two maps is shown in the $5^{th}$ and $6^{th}$ rows in Fig.~\ref{fig:filter_replacement_effect}.
We can observe that for different versions (\ie, blurry or sharp) of $h_{t-1}$, SCA indeed learns to select the proper information with end-to-end training.
Besides, we compare these two maps and get the binary map shown in the $7^{th}$ row in Fig.~\ref{fig:filter_replacement_effect}. The binary map indicates the region where the value of blurry attention map is larger than the sharp one. It can be observed that \textit{for the regions with artifacts, SCA prefers the blurry representations.}
\subsection{Comparison with State-of-the-Art Methods}
We compare our FastRealVSR method with seven algorithms: DAN~\cite{dan}, RealSR~\cite{real-sr}, BSRGAN~\cite{bsrgan}, Real-ESRGAN~\cite{wang2021real}, RealVSR~\cite{realvsr}, DBVSR~\cite{dbvsr} and RealBasicVSR~\cite{chan2022investigating} on the VideoLQ~\cite{chan2022investigating} dataset. VideoLQ consists of $50$ real-world low-quality videos that are collected from various video hosting sites such as Flickr and YouTube.
The quantitative results on VideoLQ are shown in Tab.~\ref{tab:cmp}. Compared with other methods, FastRealVSR achieves a better speed-performance trade-off. In particular, compared to RealBasicVSR, FastRealVSR achieves comparable performance with $2$ $\times$ faster speed and smaller model size. Qualitative results are presented in Fig.~\ref{fig:cmp}. Compared to previous methods, FastRealVSR can effectively remove annoying artifacts and preserve sharp details.
\section{Acknowledgments}
This work was supported in part by the National Natural Science Foundation of China under Grant (62276251,U1913210), the Joint Lab of CAS-HK, in part by the Shanghai Committee of Science and Technology, China (Grant No. 20DZ1100800).
\section{Appendix}
In this section, we first provide detailed descriptions of adopted blur filters and sharpening filters. Then we show more quantitative and qualitative comparisons with previous works.
\begin{figure*}[t]
\centering
\includegraphics[width=2.0\columnwidth]{figs/cmp1.pdf}
\vspace{-0.4cm}
\caption{Qualitative comparison on the VideoLQ dataset for $\times 4$ video SR. Compared against other methods, FastRealVSR can better remove the noise around texts and buildings, with clearer edges. (\textbf{Zoom-in for best view})}
\label{fig:cmp1}
\end{figure*}
\subsection{Details of Adopted Filters}
\textbf{Blurry filters:} In our experiment, we adopt two mean filters and one Gaussian filter to generate three blurry hidden states. Their specific kernels are listed in Tab.~\ref{tab:blurry_filters}.
\begin{table}[th]
\centering
\caption{The specific kernel of adopted blur filters.}
\label{tab:blurry_filters}\resizebox{0.7\linewidth}{!}{
\begin{tabular}{|c|c|c|} \hline
Index & Filter type & Filter weight \\ \hline
\ding{172} & Mean filter & $\begin{bmatrix}
\frac{1}{9} & \frac{1}{9} & \frac{1}{9} \\
\frac{1}{9} & \frac{1}{9} & \frac{1}{9} \\
\frac{1}{9} & \frac{1}{9} & \frac{1}{9}
\end{bmatrix}$ \\ \hline
\ding{173} & Mean filter & $\begin{bmatrix}
\frac{1}{25} & \frac{1}{25} & \frac{1}{25} & \frac{1}{25} & \frac{1}{25}\\
\frac{1}{25} & \frac{1}{25} & \frac{1}{25} & \frac{1}{25} & \frac{1}{25}\\
\frac{1}{25} & \frac{1}{25} & \frac{1}{25} & \frac{1}{25} & \frac{1}{25} \\
\frac{1}{25} & \frac{1}{25} & \frac{1}{25} & \frac{1}{25} & \frac{1}{25} \\
\frac{1}{25} & \frac{1}{25} & \frac{1}{25} & \frac{1}{25} & \frac{1}{25}
\end{bmatrix}$ \\ \hline
\ding{174} & Gaussian filter & $\begin{bmatrix}
0.1108 & 0.1113 & 0.1108 \\
0.1113 & 0.1117 & 0.1113 \\
0.1108 & 0.1113 & 0.1108
\end{bmatrix}$ \\ \hline
\end{tabular}
}
\end{table}
\noindent \textbf{Sharp filters:} For generating the sharp versions of the hidden state, we adopt two sharp filters. Their specific kernels are listed in Tab.~\ref{tab:sharp_filters}.
\begin{table}[th]
\centering
\caption{The specific kernel of adopted sharp filters.}
\label{tab:sharp_filters}\resizebox{0.9\linewidth}{!}{
\begin{tabular}{|c|c|c|} \hline
Index & Filter type & Filter weight \\ \hline
\ding{172} & Gaussian filter & $\begin{bmatrix}
0.1096 & 0.1118 & 0.1096 \\
0.1118 & 0.1141 & 0.1118 \\
0.1096 & 0.1118 & 0.1096
\end{bmatrix}$ \\ \hline
\ding{173} & Gaussian filter & $\begin{bmatrix}
0.0369 & 0.0392 & 0.0400 & 0.0392 & 0.0369 \\
0.0392 & 0.0416 & 0.0424 & 0.0416 & 0.0392 \\
0.0400 & 0.0424 & 0.0433 & 0.0424 & 0.0400 \\
0.0392 & 0.0416 & 0.0424 & 0.0416 & 0.0392 \\
0.0400 & 0.0424 & 0.0433 & 0.0424 & 0.0400
\end{bmatrix}$ \\ \hline
\end{tabular}
}
\end{table}
\subsection{Comparison with other methods}
We compare our FastRealVSR method with seven algorithms: DAN~\cite{dan}, RealSR~\cite{real-sr}, BSRGAN~\cite{bsrgan}, Real-ESRGAN~\cite{wang2021real}, RealVSR~\cite{realvsr}, DBVSR~\cite{dbvsr} and RealBasicVSR~\cite{chan2022investigating} on the VideoLQ~\cite{chan2022investigating} dataset. VideoLQ consists of $50$ real-world low-quality videos that are collected from various video hosting sites such as Flickr and YouTube.
To better evaluate the performance of these methods, we adopt another metric MANIQA~\cite{yang2022maniqa} to evaluate the visual quality. MANIQA is the champion of the NTIRE 2022 Perceptual Image Quality Assessment Challenge and its performance outperforms state-of-the-art methods by a large margin. Therefore, this metric is more convincing for comparison of the results of these methods. As mentioned in RealBasicVSR~\cite{chan2022investigating}, to save the time, NIQE and BRISQUE metrics are adopted to evaluate the quality score of the first, middle and last frames of each sequence. This evaluation strategy can not comprehensively evaluate an algorithm. For MANIQA, we adopt two evaluation strategies, the first is to evaluate three frames as RealBasicVSR, while the second is to evaluate all frames of each sequence. Both two evaluation strategies evaluate the quality of frames on the RGB channel.
The quantitative results on VideoLQ are shown in Tab.~\ref{tab:cmp}. For RealBasicVSR, we use the official model and testing strategy to test VideoLQ dataset, the evaluation result of PI is different from the value that represented in their paper. From this table, we can observe that FastRealVSR achieves comparable performance with RealBasicVSR, with $2$ $\times$ faster speed and smaller model size. For the state-of-the-art image quality assessment, MANIQA, we can observe that the performance of FastRealVSR outperforms other methods.
Qualitative results are presented in Fig.~\ref{fig:cmp1}. Compared to previous methods, FastRealVSR can effectively remove annoying artifacts and preserve sharp details.
\begin{table*}[t]
\centering
\caption{Quantitative comparison on VideoLQ dataset. FastRealVSR is capable of achieving a better speed-performance trade-off. Runtime is computed with an output size of $720 \times 1280$, with an NVIDIA V100 GPU. \redbf{Red} and \blueud{blue} indicates the best and second best performance. Following RealBasicVSR, NIQE and BRISQUE are computed on the Y-channel of the first, middle, and last frames of each sequence. For MANIQA, we adopt two evaluation strategies, the first is to evaluate three frames as what RealBasicVSR does, while the second is to evaluate all frames of each sequence. Both two evaluation strategies evaluate the quality of frames on the RGB channel.}
\vspace{-0.3cm}
\label{tab:cmp}\resizebox{\linewidth}{!}{
\begin{tabular}{l|c|c|c|c|c|c|c|c} \hline
& Bicubic & DAN & RealSR & BSRGAN & Real-ESRGAN & RealVSR & RealBasicVSR & \textbf{FastRealVSR} \\ \hline
Params (M) & - & 4.3 & 16.7 & 16.7 & 16.7 & \redbf{2.7} & 6.3 & \blueud{3.8} \\
Runtime (ms) & - & 185 & 149 & 149 & 149 & 1082 & \blueud{63} & \redbf{30} \\ \hline
NIQE~\cite{niqe} $\boldsymbol{\downarrow}$ (\textit{3 frames}) & 8.0049 & 7.1230 & 4.1482 & 4.2460 & 4.2091 & 8.0606 & \blueud{3.7662} & \redbf{3.7658} \\
PI~\cite{pi} $\boldsymbol{\downarrow}$ (\textit{3 frames}) & 7.6017 & 6.8942 & 4.2648 & 4.2644 & 4.2492 & 7.7824 & \blueud{3.9152*} & \redbf{3.8975} \\
BRISQUE~\cite{brisque} $\boldsymbol{\downarrow}$ (\textit{3 frames}) & 54.899 & 51.563 & 30.542 & 30.213 & 32.103 & 54.988 & \redbf{29.030} & \blueud{29.374} \\
MANIQA~\cite{yang2022maniqa} $\boldsymbol{\uparrow}$ (\textit{\textbf{3} frames}) & 0.3019 & 0.2529 & 0.2896 & 0.4265 & 0.4347 & 0.2560 & \blueud{0.4387} & \redbf{0.4483} \\ \hline
MANIQA~\cite{yang2022maniqa} $\boldsymbol{\uparrow}$ (\textit{\textbf{all} frames}) & 0.3011 & 0.2522 & 0.2827 & 0.4233 & 0.4302 & 0.2567 & \blueud{0.4385} & \redbf{0.4637} \\ \hline
\end{tabular}
}
\end{table*}
|
\section{Introduction}
In recent years, conditional image generation has received significant attention in the computer vision community. Some applications that make use of conditional image generation include text-to-image generation (\eg DALLE-2~\cite{ramesh2022hierarchical}) and image restoration (\eg SR3~\cite{saharia2021image}). The most challenging part of these restoration applications comes from the ill-posedness, \ie, the same degraded images may come from multiple different ground truth images.
The ill-posedness affects the performance of traditional methods like sparse coding~\cite{mairal2007sparse, mairal2009non} and makes it difficult for the learning-based algorithms to solve this problem.
Although recent learning-based methods have made impressive progress ~\cite{ledig2017photo}, there remains a significant quality gap between the prediction and natural images.
Recent works that utilize pretrained generative networks have shown the superior visual quality of conditional generation compared to the aforementioned end-to-end learning methods.
Generative models have shown impressive image generation results in terms of sample quality and diversity, indicating their capacity for encapsulating rich photorealistic priors.
Some representative methods include Generative Adversarial Networks (GANs)~\cite{goodfellow2014generative}, Variational Autoencoders (VAEs)~\cite{kingma2013auto}, and Autoregressive models~\cite{larochelle2011neural}.
Their generation process generally starts from the standard normal distribution from which diverse high-fidelity images sampled ~\cite{karras2019style, karras2020analyzing}.
Recent work~\cite{richardson2021encoding} has shown that the \emph{continuity} in the normal distribution remains preserved in the sampled results.
For example, the results produced from two different Gaussian noises with the same model will be close to each other if the two noises are close to each other in Euclidean space.
The continuity allows one to perform conditional image generation in an inversion manner that inverts degraded images into standard noises.
This inverted noise can then generate clear images by projecting the noise with generative models.
Following the protocol, multiple GAN-based generative priors, including optimization-based~\cite{menon2020pulse} and learning-based~\cite{richardson2021encoding} schemes have been proposed for various real-world tasks~\cite{wang2021towards}.
\begin{figure*}[ht!]
\center
\includegraphics[width=\linewidth]{figures/bi-noising.pdf}
\vspace{-1.5\baselineskip}
\caption{The graphical model showing the difference between the previous diffusion sampling process and ours with bi-noising guidance for colorization, where $\mathbf{x}_t$ is the noise of each diffusion process at timestep $t$, $p(\mathbf{x}_0)$ is the predicted noise-free start point of $\mathbf{x}_0$, and arrows indicate the denoising results of the diffusion models at each denoising process.
Top figure shows how the conditional denoising process for colorization gradually accumulates the incorrect noise and results in artifacts. Instead, as shown on the bottom figure, the proposed additional noising and denoising steps diminish the incorrect noise and help in achieving better results.}
\label{fig:teaser}
\vspace{-.5\baselineskip}
\end{figure*}
Denoising diffusion probabilistic models~\cite{sohl2015deep, ho2020denoising} are the most recent deep generative models.
They have shown comparable and even better performance at image synthesis than GANs with delicate guidance~\cite{dhariwal2021diffusion}.
These models learn to sequentially denoise stochastic noise map starting from the normal distribution $\mathcal{N}(0, \mathbf{I})$ to clean images.
However, the generation process is stochastic, and the continuity cannot be preserved from the initial sampled noise.
For instance, two sampled noises from the same normal distribution with a small divergence may generate significantly different clear images.
Such a noncontinuous generation process prevents the generative priors from being applied along with the denoising process like the inversion GANs~\cite{richardson2021encoding}.
Hence, despite their impressive synthesis capacity, diffusion model-based priors have not been explored before.
In this work, we introduce a new method, named bi-noising diffusion, for utilizing rich priors encapsulated in the unconditional pretrained diffusion models.
Inspired by implicit sampling that was first developed in the denoising diffusion implicit models~\cite{song2020denoising} for acceleration, we show that the implicit sampling using an unconditional pretrained diffusion model has a capacity for correcting the divergence of distributions modeled by the conditional diffusion models. Specifically, we make a coarse implicit prediction at each intermediate diffusion time step by sampling from the conditional model.
We then sample the prediction back to the intermediate step with the forward diffusion process. Finally, we make a refined prediction by utilizing an unconditional model.
Fig.~\ref{fig:teaser} visualizes the bi-noising procedure and the error by predicting the noiseless start-point image $p(\mathbf{x}_0)$ of the noise image $\mathbf{x}_t$.
Using this two-step procedure, one can utilize the embedded rich priors learned by the unconditional model and produce better-quality images. This hypothesis is further validated through extensive experiments demonstrating that the introduced method performs favorably against state-of-the-art conditional diffusion models.
\section{Related Work}
\noindent\textbf{Iterative methods.} Finding the corresponding latent code~\cite{creswell2018inverting, abdal2019image2stylegan, abdal2020image2stylegan++, menon2020pulse, gu2020image} or sampled noise~\cite{choi2021ilvr, meng2022sdedit, lugmayr2022repaint} of distorted images for restoration is one of the most straightforward ways of utilizing the generative priors.
The intuition is that the pretrained generative models tend to produce natural results from their initial distribution. Thus the corresponding latent code or sampled noise can be projected to the restored images without additional optimization or learning.
Menon et al.~\cite{menon2020pulse} proposed optimizing the latent code based on the difference between the generative results and the distorted images.
Gu et al.~\cite{gu2020image} proposed to optimize multiple latent codes and compose them together for better visual quality.
Similar iterative methods based on diffusion models have also been explored.
Choi et al.~\cite{choi2021ilvr} proposed to refine the sampled noise at each reverse diffusion step with the residual of distorted images.
However, the applied stochastic iterative process tends to produce significantly different results though slightly changing its input.
Therefore, these methods can only be applied to applications that do not require preserving the image identity.
\noindent\textbf{Learning-based methods.} Employing additional encoders~\cite{zhu2016generative, richardson2021encoding, wang2021towards, chan2021glean} to predict the latent code is another promising way that can bypass the stochastic optimization issues.
However, such a method is incompatible with the diffusion models since it is impossible to encode the distribution of each reverse diffusion process for models that employ many sampling timesteps.
Existing works learn to model conditional generative restoration~\cite{saharia2021image, whang2021deblurring} instead.
Richardson et al.~\cite{richardson2021encoding} proposed to encode images with a ResNet backbone into an extended $\mathcal{W}+$ latent space, which defines upon features of each input layer of the generative networks.
Wang et al.~\cite{wang2021towards} proposed to encode images with a U-Net backbone and modulate the features of each generative layer of the generative networks.
Saharia et al.~\cite{saharia2021image} proposed to learn the noise distribution with the distorted image as the condition.
Whang et al.~\cite{whang2021deblurring} proposed to learn the generative process of residual given restored images.
Compared with these GAN-based learning methods, a large number of sampling timesteps significantly increases the complexity of designing the corresponding encoders and thus makes the priors difficult to be learned.
\noindent\textbf{Classifier guidance.}
Diffusion models have been using class information heavily to perform truncated or low-temperature sampling to increase the sample quality.
The initial attempt~\cite{sohl2015deep, song2020score, dhariwal2021diffusion, nichol2021glide} is to incorporate a pre-trained classifier by using its gradients to guide the diffusion sampling process.
However, it complicates the diffusion model because additional training is required for the classifier on noisy data.
Classifier-free guidance~\cite{ho2022classifier, wang2022semantic} is another approach for addressing the complexity issue.
It alleviates the complexity by combining the existing network with the classifier for guidance, \eg, Ho et al.~\cite{ho2022classifier} use conditional diffusion network with an empty condition, and Wang et al.~\cite{wang2022semantic} use pretrained segmentation with a null label.
Nevertheless, the classifier fails at natural images, and its gradient is meaningless for restoration.
Its strength parameters also become less reasonable for the almost definite restoration sampling process.
In this paper, we are interested in incorporating the sampling quality superiority of the empty condition and the sampling guidance ability of degraded images.
We show that the empty condition can bring the incorrect noisy image back into the high-quality manifold.
Compared with the classifier and classifier-free guidance,
our bi-noising guided diffusion process keeps the same complexity but better fits the restoration task.
\section{Proposed Method}
In this section, we discuss the proposed mechanism to add the embedded priors to diffusion models.
For consistency, we denote the intermediate output of the unconditional diffusion model as $\boldsymbol{\epsilon_\theta}(\cdot)$, parameterized by $\theta$ in the upcoming discussions following Denoising Diffusion Probabilistic Models~\cite{ho2020denoising} (DDPM). The additional, conditional diffusion model is denoted by $\boldsymbol{f_\phi}(\cdot)$, the condition (\ie degraded images) and natural image pairs are denoted by $\{\mathbf{x}_0, \mathbf{y}_0\}$, where the conditional diffusion model $\boldsymbol{f_\phi}(\cdot)$ with parameters $\phi$ denoises noisy image $\mathbf{x}_t$ at timestep $t$ with the concatenated condition $\mathbf{y}_0$.
\subsection{Preliminaries}
Diffusion probabilistic models belong to a new family of generative models~\cite{sohl2015deep, song2019generative, ho2020denoising, nichol2021improved, dhariwal2021diffusion} that can effectively model intractable distributions~\cite{sohl2015deep}. A diffusion process consists of two parts, \ie, the forward process and the reverse diffusion process. In the forward diffusion process, a clean image is sampled from its data distribution and destroyed in $T$ timesteps by repetitive noising using Gaussians of very small variances. Specifically, the forward process can be formulated as
\setlength{\belowdisplayskip}{0pt} \setlength{\belowdisplayshortskip}{0pt}
\setlength{\abovedisplayskip}{0pt} \setlength{\abovedisplayshortskip}{0pt}
\begin{align}
\begin{split}
q(\mathbf{y}_t | \mathbf{y}_{t-1}) &= \mathcal{N}\left(\mathbf{y}_t ; \sqrt{{\beta}_{t}} \mathbf{y}_0,\left(1-{\beta}_{t}\right) \mathbf{I}\right) \\
&= \sqrt{{\beta}_{t}} \mathbf{y}_0 + \epsilon \sqrt{1-{\beta}_{t}}, \epsilon \sim \mathcal{N}(0, \mathbf{I}),
\end{split}
\end{align}
or
\begin{align}
\begin{split}
q(\mathbf{y}_t | \mathbf{y}_{0}) &= \mathcal{N}\left(\mathbf{y}_t ; \sqrt{\bar{\alpha}_{t}} \mathbf{y}_0,\left(1-\bar{\alpha}_{t}\right) \mathbf{I}\right) \\
&= \sqrt{\bar{\alpha}_{t}} \mathbf{y}_0 + \epsilon \sqrt{1-\bar{\alpha}_{t}}, \epsilon \sim \mathcal{N}(0, \mathbf{I}),
\end{split}
\end{align}
where $\alpha_t =1 - \beta_t$ and $\bar \alpha_t := \prod^t_{s=1} \alpha_s$ come from the variance schedule $\{\beta_1, \dots, \beta_T\}$ . The key idea here is that for large values of $T$, repetitive noising using Gaussians of small variances lead to a standard Gaussian, \ie,
\begin{align}
q(\mathbf{y}_T | \mathbf{y}_{0}) &= \mathcal{N}\left(\mathbf{y}_T ; 0, \mathbf{I}\right).
\end{align}
Now at each reverse timestep $t$, we attempt to reconstruct the noisy $y_{t-1}$ from $y_t$ using a distribution $p$ modeled by a neural network with parameters $\theta$. The parameters of the distribution $p_{\theta}(\cdot)$, found by optimizing variational lower bound of log-likelihood of $p_{\theta}(y_0)$, which is simplified by Ho et al.~\cite{ho2020denoising} by claiming that the major component in the objective comes from $L_{t-1}$, and the simplified loss is
\begin{align}
L_{t-1} &= E_{t \sim[1, T], \epsilon \sim \mathcal{N}(0, \mathbf{I})}\left[\left\|\epsilon-\epsilon_{\theta}\left(\mathbf{y}_t, t\right)\right\|^{2}\right].
\end{align}
Here network $\epsilon_{\theta}(\cdot)$ models the noise $\epsilon \sim \mathcal{N}(0, \mathbf{I})$ at each timestep $t$ with the denoised one $\mathbf{y}_t$, which can be seen as the process of learning the gradient of distributions with score matching according to Song et al.~\cite{song2019generative}.
Therefore, we can learn the impressive perceptual synthesizing capacity with the simplified loss function between noises.
\subsection{Learning to Refine Diffusion Process}
In our experiments, we denote the recent diffusion models~\cite{saharia2021image, whang2021deblurring} that learn the diffusion process with conditions as the way of Learning to Refine Diffusion Process (LRDP).
LRDP models the conditional distribution of a clean image given a degraded image for restoration learning, and thus it requires separate training for different tasks or datasets.
The objective for this learning process is formulated as
\begin{align}
L_{\text {vlb}}:=E_{t \sim[1, T], \epsilon \sim \mathcal{N}(0, \mathbf{I})}\left[\left\|\epsilon-\epsilon_{\theta}\left(\mathbf{x}_0, \mathbf{y}_t, t\right)\right\|^{2}\right],
\end{align}
where $\mathbf{y}_t \sim \mathcal{N}(\mathbf{y}_t | \sqrt{\bar \alpha_t} \mathbf{y}_0, (1-\bar \alpha_t)\mathbf{I})$.
The network architecture in LRDP is a slightly changed version from the original U-Net found in DDPM, and the additional input $\mathbf{x}_0$ and $\mathbf{y}_0$ are concatenated and passed to the input layer.
Similarly, the reverse diffusion process of LRDP is slightly changed from the original one and formulated as
\begin{align}
\mathbf{y}_{t-1}=\frac{1}{\sqrt{\alpha_{t}}}\left(\mathbf{y}_{t}-\frac{1-\alpha_{t}}{\sqrt{1-\bar{\alpha}_{t}}} \boldsymbol{\epsilon}_{\theta}\left(\mathbf{x}_0, \mathbf{y}_{t}, t\right)\right)+\sqrt{1-\alpha_t} \mathbf{z},
\end{align}
where $\mathbf{z} \sim \mathcal{N}(0, \mathbf{I})$ and $\alpha_t, \bar \alpha_t$ is the variant of the pre-defined variance schedule $\{\beta_1, \dots, \beta_T\}$, that is $\alpha_t := 1 - \beta_t$ and $\bar \alpha_t := \prod^t_{s=1} \alpha_s$.
Since the diffusion process conditions on the specific type of degradation $p(\cdot)$ that produces degraded image $\mathbf{x}_0$ given clear image $\mathbf{y}_0$ as $p(\mathbf{x}_0|\mathbf{y}_0)$, LRDP needs re-training from scratch for different restoration tasks, which further heightens the training cost.
We experimentally find that such a protocol degrades the visual quality of the generation compared with the one without $\mathbf{x}_0$ conditioned.
The most straightforward assumption towards the performance drop is that the assumed posterior $p(\mathbf{x}_0|\mathbf{y}_0)$ contrasts with the diffusion process $p_\theta(\mathbf{y}_0 | \mathbf{x}_0) \propto p_\theta(\mathbf{y}_0) p_\theta(\mathbf{x}_0|\mathbf{y}_0)$ due to the ambiguous property of the degradation models.
Thus, we claim that decomposing the diffusion generation process into different protocols should be a more promising way to handle restoration tasks.
\subsection{Conditioning on Diffusion Process}
The recent work~\cite{choi2021ilvr, lugmayr2022repaint} falls into another category of utilizing the diffusion process, which uses a pretrained DDPM and changes its reverse diffusion process with distorted images by Conditioning on Diffusion Process (CDP).
A similar way was previously explored in the other generative models, \eg, mGANprior~\cite{gu2020image} and PULSE~\cite{menon2020pulse} invert a trained GAN by optimizing its latent code.
However, CDP does not require optimization compared with the previously mentioned GAN-based methods.
In contrast, it ensembles the conditions during sampling as
\begin{align}
\hat \mathbf{y}_{t-1} &=\frac{1}{\sqrt{\alpha_{t}}}\left(\mathbf{y}_{t}-\frac{1-\alpha_{t}}{\sqrt{1-\bar{\alpha}_{t}}} \boldsymbol{\epsilon}_{\theta}\left(\mathbf{y}_{t}, t\right)\right)+\sqrt{1-\alpha_t} \mathbf{z} \\
\mathbf{y}_{t-1} &= \hat \mathbf{y}_{t-1} + \boldsymbol{\sigma}(x_0, \hat \mathbf{y}_{t-1}),
\end{align}
where $\boldsymbol{\sigma}(\cdot)$ is a handcrafted transformation which aims at combining $\mathbf{x}_0$ with $\hat \mathbf{y}_{t-1}$ for accurate restoration.
For example, Choi et al.~\cite{choi2021ilvr} proposed to downsample $\mathbf{x}_0$ and $\hat \mathbf{y}_{t-1}$ and take their residual as the conditioning, while Lugmayr et al.~\cite{lugmayr2022repaint} proposed to sum the visible region of $\mathbf{x}_0$ with the invisible region $\hat \mathbf{y}_{t-1}$ for the inpainting task.
Though CDP avoids the heavy training cost and is suitable for some conditional generation tasks like restoration with minimal modifications, its performance highly depends on the amount of degradation in the conditioned images.
For example, when the conditioned images suffer from high amounts of distortion for face image restoration, CDP cannot preserve the face identity and tends to generate pseudo-sharp results with fake details.
These fake details introduce further ill-posedness to the restored images and greatly limit the applications of such methods.
Therefore, we propose refining the denoised results for correcting such artifacts at each step.
\subsection{Implicit Error-feedback Diffusion Priors}
\label{section:forward}
Since the diffusion models follow a time-sequential process, the error in each step and the visual artifacts propagate and add up, hence severely degrading the quality of some CDP results. However, such issues are rarely observed in the unconditional diffusion models.
We argue that the difference comes from conditioning breaking the inherent probabilistic distribution of noises at each sampling timestep, causing them to deviate from the manifold of natural images. Therefore, we propose to apply generative priors embedded in a pretrained unconditional model to regularize the noise predicted at each timestep from the conditional model.
The trained diffusion model with conditioning denoted as $\boldsymbol{f_\phi}(\cdot)$ takes as input the predicted image of the previous timestep and makes an implicit prediction $\tilde{\mathbf{y}}_0$ defined by
\begin{align}
\tilde\mathbf{y}_0 = (\mathbf{y}_t - \sqrt{1 - \bar \alpha_t} \boldsymbol{f_\phi}(\mathbf{x}_0, \mathbf{y}_t, t)) / \sqrt{\bar \alpha_t}.
\end{align}
Here $y_t$ denotes the prediction at the previous timestep. We then estimate the noisy version of the implicit prediction, which undergoes further regularization from an unconditional diffusion model. Please note that the unconditional diffusion model that fits the inherent probabilistic distribution.
The diffusion process $\mathbf{y}_t \sim q(\mathbf{y}_t | \tilde{\mathbf{y}}_0)$ with $\boldsymbol{\epsilon_\theta}(\cdot)$ is formulated as
\begin{align}
\begin{split}
q(\mathbf{y}_t | \tilde{\mathbf{y}}_0) &:= \mathcal{N}(\mathbf{y}_t | \sqrt{\bar \alpha_t} \tilde \mathbf{y}_0, (1-\bar \alpha_t) \mathbf{I})\\
&= \sqrt{\bar{\alpha}_t} \tilde{\mathbf{y}}_0 + \epsilon \sqrt{1-\bar{\alpha}_{t}}, \epsilon \sim \mathcal{N}(0, \mathbf{I}),
\end{split}
\end{align}
and
\begin{align}
\mathbf{y}_{t-1} = \frac{1}{\sqrt{\alpha_t}}\left(\mathbf{y}_t-\frac{1-\alpha_t}{\sqrt{1-\bar{\alpha}_t}} \boldsymbol{\epsilon_\theta}\left(\mathbf{y}_{t}, t\right)\right)+\sigma_{t} \mathbf{z}.
\label{eq:predict}
\end{align}
Following this procedure brings in an inherent regularization to the output of the conditional model during the reverse diffusion process.
Note that Equation~\eqref{eq:predict} takes the noised version $\mathbf{y}_t$ sampled from $\tilde \mathbf{y}_0$ as input. It is similar to the original reverse diffusion process, which takes the noised version of natural images as input.
In summary, we utilize two diffusion models for conditional image generation. The unconditional diffusion model regularizes the predicted outliers at each prediction timestep of the conditional diffusion model in an error-feedback way.
Moreover, for the complex real-world application like draining where domain gaps may exist, we further discuss the details of applying our bi-noising diffusion with slight modifications to achieve better performance.
\subsection{Complex Application: Deraining}
Here we discuss one of the applications where we apply our introduced bi-noising diffusion for further clarification.
The diffusion model is trained for the task of deraining using the rainy image as a condition for generating rain-free results. Motivated by the power of diffusion models to learn the distribution of clean natural images, we train a diffusion model to learn the distribution of rain and, at the same time, make the model aware of the distribution of real-world rain-free images. For this, we first train a diffusion model for image generation using the ImageNet dataset, and we then train another diffusion model by ensuring that the weights of the model trained for deraining are aligned to the weights learned for real-world rain-free images.
Let $W_r$ denote the weights of the first model and $\theta$ denote the weights of the diffusion decoder estimated after each iteration through backpropagation. Then the weights of the second model are updated after each iteration of training according to,
\begin{equation}
W' = \alpha.\theta + (1-\alpha).W_r,
\end{equation}
where $\alpha$ denotes the rate of Exponential Moving Average (EMA) for updating the decoder weights. The encoder weights are updated as such.
\input{content/colorization.tex}
One observation from our experiments on image deraining while training by direct conditioning like in SR3 \cite{saharia2021image} was that the restored images suffered from artifacts and color channel shift which can be seen in Fig.~\ref{fig:rain800}. On further investigation, we found that this is due to incorrect conditioning of input during the training process. Specifically, for the task of image restoration with source-target pairs denoted as $(\mathbf{x}_0,\mathbf{y}_0)$, existing methods optimize the weights of the network $\boldsymbol{\epsilon_\theta}(\cdot)$ modelling the reverse process of diffusion, by minimizing the $L_{simple}$ function defined in \cite{ho2020denoising} as
\begin{align}
L_{\text {simple}}:=E_{t \sim[1, T], \epsilon \sim \mathcal{N}(0, \mathbf{I})}\left[\left\|\epsilon-\epsilon_{\theta}\left(\mathbf{x}_0, \mathbf{y}_t, t\right)\right\|^{2}\right].
\end{align}
The training objective $L_{simple}$ holds the inherent assumption that during inference time $\mathbf{y}_t$, (i.e. the reconstructed image at time $t$ ) is close to the clean target. But for extreme cases where the intermediate diffusion outputs are not accurate during the initial steps of diffusion, the rain streaks continue to propagate through the diffusion process, as can be seen in Fig. \ref{fig:rain800}. This is because, inherently, the diffusion model works by predicting the noise present in $\mathbf{y}_t$ than the amount of degradation in it. To account for this, we add a correction prior $L_{corr}$ so that the network can give equally good output for high distortion levels. This term is defined by,
\begin{align}
L_{\text {corr}}:=\alpha_t \left\|\epsilon_{\theta}\left(\mathbf{x}_t, \mathbf{x}_0, t\right)-\epsilon_{\theta}\left(\mathbf{y}_t, \mathbf{x}_0, t\right)\right\|^{2}.
\end{align}
The final objective for training the network is,
\begin{align}
L_{\text {final}}=L_{\text {simple}}+\lambda_{corr}L_{\text {corr}}.
\end{align}
The value of $\lambda_{corr}$ is empirically set equal to $0.001$ for all experiments.
\section{Experiments}
To demonstrate the restoration capacity of our method, we evaluate our method with several experimental settings following the most representative diffusion models, \ie, ILVR~\cite{choi2021ilvr} and SR3~\cite{saharia2021image} based on the Guided-diffusion architecture~\cite{nichol2021glide}.
Following the common practice that pixel-wise metrics, \ie, PSNR and SSIM cannot comprehensively denote the visual quality of restored results, we utilize FID
and LPIPS as the additional metrics for evaluation. The tasks in which we evaluate our method on are
\begin{itemize}[noitemsep]
\item Conditional image restoration which is trained on the FFHQ~\cite{russakovsky2015imagenet} dataset (70000 images) and evaluated on the CelebA-HQ~\cite{liu2015deep,karras2017progressive} dataset (first 3000 images) with a resolution of $256\times 256$ pixels.
\item Conditional image restoration which is $4\times$ face super-resolution trained on the FFHQ~\cite{russakovsky2015imagenet} dataset and evaluated on the CelebA-HQ~\cite{liu2015deep,karras2017progressive} dataset (first 3000 images) with a resolution of $256\times 256$ pixels.
\item Image turbulence removal follows the turbulence simulation settings~\cite{nair2022ddpm} on the FFHQ dataset and conducts evaluation on the real long-range imaging images~\cite{miller2019data}.
\item Image deraining which is conducted on the Rain800~\cite{zhang2019image} dataset and Jorder 200L~\cite{yang2017deep} dataset with their respective train sets. The diffusion models conduct in a resolution of $256\times 256$ pixels.
\end{itemize}
Note that for the first three tasks, the diffusion models are trained on the FFHQ dataset for face generation.
For the last task, the diffusion models are trained on the ImageNet dataset for natural image generation. The unconditional model utilized has never seen the validation dataset during its training process for all of these cases.
\subsection{Colorization}
Colorization aims at reconstructing grayscale images with colors that are fitted to natural statistics and image semantics.
The grayscale image is obtained by averaging the values at red, green, and blue channels of the corresponding colour image.
We empirically observed that conditional denoising diffusion models fail at colorization. Even though they can preserve the fine-grained details, unnatural colors always exist in their reconstructed results.
In contrast, the method that adopts our proposed bi-noising diffusion is capable of correcting the reconstruction with more semantics and accurate color descriptions.
The quantitative performance comparison is shown in Tab.~\ref{tab:colorizarion_performance}, where our method achieves 7.906dB higher PSNR than the one without pretraining.
The visual results in Fig.~\ref{fig:colorization} further clarify the improvements that come from more globally consistent colors and tones of our results, even though the pretraining had never seen the ground truth before.
In contrast, a similar method, i.e., ILVR cannot deal with the colorization task even though it also utilizes a pretrained unconditional model, which demonstrates the superiority of our proposed DDRP in such tasks. Therefore, we argue that utilizing the priors plays a crucial role in ensuring the color naturalism.
\begin{table}[ht!]
\centering
\resizebox{1.0\linewidth}{!}{
\begin{tabular}{lcccc}
\toprule
~Method &~PSNR $\uparrow$~&~SSIM $\uparrow$~&~FID $\downarrow$~&~LPIPS $\downarrow$~\\
\midrule
ILVR Diffusion~\cite{choi2021ilvr} & 18.3936 & 0.5674 & 86.2642 & 0.5008 \\
SR3 Diffusion~\cite{saharia2021image} & \emph{\underline{19.1647}} & \emph{\underline{0.8680}} & \emph{\underline{13.8126}} & \emph{\underline{0.2959}} \\
Bi-Noising (Ours) & \textbf{27.0707} & \textbf{0.9531} & \textbf{12.6796} & \textbf{0.1417} \\
\bottomrule
\end{tabular}}
\vspace{-.1in}
\caption{\textbf{Colorization results corresponding to the CelebAHQ dataset.} The best and second-best performace is indicated with \textbf{blod} and \emph{\underline{italic}} respectively. We use $\uparrow$ and $\downarrow$ to suggest high/lower score should be achieved by better methods.}
\label{tab:colorizarion_performance}
\vspace{-.1in}
\end{table}
\subsection{Face Super-resolution}
Face super-resolution is the other representative task in image restoration, and it is widely evaluated in the other denoising diffusion-based restoration works.
We follow the experimental settings of SR3 and ILVR, i.e., restore $256\times 256$ face images from $64\times 64$ face images downsampled by Bicubic interpolation.
The implementation details of PULSE~\cite{menon2020pulse}, ILVR~\cite{choi2021ilvr}, SR3~\cite{saharia2021image} are presented in the supplementary file.
From Fig.~\ref{fig:facesr}, one can notice that our method achieves the best visual quality compared with the other methods.
Compared with the state-of-the-art face super-resolution method based on GAN priors, our method better preserves the identity of the restored face images.
As can be seen from Tab.~\ref{tab:sr-performance}, our method significantly outperforms the other methods in terms of the distortion measures, i.e., PSNR and SSIM with 4.8316 dB and 0.04 better than the second one.
Though our results in the FID metric are not better than ILVR, FID doesn't denote the reconstruction accuracy that is curcial for super-resolution.
Therefore, the above results demonstrate our performance superiority.
\input{content/face-sr.tex}
\begin{table}[htbp]
\centering
\resizebox{1.0\linewidth}{!}{
\begin{tabular}{lcccc}
\toprule
Method &~PSNR $\uparrow$~&~SSIM $\uparrow$~&~FID $\downarrow$~&~LPIPS $\downarrow$~\\
\midrule
PULSE~\cite{menon2020pulse} & \emph{\underline{23.5769}} & \emph{\underline{0.6794}} & 31.2309 & 0.3832 \\
ILVR Diffusion ~\cite{choi2021ilvr} & 22.5374 & 0.6150 & \textbf{20.4621} & 0.3393 \\
SR3 Diffusion~\cite{saharia2021image} & 22.8290 & 0.6442 & 29.8932 & \emph{\underline{0.3350}} \\
Bi-Noising (Ours) & \textbf{29.3996} & \textbf{0.8414} & \emph{\underline{24.5632}} & \textbf{0.1809} \\
\bottomrule
\end{tabular}}
\vspace{-.5\baselineskip}
\caption{\textbf{$4\times$ super-resolution results corresponding to the CelebAHQ dataset.}}
\label{tab:sr-performance}
\end{table}
\input{content/rain800.tex}
\input{content/jorder100.tex}
\subsection{Image Deraining}
We perform single image deraining on two popular draining datasets. Namely, the Jorder 200L dataset which contains large rain streaks, and the Rain800 dataset which contains realistic rain. Since no diffusion-based deraining method has been proposed in literature before, we perform comparisons after retraining the models proposed for super-resolution in the literature. Specifically, we perform comparisons with ILVR diffusion~\cite{choi2021ilvr} and conditional diffusion models, and we include the improvements brought about by our modules. To evaluate the reconstruction quality, we use the PSNR and SSIM metrics. To assess the quality of images produced by various methods, we use LPIPS and NIQE as metrics.
As we can see from Tab.~\ref{tab:j200}, the proposed conditioning loss functions bring significant improvement for all metrics in the JORDER 200L dataset \cite{yang2017deep}, obtaining about $2.45$ dB PSNR over the exiting method as well as giving realistic natural images.
The visual comparisons in Fig.~\ref{fig:rain800} and Fig.~\ref{fig:jorder200} further demonstrate our method on the visual quality compared with the other methods.
\begin{table}[htbp]
\centering
\resizebox{1.0\linewidth}{!}{
\begin{tabular}{lcccc}
\toprule
& \multicolumn{4}{c}{Jorder 200L dataset} \\
\cmidrule(r){2-5}
Method & PSNR $\uparrow$ &SSIM $\uparrow$ & LPIPS $\downarrow$ & NIQE $\downarrow$ \\
\midrule
Rain Images (Input) & $26.70$ &$0.8439$&$0.2411$ &$4.131$ \\
ILVR Diffusion~\cite{choi2021ilvr} & $21.22$ & $0.4942$ & \textbf{0.0972} & $6.467$ \\
SR3 Diffusion~\cite{saharia2021image} & \emph{\underline{31.45}} & \emph{\underline{0.9091}} & \emph{\underline{0.1779}} & \emph{\underline{3.588}} \\
Bi-Noising (Ours) & \textbf{33.90} & \textbf{0.9555} & \textbf{0.0972} & \textbf{3.232} \\
\bottomrule
\end{tabular}}
\vspace{-.5\baselineskip}
\caption{\textbf{Restoration results comparison on the Jorder 200L dataset with the other re-trained diffusion models.}}
\label{tab:j200}
\vspace{-.5\baselineskip}
\end{table}
\begin{figure*}[ht!]
\centering
\begin{minipage}[c]{.73\linewidth}
\resizebox{\linewidth}{!}{
\begin{tabular}{rccccccc}
\toprule
& \multicolumn{4}{c}{Methods} & \multicolumn{3}{c}{Ours} \\
\cmidrule(r){2-5} \cmidrule(r){6-8} ~Settings & Ho et al.~\cite{ho2020denoising} & Dhariwal et al.~\cite{dhariwal2021diffusion} & Nichol et al.~\cite{nichol2021glide} & Ho et al.~\cite{ho2022classifier} & w/o parametric & w/o full guidance & Bi-Noising \\
\midrule
classifier guidance~\cite{dhariwal2021diffusion} & & \checkmark \\
CLIP guidance~\cite{nichol2021glide} & & & \checkmark \\
classifier-free guidance~\cite{ho2022classifier} & & & & \checkmark & \checkmark \\
alternative guidance & & & & & & \checkmark \\
Bi-Noising & & & & & \checkmark & \checkmark & \checkmark \\
\midrule
PSNR $\uparrow$ & 19.16 & 20.10 & 23.14 & 25.91 & 26.46 & \emph{\underline{26.81}} & \textbf{27.07} \\
Parameters (M) $\downarrow$ & \textbf{93.6} & 147.7 & 243.2 & \textbf{93.6} & \textbf{93.6} & 187.2 & 187.2 \\
Running Time (s) $\downarrow$ & \textbf{1.6} & 4.9 & 3.4 & 3.1 & 3.1 & \emph{\underline{2.3}} & 3.1 \\
\bottomrule
\end{tabular}
}
\vspace{-.5\baselineskip}
\captionof{table}{\textbf{Result comparisons between different prior parameterizations.}}
\label{tab:param}
\end{minipage}
\begin{minipage}[c]{.26\linewidth}
\includegraphics[width=\linewidth]{content/efficiency.pdf}
\vspace{-1.8\baselineskip}
\captionof{figure}{\textbf{FID v.s. Time.}}
\label{fig:fid}
\end{minipage}
\vspace{-1\baselineskip}
\end{figure*}
\subsection{Turbulence Removal}
We plug the proposed bi-noising approach into the recent diffusion restoration work~\cite{nair2022ddpm} to demonstrate the applicability of our method on an extremely ill-posed atmospheric turbulence mitigation problem.
Compared with the diffusion network with single noise conditioning, the results shown in Fig.~\ref{fig:turb} validate that our bi-noising method is able to remove the unnatural textures from the face images resulting from the incorrect denoising results.
\begin{figure}[htbp]
\centering
\resizebox{\linewidth}{!}{
\setlength{\tabcolsep}{1pt}
{\small
\renewcommand{\arraystretch}{0.5}
\begin{tabular}{c c c c c}
\captionsetup{type=figure, font=scriptsize}
\raisebox{0.3in}{\rotatebox[origin=t]{90}{\scriptsize \emph{Input}}}&
\includegraphics[width=0.22\linewidth]{results/turbulence/043/input.png} &
\includegraphics[width=0.22\linewidth]{results/turbulence/064/input.png} &
\includegraphics[width=0.22\linewidth]{results/turbulence/088/input.png} &
\includegraphics[width=0.22\linewidth]{results/turbulence/070/input.png}
\tabularnewline
\raisebox{0.3in}{\rotatebox[origin=t]{90}{\scriptsize \emph{AT-DDPM}}}&
\includegraphics[width=0.22\linewidth]{results/turbulence/043/nithin.png} &
\includegraphics[width=0.22\linewidth]{results/turbulence/064/nithin.png} &
\includegraphics[width=0.22\linewidth]{results/turbulence/088/nithin.png} &
\includegraphics[width=0.22\linewidth]{results/turbulence/070/nithin.png}
\tabularnewline
\raisebox{0.3in}{\rotatebox[origin=t]{90}{\scriptsize \emph{Bi-noising (Ours)}}}&
\includegraphics[width=0.22\linewidth]{results/turbulence/043/ours.png} &
\includegraphics[width=0.22\linewidth]{results/turbulence/064/ours.png} &
\includegraphics[width=0.22\linewidth]{results/turbulence/088/ours.png} &
\includegraphics[width=0.22\linewidth]{results/turbulence/070/ours.png}
\tabularnewline
\raisebox{0.3in}{\rotatebox[origin=t]{90}{\scriptsize \emph{Ground Truth}}}&
\includegraphics[width=0.22\linewidth]{results/turbulence/043/gt.png} &
\includegraphics[width=0.22\linewidth]{results/turbulence/064/gt.png} &
\includegraphics[width=0.22\linewidth]{results/turbulence/088/gt.png} &
\includegraphics[width=0.22\linewidth]{results/turbulence/070/gt.png}
\end{tabular}
}
}
\vspace{-.5\baselineskip}
\caption{\textbf{Atmospheric turbulence mitigation results corresponding to the LRFID dataset \cite{nair2022ddpm}.}}
\label{fig:turb}
\vspace{-1.5\baselineskip}
\end{figure}
\subsection{Design Analysis}
\noindent\textbf{Nonparametric v.s. Parametric Priors.}
Inspired by Ho et al.~\cite{ho2022classifier}, here we analysis the effect of alleviating complexity by parameterizing unconditional models into the conditional restoration model, denoted as \emph{Nonparametric Prior} in Tab.~\ref{tab:param}.
The compared methods use the same diffusion model but different guidance settings as the prior for fair comparisons. We showcase their performance and efficiency difference in the colorization task that was conducted using a single NVIDIA A6000 GPU.
Specifically, we use a single diffusion model that takes conditions for restoration, and it takes a null token $\varnothing$ for unconditional generation.
From Tab.~\ref{tab:param}, we can conclude that the nonparametric prior, \ie, w/o parametric, significantly reduces half of parameters for diffusion sampling, while the model suffers 0.26 dB performance drop compared with the parametric prior, \ie, Bi-Noising, that is our final setting.
The reason is that the null token increases the diffusion model training difficulty and thus the model fits worse than the unconditional model used in our final setting.
Compared with other concurrent works that utilize classifier guidance~\cite{dhariwal2021diffusion} and clip guidance~\cite{nichol2021glide}, our method outperforms them significantly with a slight increase in the number of parameters and running time. This clearly demonstrates the benefits of our parametric prior that can encapsulate the low-level information distribution for restoration.
Fig.~\ref{fig:fid} further demonstrates the efficiency of our method that achieves better FID score after 2.0s denoising process.
\noindent\textbf{Priors Correlation.}
For the complex applications like deraining, we introduce additional correlation priors to further boost the final results.
In Tab.~\ref{tab:abj200}, we present the ablation study of the introduced correlation priors to demonstrate its effectiveness.
Since the rain streaks in the JORDER 200L dataset are relatively small, we choose the Rain800 dataset for a fair experiment.
The ablation starts with the base model and then adds the two priors one by one to show the improvements.
From the improved results due to $\mathrm{L}_{\mathrm{corr}}$, we can conclude that the introduced correlation priors allow our diffusion priors to better fit the probabilistic distribution of complex images, which ultimately benefits conditional generation with more realistic results.
The comparisons presented in Fig.~\ref{fig:rain800_supp} also visually validate the conclusion.
\begin{table}[]
\centering
\resizebox{1.0\linewidth}{!}{
\begin{tabular}{lcccc}
\toprule
Method & PSNR $\uparrow$ &SSIM $\uparrow$ & LPIPS $\downarrow$ & NIQE $\downarrow$ \\
\midrule
Rain Images (Input) & $26.70$ &$0.8439$&$0.2411$ &$4.131$ \\
SR3 Diffusion~\cite{saharia2021image} &$31.45$ & $0.9091$ & $0.1779$&$3.588$ \\
Ours & \emph{\underline{33.23}} & \emph{\underline{0.9505}} & \emph{\underline{0.1043}} & \emph{\underline{3.285}} \\
Ours + $\mathrm{L}_{\mathrm{corr}}$ & \textbf{33.90} & \textbf{0.9555} & \textbf{0.0972} & \textbf{3.232} \\
\bottomrule
\end{tabular}}
\vspace{-.1in}
\caption{\textbf{Ablation study on the improvements brought by each introduced component.}}
\label{tab:abj200}
\vspace{-.5\baselineskip}
\end{table}
\input{content/jorder_supp}
\section{Conclusion}
We explored ways in which one can utilize denoising diffusion probability model priors for improving image enhancement and restoration tasks.
The proposed way of integrating the stochastic priors into the deterministic conditioning denoising diffusion restoration model showed its superiority in colorization, face super-resolution, natural image super-resolution, and deraining tasks.
Compared with similar denoising diffusion-based restoration methods, the restored results of our introduced method achieved better color consistency and contain more fine-grained details.
\section{Acknowledgement}
This research is based upon work supported in part by the Office of the Director of National Intelligence (ODNI), Intelligence Advanced Research Projects Activity (IARPA), via [2022-21102100005]. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies, either expressed or implied, of ODNI, IARPA, or the U.S. Government. The US. Government is authorized to reproduce and distribute reprints for governmental purposes notwithstanding any copyright annotation therein.
{\small
\bibliographystyle{ieee_fullname}
\section{Demo}
In order to provide a straightforward overview of our method, we have provided an online colorization demo and compared our bi-denoising process with the \emph{naive-diffusion} by visualizing their intermediate predicted $\mathbf{x}_0$.
The online demo is accessible at \href{http://bi-noising.demohub.cc}{http://bi-noising.demohub.cc}.
\begin{figure}[htbp]
\centering
\includegraphics[width=.72\linewidth]{figures/screenshot.png}
\caption{Screenshot of our online demo.}
\end{figure}
\begin{figure}[htbp]
\centering
\includegraphics[width=.72\linewidth]{figures/result_screenshot.png}
\caption{Screenshot of our online demo that shows artifacts in naive diffusion results, highlighted with red arrow.}
\end{figure}
\section{Manifold Correction using Diffusion Priors}
An alternate advantage of the proposed bi-noising diffusion is the effect of manifold correction. Consider any general restoration model$f(.))$ which maps from a domain $\mathcal{B}$ of the degraded image to a domain $\mathcal{A}$ of all natural images. The desired mapping function of this model for any restoration task is to learn the mapping
\begin{align}
b \in \mathcal{B}\xRightarrow[]{h(.)} a \in\mathcal{A}
\end{align}
\noindent But in real scenarios, because the restoration task is ill-posed, rather than learning the natural image manifold, a deep network learns an inverse function that merely removes the degradation effect. Let this manifold be denoted by $\mathcal{C}$. The mapping function hence learned is
\begin{align}
b \in \mathcal{B}\xRightarrow[]{f(.)} a \in\mathcal{C}
\end{align}
For example, for a restoration model trained for the task of face super-resolution, for an input (b), the network could create an output (c) that is the image of a disoriented face rather than an image in the manifold of faces. Theoretically, if we utilize any generative model for the restoration, the model should be able to achieve the ideal mapping. But often, the model learns the more complex problem of removing the degradation than learning to map to the domain of natural images. This is because it is difficult to reach the solution corresponding to the global optimum. In any generic restoration method, this deviation from the natural manifold can be corrected by adding a correction network that learns the mapping from domain $\mathcal{C}$ to $\mathcal{A}$. Unlike all other models, diffusion models contain a flexible model structure where intermediate latent variables can be accessed. This enables a manifold correction during inference time alone with explictly training a network to map from the generated manifold $\mathcal{C}$ to the natural manifold $\mathcal{A}$. Hence in our work, we exploit this property and perform the manifold correction to the domain of natural images through an additional step that utilizing an unconditional model. Consider a CDP trained for any restoration task denoted by $f_{\phi}(c_t,b,t)$. During inference, the restored sample $c_T$ is generated through the cascade of steps
\begin{align}
f_{\phi}(c_0,b,t)\rightarrow f_{\phi}(c_1,b,t),.... \rightarrow f_{\phi}(c_T,b,t)
\end{align}
or equivalently,
\begin{align}
c_0 \xrightarrow[]{f_{\phi}(.)} c_1 ,... \xrightarrow[]{f_{\phi}(.)} c_T
\end{align}
Here, ${c_1,...,c_T}$ denotes the intermediate diffusion outputs of an image $c_T$in the manifold $\mathcal{C}$ that can be reconstructed from a degraded b. As mentioned before, the function $f$ will not always map to the domain $\mathcal{A}$ of natural images. Hence we add an unconditional model $g_{\theta}(.)$ that does the task of aligning the manifold of the generated image to the manifold of natural images. The sequence of operations is as follows
\begin{align}
f_{\phi}(c_0,b,t)\rightarrow g_{\theta}(c_1,t),.... f_{\phi}(a_{T-1},b,t)\rightarrow g_{\theta}(c_T,t) \\
c_0 \xrightarrow[]{f_{\phi}(.)} c_1\xrightarrow[]{g_{\theta}(.)} a_1\xrightarrow[]{f_{\phi}(.)}c_2,... \xrightarrow[]{f_{\phi}(.)} c_T \xrightarrow[]{g_{\theta}(.)} a_T
\end{align}
${a_1,...,a_T}$ denotes the intermediate diffusion outputs of an image $a_T$ in manifold $\mathcal{A}$.
\section{Turbulence Removal}
Here we provide the quantitative evaluation of our method on the turbulence removal benchmark LRFID dataset.
Compared with the other methods, ours not only has achieved better performance in the sample quality in terms of LPIPS but also better fidelity in terms of face recognition accuracy Top-1 and Top-3.
\input{content/lrfid_metrics}
|
\section{Introduction}
The interaction of composite particles with electromagnetic, QCD energy-momentum and weak currents are parameterized in terms of the corresponding form factors (FFs) as the building blocks. Having determined these FFs, one can construct different observables in terms of these non-perturbative objects to determine the nature and internal quark-gluon organizations of hadrons. In the case of the energy-momentum tensor (EMT) or gravitational form factors (GFFs), which represent the gravitational interaction between a graviton and the hadron, these FFs provide information on the mechanical properties of the hadron such as the mass, spin, pressure, shear force, radius, etc. Understanding the mechanical structure of hadrons is of great importance as it provides us with the fundamental information on the internal structure and geometric shapes of the hadrons.
The GFFs for the spin-1/2 particles entering the matrix elements of the
EMT current were first parameterized in Refs. \cite{Kobzarev:1962wt,Pagels:1966zza,Kobsarev:1970qm,Ng:1993vh}. Based on these parametrizations, the GFFs of the nucleon as a basic baryonic structure have been intensively
investigated in various approaches~\cite{Polyakov:2002wz,Ji:1997gm,Schweitzer:2002nm,Jung:2013bya,Hagler:2003jd,Gockeler:2003jfa,Pasquini:2007xz,Hwang:2007tb,Abidin:2008hn,Brodsky:2008pf,Pasquini:2014vua,Chakrabarti:2015lba,Lorce:2018egm,Teryaev:2016edw,Shanahan:2018nnv, Shanahan:2018pib,Neubelt:2019sou,Anikin:2019kwi,Alharazin:2020yjv,Gegelia:2021wnj,Varma:2020crx,Fujita:2022jus,Mamo:2022eui,Pefkou:2021fni,Azizi:2019ytx,Polyakov:2018exb,Freese:2021mzg,Freese:2021qtb,Freese:2021czn, Burkert:2021ith,Burkert:2018bqq}. Compared to the nucleon, the GFFs of other members of the baryon octet were much less studied~\cite{Ozdem:2020ieh,Won:2022cyy}.
The calculations of EMT FFs have recently been extended to the parity flip transitions, $ N^*\rightarrow N $~\cite{Polyakov:2020rzq, Azizi:2020jog, Ozdem:2019pkg}. They were also generalized to the higher-spin particles~\cite{Cotogno:2019vjb} in a systematic way. For a spin-1 particle, the
the corresponding GFFs were studied in Refs.~\cite{Polyakov:2019lbq,Cosyn:2019aio,Kim:2022wkc,Freese:2022yur,Freese:2022ibw,Freese:2019bhb,Sun:2020wfo,Epelbaum:2021ahi,Pefkou:2021fni} using different models and approaches.
Very recently, the $N \to \Delta$ transition matrix elements of the EMT current were also parametrized in Ref. \cite{Kim:2022bwn}.
The GFFs for the spin-3/2 particle were also examined in Refs. \cite{Pefkou:2021fni,Fu:2022rkn,Alharazin:2022wjj,Panteleeva:2020ejw}, recently.
In our study on the GFFs of the nucleon \cite{Azizi:2019ytx}, which were calculated in a large range of transferred momentum square $ Q^2 $, we compared our predictions on the GFFs $M_2^q(Q^2)$, $J^q(Q^2)$ and $d_1^q(Q^2)$ (D-term) at small values of $ Q^2 $ with the existing Lattice QCD predictions \cite{Hagler:2007xi} as well as the JLab data for the D-term \cite{Burkert:2018bqq}. For the form factor $M_2^q(Q^2)$, our predictions were consistent with most of the Lattice QCD data points considering the errorbars. In the case of $J^q(Q^2)$ and $d_1^q(Q^2)$ form factors, the Lattice results suffer from large uncertainties at small values of $ Q^2 $ that should be computed more accurately. It was also obtained that our predictions reproduce most of the JLab data at small values of $ Q^2 $.
The GFFs of nucleon can also be calculated from its generalized parton distributions (GPDs) at zero skewness \cite {Polyakov:2002yz}. Very recently, the GFFs $M_2(Q^2)$ and $J(Q^2)$ were calculated at zero skewness limit by integration of the GPDs $H^q(x,Q^2) $ and $E^q(x,Q^2) $ over Bjorken-$ x $ in Refs. \cite{ Hashamipour:2022noy,Hashamipour:2021kes} and compared with our previous light cone QCD sum rules predictions presented in Ref. \cite{Azizi:2019ytx}. Very surprisingly, for special sets, the obtained results via the GPDs that themselves are extracted from the pure experimental data of different collaborations are in a nice consistency with the predictions of our previous study, Ref. \cite{Azizi:2019ytx}, for the GFFs $M_2(Q^2)$ and $J(Q^2)$ at a vide range of $ Q^2 $. Our predictions on the GFFs of the negative parity excited nucleon $N(1535) $ or $N^* $ and the transition GFFs of the $ N^*\rightarrow N $ \cite{ Azizi:2020jog, Ozdem:2019pkg} can be used to be compared with the future Lattice QCD results as well as experimental data.
In the present article, we study the transition GFFs of the $N \to \Delta$ for the first time. We use the light cone QCD sum rule formalism to find the five independent conserved and four independent non-conserved GFFs defining this transition by considering the nucleon as the on-shell state, which allows us to use the DAs of the nucleon in terms of wavefunctions of different twists. We use two sets of parameters entering the wavefunctions to numerically analyze the obtained sum rules for the FFs. We also give the fit functions of the transition form factors describing the behavior of the FFs in terms of $ Q^2 $.
The article is structured as follows. In next section we derive the desired sum rules for the transition GFFs of the $N \to \Delta$. In section III, we numerically analyze the sum rules for the form factors to find their $ Q^2 $ behavior and their values at static limit . The last section is reserved for our concluding remarks.
\section{Formalism} \label{secII}
In light-cone sum rule (LCSR) approach, the following two-point correlation function in the presence of the on-shell nucleon state is responsible for the calculations of the GFFs of the $N \to \Delta$ transition:
\begin{equation}\label{corf}
\Pi_{\alpha\mu\nu}(p,q)=i\int d^4 x e^{iqx} \langle 0 |\mathcal{T}[J_{\alpha}^{\Delta}(0)T_{\mu \nu}^q(x)]|N(p)\rangle,
\end{equation}
where $ \mathcal{T} $ is the time ordering operator, $p$ is the nucleon's four-momentum, $q$ is the transferred momentum,
$J_{\alpha}^{\Delta}(0)$ is the $\Delta$'s interpolating current and $T_{\mu\nu}^q(x)$ is the quark part of the EMT current at point $ x $. We insert a complete set of the intermediate $\Delta(p',s') $ with momentum $ p' $ and $ ,s' $ to the correlation function and perform the four-integral over $ x $. This ends up in
\begin{align}\label{phys}
\Pi_{\alpha\mu\nu}^{Had}(p,q) &=\sum_{s{'}} \frac{\langle0|J_{\alpha}^{\Delta}|{\Delta(p',s')}\rangle\langle {\Delta(p',s')}
|T_{\mu \nu}^q|N(p,s)\rangle}{m^2_{\Delta}-p'^2}
+...,
\end{align}
where dots stand for the contributions of the higher states and continuum and we choose the threshold such that the correlation function includes only the ground state $\Delta$. To go further, we need to define the following matrix element in terms of the residue of the $\Delta$ baryon ($ \lambda_\Delta $):
\begin{align}\label{resi1}
\langle0|J_{\alpha}^{\Delta}|{\Delta(p',s')}\rangle &= \lambda_{\Delta} u_{\alpha}(p',s'),
\end{align}
where, $ u_{\alpha}(p',s') $ is the Rarit-Schwinger spinor. To proceed, we define the matrix element of the quark part of the EMT current sandwiched between the nucleon and $\Delta$ state. As the quark part of the EMT current solely is not a conserved quantity, the transition matrix element is decomposed in terms of nine form factors (five independent conserved and four independent non-conserved) by demanding the criteria of the Lorentz invariance, discrete space-time symmetries and the related equations of motions \cite{Kim:2022bwn}:
\begin{widetext}
\begin{align}\label{GFFs}
\langle \Delta(p^\prime,s')|T_{\mu\nu}^q|N(p,s)\rangle &=
\bar{u}_{\beta}(p^\prime,s')
\Bigg[F_1^{N\Delta}(Q^2)\bigg\{g_{\beta\{\mu}P_{\nu\}}+\frac{(m^2_{\Delta}-m^2_N)}{\Delta^2}g_{\mu\nu} \Delta_{\beta}- \frac{(m^2_{\Delta}-m^2_N)}{2 \Delta^2}g_{\beta\{ \mu}\Delta_{\nu\}}-\frac{\Delta_\beta P_{\{\mu} \Delta_{\nu\}}}{\Delta^2}
\bigg\}\nonumber\\
&+F_2^{N\Delta}(Q^2)\bigg\{ P_\mu P_\nu \Delta_\beta + \frac{(m_\Delta^2-m_N^2)^2}{4 \Delta^2}g_{\mu\nu} \Delta_\beta- \frac{(m^2_{\Delta}-m^2_N)}{2 \Delta^2}P_{\{\mu}\Delta_{\nu\}}\Delta_\beta
\bigg\}\nonumber\\
&+F_3^{N\Delta}(Q^2)\bigg\{ (\Delta_{\mu}\Delta_{\nu}- \Delta^2 g_{\mu\nu})\Delta_\beta \bigg\}
\nonumber\\
&+F_4^{N\Delta}(Q^2) \bigg\{ \gamma_{\{\mu} g_{ \nu\}\beta} + \frac{2(m_\Delta+m_N)}{\Delta^2} g_{\mu\nu} \Delta_\beta - \frac{(m_\Delta+m_N)}{\Delta^2} g_{\beta\{\mu}\Delta_{\nu\}}-\frac{1}{\Delta^2} \gamma_{\{\mu}\Delta_{\nu\}}\Delta_\beta
\bigg\}\nonumber\\
&+ F_5^{N\Delta}(Q^2)\bigg\{ \gamma_{\{\mu}P_{\nu\}}\Delta_\beta +\frac{(m^2_{\Delta}-m^2_N)(m_\Delta+m_N)}{\Delta^2} g_{\mu\nu}\Delta_\beta - \frac{(m_\Delta+m_N)}{\Delta^2}\Delta_{\{\mu}P_{\nu\}}\Delta_\beta \nonumber\\
&- \frac{(m^2_{\Delta}-m^2_N)}{2\Delta^2}\gamma_{\{\mu}\Delta_{\nu\}}\Delta_\beta
\bigg\}
+ \bar C_1^{N\Delta} (Q^2) g_{\mu\nu}\Delta_\beta
+\bar C_2^{N\Delta} (Q^2) \Delta_{\{\mu} P_{\nu\}}\Delta_\beta
\nonumber\\
&
+\bar C_3^{N\Delta} (Q^2) \gamma_{\{\mu}\Delta_{\nu\}} \Delta_\beta
+\bar C_4^{N\Delta} (Q^2) g_{\beta\{ \mu}\Delta_{\nu\}}
\Bigg] \gamma_5 u_{N}(p,s),
\end{align}
\end{widetext}
where $ P= (p'+p)/2$,
$\Delta = p'-p$,
$X_{\{\mu}Y_{\nu\}} = (X_{\mu}Y_{\nu}+ X_{\nu}Y_{\mu})/2$ and $Q^2=- \Delta^2$.
Here, $F_1^{N\Delta}(Q^2) $, $F_2^{N\Delta}(Q^2) $, $F_3^{N\Delta}(Q^2) $, $F_4^{N\Delta}(Q^2) $, $F_5^{N\Delta}(Q^2) $, $\bar C_1^{N\Delta} (Q^2) $, $\bar C_2^{N\Delta} (Q^2) $, $\bar C_3^{N\Delta} (Q^2) $, and $\bar C_4^{N\Delta} (Q^2) $ are the transition GFFs.
To further simplify, the summation over the spin of the the Rarita-Schwinger spinor for the $\Delta$ baryon is introduced:
\begin{equation}\label{sspin}
\sum_{s'}{u_\alpha}(p',s'){\bar u_{\beta}}(p',s')=-(\not\!p'+m_{\Delta})\{g_{\alpha\beta}-\frac{1}{3}\gamma_{\alpha}\gamma_{\beta}-\frac{2p'_{\alpha}p'_{\beta}}
{3m_{\Delta}^{2}}+\frac{p'_{\alpha}\gamma_{\beta}-p'_{\beta}\gamma_{\alpha}}{3m_{\Delta}}\}.
\end{equation}
Using Eqs. (\ref{resi1}), \eqref{GFFs} and \eqref{sspin} in Eq. \eqref{phys}, we recast the phenomenological or physical form of the correlation function in terms of the GFFs and other corresponding hadronic parameters as:
\begin{widetext}
\begin{align}\label{physson}
\Pi_{\alpha \mu\nu}^{Had} (p,q)&= \frac{\lambda_{\Delta}}{m^2_{\Delta}-p'^2}\Bigg[ -(\not\!p'+m_{\Delta})\{g_{\alpha\beta}-\frac{1}{3}\gamma_{\alpha}\gamma_{\beta}-\frac{2p'_{\alpha}p'_{\beta}}
{3m_{\Delta}^{2}}+\frac{p'_{\alpha}\gamma_{\beta}-p'_{\beta}\gamma_{\alpha}}{3m_{\Delta}}\}\Bigg]\nonumber\\
&\times \Bigg[F_1^{N\Delta}(Q^2)\bigg\{g_{\beta\{\mu}P_{\nu\}}+\frac{(m^2_{\Delta}-m^2_N)}{\Delta^2}g_{\mu\nu} \Delta_{\beta}- \frac{(m^2_{\Delta}-m^2_N)}{2 \Delta^2}g_{\beta\{ \mu}\Delta_{\nu\}}-\frac{\Delta_\beta P_{\{\mu} \Delta_{\nu\}}}{\Delta^2}
\bigg\}\nonumber\\
&+F_2^{N\Delta}(Q^2)\bigg\{ P_\mu P_\nu \Delta_\beta + \frac{(m_\Delta^2-m_N^2)^2}{4 \Delta^2}g_{\mu\nu} \Delta_\beta- \frac{(m^2_{\Delta}-m^2_N)}{2 \Delta^2}P_{\{\mu}\Delta_{\nu\}}\Delta_\beta
\bigg\}\nonumber\\
&+F_3^{N\Delta}(Q^2)\bigg\{ (\Delta_{\mu}\Delta_{\nu}- \Delta^2 g_{\mu\nu})\Delta_\beta \bigg\}
\nonumber\\
&+F_4^{N\Delta}(Q^2) \bigg\{ \gamma_{\{\mu} g_{ \nu\}\beta} + \frac{2(m_\Delta+m_N)}{\Delta^2} g_{\mu\nu} \Delta_\beta - \frac{(m_\Delta+m_N)}{\Delta^2} g_{\beta\{\mu}\Delta_{\nu\}}-\frac{1}{\Delta^2} \gamma_{\{\mu}\Delta_{\nu\}}\Delta_\beta
\bigg\}\nonumber\\
&+ F_5^{N\Delta}(Q^2)\bigg\{ \gamma_{\{\mu}P_{\nu\}}\Delta_\beta +\frac{(m^2_{\Delta}-m^2_N)(m_\Delta+m_N)}{\Delta^2} g_{\mu\nu}\Delta_\beta - \frac{(m_\Delta+m_N)}{\Delta^2}\Delta_{\{\mu}P_{\nu\}}\Delta_\beta \nonumber\\
&- \frac{(m^2_{\Delta}-m^2_N)}{2\Delta^2}\gamma_{\{\mu}\Delta_{\nu\}}\Delta_\beta
\bigg\}
+ \bar C_1^{N\Delta} (Q^2) g_{\mu\nu}\Delta_\beta
+\bar C_2^{N\Delta} (Q^2) \Delta_{\{\mu} P_{\nu\}}\Delta_\beta
\nonumber\\
&
+\bar C_3^{N\Delta} (Q^2) \gamma_{\{\mu}\Delta_{\nu\}} \Delta_\beta
+\bar C_4^{N\Delta} (Q^2) g_{\beta\{ \mu}\Delta_{\nu\}}
\Bigg] \gamma_5 u_{N}(p,s),
\end{align}
where we only kept the contribution of the spin-$3/2$ $\Delta$ baryon and omitted the pollution coming from the spin-$1/2$ particles.
In principle, the correlation function can also include the contributions from spin-$1/2$ particles.
The overlap of the spin-$1/2$ particles with the $J_\alpha^\Delta$ current can be written as
\begin{equation}
\langle 1/2(p') \vert J^\Delta_\alpha \vert 0 \rangle = \left(A p'_\alpha + B \gamma_\alpha \right) u(p')
\end{equation}
where $u(p')$ is the Dirac spinor describing the spin-$1/2$ particles. Hence, when the gamma matrices are put into the order $\gamma_\alpha\gamma_\mu\gamma_\nu q\!\!\!\slash p\!\!\!\slash ' \gamma_5$ in the related correlation function,
the spin-$1/2$ states contribute only to the structures which have $\gamma_\alpha$ at the beginning or those that are proportional to $p'_\alpha$. Then, the contributions of the spin-$1/2$ states in the correlation
function are eliminated by ignoring the structures proportional to
$p'_\alpha$ and the structures that contain a $\gamma_\alpha$ at the beginning. By this way, only
the contributions from spin-$3/2$ states are kept (see also Refs. \cite{Belyaev:1993ss, Belyaev:1982cd}).
\end{widetext}
As a result, one can decompose the hadronic representation of the correlation function in terms of the various Lorentz structures entering the calculations:
\begin{align}
\Pi_{\alpha\mu\nu}^{Had} (p,q) &=
\Pi_1^{Had}(Q^2)\, q_\mu g_{\alpha \nu} q\!\!\!\slash \gamma_5
+ \Pi_2^{Had}(Q^2)\, p^{\prime}_\mu p^{\prime}_\nu q_\alpha \gamma_5
+ \Pi_3^{Had}(Q^2)\, q_\mu q_\nu q_\alpha q\!\!\!\slash \gamma_5
+ \Pi_4^{Had}(Q^2)\,g_{\alpha \mu}\gamma_\nu q\!\!\!\slash \gamma_5\nonumber\\
&+ \Pi_5^{Had}(Q^2)\, q_\alpha q_\mu \gamma_\nu q\!\!\!\slash \gamma_5
+ \Pi_6^{Had}(Q^2)\, g_{\mu\nu} q_\alpha \gamma_5
+ \Pi_7^{Had}(Q^2)\, p^{\prime}_\mu q_\alpha q_\nu \gamma_5
+ \Pi_8^{Had}(Q^2)\, q_\alpha q_\nu \gamma_\mu \gamma_5 \nonumber\\
&+ \Pi_9^{Had}(Q^2)\, g_{\alpha\mu}q_\nu \gamma_5 +...,
\end{align}
where the invariant functions $ \Pi_i^{Had}(Q^2) $ are functions of GFFs.
The next step is to calculate the correlation function in quark-gluon language and in terms of the QCD fundamental degrees of freedom. For this purpose, we insert the explicit forms of the currents, which are given in terms of the corresponding quark fields, into the correlation function. These currents are given as:
\begin{align}
\label{intpol}
J_{\alpha}^{\Delta}(0)=&\frac{1}{\sqrt{3}}\epsilon^{abc}[2(u^{aT}(0) C\gamma_\alpha d^b(0)) u^c(0)+ (u^{aT}(0)C\gamma_\alpha u^b(0))d^c(0)],\nonumber\\
T_{\mu\nu}^q (x) &= \frac{i}{2}\bigg[\bar{u}^d(x)\overleftrightarrow{D}_\mu(x) \gamma_\nu u^d(x)
+ \bar{d}^e(x)\overleftrightarrow{D}_\mu(x) \gamma_\nu d^e(x)
+(\mu \leftrightarrow \nu) \bigg],
\end{align}
where $C$ is the charge conjugation operator; and $a$, $b$, $c$, $d$, $e$ are color indices.
The covariant derivative, $ \overleftrightarrow{D}_\mu(x) $, is defined as
\begin{align}
\overleftrightarrow{D}_\mu(x) &=\frac{1}{2} \Big[ \overrightarrow{D}_\mu(x) - \overleftarrow{D}_\mu(x) \Big],
\end{align}
where
\begin{align}
\overrightarrow{D}_\mu(x) &= \overrightarrow{\partial}_\mu(x)+igA_\mu(x), \\
\overleftarrow{D}_\mu(x) &= \overleftarrow{\partial}_\mu(x) -igA_\mu(x),
\end{align}
with $A_\mu$ being the gluon field. By using the explicit forms of the interpolating currents in the correlation function and doing all the possible contractions among the quark fields via the Wick's theorem, we get the QCD side of the correlation function in terms of the light-quark propagators and DAs of the nucleon:
\begin{widetext}
\begin{eqnarray}\label{corrfunc}
( \Pi_{\alpha \mu\nu}^{QCD} )_{\lambda\eta}(p,q)&=&-\frac{1}{8\sqrt3}\int d^4 x e^{iqx}~(C\gamma_\alpha)_{\alpha\beta}( \overleftrightarrow{D}_\mu (x)\gamma_\nu)_{\rho\sigma} \bigg\{4\epsilon^{abc}\langle \nonumber
0|{u}_{\sigma}^a(0) {u}_{\theta}^b(x) {d}_{\phi}^c(0)|N(p,s) \rangle \\ \nonumber
& & \bigg[2\delta_\alpha^{\eta}\delta_\sigma^{\theta} \delta_\beta^{\phi}S_q(-x)_{\lambda\rho}+
2\delta_\lambda^{\eta} \delta_\sigma^{\theta} \delta_\beta^{\phi}S_q(-x)_{\alpha\rho}\nonumber
+ \delta_\alpha^{\eta} \delta_\sigma^{\theta} \delta_\lambda^{\phi} S_q(-x)_{\beta\rho}+\delta_\beta^{\eta} \delta_\sigma^{\theta} \delta_\phi^{\lambda} S_q(-x)_{\alpha\rho}\bigg]\nonumber \\
& &
- 4\epsilon^{abc}\langle 0|{u}_{\sigma}^a(0) {u}_{\theta}^b(0) {d}_{\phi}^c(x)|N(p,s) \rangle
\bigg[ 2\delta_\alpha^{\eta}\delta_\lambda^{\theta} \delta_\sigma^{\phi} S_q(-x)_{\beta\rho}
+\delta_\alpha^{\eta}\delta_\beta^{\theta} \delta_\sigma^{\phi} S_q(-x)_{\lambda\rho}\bigg]\bigg\}+ (\mu \leftrightarrow \nu) ,
\end{eqnarray}
\end{widetext}
where $S_q(x)$ is the light quark propagator defined as
\begin{align}
\label{edmn09}
S_{q}(x) &=
\frac{1}{2 \pi^2 x^2}\Big( i \frac{{x\!\!\!\slash }}{x^{2}}-\frac{m_{q}}{2 } \Big)
- \frac{\langle \bar qq \rangle }{12} \Big(1-i\frac{m_{q} x\!\!\!\slash }{4} \Big)
- \frac{\langle \bar q \sigma.G q \rangle }{192}x^2 \Big(1-i\frac{m_{q} x\!\!\!\slash }{6} \Big)
-\frac {i g_s }{32 \pi^2 x^2} ~G^{\mu \nu} (x) \bigg[\rlap/{x}
\sigma_{\mu \nu} + \sigma_{\mu \nu} \rlap/{x}
\bigg].
\end{align}
We set $m_q = 0$, and the terms $ \sim $ $\langle \bar qq \rangle$ and $\langle \bar q \sigma.G q \rangle$ are omitted following the Borel transformation, which is applied to suppress the contributions of the higher states and continuum. Hence, only the first term in the light quark propagator gives contribution to the calculations.
To proceed with the calculation of the correlation function, the matrix element of the local three-quark operator
$4\epsilon^{abc}\langle 0|q_{1\alpha}^a(a_1 x) q_{2\beta}^b(a_2 x) q_{3\gamma}^c(a_3 x)|N(p,s)\rangle$
is needed.
The light-cone distribution amplitudes of the nucleon, which we use in our study to extract the GFFs, are presented in Ref.~\cite{Braun:2006hz} up to twist six on the basis of QCD conformal partial wave expansion. All parameters inside the DAs are also borrowed from this reference.
By using the DAs of the nucleon and applying the Fourier transformations, the QCD representation of the correlation function is acquired in terms of different Lorentz structures in the following form:
\begin{align}
\Pi_{\alpha\mu\nu}^{QCD} (p,q) &=
\Pi_1^{QCD}(Q^2)\, q_\mu g_{\alpha \nu} q\!\!\!\slash \gamma_5
+ \Pi_2^{QCD}(Q^2)\, p^{\prime}_\mu p^{\prime}_\nu q_\alpha \gamma_5
+ \Pi_3^{QCD}(Q^2)\, q_\mu q_\nu q_\alpha q\!\!\!\slash \gamma_5
+ \Pi_4^{QCD}(Q^2)\,g_{\alpha \mu}\gamma_\nu q\!\!\!\slash \gamma_5\nonumber\\
&+ \Pi_5^{QCD}(Q^2)\, q_\alpha q_\mu \gamma_\nu q\!\!\!\slash \gamma_5
+ \Pi_6^{QCD}(Q^2)\, g_{\mu\nu} q_\alpha \gamma_5
+ \Pi_7^{QCD}(Q^2)\, p^{\prime} _\mu q_\alpha q_\nu \gamma_5
+ \Pi_8^{QCD}(Q^2)\, q_\alpha q_\nu \gamma_\mu \gamma_5 \nonumber\\
&+ \Pi_9^{QCD}(Q^2)\, g_{\alpha\mu}q_\nu \gamma_5 +... ~,
\end{align}
where $ \Pi_{1,2,...,9}^{QCD}(Q^2) $ are invariant functions corresponding to the coefficients of the selected structures.
The LCSR for the $N \rightarrow \Delta $ transition GFFs are obtained by matching the coefficients of
the chosen Lorentz structures from both the physical and QCD representations of the correlation function. To abolish the contributions coming from the higher states and continuum, Borel transformation as well as continuum subtraction supplied by the quark-hadron duality assumption are applied.
We should stress that, we use the Lorentz structures $q_\mu g_{\alpha \nu} q\!\!\!\slash \gamma_5 $, $p^{\prime}_\mu p^{\prime}_\nu q_\alpha \gamma_5$,
$q_\mu q_\nu q_\alpha q\!\!\!\slash \gamma_5$, $g_{\alpha \mu}\gamma_\nu q\!\!\!\slash \gamma_5$, $q_\alpha q_\mu \gamma_\nu q\!\!\!\slash \gamma_5$, $g_{\mu\nu} q_\alpha \gamma_5$, $ p^{\prime}_\mu q_\alpha q_\nu \gamma_5$, $q_\alpha q_\nu \gamma_\mu \gamma_5$ and $g_{\alpha\mu}q_\nu \gamma_5 $ to find the LCSR for the $N \rightarrow \Delta $ transition GFFs, $F_1^{N\Delta}(Q^2)$, $F_2^{N\Delta}(Q^2) $, $F_3^{N\Delta}(Q^2) $, $F_4^{N\Delta}(Q^2) $, $F_5^{N\Delta}(Q^2) $, $\bar C_1^{N\Delta} (Q^2) $, $\bar C_2^{N\Delta} (Q^2) $, $\bar C_3^{N\Delta} (Q^2) $, and $\bar C_4^{N\Delta} (Q^2) $, respectively.
Hence,
\begin{align}
&\label{F1FF} F_1^{N\Delta}(Q^2) \frac{5\, \lambda_\Delta}{6\, (m_\Delta^2-{p^{\prime}}^2)} = \rho_1^{QCD}(v,y,x_1,x_2,x_3),~~~~ ~~~~ ~~~~ ~~
F_2^{N\Delta}(Q^2) \frac{\lambda_\Delta (m_\Delta+ 5m_N) }{3\,(m_\Delta^2-{p^{\prime}}^2)}= \rho_2^{QCD}(v,y,x_1,x_2,x_3),\\ \label{F2FF}
& F_3^{N\Delta}(Q^2) \frac{5\,\lambda_\Delta}{3\,(m_\Delta^2-{p^{\prime}}^2)} =- \rho_3^{QCD}(v,y,x_1,x_2,x_3),
~~~~ ~~~~~~~~
F_4^{N\Delta}(Q^2) \frac{\lambda_\Delta}{(m_\Delta^2-{p^{\prime}}^2)} =- \rho_4^{QCD}(v,y,x_1,x_2,x_3), \\
& F_5^{N\Delta}(Q^2) \frac{\lambda_\Delta (m_\Delta+ m_N)}{3\,(m_\Delta^2-{p^{\prime}}^2)} = \rho_5^{QCD}(v,y,x_1,x_2,x_3), ~~~~ ~~~~ ~~~~
\bar C_1^{N\Delta} (Q^2) \frac{\lambda_\Delta (m_\Delta+5 m_N)}{3\,(m_\Delta^2-{p^{\prime}}^2)} = \rho_6^{QCD}(v,y,x_1,x_2,x_3),\\
&\bar C_2^{N\Delta} (Q^2) \frac{2\,\lambda_\Delta}{3\,(m_\Delta^2-{p^{\prime}}^2)} = \rho_7^{QCD}(v,y,x_1,x_2,x_3), ~~~~ ~~~~~~~~ ~~~
\bar C_3^{N\Delta} (Q^2) \frac{\lambda_\Delta (m_\Delta+ m_N)}{3\,(m_\Delta^2-{p^{\prime}}^2)} = \rho_8^{QCD}(v,y,x_1,x_2,x_3), \\
& \bar C_4^{N\Delta} (Q^2) \frac{2 \,\lambda_\Delta}{m_N m_\Delta\,(m_\Delta^2-{p^{\prime}}^2)} = \rho_9^{QCD}(v,y,x_1,x_2,x_3).\label{C4FF}
\end{align}
Since the explicit forms of $\rho_i^{QCD}(v,y,x_1,x_2,x_3)$ functions are very lengthy, they will not be given in full here, but in order for the reader to understand what these functions look like, we give the explicit expression of the $\rho_1^{QCD}(v,y,x_1,x_2,x_3)$ function as an example,
\begin{align}
\rho_1^{QCD}(v,y,x_1,x_2,x_3)&=\frac{m_N^3}{4 \sqrt{3}}\int_0^1 \frac{(1+v)}{(q-p v)^4} dv \int_0^v dy \int_y^1 d{x_2} \int_0^{1-x_2} d{x_1} \big[\big(T_1-T_2-T_5+T_6 -2T_7-2T_8\nonumber\\
&+V_1-V_2-V_3-V_4-V_5+V_6\big)(x_1,x_2,1-x_1-x_2)\big
\nonumber\\
&+\frac{m_N^3}{4 \sqrt{3}}\int_0^1 \frac{(1+v)}{(q-p v)^4} dv \int_0^v dy \int_y^1 d{x_3} \int_0^{1-x_3} d{x_1}\big[\big(-A_1+A_2-A_3-A_4+A_5-A_6-4T_1 \nonumber\\
&+2T_2+2T_3+2T_4+2T_5-4T_6+2T_7+2T_8+V_1-V_2-V_3-V_5+V_6\big)(x_1,1-x_1-x_3,x_3)\big]
\nonumber\\
& +\frac{m_N}{4\sqrt{3}}\int_0^1 \frac{1}{(q-p y)^2} dy \int_y^1 d{x_2} \int_0^{1-x_2}d{x_1} \big[\big(-2A_1+2A_2-2A_3-T_1+T_2+2T_7+2V_1-2V_2 \nonumber\\&
-2V_3 \big)(x_1,x_2,1-x_1-x_2)\big]\nonumber\\
&+\frac{m_N}{\sqrt{3}}\int_0^1 \frac{1}{(q-p y)^2} dy \int_y^1 d{x_3} \int_0^{1-x_3}d{x_1} \big[\big(T_1+T_3-T_7 \big)(x_1,1-x_1-x_3,x_3)\big]\nonumber\\
&+\frac{m_N^3}{2\sqrt{3}}\int_0^1 \frac{(1+x_2)}{(q-p x_2)^4}dx_2 \int_0^{1-x_2}d{x_1} \big[\big( A_1^M-T_1^M-2V_1^M\big)(x_1,x_2,1-x_1-x_2)\big]\nonumber\\
&-\frac{m_N^3}{2\sqrt{3}}\int_0^1 \frac{(1+x_3)}{(q-p x_3)^4}dx_3 \int_0^{1-x_3}d{x_1} \big[T_1^M(x_1,1-x_1-x_3,x_3)\big]\nonumber\\
&-\frac{m_N}{4 \sqrt{3}}\int_0^1 \frac{(1+x_2)}{(q-p x_2)^2}dx_2 \int_0^{1-x_2}d{x_1} \big[\big( 2A_1+A_3-2V_1+V_3\big)(x_1,x_2,1-x_1-x_2)\big]\nonumber\\
&+\frac{m_N}{4 \sqrt{3}}\int_0^1 \frac{(1+x_3)}{(q-p x_3)^2}dx_3 \int_0^{1-x_3}d{x_1} \big[\big( A_3+2T_1+V_3\big)(x_1,1-x_1-x_3,x_3)\big].
\end{align}
The Borel transformation as well as the continuum subtraction are performed making use of the following replacements \cite{Braun:2006hz}:
\begin{align}
\int dx \frac{\rho(x)}{(q-xp)^2}&\rightarrow -\int_{x_0}^1\frac{dx}{x}\rho(x) e^{-s(x)/M^2}, \nonumber \\
\int dx \frac{\rho(x)}{(q-xp)^4}&\rightarrow \frac{1}{M^2} \int_{x_0}^1\frac{dx}{x^2}\rho(x) e^{-s(x)/M^2}
+\frac{\rho(x_0)}{Q^2+x_0^2 m_{N}^2} e^{-s_0/M^2},\nonumber\\
\int dx \frac{\rho(x)}{(q-xp)^6}&\rightarrow -\frac{1}{2M^4}\int_{x_0}^1\frac{dx}{x^3}\rho(x) e^{-s(x)/M^2}
-\frac{1}{2M^2}\frac{\rho(x_0)}{x_0(Q^2+x_0^2m_{N}^2)}e^{-s_0/M^2}\nonumber\\
&+\frac{1}{2}\frac{x_0^2}{Q^2+x_0^2m_{N}^2}\bigg[\frac{d}{dx_0}\frac{\rho(x_0)}{x_0(Q^2+x_0^2m_{N}^2)}\bigg]e^{-s_0/M^2},
\label{subtract3}
\end{align}
where
\begin{eqnarray}
s(x)=(1-x)m_{N}^2+\frac{1-x}{x}Q^2,
\end{eqnarray}
and $M^2$ is the Borel mass parameter, which enters in the calculations after applying the Borel transformation in terms of $p^{\prime 2}$. Here, $x_0$ is found by solving the equation, $s(x)=s_0$, i.e.,
\begin{eqnarray}
x_0&=&\Big[\sqrt{(Q^2+s_0-m_{N}^2)^2+4m_{N}^2 Q^2}-(Q^2+s_0-m_{N}^2)\Big]/2m_{N}^2,
\end{eqnarray}
with $s_0$ being the continuum threshold parameter.
\section{Numerical Results}\label{secIII}
The sum rules obtained for the GFFs in the previous section contain different variables: Hadronic and QCD input parameters, input parameters inside the DAs of the nucleon, some auxiliary or helping parameters and transferred momentum square $ Q^2 $. The main purpose in this section is to discuss the $ Q^2 $ behavior of the GFFs.
To this end, the input parameters of the quarks, nucleon and $\Delta $ baryon entering the sum rules are selected as $m_u =m_d =0$, $m_N=0.94$~GeV, $m_\Delta = 1.23$~GeV and $\lambda_\Delta = 0.038$~GeV$^3$~\cite{Aliev:2007pi,Aliev:2004ju}. Various input parameters inside the DAs of the nucleon in two sets are borrowed from Ref.~\cite{Braun:2006hz} and presented in table \ref{parameter_table}.
\begin{table}[htp]
\renewcommand{\arraystretch}{1.3}
\addtolength{\arraycolsep}{-0.5pt}
\small
$$
\begin{array}{|l|c|c|}
\hline \hline
& \mbox{Set-I} & \mbox{Set-II} \\ \hline \hline
f_N & ( 5.0 \pm 0.5)\times 10^{-3}~GeV^2 & ( 5.0 \pm 0.5)\times 10^{-3}~GeV^2 \\
\lambda_1 & (-2.7 \pm 0.9)\times 10^{-2}~GeV^2 & (-2.7 \pm 0.9)\times 10^{-2}~GeV^2 \\
\lambda_2 & ( 5.4 \pm 1.9)\times 10^{-2}~GeV^2 & ( 5.4 \pm 1.9)\times 10^{-2}~GeV^2 \\ \hline \hline
A_1^u & 0.38 \pm 0.15 & 0 \\
V_1^d & 0.23 \pm 0.03 & 1/3 \\
f_1^d & 0.40 \pm 0.05 & 1/3 \\
f_2^d & 0.22 \pm 0.05 & 4/15 \\
f_1^u & 0.07 \pm 0.05 & 1/10 \\ \hline \hline
\end{array}
$$
\caption{The numerical values of the main input parameters entering the expressions of the nucleon's DAs. The upper panel shows the dimensionfull parameters the nucleon. In the lower panel we list the values of the five dimensionless parameters that determine the shape of the DAs.}
\renewcommand{\arraystretch}{1}
\addtolength{\arraycolsep}{-1.0pt}
\label{parameter_table}
\end{table}
The next step is to fix the auxiliary parameters: The Borel parameter $ M^2 $ and the continuum threshold $ s_0 $. To this end, we refer to the standard criteria of the sum rule method: The pole dominance over the contributions of the higher states and continuum as well as the higher the twist of the nucleon DAs, the lower its contribution. These criteria are satisfied in the regions that the GFFs, as physical quantities, depend relatively weakly on the auxiliary parameters and show relatively good stability in their working windows. We acquire the following working windows for the $ s_0 $ and $ M^2 $ from the analyses:
\begin{align}
2.0 ~\mbox{GeV}^2 \leq s_0 \leq 2.50 ~\mbox{GeV}^2, \nonumber\\
2.0 ~\mbox{GeV}^2 \leq M^2 \leq 3.0 ~\mbox{GeV}^2. \nonumber
\end{align}
Figures \ref{Msqfigs1} and \ref{Msqfigs2} present the dependence of the GFFs for $ N-\Delta $ transition on the auxiliary parameters at $Q^2 = 1.0$ GeV$^{2}$ for
two sets of the nucleon DAs parameters. These figures demonstrate good stability of the GFFs with respect to the changes of the auxiliary parameters in their working regions: The residual dependencies appear as the uncertainties in the results.
\begin{table}[htp]
\hspace*{-0.5cm}
\begin{tabular}{ |l|c|c|c|c|c|l|}
\hline\hline
&\multicolumn{3}{|c|}{Results of set-I} &\multicolumn{3}{|c|}{Results of set-II}\\
\hline\hline
Form Factors & ~~${ f}(0)$~~ & ~~${\cal M}$(GeV)~~ & ~~$\alpha$~~
& ~~${ f}(0)$~~ &~~ ${\cal M}$(GeV)~~ & ~~$\alpha$~~ \\
\hline\hline
$F_1^{N\Delta}$($Q^2$) &$0.52 \pm 0.04$ &$ 1.20 \pm 0.10$ &$2.1-2.3$
&$0.70 \pm 0.06$ &$ 1.18 \pm 0.11$ &$2.8-3.0$\\
$F_2^{N\Delta}$($Q^2$) &$0.22 \pm 0.04$ &$1.10 \pm 0.05$ &$2.4-2.6$
&$0.29 \pm 0.03$ &$1.24 \pm 0.12$ &$2.2-2.4$\\
$F_3^{N\Delta}$($Q^2$) &$-1.62 \pm 0.26$ &$1.22 \pm 0.12$ &$2.3-2.5$
&$-1.00 \pm 0.12$ &$1.14 \pm 0.11$ &$3.4-3.6$ \\
$F_4^{N\Delta}$($Q^2$) &$0.14 \pm 0.02$ &$1.14 \pm 0.06$ &$2.4-2.6$
&$0.10 \pm 0.02$ &$1.13 \pm 0.06$ &$2.4-2.6$\\
$F_5^{N\Delta}$($Q^2$) &$-0.32 \pm 0.02$ & $1.20 \pm 0.10$ &$2.2-2.4$
&$-0.46 \pm 0.06$ & $1.09 \pm 0.12$ &$2.2-2.4$\\
$\bar C_1^{N\Delta}$($Q^2$) &$-0.082 \pm 0.002$ &$ 1.22 \pm 0.10$ &$2.4-2.6$
&$-0.091 \pm 0.003$ &$ 1.24 \pm 0.11$ &$2.4-2.6$ \\
$\bar C_2^{N\Delta}$($Q^2$) &$1.00 \pm 0.10$ &$1.29 \pm 0.11$ &$2.3-2.5$
&$1.73 \pm 0.15$ &$1.16 \pm 0.10$ &$2.1-2.3$ \\
$\bar C_3^{N\Delta}$($Q^2$) &$0.52 \pm 0.02$ &$1.25 \pm 0.05$ &$1.8-2.0$
&$0.79 \pm 0.05$ &$1.19 \pm 0.10$ &$2.7-2.9$\\
$\bar C_4^{N\Delta}$($Q^2$) &$-$ &$-$ &$-$ &$-$ &$-$ &$-$ \\
\hline \hline
\end{tabular}
\caption{Numerical values of the fitting parameters for the transition GFFs of $N-\Delta$.
}
\label{fit_table}
\end{table}
Now, we proceed to discuss the $Q^2 $-behavior of the GFFs. The light cone QCD sum rules give reliable results for $Q^2 \geq 1$ ~GeV$^2$. Hence to find the values of the GFFs at static limit, $Q^2 =0$, we need to extrapolate the results to small values of $Q^2 $. The following $ \alpha $-pole fit functions do this job well and produce all our sum rules results for $Q^2 \geq 1$:
\begin{align}
F_i^{N\Delta}(Q^2)[\bar C_i^{N\Delta}(Q^2)]= f(0)\bigg[1+ \frac{Q^2}{{\cal M}^2}\bigg]^{-\alpha},
\end{align}
where the fit parameters including the values of GFFs at static limit, $ f(0) $, are given in table \ref{fit_table}. The uncertainties in the presented values belong to the errors of all input parameters and those related to the determination of working intervals for the auxiliary parameters. As we previously mentioned, the quark part of the EMT current alone is not conserved and matrix element of the EMT current is parameterized in terms of five independent conserved and four independent non-conserved form factors. The $ C_i^{N\Delta}(Q^2) $ form factors depict the orders of breaking the conservation of quark part of the EMT current. Our analyses do not give reliable results for $ C_4^{N\Delta}(Q^2) $ for both the sets of nucleon DAs parameters.
To our best knowledge, this is the first study in the literature dedicated to the investigation of the $N \rightarrow \Delta$ transition GFFs. Therefore, any experimental data or theoretical estimations are not yet available to compare the results for the considered GFFs. Using the above fit functions, the $Q^2 $-behavior of the GFFs in a wide range, $ Q^2\in[0-10]$ ~GeV$^2$, are depicted in figures \ref{Qsqfigs1} and \ref{QsqfigsII}. Our results may be checked by future related experiments, Lattice QCD and other phenomenological models.
\begin{figure}[htp]
\centering
\subfloat[]{\includegraphics[width=0.32\textwidth]{F1MsqsetI.eps}}~~~~
\subfloat[]{\includegraphics[width=0.32\textwidth]{F2MsqsetI.eps}}\\
\subfloat[]{\includegraphics[width=0.32\textwidth]{F3MsqsetI.eps}}~~~~
\subfloat[]{\includegraphics[width=0.32\textwidth]{F4MsqsetI.eps}}\\
\subfloat[]{\includegraphics[width=0.32\textwidth]{F5MsqsetI.eps}}~~~~
\subfloat[]{\includegraphics[width=0.32\textwidth]{C1barMsqsetI.eps}}\\
\subfloat[]{\includegraphics[width=0.32\textwidth]{C2barMsqsetI.eps}}~~~~
\subfloat[]{\includegraphics[width=0.32\textwidth]{C3barMsqsetI.eps}}
\caption{The dependence of the $N \rightarrow \Delta$ transition GFFs on $M^2$ at $Q^2 = 1.0$~GeV$^2$ and three fixed values of the $s_0$ and set-I parameters.}
\label{Msqfigs1}
\end{figure}
\begin{figure}[htp]
\centering
\subfloat[]{\includegraphics[width=0.32\textwidth]{F1MsqsetII.eps}}~~~~
\subfloat[]{\includegraphics[width=0.32\textwidth]{F2MsqsetII.eps}}\\
\subfloat[]{\includegraphics[width=0.32\textwidth]{F3MsqsetII.eps}}~~~~
\subfloat[]{\includegraphics[width=0.32\textwidth]{F4MsqsetII.eps}}\\
\subfloat[]{\includegraphics[width=0.32\textwidth]{F5MsqsetII.eps}}~~~~
\subfloat[]{\includegraphics[width=0.32\textwidth]{C1barMsqsetII.eps}}\\
\subfloat[]{\includegraphics[width=0.32\textwidth]{C2barMsqsetII.eps}}~~~~
\subfloat[]{\includegraphics[width=0.32\textwidth]{C3barMsqsetII.eps}}
\caption{The dependence of the $N \rightarrow \Delta$ transition GFFs on $M^2$ at $Q^2 = 1.0$~GeV$^2$ and three fixed values of the $s_0$ and set-II parameters.}
\label{Msqfigs2}
\end{figure}
\begin{figure}[htp]
\centering
\subfloat[]{\includegraphics[width=0.32\textwidth]{F1QsqsetI.eps}}~~~~
\subfloat[]{\includegraphics[width=0.32\textwidth]{F2QsqsetI.eps}}\\
\subfloat[]{\includegraphics[width=0.32\textwidth]{F3QsqsetI.eps}}~~~~
\subfloat[]{\includegraphics[width=0.32\textwidth]{F4QsqsetI.eps}}\\
\subfloat[]{\includegraphics[width=0.32\textwidth]{F5QsqsetI.eps}}~~~~
\subfloat[]{\includegraphics[width=0.32\textwidth]{C1barQsqsetI.eps}}\\
\subfloat[]{\includegraphics[width=0.32\textwidth]{C2barQsqsetI.eps}}~~~~
\subfloat[]{\includegraphics[width=0.32\textwidth]{C3barQsqsetI.eps}}
\caption{The dependence of the $N \rightarrow \Delta$ transition GFFs on $Q^2$ at fixed values of the $s_0$, average $M^2$ and set-I parameters.}
\label{Qsqfigs1}
\end{figure}
\begin{figure}[htp]
\centering
\subfloat[]{\includegraphics[width=0.32\textwidth]{F1QsqsetII.eps}}~~~~
\subfloat[]{\includegraphics[width=0.32\textwidth]{F2QsqsetII.eps}}\\
\subfloat[]{\includegraphics[width=0.32\textwidth]{F3QsqsetII.eps}}~~~~
\subfloat[]{\includegraphics[width=0.32\textwidth]{F4QsqsetII.eps}}\\
\subfloat[]{\includegraphics[width=0.32\textwidth]{F5QsqsetII.eps}}~~~~
\subfloat[]{\includegraphics[width=0.32\textwidth]{C1barQsqsetII.eps}}\\
\subfloat[]{\includegraphics[width=0.32\textwidth]{C2barQsqsetII.eps}}~~~~
\subfloat[]{\includegraphics[width=0.32\textwidth]{C3barQsqsetII.eps}}
\caption{The dependence of the $N \rightarrow \Delta$ transition GFFs on $Q^2$ at fixed values of the $s_0$, average $M^2$ and set-II parameters.}
\label{QsqfigsII}
\end{figure}
\section{Summary and Concluding Remarks}\label{secIV}
We investigated the transition GFFs of the $N \rightarrow \Delta$ for the first time. The quark part of the EMT current are parameterized in terms of
nine independent (five conserved and four non-conserved) form factors. To calculate these GFFs, we applied the light-cone QCD sum rule technique using the DAs of the on-shell nucleon. The nine independent Lorentz structures entering to the calculations in both the hadronic and QCD sides allowed us to extract the sum rules for the desired GFFs. We numerically analyzed the obtained sum rules for two sets of input parameters inside the DAs of the nucleon. We used multipole fit functions to extrapolate the results to the small values of $Q^2$, $0\leq Q^2<1$~GeV$^2$, to find the values of the GFFs at static limit. We presented the $Q^2$-behaviors of the form factors in the interval $ [0-10]$~GeV$^2$, using the working windows of the auxiliary parameters.
As we previously mentioned, investigation of the EMT current interactions of hadrons besides their electromagnetic, weak and strong interactions can give us valuable information about their mass and spin as well as the pressure and shear force their inside. The transition GFFs of $N \rightarrow \Delta$ calculated in the present study include useful knowledge on the $N-\Delta$ system. Our results can be checked in future experiments. Comparison of our results with the future Lattice QCD and other phenomenological predictions will be of great importance as well.
The direct measurement of the $N-\Delta$ GFFs may not be possible with the present facilities. However, in principle, these GFFs can be extracted from the $N-\Delta$ GPDs by considering them as the second Mellin moment of the transition GPDs. The project of extracting the transition GPDs from the CLAS data is ongoing \cite{Kim:2022bwn,Proceedings:2020fyd}.
\section*{ACKNOWLEDGEMENTS}
K. Azizi is grateful to Iran Science Elites Federation (Saramadan)
for the partial financial support provided under the grant number ISEF/M/401385.
|
\section{Introduction}\label{sec:intro}
To protect accounts from unauthorized access, login notifications are intended to inform users about recent sign-ins.
Depending on the service, the notifications are only sent if the login occurred from an \textit{unknown location} or \textit{new device}, which could indicate malicious activity.
Notifications are often delivered via email and include details about the device (browser and OS), approximate location, date, and time of the sign-in.
Users need to decide whether the reported login is legitimate or malicious and are recommended to
change the password in case the login is unfamiliar.
Logins can be confused to be malicious when users \textit{share accounts} and friends or family log in unknowingly.
While the notification is intended to protect users and provide a feeling of security,
it can also be perceived as burdening and overwhelming by requiring a decision based on technical jargon and highlighting negative consequences.
Previous work~\cite{redmiles-19-should-worry} focused on challenge-based notifications and studied incident-response information-seeking and mental models about attackers.
In contrast, we focus on \emph{granted access} notifications informing users about a recent sign-in and analyze users' comprehension, expectations, and reaction to the notification.
In this work, we collected and analyzed 66~login notifications sent by real-world services and developed a \emph{baseline} notification that we employed in two user studies.
In the first study ($n = 220$), we asked participants to imagine having an account they recently have or have \emph{not} signed in to (depending on their treatment) and then showed participants a login notification that either contained their current \emph{real} sign-in information or was prefilled with information from an attacker relying on statistical data (i.\,e., California, USA; Chrome on Windows).
The second study ($n = 229$), disguised as a psychological test, let users create an account they had to sign in to during the multi-stage study.
Participants then either received a legitimate login notification to their email upon signing in themselves (\emph{Legit}) or unexpectedly received a notification prefilled with statistical sign-in information after around one week (\emph{Malicious}).
We sought to answer the following research questions:
\begin{enumerate}[wide,format=\bfseries,topsep=0pt,itemsep=0pt,labelindent=0pt,partopsep=0pt,parsep=0pt,labelwidth=0pt,label=\textbf{RQ\arabic*}, ref=\textbf{RQ\arabic*}]
\item\label{RQ:reaction-comprehension} \emph{[Comprehension \& Reaction]
Do users understand why they received the notification and which factors may have caused receiving it?
Which actions do users take in response, and is resolving the situation a priority?}
We find that participants correctly understand that ``a login'' caused receiving the notification.
However, they are unaware of or misinterpret the trigger and are thus unsure how to react appropriately.
Unique to Study~1, participants either claim they would change the password (Legit) or report they would mistakenly ignore the notification (Malicious).
Comparing both studies, we find a large gap between participants' intended and actual behavior regarding password changes.
\item\label{RQ:decision-making} \emph{[Decision-Making \& Execution]
Do the currently employed notifications help users distinguish malicious and legitimate logins?
Which information helps account owners with their decision, and do current notifications appropriately guide users in resolving the situation?}
Based on device and location, participants can correctly attribute notifications caused by their own logins, but they are confused when the notification is unexpected (Malicious) and struggle to identify the correct reaction.
\item\label{RQ:perception-expectation} \emph{[Perception \& Expectation]
How do login notifications make users feel?
When do they expect notifications to be sent, and how does prior experience affect their perception and decision?}
Notifications about malicious logins evoke (more) negative emotions, but participants who changed their password also felt empowered by taking action to protect their account. Interestingly, more than 90\% of the participants
expect services to send login notifications because it makes
them feel protected.
\end{enumerate}
Our results suggest that login notifications are not, \emph{per~se}, a net positive for online account security.
We find only $\frac{1}{5}$ of the participants who should have changed their password to protect their account actually did.
While participants appreciate when companies decide to monitor their accounts for incidents, services that send notifications for every login can cause warning fatigue.
We find malformed login notifications and current anti-phishing advice problematic and give clear recommendations for service providers to improve their notifications.
While login notifications can help to reinforce account security, we think that protecting their accounts by identifying malicious logins should not be the sole responsibility of the user and highlight the need for better notifications.
\section{Related Work}\vspace{-.2em}
Next, we outline how our research extends related work.
\subsection{Login Notifications \& Challenges}\vspace{-.2em}
Related to our work is a qualitative interview study ($n=67$) by Redmiles~\cite{redmiles-19-should-worry}.
It explores the account security incident response at Facebook by interviewing users who experienced a login incident.
Different from our work, Redmiles focused on ``secondary authentication'' notifications that prompt users to enter a code to access their accounts.
Redmiles interviewed participants from 5~countries and reported on incident-response information-seeking and mental models about attackers.
Regarding the effectiveness of notifications, Redmiles identified a lack of key information, like the likelihood that the notification is about a legitimate threat, to be problematic.
In contrast, our work studies a different kind of notification (see Section~\ref{sec:background}) and focuses on users' comprehension, expectations, and reaction to the notification and not on regaining access or mental models about attackers.
Markert et al.~\cite{markert-22-rba-admin} studied administrators' configuration of risk-based authentication~(RBA).
The predefined notifications were slightly customized by the administrators, and only few opted to disable them completely.
Administrators lacked consensus about which information to include.
Further aspects were the need for more context and explanation to prevent phishing attacks and to mention the inaccuracy of IP-based location estimation.
Wardle~\cite{wardle-19-get-owned} measured the time it takes for leaked credentials to be abused.
For this, Wardle created accounts on web services, intentionally leaked them online, and then used login notifications, among other signals, to measure the time between the leak of the credentials and their first malicious abuse.
A study by Doerfler et al.~\cite{doerfler-19-login-challenges} evaluated the efficacy of login challenges in preventing account takeovers, finding that up to 94\% of phishing-rooted hijacking attempts and 100\% of automated hijacking attempts can be prevented.
Wiefling et al.~\cite{wiefling-19-rba-in-the-wild} showed that verification codes sent via email are the de~facto standard for login challenges enforced by RBA.
In a subsequent study, they demonstrated that providing this code in the subject can reduce the login time~\cite{wiefling-20-rba-evaluation}.
Jubur et al.~\cite{jubur-21-bypassing-2fa} presented an attack to bypass push-based two-factor authentication~(2FA) notifications.
\subsection{Security Warning \& Notification Design}
There is a large body of literature on security warning design~\cite{bauer-13-warning-guidelines, reeder-18-warning-reaction, walkington-19-better-warnings}.
The most prominent applications are notifications in the context of TLS~\cite{akhawe-13-alice, felt-15-ssl-warnings}, phishing~\cite{petelka-19-phishing-warnings}, and malware~\cite{almuhimedi-14-malware-warning}, as well as domains like warnings for developers~\cite{gorski-20-pd-for-crypto-apis} or countering misinformation~\cite{kaiser-20-warning-disinformation}.
For user authentication, there is work on breach notifications~\cite{huh-17-linkedin, zou-19-might-be-affected},
password-reuse notifications~\cite{golla-18-reuse-notification, thomas-19-pw-checkup}, notifications to promote the use of 2FA~\cite{redmiles-17-2fa-msg-design, golla-21-2fa-adoption}, or FIDO2~\cite{lassak-21-webauthn-misconceptions}, or protect users from using common PINs~\cite{markert-21-pin-unlock}.
While certain design patterns discussed in these works, like the use of opinionated design, proved to be effective for certain applications~\cite{felt-15-ssl-warnings}, we did not observe such patterns in login notifications.
This is likely owed to the uncertainty of whether a true threat is present.
Similarly, the use of alarming language or wording highlighting urgency was low.
Interestingly, only a handful of notifications in our analysis (see Section~\ref{sec:background}) tried to explain why the notification is sent and attempted to address contextual misunderstandings~\cite{reeder-18-warning-reaction, markert-22-rba-admin}.
Personalization was the only best practice strictly applied~\cite{golla-21-2fa-adoption} by referring to ``your account'' and including the account name to create trust and direct the notification.
\section{Login Notifications in the Wild}\label{sec:background}
Login notifications intend to inform users about recent sign-ins and often include technical details such as the login time, used device, or approximate sign-in location.
However, they are not sent for every login.
While theoretically significant location or device changes trigger notifications, the probabilistic nature involving factors like sign-in history and user behavior makes it difficult to predict when notifications are actually sent.
Some services sent notifications for every login, others only in case of significant location and device changes causing a higher risk level.
For example, we noticed receiving fewer sign-in notifications if the affected account had two-factor authentication enabled.
\figNotificationStatistics
\subsection{Notification Types}\label{sec:background:notification-types}
Based on the type of information that is conveyed about a sign-in, one can distinguish three different notifications that are commonly sent by online services:
\renewcommand{\labelenumi}{(\theenumi)}
\begin{enumerate}[noitemsep,nolistsep,leftmargin=1.7em]
\item \textbf{Granted Access:} These notifications inform about a \textit{granted access}. Some services send such notifications for every sign-in, while others follow a risk-based approach (sending only if one or multiple aspects of a sign-in differ from prior observations).
\item \textbf{Additional Challenge:} These notifications inform about a new sign-in attempt for which an \textit{additional challenge} needs to be solved to complete the sign-in (i.\,e., insert a code or click a link).
\item \textbf{Blocked Access:} The notification informs users about a \textit{blocked access}, which can happen because the risk-based authentication system ranks the sign-in as too risky or because a challenge was not solved successfully.
\end{enumerate}
For the remainder of this work, we focus on the first type, i.\,e., notifications informing the user about a \textit{granted access}.
Granted access notifications are by far the most widespread notifications.
Sending this type can be done by every organization, as it does not require an advanced risk assessment (i.\,e., basic logic and the ability to display login details are enough).
While notifications can be sent via different channels (i.\,e., email, SMS, or push notifications), we limited our dataset to the most prominent type, which is email-based notifications.
\figBaselineEmail
\vspace{0.5em}
\subsection{Analysis}
To find a representative \emph{baseline} notification, we collected more than $80$ login-related emails sent by real-world services by actively enumerating over $500$ existing accounts we had on various websites.
For this, we signed in using the Tor browser, which is often classified as suspicious activity, and monitored our inbox for notifications.
Our collection is limited to the top Tranco list~\cite{lepochat-19-tranco} websites (as of November~2022), with about \(\frac{1}{3}\) being in the top~$100$, top~$1000$, and top~$50,000$ respectively.
The dataset is biased towards English language websites (few non-English notifications have been translated).
For analysis, two authors categorized $66$ of the emails as \emph{granted access} notifications.
They then independently analyzed various aspects including: which sign-in information the notifications includes (i.\,e., login time, location, device), what the main components are (i.\,e., headline, malicious instructions), salient design and wording decisions (i.\,e., logo, highlighting of sign-in details, neutral language), and metadata such as sender and subject.
We summarize our findings in Figure~\ref{fig:notification-statistics}.
Please refer to Appendix~\ref{app:real-world:features} for full details.
Appendix~\ref{app:real-world:emailmetadata} shows an in-depth analysis of the email sender and subject.
\vspace{0.5em}
\textbf{Sign-In Information.}
As depicted in Figure~\ref{fig:notification-statistics}, the majority of notifications included the login \CircledTop{\scriptsize\sffamily7}~\emph{Time}~($77\%$), \CircledTop{\scriptsize\sffamily5}~\emph{Account Name}~($74\%$), \CircledTop{\scriptsize\sffamily7}~\emph{Country}~($68\%$), \CircledTop{\scriptsize\sffamily7}~\emph{Browser}~($62\%$), and \CircledTop{\scriptsize\sffamily7}~\emph{Operating System}~($62\%$).
Less frequently, the notifications also included the \CircledTop{\scriptsize\sffamily7}~\emph{Time~Zone}~($59\%$) or a login \emph{IP~Address}~($44\%$).
The small number of notifications including the login \emph{City}~($36\%$) or \emph{State}~($30\%$), is explained by geographical differences between the U.S. and Europe.
Notifications about logins in the U.S. mostly focused on states; European notifications often focused on cities.
\vspace{0.5em}
\textbf{Components.} Throughout all notifications, we noted the presence of a \CircledTop{\scriptsize\sffamily4}~\emph{Headline}~($73\%$) that was often~($76\%$) different from the email subject.
Another key component is the instructions describing how users should react in response to receiving the notification.
While only $64\%$ provided instructions in the \CircledTop{\scriptsize\sffamily8}~\emph{Legitimate} case, more than $97\%$ explained how to react in the \CircledTop{\scriptsize\sffamily9}~\emph{Malicious} case if the user does not recognize the login.
The large majority ($66\%$) recommended changing the password.
Fewer (high-ranked) web services included a button to report the login as malicious or legit on a separate web page ($9\%$) displaying account remediation steps.
Similarly, a small number ($9\%$) suggested to visit the account activity page and reviewing all active sessions and devices.
Prominent among financial services was the option to contact support ($4\%$).
A dedicated \emph{Why Notification} component was included in $21\%$ of the notifications.
It primarily creates context and explains to users why they received the notification.
It often gives examples of legitimate (i.\,e., new computer/smartphone, new physical location, and private browsing) and malicious causes (``someone else/unauthorized gained access'') that might have triggered the notification.
$33\%$ included a link to a dedicated \emph{Help Page} (note: \emph{regular} support links in the email footer were not counted).
About $24\%$ of the emails tried to use the \emph{Opportune Moment}
to tell the user about other options to secure their account (i.\,e., enabling 2FA).
The dangers of \emph{Phishing} and methods to double-check the legitimacy of the notification were mentioned in $17\%$ of the emails, with the most prominent suggestion to not click the ``change password'' link and instead sign in to the website by manually pasting or typing in the URL.
About half of the notifications included a \CircledTop{\scriptsize\sffamily10}~\emph{Closing}~($50\%$) text that often thanked the user and included the name of a ``\{service\} account team.''
A footer with \CircledTop{\scriptsize\sffamily11}~\emph{Legal} information was included in $67\%$ of the emails, and an \emph{Unsubscribe} link was present in $8\%$ of the notifications.
\vspace{0.5em}
\textbf{Wording \& Design.}
The wording of the email subject was often \emph{Neutral}~($65\%$), with a strong focus on ``\emph{New login to \{service\}},'' in some cases also alarming~($23\%$), like ``\emph{Security alert}'' or a prompt~($9\%$), like ``\emph{Please review this sign in!},'' and in two cases a question~($3\%$), like ``\emph{Did you recently sign into \{service\}}?''
Almost all ($92\%$) emails referred to \CircledTop{\scriptsize\sffamily6}~``your account'' to emphasize the importance of the notification.
A few services tried to address the inaccuracies of IP-based geographic location estimation by describing it as \emph{Approximate Location}~($26\%$).
The majority~($89\%$) of the notifications were sent as \emph{HTML} emails.
Very few sent only plaintext emails.
For a ``corporate look-and-feel,'' $80\%$ of all notifications included a \CircledTop{\scriptsize\sffamily3}~\emph{Logo}, with an even split between a centered or left-aligned placement.
Interestingly, $23\%$ of the emails opted to display the sign-in information in a visually detached box, most likely to draw the user's attention to the login details.
\vspace{0.5em}
\textbf{Selecting a Representative Login Notification.}
Our data-driven \emph{baseline} login notification (see Figure~\ref{fig:baselineemail}) includes all components used by at least $50\%$ of the notifications.
It uses a neutral subject and a slightly modified headline.
We adjusted our email sender, opted for an HTML email, and included a logo.
Also, we included the affected account name and referred to ``your account.''
We included the most popular sign-in details, as well as legitimate and malicious instructions for users to take after receiving the notification.
As our study sample was U.S.-based, we included the \CircledTop{\scriptsize\sffamily7}~\emph{State} in the sign-in details.
The email also included a closing and footer with fictional legal information.
\section{Baseline Study: Method}\label{sec:baseline-method}
This section describes our baseline studies' methodology.
We first explain the study structure and treatments.
Next, we describe the recruitment process and demographics, before concluding limitations and ethical considerations.
\subsection{Study Protocol}
We ran a quantitative online study with the following structure utilizing our baseline notification (see Section~\ref{sec:background}).
\begin{enumerate}[noitemsep,nolistsep,leftmargin=1.2em]
\item \textit{Consent:} In line with ethical best practices, we briefed participants about the research project and their participation at the beginning of the study.
To start the study, participants had to consent.
\item \textit{Scenario:} Followed by the consent form, we described the scenario for the study telling participants to imagine that they are Jo Doe ({\color{RoyalBlue}jo.<EMAIL>}) and have an account with the company AcmeCo which is ``like other accounts you may have, such as for online shopping or social media.'' Based on their (randomly) assigned treatment (see Section~\ref{sec:treatments}), we told participants on a subsequent page to imagine that they recently signed into this AcmeCo account (\textit{Legit}) or have \underline{not} signed into the account for a while (\textit{Malicious}).
\item \textit{Notification:} After describing the circumstances, participants saw a login notification sent by AcmeCo (Figure~\ref{fig:baselineemail}). The login information depicted in the notification was either derived from the currently used system (\textit{Legit}) or showed a login from California using Chrome on Windows (\textit{Malicious}). From now on, up to the questions about prior experiences, the survey website used a two-column layout that displayed the login notification on the left and the questions on the right-hand side.
\item \textit{Reaction:} To understand how participants felt in reaction to being exposed to the login notification, we used the international short-form of the Positive and Negative Affect Schedule (I-PANAS-SF)~\cite{thompson-07-panas}. Afterward, we asked participants to list three actions they would take after receiving the notification (\ref{app:part1:q1}), how concerned they would feel (\ref{app:part1:q2}--\ref{app:part1:q3}), and their priority of taking action (\ref{app:part1:q4}--\ref{app:part1:q5}).
\item \textit{Understanding:} Questions \ref{app:part1:q6}--\ref{app:part1:q9} were asked to investigate if participants understand what the notification is telling them, why they received it, and whether it appropriately guides them in resolving the situation (e.\,g., changing their password).
\item \textit{Expectation:} Next, we intended to learn what participants think would be the consequence of ignoring the notification (\ref{app:part1:q10}--\ref{app:part1:q11}) and if they expect real companies to send similar emails (\ref{app:part1:q12}--\ref{app:part1:q13}). This set of questions also contained an attention check \ref{app:part1:ac}.
\item \textit{Prior Experience:} If participants described having received similar login notifications in the past (\ref{app:part1:q14}), we asked if and why they decided to read them (\ref{app:part1:q15}--\ref{app:part1:q16}) and situations where such notifications helped them to identify an unrecognized login (\ref{app:part1:q17}--\ref{app:part1:q18}).
\item \textit{Demography:} Next, we asked for demographics, including their age, gender identity, education, and any technical background (\ref{app:part1:d1}--\ref{app:part1:d4}).
\item \textit{Feedback:} At the end of the study, we asked participants if they participated honestly and whether they wanted to share any (optional) feedback regarding the study.
\end{enumerate}
\subsection{Treatments}\label{sec:treatments}
We used two treatments to cover the cases that can trigger a notification: a \textit{Legit} and a \textit{Malicious} login. While a legit login is initiated by the user (or anyone who was intentionally given access), a malicious login is initiated by an attacker who got hold of the password, e.\,g., via a breach.
\textbf{Legit} ($n=110$): To mimic a legit login, participants in this treatment should imagine that they recently logged into their AcmeCo account. The login details for the notification were obtained from the current session, i.\,e., their operating system and browser (based on the HTTP User-Agent string), as well as state and country derived from the IP address.
\textbf{Malicious} ($n=110$): Participants in this treatment were told to imagine they had not logged into their AcmeCo account for a while and were all shown a notification for a login from California using Chrome on Windows. These login details were selected as we expect an attacker to know the victim resides in the USA. Logging in from the state with the highest population~\cite{us-22-population}, as well as the browser and operating system with the highest market share~\cite{statcounter-22-browser-market, statcounter-22-os-market} minimizes the risk of being detected.
We consider more sophisticated attacks~\cite{doerfler-19-login-challenges, mirian-19-2fa-hacking} that involve relaying login details and spoofing the user's sign-in information out of scope, as it makes recognizing such logins very difficult.
\subsection{Recruitment and Demographics}
For the baseline study, we recruited $n=220$ participants, 110 for each treatment.
We required participants to be over the age of 18 and reside in the U.S. to fulfill the assumption of the attacker in the malicious treatment.
We did not exclude any participants as all of them passed the attention check (\ref{app:part1:ac}) and indicated to have answered honestly; we also did not identify any inconsistencies.
The baseline study was compensated with \$2.50~USD and took participants, on average, 12~minutes to complete.
\tableDemoBaseline
The demographics for the baseline study are depicted in Table~\ref{tab:demoOne}.
Participants split equally between male- and female-identifying participants; seven identified as non-binary, and four preferred not to disclose their gender.
In terms of age and education, 49\% of the participants were younger than 35~years, and 50\% had a Bachelor's or Master's degree.
Lastly, 73\% of the participants stated not to have a technical background.
\subsection{Limitations}
Although we carefully considered various aspects of the study, it has its limitations which we will outline in the following.
First, we conducted an online study and asked participants to describe their behavior, opinion, and perception based on a fictitious scenario.
Hence, we cannot determine if this reflects reality even if none of the participants indicated to have answered dishonestly at the end of the study.
To draw our conclusion based on measured instead of intended behavior and experienced situations, we only used this study as a preliminary and conducted a subsequent study to delve further into the topic.
Second, the baseline study's sample was comprised of younger and higher-educated participants, which is typical for Prolific or other crowdsourcing platforms.
We note that results may change for different populations, which should be investigated in succeeding studies.
Third, results may suffer from self-report biases (e.\,g., social desirability).
To mitigate this, we did not explain that this was a study about usability or security.
Fourth, like many human-subject studies, there is the potential for a bias in question wording.
To circumvent this, we piloted the study and tried to keep the questions short and clear.
The full survey instrument can be found in Appendix~\ref{app:part1}.
Fifth, we only recruited US-based participants, which can have culture-based influences on the results.
\subsection{Ethical Considerations}\label{sec:baseline-method:ethics}
When the baseline study was conducted, none of the authors worked at an institution with an Institutional Review Board~(IRB) that could oversee it.
We followed the principles of the Menlo~Report~\cite{dhs-12-menlo-report} ``Respect for Persons,'' ``Beneficence,'' and ``Justice.''
We designed the study to minimize any potential harm while maximizing benefits.
At the beginning of the study, we described the study procedure carefully, including all risks, and asked participants for their consent.
Participants could withdraw from the study at any point without risking losing their compensation.
All collected data were stored and processed in accordance with the General Data Protection Regulation (GDPR).
\section{Baseline Study: Results}\label{sec:baseline-results}
Next, we present the results of our baseline study focusing on our three research questions (see Section~\ref{sec:intro}).
The coding of qualitative answers was done by two members of the research team, who started by separately coding 10\% of the answers.
Afterward, they agreed on a joint code book and used it to code the remaining 90\%.
The agreement between the two coders was high ($\kappa = 0.82$).
When applying statistics, we first checked for normality using D'Agostino's $K^{2}$ test, followed by Levene's test to check the variance.
Based on the outcome, we applied the corresponding test for significance using Bonferroni correction for all pairwise tests.
When quoting individual participants, e.\,g., {\color{PineGreen}L}61-{\color{RoyalBlue}N}, one can derive their treatment (\textit{\textbf{\color{PineGreen}L}egit} or \textit{\textbf{\color{PineGreen}M}alicious}) and password change behavior (\textit{Would \textbf{\color{RoyalBlue}N}ot Change} or \textit{Would \textbf{\color{RoyalBlue}C}hange}).
\subsection{RQ1: Comprehension \& Reaction}\label{sec:study1-comprehension-and-reaction}
\textbf{Comprehension.}
Questions~\ref{app:part1:q6} and \ref{app:part1:q7} asked participants to describe what the notification is telling them and why they may have received it.
On the one hand, all 220~participants correctly described that a new login happened, yet, 63 (29\%), 40 in the legit and 23 in the malicious group, assumed that the login must have been abnormal.
Some respondents speculated that the device (38;~17\%) or the location is unknown (15;~7\%).
Missing cookies, e.\,g., when using private browsing, were brought up by 3~participants, and only 1~person mentioned an unusual login time as a reason.
However, all of those deviations may trigger a notification, and each service can implement its own logic.
We further discuss this discrepancy between users' understanding and the actual implementation in Section~\ref{sec:discussion}.
\noindent
\figStudyOneBreakdown
\textbf{Reaction Legit.}\label{sec:study1-comprehension-and-reaction:claimed-to-change-pw}
The legit group splits equally between participants who
would ignore the notification (56; 51\%) and those who
would change their password (54;~49\%) (\ref{app:part1:q1}).
The majority of participants who would ignore it argued that they \legendboxBlack{study1_was_me}{A} definitely (42;~75\%) or \legendboxBlack{study1_probably_was_me}{E} probably (6;~11\%) received it because they logged in themselves:
\begin{myquote}
``\emph{I live in Texas, use Chrome on Windows, and have just logged into this account. It's probably me.}''~(L61-N)
\end{myquote}
Another 8~participants (14\%) thought it was them causing the notification, but also explicitly mentioned \legendboxBlack{study1_fatigue}{C}~fatigue:
\begin{myquote}
``\emph{Delete the email because I know I am the person who signed in,
[...] I unsubscribe from Acme entirely because I am tired of receiving these emails constantly.}''~(L80-N)
\end{myquote}
Of participants who said they would change their password (unnecessarily), 40 (74\%) were \legendboxBlack{study1_unsure}{D} unsure whether it was them logging in. They said they would change the password to eliminate the possibility of someone having access:
\begin{myquote}
``\emph{It looks like someone else logged in. [...] Even if I'm not sure, might do this just in case.}''~(L87-C)
\end{myquote}
The remaining \legendboxBlack{someone_else}{B} 14 (26\%) do not describe being uncertain that someone else knows their password:
\begin{myquote}
``\emph{I would think my password was compromised.}''~(L56-C)
\end{myquote}
Taking action (see Figure~\ref{fig:LikertStudyOneSectionOne}) has a significantly higher priority for participants who said that they would change their password than for those who do not ($\chi^{2}(2)= 31.3, p < 0.001$): 39, i.\,e., 72\% describe the priority to be \emph{high} or \emph{very high} which aligns with their assumption of being compromised.
Likewise, their agreement to \ref{app:part1:q10} saying ``that ignoring this email from AcmeCo would have consequences'' is also significantly higher ($\chi^{2}(2)= 29.0, p < 0.001$): 67\% \emph{agree} or \emph{strongly agree} compared to only 25\% for participants who stated they would ignore the notification.
Generally, the ratings of participants who reported that they would change their password in the legit group are more similar to those participants in the malicious group who reported that they would change their password, which highlights the strong influence of a misinterpreted notification.
\figLikertStudyOneSectionOne
\textbf{Reaction Malicious.} In the malicious group, the majority (97; 88\%) described that they would change their password, and 13 (12\%) said they would ignore it. Of the prior, \legendboxBlack{study1_someone_else}{B} 73 (75\%) justified their reaction by saying that someone else just logged into their account:
\begin{myquote}
``\emph{Someone else got into my account, so I need to change the password right away.}''~(M56-C)
\end{myquote}
Again, there are also \legendboxBlack{study1_unsure}{D} participants who expressed some form of uncertainty (24;~25\%), yet, in contrast to the legit group, they mostly questioned the legitimacy of the email:
\begin{myquote}
``\emph{I would hover over the email to ensure it wasn't a phishing email [...] after ensuring it was legit and safe, I would change my password}''~(M87-C)
\end{myquote}
Of the 13 participants in the malicious group who said they would ignore it, \legendboxBlack{study1_was_me}{A} 11 described that it must have been them logging in, misinterpreting the situation:
\begin{myquote}
``\emph{It is safe to assume that my account is not compromised. [\dots] It is good to see that my safety is a concern.}''~(M92-N)
\end{myquote}
Only \legendboxBlack{study1_other}{F} 2~participants argued differently: M25-N mentions phishing, similar to many other participants, but would ``send the email to spam'' and not do anything else. M51-N, on the other hand, would ``ignore it. It is my business, not theirs.''
Similar to the legit treatment, the intend to take action in response to the notification was of significantly higher priority for participants who supposed to be compromised than those who did not ($\chi^{2}(2)= 23.4, p < 0.001$).
Likewise, of those who intended to change their password, 87\% \emph{agreed} or \emph{strongly agreed} that ignoring the notification would have consequences compared to only 23\% among those who said they would not change it ($\chi^{2}(2)= 24.9, p < 0.001$).
\textit{\textbf{Summary.}} While users understand that the notification informs them about a new login, their comprehension of potential triggers deviates from what can be actual reasons.
We find a tendency to misinterpret the notification: participants either stated they would unnecessarily change the password (\textit{Legit}) or they would mistakenly ignore the notification (\textit{Malicious}). Note this study is limited to self-reported, hypothetical reactions. Thus, we further investigate these findings in the subsequent study (see Section~\ref{sec:measurment-method}).
\subsection{RQ2: Decision-Making \& Execution}
\label{sec:study1-decision-and-execution}
\textbf{Decision-Making.}
To identify the type of information (\textit{location}, \textit{date}, \textit{device}) participants based their decision on, we specifically checked the open responses.
Most participants (122; 55\%) consider the device as a factor (explicitly the browser or operating system).
However, percentages vary noticeably between the 4~groups.
While 79\% in the \textit{Legit (Would Not Change)} group mentioned the device when describing how they made their decision, only 67\% did in the \textit{Legit (Would Change)} group.
Participants in the malicious treatment mentioned the device less, 46\% of the time by those who would not change their password and only 37\% by those who stated they would.
Although one might think that the \textit{location} is the easiest, hence, most popular factor to put into context, it is only the second most frequent one mentioned by 87~participants across all groups (40\%).
Again, participants in the legit treatment who reported they would not change their password included it most often (31; 55\%), followed by \textit{Legit (Would Change)} (21; 39\%).
In the malicious treatment, the differences are only marginal: 31\% vs. 32\% when contrasting the \textit{Would Not Change} and \textit{Would Change} groups.
Lastly, 49 participants (22\%) mentioned the \textit{date} and/or \textit{time} in their responses.
When comparing the groups, ratios range from 10\% for the \textit{Malicious (Would Change)}
group to 43\% for the \textit{Legit (Would Not Change)} group, which fits the overall picture: participants in the legit group who expressed not to change their password refer to the login factors the most, followed by the \textit{Legit (Would Change)} and the \textit{Malicious (Would Not Change)} group.
The participants who refer to any of the 3~types of information the least are those in the \textit{Malicious (Would Change)} group.
A second view on this is given by \ref{app:part1:q18}, where we asked participants if they ever had a real-world situation in which they learned about an unrecognized login through such a notification.
Of those who did (134; 61\%), 51\% described that the \textit{location} was the determining factor.
Only 13\% say that the stated information about the \textit{device} helped them, and only twice it was the \textit{time}.
\textbf{Execution.}
Question~\ref{app:part1:q8} asked participants if the notification explained how to resolve the situation. The majority across all groups \textit{agrees} or \textit{strongly agrees}, ranging from 81\% (\textit{Legit (Would Change)}) to 89\% (\textit{Legit (Would Not Change)}).
When explaining their answer (\ref{app:part1:q9}), most participants (61\%; 133) focus on the information on how to resolve the situation for a malicious login, even if they do not intend to change their password:
\begin{myquote}
``\emph{It gave me steps to take if this was not me
}''~(L41-N)
\end{myquote}
Another 25\% (53) explicitly mention both outcomes:
\begin{myquote}
``\emph{It explains to ignore the email if it was me, and change my password if it wasn't.}''~(L101-N)
\end{myquote}
Participants who criticize the notification report that they miss information in some form (15\%; 33):
\begin{myquote}
``\emph{Changing my password is the very first line of defense, but I wish there was more information about what steps to take, or further assistance from AcmeCo.}''~(M26-C)
\end{myquote}
This highlights an important aspect, as account remediation usually consists of more steps than just changing the password.
Related work similarly observed a lack of advice provided by websites~\cite{neil-21-acc-remediation-adv, walsh-21-intercultural-analysis}. In Section~\ref{sec:discussion}, we further discuss the risk that arises from this lack of information and the opportunity of the moment.
Lastly, 32 participants referred to the \textit{change your password} link, which is provided in the notification. On the one hand, 22 (10\%) noted it positively:
\begin{myquote}
``\emph{The steps to take were very clear and easy to follow. It also provided a link to make it easy to change my password if I needed to.}''~(L42-C)
\end{myquote}
In contrast, 10 participants (5\%) mention the link negatively, saying that they do not trust it and describe it as bad practice:
\begin{myquote}
``\emph{It shouldn't be suggesting that I click the link in the email as this is a trick often used by scammers.}''~(M35-C)
\end{myquote}
Both positions provide reasonable arguments which come down to a weighting of usability and security. We further contrast both perspectives in Section~\ref{sec:discussion}.
\textit{\textbf{Summary.}} Based on the open-ended questions, the \textit{device} and the \textit{location} appear to be influential types of information when participants make their decision.
The \textit{Legit (Would Not Change)} group generally refers to login information the most, \textit{Malicious (Would Change)} participants the least.
As we identified differences between responses regarding the study and real-world experiences, we will employ the second study to investigate to what extent the different types drive their decision processes.
In terms of the execution, most participants, irrespective of the treatment and the described action, agree that the notification provides them with sufficient information.
Participants who commented negatively mentioned a lack of information or disliked the \textit{change your password} link.
\figLikertStudyOneSectionThree
\subsection{RQ3: Perception \& Expectation}
\label{sec:study1-perception-and-expectation}
\textbf{Perception.}
Using the PANAS, we investigated how the notifications make participants feel and also asked them about their concern (\ref{app:part1:q2}).
Participants in the \textit{Malicious (Would Change)} group feel the most positive, averaging 15.6 (SD:~3.2), followed by \textit{Legit (Would Change)} with a mean of 14.6 (SD:~4.2).
\textit{Malicious (Would Not Change)}, 13.0 (SD:~4.6), and \textit{Legit (Would Not Change)}, 11.8 (SD:~3.9), rank the lowest.
Differences within the treatments, i.\,e., \textit{Would Not Change} vs. \textit{Would Change}, are also significant (\textit{Legit}: $F_{55,53}=13.0, p<0.05$, \textit{Malicious}: $F_{96,12}=42.6, p<0.001$), which shows that participants feel more positive if they do something in reaction to the notification.
This may explain why participants (\textit{Legit}) would tend to change their password even if it may not be necessary.
The situation for the negative \emph{affect} is more distinct. The mean for participants in the malicious group who described to change their password, 12.7 (SD: 4.1), is significantly higher than for all other groups ($p<0.001$).
This stark contrast may be explained by the fact that 75\% of them report being hacked in the scenario, while only 26\% in the \textit{Legit (Would Change)} have a similar assumption and all others just try to ``better be safe than sorry.''
Scores of the \textit{Legit (Would Not Change)}, 6.8 (SD: 3.0), and \textit{Malicious (Would Not Change)} group, 6.7 (SD: 2.3), are significantly lower, even compared to \textit{Legit (Would Change)}, $\chi^{2}(2)=21.7, p<0.001$ and $\chi^{2}(2)=7.94, p<0.05$, respectively.
This aligns with the participant's understanding of the notification as information about their own login.
Participants' responses regarding the concern (\ref{app:part1:q2}--\ref{app:part1:q3}) match the previous observations. As can be seen in Figure~\ref{fig:LikertStudyOneSectionThree}, the \textit{Malicious (Would Change)} group has the highest concern level, with 72\% being \textit{moderately} or \textit{extremely concerned}.
This is significantly higher compared to \textit{Legit (Would Change)} ($\chi^{2}(2)=1702.0, p<0.001$), where 31\% are \textit{moderately}, and 20\% are \textit{extremely concerned}.
The lowest concern is reported by participants who described not to change their password; 81\% (\textit{Legit (Would Not Change)}) and 69\% (\textit{Malicious (Would Not Change)}) are only \textit{slightly} or \textit{not at all concerned}.
\textbf{Expectation.}
In response to \ref{app:part1:q12}, nearly all participants \textit{agree} or \textit{strongly agree} that login notifications should be sent by real companies.
They feel that the notifications \textit{protect accounts} (111; 51\%),
\textit{alert customers} (95; 43\%),
and show that companies are invested in security (15; 7\%).
The only negative aspect mentioned was \textit{annoyance}:
\begin{myquote}
``\emph{People who log into the site a lot are going to get spammed. It would be better if they only sent you a notification when they detect a strange login.
}''~(L32-N)
\end{myquote}
Still, this is only reported by 10 participants (5\%). Likewise, only 9\% (20) report to receive such notifications \textit{regularly} (\ref{app:part1:q14}). More than three times as many (74; 33\%) said they get them \textit{many times}, and the largest portion (93; 42\%) described to receive them \textit{occasionally}. Similarly, 63\% say they read them \textit{always}, 17\% \textit{often}, 13\% \textit{sometimes}. Hence, there appears to be some decline which could be explained by participants differentiating between \textit{expected} and a \textit{unexpected} notifications:
\begin{myquote}
``\emph{If I
signed in from a second device and then received the email, I knew it was me, so I would
ignore it. But
when I didn't sign in, I would pay attention and read the email.}''~(L16-N)
\end{myquote}
\textit{\textbf{Summary.}}
Participants who would change their password have a higher positive \emph{affect}, which may be explained by them securing their accounts.
Yet, those that assumed to be hacked, also felt the most negative.
If participants reported they would not change their password, their reaction was significantly more indifferent.
We use Study~2 to observe participants' feelings when confronted with a real notification.
More than 90\% of the participants expect services to send login notifications, leaving them feel protected.
They claimed to always read them or at least when the notification arrives unexpectedly.
About 10\% reported being annoyed by such notifications.
\section{Measurement Study: Method}\label{sec:measurment-method}
In the second study, we expand on the self-reported results from our baseline study by measuring user behavior.
The following outlines the study protocol, treatments, recruitment, ethical considerations, and limitations.
\subsection{Study Protocol}
In contrast to Study~1, participants in this study received a notification for an account they created.
To resemble a real-world setting, the protocol had to fulfill four criteria: \begin{enumerate*}
\item a \emph{real account} gets created
\item participants are \emph{unaware} that the study is about login notifications
\item participants receive the notification in their \emph{personal email account}
\item reactions to login notifications are \emph{measurable}.
\end{enumerate*}
To achieve this, we invited participants to take part in a multi-stage study about changes in the cognitive ability of mental rotation over time~\cite{shepard-71-mental-rotation, vandenberg-78-mental-rotation}.
Framing the study this way allowed us to inform people about the length of the commitment without revealing our interest and justified the necessity of creating an account.
The task also was a strong cognitive distractor that prevented participants from drawing too much attention to the authentication task.
Similar to Study~1, we used two treatments:
The legit group ($n=110$) received a notification only after they logged in themselves.
The location, date, and device information in the notification were derived from the metadata of their login.
The malicious group ($n=119$) received a notification unexpectedly at a time when they had not interacted with the account for multiple days.
This resembled a login attempt by a malicious actor from ``California, USA'' using ``Chrome on Windows.''
Location and device were selected to have the highest statistical chance of matching any user in our U.S.-based sample~\cite{us-22-population, statcounter-22-browser-market, statcounter-22-os-market}.
We did not allow mobile devices.
Next, we detail the study procedure.
\noindent \textbf{Stage~1:}\hspace{.5em} The study started by explaining the mental rotation test.
To ensure participants would regularly check their email and understand the value of the account, after giving their consent, they saw a privacy notice, which highlighted the importance of the account as it would be used to store the study data, name, and email address.
It also explained that the email would be used to send invitations to subsequent stages, and the compensation in form of Amazon gift cards.
After the account creation, participants solved 5~mental rotation tests and provided demographic information (\ref{app:part2:d1}--\ref{app:part2:d4}).
At the end, participants in the legit treatment were informed that invitations to Stage~2 would be sent in approx. 7~days; in the malicious group, the note said 14~days.
\noindent \textbf{Stage~2:}\hspace{.5em} After 7~days, participants in the legit group received an email inviting them to conduct another mental rotation test.
To do so, they had to log into their account, which triggered a login notification.
Participants in the malicious group expected their next email after 14~days.
However, to imitate a malicious login, we sent them an (unexpected) login notification filled with our statistical sign-in data 7~days after they completed the first stage.
\noindent \textbf{Stage~3:}\hspace{.5em} For the legit group, invitations to the final Stage~3 were sent 48~hours after they completed Stage~2; in the malicious group, 48~hours after they received a notification for a login they did not initiate.
We chose this time frame to give participants enough time to react to the notification.
After logging into Stage~3, participants were debriefed and told about the actual purpose of the study.
This was followed by our questionnaire (see Appendix~\ref{app:part2}).
From then on, the notification we sent to the participant was shown on the left side of their screen for reference.
\begin{enumerate}[noitemsep,nolistsep,leftmargin=1.2em]
\item \textit{Email:} First, we asked participants if they remember receiving the notification (\ref{app:part2:s1}); if not, they were forwarded to a different section (see Appendix~\ref{app:part2}). Participants who triggered the tracking pixel in the notification or who changed their password skipped this question.
\item \textit{I-PANAS-SF:} To learn about the feelings and emotions in reaction to the notification, we again utilized the Positive and Negative Affect Schedule (I-PANAS-SF)~\cite{thompson-07-panas}.
\item \textit{Reaction:} Next, we asked how thoroughly participants read the notification (\ref{app:part2:q1}) and how and why they chose to react to it (\ref{app:part2:q2a}--\ref{app:part2:q3a}).
Participants who changed their password were specifically asked about any other actions (\ref{app:part2:q2b}--\ref{app:part2:q3b}).
\item \textit{Content \& Design:} To better understand the reactions, \ref{app:part2:q4} asked about influencing factors like metadata, content, and design.
\ref{app:part2:q5} specifically asked about the helpfulness of the account name, location, date, and device.
\item \textit{Time \& Location:} \ref{app:part2:q6}--\ref{app:part2:q10} investigated the time when and location where the notification was read.
With \ref{app:part2:q7}, we verified if the location, which had been derived automatically, was actually accurate or could have led to confusion, and \ref{app:part2:ac2} was an attention check.
\item \textit{Comprehension \& Expectation:} With \ref{app:part2:q11}, we captured if participants understood why they received the notification.
\ref{app:part2:q12} and \ref{app:part2:q13} asked participants when they expect real companies to send notifications.
\item \textit{Prior Experience:} We concluded with three questions covering negative experiences with security incidents (\ref{app:part2:q14}), as well as their opinion on regular (\ref{app:part2:q15}) and event-driven password changes (\ref{app:part2:q16}).
\end{enumerate}
\subsection{Recruitment \& Demographics}
We recruited 625~participants for Stage~1.
After filtering 12~participants who failed the attention check (\ref{app:part2:ac1}).
At the end of Stage~3, we had 252~completions.
However, we removed 23 as they provided unrelated answers or failed the second attention check (\ref{app:part2:ac2}).
The final number of participants was $n=229$.
The relatively high number of dropouts is not owed to the login notification, as the participants already left the study before any notification was sent (before the login at Stage~2).
Instead, it is primarily owed to the fact that it was a multi-stage study, and dropout rates like ours are within our panel provider's expected range.
Stage~1 took, on average, 2.5~minutes and was compensated with \$3.00~USD.
Stage~3 took, on average, 6~minutes and was compensated with \$4.00~USD.
Participants in the legit group received an additional \$1.00~USD for the completion of Stage~2, which took 2~minutes on average.
\tableDemoMeasurement
Table~\ref{tab:demoTwo} shows the participants' demographics. While we observe a shift towards male-identifying participants (65\%), the age distribution is diverse, ranging from 14\% to 21\% for all age groups between 25 and 74.
Most participants had a high school (33\%), Bachelor's (26\%), or trade degree (23\%) and did not have a technical background (82\%).
\subsection{Ethical Considerations}\label{sec:measurment-method:ethics}
As with our baseline study, at the time we conducted the measurement study, none of the authors worked at an institution with an IRB.
However, as discussed in Section~\ref{sec:baseline-method:ethics}, we carefully followed the guidelines provided in the Menlo~Report, including a risk-benefit evaluation, following the legal requirements of the GDPR, and discussed and tested the protocol with peers familiar with conducting user studies.
The study included deception and sent a login notification to participants' personal email accounts, which could have caused more anxiety than just imagining to have received a login notification.
We actively tried to avoid this situation by conducting our baseline study first.
However, motivated by its results, i.\,e., 49\% stated they would (unnecessarily) change their password, we decided to verify this potentially user-burdening behavior by conducting a second study.
To protect participants from unnecessary risks, we implemented several safeguards:
\begin{enumerate*}[label=\roman*)]
\item Our panel provider offered the study only to participants that agreed to studies that might involve deception.
\item The affected spatial reasoning account had no subjective value to the participants and only allowed to access the name and email address.
\item All participants have been debriefed (also the ones that decided to withdraw early or drop out). In particular, we told them about the true purpose of the study, and in case they belonged to the \emph{malicious} treatment that ``This sign-in did not take place; at no time was your account at risk,'' and asked them whether they prefer to leave the study early (while being fully compensated), which nobody did.
\item We provided an optional contact address and feedback form that we closely monitored (we have not received any complaints).
\item We shared a website (also accessible from outside of the study), that participants could visit and share with their friends to learn more about login notifications and related account security measures.
\item We created a distinct email account for sending the notifications that applied all state-of-the-art email security features, which can prevent email spoofing attacks. We also allowed participants to reply to the notification and ask for assistance. Finally, all email addresses were only stored encrypted, separated from the study responses, and were deleted after the study in accordance with GDPR.
\end{enumerate*}
\subsection{Limitations}\label{sec:measurment-method:limitations}
In addition to the limitations from Study~1, we relied on a controllable artificial account setting for this study which might lack ecological validity.
However, only 7~participants mention the non-real-world setting as a reason for not reacting to the notification. We expect more participants to change their password if the notification was sent for an account with a higher subjective value.
\section{Measurement Study: Results}\label{sec:measurment-results}
Next, we present the results of the measurement study.
We have applied the same structure, coding methods, and statistical testing as with out baseline study (see Section~\ref{sec:baseline-results}).
\subsection{RQ1: Comprehension \& Reaction}
\label{sec:study2-comprehension-and-reaction}
\textbf{Comprehension.}
When asked to describe why they have received the notification (\ref{app:part2:q11}), 85\% (93) of the participants in the legit and 79\% (94) in the malicious treatment realized that a new login happened to their account. Very few participants who gave a different explanation believed it was a phishing attempt (3; 1\%), most simply did not understand what has happened at all (39; 17\%):
\begin{myquote}
``\emph{I had no idea, which is why I deleted it.}''~(M93-N)
\end{myquote}
Those in the legit treatment who mapped the notification to a new login usually perceived it as a simple info email (42; 38\%), followed by those who saw it as a prompt to review the login (28, 26\%). Fewer responses (15; 13\%) explicitly mention that the login must have been abnormal. In the malicious treatment, most participants, who understood that a new login happened, described that they were (potentially) compromised (46; 36\%). Another 19\% (22) perceived it as an informative but non-critical email. The remainder (13; 11\% each) either mentioned that the system rated the login as unusual or wants them to review the login.
\figStudyTwoBreakdown
In contrast to the first study, we observed a lower comprehension of what might have caused the notification, especially in the malicious group. One explanation might be the temporal connection between logging in and receiving the notification. From \ref{app:part2:q6}, we know that about two-thirds read it immediately, most of the others within one or multiple hours. Hence, participants in the legit treatment had indeed a connection, and their understanding was substantially better. Similarly, we did not observe differences between treatments in the baseline study, where both had identical circumstances.
This influence of contextual factors was already observed by related work on security warning design~\cite{reeder-18-warning-reaction, gorski-20-pd-for-crypto-apis} and could be achieved by including a \emph{Why Notification} section. Some websites already do (see Section~\ref{sec:background}), and we will further elaborate on this in the discussion.
\textbf{Reaction General.}
Out of the total 229 participants,
48 participants, 23 in the legit and 25 in the malicious treatment, \legendboxBlack{dont_remember}{F}~cannot remember the notification. Still, 26 of them triggered the tracking pixel, so they must have at least opened the notification.
Among the large majority of participants who saw the login notification (181; 79\%), it was very rare that they completely ignored its content.
In response to \ref{app:part2:q1}, just 6\% said that they only read the subject.
About 90\% read the notification completely or at least skimmed the body.
\figHelpfullness
\textbf{Reaction Legit.}
Unlike the baseline study suggested, where 49\% of participants in legit treatment said they would change their password (see Section~\ref{sec:study1-comprehension-and-reaction:claimed-to-change-pw}), none in the measurement study actually did.
As shown in Figure~\ref{fig:StudyTwoBreakdown}, the majority of participants (60; 55\%) explained their reaction by saying \legendboxBlack{was_me}{A} it was their own login. Another 12, i.\,e., 11\%, described it as a \legendboxBlack{spontaneous}{E} \textit{spontaneous} reaction, e.\,g., M42-N: ``I just didn't think much of it.''
Previously, we saw that some participants do not understand what the notification is saying, which was the driving reason for \legendboxBlack{didnt_understand}{H}~6\% (6) to ignore it.
Finally, we recorded themes aligning with the findings from the baseline study with participants who were \legendboxBlack{suspicious}{I} \textit{suspicious} about the legitimacy of the notification (4; 4\%), felt \legendboxBlack{fatigue}{C} \textit{fatigued} (3; 3\%), or \legendboxBlack{feel_protected}{G} \textit{protected} (3; 3\%).
\textbf{Reaction Malicious.}
In the malicious group, only 26 of the 119 participants, i.\,e., 22\%, changed their password; all of them correctly saying \legendboxBlack{someone_else}{B} it was not them but someone else logging in.
The reasons given by the other 78\% (93) for not changing their password have mostly been given by participants in the legit treatment: \legendboxBlack{spontaneous}{E} \textit{spontaneous} reaction (15; 13\%), notification looked \legendboxBlack{suspicious}{I} \textit{suspicious} (14; 12\%), or was \legendboxBlack{didnt_understand}{H} \textit{not understood} (8; 7\%), \legendboxBlack{fatigue}{C} being \textit{fatigued} (6; 5\%) or \legendboxBlack{unsure}{D} \textit{unsure} how to react (6; 5\%).
Finally, there are two justifications that are owed to the study design: participants describing they \legendboxBlack{was_me}{A} logged in themselves although they did not (11; 9\%), likely an example of social desirability, and those who assigned a low value to the account (7; 6\%):
\begin{myquote}
``\emph{This account has no value, it was not a streaming or banking account or amazon account}''~(M74-N)
\end{myquote}
This justification can be reasonable, but users need to keep in mind that an attacker can also target other accounts that verbatim or partially reused the compromised password~\cite{pal-19-personal-psm}.
\textit{\textbf{Summary.}}
About 80\% saw the notification. Participants in the legit treatment who triggered it themselves understood what it was telling them and reacted accordingly. In the malicious treatment where participants did not have this context, only 22\% changed their password, and they had more difficulties explaining the circumstances. Hence, the number of password changes in the malicious treatment is substantially lower than expected, while we could not prove the tendency to unnecessary password changes.
\figEffect
\subsection{RQ2: Decision-Making \& Execution}
\label{sec:study2-decision-and-execution}
\vspace{-.5em}
Participants in the baseline study reported that the notification tells them how to resolve the situation, irrespective of their assigned treatment and described reaction. Hence, we did not further test the execution and focused on the decision-making process instead, as participants seem to struggle when determining whether it was them or not, especially in case of a malicious login.
\textbf{Helpfulness of Login Information.}
Foremost, we wanted to get more insights into the helpfulness of the displayed login information.
We know from the baseline study that information about the \textit{device} and \textit{location} are considered by participants when deciding how to react to a login notification.
In Figure~\ref{fig:helpfulnessInformation}, we can see that for those in the \textit{Legit} and \textit{Malicious (Change)}
group, all information is about equally helpful: 22--35\% find the different types \textit{moderately} and 38--62\% even \textit{extremely} helpful.
Participants in the \textit{Malicious (No Change)} group,
in contrast, appear to have a less distinct opinion as ratings are more equally distributed, ranging from 8--30\%.
A Kruskal-Wallis test also showed significant differences for all types of information when comparing \textit{Malicious (No Change)} to \textit{Legit} and \textit{Malicious (Change)}, respectively.
This uncertainty of participants in the \textit{Malicious (No Change)} group regarding the displayed information aligns with the previous section, where we found that those participants misattributed or did not understand the cause of the notification.
\textbf{Effect of Other Factors.}
In addition to the already-known influence of the login information, we were also interested in the effect of other hard and soft factors. Figure~\ref{fig:effectReaction} gives an overview.
Generally speaking, the content (e.g., provided information, instructions, wording) and prior experience in dealing with such notifications had the highest effect on participants' reactions, with 42\% expressing a \textit{moderate} or \textit{major effect} on average.
Followed by that is the metadata (e.g., sender, subject, time of arrival) with 29\%.
All other factors seemed to have a less distinct influence, with 18\% (appeared to be phishing) to 23\% (was expected) of the participants reporting a \textit{moderate} or \textit{major effect}.
When comparing the groups, \textit{Legit} is the one where most participants reported a factor having no effect. The \textit{Malicious (Change)} group, on the other hand, is the one where participants describe the strongest influences of the factors. Using a Kruskal-Wallis test with Bonferroni-correction for pairwise comparisons, we found that the metadata had a significantly higher effect for \textit{Malicious (Change)} participants
compared to \textit{Malicious (No Change)} participants ($\chi^{2}(2) = 6.65, p < 0.05$).
The same is true for the email content ($\chi^{2}(2) = 7.73, p < 0.05$).
These findings suggest that to nudge more users to change their password upon receiving potentially malicious login notifications, focusing on properly designing the content and metadata is vital.
\textbf{Influence of Negative Experiences.}
Overall, 30\% of participants described falling victim to a security breach within the last two years (\ref{app:part2:q14}).
In the malicious treatment, 42\% of those who changed their password reported prior negative experiences. Only 32\% of those who did not change their password said so.
The difference is not statistically significant, $\chi^{2}(2) = 2.61, p = 0.271$ but suggests that prior breach experience increases the likelihood of users changing their password upon receiving a notification.
\textit{\textbf{Summary.}}
Based on the first study, we assumed the device and location to be more helpful than the account name and the date. However, when compared side-by-side, we can conclude that all factors are equally essential. What we did observe is that the helpfulness of the information for the \textit{Malicious (No Change)} participants is significantly lower, which further explains the issues of this group when determining what happened.
Regarding other factors, the content of the notification, its metadata, and prior experience in dealing with it had the highest effect across all treatments. Negative experience tends to influence the reaction as well; other aspects appeared to be less crucial.
\subsection{RQ3: Perception \& Expectation}
\label{sec:study2-perception-and-expectation}
\textbf{Perception.}
In the baseline study, the PANAS revealed that participants who would change their password felt significantly more positive (M: 15.6, SD: 3.2) but also more negative (M: 12.7, SD: 4.1).
Participants who described ignoring the notification scored the lowest.
The second study backs up both findings. The average positive \emph{affect} of the \textit{Malicious (Change)}
group is 15.0 (SD: 4.5) but only 11.6 (SD: 5.0) and 12.6 (SD: 5.6) for the \textit{Malicious (No Change)}
and \textit{Legit}, respectively.
Using a Kruskal-Wallis test (Bonferroni corrected), we were also able to confirm the significance between the two malicious groups, $\chi^{2}(2) = 8.29, p < 0.05$. For the negative \emph{affect}, \textit{Malicious (Change)} averages 9.8 (SD: 4.1), \textit{Malicious (No Change)}
(M: 8.1, SD: 4.3), and \textit{Legit} (M: 5.7, SD: 1.6).
Again, Kruskal-Wallis was used yielding significance between both malicious groups and \textit{Legit} ($p < 0.01$); the comparison between the two malicious groups nearly did, $\chi^{2}(2) = 5.26, p = 0.0654$.
\textbf{Expectation.}
More than 90\% of the participants in the baseline study expect real-world services to send login notifications. Participants who expressed disagreement reported receiving them constantly and being annoyed. Moreover, the second study showed so far that there is a substantial number of participants who have not changed their password although they should, some of them mentioning that it was a spontaneous reaction which this fatigue may also explain. Hence, we used \ref{app:part2:q12} to further understand when users expect to receive login notifications.
A majority of participants (151; 66\%) expressed they want to receive notifications after suspicious account activity.
On average, 60\% want to be notified if a login takes place from a new device, 47\% for logins from a new location, 31\% if they have not logged in for a while, and 22\% for logins that take place at an unusual time of the day. Only 9\% want to receive a login for \textit{every} login, and even less (9, 3\%) do not want to receive login notifications at all.
\textit{\textbf{Summary.}}
We can confirm that participants who changed their password felt both more positively and negatively, probably because they assumed some form of compromise but also had a sense of achievement after preventing it by changing the password. The other groups had lower scores, aligning with them not expecting any harm.
The baseline study showed that participants expect services to send login notifications. With the new findings, we can further specify this by saying that participants want to be notified after suspicious logins, logins from new devices, and logins from new locations. Fewer participants expect to receive notifications based on temporal deviations.
\section{Discussion \& Recommendations}\label{sec:discussion}\vspace{-.5em}
Next, we discuss the takeaways of our analyses of 66~notifications and the main findings of our two subsequent user studies and give recommendations for service providers.
\subsection{Best Practice} \vspace{-.5em}
Overall, the baseline notification we developed achieved its primary goal: informing participants about a new login.
Yet, understanding the trigger and correctly assessing the situation remained an issue.
Below, we describe best practice based on what we learned for each of the components.
\textbf{Metadata.}
We found that the metadata is an influential factor, and 75\% of the participants paid attention to the subject.
Hence, in addition to the most important information, the subject should already provide context for deciding how to react.
Generic clauses, e.\,g., ``Your account has been logged into.'' as used by Tumblr, do not achieve this, which is why email subjects like ``New login to Twitter from \{browser\} on \{OS\}'' are preferable.
Similarly, websites should make use of the email sender's name so that recipients can quickly parse the information about the sender; in our initial analysis, we found 5 services that did not.
\textbf{What Happened.}
What triggers a notification, e.\,g., an ``unusual login'' is often unclear to participants.
Services could easily address this issue
by explaining what triggered the notification, yet, only 12\% of the evaluated emails currently provide examples of common triggers.
Explaining the circumstances would also help to create context, which is especially important when users receive unexpected notifications and struggle to assess the situation correctly.
\textbf{Login Information.}
We found that all types of information (account name, location, time, and device) have a positive influence.
IP~addresses are a technical detail, not mentioned by any of the participants when referring to prior experiences.
Tech-savvy users may argue
being able to derive information from it, yet, the shift to IPv6 mostly eliminates this argument. Thus, we dissuade displaying it.
For the location, one needs to keep in mind that IP-based localization can be inaccurate, and potential issues like the reporting of default locations~\cite{rawlinson-16-kansas-lawsuit} were already noted by prior work~\cite{markert-22-rba-admin}.
To circumvent this problem, some services add notes or an ``Approximate'' to the location.
Another approach would be to only report the location on a country or state level, as none of the participants mentioned cases where a login from a different city attracted their attention; it was always a different state, country, or continent.
When the date is reported, the month should be spelled out, as we found multiple examples where the DD/MM vs.\ MM/DD representation caused confusion.
Similarly, the login time should follow the user's local conventions to prevent confusion due to the 12/24-hour clock format.
Finally, it is important to include the timezone.
Concerningly, we found 26~services that did not report any timezone.
Including the browser and operating system proved to be helpful.
To fully benefit, services should avoid technical details, as most services already do for browsers, when they list them with just their name.
Negative examples are sign-ins on smartphones displayed using the model number instead of the device name, e.g., ``SM-S908B/DS'' instead of ``Samsung Galaxy S22.'' Operating systems are also usually reported with their full version number, e.g., ``iOS 16.0.1.''
The least recommendable example is the inclusion of the raw HTTP User-Agent string, which we found in 4~emails.
\textbf{Instructions.}
Notifications need to include instructions for both outcomes, i.\,e., legit and malicious logins. For the legit case, most services describe to ignore the message.
For malicious logins, the recommendation needs to prompt users to visit the website and change or reset the password.
Most services facilitate this by including a link which is a controversial practice.
Of course, it simplifies the process, and 10\% in our study explicitly appreciated the shortcut.
However, also 10\% of the participants were suspicious due to the presence of a link and did not change their password.
Hence, services should carefully evaluate including a link.
\textbf{Avoidable Pitfalls.}
Throughout our analysis of real-world notifications, we found numerous shortcomings that can make legit notifications appear suspicious.
We saw that assessing a notification as suspicious can make the difference between changing the password and ignoring the notification.
Luckily, all of the identified issues are easily avoidable, so services should make sure to thoroughly test their implementation.
Examples we found include:
incorrect, e.\,g., ``\Verb"<b>"If you did not do this\Verb"<b\>"''
or escaped HTML ``\Île-de-France'' instead of ``Île-de-France,''
and missing placeholders, e.\,g., ``Browser: N/A.''
\textbf{Privacy.}
The vast majority of services employed tracking technologies in their emails like tracking pixels, URL parameters, and third-party redirects.
The only exception to this are the 7~plaintext notifications we received from services like Cloudflare, GitHub, and Nintendo.
Some participants reported to be concerned about this and wondered ``what else they're tracking.''
\subsection{Expectations \& Fatigue}\vspace{-0.5em}
More than 90\% of the participants expect services to send login notifications, and finding a balance
between sending them too often and too rarely is crucial.
We found that participants want to be notified when a login takes place from a new device or location, but also if a login appears ``suspicious,'' which can be accomplished with advanced logic provided by risk-based algorithms~\cite{wiefling-19-rba-in-the-wild, wiefling-20-rba-evaluation}.
Time-related notifications (i.\,e., login after a long or at an unusual time) are less demanded.
For service providers, sending notifications more often than necessary, following a ``better safe than sorry''- mentality may be tempting.
Yet, for users, this leads to a well-documented phenomenon in the area of security warnings:
fatigue~\cite{egelman-08-phishing-warnings, sunshine-09-crying-wolf, akhawe-13-alice}.
This fatigue, which affects all services, is most likely caused by unnecessary login notifications, i.\,e., those that do not convey a real risk teach users that all notifications are unimportant.
The situation is further aggravated by services like Mozilla, Tumblr, Etsy, and others that send notifications for every single login.
\subsection{Phishing Warning \& Questionable Advice}\vspace{-0.5em}
Throughout both studies, about 15\% of the participants questioned the legitimacy of the notification or referred to it as phishing.
Related work showed how to best advice users~\cite{secuso-22-detect-phishing},
yet, most of the 10~notifications that include some information about phishing, do not follow it.
For example, questionable advice is given by Twitter (and three other major services) which suggests that the presence of a padlock icon will ``let you know a site is secure'' and that users should check for the presence of ``https://'' and ``\{domain\}'' in the hyperlink.
Similarly, Amazon suggests to better copy and paste the ``It wasn't me''-link into a browser, ``just to be safe.''
Spotify advises users to verify the email was sent from <EMAIL>,'' which is only expedient, if the email server and DNS are configured correctly.
Related work demonstrated how vulnerable this email ecosystem is~\cite{shen-21-email-spoofing}.
A good example is some of PayPal's advice~\cite{paypal-22-phishing} that explicitly mentions to ``not rely on the padlock symbol and the `s' in https.''
Interestingly, LinkedIn opted to add a security footer message~\cite{linkedin-22-security-footer} to their login notification that includes the name and profession of the affected user in order to authenticate official emails.
\section{Appendix}\label{sec:app}
\input{94-survey-instrument-baseline}
\clearpage
\input{95-survey-instrument-experiment}
\clearpage
\onecolumn
\subsection{Real-World Notifications: Features}\label{app:real-world:features}
\tableNotificationsGrantedAccess
\clearpage
\subsection{Real-World Notifications: Email Metadata}\label{app:real-world:emailmetadata}
\tableNotificationsSenderSubject
\ifextended
\input{96-codebook.tex}
\fi
\subsection{Baseline Study: Survey Instrument}\label{app:part1}
\setlist{nolistsep}
\newcommand{\mychoice}[1]{{$\circ$}~#1 \, }
\newcommand{\mymultchoice}[1]{{$\square$}~#1 \, }
\newcommand{\myspecial}[1]{{$\bowtie$}~#1 \, }
\footnotesize
\noindent\textbf{Scenario Description}
\noindent In the following survey, you will be asked to imagine that your name is \textbf{Jo Doe} (<EMAIL>). You have an online account with a major company called \textbf{AcmeCo} which you \textbf{regularly access with all the devices you own}. Imagine that \textbf{this account is important to you}, and that it is like other accounts you may have, such as for online shopping or social media.
\noindent\textbf{A Notification from AcmeCo} \\
Imagine, you \{\textbf{recently signed into} your AcmeCo account. / \textbf{have not signed into} your AcmeCo account for a while.\}\\
\noindent In your inbox, you see the following email.\\
\noindent\textbf{Notification}
\begin{center}\textit{{\color{teal} We displayed an interactive mockup email interface with our baseline notification shown in Figure~\ref{fig:baselineemail}.}}\end{center}
\noindent\textbf{I-PANAS-SF} \\
\noindent Now we would like to know \textbf{how you feel in reaction to the email} shown on the left. The list below consists of a number of words that describe different feelings and emotions. Read each item and then mark the appropriate answer on the list. Indicate to what extent you feel this way right now, that is, at the present moment.
\begin{table}[htbp]
\centering
\footnotesize
\setlength{\tabcolsep}{.4\tabcolsep}
\resizebox{\linewidth}{!}{\begin{tabular}{lccccc}
\toprule
& Very slightly \\
& or not at all & A little & Moderately & Quite a bit & Extremely \\
& (1) & (2) & (3) & (4) & (5) \\
\midrule
Upset & $\bigcirc$ & $\bigcirc$ & $\bigcirc$ & $\bigcirc$ & $\bigcirc$ \\
Hostile & $\bigcirc$ & $\bigcirc$ & $\bigcirc$ & $\bigcirc$ & $\bigcirc$ \\
Alert & $\bigcirc$ & $\bigcirc$ & $\bigcirc$ & $\bigcirc$ & $\bigcirc$ \\
Ashamed & $\bigcirc$ & $\bigcirc$ & $\bigcirc$ & $\bigcirc$ & $\bigcirc$ \\
Inspired & $\bigcirc$ & $\bigcirc$ & $\bigcirc$ & $\bigcirc$ & $\bigcirc$ \\
Nervous & $\bigcirc$ & $\bigcirc$ & $\bigcirc$ & $\bigcirc$ & $\bigcirc$ \\
Determined & $\bigcirc$ & $\bigcirc$ & $\bigcirc$ & $\bigcirc$ & $\bigcirc$ \\
Attentive & $\bigcirc$ & $\bigcirc$ & $\bigcirc$ & $\bigcirc$ & $\bigcirc$ \\
Afraid & $\bigcirc$ & $\bigcirc$ & $\bigcirc$ & $\bigcirc$ & $\bigcirc$ \\
Active & $\bigcirc$ & $\bigcirc$ & $\bigcirc$ & $\bigcirc$ & $\bigcirc$ \\
\bottomrule
\end{tabular}}
\end{table}
\noindent\textbf{Reaction}
\begin{enumerate}[leftmargin=3em, label=\textbf{BQ\arabic*},noitemsep]
\item \textbf{Please list 3 actions} you might take after receiving this email \textbf{and explain why} you might take those actions.
\begin{enumerate}[leftmargin=1em, label=\arabic*,noitemsep,nolistsep]
\item Action: \rule{3.5cm}{.1pt} \\ Explanation: \rule{3.5cm}{.1pt}
\item Action: \rule{3.5cm}{.1pt} \\ Explanation: \rule{3.5cm}{.1pt}
\item Action: \rule{3.5cm}{.1pt} \\ Explanation: \rule{3.5cm}{.1pt}
\end{enumerate}
\label{app:part1:q1}
\item I would \textbf{feel} \rule{1.25cm}{.1pt} \textbf{about receiving} this email from AcmeCo. \\
\mychoice{Not at all concerned}
\mychoice{Slightly concerned} \\
\mychoice{Somewhat concerned}
\mychoice{Moderately concerned} \\
\mychoice{Extremely concerned}
\label{app:part1:q2}
\item Why? \\
Answer: \rule{3.5cm}{.1pt}
\label{app:part1:q3}
\item For me, \textbf{taking action in response} to this email from AcmeCo \textbf{would~be} \\
\mychoice{Not a priority}
\mychoice{Low priority}
\mychoice{Medium priority}
\mychoice{High priority}
\mychoice{Very high priority}
\label{app:part1:q4}
\item Why? \\
Answer: \rule{3.5cm}{.1pt}
\label{app:part1:q5}
\end{enumerate}
\noindent\textbf{Understanding}
\begin{enumerate}[leftmargin=3em, label=\textbf{BQ\arabic*},noitemsep,nolistsep]
\setcounter{enumi}{5}
\item In your own words, please \textbf{describe what this email is telling} you. \\
Answer: \rule{3.5cm}{.1pt}
\label{app:part1:q6}
\item In your own words, please \textbf{describe all of the factors that may have caused you to receive this email}. \\
Answer: \rule{3.5cm}{.1pt}
\label{app:part1:q7}
\newpage
\item I feel that this email from AcmeCo explained to me \textbf{how to resolve the situation}. \\
\mychoice{Strongly disagree}
\mychoice{Disagree}
\mychoice{Neither agree or disagree} \\
\mychoice{Agree}
\mychoice{Strongly agree}
\label{app:part1:q8}
\item Why? \\
Answer: \rule{3.5cm}{.1pt}
\label{app:part1:q9}
\end{enumerate}
\noindent\textbf{Expectation}
\begin{enumerate}[leftmargin=3em, label=\textbf{BQ\arabic*},noitemsep,nolistsep]
\setcounter{enumi}{9}
\item I feel that \textbf{ignoring this email} from AcmeCo \textbf{would have consequences}. \\
\mychoice{Strongly disagree}
\mychoice{Disagree}
\mychoice{Neither agree or disagree} \\
\mychoice{Agree}
\mychoice{Strongly agree}
\label{app:part1:q10}
\item Why? \\
Answer: \rule{3.5cm}{.1pt}
\label{app:part1:q11}
\end{enumerate}
\begin{enumerate}[leftmargin=3em, label=\textbf{BAC},noitemsep,nolistsep]
\item Please select `Agree' as the \textbf{answer} to this question. \\
\mychoice{Strongly disagree}
\mychoice{Disagree}
\mychoice{Neither agree or disagree} \\
\mychoice{Agree}
\mychoice{Strongly agree}
\label{app:part1:ac}
\end{enumerate}
\begin{enumerate}[leftmargin=3em, label=\textbf{BQ\arabic*},noitemsep,nolistsep]
\setcounter{enumi}{11}
\item I think \textbf{real companies should send emails like this one} when necessary. \\
\mychoice{Strongly disagree}
\mychoice{Disagree}
\mychoice{Neither agree or disagree} \\
\mychoice{Agree}
\mychoice{Strongly agree}
\label{app:part1:q12}
\item Why? \\
Answer: \rule{3.5cm}{.1pt}
\label{app:part1:q13}
\end{enumerate}
\noindent\textbf{Prior Experience}
\begin{enumerate}[leftmargin=3em, label=\textbf{BQ\arabic*},noitemsep,nolistsep]
\setcounter{enumi}{13}
\item \textbf{I have received emails similar to this one} in the past. \\
\mychoice{Never}
\mychoice{A few times}
\mychoice{Occasionally}
\mychoice{Many times}
\mychoice{Regularly}
\label{app:part1:q14}
\end{enumerate}
\noindent \textit{{\color{teal}If participant stated to have received similar notifications in \ref{app:part1:q14}: }}
\noindent We will now ask you about \textbf{your personal experiences with similar emails} you have received in the past. \textbf{Based on these experiences}, please answer the questions below:
\begin{enumerate}[leftmargin=3em, label=\textbf{BQ\arabic*},noitemsep,nolistsep]
\setcounter{enumi}{14}
\item I \rule{1.25cm}{.1pt} read them. \\
\mychoice{Never}
\mychoice{Rarely}
\mychoice{Sometimes}
\mychoice{Often}
\mychoice{Always}
\label{app:part1:q15}
\item Why? \\
Answer: \rule{3.5cm}{.1pt}
\label{app:part1:q16}
\item Have you \textbf{ever received a similar email} that helped you to \textbf{learn about an unrecognized sign-in}? For example, a sign-in that concerned you. \\
\mychoice{Yes}
\mychoice{No}
\mychoice{Don't know}
\label{app:part1:q17}
\begin{center}\textit{{\color{teal}If participant answered `Yes' in \ref{app:part1:q17}: }}\end{center}
\item Referring to the situation where a similar email helped you to learn about an unrecognized sign-in. Please \textbf{describe the situation}, how you knew that something was wrong, and \textbf{how you solved it}. \\
Answer: \rule{3.5cm}{.1pt}
\label{app:part1:q18}
\end{enumerate}
\noindent\textbf{Demography}
\begin{enumerate}[leftmargin=3em, label=\textbf{BD\arabic*},noitemsep,nolistsep]
\item Select your age. \\
\mychoice{18--24}
\mychoice{25--34}
\mychoice{35--44}
\mychoice{45--54}
\mychoice{55--64}
\mychoice{65--74}
\mychoice{75+}
\mychoice{Prefer not to answer}
\label{app:part1:d1}
\item Which of these best describes your current gender identity? \\
\mychoice{Woman}
\mychoice{Men}
\mychoice{Non-binary} \\
\mychoice{Prefer to self-describe: \rule{1.5cm}{.1pt}} \\
\mychoice{Prefer not to answer}
\label{app:part1:d2}
\item What is the highest degree or level of school you have completed? \\
\mychoice{No schooling completed}
\mychoice{Some high school, no diploma} \\
\mychoice{High school graduate, diploma, or equivalent}
\mychoice{Some college} \\
\mychoice{Trade, technical, or vocational training}
\mychoice{Associate's degree}
\mychoice{Bachelor's degree}
\mychoice{Master's degree}
\mychoice{Professional degree}
\mychoice{Doctorate}
\mychoice{Prefer not to answer}
\label{app:part1:d3}
\item Which of the following best describes your educational background or job field? \\
\mychoice{I have an education in, or work in, the field of computer science, computer engineering or IT.} \\
\mychoice{I do not have an education in, nor do I work in, the field of computer science, computer engineering or IT.} \\
\mychoice{Prefer not to answer}
\label{app:part1:d4}
\end{enumerate}
\noindent{\textbf{One More Thing}} \\
Please indicate if you've honestly participated in this survey and followed instructions completely. You will not be penalized/rejected for indicating `No' but your data may not be included in the analysis: \\
\mychoice{Yes}
\mychoice{No}
\subsection{Measurement Study: Survey Instrument}\label{app:part2}
\setlist{nolistsep}
\footnotesize
\noindent\small{\textbf{Stage 1: Enrollment}}
\footnotesize
\begin{center}
\textit{{\color{teal}Participants solved 5x spatial reasoning tests.}}
\end{center}
\noindent\textbf{Demography}
\begin{enumerate}[leftmargin=3em, label=\textbf{MD\arabic*},noitemsep,nolistsep]
\item Select your age. \\
\mychoice{18--24}
\mychoice{25--34}
\mychoice{35--44}
\mychoice{45--54}
\mychoice{55--64}
\mychoice{65--74}
\mychoice{75+}
\mychoice{Prefer not to answer}
\label{app:part2:d1}
\item Which of these best describes your current gender identity? \\
\mychoice{Woman}
\mychoice{Men}
\mychoice{Non-binary} \\
\mychoice{Prefer to self-describe: \rule{1.5cm}{.1pt}} \\
\mychoice{Prefer not to answer}
\label{app:part2:d2}
\item What is the highest degree or level of school you have completed? \\
\mychoice{No schooling completed}
\mychoice{Some high school, no diploma} \\
\mychoice{High school graduate, diploma, or equivalent}
\mychoice{Some college} \\
\mychoice{Trade, technical, or vocational training}
\mychoice{Associate's degree}
\mychoice{Bachelor's degree}
\mychoice{Master's degree}
\mychoice{Professional degree}
\mychoice{Doctorate}
\mychoice{Prefer not to answer}
\label{app:part2:d3}
\end{enumerate}
\begin{enumerate}[leftmargin=3em, label=\textbf{MAC1},noitemsep,nolistsep]
\item Please select `Agree' as the answer to this question. \\
\mychoice{Strongly disagree}
\mychoice{Disagree}
\mychoice{Neither agree or disagree} \\
\mychoice{Agree}
\mychoice{Strongly agree}
\label{app:part2:ac1}
\end{enumerate}
\begin{enumerate}[leftmargin=3em, label=\textbf{MD\arabic*},noitemsep,nolistsep]
\setcounter{enumi}{3}
\item Which of the following best describes your educational background or job field? \\
\mychoice{I have an education in, or work in, the field of computer science, computer engineering or IT.} \\
\mychoice{I do not have an education in, nor do I work in, the field of computer science, computer engineering or IT.} \\
\mychoice{Prefer not to answer}
\label{app:part2:d4}
\end{enumerate}
\noindent\small{\textbf{Stage 2: Recall}}
\footnotesize
\begin{center}
\textit{{\color{teal}Participants solved 5x spatial reasoning tests.}}
\end{center}
\noindent\small{\textbf{Stage 3: Questionnaire}} \\
\footnotesize
\noindent\textbf{Debriefing} \\
\noindent Participants were debriefed and told about the actual purpose of the study.
\noindent\textbf{Email} \\
\textit{{\color{teal} The \underline{individual} login notification we sent to the participant is displayed for later reference (see Figure~\ref{fig:baselineemail}, but re-branded to match the SRS~study).}}
\noindent \textit{{\color{teal}If participant has not changed their password.}}
\begin{enumerate}[leftmargin=3em, label=\textbf{MQ\arabic*},noitemsep]
\setcounter{enumi}{-1}
\item Do you \textbf{remember} receiving \textbf{this email}?\\
\mychoice{Yes}
\mychoice{No}
\label{app:part2:s1}
\end{enumerate}
{\centering
\textit{{\color{teal}Participants who selected `No' in \ref{app:part2:s1} were forwarded to \ref{app:part2:q10}.}}
}
\noindent\textbf{I-PANAS-SF} \\
\noindent Now we would like to know how you felt in reaction to the email. The list below consists of a number of words that describe different feelings and emotions. Read each item and then mark the appropriate answer on the list. \textbf{Indicate to what extent you felt this way when you noticed the email}.\\
\textit{{\color{teal}We asked I-PANAS-SF from the Baseline Study again.}}
\noindent\textbf{Reaction}
\begin{enumerate}[leftmargin=3em, label=\textbf{MQ\arabic*},noitemsep]
\item Did you \textbf{read} this \textbf{email} when you received it? (email as shown on the left)\\
\mychoice{I did not read it at all}
\mychoice{I only read the subject but not the body}\\
\mychoice{I read the subject and skimmed the body}
\mychoice{I fully read it}
\label{app:part2:q1}
\end{enumerate}
\begin{enumerate}[leftmargin=3em, label=\textbf{MQ2a},noitemsep]
\item In reaction to this email, you decided to change your password.\\Please \textbf{describe} any \textbf{other actions} you took.\\
Answer: \rule{3.5cm}{.1pt}
\label{app:part2:q2a}
\end{enumerate}
\begin{enumerate}[leftmargin=3em, label=\textbf{MQ3a},noitemsep]
\item \textbf{Why} did you react this way, i.e., change your password and take the other actions you described.\\
Answer: \rule{3.5cm}{.1pt}
\label{app:part2:q3a}
\end{enumerate}
\noindent \textit{{\color{teal}If participant has not changed their password.}}
\begin{enumerate}[leftmargin=3em, label=\textbf{MQ2b},noitemsep]
\item \textbf{What} did you do \textbf{in reaction} to it?\\
Answer: \rule{3.5cm}{.1pt}
\label{app:part2:q2b}
\end{enumerate}
\begin{enumerate}[leftmargin=3em, label=\textbf{MQ3b},noitemsep]
\item \textbf{Why} did you react this way?\\
Answer: \rule{3.5cm}{.1pt}
\label{app:part2:q3b}
\end{enumerate}
\noindent\textbf{Content \& Design}
\begin{enumerate}[leftmargin=3em, label=\textbf{MQ\arabic*},noitemsep]
\setcounter{enumi}{3}
\item How much did the following \textbf{factors influence} your \textbf{reaction}?\\
\textit{{\color{teal}Answer choice per item: No effect (1) -- Major effect (5).}}\\
\myspecial{Email metadata (e.g., sender, subject, time of arrival)}\\
\myspecial{Email content (e.g., information, instructions, wording)}\\
\myspecial{Email design (e.g., structure, color, font size)}\\
\myspecial{Experience in dealing with such emails}\\
\myspecial{Negative experience with security and privacy incidents\\(e.g., data breach, identity theft)}
\myspecial{Email appeared to be phishing}\\
\myspecial{Expected to receive such an email}\\
\myspecial{Other: \rule{1.5cm}{.1pt}}\\
\textit{{\color{teal}Answer choices were randomly ordered.}}
\label{app:part2:q4}
\item Please rate how \textbf{helpful} the following \textbf{information} was for \textbf{deciding how to react} to this email?\\
\textit{{\color{teal}Ans. choice per item: Not at all helpful (1) -- Extremely helpful (5).}}\\
\myspecial{Affected account name (i.e., email address)}\\
\myspecial{Location}
\myspecial{Date}
\myspecial{Device}
\label{app:part2:q5}
\end{enumerate}
\noindent \textbf{Time \& Location}
\begin{enumerate}[leftmargin=3em, label=\textbf{MQ\arabic*},noitemsep]
\setcounter{enumi}{5}
\item \textbf{When} did you read the email?\\
\mychoice{I never read it}
\mychoice{Immediately after I noticed it}\\
\mychoice{Less than 1 hour after I noticed it}
\mychoice{A few hours after I noticed it}
\mychoice{One day after I noticed it}
\mychoice{More than one day after I noticed it}
\mychoice{I do not remember}
\label{app:part2:q6}
\noindent \textit{{\color{teal}If participant has not selected ``Never'' in \ref{app:part2:q6}: }}
\item \textbf{In which US state} have you been when you \textbf{read} the \textbf{email}?\\
\textit{Dropdown with all 50 US states + District of Columbia} \\
If somewhere outside the USA: \rule{1.5cm}{.1pt}
\label{app:part2:q7}
\item \textbf{Where} did you read the email?\\
\mychoice{At home}
\mychoice{At work}
\mychoice{On the go}\\
\mychoice{Somewhere else: \rule{1.5cm}{.1pt}}
\mychoice{I do not remember}
\label{app:part2:q8}
\end{enumerate}
\begin{enumerate}[leftmargin=3em, label=\textbf{MAC2},noitemsep,nolistsep]
\item Please select `Agree' as the \textbf{answer} to this question. \\
\mychoice{Strongly disagree}
\mychoice{Disagree}
\mychoice{Neither agree or disagree} \\
\mychoice{Agree}
\mychoice{Strongly agree}
\label{app:part2:ac2}
\end{enumerate}
\begin{enumerate}[leftmargin=3em, label=\textbf{MQ\arabic*},noitemsep]
\setcounter{enumi}{8}
\item In case you received the email at a \textbf{different location or different time}, would your \textbf{reaction} to it been any \textbf{different}?\\
\mychoice{Yes}
\mychoice{No}
\mychoice{Do not know}
\label{app:part2:q9}
\noindent \textit{{\color{teal}If participant selected ``Yes'' in \ref{app:part2:q8}: }}
\item \textbf{What} would you have \textbf{done differently}, if you had received the email at a \textbf{different location or different time}?\\
Answer: \rule{3.5cm}{.1pt}
\label{app:part2:q10}
\end{enumerate}
\noindent\textbf{Comprehension}
\begin{enumerate}[leftmargin=3em, label=\textbf{MQ\arabic*},noitemsep]
\setcounter{enumi}{10}
\item In your opinion, \textbf{why} have you \textbf{received} this email?\\
Answer: \rule{3.5cm}{.1pt}
\label{app:part2:q11}
\end{enumerate}
\noindent \textbf{Expectation}
\begin{enumerate}[leftmargin=3em, label=\textbf{MQ\arabic*},noitemsep]
\setcounter{enumi}{11}
\item In your opinion, \textbf{when} should real companies \textbf{send emails} like this one? (Select all that apply)\\
\mymultchoice{Never}
\mymultchoice{After every detected sign-in which suggests that something is suspicious or wrong}
\mymultchoice{After every detected sign-in when I have not signed in for a while}
\mymultchoice{After every detected sign-in from a new device}
\mymultchoice{After every detected sign-in at an unusual time of the day (e.g., in the middle of the night)}
\mymultchoice{After every detected sign-in from a new location}
\mymultchoice{After every detected sign-in}
\mymultchoice{Other: \rule{1.5cm}{.1pt}}
\label{app:part2:q12}
\noindent \textit{{\color{teal}If participant selected ``Never'' in \ref{app:part2:q12}: }}
\item In your opinion, \textbf{why} do you think real companies should \textbf{never send emails} like this one?\\
Answer: \rule{3.5cm}{.1pt}
\label{app:part2:q13}
\end{enumerate}
\noindent \textbf{Prior Experience}
\begin{enumerate}[leftmargin=3em, label=\textbf{MQ\arabic*},noitemsep]
\setcounter{enumi}{13}
\item Have you had any \textbf{negative experiences} with a \textbf{security or privacy} incident within the \textbf{last two years} (e.g., data breach, identity theft)?\\
\mychoice{Yes}
\mychoice{No}
\label{app:part2:q14}
\item \textbf{Regularly changing} my \textbf{password} (e.g., every 90 days) \textbf{increases} the \textbf{security} of my account.\\
\mychoice{Strongly disagree}
\mychoice{Disagree}
\mychoice{Neither agree or disagree} \\
\mychoice{Agree}
\mychoice{Strongly agree}
\label{app:part2:q15}
\item \textbf{Changing} my \textbf{password} after it \textbf{has been breached} \textbf{increases} the \textbf{security} of my account.\\
\mychoice{Strongly disagree}
\mychoice{Disagree}
\mychoice{Neither agree or disagree} \\
\mychoice{Agree}
\mychoice{Strongly agree}
\label{app:part2:q16}
\end{enumerate}
\noindent{\textbf{One More Thing}}\\
Please indicate if you've honestly participated in this survey and followed instructions completely. You will not be penalized/rejected for indicating `No' but your data may not be included in the analysis: \\
\mychoice{Yes}
\mychoice{No}
\subsection{Codebook}\label{sec:codebook}
\begin{table}[!htbp]
\footnotesize
\centering
\caption{Codebook for \ref{app:part1:q1}, \ref{app:part1:q5}, and \ref{app:part1:q6} used in Section~\ref{sec:study1-comprehension-and-reaction} \nameref{sec:study1-comprehension-and-reaction}.}
\label{tab:codebook:study1-rq1-1}
\begin{tabularx}{\linewidth}{@{}L{.55}R{.15}L{1.5}L{1.8}@{}}
\toprule
\textbf{Code} & \textbf{Freq.} & \textbf{Description} & \multicolumn{1}{l}{\textbf{Example}} \\
\midrule
\multicolumn{4}{c}{\ref{app:part1:q1}: Please list 3 actions you might take after receiving this email and explain why you might take those actions.} \\
\midrule
Change PW & 151 & Participant would change the password. & \emph{``Someone else got into my account, so I need to change the
password right away.''} (M56-C) \\[.5em]
Ignore & 69 & Participant would ignore the email. & \emph{``Ignore the email. It was me that signed in.''} (L78-N) \\[.5em]
Contact AcmeCo & \hspace{.5em}64 & Participant would contact AcmeCo. & \emph{``I would try and contact them and get a solution to this urgent problem.''} (M54-C) \\[.5em]
Review Activity & \hspace{.5em}59 & Participant would check what has been done with the account. & \emph{``I would look for strange activity on the account.''} (M47-C) \\[.5em]
Review Details & \hspace{.5em}37 & Participant would check the shown login information. & \emph{``Check location, date, device. If it is my own location, date, and device, then I can safely ignore this message.''} (L12-N) \\[.5em]
Prevent Phishing & \hspace{.5em}33 & Participant would check the legitimacy of the notification. & \emph{``Check the sender of the email. I would want to be sure the email was legitimate.''} (M16-C) \\[.5em]
Change Other PWs & \hspace{.5em}25 & Participant would change passwords of other accounts. & \emph{``Change passwords for other important accounts as well''} (M87-C) \\[.5em]
Check Other Accounts & \hspace{.5em}22 & Participant would check the other accounts. & \emph{``I would look to see if other accounts of mine had been compromised.''} (M28-C) \\[.5em]
Add Additional Security & \hspace{.5em}19 & Participant would try to add security measures. & \emph{``Look for additional security options e.g. 2FA''} (L8-C) \\[.5em]
Ask Family/Friends & \hspace{.5em}12 & Participant would ask family and/or friends if they have used the account. & \emph{``Ask my husband if he signed in.''} (L42-C) \\[.5em]
Delete Account & \hspace{.5em}10 & Participant would consider deleting the account. & \emph{``My account has been taken over, and might as well make a new account and terminate this account.''} (M7-C) \\[.5em]
Log Out (All) Devices & \hspace{.5em}10 & Participant would try to end the session of some or all devices. & \emph{``I would logout all devices if the option is available.''} (M61-C) \\[.5em]
Remember Recent Logins & \hspace{1em}9 & Participant would try to recall past actions. & \emph{``I will do my best to remember whether I have login recently.''} (L71-C) \\[.5em]
Verify Old PW & \hspace{1em}7 & Participant would log in to check if the password still works. & \emph{``Check to see if my login still works to see if my password's been changed.''} (M16-C) \\[.5em]
\midrule
\multicolumn{4}{c}{\ref{app:part1:q5}: Why? (\textit{\ref{app:part1:q4}: For me, taking action in response to this email from AcmeCo would be [Priority Level]}) } \\
\midrule
Secure Account & 105 & Participant would want to secure the account. & \emph{``To make my information secure if it wasn't me logging in.''} (M46-C) \\[.5em]
It Was Me & 37 & Participant would know why it was sent. & \emph{``It wouldn't be a priority if I recognize the device, data, and location.''} (L1-N) \\[.5em]
Serious Consequences & 31 & Participant would assume that an attacker could cause serious damage. & \emph{``It could have serious consequences.''} (M53-C) \\[.5em]
Security Is Important & 11 & Behaving securely is generally important to the participant. & \emph{``Protecting my accounts is important''} (M31-C) \\[.5em]
Just Info & 8 & Participant would see the notification as a heads-up. & \emph{``I assume it's simply a courtesy email.''} (L95-N) \\[.5em]
Account Is Important & 9 & The account is important to the participant. & \emph{``This account is important to me, so I want to make sure it's not compromised.''} (M83-C) \\[.5em]
Annoyed by Notification & 4 & Participant would be annoyed by the notification. & \emph{``I will not waste my time pursuing something that is not a real security threat to my account.''} (L83-N) \\[.5em]
Identify Damage & 4 & Participant would want to understand what happened. & \emph{``I would need to get to the bottom of the notification''} (L37-C) \\[.5em]
\midrule
\multicolumn{4}{c}{\ref{app:part1:q6}: In your own words, please describe what this email is telling you.} \\
\midrule
New Login & 157 & Participant describes that there was a new login. & \emph{``It's alerting me that my account has been signed into.''} (L61-N) \\[.5em]
Abnormal Login & 63 & Participant describes that there was a new login which deviates from previous ones. & \emph{``The main factor is a security issue associated with an unknown login to my account.''} (L72-N) \\[.5em]
Abnormal Login: Device & 38 & Participant describes that there was a new login from an unusual device. & \emph{``There was a sign in to my account not recognized as me or one of my devices''} (L6-N) \\[.5em]
Abnormal Login: Location & 15 & Participant describes that there was a new login from an unusual location. & \emph{``someone signed in from California and I'm not normally in that area''} (M91-N) \\[.5em]
Abnormal Login: Cookies & 3 & Participant describes that there was a new login without cookies. & \emph{``I just logged in after not logging in for a while, so the cached session was probably expired, and it thought I was a new login.''} (M87-N) \\[.5em]
Abnormal Login: Time & 1 & Participant describes that there was a new login at an unusual time. & \emph{``A new sign-in from a different location or a different time signing in.''} (L91-N) \\[.5em]
\bottomrule
\end{tabularx}
\end{table}
\begin{table}[!htbp]
\footnotesize
\centering
\caption{Codebook for \ref{app:part1:q7} and \ref{app:part1:q11} used in Section~\ref{sec:study1-comprehension-and-reaction} \nameref{sec:study1-comprehension-and-reaction}.}
\label{tab:codebook:study1-rq1-2}
\begin{tabularx}{\linewidth}{@{}L{.6}R{.1}L{1.5}L{1.8}@{}}
\toprule
\textbf{Code} & \textbf{Freq.} & \textbf{Description} & \multicolumn{1}{l}{\textbf{Example}} \\
\midrule
\multicolumn{4}{c}{\ref{app:part1:q7}: In your own words, please describe all of the factors that may have caused you to receive this email.} \\
\midrule
Hack & 134 & An unauthorized person logged into the account. & \emph{``There could have been a breach into my account and jeopardized my security.''} (M54-C) \\[.5em]
New Device & 90 & Login with a new device. & \emph{``Signing in on a different device than normally used''} (L40-N) \\[.5em]
It Was Me & 53 & Participant logged in. & \emph{``I logged in and triggered the email.''} (L20-C) \\[.5em]
New Location & 32 & Login from a new location. & \emph{``Because someone signed into my account from a new location.''} (M38-C) \\[.5em]
Account Inactivity & 12 & Login after a longer period of inactivity. & \emph{``I logged in to an account that I haven't tried in a long time and it wanted to give me notice''} (M62-N) \\[.5em]
Deleted Cookies & 11 & Previously set cookies were deleted. & \emph{``I probably have cleared all of my cookies so the site didn't recognize this device and though it was a new device.''} (L28-N) \\[.5em]
Phishing & 11 & Notification is a phishing attempt. & \emph{``It is a phishing email... they probably just got a list of everyone who has AcmeCo emails and sent it out to all of them to try to get passwords.''} (M25-N) \\[.5em]
Unknown Login & 9 & An unknown login happened & \emph{``The main factor is a security issue associated with an unknown login to my account.''} (L72-C) \\[.5em]
Shared Account & 9 & Someone who the account is shared with logged in. & \emph{``If it wasn't me, then it might have been a loved one with my account access.''} (L93-C) \\[.5em]
Error & 8 & An error occurred on AcmeCo's side. & \emph{``It could be an error, the email may have been sent to me by a glitch in the system.''} (M57-C) \\[.5em]
Notifications Enabled & 8 & The participant opted in to receive such notifications. & \emph{``I have alerts set up to send this notice to me automatically every time my AcmeCo account is accessed (even if it's me who logged in)''} (L41-N) \\[.5em]
\midrule
\multicolumn{4}{c}{\ref{app:part1:q11}: Why? (\textit{\ref{app:part1:q10}: I feel that ignoring this email from AcmeCo would have consequences. [Agreement Level]}) } \\
\midrule
Someone Has Access & 115 & An attacker has access to the account. & \emph{``Someone potentially may have hacked my account''} (L6-C) \\[.5em]
It Was Me & 37 & It was the participant logging in. & \emph{``There would be no consequences of ignoring it if the sign-in was me.''} (L1-N) \\[.5em]
Identity Compromised & 36 & Own identity could be compromised. & \emph{``My personal info could be stolen and used for nefarious purposes.''} (M7-C) \\[.5em]
Depends & 27 & It depends on whether the login was legit or malicious. & \emph{``It depends on whether this was me or not.''} (L90-N) \\[.5em]
Unclear & 9 & Participant has no clear opinion on the consequences of ignoring. & \emph{``There is no telling what is happening.''} (M94-C) \\[.5em]
Compromise Other Accounts & 7 & Ignoring the notification could also compromise other accounts. & \emph{``Someone else has my account and can use it to steal other accounts associated with this email.''} (M11-C) \\[.5em]
Lock Out & 7 & Access to the account could be lost entirely. & \emph{``I feel that someone could get me locked out of my account.''} (M28-C) \\[.5em]
Just Info & 6 & Notification is just a heads-up. & \emph{``It is a courtesy email.''} (L95-N) \\[.5em]
Check Login & 5 & Login needs to be checked. & \emph{``I should at least ensure there was no fraud.''} (L55-C) \\[.5em]
\bottomrule
\end{tabularx}
\end{table}
\begin{table}[!htbp]
\footnotesize
\centering
\caption{Codebook for \ref{app:part1:q9} and \ref{app:part1:q18} used in Section~\ref{sec:study1-decision-and-execution} \nameref{sec:study1-decision-and-execution}.}
\label{tab:codebook:study1-rq2}
\begin{tabularx}{\linewidth}{@{}L{.6}R{.1}L{1.5}L{1.8}@{}}
\toprule
\textbf{Code} & \textbf{Freq.} & \textbf{Description} & \multicolumn{1}{l}{\textbf{Example}} \\
\midrule
\multicolumn{4}{c}{\ref{app:part1:q9}: Why? (\textit{\ref{app:part1:q8}: I feel that this email from AcmeCo explained to me how to resolve the situation. [Agreement Level]}) } \\
\midrule
Explains Malicious & 133 & Participant appreciates that instructions for the malicious case are explained. & \emph{``It gives me the option to change my password.''} (L41-N) \\[.5em]
Explains Both & 53 & Participant appreciates that instructions for the legit and malicious case are explained. & \emph{``It explains to ignore the email if it was me, and change my password if it wasn't.''} (L101-N) \\[.5em]
Missing & 33 & Participant misses information or advice. & \emph{``Changing my password is the very first line of defense, but I wish there was more information about what steps to take, or further assistance from AcmeCo.''} (M26-C) \\[.5em]
Provides Link & 22 & Participant appreciates that a link is provided to change the password. & \emph{``The steps to take were very clear and easy to follow. It also provided a link to make it easy to change my password if I needed to.''} (L42-C) \\[.5em]
Notification Untrusted & 10 & Participant describes the notification to be untrustworthy. & \emph{``It shouldn't be suggesting that I click the link in the email as this is a trick often used by scammers.''} (M35-C) \\[.5em]
\midrule
\multicolumn{4}{c}{\ref{app:part1:q18}: Referring to the situation where a similar email helped you to learn about an unrecognized sign-in.} \\
\multicolumn{4}{c}{Please describe the situation, how you knew that something was wrong, and how you solved it.} \\
\midrule
Malicious & 134 & It was not the participant signing in. & \emph{``I changed my password for my email once because someone got a hold of it somehow.''} (L42-C) \\[.5em]
Location & 68 & Login took place at an unknown location. & \emph{``I knew it because the login happened from a place I did not even visit and immediately changed my password.''} (L43-C) \\[.5em]
Location: Continent & 19 & Login took place on a different continent. & \emph{``I received a notification that someone from another continent logged into my account and I immediately knew this was not me.''} (L50-C) \\[.5em]
Location: Country & 17 & Login took place in a different country. & \emph{``I received an email that someone logged into my account from another country so I changed my password''} (L69-N) \\[.5em]
Location: State & 8 & Login took place in a different state. & \emph{``I had a login notification from California when I live in Texas. Immediately I knew someone hacked into my account.''} (M61-C) \\[.5em]
Unexpected & 32 & Notification came unexpected. & \emph{``I knew something was wrong because I received a sign-in attempt notification from a website I hadn't used in a while, and hadn't even tried signing in to lately.''} (M95-N) \\[.5em]
Device & 17 & Login was done with an unknown device. & \emph{``It was my instagram account, I hadn't logged in on that device they described, I changed my password''} (M20-C) \\[.5em]
Account Sharing & 10 & Login was done by someone who the account is shared. & \emph{``My kids sometimes use my accounts - Amazon, Netflix, Disney plus - I ask them to confirm it was them signing in - which is fine.''} (L89-C) \\[.5em]
It Was Me & 10 & It was actually the participant logging in. & \emph{``I received a similar email from Amazon about a new sign in. It turned out that I used a new device that it didn't recognize.''} (L10-C) \\[.5em]
Time & 2 & Login took place at an unusual time. & \emph{``I knew it was me because the sign in came at a time, and from a location, which did not at all correlate to my real world location, and I dealt with the situation as described in this study - I changed my password and ensured that 2FA was enabled.''} (M27-C) \\[.5em]
\bottomrule
\end{tabularx}
\end{table}
\begin{table}[!htbp]
\footnotesize
\centering
\caption{Codebook for \ref{app:part1:q3}, \ref{app:part1:q13} and \ref{app:part1:q16} used in Section~\ref{sec:study1-perception-and-expectation} \nameref{sec:study1-perception-and-expectation}.}
\label{tab:codebook:study1-rq3}
\begin{tabularx}{\linewidth}{@{}L{.6}R{.15}L{1.5}L{1.75}@{}}
\toprule
\textbf{Code} & \textbf{Freq.} & \textbf{Description} & \multicolumn{1}{l}{\textbf{Example}} \\
\midrule
\multicolumn{4}{c}{\ref{app:part1:q3}: Why? (\textit{\ref{app:part1:q2}: I would feel [Concern Level] about receiving this email from AcmeCo.}) } \\
\midrule
Someone Has Access & 99 & The account is accessed by someone else. & \emph{``Cause someone else is using my account''} (M60-C) \\[.5em]
It Was Me & 41 & Participant knows why it was sent. & \emph{``Because I already know that I just signed in''} (L90-N) \\[.5em]
Privacy Invasion & 28 & Participant mentions the private information which may be saved on the account. & \emph{``i don't want someone to have access to my personal info''} (M39-C) \\[.5em]
PW Exposed & 13 & Participant refers to the risk of an exposed password. & \emph{``Someone is logging into my account which means they have my login info''} (M31-C) \\[.5em]
Someone Tried & 12 & Participant focuses on someone trying to sign into the account. & \emph{``Somebody else tried to log in to my account''} (L85-C) \\[.5em]
Only Information & 8 & Participant sees the notification as a heads-up. & \emph{``It seems like a normal security warning''} (L56-N) \\[.5em]
Feeling of Security & 5 & Participant appreciates the information. & \emph{``Because I like the added security function and appreciate the notice.''} (L28-N) \\[.5em]
Annoyed & 5 & Participant is annoyed by the notification. & \emph{``It's not a problem. It is just annoying.''} (L32-N) \\[.5em]
Phishing & 4 & Participant sees the notification as phishing. & \emph{``I would feel moderately concerned because it is a phishing email.''} (M25-N) \\[.5em]
\midrule
\multicolumn{4}{c}{\ref{app:part1:q13}: Why? (\textit{\ref{app:part1:q12}:I think real companies should send emails like this one when necessary. [Agreement Level]}) } \\
\midrule
Protect Accounts & 111 & Notifications are a form of protection. & \emph{``It's helpful and protects the persons data.''} (M45-C) \\[.5em]
Alert Users & 95 & Notifications alert users to unknown logins. & \emph{``To alert customers about possible security breach''} (M67-C) \\[.5em]
Show Protection & 15 & Sending notifications shows that companies are security conscious. & \emph{``It shows that they care about users security.''} (L22-C) \\[.5em]
Annoying & 10 & Notifications are annoying. & \emph{``These emails are incredibly annoying.''} (L90-N) \\[.5em]
Quick Action & 9 & Notifications allow to take a quick action in case something unauthorized happens. & \emph{``You can immediately stop access to your account from people who exhibit criminal behavior''} (M91-C) \\[.5em]
\midrule
\multicolumn{4}{c}{\ref{app:part1:q16}: Why? (\textit{\ref{app:part1:q15}: I [Frequency Level] read them.}) } \\
\midrule
Account Security & 110 & Reading the notification assure that the account is secure. & \emph{``Because it keeps my accounts safe and protected by monitoring this.''} (L82-C) \\[.5em]
Know Sign-in & 33 & Reading the notification depends on whether the login is known. & \emph{``If I signed in from a second device and then received the email, I knew it was me, so I would ignore it. But when I didn't sign in, I would pay attention and read the email.''} (L16-N) \\[.5em]
Important & 46 & Reading the notifications is important. & \emph{``I read them, cos they are important.''} (M36C) \\[.5em]
Check Details & 36 & Notification is read to check the login details. & \emph{``Just in case it wasn't me. I look at the date, location, and time to determine.''} (L20-N) \\[.5em]
\bottomrule
\end{tabularx}
\end{table}
\begin{table}[!htbp]
\footnotesize
\centering
\caption{Codebook for \ref{app:part2:q2a}, \ref{app:part2:q2b}, \ref{app:part2:q3a}, \ref{app:part2:q3b}, and \ref{app:part2:q11} used in Section~\ref{sec:study2-comprehension-and-reaction} \nameref{sec:study2-comprehension-and-reaction}.}
\label{tab:codebook:study2-rq1}
\vspace{-1em}
\begin{tabularx}{\linewidth}{@{}L{.65}R{.1}L{1.65}L{1.6}@{}}
\toprule
\textbf{Code} & \textbf{Freq.} & \textbf{Description} & \multicolumn{1}{l}{\textbf{Example}} \\
\midrule
\multicolumn{4}{c}{\ref{app:part2:q2a}: In reaction to this email, you decided to change your password. Please describe any other actions you took. } \\
\multicolumn{4}{c}{\ref{app:part2:q2b}: What did you do in reaction to it? } \\
\midrule
Nothing & 121 & Participant did nothing. & \emph{``After I read it, I didn't do anything as it was me who signed in.''} (L17-N) \\
Change PW & 26 & Participant changed the password. & \emph{``I took no other actions than to change my password as directed because I had not signed in.''} (M71-C) \\
Check Details & 10 & Participant checked the login details in the notification. & \emph{``I just made sure it was my device, and on the day I accessed''} (L69-N) \\
Reaction Unclear & 10 & Participant did not know how to react. & \emph{``I was confused and decided to wait and see.''} (M93-N) \\
Archive Email & 6 & Participant archived the notification. & \emph{``save it in my personal files in gmail''} (M8-N) \\
Mark as Spam & 4 & Participant marked the notification as spam. & \emph{``Put it in my spam folder''} (M25-N) \\
Understand & 4 & Participant tried to understand the notification. & \emph{``I thought about it for a couple of minutes and then deleted it.''} (M105-N) \\
\midrule
\multicolumn{4}{c}{\ref{app:part2:q3a}: Why did you react this way, i.e., change your password and take the other actions you described. } \\
\multicolumn{4}{c}{\ref{app:part2:q3b}: Why did you react this way? } \\
\midrule
Was Me & 71 & Participant described the own login being the reason. & \emph{``because it was me that logged in''} (L10-N) \\
Spontaneous & 27 & Participant reacted spontaneously. & \emph{``I just didn't think much of it''} (M51-N) \\
Not Me & 26 & Participant was not the one signing in. & \emph{``Because the wrong state especially the opposite coast is a huge red flag.''} (M77-C) \\
Suspicious & 18 & Participant questioned the legitimacy the notification. & \emph{``I hadn't logged in and the location was California so I was afraid it was a phishing attempt.''} (M107-N) \\
Don't Understand & 14 & Participant did not understand the notification. & \emph{``Wasn't sure what it was for''} (L30-N) \\
Fatigue & 9 & Participant felt fatigued by seeing the notification. & \emph{``it is good for security but I get these all the time.''} (M74-N) \\
Low Value & 7 & Account has a low value for the participant. & \emph{``Why should I care if someone accesses my SRS survey?''} (M70-N) \\
Unsure & 6 & Participant did not know how to react. & \emph{``I unsure it was me why I received it''} (L40-N) \\
Feel Protected & 3 & Participant felt protected by receiving the notification. & \emph{``I was glad that they sent me this in case there was anything out of the ordinary going on.''} (L17-N) \\
\midrule
\multicolumn{4}{c}{\ref{app:part1:q11}: In your opinion, why have you received this email? } \\
\midrule
Inform About Login & 64 & Notification informed about a new login. & \emph{``Because your system recognized that a device signed into my account.''} (L47-N) \\
Check Login & 41 & Notification was a prompt to check the login that just happened. & \emph{``To make sure that it was in fact you who had signed in to the account.''} (M104-N) \\
(Potential) Compromise & 46 & Notification informed about an actual or a potential compromise. & \emph{``My reaction was that someone from California somehow got into my account.''} (M116-C) \\
Don't Know & 39 & Participant did not know why the notification was sent. & \emph{``I had no idea, which is why I deleted it.''} (M93-N) \\
Unusual Login & 28 & Notification informed about a login that was somehow unusual. & \emph{``It sounded like someone other than my typical device had logged into my account.''} (M45-N) \\
Security & 8 & Notification was sent for security reasons. & \emph{``Security purposes.''} (L9-N) \\
Phishing & 3 & Notification was phishing. & \emph{``I thought it was phishing''} (L30-N) \\
\bottomrule
\end{tabularx}
\end{table}
\begin{table}[!htbp]
\footnotesize
\centering
\caption{Codebook for \ref{app:part2:q10} used in Section~\ref{sec:study2-decision-and-execution} \nameref{sec:study2-decision-and-execution}.}
\label{tab:codebook:study2-rq2}
\vspace{-1em}
\begin{tabularx}{\linewidth}{@{}L{.65}R{.1}L{1.65}L{1.6}@{}}
\toprule
\textbf{Code} & \textbf{Freq.} & \textbf{Description} & \multicolumn{1}{l}{\textbf{Example}} \\
\midrule
\multicolumn{4}{c}{\ref{app:part2:q10}: What would you have done differently, if you had received the email at a different location or different time? } \\
\midrule
Pay More Attention & 12 & Participant would have payed more attention to the email. & \emph{``I might have taken a closer look at it.''} (L56-N) \\
Change PW & 6 & Participant would have changed the password. & \emph{``I would have done as the email said and changed my password ''} (L104-N) \\
Contact Support & 4 & Participant would have contacted the support. & \emph{``Read it very carefully. If anything didn't look right I’d have contacted your organization''} (L19-N) \\
Panic & 2 & Participant would have panicked because then someone else would have been signing in. & \emph{``If i was outside I might panic a bit more, or if the email came at a weird or random time''} (L69-N) \\
\bottomrule
\end{tabularx}
\end{table}
\begin{table}[!hb]
\footnotesize
\centering
\caption{Codebook for \ref{app:part2:q13} used in Section~\ref{sec:study2-perception-and-expectation} \nameref{sec:study2-perception-and-expectation}.}
\label{tab:codebook:study2-rq3}
\vspace{-1em}
\begin{tabularx}{\linewidth}{@{}L{.65}R{.1}L{1.65}L{1.6}@{}}
\toprule
\textbf{Code} & \textbf{Freq.} & \textbf{Description} & \multicolumn{1}{l}{\textbf{Example}} \\
\midrule
\multicolumn{4}{c}{\ref{app:part2:q13}: In your opinion, why do you think real companies should never send emails like this one? } \\
\midrule
Feels Like Scam & 4 & Email notification in the current form feels like scam. & \emph{``I got very concern, since include a link in the email instead of suggesting go to the website.''} (M10-N) \\
Annoying & 2 & Receiving the email notifications is annoying. & \emph{``They take too much time''} (M107-N) \\
\bottomrule
\end{tabularx}
\end{table}
|
\section{Introduction}
\label{sec:01}
\input{sections/01_Intro}
\section{Liquidation Mechanics}
\label{sec:02}
\input{sections/02_Liquidation_Mechanics}
\section{Toxic Liquidation Spirals}
\label{sec:03}
\input{sections/03_Toxic_Liquidation_Spiral}
\section{Mitigation Measures}
\label{sec:04}
\input{sections/04_Mitigation_Measures}
\section{Discussion and Conclusion}
\label{sec:05}
\input{sections/05_Conclusion}
\subsection{Dynamic incentives}
The result of halting liquidations is enlightening in its simplicity, but the simulated bad debt can be improved even further with slight tweaks to the liquidation logic. Ultimately, Equation~\ref{eq:UC_condition} can be flipped on its head to obtain the largest allowable liquidation incentive given the user's $LTV$ at the moment of being liquidated, such that the liquidation is not toxic. The condition reads:
\begin{equation}
\label{eq:safety_condition}
i < \frac{1}{LTV}-1
\end{equation}
As long as the liquidation incentive satisfies this condition, user liquidations will always proceed in a healthy manner. It is important to note that when the user's portfolio becomes borderline undercollateralized ($LTV=1$) the liquidation incentive vanishes altogether. This is a safer condition than halting liquidations altogether as some liquidators may still find it profitable to liquidate a position by arbitraging the lending market's oracle's price feed. To avoid complications leading to nonsensically negative incentives when $LTV>1$, and to impose some maximal protocol-set incentive $i_0$ for a given collateral asset, the full model for healthy liquidation incentives can be written as:
\begin{equation}
\label{eq:dyn_incentive}
i(LTV,i_0) = \max\left[\min\left[i_0, \frac{1}{LTV}-1-\epsilon\right],0\right]
\end{equation}
where, for extra safety, we have introduced a static modulation parameter $\epsilon$ to guarantee that the incentive is strictly less than the right-hand side of condition~\ref{eq:safety_condition}. For practical purposes, $\epsilon$ can be any arbitrarily small, non-zero number.
\subsection{Dynamic closing factors}
The dynamic liquidation incentive~\ref{eq:dyn_incentive} can be further paired with a dynamic closing factor which increases as the user's $LTV$ inches towards unity. The idea here is that progressively larger portions of a user's portfolio should be allowed to be closed as the user's portfolio comes progressively closer to becoming undercollateralized. Whenever $LTV\geq 1$, liquidators (or protocol safety modules) should be allowed to liquidate entire asset positions of the user's portfolio in one go. There are many ways this can be expressed mathematically. For concreteness and simplicity we have tested the following linear model:
\begin{equation}
\label{eq:dyn_closing}
c(LTV,c_0) = \min\left[c_0\cdot\frac{1-LTV}{1-LTV_{liq}}+\frac{LTV-LTV_{liq}}{1-LTV_{liq}}, 1\right]
\end{equation}
where we introduce the minimum protocol-set closing factor $c_0$ similarly to what was done for liquidation incentives earlier, as well as a $\min[\cdot,1]$ operation to guarantee that closing factors are always a number $c\leq1$. The reader can verify on their own that whenever $LTV=LTV_{liq}$ the closing factor becomes $c(LTV_{liq},c_0)=c_0$, while when $LTV\geq1$ the closing factor becomes $c(1,c_0)=1$.
The choice of expressions for~\ref{eq:dyn_incentive} and~\ref{eq:dyn_closing} is not unique. They can be modified in a number of different ways and optimized for different purposes according to protocol prerogatives. Our choice is meant solely for demonstrative purposes (where the prerogative is simplicity).
Simultaneous use of dynamic incentives and closing factors should allow the protocol to compensate for the decreasing incentives by offering liquidators more absolute liquidity to profit from as a user's $LTV$ becomes more risky to the protocol. In Figure~\ref{fig:3}, we simulate dynamic incentives both with (orange) and without (yellow) dynamic closing factors. The reader can appreciate how dynamical incentives significantly improve on simple liquidation halting, while the additional inclusion of dynamical closing factors leads to virtually identical results. The orange/gold line (and its shaded confidence interval) shows an expectation of only $\sim$\$250k of total bad debt created 24 hours into the future. This is a significant improvement in risk management that any lending market should consider adopting.
\begin{figure}[H]
\centering
\includegraphics[scale=0.5,width=\textwidth]{figures/Bad_Debt_Dist.png}
\caption{Distribution of bad debt after 24 hours of evolution across the various liquidation policies discussed: toxic spirals (green), halted liquidations (blue), dynamic incentives only (orange), and dynamic incentives + closing factors (gold). The x-axis is broken into two sections to highlight the magnitude of the events involving no bad debt (left vertical axis), while still offering insight into the distribution of tail events (right vertical axis).}
\label{fig:4}
\end{figure}
The average bad debt shown in~\ref{fig:3} does not however tell the entire story of the performance of the different policies shown. For a more nuanced insight, one must look in detail at the distribution of bad debt across the 20k price trajectories used. In Figure~\ref{fig:4} we show a detailed histogram of such bad debt distributions taken at the end of the 24hr simulation run. In Figure~\ref{fig:4}, we plot the histograms of bad debt incurred across all our policy simulations, at the end of the 24-hour simulation timeframe. Whereas the toxic liquidation spirals trade off higher chances of generating bad debt ($\sim 85\%$) for more certainty on its size (tail mean $=\$820k$), the mitigation policies discussed in the text offer significantly lower chances of generating bad debt ($\sim 19\%$) with slightly larger worst-case outcomes (tail mean$=\$1.4M$). This however assumes that no other open market interventions take place. The mitigation policies discussed in the text offer the protocol optionality on how they wish to handle toxic users on a case-by-case basis, whereas toxic liquidation spirals do not.
\subsection{Liquidation Modelling}
The first condition on $q_{repay}$ is set by the protocol's closing factor $c$:
\begin{equation}
q_{repay}< c\cdot B
\end{equation}
Where $B$ is the total dollar amount of outstanding user loans as described in the main text. Once $q_{repay}$ is repaid, the liquidator is allowed to repossess an amount of collateral $\Delta C = (1+i)\cdot q_{repay}$. Since a liquidator cannot repossess more collateral than the total amount $C$ which the user actually owns. This leads to a second condition on $q_{repay}$:
\begin{equation}
(1+i)\cdot q_{repay} < C.
\end{equation}
Once the collateral has been repossessed, the liquidator will swap some amount $x$ to repay the initial flash-loan, incurring some net slippage due to swap routes and trading fees $s(x)$:
\begin{eqnarray}
x\cdot (1-s(x)) &=& q_{repay} \label{eq:slippage_condition}\\
x &<& (1+i)\cdot q_{repay}. \nonumber
\end{eqnarray}
The liquidator's profit $\Pi$ is whatever is leftover from the operation:
\begin{equation}
\Pi(q_{repay}) = (1+i)\cdot q_{repay}-x(q_{repay}),
\end{equation}
where $x(q_{repay})$ requires inverting Equation~\ref{eq:slippage_condition} first.
The final condition on $q_{repay}$ is that it be less than the amount $q_{repay}\leq q_{opt}$ which maximizes liquidator profit:
\begin{equation}
\label{eq:opt_liquidity_condition}
\partial_{q}\Pi(q)\rvert_{q=q_{opt}}= 1+i - \partial_{q} x(q)\rvert_{q=q_{opt}} = 0
\end{equation}
From which the three constraints defining $q_{repay}$ can be written together as:
\begin{equation}
\label{eq:q_repay}
q_{repay} = \textrm{min}\{q_{opt}, c\cdot B, \frac{C}{1+i}\}
\end{equation}
\subsection{Linear Slippage Model}
In the linear slippage model approximation, one has:
\begin{equation}
s(x) = \gamma + \sigma \frac{x}{L},
\end{equation}
where $gamma$ is the trading fee, $\sigma$ is the \emph{linear slippage factor}, $x$ is the amount being swapped, and $L$ the total available swap liquidity used for normalization.
\newline
Inverting equation~\ref{eq:slippage_condition} for $x$, one gets:
\begin{equation}
x(q) = L\frac{1-\gamma}{2\sigma}\left[1-\sqrt{1-\frac{4\sigma q}{L\cdot (1-\gamma)^2}}\right],
\end{equation}
whose derivative computes to:
\begin{equation}
\label{eq:dxq_linear}
\partial_{q}x(q) = \frac{1}{1-\gamma}\left[1-\frac{4\sigma q}{L\cdot (1-\gamma)^2}\right]^{-1/2}=\frac{1}{1-\gamma}\left[1-\frac{4\sigma q}{L(1-\gamma)^2}\right]^{-1/2}.
\end{equation}
The optimal repay amount $q_{opt}$ can then be obtained by plugging~\ref{eq:dxq_linear} into~\ref{eq:opt_liquidity_condition} and solving. One gets:
\begin{equation}
\label{eq:q_opt}
q_{opt} = L\cdot\frac{(1+i)^2(1-\gamma)^2-1}{4\sigma(1+i)^2}
\end{equation}
\subsection{Empirical Slippage Factors}
Slippage factors $\sigma$ are model parameters that must be extracted from real-world data. Ideally, they require linearly approximating the real-world slippage curve as extracted from aggregators. Due to the unavailability of this historical data, we approach the slippage modelling in reverse.
Upon collecting all liquidation calls made on November 22nd, they can be classified according to whether $q_{repay}=q_{opt}$ or not. For each such liquidation, the slippage factor $\sigma$ can be obtained empirically from Equation~\ref{eq:q_opt} to give:
\begin{equation}
\label{eq:empirical_sigma}
\sigma = \frac{(1+i)^2(1-\gamma)^2-1}{4(1+i)^2}\cdot \frac{L}{q_{repay}},
\end{equation}
where we set $\gamma=0.003$ in line with typical on-chain AMM trading fees, and $L=\$ 190M$ in line with normally available liquidity.
\begin{figure}[H]
\centering
\includegraphics[scale=0.5,width=\textwidth]{figures/slippage_factor_dist_colorcoded.png}
\caption{Independent histogram of linear slippage factors as derived from real-life liquidation events occurring before (blue) and after (red) Avi's LTV crossed the UC frontier along with their respective medians (vertical lines). The vertical black line represents the computed median of the entire data set of slippage factors, which was used as the static slippage factor in the simulator's liquidation module.}
\label{fig:5}
\end{figure}
In Figure~\ref{fig:5} we show two histograms charting the empirical distribution of liquidation slippage factors as computed through~\ref{eq:empirical_sigma}. Following the same color-coding scheme used in Figure~\ref{fig:2}, we show the slippage factor distribution both before (blue) and after (red) Avi's portfolio crossed the UC frontier $LTV>LTV_{UC}$. Vertical dashed lines show the median slippage factor of each distribution (dashed red/blue vertical lines), as well as for the entire data as a whole (black dashed vertical line). We use the median of the entire dataset as the slippage factor in the liquidation module of the $\emptyset\textrm{VIX}$ protocol simulator~\cite{chaudhary2022market} when running simulations on Avi's portfolio's bad debt.
For readability, Figure~\ref{fig:5} cuts the x-axis off at values of $\sigma=600$ while $22$ larger values (out of $318$ total events) were observed.
|
\section{Conclusion}
\noindent We build one of the first generative adversarial networks 3DHumanGAN that achieves 3D-aware synthesis of full-body human images. Our 2D-3D hybrid generator effectively and efficiently uses 3D geometric prior of the human body and achieves consistent appearance under varying poses and view-angles with competitive image quality. Our segmentation-based GAN loss plays an irreplaceable role in teaching the generator to parse and condition on the 3D human body prior. By examining ablated models and visualizing internal representation, we find that our pixel-independent backbone together with the spatial adaptive batch normalization technique preserve consistency by keeping the underlying structures inferred from the 3D geometric priors intact.
\section{Experiments}
\noindent\textbf{Dataset}. We compare 3D-aware human image generation performance on SHHQ \cite{fu2022styleganhuman}, a collection of $\sim220$k real-world full-body human images. All compared models are trained on this dataset. We use an off-the-shelf model PARE \cite{kocabas2021pare} to register a SMPL mesh for each of the training images. The semantic label maps are then obtained by rasterizing the conditioning posed mesh (mapping between mesh vertices and semantic labels is known). We augment the semantic label maps with estimations from a 2D semantic segmentation method DensePose \cite{Guler2018DensePose}. The latter's results are better aligned with the images compared to the estimations from PARE. During training, pose, view and appearance are independently sampled. Poses are directly sampled from data. The view angle is sampled from a normal distribution set according to dataset statistics i.e. yaw $\sim \mathcal{N}(0, 0.4)$ and pitch $\sim \mathcal{N}(0, 0.06)$).
\noindent\textbf{Baselines}. We compare with a 2D unconditional method StyleGAN2 \cite{stylegan2}, a 2D conditional method OASIS \cite{oasis2021} and 3D-aware unconditional methods StyleNeRF \cite{gu2021stylenerf} and EG3D \cite{chan2021efficient}. Note that we do not compare with purely-3D methods such as \cite{chan2021pi,schwarz2020graf,niemeyer2021giraffe} because these methods do not support training at a reasonable resolution. To evaluate and compare the pose-conditioning capability, we implement pose conditioning for the unconditional methods by first inverting a human image with desired pose to the style-space and then mixing these style vectors with the ones calculated from the randomly sampled $\mathbf{z}$. For StyleGAN2, the inversion method used is e4e\cite{e4e}. For StyleNeRF and EG3D, we train a ResNet \cite{he2016deep} encoder to infer the latent code given camera parameters.
\noindent\textbf{Evaluation Metrics}. We evaluate three main criteria of the generated results, i.e. fidelity, accuracy of pose conditioning and consistency. We measure fidelity with Fréchet inception distance \cite{heusel2017fid} and kernel inception distance \cite{binkowski2018demystifying}. Two sets of fidelity metrics are calculated for unconditional generation and pose-conditional generation, denoted as (\textit{FID-u}, \textit{KID-u}) and (\textit{FID-p}, \textit{KID-p}), respectively. These values are calculated between $50,000$ real and generated images using an alias-free implementation \cite{parmar2022aliased}. To evaluate the accuracy of generated pose under conditioning, we use PARE \cite{Kocabas_PARE_2021} to re-estimate pose from generated images and calculate the 2D mean per-joint position error \cite{ionescu2013human36m} in normalized screen space ranged $[-1, 1]$. We measure the consistency of appearance under varying pose and view-angle using the signal restoration metric peak signal-to-noise ratio (PSNR) in decibels (dB), following \cite{karras2021alias,zhang2019making}. Specifically , we use the conditioning 3D SMPL mesh to perform a reverse mapping from pixels to mesh vertices. The PSNR is then calculated between corresponding vertices on these colored meshes. The \textit{Con-v} and \textit{Con-p} respectively denote PSNR calculated under view change and pose change.
\subsection{Qualitative Results}
\noindent Figure \ref{fig:qualitative_ours}a shows 3D-aware human image generation results in multiple poses and view-angles. The appearances of the humans generated with our method is consistent under view and pose changes. Figure \ref{fig:qualitative_ours}b and \ref{fig:qualitative_ours}c show the interpolation results. Our model is able to generate smooth interpolation of appearance and pose.
\begin{figure*}
\centering
\vspace{-1ex}
\includegraphics[width=0.875\linewidth]{figures/fig_qualitative.pdf}
\caption{\textbf{Qualitative Comparison.} We show two cases separated by the dotted line. For each case, the first row shows unconditional generation results, the second row shows pose-conditional generation results. We show three view angles, from $-30^{\circ}$ to $30^{\circ}$.
}
\label{fig:qualitative}
\vspace{-1ex}
\end{figure*}
\begin{table}
\caption{\textbf{Quantitative Comparisons} on human images generated by 2D and 3D methods. \textit{KID} values are reported in $10^{-3}$ units; \textit{Pose} values are reported in $10^{-2}$ units. Our method achieves the lowest fidelity scores and the highest score of appearance consistency under different body-poses.}
\vspace{-1ex}
\begin{center}
\resizebox{\linewidth}{!}{
\begin{tabular}{l|cccc|cc}
\multirow{2}{*}{Metrics} & \multicolumn{1}{l}{\multirow{2}{*}{\begin{tabular}[c]{@{}l@{}}StyleGAN-\\ Human, 512\end{tabular}}} & \multicolumn{1}{l}{\multirow{2}{*}{\begin{tabular}[c]{@{}l@{}}OASIS,\\ 256\end{tabular}}} & \multicolumn{1}{l}{\multirow{2}{*}{\begin{tabular}[c]{@{}l@{}}StyleNeRF,\\ 512\end{tabular}}} & \multicolumn{1}{l}{\multirow{2}{*}{\begin{tabular}[c]{@{}l@{}}EG3D,\\ 512\end{tabular}}} & \multicolumn{1}{|l}{\multirow{2}{*}{\begin{tabular}[c]{@{}l@{}}Ours,\\ 256\end{tabular}}} & \multicolumn{1}{l}{\multirow{2}{*}{\begin{tabular}[c]{@{}l@{}}Ours,\\ 512\end{tabular}}} \\
& \multicolumn{1}{l}{} & \multicolumn{1}{l}{} & \multicolumn{1}{l}{} & \multicolumn{1}{l}{} & \multicolumn{1}{|l}{} & \multicolumn{1}{l}{} \\
\Xhline{1.2pt}
FID-p & 6.68 & 10.74 & 13.13 & 26.75 & \textbf{6.61} & 9.31 \\
KID-p & 3.60 & 8.82 & 7.01 & 19.70 & \textbf{3.01} & 5.16 \\
FID-u & \textbf{2.83} & - & 7.6 & 17.83 & - & - \\
KID-u & \textbf{1.50} & - & 3.96 & 11.74 & - & - \\
\hline
Pose & 4.48 & \textbf{1.82} & 6.64 & 8.22 & 2.21 & 2.08 \\
\hline
Con-v & - & - & 18.29 & 19.93 & 21.83 & \textbf{22.66} \\
Con-p & 12.77 & 13.19 & 13.74 & 10.66 & 17.72 & \textbf{18.58} \\
\Xhline{1.2pt}
\end{tabular}
}
\end{center}
\vspace{-2ex}
\label{table:sota}
\end{table}
\begin{table}
\caption{\textbf{Ablation study} quantitative results. \textit{KID} values are reported in $10^{-3}$ units; \textit{Pose} values are reported in $10^{-2}$ units. }
\begin{center}
\vspace{-1ex}
\resizebox{\linewidth}{!}{
\small
\begin{tabular}{l|cccc|c}
Metrics & VAE & Feed Forward & Upsampling & Instance Norm & Full \\
\Xhline{1.2pt}
FID-p & 8.00 & 9.42 & 8.28 & 8.90 & \textbf{6.61} \\
KID-p & 4.02 & 4.57 & 5.42 & 4.04 & \textbf{3.01} \\
\hline
Pose & 9.54 & 2.60 & \textbf{1.89} & 1.93 & 2.21 \\
\hline
Con-v & 19.22 & 21.63 & 16.03 & 17.73 & \textbf{21.83} \\
Con-p & 15.18 & \textbf{17.89} & 13.16 & 13.36 & 17.72 \\
\Xhline{1.2pt}
\end{tabular}
}
\end{center}
\vspace{-1ex}
\label{table:ablation}
\end{table}
\subsection{Comparisons}
\noindent Quantitative results are presented in Table \ref{table:sota}. Qualitative results are shown in Figure \ref{fig:qualitative}. Our method demonstrates the best overall performance, with an FID comparable to the 2D state-of-the-art method StyleGAN-Human and the best consistency. With GAN inversion techniques applied, the unconditional StyleGAN2, StyleNeRF and EG3D are able to generate images where the target pose is close to the condition but with some gaps. This suggests that the representations learned by these methods are entangled in appearance and pose. Style mixing is not enough to achieve clean separation in the two domains. Our method is able to generate images that accurately matches the conditioning pose. Besides, all of the compared baselines show different extents of inconsistencies under view and pose variation. This is an indicator of entanglement of appearance and pose in the learned latent spaces of the baselines. In contrast, the appearance of the humans generated by our method remains consistent under pose and view variation.
\subsection{Ablation Study}
\noindent To assess the effectiveness of each designed module in our 3d-aware generator, we evaluate a set of ablated models where the modules to be evaluated are replaced with approaches used by previous works. Specifically,
to examine the effects of passing the 3D geometric prior into the 2D backbone as styles, i.e. the 2D-3D hybrid generator, we compare with the feed-forward approach used by \cite{gu2021stylenerf, chan2021efficient, zhou2021CIPS3D}, denoted as (\textit{Feed Forward}).
To evaluate the contribution of the segmentation-based GAN loss, we train a model in which this loss is replaced with the traditional GAN loss with binary discrimination. This results in a model trained under VAE-style supervision as in \cite{hong2021headnerf}, denoted as \textit{VAE}.
To examine the effectiveness of the pixel-independent backbone design, we compare with a variant that uses traditional upsampling convolution network like in \cite{oasis2021, stylegan2}, which is denoted as \textit{Upsampling}.
To examine the effectiveness of spatial adaptive batch normalization, we compare with a variant that uses spatial adaptive instance normalization instead, denoted as \textit{Instance Norm}. All ablation models are trained and evaluate in the resolution of $256 \times 128$. Quantitative results are presented in Table \ref{table:ablation} and qualitative results are presented in the appendix. Our full model performs best in fidelity and view-consistency. The \textit{Feed-Forward} approach slightly outperforms the full model in consistency, but with no small sacrifice in image quality. In other terms, passing the geometric prior in the form of low-level styles has a small toll on pose-accuracy and consistency, but completely within the acceptable range. Meanwhile, the image quality is greatly improved. By examining the results of the \textit{VAE} setting, we can see that segmentation-based GAN loss is crucial to pose accuracy. It also improves fidelity, possibly because this loss helps the model to make better use of the 3D Human Prior. The \textit{Upsampling} and \textit{Instance Norm} configurations show the best accuracy of generated pose but with impaired image consistency in different extents. These two configurations show that the pixel-wise independent backbone and spatial adaptive batch normalization indeed preserves consistency.
\begin{figure}
\centering
\includegraphics[width=0.9\linewidth]{figures/fig_internal.pdf}
\vspace{-1ex}
\caption{\textbf{Internal Representation Visualization.} Two identities in two different poses are visualized for each model. For the same model, the same feature channel are shown for the two identities.}
\label{fig:internal}
\vspace{-1ex}
\end{figure}
\subsection{Internal Representation}
\noindent Figure \ref{fig:internal} visualizes typical internal representations from our networks. One interesting obervation is that in the model that uses batch normaliztion, we are able to find certain feature activations that are consistent across different poses, view-angles and appearances, such as the ones depicted. In contrast, activations of such kind cannot be found in the model that uses instance normalization. One explanation is that the feature activations in the instance normalization model mainly controls the presence of image features rather the location, which is similarly reported in \cite{karras2021alias}. We believe an internal coordinate system is preserved when using batch normalization and this helps locate different image textures on the 3D surface, which certainly helps image consistency.
\subsection{Limitations and Discussion}
\noindent Our network still generates some artifacts (see Figure \ref{fig:qualitative}) when we rotate by a large angle. Part of the reason is that the dataset features a narrow distribution of view-angles. Another observation is that there are small inconsistencies in the appearance when the view-angle changes. (see Figure \ref{fig:qualitative_ours}). Therefore, our method can still be improved in terms of view-consistency.
\section{Introduction}
Human image generation is a long-standing topic in computer vision and graphics with countless applications across multiple areas of interest including movie production, social networking and e-commerce. Compared to physically-based methods, data-driven approaches are preferred due to the photo-realistcness of their results, versatility and ease of use \cite{tewari2020state}. In this work, we are interested in synthesizing full-body human images with a 3D-aware generative adversarial network (GAN) that produces appearance-consistent images under different view-angles and body-poses.
Rapid developments have been seen in using 3D-aware GANs to generate view-consistent images of human faces \cite{niemeyer2021giraffe, schwarz2020graf, chan2021pi, chan2021efficient, gu2021stylenerf, zhou2021CIPS3D, hong2021headnerf}. However, these methods have limited capacity when dealt with complex and articulated objects such as human bodies. To begin with, methods based solely on neural volume rendering \cite{chan2021pi, niemeyer2021giraffe, schwarz2020graf} are too memory inefficient. Rendering human bodies requires a volumetric representation that is much more dense than that of faces, which makes it computationally infeasible. A line of work improves the computational efficiency and rendering quality of 3D-aware GANs by upsampling the rendered result with a convolutional neural network \cite{zhou2021CIPS3D, gu2021stylenerf, chan2021efficient, hong2021headnerf}.
We argue that this feed-forward approach is not well-suited for generating full-body human images, since the 3D representation is tasked with simultaneously modeling the articulated geometry and the appearance of human bodies, which poses great challenges on representational capacity, while the capacity of the 2D-network is not fully utilized.
As depicted in Figure \ref{fig:gen}, our work seeks to solve these problems by introducing a novel generator architecture in which a 2D convolutional backbone is modulated by a 3D pose mapping network. This design is motivated by the observation that in a StyleGAN2~\cite{stylegan2} model trained on human images certain layers of styles correlate strongly with the pose of the generated human \cite{fu2022styleganhuman} while others correlate more apparently with the appearance.
The 3D Pose Mapping Network is formulated as a renderable implicit function conditioned on a posed 3D human mesh derived with a parametric model \cite{SMPL:2015}. In this way, the 3D representation handles the simplified task of parsing a geometric prior while the network no longer has to learn the articulated body geometry by itself. As an additional benefit of explicitly conditioning on posed human mesh, the pose of the generated human can be specified. The output of the 3D pose mapping network is used to render a 2D low-resolution style map through ray integration \cite{mildenhall2020nerf}. The style map is passed into the first few layers of our backbone network.
Then, we introduce the Appearance Mapping Network to control the appearance of the generated human. The appearance mapping network is formulated as an MLP, following a common practice of style-based generators \cite{stylegan, stylegan2}. For the network to learn to parse the 3D geometric prior, we use a segmentation-based GAN loss \cite{sushko2020you} calculated using a U-Net \cite{ronneberger2015unet} discriminator. This design enables the network to establish one-to-many mapping from 3D geometry to synthesized 2D textures using only collections of single-view 2D photographs without manual annotations.
Although capable of producing photo-realistic results, CNN-based generator networks often sacrifice consistency under geometric transformations \cite{chan2021efficient}, \ie the appearance of the generated human may change with varying pose and view-angle. To preserve consistency, we propose a network design with two key aspects: 1) Our backbone network is built entirely with $1 \times 1$ convolutions. This helps eliminate positional reference and promotes equalvariance \cite{karras2021alias}. 2) Modulation from the pose mapping network is passed into the backbone by means of spatial adaptive batch normalization instead of the commonly used instance
normalization \cite{stylegan2, park2019spade}, so that underlying structure from the geometric information parsed by our pose mapping network is not impaired.
Our contributions can be summarized as follows:
1) Our work is one of the first to achieve 3D-aware generation of photo-realistic full-body human images.
2) We propose a 2D-3D hybrid generator which is both efficient and expressive. The model is supervised with segmentation-based GAN loss which helps establish a mapping between 3D coordinates and 2D human body semantics.
3) Our generator is carefully designed to preserve the consistency of appearance when pose and view-angle vary.
\section{Generator Architecture}
\noindent We build a generative adversarial network that synthesizes 3D-aware full-body human images with specified pose $\mathbf{p}$ and view-angle $\mathbf{v}$. Pose is specified as a set of coordinates of the vertices in a posed body mesh $\mathbf{p} = \{\mathbf{v}_i \in \mathbb{R}^{3}\}_{i=1..6890}$ derived using a parametric human body model SMPL \cite{SMPL:2015} $\mathbf{p} = \text{SMPL}(\boldsymbol{\theta}, \mathbf{v})$. $\boldsymbol{\theta} \in \mathbb{R}^{K \times 4}$ denotes the rotation at each body joint in the quaternion format. The appearance of the generated human is randomly sampled, represented as a vector $\mathbf{z} \in \mathbb{R}^{N_z}$, but required to stay consistent when pose and view-angle vary. To this end, we design a 2D-3D-hybrid generator architecture and a training strategy to achieve partially conditioning on pose.
\noindent\textbf{Architecture.} An overview of our generator is presented in Figure \ref{fig:gen}. We use a style-based generator but with a major difference from prior works: different layers of the convolutional backbone are separately modulated by two mapping networks, \ie, a pose mapping network that handles 3D human-body geometry and an appearance mapping network which resembles that of other style-based generators. This design is motivated by the observation that in a StyleGAN2 \cite{stylegan2} model trained on human images. The low-level styles correlate strongly with the pose and orientation of the generated human \cite{fu2022styleganhuman} while others correlate more apparently with the appearance, as shown in Figure \ref{fig:stylemix}. This inspires us to inject the human geometric prior by calculating the low-level styles from the conditioning posed mesh.
\noindent\textbf{3D Pose Mapping Network.} To parse the 3D geometric information, the pose mapping network is formulated as a renderable 3D implicit function $f_p(\psi(\mathbf{x}|\mathbf{p}, \boldsymbol{\theta}), \mathbf{d}) = (\sigma, \mathbf{s}_\mathbf{x})$ where $\mathbf{x} \in \mathbb{R}^3$ is a coordinate in the 3D camera space; $\mathbf{d} \in \mathbb{R}^3$ denotes the orientation of the camera ray; $\mathbf{p} \in \mathbb{R}^{6890 \times 3}$ denotes the coordinates of the body vertices in camera space; $\psi(\mathbf{x}|\mathbf{p}, \boldsymbol{\theta})$ transforms the coordinate $\mathbf{x}$ from the camera space to a canonical space, added to facilitate learning. Following literature in 3D human body reconstruction \cite{zhao2021humannerf,peng2021animatable,mihajlovic2021leap}, $\psi$ is defined as the inverse process of linear blend skinning,
\begin{align*}
\psi(\mathbf{x}|\mathbf{p}, \boldsymbol{\theta}) = {\left ( \sum_{k=1}^{K} \omega_k(\mathbf{x}_\mathbf{p}) \mathbf{G}_k (\boldsymbol{\theta}) \right )} ^ {-1} \mathbf{x}
\end{align*}
$\omega_k(\mathbf{x}, \mathbf{p})$ returns the blend weights corresponding to the vertex nearest to the querying point $\mathbf{x}$; $\mathbf{G}_k (\boldsymbol{\theta}) \in \text{SE}(3)$ denotes the cumulative linear transformation at the $k^{th}$ skeleton joint. The outputs of the implicit function are the opacity $\sigma \in \mathbb{R}$ and a style vector $\mathbf{s} \in \mathbb{R}^{N_s}$, which are used to render a 2D low-resolution pose style map $\mathbf{w}_\mathbf{s} \in \mathbb{R}^{H_s \times W_s \times N_s}$ via ray integration \cite{mildenhall2020nerf}. The implicit function is parameterized as an MLP with periodic activation \cite{sitzmann2020siren} to handle low-dimensional coordinate input.
\begin{figure}
\centering
\includegraphics[width=0.8\linewidth]{figures/fig_stylemix.pdf}
\vspace{-2ex}
\caption{\textbf{Style mixing with StyleGAN2.} The source and reference images are randomly generated using StyleGAN2. Style mixing is to replace to part of the style vectors from the source image with their counterparts from the reference images. \textit{Low}-level styles modulate the first few layers of the StyleGAN2 generator while \textit{high}-level styles modulate the last few layers.}
\label{fig:stylemix}
\vspace{-1ex}
\end{figure}
\noindent\textbf{Neural Rendering.} We use classic volume rendering technique to aggregate predicted style vectors across space. We start by evenly sampling $N$ points $\{ \mathbf{x}_i = \mathbf{o} + t_i \mathbf{d} \}$ within near and far bounds $[t_n, t_f]$ along each camera ray $\mathbf{r}(t) = \mathbf{o} + t \mathbf{d}$. $\mathbf{o}$ denotes the camera center. The style vector on each 2D spatial location is then estimated via
\begin{align*}
\mathbf{S}(\mathbf{r}) & =
\sum_{i=1}^N ~
T_i \left( 1 - \exp \left( - \sigma \left( \mathbf{x}_i \right) \delta_i \right) \right) \mathbf{s} (\mathbf{x}_i, \mathbf{d}), \\
T_i & = \exp \left(- \sum_{j=1}^{i-1} \sigma(\mathbf{x}_j) \delta_j \right),
\end{align*}
where $\delta_i = \left | \mathbf{x}_{i+1} - \mathbf{x}_{i} \right |$ denotes the distance between adjacent samples.
\noindent\textbf{Appearance Mapping Network}. The appearance mapping network is an MLP $f_a(\mathbf{z})=\mathbf{w} \in \mathbb{R}^{N_w}$, following common practice of style-based generators \cite{stylegan, stylegan2, karras2021alias}. The output style vectors $\mathbf{w}$ do not have spatial dimensions. The final output image is synthesized by the convolutional backbone under modulation from both mapping networks.
\section{Training}
\noindent Our training pipeline is illustrated in Figure \ref{fig:train}. For the network to learn to parse the 3D geometric prior and synthesize images with specified pose, a mapping from 3D geometric information to 2D textures needs to be established. Inspired by \cite{oasis2021}, we use a U-Net \cite{ronneberger2015unet} discriminator architecture together with a segmentation-based GAN loss to establish a one-to-many mapping between geometry and textures. Specifically, the U-Net discriminator classifies each pixel as \textit{fake}, \textit{background} or one of 25 semantic classes (e.g. \textit{head}, \textit{torso} ...), unlike the binary classification on entire images used in traditional GAN training. This approach simultaneously enables free appearance sampling and pose conditioning. A popular practices in pose-conditioned human image generation it to use VAE-style supervision, but the quality of generated image is often outperformed by that of GAN-style training \cite{fruhstuck2022insetgan}. When updating the discriminator, triplets of real images, corresponding SMPL meshes and semantic label maps $(\mathbf{I}, \mathbf{p}_\mathbf{I}, \mathbf{m}_\mathbf{I})$ are needed to calculate the loss.
\begin{align*}
\mathcal{L}^D_\text{adv} = &- \mathbb{E}_{\mathbf{I}}
\left[
\sum_{c}^{N_c} \alpha_c \sum_{x,y}^{H \times W} {\mathbf{m}_\mathbf{I}}_{x,y,c} \log D(\mathbf{I})_{x,y,c}
\right] \\
&- \mathbb{E}_{\mathbf{p},\mathbf{z}}
\left[
\sum_{c}^{N_c} \alpha_c \sum_{x,y}^{H \times W} {\mathbf{m}_\mathbf{0}}_{x,y,c} \log D(G(\mathbf{p}, \mathbf{z}))_{x,y,c}
\right]
\end{align*}
where $x$, $y$ and $c$ are subscripts of the height, width and class dimension. $\alpha_c$ denotes class weights, calculated as the inverse of the per-pixel class frequency. $\mathbf{m}_\mathbf{0}$ denotes a semantic label map in which everywhere is $0$, corresponding to the \textit{fake} class. When updating the generator, pairs of SMPL meshes and semantic label maps $(\mathbf{p}, \mathbf{m}_\mathbf{p})$ are required.
\begin{equation*}
\mathcal{L}^G_\text{adv} =
- \mathbb{E}_{\mathbf{p},\mathbf{z}}
\left[
\sum_{c}^{N_c} \alpha_c \sum_{x,y}^{H \times W} {\mathbf{m}_\mathbf{p}}_{x,y,c} \log D(G(\mathbf{p}, \mathbf{z}))_{x,y,c}
\right]
\end{equation*}
The segmentation-based GAN loss is sufficient for training the model to generate consistent images with desired pose and view-angle. In practice, we find adding two additional loss terms improves image quality. The first is a perceptual loss \cite{johnson2016perceptual} which minimizes the difference between feature activations extracted from real and generated images using a pretrained VGG \cite{simonyan2014vgg} network.
\begin{equation*}
\mathcal{L}^G_\text{perceptual} =
\left | \text{VGG}(G(\mathbf{p}_\mathbf{I}, \hat{\mathbf{z}}_\mathbf{I})) - \text{VGG}(\mathbf{I}) \right |
\end{equation*}
where $\hat{\mathbf{z}}_\mathbf{I}$ is an appearance latent code corresponding with the groundtruth image $\mathbf{I}$. These appearance codes are learnable parameters in the model, initialized by inverting a pretrained StyleGAN2 model \cite{fu2022styleganhuman} with e4e \cite{e4e}.
During training, the collection of appearance latent codes are optimized but anchored by a latent loss.
\begin{equation*}
\mathcal{L}^G_\text{latent} =
\left | \hat{\mathbf{z}_\mathbf{I}} - \text{e4e}(\mathbf{I}) \right |
\end{equation*}
It is reported that gradient penalty on the discriminator facilitates stable training and convergence \cite{mescheder2018training, stylegan2}. We implement an R1 regularization for our segmentation-based GAN loss:
\begin{equation*}
\mathcal{L}^D_{R1} = \frac{\gamma}{2} \mathbb{E}_\mathbf{I} { \left \| \nabla_\mathbf{I} D(\mathbf{I}) \right \| }^ 2
\end{equation*}
The total loss for training the generator is the sum of the above terms.
\begin{equation*}
\mathcal{L}^G = \lambda_\text{adv} \mathcal{L}^G_\text{adv} + \lambda_\text{perceptual} \mathcal{L}^G_\text{perceptual} + \lambda_\text{latent} \mathcal{L}^G_\text{latent} + \gamma \mathcal{L}^D_{R1}
\end{equation*}
\section{Preserving Consistency}
\noindent Although capable of producing high-quality results, CNN-based generator networks often sacrifice consistency under geometric transformations \cite{chan2021efficient}. This is a crucial challenge is our approach. Our network is carefully designed to preserve the consistency of appearance against varying pose and view-angle.
\begin{figure}
\centering
\includegraphics[width=0.8\linewidth]{figures/fig_sabn.pdf}
\caption{\textbf{Spatial adaptive batch normalization} first performs a statistics-based per-channel normalization on the feature maps and then a re-normalization with mean and variance values calculated from the style maps.}
\label{fig:san}
\vspace{-3ex}
\end{figure}
\noindent\textbf{Pixel-wise independent convolutional backbone}. Our backbone network is based entirely on $1 \times 1$ convolutions and no upsampling is used. The input to the backbone network is a constant coordinate grid $\mathbf{g} \in \mathbb{R}^{H \times W \times 2}$ containing normalized screen coordinates ranged $[-1, 1]$. Thereby, each pixel is synthesized independent of others. The purpose of this is to eliminate unwanted positional references so that the network is equal-variant to geometric transformations \cite{karras2021alias}.
\noindent\textbf{Spatial-adaptive batch normalization}. The spatial adaptive batch normalization which is used to pass the style maps into the backbone in a way that does not impair underlying structure. Let us denote by $\mathbf{f}^{(i)}$ the output feature map of the $i^{th}$ layer of the backbone. A spatial adaptive normalization \cite{park2019spade} function $\phi(\mathbf{f}^{(i)}, \mathbf{w}_\mathbf{s})=\hat{\mathbf{f}}^{(i)}$ is needed to inject the style maps rendered by our pose mapping network into the feature maps. As shown in Figure \ref{fig:san}, spatial adaptive normalization first performs a statistics-based per-channel normalization on the feature maps and then a re-normalization with mean and variance values calculated from the style maps. In the first normalization step, how the statistics are calculated makes a crucial difference. When the statistics are calculated on individual feature maps in a mini-batch, \ie instance normalization \cite{ulyanov2016instance}, which is used by \cite{stylegan, park2019spade}, consistency is impaired. We hypothesize that this is because instance normalization removes global affine transformation applied by the previous spatial normalization and forces the normalized layer to resort to controlling only the presence of finer features rather than their precise positions, as similarly observed in \cite{karras2021alias}.
We propose to use \textit{spatial adaptive batch normalization}:
\begin{equation*}
\phi(\mathbf{f}^{(i)}_{x, y, c}, \mathbf{s}) =
\gamma^{(i)}(\mathbf{s}) _{x, y, c} \cdot
\frac
{\mathbf{f}^{(i)}_{x, y, c} - \mathbb{E}_{\mathbf{p}, \mathbf{z}}[\mathbf{f}^{(i)}_{x, y, c}]}
{\sqrt{\mathbb{V}_{\mathbf{p}, \mathbf{z}}[\mathbf{f}^{(i)}_{x, y, c}]}}
+ \beta^{(i)}(\mathbf{s})_{x, y, c}
\end{equation*}
where $x$, $y$ and $c$ are the subscripts for the width, height and channel dimension. $\gamma^{(i)}$ and $\beta^{(i)}$ are learned affine transformations. $\mathbb{E}$ and $\mathbb{V}$ estimates the means and variances of the activations of a feature channel across all pixels and all combinations of poses, views and appearances. In practice, this is done by keeping track of the mini-batch statistics of the per-channel activations through exponential moving average. Spatial adaptive batch normalization is used throughout the generation pipeline. For the style vectors $\mathbf{w}$ calculated with the appearance network, the same re-normalization is broadcasted to all spatial locations.
\section{Related Works}
\noindent\textbf{3D-Aware Image Generation}. Generative Adversarial Networks \cite{goodfellow2014generative} are able to generate images of human/animal faces, cars, indoor and natural scenes in eye-fooling photo-realisticness \cite{stylegan,stylegan2,karras2021alias}. This motivates researchers to explore adversarial learning's potential in achieving 3D-aware image synthesis. Earlier efforts devised different network structures to process meshes \cite{liao2020towards, szabo2019unsupervised}, voxels \cite{zhu2018visual, gadelha20173d, henzler2019escaping} and block-based representations \cite{hao2021gancraft, liu2020neural}. These methods suffer from the common problem of insufficient 3D inductive bias \cite{chan2021efficient}. Neural implicit representations rendered via ray integration have been found to be an effective representation for data-driven 3D synthesis \cite{lombardi2019neural}. Progress was gained in generating objects with simple structure such as human faces \cite{chan2021pi, niemeyer2021giraffe, schwarz2020graf}. These approaches are burdened by heavy computational cost and limited to generating low-resolution images. Several works propose to remedy this by upsampling the rendered results with a 2D convolutional network and achieved impressive quality \cite{zhou2021CIPS3D, gu2021stylenerf, chan2021efficient, hong2021headnerf}, yet still not ready for the challenging 3d-aware human image generation. This is because the complexity of the articulated structure of human bodies challenges the 3D representations used by these works.
\begin{figure*}[!t]
\centering
\begin{minipage}[b]{0.648\textwidth}
\includegraphics[width=\linewidth]{figures/fig_gen.pdf}
\vspace{-2ex}
\subcaption{\small{Generator Architecture}}
\label{fig:gen}
\end{minipage}%
\begin{minipage}[b]{0.252\textwidth}
\includegraphics[width=\linewidth]{figures/fig_train.pdf}
\vspace{-2ex}
\subcaption{\small{Training Pipeline}}
\label{fig:train}
\end{minipage}%
\vspace{-2ex}
\caption{\textbf{Architecture.} \textbf{(a)} Our generator uses a pixel-wise independent convolutional backbone modulated by a 3D pose mapping network and an appearance mapping network. \textbf{(b)} Our model is supervised with a segmentation-based GAN loss where the semantic labels are rasterized from the conditioning mesh. }
\end{figure*}
\noindent\textbf{Human Image Generation}. Previous works on person image generation focus on re-synthesizing the reference image, such as rendering humans in a novel pose~\cite{ma2017pose,PATN,posewithstyle,cheng2022generalizable}, with different garments~\cite{han2018viton,tryongan} or with text~\cite{jiang2022text2human}. Earlier methods condition sparse key-points~\cite{ma2017pose,ma2018disentangled,PATN,GFLA,styleposegan} or semantic maps~\cite{dong2018soft,han2019clothflow,transmomo2020} to manipulate the 2D images. To better preserve the appearance consistency of the source, surface-based methods~\cite{li2019dense,lwb2019,posewithstyle} are introduced to establish the dense correspondences between pixel and human prior surface, \ie , DensePose~\cite{Guler2018DensePose} and SMPL~\cite{SMPL:2015}. A limited amount of work focuses on human synthesis without a reference image, where \cite{vunet2018} and \cite{sarkar2021humangan} map the appearance into the Gaussian variables, enabling the synthesis of novel persons. Inspired by the success of StyleGAN~\cite{stylegan,stylegan2,karras2021alias} on unconditional generation, several works~\cite{fu2022styleganhuman, fruhstuck2022insetgan} explore the capacity of the model for full-body synthesis from different aspects. Specifically, StyleGAN-Human~\cite{fu2022styleganhuman} takes a data-centric perspective toward human generation and discusses the effect of manipulated style embedding on human images. InsetGAN~\cite{fruhstuck2022insetgan} proposes a multi-GAN optimization framework to synthesize more plausible-looking humans. All methods mentioned above are 2D methods.
Three concurrent works~\cite{bergman2022generative,zhang2022avatargen,hong2022eva3d} tackle 3D-aware human image generation based on different 3D representations. \cite{bergman2022generative} and \cite{zhang2022avatargen} render a low-resolution image with a 3D triplane representation \cite{chan2021efficient} and uses a 2D convolutional network to enlarge the image in a feed-forward manner. \cite{hong2022eva3d} render images directly with an efficient compositional 3D neural field.
Our work is different from these works in the following aspects: 1) We use a 2D generator modulated by 3D human body prior instead of directly rendering the image from a 3D representation. 2) We use segmentation-based adversarial supervision which encourages not only fidelity but also semantic correspondence.
|
\subsubsection*{Acknowledgements}}
\long\def\longthanks#1{\subsubsection*{Acknowledgements}
#1}%
\def\enddoc@text{}
\def\@makePagedeGarde
\ifcdr@pagedegarde
{\makeatletter\input{CR-pagedegarde}}%
\newpage\pagecolor{white}\gdef\GPT@pageliteral{}\fi}
\ifcdr@sommaire
\let\cdr@postabstracthook\@empty
\def\ps@firstpage{\count@=0000
\ps@plain
\let\@oddfoot\@empty
\let\@evenfoot\@oddfoot
\def\@oddhead{\@serieslogo\hss}%
\let\@evenhead\@oddhead
}
\let\ps@headings\ps@empty
\let\ps@plain\ps@empty
\def\article@logo{%
\set@logo{%
\hsize\textwidth
{\Large\textit{Comptes Rendus\\\CDRcrasseries}\\[2pt]}%
\ifx\@empty\currentvolume
\textbf{Draft}
\else
\textbf{0000},
\currentvolume
\ifx\currentissue\@empty\else, \no\currentissue
\fi\fi\\
\mbox{}%
\hfill
\raisebox{-3mm}[0pt][0pt]{\includegraphics[height=16mm]{Logo_Academie_grisfonce.pdf}}\\[18bp]
\centerline{\rule{.8\textwidth}{1.5bp}}
}%
}
\def\vspace*{-17mm}{\vspace*{-17mm}}
\def\lastname#1{#1}
\def\junior#1{\unskip, #1}%
\renewcommand{\@pnumwidth}{4.5em}
\def\@adminfootnotes{}
\def\enddoc@text{}
\fi
\ifcdr@volume
\def\@makefront{%
\IncludeSom{\cdr@journal -Sommaire}%
\cleardoublepage
\ifcdr@SOM
\@makeSOM
\begingroup\makeatletter
\newcommand\SomLine[9]{%
\gdef\cdr@shortresps{##1}%
\gdef\cdr@resps{##2}%
\gdef\cdr@specialtitle@french{##5}%
\gdef\cdr@specialtitle@english{##4}%
\gdef\cdr@specialabstract@french{##7}%
\gdef\cdr@specialabstract@english{##6}%
}%
\InputIfFileExists{./\CDRidvol.cdrsom}{%
\message{** Fichier \CDRidvol.cdrsom charge **^^J}}{%
\message{** Pas de fichier \CDRidvol.cdrsom pour ce volume !**^^J}}%
\makeatother\endgroup
\fi
}
\def\@makebody{%
\ifcdr@newsom\ShellEscape{rm -f \[email protected]}\fi
\pagenumbering{arabic}%
\setcounter{page}{\@FirstPage}%
\tkkp={\setpage}%
}
\def\enddoc@text{}
\else
\ifcdr@sommaire
\AtEndOfClass{%
\begingroup\makeatletter
\newcommand\SomLine[9]{%
\gdef\cdr@shortresps{#1}%
\gdef\cdr@resps{#2}%
\gdef\cdr@specialtitle@french{#5}%
\gdef\cdr@specialtitle@english{#4}%
\gdef\cdr@specialabstract@french{#7}%
\gdef\cdr@specialabstract@english{#6}%
}
\InputIfFileExists{./\CDRidvol.cdrsom}{%
\message{** Fichier \CDRidvol.cdrsom charge **^^J}%
}%
{\message{** Pas de fichier \CDRidvol.cdrsom pour ce volume !**^^J}}%
\makeatother\endgroup
}
\AtBeginDocument{%
\newcommand\SomLine[9]{%
\def\lng{#3}%
\def\LNG{french}%
\par \addpenalty\@secpenalty\addvspace{\medskipamount}%
\begingroup
\hyphenpenalty50\def\\{\allowbreak}%
\parindent0pt
\rightskip\@pnumwidt
{\bfseries\ignorespaces#2\unskip}\@par
\parfillskip-\@pnumwidth
\href{\CDRidvol.pdf\#page.#8}{\ifx\lng\LNG#5\else#4\fi}\unskip\leavevmode
\nobreak\hskip2pt\nobreak \leaders\hbox{$\m@th \mkern 1.5mu\hbox{.}\mkern 1.5mu$}\hfill \nobreak
\hb@xt@\@pnumwidth{\@tocpagenum{\ignorespaces#8\unskip-\ignorespaces#9\unskip}}%
\par
\endgroup
}
}
\else
\AtEndOfClass{%
\begingroup\makeatletter
\newcommand\SomLine[9]{%
\gdef\cdr@shortresps{#1}%
\gdef\cdr@resps{#2}%
\gdef\cdr@specialtitle@french{#5}%
\gdef\cdr@specialtitle@english{#4}%
\gdef\cdr@specialabstract@french{#7}%
\gdef\cdr@specialabstract@english{#6}%
}%
\InputIfFileExists{../\CDRidvol.cdrsom}{%
\message{** Fichier \CDRidvol.cdrsom charge pour cet article **^^J}}{%
\message{** Pas de fichier \CDRidvol.cdrsom pour ce volume !**^^J}}%
\makeatother\endgroup
}
\fi
\fi
\NumberTheoremsIn{}
\setcounter{tocdepth}{3}
\ifcdr@ThmDefs\OneNumberAllTheorems\fi
\def\th@plain{%
\thm@bodyfont{\normalfont\itshape}%
\thm@headfont{\normalfont\bfseries}%
\thm@notefont{\normalfont\bfseries}%
\thm@headpunct{.\kern.2em\ignorespaces}%
\let\thm@indent\noindent
\def\thm@space@setup{%
\thm@preskip=\smallskipamount
\thm@postskip=\thm@preskip}%
\itshape
}
\def\th@remark{%
\thm@bodyfont{\normalfont}%
\thm@headfont{\normalfont\bfseries}%
\thm@notefont{\normalfont\bfseries}%
\thm@headpunct{.\kern.2em}%
\let\thm@indent\noindent
\def\thm@space@setup{%
\thm@preskip=\smallskipamount
\thm@postskip=\thm@preskip}%
}
\let\th@definition\th@plain
\renewenvironment{proof}[1][\proofname]{\par
\pushQED{\qed}%
\normalfont \topsep6\p@\@plus6\p@\relax
\trivlist
\item[\hskip\cdr@proofindent
{\providecommand\printref{\relax}%
\bfseries
#1\@addpunct{\cdr@proofpunct}}]\ignorespaces
}{%
\popQED\endtrivlist\@endpefalse
}
\def\@startsection#1#2#3#4#5#6{%
\if@noskipsec \leavevmode \fi
\par \@tempskipa #4\relax
\@afterindenttrue
\ifdim \@tempskipa <\z@ \@tempskipa -\@tempskipa \@afterindentfalse\fi
\if@nobreak \everypar{}\else
\addpenalty\@secpenalty\addvspace\@tempskipa\fi
\@ifstar{\@dblarg{\@sect{#1}{\@m}{#3}{#4}{#5}{#6}}}%
{\@dblarg{\@sect{#1}{#2}{#3}{#4}{#5}{#6}}}%
}
\def\@seccntformat#1{%
\protect\textup{\protect\@secnumfont
\csname the#1\endcsname
\protect\@secnumpunct
}%
}
\def\@sectionfamily{\normalfont}
\def\@sectionfont{\@sectionfamily\large\bfseries}
\let\contentsnamefont\@sectionfont
\def\@secnumfont{\selectfont}
\def\@subsectionfont{\@sectionfamily\large\itshape}
\def\@subsubsectionfont{\@sectionfamily\itshape}
\def\@paragraphfont{\normalfont\bfseries}
\def\section{\@startsection{section}{1}%
\z@{-1.5\linespacing\@plus-.5\linespacing\@minus-.5\linespacing}{.8\linespacing\@plus.2\linespacing\@minus.3\linespacing}%
{\@sectionfont}}
\def\subsection{\@startsection{subsection}{2}%
\z@{-1.5\linespacing\@plus-.5\linespacing\@minus-.5\linespacing}{.8\linespacing\@plus.2\linespacing\@minus.3\linespacing}%
{\normalfont\@subsectionfont}}
\def\subsubsection{\@startsection{subsubsection}{3}%
\z@{.5\linespacing\@plus.5\linespacing}{.5\linespacing}%
{\normalfont\@subsubsectionfont}}
\def\paragraph{\@startsection{paragraph}{4}%
\z@{.5\linespacing\@plus.5\linespacing}{-\fontdimen2\font}%
{\normalfont\@paragraphfont}}
\newcount\tochyphenpenalty
\tochyphenpenalty\@M
\def\@tocline#1#2#3#4#5#6#7{\relax
\ifnum #1>\c@tocdepth
\else
\par \addpenalty\@secpenalty\addvspace{#2}%
\begingroup \hyphenpenalty\tochyphenpenalty
\@ifempty{#4}{%
\@tempdima\csname r@tocindent\number#1\endcsname\relax
}{%
\@tempdima#4\relax
}%
\parindent\z@ \leftskip#3\relax
\advance\leftskip\@tempdima\relax
\advance\rightskip\@pnumwidth
\parfillskip-\@pnumwidth
{#5{\leavevmode\hskip-\@tempdima #6}}%
\nobreak
\xleaders\hbox to1ex{\normalfont\hss.\hss}\hfil\nobreak
\hbox to\@pnumwidth{\@tocpagenum{#7}}\par
\nobreak
\endgroup
\fi}
\renewcommand{\tocpart}[3]{%
\indentlabel{\@ifnotempty{#2}{\ignorespaces#1 #2.\kern1ex}}#3}
\let\tocsection\tocpart
\let\tocsubsection\tocsection
\let\tocsubsubsection\tocsection
\let\tocparagraph\tocsection
\def\l@part{\@tocline{-1}{12pt plus2pt}{0pt}{}{\bfseries\smf@boldmath}}
\def\l@section{\@tocline{1}{0pt}{0pc}{}{}}
\def\l@subsection{\@tocline{2}{0pt}{1pc}{}{}}
\def\l@subsubsection{\@tocline{3}{0pt}{2pc}{}{}}
\def\l@paragraph{\@tocline{5}{0pt}{3pc}{}{}}
\def\l@figure{\@tocline{0}{3pt plus2pt}{0pt}{}{}}
\let\l@table\l@figure
\def\@smfprotect{\let\smfcr\\\def\\{\protect\smfcr}}
\DeclareRobustCommand{\MakeUppercase}[1]{{%
\@smfprotect\upchars@
\protected@edef\reserved@a{#1}\uppercasenonmath\reserved@a\reserved@a}}
\protected@edef\MakeUppercase#1{\MakeUppercase{#1}}
\AtBeginDocument{%
\csname url@samestyle\endcsname
\begingroup
\ifx\@dateposted\@empty
\ifnum\currentvolume=\z@%
\def\setpdftimes#1#2{\gdef\@dateposted{#2}}%
\ifcdr@published
\InputIfFileExists{\jobname-pdftimestamp.txt}{}{
\else
\cdr@mkfile{\jobname-pdftimestamp.txt}{\string\setpdftimes{\pdftex@now}{\@dateprod}}
\edef\cdr@tmp@{\pdftex@now}
\expandafter\setpdftimes\expandafter{\cdr@tmp@}{\@dateprod}%
\fi
\cdr@mkfile{\jobname-dateposted.txt}{\string\dateposted{\@dateposted}}%
\else
\InputIfFileExists{\jobname -dateposted.txt}{}{
\fi
\fi
\endgroup
}
\emergencystretch2em
\arraycolsep 1.5pt
\def{\normalfont\itshape et al.}{{\normalfont\itshape et al.}}
\def\vadjust{\pagebreak}{\vadjust{\pagebreak}}
\newcommand\PB{\pagebreak}
\newcommand\SP{\vadjust{\pagebreak}}
\newcommand\SPB{\vadjust{\pagebreak}\break}
\newcommand\PBI{\pagebreak\noindent}
\ExplSyntaxOn
\NewDocumentCommand {\displayissue} {mm} {
\ifx\currentissue\@empty\else
\exp_args:NV \str_if_in:nnF \currentissue {G}
{\exp_args:NV \str_if_in:nnTF \currentissue {S} {#2} {#1} \currentissue}
\fi
}
\ExplSyntaxOff
\endinput
\let\le\leqslant
\let\ge\geqslant
\let\leq\leqslant
\let\geq\geqslant
\let\emptyset\varnothing
\def\sfrac#1#2{{#1}/{#2}}
\defloc.\kern3pt cit.\kern.3em{loc.\kern3pt cit.\kern.3em}
\defcf.\kern.3em{cf.\kern.3em}
\defe.g.\kern.3em{e.g.\kern.3em}
\defi.e.\kern.3em{i.e.\kern.3em}
\def\text{resp.}\kern.3em{\text{resp.}\kern.3em}
\defetc.\kern.3em{etc.\kern.3em}
\let\eccetc.\kern.3em
\let\oldsetminus\setminus
\let\setminus\smallsetminus
\newcommand{\arXiv}[1]{\href{http://arxiv.org/abs/#1}{\texttt{arXiv\string:\allowbreak#1}}}
\newcommand{\eprintHal}[1]{\href{https://hal.archives-ouvertes.fr/hal-#1}{\texttt{hal-#1}}}
\AtEndOfClass{%
\let\oldbigoplus\bigoplus
\let\oldbigcup\bigcup
\let\oldbigcap\bigcap
\let\oldbigotimes\bigotimes
\let\oldbigwedge\bigwedge
\let\oldcoprod\coprod
\def\vadjust{\pagebreak}{\vadjust{\pagebreak}}
\newcommand\PB{\pagebreak}
\newcommand\SP{\vadjust{\pagebreak}}
\newcommand\SPB{\vadjust{\pagebreak}\break}
\newcommand\PBI{\pagebreak\noindent}
}
\AtBeginDocument{%
\url@samestyle
\def\bigoplus{\mathop{\textstyle\oldbigoplus}\displaylimits}%
\def\bigcup{\mathop{\textstyle\oldbigcup}\displaylimits}%
\def\bigcap{\mathop{\textstyle\oldbigcap}\displaylimits}%
\def\bigotimes{\mathop{\textstyle\oldbigotimes}\displaylimits}%
\def\bigwedge{\mathop{\textstyle\oldbigwedge}\displaylimits}%
\def\coprod{\mathop{\textstyle\oldcoprod}\displaylimits}%
}
\section{Introduction, meta-data commands}
This is the beginning of our article.
\subsection{Title}
The command for the title is: \verb#\title#.
The \verb#\maketitle# command must be put after the abstract.
\subsection{Citations}
The bibliography must be built using bibtex. A sample of a bibtex file
\texttt{samplebib.bib} is with this sample.
The references must be refered to in the article by using the
\verb#\cite# command, which produces for example \cite{KNU} or
\cite{LL} (see also the comments in \texttt{samplebib.bib}).
\section{Figures}
The article being compiled with pdflatex, the figures should also be
in PDF (esp. for vector graphics, bitmap graphics should be of good
enough quality and can be PNG or JPEG). The inclusion of the figure is
done using the following commands.
The parameter \emph{xxx}, a real number between $0.0$ and $1.0$, indicates the
width the figure should take in the page.
One can refer to the figure with \verb+\ref{refname}+, which gives
for example:
Figure \ref{fig:exmp} is an example of figure.
\begin{figure}[tbp]
\includegraphics[width=0.4\linewidth]{Logo_Academie_grisfonce.pdf}
\caption{Example of figure.}
\label{fig:exmp}
\end{figure}
To refer to a specific definition, theorem, etc.,
put \verb+\label{labelname}+ inside the corresponding environment and use
\verb+\ref{labelname}+ in text to point to this definition, theorem,
etc.
Here is an example:
\begin{theo}\label{th:true}
Most theorems are true.
\end{theo}
\begin{proof}
Th. \ref{th:true} is obviously true.
\end{proof}
\begin{exam}\label{ex:good}
This should look like a good example.
\end{exam}
\begin{rema}
Can an example like Ex.~\ref{ex:good} give some insight in
Th.~\ref{th:true}'s proof?
\end{rema}
\bibliographystyle{crunsrt}
\nocite{*}
\section{Introduction}
Analog gravity \cite{Barcelo2005} is based on the use of table-top set-ups
to simulate the physics of quantum fields on curved spacetimes. This approach is based on the formal equivalence of the equations of motion governing the dynamics of the two systems under suitable conditions. Since the original proposal by Unruh \cite{unruh1981}, curved space-time geometries have been realized in a large variety of systems, including surface waves on classical fluids \cite{rousseaux2008}, ultra-cold atoms \cite{lahav2010}, polariton fluids \cite{nguyen2015} and optical systems \cite{philbin2008, vocke2018}.
The phenomena considered in this context typically involve strong gravity environments that escape direct astrophysical investigations, as in the vicinity of a black hole or during the fast expansion of the early universe. Most remarkable examples are Hawking emission from black hole horizons~\cite{belgiorno2010, weinfurtner2011, euve2016, steinhauer2016, munoz2019, drori2019}, cosmological particle creation and dynamical Casimir effect~\cite{wilson2011, jaskula2012,eckel2018,hung2013,wittemer2019,steinhauer2022analogue}, Penrose rotational superradiance \cite{torres2017, solnyshkov2019,cromb2020,braidotti2022measurement}.
In this context, the term rotational superradiance refers to the amplified scattering of waves from a rotating object at the expenses of the rotational energy of the object itself~\cite{Brito_2020}. It is a fully classical phenomenon and is predicted to occur in a variety of spacetime geometries, in particular in the surroundings of massive rotating objects. In this case, dragging of the spacetime by the rotating object gives rise to a so-called ergoregion, namely a spatial region supporting negative-energy light and matter modes.
The process of superradiance can be microscopically explained as the amplified reflection of a positive-energy wave impinging onto the boundary of the ergoregion, accompanied by the transmission of a negative-energy wave inside the ergoregion. In the presence of some reflecting element, self-amplification of one of the outgoing modes may take place, leading to an exponentially growing perturbation. Depending on the location of the reflecting element on either the inner or the outer side of the ergosurface, the instability takes the name of ergoregion instability~\cite{comins1978, friedman1978} or black hole-bomb instability~\cite{cardoso2004}.
Rotating Bose-Einstein condensates (BEC) are promising platforms to study superradiant phenomena~\cite{oliveira2018,Brito_2020}.
A number of theoretical studies have investigated superradiance in configurations featuring a single vortex located at the center of a cylindrically symmetric trap as well as in planar geometries involving synthetic magnetic fields~\cite{giacomelli2021understanding}. In particular, the analog gravity perspective has offered transparent physical explanations for the hydrodynamic instability of multiply-charged single vortex configurations~\cite{fetter2001vortices, fetter2009rotating,fedichev1999dissipative, rokhsar1997vortex}
by connecting them to ergoregion instabilities of rotating massive objects in gravitation~\cite{giacomelli2020ergoregion}.
In the last decade, several authors have started investigating two-component BECs as a promising new platform for analog gravity~\cite{fischer2004quantum,visser2005massive,liberati2006analogue,butera2017black}.
Such systems are characterized by the existence of two independent branches of collective excitations, associated to perturbations of the total density and of the spin density, featuring very different values of the speed of sound and of the healing length~\cite{cominotti2022observation}.
As a consequence of the much faster speed of density-sound, configurations displaying horizons for spin-sound waves can be engineered while keeping the superflow sub-sonic with respect to density-sound and therefore robust against perturbations coupling to the density. Furthermore, the direct experimental observation of the system dynamics is facilitated by the significantly larger size of the spin-healing length and by the availability of interferometric techniques based on a coherent mixing of the two components which allow to image all quadratures of the quantum field. On the long run, all this suggests the possibility of investigating quantum entanglement features of the superradiant emission~\cite{finazzi2014entangled}.
Within this general framework, the goal of this paper is to explore ergoregion instability phenomena in two-component BECs displaying a single quantized vortex in both components. Thanks to the very different value of the density- and spin-sound speeds, the ergosurface for spin waves is no longer bound to sit inside or in the close vicinity of the vortex core as it instead happens in single component BECs, but can be pushed far away from the vortex core into the external region where the density is approximately constant. This is of great interest for analog gravity as it leaves a wider space to investigate the dynamics of the quantum field in the ergoregion and disentangle the different effects at play
Beyond analog gravity, an active interest for this physics is also coming from a purely quantum gas perspective: a recent numerical work~\cite{kuopanportti2019} has pointed out the rich phenomenology of vortex splitting in harmonically trapped two-component BECs: singly-charged vortices are dynamically unstable for sufficiently strong repulsive inter-component interactions, whereas doubly-charged vortices dispose of several decay channels. Experimentally, the splitting of a singly-charged vortex into a pair of half-quantized vortices has been observed in spinor exciton-polariton superfluids \cite{manni2012dissociation} and antiferromagnetic spinor BECs \cite{seo2015half}.
Here, we take advantage of the analog gravity point of view to develop a transparent physical interpretation of vortex physics in two-component BECs in terms of superradiance effects. Use of this formalism allows us to highlight a number of features that are specific to two-components BECs and characterize them in view of forthcoming experimental studies. In addition to recovering the main predictions of previous works, e.g. on vortex splitting instabilities, we identify regimes where a clean analog of the ergoregion instability of massive rotating objects is visible.
The structure of the paper is the following: in Secs.\ref{sec:bogo} and \ref{sec:vortices} we give a brief review of the theoretical tools that we are going to use in the following of the paper. In Sec. \ref{sec:lda} we tackle the vortex stability problem within the local density approximation (LDA) that provides a clear qualitative understanding of the physical mechanisms. A more quantitative analysis using a Bogoliubov characterization of the collective modes and a Gross-Pitaevskii (GP) simulation of the time-dependence is then provided in Secs.\ref{sec:spectra} and \ref{sec:GP}, respectively. A summary of our conclusions and of further extensions of our work is finally given in Sec.\ref{sec:concl}. Additional information on the numerical techniques adopted and on the independence of the results on the shape of the confinement potential are given in the Apppendices.
\section{Two-component condensates and Bogoliubov theory of density and spin collective excitations}
\label{sec:bogo}
In this work, we are interested in the dynamics of a two-component condensate subject to an external potential $V(r)$ and to a resonant coherent coupling of strength $\Omega\ge 0$ between the two components. Let us focus on a spin-symmetric configuration of equal masses $m_1=m_2\equiv m$ and intraspecies interactions $g_1 = g_2 \equiv g>0$. Moreover, we assume to deal with a miscible mixture, $|g_{12}|< g$, where $g_{12}$ defines the interspecies interaction strength.
The dynamics of such a system is ruled by two coupled GP equations \cite{stringari_2016}:
\begin{align}
\begin{split}
i \hbar \partial_t \psi_1 &= \Big( K + V + g n_1 + g_{12} n_2 \Big) \psi_1 - \frac{\hbar \Omega}{2}\psi_{2} \\
i \hbar\partial_t \psi_2 &= \Big(K + V + g n_2 + g_{12}n_1 \Big)\psi_2 - \frac{\hbar \Omega}{2} \psi_1
\end{split},
\label{eq:GP}
\end{align}
where $K = -\hbar^2 \nabla^2/2m $ is the kinetic energy operator, $\psi_j$ is the order parameter of the $j$-th component, and $n_j = |\psi_j|^2$ is the associated atomic density, normalized to the number of particles in the $j$-th state, $N_j$. As long as $N_1 = N_2$, the ground state of the system is everywhere unpolarized, with identical order parameters $\psi_1^o(r)=\psi_2^o(r)$ and density profiles $n_1(r)=n_2(r)$ for the two components of the mixture.
In addition to the $\mathbb Z_2$ spin-symmetry related to the exchange of the two components, for a vanishing coherent coupling $\Omega = 0$ the system has two additional $U(1)$ symmetries associated to the independent conservation of the number of particles $N_{1,2}$ in the two states. The presence of the coherent coupling allows particles to be transferred from one component to the other and therefore breaks one of these two $U(1)$ symmetries. In this case, we are left with a single $U(1)$ symmetry related to the conservation of the total atom number $N=N_1+N_2$ only.
A deep insight on the physical properties of the system is provided by the Bogoliubov theory of small perturbations of the collective excitations on top of the ground state. Thanks to the $\mathbb Z_2$ spin-symmetry of the problem under $\psi_1\leftrightarrow \psi_2$, excitations can be classified as symmetric, density ($d$) modes, and antisymmetric, spin ($s$) modes~\cite{Abad_2013}.
\subsection{Uniform condensate}
For a uniform condensate of total density $n$, the energy of collective excitations is given by the Bogoliubov dispersion relations as a function of the excitation wavevector $k$~\cite{Abad_2013}:
\begin{align}
(\hbar \omega_d)^2 &= \frac{\hbar^2 k^2}{2m} \left(\frac{\hbar^2 k^2}{2m} + 2\mu_d \right) \\
(\hbar \omega_s)^2 &= \left(\frac{\hbar^2 k^2}{2m} + \hbar \Omega \right)\left(\frac{\hbar^2 k^2}{2m} + \hbar \Omega + 2\mu_s \right)\, ,\label{spin_disp}
\end{align}
where $\mu_d = (g+g_{12})n/2$ and $\mu_s = (g-g_{12})n/2$ are the density and spin chemical potentials.
In the absence of Rabi coupling $\Omega=0$, both dispersions are linear at low momenta, with speed of sound equal to $c_d = \sqrt{\mu_d/m}$ and $c_s = \sqrt{\mu_s/m}$ in the density and spin channel, respectively; the presence of two gapless Goldstone modes is associated with the conservation of the total and relative numbers of particles, $N_1\pm N_2$. For a finite Rabi coupling $\Omega\ne 0$, instead, the breaking of the $U(1)$ symmetry related to the relative phase of the order parameters of the two components is responsible for the opening of a gap of size $\omega_p = \sqrt{\Omega (2\mu_s/\hbar + \Omega)}$ in the dispersion of spin modes, which acquire a finite effective mass~\cite{cominotti2022observation}.
\subsection{Spatially inhomogeneous condensate}
For spin-symmetric, but spatially inhomogeneous systems, the Bogoliubov spectrum can only be obtained numerically (details are given in the Appendix): it is first neceassary to compute the density profile of a stationary state, that is, of an eigensolution of the stationary GP equation,
\begin{equation}
\mu \Phi(\vect r) = \left[ K + V(\mathbf r) + (g+g_{12})\frac{n(\vect r)}{2} -\frac{\hbar\Omega}{2} \right] \Phi(\vect r)\,:
\label{stat_GP}
\end{equation}
where $\Phi({\vect r})$ is the order parameter of both components, $\psi_1^o(\vect r)=\psi_2^o(\vect r)=\Phi(\vect r)$ and $\mu$ is the oscillation frequency of the matter field. Of course $\mu$ takes the meaning of chemical potential when we consider the lowest-order stationary state of \eqref{stat_GP} corresponding to the ground state of the condensate. In practice, the stationary state is obtained through imaginary-time evolution.
The frequencies $\omega_i$ and spatial shape $[u_i({\vect r}),v_i({\vect r})]$ of the collective excitation modes on top of the stationary state are then computed as the eigenvalues and eigenvectors of the Bogoliubov operators in the density $d$ and spin $s$ sectors:
\begin{align}
\mathscr{L}_{d,s} &= \begin{pmatrix}
\mathcal D_{d,s} & \mu_{d,s}(\vect r) \\
-\mu_{d,s}(\vect r) & - \mathcal D_{d,s}
\end{pmatrix},
\label{Bogomat}
\end{align}
where the operators on the diagonal read:
\begin{align}
\mathcal D_d &= K + V(\vect r) + 2\mu_d(\vect r) - \mu -\hbar \Omega/2\\
\mathcal D_s &= K + V(\vect r) + \mu_s(\vect r) +\mu_d(\vect r) -\mu+\hbar \Omega/2
\end{align}
and the interaction energies acquire a spatial dependence due to the inhomogeneity of the density profile, $\mu_{d,s}(\vect r)={(g\pm g_{12})}\,n({\vect r})/2$.
For each collective mode, the time-evolving perturbation of the order parameter of each species then has the form:
\begin{equation}
\psi_i(\vect r,t) = \Phi(\vect r,t)+\delta\psi_i(\vect r,t)=
\Phi(\vect r,t) + \Big[ u_i(\vect r) e^{-i\omega t}+ v_i^*(\vect r) e^{i\omega t} \Big] e^{-i\mu t/\hbar}\,, \label{bogoperturb}
\end{equation}
where the two components $u_i$ and $v_i$ of the eigenvector are often referred to as particle and antiparticle components. Depending on which of the two dominates over the other, excitation can have positive, negative or zero Bogoliubov norm:
\begin{equation}
\|\delta\psi_i\|^2 = \int \sum_{i=1,2}\Big(|u_i(\vect r)|^2 - |v_i(\vect r)|^2 \Big) \text{d} \vect r\,.
\end{equation}
In this work we exploit the analysis of the Bogoliubov spectrum~\cite{castin2001bose} to study the stability of stationary states, in particular vortex configurations. In stable configurations, all frequencies are real-valued and, for each excitation mode, have the same sign as the norm.
Modes with positive frequency and negative norm (or vice versa) have instead a negative energy and their existence is a signature of an energetic (or thermodynamic) instability, whereas the configuration is dynamically unstable if some modes have a complex frequency. Thanks to the pseudo-hermiticity of the Bogoliubov matrix, these zero-norm modes always come in pairs, whose frequencies share the same real part and have opposite imaginary parts, $\omega = \eta \pm i\Gamma$. The imaginary part $\Gamma$ represents the rate of exponential growth of the dynamically unstable mode.
\section{Analog gravity and vortices in two-component BECs}
\label{sec:vortices}
Analog gravity is based on the formal equivalence between the Klein-Gordon equation for the dynamics of a scalar field on a curved spacetime and the linearized GP equations in the hydrodynamic regime in which all excitations propagate with the same speed of sound. In the particular case of excitations on top of a two-component BEC, the effective metric tensors of density and spin excitations are different, but share
the general form \cite{Barcelo2005}:
\begin{equation}
(g_{d,s})_{\mu\nu} = \frac{n}{c_{d,s}}\begin{pmatrix}
-(c_{d,s}^2 - |\vect v|^2) & -\vect v^T \\ -\vect v & \mathbb I
\end{pmatrix}
\label{analogmetric}
\end{equation}
where $\vect v(\vect r)$ is the (irrotational) velocity field associated to the background condensate flow:
\begin{equation}
\vect v = \frac \hbar m \nabla \big( \text{arg}(\Phi) \big)\,.
\end{equation}
Let us now focus on the specific case of two-dimensional systems. Even though condensation is strictly speaking not possible in 2D, effectively two-dimensional atomic superfluids can be realized experimentally by tightly confining the condensate along one direction, say $z$, imposing a much weaker confinement in the orthogonal plane $(x,y)$, and working at a sufficiently low temperature for the coherence length to exceed the finite size of the system. The main effect of the vertical confinement is a geometrical renormalization of the interaction strength.
In the following we focus our attention on a radially symmetric two-dimensional BEC confined in the $(x,y)$ plane. Thus $\vect r = (x,y)$ and polar coordinates are defined as $\theta = \arctan(y/x)$ and $r = \sqrt{x^2+y^2}$. For such configurations, the analog spacetime defined by \eqref{analogmetric} may display a density and a spin ergosurface, defined by $|\vect v| = c_{d,s}$, while analog event horizons are found if the radial fluid velocity equals the speed of sound, $|\vect v\cdot \vect u_r| = c_{d,s}$ and then exceeds it. Of course, a radial flow requires the presence of some particle non-conservation mechanism. A study of this physics goes beyond this paper, which is restricted to purely tangential flows with no radial component.
The ground state of a non-rotating condensate is well approximated by the Thomas-Fermi (TF) solution:
\begin{equation}
\frac{n_\text{TF}(r)}{2} = |\Phi_\text{TF}(r)|^2 = \frac{\mu+\hbar\Omega/2 - V(r)}{g+g_{12}}\,.
\end{equation}
In the presence of rotation, a spin-symmetric quantized vortex is a solution of the stationary Gross-Pitaevski equation \eqref{stat_GP} of the form:
\begin{equation}
\Phi(\vect r,t) = f(r) e^{iL\theta} e^{-i\mu t/\hbar}
\label{eq:Phi_L}
\end{equation}
where the radial function $f(r)$ defines the vortex profile and $L$ is the integer-valued vortex charge.
The density profile of a vortex differs from the TF one only in the region of the vortex core, while $n(r) \to n_\text{TF}(r)$ at larger distances. For a charge-$L$ vortex, the core has a size on the order of $\sqrt{L}\,\xi_d$, where $\xi_{d}=\sqrt{\hbar/mc_{d}}$ is the density healing length. The spin healing length is defined in an analogous way as $\xi_{s}=\sqrt{\hbar/mc_{s}}$.
As usual, the velocity flow of the vortex is tangential and inversely proportional to the distance $r$ from the center:
\begin{equation}
\vect v(r) = \frac{\hbar L}{mr} \vect u_\theta
\end{equation}
Assuming the two speeds of sound $c_{d,s}$ are slowly varying, the analog ergosurfaces for density and spin modes, defined by $|{\vect v}|=c_{d,s}$, are approximately located at $r\sim L\xi_{d,s}$.
An example of promising configuration is shown in Fig.\ref{fig:speed}, where the local values of $c_{d,s}$ have been computed using the exact density profile of an $L=2$ vortex in a very wide trap and the interactions constants have been chosen such that the large-$r$ asymptotic values $c_s\ll c_d$ and $\xi_s \gg \xi_d$. This condition guarantees that the the spin ergoregion is located well outside the vortex core, where the density profile is approximately flat.
The absence of a radial flow imposed by particle condensation prevents the system from having an analog event horizon, that would act as an absorbing boundary: the vortex is thus the ideal background to study analogs of the ergoregion instabilities of rotating space-times.
\begin{figure}[t
\centering
\includegraphics[width = 0.55 \linewidth]{Fig1.pdf}
\caption{Comparison between the density and spin sound speed (red and blue line respectively), and the flow velocity (dashed black line) in a vortex configuration. The location of the ergosurface for density (spin) excitations, indicated by the dotted red (blue) line, is found as the intersection between $|\vect v(r)|$ and $c_d(r)$ [$c_s(r)$]. All velocities are normalized to the large-distance value of the speed of density sound, $c_d^\infty \equiv c_d(r\to\infty)$.
The plot is obtained using the numerically calculated density profile of a vortex of charge $L=2$; interactions are set to $g_{12}=0.93 g$, giving $c_d/c_s=\xi_s/\xi_d \sim 5.25$. }
\label{fig:speed}
\end{figure}
Taking advantage of the cylindrical symmetry, the Bogoliubov excitations of frequency $\omega$ and angular momentum $M$ on top of a charge-$L$ vortex can be written in the factorized form
\begin{align}
u_i(\vect r) &= u_i(r) e^{iM\theta} e^{iL\theta}\\
v_i(\vect r) &= v_i(r) e^{iM\theta} e^{-iL\theta}
\end{align}
and the stability of the vortex configuration can be assessed by solving the effectively 1D radial Bogoliubov problem for $u_i(r)$ and $v_i(r)$.
The Bogoliubov matrices \eqref{Bogomat} involve the kinetic energy operator. In polar coordinates this reads:
\begin{equation}
K = -\frac{\hbar^2}{2m} \left( \dpar{^2}{r^2} + \frac{1}{r}\dpar{}{r} - \frac{(L\pm M)^2}{r^2} \right)
\end{equation}
where the $+$ ($-$) sign must be chosen when applying the operator to the particle (antiparticle) component $u$ ($v$).
\section{Local density approximation}
\label{sec:lda}
\begin{figure*}[t]
\centering
\includegraphics[width = \linewidth]{Fig2.pdf}
\caption{Properties of the dispersion relation in the LDA \eqref{bogodispM} (obtained with $L=2,M=1,g_{12}/g= 0.93,\Omega = 0$, $\mu_d/\mu_s = 27.5$). (a-c) Plot of the dispersion at different radii, showing the availability of positive and negative norm modes at a generic frequency $\omega$, indicated by the red horizontal line: solid (dotted) black lines refer to the upper (lower) branch. (d) Plot of the maximum frequency of the lower branch (black line) and minimum frequency of the upper branch (red line); the red (gray) area indicates the region of the plot where positive (negative) energy modes are available. The white area represents the centrifugal gap. The dashed black line indicates the spin ergosurface location, given by \eqref{ergoposition}.}
\label{bogodisp}
\end{figure*}
In general, the Bogoliubov problem on top a spatially inhomogeneous vortex configuration can hardly be solved analytically.
Before proceeding with the numerical diagonalization of the Bogoliubov matrices which will be the subject of the next Section, we start here with an intuitive discussion of the physics on the basis of a local density approximation (LDA).
Let us assume that the condensate is enclosed in a large box trap, so that the density can be considered approximately uniform far from the vortex core, $n(r\gg L\xi_d) \simeq n^\infty$.
Within a small spatial region around $\vect r$, excitations of angular momentum $M$ can be approximated as plane waves with wavevector $\vect k = k\vect u_r + (M/r)\vect u_\theta$, so that the Bogoliubov wavefunctions $u(\vect r),v(\vect r) \propto e^{ikr}e^{iM\theta}$. By exploiting the LDA, the $k$-dependent Bogoliubov dispersion relation in the radial direction is obtained by taking the uniform result \eqref{spin_disp} and replacing $k^2 \to k^2+M^2/r^2$. The effect of the finite $M$ is thus to open a gap in the dispersion relation and is equivalent to having an additional space-dependent coherent coupling $\Omega_M(r) = \hbar M^2/2m r^2$.
On top of this, one needs to include the overall $M$-dependent Doppler shift $\delta(r)\equiv \vect k\cdot\vect v = \hbar LM/m r^2$ due to the tangential velocity flow. Due to the absence of in-going flow, this Doppler term does not tilt the dispersion relation as a function of the radial $k$, but only rigidly shifts it vertically along the frequency axis. This shift is crucial to bring negative-norm (positive-norm) modes up (down) to positive (negative) frequencies.
Combining all these elements together, we obtain the final formula
\begin{equation}
\omega_{s}^\pm (k) = \delta \pm \sqrt{\left(\frac{\hbar k^2}{2m} + \tilde\Omega\right) \left(\frac{\hbar k^2}{2m} + \tilde\Omega + \frac{2\mu_s^\infty}{\hbar} \right) }
\label{bogodispM}
\end{equation}
where the sign $\pm$ refers to the positive and negative branches, respectively, and the effective chemical potential and coupling strength are $\mu_{s}^\infty = (g-g_{12})n^\infty/2$ and $\tilde\Omega(r) = \Omega + \Omega_M(r)$. From this formula, the space-dependent centrifugal gap turns out to be $\Delta(r) \equiv \sqrt{ \tilde\Omega(r) \big[\tilde\Omega(r) +2\mu_s^\infty/\hbar\big] }$.
Because of the Doppler shift and the centrifugal gap, for a generic positive frequency $\omega$, there can be positive norm, negative norm or no modes available depending on the radial position $r$. Three specific examples for different radial positions $r$ are shown in Fig.\ref{bogodisp}(a-c) for the most relevant $\Omega=0$ case; including a finite $\Omega$ would simply result in a larger centrifugal gap. A summary of the results for generic $(r,\omega>0)$ is displayed in panel (d): here, the red and gray shaded areas indicate the $(r,\omega)$ regions where positive and negative norm modes are available, respectively. The boundaries of these regions identify two frequency-dependent positions in real space, $r_\pm(\omega)$, indicated by the red and black solid lines in Fig.\ref{bogodisp}(d). In particular, $r_-(\omega)$ can be thought as a sort of effective frequency-dependent ergosurface position.
An ergoregion instability occurs when a negative norm mode in the inner part of the system is coupled to a positive norm mode living in the outer region, so that the two combine into a pair of zero-norm modes with complex frequencies $\omega \pm i\Gamma$~\cite{giacomelli2020ergoregion}.
In other words, the dynamical instability can be thought as resulting from a tunnelling process between the red- and gray-shaded regions in Fig.\ref{bogodisp}(d), with a tunneling-mediated instability rate determined by the real-space width of the forbidden white region separating them. According to this interpretation, modes with smaller frequency, that is, hydrodynamic excitations, have a more extended antiparticle component, but also a smaller instability rate since they have to cross a much wider forbidden region.
In general, a necessary (but not sufficient) requirement for the occurrence of an ergoregion instability is the presence of negative norm modes in the inner region, which translates in the condition $\delta(r)\ge \Delta(r)$ between the Doppler shift and the centrifugal gap. The limiting radius $r_{E}$ for which such condition is satisfied can be thought as the boundary of the ergoregion, that is the ergosurface.
In the absence of Rabi coupling ($\Omega=0$), the effective ergosurface for angular momentum $M$ modes is located at a radial position
\begin{equation}
r_{E} = L\xi_s \sqrt{1-\frac{M^2}{4L^2}}\,. \label{ergoposition}
\end{equation}
This value is slightly reduced with respect to the hydrodynamic prediction $L\xi_s$, which is recovered in the $M\to 0$ limit, with the correction being due to the non-linear behaviour of the dispersion relation. For larger $M$ values, the ergoregion shrinks in space and eventually disappears for large angular momentum modes $M\ge 2L$.
As mentioned above, the presence of a coherent coupling $\Omega \neq 0$ results in a larger centrifugal gap [white region in Fig.\ref{bogodisp}(d)], so the matching between the external positive-norm modes (red-shading) and the negative-norm ones in the ergoregion (gray-shading) modes is harder to obtain and the unstable modes, if present, are more localized and display a weaker instability rate.
While this discussion based on the LDA provides an intuitive understanding of the microscopic process underlying the ergoregion instability, it completely misses all those features that stem from the quantization of the radial wavevector $k$ in a finite-size system\footnote{Differently from the planar configurations with synthetic vector potentials considered in~\cite{giacomelli2021understanding}, in our vortex configuration the irrotational nature of the superfluid flow fixes the characteristic spatial length on which the variation of ${\vect v}(r)$ can occur. As a result, it is not possible to find a rigorous limiting procedure in which the LDA is exactly verified.}. Since the radial wavevector is associated to the radial kinetic energy contribution $\hbar^2 k^2/2m$, we expect that for tight enough ergoregions, the minimum value of the kinetic energy set by the quantization of the negative norm mode may increase the size of the effective gap so much that it prevents the development of the instability. Assessing all these features in a quantitative way requires going beyond the LDA and solving the full Bogoliubov problem with numerical tools. This will be the subject of the next Section.
\section{Bogoliubov spectrum}
\label{sec:spectra}
The numerical solution of the Bogoliubov problem is performed as follows (more details can be found in the Appendix). For the chosen value of the vortex charge $L$, we first find the radial function $f(r)$ that describes the density profile of the stationary vortex and the associated oscillation frequency $\mu$. These are then used to build the Bogoliubov matrices for the density and spin excitations, whose diagonalization gives the eigenstates and the corresponding eigenvalues, that is, the Bogoliubov spectrum on top of the stationary state.
In what follows we present results for harmonically trapped systems, but no qualitative difference is observed for box traps (see Appendix). In both cases we choose the parameters so that the radial size of the condensate $R$ is much larger than the vortex size.
Fig.\ref{fig:L1spectrum} and Fig.\ref{fig:L2spectrum} show the Bogoliubov spectra in the spin channel for different angular momenta $M$ as a function of the interaction strength ratio $g_{12}/g$ for vortices of charge $L=1,2$ in the $\Omega=0$ case: in correspondence of the crossing points between a positive and a negative energy mode, we observe the appearance of dynamically unstable modes (red dots) with the associated imaginary frequency bubble. The insets of Fig.\ref{fig:L1spectrum}(panels c, f) and Fig.\ref{fig:L2spectrum}(panels c,d,g) show examples of the spatial profiles of the particle and antiparticle components of the dynamically unstable modes: the common feature consists in the localization of the antiparticle component (red solid line) in the inner portion of the system, while the particle component (black solid line) is spread throughout the whole volume occupied by the fluid.
\begin{figure*}[t]
\centering
\includegraphics[width = \linewidth]{Fig3.pdf}
\caption{Bogoliubov spectrum for $M=1$ spin excitations on top of a $L=1$ (panels a-c) or $L=2$ (panels d-f) vortex located at the center of
a harmonically trapped mixture of TF radius $R=30\xi_d$. Panels (c) and (f) show, together with the radial profile of the BEC density (grey dashed line), an example of the real-space profile of the particle $|u(x)|$ (black solid line) and antiparticle $|v(x)|$ (red solid line) components of the dynamically unstable mode for $g_{12}=0.75 g$.
}
\label{fig:L1spectrum}
\end{figure*}
In the case of singly-charged $L=1$ vortices, in the density channel a negative energy density mode is always present, but its frequency remains below that of positive norm modes (not shown); hence a dynamical instability never develops, in agreement with previous works~\cite{giacomelli2020ergoregion}. On the other hand, the spin channel shows energetic stability for attractive interspecies interactions $g_{12}<0$ (not shown) and alternated intervals of dynamical stability and instability for repulsive interactions $g_{12}>0$ [see Fig.\ref{fig:L1spectrum}(a-b)]. Similarly to what happens for density fluctuations of multiple-charged vortices in single-component BECs, this originates from the crossing of a spectrally isolated negative norm mode with the band of positive norm ones; the discreteness of the band, due to the finite size of the cloud, explains the existance of instability bubbles separated by intervals of dynamical stability\cite{giacomelli2020ergoregion}. Moreover, since the frequency of the isolated mode can be arbitrarily high, the associated instabilities are typically well localized in the vortex core [see Fig.\ref{bogodisp}(d)].
\begin{figure*}[t]
\centering
\includegraphics[width = \linewidth]{Fig4.pdf}\caption{Bogoliubov spectrum for $M=2$ (panels a-d) and $M=3$ (panels e-g) spin excitations on top of an $L=2$ vortex in a harmonically trapped mixture of TF radius $R=30\xi_d$.
Two dynamical instabilities are present in the $M=2$ channel: the real-space profile of the particle (black solid line) and antiparticle (red solid line) components of these modes is shown in panels (c) and (d) for $g_{12}=0.75 g$ and $g_{12}=0.95$, respectively. The main difference between the two modes is in the localization of the antiparticle component, which dominates up to $r_- \sim 3\xi_d$ for the former, and up to $r_-\sim 14\xi_d$ for the latter; indeed, according to our LDA treatment, lower frequency modes are more extended. Remarkably, the low frequency mode is only present for $g_{12}\sim g$. The $M=3$ channel shows a single instability: an example of the real-space profile of the mode is shown in panel (g) for $g_{12}=0.81 g$.
}
\label{fig:L2spectrum}
\end{figure*}
As it is well known~\cite{fetter2001vortices}, when multiply-charged $L>1$ vortices are considered, dynamical instabilities of the same nature appear in the density channel too: since the energy of a vortex with charge $L$ is larger than that of $L$ singly-charged vortices, multiply charged vortices may be unstable against splitting into several lower-charge vortices. For instance, for $L=2$ vortices, we find, as a function of the interparticle interaction, alternate intervals of dynamical stability and instability in the $M=2$ channel of the density Bogoliubov problem (not shown), again in agreement with previous work~\cite{giacomelli2020ergoregion}.
An additional class of dynamical instabilities occur in the spin channels: besides the high-frequency isolated modes crossing the positive energy band discussed above, which are present for all angular momenta $1 \le M < 2L$ [see Fig.\ref{fig:L1spectrum}(d), Fig.\ref{fig:L2spectrum}(a,e)] and which are related to the deformation of the vortex core and/or its splitting, low-frequency and spatially extended instabilities may also appear. As visible in Fig.\ref{fig:L2spectrum}(a), such an instability may originate from the crossing of the two bands of closely spaced positive and negative norm modes that extend throughout the ergoregion. As such it belongs to the class of superradiant ergoregion instabilities closer to the hydrodynamic regime. The possibility of observing instabilities of this superradiant kind is one of the main interests of replacing single-component systems with spin mixtures with a much wider ergoregion.
Due to their different origin, these additional modes appear only around zero frequency for $g_{12}\lesssim g$ close to the demixing point: most importantly, their antiparticle component extends well outside the vortex core [see Fig.\ref{fig:L2spectrum}(d) as compared with Fig.\ref{fig:L2spectrum}(c)]. Physically, this means that long-wavelength spin waves are involved in the instability mechanism. From a quantum gas perspective, the instability can be understood as a demixing instability whose appearance is facilitated by the presence of a vortex which lowers its threshold. From the point of view of the gravitational analogy, this instability is of superradiant nature: as it happens for ergoregion instabilities around massive rotating objects, the positive energy of the wave in the outer region is compensated by the negative energy of the wave inside the ergoregion.
Analogously, in the presence of a coherent coupling, low-frequencies instabilities appear close to the ferromagnetic phase transition point $g_{12} \lesssim g + 2\Omega/n$, where the massive gap in the spin dispersion closes and the spin healing length diverges. This will be the subject of future work.
In general, depending on the parameters' values, multiple dynamically unstable modes can be found for the same configuration. When observing the time evolution of the system starting from an unperturbed $L$-charged vortex, after a short transient in which all the unstable modes compete, the mode with the largest imaginary frequency will eventually win over the others and dominate the intermediate-time evolution. In most cases, the dominating mode is a high-frequency localized one. However, it is possible to play with the system size $R$ and with the interaction constants ratio $g_{12}/g$ to make all localized modes dynamically stable, while keeping a spatially extended mode unstable via superradiant mechanisms. In this case the dynamical instability is associated to the generation of long-wavelength phonon within the ergoregion, as expected from the gravitational analogy with ergoregion instabilities of massive rotating objects~\cite{Brito_2020}.
At very long times, nonlinear effects set in, resulting in complex mode mixing phenomena and additional instabilities. A study of this physics will be the subject of the next Section.
\section{Long-time dynamics}
\label{sec:GP}
It is well-known that the linearized Bogoliubov theory holds as long as the the excited modes are small perturbations on top of the stationary vortex state. When dealing with a dynamical instability, this approximation is valid for a limited amount of time, roughly given by the inverse of the growth rate $\Gamma$ of the unstable mode. After that, the cylindrical symmetry of the problem is likely to be broken by complex nonlinear processes, and mixing phenomena start occurring.
In order to access the long-time dynamics of the spin mixture, it is necessary to simulate the full GP equations \eqref{eq:GP} in two dimensions. The numerical protocol is the following (more details can be found in the Appendix): the stationary state corresponding to a vortex of charge $L$ sitting at $r=0$ is first obtained via a conjugate gradient algorithm \cite{antoine2017efficient}. In order to trigger the instability, we perturb the stationary state with some weak random noise, independently applied on the total and relative density. The temporal evolution of the system under the GPE \eqref{eq:GP} is then simulated using a split-step method.
Examples of the simulated time-evolution are presented in Figs.\ref{fig:L1splitting}-\ref{fig:L2evol} for vortices of charge $L=1,2$. As already discussed, $L=1$ vortices have a single $M=1$ potentially unstable mode in the spin channel.
On the other hand, $L=2$ vortices can feature up to five unstable modes, four of which in the spin channel (one for $M=1$, two for $M=2$, one for $M=3$). For the sake of simplicity, the simulation parameters are chosen in a way to have a single dynamical instability in the spin channel: in practice, this is done by keeping the interaction ratio $g_{12}/g$ fixed and tuning the system size $R$ to select the desired mode.
\begin{figure*}[ht]
\centering
\includegraphics[width = \linewidth]{Fig5.pdf}
\caption{Dynamics of the vortex splitting instability.
Real time evolution of a $L=1$ vortex with $R=21\xi_d$ and $g_{12}=0.8g$. For these parameters, the growth rate of the $M=1$ unstable mode, whose real-space profile is shown in panel (c), is $\Gamma/\mu \simeq 0.014$. Panels (a1-a6): total density $n = n_1+n_2$, measured in units of the peak Thomas-Fermi density $2 n_\text{TF}(r=0) = 2\mu/(g+g_{12})$. Panels (b1-b6): polarization, $Z = (n_1-n_2)/n$. Each column is computed at the time indicated by one of red dots in panel (d), which shows the temporal evolution of the standard deviation of the spin density $\delta n= n_1-n_2$ (black solid line), compared with the analytical exponential growth $\exp(\Gamma t)$ (blue dashed line).
}
\label{fig:L1splitting}
\end{figure*}
\begin{figure*}[ht]
\centering
\includegraphics[width = \linewidth]{Fig6.pdf}
\caption{Dynamics of the vortex splitting instability. Same plots as in Fig.\ref{fig:L1splitting} for an $L=2$ vortex with $R= 19\xi_d$ and $g_{12}=0.42g$. For these parameters, the instability rate of the $M=1$ unstable mode, whose real-space profile is shown in panel (c), is $\Gamma/\mu = 0.019$. }
\label{fig:L2splitting}
\end{figure*}
In all these figures, the short- and intermediate-time dynamics is consistent with our expectations based on the Bogoliubov theory: the instability develops initially as a well visible
perturbation characterized by $2M$ lobes in the azymuthal direction, where $M$ is the angular momentum of the unstable mode. For high-frequency instabilities, this excitation has a larger-amplitude part localized around $r=0$ and a smaller-amplitude one visible as interference fringes at larger radii [see panel (b2) of Figs.\ref{fig:L1splitting}-\ref{fig:L1evol}]; for low-frequencies instabilities instead, the localized part of the mode is itself spread over a larger portion of the system (see panel (b2) of Fig.\ref{fig:L2evol}).
In panel (d) of Figs.\ref{fig:L1splitting}-\ref{fig:L2evol}, we show the standard deviation (STD) of the spin density $\delta n = n_1-n_2$ from its ground state vanishing value, defined as
\begin{equation}
\text{std}(\delta n) = \sqrt{\frac{1}{Q^2-1}\sum_{i,j=1}^Q \delta n^2(x_i, y_j) }
\end{equation}
where $Q\times Q$ is the number of points in the 2D numerical grid.
As expected, this quantity is exponentially growing at a rate equal to the imaginary part $\Gamma$ of the frequency of the dynamically unstable mode, whereas the total density remains almost unperturbed.
At longer times, $t\Gamma \gg 1$, for the intermediate values of $g_{12}/g$ considered in Figs.\ref{fig:L1splitting}-\ref{fig:L1evol}, the high-frequency dynamically unstable modes lead to the splitting of the original vortex into smaller vortices and/or pairs of half-quantized vortices. Two examples are shown in Figs.\ref{fig:L1splitting}-\ref{fig:L2splitting} for $L=1$ and $L=2$ vortices, respectively, and closely resemble the results of Ref.\cite{kuopanportti2019}. Interestingly, the splitting is often followed by a recombination process; similar dynamics was predicted for multiply-charged vortices in a single-component BEC in \cite{patrick2022origin}. We attribute this phenomenon to energy conservation and to the finite size of the system: once the vortex is split, the excess of energy is released in the form of spin excitations, whose interference, after bouncing on the trap walls, reverses the splitting process. This typically occurs multiple times, before, in the absence of a dissipation mechanism, the evolution becomes turbulent, with additional vortices being nucleated from the boundaries and/or deformation of the (otherwise circular) cloud [see panels (a6-b6) of Fig.\ref{fig:L2splitting}].
All these complex nonlinear features go beyond this work and will be the subject of future studies.
\begin{figure*}[t]
\centering
\includegraphics[width = \linewidth]{Fig7.pdf}
\caption{ Real time evolution of a $L=1$ vortex with $R=22\xi_d$ and $g_{12}=0.97g$. For these parameters, the growth rate of the $M=1$ unstable mode, whose real-space profile is shown in panel (c), is $\Gamma/\mu \simeq 0.0016$. Same plots as in Fig.\ref{fig:L1splitting}. }
\label{fig:L1evol}
\end{figure*}
\begin{figure*}[ht]
\centering
\includegraphics[width = \linewidth]{Fig8.pdf}
\caption{ Real time evolution of a $L=2$ vortex with $R=23\xi_d$ and $g_{12}=0.97g$. For these parameters, the growth rate of the $M=2$ unstable mode, whose real-space profile is shown in panel (c), is $\Gamma/\mu \simeq 0.0019$. Same plots as in Fig.\ref{fig:L1splitting}; the dotted black line in panel (b2) indicates the size of the region where the antiparticle component of the unstable mode dominates. }
\label{fig:L2evol}
\end{figure*}
The physics is more intriguing in the $g_{12}\lesssim g$ regime close to the demixing point, where
vortex splitting is much harder to observe and, if it does so, only occurs at much longer times scales: in addition to the reduced value of the instability rate $\Gamma$, the softness of spin modes prevents in fact the cloud from absorbing the excess energy that would derive from vortex splitting. This result is confirmed by looking at the Bogoliubov spectra in Figs.\ref{fig:L1spectrum}-\ref{fig:L2spectrum}: for $g_{12}/g\lesssim 1$, the instability bubbles disappear or become tiny, and instability rates drop by roughly an order of magnitude.
As a consequence, when approaching the demixing point, vortices in condensate mixtures remain stable for a very long time, while their spin dynamics is ruled by linear Bogoliubov theory.
Two examples of the real-time dynamics in this regime are shown in Figs.\ref{fig:L1evol}-\ref{fig:L2evol} for $L=1,2$ vortices, respectively, and $g_{12}/g=0.97$. In the former case, the unstable Bogoliubov mode is spatially localized around the vortex and corresponds to a displacement of the cores in the two components,
with the consequent appearance of a net polarization. This process, however, does not lead to a full splitting of the vortex. We rather observe a sequence of intervals of suppression and revival of the instability, as witnessed by the oscillation in time of the standard deviation (STD) of the relative density. The underlying mechanism is rooted in the finite-size of the system and is analogous to the one leading to the sequence of vortex splitting and recombination events seen above.
The temporal evolution associated to a superradiant instability is finally visible in Fig.\ref{fig:L2evol}. At intermediate times [panel (b2)], a spatially extended $M=2$ modulation is clearly visible in association to the vortex core deformation. Its characteristic size corresponds to the spatial profile of the unstable mode shown in panel (c) and, in particular, its antiparticle component matches the extension of the ergoregion. This is the smoking gun of the superradiant nature of the instability.
At later times, the dynamics develops instead a more complicated behaviour that involves strong nonlinear mode-mixing effects: the perturbation of the spin density, due to the development of the instability, is large enough to excite the $M=2$ waves with higher radial momentum and several radial nodes that are visible in panel (b3). These different excitation patterns are in competition, eventually leading to a marked deformation of the vortex and a significant modulation also of the total density [see panel (a6)]. Once again, the accurate analysis of these complex mode-mixing processes goes beyond the scope of the present work.
\section{Conclusions and perspectives}
\label{sec:concl}
In this work, we have theoretically analysed the stability of quantized vortices in symmetric two-component atomic Bose-Einstein condensates (BEC) and we have interpreted the results within an analog gravity context in terms of ergoregion instabilities. In addition to instabilities related to high-frequency Bogoliubov modes localized in the vortex core and associated to the distortion or the splitting of the vortex core like in single-component BECs~\cite{giacomelli2020ergoregion}, suitable regimes are found where the physics is determined by low-frequency and long-wavelength spin excitations. In this case, the superradiant scattering process underlying the ergoregion instability involves the excitation of a Doppler-shifted, negative-energy spin-sound wave spread over the ergoregion and the simultaneous emission of positive-energy spin waves into the outer part of the BEC, in close analogy with the ergoregion instability of space-time around a rotating massive object.
On the theoretical side, the most challenging open questions concern the late time dynamics of unstable configurations. First studies of this physics focused on the simpler case of black hole laser instabilities~\cite{deNova:PRA2016} and first hints of the remarkable complexity of the superradiant case are visible in the simulations reported in this work. In this long-term adventure, a special attention has to be paid to the role of nonlinear processes in saturating the instability, in analogy with related phenomena predicted in the gravitational context such as the growth of the so-called black hole ``hair''~\cite{Bos16,San16,Eas17,Ike19}.
On the experimental side, two-component condensates can be obtained in fluids of atoms and of light. In the atomic case, vortices can be generated by means of suitably chosen stirring potential~\cite{Madison:PRL2000,fetter2009rotating} and the different components can be chosen within the atomic hyperfine structure in a way to obtain slow spin-sound waves~\cite{cominotti2022observation}.
In the optical case, the polarization degree of freedom can be used to obtain two-component condensates and arbitrary velocity patterns can be imprinted onto the fluid by suitably designing the phase profile of the pump beam so to generate rotating configurations~\cite{Carusotto:RMP2013}.
In both cases, the control of the coherent coupling strength via either an external electromagnetic dressing of the atoms or the optical birefringence of the cavity material may be used to tune the mass of the quantum field.
The remarkable experimental advances in both fields make us confident that it will be soon possible to validate our conclusions using state-of-the-art set-ups.
As a more ambitious challenge, the analysis of correlations in the spirit of \cite{carusotto2008numerical,recati2009bogoliubov,steinhauer2016}, will allow to investigate superradiant processes at the quantum level, so to prove superradiant amplification is intrinsically connected to the spontaneous creation of correlated pairs of Bogoliubov modes with opposite energy at the ergosurface.
\section{Acknowledgements}
AB aknowledges financial support from Q@TN, the joint lab between University of Trento, FBK- Fondazione Bruno Kessler, INFN- National Institute for Nuclear Physics and CNR- National Research Council.
IC acknowledges financial support from the European Union H2020-FETFLAG-2018-2020 project ``PhoQuS'' (n.820392), from the Provincia Autonoma di Trento, and from the Q@TN initiative.
\section{Appendix}
\subsection{Numerical details}
The numerical calculation of the Bogoliubov spectrum is performed as follows. Given a value of the vortex charge $L$, we first find the radial function $f(r)$ that describes the density profile of the stationary vortex and the associated oscillation frequency $\mu$. This is achieved by means of an imaginary-time evolution .
Once the vortex profile $f(r)$ is known, for each value of angular momentum $1\leq M<2L$, we build the density and spin Bogoliubov matrices and diagonalize them: the eigenvalues form the Bogoliubov spectrum of the vortex, while the eigenvectors give the real-space profiles of the modes.
The free parameters of the simulation are the interaction ratio $g_{12}/g$, the Rabi frequency $\Omega$, the radial size of the system $R$ and the expected chemical potential in the TF regime (i.e. without vortices) $\mu_{TF}$. The remaining parameters, i.e. the number of particles $N$ and the trapping frequency $\omega_0$, are chosen accordingly; in the case of a uniform mixture in a box of radius $R$, we set:
\begin{equation}
\frac{N}{\pi R^2} = \frac{2\mu_{TF} + \hbar\Omega}{g+g_{12}}
\end{equation}
while, for a system in an harmonic trap of TF radius $R$, we choose:
\begin{equation}
\omega_0 R = \sqrt{ \frac{2\mu_{TF} + \hbar\Omega}{m}} \qquad \textrm{and} \qquad \frac{2N}{\pi R^2} = \frac{2\mu_{TF} + \hbar\Omega}{g+g_{12}}
\end{equation}
While we can take advantage of the axial symmetry of the system to compute the Bogoliubov spectra, the same strategy is not applicable to the problem of determining the long-time dynamics of the mixture: we expect indeed the cylindrical symmetry to be broken by the development of the instability.
We thus solve the full GP equations in 2D, according to the following numerical protocol: we first identify, thanks to the Bogoliubov spectra, a set of parameters leading to a dynamical instability; we calculate the 2D profile of the stationary vortex via a conjugate gradient algorithm \cite{antoine2017efficient}, where the azymuthal profile of the wavefunction phase is enforced at every step.
The stationary state is perturbed with some weak random noise, applied independently on the total and relative density, to trigger the instability. We then simulate the time evolution of the system given by the GPE \eqref{eq:GP} using a split-step method.
\subsection{Dependence on the size and on the external potential}
In order to assess the dependence of dynamical instabilities on the size of the system and on the external potential applied to the atoms, we performed a calculation of the Bogoliubov spectra as a function of $R$ for two experimentally relevant cases: the harmonic trap and the box.
The results are shown in Fig.\ref{fig:appendix1} for $L=M=1$: in agreement with Ref.\cite{giacomelli2020ergoregion}, for both cases we find alternate intervals of dynamical stability and instability; as shown in panels (b) and (e), the instability bubbles acquire a larger width but smaller amplitude as $R$ increases.
This is the reason why in the main text, if possible, we simulate the real-time dynamics of systems with relatively small size ($R\sim 20\xi_d$): in addition to the advantage of allowing for a better resolution in space, such configurations also feature a faster development of the instabilities. As a further check, panels (c) and (f) of Fig.\ref{fig:appendix1} show that there is no qualitative difference in the real-space profiles of the unstable modes given \iac{in} the two examined configurations. We verified that this holds true also for other values of vortex charge $L$ and angular momentum $M$ (not shown).
\begin{figure}
\centering
\includegraphics[width = \linewidth]{appendix1.pdf}
\caption{Bogoliubov spectra for $M=1$ spin excitations over a vortex of charge $L=1$ with $g_{12}=0.93g$. Panels (a-c) refer to an harmonically trapped system of TF radius $R$, while panels (d-f) refer to a uniform mixture in a box of size $R$. Panels (c) and (f) show the real-space profile of the particle (black solid line) and antiparticle (red solid line) components of the dynamically unstable mode for $R=23\xi_d$; the dashed line represents the rescaled vortex profile $f(r)$. }
\label{fig:appendix1}
\end{figure}
\bibliographystyle{crunsrt}
\nocite{*}
|
\section{Introduction}\lb{sec1}\lb{sec1}
The post-Newtonian (PN) approximation (see, e.g., \cite{1987thyg.book..128D,1997PThPS.128..123A,2003trso.conf..411B,2006LRR.....9....4B,2007LRR....10....2F,2018tegp.book.....W} and references therein) is a computational scheme for solving the Einstein's field equations of his General Theory of Relativity (GTR) relying upon the assumptions that the characteristic speeds of the bodies under consideration are smaller than the speed of light $c$ and that the gravitational fields inside and around them are weak. Nonetheless, as pointed out in
\cite{2011PNAS..108.5938W}, such a scheme turned out to be remarkably effective in describing also certain strong-field and fast motion systems such as compact binaries made of at least one dense neutron star and inspiralling pairs of black holes emitting gravitational waves; the reasons for that are largely unknown \cite{2011PNAS..108.5938W}. Thus, putting the PN approximation to the test in as many different scenarios and at the highest order of approximation as possible is of paramount importance to gain an ever increasing confidence in it.
In its technical realm of validity, the PN approximation has been successfully tested so far only to the first post-Newtonian (1PN) order with, e.g., the orbital motions of planets, asteroids and spacecraft in the Solar System \cite{1968PhRvL..20.1517S,1972PhRvL..28.1594S,1971AJ.....76..588S,1990grg..conf..313S,2010PhRvL.105w1103L,2014PhRvD..89h2002L}, being its 2PN effects deemed too small to be currently measurable. The 1PN precession of the pericentre $\omega$ was measured also with binary pulsars \cite{2006Sci...314...97K,2021PhRvX..11d1050K} and stars orbiting the supermassive black hole at the centre of the Galaxy in Sgr A$^\ast$ \cite{2020A&A...636L...5G}.
The 2PN precession of the pericentre of a two-body system, recently worked out\footnote{Among the several calculation existing in the literature with different computational schemes and parameterizations like, e.g., \cite{1993PhLA..174..196S,1995CQGra..12..983W,2019CQGra..36k5001T}, see \cite{1988NCimB.101..127D} for a derivation based on the Hamilton-Jacobi method and the Damour-Deruelle parameterization \cite{1985AIHS...43..107D}.} \cite{2021Univ....7..443I} in a perturbative way with the Gauss equations \cite{2011rcms.book.....K} and the standard osculating Keplerian orbital elements \cite{1994ApJ...427..951K}, was investigated in \cite{2020MNRAS.497.3118H} for the double pulsar PSR J0737-3039A/B \cite{2003Natur.426..531B,2004Sci...303.1153L} since it is viewed as a major source of systematic error in the expected future determination of the spin-orbit Lense-Thirring periastron precession \cite{1988NCimB.101..127D} since it should fall within the envisaged sensitivity level.
Here, we explore the perspectives of including, a day, the 2PN effects in the dynamics of the Solar System by looking at the perihelion of Mercury and the present and future level of accuracy in knowing its orbit. For the light propagation to the $\mathcal{O}\ton{c^{-4}}$ order and its possible astrometric measurements in the Solar System, see, e.g., \cite{2015IJMPD..2450056D} and references therein. Also various consequences of modified models of gravity were investigated to the 2PN order; see, e.g., \cite{2008PhRvD..77l4049X,2009AdSpR..43..171X,2012PhRvD..86d4007D,2015Deng,2016IJMPD..2550082D}.
The paper is organized as follows. In Section\,\ref{sec2}, we review the calculational strategy put forth in \cite{2021Univ....7..443I} to obtain the total 2PN precession of the pericenter of a two-body system. The results are applied in Section\,\ref{sec3} to Mercury and compared to the latest figures for the uncertainty in determining its perihelion rate. The role of the ongoing BepiColombo mission in improving the Hermean orbit determination is discussed as well. Section\,\ref{sec4} summarizes our findings and offers our conclusions.
\section{The 2PN precession of the pericenter}\lb{sec2}
The \textit{total} 2PN net\footnote{Here and in the following, the angular brackets $\ang{\ldots}$ denoting the orbital average are omitted.} \textit{precession} of the argument of pericenter $\omega$ of a binary system made of two static, spherically symmetric bodies A and B, written in terms of the usual osculating Keplerian orbital elements, is \cite[eq.\, (18),\,pag.\,4]{2021Univ....7..443I}
\begin{align}
\dot\omega^\mathrm{2PN} \nonumber \lb{o2PN}& = \rp{3\,\mu^{5/2}}{8\,c^4\,a^{7/2}\,\ton{1-e^2}^3}\,\grf{
-68 + 8\,\nu + e^4\,\ton{-26 + 8\,\nu} + \right.\\ \nonumber \\
\nonumber &\left. + 2\,e^2\,\ton{-43 + 52\,\nu} + e\,\qua{8\,\ton{-29 + 13\,\nu} + e^2\,\ton{-8 + 61\,\nu}}\,\cos f_0 + \right.\\ \nonumber \\
&\left. + 3\,e^2\,\qua{4\,\ton{-5 + 4\,\nu}\,\cos 2 f_0 + e\,\nu\,\cos 3 f_0}}.
\end{align}
where $\mu\doteq G\,M$ is the gravitational parameter of the two-body system given by the product of the sum of its masses $M\doteq M_\mathrm{A} + M_\mathrm{B}$ by the Newtonian constant of gravitation $G$, the dimensionless parameter $\nu$ is given by $\nu\doteq M_\mathrm{A}\,M_\mathrm{B}/M^2$, $a,\,e$, and $f_0$ are the osculating values of the semimajor axis, eccentricity and true anomaly, respectively, at the same arbitrary moment of time $t_0$ \cite{1994ApJ...427..951K}.
By expanding \rfr{o2PN} in powers of the eccentricity $e$, one gets
\begin{align}
\dot\omega^\mathrm{2PN} \nonumber \lb{oappr} &\simeq \rp{3\,\mu^{5/2}\,\ton{-17 +2\,\nu}}{2\,c^4\,a^{7/2}}+ \rp{3\,\mu^{5/2}\,\ton{-29 + 13\,\nu}\,\cos f_0}{c^4\,a^{7/2}}\,e + \\ \nonumber \\
& + \rp{3\,\mu^{5/2}\,\qua{-145 + 64\,\nu + 6\,\ton{-5 + 4 \nu}\,\cos 2 f_0}}{4\,c^4\,a^{7/2}}\,e^2 + \mathcal{O}\ton{e^3}.
\end{align}
From \rfr{oappr} it can be noted that, to the order of zero in $e$, the 2PN pericentre precession is independent of $f_0$.
It should be recalled that \rfr{o2PN} does not come only from the \textit{direct}\footnote{It is calculated perturbatively in the usual way by evaluating the right hand side of the Gauss equation for $\mathrm{d}\omega/\mathrm{d}t$ \cite{2011rcms.book.....K}, calculated with ${\bds A}^\mathrm{2PN}$, onto a \textit{fixed} Keplerian ellipse, and integrating it over a \textit{Keplerian} orbital period.} effect of the 2PN acceleration ${\bds A}^\mathrm{2PN}$ entering the equation of motion (see, e.g., \cite[eq.\,(38), pag.\,8]{2021Univ....7..443I}). Indeed, also two \textit{indirect}, or \textit{mixed}, effects related to the 1PN acceleration ${\bds A}^\mathrm{1PN}$ (see, e.g., \cite[eq.\,(37), pag.\,7]{2021Univ....7..443I}) subtly concur to yield the net \textit{shift per orbit} $\Delta\omega$ to the 2PN level from which (a part of) the \textit{precession} follows by dividing it by the (Keplerian) orbital period $P_\mathrm{b}$; see \cite{2015IJMPD..2450067I} for how to calculate such \textit{indirect} effects in a different scenario. They are due to the following facts. On the one hand, during an orbital revolution, all the orbital elements entering the right hand side of the Gauss equation for $\mathrm{d}\omega/\mathrm{d}f$ \cite{1979AN....300..313M}, calculated with ${\bds A}^\mathrm{1PN}$, in principle, undergo instantaneous variations due to ${\bds A}^\mathrm{1PN}$ itself changing their values with respect to their fixed Keplerian ones referred to some reference epoch $t_0$. On the other hand, when the integration over $f$ is performed calculating $\Delta\omega$, the fact that the true anomaly is reckoned from a generally varying line of apsides because of ${\bds A}^\mathrm{1PN}$ should be taken into account as well.
Such features yield two additional corrections to $\Delta\omega$ with respect to the usual 1PN one\footnote{It is obtained by keeping $a$ and $e$ \textit{fixed} during the integration of the right hand side of the Gauss equation for $\mathrm{d}\omega/\mathrm{d}f$, calculated with ${\bds A}^\mathrm{1PN}$, over a \textit{Keplerian} orbital period.}
\eqi
\Delta\omega^\mathrm{1PN} = \rp{6\,\uppi\,\mu}{c^2\,a\,\ton{1-e^2}}\lb{1PN}
\eqf
which, in the case of ${\bds A}^\mathrm{1PN}$, are just of the order of $\mathcal{O}\ton{c^{-4}}$. Finally, as shown in \cite{2021Univ....7..443I}, when the \textit{total} 2PN net \textit{precession} has to be calculated, the 1PN \textit{fractional shift per orbit} $k^\mathrm{1PN}\doteq\Delta\omega^\mathrm{1PN}/2\uppi$ must be multiplied by the
1PN mean motion $\nk^\mathrm{1PN}$, and an expansion in powers of $c^{-1}$ to the order of $\mathcal{O}\ton{c^{-4}}$ must be taken.
Thus, the precession of \rfr{o2PN} comes from the sum of the latter contribution plus the \textit{direct} rate induced by ${\bds A}^\mathrm{2PN}$ and the two \textit{indirect} terms due to ${\bds A}^\mathrm{1PN}$.
In \cite{2021Univ....7..443I}, it is shown that \rfr{o2PN} agrees with other calculations existing in the literature performed with different computational strategies \cite{1988NCimB.101..127D,1994ARep...38..104K}. In particular, \rfr{o2PN} is in agreement with the expression for the total 2PN pericentre
precession, written in terms of the osculating Keplerian orbital elements, which can be inferred from \cite[eq.\,(5.18),\,pag.\,158]{1988NCimB.101..127D} based on the Damour-Deruelle parameterization \cite{1985AIHS...43..107D}.
\section{The 2PN perihelion precession of Mercury}\lb{sec3}
In the case of Sun and Mercury\footnote{For a recent comparative study of Mercury's perihelion advance induced by some classical dynamical effects, see \cite{2022CeMDA.134...33P}.}, Figure\,\ref{fig1}, displaying the plot of \rfr{o2PN} as a function of $f_0$, shows that the Hermean 2PN precession ranges from $-18$ to $-4$ microarcseconds per century $\ton{\mu\mathrm{as\,cty}^{-1}}$.
\begin{figure}[ht!]
\centering
\begin{tabular}{c}
\includegraphics[width = 12 cm]{2pnmerc.eps}\\
\end{tabular}
\caption{
Total 2PN perihelion precession $\dot\omega^\mathrm{2PN}$ of Mercury, in $\mu\mathrm{as\,cty}^{-1}$, as a function of the true anomaly at epoch $f_0$ according to \rfr{o2PN}. It turns out that $-18\,\mu\mathrm{as\,cty}^{-1}\lesssim \dot\omega^\mathrm{2PN}\lesssim -4\,\mu\mathrm{as\,cty}^{-1}$.
}\label{fig1}
\end{figure}
About the current accuracy in determining observationally the perihelion precession $\dot\omega_\mathrm{exp}$ of Mercury, the present author in \cite{2019AJ....157..220I} tentatively inferred a \textit{formal} uncertainty as little as
\eqi
\sigma_{\dot\omega_\mathrm{exp}}\simeq 8\,\mu\mathrm{as\,cty}^{-1}\lb{erro}
\eqf
from the planetary ephemerides\footnote{EPM stands for Ephemeris of Planets and the Moon.} EPM2017 \cite{2018AstL...44..554P}. At first glance, \rfr{erro} might seem interesting since it is of the same order of magnitude of the 2PN contribution to the Hermean perihelion precession. The dynamical models of\footnote{See https://iaaras.ru/en/dept/ephemeris/epm/2017/ for details.} EPM2017 are accurate to the 1PN level, including, among other things, also the Lense-Thirring effect induced by the Sun's angular momentum $S_\odot$. As far as Mercury is concerned, they are based only on radio tracking data resulting in 1556 normal points over a temporal interval 51 years long (1964-2015); in particular, data collected by the MESSENGER (Mercury Surface, Space Environment, Geochemistry and Ranging) spacecraft from 2011 to 2015 were analyzed. As pointed out in \cite{2018AstL...44..554P,2019AJ....157..220I}, realistic accuracies at the time of writing such papers may be $\simeq 10-50$ times larger than \rfr{erro}. On the other hand, a new model of solar plasma affecting the spacecraft ranging observations was recently published \cite{2022MNRAS.514.3191A}; it can now be used to reprocess the raw MESSENGER data\footnote{Until now, the simpler model of the NASA Jet Propulsion Laboratory (JPL) was pre-applied to the normal points published by it; see https://ssd.jpl.nasa.gov/dat/planets/messenger.txt} which were recently released\footnote{D. Pavlov, private communication to the present author, November 2022.}. This should improve the accuracy in our knowledge of the Mercury's orbit. As noted in \cite{2019AJ....157..220I}, \rfr{erro} corresponds to an uncertainty as little as $2\times 10^{-7}$ in the combination $\ton{2 + 2\,\gamma - \beta}/3$ of the PPN parameters $\gamma$ and $\beta$ in front of \rfr{1PN}. The ongoing mission to Mercury BepiColombo \cite{2021SSRv..217...90B} aims, among other things, to accurately determine $\beta$ and $\gamma$; according to \cite[Table\,5,\,pag.\,21]{2021SSRv..217...21I}, an extended mission may reach just the $\simeq 10^{-7}$ accuracy level in constraining such PPN parameters. The same conclusion is shown also in
\cite[Table\,2,\,pag.\,12]{2022RemS...14.4139V}; see also references therein. However, it should be remarked that most of the scenarios examined in \cite{2021SSRv..217...90B,2022RemS...14.4139V} envisage an accuracy in constraining $\beta$ and $\gamma$ of the order of $\simeq 10^{-6}$.
Be that as it may, perhaps, we may not be so far away from having to include, one day, also the 2PN terms in Solar System's dynamics.
In principle, a source of major systematic error which may overwhelm the 2PN perihelion precession is represented by the competing classical effect due to tidal distortion involving the Hermean Love number $k_2$ \cite{1911spge.book.....L,1959cbs..book.....K}. Such a parameter measures the mass concentration toward the centre of a fluid body assumed in hydrostatic equilibrium like, e.g., a main sequence star. Its possible values range from 0 for the mass point
approximation to $3/4=0.75$ for a fully homogeneous fluid body \cite{2014grav.book.....P}.
For a binary system, the periastron precession of tidal origin is \cite[eq.\,(3.100),\,pag.\,170]{2014grav.book.....P}
\eqi
\dot\omega_\mathrm{tid} = 15\,\nk\,\ton{1 + \rp{3}{2}\,e^2 + \rp{1}{8}\,e^4}\,\qua{k_2^\mathrm{A}\,\rp{M_\mathrm{B}}{M_\mathrm{A}}\,\ton{\rp{R_\mathrm{A}}{p}}^5 + \mathrm{A}\rightleftarrows\mathrm{B}},\lb{otid}
\eqf
where $p\doteq a\ton{1-e^2}$ is the semilatus rectum of the two-body relative orbit, and $R_\mathrm{A/B}$ is the equatorial radius of the body A or B.
A recent determination of the Love number of Mercury relying upon the analysis of the complete four years of MESSENGER tracking data from March 2011 to April 2015 yields\footnote{It has to be meant as the geophysicists' Love number, which is twice the astronomers' one, known also as apsidal constant, entering \rfr{otid} \cite[pag.\,115]{2014grav.book.....P}.} \cite{2020Icar..33513386K}
\eqi
k_2=0.53\pm 0.03\lb{k2}.
\eqf
The nominal value of the Hermean contribution\footnote{The one due to the Sun's Love number is much smaller, being, thus, of no concern.} to \rfr{otid}, calculated with \rfr{k2}, is of the order of a few $\mu\mathrm{as}\,\mathrm{cty}^{-1}$. Nonetheless, by assuming to model the tidal effects to the level of accuracy of \rfr{k2}, the resultant mismodeled perihelion precession would be well below the 2PN one.
From a practical point of view, experts in planetary data reductions should clarify which part of \rfr{o2PN} could be, actually, measured and how. Indeed, given that the 1PN equations of motion are currently included in the dynamical force models of the softwares routinely adopted to process the data, it may be argued that the indirect components of \rfr{o2PN} should have already been measured along with the 1PN precession; if so, only the direct part due to ${\bds A}^\mathrm{2PN}$ could be detected by explicitly modeling it and estimating, say, a dedicated scaling solve-for parameter. Otherwise, ${\bds A}^\mathrm{2PN}$ should be modeled in terms of $\beta$ and $\gamma$ whose newly estimated values, if known at the $10^{-7}$ level, would account for the 2PN component of the equations of motion as well.
\section{Summary and conclusions}\lb{sec4}
The 2PN pericentre precession $\dot\omega^\mathrm{2PN}$ of a two-body system made of two static, spherically symmetric masses, obtained with the standard Gauss perturbative equation in terms of the usual osculating Keplerian orbital elements, is reviewed. Both the exact expression (\rfr{o2PN}) and an approximated expansion in powers of the eccentricity $e$ up to the order of $\mathcal{O}\ton{e^2}$ (\rfr{oappr}) are given. It is recalled that $\dot\omega^\mathrm{2PN}$ consists of four contributions: a direct term straightforwardly arising from the 2PN acceleration ${\bds A}^\mathrm{2PN}$ in the equations of motion, and three indirect parts due to the 1PN acceleration ${\bds A}^\mathrm{1PN}$ itself. In particular, one of the latter ones arises from the expansion to the order of $\mathcal{O}\ton{c^{-4}}$ of the product of the well known fractional 1PN shift per orbit (the ratio of \rfr{1PN} to $2\,\uppi$) by the 1PN mean motion. The other two indirect contributions come from taking into account also the instantaneous variations of the order of $\mathcal{O}\ton{c^{-2}}$ of the orbital elements and the fact that the anomalistic period over which the 1PN shift is integrated is the time interval between two successive crossings of an actually moving pericenter due to ${\bds A}^\mathrm{1PN}$ itself. The resulting total 2PN precession of \rfr{o2PN} depends on the true anomaly at epoch $f_0$. It is remarked that \rfr{o2PN} agrees with other expressions for $\dot\omega^\mathrm{2PN}$ in the literature obtained with different parameterizations and calculational schemes.
In the case of the Sun and Mercury, the Hermean 2PN perihelion precession, calculated with \rfr{o2PN}, ranges from $-18$ to $-4\,\mu\mathrm{as\,cty}^{-1}$ depending on $f_0$. A recent guess for the formal experimental uncertainty in determining the Mercury's perihelion precession with the EPM2017 planetary ephemerides is $\sigma_{\dot\omega_\mathrm{exp}}\simeq 8\,\mu\mathrm{as\,cty}^{-1}$, although the realistic uncertainty may be up to $\simeq 10-50$ times larger. Nonetheless, the raw data collected during the past MESSENGER mission are now available, and a new, accurate model of the solar corona, usually a major bias impacting the accuracy of ranging measurements, was recently published. Thus, reprocessing the MESSENGER observations with such a new model should improve our knowledge of Mercury in the near future. A $\simeq\mu\mathrm{as\,cty}^{-1}$ level corresponds to an uncertainty of the order of $\simeq 10^{-7}$ in the PPN combination $\ton{2+2\gamma-\beta}/3$
scaling the 1PN precession. In a few scenarios encompassing an extended mission profile of the BepiColombo spacecraft, currently en route to Mercury, such a level of accuracy in constraining the PPN parameters $\gamma$ and $\beta$ may be reached, although $\simeq 10^{-6}$ seems more plausible according to the majority of the simulations performed so far in the literature. Anyway, perhaps, the time when it will be necessary to model the dynamics of the Solar System at the 2PN order might not be that far away, after all.
\section*{Acknowledgements}
I am indebted to M. Efroimsky for his careful and attentive comments and remarks, and to R.S. Park for clarifications about the Love number of Mercury.
I gratefully thanks also G. Tommei and D. Pavlov for useful information.
|
\section{Introduction}
Mobility on demand (MoD) systems, in which a fleet of free-floating vehicles serves customers' ad hoc requests for point-to-point transportation, have transformed urban mobility in recent years. Companies like Uber, Lyft, and DiDi, made MoD more accessible compared to taxi-based ride hailing services. Autonomous vehicles will further transform MoD systems; besides much lower prices, a major benefit of autonomous MoD (AMoD) is its improved potential for advanced control strategies, as a central operator obtains full control over the entire fleet. This transformation changes the fleet operator's control problem substantially: MoD operators focus primarily on revenue maximization, as human drivers' income is (almost) a fixed cost that dominates mileage-dependent operational cost. Contrarily, AMoD operators focus on the maximization of their operating profit, because operational costs dominate their total cost balance. In this context, the central operator can leverage its full knowledge about the system state and fleet control to make improved (proactive) dispatching decisions, i.e., request to vehicle assignment and rejection, to maximize its profit.
Since an operator does not have profound knowledge about future trip requests, it faces an online decision-making problem in a stochastic environment. Hence, it seems promising to apply deep reinforcement learning (DRL) to this problem. However, AMoD systems entail many vehicles and trip requests, such that an operator's action space is very large and possibly time-varying as the number of requests changes over time. It is thus infeasible to apply off-the-shelf single-agent DRL algorithms. To solve this problem, we propose a novel combination of a multi-agent DRL algorithm with optimization-based centralized decision-making through weighted bipartite matching, thereby combining the advantages of multi-agent approaches, DRL, and combinatorial optimization.
\subsection{Related Work}
To keep this literature overview concise, we focus on literature for controlling (autonomous) MoD systems in the following. For a review of multi-agent DRL, we refer to \cite{Gronauer2022} and further elaborate on how we build on the multi-agent DRL literature in Section~\ref{sec:method}.
Classical approaches for dispatching and explicit rebalancing decisions focused on greedy or hand-crafted feature-based policies \citep{Liao2003,Zhang2017}, queueing theoretical approaches \citep{Zhang2016}, and model predictive control (MPC) \citep{AlonsoMora2017}.
Recently, many works applied DRL in the context of (autonomous) MoD, often including or purely focusing on explicit rebalancing \citep[e.g.,][]{Jiao2021, Gammelli2021, Skordilis2022, Liang2022}. Contrarily, other works focused on DRL for non-myopic dispatching, which entails an implicit rebalancing decision that avoids additional costs due to empty driving. Early approaches \citep[cf.][]{Xu2018, Wang2018} were shown to be inferior to at least one of the subsequent works: \cite{Li2019} proposed a mean field multi-agent actor-critic algorithm. \cite{Tang2019} used bipartite matching based on learned $V$-values. \cite{Zhou2019} combined a multi-agent Deep Q-Network with minimization of the Kullback-Leibler divergence (KL-divergence) between the vehicle and the request distribution. Finally, \cite{SadeghiEshkevari2022} described how DiDi recently rolled out DRL for dispatching in practice.
Since these works strive to improve the operations of today's MoD systems, they aim at maximizing the drivers' revenue or the number of orders served, rather than at maximizing the profit of an AMoD system. While \cite{Xu2018, Wang2018, Tang2019, SadeghiEshkevari2022, Liang2022} also use a combination of multi-agent DRL and weighted matching, they all employ value-based algorithms. Contrarily, we use an actor-critic algorithm, enabling more advanced strategies to mitigate problems arising from multi-agent learning, in particular, decentralized actors with centralized critics, see Section~\ref{sec:method}. By using Soft Actor-Critic (SAC) \citep{Haarnoja2018}, we can nevertheless benefit from the improved sample-efficiency of off-policy algorithms.
\subsection{Contributions}
To the best of our knowledge, we are the first to consider the problem of making proactive dispatching decisions for a profit-maximizing AMoD system operator with DRL. We propose a novel method that combines multi-agent SAC with centralized final decision-making through weighted matching. We perform experiments based on real-world data and, similar to related works, benchmark our method against a greedy policy. In addition, we are the first to compare our method against an MPC approach. We show that our method outperforms the greedy policy on all instances by up to 5\%. Moreover, we outperform the MPC approach in most cases. Our DRL method shows a significantly more stable performance across varying instances, while MPC may perform arbitrarily bad---in single cases up to 60\% worse than the greedy policy.
Our code can be found at \url{https://github.com/tumBAIS/HybridMADRL-AMoD}.
\section{Problem Formulation: Markov Decision Process}\label{sec:MDP}
We consider a profit-maximizing operator who centrally controls a fixed-size fleet of vehicles to serve customer trip requests revealed over time within an operating area. The operator can accept or reject requests and dispatches accepted requests to vehicles. These decisions must be made in real-time and immediately, i.e., the operator cannot defer requests to a later time step, as customers are not willing to wait for feedback. If the operator accepts a request, customers must be picked up within a known maximum waiting time $\omega^\text{max}\in\mathbb{N}_0$ after the request was placed. We formalize this control problem as a Markov decision process (MDP) as follows.
\textbf{Preliminaries.} We consider a discrete time horizon $\mathcal T=\left\{0,1,...,T\right\}$. During one time step, multiple requests can enter the system. The operator makes one decision per time step for multiple requests simultaneously, which allows to optimize over a batch of requests. We represent the operating area as a graph $G=(V,E)$ with weight vectors $\prescript{e}{}{\bm w}=\left(\prescript{e}{}{w}^1, \prescript{e}{}{w}^2\right)\in\mathbb R_{>0}\times\mathbb N$, denoting the distance ($\prescript{e}{}{w}^1$) of and the time steps ($\prescript{e}{}{w}^2$) to traverse an edge $e\in E$.
\textbf{States.} We describe the state of the system at time $t\in\mathcal T$ by
\begin{equation*}
\bm S_t=\left(t,\ \left(\prescript{t}{}{\bm r}^i\right)_{i\in\left\{1,...,R_t\right\}},\ \bigl(\bm k_t^j\bigr)_{j\in\left\{1,...,K\right\}}\right),
\end{equation*}
with $R_t$ being the variable number of new requests $\prescript{t}{}{\bm r}^i$, $i\in\{1,...,R_t\}$, at time step $t$, and $K$ vehicles $\bm k_t^j$, $j\in\{1,...,K\}$. A request $\bm r=\left(\omega,o,d\right)$ consists of a waiting time $\omega\in\mathbb N_0\cup\emptyset$, an origin $o\in V$, and a destination $d\in V\setminus \{o\}$; $\omega$ tracks the elapsed time from request placement to pickup, where we set $\omega\leftarrow\emptyset$ at pickup. We denote a vehicle by $\bm k = (v,\tau,\bm r^1,\bm r^2)$, with position $v\in V$ and the number of time steps $\tau\in\mathbb N_0$ left to reach this position. Here, $v$ can either be the current node if the vehicle idles or the next node that will be reached if the vehicle travels. Furthermore, slightly abusing notation, a vehicle can have at most two assigned requests $\bm r^1,\bm r^2$. Assigning more requests to one vehicle is unreasonable for realistic trip lengths and maximum waiting times. We denote the position of vehicle $\bm k_t^j$ by $\prescript{k^j}{}{\! v}_t$ and denote other components of the vehicle vector likewise.
\textbf{Actions.} The action space describing feasible decisions of the operator is
\begin{align}
\mathcal A\left(\bm S_t\right)=\Biggl\{ \left(a_t^1,...,a_t^{R_t}\right)\Biggl\lvert\ & a_t^i=0\ \lor\ \left(a_t^i=j\in\left\{1,...,K\right\}\ \wedge\ \prescript{k^j}{}{\! \bm r}_t^2=\emptyset\right)\ \forall\ i\in\left\{1,...,R_t\right\}, \nonumber \\
& \sum_{i=1}^{R_t}\mathds 1\!\left(a_t^i=j\right)\leq 1\ \ \forall j\in\left\{1,...,K\right\} \Biggr\} . \label{eq:action space}
\end{align}
The operator can take one decision $a_t^i$ per request $\prescript{t}{}{\bm r}^i$, $i\in\left\{1,...,R_t\right\}$, either rejecting it ($a_t^i=0$), which means that the request leaves the system, or assigning it to vehicle $\bm k^j$ ($a_t^i=j$), which is only possible if the vehicle does not already have two assigned requests, i.e., if $\prescript{k^j}{}{\! \bm r}_t^2=\emptyset$ holds. The final condition in (\ref{eq:action space}) implies that at most one new request is assigned to each vehicle in each time step, which is a realistic simplification facilitating the application of a matching algorithm.
\textbf{Transitions.} We first describe the action-dependent transition from the pre-decision to post-decision state. Then, we describe the transition from the post-decision state to the next pre-decision state, which is independent of the action and only determined by the system dynamics.
A reject decision has no impact on the state. When $a^i=j$, we add the request to the vehicle state, i.e., if $\prescript{k^j}{}{\! \bm r}^1=\emptyset$, then $\prescript{k^j}{}{\! \bm r}^1\leftarrow\prescript{t}{}{\bm r}^i$, and $\prescript{k^j}{}{\! \bm r}^2\leftarrow\prescript{t}{}{\bm r}^i$ otherwise.
The following transitions apply to all vehicles: if the vehicle picks up a customer, i.e., if $\bm r^1\neq\emptyset\ \wedge\ \tau=0\ \wedge\ v=o\!\left(\bm r^1\right)$, where $o\!\left(\bm r^1\right)$ denotes the origin of request $\bm r^1$, then $\omega\!\left(\bm r^1\right)\leftarrow\emptyset$. If the vehicle moves between two nodes, i.e., if $\tau>0$, then $\tau\leftarrow\tau-1$. If the vehicle is at a node but moves to serve a request, i.e., if $\tau=0\ \wedge\ \bm r^1\neq\emptyset$, then $v$ is replaced by the next node $v'$ on the vehicle's route to serve the request, going to origin or from origin to destination, and $\tau\leftarrow\prescript{(v,v')}{}{w}^2-1$. If a vehicle drops off a customer before the next decision is made, i.e., if $\bm r^1\neq\emptyset\ \wedge\ \omega\!\left(\bm r^1\right)=\emptyset\ \wedge\ \tau=0\ \wedge\ v=d\!\left(\bm r^1\right)$, we shift requests: $\bm r^1\leftarrow\bm r^2$ and $\bm r^2\leftarrow\emptyset$. We increment the waiting times $\omega\neq\emptyset$ of requests that have not been picked up yet, i.e., $\omega\leftarrow\omega+1$, where $\omega$ refers to $\omega\!\left(\bm r^1\right)$ and/or $\omega\!\left(\bm r^2\right)$. Moreover, independent of the vehicles' states, customers place new requests, i.e., $\left(\prescript{t}{}{\bm r}^i\right)_{i\in\left\{1,...,R_t\right\}}$ is replaced by $\left(\prescript{t+1}{}{\bm r}^i\right)_{i\in\left\{1,...,R_{t+1}\right\}}$. Note that we do not know the underlying time-dependent probability distribution which generates new requests, but we can simulate the resulting requests by replaying historic data. Finally, $t\leftarrow t+1$.
\textbf{Rewards.} Since the operator maximizes its profit and fixed costs are independent of the control problem, our reward function focuses on the operating profit, which is the revenue from serving requests minus operational costs, e.g., for fuel and maintenance. The operator obtains the revenue for a request $\bm{r}$ when a vehicle picks up the request within the maximum waiting time. The revenue is given by a function $\mathrm{rev}(\bm r)\in\mathbb R_{>0}$, representing the operator's pricing model. For improved readability, we express the profit components as functions of the post-decision state $\bm S_{t^+}$ and write $t$ for $t^+$. Then, the total revenue at time $t$ is
\begin{equation*}
\mathrm{Rev}(\bm S_t)=\sum_{j=1}^K \mathds 1\!\left( \prescript{k^j}{}{\!\bm r}_t^1\neq\emptyset\ \wedge\ \prescript{k^j}{}{\!\tau}_t=0\ \wedge\ \prescript{k^j}{}{\! v}_t=o\!\left(\prescript{k^j}{}{\!\bm r}_t^1\right)\ \wedge\ \omega\!\left(\prescript{k^j}{}{\!\bm r}_t^1\right)\leq\omega^\text{max} \right)\cdot \mathrm{rev}\!\left(\prescript{k^j}{}{\!\bm r}_t^1\right).
\end{equation*}
When a vehicle starts to move from $v$ to $v'$, the operator incurs operational costs $c\in\mathbb R_{>0}$ per distance unit, as commonly assumed \citep[see, e.g.,][]{Boesch2018}. Thus, the total cost at time $t$ is
\begin{equation*}
\mathrm{Cost}(\bm S_t)=c\cdot\sum_{j=1}^K \mathds 1\!\left(\prescript{k^j}{}{\!\tau}_t=0\ \wedge\ \prescript{k^j}{}{\!\bm r}_t^1\neq\emptyset\right)\cdot\prescript{\bigl(\prescript{k^j}{}{\! v}_t,\ \prescript{k^j}{}{\! v}_t'\bigr)}{}{w}^1 .
\end{equation*}
The total profit at time $t^+$ is $\mathrm{Profit}(\bm S_{t^+})=\mathrm{Rev}(\bm S_{t^+})-\mathrm{Cost}(\bm S_{t^+})$. Note that $\bm S_{t^+}$ is a function of $\bm S_t$ (pre-decision) and $\bm a_t\in\mathcal A\left(\bm S_t\right)$, such that we write $\mathrm{Profit}(\bm S_{t^+})=\mathrm{Profit}(\bm S_t, \bm a_t)$.
The AMoD operator wants to find a policy $\pi: \bm S_t\rightarrow\bm a_t \in\mathcal A\left(\bm S_t\right)$ that maximizes the expected total reward over all time steps, given the initial state $\bm S_0$:
\begin{equation*}
\mathrm{Profit}^\ast\!\left(\bm S_0\right)=\max_\pi\, \mathbb E\left[\, \left. \sum_{t=0}^{T-1}\mathrm{Profit}\left(\bm S_t,\pi\left(\bm S_t\right)\right)\, \right\rvert\, \bm S_0\, \right] .
\end{equation*}
To do so, we propose a hybrid DRL algorithm in the following section.
\section{Method: Multi-agent Soft Actor-Critic with Global Matching}\label{sec:method}
Analyzing our problem setting, we identify two key requirements to develop an algorithm that constructs an effective control policy: first, it should leverage information patterns that can be observed from historic trip data to make non-myopic decisions. Second, it should be scalable to a realistic system size to coordinate a large number of vehicles and requests. To account for the second requirement, we formalize the centralized dispatching of vehicles to requests as a bipartite matching problem (BMP). This BMP should be weighted to allow for non-myopic dispatching decisions, anticipating the downstream impact of decisions in a stochastic environment. The choice of weights heavily impacts the policy's performance. Therefore, we use DRL to parameterize these weights, as it accounts for the downstream impact of decisions in stochastic environments by design and allows to extract and use information from historic data---thus, covering the first requirement.
However, single-agent DRL is not suitable for our problem setting, as the central operator's action space scales exponentially with the number of vehicles and requests per time step and becomes intractable very quickly. Thus, we leverage multi-agent DRL to factorize the action space at the price of increased complexity, caused by having to coordinate the actions of multiple DRL agents to finally take a centralized decision. Thereby, we combine the advantages of multi-agent DRL with those of combinatorial optimization: we use DRL agents as estimators to compute non-myopic weights, serving as the input to a weighted bipartite matching algorithm, which then makes a globally optimal and coordinated decision.
\subsection{Overview}\label{sec:method_overview}
Figure~\ref{fig:methodology_overview} provides an overview of our method in which we leverage a DRL algorithm to parameterize a weighted bipartite matching to take anticipatory global dispatching decisions. To obtain a weight for each request-vehicle combination, we consider each combination as one agent. We represent these agents by an actor network, which we train using the SAC algorithm \citep{Haarnoja2018}. To obtain the weights for our BMP, we post-process the actors' outputs, such that from the perspective of the DRL agents and the computation of policy parameter gradients, post-processing and matching are part of the environment.
\begin{figure}
\includegraphics[width=\textwidth]{methodology_overview_v3.pdf}
\caption{Overview of our method (gray text refers to parts which we use only during training).}
\label{fig:methodology_overview}
\end{figure}
Since the DRL agents are a means to factorize the action space of the central operator, rather than ``real" individual agents, they can observe the global system state. Although they should take cooperative decisions that eventually benefit the central operator's profit, the agents observe their own (egoistic) rewards, not the global system reward, to avoid a credit assignment problem \citep[e.g.,][]{Agogino2004} that otherwise occurs, in particular with many agents. Here, we enforce coordination of the agents through the BMP and note that varying the credit assignment scheme remains an interesting question for future work.
Individual rewards imply a need for per-agent critic values. From one agent's perspective, the other agents are part of the environment. Since we train all agents concurrently, their policies change simultaneously and the perceived environment is non-stationary. To mitigate this, the critic gets the other agents' actions as an additional input (``centralized critic"), such that the policy evaluation can explicitly account for other agents' behavior \citep[e.g.,][]{Lowe2017, Iqbal2019}.
All agents represent a request-vehicle combination and are thus homogeneous. Accordingly, they can share parameters and we need only one actor and one critic network for all agents \citep[cf.][]{Iqbal2019}. We can train those centrally, i.e., the total parameter update is given by the sum of per-agent updates. Still, the forward pass of the actor network is independent across agents, allowing for decentralized and parallelized execution, which is important for scalability.
Our method can handle a variable number of requests and thus a variable global action space size, as we can use neural networks with parameter sharing for any number of agents in parallel and the BMP does not require a fixed number of requests. The same holds true for the vehicles, such that the system size when testing may differ from training (see appendix).
We provide details on the individual components of our method in the following.
\subsection{Per-agent Post-processing and Global Matching}
The actor network parameterizes a categorical probability distribution over the two actions that can be taken for a request-vehicle combination: reject or accept. We post-process the actor output per agent to transform it into a per-agent score, that we then use in the global weighted matching.
Algorithm~\ref{alg:post-processing} defines the post-processing. First, we mask infeasible actions by setting the accept probability $p_\text{a}$ to zero if the vehicle already has two assigned requests. Then, we sample a reject/assign decision from the masked probability distribution; when testing, we instead take the argmax of the probabilities. A reject decision ($\delta=0$) at the per-agent level implies a request-to-vehicle reject decision at the global level, such that we set the respective score to zero. For an accept decision, we use the accept probability as score\footnote{At first sight, it might seem more intuitive to let the actor parameterize a continuous distribution, from which we can take a sample to directly obtain the score, and/or have separate outputs for the reject/accept probability and the score. We tested both approaches, but empirically observed that they perform worse than the variant described here. If we choose one of those approaches, we cannot compute the terms in the loss functions (see Section~\ref{sec:loss_functions}) which are an expectation w.r.t. the policy, i.e., $\pi_\phi(a\,\lvert\, s,i)^T\cdot(...)$, since the action space is not (purely) categorical anymore. Then, we need to sample to estimate the expectation, as in the version of SAC for continuous action spaces, which increases the variance. We hypothesize that this harms the algorithm's performance and explains our empirical observation.}, such that a higher probability leads to a higher score for the weighted matching. An accept decision at the per-agent level does not always imply an accept decision at the global level, as the matching might assign the request to a different vehicle.
\begin{algorithm2e}
\caption{Per-agent post-processing}
\label{alg:post-processing}
\DontPrintSemicolon
\KwIn{$p_\text{r}, p_\text{a}\in[0,1] \text{ s. t. } p_\text{r}+p_\text{a}=1 \ \ ; \ \ \bm k^j$}
\KwOut{score $s$}
\lIf(\tcp*[f]{\small reject if already two assigned requests}){$\prescript{k^j}{}{\!\bm r}^2\neq\emptyset$}
{$p_\text{r}\gets 1\ ,\ p_\text{a}\gets 0$}
\leIf{training}
{$\delta\sim\mathrm{Categorical}\bigl((p_\text{r},p_\text{a})\bigr)$ \tcp*{\small sample $\delta\in\{0,1\}$ when training}}
{$\delta\gets 0\text{ if } p_\text{r}\geq p_\text{a},\ \delta\gets1\text{ if } p_\text{r}< p_\text{a}$ \tcp*[f]{\small argmax when testing\hspace{-0.1cm}}}
\leIf{$\delta=0$}
{$s\gets 0$ \tcp*{\small score is zero if rejected}}
{$s\gets p_\text{a}$ \tcp*[f]{\small score is accept probability if accepted\hspace{-0.1cm}}}
\end{algorithm2e}
We use all agents' scores to create a bipartite graph, with vehicles and requests as nodes, and edges between all vehicle and request nodes for which we obtain per-agent accept decisions (i.e., $s>0$). The edges' weights correspond to the respective scores. We solve the resulting maximum weighted BMP (formally defined in the appendix) using the Hungarian algorithm \citep{Kuhn1955} to get a globally coordinated decision, where each request is assigned at most once.
\subsection{Multi-agent Soft Actor-Critic}\label{sec:loss_functions}
SAC is an entropy-regularized, off-policy actor-critic algorithm. It trains a stochastic policy $\pi\left(\bm a_t\lvert\bm S_t\right)$ with entropy maximization, incentivizing exploration through a random policy:
\begin{equation*}
\pi^\ast=\arg\max_\pi\, \mathbb E_{\left(\bm S_t, \bm a_t\right)\sim\pi}\left[\, \sum_{t=0}^{T-1} \mathrm{Profit}\left(\bm S_t,\bm a_t\right) + \alpha\, H\!\left(\pi\left(\, \cdot\, \lvert\, \bm S_t\right)\right) \right] .
\end{equation*}
The entropy of the policy is defined as $H\!\left(\pi\left(\, \cdot\, \lvert\, \bm S_t\right)\right)=-\mathbb E_{\bm a_t\sim\pi}\log\pi\left(\bm a_t\lvert\bm S_t\right)$ and the entropy coefficient $\alpha\in\mathbb R_{\geq 0}$ is a hyperparameter that controls the exploitation/exploration trade-off. While SAC was originally developed for continuous action spaces in \cite{Haarnoja2018}, it can also be applied to discrete actions \citep{Christodoulou2019}.
We chose to use an actor-critic algorithm to enable our multi-agent approach of decentralized actors with centralized critics as explained in Section~\ref{sec:method_overview}. Since exploration is paramount for our problem setting, we use SAC, which lets us explicitly tune how much the policy explores.
We parameterize the actor network with parameters $\phi$ and the critic with $\theta$. SAC uses two critic networks, $Q\in\left\{Q^1,Q^2\right\}$, as well as corresponding target networks with parameters $\bar\theta$, which are an exponential moving average of the primary parameters $\theta$. Based on the multi-agent approach as described in Section~\ref{sec:method_overview}, the loss function for the actor with shared parameters is
\begin{equation*}
J_\pi(\phi)=\mathbb E_{s\sim D}\left[ \sum_i \pi_\phi(a\,\lvert\, s,i)^T\cdot \left( \alpha\log\pi_\phi(a\,\lvert\, s,i) - \min_{j\in\{1,2\}}\left\{ Q_\theta^j\left( a\,\lvert\, s,i,\bar a_{-i} \right) \right\} \right) \right] .
\end{equation*}
Here, we use a simplified notation for improved readability: we denote a transition by $\left(s,\bar a,r,s'\right)$, with global states $s,s'$, global action $\bar a$ (after the matching), and rewards $r$. For agent $i$, $r_i$ denotes its reward, $\bar a_{-i}$ is the global action except for agent $i$'s action, and $a$ is a per-agent action (reject/assign), such that $\pi_\phi(a\,\lvert\, s,i)\in[0,1]^2$ and $Q_\theta^j\left( a\,\lvert\, s,i,\bar a_{-i} \right)\in\mathbb R^2$. We sample states (or transitions) from the replay buffer $D$ and denote the discount factor by $\gamma$. For the actor loss, we do not sample the global action from the replay buffer, but compute it based on the state $s$ and the current policy, as in \cite{Iqbal2019}. For each of the two critics $Q\in\left\{Q^1,Q^2\right\}$, the loss function is
\begin{align*}
J_Q(\theta)&=\mathbb E_{(s, \bar a, r, s')\sim D}\left[ \sum_i\frac12 \left( Q_\theta\left(a\,\lvert\, s,i,\bar a_{-i}\right)\Big\rvert_{\bar a_i}-y_i \right)^2 \right] \text{, with} \\
y_i&=r_i+\gamma\cdot\pi_\phi\left(a'\,\big\lvert\, s', i\right)^T\cdot\left( \min_{j\in\{1,2\}} \left\{ Q_{\bar\theta}^j\left( a'\,\big\lvert\, s',i,\bar a'_{-i} \right) \right\}-\alpha\log\pi_\phi\left(a'\,\big\lvert\, s',i\right) \right) .
\end{align*}
Here, the notation $\big\rvert_{\bar a_i}$ means ``evaluated at $\bar a_i$", i.e., of the two $Q$-values that we compute for the two possible actions of agent $i$, we use the one corresponding to the global decision $\bar a$. The term after $\gamma$ is the $V$-value estimate for $s'$ based on $\bar\theta$, for which we compute the next global action $\bar a'$ with the current policy. The number of requests and thus the number of agents can change between subsequent time steps. This poses a numerical problem for the critic loss computation, which requires the same number of agents for $s$ and $s'$. We solve this problem by amending the requests in $s'$ when saving a transition to the replay buffer and provide details on this in the appendix.
The actor network obtains all vehicle states and requests for which a decision must be made in the current time step as an input. To deal with these (potentially) many inputs, we train a single request embedding and a single vehicle embedding to encode all requests and vehicles, respectively. To account for the variable number of requests and to let each agent focus on the parts of the input that are important for this particular agent, we equip the neural network with an attention mechanism \citep[cf.][]{Holler2019, Kullman2022}. Together with the request and vehicle embeddings for the agent and additional features, we pass the context computed by the attention mechanism to a sequence of feedforward layers. The critic network has the same architecture, but receives the global action as an additional input. We remove the action of the agent from this input, since the critic outputs $Q$-values for both possible actions. Further details on the neural networks, e.g., a formal description of the attention mechanism and hyperparameters, can be found in the appendix.
\section{Experiments}\label{sec:experiments}
To validate our method, we perform experiments based on historic taxi data that is publicly available for New York City \citep{NYCTLC2015}. We use a hexagon grid for spatial discretization and consider two different instances: one with 11 small zones (approx. 500 meters distance between neighboring zones) and one with 38 large zones (approx. 1 km distance), both in Manhattan. We consider the time interval from 8:30 am to 9:30 am during morning rush hour as one episode. Our data set contains data for 245 different dates in 2015, which we split into 200 training dates, 25 validation dates, and 20 test dates. We use a time step size of one minute and choose revenue and cost parameters such that a vehicle that serves a customer without empty driving achieves an operating profit margin of 10\%. We consider different numbers of vehicles to simulate different degrees of supply shortage; only cases with supply shortage are interesting, as the operator can serve each request immediately in the case of infinite supply, such that a myopic policy would be sufficient. For additional details on the data set, system setup, and hyperparameters, we refer to the appendix.
To benchmark our method, we compare its test performance against two ``classical", non RL-based, algorithms: a greedy policy and an MPC approach. The greedy policy accepts any request that can be served with a positive profit (accounting for the cost for empty driving to the pickup location) and rejects all others. It is a reasonable choice when there is no reliable estimate of future requests. If we have such an estimate, it is promising to apply MPC \citep[see, e.g.,][]{AlonsoMora2017}. We adapt this approach to our setting, using a request distribution estimate for mixed-integer-based receding-horizon optimization. Details on both benchmarks can be found in the appendix.
\section{Results and Discussion}\label{sec:results}
We provide plots illustrating the training process in the appendix. Figure~\ref{fig:box_plots} summarizes the performance of greedy, MPC, and our RL method on the test data for all considered instances. On average, our RL method always outperforms the greedy policy, by up to 5\% over the 20 test dates. For individual dates, RL outperforms greedy by up to 17\%. It performs by at most 6\% worse than greedy for less than 20\% of the individual dates. MPC is (substantially) worse than greedy and RL in many cases, although it sometimes outperforms the RL method. This means that MPC can provide a benefit in certain situations, but comes with an unstable performance across instances, which limits its practical applicability. In particular, MPC does not perform well in situations where there is a large shortage of vehicles, which are handled well by our RL method. Thus, our method provides a stable alternative, that always achieves at least the greedy performance and outperforms it by a substantial margin in many cases. Note that the order of magnitude of this performance improvement is significant for our application area \citep[cf.][]{SadeghiEshkevari2022}. Given the large scale at which AMoD systems operate, the seemingly small percentage improvements translate into significant monetary value for the operator.
\begin{figure}
\centering
\includegraphics[width=\textwidth]{plots/box_plots.pdf}
\caption{Test performance of MPC and our RL method compared to greedy (greedy is 0\%, values $<0\%$ indicate a performance worse than greedy). Each dot represents one test date.}
\label{fig:box_plots}
\end{figure}
Figure~\ref{fig:kl_divergence} shows the performance of MPC and RL for the instance with 38 large zones and different amounts of training data, i.e, different estimation qualities of the request probability distribution. Since our problem setting excludes fixed costs, additional resources, i.e., vehicles, are free of charge. Thus, the greedy policy performs better with very few or many vehicles, compared to instances with a medium number of vehicles. With few vehicles, many requests are available for each vehicle, such that vehicles are rarely idle or drive without a customer. With many vehicles, most requests can be served quickly without empty driving because vehicles are usually available. Consequently, with sufficient training data, for both MPC and our RL approach, the performance gain vs. greedy is largest for a medium number of vehicles. However, with few vehicles, MPC performs worse than greedy, as it is not robust against mistakes when sampling future requests. Such errors have a larger effect with fewer vehicles. With less training data, the performance gain of our method decreases by about one percentage point, but it remains reliably better than greedy. For all instances except the 250 vehicles case, the performance loss is much larger for MPC; it is not always able to sustain a performance better than greedy, even for instances where it outperforms greedy with more training data. With 250 vehicles, there are many resources free of charge, such that the mistakes made by MPC have such a small effect that it is robust against a poor estimation quality. Based on these observations, we conclude that our RL method is more robust against a poor estimation quality due to insufficient training data than MPC. These results might seem surprising, as RL is in general not very sample-efficient---although SAC has better sample-efficiency than most policy gradient-based algorithms, since it is an off-policy algorithm and uses a replay buffer. However, for our problem setting, less training data does not mean that the RL agents must learn from fewer samples, as the available training data can be replayed multiple times in the simulated environment. The performance loss that we observe for the RL method is more likely due to the decreasing diversity of the training data to which the RL agents are exposed, leading to less generalization.
\begin{figure}
\centering
\includegraphics[width=\textwidth]{plots/kl_divergence.pdf}
\caption{Mean test performance of MPC and RL vs. greedy as a function of the KL-divergence between the true request distribution and the distribution estimated from the training data (with different amounts of training data, resulting in different KL-divergence values). We use the request distribution estimated from the real data for 38 large zones as the true distribution and run the experiments with synthetic data sampled from this distribution.}
\label{fig:kl_divergence}
\end{figure}
Finally, a major advantage of our RL method over MPC is its shorter computational time during execution. We can train the network parameters offline in advance and easily scale the online execution, because the per-agent actor computations are fast and straightforward to parallelize. Figure~\ref{fig:runtime} shows that although we solve a combinatorial optimization problem in each time step, the computational time of our method is very short even for large system sizes. On the other hand, the computational time of MPC increases quickly for large instances: for 3000 vehicles, the action computation for a single step takes more than 30 seconds with MPC, while our RL method (including the matching) takes less than 0.2 seconds. Thus, the practical application of MPC at scale is greatly limited by its computational time, while our RL method can be scaled to much larger system sizes.
\begin{figure}
\centering
\includegraphics{plots/runtime.pdf}
\caption{Time to compute one action (mean over 60 steps, based on 38 large zones, number of requests scaled with number of vehicles). MPC solve refers to solving the mixed integer program (MIP), MPC total also includes the generation of the MIP instance in each step.}
\label{fig:runtime}
\end{figure}
\section{Conclusion}\label{sec:conclusion}
We consider the dispatching problem of a profit-maximizing AMoD operator with centralized control over a fleet of autonomous vehicles, who accepts (and serves) or proactively rejects requests in real-time. To solve this problem, we use a combination of multi-agent SAC with centralized final decision-making through weighted matching. Our experiments based on real-world data show that our method outperforms two strong benchmarks on most problem instances, that it is stable across these instances and robust against a poor estimation of the request distribution, and that it can be easily scaled to large system sizes. In future work, we will investigate the use of global instead of ego rewards. Furthermore, we will extend our framework to more complex use cases, e.g., dispatching and charge scheduling.
\newpage
|
\section{Introduction}\label{sec-1}
Let $V$ be a vector space over a field $K$. Recall that a linear map $R\colon V\otimes V\to V\otimes V$ is said to
be a solution of the Yang--Baxter equation if
\[(R\otimes{\id})({\id}\otimes R)(R\otimes{\id})=({\id}\otimes R)(R\otimes{\id})({\id}\otimes R).\]
This equation originates from papers by Baxter \cite{Ba} and Yang \cite{Ya} on statistical physics. The search for
solutions has attracted numerous studies in both mathematical physics and pure mathematics. Topics of interest include
quantum groups, Hopf algebras, Galois theory, group algebras, knot theory, radical rings, pre-Lie algebras, solvable
groups and Garside groups and algebraic structure of quadratic algebras (see for example
\cite{AngionoVendramin,CJDR,CJORet,Chouraqui,Dehornoy,EPGS,ESS99,FCC,GVdB,KasselBook,LeVe17,RumpRR,RumpClassical,SmokPre,SmokPre1}).
Because of the complexity of the solutions, Drinfeld in \cite{Dri1992} initiated the investigations of set-theoretic
solutions of the Yang--Baxter equation. These are solutions which are induced by a linear extension of a map
$r\colon X\times X\to X\times X$, where $X$ is a basis of $V$. In this case, $r$ satisfies
\[(r\times{\id})({\id}\times r)(r\times{\id})=({\id}\times r)(r\times{\id})({\id}\times r),\]
and one says that $(X,r)$ is a set-theoretic solution of the Yang--Baxter equation (for simplicity, throughout the paper,
we will simply call this a solution). For any $x,y \in X$, we put \[r(x,y)=(\lambda_x(y),\rho_y(x)).\] The solution
is said to be bijective if $r$ is bijective (involutive if $r^{2}=\id$), left non-degenerate (respectively right non-degenerate)
if all maps $\lambda_x$ (respectively all maps $\rho_x$) are bijective (and non-degenerate if it is left and right non-degenerate).
The work of Etingoff, Soloviev, Schedler \cite{ESS99} and Gateva-Ivanova and Van den Bergh \cite{GVdB} brought renewed
attention to Drinfeld's problem in the particular case of involutive solutions. In \cite{GVdB} the associated structure
monoid $M(X,r)$ and structure algebra $K[M(X,r)]$ (an algebra with defining quadratic relations naturally determined
by the solution $(X,r)$) were investigated, and it was shown that the structure algebra shares many ring and homological
properties of polynomial algebras in commuting variables rovided $X$ also is finite. The ring theoretical properties of
the structure algebra of an arbitrary finite bijective non-degenerate solution have been extensively further investigated,
see for example \cite{GIMa08,MR2927367,MR4278764,MR4213053,MR4024551,MR4105532}. In particular, it is shown that the algebraic
structure determines information of the type of solutions; for example a bijective finite non-degenerate solution is involutive
precisely when its structure algebra is prime (or equivalently, a domain). The introduction of algebraic structures, such
as braces and skew braces \cite{Rump2007, GV17}, breathes new life into the study of involutive and, more generally, bijective
non-degenerate solutions. In fact, any skew brace gives rise to a bijective non-degenerate solution to the Yang--Baxter equation
and conversely, one can always associate a skew brace to such a solution \cite{SmVe18}. In the same vein, in \cite{CoJeVAVe21x},
it is shown that arbitrary left non-degenerate solutions can be studied via the so-called YB-semitrusses (a structure that
specialises to skew braces in the case of bijective solutions).
Jones \cite{MR908150} and Turaev \cite{MR939474} investigated solutions $R\colon V\otimes V \to V\otimes V$ of the Yang--Baxter
equation, where $V$ is a finite-dimensional linear space, which are algebraic of degree two, that is $R^2=aR+b\id$. In case they
are bijective, they give rise to polynomial invariants of knots. In the combinatorial setting, such class of solutions
restricts to set-theoretic solutions $(X,r)$ satisfying $r^2=r$ or $r^2=\id$, since one does not have an addition (as in $\End(V\otimes V)$).
As discussed before, involutive solutions have been widely studied. Whereas, Lebed in \cite{Leb2017,MR4146852}
initiated to investigate idempotent solutions in order to identify the Hochschild cohomology of the plactic monoid and
to show that the quantum symmetrizer yields a quasi-isomorphism between a certain quotient of the braided chain complex for
an idempotent solution, and the Hochschild chain complex for its monoid, with the same coefficients.
Stanovsk\'{y} and Vojt\v{e}chovsk\'{y} in \cite{StVo21} proved that idempotent left non-degenerate solutions $(X,r)$ are in
one-to-one correspondence with twisted left Ward quasigroups, i.e., left quasigroups $(X,*)$ satisfying the identity $(x*y)*(x*z)=(y*y)*(y*z)$.
The latter are completely described in case $X$ has prime cardinality. They also showed that twisted Ward quasigroups are obtained by
twisting the left division operation in groups (that is $x*y=\psi^{-1}(x^{-1}\cdot y)$ for a group $(X,\cdotp)$), and these
correspond to idempotent latin solutions, i.e., such solutions are determined by a group structure on $X$ and an automorphism of this group.
In this paper, we investigate arbitrary finite left non-degenerate idempotent solutions $(X,r)$ and show that they are determined by a
left simple semigroup structure on $X$, i.e., by finitely many isomorphic finite groups and some mappings on $X$.
Our approach is different and makes use of two results. Firstly, Majid and Gateva-Ivanova \cite{GIMa08} proved that
any solution (without any restriction) can be extended to a solution on its structure monoid.
Hence the latter contains all information on solutions. Secondly, Ced\'o, Jespers and Verwimp \cite{CJV21} showed
that the structure monoid of any left non-degenerate solution is embeddable in the holomorph of the associated
derived structure monoid. The latter is the structure monoid of the associated rack solution. Its special
algebraic structure can then be exploited to obtain information on the original solution.
We prove that for left non-degenerate finite idempotent solutions $(X,r)$ we have much more rigid
control and, surprisingly, such solutions again are determined by a group and some maps on finitely many isomorphic
copies of this group (Theorem~\ref{solgeneral}). While the structure algebra $K[M(x,r)]$ always is left Noetherian,
it also is right Noetherian only if the number of these isomorphic copies is precisely one, or equivalently the
structure algebra is semiprime in case the field $K$ has characteristic zero (Theorem~\ref{algebrastructure}).
In Section \ref{sec-2}, we recall the necessary background on left non-degenerate solutions $(X,r)$ and the
associated structure monoid $M(X,r)$ and structure semigroup $S=S(X,r)$. In Section \ref{sec-3}, we deal
with such idempotent finite solutions. First, we show that $S$ has a natural decomposition into cancellative
semigroups $S_u$, indexed by elements $u$ in the diagonal $\Lambda =\{\lambda_x^{-1}(x):x\in X\}$. We show
that each $S_u$ is a subsemigroup of its group of quotients $G_u$, which turns out to be a finite conjugacy
group and thus has a finite torsion part, denoted $T(G_u)$. Furthermore, the union of these torsion-groups
is a left simple semigroup and it is in bijective correspondence with $X$. Hence, in the main theorem, we
show that the considered solutions $(X,r)$ are determined by such a left simple semigroup $(X,\cdot)$ and
some maps satisfying some identities. It is then shown that having a unique component $S_u$ (i.e., $S(X,r)$
is embeddable in a group) is equivalent with $(X,\cdot)$ being a group and having an automorphism $\varphi$
so that $\lambda_x (y) =x\cdot \varphi (y)$, and this on its turn is equivalent with its structure algebra
$K[M(X,r)]$ being right Noetherian (or, equivalently, sempiprime if $K$ has characteristic zero).
It turns out that $K[M(X,r)]$ always is a left Noetherian representable algebra (and hence satisfies
a polynomial identity) of Gelfand--Kirillov dimension one. So the algebraic structure of the structure algebra
determines precise information of the solution $r$. We finish the paper with three consequences. First the case
that $|\Lambda |=1$, or equivalently $S$ is embeddable in a group, also corresponds to $(X,r)$ being a latin
solution and such solutions are determined by a group structure on $X$ and an automorphism on this group
(Corollary~\ref{idempotentlatin}). Second the case that $|\Lambda|=|X|$ is simply determined by a permutation
on $X$ (Corollary~\ref{qbijective}). Third, there are only two types of left non-degenerate idempotent
solutions of prime cardinality (Corollary~\ref{primecase}). In Corollary~\ref{idempotentlatin} and
Corollary~\ref{primecase} we recover two results of \cite{StVo21}.
\section{Preliminaries}\label{sec-2}
Throughout we denote by $(X,r)$ a left non-degenerate set-theoretic solution of the Yang--Baxter equation
and we put $r(x,y)=(\lambda_x(y), \rho_y(x))$. So each $\lambda_x \colon X\to X$ is a bijection.
Being a solution of the Yang--Baxter equation translates into the following equalities to hold for all $x,y,z\in X$:
\begin{align}
\lambda_x(\lambda_y(z)) &=\lambda_{\lambda_x(y)}(\lambda_{\rho_y(x)}(z)),\label{YB1}\\
\lambda_{\rho_{\lambda_y(z)}(x)}(\rho_z(y)) &=\rho_{\lambda_{\rho_y(x)}(z)}(\lambda_x(y)),\label{YB2}\\
\rho_z(\rho_y(x)) &=\rho_{\rho_z(y)}(\rho_{\lambda_y(z)}(x)).\label{YB3}
\end{align}
Note that
\[r^2(x,y)=(\lambda_{\lambda_x(y)}(\rho_y(x)),\rho_{\rho_y(x)}(\lambda_x(y))).\]
Hence the solution $(X,r)$ is idempotent, i.e., $r^2=r$, if and only if
\begin{equation}
\lambda_{\lambda_x(y)}(\rho_y(x))=\lambda_x(y)\qquad\text{and}\qquad\rho_{\rho_y(x)}(\lambda_x(y))=\rho_y(x).\label{r2=r}
\end{equation}
The associated structure monoid of the solution $(X,r)$ is
\[M=M(X,r)=\free{x\in X\mid x \circ y = \lambda_x(y) \circ \rho_y(x),\; x,y\in X}\]
The structure semigroup is \[S=S(X,r)=M(X,r) \setminus \{1\}.\]
The associated derived solution is \[s\colon X^2 \to X^2\colon (x,y) \mapsto (y, \sigma_y(x)),\]
where $\sigma_y(x) = \lambda_y \rho_{\lambda^{-1}_x (y)} (x)$ and its associated structure monoid
(called the derived structure monoid) is denoted $A(X,r)$. Its operation is denoted additively.
Let $\pi\colon M\to A$ be the bijective $1$-cocycle described in \cite{CJV21}. So
\[A=A(X,r)=\free{\pi(x)\text{ for }x\in X\mid \pi(x)+\pi(y)=\pi(y)+\pi(\sigma_y(x)),\; x,y\in X}.\]
Furthermore, \[f\colon M\to A\rtimes{\Img\lambda}\colon m\mapsto (\pi (m), \lambda_{m})\]
is a monoid embedding, where \[\lambda\colon (M,\circ) \to \Aut(A,+)\colon m \mapsto \lambda_m\]
and $\lambda_m =\lambda_x$ if $m=x$. Abusing notation, we will identify $m$ with $f(m)$, i.e.,
we will write $m=(\pi(m),\lambda_m)$. For simplicity reasons, for $a\in A$, we will write $\lambda_a$
for $\lambda_{\pi^{-1}(a)}$. So we simply may write \[M=\{(a,\lambda_a):a\in A\}=\free{x=(x,\lambda_{x}):x\in X}.\]
Therefore, we have a mapping \[\lambda\colon A\to\Aut(A,+)\colon a\mapsto\lambda_a,\]
and, for $a,b\in A$, \[\lambda_{a+\lambda_a(b)}=\lambda_a\circ\lambda_b .\]
We also have a monoid antihomomorphism $\sigma\colon (A,+) \to \End (A,+)\colon a\mapsto \sigma_a$
(``extending'' the maps $\sigma_x$ for $x\in X$). So for $a,b\in A$, we have $a+b=b +\sigma_b (a)$
and thus $A+b\subseteq b+A$. In particular, each right ideal of $A$ is a two-sided ideal of $A$.
For a subset $B$ of $A$ we put, as in \cite{MR4024551}, \[B^e=\{(a,\lambda_a):a\in B\}\subseteq M.\]
Moreover, if $Z$ is a subset of a monoid $T$ then $\free{Z}$ denotes the submonoid of $T$ generated by $Z$.
In case $T$ is a group, $\gr(Z)$ stands for the subgroup of $T$ generated by $Z$.
A fundamental theorem of Gateva-Ivanova and Majid says that set-theoretic solutions are determined
by the solutions associated to their structure monoids.
\begin{thm}[{see \cite[Theorem 3.6]{GIMa08}}]\label{main}
Assume $(X,r)$ is a solution of the Yang--Baxter equation. If $M=M(X,r)$ then there exist a monoid morphism
$\lambda\colon(M,\circ)\to\Map(M,M)$ and a monoid antimorphism $\rho\colon(M,\circ)\to\Map(M,M)$, both naturally
extending the maps $\lambda_x$ and $\rho_x$, respectively, such that, for all $a,b,c\in M$,
\begin{align}
\rho_b(c\circ a) &= \rho_{\lambda_a(b)}(c)\circ \rho_b(a), \label{eqrho} \\
\lambda_b(a \circ c) &= \lambda_b(a) \circ \lambda_{\rho_a(b)}(c),
\label{eqlambda}
\end{align}
and
\begin{equation}\label{eqproduct}
a\circ b=\lambda_{a}(b)\circ\rho_{b}(a).
\end{equation}
Let $r_M\colon M\times M\to M\times M$ be defined by $r_M(a,b)=(\lambda_a(b),\rho_b(a))$, for all $a,b\in M$.
Then $(M,r_M)$ is a set-theoretic solution of the Yang--Baxter equation. Obviously, $r_M$ extends the solution $r$.
\end{thm}
So set-theoretic solutions of the Yang--Baxter equation are determined by the presentation of their structure monoid.
In this paper, we will describe the presentations of the structure semigroup of left non-degenerate idempotent solutions,
and as a consequence, we obtain a description of all such solutions.
So from now on all solutions $(X,r)$ are idempotent and left non-degenerate. The diagonal map
\[q\colon X\to X\colon x\mapsto \lambda_x^{-1}(x)\] will play a crucial role. Put \[\Lambda =\Lambda(X,r)=\Img(q).\]
Clearly $r^2=r$ means that $\rho_y(x) = q(\lambda_{x}(y))$ and thus \[r(x,y)=(\lambda_x(y), q(\lambda_x(y)).\]
Furthermore, the left non-degeneracy yields \[A(X,r)=\free{x\in X\mid x+y =y+y,\;x,y\in X}=\bigcup_{x\in X}\free{x};\]
clearly $\free{x}=\{nx:n\ge 0\}$. Note that if $(X,r)$ is idempotent and left non-degenerate then each element of
$A(X,r)$ is uniquely determined by its length and its last letter; in particular $A(X,r)$ is left cancellative.
We will use this observation freely in the rest of the paper without further reference. The natural length function
on $A$ and $M$ is denoted $|a|$, for $a\in A$, or $a\in M$. All mapping $\lambda_a$ are length preserving.
\section{Main results}\label{sec-3}
In this section, we deal with arbitrary finite left non-degenerate idempotent solutions $(X,r)$. First, we show that
the structure semigroup can be decomposed as a finite union of cancellative semigroups $S_u$, in such a way that it
is graded by $\Lambda$. Next, we prove that each $S_u$ has a group of quotients $G_u$ which is a finite conjugacy group
and thus that its periodic subgroup $T(G_u)$ is finite. This will allow us to obtain a left simple semigroup structure
on $X$ and prove the main results.
\begin{notation}
Throughout the paper $d$ denotes a multiple of the exponent of the permutation group $\mathcal{G}(X,r)=\gr(\lambda_x:x\in X)$
of the solution $(X,r)$. In particular, $\lambda_a^d=\id$ for all $a\in A(X,r)$.
\end{notation}
\begin{prop}\label{decomposition}
Let $(X,r)$ be a left non-degenerate idempotent solution of the Yang--Baxter equation. Let $S=S(X,r)$
and $\Lambda=\Lambda(X,r)$. For $u\in\Lambda$ put \[S_u=\{(a,\lambda_a)\in S:\lambda_a^{-1}(a)=|a|u\}.\]
Then:
\begin{enumerate}
\item each $S_u$ is a cancellative subsemigroup of $S$.
\item $S=\bigcup_{u\in\Lambda} S_u$, a disjoint union.
\item $S_u \circ S_v \subseteq S_v$ for all $u,v\in\Lambda$.
\item if $s,t\in S$ then $s\circ t\in S_u$ if and only if $t\in S_u$.
\end{enumerate}
In particular, equipping $\Lambda$ with the band structure via $uv=v$ the semigroup $S$
is a band graded semigroup with homogeneous components that are cancellative semigroups.
\begin{proof}
First note that if $u=q(x)\in\Lambda$ then $x\in S_u$. So all $S_u$ are non-empty.
Let $u,v\in\Lambda$ and assume $(a,\lambda_a)\in S_u$ and $(b,\lambda_b)\in S_v$. Then
\begin{align*}
\lambda_{a+\lambda_a(b)}^{-1}(a+\lambda_a(b)) & =(\lambda_a\lambda_b)^{-1}(a+\lambda_a(b))
=\lambda_b^{-1}(\lambda_a^{-1}(a+\lambda_a(b)))\\
& =\lambda_b^{-1}(\lambda_a^{-1}(a)+b)=\lambda_b^{-1}(\lambda_a^{-1}(a))+\lambda_b^{-1}(b)\\
& =\lambda_b^{-1}(|a|u)+|b|v=|a+\lambda_a(b)|v.
\end{align*}
Hence $(a,\lambda_a)\circ (b,\lambda_b)=(a+\lambda_a(b), \lambda_{a+\lambda_a(b)})\in S_v$ and thus
$S_u\circ S_v\subseteq S_v$. In particular, each $S_u$ is a subsemigroup of $S$. Moreover, $S_u\cap S_v=\varnothing$ for $u\ne v$.
Assume $(a,\lambda_a),(b,\lambda_b),(c,\lambda_c)\in S_u$. If $(a,\lambda_a)\circ (c,\lambda_c)=(b,\lambda_b)\circ(c,\lambda_c)$
then $a+\lambda_a(c)=b+\lambda_b(c)$ and $\lambda_a=\lambda_b$. This implies $|a|=|b|$ and thus
\[\lambda_a^{-1}(a)=|a|u=|b|u=\lambda_b^{-1}(b)=\lambda_a^{-1}(b),\] which yields $a=b$. Similarly, we show that
$(c,\lambda_c)\circ(a,\lambda_a)=(c,\lambda_c)\circ (b,\lambda_b)$ leads to $a=b$. Hence, $S_u$ is a cancellative semigroup.
To prove that $S=\bigcup_{u\in\Lambda}S_u$, it is enough to show that for each positive integer $n$ and $x\in X$ we have that
$u=\lambda_{nx}^{-1}(x)\in\Lambda$, because then $\lambda_{nx}^{-1}(nx)=n\lambda_{nx}^{-1}(x)=|nx|u$ and thus
$(nx,\lambda_{nx})\in S_u$. We prove this by induction on $n$. Clearly, by definition of $\Lambda$, the result holds for $n=1$.
So, assume that $n>1$ and the result holds for $n-1$. Put $y=\lambda_x^{-1}(x)\in\Lambda\subseteq X$. Because
\[\lambda_{nx}=\lambda_{x+\lambda_x(\lambda_x^{-1}((n-1)x))}=\lambda_{x+\lambda_x((n-1)y)}=\lambda_x\lambda_{(n-1)y},\]
we get \[u=\lambda_{nx}^{-1}(x)=\lambda_{(n-1)y}^{-1}\lambda_x^{-1}(x)=\lambda_{(n-1)y}^{-1}(y)\in\Lambda,\] as desired.
This finishes the proof of (1), (2) and (3).
Part (4) follows at once from part (3).
\end{proof}
\end{prop}
\begin{rem}\label{qinlambda}
In the proof of Proposition~\ref{decomposition} we have shown that $\lambda_{nx}^{-1}(x)\in\Lambda$
for all $x\in X$ and positive integers $n$.
\end{rem}
\begin{lem}\label{quotientgroup}
If $u\in\Lambda$ then:
\begin{enumerate}
\item $c_u\coloneqq(du,\id) \in S_u$ is a central element in $S_u$.
\item $S_u =\free{c_u}\circ F_u$, where $F_u\coloneqq\{(a,\lambda_a)\in S_u:0<|a|<d\}$ is a finite set.
\item $S_u$ has a group of quotients, denoted by $G_u$, that is a finite conjugacy group (FC-group).
In particular, its elements of finite order form a finite group, denoted by $T(G_u)$.
\item $G_u/T(G_u)\cong \mathbb{Z}$ and there exists an element $x_u\in X$ so that $(x_u,\lambda_{x_u})\in S_u$
and $G_u/T(G_u)$ is generated by $(x_u,\lambda_{x_u}) \circ T(G_u)\in G_u/T(G_u)$.
\end{enumerate}
In particular, $G_u \cong T(G_u) \rtimes_{\sigma_u} {\gr(x_u)}\cong T(G_u)\rtimes\mathbb{Z}$, a semidirect product,
and the action $\sigma_u$ is given by conjugation by $x_u$.
Furthermore, $S=S(X,r)$ is cancellative if and only if $|\Lambda|=1$.
\begin{proof}
As $\lambda_{du}=\id$, we have $\lambda_{du}^{-1}(du)=du$ and thus $c_u\in S_u$. Moreover, if $(nx,\lambda_{nx})\in S_u$
for some positive integer $n$ and some $x\in X$ then $n\lambda_{nx}(u)=\lambda_{nx} (nu)=nx$. Hence $\lambda_{nx}(u)=x$,
which leads to
\begin{align*}
(nx,\lambda_{nx})\circ c_u & =(nx,\lambda_{nx})\circ(du,\id)=(nx+\lambda_{nx}(du),\lambda_{nx})\\
& =(nx+d\lambda_{nx}(u),\lambda_{nx})=(nx+dx,\lambda_{nx})=((n+d)x,\lambda_{nx}).
\end{align*}
Therefore,
\begin{align*}
c_u\circ(nx,\lambda_{nx}) & =(du,\id)\circ (nx,\lambda_{nx})=(du+nx,\lambda_{nx})\\
& = ((d+n)x,\lambda_{nx})=(nx,\lambda_{nx})\circ c_u.
\end{align*}
So part (1) follows.
Let $(nx,\lambda_{nx})\in S_u$ and write $n=kd+l$ for some non-negative integer $k$ and some integer $0<l<d$. Then
\begin{align*}
c_u^k\circ(lx,\lambda_{lx}) & =(du,\id)^k\circ(lx,\lambda_{lx})=(kdu,\id)\circ(lx,\lambda_{lx})\\
& =(kdu+lx,\lambda_{lx})=(kdx+lx,\lambda_{lx})=(nx,\lambda_{lx}).
\end{align*}
In particular, $\lambda_{nx}=\lambda_{lx}$. Moreover, $x=\lambda_{nx}(u)=\lambda_{lx}(u)$, which implies
$(lx,\lambda_{lx})\in F_u$, and part (2) follows.
Next, observe that for any $(a,\lambda_a)\in S_u$, we have
\begin{equation}
(a,\lambda_a)^d=(d\lambda_a^{-1}(a),\id)=(d|a|u,\id)=(du,\id)^{|a|}=c_u^{|a|}.\label{2.4-1}
\end{equation}
Thus, $G_u\coloneqq S_u\circ\free{c_u}^{-1}$ is the group of quotients of $S_u$ and the central subgroup
$\gr(c_u)=\free{c_u}\circ\free{c_u}^{-1}$ of $G_u$ is such that the group $G_u/\gr(c_u)$ is finite (of order
not exceeding $|F_u|\le d|X|$). Hence $G_u$ has a central (infinite cyclic) subgroup of finite index, and
thus $G_u$ is an FC-group. Consequently, its torsion subgroup $T(G_u)$ is finite and $G_u/T(G_u)$ is a
torsion-free abelian group. In particular, part (3) follows.
As $u\in\Lambda$, there exists $x_u\in X$ such that $u=\lambda_{x_u}^{-1}(x_u)$. In particular,
$x_u=(x_u,\lambda_{x_u})\in S_u$ and $x_u^d=c_u$. Now, suppose $(nx,\lambda_{nx})\in S_u$. Then \eqref{2.4-1}
implies that $(nx,\lambda_{nx})^d=c_u^n=x_u^{dn}=(x_u^n)^d$. As the group $G_u/T(G_u)$ is torsion-free abelian,
it follows that $(nx,\lambda_{nx})\in T(G_u)\circ x_u^n$. So, indeed
$G_u\cong T(G_u)\rtimes_{\sigma_u}{\gr((x_u,\lambda_{x_u}))}\cong T(G_u)\rtimes\mathbb{Z}$ and the action of
$\sigma_u$ is conjugation by $(x_u,\lambda_{x_u})$. This finishes the proof of part (4).
The last part of the result follows from the fact that if $(x,\lambda_x)\in S_u$ and $(y,\lambda_y)\in S_v$,
with $x,y\in X$ and distinct $u,v\in\Lambda$ then
$(x,\lambda_x)^{d} \circ (y,\lambda_y)^{d} =c_u \circ c_v =c_v^{2}=(y,\lambda_y)^d\circ (y,\lambda_y)^{d}$,
while $(x,\lambda_x)^{d}=c_u \ne (y,\lambda_y)^d=c_v$.
\end{proof}
\end{lem}
Next, we describe the periodic elements of $G_u$. To do so we need to introduce some notation, which will be useful later as well.
Namely, for $u\in\Lambda$ define \[X_u\coloneqq\{x\in X:\lambda_{dx}(u)=x\}=\{x\in X:(dx,\lambda_{dx})\in S_u\}.\]
As $S(X,r)=\bigcup_{u\in\Lambda}S_u$, it is clear that we have a decomposition $X=\bigcup_{u\in\Lambda}X_u$.
Moreover, for $x\in X_u$ put \[t_{u,x}\coloneqq(dx,\lambda_{dx})\circ(du,\id)^{-1}=(dx,\lambda_{dx})\circ c_u^{-1}\in G_u.\]
\begin{lem}\label{lemmatorsion}
If $u\in\Lambda$ then $T(G_u)=\{t_{u,x}:x\in X_u\}$. Furthermore, if $x,y\in X_u$ then:
\begin{enumerate}
\item $t_{u,u}$ is the identity of $T(G_u)$.
\item $t_{u,x}\ne t_{u,y}$ provided $x\ne y$. In particular, $|T(G_u)|=|X_u|$.
\item $t_{u,x}\circ t_{u,y}=t_{u,\lambda_{dx}(y)}$.
\item $t_{u,x}=(kdx,\lambda_{kdx})\circ(kdu,\id)^{-1}$ for each positive integer $k$.
\item the order of the element $t_{u,x}$ is a divisor of $d$.
\item $\lambda_x=\lambda_{dx}\lambda_u$ and if $q(u)=u$ then $t_{u,x}=(x,\lambda_x) \circ (u,\lambda_u)^{-1}$.
\item $X=\bigcup_{u\in\Lambda}(\{ (x,\lambda_x): x\in X \} \cap S_u)$ and if $(x,\lambda_x)\in S_u$
then $(x,\lambda_x)\circ(x_u,\lambda_{x_u})^{-1}\in T(G_u)$.
\end{enumerate}
\begin{proof}
Because $\lambda_{du}=\id$, we get $u\in X_u$ and $t_{u,u}=1$. So part (1) is obvious. Since $dx\ne dy$, part (2) also follows.
Next, write as before $c_u=(du,\id)\in G_u$ and observe that
\begin{align*}
t_{u,x} \circ t_{u,y} &=(dx,\lambda_{dx})\circ c_u^{-1} \circ (dy,\lambda_{dy})\circ c_u^{-1}
=(dx+d\lambda_{dx}(y),\lambda_{dx}\lambda_{dy})\circ c_u^{-2}\\
& =(du+d\lambda_{dx}(y),\lambda_{dx}\lambda_{dy})\circ c_u^{-2}
=c_u\circ(d\lambda_{dx}(y),\lambda_{dx}\lambda_{dy})\circ c_u^{-2}\\
& =(d\lambda_{dx}(y),\lambda_{d\lambda_{dx}(y)})\circ c_u^{-1}=t_{u,\lambda_{dx}(y)}.
\end{align*}
Hence part (3) follows as well.
Choose $(a,\lambda_a)\in S_u$ with $a=kdx$ for some positive integer $k$ and some $x\in X$. Then
\begin{align}
\begin{aligned}
c_u\circ(a,\lambda_a) & =(du,\id)\circ(kdx,\lambda_{kdx})=(du+kdx,\lambda_{kdx})=((k+1)dx,\lambda_{kdx}),\\
c_u^k\circ(dx,\lambda_{dx}) & =(kdu,\id)\circ(dx,\lambda_{dx})=(kdu+dx,\lambda_{dx})=((k+1)dx,\lambda_{dx}).
\end{aligned}\label{2.5-1}
\end{align}
In particular, $\lambda_{dx}=\lambda_{kdx}=\lambda_a$. Hence $\lambda_{dx}(u)=\lambda_a(u)=u$ because $(a,\lambda_a)\in S_u$,
and thus $x\in X_u$. It also follows from \eqref{2.5-1} that $(a,\lambda_a)\circ c_u^{-k}=(dx,\lambda_{dx})\circ c_u^{-1}=t_{u,x}$.
So part (4) follows. As $S_u$, and in consequence, its central localization $G_u=S_u\free{c_u}^{-1}$, inherits gradation by length
from $S$, we know that elements of $T(G_u)$ are of length zero, and thus each such element may be written in the form
$(a,\lambda_a)\circ c_u^{-k}$, where $(a,\lambda_a)\in S_u$ is of length $kd$. This yields $T(G_u)=\{t_{u,x}:x\in X_u\}$.
Since $(dx,\lambda_{dx})^d=c_u^d$ (see \eqref{2.4-1} in the proof of Lemma \ref{quotientgroup}), we get $t_{u,x}^d=1$.
In particular, part (5) follows.
We prove part (6). Because $\lambda_{dx}(u)=x$, we obtain $du+x=dx+\lambda_{dx}(u)$, and thus
\begin{align*}
c_u\circ(x,\lambda_x) & =(du,\id)\circ(x,\lambda_x)=(du+x,\lambda_x)=(dx+\lambda_{dx}(u),\lambda_x)\\
& =(dx,\lambda_{dx}) \circ (u,\lambda_u)=(dx+\lambda_{dx}(u),\lambda_{dx}\lambda_u).
\end{align*}
Therefore, $\lambda_x=\lambda_{dx}\lambda_u$. Furthermore, if $q(u)=u$ then $(u,\lambda_u)\in S_u$ and thus
$t_{u,x}=(dx,\lambda_{dx})\circ c_u^{-1}=(x,\lambda_{x}) \circ (u,\lambda_u)^{-1}$. Hence (6) follows.
Finally part (7) follows from the previous grading argument, which says that $T(G_u)$ consists of the elements of degree $0$.
\end{proof}
\end{lem}
\begin{rem}\label{fixedpoint}
We know that if $(X,r)$ is a left non-degenerate idempotent solution of the Yang--Baxter equation and $a,b\in A(X,r)$
are of equal length then $\lambda_a=\lambda_b$ or $\lambda_a\lambda_b^{-1}$ has no fixed-points. Indeed, if
$\lambda_a(x)=\lambda_b(x)$ for some $x\in X$ then
\[\lambda_a\lambda_x=\lambda_{a+\lambda_a(x)}=\lambda_{b+\lambda_b(x)}=\lambda_b\lambda_x,\]
and thus $\lambda_a=\lambda_b$. Since $\lambda_{dx}=\lambda_x\lambda_u^{-1}$ for $u\in\Lambda$ and $x\in X_u$,
it follows $\lambda_{dx}=\id$ or $\lambda_{dx}$ does not have fixed-points.
\end{rem}
Recall that a finite simple semigroup is isomorphic with a matrix semigroup $\mathcal{M}(G,n,m,P)$ consisting of elements
$(g,i,j)$ with $g\in G$, $1\le i\le n$ and $1\le j\le m$. The multiplication is given by the $m\times n$ sandwich matrix
$P=(p_{ji})$ with entries in $G$ as follows: $(g,i,j)(h,k,l) =(gp_{jk}h,i,l)$. Conversely, any such semigroup (even if $G$
is infinite) is a simple semigroup. Furthermore, such a semigroup is left cancellative if and only if $n=1$ and in such a
case, one may assume that each entry of $P$ is the identity element of the group. We will denote such a matrix as $J$.
Clearly $\mathcal{M}(G,1,n,J) =\bigcup_{j=1}^{m}G_j$, where $G_j=\{(g,1,j):g\in G\}$. Moreover, in this case, $G_iG_j=G_j$ and
all $G_i$ are isomorphic and they are the maximal subgroups. Note that every idempotent of this semigroup is a left identity.
For details we refer the reader to \cite{CP}.
We now prove that \[T(G^e)\coloneqq\bigcup_{u\in\Lambda}T(G_u)\]
is equipped with a semigroup structure that extends each group $T(G_u)$.
\begin{lem}\label{infotorsioncover}
$T(G^e)$ is a left cancellative simple semigroup for the binary operation $\circ$ defined as
\begin{equation}
t_{u,x}\circ t_{v,y}\coloneqq t_{v,\lambda_{dx}(y)}\label{defproducttorsion}
\end{equation}
for $u,v\in\Lambda$, $x\in X_u$ and $y\in X_v$. Its maximal subgroups are the groups $T(G_u)$. Hence:
\begin{enumerate}
\item $T(G_u) \cong T(G_v)$ for all $u,v\in\Lambda$.
\item $T(G^e) \cong \mathcal{M}(T(G_u),1,|\Lambda|,J)$ for some (and thus any) $u\in\Lambda$.
\item $|X| = |\Lambda|\cdot |T(G_u)|$ for any $u\in\Lambda$.
\item $X=\bigcup_{u\in\Lambda}X_u$ is a disjoint union and $|X_u|=|X_v|$ for all $u,v\in\Lambda$.
\end{enumerate}
\begin{proof}
First notice that
\[(dx,\lambda_{dx})\circ(dy,\lambda_{dy})=(dx+\lambda_{dx}(dy),\lambda_{dx}\lambda_{dy})=(du+\lambda_{dx}(dy),\lambda_{dx}\lambda_{dy})\]
and also that \[(du,\id)\circ(\lambda_{dx}(dy),\lambda_{\lambda_{dx}(dy)})=(du+\lambda_{dx}(dy),\lambda_{\lambda_{dx}(dy)}).\]
Hence, it follows that
\begin{equation}
\lambda_{dx} \lambda_{dy} =\lambda_{\lambda_{dx}(dy)}=\lambda_{d\lambda_{dx}(y)}.\label{2.7-1}
\end{equation}
In particular, $\lambda_{d\lambda_{dx}(y)}(v)=\lambda_{dx}\lambda_{dy}(v)=\lambda_{dx}(y)$,
which assures that $\lambda_{dx}(y)\in X_v$ and thus $t_{v,\lambda_{dx}(y)}\in G_v$.
Next we verify that the operation \eqref{defproducttorsion} is associative. Hence, if additionally
$w\in\Lambda$ and $z\in X_w$ then using \eqref{2.7-1} we get
\begin{align*}
t_{u,x}\circ(t_{v,y}\circ t_{w,z}) & =t_{u,x}\circ t_{w,\lambda_{dy}(z)}=t_{w,\lambda_{dx}\lambda_{dy}(z)}\\
& =t_{w,\lambda_{d\lambda_{dx}(y)}(z)}=t_{v,\lambda_{dx}(y)}\circ t_{w,z}=(t_{u,x}\circ t_{v,y})\circ t_{w,z}.
\end{align*}
Consequently, we have shown that $T(G^e)$ is semigroup that contains each $T(G_u)$ as a subgroup and $T(G_u)\circ T(G_v)=T(G_v)$
for all $u,v\in\Lambda$. It follows that $T(G^e)$ is a left cancellative simple semigroup with maximal subgroups $T(G_u)$.
Hence all $T(G_u)$ are isomorphic, and in particular they all have the same cardinality and
\[T(G^e)\cong\mathcal{M}(H,1,|\Lambda|,J),\] where $H$ is a finite group isomorphic to each $T(G_u)$.
The remaining statements of the result are now obvious.
\end{proof}
\end{lem}
The next step is to prove that \[G^e\coloneqq\bigcup_{u\in\Lambda}G_u\] may be equipped with a semigroup
structure so that each $G_u$ is a subgroup and $S(X,r)$ is a subsemigroup. From Lemma~\ref{quotientgroup}
we know that an arbitrary element of $G_u$ can be written in the form
$(a,\lambda_a)\circ c_u^n$ with $(a,\lambda_a)\in S_u$ and $n\in\mathbb{Z}$. Recall that $c_u=(du,\id)\in Z(S_u)$.
\begin{thm}\label{descriptionsimplecover}
$G^e$ is left cancellative simple semigroup for the binary operation $\circ$ defined as
\[((a,\lambda_a)\circ c_u^n)\circ((b,\lambda_b)\circ c_v^m)=(a+\lambda_a(b),\lambda_{a+\lambda_a(b)})\circ c_v^{n+m}\]
for $u,v\in\Lambda$, $(a,\lambda_a)\in S_u$, $(b,\lambda_b)\in S_v$ and $n,m\in\mathbb{Z}$. Moreover, $S=S(X,r)$ is the
subsemigroup of
\[G^e=\bigcup_{u\in \Lambda}G_u =\bigcup_{u\in \Lambda}(T(G_u)\rtimes_{\sigma_u}\gr((x_u,\lambda_{x_u})))
=\bigcup_{u\in \Lambda}(T(G_u)\rtimes_{\sigma_{u}}\mathbb{Z})\cong\mathcal{M}(G_u,1,|\Lambda|,J),\]
generated by the elements $t\circ (x_{u},\lambda_{x_u})$, with $t\in T(G_u)$ and $u\in \Lambda$.
Furthermore, \[T(G_u)=\{(x,\lambda_x)\circ(x_u,\lambda_{x_u})^{-1}:x\in X \text{ with } q(x)=u\}.\]
In particular, the structural algebra $K[S(X,r)]$ is a subalgebra of the $\Lambda$-graded ring
\[\bigoplus_{u\in \Lambda} K[G_{u}]=\bigoplus_{u\in \Lambda}K[T(G_u)][t_u^{\pm 1};\sigma_u],\]
where each summand is a skew Laurent polynomial algebra over the group algebra $K[T(G_u)]$ of the finite group $T(G_u)$.
\begin{proof}
Clearly, if the proposed binary operation yields a semigroup then it contains $S$ as a subsemigroup
(as one then multiplies elements with $n=m=0$ precisely as in $S$).
We need to prove that the operation is well-defined. Notice that
\[(a+\lambda_a(b),\lambda_{a+\lambda_a(b)})=(a,\lambda_a)\circ(b,\lambda_b)\in S_u\circ S_v\subseteq S_v,\]
and thus indeed $(a+\lambda_a(b),\lambda_{a+\lambda_a(b)})\circ c_v^{n+m}\in G_v$.
Now, we verify that the product $\circ$ defined in $G^e$ is independent of representatives in $G_u$ and $G_v$.
So, suppose first that $(a,\lambda_a)\circ c_u^n=(a',\lambda_{a'})\circ c_u^{n'}$ for some $(a',\lambda_{a'})\in S_u$
and some $n'\in\mathbb{Z}$. Without loss of generality we may assume that $n\ge n'$. Then, in $S_u$, we have
$(a',\lambda_{a'})=(a,\lambda_a)\circ c_u^{n-n'}=c_u^{n-n'}\circ (a,\lambda_a)$. In particular,
$\lambda_{a'}=\lambda_a$ and $a'=(n-n')du+a$. Since $(a+\lambda_a(b),*)\in S_v$, we get
\begin{align*}
(a'+\lambda_{a'}(b),*)\circ c_v^{n'+m} & =((n-n')du+a+\lambda_a(b),*)\circ c_v^{n'+m}\\
& =c_u^{n-n'}\circ(a+\lambda_a(b),*)\circ c_v^{n'+m}\\
& =c_v^{n-n'}\circ(a+\lambda_a(b),*)\circ c_v^{n'+m}\\
& =(a+\lambda_a(b),*)\circ c_v^{n-n'}\circ c_v^{n'+m}\\
& =(a+\lambda_a(b),*)\circ c_v^{n+m}.
\end{align*}
So the product is independent of the choice of the representative of the left factor.
Next, we verify the independence of the right factor. So, suppose $(b,\lambda_b)\circ c_v^m=(b',\lambda_{b'})\circ c_v^{m'}$
for some $(b',\lambda_{b'})\in S_v$ and some $m'\in\mathbb{Z}$. Again, without loss of generality, we may assume that $m\ge m'$.
Then, in $S_v$, we have $(b',\lambda_{b'})=(b,\lambda_b)\circ c_v^{m-m'}=c_v^{m-m'}\circ(b,\lambda_b)$. In particular,
$b'=(m-m')dv+b$. Since $(a+\lambda_a(b),*)\in S_v$, we get
\begin{align*}
(a+\lambda_a(b'),*)\circ c_v^{n+m'} & =(a+\lambda_a((m-n')dv+b),*)\circ c_v^{n+m'}\\
& =(a+(m-m')d\lambda_a(v)+\lambda_a(b),*)\circ c_v^{n+m'}\\
& =((m-m')dv+a+\lambda_a(b),*)\circ c_v^{n+m'}\\
& =c_v^{m-m'}\circ(a+\lambda_a(b),*)\circ c_v^{n+m'}\\
& =(a+\lambda_a(b),*)\circ c_v^{m-m'}\circ c_v^{n+m'}\\
& =(a+\lambda_a(b),*)\circ c_v^{n+m}.
\end{align*}
So the product is indeed well-defined. Notice that the product of $(a,\lambda_a)\circ c_u^n$
and $(b,\lambda_b)\circ c_v^m$ is determined by the product of $(a,\lambda_a)$ and $(b,\lambda_b)$
in $S$ and the product of the elements $c_u^n\circ c_v^m=c_v^{n+m}$. Clearly this definition of the
product on the elements of type $c_u^n$, for various $u\in\Lambda$ and $n\in\mathbb{Z}$, is associative.
Hence, the proposed product on $G^e$ yields an associative product.
It is readily verified that the elements $t_{u,u}$ are left identities of $G^e$. Since each $G_u$ is a
group we obtain that $G^e$ is a left cancellative simple semigroup with maximal subgroups the groups $G_u$;
in particular all groups $G_u$ are isomorphic. So $G^e \cong\mathcal{M}(H,1,|\Lambda|,J)$, where $H$ is a group
that is isomorphic to $G_u\cong T(G_u)\rtimes{\gr(x_u)}$ and $\gr(x_u)\cong\mathbb{Z}$. It follows that
\[G^e \cong \mathcal{M}(G_u,1,|\Lambda|,J).\] Next we prove that the generators $(x,\lambda_x)$ are of the
form $t\circ (x_u,\lambda_{x_u})$, with $t\in G_u$. Indeed, let $u\in \Lambda$ be such that
$(x,\lambda_x)\in S_u$, i.e., $q(x)=u$. By Lemma~\ref{lemmatorsion},
$t\coloneqq(x,\lambda_x)\circ(x_u,\lambda_{x_u})^{-1}\in T(G_u)$.
Hence $(x,\lambda_x) =t\circ (x_{u},\lambda_{x_u})\in T(G_u)\circ (x_u,\lambda_{x_u})$. Finally,
since the cardinality of $\bigcup_{u\in \Lambda} T(G_u)\circ (x_u, \lambda_{x_u})$ is the same as
the cardinality of $X$, it follows that all elements in the union belong to the generating set
$\{(x,\lambda_x):x\in X\}$ of $S(X,r)$ and the result follows.
\end{proof}
\end{thm}
Recall that $r(x,y) =(\lambda_x (y), q(\lambda_x(y))$, for $x,y\in X$. From Lemma~\ref{lemmatorsion}
we also know that $\lambda_x = \lambda_{dx}\lambda_u$ for $x\in X_u$. Furthermore, from Lemma~\ref{infotorsioncover},
$t_{u,x} \circ t_{v,y}=t_{v,\lambda_{dx}(y)}$. Hence, via the bijective mapping (see Lemma~\ref{infotorsioncover})
\[f\colon X\to T(G^e)\colon x\mapsto t_{u,x}\text{ if $x\in X_u$ for some }u\in\Lambda,\]
we obtain a semigroup structure $(X,\cdotp)$ on $X$ for the binary operation $\cdot$ defined for $x,y\in X$ as
\[x\cdot y =\lambda_{dx}(y).\] Then $(X,\cdotp)$ is a simple left cancellative semigroup with maximal subgroups
isomorphic with $f^{-1}(G_u)$ for $u\in\Lambda$. Note that the elements $u\in\Lambda$ correspond to the idempotents
$t_{u,u}$ of $T(G^e)$ and hence the elements $u\in\Lambda$ are left identities in $(X,\cdotp)$. Furthermore, for
$x\in X_u$ and $y\in X$, \[\lambda_x(y) = x\cdot \lambda_u (y).\] Clearly we then get, for $x\in X_u$ and $y\in X$,
\[r(x,y) =(x\cdot \lambda_u (y),q(x\cdot \lambda_u (y)) ).\]
\begin{lem}\label{qq}
We have $q^k(x)=\lambda_{kx}^{-1}(x)$ for each $x\in X$ and each positive integer $k$. Moreover, $q^{d+1}=q$.
\begin{proof}
We prove this by induction on $k$. For $k=1$ this is clear. Assume now that it holds for $k$.
Put $y=\lambda_{kx}^{-1}(x)=q^k(x)\in X$. Then we have $\lambda_{(k+1)x}=\lambda_{kx+\lambda_{kx}(y)}=\lambda_{kx}\lambda_y$
and thus \[\lambda_{(k+1)x}^{-1}(x)=\lambda_y^{-1}\lambda_{kx}^{-1}(x)=\lambda_y^{-1}(y)=q(y)=q^{k+1}(x),\] as desired.
Next, notice that if $u=q^d(x)\in X$ then we know by the first paragraph that $u\in\Lambda$ and $x\in X_u$.
Thus, by Lemma~\ref{lemmatorsion}, we get $\lambda_x=\lambda_{dx}\lambda_u$.
Therefore, $\lambda_{(d+1)x}=\lambda_{dx+\lambda_{dx}(u)}=\lambda_{dx}\lambda_u=\lambda_x$. Consequently,
$q^{d+1}(x)=\lambda_{(d+1)x}^{-1}(x)=\lambda_x^{-1}(x)=q(x)$.
\end{proof}
\end{lem}
Note that Lemma \ref{qq} implies that the map $q$ is invertible if and only if $q^d=\id$, or equivalently $\Lambda=X$.
We introduce one more notation. Namely, for $x\in X$, put \[\varphi_x\coloneqq\lambda_{q^d(x)}.\]
In particular, $\varphi_x=\varphi_u$ for each $x\in X_u$. So \[\lambda_x (y) = x\cdot\lambda_{q^d(x)}(y)=x\cdot \varphi_x(y).\]
Hence the solution $r$ has the form \[r(x,y)=(x\cdot\varphi_x(y),q(x\cdot\varphi_x(y))).\]
As $r^2=r$, the idempotent equalities \eqref{r2=r} become
\[\rho_y(x)=q(\lambda_x(y))\qquad\text{and}\qquad q(\lambda_{\lambda_x(y)}(q(\lambda_x(y))))=q(\lambda_x(y)).\]
As $z=\lambda_x(y)$ may be an arbitrary element of $X$, the latter equality reduces to \[q(\lambda_z(q(z)))=q(z).\]
Note that
\begin{align*}
\lambda_x\lambda_y(z) & =x\cdot\varphi_x(y\cdot\varphi_y(z)),\\
\lambda_{\lambda_x(y)}\lambda_{\rho_y(x)}(z) & =\lambda_{\lambda_x(y)}\lambda_{q(\lambda_x(y))}(z)\\
& =x\cdot\varphi_x(y)\cdot\varphi_{x\cdot\varphi_x(y)}(q(x\cdot\varphi_x(y))\cdot\varphi_{q(x\cdot\varphi_x(y))}(z)).
\end{align*}
Since the semigroup $(X,\cdotp)$ is left cancellative and because elements of $\Lambda$ are left identities for $(X,\cdotp)$,
the Yang--Baxter equation \eqref{YB1} becomes
\[\varphi_x(y\cdot\varphi_y(z))=\varphi_x(y)\cdot\varphi_{x\cdot\varphi_x(y)}\varphi_{q(x\cdot\varphi_x(y))}(z).\]
Note that if all $\varphi_x =\varphi$ (for example if $\Lambda$ is a singleton then this equation becomes
$\varphi(y\cdot\varphi(z))=\varphi(y)\cdot\varphi(\varphi(z))$ and thus writing $t=\varphi(z)$ we obtain
$\varphi(y\cdot t)=\varphi(y)\cdot\varphi(t)$, i.e., $\varphi$ is a homomorphism of the semigroup $(X,\cdotp)$.
Similarly, taking into account that
\begin{align*}
\lambda_{\rho_{\lambda_y(z)}(x)}(\rho_z(y))
& =\lambda_{q(\lambda_x\lambda_y(z))}(q(\lambda_y(z)))\\
& =q(x\cdot\varphi_x(y\cdot\varphi_y(z)))\cdot\varphi_{q(x\cdot\varphi_x(y\cdot\varphi_y(z)))}(q(y\cdot\varphi_y(z))),\\
\rho_{\lambda_{\rho_y(x)}(z)}(\lambda_x(y)) & =q(\lambda_{\lambda_x(y)}\lambda_{\rho_y(x)}(z))\\
& =q(x\cdot\varphi_x(y)\cdot\varphi_{x\cdot\varphi_x(y)}(q(x\cdot\varphi_x(y))\cdot\varphi_{q(x\cdot\varphi_x(y))}(z))),
\end{align*}
one can check that the Yang--Baxter equation \eqref{YB2} simplifies to
\[\varphi_{q(x\cdot\varphi_x(y\cdot\varphi_y(z)))}(q(y\cdot\varphi_y(z)))=
q(x\cdot\varphi_x(y)\cdot\varphi_{x\cdot\varphi_x(y)}(q(x\cdot\varphi_x(y))\cdot\varphi_{q(x\cdot\varphi_x(y))}(z))).\]
The latter simplifies further because elements $\Lambda$ are left identities to
\[\varphi_{q(x\cdot\varphi_x(y\cdot\varphi_y(z)))}(q(y\cdot\varphi_y(z)))=
q(x\cdot\varphi_x(y)\cdot\varphi_{x\cdot\varphi_x(y)}(\varphi_{q(x\cdot\varphi_x(y))}(z))).\]
In case $\Lambda$ is a singleton then this equation is trivially satisfied, and it thus not give any extra information.
Finally, since
\begin{align*}
\rho_z(\rho_y(x))
& =q(\lambda_{\rho_y(x)}(z))\\
& =q(\lambda_{q(\lambda_x(y))}(z))\\
& =q(q(x\cdot\varphi_x(y))\cdot\varphi_{q(x\cdot \varphi_x (y))}(z)),\\
\rho_{\rho_z(y)}(\rho_{\lambda_y(z)}(x))
& =q(\lambda_{\rho_{\lambda_y(z)}(x)}(\rho_z(y)))\\
& =q(\lambda_{q(\lambda_x\lambda_y(z))}(q(\lambda_y(z))))\\
& =q(q(\lambda_x\lambda_y(z))\cdot\varphi_{q(\lambda_x\lambda_y(z))}(q(\lambda_y(z))))\\
& =q(q(x\cdot\varphi_x(y\cdot\varphi_y(z)))\cdot\varphi_{x\cdot\varphi_x(y\cdot\varphi_y(z))}(q(y\cdot\varphi_y(z)))),
\end{align*}
the Yang--Baxter equation \eqref{YB3} is equivalent to
\[q(\varphi_{q(x\cdot\varphi_x(y))}(z))= q(\varphi_{x\cdot\varphi_x(y\cdot\varphi_y(z))}(q(y\cdot\varphi_y(z)))).\]
Also this equation is trivially satisfied if $\Lambda$ is a singleton.
Hence we have shown the following theorem.
\begin{thm}\label{solgeneral}
Assume $(X,r)$ is a finite left non-degenerate idempotent solution of the Yang--Baxter equation.
Then the following properties are satisfied:
\begin{enumerate}
\item there exists a binary operation $\cdot$ on $X$ so that $(X,\cdotp)$ is a left cancellative simple semigroup.
\item there exits a surjective mapping $q\colon X\to\Lambda$, where $\Lambda$ is the set of idempotents of $(X,\cdotp)$.
\item for each $x\in X$ there exists a permutation $\varphi_x\in\Sym(X)$ so that the following identities hold:
\begin{align}
\varphi_x(y\cdot\varphi_y(z)) & =\varphi_x(y)\cdot\varphi_{x\cdot\varphi_x(y)}\varphi_{q(x\cdot\varphi_x(y))}(z),\label{fineq1}\\
\varphi_{q(x\cdot\varphi_x(y\cdot\varphi_y(z)))}(q(y\cdot\varphi_y(z)))
& =q(x\cdot\varphi_x(y)\cdot\varphi_{x\cdot\varphi_x(y)}(\varphi_{q(x\cdot\varphi_x(y))}(z))),\label{fineq2}\\
q(\varphi_{q(x\cdot\varphi_x(y))}(z)) & =q(\varphi_{x\cdot\varphi_x(y\cdot\varphi_y(z))}(q(y\cdot\varphi_y(z)))),\label{fineq3}\\
q(x\cdot\varphi_x(q(x))) & =q(x).\label{fineq4}
\end{align}
\item the solution $r$ has the form
\begin{equation}
r(x,y)=(x\cdot\varphi_x(y),q(x\cdot\varphi_x(y))).\label{sol-form}
\end{equation}
\end{enumerate}
Conversely, if $(X,\cdotp)$ is a finite left cancellative simple semigroup with the set of idempotents $\Lambda$,
and there exist a surjective mapping $q\colon X\to\Lambda$ and permutations $\varphi_x\in\Sym(X)$, for $x\in X$,
satisfying conditions \eqref{fineq1}--\eqref{fineq4}, then the formula \eqref{sol-form} yields a finite left
non-degenerate idempotent solution $(X,r)$ of the Yang--Baxter equation.
\begin{proof}
The necessity has been proven before the statement. The proof of sufficiency can be obtained by reversing
the order in argumentation. Namely, define $\lambda_x(x)=x\cdot\varphi_x(y)$ and $\rho_y(x)=q(x\cdot\varphi_x(y))$
for $x,y\in X$. Because the maps $q$ and $\varphi_x$ for $x\in X$ satisfy \eqref{fineq1}, \eqref{fineq2} and
\eqref{fineq3}, it follows that the equalities \eqref{YB1}, \eqref{YB2} and \eqref{YB3} hold. Hence $(X,r)$
is a solution of the Yang--Baxter equation of the form \eqref{sol-form}. One also easily verifies that the equality
\eqref{fineq4} translates to $r^2=r$. Finally, as the semigroup $X$ is left cancellative and $\varphi_x$ is bijective,
the map $\lambda_x$ is injective, and thus bijective as $X$ is finite. Therefore, the proof is complete.
\end{proof}
\end{thm}
In the next main result we show that $\Lambda$ is a singleton, i.e., $S(X,r)$ is contained in a group, precisely when
the structure algebra $K[M(X,r)$ is right Noetherian. So also for idempotent left non-degenerate solutions the structure
algebra determines crucial information of solutions. First we show that structure algebras of finite left non-degenerate
idempotent solutions are always left Noetherian.
\begin{prop}\label{stralg}
Assume $(X,r)$ is a finite left non-degenerate idempotent solution of the Yang--Baxter equation. If $M=M(X,r)$ and $K$
is a field then $K[M]$ is a finite module over a $K$-subalgebra that is isomorphic to a polynomial algebra in one variable.
In particular, $K[M]$ is a left Noetherian representable algebra (and hence satisfies a polynomial identity)
of Gelfand--Kirillov dimension one.
\begin{proof}
Fix, arbitrarily, an element $t\in X$. We claim that each element $w\in S(X,r)$ may be written (uniquely) as
$w=(t,\lambda_t)^{n-1}\circ(x,\lambda_x)$ for some $x\in X$ and some positive integer $n$. We shall prove
this by induction on $n=|w|$, the length of $w$. If $n=1$ then this is obvious. So, let $n>1$
and assume the result holds for elements of $S(X,r)$ of length $<n$. Write $w=v\circ(z,\lambda_z)$
for some $v\in S(X,r)$ of length $n-1$ and some $z\in X$. Then, by the induction hypothesis,
$v=(t,\lambda_t)^{n-2}\circ(y,\lambda_y)$ for some $y\in X$. Let $x=\lambda_t^{-1}(\lambda_y(z))\in X$.
Then, taking into account that $t+\lambda_y(z)=y+\lambda_y(z)$ in $A(X,r)$, we obtain
\begin{align*}
(t,\lambda_t)\circ(x,\lambda_x) & =(t+\lambda_t(x),\lambda_{t+\lambda_t(x)})=(t+\lambda_y(z),\lambda_{t+\lambda_y(z)})\\
& =(y+\lambda_y(z),\lambda_{y+\lambda_y(z)})=(y,\lambda_y)\circ(z,\lambda_z).
\end{align*}
Therefore,
\begin{align*}
w & =v\circ(z,\lambda_z)=(t,\lambda_t)^{n-2}\circ(y,\lambda_y)\circ(z,\lambda_z)\\
& =(t,\lambda_t)^{n-2}\circ(t,\lambda_t)\circ(x,\lambda_x)=(t,\lambda_t)^{n-1}\circ(x,\lambda_x),
\end{align*}
as claimed. In particular, \[K[M]=R+\sum_{x\in X}R\circ(x,\lambda_x)\] is a finite left $R$-module over the polynomial
algebra in one variable $R=K[(t,\lambda_t)]$. Thus, clearly, $K[M]$ is a left Noetherian algebra of Gelfand--Kirillov
dimension equal to one (see, e.g., \cite{MR}), while the representability of $K[M]$ is a consequence of \cite{Ana1989}.
\end{proof}
\end{prop}
{It is worth mentioning that as a consequence of the previous result and a result of Anan'in \cite{Ana1989} the structure
monoid $M(X,r)$ of a finite left non-degenerate idempotent solution $(X,r)$ is a linear monoid, i.e., a submonoid of the
multiplicative semigroup of a matrix ring over a field, and thus
\cite[Proposition 1, p.~221, Proposition 7, p.~280--281, and Theorem 14, p.~284]{Okn1991} yields
\[\clK K[M(X,r)]=\GK K[M(X,r)]=\rk M(X,r)=1,\] where $\rk M(X,r)$ denotes the rank of the semigroup
$M(X,r)$, i.e., the rank of the largest free commutative subsemigroup of $M(X,r)$.
On the other hand, the structure algebras of left non-degenerate finite idempotent solutions also are
right Noetherian if and only if the structure monoid is cancellative.
\begin{thm} \label{algebrastructure}
Assume $(X,r)$ is a finite left non-degenerate idempotent solution of the Yang--Baxter equation.
If $M=M(X,r)$ and $K$ is a field, then the following conditions are equivalent:
\begin{enumerate}
\item $|\Lambda|=1$.
\item the monoid $M$ is cancellative.
\item there exists an operation $\cdot$ on $X$ so that $(X,\cdotp)$ is a group and there exists
a homomorphism $\varphi$ on this group so that $\lambda_x(y)=x\cdot\varphi(y)$.
\item the algebra $K[M]$ is right Noetherian.
\item the algebra $K[M]$ is not central (that is, its center is strictly bigger than $K$).
\end{enumerate}
Furthermore, if $\ch K$ does not divide $|X|$ these conditions are equivalent with
\begin{enumerate}
\item[(6)] the algebra $K[M]$ is semiprime.
\end{enumerate}
For arbitrary fields $K$, $(6)$ implies $(1)$. Moreover, if any of these conditions is satisfied
then $K[M]$ is a subalgebra of $K[G_u]\cong K[T(G_u)][t^{\pm 1};\sigma]$, the skew Laurent polynomial
algebra in the variable $t$ over the group algebra $K[T(G_u)]$ of the finite group $T(G_u)$.
\begin{proof}
The equivalence of (1) and (2) follows from Lemma~\ref{quotientgroup}.
Hence, from Theorem~\ref{solgeneral}, (2) is equivalent with $X$ having a group structure for some binary
operation~$\cdot$ so that there exists a permutation $\varphi\in\Sym(X)$ that satisfies the equations
\eqref{fineq1}--\eqref{fineq4} (note that all $\varphi_x=\varphi$). It is readily verified that these
equations reduce to $\varphi$ being an automorphism of the group $(X,\cdotp)$. This shows (3) holds.
Conversely, if (3) holds then $\lambda_{x}^{-1}(x)= \varphi^{-1}(x^{-1} \cdot x)=1$, the identity of
the group $(X,\cdotp)$. Hence, $|\Lambda |=1$ and thus (1) holds.
Assume (1) holds, say $\Lambda =\{ u\}$. If $S=S(X,r)$ then Lemma~\ref{quotientgroup} yields that
$S=S_u= \free{c_u}\circ F_u$, and $F_u$ is a finite set. Furthermore, since $\Lambda $ is a singleton,
the element $c_u$ is central in $S_u$. Hence $K[M]=K+K[S]=\sum_{f\in F_u}f\circ K[\free{c_u}]$,
a finite right (and left) module over a polynomial algebra in one variable. Hence, $K[M]$ is right Noetherian.
So (4) holds.
Assume (4) holds. We need to prove that $\Lambda$ is a singleton. We prove this by contradiction.
So suppose $|\Lambda|>1$, i.e., there exist distinct $u,v\in\Lambda$ and $s_x\coloneqq(x,\lambda_x)\in S_u$
and $s_y\coloneqq(y,\lambda_y)\in S_v$. Then, by Proposition~\ref{decomposition}, $s_x^d \circ m =s_y^d \circ m$
for any $1\ne m\in M$. Consequently, $(s_x^d-s_y^d)\circ K[M]=(s_x^d-s_y^d)K$, and
\[I\coloneqq K[M]\circ(s_x^d-s_y^d)\circ K[M]=K[M]\circ(s_x^d-s_y^d).\]
We now show that the principal right ideal $I$ is not a finite right $K[M]$-module. Indeed, it were
$I=\sum_{i=1}^nf_i\circ K[M]$ for some $f_1,\dotsc,f_n\in I$ then $I=\sum_{i=1}^nf_iK$ would be a finite-dimensional
ideal. But the infinite set $\{(s_x^d)^j-(s_y^d)^j:j\ge 1\}$ contained in $I$ is linearly independent over $K$.
This proves that indeed $\Lambda$ is a singleton.
Also notice that if $\Lambda$ is a singleton then $c_u\in Z(K[M])$ and thus $Z(K[M])\ne K$ and (5) holds.
To prove the converse, we need to show that $|\Lambda|>1$ implies $Z(K[M])=K$. Since $Z(K[M])$ is $\mathbb{N}$-graded,
it is enough to show that there are no non-zero central homogeneous elements of $K[M]$ of positive degree.
Suppose, on the contrary, that $0\ne\alpha=\sum_{x\in X}\alpha_x(nx,\lambda_{nx})\in Z(K[M])$ is a homogeneous
element of degree $n>0$. Let $y,z\in X$ be such that $(y,\lambda_y)\in S_u$ and $(z,\lambda_z)\in S_v$ with $u\ne v$.
Then $(s_y^d-s_z^d)\circ\alpha=0$ implies that $\alpha\circ(s_y^d-s_z^d)=0$. Hence $\beta=\alpha\circ s_y^d=\alpha\circ s_z^d$.
Moreover, \[\beta=\alpha\circ s_y^d=s_y^d\circ\alpha=\sum_{x\in X}\alpha_x(du,\id)\circ(nx,\lambda_{nx})
=\sum_{x\in X}\alpha_x((n+d)x,\lambda_{(n+d)x}).\]
So also $\beta\ne 0$. Define $\mathcal{H}=\{\lambda_{nx}:x\in X\}\subseteq\mathcal{G}(X,r)$ and $X_\sigma=\{x\in X:\lambda_{nx}=\sigma\}$
for $\sigma\in \mathcal{H}$. Then
\begin{align*}
\beta=\alpha\circ s_y^d & =\sum_{x\in X}\alpha_x(nx,\lambda_{nx})\circ(du,\id)\\
& =\sum_{x\in X}\alpha_x (\lambda_{nx} ((n+d)u),\lambda_{nx})\\
& =\sum_{\sigma\in \mathcal{H}}\Big(\sum_{x\in X_\sigma}\alpha_x\Big)(\sigma((n+d)u),\sigma).
\end{align*}
Thus $\beta\ne 0$ yields that $(\sigma( (n+d)u,\sigma)$ lies in the support of $\beta$ for some $\sigma\in\mathcal{H}$,
i.e., $0\ne \sum_{x\in X_\sigma}\alpha_x$. Because also
\[\beta=\alpha\circ s_z^d=\sum_{\tau\in\mathcal{H}}\Big(\sum_{x\in X_\tau}\alpha_x\Big)(\tau((n+d)v),\tau),\]
we get $(\sigma((n+d)u),\sigma)=(\tau((n+d)v),\tau)$ for some $\tau\in \mathcal{H}$. Therefore, we have $\sigma=\tau$
and $\sigma((n+d)u)=\tau ((n+d)v)$, which leads to $\sigma(u)=\sigma(v)$, and consequently to $u=v$, a contradiction.
This finishes the proof of (5) implies (1).
To prove (6), assume $\ch K$ does not divide $|X|$. If (1) holds then $S=S_u$ is cancellative and hence so is $M$.
By Lemma~\ref{quotientgroup}, $S$, and thus $M$ has a group of quotients $G_u\cong T(G_u)\rtimes_{\sigma_u} \mathbb{Z}$,
that is a central localisation with respect the central monoid $\free{c_u}$, which is also a regular subset. It is
well-known that $K[G_u]$ is a semiprime algebra (see, e.g., \cite[Theorem 4.2.13]{MR798076}) because $\ch K$ does
not divide $|T(G_u)|=|X|$. Hence also $K[M]$ is semiprime. For the converse, it follows from the proof of the
equivalence of (4) and (1) that if $|\Lambda|\ne 1$ that the ideal generated by the elements $s_x^{d}-s_y^{d}$
is a nilpotent ideal of nilpotency class $2$. So the algebra $K[M]$ is not semiprime.
The last part of the result is an immediate consequence of Lemma~\ref{quotientgroup}.
\end{proof}
\end{thm}
Now, we give two remarks about the algebraic properties of the structure algebra determined
by a finite left non-degenerate idempotent solution.
\begin{rem}
The structure algebra $K[M(X,r)]$ of a finite left non-degenerate idempotent solution $(X,r)$
is prime only if $|X|=1$. Indeed if $K[M(X,r)]$ is prime then Theorem~\ref{algebrastructure}
implies $|\Lambda|=1$. Hence, by Theorem~\ref{algebrastructure} and because $K[G_u]$ is a central
localisation of $K[S_u]$, it follows that $K[M(X,r)]$ is prime if and only if $K[G_u]$ is prime.
The latter only holds if $|G_u|=1$. So indeed $|X|=1$.
\end{rem}
\begin{rem}
Since each $K$-algebra which is a finite module over a commutative finitely generated $K$-subalgebra
is automaton (see \cite{COAut}), it follows by Proposition~\ref{stralg} that $K[M(X,r)]$ is an automaton
algebra in case $(X,r)$ is a finite left non-degenerate idempotent solution. It also is known
(see, e.g., \cite{Ufna}) that each finitely generated $K$-algebra admitting a finite Gr\"obner basis
is automaton. Because there are numerous results indicating that computational properties of such algebras
as well as several of their algebraic and structural properties behave better than arbitrary finitely
generated (or even finitely presented) algebras (see, e.g., \cite{MR2497577,Okn1991,MR2450723,Ufna,BBL}),
it would be of interest to decide whether $K[M(X,r)]$ always admit a finite Gr\"obner basis.
\end{rem}
In case all $\lambda_x$ are equal we show in the following example that $K[M(X,r)]$ indeed has a finite Gr\"obner basis.
\begin{ex}
Let $(X,r)$ be a finite idempotent solution defined as $r(x,y)=(\lambda(y),y)$, where $\lambda\colon X\to X$
is an arbitrary map. (Note that if $(X,r)$ is an arbitrary finite left non-degenerate idempotent solution,
written as $r(x,y)=(\lambda_x(y),\rho_y(x))$, such that $\lambda_x=\lambda_y$ for all $x,y\in X$ (call this map $\lambda$)
then $(X,r)$ is of the above form.) Let us choose (arbitrarily) a well ordering $\le$ on $X$ (e.g., $x_1<\dotsb<x_n$ provided
$X=\{x_1,\dotsc,x_n\}$) and extend it to the degree-lexicographic ordering on the free monoid $X^*$ freely generated by $X$.
As \[M(X,r)=\free{X\mid xy=\lambda (y)y,\; x,y\in X}=\free{X\mid xy=yy,\; x,y\in X},\]
we get that $K[M(X,r)]\cong K[X^*]/I$, where the ideal $I$ of $K[X^*]$ is generated by the set $G$
consisting of all binomials of the form:
\begin{enumerate}
\item $yx-xx$ for $x,y\in X$ with $x<y$,
\item $yy-xy$ for $x,y\in X$ with $x<y$,
\item $yz-xz$ for $x,y,z\in X$ with $x<y$.
\end{enumerate}
Moreover, one may check that $G$ is a Gr\"obner basis of the ideal $I$ (this can be done easily by using
Bergman's Diamond Lemma \cite{MR506890}). In particular, the algebra $K[M(X,r)]$ admits a finite Gr\"obner basis.
\end{ex}
As an application, we obtain one of the main results, Theorem 3.10, proven in \cite{StVo21}.
Recall that a left non-degenerate solution $(X,r)$ is said to be latin if for each $x\in X$ the map
$X\to X\colon y\mapsto \lambda_y (x)$ is bijective. In the terminology of the authors this is formulated
as $(X,*)$ is a quasigroup, that is the operation $*$ on $X$ defined by $x*y=\lambda_x (y)$ defines a left
and right quasigroup, i.e., for each $x\in X$, left and right multiplications by $x$ are bijective.
\begin{cor}\label{idempotentlatin}
Assume $(X,r)$ is a finite left non-degenerate idempotent solution of the Yang--Baxter equation.
Then the following conditions are equivalent:
\begin{enumerate}
\item $(X,r)$ is a latin solution.
\item $X$ can be equipped with a binary operation $\cdot$ so that $(X,\cdotp)$ is a group
and there exists $\varphi\in \Aut(X,\cdotp)$ so that $\lambda_x (y) =x\cdot\varphi (y)$.
\end{enumerate}
\begin{proof}
Assume $(X,r)$ is a latin solution. Let $S=S(X,r)$ and fix $y\in X$. Because of Proposition~\ref{decomposition},
let $u\in\Lambda$ be so that $(y,\lambda_y)\in S_u$. Also in $S$ we have that
\[(x,\lambda_x)\circ(y,\lambda_y)=(\lambda_x(y),\lambda_{\lambda_x(y)})\circ(q(\lambda_x(y)),\lambda_{q(\lambda_x(y))})\]
and thus $(q(\lambda_x(y)),\lambda_{q(\lambda_x(y))})\in S_u$ for all $x\in X$. Because of the assumption this yields
$(q(z),\lambda_{q(z)})\in S_u$ for all $z\in X$. Hence $q^2$ is a constant map and thus, by Lemma \ref{qq}, it follows
that the map $q$ is constant. As $\Img(q)=\Lambda$, we get that $\Lambda$ is a singleton. By Theorem~\ref{algebrastructure}
it follows that (2) holds.
The converse is obvious.
\end{proof}
\end{cor}
In Theorem~\ref{algebrastructure} we dealt with idempotent left non-degenerate solutions $(X,r)$ with
$q\colon X \to \Lambda$ a constant map. Next we consider the case that the mapping $q$ is bijective, i.e., $X=\Lambda$.
\begin{cor}\label{qbijective}
Assume $(X,r)$ is a finite left non-degenerate idempotent solution of the Yang--Baxter equation.
Then $\Lambda=X$ if and only if $r(x,y)=(\varphi(y),y)$ for some permutation $\varphi\in\Sym(X)$.
Such a solution we will simply denote as $(X,r_{\varphi})$.
In particular, the number of non-isomorphic finite left non-degenerate idempotent solutions on a set
$X$ with bijective diagonal map equals the number of partitions of $|X|$.
\begin{proof}
Because of the assumption and Theorem~\ref{solgeneral}, $X$ can be equipped with a left simple band
structure $(X,\cdotp)$ and there exist bijective maps $q\colon X\to X$ and $\varphi_x\colon X \to X$,
for $x\in X$, satisfying \eqref{fineq1}, \eqref{fineq2}, \eqref{fineq3} and \eqref{fineq4}.
Because $q$ is bijective and all elements of $X$ are left identities these equations become:
\begin{align*}
\varphi_x \varphi_y (z)&= \varphi_{\varphi_x(y)} \varphi_{q\varphi_{x}(y)}(z),\\
\varphi_{q\varphi_x \varphi_y (z)}q\varphi_{y}(z)&=q\varphi_{\varphi_{x}(y)} \varphi_{q\varphi_x(y)}(z),\\
\varphi_{q\varphi_{x}(y)}(z)&= \varphi_{\varphi_x \varphi_y (z)} q\varphi_y(z),\\
q(x) &= \varphi_x^{-1}(x).
\end{align*}
Put $t\coloneqq \varphi_x \varphi_y (z)$. Then the second equation becomes $\varphi_{q(t)}q\varphi_x^{-1}(t)=q(t)$
or equivalently $q\varphi_x^{-1}(t) =\varphi_{q(t)}^{-1} q(t)$. It follows that all $\varphi_x$ are equal; let us
denote this map simply as $\varphi$. The equations then simply say that $q=\varphi^{-1}$ and thus
$x\cdot\varphi_x(y)=\varphi(y)$ and $q(x\cdot\varphi_x(y))=q(\varphi(y))=y$. Hence indeed $r(x,y)=(\varphi(y),y)$
for $x,y\in X$. This proves the necessity of the result. The converse is obvious.
For the last part of the statement it is sufficient to notice that solutions $(X,r_{\varphi})$ and $(X,r_{\psi})$
are isomorphic if and only if the permutations $\varphi$ and $\psi$ are conjugate.
\end{proof}
\end{cor}
As an application we obtain another main result, Theorem 4.7, proven in \cite{StVo21}.
\begin{cor}\label{primecase}
The finite left non-degenerate idempotent solutions of the Yang--Baxter equation of prime cardinality $p$
are the solutions of the following type:
\begin{enumerate}
\item $(X,r)$ with $X$ a set of cardinality $p$, $\varphi\in\Sym(X)$ and $r(x,y)=(\varphi(y),y)$.
\item $(X,r)$ with $X$ having a group structure $(X,\cdotp)\cong(\mathbb{Z}_p,+)$, $\varphi\in\Aut(X,\cdotp)$
and $r(x,y)=(x\cdot\varphi(y),1)$, where $1$ is the identity of the group $(X,\cdotp)$.
\end{enumerate}
\begin{proof}
Let $(X,r)$ be a finite left non-degenerate idempotent solution of the Yang--Baxter equation such that $|X|=p$.
Then Lemma~\ref{infotorsioncover} yields that $|\Lambda|=1$ or $|\Lambda|=p$. The latter case is dealt with
in Corollary~\ref{qbijective} and hence $(X,r)=(X,r_{\varphi})$ for some $\varphi\in\Sym(X)$. The former case
is dealt with in Theorem~\ref{algebrastructure} and thus one has a group structure on $X$. Because $p$ is prime,
we have that $(X,\cdotp) \cong (\mathbb{Z}_p,+)$, $\lambda_{x}(y)=x\cdot\varphi(y)$ and $r(x,y)=(x\cdot\varphi(y),1)$,
with $\varphi\in\Aut(X,\cdotp)$.
\end{proof}
\end{cor}
We now give an example of a finite left non-degenerate idempotent solution $(X,r)$ that is not covered by
Corollary~\ref{idempotentlatin} and Corollary~\ref{qbijective}, i.e., $\Lambda \ne X$ and $|\Lambda|>1$.
For this it is convenient to determine when all maps $\varphi_x$ in Theorem~\ref{solgeneral} are equal.
Note that this is the case if $|\Lambda|=1$.
\begin{cor}\label{allphiequal}
Assume $(X,r)$ is a finite left non-degenerate idempotent solution of the Yang--Baxter equation
as in Theorem~\ref{solgeneral}. If all $\varphi_x$ are equal, and we denote this map by $\varphi$,
then the conditions \eqref{fineq1}, \eqref{fineq2}, \eqref{fineq3} and \eqref{fineq4}
are equivalent, respectively, with:
\begin{enumerate}
\item $\varphi\in\Aut(X,\cdotp)$.
\item $\varphi q=q^2$.
\item $q=q^4$.
\item $q(x\cdot q^{2}(x)) =q(x)$ for all $x\in X$.
\end{enumerate}
\begin{proof}
That $\varphi$ is an automorphism of $(X,\cdotp)$ is equivalent with the identity \eqref{fineq1}.
The identity \eqref{fineq2} is equivalent with $\varphi q(y\cdot \varphi(z))=q(x\cdot\varphi(y)\cdot\varphi\varphi(z))$
for all $x,y,z\in X$. Equivalently, because $\varphi$ is an automorphism,
$\varphi q(y\cdot t)=q(x\cdot\varphi(y)\cdot\varphi(t))=q(x\cdot\varphi(y\cdot t))$. Taking into account that
the mapping $t\mapsto y\cdot t$ is bijective, $\varphi q(z)=q(x\cdot z)$ for all $x,z\in X$. If $z\in S_u$,
that is $q(z)=u\in\Lambda$, then, because $(X,\cdotp)$ is a left cancellative simple semigroup, there exists
$x\in X$ so that $x\cdot z =q(z)$ (note that $q(z)$ is the identity elements of $G_u$). Hence, $\varphi q (z) =qq(z)$.
This proves part (2).
The identity \eqref{fineq3} is equivalent with $q\varphi(z)=q\varphi q(y\cdot\varphi(z))$, for all $y,z\in X$.
Because $\varphi (z) \in G_{q(\varphi (z))}$, this is equivalent with $q\varphi (z)= q \varphi q q(\varphi (z))$,
i.e., $q =q \varphi qq$. So by part (2), this is the same as $q= q^4$.
Finally, the identity \eqref{fineq4} is equivalent with $q(x\cdot \varphi q(x)) =q(x)$, and thus by (2),
this is the same as $q(x\cdot q^{2}(x)) =q(x)$.
\end{proof}
\end{cor}
\begin{ex}\label{examplespecial}
Let $G$ be a finite group, with identity $1$, and $X=\mathcal{M}(G,1,n,J)$, a finite left cancellative
simple semigroup, with $n$ a positive even integer. Denote the elements of $X$ as $(g,i)$, with
$g\in G$ and $1\le i\le n$. Let $f$ be an automorphism of the group $G$. Let $A\cup B$ be a partition
of $\{1,\dotsc,n\}$ with $|A|=|B|$ and let $t\colon B \to A$ be a bijection. Define $\theta\colon A\cup B\to A\cup B$
by $\theta(a)=a$ for $a\in A$ and $\theta(b)=t(b)$ for $b\in B$. So $\theta^2=\theta$. Define $q\colon X \to X$ as
$q(g,i)=(1,\theta(i))$ and $\varphi(g,i)=(f(g),\psi(i))$, where $\psi\in\Sym(A\cup B)$ is such that $\psi|_A=\id_A$.
Then all conditions listed in Corollary~\ref{allphiequal} are satisfied and hence it yields a solution with
$\Lambda=\Img(q)\ne X$ and $\Lambda$ is not a singleton.
\end{ex}
\section*{Acknowledgement}
The first named author is supported by the Engineering and Physical Sciences Research Council [grant number EP/V005995/1].
The third author is supported by National Science Centre grant 2020/39/D/ST1/01852 (Poland).
The fourth author is supported by Fonds voor Wetenschappelijk Onderzoek (Flanders), via an FWO post-doctoral fellowship, grant 12ZG221N.
The fifth author is supported by Fonds voor Wetenschappelijk Onderzoek (Flanders), via an FWO Aspirant-fellowship, grant 11C9421N.
\bibliographystyle{amsplain}
|
\section{Introduction}
\label{sec_intro}
One of the fundamental observables both in classical and quantum field theory is the stress tensor (or stress-energy or energy-momentum tensor) $T_{\mu\nu}$. It enters the Einstein equations in General Relativity, linking the distribution of matter to the curvature of spacetime. In this context, the stress tensor should fulfill certain positivity conditions, such as the \emph{weak energy condition}, the \emph{strong energy condition}, or the \emph{null energy condition}, which imply constraints on exotic spacetime geometries. Assuming such conditions, various important results were derived such as the Penrose and Hawking singularity theorems~\cite{hawkingpenrose1970}, the Schoen--Yau positive mass theorem~\cite{schoenyau1979,schoenyau1981}, and Hawking's chronology protection results~\cite{hawking1992}.
However, in quantum field theory (QFT) these energy conditions are violated, and moreover one can find states in which the energy density at any single point can have arbitrarily negative expectation values~\cite{epsteinglaserjaffe1965,fewster2006}. Therefore, one might ask whether quantum fields are compatible with the energy conditions in the singularity theorems, or whether they allow for exotic spacetime geometries, such as wormholes. Moreover, unbounded negative energy densities have physically undesirable consequences regarding the validity of the second law of thermodynamics at macroscopic scales. To avoid these, there must be constraints on how negative energy densities of quantum fields can be. These constraints are called \emph{Quantum Energy Inequalities} (QEIs), and were first investigated by Ford~\cite{ford1991}. They are reminiscent of the classical energy conditions, suggesting that singularity or other theorems might still be obeyed by quantum fields, possibly in some weakened condition~\cite{brownfewsterkontou2018,fewsterkontou2020,freivogelkontoukrommydas2022,fewsterkontou2022}.
Technically, these bounds are of the following type: consider the energy density operator $T^{00}(t,\vec{x})$ averaged in time with the square of a real-valued smooth function:
\begin{equation}
T^{00}(f^2) \equiv \int T^{00}(t,\vec{x}) f^2(t) \mathop{}\!\mathrm{d} t \eqend{.}
\end{equation}
A QEI is an inequality of the form
\begin{equation}
\label{eq:masterqei}
\omega\left( T^{00}(f^2) \right) \geq - K_\omega(f)
\end{equation}
for all sufficiently regular quantum states $\omega$ of the system, where the constant $K_\omega$ depends on the smearing function $f$ and possibly on the state $\omega$. The QEI is called \emph{state-independent} or \emph{absolute} if $K$ is actually independent of $\omega$ and only depends on $f$. However, in many cases the dependence on the state is such that $K_\omega$ only depends mildly on the energy content of $\omega$~\cite{fewsterosterbrink2008}.
State-independent QEIs in the form~\eqref{eq:masterqei} have been established for \emph{free} quantum field theories in flat and curved spacetimes in a wide range of physical situations, including the free scalar, Dirac fermion, vector, and Rarita-Schwinger field, see the reviews~\cite{fewster2012,kontousanders2020} and references therein, as well as in conformal field theories (CFTs) in $1+1$-dimensions~\cite{fewsterhollands2005}. Moreover, they have been established also for smearings along a generic time-like worldline of an observer. In a more general setting, it was shown that analogues of QEIs are fulfilled by observables that arise from the operator product expansions of classically positive expressions~\cite{bostelmannfewster2009}. However, these observables do not necessarily have a direct physical interpretation (such as the energy density of a system).
These inequalities are also related to the Averaged Null Energy Condition (ANEC)~\cite{flanaganwald1996}, which is formally obtained from a QEI in the limit when the averaging function $f$ tends to $1$ and the integration is over a null geodesic. There is recent interest in the ANEC due to its connection to holography and the quantum information carried by black hole horizons, see for example~\cite{fukoellermarolf2017,boussoetal2016a,boussoetal2016b}.
However, the analysis of QEIs in theories with \emph{self-interaction} (apart from the case of CFTs) is a long-standing problem due to the complicated structure that local quantum fields have in the presence of non-trivial interactions. This is unsatisfactory: the dependence of the energy density on self-interaction is clearly of physical importance, and its lower bounds are related to the stability of spacetime, as explained above. A first result on QEIs in interacting massive theories was obtained by the second author, Bostelmann and Fewster~\cite{bostelmanncadamurofewster2013}, who derived a state-independent QEI for the massive Ising model. This model belongs to the class of integrable quantum field theories on $1+1$-dimensional Minkowski spacetime, and describes scalar bosons whose two-particle scattering matrix is given by $S_2 = - \mathbbm{1}$. More generally, integrable models are characterised by a \emph{factorizing scattering marix}, where two-particle scattering processes determine the behaviour of particles completely. A proof of QEIs in a larger class of scalar integrable models (including the Sinh--Gordon model) was obtained in~\cite{bostelmanncadamuro2016}, on the level of expectation values of the energy density in one-particle states. Presently, the second author, Bostelmann and Mandrysch are also establishing QEIs in integrable models with momentum-independent two-particle scattering matrix in states of arbitrary particle number. Moreover, they are extending the analysis of QEIs at the one-particle level to integrable models with several particle species and with bound states. This includes several relevant examples, such as the Bullough-Dodd model, the $\mathrm{O}(N)$ nonlinear sigma model and the Federbush model. At the same time, the analysis extends to integrable models describing asymptotic fermions, first described in the non-perturbative algebraic quantum field theory (AQFT) framework by~\cite{bostelmanncadamuro2021}.
In this paper, we establish a state-independent or absolute QEI for the massless Sine--Gordon model in the finite regime of $\beta^2 < 4\pi$. For this, we use methods from the perturbative AQFT framework reviewed in~\cite{hollandswald2015,fewsterverch2015,fredenhagenrejzner2016}, in particular employing the Bogoliubov formula to define interacting field operators. To our knowledge, this is the first QEI that has been established in a self-interacting model without the use of any special symmetry (such as conformal symmetry) or the integrability of the model (like the Ising model as explained above). We build on our previous paper~\cite{froebcadamuro2022}, where we constructed the stress tensor $T_{\mu\nu}$ in the same model using an alternative approach based on the Gell--Mann-Low formula. There, we showed that the renormalised expectation value of the stress tensor is well-defined in the sense of distributions at each order in perturbation theory after removal of the infrared (IR) and ultraviolet (UV) cutoffs. Moreover, we proved that the renormalised perturbative series converges both for the Euclidean and the Lorentzian theory, for the latter in a large class of quasi-free Hadamard states. We also showed that the stress tensor is conserved after adding a quantum correction proportional to $\hbar$, yielding the expression
\begin{equation}
\label{eq:stresstensor}
\hat{T}_{\mu\nu} \equiv \mathcal{O}_{\mu\nu} - \frac{1}{2} \eta_{\mu\nu} \mathcal{O}_\rho{}^\rho + g \left( 1 - \frac{\beta^2}{8 \pi} \right) \eta_{\mu\nu} ( V_\beta + V_{-\beta} )
\end{equation}
with
\begin{equation}
\mathcal{O}_{\mu\nu} \equiv \partial_\mu \phi \, \partial_\nu \phi \eqend{.}
\end{equation}
Previously, various aspects of the Sine--Gordon model had also been studied in the pAQFT framework by Bahns, Fredenhagen, Pinamonti and Rejzner~\cite{bahnsrejzner2018,bahnsrejznerfredenhagen2021,bahnspinamontirejzner2021}.
The main results of the present work are as follows: As a first step, we show that the perturbative series for the renormalised interacting stress-energy tensor given by the Bogoliubov formula converges in expectation in an arbitrary quasi-free Hadamard state $\omega$. This requires almost the same conditions as in~\cite[Theorem~5]{froebcadamuro2022}, with only a slightly stronger assumption on the state-dependent part $W$ of the two-point function of $\omega$, namely the requirement of \emph{conditional positive semidefiniteness} explained in Remark~\ref{remark1} below.
\begin{theorem}[Renormalisation and convergence]
\label{thm_renorm}
Consider the massless Lorentzian Sine--Gordon model in the finite regime $\beta^2 < 4\pi$ and a quasi-free state $\omega^{\Lambda,\epsilon}$ in the vacuum sector whose two-point function has an IR cutoff $\Lambda$ and UV cutoff $\epsilon$. We furthermore require that the state-dependent part $W$ of the two-point function of the state $\omega$ satisfies:
\begin{enumerate}
\item $W(x,y)$ and its first and second derivatives grow at most polynomially,
\item $W(x,y)$ is conditionally positive semidefinite.
\end{enumerate}
Then there exists a redefinition of time-ordered products $\mathcal{T}\left[ \mathcal{O}_{\mu\nu}(x) \otimes V_\alpha(y) \right] \to \mathcal{T}\left[ \mathcal{O}_{\mu\nu}(x) \otimes V_\alpha(y) \right] + \delta \mathcal{T}\left[ \mathcal{O}_{\mu\nu}(x) \otimes V_\alpha(y) \right]$ with $\delta \mathcal{T}$ a local term proportional to $\delta^2(x-y) \mathcal{T}\left[ V_\alpha(y) \right]$, such that for any adiabatic cutoff function $g \in \mathcal{S}(\mathbb{R}^2)$ the interacting modified stress tensor~\eqref{eq:stresstensor} defined by the Bogoliubov formula
\begin{equation}
\label{eq:interacting_stresstensor}
\mathcal{T}_\text{int}\left[ \hat{T}_{\mu\nu}(z) \right] \equiv \mathcal{T}\left[ \mathrm{e}_\otimes^{\mathrm{i} S_\text{int}} \right]^{\star (-1)} \star \mathcal{T}\left[ \hat{T}_{\mu\nu}(z) \otimes \mathrm{e}_\otimes^{\mathrm{i} S_\text{int}} \right]
\end{equation}
has a finite expectation value in the state $\omega$ in the physical limit $\Lambda, \epsilon \to 0$:
\begin{equation}
\lim_{\Lambda,\epsilon \to 0} \omega^{\Lambda,\epsilon}\left( \mathcal{T}_\text{int}\left[ \hat{T}_{\mu\nu}(f) \right] \right) < \infty \eqend{,}
\end{equation}
where the limit is taken termwise and $f \in \mathcal{S}(\mathbb{R}^2)$. Moreover, the interacting modified stress tensor is conserved: we have
\begin{equation}
\label{eq:thm_renorm_cons_eq}
\lim_{\Lambda,\epsilon \to 0} \omega^{\Lambda,\epsilon}\left( \mathcal{T}_\text{int}\left[ \hat{T}_{\mu\nu}(\partial^\mu f) \right] \right) = 0
\end{equation}
for all $f \in \mathcal{S}(\mathbb{R}^2)$ such that $g$ is constant on the support of $f$.
\end{theorem}
\begin{remark}
\label{remark1}
The condition of conditional positive semidefiniteness of $W$ is
\begin{equation}
\label{eq:thm_renorm_wpos_cont}
W(f,f^*) = \iint W(x,y) f(x) f^*(y) \mathop{}\!\mathrm{d}^2 x \mathop{}\!\mathrm{d}^2 y \geq 0
\end{equation}
for a complex test function $f \in \mathcal{S}(\mathbb{R}^2)$ with vanishing mean $\int f(x) \mathop{}\!\mathrm{d}^2 x = 0$. This condition is clearly related to the positive definiteness of the state $\omega^{\Lambda,\epsilon}$ for derivatives of $\phi$, since
\begin{equation}
(\partial_\mu \phi)(f^\mu) = \int \partial_\mu \phi(x) f^\mu(x) \mathop{}\!\mathrm{d}^2 x = - \int \phi(x) \partial_\mu f^\mu(x) \mathop{}\!\mathrm{d}^2 x = - \phi(\partial_\mu f^\mu) \eqend{,}
\end{equation}
the test function $f(x) = \partial_\mu f^\mu(x)$ has vanishing mean, and positivity of the state in this case is
\begin{equation}
\label{eq:thm_renorm_pos_hplusw}
\omega^{\Lambda,\epsilon}\left( (\partial^\mu \phi)(f_\mu^*) \star (\partial^\nu \phi)(f_\nu) \right) = \mathrm{i} H^+\left( \partial^\mu f_\mu^*, \partial^\nu f_\nu \right) + W\left( \partial^\mu f_\mu^*, \partial^\nu f_\nu \right) > 0
\end{equation}
for $f_\mu \neq 0$, where $H^+$ is the Hadamard parametrix. Since $\mathrm{i} H^+$ is already positive definite for test functions with vanishing mean, the condition~\eqref{eq:thm_renorm_wpos_cont} is slightly stronger than would be strictly required, but seems to be necessary for the convergence of the perturbative series.
In our previous work~\cite{froebcadamuro2022}, we instead required the discrete condition
\begin{equation}
\label{eq:thm_renorm_wpos_disc}
\sum_{i,j=1}^n \Big[ W(x_i,x_j) - W(y_i,x_j) - W(x_i,y_j) + W(y_i,y_j) \Big] \geq 0
\end{equation}
for any configuration of points $x_i$ and $y_i$ and any $n \in \mathbb{N}$. However, this can be derived from~\eqref{eq:thm_renorm_wpos_cont} as follows: we take a sequence of test functions $\delta_N \in \mathcal{S}(\mathbb{R}^2)$ approximating the $\delta$ distribution as $N \to \infty$, and $f$ to be the sum
\begin{equation}
f(x) = \sum_{i=1}^n \Big[ \delta^2_N(x-x_i) - \delta^2_N(x-y_i) \Big] \eqend{.}
\end{equation}
Taking the limit $N \to \infty$ after integration, we obtain~\eqref{eq:thm_renorm_wpos_disc} from~\eqref{eq:thm_renorm_wpos_cont}; such approximations and the existence of the limit for smooth $W$ are well known.
For the example given in our previous work (a state that is thermal in a wide range of energies) we have~\cite[Eq.~(14)]{froebcadamuro2022}
\begin{splitequation}
\label{eq:thm_renorm_w_example}
W(x,y) &= \int \mathrm{e}^{\mathrm{i} p (x-y)} \Theta\left( \abs{p^0} \in [E_0, E_1] \right) \frac{\pi \mathrm{e}^{- \beta \abs{p^0}}}{\abs{p^0} \left( 1 - \mathrm{e}^{- \beta \abs{p^0}} \right)} \\
&\quad\times [ \delta(p^1+p^0) + \delta(p^1-p^0) ] \frac{\mathop{}\!\mathrm{d}^2 p}{(2\pi)^2} \eqend{,}
\end{splitequation}
and since this is positive semidefinite for all $f \in \mathcal{S}(\mathbb{R}^2)$ independently of their mean, it fulfills the condition~\eqref{eq:thm_renorm_wpos_cont}.
\end{remark}
Secondly, we show that the expectation value of the perturbative series for the renormalised interacting stress tensor in an arbitrary Hadamard state converges also when the smearing is only along a one-dimensional time-like worldline, and not in spacetime as in the previous theorem:
\begin{theorem}[Renormalisation of the quantum energy density]
\label{thm_renorm_qed}
We assume the same conditions as in Theorem~\ref{thm_renorm} and take a smearing function $f \in \mathcal{S}(\mathbb{R})$ supported on a smooth time-like curve $z^\mu(\tau)$, $\dot z^\mu(\tau) \dot z_\mu(\tau) = -1$, $\tau \in \mathbb{R}$. The quantum energy density
\begin{equation}
\label{eq:energy_density}
E_\omega(f) \equiv \lim_{\Lambda,\epsilon \to 0} \omega^{\Lambda,\epsilon}\left( \mathcal{T}_\text{int}\left[ \hat{T}_{\mu\nu}(F^{\mu\nu}) \right] \right)
\end{equation}
seen by an observer following the curve is finite, where
\begin{equation}
F^{\mu\nu}(z) \equiv \dot z^\mu(\tau) \dot z^\nu(\tau) f(\tau) \eqend{.}
\end{equation}
\end{theorem}
Finally, we show that the interacting energy density, as seen by an observer following a one-dimensional time-like worldline, satisfies a lower bound which is independent of the quantum state. Namely, we establish a state-independent QEI for the massless Sine--Gordon model:
\begin{theorem}[Quantum energy inequality]
\label{thm_qei}
We assume the same conditions as in Theorem~\ref{thm_renorm_qed}, but smear with the square of a test function $f \in \mathcal{S}(\mathbb{R})$. The smeared quantum energy density~\eqref{eq:energy_density} seen by an observer is then bounded from below independently of the quantum state $\omega$:
\begin{equation}
\label{eq:thm_qei}
E_\omega(f^2) \geq - K(z,f,\beta,g) \eqend{.}
\end{equation}
\end{theorem}
\begin{remark}
As we will see in the proof, there are three different contributions to the constant $K$ on the right-hand side of the inequality~\eqref{eq:thm_qei}, $K = K_0 + K_V + K_H$. The first one is the contribution from the free theory, which can be calculated exactly~\eqref{eq:qei_h_result}:
\begin{equation}
\label{eq:thm_qei_k0}
K_0 = \frac{1}{24 \pi} \int \left[ 6 \big[ f'(\tau) \big]^2 + f^2(\tau) \frac{[ \ddot z^1(\tau) ]^2}{1 + [ \dot z^1(\tau) ]^2} \right] \mathop{}\!\mathrm{d} \tau \eqend{.}
\end{equation}
It depends on the smearing function $f$ and the trajectory $z^\mu(\tau)$, but obviously not on $\beta$ nor $g$. It consists of two parts: the first one that arises for a straight trajectory with $z^\mu(\tau) = \delta^\mu_0 \tau$, $\ddot z^\mu = 0$ and coincides with old results~\cite{fewstereveson1998} in two-dimensional Minkowski space, and the second one which arises from the difference of the actual trajectory and a straight one, and consequently depends on the acceleration $\ddot z^\mu$.\footnote{We do not claim that this second term is new, but we haven't been able to find an explicit expression for it in the literature.} Moreover, $K_0$ stays finite in the limit where the trajectory becomes light-like $\dot z^1 \to \pm \infty$, at least if it is taken in such a way that $\ddot z^1/\dot z^1$ stays bounded. The second contribution $K_V$ is the one from the vertex operators $V_{\pm \beta}$ in the (quantum-corrected) stress tensor~\eqref{eq:stresstensor}, while the third one $K_H$ comes from derivatives of Hadamard parametrices that arise from the $\mathcal{O}_{\mu\nu}$ terms in the stress tensor~\eqref{eq:stresstensor}. Both of them clearly depend also on $\beta$ and $g$, but while $K_V$ is also finite in the light-like limit, the bounds that we derived to obtain $K_H$ diverge in this limit, see for example~\eqref{eq:mink_conv_ddf_bound} and the subsequent comments. The QEI~\eqref{eq:thm_qei} therefore only holds in this form for time-like trajectories $z^\mu(\tau)$, and we leave the derivation of a null QEI for the Sine--Gordon model to future work.
We do not believe our bound~\eqref{eq:thm_qei} to be optimal. Indeed, it is known~\cite{fewstereveson1998} that already the free-theory bound~\eqref{eq:thm_qei_k0} for a straight trajectory overestimates the optimal bound~\cite{flanagan1997} by a factor $3/2$. However, since we prove that the perturbative series converges, for sufficiently small coupling $g$ the contribution from interactions ($K_V$ and $K_H$) will be subdominant to the free-theory contribution from $K_0$. Unfortunately, through estimates such as~\eqref{eq:mink_conv_ddf_bound}, what is sufficiently small depends also on the trajectory $z^\mu(\tau)$ (except if it is straight).
\end{remark}
\section{Proof of Theorem~\ref{thm_renorm} (Renormalisation and convergence)}
\label{sec_renorm}
We again use the framework of perturbative algebraic quantum field theory, but refer the reader to~\cite{hollandswald2015,fewsterverch2015,fredenhagenrejzner2016} for reviews and our previous work~\cite{froebcadamuro2022} for a short introduction and all relevant formulas. However, instead of defining interacting expectation values using the Gell-Mann--Low formula, we define interacting time-ordered products $\mathcal{T}_\text{int}$ (and interacting operators) using the Bogoliubov formula, and then take their expectation value.
Using that
\begin{equation}
\label{eq:t_starinverse}
\mathcal{T}\left[ \mathrm{e}_\otimes^{\mathrm{i} S_\text{int}} \right]^{\star (-1)} = \overline{\mathcal{T}}\left[ \mathrm{e}_\otimes^{ - \mathrm{i} S_\text{int} } \right]
\end{equation}
with the anti-time-ordered products $\overline{\mathcal{T}}$, the Bogoliubov formula defines interacting time-ordered products by
\begin{equation}
\label{eq:tint_def}
\mathcal{T}_\text{int}\left[ \mathcal{O}_1(f_1) \cdots \mathcal{O}_k(f_k) \right] \equiv \overline{\mathcal{T}}\left[ \mathrm{e}_\otimes^{ - \mathrm{i} S_\text{int} } \right] \star \mathcal{T}\left[ \mathrm{e}_\otimes^{ \mathrm{i} S_\text{int} } \otimes \mathcal{O}_1(f_1) \otimes \cdots \otimes \mathcal{O}_k(f_k) \right] \eqend{.}
\end{equation}
Here the operators $\mathcal{O}_i$ are classical polynomials of the basic field $\phi$ and its derivatives or vertex operators $V_\alpha \equiv \mathrm{e}^{\mathrm{i} \alpha \phi}$, smeared with test functions $f_i \in \mathcal{S}(\mathbb{R}^2)$, and $S_\text{int}$ is the interaction with adiabatic cutoff. The time-ordered products $\mathcal{T}$ and anti-time-ordered products $\overline{\mathcal{T}}$ are constructed inductively, with the ones with single entries equal to the Hadamard-normal-ordered products
\begin{equation}
\label{eq:timeordered_single}
\mathcal{T}\left[ \mathcal{O}(x) \right] = \mathcal{N}_H\left[ \mathcal{O}(x) \right] = \overline{\mathcal{T}}\left[ \mathcal{O}(x) \right] \eqend{.}
\end{equation}
Outside the diagonal, the higher ones are defined by (anti-)causal factorisation:
\begin{splitequation}
\label{eq:timeordered_causal}
&\mathcal{T}\left[ \mathcal{O}_1(x_1) \otimes \cdots \otimes \mathcal{O}_n(x_n) \right] \\
&\quad= \mathcal{T}\left[ \mathcal{O}_1(x_1) \otimes \cdots \otimes \mathcal{O}_k(x_k) \right] \star \mathcal{T}\left[ \mathcal{O}_{k+1}(x_{k+1}) \otimes \cdots \otimes \mathcal{O}_n(x_n) \right]
\end{splitequation}
and
\begin{splitequation}
\label{eq:timeordered_anticausal}
&\overline{\mathcal{T}}\left[ \mathcal{O}_1(x_1) \otimes \cdots \otimes \mathcal{O}_n(x_n) \right] \\
&\quad= \overline{\mathcal{T}}\left[ \mathcal{O}_{k+1}(x_{k+1}) \otimes \cdots \otimes \mathcal{O}_n(x_n) \right] \star \overline{\mathcal{T}}\left[ \mathcal{O}_1(x_1) \otimes \cdots \otimes \mathcal{O}_k(x_k) \right]
\end{splitequation}
if none of the $x_1,\ldots,x_k$ lie in the past light cone of any of the $x_{k+1},\ldots,x_n$; the extension to the total diagonal $x_1 = \cdots = x_n$ corresponds to renormalisation. The interacting time-ordered products~\eqref{eq:tint_def} with a single entry define interacting operators in the quantum theory, and expanding the exponentials we obtain
\begin{equation}
\label{eq:opint_def}
\mathcal{T}_\text{int}\left[ \mathcal{O}(x) \right] = \sum_{n=0}^\infty \sum_{k=0}^n \frac{(-1)^k \mathrm{i}^n}{k! (n-k)!} \overline{\mathcal{T}}\left[ S_\text{int}^{\otimes k} \right] \star \mathcal{T}\left[ S_\text{int}^{\otimes (n-k)} \otimes \mathcal{O}(x) \right] \eqend{.}
\end{equation}
For the Sine--Gordon model, the interaction is given by
\begin{equation}
\label{eq:sint}
S_\text{int} = \int g(x) [ V_\beta(x) + V_{-\beta}(x) ] \mathop{}\!\mathrm{d}^2 x \eqend{,}
\end{equation}
where $g \in \mathcal{S}(\mathbb{R}^2)$ is the adiabatic cutoff and we consider the case where $\beta^2 < 4 \pi$. We are interested in the interacting stress tensor, and from our previous work~\cite{froebcadamuro2022} we know that the classical expression needs to be amended to obtain a conserved tensor in the quantum theory. The corrected stress tensor is given by
\begin{equation}
\label{eq:stresstensor_def}
\hat{T}_{\mu\nu} \equiv \mathcal{O}_{\mu\nu} - \frac{1}{2} \eta_{\mu\nu} \mathcal{O}_\rho{}^\rho + g \left( 1 - \frac{\beta^2}{8 \pi} \right) \eta_{\mu\nu} ( V_\beta + V_{-\beta} )
\end{equation}
with the composite operator
\begin{equation}
\label{eq:opmunu_def}
\mathcal{O}_{\mu\nu} \equiv \partial_\mu \phi \, \partial_\nu \phi \eqend{,}
\end{equation}
such that we need to determine $\mathcal{T}_\text{int}\left[ \mathcal{O}_{\mu\nu} \right]$ and $\mathcal{T}_\text{int}\left[ V_{\pm \beta} \right]$. Moreover, we also need $\mathcal{T}_\text{int}\left[ \partial_\mu \phi \right]$ to prove the quantum energy inequality, so we first have to derive explicit expressions for these three interacting operators:
\begin{lemma}
\label{lemma1}
The interacting operators $\mathcal{T}_\text{int}\left[ V_\alpha(z) \right]$, $\mathcal{T}_\text{int}\left[ \partial_\mu \phi(z) \right]$ and $\mathcal{T}_\text{int}\left[ \mathcal{O}_{\mu\nu}(z) \right]$ are given by the formal power series~\eqref{eq:tint_vertex_def}, \eqref{eq:tint_1phi_def} and \eqref{eq:tint_2phi_def}.
\end{lemma}
\begin{proof}
We recall from~\cite{froebcadamuro2022} that the two-point function $G$ of the quasi-free state $\omega$ is given by
\begin{equation}
\label{eq:twopf_hadamard}
G^+(x,y) \equiv - \mathrm{i} \omega^{\Lambda,\epsilon}\left( \phi(x) \star \phi(y) \right) = H^+(x,y) + \frac{\mathrm{i}}{2 \pi} \ln\left( \frac{\Lambda}{\mu} \right) - \mathrm{i} W(x,y)
\end{equation}
with the Hadamard parametrix
\begin{equation}
\label{eq:hadamard}
H^+(x,y) \equiv \frac{\mathrm{i}}{4 \pi} \ln\left[ \mu^2 ( \epsilon + \mathrm{i} u ) ( \epsilon + \mathrm{i} v ) \right]
\end{equation}
containing the singular part of the two-point function and the state-dependent part $W(x,y)$ which is a real smooth and symmetric bisolution of the massless Klein--Gordon equation $\partial_x^2 W(x,y) = \partial_y^2 W(x,y) = 0$. We introduced the light cone coordinates
\begin{equation}
\label{eq:lightconecoords}
u = u(x,y) \equiv (x^0-y^0) - (x^1-y^1) \eqend{,} \quad v = v(x,y) \equiv (x^0-y^0) + (x^1-y^1) \eqend{,}
\end{equation}
$\Lambda$ is an IR cutoff which we ultimately take to vanish, and we keep $\epsilon > 0$ as a UV cutoff. That is, the physical two-point function is obtained as the distributional boundary value (in the limit $\epsilon \to 0$) from the function~\eqref{eq:twopf_hadamard} which is analytic for all $\epsilon > 0$.
From the proof of~\cite[Lemma 2]{froebcadamuro2022}, we know that
\begin{splitequation}
\label{eq:timeordered_exponential}
\mathcal{T}\left[ \bigotimes_{j=1}^n V_{\alpha_j}(x_j) \right] &= \exp\left[ - \mathrm{i} \sum_{1 \leq i < j \leq n} \alpha_i \alpha_j H^\text{F}(x_i,x_j) \right] \mathcal{N}_G\left[ \prod_{j=1}^n V_{\alpha_j}(x_j) \right] \\
&\quad\times \exp\left[ - \frac{1}{2} \sum_{i,j=1}^n \alpha_i \alpha_j W(x_i,x_j) \right] \left( \frac{\Lambda}{\mu} \right)^\frac{\left( \sum_{k=1}^n \alpha_k \right)^2}{4 \pi}
\end{splitequation}
and
\begin{splitequation}
\label{eq:timeordered_exponential_2phi}
&\mathcal{T}\left[ \bigotimes_{j=1}^n V_{\alpha_j}(x_j) \otimes \mathcal{O}_{\mu\nu}(z) \right] = \exp\left[ - \mathrm{i} \sum_{1 \leq i < j \leq n} \alpha_i \alpha_j H^\text{F}(x_i,x_j) \right] \\
&\quad\times \Bigg[ \sum_{i,j=1}^n \alpha_i \alpha_j \partial_\mu G_\text{F}(z,x_i) \partial_\nu G_\text{F}(z,x_j) \, \mathcal{N}_G\left[ \prod_{j=1}^n V_{\alpha_j}(x_j) \right] \\
&\qquad- 2 \sum_{i=1}^n \alpha_i \partial_{(\mu} G_\text{F}(z,x_i) \, \mathcal{N}_G\left[ \partial_{\nu)} \phi(z) \prod_{j=1}^n V_{\alpha_j}(x_j) \right] \\
&\qquad+ \mathcal{N}_G\left[ \mathcal{O}_{\mu\nu}(z) \prod_{j=1}^n V_{\alpha_j}(x_j) \right] + \lim_{z' \to z} \partial^z_\mu \partial^{z'}_\nu W(z,z') \, \mathcal{N}_G\left[ \prod_{j=1}^n V_{\alpha_j}(x_j) \right] \Bigg] \\
&\qquad\times \exp\left[ - \frac{1}{2} \sum_{i,j=1}^n \alpha_i \alpha_j W(x_i,x_j) \right] \left( \frac{\Lambda}{\mu} \right)^\frac{\left( \sum_{k=1}^n \alpha_k \right)^2}{4 \pi} \eqend{,} \raisetag{2.4em}
\end{splitequation}
where $\mathcal{N}_G$ denotes normal-ordering with respect to the full two-point function~\eqref{eq:twopf_hadamard}, $H^\text{F}$ is the Feynman parametrix
\begin{splitequation}
\label{eq:hf_def}
H^\text{F}(x,y) &\equiv \Theta(x^0-y^0) H^+(x,y) + \Theta(y^0-x^0) H^+(y,x) \\
&= \frac{\mathrm{i}}{4 \pi} \ln\left[ \mu^2 ( - u v + \mathrm{i} \epsilon \abs{u + v} + \epsilon^2 ) \right]
\end{splitequation}
which is a fundamental solution of the massless Klein--Gordon equation:
\begin{splitequation}
\label{eq:hf_fundamental_solution}
\partial^2 H^\text{F}(x,y) &= - 4 \partial_u \partial_v H^\text{F}(x,y) = \frac{2 \epsilon}{\pi ( u^2 + \epsilon^2 )} \delta(u+v) \\
&\to 2 \delta(u) \delta(v) = \delta^2(x-y) \quad (\epsilon \to 0) \eqend{,}
\end{splitequation}
and
\begin{equation}
\label{eq:twopf_feynman}
G^\text{F}(x,y) \equiv H^\text{F}(x,y) + \frac{\mathrm{i}}{2 \pi} \ln\left( \frac{\Lambda}{\mu} \right) - \mathrm{i} W(x,y)
\end{equation}
is the time-ordered two-point function or Feynman propagator. Completely analogous to the proof of~\cite[Lemma 2]{froebcadamuro2022}, we also compute
\begin{splitequation}
\label{eq:timeordered_exponential_1phi}
&\mathcal{T}\left[ \bigotimes_{j=1}^n V_{\alpha_j}(x_j) \otimes \phi(z) \right] = \exp\left[ - \mathrm{i} \sum_{1 \leq i < j \leq n} \alpha_i \alpha_j H^\text{F}(x_i,x_j) \right] \\
&\qquad\times\Bigg[ \mathcal{N}_G\left[ \phi(z) \prod_{j=1}^n V_{\alpha_j}(x_j) \right] - \sum_{i=1}^n \alpha_i G^\text{F}(x_i,z) \mathcal{N}_G\left[ \prod_{j=1}^n V_{\alpha_j}(x_j) \right] \Bigg] \\
&\qquad\times \exp\left[ - \frac{1}{2} \sum_{i,j=1}^n \alpha_i \alpha_j W(x_i,x_j) \right] \left( \frac{\Lambda}{\mu} \right)^\frac{\left( \sum_{k=1}^n \alpha_k \right)^2}{4 \pi} \eqend{,}
\end{splitequation}
as well as (using now the anti-causal factorisation~\eqref{eq:timeordered_anticausal})
\begin{splitequation}
\label{eq:antitimeordered_exponential}
\overline{\mathcal{T}}\left[ \bigotimes_{j=1}^n V_{\alpha_j}(x_j) \right] &= \exp\left[ - \mathrm{i} \sum_{1 \leq i < j \leq n} \alpha_i \alpha_j H^\text{D}(x_i,x_j) \right] \mathcal{N}_G\left[ \prod_{j=1}^n V_{\alpha_j}(x_j) \right] \\
&\quad\times \exp\left[ - \frac{1}{2} \sum_{i,j=1}^n \alpha_i \alpha_j W(x_i,x_j) \right] \left( \frac{\Lambda}{\mu} \right)^\frac{\left( \sum_{k=1}^n \alpha_k \right)^2}{4 \pi} \eqend{,}
\end{splitequation}
where
\begin{equation}
\label{eq:hd_def}
H^\text{D}(x,y) \equiv \Theta(x^0-y^0) H^+(y,x) + \Theta(y^0-x^0) H^+(x,y)
\end{equation}
is the anti-time-ordered or Dyson parametrix. As in~\cite[Lemma 2]{froebcadamuro2022}, these expressions are smooth functions of the $x_i$ if $\epsilon > 0$, but for $\epsilon = 0$ need renormalisation (which we will do later).
To determine the interacting operators that we need, we use the product formula~\cite[Eq.~(126)]{froebcadamuro2022}
\begin{equation}
\label{eq:normal_ordering_starproduct}
\mathcal{N}_G\left[ \mathrm{e}^{\mathrm{i} (J,\phi)} \right] \star \mathcal{N}_G\left[ \mathrm{e}^{\mathrm{i} (K,\phi)} \right] = \exp\left[ - \mathrm{i} (J, G^+ \ast K) \right] \, \mathcal{N}_G\left[ \mathrm{e}^{\mathrm{i} (J+K,\phi)} \right] \eqend{,}
\end{equation}
where $(J,\phi) \equiv \int J(x) \phi(x) \mathop{}\!\mathrm{d}^2 x$ and $(J, G^+ \ast K) \equiv \iint J(x) G^+(x,y) K(y) \mathop{}\!\mathrm{d}^2 x \mathop{}\!\mathrm{d}^2 y$. Taking functional derivatives thereof with respect to $J$ or $K$ and choosing $J$ and $K$ appropriately, we obtain the product for terms involving powers of $\phi$ and its derivatives, while for $J$ or $K$ proportional to sums of $\delta$ distributions we obtain the product for vertex operators. From the Bogoliubov formula~\eqref{eq:opint_def} and the Sine--Gordon interaction~\eqref{eq:sint}, we first compute
\begin{splitequation}
&\mathcal{T}_\text{int}\left[ \mathcal{O}(z) \right] = \sum_{n=0}^\infty \sum_{k=0}^n \frac{(-1)^k \mathrm{i}^n}{k! (n-k)!} \int\dotsi\int \sum_{\sigma_j = \pm 1} \overline{\mathcal{T}}\left[ \bigotimes_{i=1}^k V_{\sigma_i \beta}(x_i) \right] \\
&\qquad\qquad \star \mathcal{T}\left[ \bigotimes_{j=k+1}^n V_{\sigma_j \beta}(x_j) \otimes \mathcal{O}(z) \right] \prod_{j=1}^n g(x_j) \mathop{}\!\mathrm{d}^2 x_j \\
&\quad= \sum_{n=0}^\infty \sum_{k=0}^n \sum_{\ell=0}^k \sum_{m=0}^{n-k} \frac{(-1)^k \mathrm{i}^n}{\ell! (k-\ell)! m! (n-k-m)!} \int\dotsi\int \\
&\qquad\times \overline{\mathcal{T}}\left[ \bigotimes_{i=1}^\ell V_\beta(x_i) \bigotimes_{j=1}^{k-\ell} V_{-\beta}(y_j) \right] \star \mathcal{T}\left[ \bigotimes_{i=\ell+1}^{\ell+m} V_\beta(x_i) \bigotimes_{j=k-\ell+1}^{n-\ell-m} V_{-\beta}(y_j) \otimes \mathcal{O}(z) \right] \\
&\qquad\times \prod_{i=1}^{\ell+m} g(x_i) \mathop{}\!\mathrm{d}^2 x_i \prod_{j=1}^{n-\ell-m} g(y_j) \mathop{}\!\mathrm{d}^2 y_j \eqend{,}
\end{splitequation}
where we used that the (anti-)time-ordered products are symmetric in their arguments, and that there are $k!/[ \ell! (k-\ell)! ]$ ways to choose $\ell$ vertex operators $V_\beta$ with positive sign from a total of $k$ ones $V_{\pm \beta}$ with either sign; we also renamed the insertion points of the $V_{-\beta}$ operators to $y_j$.
Using then the above results for the (anti-)time-ordered products~\eqref{eq:antitimeordered_exponential} and \eqref{eq:timeordered_exponential_1phi} and the product formula~\eqref{eq:normal_ordering_starproduct} as well as the decomposition of the two-point function~\eqref{eq:twopf_hadamard}, a long but straightforward computation results in
\begin{splitequation}
\label{eq:tint_1phi_def}
\mathcal{T}_\text{int}\left[ \partial_\mu \phi(z) \right] &= - \sum_{n=0}^\infty \sum_{k=0}^n \sum_{\ell=0}^k \sum_{m=0}^{n-k} \frac{(-1)^k \mathrm{i}^n}{\ell! (k-\ell)! m! (n-k-m)!} \int\dotsi\int \mathcal{E}_{k,\ell,n,m}(\vec{x};\vec{y}) \\
&\quad\times\Bigg[ \beta \partial^z_\mu \mathcal{G}_{k,\ell,n,m}(\vec{x};\vec{y};z) \, \mathcal{N}_G\left[ \prod_{i=1}^{\ell+m} V_\beta(x_i) \prod_{j=1}^{n-\ell-m} V_{-\beta}(y_j) \right] \\
&\qquad\quad- \mathcal{N}_G\left[ \partial_\mu \phi(z) \prod_{i=1}^{\ell+m} V_\beta(x_i) \prod_{j=1}^{n-\ell-m} V_{-\beta}(y_j) \right] \Bigg] \\
&\quad\times \left( \frac{\Lambda}{\mu} \right)^\frac{\beta^2 (2(\ell+m)-n)^2}{4 \pi} \prod_{i=1}^{\ell+m} g(x_i) \mathop{}\!\mathrm{d}^2 x_i \prod_{j=1}^{n-\ell-m} g(y_j) \mathop{}\!\mathrm{d}^2 y_j \eqend{,} \raisetag{2.3em}
\end{splitequation}
where we defined
\begin{splitequation}
\label{eq:calg_def}
\mathcal{G}_{k,\ell,n,m}(\vec{x};\vec{y};z) &\equiv \sum_{i=1}^\ell H^+(x_i,z) + \sum_{i=\ell+1}^{\ell+m} H^\text{F}(x_i,z) - \sum_{j=1}^{k-\ell} H^+(y_j,z) \\
&\quad- \sum_{j=k-\ell+1}^{n-\ell-m} H^\text{F}(y_j,z) - \mathrm{i} \sum_{i=1}^{\ell+m} W(x_i,z) + \mathrm{i} \sum_{j=1}^{n-\ell-m} W(y_j,z)
\end{splitequation}
and
\begin{splitequation}
\label{eq:cale_def}
&\mathcal{E}_{k,\ell,n,m}(\vec{x};\vec{y}) \equiv \exp\Bigg[ - \mathrm{i} \beta^2 \sum_{1 \leq i < j \leq \ell} H^\text{D}(x_i,x_j) + \mathrm{i} \beta^2 \sum_{i=1}^\ell \sum_{j=1}^{k-\ell} H^\text{D}(x_i,y_j) \\
&\qquad- \mathrm{i} \beta^2 \sum_{1 \leq i < j \leq k-\ell} H^\text{D}(y_i,y_j) - \mathrm{i} \beta^2 \sum_{\ell+1 \leq i < j \leq \ell+m} H^\text{F}(x_i,x_j) \\
&\qquad+ \mathrm{i} \beta^2 \sum_{i=\ell+1}^{\ell+m} \sum_{j=k-\ell+1}^{n-\ell-m} H^\text{F}(x_i,y_j) - \mathrm{i} \beta^2 \sum_{k-\ell+1 \leq i < j \leq n-\ell-m} H^\text{F}(y_i,y_j) \\
&\qquad- \mathrm{i} \beta^2 \sum_{i=1}^\ell \sum_{j=\ell+1}^{\ell+m} H^+(x_i,x_j) + \mathrm{i} \beta^2 \sum_{i=1}^\ell \sum_{j=k-\ell+1}^{n-\ell-m} H^+(x_i,y_j) \\
&\qquad+ \mathrm{i} \beta^2 \sum_{i=1}^{k-\ell} \sum_{j=\ell+1}^{\ell+m} H^+(y_i,x_j) - \mathrm{i} \beta^2 \sum_{i=1}^{k-\ell} \sum_{j=k-\ell+1}^{n-\ell-m} H^+(y_i,y_j) \Bigg] \\
&\times \exp\left[ - \frac{\beta^2}{2} \sum_{i,j=1}^{\ell+m} W(x_i,x_j) + \beta^2 \sum_{i=1}^{\ell+m} \sum_{j=1}^{n-\ell-m} W(x_i,y_j) - \frac{\beta^2}{2} \sum_{i,j=1}^{n-\ell-m} W(y_i,y_j) \right] \eqend{.}
\end{splitequation}
Analogously, we obtain
\begin{splitequation}
\label{eq:tint_2phi_def}
&\mathcal{T}_\text{int}\left[ \mathcal{O}_{\mu\nu}(z) \right] = \sum_{n=0}^\infty \sum_{k=0}^n \sum_{\ell=0}^k \sum_{m=0}^{n-k} \frac{(-1)^k \mathrm{i}^n}{\ell! (k-\ell)! m! (n-k-m)!} \int\dotsi\int \mathcal{E}_{k,\ell,n,m}(\vec{x};\vec{y}) \\
&\quad\times \Bigg[ \beta^2 \partial^z_\mu \mathcal{G}_{k,\ell,n,m}(\vec{x};\vec{y};z) \partial^z_\nu \mathcal{G}_{k,\ell,n,m}(\vec{x};\vec{y};z) \, \mathcal{N}_G\left[ \prod_{i=1}^{\ell+m} V_\beta(x_i) \prod_{j=1}^{n-\ell-m} V_{-\beta}(y_j) \right] \\
&\qquad\quad- 2 \beta \partial^z_{(\mu} \mathcal{G}_{k,\ell,n,m}(\vec{x};\vec{y};z) \, \mathcal{N}_G\left[ \partial_{\nu)} \phi(z) \prod_{i=1}^{\ell+m} V_\beta(x_i) \prod_{j=1}^{n-\ell-m} V_{-\beta}(y_j) \right] \\
&\qquad\quad+ \mathcal{N}_G\left[ \mathcal{O}_{\mu\nu}(z) \prod_{i=1}^{\ell+m} V_\beta(x_i) \prod_{j=1}^{n-\ell-m} V_{-\beta}(y_j) \right] \\
&\qquad\quad+ \lim_{z' \to z} \partial^z_\mu \partial^{z'}_\nu W(z,z') \, \mathcal{N}_G\left[ \prod_{i=1}^{\ell+m} V_\beta(x_i) \prod_{j=1}^{n-\ell-m} V_{-\beta}(y_j) \right] \Bigg] \\
&\quad\times \left( \frac{\Lambda}{\mu} \right)^\frac{\beta^2 (2(\ell+m)-n)^2}{4 \pi} \prod_{i=1}^{\ell+m} g(x_i) \mathop{}\!\mathrm{d}^2 x_i \prod_{j=1}^{n-\ell-m} g(y_j) \mathop{}\!\mathrm{d}^2 y_j \raisetag{2.3em}
\end{splitequation}
and
\begin{splitequation}
\label{eq:tint_vertex_def}
&\mathcal{T}_\text{int}\left[ V_\alpha(z) \right] = \sum_{n=0}^\infty \sum_{k=0}^n \sum_{\ell=0}^k \sum_{m=0}^{n-k} \frac{(-1)^k \mathrm{i}^n}{\ell! (k-\ell)! m! (n-k-m)!} \int\dotsi\int \mathcal{E}_{k,\ell,n,m}(\vec{x};\vec{y}) \\
&\qquad\times \exp\left[ - \mathrm{i} \alpha \beta \, \mathcal{G}_{k,\ell,n,m}(\vec{x};\vec{y};z) - \frac{\alpha^2}{2} W(z,z) \right] \\
&\qquad\times \mathcal{N}_G\left[ \prod_{i=1}^{\ell+m} V_\beta(x_i) \prod_{j=1}^{n-\ell-m} V_{-\beta}(y_j) V_\alpha(z) \right] \\
&\qquad\times \left( \frac{\Lambda}{\mu} \right)^\frac{\left[ ( 2(\ell+m)-n ) \beta + \alpha \right]^2}{4 \pi} \prod_{i=1}^{\ell+m} g(x_i) \mathop{}\!\mathrm{d}^2 x_i \prod_{j=1}^{n-\ell-m} g(y_j) \mathop{}\!\mathrm{d}^2 y_j \eqend{.} \raisetag{2.3em}
\end{splitequation}
\end{proof}
The proofs of Theorems~\ref{thm_renorm} and~\ref{thm_renorm_qed} then proceeds analogously to the proof of~\cite[Thms.~4--6]{froebcadamuro2022} in our previous work, where we showed renormalisability and convergence of the Gell-Mann--Low series for the interacting expectation value of the stress tensor as well as its conservation. However, since now we are considering the Bogoliubov definition of the interacting stress tensor, as well as smearing functions that are supported on a smooth time-like curve (instead of space-time smearings), there are some small differences. To shorten the proof, we will only explain the differences and refer the reader for all other details to~\cite{froebcadamuro2022}.
As in~\cite{froebcadamuro2022}, we see that taking the expectation value in the quasi-free state $\omega^{\Lambda,\epsilon}$ with two-point function $G^+$ and taking the limit $\Lambda \to 0$, only neutral configurations contribute at each order in perturbation theory. For the (smeared) expectation value of the interacting modified stress tensor $\hat{T}_{\mu\nu}$, using the results~\eqref{eq:tint_2phi_def} and~\eqref{eq:tint_vertex_def}, we thus obtain the series
\begin{splitequation}
\label{eq:tint_tmunu_series}
&\lim_{\Lambda \to 0} \omega^{\Lambda,\epsilon}\left( \mathcal{T}_\text{int}\left[ \hat{T}_{\mu\nu}(f) \right] \right) = \sum_{n=0}^\infty \sum_{k=0}^n \sum_{\ell=0}^k \sum_{m=0}^{n-k} \frac{(-1)^k \mathrm{i}^n}{\ell! (k-\ell)! m! (n-k-m)!} \int\dotsi\int \\
&\qquad\times \mathcal{E}_{k,\ell,n,m}(\vec{x};\vec{y}) \Theta^{k,\ell,n,m}_{\mu\nu}(\vec{x};\vec{y};z) f(z) \mathop{}\!\mathrm{d}^2 z \prod_{i=1}^{\ell+m} g(x_i) \mathop{}\!\mathrm{d}^2 x_i \prod_{j=1}^{n-\ell-m} g(y_j) \mathop{}\!\mathrm{d}^2 y_j
\end{splitequation}
with
\begin{splitequation}
\label{eq:theta_munu_def}
&\Theta^{k,\ell,n,m}_{\mu\nu}(\vec{x};\vec{y};z) \equiv \delta_{2(\ell+m),n} \Bigg[ \beta^2 \partial^z_\mu \mathcal{G}_{k,\ell,n,m}(\vec{x};\vec{y};z) \partial^z_\nu \mathcal{G}_{k,\ell,n,m}(\vec{x};\vec{y};z) \\
&\hspace{6em}- \frac{\beta^2}{2} \eta_{\mu\nu} \partial_z^\rho \mathcal{G}_{k,\ell,n,m}(\vec{x};\vec{y};z) \partial^z_\rho \mathcal{G}_{k,\ell,n,m}(\vec{x};\vec{y};z) \\
&\hspace{6em}+ \lim_{z' \to z} \left( \partial^z_\mu \partial^{z'}_\nu W(z,z') - \frac{1}{2} \eta_{\mu\nu} \partial_z^\rho \partial^{z'}_\rho W(z,z') \right) \Bigg] \\
&\quad+ \left( 1 - \frac{\beta^2}{8 \pi} \right) \eta_{\mu\nu} g(z) \delta_{2(\ell+m),n-1} \exp\left[ - \mathrm{i} \beta^2 \mathcal{G}_{k,\ell,n,m}(\vec{x};\vec{y};z) - \frac{\beta^2}{2} W(z,z) \right] \\
&\quad+ \left( 1 - \frac{\beta^2}{8 \pi} \right) \eta_{\mu\nu} g(z) \delta_{2(\ell+m),n+1} \exp\left[ \mathrm{i} \beta^2 \mathcal{G}_{k,\ell,n,m}(\vec{x};\vec{y};z) - \frac{\beta^2}{2} W(z,z) \right] \eqend{.}
\end{splitequation}
Inserting the Wightman, Feynman and Dyson Hadamard parametrices~\eqref{eq:hadamard}, \eqref{eq:hf_def} and~\eqref{eq:hd_def} into $\mathcal{E}_{k,\ell,n,m}$~\eqref{eq:cale_def}, we obtain a product of terms of the form
\begin{equation}
\left[ \mu^2 ( - u(x_i,x_j) v(x_i,x_j) + \mathrm{i} \epsilon ) \right]^{\pm \frac{\beta^2}{4\pi}}
\end{equation}
as well as such terms with $x_i$ or $x_j$ replaced by $y_i$ or $y_j$, and terms with different $\mathrm{i} \epsilon$ prescriptions: $- \mathrm{i} \epsilon$ instead of $+ \mathrm{i} \epsilon$ for the Dyson parametrices, and $+ \mathrm{i} \epsilon \sgn(x_i^0 - x_j^0)$ for the Wightman parametrices. For a negative exponent, these are singular as $\epsilon \to 0$, but the singularity is absolutely integrable since we are in the finite regime $\beta^2 < 4\pi$. We can thus take their absolute value and ignore the $\mathrm{i} \epsilon$ prescription, and then these terms are bounded in the same way as in the proof of~\cite[Thm.~5]{froebcadamuro2022}. The same applies to the exponentials containing parametrices in $\Theta^{k,\ell,n,m}_{\mu\nu}$~\eqref{eq:theta_munu_def} which are contained in $\mathcal{G}_{k,\ell,n,m}$~\eqref{eq:calg_def}, and which come from the vertex operators in the stress tensor.
We thus only need to potentially renormalise the products of derivatives of $\mathcal{G}_{k,\ell,n,m}$~\eqref{eq:calg_def} in the first contribution to $\Theta^{k,\ell,n,m}_{\mu\nu}$~\eqref{eq:theta_munu_def}. Expanding those, we have three different types of terms: products of two $W$ which are smooth, products of $W$ and one parametrix, and products of two parametrices. For the products of $W$ and one parametrix, we integrate by parts to obtain
\begin{equation}
\int \partial_\mu W(x,z) \partial_\nu H(y,z) f(z) \mathop{}\!\mathrm{d}^2 z = - \int H(y,z) \partial_\nu \left[ \partial_\mu W(x,z) f(z) \right] \mathop{}\!\mathrm{d}^2 z \eqend{,}
\end{equation}
where the points $x$ and $y$ may be the same, and where the parametrix $H$ is either a Wightman one $H^+$ or a Feynman one $H^\text{F}$. Since $W$ and $f$ are smooth and the parametrix is only logarithmically divergent, the singularity is then integrable and we can bound this integral in the same way as in the proof of~\cite[Thm.~5]{froebcadamuro2022}. We now consider the terms with two parametrices. If they are supported at different points, i.e., for the terms of the form $\partial_\mu H(x,z) \partial_\nu H(y,z)$ with $x \neq y$, no renormalisation is needed and we can bound them in the same way as in the proof of~\cite[Thm.~5]{froebcadamuro2022}. For this, we use that the mixed derivatives can be rewritten in the form
\begin{splitequation}
\int \partial^z_{(u} H^\text{F}(x,z) \partial^z_{v)} H^\text{F}(y,z) f(z) \mathop{}\!\mathrm{d}^2 z &= \frac{1}{8} H^\text{F}(x,y) [ f(x) + f(y) ] \\
&\quad+ \frac{1}{2} \int H^\text{F}(x,z) H^\text{F}(y,z) \partial_u \partial_v f(z) \mathop{}\!\mathrm{d}^2 z \eqend{,}
\end{splitequation}
which is~\cite[Eq.~(187)]{froebcadamuro2022}, and which follows straightforwardly from an integration by parts and the fact that the Feynman parametrix $H^\text{F}$ is a fundamental solution of the massless Klein--Gordon equation~\eqref{eq:hf_fundamental_solution}. Since the Wightman parametrix $H^+$ is a solution, for their products we have instead
\begin{equation}
\int \partial^z_{(u} H^+(x,z) \partial^z_{v)} H^+(y,z) f(z) \mathop{}\!\mathrm{d}^2 z = \frac{1}{2} \int H^+(x,z) H^+(y,z) \partial_u \partial_v f(z) \mathop{}\!\mathrm{d}^2 z \eqend{,}
\end{equation}
and for the mixed products we obtain
\begin{splitequation}
\int \partial^z_{(u} H^+(x,z) \partial^z_{v)} H^\text{F}(y,z) f(z) \mathop{}\!\mathrm{d}^2 z &= \frac{1}{8} H^+(x,y) f(y) \\
&\quad+ \frac{1}{2} \int H^+(x,z) H^\text{F}(y,z) \partial_u \partial_v f(z) \mathop{}\!\mathrm{d}^2 z \eqend{.}
\end{splitequation}
On the other hand, for the same derivatives it is shown in~\cite[Eqs.~(195)--(198)]{froebcadamuro2022} that
\begin{equation}
\partial^z_u H^\text{F}(x,z) \partial^z_u H^\text{F}(y,z) = \frac{1}{32 \pi^2} \frac{\partial^2}{\partial u(z)^2} \ln\left[ \mu^2 u(z,a)^2 \right] \eqend{,}
\end{equation}
where $a$ is some point such that $\min(u(x),u(y)) \leq u(a) \leq \max(u(x),u(y))$, and the analogous result for $v$ derivatives, and the same computation establishes this result also for the Wightman and mixed parametrices. All these can now be bounded in the same way as in the proof of~\cite[Thm.~5]{froebcadamuro2022}.
On the other hand, the terms with products of two parametrices at the same point need renormalisation. We know that~\cite[Eq.~(150)]{froebcadamuro2022}
\begin{splitequation}
\label{eq:thm_renorm_hmuhnu}
\partial_\mu H^\text{F}(x,y) \partial_\nu H^\text{F}(x,y) &= \left[ \partial_\mu H^\text{F}(x,y) \partial_\nu H^\text{F}(x,y) \right]^\text{ren} \\
&\quad- \frac{\mathrm{i}}{4 \pi} \, \eta_{\mu\nu} \ln(2 \mu \epsilon) \, \delta^2(x-y) + \bigo{\epsilon} \eqend{,}
\end{splitequation}
with $\delta^2(x-y) = 2 \delta(u) \delta(v)$ and
\begin{equations}[eq:thm_renorm_hmuhnu_ren]
\left[ \partial_u H^\text{F}(x,y) \partial_u H^\text{F}(x,y) \right]^\text{ren} &= - \frac{\mathrm{i}}{4 \pi} \, \partial_u^2 H^\text{F}(x,y) - \frac{\mathrm{i}}{16 \pi} \delta^2(x-y) \eqend{,} \\
\left[ \partial_v H^\text{F}(x,y) \partial_v H^\text{F}(x,y) \right]^\text{ren} &= - \frac{\mathrm{i}}{4 \pi} \, \partial_v^2 H^\text{F}(x,y) - \frac{\mathrm{i}}{16 \pi} \delta^2(x-y) \eqend{,} \\
\left[ \partial_u H^\text{F}(x,y) \partial_v H^\text{F}(x,y) \right]^\text{ren} &= \frac{1}{2} \partial_u \partial_v \left[ H^\text{F}(x,y) \right]^2 \eqend{.}
\end{equations}
Moreover, in the proof of~\cite[Thm.~6]{froebcadamuro2022} it is shown that there exists a redefinition of time-ordered products~\cite[Eqs.~(220)--(224)]{froebcadamuro2022} such that the last term is removed from $\left[ \partial_u H^\text{F}(x,y) \partial_u H^\text{F}(x,y) \right]^\text{ren}$ and $\left[ \partial_v H^\text{F}(x,y) \partial_v H^\text{F}(x,y) \right]^\text{ren}$. It is also shown that there exists another redefinition of time-ordered products~\cite[Eqs.~(152)--(155)]{froebcadamuro2022} such that the divergent term $\sim \ln \epsilon$ is removed; however, this is actually not necessary for the stress tensor, since it anyway cancels out in $\Theta^{k,\ell,n,m}_{\mu\nu}$~\eqref{eq:theta_munu_def}. We now only need to determine the corresponding formulas for the product of two Wightman parametrices $H^+$, since the product of a Feynman parametrix and a Wightman one at the same point never appears. Namely, using the explicit form of the parametrix~\eqref{eq:hadamard}, we compute
\begin{splitequation}
\label{eq:thm_renorm_hadamard_uu_ren}
\partial_u H^+(x,y) \partial_u H^+(x,y) &= - \frac{1}{16 \pi^2} \frac{1}{(u - \mathrm{i} \epsilon)^2} = \frac{1}{16 \pi^2} \partial_u^2 \ln (u - \mathrm{i} \epsilon) \\
&= - \frac{\mathrm{i}}{4 \pi} \partial_u^2 H^+(x,y)
\end{splitequation}
and the analogous equation for $\partial_v H^+(x,y) \partial_v H^+(x,y)$, as well as
\begin{equation}
\label{eq:thm_renorm_hadamard_uv_ren}
\partial_u H^+(x,y) \partial_v H^+(x,y) = - \frac{1}{16 \pi^2} \frac{1}{(u - \mathrm{i} \epsilon) (v - \mathrm{i} \epsilon)} = \frac{1}{2} \partial_u \partial_v \Big[ H^+(x,y) \Big]^2
\end{equation}
for the mixed derivatives. For these terms we thus do not need any redefinition of time-ordered products, and can also bound them in the same way as the product of Feynman parametrices, as in the proof of~\cite[Thm.~5]{froebcadamuro2022}.
It then follows as in the proof of~\cite[Thm.~5]{froebcadamuro2022} that we have the bounds
\begin{splitequation}
\label{eq:thm_renorm_bound}
&\abs{ \int\dotsi\int \mathcal{E}_{k,\ell,n,m}(\vec{x};\vec{y}) \Theta^{k,\ell,n,m}_{\mu\nu}(\vec{x};\vec{y};z) f(z) \mathop{}\!\mathrm{d}^2 z \prod_{i=1}^{\ell+m} g(x_i) \mathop{}\!\mathrm{d}^2 x_i \prod_{j=1}^{n-\ell-m} g(y_j) \mathop{}\!\mathrm{d}^2 y_j } \\
&\leq \delta_{2(\ell+m),n} C n^2 K^n \left[ (\ell+m)! \right]^{1+\frac{\beta^2}{4 \pi}} + \delta_{2(\ell+m),n} C' (K')^n \left[ (\ell+m)! \right]^{1+\frac{\beta^2}{4 \pi}} \\
&\quad+ \left( \delta_{2(\ell+m),n-1} + \delta_{2(\ell+m),n+1} \right) C'' (K'')^{n+1} \left[ (\ell+m+1)! \right]^{1+\frac{\beta^2}{4 \pi}} \eqend{,} \raisetag{2em}
\end{splitequation}
with the constants $C$, $C'$, $C''$, $K$, $K'$ and $K''$ depending on $f$, $g$, $\beta$ and $W$, and using that
\begin{equation}
(\ell+m+1)^{1+\frac{\beta^2}{4 \pi}} \leq (\ell+m+1)^2
\end{equation}
for $\beta^2 < 4 \pi$, the expectation value of the interacting stress tensor~\eqref{eq:tint_tmunu_series} is bounded by
\begin{splitequation}
\label{eq:renorm_stress_bound}
&\abs{ \lim_{\Lambda,\epsilon \to 0} \omega^{\Lambda,\epsilon}\left( \mathcal{T}_\text{int}\left[ \hat{T}_{\mu\nu}(f) \right] \right)} \leq \hat{C} \sum_{n=0}^\infty \hat{K}^n \sum_{k=0}^n \sum_{\ell=0}^k \sum_{m=0}^{n-k} \frac{\left[ (\ell+m)! \right]^{1+\frac{\beta^2}{4 \pi}}}{\ell! (k-\ell)! m! (n-k-m)!} \\
&\quad\times \Big[ \delta_{2(\ell+m),n} (1+n^2) + \Big( \delta_{2(\ell+m),n-1} + \delta_{2(\ell+m),n+1} \Big) (\ell+m+1)^2 \Big] \eqend{,} \raisetag{1.6em}
\end{splitequation}
where $\hat{C} = \max(C,C',C'' K'')$ and $\hat{K} = \max(K,K',K'')$. It remains to estimate the sums, and we start with the first term where only terms with $2(\ell+m) = n$ contribute. Changing $n \to 2n$, we thus have to bound
\begin{splitequation}
\label{eq:renorm_stress_bound1}
&\hat{C} \sum_{n=0}^\infty [ 1 + (2n)^2 ] \hat{K}^{2n} \sum_{k=0}^{2n} \sum_{\ell=0}^k \sum_{m=0}^{2n-k} \frac{\left[ (\ell+m)! \right]^{1+\frac{\beta^2}{4 \pi}}}{\ell! (k-\ell)! m! (2n-k-m)!} \delta_{(\ell+m),n} \\
&\leq 2 \hat{C} \sum_{n=0}^\infty (n+1)^2 \hat{K}^{2n} (n!)^{1+\frac{\beta^2}{4 \pi}} \sum_{k=0}^{2n} \sum_{\ell=\max(0,k-n)}^{\min(k,n)} \frac{1}{\ell! (k-\ell)! (n-\ell)! (n-k+\ell)!} \eqend{.}
\end{splitequation}
We split the sum over $k$ in two and shift $k \to k+n$ in the second half, such that
\begin{splitequation}
&\sum_{k=0}^{2n} \sum_{\ell=\max(0,k-n)}^{\min(k,n)} \frac{1}{\ell! (k-\ell)! (n-\ell)! (n-k+\ell)!} \\
&= \sum_{k=0}^n \sum_{\ell=0}^k \frac{1}{\ell! (k-\ell)! (n-\ell)! (n-k+\ell)!} + \sum_{k=1}^n \sum_{\ell=k}^n \frac{1}{\ell! (k+n-\ell)! (n-\ell)! (\ell-k)!} \\
&= \sum_{\ell=0}^n \sum_{k=\ell}^n \frac{1}{\ell! (k-\ell)! (n-\ell)! (n-k+\ell)!} + \sum_{\ell=1}^n \sum_{k=1}^\ell \frac{1}{\ell! (k+n-\ell)! (n-\ell)! (\ell-k)!} \\
&= \sum_{\ell=0}^n \sum_{k=\ell}^n \frac{1}{\ell! (n-k)! (n-\ell)! k!} + \sum_{\ell=1}^n \sum_{k=0}^{\ell-1} \frac{1}{\ell! (n-k)! (n-\ell)! k!} \eqend{,} \raisetag{2.1em}
\end{splitequation}
where we switched the order of summation of both sums in the second equality, and replaced $k \to n-k+\ell$ in the first sum and $k \to \ell-k$ in the second sum in the third equality. We can then extend the summation in the second sum to include $\ell = 0$, and both sums combine to give
\begin{splitequation}
&\sum_{k=0}^{2n} \sum_{\ell=\max(0,k-n)}^{\min(k,n)} \frac{1}{\ell! (k-\ell)! (n-\ell)! (n-k+\ell)!} \\
&= \left[ \sum_{\ell=0}^n \frac{1}{\ell! (n-\ell)!} \right] \sum_{k=0}^n \frac{1}{(n-k)! k!} = \frac{2^{2n}}{(n!)^2} \eqend{.}
\end{splitequation}
It follows that~\eqref{eq:renorm_stress_bound1} is bounded by
\begin{equation}
4 \hat{C} \sum_{n=0}^\infty \frac{(n+1)^2 (2\hat{K})^{2n}}{(n!)^{1-\frac{\beta^2}{4 \pi}}} < \infty \eqend{,}
\end{equation}
since we are in the finite regime $\beta^2 < 4 \pi$.
For the second term in the sum~\eqref{eq:renorm_stress_bound} where only terms with $2(\ell+m) = n-1$ contribute, we change $n \to 2n+1$ and have to bound
\begin{splitequation}
\label{eq:renorm_stress_bound2}
&\hat{C} \sum_{n=0}^\infty \hat{K}^{2n+1} \sum_{k=0}^{2n+1} \sum_{\ell=0}^k \sum_{m=0}^{2n+1-k} \frac{\left[ (\ell+m)! \right]^{1+\frac{\beta^2}{4 \pi}}}{\ell! (k-\ell)! m! (2n+1-k-m)!} (\ell+m+1)^2 \delta_{(\ell+m),n} \\
&= \hat{C} \hat{K} \sum_{n=0}^\infty (n+1)^2 \hat{K}^{2n} \sum_{k=0}^{2n+1} \sum_{\ell=\max(0,k-n-1)}^{\min(k,n)} \frac{(n!)^{1+\frac{\beta^2}{4 \pi}}}{\ell! (k-\ell)! (n-\ell)! (n+1-k+\ell)!} \eqend{.}
\end{splitequation}
Similarly to before, we split the sum over $k$, shift $k$ in the second sum and switch the order of summation to obtain
\begin{splitequation}
&\sum_{k=0}^{2n+1} \sum_{\ell=\max(0,k-n-1)}^{\min(k,n)} \frac{1}{\ell! (k-\ell)! (n-\ell)! (n+1-k+\ell)!} \\
&= \sum_{\ell=0}^n \sum_{k=\ell}^n \frac{1}{\ell! (k-\ell)! (n-\ell)! (n+1-k+\ell)!} \\
&\quad+ \sum_{\ell=0}^n \sum_{k=0}^\ell \frac{1}{\ell! (k+n+1-\ell)! (n-\ell)! (\ell-k)!} \eqend{.}
\end{splitequation}
We now replace $k \to n+1-k+\ell$ in the first sum and $k \to \ell-k$ in the second sum, and obtain
\begin{splitequation}
&\sum_{k=0}^{2n+1} \sum_{\ell=\max(0,k-n-1)}^{\min(k,n)} \frac{1}{\ell! (k-\ell)! (n-\ell)! (n+1-k+\ell)!} \\
&= \left[ \sum_{\ell=0}^n \frac{1}{\ell! (n-\ell)!} \right] \sum_{k=0}^{n+1} \frac{1}{(n+1-k)! k!} = \frac{2^{2n+1}}{n! (n+1)!} \eqend{,}
\end{splitequation}
such that~\eqref{eq:renorm_stress_bound2} is bounded by
\begin{equation}
2 \hat{C} \hat{K} \sum_{n=0}^\infty \frac{(n+1) (2\hat{K})^{2n}}{(n!)^{1-\frac{\beta^2}{4 \pi}}} < \infty \eqend{.}
\end{equation}
The last term in the sum~\eqref{eq:renorm_stress_bound} is bounded analogously, such that the expectation value of the interacting stress tensor~\eqref{eq:renorm_stress_bound} has a finite bound; in particular, the perturbative sum from the Bogoliubov formula~\eqref{eq:opint_def} is absolutely convergent in this case.
Conservation of the interacting (modified) stress tensor then follows exactly as in the proof of~\cite[Thm.~6]{froebcadamuro2022}, and we omit the details. \hfill\squareforqed
\section{Proof of Theorem~\ref{thm_renorm_qed} (Renormalisation of the quantum energy density)}
\label{sec_renorm_qed}
While Theorem~\ref{thm_renorm} shows that the expectation value of the stress tensor is finite when smeared with a test function of space-time, for the QEI we need to smear the energy density with a test function supported on a time-like curve. We thus need to derive suitable bounds to show that also this smearing is finite.
We consider a smooth time-like curve $z^\mu(\tau)$ with $\tau \in \mathbb{R}$ and define the associated relativistic velocity $v^\mu(\tau) \equiv \dot z^\mu(\tau)$, which we assume to be normalised: $v^\mu v_\mu = -1$. We also define the associated space-like vector $w_\mu(\tau) \equiv - \epsilon_{\mu\nu} v^\nu(\tau)$, which is also normalised: $w^\mu w_\mu = 1$. Using that $\epsilon_{\mu\nu} \epsilon_{\rho\sigma} = - 2 \eta_{\mu[\rho} \eta_{\sigma]\nu}$, it follows that the Minkowski metric can be written as
\begin{equation}
\label{eq:metric_vv_ww}
\eta_{\mu\nu} = - v_\mu v_\nu + w_\mu w_\nu \eqend{,} \quad \eta^{\mu\nu} = - v^\mu v^\nu + w^\mu w^\nu \eqend{.}
\end{equation}
The energy density as seen by an observer following this curve is then given by the contraction of the stress tensor with $v^\mu v^\nu$, and the (smeared) expectation value of the quantum energy density is~\eqref{eq:energy_density}
\begin{equation}
E_\omega(f) \equiv \lim_{\Lambda,\epsilon \to 0} \omega^{\Lambda,\epsilon}\left( \mathcal{T}_\text{int}\left[ \hat{T}_{\mu\nu}(v^\mu v^\nu f) \right] \right)
\end{equation}
with $f \in \mathcal{S}(\mathbb{R})$, where we integrate over the one-dimensional submanifold $z(\tau)$.
Analogously to the previous result~\eqref{eq:tint_tmunu_series} for a space-time smearing, using the results~\eqref{eq:tint_2phi_def} and~\eqref{eq:tint_vertex_def} we obtain the series
\begin{splitequation}
\label{eq:eomega_series}
E_\omega(f) &= \sum_{n=0}^\infty \sum_{k=0}^n \sum_{\ell=0}^k \sum_{m=0}^{n-k} \frac{(-1)^k \mathrm{i}^n}{\ell! (k-\ell)! m! (n-k-m)!} \int\dotsi\int \mathcal{E}_{k,\ell,n,m}(\vec{x};\vec{y}) \\
&\qquad\times \Theta^{k,\ell,n,m}(\vec{x};\vec{y};\tau) f(\tau) \mathop{}\!\mathrm{d} \tau \prod_{i=1}^{\ell+m} g(x_i) \mathop{}\!\mathrm{d}^2 x_i \prod_{j=1}^{n-\ell-m} g(y_j) \mathop{}\!\mathrm{d}^2 y_j
\end{splitequation}
for the (smeared) expectation value of the interacting energy density, where
\begin{splitequation}
\label{eq:theta_00_def}
&\Theta^{k,\ell,n,m}(\vec{x};\vec{y};\tau) \equiv v^\mu(\tau) v^\nu(\tau) \Theta^{k,\ell,n,m}_{\mu\nu}(\vec{x};\vec{y};z(\tau)) \\
&\quad= \frac{1}{2} \left( v^\mu v^\nu + w^\mu w^\nu \right) \delta_{2(\ell+m),n} \Bigg[ \beta^2 \partial^z_\mu \mathcal{G}_{k,\ell,n,m}(\vec{x};\vec{y};z) \partial^z_\nu \mathcal{G}_{k,\ell,n,m}(\vec{x};\vec{y};z) \\
&\hspace{14em}+ \lim_{z' \to z} \partial^z_\mu \partial^{z'}_\nu W(z,z') \Bigg] \\
&\qquad- \left( 1 - \frac{\beta^2}{8 \pi} \right) g(z) \delta_{2(\ell+m),n-1} \exp\left[ - \mathrm{i} \beta^2 \mathcal{G}_{k,\ell,n,m}(\vec{x};\vec{y};z) - \frac{\beta^2}{2} W(z,z) \right] \\
&\qquad- \left( 1 - \frac{\beta^2}{8 \pi} \right) g(z) \delta_{2(\ell+m),n+1} \exp\left[ \mathrm{i} \beta^2 \mathcal{G}_{k,\ell,n,m}(\vec{x};\vec{y};z) - \frac{\beta^2}{2} W(z,z) \right] \eqend{,}
\end{splitequation}
and we evaluate it on the observer's worldline $z = z(\tau)$. To simplify the expression for $\Theta^{k,\ell,n,m}$, we have used the expression for the inverse metric in terms of $v^\mu$ and $w^\mu$~\eqref{eq:metric_vv_ww}, from which it follows that for any $A_\mu$ and $B_\mu$
\begin{equation}
v^\mu v^\nu \left( A_\mu B_\nu - \frac{1}{2} \eta_{\mu\nu} A^\rho B_\rho \right) = \frac{1}{2} \left( v^\mu v^\nu + w^\mu w^\nu \right) A_\mu B_\nu \eqend{.}
\end{equation}
To show that $E_\omega(f)$ is renormalisable and the series~\eqref{eq:eomega_series} convergent, we need to show bounds on the smearing in $\tau$ which are of the same form as for a space-time smearing. Afterwards, we can proceed exactly as in the proofs of Theorem~\ref{thm_renorm} (the previous section) and~\cite[Thms.~4 and 5]{froebcadamuro2022}, and we omit the details.
For the required bounds, we first consider the last two terms in~\eqref{eq:theta_00_def}, which contain exponentials of the parametrix. As in the previous section, inserting the various parametrices into $\mathcal{E}_{k,\ell,n,m}$ and $\mathcal{G}_{k,\ell,n,m}$, we obtain after taking the absolute value a product of terms of the form
\begin{equation}
\abs{ \mu^2 u(x_i,x_j) v(x_i,x_j) }^{\pm \frac{\beta^2}{4\pi}}
\end{equation}
as well as such terms with $x_i$ or $x_j$ replaced by $y_i$ or $y_j$ or $z$. As in the proof of~\cite[Thm.~5]{froebcadamuro2022}, we can write them using the Cauchy determinant formula as a determinant, which can then be estimated by the absolute value of the sum of individual terms. For this, we need to combine the exponentials $\exp\left[ \pm \mathrm{i} \beta^2 \mathcal{G}_{k,\ell,n,m}(\vec{x};\vec{y};z) - \beta^2/2 W(z,z) \right]$ with $\mathcal{E}_{k,\ell,n,m}(\vec{x};\vec{y})$ analogously to~\cite[Eq.~(202)]{froebcadamuro2022}, renaming $z$ to $x_{\ell+m+1}$ for the second-to-last term in~\eqref{eq:theta_00_def} or $y_{n-\ell-m+1}$ for the last term in~\eqref{eq:theta_00_def}. Since both cases are bounded in the same way, we only show the first explicitly. We thus need to bound
\begin{splitequation}
\label{eq:thm_renorm_vertex_bound}
&\int\dotsi\int \left[ \sum_{\pi} \left( \prod_{j=1}^k \frac{1}{\mu \abs{u(x_j,y_{\pi(j)})}} \right) \frac{1}{\mu \abs{u(z(\tau),y_{\pi(k+1)})}} \right]^\frac{\beta^2}{4 \pi} \\
&\qquad\times \left[ \sum_{\pi} \left( \prod_{j=1}^k \frac{1}{\mu \abs{v(x_j,y_{\pi(j)})}} \right) \frac{1}{\mu \abs{v(z(\tau),y_{\pi(k+1)})}} \right]^\frac{\beta^2}{4 \pi} \\
&\qquad\times \abs{g(z(\tau))} \abs{f(\tau)} \mathop{}\!\mathrm{d} \tau \prod_{i=1}^k \abs{ g(x_i) } \mathop{}\!\mathrm{d} u(x_i) \mathop{}\!\mathrm{d} v(x_i) \prod_{i=1}^{k+1} \abs{ g(y_i) } \mathop{}\!\mathrm{d} u(y_i) \mathop{}\!\mathrm{d} v(y_i) \eqend{,}
\end{splitequation}
where $k = \ell+m$ and the sums run over all permutations $\pi$ of $\{1,\ldots,k+1\}$. The only difference to~\cite[Eq.~(163)]{froebcadamuro2022} is then that instead of a space-time integral over $x_{k+1}$, we only have a one-dimensional integral over $\tau$. We thus proceed exactly as in the proof of~\cite[Thm.~5]{froebcadamuro2022}, producing IR convergence factors $\big[ 1 + \mu^2 u(x)^2 \big]^{-2} \big[ 1 + \mu^2 v(x)^2 \big]^{-2}$ by bounding their inverse together with the adiabatic cutoff functions $g(x)$, using Hölder's inequality to bound the resulting integrals which factorise in $u$ and $v$, and extracting the sum over permutations (which gives an additional factor of $[(k+1)!]^{2/\rho}$ with $\rho = 8 \pi/(4 \pi + \beta^2)$). The only new bound that we need to prove is a bound on
\begin{splitequation}
&\iiint \Big[ \mu^2 \abs{u(z(\tau),y_{k+1}) v(z(\tau),y_{k+1})} \Big]^{- \rho \frac{\beta^2}{4 \pi}} \abs{f(\tau)} \mathop{}\!\mathrm{d} \tau \\
&\quad\times \frac{\mathop{}\!\mathrm{d} u(y_{k+1})}{[ 1 + \mu^2 u(y_{k+1})^2 ]^\rho} \frac{\mathop{}\!\mathrm{d} v(y_{k+1})}{[ 1 + \mu^2 v(y_{k+1})^2 ]^\rho} \eqend{,}
\end{splitequation}
since the integral over $x_{k+1} = z(\tau)$ is now only one-dimensional. We shift $u(y_{k+1}) \to u(y_{k+1}) + u(z(\tau))$, $v(y_{k+1}) \to v(y_{k+1}) + v(z(\tau))$ and perform these integrals first. We split the integral over $u(y_{k+1})$ into two parts: If $\mu \abs{u(y_{k+1})} \leq 1$, we estimate the factor $[ 1 + \mu^2 [ u(y_{k+1}) + u(z(\tau)) ]^2 ]^{-\rho}$ by $1$ and obtain
\begin{equation}
\int_{\mu \abs{u(y_{k+1})} \leq 1} \Big[ \mu \abs{u(y_{k+1})} \Big]^{- \rho \frac{\beta^2}{4 \pi}} \frac{\mathop{}\!\mathrm{d} u(y_{k+1})}{[ 1 + \mu^2 [ u(y_{k+1}) + u(z(\tau)) ]^2 ]^\rho} \leq \frac{8 \pi}{4 \pi - \rho \beta^2} \mu^{-1} \eqend{,}
\end{equation}
while for $\mu \abs{u(y_{k+1})} > 1$ we estimate
\begin{splitequation}
&\int_{\mu \abs{u(y_{k+1})} > 1} \Big[ \mu \abs{u(y_{k+1})} \Big]^{- \rho \frac{\beta^2}{4 \pi}} \frac{\mathop{}\!\mathrm{d} u(y_{k+1})}{[ 1 + \mu^2 [ u(y_{k+1}) + u(z(\tau)) ]^2 ]^\rho} \\
&\quad\leq \int \frac{\mathop{}\!\mathrm{d} u(y_{k+1})}{[ 1 + \mu^2 [ u(y_{k+1}) + u(z(\tau)) ]^2 ]^\rho} = \frac{\sqrt{\pi}}{\mu} \frac{\Gamma\left( \rho - \frac{1}{2} \right)}{\Gamma(\rho)} \eqend{.}
\end{splitequation}
The integral over $v(y_{k+1})$ is bounded in the same way, and finally the integral over $\tau$ gives $\norm{ f }_1$. It follows that for the contribution of the last two terms in~\eqref{eq:theta_00_def} to the expectation value of the energy density we have a bound of the same form as for a space-time smearing, namely
\begin{splitequation}
\label{eq:thm_renorm_vertex_bound2}
&\int\dotsi\int \abs{ \mathcal{E}_{k,\ell,n,m}(\vec{x};\vec{y}) } \left( 1 - \frac{\beta^2}{8 \pi} \right) \delta_{2(\ell+m),n+1} \abs{ g(z) } \abs{ f(\tau) } \\
&\times \abs{ \exp\left[ \mathrm{i} \beta^2 \mathcal{G}_{k,\ell,n,m}(\vec{x};\vec{y};z) - \frac{\beta^2}{2} W(z,z) \right] } \mathop{}\!\mathrm{d} \tau \prod_{i=1}^{\ell+m} \abs{ g(x_i) } \mathop{}\!\mathrm{d}^2 x_i \prod_{j=1}^{n-\ell-m} \abs{ g(y_j) } \mathop{}\!\mathrm{d}^2 y_j \\
&\leq \delta_{2(\ell+m),n-1} C K^{n+1} \left[ (\ell+m+1)! \right]^{1+\frac{\beta^2}{4 \pi}} \raisetag{1.4em}
\end{splitequation}
with constants $C$ and $K$ depending on $f$, $g$, $\beta$ and $z(\tau)$; compare~\eqref{eq:thm_renorm_bound} for the space-time smearing. We note that due to our condition~\eqref{eq:thm_renorm_wpos_disc}
\begin{equation}
\sum_{i,j=1}^n \Big[ W(x_i,x_j) - W(x_i,y_j) - W(y_i,x_j) + W(y_i,y_j) \Big] \geq 0
\end{equation}
for all $n \in \mathbb{N}$, the $W$ dependence in the exponential in $\mathcal{E}_{k,\ell,n,m}(\vec{x};\vec{y})$~\eqref{eq:cale_def}, combined with $W(z,z)$, is simply estimated by $1$, and thus the constants $C$ and $K$ do not depend on $W$.
To bound the first term in~\eqref{eq:theta_00_def}, we first consider the second derivative of the state-dependent part $W$. A straightforward bound is
\begin{splitequation}
&\abs{ \int \left. \frac{1}{2} \left( v^\mu v^\nu + w^\mu w^\nu \right) \lim_{z' \to z} \partial^z_\mu \partial^{z'}_\nu W(z,z') \right\rvert_{z = z(\tau)} f(\tau) \mathop{}\!\mathrm{d} \tau } \\
&\quad\leq \sup_{\mu \in \{0,1\} } \norm{ \sup(v^\mu) }_\infty^2 \sup_{\mu,\nu \in \{0,1\}} \norm{ \partial^z_\mu \partial^{z'}_\nu W(z,z') }_\infty \norm{ f }_1 \eqend{,}
\end{splitequation}
where we used that $w^1 = v^0$, $w^0 = v^1$. For the terms contained in $\mathcal{G}_{k,\ell,n,m}$~\eqref{eq:calg_def}, we again have three different types of terms: products of two $W$, products of $W$ and one parametrix, and products of two parametrices. For products of two $W$, we again have a straightforward bound
\begin{splitequation}
&\abs{ \int \left. \frac{1}{2} \left( v^\mu v^\nu + w^\mu w^\nu \right) \partial^z_\mu W(x,z) \partial^z_\nu W(y,z) \right\rvert_{z = z(\tau)} f(\tau) \mathop{}\!\mathrm{d} \tau } \\
&\quad\leq \sup_{\mu \in \{0,1\} } \norm{ \sup(v^\mu) }_\infty^2 \sup_{\mu \in \{0,1\}} \norm{ \partial^z_\mu W(x,z) }_\infty^2 \norm{ f }_1 \eqend{,}
\end{splitequation}
where $x$ and $y$ can be equal or distinct. For the products of $W$ and one parametrix, we compute instead
\begin{splitequation}
\label{eq:thm_renorm_qed_whder}
&\int \left. \frac{1}{2} \left( v^\mu v^\nu + w^\mu w^\nu \right) \partial^z_\mu W(x,z) \partial^z_\nu H(y,z) \right\rvert_{z = z(\tau)} f(\tau) \mathop{}\!\mathrm{d} \tau \\
&= \int \Big[ (v^0 - v^1)^2 \partial_{u(z)} W(x,z) \partial_{u(z)} H(y,z) \\
&\qquad+ (v^0 + v^1)^2 \partial_{v(z)} W(x,z) \partial_{v(z)} H(y,z) \Big]_{z = z(\tau)} f(\tau) \mathop{}\!\mathrm{d} \tau \eqend{,}
\end{splitequation}
using that
\begin{equation}
v^u = v^0 - v^1 \eqend{,} \quad v^v = v^0 + v^1 \eqend{,} \quad w^u = - v^u \eqend{,} \quad w^v = v^v \eqend{.}
\end{equation}
In particular, the mixed terms where one derivative is with respect to $u$ and one with respect to $v$ have dropped out. We now need to change the derivatives acting on the parametrix into the $\tau$ derivative, for which we first compute
\begin{equations}[eq:dudv_dtau]
\partial_\tau u(z(\tau)) &= \partial_\tau \Big[ z^0(\tau) - z^1(\tau) \Big] = v^0 - v^1 = \sqrt{1 + (v^1)^2} - v^1 > 0 \eqend{,} \\
\partial_\tau v(z(\tau)) &= \partial_\tau \Big[ z^0(\tau) + z^1(\tau) \Big] = v^0 + v^1 = \sqrt{1 + (v^1)^2} + v^1 > 0 \eqend{,}
\end{equations}
using that $z(\tau)$ is a future-directed time-like curve. For the Wightman parametrix \eqref{eq:hadamard}, it follows that
\begin{splitequation}
\label{eq:hadamard_der_u_tau}
\partial_{u(z)} H^+(y,z) \Big\rvert_{z = z(\tau)} &= - \frac{\mathrm{i}}{4 \pi} \frac{1}{u(y,z) - \mathrm{i} \epsilon} \bigg\rvert_{z = z(\tau)} \\
&= \frac{\mathrm{i}}{4 \pi} \frac{1}{v^0 - v^1} \partial_\tau \ln\left[ \mu ( \epsilon + \mathrm{i} u(y,z(\tau)) ) \right]
\end{splitequation}
and the analogous equation with $v^0 + v^1$ for the $v(z)$ derivative, while for the Feynman parametrix~\eqref{eq:hf_def} we obtain
\begin{splitequation}
\label{eq:hf_der_u_tau}
&\partial_{u(z)} H^\text{F}(y,z) \Big\rvert_{z = z(\tau)} = - \frac{\mathrm{i}}{4 \pi} \left[ \frac{\Theta(u(y,z)+v(y,z))}{u(y,z) - \mathrm{i} \epsilon} + \frac{\Theta(-u(y,z)-v(y,z))}{u(y,z) + \mathrm{i} \epsilon} \right]_{z = z(\tau)} \\
&\quad= \frac{\mathrm{i}}{4 \pi} \frac{1}{v^0 - v^1} \partial_\tau \Big[ \Theta(u(y,z(\tau))+v(y,z(\tau))) \ln\left[ \mu ( \epsilon + \mathrm{i} u(y,z(\tau)) ) \right] \\
&\hspace{8em}+ \Theta(-u(y,z(\tau))-v(y,z(\tau))) \ln\left[ \mu ( \epsilon - \mathrm{i} u(y,z(\tau)) ) \right] \Big] \\
&\qquad- \frac{\mathrm{i}}{4 \pi} \frac{v^0}{v^0 - v^1} \delta(y^0-z^0(\tau)) \mathrm{i} \pi \sgn[ y^1 - z^1(\tau) ] \eqend{,} \raisetag{1.6em}
\end{splitequation}
and the analogous equation with $u$ replaced by $v$, $v^0 - v^1$ by $v^0 + v^1$ and the $\sgn$ function replaced by its negative. Integrating the $\tau$ derivative by parts, and changing coordinates to either $u(z(\tau))$ or $v(z(\tau))$ according to the argument of the logarithm, for the Wightman parametrix we thus obtain
\begin{splitequation}
\label{eq:thm_renorm_qed_bound_wh1}
&\int \left. \frac{1}{2} \left( v^\mu v^\nu + w^\mu w^\nu \right) \partial^z_\mu W(x,z) \partial^z_\nu H^+(y,z) \right\rvert_{z = z(\tau)} f(\tau) \mathop{}\!\mathrm{d} \tau \\
&= - \frac{\mathrm{i}}{4 \pi} \int \bigg[ f(\tau) \partial_\tau \Big[ \partial_{u(z)} W(x,z) \Big]_{z = z(\tau)} + f(\tau) \partial_\tau \ln(v^0 - v^1) \Big[ \partial_{u(z)} W(x,z) \Big]_{z = z(\tau)} \\
&\qquad\qquad+ \partial_\tau f(\tau) \Big[ \partial_{u(z)} W(x,z) \Big]_{z = z(\tau)} \bigg] \ln\left[ \mu ( \epsilon + \mathrm{i} u(y,z(\tau)) ) \right] \mathop{}\!\mathrm{d} u(z(\tau)) \\
&\quad- \frac{\mathrm{i}}{4 \pi} \int \bigg[ f(\tau) \partial_\tau \Big[ \partial_{v(z)} W(x,z) \Big]_{z = z(\tau)} + f(\tau) \partial_\tau \ln (v^0 + v^1) \Big[ \partial_{v(z)} W(x,z) \Big]_{z = z(\tau)} \\
&\qquad\qquad+ \partial_\tau f(\tau) \Big[ \partial_{v(z)} W(x,z) \Big]_{z = z(\tau)} \bigg] \ln\left[ \mu ( \epsilon + \mathrm{i} v(y,z(\tau)) ) \right] \mathop{}\!\mathrm{d} v(z(\tau)) \eqend{,}
\end{splitequation}
where now $\tau$ (and thus also $z^\mu(\tau)$) is a function of $u(z(\tau))$ in the first integral and of $v(z(\tau))$ in the second integral. This identification proceeds through the identities
\begin{equation}
\label{eq:thm_renorm_udot}
u(\tau) = z^0(\tau) - z^1(\tau) \eqend{,} \quad \dot u(\tau) = v^0(\tau) - v^1(\tau) = \sqrt{1 + [ v^1(\tau) ]^2} - v^1(\tau) \eqend{,}
\end{equation}
where we used the normalisation $v^\mu v_\mu = -1$. It follows that
\begin{equation}
v^1(\tau) = \frac{1 - \dot u(\tau)^2}{2 \dot u(\tau)} \eqend{,} \quad v^0(\tau) = \frac{1 + \dot u(\tau)^2}{2 \dot u(\tau)} \eqend{,}
\end{equation}
where we used that $z^\mu(\tau)$ is a future-directed curve such that $v^0 > 0$ and $\dot u(\tau) > 0$, and from this finally
\begin{equation}
\label{eq:zmu_tau_of_uv}
z^\mu(\tau) = z^\mu(0) + \int_0^\tau v^\mu(\sigma) \mathop{}\!\mathrm{d} \sigma \eqend{,} \quad v(z(\tau)) = v(u(\tau)) = v(0) + \int_0^\tau \frac{1}{\dot u(\sigma)} \mathop{}\!\mathrm{d} \sigma \eqend{.}
\end{equation}
Analogously, for the second integral we have to use
\begin{equation}
u(z(\tau)) = u(v(\tau)) = u(0) + \int_0^\tau \frac{1}{\dot v(\sigma)} \mathop{}\!\mathrm{d} \sigma \eqend{.}
\end{equation}
To bound~\eqref{eq:thm_renorm_qed_bound_wh1}, we estimate the terms in brackets by their supremum:
\begin{splitequation}
&\bigg\lvert f(\tau) \partial_\tau \Big[ \partial_{u(z)} W(x,z) \Big]_{z = z(\tau)} + f(\tau) \partial_\tau \ln(v^0 - v^1) \Big[ \partial_{u(z)} W(x,z) \Big]_{z = z(\tau)} \\
&\quad+ \partial_\tau f(\tau) \Big[ \partial_{u(z)} W(x,z) \Big]_{z = z(\tau)} \bigg\rvert \\
&\leq \frac{1}{1 + \mu^2 u(z(\tau))^2} \sup_\tau \bigg\lvert [ 1 + \mu^2 u(z(\tau))^2 ] \bigg[ f(\tau) \partial_\tau \Big[ \partial_{u(z)} W(x,z) \Big]_{z = z(\tau)} \\
&\quad+ f(\tau) \partial_\tau \ln(v^0 - v^1) \Big[ \partial_{u(z)} W(x,z) \Big]_{z = z(\tau)} + \partial_\tau f(\tau) \Big[ \partial_{u(z)} W(x,z) \Big]_{z = z(\tau)} \bigg] \bigg\rvert
\end{splitequation}
and the analogous estimate for the second integral, as well as
\begin{equation}
\abs{ \ln\left[ \mu ( \epsilon + \mathrm{i} u(y,z(\tau)) ) \right] } \leq \abs{ \ln\left[ \mu \abs{ u(y,z(\tau)) } \right] } + \frac{\pi}{2} \eqend{,}
\end{equation}
which holds in the limit $\epsilon \to 0$. The remaining dependence on $x$ can then be absorbed in the adiabatic cutoff functions $g(x)$ as in the proof of~\cite[Thm.~5]{froebcadamuro2022}. For the remaining integral over $u(z(\tau))$ or $v(z(\tau))$, we use the bound~\cite[Eq.~(180)]{froebcadamuro2022}
\begin{equation}
\label{eq:mink_conv_num_2_int}
\int \frac{\abs{ \ln\abs{ \mu u(z,x) } }^k}{1 + \mu^2 u(z)^2} \mathop{}\!\mathrm{d} u(z) \leq \frac{2}{\mu} c_k \ln^k\left( 2 + \mu \abs{u(x)} \right)
\end{equation}
and $\int [ 1 + \mu^2 u(z)^2 ]^{-1} \mathop{}\!\mathrm{d} u(z) = \pi/\mu$. As in the proof of~\cite[Thm.~5]{froebcadamuro2022}, the logarithm can again be absorbed in the adiabatic cutoff functions $g(x)$.
For the Feynman parametrices $H^\text{F}$, using~\eqref{eq:hf_der_u_tau} instead of~\eqref{eq:hadamard_der_u_tau} the same procedure leads to
\begin{splitequation}
\label{eq:thm_renorm_qed_bound_wh2}
&\int \left. \frac{1}{2} \left( v^\mu v^\nu + w^\mu w^\nu \right) \partial^z_\mu W(x,z) \partial^z_\nu H^\text{F}(y,z) \right\rvert_{z = z(\tau)} f(\tau) \mathop{}\!\mathrm{d} \tau \\
&= - \frac{\mathrm{i}}{4 \pi} \int \Big[ f(\tau) \partial_\tau \Big[ \partial_{u(z)} W(x,z) \Big]_{z = z(\tau)} + f(\tau) \partial_\tau \ln(v^0 - v^1) \Big[ \partial_{u(z)} W(x,z) \Big]_{z = z(\tau)} \\
&\qquad\qquad+ \partial_\tau f(\tau) \Big[ \partial_{u(z)} W(x,z) \Big]_{z = z(\tau)} \Big] \Big[ \Theta(y^0+z^0(\tau)) \ln\left[ \mu ( \epsilon + \mathrm{i} u(y,z(\tau)) ) \right] \\
&\qquad\qquad+ \Theta(-y^0-z^0(\tau)) \ln\left[ \mu ( \epsilon - \mathrm{i} u(y,z(\tau)) ) \right] \Big] \mathop{}\!\mathrm{d} u(z(\tau)) \\
&\quad- \frac{\mathrm{i}}{4 \pi} \int \Big[ f(\tau) \partial_\tau \Big[ \partial_{v(z)} W(x,z) \Big]_{z = z(\tau)} + f(\tau) \partial_\tau \ln(v^0 + v^1) \Big[ \partial_{v(z)} W(x,z) \Big]_{z = z(\tau)} \\
&\qquad\qquad+ \partial_\tau f(\tau) \Big[ \partial_{v(z)} W(x,z) \Big]_{z = z(\tau)} \Big] \Big[ \Theta(y^0+z^0(\tau)) \ln\left[ \mu ( \epsilon + \mathrm{i} v(y,z(\tau)) ) \right] \\
&\qquad\qquad+ \Theta(-y^0-z^0(\tau)) \ln\left[ \mu ( \epsilon - \mathrm{i} v(y,z(\tau)) ) \right] \Big] \mathop{}\!\mathrm{d} v(z(\tau)) \\
&\quad- \frac{1}{4} \int v^0 \delta(y^0-z^0(\tau)) \sgn[ y^1 - z^1(\tau) ] \\
&\qquad\quad\times \Big[ (v^0 + v^1) \partial_{v(z)} W(x,z) - (v^0 - v^1) \partial_{u(z)} W(x,z) \Big]_{z = z(\tau)} f(\tau) \mathop{}\!\mathrm{d} \tau \eqend{,} \raisetag{1.8em}
\end{splitequation}
and the first two terms can be bounded in the same way as for the Wightman parametrix, using that $\abs{ \Theta(x) } \leq 1$. For the last term, we use the $\delta$ to perform the integral over $\tau$, using~\eqref{eq:dudv_dtau} to change variables. This gives
\begin{equation}
v^0 \mathop{}\!\mathrm{d} \tau = \mathop{}\!\mathrm{d} z^0(\tau) \eqend{,}
\end{equation}
and then we can estimate this term by its supremum over all $\tau$:
\begin{splitequation}
&\bigg\lvert - \frac{1}{4} \int v^0 \delta(y^0-z^0(\tau)) \sgn[ y^1 - z^1(\tau) ] \\
&\qquad\times \Big[ (v^0 + v^1) \partial_{v(z)} W(x,z) - (v^0 - v^1) \partial_{u(z)} W(x,z) \Big]_{z = z(\tau)} f(\tau) \mathop{}\!\mathrm{d} \tau \bigg\rvert \\
&\leq \frac{1}{4} \sup_\tau \abs{ \Big[ (v^0 + v^1) \partial_{v(z)} W(x,z) - (v^0 - v^1) \partial_{u(z)} W(x,z) \Big]_{z = z(\tau)} f(\tau) } \eqend{.}
\end{splitequation}
Again, the dependence on $x$ can be absorbed in the adiabatic cutoff functions $g(x)$ as in the proof of~\cite[Thm.~5]{froebcadamuro2022}.
Lastly, we need to consider terms with products of two Hadamard parametrices appearing in the first term of~\eqref{eq:theta_00_def}, either at the same point or at two different points. We first compute the analogue of~\eqref{eq:thm_renorm_qed_whder}
\begin{splitequation}
\label{eq:thm_renorm_qed_hhder}
&\int \left. \frac{1}{2} \left( v^\mu v^\nu + w^\mu w^\nu \right) \partial^z_\mu H(x,z) \partial^z_\nu H(y,z) \right\rvert_{z = z(\tau)} f(\tau) \mathop{}\!\mathrm{d} \tau \\
&= \int \Big[ (v^0 - v^1)^2 \partial_{u(z)} H(x,z) \partial_{u(z)} H(y,z) \\
&\qquad+ (v^0 + v^1)^2 \partial_{v(z)} H(x,z) \partial_{v(z)} H(y,z) \Big]_{z = z(\tau)} f(\tau) \mathop{}\!\mathrm{d} \tau \eqend{,}
\end{splitequation}
and then start with the case of equal points $x = y$. As in the space-time smearing case, only products of two Wightman parametrices or two Feynman parametrices appear, for which we have computed the relevant expressions~\eqref{eq:thm_renorm_hmuhnu_ren} and~\eqref{eq:thm_renorm_hadamard_uu_ren} for products of their derivatives. Again, the logarithmically divergent term appearing for the Feynman parametrices~\eqref{eq:thm_renorm_hmuhnu} drops out since it is proportional to the metric $\eta_{\mu\nu}$ and thus vanishes when contracted with $v^\mu v^\nu + w^\mu w^\nu$, and we also recall that we have used a redefinition of time-ordered products to get rid of the local terms in~\eqref{eq:thm_renorm_hmuhnu_ren}. It follows that both for the Wightman and Feynman parametrices, we have
\begin{splitequation}
&\int \left. \frac{1}{2} \left( v^\mu v^\nu + w^\mu w^\nu \right) \left[ \partial^z_\mu H(x,z) \partial^z_\nu H(x,z) \right]^\text{ren} \right\rvert_{z = z(\tau)} f(\tau) \mathop{}\!\mathrm{d} \tau \\
&= \int \left. \frac{1}{2} \left( v^\mu v^\nu + w^\mu w^\nu \right) \partial^z_\mu H(x,z) \partial^z_\nu H(x,z) \right\rvert_{z = z(\tau)} f(\tau) \mathop{}\!\mathrm{d} \tau \\
&= - \frac{\mathrm{i}}{4 \pi} \int \Big[ (v^0 - v^1)^2 \partial_{u(z)}^2 H(x,z) + (v^0 + v^1)^2 \partial_{v(z)}^2 H(x,z) \Big]_{z = z(\tau)} f(\tau) \mathop{}\!\mathrm{d} \tau \eqend{,}
\end{splitequation}
and analogously to~\eqref{eq:hadamard_der_u_tau} and~\eqref{eq:hf_der_u_tau} we convert the $u$ and $v$ derivatives into $\tau$ derivatives. We compute
\begin{splitequation}
&\left[ (v^0-v^1)^2 \partial_u^2 H(x,z) + (v^0+v^1)^2 \partial_v^2 H(x,z) \right]_{z = z(\tau)} \\
&\quad= \partial_\tau^2 H(x,z(\tau)) - 2 \left[ \partial_{u(z)} \partial_{v(z)} H(x,z) \right]_{z = z(\tau)} \\
&\qquad- [ \partial_\tau (v^0-v^1) ] \left[ \partial_{u(z)} H(x,z) \right]_{z = z(\tau)} - [ \partial_\tau (v^0+v^1) ] \left[ \partial_{v(z)} H(x,z) \right]_{z = z(\tau)} \eqend{,}
\end{splitequation}
and for a Wightman parametrix we have $\partial_{u(z)} \partial_{v(z)} H^+(x,z) = 0$, while for a Feynman parametrix $\partial_{u(z)} \partial_{v(z)} H^\text{F}(x,z) = - \frac{1}{4} \delta^2(x-z)$~\eqref{eq:hf_fundamental_solution}. The remaining $u$ and $v$ derivatives we can convert into $\tau$ derivatives using~\eqref{eq:hadamard_der_u_tau} and~\eqref{eq:hf_der_u_tau}, partially integrate all $\tau$ derivatives to act on the test function $f$ or the $v^\mu(\tau)$, and then bound the result completely analogous to the previous case of a product of $W$ and one parametrix.
The only potentially problematic term is the local one $\sim \delta^2(x-z(\tau))$ that appears for a product of Feynman parametrices, whose contribution to~\eqref{eq:thm_renorm_qed_hhder} (for the Feynman parametrix) is
\begin{equation}
- \frac{\mathrm{i}}{8 \pi} \int \delta^2(x-z(\tau)) f(\tau) \mathop{}\!\mathrm{d} \tau \eqend{,}
\end{equation}
and whose contribution to the first term of~\eqref{eq:theta_00_def} and thus the expectation value of the interacting energy density~\eqref{eq:eomega_series} reads
\begin{splitequation}
\label{eq:thm_renorm_qed_dhdh_local}
&- \frac{\mathrm{i}}{8 \pi} \beta^2 \sum_{n=0}^\infty \sum_{k=0}^{2n} \sum_{\ell=\max(0,k-n)}^{\min(k,n)} \frac{(-1)^{k+n}}{\ell! (k-\ell)! (n-\ell)! (n-k+\ell)!} \int\dotsi\int \\
&\qquad\times \int \left[ \sum_{i=\ell+1}^n \delta^2(x_i-z(\tau)) + \sum_{j=k-\ell+1}^n \delta^2(y_j-z(\tau)) \right] f(\tau) \mathop{}\!\mathrm{d} \tau \\
&\qquad\times \mathcal{E}_{k,\ell,2n,n-\ell}(\vec{x};\vec{y}) \prod_{i=1}^n g(x_i) g(y_i) \mathop{}\!\mathrm{d}^2 x_i \mathop{}\!\mathrm{d}^2 y_i \eqend{,}
\end{splitequation}
where because of the Kronecker $\delta_{2(\ell+m),n}$ we replaced $n \to 2n$ and then performed the sum over $m$. Let us thus consider one of the terms, say with $\delta^2(x_n-z(\tau))$ which allows us to perform the integral over $x_n$ and bound the result by
\begin{splitequation}
&\frac{\beta^2}{8 \pi} \sum_{n=0}^\infty \sum_{k=0}^{2n} \sum_{\ell=\max(0,k-n)}^{\min(k,n)} \frac{1}{\ell! (k-\ell)! (n-\ell)! (n-k+\ell)!} \int\dotsi\int \abs{ \mathcal{E}_{k,\ell,2n,n-\ell}(\hat{\vec{x}};\vec{y}) } \\
&\quad\times \prod_{i=1}^{n-1} \abs{ g(x_i) } \abs{ g(y_i) } \mathop{}\!\mathrm{d}^2 x_i \mathop{}\!\mathrm{d}^2 y_i \, \abs{ g(z(\tau)) } \abs{ f(\tau) } \mathop{}\!\mathrm{d} \tau \, \abs{ g(y_n) } \mathop{}\!\mathrm{d}^2 y_n \eqend{,} \raisetag{2.4em}
\end{splitequation}
where $\hat{\vec{x}} = \{x_1,\ldots,x_{n-1},z(\tau)\}$. However, after the shift $n \to n+1$ this is now of the same form as the bounds~\eqref{eq:thm_renorm_vertex_bound} for the vertex operators in~\eqref{eq:theta_00_def}, which are given by~\eqref{eq:thm_renorm_vertex_bound2}, and so also the local terms are suitably bounded, taking into account that there $n-\ell + n-(k-\ell) = 2n-k \leq 2n$ of them~\eqref{eq:thm_renorm_qed_dhdh_local}.
Finally, we need to consider terms with products of two Hadamard parametrices at two different points, for which we still use~\eqref{eq:thm_renorm_qed_hhder}. However, now we use that~\cite[Eq.~(198)]{froebcadamuro2022}
\begin{equation}
\partial_{u(z)} H^\text{F}(z,x) \partial_{u(z)} H^\text{F}(z,y) = \frac{1}{32 \pi^2} \partial_{u(z)}^2 \ln\left[ \mu^2 u(z,a)^2 \right]
\end{equation}
for some point $a$ such that $\min(u(x),u(y)) \leq u(a) \leq \max(u(x),u(y))$, and the analogous result for $v$ or for the product of two Wightman parametrices. We can then convert the $u(z)$ and $v(z)$ derivatives into $\tau$ derivatives using~\eqref{eq:dudv_dtau}, and obtain
\begin{splitequation}
&\int \left. \frac{1}{2} \left( v^\mu v^\nu + w^\mu w^\nu \right) \partial^z_\mu H(x,z) \partial^z_\nu H(y,z) \right\rvert_{z = z(\tau)} f(\tau) \mathop{}\!\mathrm{d} \tau \\
&= \frac{1}{32 \pi^2} \int \Big[ \partial_\tau^2 \ln\left[ \mu^2 u(z(\tau),a)^2 \right] - \partial_\tau \ln(v^0-v^1) \partial_\tau \ln\left[ \mu^2 u(z(\tau),a)^2 \right] \\
&\qquad+ \partial_\tau^2 \ln\left[ \mu^2 v(z(\tau),a')^2 \right] - \partial_\tau \ln(v^0+v^1) \partial_\tau \ln\left[ \mu^2 v(z(\tau),a')^2 \right] \Big] f(\tau) \mathop{}\!\mathrm{d} \tau \eqend{,}
\end{splitequation}
where $a'$ may be different from $a$. Integrating the $\tau$ derivatives by parts and changing integration variables to either $u(z(\tau))$ or $v(z(\tau))$ depending on the argument of the logarithm [using~\eqref{eq:dudv_dtau}], this gives
\begin{splitequation}
&\int \left. \frac{1}{2} \left( v^\mu v^\nu + w^\mu w^\nu \right) \partial^z_\mu H(x,z) \partial^z_\nu H(y,z) \right\rvert_{z = z(\tau)} f(\tau) \mathop{}\!\mathrm{d} \tau \\
&= \frac{1}{32 \pi^2} \int \frac{1}{v^0 - v^1} \Big[ f''(\tau) + \partial_\tau \left[ f(\tau) \partial_\tau \ln(v^0-v^1) \right] \Big] \ln\left[ \mu^2 u(z(\tau),a)^2 \right] \mathop{}\!\mathrm{d} u(z(\tau)) \\
&\quad+ \frac{1}{32 \pi^2} \int \frac{1}{v^0 + v^1} \Big[ f''(\tau) + \partial_\tau \left[ f(\tau) \partial_\tau \ln(v^0+v^1) \right] \Big] \ln\left[ \mu^2 v(z(\tau),a')^2 \right] \mathop{}\!\mathrm{d} v(z(\tau)) \eqend{,}
\end{splitequation}
where now $\tau$ is a function of $u(z(\tau))$ or $v(z(\tau))$. We then estimate the terms apart from the logarithm by their supremum:
\begin{splitequation}
\label{eq:mink_conv_ddf_bound}
&\abs{ \frac{1}{v^0 - v^1} \Big[ f''(\tau) + \partial_\tau \left[ f(\tau) \partial_\tau \ln(v^0-v^1) \right] \Big] } \\
&\leq \frac{1}{1 + \mu^2 u(z(\tau))^2} \sup_\tau \abs{ \frac{1 + \mu^2 u(z(\tau))^2}{v^0 - v^1} \Big[ f''(\tau) + \partial_\tau \left[ f(\tau) \partial_\tau \ln(v^0-v^1) \right] \Big] } \eqend{,}
\end{splitequation}
and then use~\eqref{eq:mink_conv_num_2_int} to perform the integral over $u(z(\tau))$, and the analogous estimate for the integral over $v(z(\tau))$. Finally, using that $\min(u(x),u(y)) \leq u(a) \leq \max(u(x),u(y))$ we estimate the resulting logarithm by
\begin{equation}
\ln\left( 2 + \mu \abs{u(a)} \right) \leq \ln\left( 2 + \mu \abs{u(x)} \right) + \ln\left( 2 + \mu \abs{u(y)} \right) \eqend{,}
\end{equation}
and the logarithms can then be absorbed into the adiabatic cutoff functions $g(x)$. We see at this point that the bound~\eqref{eq:mink_conv_ddf_bound} diverges in the limit where the trajectory becomes light-like, $v^1 \to \pm v^0$. This precludes the direct extension of our result to the null case, but for arbitrary time-like trajectories $z(\tau)$ that we consider in this work we of course have a finite bound.
We then follow the same steps as in the proof of~\cite[Thm.~5]{froebcadamuro2022} and use that there are in total $4 n^2$ products of Feynman or Wightman parametrices and $W$ in the product of derivatives of $\mathcal{G}_{k,\ell,n,m}$~\eqref{eq:calg_def} in $\Theta^{k,\ell,n,m}$~\eqref{eq:theta_00_def}. We can thus bound the contribution of these terms to the expectation value of the energy density at order $n$ by $C n^2 K^n (n!)^{1 + \frac{\beta^2}{4\pi}}$ as required, where the constants $C$ and $K$ depend on $f$, $g$, $\beta$, $W$ and $z(\tau)$. We note however that the dependence on $W$ only comes from the explicit $W$ contained in $\mathcal{G}_{k,\ell,n,m}$, since due to our condition on $W$~\eqref{eq:thm_renorm_wpos_disc} the $W$ dependence in $\mathcal{E}_{k,\ell,n,m}$ is bounded by $1$ analogously to the bounds~\eqref{eq:thm_renorm_vertex_bound2} for the contribution from the vertex operators. Taking all together, we obtain a bound of the same form as for a space-time smearing~\eqref{eq:thm_renorm_bound}, and thus following the same steps to estimate the sums we obtain the bound
\begin{equation}
\abs{ E_\omega(f) } \leq C' \sum_{n=0}^\infty \frac{(n+1)^2 (K')^{2n}}{(n!)^{1-\frac{\beta^2}{4 \pi}}} < \infty
\end{equation}
on the smeared quantum energy density, where $C'$ and $K'$ are new constants depending on $f$, $g$, $\beta$, $W$ and $z(\tau)$. \hfill\squareforqed
\section{Proof of Theorem~\ref{thm_qei} (Quantum energy inequality)}
\label{sec_qei}
To obtain the quantum energy inequality~\eqref{eq:thm_qei} in the interacting theory, we suitably extend methods known from the free theory~\cite{flanagan1997,fewster2000,fewstersmith2008}, namely writing the energy density in point-split form and then separating it into a sum of a positive and a bounded part, with the bounded part being state-independent.
We first note that the energy density when smeared with the square of a test function $f \in \mathcal{S}(\mathbb{R})$ can be written in a point-split form
\begin{equation}
E_\omega(f^2) = \lim_{\Lambda,\epsilon \to 0} \omega^{\Lambda,\epsilon}\left( \mathcal{T}_\text{int}\left[ \hat{T}_{\mu\nu}( v^\mu v^\nu f^2 ) \right] \right) = \int e_{\omega,f}(\tau,\tau) \mathop{}\!\mathrm{d} \tau \eqend{,}
\end{equation}
where
\begin{equation}
\label{eq:qei_pointsplit_e_def}
e_{\omega,f}(\tau,\tau') \equiv v^\mu(\tau) v^\nu(\tau') f(\tau) f(\tau') \lim_{\Lambda,\epsilon \to 0} \omega^{\Lambda,\epsilon}\left( \mathcal{T}_\text{int}\left[ \hat{T}^\parallel_{\mu\nu}( z(\tau), z(\tau') ) \right] \right)
\end{equation}
with the point-split stress tensor
\begin{splitequation}
\label{eq:qei_pointsplit_tmunu_pre}
\hat{T}^\parallel_{\mu\nu}(z,z') &\equiv \partial_\mu \phi(z) \partial_\nu \phi(z') - \frac{1}{2} \eta_{\mu\nu} \partial^\rho \phi(z) \partial_\rho \phi(z') \\
&\quad+ \frac{1}{2} \left( 1 - \frac{\beta^2}{8 \pi} \right) \eta_{\mu\nu} \Big[ g(z) ( V_\beta(z) + V_{-\beta}(z) ) + g(z') ( V_\beta(z') + V_{-\beta}(z') ) \Big] \eqend{.}
\end{splitequation}
Actually, since the time-ordered products involve renormalisation, and in particular the subtraction of the Hadamard parametrix which is singular as the two points approach each other, it is more appropriate to define the point-split value of the interacting time-ordered product of the point-split stress tensor instead of the point-split stress tensor itself. So instead of Lemma~\ref{lemma1}, in particular the interacting time-ordered product $\mathcal{T}_\text{int}\left[ \mathcal{O}_{\mu\nu}(z) \right]$~\eqref{eq:tint_2phi_def}, we define a point-split time-ordered product by
\begin{splitequation}
\label{eq:tint_2phi_pointsplit_def}
&\mathcal{T}_\text{int}\left[ \mathcal{O}^\parallel_{\mu\nu}(z,z') \right] \equiv \sum_{n=0}^\infty \sum_{k=0}^n \sum_{\ell=0}^k \sum_{m=0}^{n-k} \frac{(-1)^k \mathrm{i}^n}{\ell! (k-\ell)! m! (n-k-m)!} \int\dotsi\int \mathcal{E}_{k,\ell,n,m}(\vec{x};\vec{y}) \\
&\quad\times \Bigg[ \beta^2 \partial^z_\mu \mathcal{G}_{k,\ell,n,m}(\vec{x};\vec{y};z) \partial^{z'}_\nu \mathcal{G}_{k,\ell,n,m}(\vec{x};\vec{y};z') \, \mathcal{N}_G\left[ \prod_{i=1}^{\ell+m} V_\beta(x_i) \prod_{j=1}^{n-\ell-m} V_{-\beta}(y_j) \right] \\
&\qquad\quad- \beta \partial^z_\mu \mathcal{G}_{k,\ell,n,m}(\vec{x};\vec{y};z) \, \mathcal{N}_G\left[ \partial_\nu \phi(z') \prod_{i=1}^{\ell+m} V_\beta(x_i) \prod_{j=1}^{n-\ell-m} V_{-\beta}(y_j) \right] \\
&\qquad\quad- \beta \partial^{z'}_\nu \mathcal{G}_{k,\ell,n,m}(\vec{x};\vec{y};z') \, \mathcal{N}_G\left[ \partial_\mu \phi(z) \prod_{i=1}^{\ell+m} V_\beta(x_i) \prod_{j=1}^{n-\ell-m} V_{-\beta}(y_j) \right] \\
&\qquad\quad+ \mathcal{N}_G\left[ \partial_\mu \phi(z) \partial_\nu \phi(z') \prod_{i=1}^{\ell+m} V_\beta(x_i) \prod_{j=1}^{n-\ell-m} V_{-\beta}(y_j) \right] \\
&\qquad\quad+ \partial^z_\mu \partial^{z'}_\nu W(z,z') \, \mathcal{N}_G\left[ \prod_{i=1}^{\ell+m} V_\beta(x_i) \prod_{j=1}^{n-\ell-m} V_{-\beta}(y_j) \right] \Bigg] \\
&\quad\times \left( \frac{\Lambda}{\mu} \right)^\frac{\beta^2 (2(\ell+m)-n)^2}{4 \pi} \prod_{i=1}^{\ell+m} g(x_i) \mathop{}\!\mathrm{d}^2 x_i \prod_{j=1}^{n-\ell-m} g(y_j) \mathop{}\!\mathrm{d}^2 y_j \eqend{,} \raisetag{2.4em}
\end{splitequation}
and the point-split interacting time-ordered product of the stress tensor by
\begin{splitequation}
\label{eq:qei_pointsplit_tmunu}
&\mathcal{T}_\text{int}\left[ \hat{T}^\parallel_{\mu\nu}(z,z') \right] \equiv \mathcal{T}_\text{int}\left[ \mathcal{O}^\parallel_{\mu\nu}(z,z') \right] - \frac{1}{2} \eta_{\mu\nu} \eta^{\rho\sigma} \mathcal{T}_\text{int}\left[ \mathcal{O}^\parallel_{\rho\sigma}(z,z') \right] \\
&\qquad+ \frac{1}{2} \left( 1 - \frac{\beta^2}{8 \pi} \right) \eta_{\mu\nu} \Big[ g(z) ( \mathcal{T}_\text{int}\left[ V_\beta(z) \right] + \mathcal{T}_\text{int}\left[ V_{-\beta}(z) \right] ) + (z \leftrightarrow z') \Big] \eqend{.}
\end{splitequation}
Since the normal-ordered products of operators are continuous functions of the insertion points and $W$ is smooth, almost all terms in the point-split time-ordered product~\eqref{eq:tint_2phi_pointsplit_def} have a finite limit as $z' \to z$. Only the products of derivatives of $\mathcal{G}_{k,\ell,m,n}$~\eqref{eq:calg_def} which contain Hadamard parametrices need renormalisation in the limit where the UV cutoff $\epsilon \to 0$, but as we have seen in the proof of Theorem~\ref{thm_renorm}, the required counterterms are proportional to $\eta_{\mu\nu}$~\eqref{eq:thm_renorm_hmuhnu} and thus cancel in the stress tensor. We thus have
\begin{equation}
\lim_{z' \to z} \mathcal{T}_\text{int}\left[ \hat{T}^\parallel_{\mu\nu}(z,z') \right] = \mathcal{T}_\text{int}\left[ \hat{T}_{\mu\nu}(z) \right]
\end{equation}
with the point-split stress tensor~\eqref{eq:qei_pointsplit_tmunu}, and moreover the point-split quantity $e_{\omega,f}(\tau,\tau')$~\eqref{eq:qei_pointsplit_e_def} is continuous in $\tau$ and $\tau'$. All the hypotheses of~\cite[Lemma~2.12]{fewstersmith2008} are thus fulfilled and we can apply it and obtain
\begin{equation}
\label{eq:qei_e_fourier}
E_\omega(f^2) = \int e_{\omega,f}(\tau,\tau) \mathop{}\!\mathrm{d} \tau = \lim_{\delta \to 0^+} \int \mathrm{e}^{- \delta \xi^2} (\mathcal{F} e_{\omega,f})(-\xi,\xi) \frac{\mathop{}\!\mathrm{d} \xi}{2 \pi} \eqend{,}
\end{equation}
where the Fourier transform is defined by
\begin{equation}
\label{eq:qei_fourier_def}
(\mathcal{F} g)(\xi,\zeta) \equiv \iint \mathrm{e}^{\mathrm{i} \xi \tau + \mathrm{i} \zeta \sigma} g(\tau,\sigma) \mathop{}\!\mathrm{d} \tau \mathop{}\!\mathrm{d} \sigma \eqend{.}
\end{equation}
Since~\cite[Lemma~2.12]{fewstersmith2008} is only applicable to functions of compact support, we temporarily restrict to test functions $f$ of compact support, and obtain the final result for arbitrary Schwartz functions $f$ by a limit argument. Moreover, since $e_{\omega,f}(\tau,\tau')$ is by construction symmetric in $\tau$ and $\tau'$, the integrand in~\eqref{eq:qei_e_fourier} is invariant under the change $\xi \to -\xi$, and so we may restrict the integration to positive $\xi$, obtaining a factor of 2. That is, to prove the QEI~\eqref{eq:thm_qei} we need to show that
\begin{equation}
\label{eq:thm_qei_N}
E^N_\omega(f^2) = 2 \lim_{\delta \to 0^+} \int_0^\infty \mathrm{e}^{- \delta \xi^2} (\mathcal{F} e_{\omega,f_N})(-\xi,\xi) \frac{\mathop{}\!\mathrm{d} \xi}{2 \pi} \geq - K(z,f_N)
\end{equation}
and that $\lim_{N \to \infty} K(z,f_N) < \infty$, where $f_N \in \mathcal{S}_\text{c}(\mathbb{R})$ is a sequence of test functions of compact support converging to $f$.\footnote{We could also restrict to negative $\xi$, but this would ultimately result in a trivial QEI with $\lim_{N \to \infty} K(z,f_N) = \infty$~\cite{fewstersmith2008}.}
However, we will need~\cite[Lemma~2.12]{fewstersmith2008} only for some part of the bound, where it is technically difficult to take the limit $\tau' \to \tau$ directly in position space and then integrate over $\tau$. To show the quantum energy inequality~\eqref{eq:thm_qei}, we first split the energy density $e_{\omega,f}$~\eqref{eq:qei_pointsplit_e_def} into a positive part and a part that we can bound independently of the state. The latter part is furthermore split into several terms, one of which requires the use of~\cite[Lemma~2.12]{fewstersmith2008} in the form~\eqref{eq:thm_qei_N}, while the other ones can be bounded directly. We will see that the positive part is given by
\begin{equation}
\frac{1}{2} \Big[ v^\mu(\tau) v^\nu(\tau') + w^\mu(\tau) w^\nu(\tau') \Big] \, \Big( \mathcal{T}_\text{int}\left[ \partial_\mu \phi(z(\tau)) \right] \Big)^\dagger \star \mathcal{T}_\text{int}\left[ \partial_\nu \phi(z(\tau')) \right] \eqend{,}
\end{equation}
the star product of two interacting time-ordered products of derivatives of the field $\partial_\mu \phi$, for which we could use the result~\eqref{eq:tint_1phi_def}. However, the computation is shortened by inserting the Bogoliubov formula~\eqref{eq:tint_def} and using that
\begin{equation}
\Big( \overline{\mathcal{T}}\left[ \mathcal{O}_1(f_1) \otimes \cdots \otimes \mathcal{O}_k(f_k) \right] \Big)^\dagger = \mathcal{T}\left[ \mathcal{O}_1^\dagger(f_1^*) \otimes \cdots \otimes \mathcal{O}_k^\dagger(f_k^*) \right]
\end{equation}
as well as the star inverse~\eqref{eq:t_starinverse}, such that
\begin{equation}
\Big( \mathcal{T}_\text{int}\left[ \partial_\mu \phi(z) \right] \Big)^\dagger \star \mathcal{T}_\text{int}\left[ \partial_\nu \phi(z') \right] = \overline{\mathcal{T}}\left[ \mathrm{e}_\otimes^{ - \mathrm{i} S_\text{int} } \otimes \partial_\mu \phi(z) \right] \star \mathcal{T}\left[ \mathrm{e}_\otimes^{ \mathrm{i} S_\text{int} } \otimes \partial_\nu \phi(z') \right] \eqend{.}
\end{equation}
A quite long but straightforward computation using the result~\eqref{eq:timeordered_exponential_1phi}, the analogously determined
\begin{splitequation}
\label{eq:antitimeordered_exponential_1phi}
&\overline{\mathcal{T}}\left[ \bigotimes_{j=1}^n V_{\alpha_j}(x_j) \otimes \phi(z) \right] = \exp\left[ - \mathrm{i} \sum_{1 \leq i < j \leq n} \alpha_i \alpha_j H^\text{D}(x_i,x_j) \right] \\
&\qquad\times\Bigg[ \mathcal{N}_G\left[ \phi(z) \prod_{j=1}^n V_{\alpha_j}(x_j) \right] - \sum_{i=1}^n \alpha_i G^\text{D}(x_i,z) \mathcal{N}_G\left[ \prod_{j=1}^n V_{\alpha_j}(x_j) \right] \Bigg] \\
&\qquad\times \exp\left[ - \frac{1}{2} \sum_{i,j=1}^n \alpha_i \alpha_j W(x_i,x_j) \right] \left( \frac{\Lambda}{\mu} \right)^\frac{\left( \sum_{k=1}^n \alpha_k \right)^2}{4 \pi} \eqend{,}
\end{splitequation}
the formula~\eqref{eq:normal_ordering_starproduct} for the star product and the explicit form of the two-point function~\eqref{eq:twopf_hadamard} then gives
\begin{splitequation}
\label{eq:qei_tint_dphi2}
&\Big( \mathcal{T}_\text{int}\left[ \partial_\mu \phi(z) \right] \Big)^\dagger \star \mathcal{T}_\text{int}\left[ \partial_\nu \phi(z') \right] \\
&= \sum_{n=0}^\infty \sum_{k=0}^n \sum_{\ell=0}^k \sum_{m=0}^{n-k} \frac{(-1)^k \mathrm{i}^n}{\ell! (k-\ell)! m! (n-k-m)!} \int\dotsi\int \mathcal{E}_{k,\ell,n,m}(\vec{x};\vec{y}) \\
&\quad\times\Bigg[ \beta^2 \partial^z_\mu \overline{\mathcal{G}}_{k,\ell,n,m}(\vec{x};\vec{y};z) \partial^{z'}_\nu \mathcal{G}_{k,\ell,n,m}(\vec{x};\vec{y};z') \, \mathcal{N}_G\left[ \prod_{i=1}^{\ell+m} V_\beta(x_i) \prod_{j=1}^{n-\ell-m} V_{-\beta}(y_j) \right] \\
&\qquad\quad- \beta \partial^z_\mu \overline{\mathcal{G}}_{k,\ell,n,m}(\vec{x};\vec{y};z) \, \mathcal{N}_G\left[ \partial_\nu \phi(z') \prod_{i=1}^{\ell+m} V_\beta(x_i) \prod_{j=1}^{n-\ell-m} V_{-\beta}(y_j) \right] \\
&\qquad\quad- \beta \partial^{z'}_\nu \mathcal{G}_{k,\ell,n,m}(\vec{x};\vec{y};z') \, \mathcal{N}_G\left[ \partial_\mu \phi(z) \prod_{i=1}^{\ell+m} V_\beta(x_i) \prod_{j=1}^{n-\ell-m} V_{-\beta}(y_j) \right] \\
&\qquad\quad+ \mathcal{N}_G\left[ \partial_\mu \phi(z) \partial_\nu \phi(z') \prod_{i=1}^{\ell+m} V_\beta(x_i) \prod_{j=1}^{n-\ell-m} V_{-\beta}(y_j) \right] \\
&\qquad\quad+ \mathrm{i} \partial^z_\mu \partial^{z'}_\nu G^+(z,z') \, \mathcal{N}_G\left[ \prod_{i=1}^{\ell+m} V_\beta(x_i) \prod_{j=1}^{n-\ell-m} V_{-\beta}(y_j) \right] \Bigg] \\
&\quad\times \left( \frac{\Lambda}{\mu} \right)^\frac{\beta^2 (2(\ell+m)-n)^2}{4 \pi} \prod_{i=1}^{\ell+m} g(x_i) \mathop{}\!\mathrm{d}^2 x_i \prod_{j=1}^{n-\ell-m} g(y_j) \mathop{}\!\mathrm{d}^2 y_j \eqend{,} \raisetag{2.3em}
\end{splitequation}
where we used that the (anti-)time-ordered products are symmetric in their arguments, and that there are $k!/[ \ell! (k-\ell)! ]$ ways to choose $\ell$ vertex operators $V_\beta$ with positive sign from a total of $k$ ones $V_{\pm \beta}$ with either sign. We also renamed the insertion points of the $V_{-\beta}$ operators to $y_j$, and recall that $\mathcal{E}_{k,\ell,n,m}$ and $\mathcal{G}_{k,\ell,n,m}$ are defined by~\eqref{eq:cale_def} and~\eqref{eq:calg_def}, while
\begin{splitequation}
\label{eq:barcalg_def}
\overline{\mathcal{G}}_{k,\ell,n,m}(\vec{x};\vec{y};z) &\equiv \sum_{i=1}^\ell H^\text{D}(z,x_i) + \sum_{i=\ell+1}^{\ell+m} H^+(z,x_i) - \sum_{j=1}^{k-\ell} H^\text{D}(z,y_j) \\
&\quad- \sum_{j=k-\ell+1}^{n-\ell-m} H^+(z,y_j) - \mathrm{i} \sum_{i=1}^{\ell+m} W(z,x_i) + \mathrm{i} \sum_{j=1}^{n-\ell-m} W(z,y_j) \eqend{.}
\end{splitequation}
Contracting with $v^\mu v^\nu$ and comparing with~\eqref{eq:tint_2phi_pointsplit_def}, we see that the point-split quantum energy density~\eqref{eq:qei_pointsplit_e_def} with the interacting time-ordered product of the stress tensor~\eqref{eq:qei_pointsplit_tmunu} can be written as
\begin{splitequation}
\label{eq:qei_e_positivebounded}
e_{\omega,f}(\tau,\tau') &= \frac{1}{2} f(\tau) f(\tau') \Big[ v^\mu(\tau) v^\nu(\tau') + w^\mu(\tau) w^\nu(\tau') \Big] \\
&\qquad\times \lim_{\Lambda,\epsilon \to 0} \omega^{\Lambda,\epsilon}\left( \Big( \mathcal{T}_\text{int}\left[ \partial_\mu \phi(z) \right] \Big)^\dagger \star \mathcal{T}_\text{int}\left[ \partial_\nu \phi(z') \right] \right) \\
&\quad+ \frac{1}{2} f(\tau) f(\tau') \Big[ v^\mu(\tau) v^\nu(\tau') + w^\mu(\tau) w^\nu(\tau') \Big] \\
&\qquad\times \sum_{n=0}^\infty \sum_{k=0}^n \sum_{\ell=0}^k \sum_{m=0}^{n-k} \frac{(-1)^k \mathrm{i}^n \delta_{2(\ell+m),n}}{\ell! (k-\ell)! m! (n-k-m)!} \int\dotsi\int \mathcal{E}_{k,\ell,n,m}(\vec{x};\vec{y}) \\
&\quad\times \bigg[ \beta^2 \partial^z_\mu \left[ \mathcal{G}_{k,\ell,n,m}(\vec{x};\vec{y};z) - \overline{\mathcal{G}}_{k,\ell,n,m}(\vec{x};\vec{y};z) \right] \partial^{z'}_\nu \mathcal{G}_{k,\ell,n,m}(\vec{x};\vec{y};z') \\
&\qquad\quad- \mathrm{i} \partial^z_\mu \partial^{z'}_\nu H^+(z,z') \bigg] \prod_{i=1}^{\ell+m} g(x_i) \mathop{}\!\mathrm{d}^2 x_i \prod_{j=1}^{n-\ell-m} g(y_j) \mathop{}\!\mathrm{d}^2 y_j \\
&\quad- \frac{1}{2} \left( 1 - \frac{\beta^2}{8 \pi} \right) f(\tau) f(\tau') \Big[ g(z(\tau)) \lim_{\Lambda,\epsilon \to 0} \\
&\qquad\times \omega^{\Lambda,\epsilon}\left( \mathcal{T}_\text{int}\left[ V_\beta(z(\tau)) \right] + \mathcal{T}_\text{int}\left[ V_{-\beta}(z(\tau)) \right] \right) + (\tau \leftrightarrow \tau') \Big] \eqend{,} \raisetag{1.6em}
\end{splitequation}
where after taking the derivatives we set $z = z(\tau)$ and $z' = z(\tau')$. Here, we used again that only neutral configurations remain in the limit $\Lambda \to 0$. The first term is the positive part, which would be obvious from the positivity of the state $\omega$, except for the IR divergence of the state. However, we will show below that it is indeed positive. The second and third term comprise the bounded part, and we will show below a state-independent bound for it.
Let us first consider the positivity of the first term of~\eqref{eq:qei_e_positivebounded}. Since it is of the form $\omega^{\Lambda,\epsilon}(A^\dagger \star A)$ and $\omega$ is a quasi-free state, it is positive if the two-point function~\eqref{eq:twopf_hadamard} of $\omega$ is positive definite. However, $\omega$ is indefinite since positivity of the two-point function~\eqref{eq:twopf_hadamard} holds only for test functions with vanishing mean~\eqref{eq:thm_renorm_pos_hplusw}: for the Hadamard parametrix $H^+$ by direct computation, and for the state-dependent part $W$ by our condition of conditional positive semidefiniteness. Therefore, to show positivity of~\eqref{eq:qei_e_positivebounded} we need to construct an auxiliary state $\tilde{\omega}$, whose two-point function has the same behaviour as $\omega$ in the limit $\Lambda \to 0$ (i.e., the difference of their two-point functions converges to $0$), but is positive for all $\Lambda > 0$. In principle, this is the massive extension of our massless state, but its construction is not fully trivial. We first use that as $\Lambda \to 0$ we have~\cite[Eq.~(10.31.2)]{dlmf}
\begin{equation}
\@ifstar{\bessel@s}{\bessel@n}*{K}{0}{2 \Lambda \mathrm{e}^{-\gamma} \sqrt{ ( \epsilon + \mathrm{i} u ) ( \epsilon + \mathrm{i} v ) } } = - \frac{1}{2} \ln \left[ \Lambda^2 ( \epsilon + \mathrm{i} u ) ( \epsilon + \mathrm{i} v ) \right] + \bigo{\Lambda^2 \ln \Lambda}
\end{equation}
for the second modified Bessel function $\mathrm{K}$, and using the integral~\cite[Eq.~(10.32.10)]{dlmf} we compute that
\begin{equation}
\@ifstar{\bessel@s}{\bessel@n}*{K}{0}{M \sqrt{ ( \epsilon + \mathrm{i} u ) ( \epsilon + \mathrm{i} v ) } } = \int_0^\infty \exp\left[ - \frac{a}{2} M ( \epsilon + \mathrm{i} u ) - \frac{M}{2 a} ( \epsilon + \mathrm{i} v ) \right] \frac{1}{2 a} \mathop{}\!\mathrm{d} a
\end{equation}
for $M > 0$. We further compute
\begin{splitequation}
\partial^2 \@ifstar{\bessel@s}{\bessel@n}*{K}{0}{M \sqrt{ ( \epsilon + \mathrm{i} u ) ( \epsilon + \mathrm{i} v ) } } &= - 4 \partial_u \partial_v \@ifstar{\bessel@s}{\bessel@n}*{K}{0}{M \sqrt{ ( \epsilon + \mathrm{i} u ) ( \epsilon + \mathrm{i} v ) } } \\
&= M^2 \int_0^\infty \exp\left[ - \frac{a}{2} M ( \epsilon + \mathrm{i} u ) - \frac{M}{2 a} ( \epsilon + \mathrm{i} v ) \right] \frac{1}{2 a} \mathop{}\!\mathrm{d} a \\
&= M^2 \@ifstar{\bessel@s}{\bessel@n}*{K}{0}{M \sqrt{ ( \epsilon + \mathrm{i} u ) ( \epsilon + \mathrm{i} v ) } } \eqend{,} \raisetag{1.5em}
\end{splitequation}
using that the $a$ integral is absolutely convergent for $\epsilon > 0$ such that we can differentiate under the integral sign. It follows that
\begin{equation}
\left( \partial^2 - 4 \Lambda^2 \mathrm{e}^{-2\gamma} \right) \@ifstar{\bessel@s}{\bessel@n}*{K}{0}{2 \Lambda \mathrm{e}^{-\gamma} \sqrt{ ( \epsilon + \mathrm{i} u ) ( \epsilon + \mathrm{i} v ) } } = 0 \eqend{,}
\end{equation}
and we see that we have obtained a suitable massive extension of our massless Hadamard parametrix $H^+(x,y)$.
For the state-dependent part $W(x,y)$, the construction is more involved. For this, we use that in two dimensions solutions of the massless Klein--Gordon equation split into functions that depend on only one light-cone coordinate: the general solution of $\partial^2 f(x) = 0$ is $f(x) = f_u(u(x)) + f_v(v(x))$ since the d'Alembertian factorises according to $\partial^2 = - 4 \partial_u \partial_v$. Since $\partial_x^2 W(x,y) = \partial_y^2 W(x,y) = 0$, this generalises to
\begin{splitequation}
W(x,y) &= W_{uu}(u(x),u(y)) + W_{uv}(u(x),v(y)) \\
&\quad+ W_{vu}(v(x),u(y)) + W_{vv}(v(x),v(y)) \eqend{,}
\end{splitequation}
where $W_{uu}$, $W_{uv}$, $W_{vu}$ and $W_{vv}$ are smooth functions. Because of the symmetry $W(x,y) = W(y,x)$, $W_{uu}$ and $W_{vv}$ are symmetric in their arguments, while $W_{vu}(u,v) = W_{uv}(v,u)$. Passing to Fourier space in each variable, we obtain
\begin{equation}
W_{uv}(u(x),v(y)) = \int \tilde{W}_{uv}(p,q) \, \mathrm{e}^{\mathrm{i} p u(x) + \mathrm{i} q v(y)} \frac{\mathop{}\!\mathrm{d} p \mathop{}\!\mathrm{d} q}{(2\pi)^2}
\end{equation}
and the analogous expressions for $W_{uu}$, $W_{vu}$ and $W_{vv}$, where the symmetry leads to $\tilde{W}_{vu}(p,q) = \tilde{W}_{uv}(q,p)$. The conditionally positive semidefiniteness~\eqref{eq:thm_renorm_wpos_cont} of $W$ reads in Fourier space
\begin{splitequation}
\label{eq:qei_w_pos}
&\iint \Big[ \tilde{W}_{uu}(p,q) \, \tilde{f}(-p,0) \tilde{f}^*(q,0) + \tilde{W}_{vu}(p,q) \, \tilde{f}(0,-p) \tilde{f}^*(q,0) \\
&\qquad+ \tilde{W}_{uv}(p,q) \, \tilde{f}(-p,0) \tilde{f}^*(0,q) + \tilde{W}_{vv}(p,q) \, \tilde{f}(0,-p) \tilde{f}^*(0,q) \Big] \frac{\mathop{}\!\mathrm{d} p \mathop{}\!\mathrm{d} q}{(2\pi)^2} \geq 0 \raisetag{4.5em}
\end{splitequation}
for all $\tilde{f} \in \mathcal{S}(\mathbb{R}^2)$ with $\tilde{f}(0,0) = 0$, where
\begin{equation}
\tilde{f}(p,q) \equiv \iint f(u,v) \mathrm{e}^{- \mathrm{i} p u - \mathrm{i} q v} \mathop{}\!\mathrm{d} u \mathop{}\!\mathrm{d} v
\end{equation}
is the Fourier transform of the test function $f$. Since $f$ is arbitrary, the condition~\eqref{eq:qei_w_pos} actually separates in individual conditions for $W_{uu}$, $W_{vv}$ and $W_{uv}$:
\begin{equations}[eq:qei_wuuvv_pos]
\iint \tilde{W}_{uu}(p,q) \, \tilde{g}(-p) \tilde{g}^*(q) \frac{\mathop{}\!\mathrm{d} p \mathop{}\!\mathrm{d} q}{(2\pi)^2} &\geq 0 \eqend{,} \\
\iint \tilde{W}_{vv}(p,q) \, \tilde{h}(-p) \tilde{h}^*(q) \frac{\mathop{}\!\mathrm{d} p \mathop{}\!\mathrm{d} q}{(2\pi)^2} &\geq 0 \eqend{,} \\
\iint \tilde{W}_{uv}(p,q) \Big[ \tilde{h}(-q) \tilde{g}^*(p) + \tilde{g}(-p) \tilde{h}^*(q) \Big] \frac{\mathop{}\!\mathrm{d} p \mathop{}\!\mathrm{d} q}{(2\pi)^2} &\geq 0 \eqend{,}
\end{equations}
where now $g,h \in \mathcal{S}(\mathbb{R})$ with $\tilde{g}(0) = \tilde{h}(0) = 0$.\footnote{For example, the condition for $\tilde{W}_{uu}$ is obtained by taking $\tilde{f}(p,q)$ such that $\tilde{f}(0,q) = 0$ and defining $\tilde{g}(p) \equiv \tilde{f}(p,0)$.} Since by assumption $W$ and thus its components $W_{uu}$ and so on are smooth functions, the Fourier coefficients $\tilde{W}_{uu}$ and so on have fast decay at infinity. We can thus define
\begin{splitequation}
W_M(x,y) &\equiv \iint \tilde{W}_{uu}(p,q) \, \mathrm{e}^{\mathrm{i} p u(x) + \mathrm{i} \frac{M^2}{4 p} v(x)} \mathrm{e}^{\mathrm{i} q u(y) + \mathrm{i} \frac{M^2}{4 q} v(y)} \frac{\mathop{}\!\mathrm{d} p \mathop{}\!\mathrm{d} q}{(2\pi)^2} \\
&\quad+ \iint \tilde{W}_{vu}(p,q) \, \mathrm{e}^{\mathrm{i} \frac{M^2}{4 p} u(x) + \mathrm{i} p v(x)} \mathrm{e}^{\mathrm{i} q u(y) + \mathrm{i} \frac{M^2}{4 q} v(y)} \frac{\mathop{}\!\mathrm{d} p \mathop{}\!\mathrm{d} q}{(2\pi)^2} \\
&\quad+ \iint \tilde{W}_{uv}(p,q) \, \mathrm{e}^{\mathrm{i} p u(x) + \mathrm{i} \frac{M^2}{4 p} v(x)} \mathrm{e}^{\mathrm{i} \frac{M^2}{4 q} u(y) + \mathrm{i} q v(y)} \frac{\mathop{}\!\mathrm{d} p \mathop{}\!\mathrm{d} q}{(2\pi)^2} \\
&\quad+ \iint \tilde{W}_{vv}(p,q) \, \mathrm{e}^{\mathrm{i} \frac{M^2}{4 p} u(x) + \mathrm{i} p v(x)} \mathrm{e}^{\mathrm{i} \frac{M^2}{4 q} u(y) + \mathrm{i} q v(y)} \frac{\mathop{}\!\mathrm{d} p \mathop{}\!\mathrm{d} q}{(2\pi)^2} \eqend{,}
\end{splitequation}
where the integrals are absolutely convergent and which by construction fulfills
\begin{equation}
\left( \partial_x^2 - M^2 \right) W(x,y) = \left( \partial_y^2 - M^2 \right) W(x,y) = 0 \eqend{.}
\end{equation}
Note that in the case where $W(x,y)$ is the Fourier transform of a positive measure, we have $W_{uv} = W_{vu} = 0$ and $\tilde{W}_{uu}(p,q) \mathop{}\!\mathrm{d} p \mathop{}\!\mathrm{d} q = \delta(p+q) \mathop{}\!\mathrm{d} \mu_{W,u}(p)$, $\tilde{W}_{vv}(p,q) \mathop{}\!\mathrm{d} p \mathop{}\!\mathrm{d} q = \delta(p+q) \mathop{}\!\mathrm{d} \mu_{W,v}(p)$, and the formulas above need to be adjusted accordingly. For example, this is the case for the state~\eqref{eq:thm_renorm_w_example}, where we compute
\begin{splitequation}
\tilde{W}_{uu}(p,q) \mathop{}\!\mathrm{d} p \mathop{}\!\mathrm{d} q &= \tilde{W}_{vv}(p,q) \mathop{}\!\mathrm{d} p \mathop{}\!\mathrm{d} q \\
&= \delta(p+q) \Theta\left( \abs{p} \in [E_0, E_1] \right) \frac{\pi \mathrm{e}^{- \beta \abs{p}}}{\abs{p} \left( 1 - \mathrm{e}^{- \beta \abs{p}} \right)} \mathop{}\!\mathrm{d} p \mathop{}\!\mathrm{d} q
\end{splitequation}
and
\begin{splitequation}
W_M(x,y) &= \frac{1}{4 \pi} \int \Theta\left( \abs{p} \in [E_0, E_1] \right) \frac{\mathrm{e}^{- \beta \abs{p}}}{\abs{p} \left( 1 - \mathrm{e}^{- \beta \abs{p}} \right)} \\
&\qquad\times \left[ \mathrm{e}^{\mathrm{i} p u(x,y) + \mathrm{i} \frac{M^2}{4 p} v(x,y)} + \mathrm{e}^{\mathrm{i} \frac{M^2}{4 p} u(x,y) + \mathrm{i} p v(x,y)} \right] \mathop{}\!\mathrm{d} p \eqend{.}
\end{splitequation}
We then define the quasi-free state $\tilde{\omega}^{\Lambda,\epsilon}$ with two-point function
\begin{equation}
\tilde{\omega}^{\Lambda,\epsilon}\left( \phi(x) \star \phi(y) \right) = \frac{1}{2 \pi} \@ifstar{\bessel@s}{\bessel@n}*{K}{0}{2 \Lambda \mathrm{e}^{-\gamma} \sqrt{ ( \epsilon + \mathrm{i} u(x,y) ) ( \epsilon + \mathrm{i} v(x,y) ) } } + W_{2 \Lambda \mathrm{e}^{-\gamma}}(x,y) \eqend{,}
\end{equation}
which is the two-point function in the massive theory with mass $M = 2 \Lambda \mathrm{e}^{-\gamma}$. For the smeared two-point function, it follows that
\begin{splitequation}
\label{eq:qei_omega_mass}
&\iint \tilde{\omega}^{\Lambda,\epsilon}\left( \phi(x) \star \phi(y) \right) f(x) f^*(y) \mathop{}\!\mathrm{d}^2 x \mathop{}\!\mathrm{d}^2 y \\
&= \frac{1}{2 \pi} \int_0^\infty \abs{ \int f(x) \exp\left[ - \mathrm{i} a \Lambda \mathrm{e}^{-\gamma} (x^0-x^1) - \mathrm{i} \frac{\Lambda \mathrm{e}^{-\gamma}}{a} (x^0+x^1) \right] \mathop{}\!\mathrm{d}^2 x }^2 \\
&\qquad\times \exp\left[ - a \Lambda \mathrm{e}^{-\gamma} \epsilon - \frac{\Lambda \mathrm{e}^{-\gamma}}{a} \epsilon \right] \frac{1}{2 a} \mathop{}\!\mathrm{d} a + \iint W_{2 \Lambda \mathrm{e}^{-\gamma}}(x,y) f(x) f^*(y) \mathop{}\!\mathrm{d}^2 x \mathop{}\!\mathrm{d}^2 y
\end{splitequation}
for an arbitrary test function $f \in \mathcal{S}(\mathbb{R}^2)$, where we could interchange the integral over $a$ with the ones over $x$ and $y$ for all $\Lambda,\epsilon > 0$ because of absolute convergence, and the first term is obviously positive. To show positivity for the second term, we pass to Fourier space and obtain
\begin{splitequation}
\label{eq:qei_wmass_ft}
&\iint W_M(x,y) f(x) f^*(y) \mathop{}\!\mathrm{d}^2 x \mathop{}\!\mathrm{d}^2 y \\
&= \iint \tilde{W}_{uu}(p,q) \, \tilde{g}(-p) \tilde{g}^*(q) \frac{\mathop{}\!\mathrm{d} p \mathop{}\!\mathrm{d} q}{(2\pi)^2} + \iint \tilde{W}_{vv}(p,q) \, \tilde{h}(-p) \tilde{h}^*(q) \frac{\mathop{}\!\mathrm{d} p \mathop{}\!\mathrm{d} q}{(2\pi)^2} \\
&\quad+ \iint \tilde{W}_{uv}(p,q) \, \left[ \tilde{g}(-p) \tilde{h}^*(q) + \tilde{g}^*(p) \tilde{h}(-q) \right] \frac{\mathop{}\!\mathrm{d} p \mathop{}\!\mathrm{d} q}{(2\pi)^2} \eqend{,}
\end{splitequation}
where
\begin{equation}
\tilde{g}(p) \equiv \tilde{f}\left( p, \frac{M^2}{4 p} \right) = \iint f(u,v) \mathrm{e}^{- \mathrm{i} p u - \mathrm{i} \frac{M^2}{4 p} v} \mathop{}\!\mathrm{d} u \mathop{}\!\mathrm{d} v \eqend{,} \quad \tilde{h}(p) \equiv \tilde{f}\left( \frac{M^2}{4 p}, p \right) \eqend{.}
\end{equation}
Since the Fourier transform is an isomorphism between Schwartz spaces, it follows that $\tilde{f} \in \mathcal{S}(\mathbb{R}^2)$, and in particular for all $M > 0$ we have
\begin{equation}
\lim_{p \to 0} \tilde{g}(p) = \lim_{p \to 0} \tilde{h}(p) = 0 \eqend{.}
\end{equation}
Using the conditions~\eqref{eq:qei_wuuvv_pos}, we see that each term in~\eqref{eq:qei_wmass_ft} is non-negative, and hence we have $\iint W_M(x,y) f(x) f^*(y) \mathop{}\!\mathrm{d}^2 x \mathop{}\!\mathrm{d}^2 y \geq 0$ for all $f \in \mathcal{S}(\mathbb{R}^2)$, so that~\eqref{eq:qei_omega_mass}
\begin{equation}
\label{eq:qei_omega_pos}
\iint \tilde{\omega}^{\Lambda,\epsilon}\left( \phi(x) \star \phi(y) \right) f(x) f^*(y) \mathop{}\!\mathrm{d}^2 x \mathop{}\!\mathrm{d}^2 y \geq 0
\end{equation}
for all $\Lambda, \epsilon > 0$. We then obtain that
\begin{equation}
\iint f(\tau) f(\tau') v^\mu(\tau) v^\nu(\tau') \, \tilde{\omega}^{\Lambda,\epsilon}\left( \Big( \mathcal{T}_\text{int}\left[ \partial_\mu \phi(z) \right] \Big)^\dagger \star \mathcal{T}_\text{int}\left[ \partial_\nu \phi(z') \right] \right) \mathop{}\!\mathrm{d} \tau \mathop{}\!\mathrm{d} \tau' \geq 0
\end{equation}
for all $\Lambda, \epsilon > 0$, and thus also in the limit where the regulators vanish, and the same with $v^\mu v^\nu$ replaced by $w^\mu w^\nu$. Since by construction
\begin{equation}
\label{eq:qei_limit_state}
\lim_{\Lambda \to 0} \omega^{\Lambda,\epsilon}(F) = \lim_{\Lambda \to 0} \tilde{\omega}^{\Lambda,\epsilon}(F)
\end{equation}
whenever the limit is finite, it follows that the first term of~\eqref{eq:qei_e_positivebounded} is positive, and integrating over $\tau$ we obtain a positive contribution to $E_\omega(f^2)$~\eqref{eq:qei_e_fourier}.
For the remaining terms in~\eqref{eq:qei_e_positivebounded}, we want to show that they are bounded. We start with the Hadamard parametrix $H^+(z,z')$, that is the series
\begin{splitequation}
\label{eq:qei_h_sum}
&- \frac{\mathrm{i}}{2} f(\tau) f(\tau') \Big[ v^\mu(\tau) v^\nu(\tau') + w^\mu(\tau) w^\nu(\tau') \Big] \\
&\quad\times \sum_{n=0}^\infty \sum_{k=0}^n \sum_{\ell=0}^k \sum_{m=0}^{n-k} \frac{(-1)^k \mathrm{i}^n \delta_{2(\ell+m),n}}{\ell! (k-\ell)! m! (n-k-m)!} \int\dotsi\int \mathcal{E}_{k,\ell,n,m}(\vec{x};\vec{y}) \\
&\quad\times \partial^z_\mu \partial^{z'}_\nu H^+(z,z') \prod_{i=1}^{\ell+m} g(x_i) \mathop{}\!\mathrm{d}^2 x_i \prod_{j=1}^{n-\ell-m} g(y_j) \mathop{}\!\mathrm{d}^2 y_j \eqend{.}
\end{splitequation}
Since the sum is independent of $z$ and $z'$, we can simplify it further. Namely, from the Sine--Gordon interaction~\eqref{eq:sint}, the results~\eqref{eq:timeordered_exponential} and~\eqref{eq:antitimeordered_exponential} for the (anti-)time-ordered products of vertex operators, the formula~\eqref{eq:normal_ordering_starproduct} for the star product and the explicit form of the two-point function~\eqref{eq:twopf_hadamard} we obtain
\begin{splitequation}
&\overline{\mathcal{T}}\left[ \mathrm{e}_\otimes^{- \mathrm{i} S_\text{int}} \right] \star \mathcal{T}\left[ \mathrm{e}_\otimes^{\mathrm{i} S_\text{int}} \right] = \sum_{n=0}^\infty \sum_{k=0}^n \frac{(-1)^k \mathrm{i}^n}{k! (n-k)!} \int\dotsi\int \sum_{\sigma_j = \pm 1} \overline{\mathcal{T}}\left[ \bigotimes_{i=1}^k V_{\sigma_i \beta}(x_i) \right] \\
&\hspace{10em} \star \mathcal{T}\left[ \bigotimes_{j=k+1}^n V_{\sigma_j \beta}(x_j) \right] \prod_{j=1}^n g(x_j) \mathop{}\!\mathrm{d}^2 x_j \\
&\quad= \sum_{n=0}^\infty \sum_{k=0}^n \sum_{\ell=0}^k \sum_{m=0}^{n-k} \frac{(-1)^k \mathrm{i}^n}{\ell! (k-\ell)! m! (n-k-m)!} \int\dotsi\int \mathcal{E}_{k,\ell,n,m}(\vec{x};\vec{y}) \\
&\qquad\times \mathcal{N}_G\left[ \prod_{i=1}^{\ell+m} V_\beta(x_i) \prod_{j=1}^{n-\ell-m} V_{-\beta}(y_j) \right] \\
&\qquad\times \left( \frac{\Lambda}{\mu} \right)^\frac{\beta^2 (2(\ell+m)-n)^2}{4 \pi} \prod_{i=1}^{\ell+m} g(x_i) \mathop{}\!\mathrm{d}^2 x_i \prod_{j=1}^{n-\ell-m} g(y_j) \mathop{}\!\mathrm{d}^2 y_j \raisetag{2.3em}
\end{splitequation}
by a long but straightforward computation. As before, we used that the (anti-)time-ordered products are symmetric in their arguments, that there are $k!/[ \ell! (k-\ell)! ]$ ways to choose $\ell$ vertex operators $V_\beta$ with positive sign from a total of $k$ ones $V_{\pm \beta}$ with either sign, and renamed the insertion points of the $V_{-\beta}$ operators to $y_j$. Taking the expectation value in the state $\omega^{\Lambda,\epsilon}$ and the limit $\Lambda \to 0$, only terms with $n = 2(\ell+m)$ give a non-vanishing contribution, and renaming $n \to 2n$ we obtain
\begin{splitequation}
\label{eq:qei_1_expansion}
&\sum_{n=0}^\infty \sum_{k=0}^{2n} \sum_{\ell=\max(0,k-n)}^{\min(k,n)} \frac{(-1)^{k+n}}{\ell! (k-\ell)! (n-\ell)! (n-k+\ell)!} \int\dotsi\int \mathcal{E}_{k,\ell,2n,n-\ell}(\vec{x};\vec{y}) \\
&\quad\times \prod_{i=1}^n g(x_i) \mathop{}\!\mathrm{d}^2 x_i \prod_{j=1}^n g(y_j) \mathop{}\!\mathrm{d}^2 y_j = \omega\left( \overline{\mathcal{T}}\left[ \mathrm{e}_\otimes^{- \mathrm{i} S_\text{int}} \right] \star \mathcal{T}\left[ \mathrm{e}_\otimes^{\mathrm{i} S_\text{int}} \right] \right) = 1 \eqend{,} \raisetag{2.4em}
\end{splitequation}
where the last equality follows because $\overline{\mathcal{T}}\left[ \mathrm{e}_\otimes^{- \mathrm{i} S_\text{int}} \right] = \mathcal{T}\left[ \mathrm{e}_\otimes^{\mathrm{i} S_\text{int}} \right]^{\star (-1)}$~\eqref{eq:t_starinverse}. Since $g$ is arbitrary, we conclude that
\begin{equation}
\label{eq:qei_vanishing_sum}
\mathbb{S}_{\vec{x};\vec{y}} \sum_{k=0}^{2n} \sum_{\ell=\max(0,k-n)}^{\min(k,n)} \frac{(-1)^{k+n}}{\ell! (k-\ell)! (n-\ell)! (n-k+\ell)!} \mathcal{E}_{k,\ell,2n,n-\ell}(\vec{x};\vec{y}) = 0
\end{equation}
for all $n \geq 1$, where $\mathbb{S}_{\vec{x};\vec{y}}$ indicates symmetrisation of the $x_i$ and $y_j$ among themselves; since $\mathcal{E}_{k,k-\ell,2n,n-k+\ell}(\vec{y};\vec{x}) = \mathcal{E}_{k,\ell,2n,n-\ell}(\vec{x};\vec{y})$ as well as $k-\max(0,k-n) = \min(k,n)$ and $k-\min(k,n) = \max(0,k-n)$, the change of summation index $\ell \to k-\ell$ shows that the sum is already symmetric under the interchange of $x_i$ and $y_i$. That is, the sum in~\eqref{eq:qei_h_sum} collapses to the single term with $n = 0$, which is the free-theory contribution
\begin{splitequation}
\label{eq:qei_h}
h(\tau,\tau') &\equiv - \frac{\mathrm{i}}{2} f(\tau) f(\tau') \Big[ v^\mu(\tau) v^\nu(\tau') + w^\mu(\tau) w^\nu(\tau') \Big] \\
&\qquad\times \partial^z_\mu \partial^{z'}_\nu H^+(z,z') \Big\vert_{z = z(\tau), z' = z(\tau')} \eqend{.}
\end{splitequation}
To bound the contribution of~\eqref{eq:qei_h} to the total energy density $E_\omega(f^2)$, we compare the actual trajectory $z^\mu(\tau)$ with a stationary one $z^\mu = \delta^\mu_0 \tau$, and bound each part separately. For this, we use that since the curve $z^\mu(\tau)$ is future-directed we have~\eqref{eq:dudv_dtau}
\begin{equation}
\label{eq:eqi_du_dtau}
\partial_\tau u(z(\tau)) = \sqrt{ 1 + [ v^1(\tau) ]^2 } - v^1(\tau) > 0 \eqend{,}
\end{equation}
and therefore we can view $\tau$ as function of $u$ using the inverse function theorem. Following~\cite{flanagan1997}, we then define
\begin{splitequation}
\label{eq:qei_deltau_def}
\Delta^u_{\mu\nu}(\tau,\tau') &\equiv \partial_\mu^z \partial_\nu^{z'} \Big[ \ln\left( \epsilon + \mathrm{i} u(z,z') \right) - \ln\left[ \epsilon + \mathrm{i} \left( \tau(u(z)) - \tau(u(z')) \right) \right] \Big]_{z = z(\tau), z' = z(\tau')} \\
&= \left( \delta_\mu^0 - \delta_\mu^1 \right) \left( \delta_\nu^0 - \delta_\nu^1 \right) \frac{\partial}{\partial u(z)} \frac{\partial}{\partial u(z')} \\
&\quad\times \Big[ \ln[ \epsilon + \mathrm{i} u(z) - \mathrm{i} u(z') ] - \ln[ \epsilon + \mathrm{i} \tau(u(z)) - \mathrm{i} \tau(u(z')) ] \Big]_{z = z(\tau), z' = z(\tau')} \eqend{.} \raisetag{4em}
\end{splitequation}
Using that as $\epsilon \to 0$ we have
\begin{equation}
\ln[ \epsilon + \mathrm{i} u(z) - \mathrm{i} u(z') ] = \ln\abs{ u(z) - u(z') } + \frac{\mathrm{i} \pi}{2} \sgn[ u(z) - u(z') ]
\end{equation}
and the analogous expression with $\tau(u)$, and that $\sgn[ \tau(u(z)) - \tau(u(z')) ] = \sgn[ u(z) - u(z') ]$ since $\tau$ is a monotone function of $u$, we can simplify~\eqref{eq:qei_deltau_def} to
\begin{splitequation}
\Delta^u_{\mu\nu}(\tau,\tau') &= \left( \delta_\mu^0 - \delta_\mu^1 \right) \left( \delta_\nu^0 - \delta_\nu^1 \right) \frac{\partial}{\partial u(z)} \frac{\partial}{\partial u(z')} \\
&\qquad\times \Big[ \ln\abs{ u(z) - u(z') } - \ln\abs{ \tau(u(z)) - \tau(u(z')) } \Big]_{z = z(\tau), z' = z(\tau')} \\
&= \left( \delta_\mu^0 - \delta_\mu^1 \right) \left( \delta_\nu^0 - \delta_\nu^1 \right) \\
&\qquad\times \bigg[ \frac{1}{\left[ u(z) - u(z') \right]^2} - \frac{\tau'(u(z)) \tau'(u(z'))}{\left[ \tau(u(z)) - \tau(u(z')) \right]^2} \bigg]_{z = z(\tau), z' = z(\tau')}
\end{splitequation}
and
\begin{splitequation}
\lim_{\tau' \to \tau} \Delta^u_{\mu\nu}(\tau,\tau') &= \left( \delta_\mu^0 - \delta_\mu^1 \right) \left( \delta_\nu^0 - \delta_\nu^1 \right) \frac{3 [ \tau''(u(z)) ]^2 - 2 \tau'(u(z)) \tau'''(u(z))}{12 [ \tau'(u(z)) ]^2} \eqend{,}
\end{splitequation}
which was already determined in~\cite[Eq.~(2.17)]{flanagan1997}. From~\eqref{eq:eqi_du_dtau}, we obtain
\begin{equation}
\tau'(u(z)) = \left[ \partial_\tau u(z(\tau)) \right]^{-1} = \frac{1}{v^0(z(\tau)) - v^1(z(\tau))} \eqend{,}
\end{equation}
and it follows that [with $v^\mu(\tau) \equiv v^\mu(z(\tau))$]
\begin{splitequation}
\label{eq:qei_deltau}
\Delta^u_{\mu\nu}(\tau,\tau) = - \frac{\left( \delta_\mu^0 - \delta_\mu^1 \right) \left( \delta_\nu^0 - \delta_\nu^1 \right)}{3 [ v^0(\tau) - v^1(\tau) ]^\frac{3}{2}} \frac{\partial^2}{\partial \tau^2} \frac{1}{\sqrt{ v^0(\tau) - v^1(\tau) }} \eqend{.}
\end{splitequation}
The analogous computation establishes that
\begin{splitequation}
\label{eq:qei_deltav}
\Delta^v_{\mu\nu}(\tau,\tau') &\equiv \partial_\mu^z \partial_\nu^{z'} \Big[ \ln\left( \epsilon + \mathrm{i} v(z,z') \right) - \ln\left[ \epsilon + \mathrm{i} \left( \tau(v(z)) - \tau(v(z')) \right) \right] \Big]_{z = z(\tau), z' = z(\tau')} \\
&\to - \frac{\left( \delta_\mu^0 + \delta_\mu^1 \right) \left( \delta_\nu^0 + \delta_\nu^1 \right)}{3 [ v^0(\tau) + v^1(\tau) ]^\frac{3}{2}} \frac{\partial^2}{\partial \tau^2} \frac{1}{\sqrt{ v^0(\tau) + v^1(\tau) }} \quad (\tau' \to \tau) \eqend{,}
\end{splitequation}
using that also
\begin{equation}
\label{eq:eqi_dv_dtau}
\partial_\tau v(z(\tau)) = \partial_\tau \left[ z^0(\tau) + z^1(\tau) \right] = v^0(\tau) + v^1(\tau) = \sqrt{ 1 + [ v^1(\tau) ]^2 } + v^1(\tau) > 0 \eqend{.}
\end{equation}
By construction, the $\Delta^{u/v}$ measure the difference between the parametrix on the actual trajectory $z^\mu(\tau)$ and an auxiliary stationary one $z^\mu = \delta^\mu_0 \tau$, and we see that this is finite, but possibly large if the trajectory is almost light-like with $v^1 \approx \pm v^0$. In particular, it diverges in the light-like limit $v^1 \to \pm v^0$, and we see again that we cannot simply take this limit to arrive at the result for a null trajectory.
Using the explicit expression for the Hadamard parametrix~\eqref{eq:hadamard}, we can thus decompose~\eqref{eq:qei_h} in the form
\begin{splitequation}
\label{eq:qei_htau_decomp}
h(\tau,\tau') &= \frac{1}{8 \pi} f(\tau) f(\tau') \Big[ v^\mu(\tau) v^\nu(\tau') + w^\mu(\tau) w^\nu(\tau') \Big] \\
&\quad\times \bigg[ \Delta^u_{\mu\nu}(\tau,\tau') + \Delta^v_{\mu\nu}(\tau,\tau') + \partial^z_\mu \partial^{z'}_\nu \ln\left[ \epsilon + \mathrm{i} \left( \tau(u(z)) - \tau(u(z')) \right) \right] \\
&\qquad+ \partial^z_\mu \partial^{z'}_\nu \ln\left[ \epsilon + \mathrm{i} \left( \tau(v(z)) - \tau(v(z')) \right) \right] \bigg]_{z = z(\tau), z' = z(\tau')} \\
&\equiv h^\Delta(\tau,\tau') + h^0(\tau,\tau') \eqend{,} \raisetag{1.2em}
\end{splitequation}
where $h^\Delta$ is the part containing $\Delta^{u/v}$. In this part, we can take the limit $\tau' \to \tau$, which results in
\begin{splitequation}
\label{eq:qei_hdeltatau}
h^\Delta(\tau,\tau) &= - \frac{1}{12 \pi} f^2(\tau) \bigg[ \sqrt{ v^0(\tau) - v^1(\tau) } \frac{\partial^2}{\partial \tau^2} \frac{1}{\sqrt{ v^0(\tau) - v^1(\tau) }} \\
&\qquad\qquad+ \sqrt{ v^0(\tau) + v^1(\tau) } \frac{\partial^2}{\partial \tau^2} \frac{1}{\sqrt{ v^0(\tau) + v^1(\tau) }} \bigg] \\
&= - \frac{1}{24 \pi} f^2(\tau) \frac{[ \dot v^1(\tau) ]^2}{1 + [ v^1(\tau) ]^2} \eqend{,}
\end{splitequation}
where we used that $w^1 = v^0$, $w^0 = v^1$ and that $v^0(\tau) = \sqrt{1 + [ v^1(\tau) ]^2}$, which follows from the normalisation $v^\mu v_\mu = -1$. For $h^0(\tau,\tau')$ which contains the contribution from the straight trajectory, we replace $f$ by a real-valued test function $f_N$ with compact support, call the resulting expression $h^0_N$ and bound the contribution of this term to the energy density using~\eqref{eq:thm_qei_N}. We thus have to compute
\begin{equation}
\label{eq:qei_htautau_fourier}
\int h^0_N(\tau,\tau) \mathop{}\!\mathrm{d} \tau = 2 \lim_{\delta \to 0^+} \int_0^\infty \mathrm{e}^{- \delta \xi^2} (\mathcal{F} h^0_N)(-\xi,\xi) \frac{\mathop{}\!\mathrm{d} \xi}{2 \pi}
\end{equation}
with the Fourier transform~\eqref{eq:qei_fourier_def}
\begin{splitequation}
\label{eq:qei_fourier_hn}
(\mathcal{F} h^0_N)(\xi,\zeta) &= \iint \mathrm{e}^{\mathrm{i} \xi \tau + \mathrm{i} \zeta \sigma} h^0_N(\tau,\sigma) \mathop{}\!\mathrm{d} \tau \mathop{}\!\mathrm{d} \sigma \\
&= \frac{1}{2 \pi} \iint \mathrm{e}^{\mathrm{i} \xi \tau + \mathrm{i} \zeta \sigma} f_N(\tau) f_N(\sigma) \frac{1}{( \tau - \sigma - \mathrm{i} \epsilon )^2} \mathop{}\!\mathrm{d} \tau \mathop{}\!\mathrm{d} \sigma \eqend{,}
\end{splitequation}
where we performed the derivatives in $h^0_N$~\eqref{eq:qei_htau_decomp} using~\eqref{eq:eqi_du_dtau} and~\eqref{eq:eqi_dv_dtau} and used again that $w^1 = v^0$, $w^0 = v^1$.
Using the Fourier transform
\begin{equation}
\frac{1}{( \tau - \mathrm{i} \epsilon )^2} = \int \left[ - 2 \pi p \, \Theta(p) \mathrm{e}^{- p \epsilon} \right] \mathrm{e}^{- \mathrm{i} p \tau} \frac{\mathop{}\!\mathrm{d} p}{2 \pi} \eqend{,}
\end{equation}
we can write~\eqref{eq:qei_fourier_hn} as
\begin{equation}
\label{eq:qei_fourier_hn2}
(\mathcal{F} h^0_N)(\xi,\zeta) = - \int \tilde{f}_N(p-\xi) \tilde{f}_N(-p-\zeta) p \, \Theta(p) \mathrm{e}^{- p \epsilon} \frac{\mathop{}\!\mathrm{d} p}{2 \pi} \eqend{,}
\end{equation}
where
\begin{equation}
\tilde{f}_N(q) \equiv \int f_N(\tau) \mathrm{e}^{- \mathrm{i} q \tau} \mathop{}\!\mathrm{d} \tau
\end{equation}
is the Fourier transform of the test function, which is also a Schwartz function, and we could interchange the integrals for $\epsilon > 0$ because everything converges absolutely. It follows that~\eqref{eq:qei_htautau_fourier}
\begin{splitequation}
\int h^0_N(\tau,\tau) \mathop{}\!\mathrm{d} \tau &= - 2 \lim_{\delta \to 0^+} \int_0^\infty \mathrm{e}^{- \delta \xi^2} \int \tilde{f}_N(p+\xi) \tilde{f}_N(-p-\xi) p \, \Theta(p) \mathrm{e}^{- p \epsilon} \frac{\mathop{}\!\mathrm{d} p}{2 \pi} \frac{\mathop{}\!\mathrm{d} \xi}{2 \pi} \\
&= - 2 \int_0^\infty \int_0^\infty \abs{ \tilde{f}_N(p+\xi) }^2 p \, \mathrm{e}^{- p \epsilon} \frac{\mathop{}\!\mathrm{d} p}{2 \pi} \frac{\mathop{}\!\mathrm{d} \xi}{2 \pi} \eqend{,} \raisetag{1.8em}
\end{splitequation}
where we used that since $f_N \in \mathcal{S}(\mathbb{R})$ is real-valued, we have $\tilde{f}_N(-p) = \tilde{f}^*_N(p)$, and where we could take the limit inside the integral because everything converges absolutely. Here we see that since $p$ is positive, for a non-trivial quantum energy inequality we need that $\xi \geq 0$~\cite{fewstersmith2008}, since otherwise the integrand would be unbounded along the diagonal $\xi = -p$ in the limit $\epsilon \to 0$. Shifting $p \to p - \xi$, we can then interchange the integrations and perform the integral over $\xi$, which gives
\begin{splitequation}
\label{eq:qei_h0n_tauint}
\int h^0_N(\tau,\tau) \mathop{}\!\mathrm{d} \tau &= - 2 \int_0^\infty \abs{ \tilde{f}_N(p) }^2 \int_0^p (p-\xi) \, \mathrm{e}^{- (p-\xi) \epsilon} \frac{\mathop{}\!\mathrm{d} \xi}{2 \pi} \frac{\mathop{}\!\mathrm{d} p}{2 \pi} \\
&= - \frac{1}{2 \pi^2} \int_0^\infty \abs{ \tilde{f}_N(p) }^2 \frac{1 - \mathrm{e}^{- p \epsilon} ( 1 + p \epsilon )}{\epsilon^2} \mathop{}\!\mathrm{d} p \eqend{.}
\end{splitequation}
Using dominated convergence, we can then take the limit $\epsilon \to 0$ inside the integral, since the integrand (apart from the test functions) is bounded by its value at $\epsilon = 0$ (which is $p^2/2$). Using finally Parseval's theorem for the Fourier transform, we obtain
\begin{equation}
\label{eq:qei_h0n_tauint2}
\int h^0_N(\tau,\tau) \mathop{}\!\mathrm{d} \tau = - \frac{1}{4 \pi^2} \int_0^\infty \abs{ p \tilde{f}_N(p) }^2 \mathop{}\!\mathrm{d} p = - \frac{1}{4 \pi} \int \big[ f'_N(\tau) \big]^2 \mathop{}\!\mathrm{d} \tau \eqend{.}
\end{equation}
Choosing now a sequence of test functions $f_N(\tau) \equiv \chi_N(\tau) f(\tau)$ with $\supp \chi_N \subset [-N-2,N+2]$, $\chi_N(\tau) = 1$ for $\tau \in [-N,N]$ and $\abs{\chi_N(\tau)}, \abs{\chi'_N(\tau)} \leq 1$, we have $\abs{ f'_N(\tau) } \leq \abs{ f(\tau) } + \abs{ f'(\tau) }$ and can use dominated convergence to take the limit $N \to \infty$ inside the integral. Taking all together, from~\eqref{eq:qei_h}, \eqref{eq:qei_htau_decomp}, \eqref{eq:qei_hdeltatau} and~\eqref{eq:qei_h0n_tauint2} we thus obtain
\begin{splitequation}
\label{eq:qei_h_result}
&\int \bigg[ - \frac{\mathrm{i}}{2} f(\tau) f(\tau) \Big[ v^\mu(\tau) v^\nu(\tau) + w^\mu(\tau) w^\nu(\tau) \Big] \partial^z_\mu \partial^{z'}_\nu H^+(z,z') \Big\vert_{z = z' = z(\tau)} \bigg] \mathop{}\!\mathrm{d} \tau \\
&\quad= - \frac{1}{24 \pi} \int \left[ 6 \big[ f'(\tau) \big]^2 + f^2(\tau) \frac{[ \dot v^1(\tau) ]^2}{1 + [ v^1(\tau) ]^2} \right] \mathop{}\!\mathrm{d} \tau \eqend{,} \raisetag{1.8em}
\end{splitequation}
which depends on the test function $f$ and the trajectory $z^\mu(\tau)$ and is clearly negative, but bounded.
It remains to bound the remaining terms in~\eqref{eq:qei_e_positivebounded}, which contain $\mathcal{G}_{k,\ell,n,m}$ and the vertex operators $V_{\pm \beta}$. For the first kind of terms, we use the definition of $\mathcal{G}_{k,\ell,n,m}$~\eqref{eq:calg_def} and $\overline{\mathcal{G}}_{k,\ell,n,m}$~\eqref{eq:barcalg_def} as well as the definitions of the Feynman~\eqref{eq:hf_def} and Dyson parametrix~\eqref{eq:hd_def} to obtain
\begin{splitequation}
&\mathcal{G}_{k,\ell,2(\ell+m),m}(\vec{x};\vec{y};z) - \overline{\mathcal{G}}_{k,\ell,2(\ell+m),m}(\vec{x};\vec{y};z) \\
&= \sum_{i=1}^\ell \Big[ H^+(x_i,z) - H^\text{D}(z,x_i) \Big] + \sum_{i=\ell+1}^{\ell+m} \Big[ H^\text{F}(x_i,z) - H^+(z,x_i) \Big] \\
&\quad\qquad- \sum_{j=1}^{k-\ell} \Big[ H^+(y_j,z) - H^\text{D}(z,y_j) \Big] - \sum_{j=k-\ell+1}^{\ell+m} \Big[ H^\text{F}(y_j,z) - H^+(z,y_j) \Big] \\
&= \sum_{i=1}^{\ell+m} \Big[ G_\text{ret}(x_i,z) - G_\text{ret}(y_i,z) \Big] \raisetag{2.2em}
\end{splitequation}
with the state-independent retarded propagator
\begin{equation}
G_\text{ret}(x,y) \equiv \Theta(x^0-y^0) \Big[ H^+(x,y) - H^+(y,x) \Big] \eqend{.}
\end{equation}
Replacing $n \to 2n$ and using the Kronecker $\delta$ for the sum over $m$, it thus follows that the contribution of terms containing $\mathcal{G}_{k,\ell,n,m}$ to~\eqref{eq:qei_e_positivebounded} can be written as
\begin{splitequation}
&\frac{\beta^2}{2} f(\tau) f(\tau') \Big[ v^\mu(\tau) v^\nu(\tau') + w^\mu(\tau) w^\nu(\tau') \Big] \\
&\quad\times \sum_{n=0}^\infty \sum_{k=0}^{2n} \sum_{\ell=\max(0,k-n)}^{\min(k,n)} \frac{(-1)^{k+n}}{\ell! (k-\ell)! (n-\ell)! (n-k+\ell)!} \int\dotsi\int \mathcal{E}_{k,\ell,2n,n-\ell}(\vec{x};\vec{y}) \\
&\quad\times \sum_{i=1}^n \partial^z_\mu \Big[ G_\text{ret}(x_i,z) - G_\text{ret}(y_i,z) \Big] \partial^{z'}_\nu \bigg[ \sum_{i=1}^\ell H^+(x_i,z') + \sum_{i=\ell+1}^n H^\text{F}(x_i,z') \\
&\quad\qquad- \sum_{j=1}^{k-\ell} H^+(y_j,z') - \sum_{j=k-\ell+1}^n H^\text{F}(y_j,z') - \mathrm{i} \sum_{i=1}^n W(x_i,z') + \mathrm{i} \sum_{j=1}^n W(y_j,z') \bigg] \\
&\quad\times \prod_{i=1}^n g(x_i) g(y_i) \mathop{}\!\mathrm{d}^2 x_i \mathop{}\!\mathrm{d}^2 y_i \eqend{,} \raisetag{2.2em}
\end{splitequation}
which contains state-dependent terms involving $W$ and state-independent ones. Since the state-dependent terms are independent of $k$ and $\ell$, and moreover symmetric under the interchange of the $x_i$ and $y_j$ among themselves, we can use again the identity~\eqref{eq:qei_vanishing_sum} to conclude that they vanish for all $n > 0$. However, since they only appear starting from $n = 1$, they do not contribute at all, and only state-independent terms remain. For those, we can take the limit $\tau' \to \tau$ and bound them as in the previous section~\ref{sec_renorm_qed}, and the bound is state-independent due to our condition~\eqref{eq:thm_renorm_wpos_disc} on $W$. In the same way, we bound the contribution of the vertex operators in~\eqref{eq:qei_e_positivebounded}, and the resulting bound is also state-independent.
Taking all together, we have shown that the first term in~\eqref{eq:qei_e_positivebounded} is state-dependent but positive, since it can be obtained as the limit of vanishing cutoff $\Lambda \to 0$~\eqref{eq:qei_limit_state} of an alternative quasi-free state with positive definite two-point function~\eqref{eq:qei_omega_pos}. For the terms involving the Hadamard parametrix $H^+(z,z')$, the only non-vanishing contribution comes from the free theory~\eqref{eq:qei_h}, which we have split into a part evaluated on a straight worldline and a part describing the difference between this and the actual worldline. The second part has a finite limit as $\tau' \to \tau$~\eqref{eq:qei_hdeltatau}, while for the first part we have used~\cite[Lemma~2.12]{fewstersmith2008} in the form~\eqref{eq:thm_qei_N}; taking both together we obtain a negative but state-independent contribution~\eqref{eq:qei_h_result} to the energy density. The remaining terms in~\eqref{eq:qei_e_positivebounded} can be simply bounded, noticing that the derivatives of the state-dependent part $W$ drop out. Since the only state dependence is then in the exponentials which we estimate by $1$ using our condition~\eqref{eq:thm_renorm_wpos_disc} on $W$, the bound on these terms is state-independent. The sum of the bounds for these terms together with the contribution~\eqref{eq:qei_h_result} from the Hadamard parametrix then gives a lower bound on the smeared energy density~\eqref{eq:thm_qei}, which depends on the worldline $z(\tau)$, the smearing $f$, $\beta$ and the adiabatic cutoff $g$. Furthermore, this bound is non-trivial since the state-dependent positive part is arbitrary and can thus certainly be larger. \hfill\squareforqed
\begin{acknowledgements}
This work is supported by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) --- project no. 396692871 within the Emmy Noether grant CA1850/1-1 and project no. 406116891 within the Research Training Group RTG 2522/1. It is a pleasure to thank Chris Fewster, Ben Freivogel, Nicola Pinamonti and Aron Wall for comments and suggestions.
\end{acknowledgements}
|
\section{Introduction} \label{sec:sgnd_introduction}
Complex datasets with outliers in the response and/or heavy-tailed errors can pose a challenge for standard modelling techniques. Classical regression procedures that assume normally distributed errors minimize the sum of squared errors (SSE) function, which can be prone to pursuing specific observations that differ from the majority of data. These influential observations can have a substantial impact and can distort parameter estimates due to their large residuals \citep{kuhn13applied}. As an example, there is evidence of outliers in the Boston housing data \citep{harrison78hprice2}, which examines the association between median house prices and numerous community characteristics. This is a well known motivating example dataset in the context of normal linear regression, though it appears that the normality assumption on the error distribution may not be appropriate. The density and the quantile-quantile (Q-Q) plot of the standardized residuals from a linear model are shown in Figure~\ref{figs:sgnd_hprice}(a.i) and \ref{figs:sgnd_hprice}(a.ii), which both suggest non-normality of the error distribution. Furthermore, a Shapiro-Wilk test on these residuals leads to a rejection of the null hypothesis of normality (p-value $< 10^{-8}$). Examining the kurtosis of the error distribution implies that the data have a leptokurtic distribution (i.e.,~heavy-tailed relative to a normal distribution). Indeed, a bootstrapped 95\% confidence interval of the excess kurtosis does not contain zero $[1.25,3.75]$.
The presence of small deviations from the model, like this heavy-tailed error distribution, can cause classical regression procedures to become biased and/or inefficient. Therefore, analysis of real data applications using classical methods can be misleading \citep{yu13stability, hanin21cavalier}. When strict model assumptions are violated, the inferential quality disimproves (e.g.,~standard errors and confidence intervals) and can become invalid \citep{ronchetti20accurate}. Robust statistical procedures are extensions of parametric statistics that remain reliable in the presence of these deviations. These methods provide a middle ground between the strictness of a parametric model and the full nonparametric approach, which can be difficult to interpret. Consequently, robust approaches benefit from the simplicity of the parametric structure, both in terms of ease of computation and interpretation, while also having the capability to protect against distributional deviations \citep{avella15robustreview}. A common approach to robust regression is to replace the SSE function in classical regression, which is highly sensitive to outliers, with the sum of the absolute errors function. This is the least absolute deviation (LAD) regression \citep{bassett78asymptotic} and is equivalent to assuming a Laplace error distribution. This is also referred to as $L_1$ regression, and is a type of quantile regression \citep{koenker78quantiles}, where the conditional median of the response variable is estimated. \citet{dielman05least} provides a review of research concerning LAD regression, with \citet{avella15robustreview} delivering a selective overview of different estimators in robust statistics. There are several software implementations available, for example, in \texttt{R} \citep{citer}. The \texttt{rlm} function in the \texttt{MASS} package \citep{massbook} fits a linear model by robust regression that is solved using iteratively reweighted least squares (IRLS). The \texttt{quantreg} package \citep{quantregbook} performs a quantile regression using the \texttt{rq} function and uses linear programming to solve the problem; the \texttt{robustbase} package also uses linear programming via the \texttt{lmrob} function \citep{robustbasepaper, maronna19robustbook}. The \texttt{L1pack} package \citep{L1pack} provides the option to use either IRLS or linear programming to fit a robust linear model using the \texttt{lad} function.
\begin{figure}[t!]
\centering
\makebox{\includegraphics[width = \textwidth]{figs/hprice_summ.pdf}}
\caption{\label{figs:sgnd_hprice}Boston Housing Data: density (a.i) and Q-Q (a.ii) plot of the standardized residuals from fitting a linear model. Standardized residuals from fitting a robust linear model plotted against the independent variables log(property-tax rate) (b.i) and log(distance) (b.ii).}
\end{figure}
Continuing with our motivating example, we fit a robust linear model to the Boston housing data (via the \texttt{rlm} function in the \texttt{MASS} package). While this will capture the aforementioned heavy tails in the error distribution, we also observe indications of heteroscedasticity in the data. Figure~\ref{figs:sgnd_hprice}(b) displays the standardized residuals from fitting a robust linear model plotted against two of the covariates: log(property-tax rate) and log(distance). These are the log of the full-value property-tax rates and the log of the distance to employment centres respectively. In both cases, it is clear that the variance of the residuals is non-constant with respect to the covariates: the variability increases with the value of log(property-tax rate) and decreases with the value of log(distance). Traditional regression methods tend to focus solely on modelling the mean, which is only one feature of the response distribution. Other potential distributional aspects, such as variability and kurtosis, are often treated as little more than nuisance parameters. This can lead to incomplete analyses, particularly when important dynamics are neglected. In the case of the Boston housing data, the presence of heteroscedasticity indicates that more complex relationships are present, with important insights to be gained by looking beyond the mean. Distributional regression is the approach of simultaneously modelling multiple (or even all) distributional parameters (i.e.,~location, scale and shape) by allowing them to depend on covariates. This reduces the risk of false conclusions and information loss that can be encountered when the conditional distribution is reduced to a single-valued functional (i.e.,~the mean) \citep{kneib13beyond, kneib21rage}. The generalized additive models for location, scale and shape (GAMLSS) model class \citep{rigby05, stas18gamlss} is one of the most popular examples of distributional regression, and is implemented in \texttt{R} using the \texttt{gamlss} package \citep{stas07}. Examination of regression effects beyond the mean, using distributional regression methods such as GAMLSS, presents an opportunity to gain a more comprehensive understanding of the problem at hand.
When there are multiple different covariates that potentially influence the response variable (e.g.,~the Boston housing data has 12), it is useful to perform variable selection to determine a smaller subset of covariates whose effects are sufficiently strong to be deemed important. Classical variable selection techniques, such as stepwise or subset selection, can become computationally expensive; this has, in part, led to the popularity of penalization methods, such as the LASSO (least absolute shrinkage and selection operator) \citep{tibshirani96}, where parameter estimation and model selection is executed simultaneously. The LAD-LASSO \citep{wang07robust} is a robust version of the LASSO, where the quadratic loss function is replaced by the LAD function. There are many extensions of the LAD-LASSO, including methods based on ranks \citep{leng10ladlasso}, adaptive penalized estimation procedures \citep{fan14robustadaptive}, weighted approaches \citep{jiang21outlier} and applications in the high dimensional setting \citep{belloni11, fan17estimation}. \citet{smucler17robust} and \citet{filzmoser21robust} present comprehensive overviews of robust variable selection methods, but current research in this area appears to deal only with the location component of the model, while the variance is assumed constant. To the best of our knowledge, \citet{stas07} were the first to address the variable selection problem with multiple distributional parameters, where they implement stepwise procedures in GAMLSS (see also \citet{stasinopoulos17gamlssbook} and \citet{ramires21gamlssstep}); although GAMLSS is not specific to robust regression, it includes the Laplace distribution. Furthermore, \citet{mayr12} employ gradient boosting techniques to fit GAMLSS models and carry out variable selection; this is implemented in the \texttt{gamboostLSS} package \citep{gamboostLSS}. \citet{hambuckers18understanding} utilize $L_1$-type penalties in the high dimensional distributional regression setting for numeric covariates. \citet{groll19} extend this work by including strategies for nominal and ordinal categorical predictors and present their corresponding package, \texttt{bamlss} \citep{umlauf18bamlsspaper}. However, since there are penalties for multiple distributional parameters, optimal tuning parameter selection (via multidimensional grid search) is computationally intensive, which is a drawback of this otherwise flexible method.
Robust distributional regression with automatic variable selection is an underdeveloped area. We propose a method that covers both classical and robust regression, includes scale regression in addition to location regression, and carries out automatic variable selection. Our starting point is the use of the generalized normal distribution (GND), which contains an additional distributional parameter to be estimated (beyond location and scale) that controls the shape of the distribution and can be regarded as a measure of the kurtosis \citep{box73pe, nelson91conditional, nadarajah05generalized}. Moreover, we adopt a distributional regression approach by allowing both the location and scale to depend on covariates. To carry out automatic variable selection, we extend the work of \citet{oneill2021smoothicarxiv}, who considered normally distributed errors only. Their approach involves the direct optimization of an information criteria, which is particularly advantageous from a computational viewpoint in the distributional regression setting, since it obviates the need for tuning parameter selection. To make our approach amenable to gradient-based optimization (e.g.,~Newton-Raphson), we employ a smooth approximation to both the $L_0$ norm in the information criterion and the absolute value function present in the GND density function. Our proposed method is implemented in the \texttt{smoothic} package \citep{smoothic_package}.
The rest of this paper is organized as follows. Section~\ref{sec:sgnd_methodology} outlines the proposed modelling framework and provides details of the estimation procedure and optimization algorithm. The results from extensive simulation studies are presented in Section~\ref{sec:sgnd_simulation_studies}, followed by the application of the method to two real datasets in Section~\ref{sec:sgnd_real_data_analyses}. Finally, Section~\ref{sec:sgnd_discussion} provides a discussion and some concluding remarks.
\section{Methodology} \label{sec:sgnd_methodology}
\subsection{Preliminaries} \label{sec:sgnd_preliminaries}
The generalized normal distribution (also known as the exponential power distribution or generalized error distribution) generalizes the Laplace and normal distributions through a shape parameter $\kappa$ \citep{box73pe, nelson91conditional, nadarajah05generalized}. The generalized normal distribution (GND) for the response variable $y$ has density function
\begin{equation*}
f(y) = \frac{\kappa}{2 s \Gamma(1 / \kappa)} e^{-\left(\frac{\lvert y-\mu \rvert}{s}\right)^{\kappa}},
\end{equation*}
where $\mu$, $s>0$, $\kappa>0$ are the location, scale and shape parameters respectively, and $\Gamma(\cdot)$ is the gamma function. This corresponds to the Laplace distribution when $\kappa = 1$ and the normal distribution when $\kappa = 2$. Note that it can be written in the general form
\begin{equation*}
f(y) = c(\kappa)\frac{1}{s} e^{-g\left(\frac{y-\mu}{s}\right)},
\end{equation*}
where $c(\kappa)=\kappa/2\Gamma(1/\kappa)$ is the normalizing constant and $g(z)=\lvert z \rvert ^{\kappa}$. We suggest using a smooth extension of the absolute value defined as $a_\tau(z) = \sqrt{z^2 + \tau^2} - \tau$, which is differentiable for $\tau > 0$ and $\lim_{\tau\to 0} a_\tau(z) = \lvert z \rvert$. Therefore, replacing $\lvert z \rvert$ with $a_\tau(z)$ in the GND density, yields the smooth GND (SGND) density
\begin{equation}\label{eq:sgnd_density}
\tilde f(y) = \tilde{c}_{\tau}(\kappa)\frac{1}{s} e^{-\tilde{g}\left(\frac{y-\mu}{s}\right)},
\end{equation}
where $\tilde g(z) = \{a_\tau(z)\}^\kappa$ and $\tilde{c}_{\tau}(\kappa)=1/\int_{-\infty}^{\infty} e^{-\tilde{g}(z)} \,dz$ is a normalizing constant that can be approximated, for example, using numerical integration.
Following the distributional regression framework, the log-likelihood function for the SGND model is
\begin{equation*}
\ell(\theta)=n\log(\tilde{c}_{\tau}(\kappa)) - \sum_{i=1}^{n} \log(s_i)-\sum_{i=1}^{n}\tilde{g}\left(\frac{y_i-\mu_i}{s_i}\right)
\end{equation*}
with $\mu_i = x_i^T \beta$, $\log(s_i^2)=x_i^T \alpha$ and $\log(\kappa - \kappa_\text{min})=\nu_0$, where the log-link functions are used to ensure the positivity of the parameters $s$ and $\kappa$. The slightly different link function for $\kappa$ is used to restrict $\kappa > \kappa_\text{min}$, for $\kappa_\text{min} > 0$ (e.g.,~$\kappa_\text{min} = 0.2$), to avoid less realistic sub-Laplace distributions with very small $\kappa$ values. (We introduce this mainly for computational reasons, as we have found that optimization can become unstable when $\kappa$ moves into this region.) The response variable is $y_i$ and $x_i=(1, x_{1i}, \ldots, x_{pi})^T$ is the vector of covariates for the $i$th individual; we assume that the covariates are scaled so that they have unit variance, since we will be carrying out penalized estimation. The vector of regression coefficients for the location and scale parameters are ${\beta}=(\beta_{0},\beta_{1},\ldots,\beta_{p})^{T}$ and ${\alpha}=(\alpha_{0},\alpha_{1},\ldots,\alpha_{p})^{T}$. Although $\kappa$ could potentially also depend on covariates, we maintain it as a covariate-independent ``distribution choosing'' parameter here. Moreover, in a somewhat related setting, \citet{burke20} have found that modelling these higher-order distributional shape parameters in terms of covariates may not be required in general (and can sometimes lead to unstable estimation). Thus, the complete parameter vector is $\theta=(\beta^T, \alpha^T, \nu_0)^T$.
In practice, we would generally expect that different sets of covariates will appear in the location and scale components, and this can be dealt with by setting various regression coefficients to zero --- which can be achieved through variable selection. Following the approach of \citet{oneill2021smoothicarxiv}, we achieve automatic variable selection by directly optimizing a ``smooth information criterion" (SIC). This formulates the estimation problem as a penalized likelihood, yielding our proposed SGND-SIC objective function
\begin{equation}
\ell^\text{SIC}(\theta)=\ell(\theta)-\frac{\lambda}{2}\left[\lvert \lvert \tilde\beta \rvert \rvert_{0, \epsilon}+ \lvert \lvert \tilde\alpha \rvert \rvert_{0, \epsilon}+3\right],
\label{eq:sgnd_sic}
\end{equation}
where fixing $\lambda=2$ or $\lambda=\log(n)$ is equivalent to the AIC and BIC (up to a factor of two), respectively, $\lvert \lvert \theta \rvert \rvert_{0, \epsilon} = \sum_{j=1}^{p}\phi_\epsilon(\theta_j)$ is a ``smooth $L_0$ norm", where $\phi_\epsilon(\theta_j) = {\theta_j^2}/({\theta_j^2 + \epsilon^2)}$ and $\epsilon$ is the penalty smoothing parameter such that smaller values of $\epsilon$ bring $\lvert \lvert \theta \rvert \rvert_{0, \epsilon}$ closer to the $L_0$ norm. In this paper, we apply a BIC-type criterion where $\lambda = \log(n)$ throughout. The intercepts are not penalized and are therefore omitted from $\tilde\beta=(\beta_1,\ldots,\beta_{p})^T$ and $\tilde\alpha=(\alpha_1,\ldots,\alpha_{p})^T$. There is an addition of three in the penalty to include the cost of estimating the intercept terms $\beta_0$, $\alpha_0$ and $\nu_0$; this constant is not required for optimization purposes but is included to keep the approach in line with an information criteria. Using a smooth approximation to both the absolute value in the GND and the $L_0$ norm in the information criterion results in a fully differentiable problem. This means that standard gradient based optimization procedures, such as Newton-Raphson, can be used to optimize our smooth objective function as described in the following section.
\subsection{Estimation Procedure}\label{sec:sgnd_estimation_procedure}
We define the penalized estimator as
\begin{equation*}
\hat{\theta} = {\arg\max}\left(\ell^{\text{SIC}}(\theta)\right),
\end{equation*}
where $\ell^{\text{SIC}}(\theta)$ is given by \eqref{eq:sgnd_sic}. The first derivatives with respect to the parameters are
\begin{align*}
\begin{split}
\frac{\partial \ell^\text{SIC}}{\partial \beta} &= \frac{\partial \ell}{\partial \beta} - \frac{\log(n)}{2}\xi_{\beta} = X^T z_{\beta} - \frac{\log(n)}{2}\xi_{\beta},\\
\frac{\partial \ell^\text{SIC}}{\partial \alpha} &= \frac{\partial \ell}{\partial \alpha} - \frac{\log(n)}{2}\xi_{\alpha} = X^T z_{\alpha} - \frac{\log(n)}{2}\xi_{\alpha},\\
\frac{\partial \ell^\text{SIC}}{\partial \nu} &= \frac{\partial \ell}{\partial \nu} = 1_n^T z_{\nu},
\end{split}
\end{align*}
where $X$ is an $n \times (p+1)$ matrix, whose $i$th row is $x_i$; $1_n = (1, \ldots, 1)^T$ is a vector of length $n$; $z_\beta$, $z_\alpha$ and $z_\nu$ are vectors of length $n$ such that
\begin{align*}
\begin{split}
z_{\beta, i} &= \frac{s_{i}^{-2}\kappa(y_i-\mu_{i})a_{i}^{\kappa-1}}{a_{i} + \tau},\\
z_{\alpha, i} &= \frac{s_{i}^{-2}\kappa(y_i-\mu_{i})^{2}a_{i}^{\kappa-1}}{2a_{i}+\tau} -\frac{1}{2},\\
z_{\nu, i} &=\frac{\partial}{\partial \nu}\log(\tilde{c}_{\tau}(\kappa))-\log (a_{i})a_{i}^{\kappa}(\kappa - \kappa_\text{min})
\end{split}
\end{align*}
where $a_{i} = \sqrt{s_{i}^{-2}(y_i-\mu_{i})^{2}+\tau^{2}}-\tau$, and $\partial \log(\tilde{c}_{\tau}(\kappa))/\partial \nu$ is given in Appendix~\ref{app:sgnd_derivatives_ctilde}; lastly, the $(j+1)$th elements of the vectors $\xi_\beta$ and $\xi_\alpha$ are given by $\phi^\prime_\epsilon(\beta_j)$ and $\phi^\prime_\epsilon(\alpha_j)$, where $\phi^\prime_\epsilon(\theta_j) = 2\theta_j \epsilon^2/(\theta_j^2+\epsilon^2)^2$, but the first elements are zero because the intercepts are not penalized, i.e., $\phi^\prime_\epsilon(\beta_0) = \phi^\prime_\epsilon(\alpha_0) = 0$. Then $-\nabla_{\theta}\nabla_{\theta}^T\ell^{\text{SIC}}(\theta)$ is given by
\begin{align*}
I(\theta)
&= I_0(\theta) +
\begin{pmatrix}
\log(n)\Sigma_{\beta}/2
& 0 & 0 \\
0 & \log(n)\Sigma_{\alpha}/2 & 0 \\
0 & 0 & 0
\end{pmatrix}\\
&= \begin{pmatrix}
X^{T} W_{\beta} X + \log(n)\Sigma_{\beta}/2
& X^{T} W_{\beta\alpha} X & {1}_n^TW_{\beta \nu}{1}_n\\
X^{T} W_{\beta\alpha} X & X^{T} W_{\alpha} X + \log(n)\Sigma_{\alpha}/2 & {1}_n^TW_{\alpha \nu}{1}_n \\
{1}_n^TW_{\beta \nu}{1}_n & {1}_n^TW_{\alpha \nu}{1}_n & {1}_n^TW_{\nu}{1}_n
\end{pmatrix}
\end{align*}
where $I_0(\theta) = -\nabla_{\theta}\nabla_{\theta}^T\ell(\theta)$ is the observed information matrix associated with the unpenalized likelihood (i.e.,~\eqref{eq:sgnd_sic} with $\lambda = 0$), $\Sigma_\beta$ and $\Sigma_\alpha$ are diagonal matrices that are present due to the penalties and whose $(j+1)$th diagonal elements are $\phi_\epsilon^{\prime\prime}(\beta_{j})$ and $\phi_\epsilon^{\prime\prime}(\alpha_{j})$, where $\phi_{\epsilon}^{\prime \prime}(\theta_j)=2\epsilon^2(\epsilon^2-3\theta_j^2)/(\theta_j^2+\epsilon^2)^3$, but the first diagonal elements are zero because the intercepts are not penalized, i.e., $\phi_{\epsilon}^{\prime \prime}(\beta_0)=\phi_{\epsilon}^{\prime \prime}(\alpha_0)=0$; $W_\beta$, $W_\alpha$, $W_\nu$, $W_{\beta \alpha}$, $W_{\beta \nu}$ and $W_{\alpha \nu}$ are $n \times n$ diagonal weight matrices whose $i$th diagonal elements are provided in Appendix~\ref{app:sgnd_derivatives_second}.
For the purpose of optimization, we use the ``RS" algorithm \citep{rigby05}, where the off-diagonal elements of the information matrix are set to zero. One justification for this approach is due to parameter orthogonality in location-scale models \citep{cox87}, where the location parameter is information orthogonal to both the scale and shape parameters. However, the scale and shape parameters in this model are not orthogonal \citep{jones11orthog}. Despite this, we have found that the algorithm works well without the cross derivatives; this finding is also supported by \citet{stas07}. Thus, the system of Newton-Raphson equations is
\begin{multline}
\hspace{-0.4cm}
\begin{pmatrix}
X^{T} W_{\beta}^{(m)} X + \log(n)\Sigma_{\beta}^{(m)}/2
& 0 & 0\\
0 & X^{T} W_{\alpha}^{(m)} X + \log(n)\Sigma_{\alpha}^{(m)}/2 & 0\\
0 & 0 & 1_n^T W_{\nu}^{(m)}1_n
\end{pmatrix}
{
\begin{pmatrix}
\beta^{(m+1)} - \beta^{(m)}\\
\alpha^{(m+1)} - \alpha^{(m)}\\
\nu^{(m+1)} - \nu^{(m)}
\end{pmatrix}}
\\=
\begin{pmatrix}
X^T z_{\beta}^{(m)} - \log(n)\xi_{\beta}^{(m)}/2\\
X^T z_{\alpha}^{(m)} - \log(n)\xi_{\alpha}^{(m)}/2\\
1_n^T z_{\nu}^{(m)}\\
\end{pmatrix},
\label{eq:sgnd_nr_compact}
\end{multline}
and this system is iteratively solved for {$\theta^{(m+1)} = ({\beta^{(m+1)}}^T, {\alpha^{(m+1)}}^T, \nu^{(m+1)})^T$}. Note that the elements superscripted by $(m)$ depend on $\theta^{(m)}$, but this dependence is suppressed for notational convenience.
Our \texttt{smoothic} package provides both the above Newton-Raphson procedure and another based on the in-built \texttt{nlm} function in \texttt{R}; we have found the former to be faster and the latter to be more stable (owing to it being a more sophisticated optimizer). We now describe our suggested initialization for the optimization procedure. For the location parameter, we use the classical ordinary least squares estimates, i.e., $\beta^{(0)} = (X^T X)^{-1} X^T Y$, where $Y = (y_1,\ldots,y_n)^T$ is the vector of response values. We then set the intercept of the scale parameter to be $\log(q^2)$, where $q^2 = \sum_{i=1}^n (y_i - x_i^T \beta^{(0)})^2/(n - p)$ is the classical residual variance estimator in normal linear regression. As in \citet{rutemiller68} and \citet{harvey76}, the remaining elements of the $\alpha^{(0)}$ parameter vector are set to zero, which yields $\alpha^{(0)}=(\log(q^2), 0, \ldots, 0)^T$. We initialize the shape parameter at $\nu_0^{(0)}=\log(2-\kappa_\text{min})$, which corresponds to a distribution with normal-like tails.
Upon convergence, the standard errors of the estimates are directly acquired by estimating the covariance of the penalized estimates for the true non-zero parameters using the sandwich formula,
\begin{equation}
\hat{\mathrm{cov}}(\hat{\theta})=\{I(\hat{\theta})\}^{-1} I_0(\hat{\theta})\{I(\hat{\theta})\}^{-1},
\label{eq:sgnd_sandwich}
\end{equation}
which is a general formula used in penalized regression that \citet{fan01, fan02} have shown to be accurate for moderate sample sizes. \citet{oneill2021smoothicarxiv} have also applied this formula successfully with heteroscedastic data and penalized estimation. Obtaining standard errors in a robust regression setting is a known problem \citep{singh98breakdown, salibian02, croux04robust}, with \citet{simpson92se} defining the concept of standard error breakdown. This occurs if the estimated standard error of the estimates are driven to zero or infinity, and, to the best of our knowledge, this remains an issue for penalized estimation with heavy-tailed heteroscedastic data. Indeed, we experience standard error breakdown for small sample sizes in the heavy-tailed heteroscedastic setting when $\kappa = 1$, with the estimated standard errors being driven to zero for small $\tau = 0.05$ (see the Supplementary Material). However, inferential performance does improve for larger sample sizes, suggesting that \eqref{eq:sgnd_sandwich} is working asymptotically as expected. The issue of standard error breakdown is not observed for larger $\kappa$ values that we have tested, implying that the problem is associated with $\kappa$ values near one (and, indeed, also the even heavier-tailed setting where $\kappa<1$ that is not shown here). Moreover, standard error breakdown is not such an issue in the homoscedastic setting (see the Supplementary Material). To improve the situation with respect to standard error breakdown, we suggest setting $\tau = 0.15$, yielding a smoother absolute value function approximation in the density function given in \eqref{eq:sgnd_density}. Although setting $\tau$ even higher than this can eliminate the issue, it also moves us further away from the desired Laplace-like distribution.
\begin{algorithm}[b!]
\caption{Implementation of the SGND-SIC $\epsilon$-telescope Method}
\label{algo:sgnd_tele}
\begin{enumerate}[label=\textbf{\arabic*}.]
\item \textbf{Initialization:} Rescale covariates to have unit variance and set $\theta^{(0)}=(\beta^{(0)T},\alpha^{(0)T}, \nu_0^{(0)})^T$, where $\beta^{(0)}$, $\alpha^{(0)}$ and $\nu_0^{(0)}$ are the initial values for the location, scale and shape parameters respectively (see Section~\ref{sec:sgnd_estimation_procedure}).
\item \textbf{$\mathbf{\epsilon}$-telescoping:} Run through the sequence of telescope values of length $T$ from $\epsilon_1$ to $\epsilon_T$, where $\epsilon_t=\epsilon_1 r^{t-1}$ for step $t=1,\ldots,T$ and $r \in (0,1)$ is the rate of exponential decay. We suggest $\epsilon_1=10$, $\epsilon_T=10^{-4}$ and $T=100$.
\begin{itemize}[leftmargin=*]
\item\textbf{For $\boldsymbol{t=1,\ldots,T}$:}\\
\textbf{Optimization:} Maximize $\ell^{\text{SIC}}(\theta)$ in \eqref{eq:sgnd_sic} to obtain $\hat{\theta}_{\epsilon_t}$ by iteratively re-solving the system of equations in \eqref{eq:sgnd_nr_compact} using initial values $\theta^{(0)}_{\epsilon_t}$ (where $\theta^{(0)}_{\epsilon_1}=\theta^{(0)}$). When $\lvert \theta^{(m+1)}_{\epsilon_t} - \theta^{(m)}_{\epsilon_t}\rvert\leq\omega$ for some small tolerance, e.g., $\omega = 10^{-8}$, convergence is achieved. The obtained estimates are used as initial values for the next step in the telescope, i.e., set $\theta^{(0)}_{\epsilon_{t+1}}$ = $\hat{\theta}_{\epsilon_t}$ to exploit warm starts.
\end{itemize}
\item \textbf{Output:} The final estimates $\hat{\theta}_{\epsilon_T}$ are acquired at $t=T$ and any estimates that are sufficiently close to zero (e.g.,~below $10^{-5}$) can be treated as being zero. The associated standard errors are calculated by evaluating \eqref{eq:sgnd_sandwich} at $\hat{\theta}_{\epsilon_T}$. The final estimates are converted back to their original scale (by dividing by the covariate sample standard deviations) so that they correspond to the original, unscaled covariates.
\end{enumerate}
\end{algorithm}
\subsection{Algorithm}\label{sec:sgnd_algorithm}
To induce sparsity, we aim for a small value of $\epsilon$ in the $L_0$ approximation. However, starting immediately with a small $\epsilon$ is challenging from an optimization perspective (since it is very close to the non-differentiable $L_0$ problem). Thus, we suggest an $\epsilon$-telescoping procedure \citep{oneill2021smoothicarxiv} where a sequence of successively smaller $\epsilon$ values are used as a way of providing good starting values (i.e.,~warm starts) towards a final small $\epsilon$. For the purpose of penalized estimation, we rescale covariates to have unit variance, i.e., we use $x_j / SD(x_j)$ in place of $x_j$ where $SD(x_j)$ is the sample standard deviation of $x_j$. Typically, for the final solution, unscaled regression coefficients are preferred. Therefore, we convert the coefficients back to their original scale. Considering this, our overall optimization procedure is summarized in Algorithm~\ref{algo:sgnd_tele}.
\section{Simulation Studies}\label{sec:sgnd_simulation_studies}
\begin{table}[b!]
\caption{\label{tab:sgnd_true_values}True parameter values
\centering
\begin{tabular}{@{}cccccccccccccc@{}}
\toprule
& & \textcolor{blue}{E} & \textcolor{red}{M} & \textcolor{mygreen}{B} & N & N & \textcolor{red}{M} & N & N & \textcolor{red}{M} & \textcolor{mygreen}{B} & \textcolor{blue}{E} & \textcolor{red}{M} \\
& $X_0$ & \textcolor{blue}{$X_1$} & \textcolor{red}{$X_2$} & \textcolor{mygreen}{$X_3$} & ${X_4}$ & ${X_5}$ & \textcolor{red}{$X_6$} & ${X_7}$ & $X_8$ & \textcolor{red}{${X_9}$} & \textcolor{mygreen}{$X_{10}$} & \textcolor{blue}{$X_{11}$} & \textcolor{red}{$X_{12}$} \\
\midrule
$\beta$ & 0 & 1 & 0.5 & 0.5 & 1 & 0.5 & 1 & 0 & 0 & 0 & 0 & 0 & 0 \\[0.05cm]
$\alpha$ & 0 & 0.5 & 1 & 0.5 & 1 & 0 & 0 & 0.5 & 1 & 0 & 0 & 0 & 0 \\
\bottomrule
\multicolumn{14}{p{0.55\textwidth}}{\footnotesize \textcolor{blue}{E = Exponential}, \textcolor{mygreen}{B = Bernoulli}, N = independent normal, \footnotesize \textcolor{red}{M = multivariate normal (correlated)}.}
\end{tabular}
\vspace{-1cm}
\end{table}
\subsection{Setup}
The performance of our proposed SGND-SIC method is investigated through simulation studies. Data are simulated from the SGND defined in \eqref{eq:sgnd_density}, with $\tau=0.15$ and $\kappa_\text{min} = 0.2$. The regression coefficients are shown in Table~\ref{tab:sgnd_true_values}, where $X$ $(X_1, \ldots, X_{12})$ is the matrix of 12 covariates. The setup of the regression coefficients enable the method to be assessed where covariates enter through both distributional parameters simultaneously, where both the $\beta$ and $\alpha$ coefficients are non-zero ($X_1$ to $X_4$), or through a single distributional parameter, where only the $\beta$ coefficient is non-zero ($X_5$ and $X_6$) or only the $\alpha$ coefficient is non-zero ($X_7$ and $X_8$). Lastly, the pure noise covariates ($X_9$ to $X_{12}$) represent the case where the covariates do not enter the model at all, i.e., both the $\beta$ and $\alpha$ coefficients are zero so that these covariates have no effect. The true values of 0.5 and 1 correspond to covariates having a weaker and stronger effect. The covariates are distributed as follows: $(X_1, X_{11}) \sim$ Exponential(1) corresponding to two skewed covariates; $(X_3, X_{10}) \sim$ Bernoulli(0.75) giving two unbalanced binary covariates; four independent normal covariates $(X_4, X_5, X_7, X_8) \sim$ $\text{N}(0,1)$; and four correlated multivariate normal covariates $(X_2, X_6, X_9, X_{12}) = (Z_1, Z_2, Z_3, Z_4) \sim$ MVN wherein $\text{corr}(Z_j, Z_k)= 0.5^{\lvert j - k \rvert}$. The shape parameter $\kappa$ is varied such that $\kappa \in (1, 1.33, 1.67, 2)$ (rounded to two decimal places), which corresponds to a Laplace-like distribution when $\kappa=1$, a normal-like distribution when $\kappa = 2$, and a distribution between these when $\kappa \in (1.33, 1.67)$. These values of $\kappa$ equate to true values of $\nu_0 \in (-0.22, 0.13, 0.38, 0.59)$ (rounded to two decimal places), where $\nu_0=\log(\kappa-\kappa_\text{min})$. Sample sizes of 500, 1000 and 5000 are considered, where each scenario is replicated 1000 times.
\subsection{Simulation Results}
We assess the variable selection performance in terms of the number of true zero coefficients correctly set to zero (C) and the probability of choosing the true model (PT). We also consider the mean squared error $\text{MSE}(\hat{\theta})=(\hat{\theta} - \theta)^T{X^TX}(\hat{\theta} - \theta)/n$, which is a measure of in-sample prediction accuracy \citep{tibshirani97}. Table~\ref{tab:sgnd_var_selection} displays these metrics averaged over simulation replicates.
\begin{table}[b!]
\caption{Simulation results: model selection metrics}
\label{tab:sgnd_var_selection}
\centering
\resizebox{\textwidth}{!}{
\begin{tabular}{@{}c@{~~~}l@{~~~} c@{~~}c@{~~}c@{~~} c@{~~} c@{~~}c@{~~}c@{~~} c@{~~} c@{~~}c@{~~}c@{~~} c@{~~} c@{~~}c@{~~}c@{}}
\toprule
{} & {} & \multicolumn{3}{c}{$\kappa = 1$} && \multicolumn{3}{c}{$\kappa = 1.33$} && \multicolumn{3}{c}{$\kappa = 1.67$} && \multicolumn{3}{c}{$\kappa = 2$} \\
\cmidrule(r){3-5} \cmidrule(r){7-9} \cmidrule(r){11-13} \cmidrule(){15-17}
{} & $n$ & C(6) & PT & MSE && C(6) & PT & MSE && C(6) & PT & MSE && C(6) & PT & MSE\\
\midrule
{$\beta$} & 500 & 5.83 & 0.84 & 0.01 && 5.88 & 0.89 & 0.01 && 5.89 & 0.90 & 0.01 && 5.90 & 0.90 & 0.01 \\
{} & 1000 & 5.89 & 0.89 & 0.01 && 5.92 & 0.93 & 0.00 && 5.94 & 0.94 & 0.00 && 5.94 & 0.94 & 0.00 \\
{} & 5000 & 5.97 & 0.97 & 0.00 && 5.97 & 0.97 & 0.00 && 5.97 & 0.97 & 0.00 && 5.97 & 0.97 & 0.00 \\[0.2cm]
{$\alpha$} & 500 & 5.90 & 0.89 & 0.12 && 5.89 & 0.90 & 0.07 && 5.89 & 0.90 & 0.05 && 5.89 & 0.90 & 0.04\\
{} & 1000 & 5.94 & 0.94 & 0.05 && 5.94 & 0.94 & 0.03 && 5.94 & 0.94 & 0.02 && 5.94 & 0.94 & 0.02\\
{} & 5000 & 5.99 & 0.99 & 0.01 && 5.99 & 0.99 & 0.01 && 5.99 & 0.99 & 0.00 && 5.99 & 0.99 & 0.00\\
\bottomrule
\multicolumn{17}{p{0.95\textwidth}}{\footnotesize C, average correct zeros; PT, the probability of choosing the true model; MSE, the average mean squared error.}\\
\end{tabular}}
\end{table}
For all values of $\kappa$, the C values are close to the true value of six for both the location and scale parameters, and this improves as the sample size increases. We also note that the method never incorrectly sets a variable to zero in the scenarios we have considered (although this is not shown in Table~\ref{tab:sgnd_var_selection}). In terms of the probability of choosing the correct model (PT) for both $\beta$ and $\alpha$, we see that this is generally above 90\% (approximately) and increases with the sample size; the only exception is the $\beta$ component when $\kappa = 1$ at the smallest sample size of $n=500$, where PT$=84\%$. The MSE values are generally small and decrease with the sample size.
\begin{table}[t!]
\caption{Simulation results: estimation and inference metrics}
\label{tab:sgnd_parameter_inference_short}
\resizebox{\textwidth}{!}{
\begin{tabular}{@{}l@{~~} c@{~~}c@{~~} c@{~~}c@{~~}c@{~~}c@{~~} c@{~~} c@{~~}c@{~~}c@{~~}c@{~~} c@{~~} c@{~~}c@{~~}c@{~~}c@{}}
\toprule
{} & {} & {} & \multicolumn{4}{c}{$n = 500$} && \multicolumn{4}{c}{$n = 1000$} && \multicolumn{4}{c}{$n = 5000$} \\%[0.05cm]
\cmidrule(r){4-7} \cmidrule(r){9-12} \cmidrule(){14-17}
{$\kappa$} & {} & $\theta$ & $\hat{\theta}$ & SE & SEE & CP && $\hat{\theta}$ & SE & SEE & CP && $\hat{\theta}$ & SE & SEE & CP \\%[0.05cm]
\midrule
1 & $\beta_{1}$ & 1.00 & 1.00 & 0.05 & 0.04 & 0.74 && 1.00 & 0.04 & 0.03 & 0.83 && 1.00 & 0.01 & 0.01 & 0.92 \\
{} & $\alpha_{1}$ & 0.50 & 0.51 & 0.10 & 0.09 & 0.94 && 0.50 & 0.07 & 0.06 & 0.94 && 0.50 & 0.03 & 0.03 & 0.94 \\
{} & $\nu_{0}$ & -0.22 & -0.22 & 0.11 & 0.10 & 0.90 && -0.22 & 0.07 & 0.07 & 0.92 && -0.22 & 0.03 & 0.03 & 0.92 \\[0.2cm]
1.33 & $\beta_{1}$ & 1.00 & 1.00 & 0.04 & 0.04 & 0.91 && 1.00 & 0.03 & 0.03 & 0.94 && 1.00 & 0.01 & 0.01 & 0.94 \\
{} & $\alpha_{1}$ & 0.50 & 0.51 & 0.08 & 0.08 & 0.94 && 0.50 & 0.06 & 0.05 & 0.94 && 0.50 & 0.02 & 0.02 & 0.94 \\
{} & $\nu_{0}$ & 0.13 & 0.16 & 0.11 & 0.10 & 0.92 && 0.14 & 0.07 & 0.07 & 0.92 && 0.13 & 0.03 & 0.03 & 0.93 \\[0.2cm]
1.67 & $\beta_{1}$ & 1.00 & 1.00 & 0.04 & 0.03 & 0.92 && 1.00 & 0.03 & 0.02 & 0.95 && 1.00 & 0.01 & 0.01 & 0.94 \\
{} & $\alpha_{1}$ & 0.50 & 0.50 & 0.07 & 0.07 & 0.94 && 0.50 & 0.05 & 0.05 & 0.95 && 0.50 & 0.02 & 0.02 & 0.94 \\
{} & $\nu_{0}$ & 0.38 & 0.43 & 0.12 & 0.11 & 0.92 && 0.40 & 0.08 & 0.07 & 0.94 && 0.38 & 0.03 & 0.03 & 0.95 \\[0.2cm]
2 & $\beta_{1}$ & 1.00 & 1.00 & 0.03 & 0.03 & 0.92 && 1.00 & 0.02 & 0.02 & 0.95 && 1.00 & 0.01 & 0.01 & 0.95 \\
{} & $\alpha_{1}$ & 0.50 & 0.50 & 0.07 & 0.06 & 0.94 && 0.50 & 0.05 & 0.04 & 0.94 && 0.50 & 0.02 & 0.02 & 0.94 \\
{} & $\nu_{0}$ & 0.59 & 0.65 & 0.13 & 0.12 & 0.91 && 0.62 & 0.08 & 0.08 & 0.93 && 0.59 & 0.03 & 0.03 & 0.95 \\
\bottomrule
\multicolumn{17}{p{0.95\textwidth}}{\footnotesize Full results for all coefficients available in the Supplementary Material; SE, standard deviation of estimates over 1000 replications; SEE, average of estimated standard errors over 1000 replications; CP, the empirical coverage probability of a nominal 95\% confidence interval.}\\
\end{tabular}}
\end{table}
The estimation and inferential performance of our method is examined in Table~\ref{tab:sgnd_parameter_inference_short}, which only shows results for $\beta_1$, $\alpha_1$ and $\nu_0$ in the interest of brevity (see the Supplementary Material for full results). More specifically, we display the average estimate over simulation replicates, the true standard error (SE) (i.e.,~the standard deviation of the estimates over the replicates), the average estimated standard error (SEE) (computed using \eqref{eq:sgnd_sandwich} in a given replicate), and the empirical coverage probability (CP) for a nominal 95\% confidence interval. In all cases, we can see that the parameter estimates are quite unbiased, apart from some slight bias in $\hat \nu_0$ for the smallest sample size. As for uncertainty quantification, we find that this is generally good: the SEs are well estimated by the SEEs and the confidence interval coverage is near 95\%. The exception to this is $\kappa = 1$ (especially at $n=500$) where the SEs are underestimated and the coverage is poor. However, the results are much improved at $n=1000$ and $n=5000$ suggesting that \eqref{eq:sgnd_sandwich} is asymptotically valid. In any case, the poorer results for $\kappa = 1$ are in line with the discussion in Section~\ref{sec:sgnd_estimation_procedure} about standard error breakdown in the context of robust regression. Of course, an alternative approach would be to estimate the standard errors in a nonparametric manner through bootstrapping; we have found that this arguably improves the situation with conservative SEE values (see the Supplementary Material).
\section{Real Data Analyses} \label{sec:sgnd_real_data_analyses}
\subsection{Overview}\label{sec:sgnd_real_data_analyses_overview}
In this section, we consider our proposed SGND-SIC estimation procedure applied to two datasets, namely, the Boston housing data (Section~\ref{sec:sgnd_real_data_analyses_hprice}) and a diabetes dataset (Section~\ref{sec:sgnd_real_data_analyses_diabetes}), both of which are accessible in the \texttt{smoothic} package \citep{smoothic_package}. More specifically, we model the location and scale parameters in terms of covariates (as described in Section~\ref{sec:sgnd_methodology}), hereafter referring to this as SGND-MPR to highlight the multi-parameter regression nature of the model. We compare this to the SGND-SPR approach, which is a so-called ``single-parameter regression" model wherein only the location parameter depends on covariates, i.e., the variance is a constant; this SGND-SPR model is analogous to standard (normal and robust) regression approaches, albeit with parameters estimated using our novel SGND-SIC procedure. Both models can be easily implemented (by using \texttt{model = "mpr"} or \texttt{"spr"}) with the following code.
\begin{lstlisting}[language=R1]
# SGND-SIC (MPR & SPR) --------------------
library(smoothic)
fit <- smoothic(formula = lcmedv ~ ., # lcmedv is the response variable
data = bostonhouseprice2,
family = "sgnd", # Smooth Generalized Normal Distribution
model = "mpr") # or model = "spr" for location only
\end{lstlisting}
We also compare our proposal to a variety of other methods, the first of which is the GAMLSS stepwise procedure via the \texttt{gamlss} package \citep{stas07}, and which we refer to as ``GAMLSS-STEP". More specifically, this approach is implemented in the \texttt{stepGAICAll.A()} function, whose argument \texttt{k} we set to $\log(n)$ corresponding to BIC as shown below.
\begin{lstlisting}[language=R1]
# GAMLSS-STEP -----------------------------
library(gamlss)
library(gamlss.dist) # for Power Exponential Distribution (PE2)
# Initial model in the stepwise search
initial_fit <- gamlss(lcmedv ~ 1,
data = bostonhouseprice2,
family = PE2)
# Strategy A stepwise-based approach
f_gamlssstep <- as.formula("~ crim + zn + indus + rm + age + rad +
ptratio + lnox + ldis + ltax + llstat + chast")
fit <- stepGAICAll.A(initial_fit,
scope = list(lower = ~1, # model with no predictors
upper = f_gamlssstep), # all predictors
k = log(nrow(bostonhouseprice2)), # BIC k = log(n)
nu.try = FALSE) # no covariates in the nu shape parameter
\end{lstlisting}
Note that \texttt{family = PE2} relates to the Power Exponential (PE) distribution (accessed via the \texttt{gamlss.dist} package \citep{rigby05}), where \texttt{PE2} corresponds to the standard parameterization of the distribution. The PE distribution is equivalent to the GND, making it comparable to our proposed SGND-MPR method.
The \texttt{gamboostLSS} package \citep{gamboostLSS} fits a GAMLSS model using distributional component-wise gradient boosting, wherein variable selection is a consequence of stopping the algorithm early. This is controlled by setting the number of iterations using the \texttt{mstop} argument, which is a vector, since the stopping point can be different for each distributional parameter. Choosing an optimal value, say, \texttt{opt\char`_mstop}, is often based on cross-validation, but we use BIC here to be more in line with our proposed BIC-based optimization procedure. (Note that we have found the results from cross-validation minimization to be broadly similar to the results from the BIC-based procedure that are presented in this section.) Due to the algorithmic nature of gradient boosting, there is no commonly accepted method to evaluate the degrees of freedom of a boosting fit \citep{gamboostLSS}; therefore, we take the degrees of freedom as the number of non-zero coefficients. We refer to this as the ``GAMLSS-BOOST" method, and it can be implemented for a specific \texttt{mstop} using the code below (which assumes that an optimal value \texttt{opt\char`_mstop} has already been determined by computing the BIC for a variety of \texttt{mstop} values).
\begin{lstlisting}[language=R1]
# GAMLSS-BOOST ----------------------------
library(gamboostLSS)
library(gamlss.dist) # for Power Exponential Distribution (PE2)
# Assume the optimal stopping iteration (opt_mstop) was found using BIC search
fit <- glmboostLSS(formula = list(mu = lcmedv ~ ., # location
sigma = lcmedv ~ ., # scale
nu = lcmedv ~ 1), # no covariates in shape
data = bostonhouseprice2,
families = as.families("PE2"),
control = boost_control(mstop = opt_mstop)) # early stopping
\end{lstlisting}
The final distributional regression method that we compare to ours is BAMLSS via the \texttt{bamlss} package \citep{umlauf18bamlsspaper}, wherein separate LASSO penalties are applied to each distributional parameter. Note that \texttt{bamlss} selects the LASSO tuning parameters by minimizing BIC using a 100-point grid search for each of the distributional parameters, which is computationally intensive since there are multiple distributional parameters; this is pronounced further by the fact that a penalty appears for the $\nu$ shape parameter even though it is covariate independent, i.e., a $100 \times 100 \times 100$ point grid search is used. The BAMLSS estimates are obtained by running the following code.
\begin{lstlisting}[language=R1]
# BAMLSS ----------------------------------
library(bamlss)
library(gamlss.dist) # for Power Exponential Distribution (PE2)
f_bamlss <- list(lcmedv ~ la(crim + zn + indus + rm + age + rad + ptratio
+ lnox + ldis + ltax + llstat + chast),
sigma ~ la(crim + zn + indus + rm + age + rad + ptratio
+ lnox + ldis + ltax + llstat + chast),
nu ~ 1)
fit <- bamlss(f_bamlss,
data = bostonhouseprice2,
family = PE2,
optimizer = opt_lasso, # LASSO optimizer using BIC grid search
criterion = "BIC",
multiple = TRUE) # separate tuning parameters for each
# distributional parameter
\end{lstlisting}
As further comparator models, we also implement the adaptive LASSO (ALASSO) and LAD-LASSO methods, which are both location regression models with constant variance that are commonly used for normal and heavy-tailed scenarios respectively. In line with the other methods, we use the BIC to select penalty tuning parameters. For further details and code, see Appendix~\ref{app:sgnd_additional_code}.
\subsection{Boston Housing Data}\label{sec:sgnd_real_data_analyses_hprice}
\citet{harrison78hprice2} collected data from 506 suburb areas of Boston, consisting of housing prices and other characteristics. Due to the presence of outliers, analysis of this dataset is prominent in research relating to robust regression \citep{wu09variable, leng10ladlasso, amin15scad, amato21penalised}. The data are available in the \texttt{mlbench} package \citep{mlbench_pkg}. The response variable is the log(corrected median value of owner-occupied homes in \$1000's). There are 12 independent variables: log(full-value property-tax rate per \$10,000) (\texttt{ltax}), average number of rooms per dwelling (\texttt{rm}), log(weighted distances to five Boston employment centres) (\texttt{ldis}), log(percentage of ``lower status" in the population) (\texttt{llstat}), pupil-teacher ratio by town (\texttt{ptratio}), per capita crime rate by town (\texttt{crim}), index of accessibility to radial highways (\texttt{rad}), log(nitric oxides concentration in parts per ten million) (\texttt{lnox}), proportion of owner-occupied units built prior to 1940 (\texttt{age}), Charles River dummy variable (=1 if tract bounds river; 0 otherwise) (\texttt{chast}), proportion of residential land zoned for lots over 25,000 sq.\ ft (\texttt{zn}) and the proportion of non-retail business acres per town (\texttt{indus}).
\begin{table}[t!]
\caption{\label{tab:sgnd_hprice_vars}Boston Housing Data: variables selected and summary metrics}
\centering
\resizebox{\textwidth}{!}{
\setlength{\tabcolsep}{4pt}{
\begin{tabular}{@{}l@{~~} c@{~} c@{~~}c@{~~}c@{~~}c@{~~} c@{~~~} c@{~~}c@{~~}c@{~~}c@{}}
\toprule
{} && \multicolumn{4}{c}{Location \& Scale} && \multicolumn{3}{c}{Location Only} & {}\\
\cmidrule(r){2-6} \cmidrule(l){7-10}
{} && {\begin{tabular}{@{}c@{}}SGND \\ -MPR\end{tabular}} & {\begin{tabular}{@{}c@{}}GAMLSS \\ -STEP\end{tabular}} & {\begin{tabular}{@{}c@{}}GAMLSS \\ -BOOST\end{tabular}} & BAMLSS && {\begin{tabular}{@{}c@{}}SGND \\ -SPR\end{tabular}} & ALASSO & {\begin{tabular}{@{}c@{}}LAD \\ -LASSO\end{tabular}} & Total \\
\midrule
\texttt{ltax} && \cellcolor{mygreen} \textcolor{white}{$\beta$} & \cellcolor{mygreen} \textcolor{white}{$\beta$} & \cellcolor{mygreen} \textcolor{white}{$\beta$} & \cellcolor{mygreen} \textcolor{white}{$\beta$} && \cellcolor{mygreen} \textcolor{white}{$\beta$} & \cellcolor{mygreen} \textcolor{white}{$\beta$} & \cellcolor{mygreen} \textcolor{white}{$\beta$} & 7 \\
\texttt{rm} && \cellcolor{mygreen} \textcolor{white}{$\beta$} & \cellcolor{mygreen} \textcolor{white}{$\beta$} & \cellcolor{mygreen} \textcolor{white}{$\beta$} & \cellcolor{mygreen} \textcolor{white}{$\beta$} && \cellcolor{mygreen} \textcolor{white}{$\beta$} & \cellcolor{mygreen} \textcolor{white}{$\beta$} & \cellcolor{mygreen} \textcolor{white}{$\beta$} & 7 \\
\texttt{ldis} && \cellcolor{mygreen} \textcolor{white}{$\beta$} & \cellcolor{mygreen} \textcolor{white}{$\beta$} & {} & \cellcolor{mygreen} \textcolor{white}{$\beta$} && \cellcolor{mygreen} \textcolor{white}{$\beta$} & \cellcolor{mygreen} \textcolor{white}{$\beta$} & \cellcolor{mygreen} \textcolor{white}{$\beta$} & 6 \\
\texttt{llstat} && \cellcolor{mygreen} \textcolor{white}{$\beta$} & \cellcolor{mygreen} \textcolor{white}{$\beta$} & \cellcolor{mygreen} \textcolor{white}{$\beta$} & \cellcolor{mygreen} \textcolor{white}{$\beta$} && \cellcolor{mygreen} \textcolor{white}{$\beta$} & \cellcolor{mygreen} \textcolor{white}{$\beta$} & \cellcolor{mygreen} \textcolor{white}{$\beta$} & 7 \\
\texttt{ptratio} && \cellcolor{mygreen} \textcolor{white}{$\beta$} & \cellcolor{mygreen} \textcolor{white}{$\beta$} & \cellcolor{mygreen} \textcolor{white}{$\beta$} & \cellcolor{mygreen} \textcolor{white}{$\beta$} && \cellcolor{mygreen} \textcolor{white}{$\beta$} & \cellcolor{mygreen} \textcolor{white}{$\beta$} & \cellcolor{mygreen} \textcolor{white}{$\beta$} & 7 \\
\texttt{crim} && \cellcolor{mygreen} \textcolor{white}{$\beta$} & \cellcolor{mygreen} \textcolor{white}{$\beta$} & {} & \cellcolor{mygreen} \textcolor{white}{$\beta$} && \cellcolor{mygreen} \textcolor{white}{$\beta$} & \cellcolor{mygreen} \textcolor{white}{$\beta$} & \cellcolor{mygreen} \textcolor{white}{$\beta$} & 6 \\
\texttt{rad} && \cellcolor{mygreen} \textcolor{white}{$\beta$} & \cellcolor{mygreen} \textcolor{white}{$\beta$} & {} & \cellcolor{mygreen} \textcolor{white}{$\beta$} && \cellcolor{mygreen} \textcolor{white}{$\beta$} & \cellcolor{mygreen} \textcolor{white}{$\beta$} & \cellcolor{mygreen} \textcolor{white}{$\beta$} & 6 \\
\texttt{lnox} && \cellcolor{mygreen} \textcolor{white}{$\beta$} & \cellcolor{mygreen} \textcolor{white}{$\beta$} & {} & \cellcolor{mygreen} \textcolor{white}{$\beta$} && \cellcolor{mygreen} \textcolor{white}{$\beta$} & \cellcolor{mygreen} \textcolor{white}{$\beta$} & \cellcolor{mygreen} \textcolor{white}{$\beta$} & 6 \\
\texttt{age} && \cellcolor{mygreen} \textcolor{white}{$\beta$} & {} & {} & \cellcolor{mygreen} \textcolor{white}{$\beta$} && {} & {} & {} & 2 \\
\texttt{chast} && \cellcolor{mygreen} \textcolor{white}{$\beta$} & \cellcolor{mygreen} \textcolor{white}{$\beta$} & {} & {} && \cellcolor{mygreen} \textcolor{white}{$\beta$} & \cellcolor{mygreen} \textcolor{white}{$\beta$} & \cellcolor{mygreen} \textcolor{white}{$\beta$} & 5 \\
\texttt{zn} && {} & {} & \cellcolor{mygreen} \textcolor{white}{$\beta$} & {} && {} & {} & \cellcolor{mygreen} \textcolor{white}{$\beta$} & 2 \\
\texttt{indus} && {} & {} & {} & {} && {} & {} & {} & 0 \\
\midrule
\texttt{ltax} && \cellcolor{mygreen} \textcolor{white}{$\alpha$} & \cellcolor{mygreen} \textcolor{white}{$\alpha$} & \cellcolor{mygreen} \textcolor{white}{$\alpha$} & \cellcolor{mygreen} \textcolor{white}{$\alpha$} && {} & {} & {} & 4 \\
\texttt{rm} && {} & {} & {} & \cellcolor{mygreen} \textcolor{white}{$\alpha$} && {} & {} & {} & 1 \\
\texttt{ldis} && \cellcolor{mygreen} \textcolor{white}{$\alpha$} & \cellcolor{mygreen} \textcolor{white}{$\alpha$} & \cellcolor{mygreen} \textcolor{white}{$\alpha$} & \cellcolor{mygreen} \textcolor{white}{$\alpha$} && {} & {} & {} & 4 \\
\texttt{llstat} && {} & {} & {} & {} && {} & {} & {} & 0 \\
\texttt{ptratio} && {} & {} & {} & {} && {} & {} & {} & 0 \\
\texttt{crim} && {} & {} & \cellcolor{mygreen} \textcolor{white}{$\alpha$} & {} && {} & {} & {} & 1 \\
\texttt{rad} && \cellcolor{mygreen} \textcolor{white}{$\alpha$} & \cellcolor{mygreen} \textcolor{white}{$\alpha$} & \cellcolor{mygreen} \textcolor{white}{$\alpha$} & \cellcolor{mygreen} \textcolor{white}{$\alpha$} && {} & {} & {} & 4 \\
\texttt{lnox} && {} & {} & {} & {} && {} & {} & {} & 0 \\
\texttt{age} && {} & {} & {} & {} && {} & {} & {} & 0 \\
\texttt{chast} && {} & {} & \cellcolor{mygreen} \textcolor{white}{$\alpha$} & {} && {} & {} & {} & 1 \\
\texttt{zn} && {} & {} & \cellcolor{mygreen} \textcolor{white}{$\alpha$} & {} && {} & {} & {} & 1 \\
\texttt{indus} && {} & {} & \cellcolor{mygreen} \textcolor{white}{$\alpha$} & {} && {} & {} & {} & 1 \\
\midrule
No. Selected && 13 & 12 & 12 & 13 && 9 & 9 & 10 & {} \\
BIC && -457 & -451 & -221 & -423 && -289 & -222 & -283 & {} \\
$\hat \kappa$ && 1.52 & 1.61 & 1.83 & 1.51 && 0.86 & - & - & {}\\
Time* (mins) && 1.42 & 2.24 & 2.80 & 559 && 0.25 & 0.002 & 1.20 & {}\\
\bottomrule
\multicolumn{11}{p{1\textwidth}}{\footnotesize *Intel(R) Core(TM) i7-10610U CPU @ 1.80GHz 2.30 GHz.}\\
\end{tabular}}}
\end{table}
Table~\ref{tab:sgnd_hprice_vars} provides summary metrics and an overview of the variables selected by each method. The SGND-MPR and GAMLSS-STEP methods select almost the same set of covariates in both the location and the scale. In the location, all covariates apart from \texttt{zn} and \texttt{indus} are selected (and \texttt{age} is also not selected by GAMLSS-STEP but is selected by SGND-MPR), while, in the scale, only \texttt{ltax}, \texttt{ldis}, and \texttt{rad} are selected. BAMLSS also yields a similar selection, but does not select \texttt{chast} in the location and additionally selects \texttt{rm} in the scale. The GAMLSS-BOOST method selects quite a different set of covariates, and has a much higher BIC value of -221 compared to the other three distributional regression approaches, which have BIC values of -457, -451, and -423, respectively. Indeed, it would appear that GAMLSS-BOOST under-selects in the location and over-selects in the scale. The single-parameter models (SGND-SPR, ALASSO and LAD-LASSO) select the same covariates (except that LAD-LASSO selects \texttt{zn}), and this set aligns with the SGND-MPR, GAMLSS-STEP, and BAMLSS methods (except that \texttt{age} is not selected).
In the single-parameter regression models, the SGND-SPR and LAD-LASSO have more similar BIC values to each other (-289 and -283) than the ALASSO, which has a higher value (-222); this to be expected given the heavy-tailedness of the residuals in the data, as discussed in Section~\ref{sec:sgnd_introduction}. However, a more pertinent point here is that these single-parameter regression models have much higher BIC values than the distributional regression models (ignoring GAMLSS-BOOST). This demonstrates the importance of accounting for heteroscedasticity by modelling the scale in addition to the location. It is also interesting to note that the SGND-SPR method estimates quite a heavy-tailed distribution with $\hat \kappa = 0.86$. In contrast, the SGND-MPR, GAMLSS-STEP, and BAMLSS methods (which capture heteroscedasticity) yield $\hat \kappa$ values approximately between 1.5 and 1.6, implying an error distribution with tails that are heavier than normal but lighter than Laplace. Therefore, it appears that the very heavy tails suggested by the single-parameter regression analysis is, in part, an artifact of the heteroscedasticity for which they fail to account.
In terms of computational expense, the SGND-MPR method is the fastest of the distributional regression methods, taking 1.42 minutes. This is approximately 1.6 times faster than the GAMLSS-STEP approach, and two times faster than the GAMLSS-BOOST procedure. The BAMLSS method performs a three-dimensional grid search and takes over nine hours, making it very computationally demanding. For the single-parameter models, the SGND-SPR method is 4.8 times faster than the LAD-LASSO search. Note that the ALASSO method is compiled in \texttt{C} code, and so is the fastest overall.
\begin{figure}[t!]
\centering
\makebox{
\begin{subfigure}[h]{\textwidth}
\centering
\includegraphics[width = 0.95\textwidth]{figs/hprice_path.pdf}%
\end{subfigure}}
\makebox{
\begin{subfigure}[h]{\textwidth}
\centering
\includegraphics[trim = {0.2cm 0 0.2cm 0.1cm}, clip, scale = 0.7]{figs/hprice_leg.pdf}%
\end{subfigure}}
\caption{\label{figs:sgnd_hprice_path}Boston Housing Data: standardized coefficient values through the $\epsilon$-telescope for the location and scale components. Coloured lines indicate the selected variables, with grey lines signifying variables that are set to zero and not included in the final model.}
\end{figure}
The standardized coefficient values with respect to the $\epsilon$-telescope of the SGND-MPR method are shown in Figure~\ref{figs:sgnd_hprice_path}. This demonstrates how the proposed method operates as $\epsilon$ moves towards zero to carry out variable selection by approximating $L_0$ penalization; indeed, by following the coefficient paths from right to left, the unimportant covariate effects tend towards zero. Two variables are eliminated from the location component, with \texttt{zn} tending to zero first, which is then followed closely by \texttt{indus} (both in grey). Although non-zero (i.e.,~selected), the \texttt{chast} and \texttt{age} effects are the weakest --- and note that BAMLSS did not select the former while GAMLSS-STEP did not select the latter. In contrast to the location, where ten variables are selected, only three variables are selected in the scale (\texttt{ltax}, \texttt{ldis} and \texttt{rad}). Interestingly, one of the last scale effects to be set to zero in the $\epsilon$-telescope is \texttt{rm}, and this variable was selected by BAMLSS.
\begin{table}[t!]
\caption{\label{tab:sgnd_dataset_estimates_hprice_mpr}Boston Housing Data: estimation metrics}
\centering
\begin{tabular}{@{}l@{~~} r@{~}c@{~}r@{~~} c@{\qquad} r@{~}c@{~}r@{~~} c@{\qquad} r@{}}
\toprule
{} & \multicolumn{9}{c}{SGND-MPR: $\hat \nu_0 = 0.28 \, (0.11)$}\\
\cmidrule(){2-10}
{} & \multicolumn{2}{c}{$\hat\beta_j$} & \multicolumn{1}{c}{$\Delta\text{BIC}_\beta$} && \multicolumn{2}{c}{$\hat\alpha_j$} & \multicolumn{1}{c}{$\Delta\text{BIC}_\alpha$} && \multicolumn{1}{c}{$\Delta\text{BIC}_{\beta \alpha}$} \\
\midrule
\texttt{intercept} &3.67 & (0.13) & && -8.30 & (1.97) & && \\
\texttt{ltax} &-0.19 & (0.02) & 46 && 0.85 & (0.33) & 146 && 266 \\
\texttt{rm} &0.23 & (0.02) & 158 && & & && 158 \\
\texttt{ldis} &-0.16 & (0.03) & 57 && -0.85 & (0.18) & 17 && 82 \\
\texttt{llstat} &-0.17 & (0.02) & 72 && & & && 72 \\
\texttt{ptratio} &-0.02 & (0.00) & 63 && & & && 63 \\
\texttt{crim} &-0.02 & (0.00) & 40 && & & && 40 \\
\texttt{rad} &0.01 & (0.00) & 12 && 0.05 & (0.02) & 6 && 27 \\
\texttt{lnox} &-0.30 & (0.13) & 11 && & & && 11 \\
\texttt{age} &-0.00 & (0.00) & 5 && & & && 5 \\
\texttt{chast} &0.05 & (0.02) & 1 && & & && 1 \\
\bottomrule
\multicolumn{10}{p{0.5\textwidth}}{\footnotesize Variables not selected: \texttt{zn}, \texttt{indus}.}\\
\end{tabular
\end{table}
For the SGND-MPR method, Table~\ref{tab:sgnd_dataset_estimates_hprice_mpr} shows the estimates and associated standard errors (in brackets) along with the $\Delta$BIC value, which is a measure of the effect of the variable and is computed as the change in BIC after removing the variable from the location ($\Delta\text{BIC}_\beta$) or scale ($\Delta\text{BIC}_\alpha$) component of the model. The overall effect of the variable (i.e.,~removal from both the location and scale components simultaneously) is denoted by $\Delta\text{BIC}_{\beta \alpha}$.\ (See the Supplementary Material for the outputs of the other methods.)\ The average number of rooms (\texttt{rm}) has a positive coefficient, indicating that more rooms in a dwelling are associated with increased house prices. Dropping this variable from the model results in an increase in the BIC of 158 units, which is substantially greater than the effect of the other variables in the location component in terms of BIC. Proximity to highways (\texttt{rad}) and houses that are situated on the banks of the Charles River (\texttt{chast}) are also linked to greater house prices, but contribute much less in terms of BIC (12 and 1 units). Communities with the following characteristics are associated with lower house prices: large property-tax rates (\texttt{ltax}), longer distances to work (\texttt{ldis}), a high percentage of ``lower status" individuals (\texttt{llstat}), high pupil teacher ratios (\texttt{ptratio}), high crime rates (\texttt{crim}), high nitric oxides concentrations (\texttt{lnox}) and the presence of older houses (\texttt{age}). Note that the top five variables in terms of BIC impact in the location component (\texttt{ltax}, \texttt{rm}, \texttt{ldis}, \texttt{llstat} and \texttt{ptratio}) appear in the location components of all of the models from Table~\ref{tab:sgnd_hprice_vars} (apart from absence of \texttt{ldis} in GAMLSS-BOOST), which is another signal of their importance.
\begin{figure}[t!]
\centering
\makebox{\includegraphics[width = 0.75\textwidth]{figs/hprice_dens.pdf}}
\caption{\label{figs:sgnd_hprice_densities}Boston Housing Data: SGND-MPR model-based conditional density curves. Keeping the covariates fixed at the median values, the red and blue densities correspond to the modification of the presented covariate as: ``low" (Q1, the first quartile) and ``high" (Q3, the third quartile).}
\end{figure}
In the case of \texttt{ltax}, although it has a negative coefficient in the location parameter, it has a positive coefficient in the scale component. Therefore, lower values of property-tax are associated with lower house prices, but also increased variability. Removal of \texttt{ltax} from the scale component results in an increase of 146 units, meaning it is the variable with the second-largest effect in terms of BIC (after \texttt{rm} in the location component with 158 units). This is also far greater than the effect of removing it from the location parameter (and all other variables in the location apart from \texttt{rm}), demonstrating the importance of modelling the scale in addition to the location in general. Removing \texttt{ltax} from both the location and scale components simultaneously results in an increase in BIC of 266 units, i.e., this is the most important covariate overall. To visualize the effect of \texttt{ltax}, Figure~\ref{figs:sgnd_hprice_densities} shows model-based conditional density curves, motivated by \citet{stadlmann21interactively}; \texttt{rm} and \texttt{ldis} are also shown and are the second and third most important variables overall. For higher values of \texttt{ltax}, there is a clear location shift to the left and greater dispersion in the associated conditional density plots. The \texttt{rm} variable only impacts the location, while the \texttt{ldis} variable has negative coefficients in both the location and scale components (reduced average house prices and variability).
\subsection{Diabetes Data}\label{sec:sgnd_real_data_analyses_diabetes}
The diabetes data come from a study where 442 diabetic patients were examined and had ten baseline variables recorded as well as the response of interest, a quantitative measure of disease progression one year after baseline. This data set is analyzed in \citet{efron04} and \citet{leng10ladlasso} and is available in the \texttt{lars} package \citep{lars_pkg}. The baseline variables are body mass index (\texttt{BMI}), average blood pressure (\texttt{BP}), sex (=1 if female, =2 if male) (\texttt{SEX}), age (\texttt{AGE}), and six blood serum measurements (\texttt{S1}, \texttt{S2}, \texttt{S3}, \texttt{S4}, \texttt{S5}, \texttt{S6}). The assumption of normality of the error distribution appears to be appropriate, as shown in Figure~\ref{figs:sgnd_diabets}(a.i) and (a.ii), which display the density and Q-Q plot of the standardized residuals from fitting a normal linear regression model. A Shapiro-Wilk test on these residuals does not provide evidence against normality (p-value = 0.6). Moreover, a bootstrapped 95\% confidence interval of the excess kurtosis contains zero [-0.52, 0.06], indicating a mesokurtic distribution, which has tails similar to that of a normal distribution. Figure~\ref{figs:sgnd_diabets}(b) shows the standardized residuals from fitting a linear model plotted against \texttt{BMI}, which highlights the possible presence of heteroscedasticity (albeit arguably not as strong as in the Boston housing data).
\begin{figure}[b!]
\centering
\makebox{\includegraphics[width = \textwidth]{figs/diabetes_summ.pdf}}
\caption{\label{figs:sgnd_diabets}Diabetes Data: density (a.i) and Q-Q (a.ii) plot of the standardized residuals from fitting a linear model. Standardized residuals plotted against the independent variable BMI (b).}
\end{figure}
\begin{table}[t!]
\caption{\label{tab:sgnd_diabetes_vars}Diabetes Data: variables selected and summary metrics}
\centering
\resizebox{\textwidth}{!}{
\setlength{\tabcolsep}{4pt}{
\begin{tabular}{@{}l@{~~} c@{~} c@{~~}c@{~~}c@{~~}c@{~~} c@{~~~} c@{~~}c@{~~}c@{~~}c@{}}
\toprule
{} && \multicolumn{4}{c}{Location \& Scale} && \multicolumn{3}{c}{Location Only} & {}\\
\cmidrule(r){2-6} \cmidrule(l){7-10}
{} && {\begin{tabular}{@{}c@{}}SGND \\ -MPR\end{tabular}} & {\begin{tabular}{@{}c@{}}GAMLSS \\ -STEP\end{tabular}} & {\begin{tabular}{@{}c@{}}GAMLSS \\ -BOOST\end{tabular}} & BAMLSS && {\begin{tabular}{@{}c@{}}SGND \\ -SPR\end{tabular}} & ALASSO & {\begin{tabular}{@{}c@{}}LAD \\ -LASSO\end{tabular}} & Total \\
\midrule
\texttt{BMI} && \cellcolor{mygreen} \textcolor{white}{$\beta$} & \cellcolor{mygreen} \textcolor{white}{$\beta$} & \cellcolor{mygreen} \textcolor{white}{$\beta$} & \cellcolor{mygreen} \textcolor{white}{$\beta$} && \cellcolor{mygreen} \textcolor{white}{$\beta$} & \cellcolor{mygreen} \textcolor{white}{$\beta$} & \cellcolor{mygreen} \textcolor{white}{$\beta$} & 7 \\
\texttt{S5} && \cellcolor{mygreen} \textcolor{white}{$\beta$} & \cellcolor{mygreen} \textcolor{white}{$\beta$} & \cellcolor{mygreen} \textcolor{white}{$\beta$} & {} && \cellcolor{mygreen} \textcolor{white}{$\beta$} & \cellcolor{mygreen} \textcolor{white}{$\beta$} & \cellcolor{mygreen} \textcolor{white}{$\beta$} & 6 \\
\texttt{S3} && \cellcolor{mygreen} \textcolor{white}{$\beta$} & \cellcolor{mygreen} \textcolor{white}{$\beta$} & {} & \cellcolor{mygreen} \textcolor{white}{$\beta$} && {} & {} & \cellcolor{mygreen} \textcolor{white}{$\beta$} & 4 \\
\texttt{BP} && \cellcolor{mygreen} \textcolor{white}{$\beta$} & \cellcolor{mygreen} \textcolor{white}{$\beta$} & {} & \cellcolor{mygreen} \textcolor{white}{$\beta$} && \cellcolor{mygreen} \textcolor{white}{$\beta$} & \cellcolor{mygreen} \textcolor{white}{$\beta$} & \cellcolor{mygreen} \textcolor{white}{$\beta$} & 6 \\
\texttt{SEX} && \cellcolor{mygreen} \textcolor{white}{$\beta$} & \cellcolor{mygreen} \textcolor{white}{$\beta$} & {} & \cellcolor{mygreen} \textcolor{white}{$\beta$} && \cellcolor{mygreen} \textcolor{white}{$\beta$} & \cellcolor{mygreen} \textcolor{white}{$\beta$} & \cellcolor{mygreen} \textcolor{white}{$\beta$} & 6 \\
\texttt{S1} && {} & {} & {} & \cellcolor{mygreen} \textcolor{white}{$\beta$} && \cellcolor{mygreen} \textcolor{white}{$\beta$} & \cellcolor{mygreen} \textcolor{white}{$\beta$} & \cellcolor{mygreen} \textcolor{white}{$\beta$} & 4 \\
\texttt{S2} && {} & {} & {} & \cellcolor{mygreen} \textcolor{white}{$\beta$} && \cellcolor{mygreen} \textcolor{white}{$\beta$} & \cellcolor{mygreen} \textcolor{white}{$\beta$} & {} & 3 \\
\texttt{S4} && {} & {} & {} & {} && {} & \cellcolor{mygreen} \textcolor{white}{$\beta$} & {} & 1 \\
\texttt{S6} && {} & {} & {} & {} && {} & {} & \cellcolor{mygreen} \textcolor{white}{$\beta$} & 1 \\
\texttt{AGE} && {} & {} & {} & {} && {} & {} & {} & 0 \\
\midrule
\texttt{BMI} && \cellcolor{mygreen} \textcolor{white}{$\alpha$} & \cellcolor{mygreen} \textcolor{white}{$\alpha$} & \cellcolor{mygreen} \textcolor{white}{$\alpha$} & {} && {} & {} & {} & 3 \\
\texttt{S5} && {} & {} & {} & \cellcolor{mygreen} \textcolor{white}{$\alpha$} && {} & {} & {} & 1 \\
\texttt{S3} && {} & {} & {} & \cellcolor{mygreen} \textcolor{white}{$\alpha$} && {} & {} & {} & 1 \\
\texttt{BP} && {} & {} & \cellcolor{mygreen} \textcolor{white}{$\alpha$} & \cellcolor{mygreen} \textcolor{white}{$\alpha$} && {} & {} & {} & 2 \\
\texttt{SEX} && {} & {} & {} & \cellcolor{mygreen} \textcolor{white}{$\alpha$} && {} & {} & {} & 1 \\
\texttt{S1} && {} & {} & {} & \cellcolor{mygreen} \textcolor{white}{$\alpha$} && {} & {} & {} & 1 \\
\texttt{S2} && {} & {} & \cellcolor{mygreen} \textcolor{white}{$\alpha$} & \cellcolor{mygreen} \textcolor{white}{$\alpha$} && {} & {} & {} & 2 \\
\texttt{S4} && {} & {} & {} & {} && {} & {} & {} & 0 \\
\texttt{S6} && {} & {} & \cellcolor{mygreen} \textcolor{white}{$\alpha$} & {} && {} & {} & {} & 1 \\
\texttt{AGE} && {} & {} & {} & {} && {} & {} & {} & 0 \\
\midrule
No. Selected && 6 & 6 & 6 & 12 && 6 & 7 & 7 & {} \\
BIC && 4819 & 4819 & 4991 & 4886 && 4828 & 4829 & 4906 & {} \\
$\hat \kappa$ && 2.41 & 2.65 & 4.78 & 2.26 && 2.13 & - & - & {} \\
Time* (mins) && 1.69 & 0.20 & 40.73 & 2115 && 0.39 & 0.01 & 0.66 & {} \\
\bottomrule
\multicolumn{11}{p{1\textwidth}}{\footnotesize *Intel(R) Core(TM) i7-10610U CPU @ 1.80GHz 2.30 GHz.}\\
\end{tabular}}}
\end{table}
The variables selected by each method and additional summary metrics are shown in Table~\ref{tab:sgnd_diabetes_vars}. Unlike in the Boston housing data, heteroscedasticity is much less of a feature for the diabetes data since, although the lowest BIC (4819) is achieved by two distributional regression models (SGND-MPR and GAMLSS-STEP, which both select the same set of covariates in both the location and the scale), the second-lowest BIC (4828) corresponds to a single-parameter regression model (SGND-SPR). Moreover, both SGND-MPR and GAMLSS-STEP only select one scale effect (\texttt{BMI}), whereas BAMLSS appears to overestimate the extent of the heteroscedasticity with six variables selected (though not \texttt{BMI}) and also has one of the higher BIC values (4886). Similar to the Boston housing data, the GAMLSS-BOOST method has the highest BIC value (4991) among the models considered and appears to be under-selecting in the location component and over-selecting in the scale component. In the location component, \texttt{BMI} is the only covariate selected by all models; \texttt{S5} is selected by all but BAMLSS, and \texttt{BP} and \texttt{SEX} are selected by all but GAMLSS-BOOST. The models also agree that \texttt{AGE} is not important, while the \texttt{S4} and \texttt{S6} variables are each selected by just one single-parameter regression approach (ALASSO and LAD-LASSO, respectively); however, there is less agreement on the variables \texttt{S1}, \texttt{S2} and \texttt{S3}. In terms of the distributional shape, all of the methods agree on a light-tailed error distribution (in line with Figure~\ref{figs:sgnd_diabets}). GAMLSS-STEP estimates $\hat \kappa = 2.65$, which indicates an error distribution with lighter-than-normal tails, and the two SGND approaches (MPR and SPR) and BAMLSS suggest tails that are closer to that of a normal distribution ($\hat \kappa = 2.41$, $2.13$ and, $2.26$ respectively). The GAMLSS-BOOST approach suggests an error distribution with tails much lighter than the normal distribution with $\hat \kappa = 4.78$. Unsurprisingly then, the LAD-LASSO, which imposes heavy tails, has the second-highest BIC value (4906).
\begin{table}[b!]
\caption{\label{tab:sgnd_dataset_estimates_diabetes_mpr}Diabetes Data: estimation metrics}
\centering
\begin{tabular}{@{}l@{~~} r@{~}c@{~}r@{~~} c@{\qquad} r@{~}c@{~}r@{~~} c@{\qquad} r@{}}
\toprule
{} & \multicolumn{9}{c}{SGND-MPR: $\hat \nu_0 = 0.79 \, (0.15)$}\\
\cmidrule(){2-10}
{} & \multicolumn{2}{c}{$\hat\beta_j$} & \multicolumn{1}{c}{$\Delta\text{BIC}_{\beta}$} && \multicolumn{2}{c}{$\hat\alpha_j$} & \multicolumn{1}{c}{$\Delta\text{BIC}_{\alpha}$} && \multicolumn{1}{c}{$\Delta\text{BIC}_{\beta \alpha}$}\\
\midrule
\texttt{intercept} & -183.89 & (36.18) & && 7.04 & (0.44) & && \\
\texttt{BMI} & 5.37 & (0.72) & 44 && 0.06 & (0.02) & 7 && 65 \\
\texttt{S5} & 40.77 & (5.77) & 43 && & & && 43 \\
\texttt{S3} & -1.03 & (0.24) & 13 && & & && 13 \\
\texttt{BP} & 0.92 & (0.22) & 12 && & & && 12 \\
\texttt{SEX} & -20.46 & (5.89) & 6 && & & && 6 \\
\bottomrule
\multicolumn{10}{p{0.5\textwidth}}{\footnotesize Variables not selected: \texttt{S1}, \texttt{S2}, \texttt{S4}, \texttt{S6}, \texttt{AGE}.}\\
\end{tabular
\end{table}
\begin{figure}[h!]
\centering
\makebox{
\begin{subfigure}[h]{\textwidth}
\centering
\includegraphics[width = 0.95\textwidth]{figs/diabetes_path.pdf}%
\end{subfigure}}
\makebox{
\begin{subfigure}[h]{\textwidth}
\centering
\includegraphics[trim = {0.2cm 0 0.2cm 0.1cm}, clip, scale = 0.7]{figs/diabetes_leg.pdf}%
\end{subfigure}}
\caption{\label{fig:sgnd_diabetes_path}Diabetes Data: standardized coefficient values through the $\epsilon$-telescope for the location and scale components. Coloured lines indicate the selected variables, with grey lines signifying variables that are set to zero and not included in the final model.}
\end{figure}
\begin{figure}[t!]
\centering
\makebox{\includegraphics[width = 0.75\textwidth]{figs/diabetes_dens.pdf}}
\caption{\label{figs:sgnd_diabetes_densities}Diabetes Data: SGND-MPR model-based conditional density curves. Keeping the covariates fixed at the median values, the red and blue densities correspond to the modification of the presented covariate as: ``low" (Q1, the first quartile) and ``high" (Q3, the third quartile).}
\end{figure}
Table~\ref{tab:sgnd_dataset_estimates_diabetes_mpr} displays the estimates (with coefficient paths visualized in Figure~\ref{fig:sgnd_diabetes_path}), standard errors and $\Delta$BIC values for the SGND-MPR method (estimation metrics for the other methods are available in the Supplementary Material). For the location component, the top two variables in terms of BIC impact are \texttt{BMI} and \texttt{S5} as previously suggested by Table~\ref{tab:sgnd_diabetes_vars}. Both of these variables are associated with increased disease progression, as is \texttt{BP}, while \texttt{S3} and \texttt{SEX} are associated with reduced disease progression. In the case of \texttt{SEX}, this translates to progression being lower in males. The only scale effect identified here is \texttt{BMI}, such that larger BMI values are associated with increased variability. However, this scale effect ($\Delta\text{BIC}_\alpha = 7$) is less important than the top four location effects, i.e., heteroscedasticity is not a strong feature of this dataset. Nevertheless, when modelled simultaneously in the location and scale, \texttt{BMI} is the most important variable ($\Delta\text{BIC}_{\beta \alpha}$ = 65). The effect of the top three variables in terms of overall BIC impact (\texttt{BMI}, \texttt{S5} and \texttt{S3}) are also visualized in Figure~\ref{figs:sgnd_diabetes_densities} using model-based conditional density curves.
\section{Discussion} \label{sec:sgnd_discussion}
Our proposed SGND-SIC method provides a flexible framework that (i) allows for the error distribution to move between (or beyond) robust and classical normal linear regression, (ii) accounts for heteroscedasticity by taking a distributional regression approach where the scale is modelled on covariates in addition to the location, and (iii) carries out automatic variable selection in both distributional regression components (location and scale) in a computationally feasible manner; we have also implemented the approach in the \texttt{smoothic} package in \texttt{R} \citep{smoothic_package}. The presence of outliers and heteroscedasticity can distort parameter estimates and make standard inferences unreliable. Through the estimation of the shape parameter in the smooth generalized normal distribution, our proposed method can move smoothly between the normal distribution and the heavier-tailed Laplace distribution. Therefore, one does not have to make an outright decision between classical and robust regression, as both are covered. Moreover, allowing the scale to depend on covariates reduces the risk of information loss, as examination of regression effects beyond the mean is key to identifying important dynamics in the data. Currently, variable selection is somewhat underdeveloped in the distributional regression setting, with existing penalization methods being particularly computationally intensive in this setting. We avoid the typical multidimensional grid search for optimal tuning parameter values, as we know from the outset that $\lambda = \log(n)$ in the BIC case. Our smooth objective function can be optimized directly using standard gradient based techniques and performs comparably to existing procedures as illustrated in the real data analyses.
Through extensive simulation studies, we have shown that our procedure performs favorably in terms of variable selection and parameter inference over the full range of $\kappa$ values that are tested. We have also demonstrated the importance of accounting for both heavy tails and heteroscedasticity when they exist. Although standard error breakdown appears to occur when the combination of heavy tails and heteroscedasticity is observed in smaller samples, the procedure performs well asymptotically --- and in smaller sample sizes when only one of heavy tails or heteroscedasticity is observed. Nevertheless, it will be useful to investigate alternative standard error calculations in our future work. Moreover, it is also of interest to extend our approach to non-symmetric distributions and other parameter types beyond location and scale; indeed, there may be merit in some settings in modelling higher-order shape parameters, such as $\kappa$ in the generalized normal distribution.
\section*{Acknowledgements}
This work was carried out within the Confirm Smart Manufacturing Research Centre (https://confirm.ie/) funded by Science Foundation Ireland (grant number: 16/RC/3918).
\bibliographystyle{apalike}
|
\section{Introduction}
The concept of cographs was unfolded independently by several researchers \cite{Corneil 1, Jung,Lerchs, Seinsche, Sumner} around $1970$. They give a simple structural decomposition with disjoint union and graph complement operations that can be described by a labelled tree which are used algorithmically to efficiently solve many hard problems, like, finding the maximum clique.
Let $P_n$, $C_n$ and $K_n$ denote the path, the cycle and the complete graph on $n$ vertices respectively. A $P_4$-free graph is called a cograph. A cograph can be constructed from a single vertex by joining another cograph and by taking compliments. The class of all cographs is closed under the following operations:
\begin{itemize}
\item
disjoint union of graphs.
\item
complementation of a graph.
\end{itemize}
In this paper we only consider those cographs which can be constructed in the following way:
Let $\alpha_1,\alpha_2,\ldots,\alpha_m$ be natural numbers and $\alpha_1+\alpha_2+\cdots+\alpha_m=n$. We recursively define a cograph, denoted by $C(\alpha_1,\alpha_2,\ldots,\alpha_m)$, which satisfies the recurrence relation:
$$C(\alpha_1,\alpha_2,\ldots,\alpha_i)=\overline{C(\alpha_1,\alpha_2,\ldots,\alpha_{i-1})\cup K_{\alpha_i}},$$
for $i=2,3,\ldots,m$, where $C(\alpha_1)=\overline{K}_{\alpha_1}$.
In other wards, to construct $C(\alpha_1,\alpha_2,\ldots,\alpha_m)$, we start with $K_{\alpha_1}$, then take its compliment. In the next step, we take disjoint union of $K_{\alpha_2}$ with the graph obtained in the first step and then take its complement. Proceeding in this way, finally, we take disjoint union of $K_{\alpha_m}$ with the graph obtained in the $(m-1)$-th step and then take its complement. A construction of the 10 vertex graph $C(3,2,2,3)$ is illustrated in the Figure \ref{Cgraph_fig1}.
\begin{figure}[h]
\includegraphics[width=\textwidth]{fig1.pdf}
\caption{Construction of the $\mathcal{C}$-graph $C(3,2,2,3)$}
\label{Cgraph_fig1}
\end{figure}
Let $\mathcal{C}$ denote the class of cographs which can be constructed in the above way, and if $G\in \mathcal{C}$ then we call $G$ a $\mathcal{C}$-graph. Throughout the paper we consider the $m$ to be even, i.e, $m=2k$ for some $k\in \mathbb{N}$, and we use $\mathcal{C}_{even}$ to represent the class of all such $\mathcal{C}$-graphs which can be constructed in even number of steps. \medskip
A graph is called a threshold graph \cite{Aguilar1,Aguilar2,Bapat} if it is $P_4,~C_4$ and $2K_2$ free. Thus if $G$ is a threshold graph, then it is a cograph, more precisely the class of threshold graphs is a subclass of the class of cographs. But in general a threshold graph $G$ is not a $\mathcal{C}$-graph, and if $G\in\mathcal{C}$ then $G$ must be the antiregular graph (a connected graph in which atmost two vertex can have the same degree) $C(1,1,1,\cdots,1)$ or $C(1,2,1,\cdots,1)$. For an undirected, simple graph $G$ with $n$ vertices, its adjacency matrix (generally denoted by $A$) is a square matrix whose $(i,j)$-th entry is $1$ if the vertices $i$ and $j$ are connected by an edge and $0$ otherwise.
We denote the $n$-vertex antiregular graph by $H_n$ and its adjacency matrix by $A_n$. \medskip
We now show that if $G\in \mathcal{C}_{even}$, then its representation is unique.
\begin{theorem}
Let $G$ be a $\mathcal{C}$-graph with n vertices. If $G\in\mathcal{C}_{even}$, then there exists an unique integer partition $\{\alpha_1,\alpha_2,\ldots,\alpha_{2k}\}$ of $n$ such that $G=C(\alpha_1,\alpha_2,\ldots,\alpha_{2k})$. In other wards, the representation of $G$ is unique.
\end{theorem}
\begin{proof}
First we observe that the largest possible order of an antiregular subgraph of $C(\alpha_1,\alpha_2,\ldots,\alpha_{2k})$ is $2k$ or $2k+1$.
Therefore, if $G=C(\alpha_1,\alpha_2,\ldots,\alpha_{2k})=C(\beta_1,\beta_2,\ldots,\beta_{2l})$, then we must have $k=l$. Now we show that $\alpha_i=\beta_i$ for $i=1,2,\ldots,2k$.
Let $f:C(\alpha_1,\alpha_2,\ldots,\alpha_{2k})\rightarrow C(\beta_1,\beta_2,\ldots,\beta_{2k})$ be an isomorphism. Let $H$ be the set of vertices which are added in the final step in the construction of $C(\alpha_1,\alpha_2,\ldots,\alpha_{2k})$. Then $H$ is an independent subset of $C(\alpha_1,\alpha_2,\ldots,\alpha_{2k})$ and $d(u)=n-\alpha_{2k}$ for all $u\in H$. Then $f(H)$ must be an independent set and every $u\notin f(H)$ is adjacent to every vertex in $f(H)$. Which implies that $\alpha_{2k}=\beta_{2k}$. Then $C(\alpha_1,\alpha_2,\ldots,\alpha_{2k})\setminus H$ is isomorphic to $C(\beta_1,\beta_2,\ldots,\beta_{2k})\setminus f(H)$. Then applying similar argument to their complement we obtain $\alpha_{2k-1}=\beta_{2k-1}$. Proceeding in this way we conclude that $\alpha_i=\beta_i$ for all $i=1,2,\ldots,k$. Which completes the proof of the theorem.
\end{proof}
In this paper we consider the adjacency matrix of a class of cograph, namely $\mathcal{C}$-graphs. Spectral characterization of cographs has been done previously by many authors \cite{Tura 3, Tura 4, Stanic, Chang, Ghorbani 1, He, Jacobs 2,Trevisan 1, Royle, Sander}. We mention some notable works below:
\begin{itemize}
\item
The rank of a cograph is equal to the number of distinct non-zero rows of $A$ (Royle \cite{Royle}).
\item
B\i y\i ko\u{g}lu et al.~\cite{Stanic} proved that $0,-1$ belong to the spectrum of a connected cograph if and only if it contains duplicate (resp. coduplicate) vertices.
\item
In \cite{Trevisan 1}, Mohhamadian and Trevisan shown that $(-1,~0)$ is an eigenvalue free interval for the adjacency matrix of a cograph.
\item
Jacobs et al. \cite{Jacobs 2} obtain a formula for the inertia of a cograph and listed a pairs of equienergetic cographs.
\item
Allem and Tura \cite{Tura 3,Tura 4} wrote several papers on spectral properties of cographs. They investigated the multiplicity of eigenvalue $\lambda\neq 0,~-1$. They obtained a family of non-cospectral and borderenergetic cographs. They also constructed a class of integral cographs.
\item
He et al.~\cite{He} established the validity of a conjecture related to the signature of chordal graphs and cographs.
\end{itemize}
In this paper, our approach to the cographs is different. We only consider $\mathcal{C}$-graphs, that can be coded by finite sequence. The sketch of the paper is as follows: in Section $2$, we determine the inertia and multiplicity of the trivial eigenvalues $-1,~0$ of $A$. In Section $3$ we obtain an eigenvalue-free interval for $\mathcal{C}$-graphs; in fact we show that any eigenvalue other than $0$ and $-1$ is not
contained in the interval $\big{[}\frac{-1-\sqrt{2}}{2}, \frac{-1+\sqrt{2}}{2}\alpha_{min} \big{]}$. Finally in Section $4$, we find the determinant of a special type of tridiagonal matrix and using that we obtain the characteristic polynomial of the adjacency matrix of a $\mathcal{C}$-graph.
\section{Eigenvalues of cographs}
Let $G=C(\alpha_1,\alpha_2,\ldots,\alpha_{2k})$ be a $\mathcal{C}$-graph on $n$ vertices. In this section, first we describe some properties of the quotient matrix corresponding to an equitable partition. Using these properties we establish multiplicity of the eigenvalues $-1$ and $0$. We also find the number of negative, zero, and positive eigenvalues of $A$.
\begin{lemma}
\label{Cgraphs_eig1_lm1}
Let $S=\{{i_1},{i_2},\ldots,{i_m}\},m\geq2$ forms a clique in a graph $G$ such that each vertex of $S$ has the same set of neighbours outside $S$. Then $-1$ is an eigenvalue of $G$ with multiplicity at least $m-1$.
\end{lemma}
\begin{proof}
For $j=1,2,\ldots,m-1$, we construct eigenvectors $X_j$ corresponding to $-1$ defined by
\begin{eqnarray*}
X_j(l)=\begin{cases}
1,&\text{if }l=i_k, k\leq j,\\
-j,&\text{if }l=i_{j+1},\\
0,&\text{otherwise.}
\end{cases}
\end{eqnarray*}
It is easy to verify that $AX_j=-X_j$ and the set $\{X_j|j=1,2,\ldots,m-1\}$ is orthogonal. So the multiplicity of $-1$ is atleast $m-1$.
\end{proof}
We can do the similar treatment for the independent sets in a graph and we can construct orthogonal set of eigenvectors corresponding to the eigenvalue 0.
\begin{lemma}
\label{Cgraphs_eig0_lm1}
Let $S=\{{i_1},{i_2},\ldots,{i_m}\},m\geq2$ be an independent set in a graph $G$ such that each vertex of $S$ has the same set of neighbours outside $S$. Then the nullity of $G$ is least $m-1$.
\end{lemma}
We now consider the vertex partition $\pi=\{C_1, C_2, C_3, \ldots,C_{2k}\}$ of a $\mathcal{C}$-graph $G=C(\alpha_1,\alpha_2,\ldots,\alpha_{2k})$ such that the set $C_i,\ 1\leq i\leq 2k$, contains $\alpha_i$ vertices that added in the $i$-th step while creating $G$. Then $\pi$ is an equitable partition, and the corresponding quotient matrix $Q_\pi(G)$ is given by
$$Q_{\pi}(G)=\begin{bmatrix}
(\alpha_1-1) &\alpha_2 &0 &\alpha_4&0 &\alpha_6& \ldots & \alpha_{2k} \\
\alpha_1&0 &0 &\alpha_4&0&\alpha_6& \ldots &\alpha_{2k}\\
0 &0&(\alpha_3-1)&\alpha_4 &0 &\alpha_6& \ldots &\alpha_{2k}\\
\alpha_1&\alpha_2&\alpha_3&0&0&\alpha_6&\ldots&\alpha_{2k}\\
0&0&0&0&(\alpha_5-1)&\alpha_6&\ldots&\alpha_{2k}\\
\alpha_1&\alpha_2&\alpha_3&\alpha_4&\alpha_5&0&\ldots&\alpha_{2k}\\
& & & & & & \ddots \\
\alpha_1&\alpha_2&\alpha_3&\alpha_4&\alpha_5&\alpha_6& \ldots & 0
\end{bmatrix}.$$
Let $P_\pi$ be the matrix of order $n\times 2k$ whose $(i,j)$-th entry is 1 if $i\in C_j$ and 0 otherwise; the matrix $P_\pi$ is called the characteristic matrix for the equitable partition $\pi$. Then $P_\pi$ and $Q_\pi(G)$ satisfy, $AP_\pi=P_\pi Q_\pi(G)$. Now let $\lambda$ be an eigenvalue of $Q_\pi(G)$ and let $X\in \mathbb{R}^{2k}$ be corresponding eigenvector. Then $A(P_\pi X)=\lambda (P_\pi X)$. Which implies that every eigenvalue of $Q_\pi(G)$ is also an eigenvalue of $A$.
Let $D=diag[\alpha_1, \alpha_2, \alpha_3, \ldots, \alpha_{2k}]$, then $Q_{\pi}(G)$ is similar to the symmetric matrix $D^{\frac{1}{2}}Q_{\pi}(G)D^{-\frac{1}{2}}$, thus $Q_{\pi}(G)$ is diagonalizable. Now we study the spectral properties of $Q_{\pi}(G)$.
\begin{theorem}
Let $G=C(\alpha_1,\alpha_2,\ldots,\alpha_{2k})$ be a $\mathcal{C}$-graph of order $n$. Then $-1$ is an eigenvalue of $Q_{\pi}(G)$ if and only if $\alpha_2=1.$ Further if $\alpha_2=1$, then $-1$ is an simple eigenvalue of $Q_{\pi}(G)$.
\end{theorem}
\begin{proof}
To prove the first part of the theorem, we show that rank of $Q_{\pi}(G)+I$ is less than $n$ if and only if $\alpha_2=1$. We have,
$$Q_{\pi}(G)+I=\begin{bmatrix}
\alpha_1 &\alpha_2 &0 &\alpha_4&0 &\alpha_6& \ldots & \alpha_{2k} \\
\alpha_1&1 &0&\alpha_4&0&\alpha_6& \ldots &\alpha_{2k}\\
0 &0&\alpha_3&\alpha_4 &0 &\alpha_6& \ldots &\alpha_{2k}\\
\alpha_1&\alpha_2&\alpha_3&1&0&\alpha_6&\ldots&\alpha_{2k}\\
0&0&0&0&\alpha_5&\alpha_6&\ldots&\alpha_{2k}\\
\alpha_1&\alpha_2&\alpha_3&\alpha_4&\alpha_5&1&\ldots&\alpha_{2k}\\
& & & & & & \ddots \\
\alpha_1&\alpha_2&\alpha_3&\alpha_4&\alpha_5&\alpha_6& \ldots & 1
\end{bmatrix}.$$
Now we perform the following row operations step by step:\\
(i) $R'_{2k-i}\leftarrow R_{2k-i}-R_{2k-i-2}$ for all $i=0,1,\ldots,2k-3$.\\
(ii) $R'_2\leftarrow R_2-R_1$.\\
(iii) $R'_1\leftarrow R_1+\sum_{i=1}^{k-1} R_{2i+1}$.\\
(iv) $R_{2k}\leftrightarrow R_{2k-1}\leftrightarrow \cdots\leftrightarrow R_2\leftrightarrow R_1\leftrightarrow R_{2k}$.\medskip
After performing above operations we obtain a matrix, say $T$, which is row equivalent to $Q_\pi(G)+I$, where
$$T=\begin{bmatrix}
0&1-\alpha_2 &0 &0&\ldots &0\\
-\alpha_1 &-\alpha_2&\alpha_3-1&0 & \ldots &0\\
0&\alpha_2-1&\alpha_3&1-\alpha_4&\ldots&0\\
0&0&-\alpha_3&-\alpha_4&\ldots&0\\
& & & & \ddots \\
0 &0 &0 &0& \ldots &-\alpha_{2k}
\end{bmatrix}.$$
Here $T$ is a tridiagonal matrix with leading entry 0. So $\det T=0$ if and only if $\alpha_2=1$. Consequently $-1$ is an eigenvalue of if and only if $\alpha_2=1$.\\
Next suppose that $\alpha_2=1$. Then eigenvector corresponding to the eigenvalue $-1$ is $X=\left[\begin{array}{ccccccc}
-1&\alpha_1&0&0&\cdots&0&0
\end{array}\right]^T$. Clearly the geometric multiplicity of the eigenvalue $-1$ is one. Thus, $Q_{\pi}(G)$ being diagonalizable, the algebraic multiplicity of the eigenvalue $-1$ is also one. Hence the result follows.
\end{proof}
\subsection{Inertia of a $\mathcal{C}$-graph}
Let $B$ be a square matrix and let $n_{-}(B),~n_{0}(B)$ and $n_{+}(B)$ denote the number of negative, zero and positive eigenvalues of $B$ respectively. The triplet $(n_{-}(B),~n_{0}(B),~n_{+}(B))$ is called the inertia of $B$. The inertia of the adjacency matrix of a graph is known as the inertia of the graph. We now recall some important well-known results (for more details we refer to the book \cite{Horn} by Horn and Johnson).
\begin{theorem}[\textbf{Weyl's inequality}]
\label{Weyl}
Let $B, ~C$ be Hermitian matrices of order $n$ and let the respective eigenvalues of $B,~C$, and $B+C$ be $\{ \lambda_i(B)\}_{i=1}^{n}$, $\{ \lambda_i(C\}_{i=1}^{n}$, $\{ \lambda_i(B+C)\}_{i=1}^{n}$, each algebraically ordered as below:
\begin{equation}
\lambda_1\leq \lambda_2 \leq \lambda_3 \leq \cdots \leq \lambda_n.
\end{equation}
Then $$\lambda_i(B)+ \lambda_1(C) \leq \lambda_i(B+C) \leq \lambda_i(B)+ \lambda_k(C),$$ where $i=1,2,\ldots,n.$
\end{theorem}
\begin{theorem}[\textbf{Sylvester's law of inertia}]
\label{Sylvester}
Let $B$ and $C$ be Hermitian matrices of equal order. Then $B$ and $C$ have the same inertia if and only if $B=SCS^{*}$ for some non-singular matrix $S$, where $S^{*}$ is the conjugate transpose of $S$.
\end{theorem}
\begin{theorem}[\textbf{Ostrowski Theorem}]
\label{Ostrowski}
Let $B,~S$ be the matrices of order $n$ with $B$ Hermitian and $S$ non-singular. Let the eigenvalues of $B,~SBS^{*}$, and $SS^{*}$ be arranged in non-decreasing order as in $(1)$. Let $\sigma_1\geq \sigma_2 \cdots \geq \sigma_n$ be the singular values of $S$. Then for each $p=1,2, \ldots n$, there is positive real number $\theta_p \in[\sigma_n^2,~\sigma_1^2]$ such that $$\lambda_p(SBS^{*})=\theta_p \lambda_p(B) $$
\end{theorem}
Now we compute the inertia of the $\mathcal{C}$-graph $G=C(\alpha_1,\alpha_2,\ldots,\alpha_{2k})$.
\begin{theorem}
Let $G=C(\alpha_1,\alpha_2,\ldots,\alpha_{2k})$ be a $\mathcal{C}$-graph. Then the inertia of the quotient matrix $Q_{\pi}(G)$ is given by
$$\big{(}n_{-}(Q_{\pi}(G)),~n_{0}(Q_{\pi}(G)),~n_{+}(Q_{\pi}(G))\big{)}=\big{(}k,~0,~k\big{)}.$$
\end{theorem}
\begin{proof}
The quotient matrix of the graph $G$ is
$$Q_{\pi}(G)=\begin{bmatrix}
(\alpha_1-1) &\alpha_2 &0 &\alpha_4&0 &\alpha_6& \ldots & \alpha_{2k} \\
\alpha_1&0 &0 &\alpha_4&0&\alpha_6& \ldots &\alpha_{2k}\\
0 &0&(\alpha_3-1)&\alpha_4 &0 &\alpha_6& \ldots &\alpha_{2k}\\
\alpha_1&\alpha_2&\alpha_3&0&0&\alpha_6&\ldots&\alpha_{2k}\\
0&0&0&0&(\alpha_5-1)&\alpha_6&\ldots&\alpha_{2k}\\
\alpha_1&\alpha_2&\alpha_3&\alpha_4&\alpha_5&0&\ldots&\alpha_{2k}\\
& & & & & & \ddots \\
\alpha_1&\alpha_2&\alpha_3&\alpha_4&\alpha_5&\alpha_6& \ldots & 0
\end{bmatrix}.$$
Consider the diagonal matrices: $$D=diag[\alpha_1, \alpha_2, \alpha_3, \ldots, \alpha_{2k} ],\ D{'}=diag\Big[\frac{\alpha_1-1}{\alpha_1}, 0, \frac{\alpha_3-1}{\alpha_3},0 \ldots, \frac{\alpha_{2k-1}-1}{\alpha_{2k-1}}, 0\Big].$$ Then $Q_{\pi}(G)$ can be expressed as
\begin{equation}
\label{Cgraphs_eq2}
Q_{\pi}(G)=D^{\frac{1}{2}}(A_{2k}+D{'})D^{\frac{1}{2}},
\end{equation}
where $A_{2k}$ denotes the adjacency matrix of the antiregular graph $H_{2k}$.
Let $C=(A_{2k}+D{'})$. By Theorem \ref{Sylvester}, $Q_{\pi}(G)$ and $C$ both have same inertia.
Let the eigenvalues of $Q_{\pi}(G),~A_{2k},~D,~D{'},$ and $C$ be arranged in non-decreasing order as in $(1)$.
It is well-known (see \cite{Bapat}) that the antiregular graph $H_{2k}$ has the inertia $\big{(}k,~0,~k\big{)}$.\\
By Theorem \ref{Weyl}, we have
$$\lambda_{k+1}(C)\geq \lambda_{k+1}(A_{2k})+\lambda_1(D{'})>0.$$
And
$$\lambda_{k}(C)\leq \lambda_{k}(A_{2k})+\lambda_k(D{'})<0,$$
since $\lambda_{k}(A_{2k})\leq-1$ and all the eigenvalues of $D{'}$ lie in the interval $[0, 1)$. \\
This completes the proof.
\end{proof}
As a consequence of this theorem we have the following corollaries.
\begin{corollary}
For the $\mathcal{C}$-graph $G=C(\alpha_1,\alpha_2,\ldots,\alpha_{2k})$, the inertia is given by
$$\big{(}n_{-}(A),~n_{0}(A),~n_{+}(A)\big{)}=\big{(}\sum_{i=1}^{k}\alpha_{2i-1},~\sum_{i=1}^{k}\alpha_{2i}-k,~k\big{)}$$
\end{corollary}
\begin{corollary}
Let $m(G)$ denote the number of distinct eigenvalue of the adjacency matrix $A$ of a $\mathcal{C}$-graph $G=C(\alpha_1,\alpha_2,\ldots,\alpha_{2k})$. Then
$ m(G) \leq 2k+2. $
\end{corollary}
\subsection{Multiplicity of the eigenvalues $0$ and $-1$}
Consider a $\mathcal{C}$-graph $G=C(\alpha_1,\alpha_2,\ldots,\alpha_{2k})$ of order $n$. Let $m_0$ and $m_{-1}$, respectively, denote the multiplicity of the eigenvalue $0$ and $-1$ of the adjacency matrix $A$ . Now we derive formulas for $m_0$ and $m_{-1}$.
\begin{theorem}
\label{Cgraphs_multiplicity_th1}
Let $G=C(\alpha_1,\alpha_2,\ldots,\alpha_{2k})$ be a $\mathcal{C}$-graph of order $n$. Then
$$m_{0}=\sum_{i=1}^{k}\alpha_{2i}-k,$$
and
$$m_{-1}=\begin{cases}
\sum_{i=1}^{k}\alpha_{2i-1}-k,&\text{if }\alpha_2\neq1,\\
\sum_{i=1}^{k}\alpha_{2i-1}-k+1,&\text{if }\alpha_2=1.
\end{cases}
$$
\end{theorem}
\begin{proof}
Firstly note that $0$ is never an eigenvalue of $Q_\pi(G)$. Now for each $i=1,2,\ldots,k$, the set $C_{2i}$ is an independent set in $G$ such that each vertex in $C_{2i}$ has the same set of neighbours. Therefore by Lemma \ref{Cgraphs_eig0_lm1} each $\alpha_{2i}$ will contribute $\alpha_{2i}-1$ in $m_0$. Consequently we obtain $m_0\geq \sum \alpha_{2i}-k.$ \medskip
Next we compute $m_{-1}$. For this, first note that $-1$ is an eigenvalue of $Q_\pi(G)$ if and only if $\alpha_2=1$. Now for each $i=1,2,\ldots,k$, the set $C_{2i-1}$ is a clique in $G$ such that each vertex in $C_{2i-1}$ has the same set of neighbours. Therefore by Lemma \ref{Cgraphs_eig1_lm1} each $\alpha_{2i-1}$ will contribute $\alpha_{2i-1}-1$ in $m_{-1}$. So $m_{-1}\geq \sum\alpha_{2i-1}-k$. Now we consider two cases:\\
\textbf{Case I.} If $\alpha_2\neq 1$. Then neither $0$ nor $-1$ is an eigenvalue of $Q_\pi(G)$. So in that case we obtain: $m_{0}=\sum_{i=1}^{k}\alpha_{2i}-k$ and $m_{-1}=\sum\alpha_{2i-1}-k$.\\
\textbf{Case II.} If $\alpha_2= 1$. The eigenvalue $-1$ is a simple eigenvalue of $Q_\pi(G)$ with corresponding eigenvector $X=\left[\begin{array}{ccccccc}
-1&\alpha_1&0&0&\cdots&0&0
\end{array}\right]^T$. Which implies $P_\pi X$ is an eigenvector corresponding to $-1$ for $A$. Now $P_\pi X$ taxes constant value in each vertex in the same vertex partition of $\pi$. Consequently $P_\pi$ is orthogonal to each eigenvector constructed in Lemma \ref{Cgraphs_eig0_lm1}. So in that case we conclude that $m_{0}=\sum_{i=1}^{k}\alpha_{2i}-k$ and $m_{-1}=\sum\alpha_{2i-1}-k+1$. This completes the proof of the theorem.
\end{proof}
\section{Eigenvalue-free interval for $\mathcal{C}$-graphs}
Here we prove a very important result of the paper. For a cograph $G$, the eigenvalues $0$ and $-1$ are called the trivial eigenvalues and all other eigenvalues are called non-trivial eigenvalues. We are interested to find out an interval which does not contain any non-trivial eigenvalue of a $\mathcal{C}$-graph. Let $\lambda_{+}(G)$ is least positive eigenvalue of $G$ and $\lambda^{-}(G)$ is the largest eigenvalue of $G$ smaller than $-1$. First we recall a result due to Aguilar et al.
\begin{lemma}{\cite{Aguilar2}}
\label{lemma1}
Let $n$ be a positive integer. Then
$$\lambda^{-}(H_n)<\frac{-1-\sqrt{2}}{2},\text{ and } \lambda_{+}(H_n)>\frac{-1+\sqrt{2}}{2}$$
That means the closed interval $\big{[}\frac{-1-\sqrt{2}}{2}, \frac{-1+\sqrt{2}}{2}\big{]}$ does not contain any non-trivial eigenvalue of the antiregular graph $H_n$.
\end{lemma}
We are now in a position to establish the main result of this section.
\begin{theorem}
Consider the $\mathcal{C}$-graph $G=C(\alpha_1,\alpha_2,\ldots,\alpha_{2k})$. Then
$$\lambda^{-}(G)<\frac{-1-\sqrt{2}}{2},\text{ and }\lambda_{+}(G)>\frac{-1+\sqrt{2}}{2}\alpha_{min},$$
where $\alpha_{min}=\displaystyle \min_{1\leq i\leq 2k}\{\alpha_i\}$.
\end{theorem}
\begin{proof}
First observe that any nontrivial eigenvalue of $G$ must be an eigenvalue of the quotient matrix $Q_\pi(G)$.
Recall the representation of $Q_\pi(G)$ described in equation (\ref{Cgraphs_eq2}):
$$Q_{\pi}(G)=D^{\frac{1}{2}}(A_{2k}+D{'})D^{\frac{1}{2}},$$
where $$D=diag[\alpha_1, \alpha_2, \alpha_3, \ldots, \alpha_{2k} ],\text{ and } D{'}=diag\bigg{[}\frac{\alpha_1-1}{\alpha_1}, 0, \frac{\alpha_3-1}{\alpha_3},0 \ldots, \frac{\alpha_{2k-1}-1}{\alpha_{2k-1}}, 0\bigg{]}.$$
Consider the expressions $(1)$ and (2). \\
By Theorem \ref{Ostrowski}, we have
$$\lambda_{k}(Q_{\pi}(G))=\theta_{k} \lambda_k (A_{2k}+D{'}),$$
where $\theta_k \in [\alpha_{min},~\alpha_{max}]$, and $\alpha_{max}=\displaystyle \max_{1\leq i\leq 2k} \{\alpha_{i}\}$. \\
Therefore by Lemma $\ref{lemma1}$ and using Theorem \ref{Sylvester},
$$\lambda_{k+1}(Q_{\pi}(G))=\theta_{k+1} \lambda_{k+1} (A_{2k}+D{'})>\frac{-1+\sqrt{2}}{2}\theta_{k+1}.$$
Thus
$$ \lambda_{k+1}(Q_{\pi}(G))>\frac{-1+\sqrt{2}}{2}\alpha_{min}. $$
That is
$$\lambda_{+}(G)>\frac{-1+\sqrt{2}}{2}\alpha_{min}.$$
\medskip
We now prove $\lambda^{-}(G)<\frac{-1-\sqrt{2}}{2}$.\\
First note that $\mathcal{C}$-graph $C(1,1,\ldots,1)$, and $C(1,2,1,\ldots,1)$ are antiregular graphs on $2k$ and $2k+1$ vertices respectively. So clearly $H_{2k}$ is a subgraph of $G$ and further if $\alpha_2\neq1$, then $H_{2k+1}$ is also a subgraphs of $G$. Let the eigenvalues of $A_{2k}$ and $A_{2k+1}$ be arranged in non-decreasing order as given in $(1)$. Then we have
$$\lambda_{k}(A_{2k})=-1,\text{ and }\lambda_{k-1}(A_{2k})<\frac{-1-\sqrt{2}}{2}.$$
Also we have,
$$\lambda_{k+1}(A_{2k+1})=0,\text{ and } \lambda_{k}(A_{2k+1})<\frac{-1-\sqrt{2}}{2}. $$
We consider two cases:\\
\textbf{Case I:} $\alpha_2=1$. Then $-1$ is the largest negative eigenvalue $G$, and so $\lambda^-(G)=\lambda_{k-1}(A)$. Since $H_{2k}$ is a subgraph of $G$, so by interlacing theorem
$$\lambda_{k-1}(A) \leq \lambda_{k-1}(A_{2k})<\frac{-1-\sqrt{2}}{2}.$$
\textbf{Case II:} $\alpha_2 \neq1$. So $\lambda_{k}(A)<-1$. Since $H_{2k+1}$ is a subgraph of $G$, by interlacing theorem
$$\lambda_{k}(A) \leq \lambda_{k}(A_{2k+1})<\frac{-1-\sqrt{2}}{2}.$$
Combining above two cases we can conclude that $\lambda^{-}(G)<\frac{-1-\sqrt{2}}{2}$.\\
Hence the theorem is proved.
\end{proof}
\begin{corollary}
For the $\mathcal{C}$-graph $G=C(\alpha_1,\alpha_2,\ldots,\alpha_{2k})$, the interval $\big{[}\frac{-1-\sqrt{2}}{2}, \frac{-1+\sqrt{2}}{2}\alpha_{min} \big{]}$ does not contain any non-trivial eigenvalue of $G$.
\end{corollary}
\section{Characteristic polynomial}
This section is about the characteristic polynomial of a $\mathcal{C}$-graph $G=C(\alpha_1,\alpha_2,\ldots,\alpha_{2k})$ of order $n$. We obtain an expression for characteristic polynomial of the quotient matrix $Q_{\pi}(G)$, and by using it we calculate the same for $A$. Suppose $\Psi(x)$ and $\Psi_\pi(x)$ respectively denote the characteristic polynomial of $A$ and $Q_\pi(G)$. Then,
$$\Psi(x)=x^{\sum_{i=1}^{k}(\alpha_{2i}-1)}(x+1)^{\sum_{i=1}^{k}(\alpha_{2i-1}-1)}\Psi_{\pi}(x).$$
Let $\beta_1,\beta_2,\ldots,\beta_n$ be real numbers. Now consider the following tridiagonal determinant:
$$T(\beta_1,\beta_2,\ldots,\beta_n)=\begin{vmatrix}
\beta_1&-1\\
1&\beta_2&-1\\
&1&\beta_3&-1\\
& &\ddots & \ddots& \ddots \\
&&&1&\beta_{n-1}&-1\\
&&&&1&\beta_n
\end{vmatrix}.$$
We can find the above determinant by three term recurrence relation for the determinant of tridiagonal matrix. In this paper we give an alternative way to find the determinant.
Let $I_n=\{1,2,\ldots,n\}$ and let $S_n$ denote the set of all permutations on $I_n$. We call a permutation $e$-transposition if it is a two cycle of the form $(l,\ l+1)$, for some $1\leq l<n$. Let $E_n$ denote the set of all permutations which are a product of disjoint $e$-transpositions. Also we consider the identity permutation to be an element of $E_n.$
First we prove the following lemma.
\begin{lemma}
\label{Cgraphs_cp_lm1}
Let $\beta_1,\beta_2,\ldots,\beta_n$ be real numbers, then
$$T(\beta_1,\beta_2,\ldots,\beta_n)=\sum_{\sigma\in E_{n}}\prod_{\substack{i\in I_n,\\ \sigma(i)=i}}\beta_i.$$
\end{lemma}
\begin{proof}
Let $T=[t_{ij}]$ be the corresponding matrix of the determinant $T(\beta_1,\beta_2,\ldots,\beta_n)$.
By the Leibniz formula of determinants:
$$T(\beta_1,\beta_2,\ldots,\beta_n)=\det T=\sum_{\sigma\in S_n}sgn(\sigma)\prod_{i\in I_n}t_{i,\sigma(i)}.$$
Now $t_{i,\sigma(i)}$ is (possibly) nonzero if $\sigma(i)=i-1$ or $i$ or $i+1$. Thus for a nonzero $t_{1,\sigma(1)}t_{2,\sigma(2)}\ldots t_{n,\sigma(n)}$, the permutation $\sigma$ must be a product of distinct $e$-transpositions. Now suppose $\sigma=(i_1,\ i_1+1)(i_2,\ i_2+1)\ldots (i_l,\ i_l+1)$ be a product of distinct $e$-transpositions. Then $sgn(\sigma)=(-1)^l$ and
\begin{equation*}\begin{split}
t_{1,\sigma(1)}t_{2,\sigma(2)}\ldots t_{n,\sigma(n)}&=(-1)^l\prod_{j\notin\{i_1, i_1+1,i_2,i_2+1,\ldots, i_l, i_l+1\}}\beta_j\\
&=(-1)^l\prod_{\sigma(j)=j}\beta_j.
\end{split}
\end{equation*}
Now taking summations over all permutations of $I_n$, we get our result.
\end{proof}
\begin{theorem}
Let $G=C(\alpha_1,\alpha_2,\ldots,\alpha_k)$ be a $\mathcal{C}$-graph of order $n$. Then the characteristic polynomial of the quotient matrix $Q_\pi(G)$ is given by $$\psi_\pi(x)=\prod_{i=1}^k(\alpha_{2i-1}-1-x)(\alpha_{2i}+x)T(\beta_1,\beta_2,\ldots,\beta_{2k}),$$
where $\beta_1=\dfrac{1+x}{\alpha_1-1-x}$, and $\beta_i=\begin{cases}\frac{\alpha_i}{\alpha_{i}-1-x}&\text{ if }i \text{ odd and }i\geq3\\
\frac{\alpha_i}{\alpha_{i}+x}&\text{ if }i \text{ even. }
\end{cases}$
\end{theorem}
\begin{proof}
The characteristic polynomial of $Q_{\pi}(G)$ is
$$\psi_\pi(x)=\det (Q_\pi(G)-xI)=\begin{vmatrix}
(\alpha_1-1-x) &\alpha_2 &0 &\alpha_4& \ldots & 0&\alpha_{2k} \\
\alpha_1&-x &0 &\alpha_4&\ldots &0&\alpha_{2k}\\
0 &0&(\alpha_3-1-x)&\alpha_4 & \ldots &0&\alpha_{2k}\\
\alpha_1&\alpha_2&\alpha_3&-x&\ldots&0&\alpha_{2k}\\
& & & & \ddots \\
0&0&0&0&\ldots&(\alpha_{2k-1}-1-x)&\alpha_{2k} \\
\alpha_1&\alpha_2&\alpha_3&\alpha_4&\ldots &\alpha_{2k-1}&-x
\end{vmatrix}.$$
We perform the some operations. We perform $R'_{2k-i}\leftarrow R_{2k-i}-R_{2k-i-2}$ for all $i=0,1,\ldots,2k-3$; and $ R_2^{'}\gets R_2-R_1$. So we obtain
$$\psi_\pi(x)\equiv \begin{vmatrix}
(\alpha_1-1-x) &\alpha_2 &0 &\alpha_4& \ldots & 0&\alpha_{2k} \\
(1+x)&-(\alpha_2+x) &0 &0&\ldots &0&0\\
-(\alpha_1-1-x) &-\alpha_2&(\alpha_3-1-x)&0 & \ldots &0&0\\
0&(\alpha_2+x)&\alpha_3&-(\alpha_4+x)&\ldots&0&0\\
& & & & \ddots \\
0&0&0&0&\ldots&(\alpha_{2k-1}-1-x)&0 \\
0&0&0&0&\ldots &\alpha_{2k-1}&-(\alpha_{2k}+x)
\end{vmatrix}$$
Again performing $R_1^{'}\gets R_1+\sum_{i=1}^{k-1}R_{2i+1}$, we obtain
$$\psi_\pi(x) \equiv \begin{vmatrix}
0 &0 &0 &0& \ldots &(\alpha_{2k-1}-1-x)&\alpha_{2k} \\
(1+x)&-(\alpha_2+x) &0 &0&\ldots &0&0\\
-(\alpha_1-1-x) &-\alpha_2&(\alpha_3-1-x)&0 & \ldots &0&0\\
0&(\alpha_2+x)&\alpha_3&-(\alpha_4+x)&\ldots&0&0\\
& & & & \ddots \\
0&0&0&0&\ldots&(\alpha_{2k-1}-1-x)&0 \\
0&0&0&0&\ldots &\alpha_{2k-1}&-(\alpha_{2k}+x)
\end{vmatrix}$$
Taking permutation on rows we obtain a tridiagonal determinant.
$$\psi_\pi(x) \equiv \begin{vmatrix}
(1+x)&-(\alpha_2+x) &0 &0&\ldots &0&0\\
-(\alpha_1-1-x) &-\alpha_2&(\alpha_3-1-x)&0 & \ldots &0&0\\
0&(\alpha_2+x)&\alpha_3&-(\alpha_4+x)&\ldots&0&0\\
0&0&-(\alpha_3-1-x)&-\alpha_4&\ldots&0&0\\
& & & & \ddots \\
0&0&0&0&\ldots &\alpha_{2k-1}&-(\alpha_{2k}+x)\\
0 &0 &0 &0& \ldots &-(\alpha_{2k-1}-1-x)&-\alpha_{2k}
\end{vmatrix}$$
Now, for $1\leq i\leq k$, we multiply $2i$-th row by $-1$. We also use the fact that sign interchange between $ij$-th and $ji$-th of a tridiagonal matrix do not change the determinant. So we obtain
\begin{equation*}
\begin{split}
\psi_\pi(x)&=(-1)^k\begin{vmatrix}
(1+x)&-(\alpha_2+x) &0 &\ldots &0&0\\
(\alpha_1-1-x) &\alpha_2&-(\alpha_3-1-x) & \ldots &0&0\\
0&(\alpha_2+x)&\alpha_3&\ldots&0&0\\
0&0&(\alpha_3-1-x)&\ldots&0&0\\
& & & \ddots \\
0&0&0&\ldots &\alpha_{2k-1}&-(\alpha_{2k}+x)\\
0 &0 &0& \ldots &(\alpha_{2k-1}-1-x)&\alpha_{2k}
\end{vmatrix}\\
&=(-1)^k(\alpha_1-1-x)(\alpha_2+x)(\alpha_3-1-x)(\alpha_4+x)\cdots(\alpha_{2k}+x)T(\beta_1,\beta_2,\ldots,\beta_{2k}).
\end{split}
\end{equation*}
The last form in the previous equation is obtained by taking $(\alpha_1-1-x)$ common from first column, $(\alpha_2+x)$ from the second column, $(\alpha_3-1-x)$ from the third column and finally $(\alpha_{2k}+x)$ from the last column.
This completes the proof of the theorem.
\end{proof}
We conclude this paper by considering an example to get the flavour of the above theorem.
\begin{example}
Consider the $\mathcal{C}$-graph $G=C(4,3,2,2)$ with $11$ vertices. Then
$$E_n=\{ id, (1, 2), (2, 3), (3, 4), (1, 2)(3, 4) \}.$$
Therefore, \\
\begin{align*}
T(\beta_1,\beta_2,\beta_3,\beta_4) =&\beta_1 \beta_2 \beta_3 \beta_4+\beta_1 \beta_2 +\beta_1 \beta_4 +\beta_3 \beta_4 +1 \\
=& \frac{12(1+x)}{(3-x)(3+x)(1-x)(2+x)} + \frac{3(1+x)}{(3-x)(3+x)} + \frac{2(1+x)}{(3-x)(2+x)}\\&+\frac{4}{(1-x)(2+x)}+1.
\end{align*}
Therefore
\begin{align*}
\Psi_{\pi}(x)=&(-1)^2(3-x)(3+x)(1-x)(2+x)T(\beta_1,\beta_2,\beta_3,\beta_4)\\=& 12(1+x)+3(1+x)(1-x)(2+x)+2(1+x)(3+x)(1-x)+4(3+x)(3-x)\\&+(3-x)(3+x)(1-x)(2+x).
\end{align*}
On simplification,
$$\Psi_{\pi}(x)= x^4-4x^3-27x^2+8x+78.$$
Therefore,
$$\Psi(x)=x^3(x+1)^4 \big{[} x^4-4x^3-27x^2+8x+78 \big{]}.$$
\end{example}
\section{Acknowledgement}
The author Santanu Mandal admits the financial support of University Grants Commission, India under the beneficiary code BININ01569755.
|
\section{Introduction}
Nowadays, counterfeiting and piracy are among the main challenges for modern economy. Existing methods of anti-counterfeiting are very diverse, ranging from watermarking techniques, special inking, holograms, electronic IDs, etc. The drawbacks of these technologies are that they can be expensive, often proprietary, and usually, authentication is performed in a non-digital way.
\begin{figure}[htbp]
\centerline{\includegraphics[scale=.87, trim={0cm, 0cm, 0cm, -0.13cm}]{images/authentication_diagram.pdf}}
\caption{Illustration of the PI channel seen as a 3-player game. The Defender (in green) generates and prints template ${\bf t}^n$ on an object package $o^n$ and sends it to the public domain. The Attacker (in red) can use $o^n$ to create a counterfeited version $c^n$ of it. Finally, the Verifier (in blue) scans the object package and authentifies the probe ${\bf y}^n$ to decide whether it is an original package or a counterfeit. The novelty of our model is adding a predictor channel (in yellow) based on a codebook $\mathcal{C}$ which is trained by the Defender and used by the Verifier to enhance the classification results at the authentication stage.}
\label{fig:defender_attacker_scheme}
\end{figure}
A newly promising emerged field in digital anti-counterfeiting technologies is the usage of Printing Unclonable Features (PUF) which are based on intrinsic forensic uncloneable features of physical objects, such as randomness of ink blots or paper micro-structures \cite{zhu2003print, adams2011study, voloshynovskiy2012towards}. Another technology is the Copy Detection Patterns (CDP) \cite{picard2004digital} which are random binary patterns of high entropy that are difficult to clone, such as very small sized QR codes. The advantages of CDP, in comparison to other technologies, are that they are cheap, easily integrable with a product into a structure of QR-code and digitally readable \cite{picard2021counterfeit}. They are also easy to integrate in a track-and-trace distribution framework. The main challenge of this technology today is that, although being mainly robust to common copy attacks when simple decision rules are used based on the similarity to the reference template blueprint, it faces significant difficulties with the advanced machine-learning (ML) copy attacks. The possibility to use powerful deep classifiers in two-class classification allows one to reliably distinguish original CDPs from fakes, if the fakes used at testing time match the statistics of those used during training. However, in the case of mismatches, the method fails to distinguish original and fakes~\cite{taran2022mobile}. In practice, the situation is further complicated by several factors:
\begin{itemize}
\item the high deviations in printing and imaging leading to large intra-class variabilities;
\item ML attacks that are able to produce blueprint estimations with an accuracy score as high as $94\%$ \cite{chaban2021machine};
\item the natural lack of exact prior knowledge for the authenticator about the fakes in field. Fakes can be produced in multiple ways and it is unknown which fake will be used at the attacking time;
\item the absence of a reliable model of printing-imaging channel that complicates the design of optimal authentication rules.
\end{itemize}
Therefore, there is a critical need in a one-class (OC) authentication scheme able to operate in the generalized setup of the above printing-imaging channel without prior knowledge of the fakes. In this paper we adress these problems by:
\begin{itemize}
\item providing a new stochastic model describing the defender Printing-Imaging (PI) channel;
\item proposing a new method of authentication based on the PI model able to perform authentication in the OC-classifier setup, i.e., under complete ignorance about the actual fakes;
\item validating the proposed approach on a real dataset of CDPs of originals and ML-based fakes based on codes designed with $1 \times 1$ symbol and produced on two industrial printers;
\item comparing the proposed method with traditional authentication techniques.
\end{itemize}
The paper is organized as follows. Section \ref{section:problem_formulation} introduces the problem formulation and presents a stochastic model of PI channel for the defender that forms the basis of the OC-classification framework. Section \ref{section:oc_classification_algo} presents the algorithm of OC-classification for CDPs in two variations. Section \ref{section:results} presents the results of performance and comparison with standard metrics on the same dataset. Finally, Section \ref{section:conclusion} concludes the paper and discusses possible extensions and perspectives. All mathematical notations used in the paper can be found in Table \ref{table:notation_table}.
\section{Problem formulation}
\label{section:problem_formulation}
\subsection{The printing-authentication scheme}
The production of an anti-counterfeit technology using CDP is best described as a 3-player game with a Defender, an Attacker and a Verifier as shown in Fig. \ref{fig:defender_attacker_scheme}.
The Defender protects his brand by using a family of digital CDP blueprints $\{ {\bf t}^n \}_{n=1}^N$ stored in the form of a binary matrix ${\bf t}^n$, which is then printed on the object package $o^n$ and sent to the public domain.
The Attacker has access to the printed version of the CDP and may use it to create a counterfeit $c^n$, through the process of scanning, post-processing and reprinting (see \cite{chaban2021machine, khermaza2021can, yadav2019estimation, taran2019clonability} for investigations of attacking techniques).
At the authentication stage, the Verifier receives an unindentified package (either $o^n$ or $c^n$) from which a digital image ${\bf y}^n$ is acquired, using any device such as a scanner or a mobile phone. We denote ${\bf x}^n$ the code acquired from $o^n$ and ${\bf f}^n$ the code acquired from $c^n$.
An authentication is then performed based on the probe ${\bf y}^n$, which might be either ${\bf x}^n$ or ${\bf f}^n$, and on the reference template ${\bf t}^n$.
\subsection{Authentication techniques}
The algorithms used for authentication evolved a lot in the last few years. At first, CDP were designed with an idea to be resistant to simple scanning \& reprinting attacks~\cite{picard2004digital}. Due to the dot gain effect of printers, a portion of the information stored in the template blueprint ${\bf t}$ is lost in the probe ${\bf y}$ through the process of printing and scanning. Various ways to measure the information loss have been proposed which can be formalized with different types of metrics:
\input{tables/notation_table}
\begin{enumerate}
\item $\ell_1$- or $\ell_2$-distance between the probe ${\bf y}$ and the template ${\bf t}$;
\item Pearson correlation between ${\bf t}$ and ${\bf y}$;
\item Hamming distance between the template ${\bf t}$ and an estimation $\tilde{{\bf t}}$ of the template, based on the probe ${\bf y}$. A very common way to perform the estimation $\tilde{{\bf t}}$ is to use Otsu's binarization algorithm and then a majority voting for each symbol. Fig. \ref{fig:otsu_majority_voting} on the next page illustrates this technique.
\end{enumerate}
Nowadays, new techniques emerge with the use of machine learning, allowing one to train deep classifiers \cite{taran2022mobile, cui2020new} and deep binarization techniques \cite{chaban2021machine, khermaza2021can, yadav2019estimation, taran2019clonability}. Although showing very promising results, these new algorithms act as black boxes and thus lack interpretability, which is paramount when working on reliability questions and security-critical applications such as the protection of pharmaceutical products.
\subsection{Stochastic model of Printing-Imaging channel}
The PI channel can be described mathematically as a Markov Chain ${\bf T} \rightarrow {\bf X} \rightarrow \tilde{{\bf T}}$, where:
\begin{itemize}
\item ${\bf T}$ is a random binary matrix of size $L \times L$ sampled from i.i.d. Bernoulli distribution: $T_{ij} \sim Bern(p)$, $p \in [0,1]$ is the probability of black symbol;
\item ${\bf X}$ is a random matrix of size $kL \times kL$, $X_{ij} \in [0,1]$ for some magnification factor\footnote{The magnification factor is related to the resolution of enrollment equipment. Nowadays, with modern scanners and mobile phones, $k \geq 1$.}
$k = 1, 2, 3, ...$;
\item $\tilde{{\bf T}}$ is a random binary matrix of size $L \times L$.
\end{itemize}
In reality, when we pass a template ${\bf t}$ through the PI channel, some distorsions occur in ${\bf x}$ due to the dot-gain effect and printing-related natural randomness. Thus, when we try to estimate $\tilde{{\bf t}}$ from ${\bf x}$, we end up with some errors, dependant on the printer, type of paper, acquisition device and chosen estimator. In this paper, we are mostly interested in understanding the probability distribution $\mathbb{P}(\tilde{{\bf T}} | {\bf T})$, which we believe to be highly correlated with the particular choices of print-acquire-estimate system and is central when trying to estimate information loss.
In \cite{voloshynovskiy2016physical}, the authors model this probability distribution as a Binary Symmetric Channel (BSC). This model assumes that each symbol $T_{ij}$ in ${\bf T}$ has a certain probability $P_b$ of bit-flipping, independently of its location $(i,j)$. We conjecture that the BSC model is too simple to capture the random behaviour of printing, as it does not take into account the local dependency of neighbouring sites and rather learns an average probability of bit-error across the whole template. Another related model with multilevel symbols has been studied in \cite{villan2006multilevel}. Inspired by the BSC model, we introduce a new stochastic model with three key assumptions:
\begin{enumerate}
\item \emph{Markovianity:} the posterior probability at a particular symbol location $(i,j)$ only depends on the local neighbourhood $\omega_{ij}$ surrounding it:
\begin{equation}
\mathbb{P}(\tilde{T}_{ij} | {\bf T}) = \mathbb{P}(\tilde{T}_{ij} | \omega_{ij}),
\label{eq:locality_axiom}
\end{equation}
where $\omega_{ij}$ is a small neighbourhood surrounding symbol $T_{ij}$, typically a square matrix centered around $(i,j)$:
\[
\omega_{ij} = \{ T_{i \pm a, j \pm b} | 0 \leq a,b < h/2 \},
\]
where $h = 1, 3, 5, ...$ is fixed by the model and defines the size of the neighbourhood.
\item \emph{Stationarity:} the posterior probability does not depend on the location inside the image. Similar patterns in ${\bf T}$ lead to similar probability values\footnote{
The printing and scanning process introduces a lot of variability. The goal of the model is not to learn the fingerprint of a particular realization but rather measure the average variability for each neighbourhood and to take advantage of this knowledge. \eqref{eq:stationarity_axiom} should be read as an equality in distribution, allowing every realisation of $\tilde{T}_{i,j}$ to be different while still following a common law, independent of the location $(i,j)$.}:
\begin{equation}
\mathbb{P}(\tilde{T}_{ij} | \omega_{ij}) = \mathbb{P}(\tilde{T}_{rs} | \omega_{rs}), \mbox{ if } \omega_{ij} = \omega_{rs}.
\label{eq:stationarity_axiom}
\end{equation}
\item \emph{Posterior independance:} the joint posterior probability factorizes as:
\begin{equation}
\mathbb{P}(\tilde{{\bf T}} | {\bf T}) = \prod_{i,j} \mathbb{P}(\tilde{T}_{ij} | {\bf T} ).
\label{eq:posterior_independance_axiom}
\end{equation}
\end{enumerate}
\begin{figure}[htbp]
\centerline{
\includegraphics[scale=.45, trim={3.cm, 2.cm, 2.5cm, 1.2cm}, clip]{images/otsu_binarization.pdf}
}
\caption{Otsu's binarization technique followed by majority voting. The first column shows different neighbourhoods $\omega$, the second column the printed originals ${\bf x}$, the third column ${\bf x}$ after Otsu's binarization and the fourth column, the estimated template $\mathbf{\tilde{t}}$ after majority voting. Red lines highlight the $3 \times 3$ patches in ${\bf x}$ corresponding to one symbol in ${\bf t}$. Different types of distorsions are illustrated leading to estimation errors in $\tilde{{\bf t}}$.}
\label{fig:otsu_majority_voting}
\end{figure}
With assumptions \eqref{eq:locality_axiom} and \eqref{eq:stationarity_axiom}, one can easily prove the expectation formula for the posterior distribution:
\begin{equation}
\mathbb{P}(\tilde{T}_{ij} | \omega_{ij}) = \mathbb{E}_{r,s: \omega_{rs} = \omega_{ij}} [ \mathbb{P}(\tilde{T}_{rs} | \omega_{rs})].
\end{equation}
This formula is a key to the proposed authentication scheme as it can be estimated directly using Monte-Carlo method from a training dataset. For each type of neighbourhood $\omega \in \Omega$ (there can be at most $2^{h^2}$), we learn the probability distribution which is highly correlated with the PI channel on which it was trained. Two measures associated with this distribution are the posterior probability of bit-flipping $P_b(\omega)$ and the positive posterior probability $P(\omega)$, which we define as:
\begin{IEEEeqnarray}{rCl}
P_b(\omega_{ij}) &:=& \mathbb{P}(\tilde{T}_{ij} \neq T_{ij} | \omega_{ij}), \\
P(\omega_{ij}) &:=& \mathbb{P}(\tilde{T}_{ij} = 1 | \omega_{ij}).
\end{IEEEeqnarray}
We can thus create a codebook in which we store all these different probability values for each type of neighbourhood and use them as references in the authentication scheme.
\subsection{Metric in the PI channel}
The introduced PI channel gives us a theoretical tool to better understanding the process of printing and acquisition of CDP. In this subsection, we show that this model comes with a very natural metric that can be easily implemented and used for authentication.
\begin{lem}
In the PI channel model, the posterior log-likelihood can be computed as:
\begin{equation}
\log \mathbb{P}(\tilde{{\bf T}} = \tilde{{\bf t}} | {\bf T}) = \sum_{i,j} \log (1 - | \tilde{t}_{ij} - P(\omega_{ij}) |).
\label{eq:lls_formula}
\end{equation}
\end{lem}
\begin{proof}
The proof relies on two steps. The first one is to use conditional independence of the symbols in $\tilde{{\bf T}}$ given ${\bf T}$ and Markovianity:
\begin{IEEEeqnarray*}{rCl}
\log \mathbb{P}(\tilde{{\bf T}} = \tilde{{\bf t}} | {\bf T}) &=& \sum_{i,j} \log \mathbb{P}(\tilde{T}_{ij} = \tilde{t}_{ij} | {\bf T}) \\
&=& \sum_{i,j} \log \mathbb{P}(\tilde{T}_{ij} = \tilde{t}_{ij} | \omega_{ij}).
\end{IEEEeqnarray*}
The second step is then a simple case study for $\tilde{t}_{ij} \in \{0,1\}$:
\begin{equation*}
\mathbb{P}(\tilde{T}_{ij} = \tilde{t}_{ij} | \omega_{ij}) = 1 - | \tilde{t}_{ij} - P(\omega_{ij}) |.
\end{equation*}
\end{proof}
\section{one-class CDP classification algorithms}
\label{section:oc_classification_algo}
The core idea of building an authentication system based on the PI channel model is to introduce \emph{the predictor channel}, which is trained using both digital templates ${\bf t}$ and acquired originals ${\bf x}$ and to learn a codebook $\mathcal{C}$ of probabilities for each neighbourhood $\omega \in \Omega$.
To train the predictor, we create two dictionaries $\mathbb{D}$ and $\mathbb{D}_b$ whose keys are the different types of neighbourhoods. For each $\omega_{ij} \in \Omega$, $\mathbb{D}[\omega_{ij}]$ lists the corresponding values of symbol $\tilde{t}_{ij}$ and $\mathbb{D}_b[\omega_{ij}]$ lists the boolean values $(\tilde{t}_{ij} \neq t_{ij})$. Finally, we compute the codebook $\mathcal{C}$, which is a database storing the statistics $P(\omega)$ and $P_b(\omega)$ for each type of neighbourhood $\omega$. A pseudo-code is given in Algorithm \ref{predictor_algo}.
\subsection{The likelihood score model}
\label{ssec:lls_model}
The first authentication scheme is a direct implementation of~\eqref{eq:lls_formula}. It starts by learning the codebook $\mathcal{C}$, running Algorithm~\ref{predictor_algo} on the training set. For the authentication of a probe ${\bf y}$, we perform the following steps:
\begin{enumerate}
\item estimate $\tilde{{\bf t}}$ from the probe ${\bf y}$;
\item with the reference template ${\bf t}$, search the probability $P(\omega_{ij})$ in $\mathcal{C}$, for each neighbourhood $\omega_{ij}$ in ${\bf t}$;
\item compute the likelihood score of $\tilde{{\bf t}}$ applying \eqref{eq:lls_formula};
\item compare the score with a chosen threshold fixed on the validation set to decide whether ${\bf y}$ is original or fake.
\end{enumerate}
It should be pointed out here that symbols $t_{ij}$ located too close to the border of the template do not have a well-defined neighbourhood $\omega_{ij}$. We propose two solutions to address this problem:
\begin{itemize}
\item the first solution is simply to ignore these symbols and run the model only on the symbols located in the inside of ${\bf t}$;
\item another solution is to consider a white padding surrounding template ${\bf t}$ as this is the natural padding for ${\bf x}$ when printing CDP on white paper.
\end{itemize}
\subsection{The attention model}
\label{ssec:attention_model}
The attention model is similar in essence to the preceding model but differs in several ways. The idea here is to use the probability bit-error map $P_b(\omega_{ij})$ as a mask, only keeping symbols that have a low probability of bit-error on the training set. In this way, we remove all regions in ${\bf y}$ that are known to produce high error for original samples ${\bf x}$. Training is done similarly to the likelihood score model above. For the authentication, we do:
\input{algo_predictor}
\begin{enumerate}
\item for each neighbourhood $\omega_{ij}$ in ${\bf t}$, search the probability of bit-flipping $P_b(\omega_{ij})$ in the codebook;
\item define an attention mask $m_{ij} := (P_b(\omega_{ij}) < \mu)$ for some fixed threshold $\mu \in [0,1]$;
\item choose any standard metric that is computed pixel-wise such as mean squared error, Hamming distance or Pearson correlation. Note that some upsampling of ${\bf t}$ might be necessary for computation;
\item weight the chosen metric $d({\bf t},{\bf y})$ by using the binary mask, upsampling it if needed:
\[
d^{m}({\bf t},{\bf y}) = \sum_{i,j} m_{ij} \cdot d(t_{ij}, y_{ij}).
\]
\end{enumerate}
\section{Results}
\label{section:results}
\subsection{Dataset choice}
For our experiments, we use the Indigo $1 \times 1$ base dataset, presented in \cite{chaban2021machine}.
It is constituted of 720 different templates~${\bf t}$ printed with two different printers: HP~Indigo 5500 DS (HPI55) and HP Indigo 7600 DS (HPI76) at 812.8 dpi, which we refer to as ${\bf x}^{55}$ and ${\bf x}^{76}$.
It also includes ML-based fakes of four different types: ${\bf f}^{55/55}$, ${\bf f}^{76/55}$, ${\bf f}^{55/76}$ and ${\bf f}^{76/76}$ where fake ${\bf f}^{mm/nn}$ is obtained from ${\bf x}^{nn}$ by the process of deepnet-based binarization, printed using HPImm and rescanned.
In this work, we only concentrate on the templates with $50 \%$ density of black symbols.
The templates ${\bf t}$ have a size of $228 \times 228$ symbols while ${\bf x}$ and ${\bf f}$ have a size of $684 \times 684$, that is a magnification by a factor $k=3$. We fix the training set size to $50$ samples, validation set to $100$ samples and test set to $500$ samples.
\input{tables/results_table_no_std}
\begin{figure*}[t]
\centerline{\includegraphics[scale=.55, trim={1.5cm, 1.5cm, 1.5cm, 1.5cm}, clip]{images/roc_results.pdf}}
\caption{Visualisation of ROC curves for original $x^{76}$ and all four types of fakes. First column is the LLS metric, second column the masked LLS, third column the Hamming distance and fourth column the masked Hamming distance. The second row shows the same plots with logarithmic scale on both axes.}
\label{fig:roc_results}
\end{figure*}
\vspace{-.8mm}
\subsection{Predictor algorithm parameters}
\label{ssec:algo_param_choice}
In order to train the predictor, we fix a certain number of parameters.
The first one is the estimator ${\bf x} \rightarrow \tilde{{\bf t}}$. As we saw in Section \ref{section:problem_formulation}, there are many different approaches to it. We decide to use Otsu's algorithm for binarization followed by majority voting on each $3 \times 3$ patch corresponding to one symbol in ${\bf t}$.
We fix the size of neighbourhoods $\omega$ in ${\bf t}$ to be of size $3 \times 3$ for the following reasons:
\begin{itemize}
\item This brings the total number of possible neighbourhoods down to $|\Omega| = 2^9 = 512$ which is small enough in comparison to the total number of neighbourhoods in a single template: $226^2 = 51'076$. We can thus expect to see every neighbourhood appear roughly $100$ times in each template.
\item The printing process can produce some random deviations as we discussed in Section \ref{section:problem_formulation}, but these deviations are local in the sense that they only affect neighbouring symbols in most cases. Thus, $3 \times 3$ neighbourhoods are sufficient to capture them. See Fig. \ref{fig:otsu_majority_voting} for an illustration.
\end{itemize}
\subsection{Discussion}
To compare all different approaches in a unified way, we test both originals ${\bf x}^{55}$ and ${\bf x}^{76}$ separately against all four kind of ML fakes ${\bf f}^{55/55}$, ${\bf f}^{55/76}$, ${\bf f}^{76/55}$ and ${\bf f}^{76/76}$. The metrics that we use are:
\begin{itemize}
\item the log-likelihood score (LLS) described in Section \ref{ssec:lls_model};
\item the mean-squared-error (MSE) between ${\bf y}$ and ${\bf t}$;
\item the Pearson correlation (PCOR) between ${\bf y}$ and ${\bf t}$;
\item the Hamming distance (HAMM) between ${\bf t}$ and $\tilde{{\bf t}}$;
\item the same four metrics mentioned above with a mask as described in Section \ref{ssec:attention_model}.
\end{itemize}
For each metric, we compute the associated ROC curves and report the AUC score. The AUC score is averaged over ten runs with randomization of training/testing set.
All results are summarized in Table \ref{table:AUC_all_metrics}.
A first observation at the results in Table \ref{table:AUC_all_metrics} shows that discriminating between originals and fakes is more accurate for ${\bf x}^{76}$ than for ${\bf x}^{55}$. In general, the results show that the metric LLS outperforms the other metrics. On average, M-LLS, its masked version, appears as the best metric with a very reliable AUC score on all types of fakes.
The masked metrics show a great improvement in AUC score over all their non-masked counterparts. This is further illustrated in Fig. \ref{fig:roc_results}, where we compare side-by-side masked and non-masked metrics for LLS and Hamming metrics.
Surprisingly, MSE proves to be the best metric for discriminating ${\bf x}^{76}$ and ${\bf f}^{55/76}$. This result should however be mitigated by the following observations:
\begin{itemize}
\item metric M-LLS performs very close to MSE and even outperformed it on certain runs;
\item the high variability in performance of MSE on different types of fakes makes it highly unreliable for authentication, as shown by its average score.
\end{itemize}
\subsection{Model stability}
Another question that we investigated is the stability of Algorithm \ref{predictor_algo} with respect to the size of the training set. We already discussed, in Section \ref{ssec:algo_param_choice}, the fact that every neighbourhood appears $100$ times on average in each template. Thus, it makes sense to run the algorithm on very small training sets. In order to measure the performance of a codebook $\mathcal{C}$ learned on a training set $\{ ({\bf t}^n,{\bf x}^n) \}$, we compare it with a reference codebook $\mathcal{C}^{ref}$ learned on the whole dataset of $720$ pairs $\{ ({\bf t}^n,{\bf x}^n) \}$. We then simply compute an average $\ell_1$-distance between the predictions:
\begin{equation}
d_1(\mathcal{C}, \mathcal{C}^{ref}) = \frac{1}{|\Omega|} \sum_{\omega \in \Omega} |P(\omega) - P^{ref}(\omega)|.
\end{equation}
Fig. \ref{fig:train_set_dependency} shows the results of this study for different training sets size with a number of samples going from $1$ to $100$. What we can see is that when using $50$ samples, the probabilities in the codebook $\mathcal{C}$ differ with the reference by less than $1 \%$ on average and the variability is very small. This explains why we decided to use $50$ training samples in our experiments.
\section{Conclusion}
\label{section:conclusion}
In this paper, we introduced a new mathematical model for the description of the Printing-Imaging channel based on local statistics.
We proposed two novel OC-authentication schemes based on this model which outperform the standard metrics used nowadays, while still maintaining full interpretability of the results. We showed that even ML-based attacks cannot fool our new authentication system. In constrast with modern deep learning approaches, our model requires very few training data and does not require much time to be run in practice, while still offering great performances against powerful ML attacks.
For future work, we aim at continuing to explore this model as the information-theoretic aspects can be deeper investigated. We also plan to replace the simple estimator with more sophisticated techniques based on neural networks and perform the comparison of the proposed approach with deep classifiers. Finally, we plan to extend the results on a new dataset acquired by several types of mobile phones which will bring more variability and new challenges for the PI channel model.
\begin{figure}[htbp]
\centerline{\includegraphics[scale=.6, trim={0cm, 0cm, 1.cm, 1.cm}]{images/train_set_dependance.pdf}}
\caption{Variability of codebook $\mathcal{C}$ with respect to the size of the training set.}
\label{fig:train_set_dependency}
\end{figure}
\bibliographystyle{IEEEtran}
|
\section{Introduction}
It is well-known that the collection of fixed points of normal
unital completely positive (UCP) maps has a von Neumann algebra
structure on introducing a new product called the Choi-Effros
product. This von Neumann algebra is known as the Poisson boundary
of the UCP map. In \cite{BhatTalwarKar} it is noted that this
product can be extended to peripheral eigenvectors, that is to
eigenvectors with eigenvalue on the unit circle. As a consequence,
the norm closed linear span of peripheral eigenvectors becomes a
$C^*$-algebra. This was termed as peripheral Poisson boundary of the
UCP map. It is a natural question as to when does the new product of
the peripheral Poisson boundary agrees with the original product we
had in the beginning. In this article we answer this question for
UCP maps on matrix algebras.
There is extensive literature on the peripheral spectrum of positive
maps \cite{FallahRansford, GrohJoT, GrohLAA, GrohMZ, EvansKrohn}. A
lot of it is motivated by the classical Perron-Frobenius theory.
However we limit ourselves to results directly relevant to the
problem posed above and try to present these results in a unified
way. In the process, we also found some statements in the literature
which are inaccurate -see \Cref{error}.
We concentrate on UCP maps and their spectrum. Instead of
considering general finite dimensional $C^*$-algebera we are
focusing on UCP maps on the full matrix algebra $\mathbb{M}_d(\mathbb{C}) :
= \mathbb{M}_d$. There is not much loss of generality in doing so, in view
of \Cref{ExtendToFinitedimensioanlC*-algebra}.
In rest of this section, we set up some basic notation and recall
some results from \cite{BhatTalwarKar}. Let $\mathbb{M}_d$ be the $C^*$-algebra \ of
all $d \times d$ complex matrices and let $\tau : \mathbb{M}_d \to \mathbb{M}_d$ be
a UCP map. A non-zero vector $X$ in $\mathbb{M} _d$ satisfying $\tau
(X)=\lambda X$ for some $\lambda \in \mathbb{T}:=\{z\in
\mathbb{C}:|z|=1\}$ is called a peripheral eigenvector of $\tau .$
We have, $$E_{\lambda }(\tau ):=\{ X\in \mathbb{M} _d: \tau(X)=\lambda
X\}.$$ The peripheral space of $\tau $ is defined as
$$\mathcal{P}(\tau) := \overline{\text{span}}\{ X \in \mathbb{M}_d : \tau(X) = \lambda X \text{ for some } \lambda \text{ with }
|\lambda| = 1 \}.$$ Observe that $\mathcal{P}(\tau)$ is a vector subspace of
$\mathbb{M} _d$, containing the identity and it is closed under taking
adjoints. In other words, it is an operator system. In general
$\mathcal{P}(\tau)$ may not be closed under matrix multiplication. However
one may define a new product by setting (See \cite{BhatTalwarKar})
$$X\circ Y = s-\lim_{n\to \infty}(\lambda \mu )^{-n}\tau ^n(XY)$$ for peripheral
eigenvectors $X,Y$ with $\lambda$ and $\mu$ as their respective
eigenvalues. Without changing the definitions of norm and adjoint,
$(\mathcal{P}(\tau), \circ)$ defines a $C^*$-algebra. Such a theory is possible for
normal UCP maps on general von-Neumann algebras. Since our focus in
this paper is on the finite dimensional $C^*$-algebra \ $\mathbb{M}_d$, we may replace the norm closure `$\overline{\text{span}}$' by `$\text{span}$' and the strong
operator limit above can be replaced with the norm limit. The
peripheral Poisson boundary of $\tau $ is defined as the
$C^*$-algebra $(\mathcal{P}(\tau) , \circ ).$ Given a UCP map one would like
to identify its peripheral Poisson boundary. The best case scenario
would be when the new product coincides with the original product of
$\mathcal{P}(\tau).$ Keeping this in mind we make the following definition.
\begin{defin}\label{PeripherallyAutomorphicDefinition}
Let $\tau :\mathbb{M} _d \to \mathbb{M} _d$ be a UCP map. Then $\tau $ is said to be
{\em peripherally automorphic\/} if $X\circ Y =XY$ for every $X, Y$ in $\mathcal{P}(\tau) .$
\end{defin}
If $X\circ Y=XY$ in $\mathcal{P}(\tau)$, we see that it is a $C^*$-algebra
under the original product and as per \cite[Theorem
2.12]{BhatTalwarKar}, restriction of $\tau$ to $\mathcal{P}(\tau)$ is a
$C\sp\ast$-automorphism. This justifies this definition. One of the
main goals of this article is to identify the collection of
peripherally automorphic UCP maps. A natural candidate is the
following well-studied class of maps.
\begin{defin}
Let $\tau : \mathbb{M} _d\to \mathbb{M} _d$ be a UCP map. Then $\tau $ is said to be
a {\em stationary} UCP map if it admits a faithful invariant state,
that is, there exists a faithful state $\phi $ on $\mathbb{M} _d$ such that
$$\phi (\tau (X))=\phi (X), ~~\forall X\in \mathbb{M} _d.$$
\end{defin}
We go through some of the existing literature on this class of UCP
maps. All these maps are peripherally automorphic. Then we look at
the notion of multiplicative domains for UCP maps and finally
characterize the collection of peripherally automorphic UCP maps. In
the last section, we study the discrete dynamics of a UCP map on $\mathbb{M}
_d$ and its decomposition into persistent and transient parts. We
are able to demonstrate this decomposition for arbitrary UCP maps on
$\mathbb{M} _d.$ We use the familiar tool of Jordan decomposition. The main
disadvantage of considering the Jordan decomposition of UCP maps is
that it does not respect the crucial positivity/algebraic structures
of $\mathbb{M} _d.$ Here we are able to overcome this obstacle to some
extent by appealing to results from the peripheral Poisson boundary
theory. For instance it shown that UCP maps on finite dimensional
$C^*$-algebras with spectrum contained in the unit circle are
$*$-automorphisms. This partially settles a conjecture of F.
Fidaleo, F. Ottomano\ and\ S. Rossi \cite{Francesco2} in the
affirmative.
\section{Stationary maps and peripherally automorphic maps}
It is well-known that every completely positive map $\tau: \mathbb{M}_d \to
\mathbb{M}_d$ is of the form $\tau(X) = \sum_{i=1}^r L_i^\ast X L_i$ for
some $d \times d$ matrices $L_i$ which are known as Kraus (or
Choi-Kraus) coefficients of $\tau .$ As trace of a positive matrix
is zero if and only if the matrix is zero, for the normalized trace
$\text{tr}$ on $\mathbb{M}_d$, the formula $\langle X,Y \rangle := \text{tr}(X^\ast Y)$
defines an inner product for every $X,Y \in \mathbb{M}_d$ (Our inner
products are anti-linear in the first variable). Then the adjoint
$\tau^\ast$ of $\tau$ is the unique linear operator satisfying
$\langle \tau(X),Y \rangle = \langle X,\tau^\ast(Y) \rangle$ for
every $X,Y \in \mathbb{M}_d$.
For a completely positive map $\tau(X) = \sum_{i=1}^kL_i^\ast X L_i$
it can be verified that $\tau^\ast(X) = \sum_{i=1}^r L_i X
L_i^\ast$.
\begin{remark}\label{adjointmap}
A UCP map $\tau :\mathbb{M} _d\to \mathbb{M} _d$ is stationary if and only if the adjoint $\tau^\ast$ has a strictly positive fixed point. A UCP
map $\tau: \mathbb{M} _d\to \mathbb{M} _d$ is trace preserving if and only if $\tau^\ast$ is unital.
\end{remark}
The importance of mixed unitary channels is well understood.
They are all trace preserving. Going further there is the notion
called trace reducing maps in literature -see \cite{NAJ}. A linear
map $\tau: \mathbb{M}_d \to \mathbb{M}_d$ is called {\em trace reducing} if
$\text{tr}(\tau(X)) \leq \text{tr}(X)$ for every positive $X$ in $ \mathbb{M}_d$. In a
similar fashion one can call $\tau$ a {\em state reducing map} if
there exists a state $\psi$ on $\mathbb{M}_d$ such that $\psi(\tau(X)) \leq
\psi(X)$ for every $0 \leq X \in \mathbb{M}_d$. The following observation
tells us that for unital linear maps these notions do not enlarge
the classes of maps under consideration. For this particular result
the state $\psi $ need not be faithful.
\begin{prop}\label{ReducingPreservingSame}
Let $\tau : \mathbb{M}_d \to \mathbb{M}_d$ be a unital linear map. Then $\tau$ is state reducing with respect to a state $\psi $ if and only if $\tau$ is $\psi $ preserving.
\end{prop}
\begin{proof}
Suppose $\tau $ is state reducing for $\psi .$ Consider any $X$ in
$\mathbb{M}_d$ with $0\leq X\leq I$. Note that we also have $0\leq I-X$. So we
get $\psi (\tau (X))\leq \psi (X)$ and $\psi (\tau (I-X))\leq \psi
(I-X)$. Since $\tau , \psi $ are unital, the second inequality means $-\psi (\tau (X))\leq -\psi (X).$ Hence $\psi (\tau(X))= \psi
(X).$ The converse statement is a triviality.
\end{proof}
\begin{remark}\label{error}
If a UCP map preserves a faithful state then it has a range of
implications.
In \cite[Theorem 5]{Shemesh} several conditions were
claimed to be equivalent to this property. In our notation, this
theorem essentially reads as follows: Let $\tau(X)=\sum_{i=1}^r
L_i^*XL_i$ be a UCP map on $\mathbb{M}_d$. Then the following are
equivalent.
\begin{enumerate}
\item $\tau$ is stationary.
\item The spaces $\mathcal{P}(\tau)$ and $E_1(\tau)$ corresponding to $\tau$ are $*$-algebras.
\item For $X\in \mathbb{M}_d$ and $\lambda \in \mathbb{T}$, $\tau (X)= \lambda X$ if
and only if $XL_i=\lambda L_iX$ for every $i.$
\item The algebra generated by the Kraus coefficients $\{L_i: 1 \leq i \leq r\}$ is a $\ast$-algebra.
\end{enumerate}
Below in \Cref{ShemeshWasWrong} we provide a UCP map $\tau$ such
that the spaces $\mathcal{P}(\tau)$ and $E_1(\tau)$ are $*$-algebras but
neither is $\tau$ a stationary map nor does the algebra generated by
Kraus coefficients form a $*$-algebra. In other words, (2) does not
imply (1) or (4).
\Cref{EquivalentConditionsForExistenceOfSpecialFaithfulState} shows
(1) is equivalent to (4) and with the help of \Cref{Ourclass} and
\Cref{MainTheorem} we will see that (2) is equivalent to (3).
\end{remark}
\begin{example}\label{ShemeshWasWrong}
Consider the UCP map $\tau :\mathbb{M} _2\to \mathbb{M} _2$ defined by
$$\tau \left(\begin{bmatrix}
x_{11} & x_{12}\\
x_{21} & x_{22}
\end{bmatrix}\right)=\begin{bmatrix}
\frac{x_{11}+ x_{22}}{2} &0\\
0 &x_{22}
\end{bmatrix},~~\forall \begin{bmatrix}
x_{11} &x_{12}\\
x_{21} &x_{22}
\end{bmatrix} \in \mathbb{M} _2.$$ Here we may take the Kraus coefficients $$L_1:=\begin{bmatrix}
\frac{1}{\sqrt 2} &0\\
0 &0
\end{bmatrix}, ~~L_2:=\begin{bmatrix}
0 &0\\
\frac{1}{\sqrt 2} &0
\end{bmatrix}, ~~L_3:=\begin{bmatrix}
0 &0\\
0 &1
\end{bmatrix}.$$
In this case $\mathcal{P}(\tau)$ and $E_1(\tau)$ are the $\ast$-algebra
$\{\lambda I:\lambda\in \mathbb{C}\}$. The adjoint map $\tau ^*$ is
given by
$$\tau^*\left(\begin{bmatrix}
x_{11} & x_{12}\\
x_{21} & x_{22}
\end{bmatrix}\right)=\begin{bmatrix}
\frac{x_{11}}{2} &0\\
0 &x_{22}+\frac{x_{11}}{2}
\end{bmatrix}.$$
Clearly $\tau^*$ has no strictly positive fixed point. Hence from
\Cref{adjointmap}, $\tau $ is not stationary.
Also note that the algebra generated by the Kraus coefficients is the
algebra of all $2\times 2$ lower triangular matrices which is not a
$*$-algebra.
\end{example}
Now we have a theorem characterizing UCP maps preserving a faithful state, i.e, the stationary maps.
Some of these results are
well-known and are here just for completeness. Recall from
\cite{EvansKrohn} that a UCP map $\tau$ on a finite dimensional $C^*$-algebra
\ $A$ is said to be irreducible if there does not exist any
projection $P \in A$, different from $0,I$ such that $\tau (P)\leq P.$
Observe that (3) implies (1) of the following theorem shows that every irreducible UCP map has a faithful invariant state.
\begin{theorem}\label{EquivalentConditionsForExistenceOfSpecialFaithfulState}
Let $\tau$ be a UCP map on $\mathbb{M}_d$ given by $\tau(X)=\sum_{i=1}^{r}
L_i^*XL_i$. Then the following are equivalent.
\begin{enumerate}
\item $\tau$ is stationary, that is, $\tau $ admits a faithful
invariant state.
\item $\tau(P) \leq P$ for a projection $P$ implies that $\tau(P) =
P$.
\item There exists a set of mutually orthogonal projections $\{ P_1, P_2, \ldots , P_l \}$ (for some $l$) in $\mathbb{M}_d$,
with $P_1+P_2+\cdots +P_l=I$,
such that $\tau_{j} := \tau_{|_{P_j \mathbb{M}_d P_j}}$ is an irreducible UCP map for every $1 \leq j \leq l$.
\item The algebra generated by Kraus coefficients $\{L_i\}_{1 \leq i \leq r}$ is a $\ast$-algebra.
\end{enumerate}
\end{theorem}
\begin{proof}
(1) implies (2): For a stationary map $\tau$, we fix $\phi$ as a
faithful state on $\mathbb{M}_d$ satisfying $\phi(\tau(X)) = \phi(X)$ for
every $X \geq 0$.
Let $P$ be a projection such that $\tau (P)\leq
P$. On applying $\phi $, we get $\phi (P-\tau (P))=\phi (P)-\phi
(P)=0$. As $\phi $ is faithful, $P=\tau(P)$ follows.
(2) implies (3): By our hypothesis,
\begin{eqnarray*}
\mathcal{P} &:= &\{ P \in \mathbb{M}_d : P \text{ is a projection and } \tau(P)
= P\}\\& =& \{ P \in \mathbb{M}_d : P \text{ is a projection and } \tau(P)
\leq P\}.\end{eqnarray*} Now if $\mathcal{P}$ is trivial, $\tau $ is
irreducible and we are done. If not, using finite dimensionality of
$\mathbb{M} _d$, we can get a non-trivial minimal projection $P_1$ such that
$\tau (P_1)=P_1.$ Clearly, $\tau$ maps ${P_1 \mathbb{M}_d P_1}$ to ${P_1
\mathbb{M}_d P_1}$ and $\tau_{|_{P_1 \mathbb{M}_d P_1}}$ is an irreducible UCP map.
The unitality of $\tau $, yields $\tau (P_1^\perp) = P_1^\perp:= I -
P_1 $ and we can repeat the process on ${P_1^\perp \mathbb{M}_d P_1^\perp
}$. Continuing this we can exhaust whole of $\mathbb{M} _d.$
(3) implies (4): Since $\tau_j$ is an irreducible UCP map on the $C^*$-algebra \
$P_j \mathbb{M}_d P_j$, there exists state $\phi_j$ on $P_j \mathbb{M}_d P_j$ such
that $\tau_j$ preserves $\phi_j$. Define a faithful linear
functional on $\mathbb{M}_d$ by $$\phi(X) := \sum_{j=1}^{l} \phi_j(P_j X
P_j).$$ Then $\phi$ is a state because $\phi(I) = \sum_{j=1}^{l}
\phi_j(P_j I P_j) = \sum_{j=1}^{l} \phi_j(P_j) = \sum_{j=1}^{l} P_j
= I$. It is well-known, and can be easily verified using the two definitions of multiplicative domain given later in this paper, that $\tau$ has block decomposition of the form
$\tau(P_j X P_j) = \tau(P_j) \tau(X) \tau(P_j) = P_j \tau(X) P_j$ for every $1 \leq j \leq l$ and $X \in \mathbb{M}_d$. Hence \[ \phi(\tau(X)) = \sum_{j=1}^{l} \phi_j(P_j \tau(X) P_j) = \sum_{j=1}^{l} \phi_j(\tau(P_j X P_j)) = \sum_{j=1}^{l} \phi_j(P_j X P_j) = \phi(X).\]
Thus, $\tau $ is a stationary UCP map and hence we can
apply the trick used later in proving (4) implies (5) of \Cref{Ourclass}
along with the fact that $\tau(P_j) = P_j$ to obtain $P_jL_i = L_i
P_j$. If $\mathbb{A}$ denotes the algebra generated by $\{ L_i \}_{1
\leq i \leq r}$, then $\mathbb{A}P_j=P_j\mathbb{A}=P_j\mathbb{A}P_j$
and hence one may write $\mathbb{A} = \sum_{j=1}^l P_j \mathbb{A} P_j$. It now suffices to prove that $\mathbb{A}P_j$ is a $\ast$-algebra.
For every fixed $1 \leq j \leq l$, we claim that $\mathbb{A}P_j =
B(\text{range}(P_j))$, the set of all linear operators on
$\text{range}(P_j)$. To prove the claim we use the Burnside's
classical theorem \cite{Burnside}. In order to apply the theorem, we
need to show that the subalgebra $\mathbb{A}P_j$ is irreducible i.e.
there is no proper non-trivial subspace of $\text{range}(P_j)$ which
is left invariant by every element of $\mathbb{A}P_j$. Let $S
\subseteq \text{range}(P_j)$ be a subspace which is invariant under
$L_iP_j$ for every $1 \leq i \leq r$. Then $L_i S = L_i P_j S \subseteq
S$ and hence $S$ is invariant under each $L_i$. If we prove that
$S$ is reducing for each $L_i$, then $P_S L_i = L_i P_S$ for every
$i$, $P_S$ being the projection onto $S$. This will imply that
$$\tau(P_S) = \sum_{i=1}^{r} L_i^*P_SL_i = \sum_{i=1}^{r}
L_i^*L_iP_S = P_S.$$ Since each $\tau_j$ is irreducible, we will
obtain that either $S=\{0\}$ or $S= \text{range}(P_j)$ proving the
claim.
Finally, we show that $S$ is reducing for each $L_i$.
It is known from \cite[Theorem 1.1]{Kribs} that there exists a projecton $P$ such that $P\leq \tau(P)$ (respectively, $P\geq \tau(P)$) if and only if range$(P)$ is invariant under $L_i$ (respectively, $L_i^\ast$) for every $1 \leq i \leq r$.
As $S$ is invariant under each $L_i$, we have $P_S \leq \tau(P_S)$.
Since $\tau$ is stationary one may use $(2)$ to obtain that $P_S = \tau(P_S)$.
This implies that $S$ is invariant under $L_i^\ast$ for every $i$, proving that $S$ is reducing for each $L_i$.
(4) implies (1): See $(4) \implies (1)$ of \cite[Theorem 5]{Shemesh}.
\end{proof}
We will study peripheral eigenvectors for UCP maps with faithful
invariant states. It is convenient to consider a larger class described
below.
Recall from (\cite{Choi}, \cite[Remark 2.7]{Rahaman}) that the
multiplicative domain $\mathcal{M}_\tau$ of a UCP map $\tau$ is the $C\sp \ast$-subalgebra \[\{X\in \mathbb{M}_d:\tau(XY)=\tau(X)\tau(Y) \text{ and
}\tau(YX)=\tau(Y)\tau(X) \text{ for every } Y \in \mathbb{M}_d\}\] which is
same as \[\{X\in \mathbb{M}_d:\tau(XX^*)=\tau(X)\tau(X^*) \text{ and
} \tau(X^*X)=\tau(X^*)\tau(X)\}.\]
Some special cases of the following result are well-known (See \cite[Theorem 2.1]{Kribs} and \cite[Theorem
5]{Burgarth}) and the proof of (4) implies (5) uses a standard trick (See for instance \cite[Theorem 4.25]{Watrous}).
\begin{theorem}\label{Ourclass}
Let $\tau :\mathbb{M} _d\to \mathbb{M} _d$ be a UCP map with a Choi-Kraus
decomposition $\tau (X)= \sum _{i=1}^rL_i^*XL_i, ~~\forall X\in \mathbb{M}
_d.$
Then the following are equivalent.
\begin{enumerate}
\item $\mathcal{P}(\tau)\subseteq \mathcal{M}_\tau.$
\item $\mathcal{P}(\tau) \subseteq \mathcal{M}_{\tau ^\infty}:=\bigcap
_{k=1}^{\infty} \mathcal{M}_{\tau ^k}$.
\item $\tau$ is peripherally automorphic.
\item For $\lambda \in \mathbb{T}$, $X \in E_\lambda(\tau) ~~\mbox{implies}~~ \tau(X^\ast X) = X^\ast X.$
\item For $\lambda \in \mathbb{T}$, $Y\in E_\lambda(\tau)$ if and only if $YL_i=\lambda L_iY$ for every $1 \leq i \leq r$.
\end{enumerate}
\end{theorem}
\begin{proof}
(1) implies (2): Consider $X\in \mathcal{P}(\tau)$ such that $\tau(X)=\lambda X$ with $\lambda\in \mathbb{T}$.
As $\mathcal{P}(\tau)\subseteq \mathcal{M}_\tau$, for every $k\in \mathbb{N}$ we have $\tau^k(XX^*)=\lambda^kX\bar{\lambda}^kX^*=\tau^k(X)\tau^k(X^*)$ and $\tau^k(X^*X)=\bar{\lambda}^kX^*\lambda^kX=\tau^k(X^*)\tau^k(X)$.
This shows that $\mathcal{P}(\tau) \subseteq \mathcal{M}_{\tau^k}$ for every $k \in \mathbb{N}$ and hence $\mathcal{P}(\tau) \subseteq \mathcal{M}_{\tau^\infty}$.
(2) implies (3) :
As $\mathcal{P}(\tau)$ is spanned by peripheral eigenvectors of $\tau$, it suffices to
prove that for $\tau(X) = \lambda X$ and $\tau(Y) = \mu Y$ with
$|\lambda| = |\mu|= 1$ we have $X \circ Y = XY$.
Using the hypothesis that $\mathcal{P}(\tau) \subseteq \mathcal{M}_{\tau^\infty}$, we obtain $$X \circ Y = s-\lim (\lambda \mu)^{-n} \tau^n(X Y) = s-\lim (\lambda \mu)^{-n} \tau^n(X) \tau^n(Y) = XY.$$
(3) implies (4) : Let $X \in E_\lambda(\tau)$ and $\lambda \in \mathbb{T}$.
From \cite[Theorem 2.12]{BhatTalwarKar} we have $\tau(X^\ast \circ X) = \tau(X^\ast) \circ \tau(X) = X^\ast \circ X$.
Using the hypothesis that $\tau$ is peripherally automorphic, one obtains $\tau(X^\ast X) = X^\ast X$.
(4) implies (5) : Suppose $\lambda \in \mathbb{T}$, $Y\in \mathbb{M} _d$ and
$\tau (Y)=\lambda Y$.
Then
\begin{eqnarray*}
\sum_{i=1}^r \left( Y L_i - \lambda L_i Y\right)^\ast \left( Y L_i - \lambda L_i Y \right)
&=& \sum_{i=1}^r \left( L_i^\ast Y^\ast Y L_i - \lambda L_i^\ast Y^\ast L_i Y - \overline{\lambda} Y^\ast L_i^\ast Y L_i + Y^\ast L_i^\ast L_i Y \right)\\
&=& \tau (Y^\ast Y) - \lambda \tau(Y^\ast) Y - \overline{\lambda} Y^\ast \tau(Y) + Y^\ast Y \\
&=& \tau(Y^\ast Y) - Y^\ast Y - Y^\ast Y + Y^\ast Y \\
&=& \tau(Y^\ast Y) - Y^\ast Y\\
&=& 0.
\end{eqnarray*}
Consequently, $YL_i=\lambda L_iY$ for all $i$.
Conversely, let $|\lambda| = 1$ and $YL_i=\lambda L_iY$ for every $1 \leq i \leq r$.
It follows from the Kraus decomposition of $\tau$ that $Y\in E_\lambda(\tau)$.
(5) implies (1) : To prove that $\mathcal{P}(\tau) \subseteq M_\tau$, it suffices to prove that for $X \in E_\lambda(\tau)$ with $|\lambda| = 1$ we have $\tau(X^\ast X) = \tau(X^\ast) \tau(X)$ and $\tau(X X^\ast) = \tau(X) \tau(X^\ast)$.
As $X \in E_\lambda(\tau)$, this is same as proving that $\tau(X^\ast X) = X^\ast X$ and $\tau(X X^\ast) = X X^\ast$.
We know from the hypothesis that $X L_i = \lambda L_iX$ and $X^\ast L_i = \overline{\lambda} L_i X^\ast$ for every $1 \leq i \leq r$.
Hence, $$X^\ast X L_i = X^\ast \lambda L_i X = \lambda \overline{\lambda} L_i X^\ast X = L_i X^\ast X$$ and $$ X X^\ast L_i = X \overline{\lambda} L_i X^\ast = \lambda \overline{\lambda} L_i X X^\ast = L_i XX^\ast $$ establishing that $\tau(X^\ast X) = X^\ast X$ and $\tau(X X^\ast) = X X^\ast$.
\end{proof}
\begin{cor}
Let $\tau$ be peripherally automorphic, $X \in E_\lambda(\tau)$
with $\lambda \in \mathbb{T}$ and let $Y \in \mathbb{M}_d$ be such that $\tau(Y)=\mu
Y$. Then $XY\in E_{\lambda \mu}(\tau)$.
\end{cor}
\begin{proof}
Since $X \in \mathcal{P}(\tau) \subseteq \mathcal{M}_\tau$, we have $\tau(XY)=\tau(X)\tau(Y)=\lambda \mu XY$.
\end{proof}
\begin{cor}\label{SIsInR}
Suppose $\tau :\mathbb{M} _d \to \mathbb{M} _d$ is a stationary UCP map with faithful invariant state $\phi$.
Then $\tau$ is peripherally automorphic.
\end{cor}
\begin{proof}
Let $\tau$ be stationary and $X \in E_\lambda(\tau) \text{ with } \lambda \in \mathbb{T}$.
Then $\tau(X) = \lambda X$, $\tau(X^\ast) = \overline{\lambda} X$ and using Kadison-Schwarz inequality we obtain $\tau(X^\ast X) \geq \tau(X^\ast) \tau(X) = X^\ast X$.
An application of the faithful state $\phi$ now shows that $\tau(X^\ast X) = X^\ast X$.
\end{proof}
The following result is particularly useful when we consider convex
sets of peripherally automorphic maps, such as the class of trace
preserving maps (which includes mixed unitary channels) or more
generally stationary UCP maps with respect to a fixed faithful
state.
\begin{cor}\label{convexity}
Suppose $\tau = \sum _{j=1}^kp_j\tau _j$ for some UCP maps $\tau_j$, $p_j>0$ for $1 \leq j \leq k$ and $\sum_{j=1}^k p_j = 1$.
If $\tau $ is peripherally automorphic, then the peripheral spectrum of $\tau $ is contained in the intersection of the peripheral spectrums of $\tau _j$'s and $\mathcal{P}(\tau) \subseteq \mathcal{P}(\tau_j)$ for every $1 \leq j \leq k$.
\end{cor}
\begin{proof}
We make use of part (5) of Theorem \ref{Ourclass}.
If $\lambda $ is in the peripheral spectrum of $\tau $, with eigenvector $X$, then $XL=\lambda LX$ for every Kraus coefficient $L$ of $\tau$.
The Kraus coefficients of $\tau _j$'s can also be considered as scalar multiples of Kraus coefficients of $\tau$.
Thus, $XL=\lambda LX$ for every Kraus coefficient $L$ of $\tau_j$.
Using this and the expression for $\tau_j$'s in terms of their Kraus coefficients, one obtains that $\lambda $ is in the peripheral spectrum of every $\tau_j$ with $X$ as corresponding eigenvector.
\end{proof}
The preceding corollary and its proof tells us that $\tau$ exhibits an extreme behavior on $\mathcal{P}(\tau)$ in the sense that for such $\tau$ and $\tau_j$'s we have $\tau(X) = \tau_j(X)$ for every $X \in \mathcal{P}(\tau)$ and $1 \leq j \leq k$.
We point it out here that in general there exists stationary maps which may be written as non-trivial convex combination of other stationary maps.
Now we have the main theorem of this Section.
\begin{theorem}\label{MainTheorem}
Let $\tau : \mathbb{M} _d\to \mathbb{M} _d$ be a UCP map.
Then $\tau$ is peripherally automorphic if and only if $\mathcal{P}(\tau)$ is closed under matrix multiplication. In this case $X\mapsto \tau (X)$ is a $C^*$-algebra automorphism on $\mathcal{P}(\tau)$.
\end{theorem}
\begin{proof}
If $\tau$ is peripherally automorphic, then \cite[Theorem 2.3]{BhatTalwarKar} implies that $\mathcal{P}(\tau)$ is closed under multiplication.
Conversely, let $\mathcal{P}(\tau)$ be closed under matrix multiplication.
Using \cite[Theorem 2.12]{BhatTalwarKar} we obtain that
$\tau(X^\ast \circ X) = \tau(X^\ast) \circ \tau(X) = X^\ast \circ X$ for $\lambda \in \mathbb{T}$ and $X \in E_\lambda(\tau)$.
In view of \Cref{Ourclass}, it suffices to prove that $X^\ast \circ X = X^\ast X$.
As $X^\ast X \in \mathcal{P}(\tau)$, one may write $X^\ast X = \sum X_i$ where $X_i \in E_{\lambda_i}(\tau)$ with each $|\lambda_i| = 1$. From \cite[Theorem 2.5 \& Lemma 2.2]{BhatTalwarKar} we obtain a sequence $k_n$ of natural numbers such that $$X^\ast \circ X = s-\lim \tau^{k_n}(X^\ast X) = s-\lim \sum \lambda_i^{k_n} X_i = \sum X_i = X^\ast X.$$
That $\tau_{|_{\mathcal{P}(\tau)}}$ is a $C^*$-automorphism is now a consequence of \cite[Theorem 2.12]{BhatTalwarKar}.
\end{proof}
It is not difficult to see that direct sum of two peripherally automorphic maps is peripherally automorphic.
Corollary \ref{convexity} suggests that perhaps one should study
convex combinations of peripherally automorphic maps. However, below
we show that convex combinations and compositions of two
peripherally automorphic maps may not be peripherally automorphic,
unlike stationary maps with respect to a fixed state.
\begin{example}\label{AverageIsNotPeripherallyAutomorphic}\label{CompositionofPAIsNotPA}
\begin{enumerate}
\item For two peripherally automorphic maps $\tau_1$ and $\tau_2$ on $\mathbb{M}_3$ given by $$\tau_1\left( \begin{bmatrix}
x_{11} &x_{12} &x_{13}\\
x_{21} &x_{22} &x_{23}\\
x_{31} &x_{32} &x_{33}
\end{bmatrix} \right) = \begin{bmatrix}
x_{11} &0 &0\\
0 &x_{22} &0\\
0 &0 &x_{11}
\end{bmatrix} \text{ and } \tau_2\left(\begin{bmatrix}
x_{11} &x_{12} &x_{13}\\
x_{21} &x_{22} &x_{23}\\
x_{31} &x_{32} &x_{33}
\end{bmatrix} \right)= \begin{bmatrix}
x_{11} &0 &0\\
0 &x_{22} &0\\
0 &0 &x_{22}
\end{bmatrix}.$$
Then for $\tau= \frac{1}{2}(\tau_1+\tau_2)$ we have
$$\tau\left(\begin{bmatrix}
1 & 0 & 0\\
0& 3 & 0\\
0 & 0 & 2
\end{bmatrix}\right)= \begin{bmatrix}
1 & 0 & 0\\
0& 3 & 0\\
0 & 0 & 2
\end{bmatrix}$$ but $$\tau\left(\begin{bmatrix}
1 & 0 & 0\\
0& 3 & 0\\
0 & 0 & 2
\end{bmatrix}^2\right) = \tau\left(\begin{bmatrix}
1 & 0 & 0\\
0& 9 & 0\\
0 & 0 & 4
\end{bmatrix}\right) = \begin{bmatrix}
1 & 0 & 0\\
0& 9 & 0\\
0 & 0 & 5
\end{bmatrix} \neq \begin{bmatrix}
0 & 0 & 0\\
0& 9 & 0\\
0 & 0 & 4
\end{bmatrix}.$$
Thus, $\tau$ is not peripherally automorphic.
\item Let $$\tau_1\left(\begin{bmatrix}
x_{11} &x_{12} &x_{13}\\
x_{21} &x_{22} &x_{23}\\
x_{31} &x_{32} &x_{33}
\end{bmatrix}\right) = \begin{bmatrix}
x_{11} &0 &0\\
0 &x_{22} &0\\
0 &0 &\frac{x_{11}+x_{33}}{2}
\end{bmatrix}$$ and $$\tau_2\left(\begin{bmatrix}
x_{11} &x_{12} &x_{13}\\
x_{21} &x_{22} &x_{23}\\
x_{31} &x_{32} &x_{33}
\end{bmatrix}\right) = \begin{bmatrix}
x_{11} &0 &0\\
0 &x_{22} &0\\
0 &0 &\frac{x_{22}+x_{33}}{2}
\end{bmatrix}.$$
Here, $$\mathcal{P}(\tau_1)=\left\{ \begin{bmatrix}
a &0 &0\\
0 &b &0\\
0 &0 &a
\end{bmatrix}:a,b\in \mathbb{C}\right\} \text{ and }
\mathcal{P}(\tau_2)=\left\{ \begin{bmatrix}
b &0 &0\\
0 &a &0\\
0 &0 &a
\end{bmatrix}:a,b\in
\mathbb{C}\right\}.$$
It follows from \Cref{MainTheorem} that $\tau_1$ and $\tau_2$ are peripherally automorphic maps.
It is verified that their composition $$\tau := (\tau_2\circ\tau_1) \left(\begin{bmatrix}
x_{11} &x_{12} &x_{13}\\
x_{21} &x_{22} &x_{23}\\
x_{31} &x_{32} &x_{33}
\end{bmatrix}\right) = \begin{bmatrix}
x_{11} &0 &0\\
0 &x_{22} &0\\
0 &0 &\frac{x_{11}+x_{33}+ 2 x_{22}}{4}
\end{bmatrix}$$
is not peripherally automorphic since
$$\tau\left(\begin{bmatrix}
3 & 0 & 0\\
0& 0 & 0\\
0 & 0 & 1
\end{bmatrix}\right)= \begin{bmatrix}
3 & 0 & 0\\
0& 0 & 0\\
0 & 0 & 1
\end{bmatrix}$$ but $$\tau\left(\begin{bmatrix}
3 & 0 & 0\\
0& 0 & 0\\
0 & 0 & 1
\end{bmatrix}^2\right) = \tau\left(\begin{bmatrix}
9 & 0 & 0\\
0& 0 & 0\\
0 & 0 & 1
\end{bmatrix}\right) = \begin{bmatrix}
9 & 0 & 0\\
0& 0 & 0\\
0 & 0 & \frac{5}{2}
\end{bmatrix} \neq \begin{bmatrix}
9 & 0 & 0\\
0& 0 & 0\\
0 & 0 & 1
\end{bmatrix}.$$
\end{enumerate}
\end{example}
\section{ A decomposition theorem for UCP maps}
Consider the discrete dynamics $\{ \tau ^n: n\geq 0\}$ of a UCP map on $\mathbb{M} _d$. It is clear that peripheral eigenvectors have periodic
or persistent behavior and other eigenvectors decay to zero under
this dynamics. So we may expect that $\mathbb{M} _d$ has some kind of direct
sum decomposition consisting of a persistent part and a transitive
part.
One may also expect that on the persistent part, $\tau$ acts as an automorphism.
However, it is to be kept in mind that in general UCP maps are
not diagonalizable and so eigenvectors may not describe the full
picture. There is extensive literature on this subject as this has a
lot of physical significance. Most papers obtain such a
decomposition under some assumption or the other on UCP map (See
\cite[Theorem 2.5]{Rahaman},\cite[Theorem 9]{CSU}, \cite{FV}) such as irreducibility, existence of faithful invariant
state etc. In this Section we have such a result in full generality.
It is applicable to all UCP maps.
\begin{theorem}\label{orthogonal}
Let $\tau$ be a UCP map on $\mathbb{M}_d$. Then the domain $\mathbb{M} _d$ has a
unique vector space direct sum decomposition:
$$\mathbb{M} _d=\mathcal{P}(\tau ) \oplus \mathcal{N}(\tau ),$$
where $\mathcal{P}(\tau )$ is the peripheral space of $\tau $ and $\mathcal{N}(\tau)=\{X\in \mathbb{M}_d:\lim_{n\to \infty}\tau ^n(X)=0\}$.
Furthermore, $\mathcal{P}(\tau ^m)=\mathcal{P}(\tau )$ and $\mathcal{N}(\tau ^m)= \mathcal{N}(\tau )$ for
every $m \geq 1.$
\end{theorem}
\begin{proof}
From the Jordan decomposition of $\tau$ we have
$$\mathbb{M}_d=\ker(\tau-\lambda_1\mathcal{I})^{\nu(\lambda_1)}\oplus \ker(\tau-\lambda_2\mathcal{I})^{\nu(\lambda_2)}\oplus \cdots \oplus \ker(\tau-\lambda_k\mathcal{I})^{\nu(\lambda_k)}$$
where $\mathcal{I}$ is the identity map on $\mathbb{M}_d$,
$\{\lambda_1,\lambda_2,\ldots, \lambda_k\}$ is the set of
eigenvalues of $\tau$ and $\nu(\lambda_i)$ is the algebraic
multiplicity of the eigenvalue $\lambda_i$. It is known that all the
Jordan blocks for peripheral eigenvalues of a UCP map are one
dimensional (see \cite{MW}, \cite{NAJ}). Hence the algebraic
multiplicity $\nu(\lambda)$ of the eigenvalue $\lambda \in \mathbb{T}$ is
same as its geometric multiplicity. Therefore the persistent part
$\oplus_{\lambda\in \sigma(\tau):|\lambda|=1}\ker(\tau-\lambda
\mathcal{I})^{\nu(\lambda)}$ is same as $\mathcal{P}(\tau)$.
For $X\in
\ker(\tau-\lambda_i\mathcal{I})^{\nu(\lambda_i)} $ with
$|\lambda_i|<1$, we have $\tau^n(X)\to 0$. Therefore taking $\mathcal{N}(\tau) :=
\oplus_{\lambda\in \sigma(\tau):|\lambda|<1}\ker(\tau-\lambda
\mathcal{I})^{\nu(\lambda)}$, we have the required decomposition.
Suppose $Z\in \mathbb{M} _d$, satisfies $\lim _{n\to \infty }\tau ^n(Z)=0$,
and $Z=X+Y$ is the decomposition of $Z$, with $X\in \mathcal{P}(\tau )$, $Y\in
\mathcal{N}(\tau )$. As $\lim _{n\to \infty }\tau ^n(Y)=0$, we get $\lim
_{n\to \infty}\tau ^n(X)=0$. But this is not possible unless $X=0$,
as $\tau $ is isometric on $\mathcal{P}(\tau )$. This proves the uniqueness of
the decomposition. The second part is clear from the Jordan
decomposition of $\tau .$
\end{proof}
In this theorem, the persistent part $\mathcal{P}(\tau )$ is always present as
$\tau $ is unital. The transient part $\mathcal{N}(\tau )$ is absent if and
only if $\tau $ is an automorphism. In other words if the spectrum
of $\tau $ is contained in the unit circle then $\tau $ is an
automorphism. This combined with Remark
\ref{ExtendToFinitedimensioanlC*-algebra} settles Conjecture 5.5 of
\cite{Francesco2} for
finite dimensional $C^*$-algebras.
As mentioned before, it follows
from \cite[Theorem 2.12]{BhatTalwarKar} that the persistent part
$\mathcal{P}(\tau)$ has a $C\sp \ast$-algebraic structure $(\mathcal{P}(\tau), \circ)$
and $\tau$ restricted to this subspace becomes an automorphism. In
addition, if the algebraic structure on $\mathcal{P}(\tau)$ is coming from
matrix multiplication, then from \Cref{MainTheorem}, $\tau$ becomes
peripherally automorphic. One might ask whether there exists a
unitary $U \in \mathbb{M}_d$ such that $\tau_{_{\mathcal{P}(\tau)}}(X) = U X U^\ast$
for every $X \in \mathcal{P}(\tau)$. This is not true in general as
established by the following example.
\begin{example}\label{RestrictionIsNotComingFromAUnitary}
Let $\tau:\mathbb{M}_3\to \mathbb{M}_3$ be the UCP map given by $$\tau\left(\begin{bmatrix}
x_{11} &x_{12} &x_{13}\\
x_{21} &x_{22} &x_{23}\\
x_{31} &x_{32} &x_{33}
\end{bmatrix}\right) = \begin{bmatrix}
x_{33} &0 &0\\
0 &x_{33} &0\\
0 &0 &\frac{x_{11}+x_{22}}{2}
\end{bmatrix}.$$
Consider a faithful state $\phi$ given by $\phi((x_{ij}))= \frac{x_{11}}{4}+\frac{x_{22}}{4}+\frac{x_{33}}{2}$ which satisfies $\phi (\tau (X))=\phi(X)$ for every $X \in \mathbb{M}_3$.
This shows that $\tau$ is stationary.
The set of peripheral eigenvalues of $\tau$ is $\{ 1,-1\}$ and $$\mathcal{P}(\tau)=\left\{\begin{bmatrix}
a &0 &0\\
0 &a &0\\
0 &0 &b
\end{bmatrix}:a,b\in \mathbb{C} \right\}.$$
It is not difficult to see that $\tau$ does not preserve trace of elements of $\mathcal{P}(\tau)$.
So the restriction of $\tau$ to $\mathcal{P}(\tau)$ can not be same as conjugation by any unitary in $\mathbb{M}_d$.
\end{example}
\begin{remark}
In \cite{FV} authors deal with decoherence problem using projection
maps. Let $$Q_\tau:=\frac{1}{2\pi i}\int_{\gamma}
(\mathcal{I}-\tau)^{-1}d\lambda,$$ where $\gamma$ is any Jordan
curve (i.e. a smooth and closed curve on a simply connected domain)
in the open unit disk containing all the non-peripheral eigenvalues
of $\tau$ in its bounded component. Let
$P_\tau:=\mathcal{I}-Q_\tau$. Therefore
$\mathbb{M}_d=P\tau\mathbb{M}_d\oplus Q_\tau \mathbb{M}_d$. Here
$P_\tau$ is an idempotent UCP map on $\mathbb{M}_d$ and hence
$(P_\tau\mathbb{M}_d,\circ)$ has $C\sp \ast$-algebraic structure
under Choi-Effros product $\circ$. Not surprisingly
$P_\tau\mathbb{M}_d$ is same as $\mathcal{P}(\tau).$
\end{remark}
\begin{theorem}
Let $\tau :\mathbb{M} _d\to \mathbb{M} _d$ be a stationary UCP map with faithful
invariant state $\phi .$ Then $\mathbb{M} _d$ is a Hilbert space with inner
product $\langle X, Y\rangle :=\phi (X^*Y), X,Y\in \mathbb{M} _d$ and with
respect to this inner product the decomposition in Theorem
\ref{orthogonal} is an orthogonal decomposition.
\end{theorem}
\begin{proof}
As $\tau $ is a stationary UCP maps, from Corollary 2.8 and part (1)
of Theorem 2.6, $\mathcal{P}(\tau )$ is in the multiplicative domain $M_\tau
$ of $\tau $. Therefore for any $X, Y \in \mathbb{M}_d$, with $X\in \mathcal{P}(\tau
)$, we have
$$\langle X, Y \rangle = \phi(X^\ast Y) = \phi(\tau(X^\ast Y)) = \phi(\tau(X)^\ast \tau(Y)) = \langle \tau(X), \tau(Y) \rangle
.$$
Let $Y \in \mathcal{N}(\tau ).$ Then, $\lim _{n\to \infty}\tau^n(Y) = 0$.
However, $\langle X, Y \rangle = \langle \tau(X), \tau(Y) \rangle =
\langle \tau^2(X), \tau^2(Y) \rangle = \langle \tau^n(X), \tau^n(Y)
\rangle$ for every $n \in \mathbb{N}$. Now the fact that $\tau$ is
contractive yields that $\langle X, Y \rangle = 0.$
\end{proof}
\begin{remark}\label{ExtendToFinitedimensioanlC*-algebra}
Suppose $\mathcal{A}$ is a finite dimensional unital $C^*$-algebra
and let $\tau : \mathcal{A}\to \mathcal{A}$ be a UCP map. Without
loss of generality, we may take $\mathcal{A}= \oplus _{j=1}^k\mathbb{M}
_{d_j}$ for some $d_1, \ldots , d_k, k\in \mathbb{N}.$ Let
$\mathcal{H}= \oplus _{j=1}^k(\mathbb{C}^{d_j})$ and let $P_j$ be
the projection of $\mathcal{H}$ to $\mathbb{C}^{d_j}$. Let
$\mathcal{B}$ be the $C^*$-algebra $\mathscr{B}(\mathcal{H})= \mathbb{M}
_{d}$ where $d= \sum _{j=1}^kd_j.$ Define $\tilde{\tau}:\mathbb{M} _d\to \mathbb{M}
_d$ by $\tilde{\tau}(X) = \tau ( \sum _{j=1}^dP_jXP_j)$. Then the
peripheral space of $\tilde{\tau }$ is identical to that of $\tau $.
This way, all the analysis done here can be extended to UCP maps on
general finite dimensional $C^*$-algebras.
\end{remark}
A UCP map $\tau : \mathbb{M} _d\to \mathbb{M} _d$ is said to be faithful if $\tau(X^\ast X)=0$ implies $X=0$. The following example shows that a faithful
UCP map need not be peripherally automorphic.
\begin{example}
Let $$\tau\left(\begin{bmatrix}
x_{11} &x_{12} &x_{13}\\
x_{21} &x_{22} &x_{23}\\
x_{31} &x_{32} &x_{33}
\end{bmatrix}\right) = \begin{bmatrix}
x_{11} &0 &0\\
0 &x_{22} &0\\
0 &0 & \frac{x_{11} + x_{22}+ x_{33}}{3}
\end{bmatrix}.$$
Then $\tau$ is a faithful UCP map on $\mathbb{M}_3$ such that $$\tau\left(\begin{bmatrix}
0 & 0 & 0\\
0& 2 & 0\\
0 & 0 & 1
\end{bmatrix}\right)= \begin{bmatrix}
0 & 0 & 0\\
0& 2 & 0\\
0 & 0 & 1
\end{bmatrix}$$ but $$\tau\left(\begin{bmatrix}
0 & 0 & 0\\
0& 2 & 0\\
0 & 0 & 1
\end{bmatrix}^2\right) = \tau\left(\begin{bmatrix}
0 & 0 & 0\\
0& 4 & 0\\
0 & 0 & 1
\end{bmatrix}\right) = \begin{bmatrix}
0 & 0 & 0\\
0& 4 & 0\\
0 & 0 &
\frac{5}{3}
\end{bmatrix} \neq \begin{bmatrix}
0 & 0 & 0\\
0& 4 & 0\\
0 & 0 & 1
\end{bmatrix}.$$
Thus, $\tau$ is not peripherally automorphic.
\end{example}
\section*{Acknowledgments}
Bhat gratefully acknowledges funding from SERB(India) through JC
Bose Fellowship No. JBR/2021/000024. Kar is thankful to NBHM (India) for funding.
Talwar thanks ISI Bangalore for financial support through Research Associate scheme.
A part of this work was completed during his research assistantship at Nazarbayev University and he appreciates their support.
|
\section{Introduction}
Jensen's square principle $\Box_\k$ at an infinite successor cardinal $\k^+$, as well as its natural weakenings $\Box_{\k, {<}\l}$, are very well studied principles in combinatorial set theory that can be naturally viewed as incompactness or anti-reflection principles. For example, and as is well-known, $\Box_\k$ implies that for every stationary $S\sub\k^+$ there is a stationary $S'\sub S$ which does not reflect (i.e., such that $S'\cap\a$ is non-stationary for every $\a<\k^+$). We recall the definition of these weakenings of $\Box_\k$.
\begin{definition}
Given a cardinal $\k\geq\o_1$ and a cardinal $\l\leq\k^+$, $\Box_{\k, {<}\l}$ holds if and only if there is a sequence $(\mtcl C_\a\,;\,\a\in\Lim(\k^+))$ with the following properties.
\begin{enumerate}
\item For every $\a\in \Lim(\k^+)$,
\begin{enumerate}
\item $\mtcl C_\a$ is a collection of clubs of $\a$,
\item $\ot(C)\leq\k$ for every $C\in\mtcl C_\a$, and
\item $|\mtcl C_\a|<\l$.
\end{enumerate}
\item (Coherence) For every $\a\in\Lim(\k^+)$, $C\in\mtcl C_\a$, and $\bar\a<\a$, if $\bar\a$ is a limit point if $C$, then $C\cap\bar\a\in \mtcl C_{\bar\a}$.
\end{enumerate}
\end{definition}
In the above definition, and throughout the paper, given an ordinal $\mu$, $\Lim(\mu)$ will denote the set of nonzero limit ordinals in $\m$.
We also write $\Box_{\k, \l}$ for $\Box_{\k, {<}\l^+}$. We will be focusing our attention on the weakest form of $\Box_{\k, \l}$ for $\k=\o_1$, i.e., on $\Box_{\o_1, \o_1}$. This principle is also called \emph{weak square} and, by a well-known result of Jensen, is equivalent to the existence of a special $\al_2$-Aronszajn tree.
Forcing axioms are reflection principles, between forcing extensions and the ground model, which tend to imply failures of square. For example, a well-known of Todor\v{c}evi\'{c} (\cite{Todorcevic}) is that the Proper Forcing Axiom ($\PFA$) implies that $\Box_\k$ fails for every uncountable cardinal $\k$.
Given a cardinal $\k$ and a class $\mtcl K$ of partial orders, we denote the forcing axiom for $\mtcl K$ with respect to families of $\k$-many dense sets by $\FA_\k(\mtcl K)$. This is the assertion that for every $\mtbb P\in\mtcl K$ and every sequence $(D_i\,:\,i<\k)$ of dense subsets of $\mtbb P$ there is a filter $G\sub\mtbb P$ such that $G\cap D_i\neq\emptyset$ for all $i<\k$.
We will say that \emph{$\FA_\k(\mtcl K)$ is the forcing axiom for $\mtcl K$ at $\k$}.
In \cite{genMA}, a certain forcing axiom $\MA^{1.5}_\k$ extending Martin's Axiom at $\k$ (i.e., $\MA_\k$) is introduced. It is proved in that paper that if $\CH$ holds, $\k$ is any closed enough cardinal, and a suitable diamond principle holds, then $\MA^{1.5}_\k$ holds in some cardinal-preserving forcing extension. The first half of the present paper is devoted to the study of the effects, on square principles at $\o_2$, of forcing axioms which are natural mild strengthenings of $\MA^{1.5}_{\al_2}$. The main result in this part of the paper is that one such axiom, which we will refer to as $\MA^{1.5}_{\o_2}(\mbox{stratified})$, implies $\Box_{\o_1, \o_1}$.
Moreover, $\MA^{1.5}_{\o_2}(\mbox{stratified})$, and in fact $\MA^{1.5}_\k(\mbox{stratified})$ for any given $\k$ fixed in advance, is consistent by essentially the same proof as in \cite{genMA}.
Certainly, weak forcing axioms such as $\MA_\k$, for any $\k$, are compatible with square principles: any cardinal-preserving forcing extension of a model with squares will preserve them, and $\MA_\k$ can always be forced by c.c.c.\ forcing. What is perhaps surprising, given that strong forcing axioms at $\o_1$ like $\PFA$ prohibit squares, is the fact that strong enough forcing axioms at $\o_2$, like $\MA^{1.5}_{\o_2}(\mbox{stratified})$, outright imply square principles at $\o_2$.
Similar results have been obtained by Neeman. One of the results in \cite{Neeman} is that a forcing axiom $\MA^{1.5}_{\o_2}(U)$, whose definition involves a certain parameter $U\sub [\o_2]^{\al_0}$, implies both $\Box_{\o_1, {<}\o}$ and the following strengthening $\Box_{\o_1, \o}^{\textrm{ta}}$ of $\Box_{\o_1, \o}$:\footnote{On the other hand, the definition of $\MA^{1.5}_{\o_2}(\mbox{stratified})$ is parameter-free.}
Given cardinals $\l\leq\k$ such that $\k\geq\o_1$, $\Box^{\text{ta}}_{\k, \l}$ holds if and only if there is a $\Box_{\k, \l}$-sequence $(\mtcl C_\a\,:\,\a\in\Lim(\k^+))$ such that for every $\a\in \Lim(\k^+)$ and for all $C$, $C'\in\mtcl C_\a$, $C$ and $C'$ agree on a tail, i.e., there is some $\b<\a$ such that $C\setminus\b=C'\setminus\b$.\footnote{The superscript \text{ta} stands for `tail agreement'.}
Neeman also points out in \cite{Neeman} that both $\Box_{\o_1, \o}$ and $\Box_{\o_1, \o}^{\textrm{ta}}$ follow from some of his strong high analogues of $\PFA$. On a related vein,\footnote{One important aspect in which Sakai's result is different from both Neeman's results and the first main result in the present paper is that Sakai's theorem involves strong forcing axiom at $\o_1$, whereas the others are implications from forcing axioms at $\o_2$.} Sakai shows in \cite{Sakai} that Martin's Maximum proves that $\Box^{\text p}_{\o_1}$ (i.e, partial square at $\o_2$) holds and that this is not the case for $\PFA$.
Given a set $N$ such that $N\cap\o_1\in \o_1$, we denote this ordinal by $\d_N$ and call it \emph{the height of $N$}. Given a sequence $\vec e=(e_\a\,:\,\a\in \o_2)$, where $e_\a:|\a|\into \a$ is a bijection for each $\a<\o_2$, we say that a set $N$ is \emph{closed under $\vec e$} if $e_\a\restr\x+1\in N$ whenever $\a\in \o_2\cap N$ and $\x\in |\a|\cap N$. We will be using the following well-known fact repeatedly, sometimes without mention.
\begin{fact}\label{agreement} Suppose $\vec e=(e_\a\,:\,\a\in \o_2)$, where $e_\a:|\a|\into \a$ is a bijection for each $\a<\o_2$, and suppose $N_0$ and $N_1$ are countable submodels of $H(\o_2)$ closed under $\vec e$ and such that $\d_{N_0}\leq\d_{N_1}$. Then $N_0\cap\a\sub N_1$ for every $\a\in N_0\cap N_1\cap\o_2$.
\end{fact}
\begin{proof}
Given any $\bar\a\in N_0\cap\a$ there is some $\xi\in N_0\cap |\a|$ such that $e_\a(\xi)=\bar\alpha$. But since $\a$ and $\xi$ are both members of $N_1$ as $|\a|\leq\o_1$, we also have that $\bar\alpha=e_\a(\xi)\in N_1$.\end{proof}
\begin{corollary}\label{agreement-0}
Suppose $\vec e=(e_\a\,:\,\a\in \o_2)$, where $e_\a:|\a|\into \a$ is a bijection for each $\a<\o_2$, and suppose $N_0$ and $N_1$ are countable submodels of $H(\o_2)$ closed under $\vec e$ of the same height. Then $N_0\cap N_1\cap\o_2$ is an initial segment of both $N_0\cap\o_2$ and $N_1\cap\o_2$.
\end{corollary}
As already mentioned, \cite{genMA} introduces a strengthening $\MA^{1.5}_\k$ of $\MA_\k$, for any given cardinal $\k$. $\MA^{1.5}_\k$ is the forcing axiom $$\FA_\k(\{\mtbb P\,:\,\mtbb P\mbox{ has the }\al_{1.5}\mbox{-c.c.}\}),$$ where having the $\al_{1.5}$-c.c.\ is defined as follows.\footnote{One can also define $\MA^{1.5}_{{<}\k}$ as $\MA^{1.5}_\l$ for all $\l<\k$.}
\begin{definition} (\cite{genMA}) A partial order $\mtbb P$ has the $\al_{1.5}$-c.c.\ if for every large enough cardinal $\t$ (i.e., every cardinal $\t$ such that $\mtbb P\in H(\t)$) there is a club $E$ of $[H(\t)]^{\al_0}$ such that for every finite $\mtcl N\sub E$ and every $N_0\in\mtcl N$, if $N_0$ has minimum height within $\mtcl N$, then for every $p_0\in N_0\cap\mtbb P$ there is some extension $p\in\mtbb P$ of $p_0$ such that $p$ is $(N, \mtbb P)$-generic for all $N\in\mtcl N$. \end{definition}
It is easy to see (s.\ \cite{genMA}) that every partial order with the countable chain condition has the $\al_{1.5}$-c.c.\ and every partial order with the $\al_{1.5}$-c.c.\ is proper and has the $\al_2$-c.c.\footnote{Hence the $\al_{1.5}$ notation.}
We can naturally strengthen $\MA^{1.5}_\k$ by restricting the definition to finite families $\mtcl N\sub E$ with some given nice structural property $P$. We may then refer to the corresponding enlargement of $$\{\mtbb P\,:\,\mtbb P\mbox{ has the }\al_{1.5}\mbox{-c.c.}\}$$ as the class of partial orders with the $\al_{1.5}$-c.c.\ with respect to finite families with property $P$, and may denote the corresponding forcing axiom by $\MA^{1.5}_\k(P)$.
The following notion of stratified family seems to give rise to a particularly useful form of $\MA^{1.5}_\k$.
\begin{definition}\label{stratified}
A collection $\mtcl N$ of countable elementary submodels of $H(\t)$, for some infinite cardinal $\t$, is \emph{stratified} in case for all $N_0$, $N_1\in\mtcl N$, if $\d_{N_0}<\d_{N_1}$, then in fact $\ot(N_0\cap\o_2)<\d_{N_1}$.
\end{definition}
Correspondingly, we say that a forcing notion $\mtbb P$ \emph{has the $\al_{1.5}$-c.c.\ with respect to finite stratified families of models} iff for every infinite cardinal $\t$ such that $\mtbb P\in H(\t)$ there is a club $E\sub [H(\t)]^{\al_0}$ such that for every finite stratified $\mtcl N\sub E$, if $p\in N_0\cap\mtbb P$, where $N_0\in\mtcl N$ is of minimal height within $\mtcl N$, then there is an extension $p^*$ of $p$ in $\mtbb P$ such that $p^*$ is $(N, \mtbb P)$-generic for every $N\in \mtcl N$. Clearly, every partial order with the $\al_{1.5}$-c.c.\ also has the $\al_{1.5}$-c.c.\ with respect to finite stratified families of models. Also, given a cardinal $\k$, we write $\MA^{1.5}_\k(\mbox{stratified})$ to denote $\FA_\k(\mtcl K)$, where $\mtcl K$ is the class of partial orders $\mtbb P$ such that $\mtbb P$ has the $\al_{1.5}$-c.c.\ with respect to finite stratified families of models. We then have that
$\MA^{1.5}_\k(\mbox{stratified})$ implies $\MA^{1.5}_\k$.
The following proposition extends the aforementioned fact that every forcing with the $\al_{1.5}$-c.c.\ is proper and has the $\al_2$-c.c.
\begin{proposition}\label{al15-2} If a forcing notion has the $\al_{1.5}$-c.c.\ with respect to finite stratified families of models, then it is proper and has the $\al_2$-c.c.\end{proposition}
\begin{proof} Suppose $\mtbb P$ is a forcing notion with the $\al_{1.5}$-c.c.\ with respect to finite stratified families of models. Let $\t$ be a cardinal such that $\mtbb P\in H(\t)$ and let $E\sub [H(\t)]^{\al_0}$ be a club witnessing, for $H(\t)$, that $\mtcl P$ has the $\al_{1.5}$-c.c.\ with respect to finite stratified families of models. Given any $N\in E$, $\{N\}$ is trivially stratified, and therefore for every $p\in \mtbb P\cap N$ there is an $(N, \mtbb P)$-generic extension of $p$. This shows that $\mtbb P$ is proper.
To prove that $\mtbb P$ has the $\al_2$-chain condition let us assume, towards a contradiction, that there is a maximal antichain $A$ of $\mtbb P$ such that $|A|\geq\al_2$, and let $(p_i\,:\,i<\l)$ be a one-to-one enumeration of $A$, for some $\l\geq\o_2$. Let $M$ be an elementary submodel of some large enough $H(\chi)$ such that
\begin{enumerate}
\item $E$, $A\in M$ and
\item $|M|=\al_1$.
\end{enumerate}
Let $i_0\in \o_2\setminus M$ and let $N\preccurlyeq H(\t)$ be countable and such that $p_{i_0}$, $E$, $M\in N$.
By correctness of $M$ we may find $i_1\in \o_2\cap M$ for which there is some $N'\in E\cap M$ such that $\d_{N'}=\d_N$ and $p_{i_1}\in N'$.
Indeed, the existence of such an $N'$ is expressed by a true sentence, as witnessed by $i_0$ and $N$, with $\d_N$, $E$ and $(p_i\,:\,i<\l)$ as parameters.
We note that $\mtcl N=\{N, N'\}$ is a stratified family of members of $E$ as $\d_N=\d_{N'}$. It follows, since $p_{i_0}\in N$ and $\d_{N}=\d_{N'}$, that we may find an $(N', \mtbb P)$-generic condition $p$ extending $p_{i_0}$. Then there must be condition $p'$ extending $p$ and extending some $\bar p\in A\cap N'$. But that is impossible since $A$ is an antichain and $\bar p\neq p_{i_0}$ as $N'\sub M$.
\end{proof}
Essentially the same forcing construction from \cite{genMA} showing the consistency of $\MA^{1.5}_{{<}\k}$, for any given closed enough $\k$, can be used to prove the following theorem.\footnote{The main point is that all relevant systems of models coming up in the proof from \cite{genMA} turn out in fact to be stratified.}
\begin{theorem}\label{mainthm000} ($\textsf{CH}$) Let $\k\geq\o_2$ be a regular cardinal such that $\m^{\al_0}< \kappa$ for all $\m < \k$ and $\diamondsuit(\{\a<\k\,:\,\cf(\a)\geq\o_1\})$ holds. Then there is a proper forcing notion $\mtcl P$ of size $\k$ with the $\al_2$-chain condition such that the following statements hold in the generic extension by $\mtcl P$.
\begin{enumerate}
\it[(1)] $2^{\al_0}=\k$
\it[(2)] For every $\l<\k$, $\MA^{1.5}_\l(\mbox{stratified})$
\end{enumerate}
\end{theorem}
Let us write $\PFA_{\al_2}(\aleph_2\mbox{-c.c.})$ to denote $\FA_{\al_2}(\mtcl K)$, where $\mtcl K$ is the class of proper forcing notions with the $\al_2$-chain condition.
Theorem \ref{mainthm000}, as well as other similar strengthenings of the main result from \cite{genMA}, motivate the following question.\footnote{This question is also motivated by the main result from \cite{Asp-Golsh}, to the effect that it is consistent, for arbitrary choice of $\kappa$, that $\FA_\k(\{\mtbb P\,:\, \mtbb P\mbox{ proper}, |\mtbb P|=\al_1\})$ holds. It is worth pointing out that the proof of this theorem is very different from the proof of the main result from \cite{genMA}.}
\begin{question} Is $\PFA_{\al_2}(\aleph_2\mbox{-c.c.})$ consistent?
\end{question}
We are not able to answer this question. However, in Section \ref{s2}
we will show that $\MM_{\al_2}(\aleph_2\mbox{-c.c.})$, a natural strengthening of $\PFA_{\al_2}(\aleph_2\mbox{-c.c.})$, is in fact inconsistent. $\MM_{\al_2}(\aleph_2\mbox{-c.c.})$ is $\FA_{\al_2}(\mtcl K^*)$, where $\mtcl K^*$ is the class of forcing notions that both preserve stationary subsets of $\o_1$ and have the $\al_2$-chain condition.
One of the ingredients of this proof will be the fact that $\MM_{\al_2}(\aleph_2\mbox{-c.c.})$ implies $\Box_{\o_1, \o_1}$ (since $\MM_{\al_2}(\aleph_2\mbox{-c.c.})$ extends $\MA^{1.5}_{\al_2}(\mbox{stratified})$).
There are other results in the literature dealing with failures of forcing axioms at $\al_2$ or above. In this respect we single out the following theorem of Shelah (\cite{Shelah}), extended by the main result in Section \ref{s2}.
\begin{theorem} (Shelah) Given any regular cardinal $\l>\o_1$, $\FA_\l(\mtcl K_\l)$ is false, where $\mtcl K_\l$ is the class of forcing notions preserving all stationary subsets of $\m$ for every uncountable regular cardinal $\m\leq\l$.
\end{theorem}
We will actually give two proofs of the failure of $\MM_{\al_2}(\aleph_2\mbox{-c.c.})$. The first one, in Section \ref{s2}, we have already referred to. The second proof, in Section \ref{s4}, will make use of a consequence of $\MA^{1.5}_{\al_2}(\mbox{stratified})$ regarding canonical functions. We recall that if $\a<\o_2$ is a nonzero ordinal and $\p:\o_1\into\a$ is a surjection, the function $g_\a:\o_1\into \o_1$ defined by letting $g_\a(\n)=\ot(\p``\n)$ is called a, or the, \emph{canonical function for $\a$}. The use the definite article when referring to canonical functions for a given $\a$ is justified by the following obvious observation, which in particular implies that $g_\a$ is uniquely determined modulo clubs.
\begin{fact} Given $\a<\o_2$ and given surjections $\p_0$, $\p_1:\o_1\into \a$ there is a club of $\n<\o_1$ such that $\p_0``\n=\pi_1``\n$. \end{fact}
Another standard fact about the canonical function for $\a$ is that it represents the ordinal $\a$ in every generic ultrapower of $V$ obtained from forcing with $\mtcl P(\o_1)/\NS_{\o_1}$. In other words, if $g$ is a canonical function for $\a$, then $\mtcl P(\o_1)/\NS_{\o_1}$ forces that, letting $M=((^{\o_1^V}V)\cap V)/\dot G$ be the generic ultrapower of $V$ obtained from $\dot G$, the set of $M$-ordinals below the class $[g]_{\dot G}$ of $g$ in $M$ is well-ordered in order type $\a$.
\emph{Club-bounding by canonical functions}, $\CB$, is the statement that every function $f:\o_1\into \o_1$ is bounded on a club by the canonical function of some nonzero $\a<\o_2$. $\CB$ is a weakening of $\NS_{\o_1}$ being saturated.
A natural weakening of $\CB$ is \emph{weak Chang's Conjecture}, $\wCC$, which is the statement that for every function $f:\o_1\into \o_1$ there is some $\a<\o_2$ such that $\{\n<\o_1\,:\, f(\n)<g(\n)\}$ is stationary for every canonical function $g$ for $\a$ (s.\ \cite{donder-levinski}). Martin's Maximum implies the saturation of $\NS_{\o_1}$ (\cite{FMS}), and hence also $\CB$. On the other hand, it is not difficult to see that not even $\wCC$ follows from $\PFA$ (s.\ e.g.\ \cite{Asp} for strong forms of this non-implication).
In Section \ref{s3} we will prove that $\MA^{1.5}_{\al_2}(\mbox{stratified})$ implies $\lnot\wCC$. Using this implication, in Section \ref{s4} we will give another proof of the inconsistency of $\MM_{1.5}(\aleph_2\mbox{-c.c.})$. Specifically, assuming this forcing axiom, we will produce a sequence $(f_n)_{n<\o}$ of functions from $\o_1$ to $\o_1$ such that for each $n$, $f_n$ dominates $f_{n+1}$ on a club. This is of course impossible as then we are able to find an infinite decreasing sequence of ordinals.
The rest of the paper is structured as follows.
In the following section we prove that $\MA^{1.5}_{\al_2}(\mbox{stratified})$ implies $\Box_{\o_1, \o_1}$. Using this result, in Section \ref{s2} we prove that $\MM_{\al_2}(\aleph_2\mbox{-c.c.})$ is false. In Section \ref{s3} we prove that $\MA^{1.5}_{\al_2}(\mbox{stratified})$ implies $\lnot\wCC$. Finally, in Section \ref{s4}, using the implication in Section \ref{s3}, we give another proof of the inconsistency of $\MM_{\al_2}(\aleph_2\mbox{-c.c.})$.
Throughout the paper, we will write $S^2_1$ for $\{\a<\o_2\,:\,\cf(\a)=\o_1\}$, and sometimes $S^2_0$ for $\{\a<\o_2\,:\,\cf(\a)=\o\}$. All undefined pieces of notation are hopefully standard and may be found in \cite{JECH} or \cite{KUNEN}.
\section{$\MA^{1.5}_{\al_2}(\mbox{stratified})$ implies $\Box_{\o_1, \o_1}$}\label{s1}
The main theorem in this section is the following.
\begin{theorem}\label{thm1} $\MA^{1.5}_{\al_2}(\mbox{stratified})$ implies $\Box_{\o_1, \o_1}$.
\end{theorem}
Let $\vec e=(e_\a\,:\, \a<\o_2)$ be such that $e_\a:|\a|\into\a$ is a bijection for each $\a<\o_2$.
We define the following forcing notion $\mtcl P$.
Conditions in $\mtcl P$ are triples $$p=(h^p, i^p, \mtcl N_p)$$ with the following properties.
\begin{enumerate}
\item $h^p$ is a function such that $\dom(h_p)\in [\Lim(\o_2)\times\o_1\times (\o_1+1)]^{{<}\o}$ and such that for each $(\a, \n, \tau)\in \dom(h^p)$, $h^p(\a, \n, \tau)\sub\tau\times\a$ is a finite function which can be extended to a strictly increasing and continuous function $f:\tau\into\a$ with range cofinal in $\a$.
\item For every $(\a, \n, \tau)\in\dom(h^p)$:
\begin{enumerate}
\item if $\cf(\a)=\o_1$, then $\tau=\o_1$ and $\n=0$;
\item if $\cf(\a)=\o$, then $\tau\in\Lim(\o_1)$.
\end{enumerate}
\item For every $\a$ and $\n$ there is at most one $\tau$ such that $(\a, \n, \tau)\in\dom(h^p)$.
\item $i^p$ is a function whose domain is the set of triples $(\a, \n, \bar\tau)$ such that $(\a, \n, \tau)\in\dom(h^p)$ for some $\tau$ and $\bar\tau\in\dom(h^p(\a, \n, \tau))\cap\Lim(\o_1)$, and $i^p(\a, \n, \bar\tau)\in\o_1$ for each $(\a, \n, \bar\tau)\in\dom(i^p)$.
\item The following holds for each $(\a, \n, \tau)\in \dom(h^p)$.
\begin{enumerate}
\item For every limit ordinal $\bar\tau\in\dom(h^p(\a, \n, \tau))$,
$$(h^p(\a, \n, \tau)(\bar\tau), i^p(\a, \n, \bar\tau), \bar\tau)\in \dom(h^p)$$ and $$h^p(h^p(\a, \n, \tau)(\bar\tau), i^p(\a, \n, \bar\tau), \bar\tau)=h^p(\a, \n, \tau)\restr\bar\tau$$
\end{enumerate}
\item $\mtcl N$ is a finite stratified collection of countable elementary submodels of $(H(\o_2); \in)$ closed under $\vec e$.
\item For every $N\in\mtcl N$ and every $(\a, \n, \tau)\in\dom(h^p)$ such that $\a$, $\n\in N$:
\begin{enumerate}
\item $\tau\in N$;
\item $h^p(\a, \n, \tau)\restr N\sub N$;\footnote{In particular, if $\cf(\a)=\o$, then $h^p(\a, \n, \tau)\in N$.}
\item If $\cf(\a)=\o_1$, then $$\d_N\in\dom(h^p(\a, \n, \o_1))$$ and $$h^p(\a, \n, \o_1)(\d_N)=\ssup(N\cap\a)$$
\item For every $\bar\tau\in\dom(h^p(\a, \n, \tau))\cap\Lim(\o_1)\cap N$, $i^p(\a, \n, \bar\tau)\in N$.
\end{enumerate}
\end{enumerate}
Given $\mtcl P$-conditions $p_0$ and $p_1$, $p_1$ extends $p_0$ if and only if:
\begin{enumerate}
\item $\dom(h^{p_0})\sub\dom(h^{p_1})$;
\item for every $(\a, \n, \tau)\in \dom(h^{p_0})$,
\begin{enumerate}
\item $h^{p_0}(\a, \n, \tau)\sub h^{p_1}(\a, \n, \tau)$, and
\item $i^{p_1}(\a, \n, \bar\tau)=i^{p_0}(\a, \n, \bar\tau)$ for each $\bar\tau\in\dom(h^{p_0}(\a, \n, \tau))\cap\Lim(\o_1)$.
\end{enumerate}
\item $\mtcl N_{p_0}\sub \mtcl N_{p_1}$.
\end{enumerate}
Given $p\in\mtcl P$, we denote $\{\a\in S^2_1\,:\,(\a, \n, \tau)\in\dom(h^p)\mbox{ for some }\n,\,\tau\}$ by $X_p$.
\begin{lemma}\label{aleph15} $\mtcl P$ has the $\al_{1.5}$-c.c.\ with respect to finite stratified families of models.
\end{lemma}
\begin{proof} Let $\t$ be such that $\mtcl P\in H(\t)$ and let $\mtcl N^*$ be a finite stratified collection of countable elementary submodels of $H(\t)$ containing $\vec e$.
We may assume that for each $N^*\in\mtcl N^*$, $N^*=\bigcup_{i<\d_{N^*}}N^*_\nu$, where $(N^*_\nu)_{\nu<\d_{N^*}}$ is a continuous $\in$-chain of countable elementary submodels of $H(\t)$ containing $\vec e$.
In fact, if there is any cardinal $\chi>\t$ such that $N^*$ is of the form $N^{**}\cap H(\t)$ for a countable $N^{**}\elsub H(\chi)$ with $\vec e$, $\t\in N^{**}$, then $N^*$ is a continuous $\in$-chain of countable elementary submodels of $H(\t)$ as above. To see this, let $(x_n)_{n<\o}$ be an enumeration of $N^*$ and let $(\b_n)_{n<\o}$ be a strictly increasing sequence converging to $\d_{N^*}$. Then, by correctness of $N^{**}$, we may build a sequence $(\vec N^*_n)_{n<\o}$ of members of $N^{**}$ such that
\begin{enumerate}
\item for each $n$, $\vec N^*_n = (N^*_{n, i})_{i\leq \b_n}$ is a continuous $\in$-chain of length $\b_n+1$ consisting of countable elementary submodesl of $H(\t)$ containing $\vec e$ and $x_n$;
\item $\vec N^*_{n+1}$ extends $\vec N^*_n$. \end{enumerate}
$\bigcup_{n<\o}\vec N^*_n$ is then as desired.
Let $\mtcl N=\{N^*\cap H(\o_2)\,:\, N^*\in\mtcl N^*\}$.
Let also $N_0\in\mtcl N$ be of minimal height and let $p_0\in N_0$ be a $\mtcl P$-condition. Given any $\a\in X_{p_0}$, let $(\a(k)\,:\,k<m_\a)$ be the strictly increasing enumeration of $$\{\ssup(N\cap\a)\,:\, N\in\mtcl N,\,\a\in N\}$$ and, for every $k<m_\a$, let $\d^\a_k=\d_N$ for any $N\in\mtcl N$ such that $\a\in N$ and $\ssup(N\cap\a)=\a(k)$. Note that by Fact \ref{agreement}, $\d^\a_k$ is well-defined for every $k<m_\a$ (i.e., $\d^\a_\k$ is independent from the choice of $N$ as long as $\a\in N$ and $\ssup(N\cap\a)=\a(k)$) as in fact $N\cap\a=N'\cap\a$ whenever $N$, $N'\in\mtcl N$ are such that $\a\in N\cap N'$ and $\d_N=\d_{N'}$. For each $\a\in X_{p_0}$ and $k<m_\a$, let $i(\a, k)\in\d^\a_k\setminus \range(i^{p_0})$ be such that $\d_N<i(\a, k)$ for each $N\in\mtcl N$ with $\d_N<\d^\a_k$.
In order to prove the lemma, it suffices to show that $$p^*=(h^{p^*}, i^{p^*}, \mtcl N_{p_0}\cup\mtcl N)$$ is an $(N^*, \mtcl P)$-generic condition for each $N^*\in\mtcl N^*$, where $$\dom(h^{p^*})=\dom(h^{p_0})\cup \{(\a(k), i(\a, k), \d^\a_k)\,:\,\a\in X_{p_0},\,k<m_\a\}$$ and where for each $(\a, \n, \tau)\in\dom(h^{p^*})$:
\begin{enumerate}
\item if $(\a, \n, \tau)\in\dom(h^{p_0})$ and $\cf(\a)=\o$, then
\begin{enumerate}
\item $h^{p^*}(\a, \n, \tau)=h^{p_0}(\a, \n, \tau)$ and
\item $i^{p^*}(\a, \n, \bar\tau)=i^{p_0}(\a, \n, \bar\tau)$ for each $\bar\tau\in\dom(h^{p_0}(\a, \n, \tau))\cap\Lim(\o_1)$;
\end{enumerate}
\item if $(\a, \n, \tau)\in\dom(h^{p_0})$ and $\cf(\a)=\o_1$,\footnote{In this case of course $\n=0$ and $\tau=\o_1$.} then
\begin{enumerate}
\item $h^{p^*}(\a, \n, \tau)=h^{p_0}(\a, \n, \tau)\cup\{(\d^\a_k, \a(k))\,:\,k<m_\a\}$,
\item $i^{p^*}(\a, \n, \bar\tau)=i^{p_0}(\a, \n, \bar\tau)$ for every $\bar\tau\in\dom(h^{p_0}(\a, \n, \tau))\cap\Lim(\o_1)$, and
\item $i^{p^*}(\a, \n, \d^\a_k)=i(\a, k)$ for each $k<m_\a$;
\end{enumerate}
\item if $\a\in X_{p_0}$ and $k<m_\a$, then
\begin{enumerate}
\item $h^{p^*}(\a(k), i(\a, k), \d^\a_k)=h^{p_0}(\a, 0, \o_1)\cup\{(\d^\a_{j}, \a(j))\,:\, j<k\}$,
\item $i^{p^*}(\a(k), i(\a, k), \bar\tau)=i^{p_0}(\a, 0, \bar\tau)$ for every limit ordinal $\bar\tau\in\dom(h^{p_0}(\a, 0, \o_1))$, and
\item $i^{p^*}(\a(k), i(\a, k), \d^\a_j)=i(\a, j)$ for each $j<k$.
\end{enumerate}
\end{enumerate}
\begin{claim}\label{cl0}
If $\a_0<\a_1$ are such that $\a_0$, $\a_1\in X_{p_0}$, then $\a_0(k_0)<\a_0< \a_1(k_1)$ for all $k_0<m_{\a_0}$ and $k_1<m_{\a_1}$.
\end{claim}
\begin{proof}
The inequality $\a_0(k_0)<\a_0$ is immediate given that $\a_0(k_0)=\ssup(M\cap\a_0)$ for some countable $M$. Also, we note that if $N\in\mtcl N$ is such that $\a_1\in N$ and $\a_1(k_1)=\ssup(N\cap\a_1)$, then $\a_0\in N$ by Fact \ref{agreement} since $\a_1\in N_0\cap N$ and $\d_{N_0}\leq\d_N$. Hence $\a_0<\ssup(N\cap\a_1)=\a_1(k_1)$.
\end{proof}
\begin{claim}\label{cl1}
For every $N\in\mtcl N$, $\a\in X_{p_0}$ and $k<m_\a$, if $\a(k)$, $i(\a, k)\in N$ and $j<k$, then $\a(j)\in N$.
\end{claim}
\begin{proof}
Since $i(\a, k)\in N$, we have that $\d_N\geq\d^\a_k$. But $\a(j)\in M\cap\a(k)$ for every $M\in\mtcl N$ such that $\a\in M$ and $\d_M=\d^\a_k$ (since $\a(j)=\ssup(e_\a``\d^\a_j)$), and $M\cap\a(k)\sub N\cap\a(\k)$ for every such $M$, where the inclusion follows from Fact \ref{agreement} since $\d_N\geq\d^\a_k$.
\end{proof}
The proof of the following claim is essentially the same.
\begin{claim}\label{cl2}
For all $N\in\mtcl N$ and $\a\in X_{p_0}$, if $\a\in N$, then $h^{p_0}(\a, 0, \o_1)\in N$ and $i^{p_0}(\a, 0, \bar\tau)\in N$ for every $\bar\tau\in\dom(h^{p_0}(\a, 0, \o_1))\cap\Lim(\o_1)$.
\end{claim}
Using the above three claims together with Fact \ref{agreement}, one can easily verify that $p^*$ is a $\mtcl P$-condition, and it obviously extends $p_0$. Let now $N^*\in\mtcl N^*$ and let us prove that $p^*$ is $(N^*, \mtcl P)$-generic. For this, let $D\in N^*$ be an open and dense subset of $\mtcl P$ and let $p\in D$ extend $p^*$. We will prove that there is a condition $r\in D\cap N^*$ compatible with $p$.
Let $(N^*_\n)_{\n<\d_{N^*}}$ be a continuous $\in$-chain of countable elementary submodels of $H(\t)$ containing $\vec e$ such that $N^*=\bigcup_{\n<\d_{N^*}}N^*_\n$. Since $\mtcl N_p$ is stratified and $N^*\cap H(\o_2)\in\mtcl N_p$, we may find some $\n_0<\d_{N^*}$ such that
\begin{enumerate}
\item $(h^p\cup i^p)\cap N^*\sub N^*_{\n_0}$,
\item there is some $\eta\in N^*_{\n_0}\cap\o_2$ such that $[\eta,\,\o_2)\cap N^*_{\n_0}\cap N=\emptyset$ for every $N\in\mtcl N_p$ with $\d_N<\d_{N^*_{\n_0}}$, and
\item for every $\a\notin N^*_{\n_0}$ such that $(\a, \n, \tau)\in \dom(h^p)$ for some $\n$, $\tau$, and such that $\a^*=\min((N^*_{\n_0}\cap\o_2)\setminus\a)$ exists, there is some $\eta_\a\in N^*_{\n_0}\cap\a^*$ with $[\eta_\a,\,\a)\cap N^*_{\n_0}\cap N=\emptyset$ for every $N\in\mtcl N_p$ such that $\d_N<\d_{N^*_{\n_0}}$.
\end{enumerate}
Given a $\mtcl P$-condition $q$, let $\mtbb M(q)$ be a structure with universe $$\mtcl U_q:=(\bigcup\{\{\a, \n, \tau, \x, \b\}\,:\, (\a, \n, \tau)\in\dom(h^q),\, (\x, \b)\in h^q(\a, \n, \tau)\})\cup i^q$$
coding $h^q$ and $i^q$ in some fixed canonical way.
Let us denote $N^*_{\n_0}$ by $N^+$. Let $R=\mtcl U_p\cap N^+$. Working in $N^+$ we may find a condition $r\in D$ such that $\mtcl U_p\cap N^+\sub\mtcl U_r$, $X_p\cap N^+\sub X_r$, and for which there is an isomorphism $$\pi:\mtbb M(p)\into\mtbb M(r)$$ which is the identity on $\mtcl U_p\cap\mtcl U_r$ and is such that the following holds for each $(\a, \n, \tau)\in \dom(h^p)$:
\begin{enumerate}
\item if $\a\geq\ssup(N^+\cap\o_2)$, then $\pi(\a)>\eta$;
\item if $\a\notin N^+$ and $\a^*=\min((N^+\cap\o_2)\setminus\a)$ exists, then $\pi(\a)>\eta_\a$;
\item if $\a\in N^+$ but $\n\notin N^+$, then $\pi(\n)>\d_N$ for each $N\in\mtcl N_p$ such that $\d_N<\d_{N^+}$;
\item $\mtcl N_p\cup\mtcl N_r$ is stratified.
\end{enumerate}
Such an $r$ can indeed be found in $N^+$ since the existence of a condition with the properties above is a true statement, as witnessed by $p$, which can be expressed over $H(\theta)$ by a sentence with parameters in $N^+$.
In order to finish the proof it suffices to show that $p$ and $r$ can be amalgamated into a condition $p'\in\mtcl P$. This condition $p'$ can be obtained as $p'=(h^{p'}, i^{p'}, \mtcl N_p\cup\mtcl N_r)$ by the following construction, very similar to that of $p^*$ from $p_0$.
For every $\a\in X_r$, let $(\a(k)\,:\,k<m_\a)$ be the strictly increasing enumeration of $$\{\ssup(N\cap\a)\,:\, N\in\mtcl N_p,\,\a\in N\}$$ and, for every $k<m_\a$, let $\d^\a_k=\d_N$ for any $N\in\mtcl N_p$ such that $\a\in N$ and $\ssup(N\cap\a)=\a(k)$. As in the construction of $p^*$ from $p_0$, each $\d^\a_k$ is well-defined. For each $\a\in X_r$ and $k<m_\a$, let $i(\a, k)\in\d^\a_k\setminus\range(i^p)$ be such that $\d_N<i(\d, k)$ for each $N\in\mtcl N_p$ with $\d_N<\d^\a_k$.
We define $h^{p'}$ and $i^{p'}$ by letting $h^{p'}$ be a function with $$\dom(h^{p'})=\dom(h^p)\cup\dom(h^r)\cup \{(\a(k), i(\a, k), \d^\a_k)\,:\,\a\in X_r,\,k<m_\a\}$$ and making the following definitions for each $(\a, \n, \tau)\in\dom(h^{p'})$ (where, given a condition $t\in\mtcl P$ and a tuple $(\a, \n, \tau)\notin \dom(h^t)$, we define $h^t(\a, \n, \tau)=\emptyset$ if $(\a, \n, \tau)\notin \dom(h^t)$, and similarly with $i^t$ in place of $h^t$):
\begin{enumerate}
\item if $(\a, \n, \tau)\in\dom(h^p)\cup\dom(h^r)$ and $\cf(\a)=\o$, then
\begin{enumerate}
\item $h^{p'}(\a, \n, \tau)=h^p(\a, \n, \tau)\cup h^r(\a, \n, \tau)$ and
\item for each $\bar\tau$ in $(\dom(h^p(\a, \n, \tau))\cup\dom(h^r(\a, \n, \tau)))\cap\Lim(\o_1)$, $i^{p'}(\a, \n, \bar\tau)=i^p(\a, \n, \bar\tau)\cup i^r(\a, \n, \bar\tau)$;
\end{enumerate}
\item if $(\a, \n, \tau)\in\dom(h^p)$ and $\cf(\a)=\o_1$,\footnote{In which case of course $\n=0$ and $\tau=\o_1$.} then
\begin{enumerate}
\item $h^{p'}(\a, \n, \tau)=h^p(\a, \n, \tau)\cup h^r(\a, \n, \tau)$,
\item $i^{p'}(\a, \n, \bar\tau)=i^p(\a, \n, \bar\tau)$ for every $\bar\tau\in\dom(h^p(\a, \n, \tau))\cap\Lim(\o_1)$, and
\item $i^{p'}(\a, \n, \bar\tau)=i^r(\a, \n, \bar\tau)$ for every $\bar\tau\in\dom(h^r(\a, \n, \tau))\cap\Lim(\o_1)$;
\end{enumerate}
\item if $(\a, \n, \tau)\in\dom(h^r)$ and $\cf(\a)=\o_1$,\footnote{Once again, in this case $\n=0$ and $\tau=\o_1$.} then
\begin{enumerate}
\item $h^{p'}(\a, \n, \tau)=h^r(\a, \n, \tau)\cup\{(\d^\a_k, \a(k))\,:\,k<m_\a\}$,
\item $i^{p'}(\a, \n, \bar\tau)=i^r(\a, \n, \bar\tau)$ for every $\bar\tau\in\dom(h^r(\a, \n, \tau))\cap\Lim(\o_1)$, and
\item $i^{p'}(\a, \n, \d^\a_k)=i(\a, k)$ for each $k<m_\a$;
\end{enumerate}
\item if $\a\in X_r$ and $k<m_\a$, then
\begin{enumerate}
\item $h^{p'}(\a(k), i(\a, k), \d^\a_k)=h^r(\a, 0, \o_1)\cup\{(\d^\a_j, \a(j))\,:\, j<k\}$,
\item $i^{p'}(\a(k), i(\a, k), \bar\tau)=i^r(\a, 0, \bar\tau)$ for every limit ordinal $\bar\tau\in \dom(h^r(\a, 0, \o_1))$, and
\item $i^{p'}(\a(k), i(\a, k), \d^\a_j)=i(\a, j)$ for each $j<k$.
\end{enumerate}
\end{enumerate}
The choice of $\eta$ and of $\eta_\a$, for $\x\in X_p\setminus N^+$ such that $\min((N^+\cap\o_2)\setminus\a)$ exists, together with the way $r$ has been fixed, immediately yields the following.
\begin{claim}\label{cl3}
For every $N\in\mtcl N_p$ and every $\a\in X_r\setminus X_p$, if $\a\in N$, then $\d_N\geq\d_{N^+}$.
\end{claim}
Using Claim \ref{cl3}, we can prove the following versions of Claims \ref{cl0} and \ref{cl2}.
\begin{claim}\label{cl4}
If $\a_0<\a_1$ are such that $\a_0$, $\a_1\in X_r$, then $\a_0(k_0)<\a_0< \a_1(k_1)$ for all $k_0<m_{\a_0}$ and $k_1<m_{\a_1}$.
\end{claim}
\begin{claim}\label{cl6}
For all $N\in\mtcl N_p$ and $\a\in X_r$, if $\a\in N$, then $h^r(\a, 0, \o_1)\in N$ and $i^r(\a, 0, \bar\tau)\in N$ for every $\bar\tau\in\dom(h^r(\a, 0, \o_1))\cap\Lim(\o_1)$.
\end{claim}
We also have the following counterpart of Claim \ref{cl1}, proved in exactly the same way.
\begin{claim}\label{cl5}
For every $N\in\mtcl N_p$, $\a\in X_r$ and $k<m_\a$, if $\a(k)$, $i(\a, k)\in N$ and $j<k$, then $\a(j)\in N$.
\end{claim}
Using Claims \ref{cl4}, \ref{cl6} and \ref{cl5}, together with Fact \ref{agreement} and the particular choice of $r$, it is not difficult to verify that $p'$ is a condition in $\mtcl P$, which finishes the proof of the lemma since then $p'$ of course extends both $p$ and $r$.
\end{proof}
We will need the following four density lemmas.
\begin{lemma}\label{density3} For every $\a<\o_1$ of countable cofinality and every $p\in\mtcl P$ there is a condition $p'\in\mtcl P$ extending $p$ and such that $(\a, 0, \o)\in\dom(h^{p'})$.
\end{lemma}
\begin{proof}
We simply let $p'=(h^p\cup\{((\a, 0, \o), \emptyset)\}, i^p, \mtcl N_p)$.
\end{proof}
\begin{lemma}\label{density0} For every $\a\in S^2_1$ and every $p\in\mtcl P$ there is a condition $p'\in\mtcl P$ extending $p$ and such that $\a\in X_{p'}$.
\end{lemma}
\begin{proof}
We may obviously assume $\a\notin X_p$. We may also assume that $\a\in N$ for some $N\in\mtcl N$ as the conclusion in the other case is immediate.
Let $(\a(k)\,:\,k<m_\a)$ be the strictly increasing enumeration of $$\{\ssup(N\cap\a)\,:\, N\in\mtcl N_p,\,\a\in N\}$$ and, for every $k<m_\a$, let $\d^\a_k=\d_N$ for any $N\in\mtcl N$ such that $\a\in N$ and $\ssup(N\cap\a)=\a(k)$. As usual, using Fact \ref{agreement} we have that each $\d^\a_k$ is well-defined. For each $\a\in X_r$ and $k<m_\a$, let $i(\a, k)\in\d^\a_k\setminus\range(i^p)$ be such that $\d_N<i(\d, k)$ for each $N\in\mtcl N_p$ with $\d_N<\d^\a_k$.
We can now easily verify that the following is a condition $p'\in\mtcl P$ as required: $p'=(h^{p'}, i^{p'}, \mtcl N_p)$, where $$\dom(h^{p'})=\dom(h^p)\cup \{(\a, 0, \o_1)\} \cup\{(\a(k), i(\a, k), \d^\a_k)\,:\,k<m_\a\}$$ and where for each $(\a, \n, \tau)\in\dom(h^{p'})$:
\begin{enumerate}
\item if $(\a, \n, \tau)\in\dom(h^p)$, then
\begin{enumerate}
\item $h^{p'}(\a, \n, \tau)=h^p(\a, \n, \tau)$ and
\item $i^{p'}(\a, \n, \bar\tau)=i^p(\a, \n, \bar\tau)$ for each $\bar\tau\in\dom(h^p(\a, \n, \tau))\cap\Lim(\o_1)$;
\end{enumerate}
\item $h^{p'}(\a, 0, \o_1)=\{(\d^\a_k, \a(i))\,:\,i<m_\a\}$ and $i^{p'}(\a, 0, \d^\a_k)=i(\a, k)$ for each $k<m_\a$;
\item for each $k<m_\a$,
\begin{enumerate}
\item $h^{p'}(\a(k), i(\a, k), \d^\a_k)=\{\d^\a_j\,:\,j<k\}$ and
\item $i^{p'}(\a(k), i(\a, k), \d^\a_j)=i(\a, j)$ for each $j<k$.
\end{enumerate}
\end{enumerate}
\end{proof}
Lemmas \ref{density1} and \ref{density2} are also easy.
\begin{lemma}\label{density1} For every $p\in\mtcl P$, $\a\in X_p$, and every $\n<\o_1$ there is a condition $p'\in\mtcl P$ extending $p$ and such that $\n\in \dom(h^{p'}(\a, 0, \o_1))$. \end{lemma}
\begin{lemma}\label{density2} For every $p\in\mtcl P$, $\a\in X_p$, every nonzero limit ordinal $\d\in \dom(h^p(\a, 0, \o_1))$, and every $\eta<h^p(\a, 0, \o_1)(\d)$ there is a condition $p'\in\mtcl P$ extending $p$ together with some $\n\in \dom(h^{p'}(\a, 0, \o_1))\cap\d$ such that $h^{p'}(\a, 0, \o_1)(\nu)>\eta$.
\end{lemma}
Given a $\mtcl P$-generic filter $G$, a limit ordinal $\a<\o_2$, and $\n<\o_1$, we define $C^G_{\a, \n}$ as $$\bigcup\{\range(h^p(\a, \n, \tau))\,:\, p\in G,\,(\a, \n, \tau)\in\dom(h^p)\mbox{ for some }\tau\}$$ Let also $$\mtcl C^G_\a=\{C^G_{\a, \n}\,:\,\n<\o_1,\,C^G_{\a, \n}\neq\emptyset\}$$
We immediately obtain the following corollary from Lemmas \ref{aleph15}, the density lemmas \ref{density0}--\ref{density2}, and the definition of $\mtcl P$.
\begin{corollary}\label{cor00} If $G$ is a $\mtcl P$-generic filter over $V$, then $$(\mtcl C^G_\a\,:\,\a\in\Lim(\o_2))$$ is a $\Box_{\o_1, \o_1}$-sequence.
\end{corollary}
Corollary \ref{cor00} yields the following.
\begin{corollary}\label{cor1} $\MA^{1.5}_{\al_2}(\mbox{stratified})$ implies $\Box_{\o_1, \o_1}$.\end{corollary}
At this point, the following question suggests itself.
\begin{question} Does $\MA^{1.5}_{\al_2}(\mbox{stratified})$ imply $\Box_{\o_1, \o}$? \end{question}
It is proved in \cite{Neeman} that $\MA^{1.5}_\k$, for any given $\k$, is consistent with $\lnot\Box_{\o_1, \o}$.
\begin{question}
Does $\MA^{1.5}_{\al_2}$ imply $\Box_{\o_1, \o_1}$? \end{question}
Finally, the following corollary is an immediate consequence of Corollary \ref{cor00}.
\begin{corollary}
$\ZFC$ proves that there is a poset $\mtcl P$ such that \
\begin{enumerate}
\item $\mtcl P$ is proper,
\item $\mtcl P$ has the $\al_2$-c.c., and
\item $\mtcl P$ forces weak square.
\end{enumerate}
\end{corollary}
\section{$\MM_{\aleph_2}(\aleph_2\mbox{-c.c.})$ is false}\label{s2}
Given a set $S$ of ordinals and a set $X$, let us denote by $\Unif_{S, X}$ the statement that for every sequence $(f_\a\,:\,\a\in S)$ of colourings $f_\a\sub\a\times X$ such that $\dom(f_\a)$ is a club of $\a$ there is a function $H:\bigcup S\into X$ such that for every $\a\in S$, $$\{\x\in \dom(f_\a)\,:\, f_\a(\x)=H(\x)\}$$ contains a club of $\a$.
Shelah proves the following theorem in (\cite{PIF}, Appendix, Chapter 3).
\begin{theorem}\label{Shelah-thm} (Shelah) $\Unif_{S^2_1, 2}$ is false.
\end{theorem}
We can also define a natural weakening $\Unif^{\,\text{c}}_{S, X}$ of $\Unif_{S, X}$ by restricting to sequences $(f_\a\,:\,\a\in S)$ of constant colourings (i.e., for every $\a\in S$, $f_\a$ is a constant function).\footnote{The superscript c is for `constant'.} It is immediate to see that for any $S\sub \Ord$ and any set $X$, $\Unif^{\,\text{c}}_{S, X}$ can be equivalently stated as the assertion that for every function $F:S\into X$ there is a function $H:\bigcup S\into X$ with the property that for every $\a\in S$ there is a club $C\sub \a$ of $\a$ such that $H(\x)=F(\a)$ for every $\x\in C$. We will say that $H$ \emph{uniformizes $F$ mod.\ clubs}.
The following is implicit in (\cite{PIF}, Appendix, Chapter 3).
\begin{theorem}\label{unif} (Shelah) If $S\sub S^2_1$ is stationary and $\Unif^{\,\text{c}}_{S, 2}$ holds, then $\CH$ holds as well.
\end{theorem}
\begin{proof}
$\Unif^{\,\text{c}}_{S, 2}$ clearly implies $\Unif^{\,\text{c}}_{S, \mtbb R}$: Given $F:S\into\, ^\o2$, let $F_n:S\into 2$ be defined by $F_n(\a)=(F(\a))(n)$ (for each $n<\o$). Applying $\Unif^{\,\text{c}}_{S, 2}$ to each $F_n$ we obtain functions $H_n:S\into 2$ and clubs $D^n_\a\sub\a$, for $\a\in S$ and $n<\o$, such that $H_n(\x)= F_n(\a)$ for all $\x\in D^n_\a$. But then, if we define $H:S\into\, ^\o2$ by letting $H(\x)=(H_n(\x)\,:\,n<\o)$, it follows that $H$ uniformizes $F$ mod.\ clubs as witnessed by the clubs $D_\a$, for $\a\in S$, where $D_\a=\bigcap_n D^n_\a$.
Thus, if $2^{\al_0}\geq\al_2$ and $\Unif^{\,\text{c}}_{S, 2}$ holds, then $\Unif^{\,\text{c}}_{S, \o_2}$ holds as well.
Now suppose $\Unif^{\,\text{c}}_{S, 2}$ holds and $2^{\al_0}\geq\al_2$. Letting $F$ be the identity function on $S$, we apply $\Unif^{\,\text{c}}_{S, \o_2}$ to $F$ and get a corresponding uniformizing function $H:\o_2\into \o_2$ and clubs $D_\a\sub \a$ for $\a\in S$. Since $S$ is stationary, we may find $\a\in S$ closed under $H$. But now we reach a contradiction since there is obviously no club $D\sub \a$ such that $H(\x)=F(\a)=\a$ for all $\x\in D$.\footnote{There is obviously not even any nonempty $D\sub\a$ like that.}
\end{proof}
\begin{remark}\label{rmk2}
Given a club-sequence $\vec C=(C_\a\,:\,\a\in S)$ such that $\ot(C_\a)=\cf(\a)$ for each $\a\in S$, we can define the following strengthening $\Unif^{\,\text{c}, \vec C,\, \text{cbd}}_{S, 2}$ of $\Unif^{\,\text{c}}_{S, 2}$: $\Unif^{\,\text{c}, \vec C,\, \text{cbd}}_{S, 2}$ is the statement that for every function $F:S\into 2$ there is a function $H:\ssup(S)\into 2$ such that for every $\a\in S$, $$\{\x\in C_\a\,:\, H(\x)=F(\a)\}$$ is co-bounded in $\a$.\footnote{We say that $H$ uniformizes $F$ on $\vec C$ modulo co-bounded sets.}
If $\CH$ holds and $\vec C=(C_\a\,:\,\a\in S)$ is as above, $\Unif^{\,\text{c}, \vec C,\, \text{cbd}}_{S^2_1, 2}$ can be forced by a $\s$-closed and $\al_2$-c.c.\ forcing, obtained as the direct limit of a long enough countable support iteration of $\s$-closed forcing notions with the $\al_2$-c.c.
At any given stage of the iteration, the corresponding iterand is the forcing $\mtcl Q_{\vec C, F}$ for adding a uniformizing function on $\vec C$ mod.\ co-bounded sets, for some given colouring $F:S\into 2$: A condition in $\mtcl Q_{\vec C, F}$ is a function $q=(b^q_\a\,:\,\a\in Z_q)$, for some countable $Z_q\sub S^2_1$, such that $b^q_\a<\a$ for each $\a\in Z_q$, and such that $\ssup(\dom(C_{\a'})\cap\a)<b^q_{\a'}$ for all $\a<\a'$ in $Z_q$ with $F(\a)\neq F(\a')$.
The extension relation is reversed inclusion.
\end{remark}
\begin{question}
Is the dependence on a fixed club-sequence in the consistency proof in Remark \ref{rmk2} necessary? In other words, is the following strengthening of $\Unif^{\,\text{c}, \vec C_\ast,\, \text{cbd}}_{S, 2}$, for a fixed club-sequence $\vec C_*=(C_\a,:\,\a\in S^2_1)$ with $\ot(C_\a)=\o_1$ for each $\a$, consistent? Suppose $\vec C=(C_\a\,:\,\a\in S)$ is a club-sequence such that $\ot(C_\a)=\o_1$ for each $\a\in S$. Then for every function $F:S\into 2$ there is a function $H:\ssup(S)\into 2$ such that for every $\a\in S$, $$\{\x\in C_\a\,:\, H(\x)=F(\a)\}$$ is co-bounded in $\a$.
\end{question}
\begin{remark}
The statement that $\Unif^{\,\text{c}}_{S, 2}$ holds for every stationary $S\sub S^2_1$ is not equivalent to $\CH$ as, for example, the assumption that $\diamondsuit(S)$ holds for every stationary $S\sub S^2_1$ implies $\lnot\Unif^{\,\text{c}}_{S, 2}$ for every such $S$: Suppose $(A_\a\,:\,\a\in S)$ is a $\diamondsuit$-sequence and let $F:S\into 2$ be such that for every $\a\in S$, $F(\a)=1-i$ if $A_\a$ codes a function $H_\a:\a\into 2$ and there are club-many $\x\in \a$ such that $H_\a(\x)=i$. It is easy to see that no function $H:\o_2\into 2$ can uniformize $F$ mod.\ clubs.
\end{remark}
Given a class $\mtcl K$ of countable models, let us say that a proper forcing $\mtbb P$ is \emph{proper with respect to $\mtcl K$} in case for every cardinal $\theta$ such that $\mtbb P\in H(\theta)$ there is a club $D\sub [H(\t)]^{\al_0}$ such that for every $N\in D\cap\mtcl K$ and every condition $p\in \mtbb P\cap N$ there is an extension $p^*\in\mtbb P$ of $p$ which is $(N, \mtbb P)$-generic.
Given a cardinal $\t$, a set $\mtcl S\sub [H(\t)]^{\al_0}$ is a \emph{projective stationary subset of $H(\t)$} in case for every stationary $S\sub\o_1$ and every club $D$ of $[H(\t)]^{\al_0}$ there is some $N\in \mtcl S\cap D$ such that $\d_N\in S$. The following proposition is standard.
\begin{proposition}\label{proj-stat}
Let $\mtcl K$ be a class of models such that $\mtcl K\cap [H(\t)]^{\al_0}$ is a projective stationary subset of $[H(\t)]^{\al_0}$ for every cardinal $\t>\o_1$ such that $\mtbb P\in H(\t)$. Let $\mtbb P$ be a forcing notion which is proper with respect to $\mtcl K$. Then $\mtbb P$ preserves stationary subsets of $\o_1$.
\end{proposition}
\begin{proof}
Let $\dot C$ be a $\mtbb P$-name for a club of $\o_1^V$, let $S\sub\o_1$ be stationary, and let $p\in\mtbb P$. Let $\t$ be large enough and, using the projective stationarity of $\mtcl K\cap H(\t)$, let $N\prec H(\t)$ be countable and such that $\mtbb P$, $\dot C$, $p\in N$ and $\d_N\in S$. Let $p^*$ be an $(N,\mtbb P)$-generic condition stronger than $p$. Then $p^*$ forces that $\d_N\in S$ is a limit of ordinals in $\dot C$ and therefore, since $\dot C$ is a $\mtbb P$-name for a closed set, that $\d_N\in\dot C$.
\end{proof}
Given a cardinal $\k$, $\MM_\k(\aleph_2\mbox{-c.c.})$ denotes $\FA_\k(\Gamma)$, where $\Gamma$ is the class of all posets $\mtbb P$ such that
\begin{itemize}
\item $\mtbb P$ preserves stationary subsets of $\o_1$ and
\item $\mtbb P$ has the $\al_2$-c.c.
\end{itemize}
The rest of this section is devoted to proving the following theorem.
\begin{theorem}\label{mainthm} $\MM_{\aleph_2}(\aleph_2\mbox{-c.c.})$ is false. \end{theorem}
Let us assume, towards a contradiction, that $\MM_{\aleph_2}(\aleph_2\mbox{-c.c.})$ holds. In particular $\FA_{\al_2}(^{{<}\o}2)$ holds and therefore $\CH$ fails.\footnote{In fact $2^{\al_0}\geq\al_3$.} Let $S=S^2_1$. It follows, by Theorem \ref{unif}, that there is a function $F:S\into 2$ for which there is no function $H:\o_2\into 2$ uniformizing $F$ mod.\ clubs.
Since $\MA^{1.5}_{\al_2}(\mbox{stratified})$ also holds, we may fix a $\Box_{\o_1, \o_1}$-sequence $\vec{\mtcl C}=(\mtcl C_\a\,:\,\a\in\Lim(\o_2))$ (by Theorem \ref{thm1}). Let also $\vec e=(e_\a\,:\,\a<\o_2)$ be such that $e_\a:|\a|\into\a$ is a bijection for each $\a<\o_2$.
Let $\mtcl K^{\vec e}_{\vec{\mtcl C}}$ be the class of countable models $N$ such that $N\cap \o_2=\bigcup_{\g\in C}e_\g``\d_N$ for some $C\in\mtcl C_\a$, where $\a=\ssup(N\cap\o_2)$.
The following is quite standard.
\begin{claim}\label{cl-proj-stat} For every cardinal $\t>\o_1$, $\mtcl K^{\vec e}_{\vec{\mtcl C}}\cap H(\t)$ is a projective stationary subset of $[H(\t)]^{\al_0}$.
\end{claim}
\begin{proof} Suppose $D$ is a club of $[H(\t)]^{\al_0}$ and $S\sub\o_1$ is stationary. Let $f:[\o_2]^{{<}\o}\into \o_2$ be a finitary function such that for every $X\in [\o_2]^{\al_0}$, if $f``[X]^{{<}\o}\sub X$, then $X=N\cap \o_2$ for some $N\in D$. Let $\a\in S^2_0$ be such that $\o_1<\a$ and $f``[\a]^{{<}\o}\sub\a$ and let $C\in\mtcl C_\a$. But now, since $E=\{M\cap\a\,:\, M\prec (H(\o_2); \in, \vec e, C)\}$ contains a club of $[\a]^{\al_0}$, we may pick some $X\in E$ closed under $f$ and such that $\d_X\in S$, and if $N\in D$ is such that $N\cap\o_2=X$, then $N$ will be a member of $\mtcl K^{\vec e}_{\vec{\mtcl C}}$ such that $\d_N\in S$.\footnote{The choice of $\a$ being of countable cofinality is inessential. We could have taken $\a$ of cofinality $\o_1$, considered $E=\{M\cap\a\,:\, M\prec (H(\o_2); \in, \vec e, \vec{\mtcl C})\}$, and continued the argument using the coherence of $\vec{\mtcl C}$.}
\end{proof}
We will show that there is a forcing notion $\mtcl Q$ which is proper with respect to $\mtcl K^{\vec e}_{\vec{\mtcl C}}$, has the $\al_2$-c.c., and forces the existence of a function $H:\o_2\into 2$ uniformizing $F$ mod.\ clubs. This will yield a contradiction since then $\mtcl Q$ will preserve stationary subsets of $\o_1$ by Proposition \ref{proj-stat} and Claim \ref{cl-proj-stat}, and so the existence of such a function $H$ will follow from an application of $\FA_{\al_2}(\{\mtcl Q\})$.
For each $\a\in\Lim(\o_2)$, let us fix an enumeration $(C_{\a, \n}\,:\,\n<\o_1)$ of $\mtcl C_\a$. If $\cf(\a)=\o_1$, we may of course take $(C_{\a, \n}\,:\,\n<\o_1)$ to be constant. Also, given a set $X$, we will write $\cl(X)$ to denote $X\cup\overline{X\cap\Ord}$, where $\overline{X\cap\Ord}$ denotes the closure of $X\cap\Ord$ in the order topology.\footnote{We stress that $X$ need not be a set of ordinals.}
Let us say that a family $\mtcl N$ of countable models is \emph{$\vec{\mtcl C}$-stratified} in case the following holds.
\begin{enumerate}
\item $\mtcl N\sub\mtcl K^{\vec e}_{\vec{\mtcl C}}$
\item For all $N_0$, $N_1\in\mtcl N$, if $\d_{N_0}=\d_{N_1}$ but $N_0\cap\o_2\neq N_1\cap \o_2$, then
\begin{enumerate}
\item $\a_i:=\min((N_i\cap \o_2)\setminus N_{1-i})$ exists for each $i\in 2$,
\item $\cf(\a_0)=\cf(\a_1)=\o_1$, and
\item there is no ordinal $\a$ above $\ssup(N_0\cap N_1\cap\o_2)$ such that $\a\in\cl(N_0\cap\o_2)\cap\cl(N_1\cap\o_2)$.
\end{enumerate}
\item For all $N_0$, $N_1\in\mtcl N$, if $\d_{N_0}<\d_{N_1}$, then $$\a:=\max(\cl(N_0\cap\o_2)\cap \cl(N_1\cap\o_2))$$ exists, $\a\in N_1$, and there is some $\n<\d_{N_1}$ such that $$N_0\cap\a=\bigcup_{\g\in C_{\a, \n}}e_\g``\d_{N_0}.$$
\end{enumerate}
The following simple remark will be quite useful.
\begin{remark}\label{r0}
Suppose $\mtcl N$ is a $\vec{\mtcl C}$-stratified family of models, $\bar\a<\o_2$, $N_0$, $N_1\in\mtcl N$, and $\a_0\in N_0\cap S$ and $\a_1\in N_1\cap S$ are such that $$\ssup(N_0\cap\a_0)=\ssup(N_1\cap\a_1)=\bar\a$$ Then $\d_{N_0}=\d_{N_1}$. Hence, if $\a_0\neq\a_1$, then $\a_0=\min((N_0\cap\o_2)\setminus N_1)$ and $\a_1=\min((N_1\cap\o_2)\setminus N_0)$.
\end{remark}
Let us also say that a $\vec{\mtcl C}$-stratified family $\mtcl N$ of models is \emph{compatible with $F$} in case for all $N_0$, $N_1\in\mtcl N$, if $\d_{N_0}=\d_{N_1}$, $N_0\cap \o_2\neq N_1\cap\o_2$, and $\a_i=\min((N_i\cap \o_2)\setminus N_{1-i})$ for each $i\in 2$, then $F(\a_0)=F(\a_1)$.
We define $\mtcl Q$ to be the forcing notion consisting of ordered pairs $$q=((\mtcl I^q_\a\,:\, \a\in X_q), \mtcl N_q)$$ with the following properties.
\begin{enumerate}
\item $X_q\in [S]^{{<}\o}$
\item For every $\a\in X_q$, $\mtcl I^q_\a$ is a finite collection of pairwise disjoint intervals of the form $[\gamma_0,\,\gamma_1)$ with $\g_0<\g_1<\a$.
\item For all $\a_0$, $\a_1\in X_q$, if $F(\a_0)\neq F(\a_1)$, then $\min(I)\neq \min(I')$ for all $I\in\mtcl I^q_{\a_0}$ and $I'\in \mtcl I^q_{\a_1}$.
\item $\mtcl N_q$ is a finite family of countable elementary submodels of the structure $(H(\o_2); \in, \vec e, (C_{\a, \n}\,:\, \a\in\Lim(\o_2),\, \n<\o_1))$ which is $\vec{\mtcl C}$-stratified and compatible with $F$.
\item The following are equivalent for every $\a\in X_q$ and every $\b<\a$.
\begin{enumerate}
\item $\b=\min(I)$ for some $I\in\mtcl I^q_\a$.
\item $\b=\ssup(N\cap\a)$ for some $N\in\mtcl N_q$ such that $\a\in N$.
\end{enumerate}
\end{enumerate}
Given conditions $q_0$, $q_1\in\mtcl Q$, $q_1$ extends $q_0$ iff
\begin{enumerate}
\item $X_{q_0}\sub X_{q_1}$,
\item for every $\a\in X_{q_0}$ and every $I\in\mtcl I^{q_0}_\a$ there is some (necessarily unique) $I'\in\mtcl I^{q_1}_\a$ such that $\min(I')=\min(I)$ and $\ssup(I')\geq\ssup(I)$, and
\item $\mtcl N_{q_0}\sub\mtcl N_{q_1}$
\end{enumerate}
We will use the two following density lemmas.
\begin{lemma}\label{dens0}
For every $\mtcl Q$-condition $q$ and every $\a\in S$ there is some $q^*\in\mtcl Q$ extending $q$ and such that $\a\in X_{q^*}$.
\end{lemma}
\begin{proof}
We may obviously assume that $\a\notin X_q$. Let $$\mtcl I=\{\{\ssup(N\cap\a)\}\,:\, N\in\mtcl N_q,\,\a\in N\}$$ Then $$q^*:=((\mtcl I^q_\b\,:\,\b\in X_q)\cup\{(\a, \mtcl I)\}, \mtcl N_q)$$ is a condition in $\mtcl Q$ as desired. To see this, let $\bar\a=\ssup(N\cap\a)$ for some $N\in\mtcl N_q$ with $\a\in N$ and suppose, towards a contradiction, that there is some $\a'\in X_q$ and some $N'\in\mtcl N_q$ such that $\a'\in N'$, $\ssup(N\cap\a')=\bar\a$, and $F(\a')\neq F(\a)$.\footnote{It is not difficult to check that this is the only way $q^*$ could fail to be a $\mtcl Q$-condition.} By $\vec{\mtcl C}$-stratification of $\mtcl N_q$ and Remark \ref{r0} we have that $\d_N=\d_{N'}$, $\a=\min((N\cap\o_2)\setminus N')$, and $\a'=\min((N'\cap\o_2)\setminus N)$. But then $F(\a)=F(\a')$ since $\mtcl N_q$ is compatible with $F$, which is a contradiction.
\end{proof}
\begin{lemma}\label{dens2} For all $q\in\mtcl Q$, $\a\in X_q$, and $\eta<\a$ there is some extension $q^*\in\mtcl Q$ together with some $I\in\mtcl I^{q^*}_\a$ such that $\min(I)>\eta$. \end{lemma}
\begin{proof}
Let $N$ be a sufficiently correct elementary submodel of $H(\o_2)$ containing $q$ and $\eta$. We build $q^*$ as $$q^*=((\mtcl I^{q^*}_\b\,:\,\b\in X_q), \mtcl N_q\cup\{N\}),$$ where $$\mtcl I^{q^*}_\b=\mtcl I^{q}_\b\cup\{\{\ssup(N\cap\b)\}\}$$ for each $\b\in X_q$.
It is clear that $q^*$ is condition in $\mtcl Q$ stronger than $q$. Also, $\eta<\ssup(N\cap\a)$ since $\eta\in N$.
\end{proof}
The main properness lemma is now the following.
\begin{lemma}\label{Q-proper} Let $\t$ be a cardinal such that $\mtcl Q\in H(\t)$ and let $M^0$ and $M^1$ be countable elementary submodels of $H(\t)$ of the same height such that $F$, $\vec C$, $\vec e\in M^0\cap M^1$ and $\{M^0, M^1\}$ is a $\vec{\mtcl C}$-stratified family\footnote{I.e., conditions (1) and (2) in the definition of $\vec{\mtcl C}$-stratified family hold for $M_0$ and $M_1$.} compatible with $F$. Then for every $q_0\in\mtcl Q\cap M^0$ there is an extension $q^*\in \mtcl Q$ of $q_0$ such that $q^*$ is $(M^i, \mtcl Q)$-generic for $i=0$, $1$.
\end{lemma}
\begin{proof}
Let $$\mtcl N=\{M^0\cap H(\o_2), M^1\cap H(\o_2)\}$$ and for every $\a\in X_{q_0}$ let $\r_\a=\ssup(M^0\cap\a)$.
The proof will be complete once we show that
$$q^*=((\mtcl I^{q^*}_\a\,:\,\a\in X_{q_0}), \mtcl N_{q_0}\cup\mtcl N)$$ is an $(M^i, \mtcl Q)$-generic condition for $i=0$, $1$,
where $\mtcl I^{q^*}_\a=\mtcl I^q_\a\cup \{\{\r_\a\}\}$.\footnote{Of course $\{\r_\a\}=[\r_\a,\,\r_\a+1)$.}
\begin{claim}\label{cl8}
$\mtcl N_{q_0}\cup\mtcl N$ is $\vec{\mtcl C}$-stratified.
\end{claim}
\begin{proof}
Since $\mtcl N_{q_0}$ and $\mtcl N$ are both $\vec{\mtcl C}$-stratified and $\mtcl N_{q_0}\sub \mtcl K^{\vec e}_{\vec{\mtcl C}}\cap M^0$, it suffices to show that if $N\in\mtcl N_{q_0}$, then $$\a:=\max(\cl(N\cap\o_2)\cap\cl(M^1\cap\o_2))$$ exists, $\a\in M^1$, and there is some $\n<\d_{M^1}$ such that $N\cap\a = \bigcup_{\g\in C_{\a, \n}} e_\g``\d_N$.
Suppose first that $N\cap\o_2\sub M^1$ and let $\a=\ssup(N\cap\o_2)$. Since $N\in M^0\cap\mtcl K^{\vec e}_{\vec{\mtcl C}}$, we then have that there is some $\n<\d_{M^0}=\d_{M^1}$ such that $N\cap\o_2=\bigcup_{\g\in C_{\a, \n}}e_\g``\d_N$. By $\vec{\mtcl C}$-stratification of $\mtcl N$ and since $\a\in M^0$, $\a$ cannot be $\ssup(M^0\cap M^1\cap\o_2)$, so it must be the case that $\a\in M^1$. Since $\a=\max(\cl(N\cap\o_2)\cap\cl(M^1\cap\o_2))$, we are done in this case.
Suppose now that $\b=\min((N\cap\o_2)\setminus M^1)$ exists. Since $\b\in M^0$, again by $\vec{\mtcl C}$-stratification of $\mtcl N$ it follows that $\cf(\b)=\o_1$. Then $C_{\b, 0}\in N$, from which we get that if $\a=\ssup(N\cap\b)$, then $N\cap\a=\bigcup_{\g\in C_{\b, 0}\cap\a}e_\g``\d_N$. Then, by coherence of $\vec{\mtcl C}$, we have that $C_{\b, 0}\cap\a\in \mtcl C_\a$. Again since $N\in M^0$, $ C_{\b, 0}\cap\a=C_{\a, \n}$ for some $\n<\d_{M^0}=\d_{M^1}$. Using once again the $\vec{\mtcl C}$-stratification of $\mtcl N$ it follows that $\a\in M^1$, which finishes the proof in this case since $\a= \max(\cl(N\cap\o_2)\cap\cl(M^1\cap\o_2))$.
\end{proof}
Also, since $\mtcl N_{q_0}$ and $\mtcl N$ are both compatible with $F$, so is $\mtcl N_{q_0}\cup\mtcl N$. In addition, for every $\a\in X_{q_0}$ and every $I\in\mtcl I^{q_0}_\a$, $\min(I)<\r_\a$ and $\r_\a\notin M^0$, and for all $\a<\a'$ in $X_{q_0}$, $\r_\a<\a<\r_{\a'}$. It thus easily follows that $q^*$ is a condition in $\mtcl Q$. Since $q^*$ of course extends $q_0$, it suffices to prove that $q^*$ is $(M^i, \mtcl Q)$-generic for each $i=0$, $1$. For this, suppose $D\in M^i$ is an open and dense subset of $\mtcl Q$ and $q$ is an extension of $q^*$ in $D$. We will find a condition in $D\cap M^i$ compatible with $q$.
Let $\Delta=\{\d_N\,:\,N\in\mtcl N_q\}\cap\d_{M^i}$. Let us note that, by $\vec{\mtcl C}$-stratification of $\mtcl N_q$ and $M^i\in\mtcl N_q$, $$R_q=\{N\cap \o_2\cap M^i\,:\,N\in\mtcl N_q,\,\d_N\in\Delta\}\in M^i.$$ Using this, and by a reflection argument as in the proof of Lemma \ref{aleph15}, we may find in $M^i$ a condition $r\in D$ such that
$$q':=((\mtcl I^q_\a\oplus\mtcl I^r_\a\,:\,\a\in X_q\cup X_r), \mtcl N_q\cup\mtcl N_r)\in\mtcl Q,$$ where $\mtcl I^q_\a\oplus\mtcl I^r_\a$ is defined as follows for each $\a\in X_q\cup X_r$.
\begin{enumerate}
\item If $\a\in X_q\setminus X_r$, then $\mtcl I^q_\a\oplus\mtcl I^r_\a=\mtcl I^q_\a$.
\item If $\a\in X_r\setminus X_q$, then $$\mtcl I^q_\a\oplus\mtcl I^r_\a=\mtcl I^r_\a\cup\{\{\ssup(N\cap\a)\}\,:\,N\in\mtcl N_q,\,\a\in N\}$$
\item If $\a\in X_q\cap X_r$, then $\mtcl I^q_\a\oplus\mtcl I^r_\a$ is the unique set $\mtcl I$ of pairwise disjoint intervals with $$\{\min(I)\,:\,I\in\mtcl I\}=\{\min(I)\,:\,I\in\mtcl I^q_\a\cup\mtcl I^r_\a\}$$ such that for every $I\in\mtcl I$, if $\g_0=\min(I)$, then
\begin{enumerate}
\item $\ssup(I)=\ssup(I_0)$ in case $I_0\in\mtcl I^q_\a$, $\min(I_0)=\g_0$, and there is no $J\in\mtcl I^r_\a$ such that $\min(J)=\g_0$;
\item $\ssup(I)=\ssup(I_1)$ in case $I_1\in\mtcl I^r_\a$, $\min(I_1)=\g_0$, and there is no $J\in\mtcl I^q_\a$ such that $\min(J)=\g_0$;
\item $\ssup(I)=\max\{\ssup(I_0), \ssup(I_1)\}$ in case $I_0\in\mtcl I^q_\a$, $I_1\in\mtcl I^r_\a$, $\min(I_0)=\g_0$, and $\min(I_1)=\g_0$.
\end{enumerate}
\end{enumerate}
More specifically, we find $r\in D\cap M^i$ with the following properties.
\begin{enumerate}
\item For all $\a\in X_r$, $I\in\mtcl I^r_\a$, $\a\in X_q$, and $I'\in\mtcl I^q_{\a'}$, if $\min(I)=\min(I')$, then $F(\a)=F(\a')$.
\item For every $N\in\mtcl N_r$ such that $\d_N\in\Delta$ there is some $\sub$-maximal $X\in\ R_q$ such that $X$ is an initial segment of $N\cap\o_2$. Moreover, $X$ is a proper initial segment of $N\cap\o_2$ if and only if there if there is some $N'\in\mtcl N_q$ such that $X$ is a proper initial segment of $N'\cap\o_2$, in which case, for every such $N'$, if $$\a_0=\min((N\cap\o_2)\setminus X)$$ and $$\a_1=\min((N'\cap\o_2)\setminus X),$$ then $\cf(\a_0)=\cf(\a_1)=\o_1$ and $F(\a_0)=F(\a_1)$. Also, for every $X\in R_q$ there is some $N\in\mtcl N_r$ such that $\d_N\in\Delta$ and such that $X$ is an initial segment of $N\cap\o_2$.
\end{enumerate}
We can indeed find such an $r\in M^i$, by correctness of $M^i$, since the existence of an $r$ with the above properties is a true fact, as witnessed by $q$, which can be expressed by a sentence with parameters in $M^i$. And given $r\in M^i$ as above, the amalgamation $q'$ of $r$ and $q$ described earlier is a condition in $\mtcl Q$. For this, it is enough to show the following claim, as all other clauses in the definition of $\mtcl Q$-condition are clear.
\begin{claim} $\mtcl N_q\cup\mtcl N_r$ is $\vec{\mtcl C}$-stratified and compatible with $F$.
\end{claim}
\begin{proof}
We can show, by an argument similar to the one in the proof of Claim \ref{cl8}, that for every $N\in \mtcl N_r$ and $M\in\mtcl N_q$, if $\d_{M^i}\leq \d_M$, then $\a:=\max(\cl(N\cap\o_2)\cap\cl(M\cap\o_2))$ exists, $\a\in M$, and $N\cap\a=\bigcup_{\g\in C_{\a, \n}}e_\g``\d_N$ for some $\n<\d_M$. And, by construction, $\{N\in\mtcl N_q\,:\,\d_N\in\Delta\}\cup\mtcl N_r$ is $\vec{\mtcl C}$-stratified and compatible with $F$. Putting these two facts together we get the conclusion.
\end{proof}
This finishes the proof of the lemma since $q'$ extends both $q$ and $r$.
\end{proof}
\begin{remark}
Unlike in the proof of Lemma \ref{aleph15}, our models $M^0$ and $M^1$ in the above proof need not be of the form $M^i=\bigcup_{i<\d_{M^i}}M^i_\nu$, with $(M^i_\nu)_{\nu<\d_{M^i}}$ being a continuous $\in$-chain of elementary submodels containing the relevant objects. This is of course thanks to the presence of clause (3) in the definition of $\vec{\mtcl C}$-stratified family.
\end{remark}
The following is an immediate corollary from Lemma \ref{Q-proper}.
\begin{corollary}\label{Q-proper-cor}
$\mtcl Q$ is proper with respect to $\mtcl K^{\vec e}_{\vec{\mtcl C}}$.
\end{corollary}
\begin{lemma}\label{Q-cc} $\mtcl Q$ has the $\al_2$-c.c. \end{lemma}
\begin{proof}
Suppose, towards a contradiction, that $(q_i\,;\,i<\l)$, for some cardinal $\l\geq\o_2$, is a one-to-one enumeration of a maximal antichain $A$ of $\mtcl Q$. Let $\t$ be a large enough cardinal. For every $i<\o_2$ let $M_i$ be a countable elementary submodel of $H(\t)$ belonging to $\mtcl K^{\vec e}_{\vec{\mtcl C}}$ and such that $p_i$, $F$, $\vec{\mtcl C}$, $\vec e$, $A\in M_i$.
Let $P$ be an elementary submodel of some higher $H(\chi)$ such that $|P|=\al_1$ and $\vec{\mtcl C}$, $((q_i, M_i)\,:\,i<\l)\in P$. Since all $q_i$ are distinct and $\l\geq\o_2$, we may find $i_0$ such that $q_{i_0}\notin P$. Now, working in $P$ and since $M_{i_0}\cap P\in P$ as $M_{i_0}\in\mtcl K^{\vec e}_{\vec{\mtcl C}}$, we may find $i_1\in P\cap\l$ such that $\d_{M_{i_0}}=\d_{M_{i_1}}$ and $\{M_{i_0}, M_{i_1}\}$ is $\vec{\mtcl C}$-stratified. By Lemma \ref{Q-proper} there is a condition $q^*\in\mtcl Q$ extending $q_{i_0}$ and such that $q^*$ is $(M_{i_1}, \mtcl Q)$-generic. But now, since $A\in M_{i_1}$ is a maximal antichain of $\mtcl Q$, we can find a common extension $q'$ of $q^*$ and some $q_{i_2}\in A\cap M_{i_1}$, which is a contradiction since $q_{i_2}\neq q_{i_0}$ yet $q'$ extends both $q_{i_0}$ and $q_{i_2}$.
\end{proof}
Let now $G$ be a $\mtcl Q$-generic filter. Given any $\a\in S$, let $$D^G_\a=\{\min(I)\,:\,I\in\mtcl I^q_\a,\,q\in G,\,\a\in X_q\}$$ By Lemma \ref{dens0}, $D^G_\a$ is an unbounded subset of $\a$.
\begin{lemma}\label{dens1}
For every $\a\in S$, $D^G_\a$ is closed in $\a$. \end{lemma}
\begin{proof}
Let $\d<\a$ be a limit ordinal forced by some $q\in \mtcl Q$ with $\a\in X_q$ to be a limit point of $D^{\dot G}_\a$ and suppose, towards a contradiction, that $\d\neq \min(I)$ for any $I\in\mtcl I^q_\a$. By the choice of $q$ we may assume that there is some $I\in\mtcl I^q_\a$ such that $\min(I)<\d$. Letting $I_0$ be the unique such $I$ with $\min(I)$ maximal within $\mtcl I^q_\a$ we may now extend $q$ to a condition $q'$ such that $[\min(I_0), \d+1)\in\mtcl I^{q'}_\a$. But $q'$ forces that $D^{\dot G}_\a\cap\d\sub \min(I_0)+1<\d$, which contradicts the assumption that $q$ forced $D^{\dot G}_\a$ to be cofinal in $\d$.
\end{proof}
It follows from Lemmas \ref{dens0}, \ref{dens2}, and \ref{dens1} together that if we aim to define $H:\o_2\into 2$ by letting $H(\eta)=F(\alpha)$ for any $\a\in S$ such that $\eta\in D^G_\a$ (and, say, $H(\eta)=0$ if there is no $\a$ as above), then $H$ is a well-defined function which uniformizes $F$ mod.\ clubs, as witnessed by $D^G_\a$ for $\a\in S$. This concludes the proof of Theorem \ref{mainthm}.
\begin{remark}
We point out that the inconsistency proof of the forcing axiom $\MM_{\al_2}(\al_2\mbox{-c.c.})$ we have given shows the impossibility of having $\FA_{\al_2}(\Gamma)$ for the class $\Gamma$ of posets which have the $\al_{1.5}$-c.c.\ with respect to families of models which are simultaneously $\vec{\mtcl C}$-stratified, for a fixed $\Box_{\o_1, \o_1}$-sequence $\vec{\mtcl C}$, and $F$-compatible for arbitrarily fixed choices of $F$. On the other hand, the methods of \cite{genMA} allow us to build models of the forcing axiom $\FA_{\al_2}(\Gamma_0)$, where $\Gamma_0$ is the class of partial orders with the $\al_{1.5}$-c.c.\ with respect to families which are $\vec{\mtcl C}$-stratified and $F$-compatible, for a fixed $\Box_{\o_1, \o_1}$-sequence $\vec{\mtcl C}$ and a fixed $F:S^2_1\into 2$.
\end{remark}
\section{$\MA^{1.5}_{\al_2}(\mbox{stratified})$ implies $\lnot\wCC$.}\label{s3}
The goal of this section is to prove the following theorem.
\begin{theorem}\label{wcc}
$\MA^{1.5}_{\al_2}(\mbox{stratified})$ implies $\lnot\wCC$.
\end{theorem}
We fix a sequence $\vec e=(e_\a\,:\,0<\a<\o_2)$,
where $e_\a:\o_1\into \a$ is a surjection for each $\a$.
We consider the following forcing notion $\mtcl R$. A condition in $\mtcl R$ is a triple $p=(f^p, (h^p_\a\,:\, \a\in X_p), \mtcl N_p)$, where:
\begin{enumerate}
\item $f^p\sub\o_1\times\o_1$ is a finite function.
\item $X_p\in [\o_2\setminus\{0\}]^{{<}\o}$
\item For each $\a\in X_p$,
\begin{enumerate}
\item $h^p_\a\sub\o_1\times\o_1$ is a finite function which can be extended to a continuous strictly increasing function $h:\o_1\into\o_1$, and
\item for each $\n\in \dom(h^p_\a)$ we have that $h^p_\a(\n)\in \dom(f^p)$ and $\ot(e_\a``h^p_\a(\n))<f^p(h^p_\a(\n))$.
\end{enumerate}
\item $\mtcl N_p$ is a finite stratified family of countable elementary submodels of $(H(\o_2); \in, \vec e)$.
\item The following holds for each $N\in \mtcl N_p$.
\begin{enumerate}
\item $f^p\restr \d_N\sub N$;
\item $\d_N\in\dom(f^p)$ and $f^p(\d_N)\geq\ot(N\cap\o_2)$;
\item for every $\a\in X_p\cap N$,
\begin{enumerate}
\item $h^p_\a\restr\d_N\sub N$,
\item $\d_N\in \dom(h^p_\a)$, and
\item $h^p_\a(\d_N)=\d_N$.
\end{enumerate}
\end{enumerate}
\end{enumerate}
Given $\mtcl R$-conditions $p_0$ and $p_1$, $p_1$ extends $p_0$ iff
\begin{enumerate}
\item $f^{p_0}\sub f^{p_1}$,
\item $X_{p_0}\sub X_{p_1}$, and
\item for every $\a\in X_{p_0}$, $h^{p_0}_\a\sub h^{p_1}_\a$.
.\end{enumerate}
The following density lemmas are easy.
\begin{lemma}\label{d0}
For every $p\in\mtcl R$ and every nonzero $\b<\o_2$ there is a $\mtcl R$-condition $p^*$ extending $p$ and such that $\b\in X_{p^*}$.
\end{lemma}
\begin{proof}
We may of course assume that $\b\notin X_p$. It then suffices to set $$p^*=(f^p, (h^p_\a\,:\, \a\in X_p)\cup\{(\b, \{(\d_N, \d_N)\,:\,N\in\mtcl N_p,\,\b\in N\})\}, \mtcl N_p)$$ To see that this is a condition in $\mtcl R$ it is enough to notice that if $N\in\mtcl N_p$ is such that $\b\in N$, then $\d_N\in \dom(f^p)$ and $$\ot(e_\b``\d_N)=\ot(N\cap \b)<\ot(N\cap\o_2)\leq f^p(\d_N)$$
\end{proof}
Lemmas \ref{d1}, \ref{d2} and \ref{d3} are obvious.
\begin{lemma}\label{d1} For every $p\in\mtcl R$, $\a\in X_p$ and $\n<\o_1$ there is some $p^*\in\mtcl R$ extending $p$ and such that $\n\in \dom(h^{p^*}_\a)$.
\end{lemma}
\begin{lemma}\label{d2} For every $p\in\mtcl R$, $\a\in X_p$, every nonzero limit ordinal $\d\in \dom(h^p_\a)$, and every $\eta<h^p_\a(\d)$ there is a condition $p^*\in\mtcl R$ extending $p$ together with some $\n\in \dom(h^{p^*}_\a)\cap\d$ such that $h^{p^*}_\a(\nu)>\eta$.
\end{lemma}
\begin{lemma}\label{d3} For every $p\in\mtcl R$ and every $\n<\o_1$ there is a condition $p^*\in\mtcl R$ extending $p$ and such that $\n\in \dom(f^{p^*})$. \end{lemma}
It follows from Lemmas \ref{d0}--\ref{d3} together that if $G$ is $\mtcl R$-generic and we set $$f^G=\bigcup\{f^p\,:\, p\in G\}$$ and $$C^G_\a=\bigcup\{\range(h^p_\a)\,:\,p\in G,\,\a\in X_p\}$$ for each nonzero $\a<\o_2$, then $f^G:\o_1^V\into \o_1^V$ is a function, each $C^G_\a$ is a club of $\o_1^V$, and $\ot(e_\a``\n)<f^G(\n)$ for each $\a$ and $\n\in C^G_\a$. Hence, if we can show that $\mtcl R$ has the $\al_{1.5}$-c.c. with respect to finite stratified families of models, an application of $\MA^{1.5}_{\al_2}(\mbox{stratified})$ to $\mtcl R$ will show that $\MA^{1.5}_{\al_2}(\mbox{stratified})$ implies $\lnot\wCC$.
\begin{lemma}\label{strat-wcc}
$\mtcl R$ has the $\al_{1.5}$-c.c.\ with respect to finite stratified families of models.
\end{lemma}
\begin{proof}
Let $\t$ be a large enough cardinal, let $\mtcl N^*$ be a finite stratified family of countable elementary submodels of $H(\t)$ containing $\vec e$ and $\vec f$, and let $p_0\in\mtcl R\cap N^*_0$, where $N^*_0$ is of minimum height within $\mtcl N^*$. We will prove that there is a condition $p^*\in\mtcl R$ stronger than $p_0$ such that $p^*$ is $(N^*, \mtcl R)$-generic for each $N^*\in\mtcl N^*$.
Let $\mtcl N=\{N^*\cap H(\o_2)\,:\,N^*\in\mtcl N^*\}$ and for every $\d\in\{\d_N\,:\,N\in\mtcl N\}$ let $$\m(\d)=\max\{\ot(N\cap\o_2)\,:\, N\in\mtcl N,\,\d_N=\d\}$$ Let $$p^*=(f^{p_0}\cup\{(\d_N, \m(\d_N))\,:\,N\in\mtcl N\}, (h^{p^*}_\a\,:\,\a\in X_{p_0}), \mtcl N_{p_0}\cup\mtcl N),$$ where $$h^{p^*}_\a=h^{p_0}_\a\cup\{(\d_N, \d_N)\,:\,N\in\mtcl N,\,\a\in N\}$$ for each $\a\in X_{p_0}$. It is easy to check that $p^*$ is a condition in $\mtcl R$ (for this it is enough to notice that $\m(\d_{N_0})<\d_{N_1}$ holds for all $N_0$, $N_1\in\mtcl N$ with $\d_{N_0}<\d_{N_1}$, which follows from the stratification of $\mtcl N$), and it of course extends $p_0$ by construction. Hence, it will be enough to show that $p^*$ is $(N^*, \mtcl R)$-generic for every $N^*\in\mtcl N^*$. Let $D\in N^*$ be a dense and open subset of $\mtcl R$ and let $p$ be an extension of $p^*$ in $D$. We will show that there is a condition in $D\cap N^*$ compatible with $p$.
As in the proof of Lemma \ref{aleph15}, we may assume that $N^*=\bigcup_{\n<\d_{N^*}}N^*_\n$, where $(N^*_\n)_{\n<\d_{N^*}}$ is a $\sub$-continuous $\in$-chain of models.
By moving to a suitable $N^*_{\n_0}$ and arguing there as in the proof of Lemma \ref{aleph15} using the stratification of $\mtcl N_p$, we may find a condition $r\in D\cap N^*_{\n_0}$ such that
\begin{enumerate}
\item for every $\a\in X_p\cap X_r$, $h^p_\a\cup h^r_\a$ can be extended to a strictly increasing and continuous function $h:\o_1\into \o_1$,
\item for every $\a\in X_r\cap X_q$ and every $N\in\mtcl N_p$ such that $\d_N<\d_{N^*}$, $\a\notin N$, and
\item $\mtcl N_p\cup\mtcl N_r$ is stratified.
\end{enumerate}
Let now $$p'=(f^r\cup (f^p\restriction (\o_1\setminus\d_{N^*})), (h^{p'}_\a\,:\,\a\in X_p\cup X_r), \mtcl N_p\cup\mtcl N_r),$$ where
\begin{enumerate}
\item for every $\a\in X_p\cap X_r$, $h^{p'}_\a=h^p_\a\cup h^r_\a$,
\item for every $\a\in X_p\setminus X_r$, $h^{p'}_\a=h^p_\a$, and
\item for every $\a\in X_r\setminus X_p$, $h^{p'}_\a=h^r_\a\cup\{(\d_N, \d_N)\,:\, N\in\mtcl N_p,\,\a\in N\}$.
\end{enumerate}
Then $p'$ is a condition in $\mtcl R$, which finishes the proof of the lemma since $p'$ is of course stronger than both $p$ and $r$.
\end{proof}
The above lemma concludes the proof of Theorem \ref{wcc}.
The following corollary is an immediate by-product of the proof of Theorem \ref{wcc}.
\begin{corollary}
$\ZFC$ proves that there is a poset $\mtcl P$ such that
\begin{enumerate}
\item $\mtcl P$ is proper,
\item $\mtcl P$ has the $\al_2$-c.c., and
\item $\mtcl P$ forces $\lnot\wCC$.\end{enumerate}
\end{corollary}
\section{Another proof of the inconsistency of $\MM_{\al_2}(\al_2\mbox{-c.c.})$}\label{s4}
In this final section of the paper we will give another proof of Theorem \ref{mainthm} in Section \ref{s2}. Our argument essentially follows an argument, due to Shelah, showing that stationary preserving forcing cannot be iterated, and in fact that there is a forcing iteration of length $\o$ each of whose iterands is forced to preserve stationary subsets of $\o_1$ but such that any limit of it collapses $\o_1$.
Let us assume that $\MM_{\al_2}(\al_2\mbox{-c.c.})$ holds. Hence, $\MA_{1.5}(\mbox{stratified})$ holds as well and so, by Theorem \ref{wcc}, there is a function $f:\o_1\into \o_1$ such that $\{\n<\o_1\,:\,g(\n)<f(\n)\}$ contains a club for every nonzero $\a<\o_2$ and every canonical function $g$ for $\a$. We will build a sequence $(f_n)_{n<\o}$ of functions from $\o_1$ to $\o_1$, together with clubs $C_n$ of $\o_1$, such that for every $n$ and every $\n\in C_n$, $f_{n+1}(\n)<f_n(\n)$. This of course will yield a contradiction since then, if $\n\in \bigcap_n C_n$, then $f_{n+1}(\n)<f_n(\n)$ for all $n$, which is impossible.
We will make sure that the construction can keep going by arranging, for every $n<\o$ and every nonzero $\a<\o_2$, that $f_n$ dominates every canonical function for $\a$ on a club. We start our construction by letting $f_0=f$.
Given $n<\o$ and assuming $f_n$ has been constructed, we will find $f_{n+1}$ by an application of $\MM_{\al_2}(\al_2\mbox{-c.c.})$ to the following slight variant $\mtcl R_{f_n}$ of the poset $\mtcl R$ in the proof of Theorem \ref{wcc}.
Let $\mtcl K^{f_n}$ be the collection of countable $N\elsub (H(\o_2); \in, \vec e, f_n)$ such that $\ot(N\cap\o_2)<f_n(\d_N)$.
A condition in $\mtcl R_{f_n}$ is a tuple $p=(f^p, d_p, (h^p_\a\,:\,\a\in X_p), \mtcl N_p)$ with the following properties.
\begin{enumerate}
\item $f^p\sub\o_1\times\o_1$ is a finite function.
\item $d_p\sub\o_1\times\o_1$ is a finite function which can be extended to a continuous strictly increasing function $d:\o_1\into\o_1$.
\item For every $\n\in\dom(d_p)$, $f^p(d_p(\n))<f_n(d_p(\n))$.
\item $X_p\in [\o_2\setminus\{0\}]^{{<}\o}$
\item For each $\a\in X_p$,
\begin{enumerate}
\item $h^p_\a\sub\o_1\times\o_1$ is a finite function which can be extended to a continuous strictly increasing function $h:\o_1\into\o_1$, and
\item for each $\n\in \dom(h^p_\a)$ we have that $h^p_\a(\n)\in \dom(f^p)$ and $\ot(e_\a``h^p_\a(\n))<f^p(h^p_\a(\n))$.
\end{enumerate}
\item $\mtcl N_p$ is a finite stratified family of members of $\mtcl K^{f_n}$.
\item The following holds for each $N\in \mtcl N_p$.
\begin{enumerate}
\item $f^p\restr \d_N\sub N$;
\item $\d_N\in\dom(f^p)$ and $f^p(\d_N)\geq\ot(N\cap\o_2)$;
\item $d_p\restr\d_N\sub N$, $\d_N\in \dom(d_p)$, and $d_p(\d_N)=\d_N$.
\item for every $\a\in X_p\cap N$,
\begin{enumerate}
\item $h^p_\a\restr\d_N\sub N$,
\item $\d_N\in \dom(h^p_\a)$, and
\item $h^p_\a(\d_N)=\d_N$.
\end{enumerate}
\end{enumerate}
\end{enumerate}
Given $\mtcl R_{f_n}$-conditions $p_0$ and $p_1$, $p_1$ extends $p_0$ iff
\begin{enumerate}
\item $f^{p_0}\sub f^{p_1}$,
\item $d_{p_0}\sub d_{p_1}$,
\item $X_{p_0}\sub X_{p_1}$, and
\item for every $\a\in X_{p_0}$, $h^{p_0}_\a\sub h^{p_1}_\a$.
\end{enumerate}
We now have the following density lemmas. Lemma \ref{d0'} is proved by the same argument as in the proof of Lemma \ref{d0} using the fact that all models of $\mtcl N_p$ are in $\mtcl K^{f_n}$, and Lemmas \ref{d1'}--\ref{d4'} are straightforward.
\begin{lemma}\label{d0'}
For every $p\in\mtcl R_{f_n}$ and every nonzero $\b<\o_2$ there is a $\mtcl R_{f_n}$-condition $p^*$ extending $p$ and such that $\b\in X_{p^*}$.
\end{lemma}
\begin{lemma}\label{d1'} For every $p\in\mtcl R_{f_n}$, $\a\in X_p$ and $\n<\o_1$ there is some $p^*\in\mtcl R_{f_n}$ extending $p$ and such that $\n\in \dom(h^{p^*}_\a)$.
\end{lemma}
\begin{lemma}\label{d2'} For every $p\in\mtcl R_{f_n}$, $\a\in X_p$, every nonzero limit ordinal $\d\in \dom(h^p_\a)$, and every $\eta<h^p_\a(\d)$ there is a condition $p^*\in\mtcl R_{f_n}$ extending $p$ together with some $\n\in \dom(h^{p^*}_\a)\cap\d$ such that $h^{p^*}_\a(\nu)>\eta$.
\end{lemma}
\begin{lemma}\label{d3'} For every $p\in\mtcl R_{f_n}$ and every $\n<\o_1$ there is a condition $p^*\in\mtcl R_{f_n}$ extending $p$ and such that $\n\in \dom(f^{p^*})$. \end{lemma}
\begin{lemma}\label{d4'} For every $p\in\mtcl R_{f_n}$ and every $\n<\o_1$ there is a condition $p^*\in\mtcl R_{f_n}$ extending $p$ and such that $\n\in \dom(d_{p^*})$. \end{lemma}
It follows from the above density lemmas that if $G$ is $\mtcl R_{f_n}$-generic and we let $$D^G=\bigcup\{\range(d_p)\,:\, p\in G\},$$ $$f^G=\bigcup\{f^p\,:\,p\in G\},$$ and $$C^G_\a=\bigcup\{\range(h^p_\a)\,:\,p\in G,\,\a\in X_p\}$$ for each nonzero $\a<\o_2$, then $D^G$ is a club of $\o_1^V$, $f^G:\o_1^V\into \o_1^V$ is a function, each $C^G_\a$ is a club of $\o_1^V$, $\ot(e_\a``\n)<f^G(\n)$ for each $\a$ and $\n\in C^G_\a$, and $f^G(\n)<f_n(\n)$ for each $\n\in D^G$. It follows that an application of $\MM_{\al_2}(\al_2\mbox{-c.c.})$ to $\mtcl R_{f_n}$ will provide us with $f_{n+1}$. Hence, we just need to prove that $\mtcl R_{f_n}$ preserves stationary subsets of $\o_1$ and has the $\al_2$-c.c. This we will prove by means of the following version of Lemma \ref{Q-proper} in Section \ref{s2}.
\begin{lemma}\label{R-proper}
Let $\t$ be a cardinal such that $\mtcl R_{f_n}\in H(\t)$ and let $M^0$, $M^1\prec H(\t)$ be countable models of the same height and such that $\mtcl R_{f_n}\in M^0\cap M^1$ and $M^0\cap H(\o_2)$, $M^1\cap H(\o_2)\in \mtcl K^{f_n}$. Then for every $p_0\in\mtcl R_{f_n}\cap M^0$ there is an extension $p^*\in \mtcl Q$ of $p_0$ such that $p^*$ is $(M^i, \mtcl R_{f_n})$-generic for $i=0$, $1$.
\end{lemma}
\begin{proof}
Let $\m=\max\{\ot(M^0\cap\o_2), \ot(M^1\cap\o_2)\}$ and $$p^*=(f^{p^*}, d_{p^*}, (h^{p^*}_\a\,:\,\a\in X_{p_0}), \mtcl N_{p_0}\cup\{M^0\cap H(\o_2), M^1\cap H(\o_2)\}),$$ where
\begin{enumerate}
\item $f^{p^*}=f^{p_0}\cup\{(\d_{M^0}, \m)\}$,
\item $d_{p^*}=d_{p_0}\cup\{(\d_{M^0}, \d_{M^0})\}$, and
\item $h^{p^*}_\a=h^{p_0}_\a\cup\{(\d_{M^0}, \d_{M^0})\}$ for each $\a\in X_{p_0}$.
\end{enumerate}
Using the fact that $M^0\cap H(\o_2)$, $M^1\cap H(\o_2)\in\mtcl K^{f_n}$, it is immediate to check that $p^*\in\mtcl R_{f_n}$, and it of course extends $p_0$. It will thus suffice to show that $p^*$ is $(M^i, \mtcl R_{f_n})$-generic for $i=0$, $1$. For this, let $D\in M^i$ be a dense and open subset of $\mtcl R_{f_n}$ and let $p\in D$ be an extension of $p^*$. We will show that there is a condition $r\in D\cap M^i$ compatible with $p$.
We can find $r$ by arguing, in $M^i$, in the same way as in the reflection argument in the proof of Lemma \ref{strat-wcc}. More specifically, and exactly as in that proof, we may assume that $M^i=\bigcup_{\n<\d_{M^i}}M^i_\n$, where $(M^i_\n)_{\n<\d_{M^i}}$ is a $\sub$-continuous $\in$-chain of models. Then, by moving to a suitable $M^i_{\n_0}$ and arguing there as in the proof of Lemma \ref{aleph15} using the stratification of $\mtcl N_p$, we may find a condition $r\in D\cap M^i_{\n_0}$ such that
\begin{enumerate}
\item for every $\a\in X_p\cap X_r$, $h^p_\a\cup h^r_\a$ can be extended to a strictly increasing and continuous function $h:\o_1\into \o_1$,
\item for every $\a\in X_r\cap X_q$ and every $N\in\mtcl N_p$ such that $\d_N<\d_{M^i}$, $\a\notin N$, and
\item $\mtcl N_p\cup\mtcl N_r$ is stratified.
\end{enumerate}
Let us define $h^{p'}_\a$, for $\a\in X_p\cup X_r$, as follows:
\begin{enumerate}
\item for every $\a\in X_p\cap X_r$, $h^{p'}_\a=h^p_\a\cup h^r_\a$;
\item for every $\a\in X_p\setminus X_r$, $h^{p'}_\a=h^p_\a$;
\item for every $\a\in X_r\setminus X_p$, $h^{p'}_\a=h^r_\a\cup\{(\d_N, \d_N)\,:\, N\in\mtcl N_p,\,\a\in N\}$.
\end{enumerate}
We then have that $$p'=(f^r\cup (f^p\restriction (\o_1\setminus\d_{M^i})), d_p\cup d_r, (h^{p'}_\a\,:\,\a\in X_p\cup X_r), \mtcl N_p\cup\mtcl N_r)$$ is a common extension in $\mtcl R_{f_n}$ of $p$ and $r$, which finishes the proof of the lemma.
\end{proof}
\begin{lemma}\label{proj-stat-wcc} $\mtcl K^{f_n}$ is projective stationary. \end{lemma}
\begin{proof}
Given a cardinal $\t\geq\o_2$, a function $F:[\o_2]^{{<}\o}\into \o_2$, and a stationary set $S\sub\o_1$, it is enough to show that there is a countable $X\sub\o_2$ closed under $F$ such that $\d:=X\cap\o_1\in S$ and such that $f_n(\d)>\ot(X)$.
In order to find such an $X$ we first pick $\a\in S^2_0$ above $\o_1$ such that $F``[\a]^{{<}\o}\sub \a$. We then let $N$ be a countable elementary submodel of some larger $H(\chi)$ containing $F$, $\a$, $f_n$, and $\vec f$ and such that $\d_N\in S$, and let $X=N\cap\a$. Then $F``[X]^{{<}\o}\sub X$, $X\cap\o_1=\d_N\in S$, and $f_n(\d_N)\geq \ot(N\cap \o_2)>\ot(N\cap \a)=\ot(X)$, where the first inequality follows from the fact that for every nonzero $\b\in N\cap \o_2$ there is a club $C\in N$ of $\o_1$ such that $\ot(e_\b``\n)<f_n(\n)$ for all $\n\in C$, which implies that $\ot(N\cap\b)=\ot(e_\b``\d)<f_n(\d)$ as $\d\in C$. \end{proof}
We now have the following corollary from Lemmas \ref{R-proper} and \ref{proj-stat-wcc}.
\begin{corollary}\label{cor000}
$\mtcl R_{f_n}$ is proper with respect to $\mtcl K^{f_n}$ and therefore it preserves stationary subsets of $\o_1$.\end{corollary}
We also have the following corollary from Lemma \ref{R-proper}.
\begin{lemma}\label{al2wcc}
$\mtcl R_{f_n}$ has the $\al_2$-c.c.
\end{lemma}
\begin{proof}
This is similar to the proof of Lemma \ref{Q-cc}. Suppose $(p_i\,;\,i<\l)$, for some cardinal $\l\geq\o_2$, is a one-to-one enumeration of a maximal antichain $A$ of $\mtcl R_{f_n}$. Let $\t$ be a large enough cardinal and for every $i<\o_2$ let $M_i$ be a countable elementary submodel of $H(\t)$ such that $p_i$, $A\in M_i$ and such that $M_i\cap H(\o_2)\in \mtcl K^{f_n}$.
Let $P$ be an elementary submodel of some higher $H(\chi)$ such that $|P|=\al_1$ and $((p_i, M_i)\,:\,i<\l)\in P$. We may then find $i_0$ such that $p_{i_0}\notin P$. Working in $P$, we may find $i_1\in P\cap\l$ such that $\d_{M_{i_0}}=\d_{M_{i_1}}$. By Lemma \ref{R-proper} there is a condition $p^*\in \mtcl R_{f_n}$ extending $p_{i_0}$ and such that $p^*$ is $(M_{i_1}, \mtcl R_{f_n})$-generic. Since $A\in M_{i_1}$ is a maximal antichain of $\mtcl R_{f_n}$, we can find a common extension $p'$ of $p^*$ and some $p_{i_2}\in A\cap M_{i_1}$, which is a contradiction since $p_{i_2}\neq p_{i_0}$ yet $p'$ extends both $p_{i_0}$ and $p_{i_2}$.
\end{proof}
Lemmas \ref{cor000} and \ref{al2wcc} complete our second proof of Theorem \ref{mainthm}.
|
\section{Introduction and related works}
Distributed data fusion (DDF) is a process in which a group of agents senses their immediate environment, communicates with other agents, and aims at inferring knowledge about a specific process collectively.
Some applications include cooperative robots mapping a room \cite{dardari2015indoor}, multisensor data fusion in internet of things\cite{ding2019survey}, navigation systems \cite{dunik2020state}, medical diagnosis, or pattern recognition to name a few\cite{hall2004mathematical}.
There are several data fusion architectures which have been proposed. Arguably, the joint
directors of laboratories (JDL) Data Fusion Group is the most widely-used taxonomy method for data fusion-related functions \cite{meng2020survey, White1991DataFL}. It defines data fusion as a ``multilevel, multifaceted process handling the automatic detection, association, correlation, estimation, and combination of data and information from several sources.'' Two recent fusion frameworks were proposed in \cite{kokar2004formalizing,castanedo2013review}, both based on category theory and are claimed to be sufficiently general to capture all kinds of fusion techniques \cite{hall1997introduction}, including raw data fusion, feature fusion, and decision fusion
Distributed data fusion research can be categorized into Bayesian or consensus-based approaches \cite{7515322}. Bayesian methods focus on preserving the full distribution of the unknowns given the data, called posterior, over the estimated process at each agent, so that sensor data can be easily and recursively merged with prior knowledge and does not need to be stored. On the other hand, consensus algorithms are designed in such a way that agents can continue to exchange information until they agree on certain parameters or quantities of interest.
In this context, many data fusion methods focus on Bayesian methodologies, such as naive Bayesian fusion, federated Kalman filtering \cite{hashemipour1988decentralized,mahmoud2013distributed}, and other different fusion methods that may leverage data-driven models \cite{tresp2000bayesian,bailey2012conservative, deisenroth2015distributed}. Recently, there has been increasing interest in the consensus area, particularly in the machine learning community, where a plethora of distributed learning or federated learning methods have been proposed \cite{mcmahan2021advances,Ji2021EmergingTI,wu2021personalized}.
In this paper, we focus on a Bayesian perspective to the data fusion problem.
There are many data fusion challenges \cite{meng2020survey,khaleghi2013multisensor}, being the data correlation problem one of the most prominent cases. In general, the performance of DDF solutions cannot outperform a centralized scheme, which is typically considered as a benchmark. Although on the other side, DDF is intrinsically more adaptive and resilient to failures. When dealing with distributed agents, some of the challenges involve both dealing with observations impacted by the same process noise \cite{bar1981track} and also non-independence of local estimates due to double counting of data \cite{bakr2017distributed,chang2010analytical,bailey2012conservative,7515322}, which essentially means that in a distributed architecture local agent estimates may be correlated. To maintain optimality and consistency, a distributed fusion method should account for such cross-correlation issues. Double counting occurs when data is utilized numerous times without the user's knowledge. This might be due to recirculation of data through cyclic channels or the same data traveling through several paths from another agent to the fusion node.
To avoid data double counting, two popular solutions are typically adopted \cite{li2019second}: arithmetic average (AA) and geometric average (GA). These two methods can be employed by most fusion methods that \textit{average} the parameters of interest across many agents. Sometimes referred to as Chernoff fusion \cite{chang2010analytical}, also known as covariance intersection \cite{julier2017general,bakr2017distributed} under Gaussian assumptions. Related to the latter are the works on \textit{mixture of experts} \cite{nguyen2016universal}, robust \textit{product of experts} \cite{deisenroth2015distributed}, and the weighted fusion of Kalman filters \cite{xing2016multisensor,alimadadi2020object}, as well as the linear fusion of partial estimators \cite{victor201803}.
DDF is a field that connects to other disciplines, for that reason related works can be found also under the umbrella of model fusion, estimator fusion, distributed estimation, and distributed learning.
Particularly, in the area of machine learning, the interest of distributed learning is growing rapidly under the so-called federated learning (FL) paradigm. Most of the research in FL focuses on frequentist approaches to inference, where data-driven neural network (NN) parameters are aggregated \cite{mcmahan2017communication,wu2021personalized,pham2021fusion} based on the AA method. Besides the fusion methodology, privacy and communication efficiency are also a key part of FL research \cite{mcmahan2021advances}.
Recently, some works explored Bayesian FL schemes \cite{achituve2021personalized, lalitha2019peer,chen2020fedbe, liu2021bayesian,Junha2022} where the model fusion problem is addressed from a Bayesian perspective, an approach that is thoroughly reviewed in \cite{Koliander22}. Also within the Bayesian approach, Bayesian committee machine (BCM) \cite{tresp2000bayesian,deisenroth2015distributed,liu2018generalized,liu2020gaussian} is another framework that relates to DDF, in this case focused on Gaussian process models.
In this paper, we investigate the use of \textit{a priori} data in distributed fusion problems. Particularly, we are interested in schemes where a parameter $\boldsymbol{\theta}$ is inferred locally by $M$ agents in a Bayesian setting. In this scheme, $\boldsymbol{\theta}$ takes values in an arbitrary $d$-dimensional space $\Theta$. The quantity of interest is therefore treated as a random variable for which an a priori distribution is available, $p(\boldsymbol{\theta})$. This prior is shared by all participating agents and updated using local data $\mathcal{D}_m$ to produce the a posteriori distribution of the unknown given the local information, $p(\boldsymbol{\theta} | \mathcal{D}_m)$. In this context, this work analyzes various Bayesian fusion methodologies that employ the $M$ local posteriors and, in particular, connects those to the optimal fusion rule whereby the full posterior would be computed at a central node collecting all data, $p(\boldsymbol{\theta} | \mathcal{D}_1,\mathcal{D}_2,\dots,\mathcal{D}_M)$. One of the main challenges of Bayesian fusion rules is the multiple use of prior data, which is generally not properly dealt with. We discuss corrections to avoid systematic bias, as well as provide a discussion on how the number of clients affects the performance of the fusion rule when $p(\boldsymbol{\theta})$ is used in excess without corrections. This work paves the way for further methodological advances in distributed Bayesian fusion, equipped with a solid theoretical justification.
The remainder of the paper is organized as follows. Section \ref{sec:fusion} introduces the optimal (global) Bayesian solution as well as different fusion rules using local updates. Section \ref{sec:GaussianFusion} provides insights for the important case where inferring $\boldsymbol{\theta}$ can be seen as a regression problem where the relevant distributions are approximated by a Gaussian distributions. Experimental results are provided and discussed in Section \ref{sec:experiments}, validating the theoretical results on earlier sections. The paper is concluded by Section \ref{sec:conclusions} with final remarks.
\section{Bayesian Data Fusion schemes}\label{sec:fusion}
\begin{figure}
\centerline{\includegraphics[width=0.5\textwidth]{figs/bayeisan_fusion_plot.jpg}
}
\caption{Distributed framework of $M$ agents sharing a priori information on a quantity, $p(\boldsymbol{\theta})$, and returning local posterior updates, $p(\boldsymbol{\theta}|\mathcal{D}_m)$.}
\label{fig:distributed}
\end{figure}
This paper considers a framework (see Fig. \ref{fig:distributed}) where observations related to an unknown quantity of interest $\boldsymbol{\theta}\in\Theta^{d}$ are obtained by $M$ distributed agents, such that the $m$-th agent has only access to its own partial data $\mathcal{D}_m$. The data is assumed to be independent and identically distributed (\textit{i.i.d.}) across agents. In addition, a Bayesian approach to the inference of $\boldsymbol{\theta}$ is considered, therefore a prior distribution is assumed $p(\boldsymbol{\theta})$ before data is observed. The main goal of Bayesian data fusion is to evaluate the posterior of the quantity given all the data $p(\boldsymbol{\theta}|\mathcal{D})$, however this might be unpractical due to the need to transmit local data to a central node in charge of processing it. Alternatively, distributed Bayesian data fusion provides a framework to combine the local inference results in a manner that approximates the optimal Bayesian solution.
This section formulates the optimal Bayesian fusion strategy, as it would be implemented at a central node using all available data $\mathcal{D} = \{\mathcal{D}_1,\mathcal{D}_2,\dots,\mathcal{D}_M \}$ to compute $p(\boldsymbol{\theta}|\mathcal{D})$. Then, we establish connections to different distributed Bayesian data fusion strategies where we observe that, unless properly accounted for, the a priori information may be used multiple times. This situation can potentially cause inconsistent approximations of the optimal Bayesian solution. We propose to quantify this deviation with the Kullback-Leibler (KL) divergence between the optimal posterior and the resulting approximation.
\noindent\textbf{Optimal Bayesian data fusion.}
Under the assumption that datasets $\mathcal{D} = \{\mathcal{D}_1,\mathcal{D}_2,\dots,\mathcal{D}_M \}$ are \textit{i.i.d.} given $\boldsymbol{\theta}$, it is possible to express the optimal Bayesian fusion rule in terms of the local likelihoods and the globally used prior distribution. We refer to this solution as the optimal Bayesian fusion rule under \emph{conditionally independent likelihoods} (CIL). The posterior $p(\boldsymbol{\theta}|\mathcal{D})$ is then
\begin{align}\label{eq:CILa}
p(\boldsymbol{\theta}|\mathcal{D})
= & \frac{p(\mathcal{D}_1,\mathcal{D}_2,\dots,\mathcal{D}_M|\boldsymbol{\theta})p(\boldsymbol{\theta})}{p(\mathcal{D}_1,\mathcal{D}_2,\dots,\mathcal{D}_M)}
= \frac{p(\boldsymbol{\theta}) \prod_{m=1}^{M}p(\mathcal{D}_m|\boldsymbol{\theta})} {p(\mathcal{D}_1,\mathcal{D}_2,\dots,\mathcal{D}_M)}\;,
\end{align}
\noindent which can be further manipulated using Bayes' rule as
\begin{align}\label{eq:CIL}
p(\boldsymbol{\theta}|\mathcal{D})
= Z \frac{\prod_{m=1}^{M} p(\boldsymbol{\theta}|\mathcal{D}_m)}{p^{M-1}(\boldsymbol{\theta})} \;,
\end{align}
where
\begin{equation}\label{eq:normalization_Z}
Z = \frac{\prod_{m=1}^{M} p(\mathcal{D}_m)}{p(\mathcal{D}_1,\mathcal{D}_2,\dots,\mathcal{D}_M)} \;,
\end{equation}
is the normalization constant and $p(\boldsymbol{\theta}|\mathcal{D}_m)$ represents the local posterior for the $m$-th agent.
This approach is optimal under the \textit{i.i.d.} data assumption, which enables the likelihood factorization in \eqref{eq:CIL}. It presents a Bayesian approach for data fusion when the same prior is used by the $M$ users. This solution enables for a distributed implementation, where prior is sent to each agent who returns their local posteriors in exchange.
It is easy to see that the optimal Bayesian fusion rule for the case where each agent uses different local prior information $p_m(\boldsymbol{\theta})$ is
\begin{equation}\label{eq:CIL_diffprior}
p(\boldsymbol{\theta}|\mathcal{D})
= Z p(\boldsymbol{\theta}) \prod_{m=1}^{M} \frac{ p(\boldsymbol{\theta}|\mathcal{D}_m)}{p_m(\boldsymbol{\theta})} \;,
\end{equation}
\noindent where the global prior $p(\boldsymbol{\theta})$ might differ from the local prior distributions.
\noindent\textbf{Approximate Bayesian data fusion.}
There are situations where distributing the priors across the collaborative agents is not possible and, instead, the data fusion node has access to the local posteriors only. These agents might have used the same priors for the quantity or different across agents. A popular fusion rule is the so-called product of experts (PoE) by which the local posteriors are multiplied together in order to produce a global posterior estimate, which is necessarily an approximation of the optimal solution in \eqref{eq:CIL}.
Following the nomenclature used in this paper, we refer to this approach as the approximate Bayesian fusion rule under \emph{conditionally independent posterior} (CIP), where the main assumption is that global posterior can be factorized into a product of the local posteriors. The resulting posterior $\widetilde p(\boldsymbol{\theta}|\mathcal{D})$ is an approximation of the \textit{true} posterior $p(\boldsymbol{\theta}|\mathcal{D})$ which is accurate when, indeed, the assumption holds. More precisely,
\begin{align}\label{eq:CIP}
\widetilde p(\boldsymbol{\theta}|\mathcal{D}) = \widetilde Z \prod_{m=1}^{M} p(\boldsymbol{\theta}|\mathcal{D}_m)\;,
\end{align}
where $\widetilde Z$ is a normalization term and $p(\boldsymbol{\theta}|\mathcal{D}_m)$ is $m$-th local posterior as computed by each of the $M$ agents.
The results in Section \ref{sec:experiments} will provide a comparison between CIP and CIL approaches. In addition, we are interested in developing analytical tools to gain understanding about their differences.
Mainly, CIP in \eqref{eq:CIP} differs from CIL in \eqref{eq:CIL} in their underlying assumptions on the independence of either the local posteriors or the data, respectively.
We will use the KL divergence between the optimal Bayesian fusion rule (referred to as CIL) and the approximate rule (or CIP), such that we can understand theoretically when it is worth accounting for the double counting of the prior. The KL divergence, for a general model where $M$ agents are fused, results in
\begin{align} \label{eq:KL}
&\mathrm{KL}_M\left(p(\boldsymbol{\theta}|\mathcal{D})||\widetilde p(\boldsymbol{\theta}|\mathcal{D})\right)
= \int p(\boldsymbol{\theta}|\mathcal{D}) \log\left(\frac{p(\boldsymbol{\theta}|\mathcal{D})}{\widetilde p(\boldsymbol{\theta}|\mathcal{D})}\right)d\boldsymbol{\theta}\\ \nonumber
&= \int \frac{\prod_{m=1}^M p(\mathcal{D}_m|\boldsymbol{\theta}) p(\boldsymbol{\theta})}{p(\mathcal{D})} \log\left(\frac{ \frac{\prod_{m=1}^M p(\mathcal{D}_m|\boldsymbol{\theta}) p(\boldsymbol{\theta})}{p(\mathcal{D})}}{ \frac{\prod_{m=1}^M p(\mathcal{D}_m|\boldsymbol{\theta}) p^M(\boldsymbol{\theta})}{p_M(\mathcal{D})} }\right)d\boldsymbol{\theta}\\ \nonumber
&= \int \frac{\prod_{m=1}^M p(\mathcal{D}_m|\boldsymbol{\theta}) p(\boldsymbol{\theta})}{p(\mathcal{D})} \log\left(\frac{p_M(\mathcal{D})}{p(\mathcal{D})}\frac{ 1 }{ p^{M-1}(\boldsymbol{\theta}) }\right)d\boldsymbol{\theta} \;,\\ \nonumber
\end{align}
\noindent where we explicitly note the number of agents as a subindex in the KL for clarity. The expression can be further manipulated as
\begin{align} \label{eq:KL}
&\mathrm{KL}_M(p(\boldsymbol{\theta}|\mathcal{D})||\widetilde p(\boldsymbol{\theta}|\mathcal{D}))
=\log\left(\frac{p_M(\mathcal{D})}{p(\mathcal{D})}\right) - (M-1) \int \frac{p(\mathcal{D}|\boldsymbol{\theta}) p(\boldsymbol{\theta})}{p(\mathcal{D})} \log(p(\boldsymbol{\theta})) d\boldsymbol{\theta},
\end{align}
with the definitions
\begin{eqnarray}
p(\mathcal{D}) &=& \int \prod_{m=1}^M p(\mathcal{D}_m|\boldsymbol{\theta}) p(\boldsymbol{\theta})d\boldsymbol{\theta}, \\
p_M(\mathcal{D}) &=& \int \prod_{m=1}^M p(\mathcal{D}_m|\boldsymbol{\theta}) p^M(\boldsymbol{\theta}) d\boldsymbol{\theta}\;.
\end{eqnarray}
\section{Fusion of Gaussian estimators}\label{sec:GaussianFusion}
Gaussian assumptions are often considered in order to allow for inferential tractability \cite{Kim08}. Furthermore, it is also common to approximate the posterior by a Gaussian distribution, for instance in the context the so-called Gaussian filters \cite{Ito00} or Laplace approximations (see for instance INLA \cite{lindgren2015bayesian}).
In this section, we analyze the relevant case where $\Theta = \mathbb{R}$, such that inferring $\boldsymbol{\theta}\in\mathbb{R}^{d}$ can be interpreted as a regression problem or a classification task when Laplace approximation is considered \cite{bishop2006pattern}.
In this context, we assume that the prior on the parameter is normally distributed, $\boldsymbol{\theta}\sim \mathcal{N}(\mathbf{\boldsymbol{\theta}}_0,\mathbf{C}_0)$, and that the distribution of the local likelihood for the $m$-th agent is
\begin{align}\label{eq:regression_example}
\mathbf{y}_{n,m} | \boldsymbol{\theta} \sim \mathcal{N}(\mathbf{f}(\boldsymbol{\theta}, \mathbf{x}_{n,m}),\mathbf{R}_m)
\end{align}
where $\mathbf{x}_{n,m}\in\mathbb{R}^{d_x}$ is the $n$-th feature input vector, $\mathbf{f}(\boldsymbol{\theta}, \mathbf{x}_{n,m})$ is a mapping function from those inputs to observed data $\mathbf{y}_{n,m}\in\mathbb{R}^{d_y}$. $n=1,\dots , N_m$ denotes the sample index for the $m$-th agent. The local dataset, as described in Section \ref{sec:fusion}, is then composed of the $N_m$ pairs $\mathcal{D}_m = \{\mathbf{y}_{n,m} , \mathbf{x}_{n,m}\}_{n=1}^{N_m}$, and $\mathbf{R}_m$ denotes the observation covariance matrix, which potentially can be different across agents.
Note that when the mapping is a linear function on $\boldsymbol{\theta}$, then the posterior is also normally distributed and its parameters can be optimally computed \cite{bishop2006pattern}. In general problems, if the function $\mathbf{f}(\cdot,
\cdot)$ is non-linear, one can still make an approximation that the posterior would be approximately Gaussian distributed and compute its mean/covariance using for instance a Laplace approximation.
Under the Gaussian approximation, the resulting local posterior $p(\boldsymbol{\theta}|\mathcal{D}_m) \approx \mathcal{N}(\boldsymbol{\theta}_{m},\mathbf{C}_m)$ with
\begin{align}\label{eq:local_posterior}
&\boldsymbol{\theta}_{m}=\mathbf{C}_m\left(\mathbf{C}_{0}^{-1} \boldsymbol{\theta}_{0} + \sum_{n=1}^{N_m} \boldsymbol{F}_{n,m}^{\top}\mathbf{R}_m^{-1} \mathbf{y}_{n,m}\right) \;,\\
&\mathbf{C}_m^{-1}=\mathbf{C}_{0}^{-1} + \sum_{n=1}^{N_m} \boldsymbol{F}_{n,m}^{\top} \mathbf{R}^{-1} \boldsymbol{F}_{n,m},
\end{align}
\noindent where $\boldsymbol{F}=\partial \mathbf{f}/\partial\boldsymbol{\theta}$ is the Jacobian of the model, resulting from its linearization (or the coefficients of the model in case it is already linear).
\noindent\textbf{Optimal Bayesian data fusion.}
The optimal Bayesian fusion rule, or CIL for short under (approximately) local Gaussian posteriors leads to a Gaussian posterior given by
\begin{align}\label{eq:gassuian_Correlated fusion}
p(\boldsymbol{\theta}|\mathcal{D}_1,\dots, \mathcal{D}_M)
= & \frac{\frac{1}{p^{M-1}(\boldsymbol{\theta})} \prod_{m=1}^{M} p(\boldsymbol{\theta}|\mathcal{D}_m)}{\int \frac{1}{p^{M-1}(\boldsymbol{\theta})} \prod_{m=1}^{M} p(\boldsymbol{\theta}|\mathcal{D}_m) d\boldsymbol{\theta}}
= \mathcal{N}(\mathbf{\boldsymbol{\mu}}, \mathbf{\Lambda}^{-1})\;,
\end{align}
where (see \ref{ap:GaussianCIL} for details) the posterior precision matrix can be obtained as
\begin{align}\label{eq:variance_close}
\mathbf{\Lambda} = \sum_{m=1}^{M}\mathbf{C}_m^{-1}-\mathbf{C}_{0}^{-1}(M-1)\;,
\end{align}
\noindent which depends on the local and a priori covariance matrices. The posterior mean is then
\begin{equation}\label{eq:mean_CIL}
\boldsymbol{\mu} = \mathbf{\Lambda}^{-1}\left(\sum_{m=1}^{M}\mathbf{C}_m^{-1} \boldsymbol{\theta}_m -(M-1) \mathbf{C}_{0}^{-1} \boldsymbol{\theta}_{0}\right)
\;,
\end{equation}
which can be rearranged as
\begin{equation}\label{eq:mean_CIL2}
\boldsymbol{\mu} = \sum_{m=1}^{M}\mathbf{\Xi}_m \boldsymbol{\theta}_m + \mathbf{\Xi}_0 \boldsymbol{\theta}_0 = \sum_{m=0}^{M}\mathbf{\Xi}_m \boldsymbol{\theta}_m\;,
\end{equation}
\noindent with the weight matrices as
\begin{equation}
\mathbf{\Xi}_m = \left\{
\begin{array}{cc}
\mathbf{\Lambda}^{-1} \mathbf{C}_m^{-1}, & m\neq 0, \\
(1-M) \mathbf{\Lambda}^{-1} \mathbf{C}_{0}^{-1}, & m=0.
\end{array}
\right.
\end{equation}
\noindent\textbf{Approximate Bayesian data fusion.}
Under the Gaussian assumption, the approximate fusion rule in \eqref{eq:CIP}, or CIP in short, results in
\begin{equation}\label{eq:gassuian_Correlated fusion_approx}
\widetilde p(\boldsymbol{\theta}|\mathcal{D}_1, \mathcal{D}_2,\dots,\mathcal{D}_M)
= \mathcal{N}(\widetilde{\mathbf{\boldsymbol{\mu}}}, \widetilde{\mathbf{\Lambda}}^{-1})\;,
\end{equation}
\noindent where
\begin{align}
\widetilde{\boldsymbol{\mu}} &= \widetilde{\mathbf{\Lambda}}^{-1} \sum_{m=1}^{M}\mathbf{C}_m^{-1} \boldsymbol{\theta}_m, \label{eq:mean_CIP}\\
\widetilde{\mathbf{\Lambda}} &= \sum_{m=1}^{M}\mathbf{C}_m^{-1}\;, \label{eq:cov_CIP}
\end{align}
\noindent as detailed in \cite{victor201803} and which can also be interpreted as a particular case of the Gaussian CIL rule \eqref{eq:gassuian_Correlated fusion} when no a priori information is considered in the fusion stage. That is, intuitively, that $\mathbf{C}_0$ takes very large values such that ${\boldsymbol{\mu}} \rightarrow \widetilde{\boldsymbol{\mu}}$ and ${\mathbf{\Lambda}} \rightarrow \widetilde{\mathbf{\Lambda}}$.
\noindent\textbf{Convergence results for CIL and CIP fusion approaches.}
We are interested in understanding when the two fusion rules are equivalent. That is, when accounting for the multiple use of the a priori information makes a difference and when, on the contrary, can be neglected thus simplifying the calculus. To that aim, we use the KL divergence to quantify the similarities of $p(\boldsymbol{\theta}|\mathcal{D}) = \mathcal{N}(\mathbf{\boldsymbol{\mu}}, \mathbf{\Lambda}^{-1})$
and
$\widetilde p(\boldsymbol{\theta}|\mathcal{D})
= \mathcal{N}(\widetilde{\mathbf{\boldsymbol{\mu}}}, \widetilde{\mathbf{\Lambda}}^{-1})$.
The KL divergence is known to be
\begin{multline}\label{eq:kl_gaussian}
\mathrm{KL}_M(p(\boldsymbol{\theta}|\mathcal{D})||\widetilde p(\boldsymbol{\theta}|\mathcal{D})) = \\
= \frac{1}{2}\left[\log\frac{|\widetilde{\mathbf{\Lambda}}^{-1}|}{|{\mathbf{\Lambda}}^{-1}|} - d + ({\mathbf{\boldsymbol{\mu}}}-\widetilde{\mathbf{\boldsymbol{\mu}}})^T \widetilde{\mathbf{\Lambda}}({\mathbf{\boldsymbol{\mu}}}
-\widetilde{\mathbf{\boldsymbol{\mu}}}) \right.
\left. + \textrm{Trace}\left\{\widetilde{\mathbf{\Lambda}}\mathbf{\Lambda}^{-1}\right\}
\right]\;,
\end{multline}
\noindent where we make explicit its dependence on the number of agents $M$ for the sake of clarity in the upcoming results.
The divergence tends to zero as the two distributions become more similar. In that sense, we provide two results showing convergence of CIP to CIL, which are stated as the following two theorems.
\begin{defn}\label{defn_thm}
Let $\mathcal{D}$ be a global dataset of \textit{i.i.d.} observations, which are related to a parameter of interest $\boldsymbol{\theta}\in\mathbb{R}^{n_\theta}$. Let $\mathcal{D}$ be processed by $M$ local agents, each observing mutually exclusive sets $\mathcal{D} = \{\mathcal{D}_1,\mathcal{D}_2,\dots,\mathcal{D}_M \}$ and sharing the same a priori information from the quantity of interest $p(\boldsymbol{\theta})$. The $m$-th agent updates the prior to compute its local posterior, $p(\boldsymbol{\theta}|\mathcal{D}_m)$, which are fused at a central node either using the optimal Bayesian fusion rule $p(\boldsymbol{\theta}|\mathcal{D})$ or the approximate Bayesian fusion rule $\widetilde p(\boldsymbol{\theta}|\mathcal{D})$.
\end{defn}
\begin{thm}\label{theorem_M}
Given Definition \ref{defn_thm}, the divergence between $p(\boldsymbol{\theta}|\mathcal{D})$ and $\widetilde p(\boldsymbol{\theta}|\mathcal{D})$ increases with $M$, that is,
$$\mathrm{KL}_{M+1} > \mathrm{KL}_M$$
when both local posteriors and a priori are normally distributed.
\end{thm}
\begin{proof}
Details in \ref{ap:thm_M}.
\end{proof}
\begin{thm}\label{theorem_q}
Given Definition \ref{defn_thm}, $\widetilde p(\boldsymbol{\theta}|\mathcal{D})$ asymptotically tends to $p(\boldsymbol{\theta}|\mathcal{D})$ as the shared prior becomes non-informative, that is,
$$\lim_{q_0\rightarrow \infty} \mathrm{KL}_M (p(\boldsymbol{\theta}|\mathcal{D})||\widetilde p(\boldsymbol{\theta}|\mathcal{D})) = 0 \;, \quad \forall M\geq 1$$
under the assumption that both local posteriors and a priori are normally distributed.
\end{thm}
\begin{proof}
Details in \ref{ap:thm_q}.
\end{proof}
Intuitively, the result in Theorem \ref{theorem_M} shows that for a fixed budget of samples $N$, increasing the number of agents $M$ increases the separation between the two global posterior solutions. That is, as more agents participate less data per agent is available, in which case the a priori distribution on $\boldsymbol{\theta}$ becomes more relevant and the issue of reusing the prior emerges.
Secondly, the main result in Theorem \ref{theorem_q} states that CIL converges to CIP as the a priori distribution on $\boldsymbol{\theta}$ becomes less informative. Notice that the result assumes that $\mathbf{C}_0 = q_0\mathbf{I}$ for the sake of simplicity.
Note that our analysis is based on a Gaussian assumption, which is widely used in the context of Bayesian inference since, under mild conditions, the posterior converges in distribution to a normal according to the Bernstein–von Mises theorem
\cite{kleijn2012bernstein}.
\section{Experiments}\label{sec:experiments}
We validate the obtained results on various relevant inference problems. Namely, $(1)$ an estimation problem, where a set of $M$ linear regression models are employed to update the prior with local data, which are subsequently fused to produce global estimates; $(2)$ a binary classification problem in which $M$ models are used locally to produce classification results that are then fused to produce a global result; $(3)$
another general classification problem, where $M$ neural networks are locally used to produce classification solutions that are then aggregated into a global Bayesian classifier; and $(4)$ an application in the context of federated learning, which in this case updates the prior recursively instead of just once, unlike the previous experiment.
\subsection{Distributed estimation of linear models}\label{sec:exa}
For the experiments presented in this subsection we generated a synthetic data set following a linear model, with respect to the parameters $\boldsymbol{\theta}$, embedded in noise:
\begin{align}\label{eq:regression_example}
\mathbf{y}_m=\boldsymbol{\theta}^\top \bm{\phi}(\mathbf{x}_m) + \mathbf{r}_m ,
\end{align}
where $\mathbf{r}_m \sim \mathcal{N}(\mathbf{0},\mathbf{R})$, with $\mathbf{R} = r \mathbf{I}$, and $\bm{\phi}:\mathbb{R}^{d_x}\to\mathbb{R}^d,\mathbf{x}\mapsto \bm{\phi}(\mathbf{x})$ is an arbitrary nonlinear function.
In our experiment, we set $r=4$, the observation dimension is $1$, and $\bm{\phi}$ is identity matrix. The $\boldsymbol{\theta}$ is a real random integer vector generated uniformly between $-10$ and $20$ with $d_x=6$ dimensions. The total number of points generated is around $700$ for training and around $300$ for testing. Then, we split the dataset into several parts, each corresponding to one of the local clients' training data.
We estimated local parameters following equation \eqref{eq:local_posterior} and fused using the approaches as discussed in Section~\ref{sec:GaussianFusion}. Furthermore, we computed the evolution of the KL divergence, and test MSE with both prior variance $q_0\in\mathbb{R}_+$, and total number of clients $M$. The results are depicted in Figures~\ref{Fig:reg_kl}-\ref{Fig:reg_client_test}. All results are obtained by average of multiple
independent Monte Carlo simulations.
\begin{figure}[t]
\centering
\begin{subfigure}[t]{0.8\textwidth}
\centering
\input{latex_figures/kl}
\label{Fig:kl}
\end{subfigure}%
\vskip 0.1cm
\begin{subfigure}[t]{0.8\textwidth}
\centering
\input{latex_figures/kl_client}
\label{Fig:kl_client}
\end{subfigure}
\caption{Evolution of the KL divergence in \eqref{eq:kl_gaussian} wrt $q_0$ (top) and number of client $M$ (bottom).}
\label{Fig:reg_kl}
\end{figure}
Figure \ref{Fig:reg_kl} shows the evolution of the KL divergence in \eqref{eq:kl_gaussian} with respect to the prior variance $q_0$ (top panel for $M=6$ and $M=26$) and the number of clients $M$ (bottom panel for $q_0=1$, $q_0=3$).
In this figure, the KL decreases wrt prior variance $q_0$. Which can also be shown in the bottom panel for $q_0=1$, $q_0=3$, where the KL increase wrt the number of clients.
These results are consistent with Theorem \ref{theorem_M} and \ref{theorem_q}, which was already predicting such behavior.
\begin{figure}[t]
\centering
\begin{subfigure}[t]{0.8\textwidth}
\centering
\input{latex_figures/reg_q_test}
\end{subfigure}
\caption{Test MSE with respect to $q_0$ for the distributed linear estimation in Section \ref{sec:exa}.}
\label{Fig:reg_q_test}
\end{figure}
Figure \ref{Fig:reg_q_test} shows the test MSE wrt the prior variance $q_0$ for a fixed number of collaborating clients of 6 and 26.
This figure shows that all tested configurations converge to similar results when the prior variance is large enough, that is when the a priori information on $\boldsymbol{\theta}$ is non-informative. However, when given $p(\boldsymbol{\theta})$ is informative, represented by smaller variance values, the impact of properly accounting for it would be more apparent. Independently of the number of users, CIL seems to exhibit stable results, while CIP has a larger impact on the result as the prior becomes narrower.
\begin{figure}[t]
\centering
\begin{subfigure}[t]{0.8\textwidth}
\centering
\input{latex_figures/reg_client_test}
\label{Fig:reg_client_test}
\end{subfigure}
\caption{The test MSE wrt $M$ under different prior information when prior variance at $1$ and $3$.}
\label{Fig:reg_client_test}
\end{figure}
Figure \ref{Fig:reg_client_test} shows the MSE of test wrt the number of clients $M$, as well as for the two representative values prior variance $q_0=1$ and $q_0=3$.
The results suggest that CIL can outperform CIP as we increasing the number of clients. In that situation, for growing $M$, CIP uses the prior information repeatedly, causing the results to be biased unless this fact is taken into consideration as in the CIL appoach. This effect is even more notorious when the a priori information is not accurate (e.g., the mean value is far from the true value of $\boldsymbol{\theta}$).
\subsection{Fusion of local class posterior probabilities}\label{sec:exb}
We are now interested in demonstrating the proposed Bayesian data fusion methodology when priors are shared across agents in a classification problem.
Similarly to BCM \cite{tresp2000bayesian}, we consider the fusion of classification results as the fusion of probability functions. Analogous to \eqref{eq:CIL}, the CIL fusion, given input $\mathbf{x}$, results in a discrete posterior probability distribution:
\begin{align}\label{eq:clf_probability}
P(C=c|\mathcal{D}_1,\dots,\mathcal{D}_M, \mathbf{x}) \propto \frac{\Pi_{m=1}^M P(C=c|\mathcal{D}_m,\mathbf{x})}{P^{M-1}(C=c)},
\end{align}
where $c$ is the classification label with $c\in\{1,\dots,L\}$. The a priori probability of the classes, shared among agents, is denoted by $P(C=c)$.
To obtain the local class posterior probabilities $P(C|\mathcal{D}_m,\mathbf{x})$, we consider an linear discriminant analysis (LDA) model, which optimally utilizes the shared prior, then fuse the local posteriors to obtain the final result in \eqref{eq:clf_probability}.
In this experiment, for the sake of clarity, we are considering an $L=2$ classes classification problem. The objective is thus to estimate their class posterior probabilities $P(C=i|\mathcal{D}_1,\dots,\mathcal{D}_M, \mathbf{x})$ with $i\in\{1,2\}$.
We generated a synthetic dataset, consisting of $1000$ data points, by sampling from two normal distributions (for class $i$, the conditional likelihood of the generated data is $p(\mathcal{D}_m|\mathbf{x},C=i) = \mathcal{N}(\mathbf{1}_i,\mathbf{I})$, where $\mathbf{1}_i$ is an all-zeroes vector except for a $1$ at the $i$-th element. The dimension of the data $\mathbf{y}$ is $10$. The prior class probabilities were $0.6$ and $0.4$ for classes $1$ and $2$, respectively. The resulting dataset was divided into $M$ subsets of the same dimension $1000/M$, such that these are \textit{i.i.d.} distributed.
\begin{figure}[t]
\centering
\begin{subfigure}[t]{0.8\textwidth}
\centering
\input{latex_figures/clf_pro_kl_q}
\end{subfigure}%
\vskip 0.1cm
\begin{subfigure}[t]{0.8\textwidth}
\centering
\input{latex_figures/clf_pro_kl_client}
\end{subfigure}
\caption{KL divergence between CIL and CIP class posteriors as a function of (top) the class $1$ prior probability $P_1$ and (bottom) number of clients $M$. Distributed classification problem from Section \ref{sec:exb}.}
\label{Fig:clf_pro_kl}
\end{figure}
Figure \ref{Fig:clf_pro_kl} shows the evolution of the KL divergence of class posterior distribution of CIP and CIL with respect to (top panel) the first class prior probability $P_1\triangleq P(C=1)$ for $M=6$ and $M=12$ clients; and (bottom panel) the number of clients $M$ for $P_1=0.1$ and $P_1=0.4$.
In the top panel, the figure shows that the KL divergence is minimized when the class prior probabilities are $P_1=P_2=1/2$, such that non-informative priors were employed.
Additionally, the KL divergence for $M=12$ clients is larger than the one for $M=6$ clients, which is also shown in the bottom panel where the KL divergence increases with $M$.
Figure \ref{Fig:clf_pro_q_acc} shows the accuracy performance of posterior class distribution of both CIP and CIL with respect to $P_1$.
Similarly to the KL analysis in Figure \ref{Fig:clf_pro_kl}, the accuracy of both approaches are similar for non-informative class priors $P_1=P_2=1/2$.
On the other hand, when the prior probability is far from the actual class prior probability, CIP's accuracy is degraded since that mismatched prior is overused, whereas for CIL the performance does not drop dramatically. When the prior probability is close to real class distribution (that is, $P_1=0.6$ and $P_2=0.4$), both of these two methods achieve relatively decent results.
Figure \ref{Fig:clf_pro_client_acc} shows the classification accuracy, this time in terms of the number of clients $M$ and two values for $P_1$.
The results show that CIL outperforms CIP, similarly to Figure \ref{Fig:clf_pro_q_acc}.
\begin{figure}[t]
\centering
\begin{subfigure}[t]{0.8\textwidth}
\centering
\input{latex_figures/clf_pro_q_acc}
\label{Fig:clf_pro_q_acc}
\end{subfigure}
\caption{The accuracy with respect to the class 1 prior probability for various numbers of clients $M$ and fixed amount of total data.}
\label{Fig:clf_pro_q_acc}
\end{figure}
\begin{figure}[t]
\centering
\begin{subfigure}[t]{0.8\columnwidth}
\centering
\input{latex_figures/clf_pro_client_acc}
\label{Fig:clf_pro_client_acc}
\end{subfigure}
\caption{Testing accuracy with respect to $M$ under class 1 prior probabilities $0.1$ and $0.4$, for the classification problem in Section \ref{sec:exb}.}
\label{Fig:clf_pro_client_acc}
\end{figure}
\subsection{Distributed learning of neural network classifiers}\label{sec:exc}
In this test we consider a set of $M$ neural networks with the same structure and network parameters $\boldsymbol{\theta}$, each trained on observed local data. The objective is therefore to fuse the trained NNs into a global NN that can then be used for classification purposes, accounting for all local data through a distributed training process. Additionally, since the focus of this paper is on Bayesian data fusion, those are Bayesian NNs (BNN) in that their weights are treated as random variables. Therefore, training of a BNN involves inferring the joint posterior of its parameters given available data. The objective of this experiment is to locally train the BNNs, then fuse the local posteriors to compute a global posterior $p(\boldsymbol{\theta}|\mathcal{D}_1,\dots, \mathcal{D}_M)$. The use of Bayesian approaches within the context of data-driven models has been investigated previously and is given attention more recently \cite{fortuin2021bayesian}, where for the sake of simplicity the weights in the BNN are typically assumed independent and normally distributed such that mean and covariance can characterize their posterior.
In this experiment, we used a similar classification dataset as in Section \ref{sec:exb}, sampling from a normal distribution for each class $p(\mathcal{D}_m|\mathbf{x},C=i) = \mathcal{N}(\mathbf{1}_i,\mathbf{I})$, where in this case the dimension of the problem was increased to $10$ classes, $i=\{1,\dots,10\}$, while size of the data was kept to $10$ as in Section \ref{sec:exb}. The BNN trained by the $M$ local agents had a hidden layer of $64$ neurons, the training epoch was set to $100$, and the learning rate was $0.05$.
It is worth mentioning that the total amount of data is fixed, such that increasing $M$ would have the impact of reducing the amount of local data available at each client.
Given a local dataset, there are several methods that can be used to calculate the posterior distribution of the BNN parameters. Without loss of generality, we considered here a Laplace approximation of the BNN parameters \cite{bishop2006pattern}, assuming the parameters are normally distributed.
\begin{figure}[t]
\centering
\begin{subfigure}[t]{0.8\textwidth}
\centering
\input{latex_figures/clf_kl_q_2}
\end{subfigure}%
\vskip 0.1cm
\begin{subfigure}[t]{0.8\textwidth}
\centering
\input{latex_figures/clf_kl_client_2}
\label{Fig:kl_client}
\end{subfigure}
\caption{KL divergence between CIL and CIP class posteriors as a function of (top) $q_0$ and (bottom) $M$ for the distributed classification training in Section \ref{sec:exc}.}
\label{Fig:clf_kl}
\end{figure}
Figure \ref{Fig:clf_kl} shows the KL divergence between the parameters' posterior of both CIP and CIL solutions, as a function of the a priori variance $q_0$ (top panel for $M=6$ and $M=16$ clients) and the number of clients $M$ (bottom panel for $q_0=4$ and $q_0=1.6$).
From the top panel of this figure, we can see that behavior of the KL divergence with respect to $q_0$ or $M$ are similar to those from \ref{sec:exa}.
\begin{figure}[t]
\centering
\begin{subfigure}[t]{0.8\textwidth}
\centering
\input{latex_figures/clf_q_acc_2}
\label{Fig:clf_q_acc}
\end{subfigure}
\caption{Testing accuracy with respect to the prior variance for a fixed amount of total data across the $M$ clients.}
\label{Fig:clf_q_acc}
\end{figure}
\begin{figure}[t]
\centering
\begin{subfigure}[t]{0.8\textwidth}
\centering
\input{latex_figures/clf_client_acc_2}
\label{Fig:clf_client_acc}
\end{subfigure}
\caption{Testing accuracy with respect to $M$ for a different values of $q_0$.}
\label{Fig:clf_client_acc}
\end{figure}
Figure \ref{Fig:clf_q_acc} shows the fusion center NN testing accuracy with respect the prior variance $q_0$ for $M=6$ and $M=16$ clients. For the sake of clarity, and to facilitate the identification of the trend behavior, the reported metric is test accuracy of CIL (and CIP) divided by global accuracy (considered as a benchmark, using the global model for testing, which is trained by using all datasets jointly). As expected, as the prior becomes more informative, CIL outperforms CIP, and as $M$ increases the local data becomes more scarce and the performance degrades.
Analogously, Figure \ref{Fig:clf_client_acc} shows the testing accuracy with respect to the number of clients $M$ for different values of $q_0$, drawing similar conclusions.
\subsection{Recursive update of classification problem using $M$ neural networks}\label{sec:exd}
Results presented in the previous subsection only considered the case where classifier's parameters are fused once, after learned using local data. In many applications, however, one aims at fusing models over time in an iterative procedure. For instance, this is the case of many distributed and/or federated methods where a common prior can be shared with several nodes and fused back recursively~\cite{liu2021bayesian}.
In this experiment, we consider the case of Bayesian federated model learning where learned distributions from distributed clients are fused by a central node in the vein of Figure~\ref{fig:distributed}.
Locally, at the $t$-th communication round each client obtains their local posterior of the BNN model parameters using Bayes theorem as
\begin{align}\label{eq:recursive}
p_t(\boldsymbol{\theta}|\mathcal{D}_m)=& \frac{p_{t-1}(\boldsymbol{\theta} ) p(\mathcal{D}_m | \boldsymbol{\theta})}{P(\mathcal{D}_m)},
\end{align}
\noindent where $p_{t-1}(\boldsymbol{\theta} )$ denotes the prior distribution of the parameters at $t$ and the $\log$-posterior can be written as
\begin{align}
\ln p_t(\boldsymbol{\theta}|\mathcal{D}_m) &=\ln p_{t-1}(\boldsymbol{\theta}) +\ln p(\mathcal{D}_m | \boldsymbol{\theta})-\ln p(\mathcal{D}_m) \, . \label{eq:recursive_last}
\end{align}
For simplicity, we consider the Gaussian assumption on the posterior $p_t(\boldsymbol{\theta}|\mathcal{D}_m) \approx \mathcal{N}(\boldsymbol{\theta}_{m,t},\mathbf{C}_{m,t})$, similarly as in \eqref{eq:local_posterior}.
Defining the loss for the $m$-th client as the log-posterior in~\eqref{eq:recursive_last}, i.e., $J_{m,t}(\boldsymbol{\theta}) = \ln p_t(\boldsymbol{\theta}|\mathcal{D}_m)$, and using the Laplace approximation for its second term, the loss can be approximated as
\begin{align}
& J_{m,t} (\boldsymbol{\theta}) \approx \frac{1}{2}(\boldsymbol{\theta}-\mathbf{\boldsymbol{\mu}}_{t-1})^{\top} \mathbf{\Lambda}_{t-1}(\boldsymbol{\theta}-\mathbf{\boldsymbol{\mu}}_{t-1})
+\frac{1}{2}(\boldsymbol{\theta}-\boldsymbol{\theta}_{m,t}^{\textrm{ML}})^{\top} \mathbf{H}_{m,t}(\boldsymbol{\theta}-\boldsymbol{\theta}_{m,t}^{\textrm{ML}})+\kappa,
\end{align}
\noindent where $p_{t-1}(\boldsymbol{\theta}) \approx \mathcal{N}(\mathbf{\boldsymbol{\mu}}_{t-1},\mathbf{\Lambda}_{t-1}^{-1})$ is the parameter prior at $t$, resulting from the CIL (or CIP) fusion of local posteriors at $k-1$.
The Laplace approximation of the likelihood
$p(\mathcal{D}_m | \boldsymbol{\theta}) \approx \mathcal{N}(\boldsymbol{\theta}_{m,t}^{\textrm{ML}} , \mathbf{H}_{m,t}^{-1})$ requires maximum likelihood (ML) estimation of the parameter, $\boldsymbol{\theta}_{m,t}^{\textrm{ML}}$, possibly through a gradient method or other numerical optimization approach. A suitable estimate of the inverse covariance $\mathbf{H}_{m,t}^{-1}$ in the Laplacian approximation is known to be \cite{bishop2006pattern} the Fisher information matrix $\mathcal{I}(\boldsymbol{\theta})$, which can be approximated \cite{liu2021bayesian} by
$\mathcal{I}_m (\boldsymbol{\theta}) \approx \frac{1}{\left|\mathcal{D}_m\right|} \sum_{(\mathbf{x}, \mathbf{y}) \in \mathcal{D}_m} \nabla_\theta \log p(\mathbf{y} | \mathbf{x}, \boldsymbol{\theta}) \nabla_\theta \log p(\mathbf{y} | \mathbf{x}, \boldsymbol{\theta})^{\top}$ for the $m$-th client.
The $\kappa$ gathers constant terms that are not related to $\boldsymbol{\theta}$ and thus do not contribute to the loss minimization.
Then, the local parameter posterior mean for the $m$-th client can be estimated by solving the following optimization problem:
\begin{equation}\label{eq:mean of weight recursively}
\boldsymbol{\theta}_{m,t} = \mathop{\arg\max}_{\boldsymbol{\theta}} J_{m,t}(\boldsymbol{\theta}) \;,
\end{equation}
\noindent and the associated local covariance is
\begin{align}\label{eq:variance of weight recursively}
\mathbf{C}_{m,t}^{-1}=\frac{\partial^2 J_m(\boldsymbol{\theta})}{\partial \boldsymbol{\theta} \partial \boldsymbol{\theta}^\top} & \approx \bar{\mathbf{F}}_{m}(\boldsymbol{\theta}_{m,t}^{\textrm{ML}})+\mathbf{\Lambda}_{t-1}\,.
\end{align}
After obtaining the local posterior parameters at $t$, $\boldsymbol{\theta}_{m,t}$ and $\mathbf{C}_{m,t}$, for $m=1,\ldots,M$, these can be fused using either CIL and CIP approaches. This fused posterior would become $p_{t}(\boldsymbol{\theta}) \approx \mathcal{N}(\mathbf{\boldsymbol{\mu}}_{t},\mathbf{\Lambda}_{t}^{-1})$ the new prior used at the next communication round $t+1$.
Compared to previous classification experiments, this experiment considers a more complex synthetic dataset generated from a Gaussian mixture with a maximum of $4$ Gaussian components, with $3$ possible classes and $10$ features dimension. A total of $600$ \textit{i.i.d.} training points and $300$ testing points were generated. The considered BNN model is a multi-layer perceptron (MLP) with two hidden layers, containing $32$ and $8$ neurons respectively, with a learning rate of $0.01$ at training.
\begin{figure}[t]
\centering
\begin{subfigure}[t]{0.8\textwidth}
\centering
\input{latex_figures/clf_recursive_acc}
\label{Fig:clf_recursive_acc}
\end{subfigure}
\caption{Testing accuracy evolution with communication round for the recursive classification learning in Section \ref{sec:exd}.}
\label{Fig:clf_recursive_acc}
\end{figure}
\begin{figure}[t]
\centering
\begin{subfigure}[t]{0.8\textwidth}
\centering
\input{latex_figures/clf_recursive_var_mean}
\label{Fig:clf_recursive_var_mean}
\end{subfigure}
\caption{Evolution over $t$ of the mean variance of model parameters.}
\label{Fig:clf_recursive_var_mean}
\end{figure}
The results in Figures \ref{Fig:clf_recursive_acc} and \ref{Fig:clf_recursive_var_mean} show the evolution of accuracy and the mean value of the variance of all parameters, as a function of the communication round $t$ at which fusion happens. $M=4$ and $M=16$ are tested.
In the experiments, the initial round is given with non-informative prior information. After several communicate rounds occur, the variance becomes very informative, as shown in Figure \ref{Fig:clf_recursive_var_mean}.
It is seen from Figure \ref{Fig:clf_recursive_acc} that accuracy generally increases with the number of communication rounds, for both CIP and CIL. Also, the accuracy decreases inversely proportional to the number of clients $M$. The latter is a consequence of having less local data samples at each node, given that the overall amount is kept constant regardless of $M$. Another conclusion from the experiment is that CIL outperforms CIP, approaching each other when enough communication rounds are run. On the other hand, the average variance of the parameters produced by \eqref{eq:variance of weight recursively} is negligible, both approaches becoming over-confident.
\section{Conclusions}\label{sec:conclusions}
This paper investigates the impact of sharing a priori information of parameters of interest in a Bayesian data fusion context. The focus is on distributed learning of models for regression and classification purposes. We discuss two fusion possibilities in a distributed system, so-called conditionally independent posterior (CIP) and conditionally independent likelihoods (CIL). CIL is an optimal Bayesian method that assumes the local datasets are mutually independent and CIP is an approximation that assumed the local posteriors to be independent. CIP is sometimes desirable due to its simplicity, however the assumption is clearly broken when prior information is shared among clients. The paper quantifies the mismatch for the Gaussian case comparing those approaches and showing that they become equivalent when the a priori information is either not informative or when the number of users decrease, in which case more data is used locally and the prior becomes less relevant.
A comparison of these two methods in different applications of distributed inference is provided.
Namely, linear regression model distributed learning; distributed classification; distributed learning of neural network classifiers; and a practical use case in the context of federated learning, where the prior information is updated recursively over time as opposed to only once. The results consistently show that CIL, which accounts for the shared prior information, generally outperforms CIP, as it was predicted by the analytical results presented in this paper.
|
\section*{Multimedia Material}
Project page: \url{https://github.com/tub-rip/event_collapse}
\section{Introduction}
Event cameras \cite{Lichtsteiner08ssc,Finateu20isscc} are an emerging technology attracting considerable attention because of their advantages, such as low latency, high dynamic range and data/power efficiency.
These advantages are especially suitable for applications that require low-latency vision-based processing, such as autonomous driving \cite{Chen20msp,Murali22aisy}.
The neuromorphic principle of operation of event cameras makes them naturally respond to the scene dynamics.
Therefore, they have been used in motion estimation tasks
\cite{Wang21aisy,Zhong22aisy}, such as Simultaneous Localization and Mapping \cite{Kim16eccv,Rebecq17ral,Zhu17cvpr,Zhou20tro,Rosinol18ral,Rebecq18ijcv},
optical flow estimation \cite{Benosman14tnnls,Zhu19cvpr,Paredes21neurips,Shiba22eccv},
particle image velocimetry \cite{Willert22expfl},
motion segmentation \cite{Stoffregen19iccv,Zhou21tnnls},
obstacle avoidance \cite{Falanga20scirob},
and space situational awareness~\cite{Cohen19jas}.
\input{floats/fig_eyecatcher}
Contrast maximization (CMax) \cite{Gallego18cvpr} is a powerful event processing framework that achieves state-of-the-art accuracy in various motion estimation tasks.
On the other hand, it may suffer from an overfitting problem called \emph{event collapse},
where the optimizer converges to an undesired global optimum \cite{Shiba22sensors}.
Prior works have tackled event collapse in several ways,
such as whitening the data \cite{Nunes21pami},
reformulating the task (e.g., by providing additional depth information \cite{Nunes21pami}),
or adding a regularizer to improve the optimization landscape \cite{Shiba22sensors}.
However, these proposals present shortcomings:
($i$) assuming known depth data is task-specific and requires an additional sensor such as a LiDAR or a stereo setup,
and ($ii$) the above regularizing techniques may not be effective \cite{Nunes21pami} or require considerable extra computation \cite{Shiba22sensors}.
Towards more practical application of event cameras, it is paramount to effectively alleviate event collapse in a computationally-efficient manner.
\input{floats/fig_block_diagram}
This paper proposes a novel, computationally efficient regularizer to mitigate event collapse in the CMax framework (\cref{fig:eyecatcher}).
From a theoretical point of view, the regularizer is designed based on geometric principles of motion field deformation (measuring area rate of change along point trajectories).
In contrast to previous methods, the regularizer does not depend on the event data; it only depends on the motion hypothesis (i.e., the warp).
This is desirable because events may not be equally distributed on the space-time image domain,
and motion hypotheses provide information even in homogeneous brightness regions (where events are scarce).
From a practical point of view, the proposed regularizer
drastically reduces computational complexity, being two to four times faster than previous solutions,
while achieving state-of-the-art results to mitigate event collapse.
The above contributions open the door to efficient and interpretable regularizers for motion estimation problems with geometrically-meaningful parametrizations.
\section{Related Work}
Motion estimation from event data is an incipient research field.
Event cameras respond to moving edges in the scene, and edges carry the essential information for motion estimation.
Hence event cameras have numerous applications in motion-related tasks
(ego-motion and/or scene-motion).
We refer to \cite{Gallego20pami} for a comprehensive review.
State-of-the-art approaches are based on the concept of \emph{event alignment} or motion compensation \cite{Gallego18cvpr,Gallego19cvpr,Nunes21pami,Gu21iccv}, estimating the motion parameters that associate events corresponding to the same scene edge.
In particular, Contrast Maximization (CMax) \cite{Gallego18cvpr} seeks the motion parameters that maximize the contrast (or focus score) of an image of warped events (IWE).
It has been used on various motion estimation tasks,
such as rotational motion \cite{Gallego17ral,Kim21ral,Gu21iccv},
feature flow \cite{Zhu17icra,Zhu17cvpr,Seok20wacv,Stoffregen19cvpr},
homographic motion \cite{Gallego18cvpr,Nunes21pami,Peng21pami},
optical flow \cite{Zhu19cvpr,Paredes21neurips,Shiba22eccv},
motion segmentation \cite{Mitrokhin18iros,Stoffregen19iccv,Zhou21tnnls,Parameshwara21icra}
and 3D reconstruction \cite{Rebecq18ijcv,Ghosh22aisy}.
However, the optimizer may converge to an undesired global optimum that pushes events into too few pixels (i.e., event collapse).
Event collapse is originally analyzed in detail by \cite{Shiba22sensors}.
It may occur or not depending on the task, the space of motion hypotheses and the data.
Hence, previous works for tackling event collapse can be categorized based on tasks.
For example, optical flow estimation has a high number of degrees-of-freedom (DOFs) ($2\numPixels$, where $\numPixels$ is the number of pixels), i.e., motion parameters, and is a collapse-enabled problem.
A common approach to mitigate collapse for this task is to add a strong regularizer, such as the classical Charbonnier loss \cite{Charbonnier97tip}, to encourage smoothness of the flow \cite{Zhu19cvpr,Stoffregen19cvpr}.
Another approach consists of increasing the well-posedness of the problem by using a tile-based motion field and a multi-reference focus loss \cite{Shiba22eccv}.
However, event collapse may still appear in some parts of the image plane.
Ego-motion estimation problems, which are the main focus of this work, parametrize motion over the image space with relatively lower DOFs.
The well-posedness of the problem is affected not only by the number of DOFs but also by the geometric meaning of the motion.
By reformulating the problem to reduce the number of DOFs, \cite{Ozawa22sensors} and \cite{Peng21pami} increase the well-posedness of the task.
Whitening of warped events is proposed in \cite{Nunes21pami} to mitigate event collapse,
while \cite{Shiba22sensors} designs effective regularizers based on the divergence or the area deformation of the motion field, at the expense of increased computational cost.
Initialization close to the solution (in the basin of attraction of the desired local optimum) can also play an important role in evading event collapse \cite{Gallego18cvpr}.
Our work is most related to \cite{Shiba22sensors} because we focus on low-DOF motion estimation tasks
and seek a principled regularizer to gauge and penalize event collapse.
Our theoretical analysis provides a formula for homographic motions (8~DOFs), which can be particularized for:
1~DOF (zoom-in/out motion), 2~DOFs (feature flow), 3~DOFs (rotational camera motion), 4~DOFs (planar similarity motion) and 6~DOFs (planar affine motion).
While both \cite{Shiba22sensors} and our proposal are interpretable and grounded on geometric principles of motion trajectories, the most important theoretical difference is that our regularizer does not depend on the event data.
As a result, the regularizer drastically improves computational complexity while achieving on par or better motion estimation results on publicly-available datasets.
\section{Methodology}
\label{sec:method}
This section first reviews how an event camera works (\cref{sec:method:eventcamera})
as well as the regularized CMax framework (\cref{sec:method:cmaxreg}).
Then, we propose the new regularizer and explain its geometrical meaning and implications (\cref{sec:method:proposedReg}).
\subsection{Event Camera}
\label{sec:method:eventcamera}
Instead of acquiring brightness images at fixed time intervals (e.g., frames), event cameras record brightness differences asynchronously, called ``events'' \cite{Lichtsteiner08ssc,Gallego20pami}.
An event $e_k \doteq (\mathbf{x}_k, t_k, p_{k})$ represents a brightness change,
and it is triggered as soon as the logarithmic brightness at the pixel $\mathbf{x}_k\doteq (x_k, y_k)^{\top}$ exceeds a preset threshold.
Here, $t_k$ is the timestamp of the event with \si{\micro\second} resolution,
and polarity $p_{k} \in \{+1,-1\}$ is the sign of the brightness change.
\Cref{fig:method:overview} shows the input events (red and blue dots ($\mathbf{x}_k,t_k$) in space-time, with color representing polarity).
\subsection{Regularized Contrast Maximization}
\label{sec:method:cmaxreg}
The regularized CMax framework \cite{Shiba22sensors} aims at finding the motion parameters $\btheta$ that optimize the objective function:
\begin{equation}
\label{eq:compositeObjective}
\btheta^\ast %
= \arg\!\min_{\btheta} \left(-G(\btheta) + \lambda \mathcal{R}(\btheta)\right).
\end{equation}
Here, $G$ is the data fidelity term and $\mathcal{R}$ is a regularizer with weight $\lambda > 0$.
The overall steps of the framework are described in \cref{fig:method:overview}, where black lines indicate the previous approaches, and blue lines indicate our proposal.
\subsubsection{Data fidelity term}
The term $G(\btheta)$ measures the alignment of the events with respect to the candidate motion $\btheta$ (\cref{fig:method:overview}, black solid line).
The original events $\cE = \{e_k\}_{k=1}^{\numEvents}$ are transformed according to the motion hypothesis into a set of warped events $\cE' = \{e'_k\}_{k=1}^{\numEvents}$:
\begin{equation}
\label{eq:warp}
e_k \doteq (\mathbf{x}_k,t_k,p_k) \quad\stackrel{\mathbf{W}}{\mapsto}\quad
e'_k \doteq (\mathbf{x}'_k,\tref,p_k).
\end{equation}
The warp function $\mathbf{x}'_k = \mathbf{W}(\mathbf{x}_k,t_k; \btheta)$ transports every event along its motion trajectory until a reference time $\tref$ is reached.
Point trajectories are parametrized by $\btheta$, which consists of motion or scene unknowns (e.g., scene depth, moving objects).
Powerful objective functions are designed based on the count of warped events \cite{Gallego19cvpr}.
The representation of event count as an image is defined by the image of warped events (IWE):
\begin{equation}
\label{eq:IWE}
\IWE(\mathbf{x};\btheta) \doteq \sum_{k=1}^{\numEvents} b_k \,\delta (\mathbf{x} - \mathbf{x}'_k(\btheta)).
\end{equation}
Each pixel of IWE counts how many events $e'_k$ are warped into pixel $\mathbf{x}$.
The event polarity can be used by setting $b_k=p_k$, and not used if $b_k=1$.
The Dirac delta $\delta$ is approximated by a Gaussian: $\delta(\mathbf{x}-\bmu)\approx\cN(\mathbf{x};\bmu,\epsilon^2)$, where $\epsilon=1$ pixel.
Finally, the objective function, such as the IWE variance is calculated:
\begin{equation}
\label{eq:IWEVariance}
G(\btheta) \equiv \operatorname{Var}\bigl(\IWE(\mathbf{x};\btheta)\bigr)
\doteq \frac{1}{|\Omega|} \int_{\Omega} (\IWE(\mathbf{x};\btheta)-\mu_{\IWE})^2 d\mathbf{x},
\end{equation}
with mean $\mu_{\IWE} \doteq \frac{1}{|\Omega|} \int_{\Omega} \IWE(\mathbf{x};\btheta) d\mathbf{x}$ and image domain $\Omega$.
The interpretation of \eqref{eq:IWEVariance} is as follows:
the larger the IWE variance (contrast of the IWE), the better the alignment of the warped events $\cE'$.
Contrast is related to sharpness and focus \cite{Gallego19cvpr}.
\subsubsection{Previous regularizers}
The regularizer term $\mathcal{R}$ in \eqref{eq:compositeObjective} penalizes event collapse in certain types of warps.
When event collapse happens, the warped events are accumulated into too few pixels or lines, resulting in an undesired global optimum $G$ (overfitting).
Two regularizers are proposed in \cite{Shiba22sensors} by averaging some collapse quantities attached to each event (\cref{fig:method:overview}, dashed line).
Specifically, they use the divergence of the flow $\cD (\cE,\btheta)$ and the area-based deformation of the warp $\cA (\cE,\btheta)$, which are given by:
\begin{equation}
\label{eq:regSensors}
\begin{split}
\cD (\cE,\btheta) &\doteq \{\nabla \cdot \mathbf{f}_k\}_{k=1}^{\numEvents}, \\
\cA (\cE,\btheta) &\doteq \bigl\{|\det(\mathtt{J}(e_k))|\bigr\}_{k=1}^{\numEvents},
\end{split}
\end{equation}
where the flow $\mathbf{f} \doteq \partial \mathbf{W}(\mathbf{x},t;\btheta) / \partial t$, and Jacobian of warp $\mathtt{J}(\mathbf{x},t;\btheta) \doteq \partial \mathbf{W}(\mathbf{x},t;\btheta) / \partial \mathbf{x}$, are the space-time derivatives of $\mathbf{W}$.
Just like \eqref{eq:IWE}, $\cD$ and $\cA$ are used to create images of average divergence and area deformation per pixel.
Finally, \cite{Shiba22sensors} computes $\mathcal{R}$ as the trimmed mean of such images.
\subsection{Proposed Motion-based Regularizer}
\label{sec:method:proposedReg}
Although \cite{Shiba22sensors} successfully mitigates overfitting, it comes at a computational cost.
The complexity of these regularizers is $O(\numEvents + \numPixels)$ because \eqref{eq:regSensors} depends linearly on the number of events $\numEvents$ and the resulting average images have $\numPixels$~pixels.
This extra complexity makes the whole pipeline more than twice slower than the original (unregularized) CMax framework, whose complexity is also $O(\numEvents + \numPixels)$ \cite{Gallego18cvpr}.
Not only the computational complexity is a burden, but also the fact that \eqref{eq:regSensors} are measured relative to a single reference time.
For example, $\cA(\cE, \btheta)$ increases as $t_k$ increases, since it measures the area deformation \emph{from $t_k$ to $\tref=t_1$}.
This scaling problem is undesirable because
($i$) events far from $\tref$ contribute more to $\mathcal{R}$ than events closer to $\tref$,
and ($ii$) this effect could be amplified depending on the temporal distribution of the events.
Intuitively, motion fields are well-posed or not (i.e., collapse-enabled) by design, regardless of the event data.
Hence, an ideal regularizer should not depend on the events, but solely on the warp parameters (\cref{fig:method:overview}, blue line).
The main idea of the proposed regularizer is to aggregate differential deformations rather than relative ones.
\Cref{fig:method:absement} shows the geometric interpretation:
$\mathcal{R}$ is obtained as the integral of the rate-of-change of the area element deformation along the space-time point trajectories $(\mathbf{x}(t),t)$ defined by the motion.
\input{floats/fig_area_deformation}
\subsubsection{Collapse-enabled warp with 1~DOF}
\label{sec:method:oneDOF}
To illustrate our approach, consider the simplest example: a 1-DOF motion that approximates the translation of camera along its optical axis~$Z$.
This is a simplified zoom-in/out motion without knowledge of scene depth, as used in \cite{Mitrokhin18iros}.
The warp $\mathbf{W}$ is given by
\begin{equation}
\label{eq:warp:hz}
\mathbf{x}'_k = (1 - t_k h_z)\, \mathbf{x}_k,
\end{equation}
where $\btheta \equiv h_z$, and the coordinate frame is at the center of the image plane. %
For simplicity, $t \in [t_1, t_{\numEvents}]$ is normalized to~$[0,1]$.
Assuming an area element attached to each point of the motion trajectory $\bgamma(t)=(\mathbf{x}(t),t)$ (\cref{fig:method:absement}),
the change of area (i.e., area deformation) from $t$ to $t + \Delta t$ is given by:
\begin{equation}
\label{eq:deform:oneDOF}
|\mJ_{t,t+\Delta t} | \doteq \left| \det \biggl( \frac{d\mathbf{x}(t + \Delta t)}{d\mathbf{x}(t)} \biggr) \right|
= \biggl( \frac{1 - t h_z}{1 - (t + \Delta t) h_z} \biggr)^2.
\end{equation}
The Taylor series expansion of \eqref{eq:deform:oneDOF} at $\Delta t = 0$ is
\begin{equation}
\label{eq:taylor}
|\mJ_{t,t+\Delta t} | = 1 + \left.\frac{d |\mJdelta|}{d \Delta t}\right|_{\Delta t = 0} \Delta t + \cdots
\end{equation}
Since the first term is always 1 (i.e., is trivial), we focus on the second term, which conveys the meaning of ``speed'' of area deformation.
The derivative of \eqref{eq:deform:oneDOF} at $\Delta t = 0$ conveys the \emph{rate of change} or \emph{differential amplification factor} of the area:
\begin{equation}
\label{eq:rateofchange:oneDOF}
\left.\frac{d |\mJdelta|}{d \Delta t}\right|_{\Delta t = 0} = \frac{2 h_z}{1 - t h_z}.
\end{equation}
Finally, the total rate of change of the deformation along the observation time window is
\begin{equation}
\label{eq:regularizer:oneDOF}
\mathcal{R} \doteq \int_0^1 \left.\frac{d |\mJdelta|}{d \Delta t}\right|_{\Delta t = 0} dt \stackrel{\eqref{eq:rateofchange:oneDOF}}{=} %
-2 \log |1 - h_z|.
\end{equation}
The regularizer \eqref{eq:regularizer:oneDOF} is plotted in \cref{fig:regularizer:onedof}.
It solely depends on $\btheta\equiv h_z$ and has computational complexity $O(1)$.
In addition, it is developed from geometric principles, and it is interpretable:
$h_z = 0$ (identity warp) gives $\mathcal{R} = 0$;
$h_z \in (0,1)$ (contraction; collapsing warp) gives large $\mathcal{R} > 0$;
and $h_z < 0$ (expansion warp) gives $\mathcal{R} < 0$.
Moreover, notice that $\mathcal{R}$ behaves like a barrier function, approaching infinity (i.e., large penalty) for values close to $h_z=1$ (maximum allowed contraction before events flip side with respect to the image center).
\input{floats/fig_reg_one_dof}
\subsubsection{Well-posed warp with 2 DOFs}
\label{sec:method:2dof}
The 2-DOF translational motion (feature flow) in image space is a well-posed warp, since collapse never happens because the motion lines are parallel.
The warp is given by $\mathbf{W}(\mathbf{x},t;\btheta) = \mathbf{x} + (t -t_\text{ref}) \btheta$ with $\btheta \equiv (v_x, v_y)^{\top}$, assuming constant velocity $\btheta$ for all pixels.
Translations do not change the area element, i.e., $\mJ_{t,t+\Delta t} = 1$,
hence $\mathcal{R} \equiv \int_0^1 \frac{d|\mathtt{J}|}{d \Delta t} = 0$.
Since the resulting regularizer vanishes, it does not affect the landscape of the composite objective function \eqref{eq:compositeObjective}, as expected.
\subsubsection{Well-posed warp with 3 DOFs}
\label{sec:method:3dof}
Angular velocity estimation of a purely rotating event camera is a common research problem \cite{Gallego17ral,Peng21pami,Nunes21pami,Gu21iccv}.
This 3-DOF motion is another well-posed warp, although it involves small amounts of deformation \cite{Shiba22sensors}.
Assuming a small time interval, the warp is parametrized by the angular velocity $\btheta \equiv \angvel = (\omega_{x},\omega_{y},\omega_{z})^{\top}$.
In calibrated homogeneous coordinates $\mathbf{x}^h$, the point trajectories are described by \cite{Gallego18cvpr}:
\begin{equation}
\label{eq:warp:rotangvel}
\mathbf{x}^{h} (t) \sim \mathtt{R}(\angvel t)\,\mathbf{x}^h (0),
\end{equation}
where $t\in [0, 1]$ is normalized
and $\mathtt{R}(\boldsymbol{\phi}) \doteq \exp(\boldsymbol{\phi}^\wedge)$ is parametrized using exponential coordinates \cite{Barfoot15book,Gallego14jmiv}.
The incremental rotation between $t$ and $t + \Delta t$ yields
\begin{equation}
\label{eq:warp:incrementalrot}
\mathbf{x}^{h} (t + \Delta t) \sim \mathtt{R}(\angvel \Delta t)\,\mathbf{x}^h (t).
\end{equation}
\input{floats/tab_MVSEC}
\input{floats/tab_DSEC}
Hence, the area element at $\mathbf{x}(t)$ deforms according to:
\begin{equation}
\label{eq:deform:threeDOF}
\det \biggl( \frac{d\mathbf{x}(t + \Delta t)}{d\mathbf{x}(t)} \biggr)
= \bigl( \br_{3}^\top (\angvel \Delta t) \mathbf{x}^h(t) \bigr)^{-3},
\end{equation}
where $\br_{3}^\top$ is the third row of $\mathtt{R}(\angvel \Delta t)$ (see \cref{sec:append3dof}).
The derivative of \eqref{eq:deform:threeDOF} at $\Delta t=0$ is given by:
\begin{equation}
\label{eq:rateofchange3DOF}
\left.\frac{d |\mJdelta|}{d \Delta t}\right|_{\Delta t = 0} = 3 \mathbf{x}^{h\top}(t) \angvel^\wedge \mathbf{e}_3.
\end{equation}
Finally, the integral of \eqref{eq:rateofchange3DOF} over the point trajectory (parametrized by the initial point $\mathbf{x}(0)$) is given by:
\begin{equation}
\label{eq:regularizer:threeDOF}
\begin{split}
\mathcal{R}_{\mathbf{x}(0)} & = \int_0^1 \left.\frac{d |\mJdelta|}{d \Delta t}\right|_{\Delta t = 0} dt \\
&= 3 \omega_x \int_0^1 y(t) dt - 3 \omega_y \int_0^1 x(t) dt.
\end{split}
\end{equation}
The integrals in \eqref{eq:regularizer:threeDOF} have units of absement.
To obtain the regularizer~$\mathcal{R}$, we threshold $\mathcal{R}_{\mathbf{x}(0)}$ at $-0.2$ and compute its mean,
which allows small amounts of natural deformation caused by rotation.
Similar to \eqref{eq:regularizer:oneDOF}, \eqref{eq:regularizer:threeDOF} does not depend on the events.
However, in contrast to \eqref{eq:regularizer:oneDOF}, \eqref{eq:regularizer:threeDOF} %
is spatially varying, providing an aggregated deformation map:
it is smaller in the center of the image and larger (in absolute value) in the periphery.
The computational complexity of $\mathcal{R}$ is $O(\numPixels)$, which can be further reduced if only a subset of the pixels is used.
Although 3-DOF rotations involve small deformations, their values \eqref{eq:rateofchange3DOF} are considerably smaller than those of collapse-enabled warps like \eqref{eq:warp:hz}, and $\mathcal{R}$ does not affect the accuracy of the angular velocity estimation (as \cref{sec:experim:rot} will show).
Also, pure rotations around the $Z$ axis $\angvel = (0, 0, \omega_z)^{\top}$ do not change the area, as expected, resulting in $\mathcal{R} = 0$.
\subsubsection{Collapse-enabled warp with 4 DOFs}
\label{sec:method:4dof}
The 1-DOF warp (\cref{sec:method:oneDOF}) is a particular case of the 4-DOF warp in \cite{Mitrokhin18iros,Nunes21pami},
which approximates a freely-moving camera (6 DOFs) by means of a similarity transformation on the image plane.
The scaling parameter $h_z$ of the similarity transformation controls the amount of zoom in/out, i.e., the amount of contraction/expansion of the warp.
Hence, we use \eqref{eq:regularizer:oneDOF} to penalize the amount of contraction.
A mathematical justification is given in Appendix~\ref{sec:appendix}.
Since the purpose of our regularizer is to discourage the collapse while allowing a small amount of natural collapse, we use $\mathcal{R} = \max \left(\alpha, -2 \log |1-h_z|\right)-\alpha$, with margin $\alpha=1$.
\section{Experiments}
\label{sec:experim}
\input{floats/fig_qualitative}
We assess the performance of our regularizer by first showing its effectiveness on collapse-enabled warps,
which naturally appear in driving sequences.
Second, a runtime analysis proves that our proposal is faster than prior work.
Finally, we also demonstrate the effect on rotational sequences, in order to show that the proposed regularizer does not harm well-posed warps.
There is no need to test feature flow since we have proved analytically that the regularizer vanishes for such a warp.
\subsection{Setup: Datasets and Evaluation Metrics}
\subsubsection{Datasets}
\label{sec:experim:datasets}
The \emph{MVSEC} dataset \cite{Zhu18ral} is a standard dataset to evaluate various tasks, such as optical flow estimation \cite{Zhu19cvpr,Gehrig21threedv,Paredes21neurips,Shiba22eccv}.
The dataset consists of events, grayscale frames and IMU data from an event camera (mDAVIS346, $346 \times 260$ pixels \cite{Taverni18tcsii}), camera poses, LiDAR data, and ground truth optical flow provided by \cite{Zhu18rss}.
We use the outdoor sequences, where the event camera is mounted on a car.
Following previous work \cite{Shiba22sensors}, we select several excerpts from the \emph{outdoor\_day1} sequence that have a dominant forward motion, which is reasonably well approximated by collapse-enabled warps such as 1~DOF and 4~DOF cases.
In total, we evaluate on 3.2 million events spanning 10 s.
The \emph{DSEC} dataset \cite{Gehrig21ral} is another recent dataset of driving sequences.
It includes more complex scenes (e.g., moving objects, higher dynamic range) with a higher resolution event camera (Prophesee Gen3, $640 \times 480$ pixels).
Ground truth optical flow is computed as the motion field, using the scene depth from a LiDAR \cite{Gehrig21threedv}.
In total, we evaluate on 380 million events spanning 40 s from the \emph{zurich\_city\_11} sequence.
The \emph{ECD} dataset \cite{Mueggler17ijrr} is widely used to assess camera ego-motion \cite{Gallego17ral,Zhu17cvpr,Rosinol18ral,Gu21iccv,Rebecq17ral,Mueggler18tro,Zhou20tro,Shiba22sensors}.
Each sequence provides events, frames, calibration information, and IMU data from a DAVIS240C camera ($240 \times 180$ pixels \cite{Brandli14ssc}), as well as ground truth camera poses from a motion capture system (at 200Hz).
We use \emph{boxes\_rotation} and \emph{dynamic\_rotation} sequences for 3-DOF rotational motion estimation, to consistently compare with previous work.
In total, we evaluate on 43 million events (10 s) of the box sequence, and on 15 million events (11 s) of the dynamic sequence.
\subsubsection{Metrics}
\label{sec:experim:metrics}
Optical flow accuracy (for MVSEC and DSEC experiments) is given by the Average Endpoint Error (AEE)
and the percentage of pixels with AEE greater than $N$ pixels (``$N$PE''), for $N=\{3,10,20\}$.
They are calculated only in pixels with valid ground truth.
We also adopt the FWL metric \cite{Stoffregen20eccv}, which is defined as the relative variance of the IWE with respect to that of the identity warp.
The FWL measures the IWE sharpness: FWL~$<1$ means that the estimation is worse than the zero-flow baseline, while FWL~$>1$ implies that the resulting IWE is sharper than the baseline.
Rotational motion accuracy is assessed as the RMS error of angular velocity estimation, following previous works \cite{Gallego19cvpr,Nunes21pami,Gu21iccv}.
Angular velocity is assumed to be constant during the time window of events, and compared with the IMU's gyroscope value (ground truth) at the midpoint: $(t_1 + t_{\numEvents}) / 2$.
We also use the FWL metric to measure the IWE sharpness.
The estimation time window spans: $dt=4$ grayscale frames (at $\approx$ 45Hz) in MVSEC (standard for MVSEC benchmark), 500k events for DSEC, and 30k events for ECD dataset, respectively.
For runtime comparison we use a fixed number of events (30k for MVSEC and 500k for DSEC), because the runtime depends on the number of events (e.g., $O(\numEvents + \numPixels)$).
We set $\lambda$ in \eqref{eq:compositeObjective} as follows:
if the data term is the IWE variance, $\lambda=1.0$ for MVSEC and ECD experiments, and $\lambda=5.0$ for DSEC experiments;
if the data term is the squared magnitude of the IWE gradient, $\lambda=0.2$ for MVSEC and $\lambda=1.0$ for DSEC experiments.
The optimization algorithm is the Tree-Structured Parzen Estimator (TPE) sampler \cite{Bergstra11nips}.
\subsection{Results on Collapse-Enabled Warps}
\label{sec:experim:main}
\Cref{tab:main_mvsec,tab:main_dsec} report the results of collapse-enabled warp experiments (1 and 4~DOFs) on MVSEC and DSEC, respectively.
They report the flow AEE, $N$PE, and FWL for the two data terms: image variance and the squared magnitude of the IWE gradient (``Gradient Magnitude'').
Throughout the experiments, both metrics are considerably high in the results of the original CMax (``No regularizer'') and whitening \cite{Nunes21pami} methods.
This indicates that the warp overfits to events (event collapse).
On the other hand, our regularizer produces better AEE values and moderately higher FWL than $1$.
These results clearly show that our regularizer successfully discourages event overfitting while producing sharper IWEs than the identity warp.
Our results are competitive with those by \cite{Shiba22sensors}, and we do not find significant accuracy differences between them.
Qualitative results are shown in \cref{fig:main_compare} (MVSEC and DSEC rows).
Our regularizer (last column) provides the best IWEs, which reveal the sharp edges of the scene, while the IWEs without regularizer produce event collapse (second column).
Although $\mathcal{R}$ in \eqref{eq:regularizer:oneDOF} is a scalar, we visualize it as an image to compare it with the area deformation map from \cite{Shiba22sensors}.
Notice that the area deformation map \cite{Shiba22sensors} shows collapse \emph{only} at pixels with warped events,
while our regularizer provides dense maps (even in pixels with no events, corresponding to homogeneous brightness regions) because it is purely geometric, based on the motion parameters.
\subsection{Runtime Comparison}
\label{sec:experim:runtimes}
\Cref{tab:runtime} reports the runtime comparison of the methods, notably with respect to the original CMax (``No regularizer'').
We use Python (3.9.12) on a CPU (Mac M1 2020, 8 Cores), and average the runtime over 400 trials.
The whitening technique \cite{Nunes21pami} is slower than the original CMax (``No regularizer'').
The runtime difference is due to an extra SVD step on the events, which is more noticeable ($2\times$ slower) in the DSEC dataset than in MVSEC because it uses more events.
The ``Deformation'' regularizer in \cite{Shiba22sensors} is also two to three times slower than the original CMax.
When both regularizers in \cite{Shiba22sensors} are combined (``Div. + Def.''), the runtime becomes even larger.
Finally, our regularized approach has almost the same runtime as the original CMax, since its complexity is $O(1)$, thus it is two to four times faster than competing methods.
\Cref{fig:runtime} visualizes the accuracy and runtime of the methods (on DSEC data).
Runtime is reported relative to the ``No regularizer'' case.
It clearly shows that the proposed regularizer is the only effective approach against event collapse that does not compromise the speed of the CMax framework.
\input{floats/tab_runtime}
\input{floats/fig_runtime}
\subsection{Results on Well-Posed Warps}
\label{sec:experim:rot}
To confirm that the proposed regularizer does not harm well-posed warps (e.g., 3-DOF rotational motion), we report results on the ECD dataset in \cref{tab:main_ecd}.
We use the variance as data fidelity and the Adam optimizer.
In both rotational sequences, the results of ``No regularizer'' and ours produce very similar RMS and FWL values.
This is because $\mathcal{R}$ values are small and do not affect the landscape of the objective function.
As qualitatively shown in \cref{fig:main_compare},
for the 3~DOF experiments
the absolute values of $\mathcal{R}_{\mathbf{x}(0)}$ are approximately less than $0.1$,
while for the collapse-enabled warps they are larger than $10$ (i.e., two orders of magnitude difference).
\input{floats/tab_ECD}
\subsection{Application: Time-to-contact}
\label{sec:ttc}
The parametrization of collapse-enabled warps has useful implications toward future application on intelligent vehicles, such as advanced driver-assistance system (ADAS).
Let us introduce another interpretation of the parameter $h_z$.
For a freely moving camera with linear and angular velocities $\linvel$ and $\angvel$, respectively, the apparent velocity $\mathbf{v}(\mathbf{x})$ on the image plane of a 3D point $\mathbf{X}=(x,y,\depthx)^\top$ (at depth $\depthx$ with respect to the camera)
can be computed using the $2\times 6$ feature sensitivity matrix \cite{Corke17book}:
\begin{equation}
\label{eq:featureSensitivity}
\mathbf{v}(\mathbf{x}) =
\begingroup %
\setlength\arraycolsep{3pt} %
\begin{pmatrix}
\frac{-1}{\depthx} & 0 & \frac{x}{\depthx} & xy & -(1+x^2) & y \\[1ex]
0 & \frac{-1}{\depthx} & \frac{y}{\depthx} & 1+y^2 & -xy & x \\
\end{pmatrix}
\endgroup
\!\!
\begin{pmatrix}
\linvel\\[1ex]
\angvel\\
\end{pmatrix},
\end{equation}
which can be used to warp events:
\begin{equation}
\label{eq:warp:flowhz}
\mathbf{x}_k' = \mathbf{x}_k - \mathbf{v}(\mathbf{x}) t_k.
\end{equation}
Assuming a vehicle with body-frame velocity $v_z$,
i.e., $\linvel \equiv (0, 0, v_z)^{\top}$, $\angvel \equiv (0, 0, 0)^{\top}$,
the motion field \eqref{eq:featureSensitivity} becomes $\mathbf{v}(\mathbf{x}) = (v_z / \depthx) \,\mathbf{x}$,
and substituting in \eqref{eq:warp:flowhz} gives
$\mathbf{x}_k' = (1 - v_z/\depthx ) t_k.$
Comparing this expression to \eqref{eq:warp:hz},
and assuming $\depthx$ is spatially invariant, we identify
\begin{equation}
\label{eq:ttc}
h_z = \frac{v_z}{Z},
\end{equation}
i.e., the parameter $h_z$ is inverse of the time-to-contact or time-to-collision (TTC) \cite{Clady14fns}. %
\input{floats/fig_ttc}
\Cref{fig:ttc} shows two examples of TTC from the MVSEC dataset.
It is remarkable that this 1-DOF warp model can be related to the popular concept in ADAS,
and our regularizer plays an important role toward real-time computation of TTC given its runtime.
Also note that \eqref{eq:ttc} establishes a relation between TTC, vehicle speed and scene depth, and that the TTC can be used to estimate the scene depth given the vehicle speed,
or vice versa, the vehicle speed given the scene depth.
We hope this connection helps future implementation of event-camera application in collision avoidance systems.
\section{Limitations}
\label{sec:limitations}
As in many regularized problems, the regularizer weight $\lambda$ is empirically set.
It depends on the paired objective function and on the data, i.e., on the scene.
The proposed method also has a heuristically selected threshold (margin) that allows us to account for small natural deformations of the motion field.
It would be desirable to develop automatic strategies to balance both data fidelity and regularizing terms (including the threshold) for best optimization convergence and results.
While we have obtained formulas to measure the rate of change of area deformation along point trajectories for low-DOF motions (up to 8-DOF homographies in Appendix~\ref{sec:appendix}),
the regularizer requires aggregation in the form of integrals (e.g., \eqref{eq:regularizer:threeDOF}),
which are evaluated via numerical integration.
Extending the ideas in this work to higher DOFs (e.g., optical flow) and developing efficient numerical approaches is an important research direction.
Our current implementation of the regularized CMax framework is not real time.
This is especially relevant as the spatial resolution of event cameras increases (VGA size \cite{Gehrig21ral} and 1Mpixel event cameras \cite{Finateu20isscc}), which also increases the number of events to be processed.
Therefore, in the future it will be important to speed up the method (both the data fidelity term and the regularizer) to enable interactive applications.
Finally, while in the time-to-contact application (\cref{sec:ttc}) we established the connection between scene depth and vehicle speed, the example warp is 1-DOF (assumes a single depth for the whole scene), which may be an over-simplification.
The warp does not consider where events happen from a user perspective (e.g., on the road or on the side of the road).
A more elaborate model would consider the location of the events and decide whether there will be an actual contact along the predicted vehicle trajectory.
\section{Conclusion}
\label{sec:conclusion}
We proposed a novel regularizer to mitigate event collapse in the CMax framework, based on aggregating differential deformations of the motion field.
The experimental results show its efficacy, achieving on-par state-of-the-art accuracy in low-DOF motion estimation problems.
Furthermore, the proposed regularizer is the only effective approach to date against event collapse that does not compromise the runtime of the CMax framework.
Since the low-DOF motion estimation forms a foundation of more complex motion models,
it would be important to analyze more complex warps, such as dense optical flow, to alleviate event collapse effectively and efficiently.
We hope this work encourages future research towards this paramount application of the CMax framework that leverages the advantages of event cameras.
\appendices
\section{}
\label{sec:appendix}
In homogeneous coordinates, a homographic warp $\mathbf{W}$ is given by \cite{Gallego18cvpr}
\begin{equation}
\label{eq:homographic:warp}
\mathbf{x}^{h\prime}_k \sim \mH^{-1}(t_k; \btheta)\,\mathbf{x}^h_k,
\end{equation}
and the point trajectories are represented by
\begin{equation}
\label{eq:homographic:motion}
\mathbf{x}^{h} (t) \sim \mH(t; \btheta)\,\mathbf{x}^h (0).
\end{equation}
Hence, the differential transformation from $t$ to $t+\Delta t$ is given also by a homography $\mH_{t,t+\Delta t}$:
\begin{equation}
\label{eq:homographic:increment}
\mathbf{x}^{h} (t+\Delta t) \sim \underbrace{\mH(t+\Delta t; \btheta)\, \mH^{-1}(t; \btheta)}_{\mH_{t,t+\Delta t}} \mathbf{x}^h (t).
\end{equation}
Using Result 1 in Appendix A of \cite{Rebecq18ijcv}, the determinant of the Jacobian $\mathtt{J}$ of the transformation (from $t$ to $t+\Delta t$) in Euclidean coordinates is
\begin{equation}
\label{eq:homography:detJ}
\det\bigl(\mJ_{t,t+\Delta t}\bigr) = \frac{\det( \mH_{t,t+\Delta t} )}{(\mathbf{e}_3^\top \mH_{t,t+\Delta t} \,\mathbf{x}^h(t))^3},
\end{equation}
where $\mathbf{e}_3=(0,0,1)^\top$, $\mathbf{x}^h(t)=(x(t),y(t),1)^\top$.
The 8-DOF motion admits several particular cases.
\subsection{Rotational Camera Motion}
\label{sec:append3dof}
Here, $\mH(t;\btheta) \equiv \mathtt{R}(t\angvel)$ in \eqref{eq:homographic:motion} is a $3\times 3$ rotation matrix.
Rotation matrices have unit determinant and simplify:
$\mH_{t,t+\Delta t} %
= \mathtt{R}((t+\Delta t)\angvel) \, \mathtt{R}^{-1}(t \angvel)
= \exp\bigl(((t+\Delta t)\angvel)^\wedge\bigr) \exp\bigl((-t \angvel)^\wedge\bigr)
= \exp((\Delta t\, \angvel)^\wedge)
= \mathtt{R}(\angvel \Delta t)$.
This holds because the rotation axis $\angvel$ is unique.
Hence \eqref{eq:homography:detJ} becomes:
\begin{equation}
\det\bigl(\mJ_{t,t+\Delta t}\bigr) = \frac1{(\mathbf{e}_3^\top \mathtt{R}(\angvel \Delta t) \,\mathbf{x}^h)^3},
\end{equation}
which is \eqref{eq:deform:threeDOF}.
Computing the derivative yields \eqref{eq:rateofchange3DOF}:
\begin{align}
\left.\frac{d |\mJdelta|}{d \Delta t}\right|_{\Delta t = 0} &= \frac{-3 \mathbf{x}^{h\top}(t)}{\bigl( \br_{3}^\top (\angvel \Delta t) \mathbf{x}^h(t) \bigr)^{4}}
\, \frac{d}{d \Delta t} \br_3 (\angvel \Delta t) \bigg|_{\Delta t = 0} \nonumber \\[0.8ex]
&= 3 \mathbf{x}^{h\top}(t) \angvel^\wedge \mathbf{e}_3,
\end{align}
because
\begin{equation}
\begin{split}
\frac{d}{d \Delta t} \br_3 (\angvel \Delta t) &= \frac{d}{d \Delta t} \mathtt{R}^{\top} (\angvel \Delta t) \mathbf{e}_3 \\
&\approx \frac{d}{d \Delta t} \bigl( \mId - (\angvel \Delta t)^\wedge \bigr) \mathbf{e}_3 \\
&= - \angvel^\wedge \mathbf{e}_3 \\
&= (-\omega_y, \omega_x, 0)^{\top}.
\end{split}
\end{equation}
\subsection{Similarity Transformation on the Image Plane, $\displaystyle \operatorname {Sim}(2)$}
A planar similarity transformation has 4~DOFs and can be parametrized using linear image velocity, angular velocity around $Z$, and scaling speed on the image plane, $\btheta = (\mathbf{v}, \omega_z, s)^\top$ \cite{Nunes21pami}.
The point trajectories are given by \eqref{eq:homographic:motion} with
\begin{equation}
\mH^S (t;\btheta) \doteq
\begin{pmatrix}
\beta(t;s) \mathtt{R}(t \omega_z) & t \mathbf{v}\\
\mathbf{0}^\top & 1
\end{pmatrix}.
\end{equation}
Using \eqref{eq:homographic:increment} it gives
$\mH_{t,t+\Delta t} = \mH^S (t+\Delta t; \btheta)\, (\mH^S)^{-1}(t; \btheta)$.
Similarities form a matrix Lie group, hence the inverse and the product of two similarities is also a similarity.
Since $\mH_{t,t+\Delta t}$ is a similarity, its third row is $\mathbf{e}_3^\top \mH_{t,t+\Delta t} = (0,0,1)$, which makes the denominator in \eqref{eq:homography:detJ} equal to one.
Substituting in \eqref{eq:homography:detJ} produces
\begin{equation}
\label{eq:deform:fourDOFsim}
\begin{split}
|\mJ_{t,t+\Delta t} |
&= \left| \det \bigl( \mH^S (t+\Delta t; \btheta)\, (\mH^S)^{-1}(t; \btheta) \bigr) \right|\\[0.5ex]
&= \left| \frac{\det \bigl(\mH^S (t+\Delta t; \btheta) \bigr)}{\det \bigl( \mH^S (t; \btheta) \bigr)} \right|\\
&= \left(\frac{\beta(t+\Delta t;s)}{\beta(t;s)}\right)^2.
\end{split}
\end{equation}
Two intuitive remarks about \eqref{eq:deform:fourDOFsim}:
($i$) it only depends on the scaling DOF (i.e., one out of the four DOFs) since $\omega_z$ and $\mathbf{v}$ do not appear;
and ($ii$) it can be used to derive the 1-DOF formula \eqref{eq:deform:oneDOF}:
the 1-DOF scaling transformation is modeled by $\mH^S$ with $\omega_z = 0, \mathbf{v}=\mathbf{0}$ and $\beta(t; h_z) = (1-t h_z)^{-1}$.
Substituting this choice of $\beta$ in \eqref{eq:deform:fourDOFsim} makes it coincide with \eqref{eq:deform:oneDOF}.
The 4-DOF transformation in \cite{Mitrokhin18iros} has a similar geometric meaning but a different parametrization.
Hence, we use the above result and penalize collapse by means of the corresponding scaling parameter in \cite{Mitrokhin18iros}.
\subsection{Affine Transformation on the Image Plane, $\displaystyle \operatorname {Aff}(2)$}
A planar affine transformation has 6 DOFs in $\btheta$.
Letting
\begin{equation}
\mH^A (t;\btheta) \doteq
\begin{pmatrix}
\mathtt{A}(t;\btheta) & t \mathbf{b} \\
\mathbf{0}^\top & 1
\end{pmatrix},
\end{equation}
and using \eqref{eq:homographic:increment} gives
$\mH_{t,t+\Delta t} = \mH^A (t+\Delta t; \btheta)\, (\mH^A)^{-1}(t; \btheta)$.
Affinities also form a matrix Lie group, hence $\mH_{t,t+\Delta t}$ is an affinity.
Moreover, its third row is also $\mathbf{e}_3^\top \mH_{t,t+\Delta t} = (0,0,1)$, and
following similar steps as those in \eqref{eq:deform:fourDOFsim} yields
\begin{equation}
\label{eq:deform:affine}
\begin{split}
|\mJ_{t,t+\Delta t} |
&= \left| \det \bigl( \mH^A (t+\Delta t; \btheta)\, (\mH^A)^{-1}(t; \btheta) \bigr) \right|\\[0.5ex]
&= \left| \frac{\det \bigl(\mathtt{A} (t+\Delta t; \btheta) \bigr)}{\det \bigl( \mathtt{A} (t; \btheta) \bigr)} \right|.
\end{split}
\end{equation}
Notice that the $2\times 2$ matrix $\mathtt{A}$ includes not only a scaling parameter but also a shear, which affects the area deformation.
\section*{Acknowledgments}
This research was funded by the German Academic Exchange Service (DAAD), Research Grant-Bi-nationally Supervised Doctoral Degrees/Cotutelle, 2021/22 (57552338) and the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under Germany’s Excellence Strategy – EXC 2002/1 ``Science of Intelligence'' – project number 390523135.
\balance
\bibliographystyle{IEEEtran}
\input{main.bbl}
\end{document}
|
\section{Introduction}
Corrosion of metallic interfaces leads to weakening of their structural integrity and has been widely estimated to cost billions of dollars annually \cite{koch2002corrosion}.
While this has inspired numerous studies on corrosion and strategies to protect against it, the fundamental atomistic mechanisms behind corrosion are difficult to identify for any given system.
Corrosion in aluminum provides an excellent example of the issues.
In aluminum, it has been experimentally demonstrated that pitting corrosion begins at potentials $\Phi > -0.50$ V relative to the standard hydrogen electrode (SHE), dependent on the salt and pH concentration of the surrounding solution \cite{dibari1971electrochemical,bessone1992eis}.
The exact mechanism that causes this pitting corrosion is still under debate, however.
Pitting corrosion is often described using the point defect model \cite{lin1981point,chao1981point,engelhardt2004unification}, which states that corrosion is initially driven by the formation of charged defects in the surface of the material, of which chloride ions are particularly damaging \cite{frankel1998pitting,natishan20182017}.
On aluminum surfaces, a several nm thick aluminum oxide layer forms \cite{evertsson2015thickness} and the insertion of chloride into oxygen vacancies in this layer has been theorized as one of the main mechanisms leading to corrosion.
Experiments have investigated the dependence of pitting corrosion on chlorine concentration in the surrounding electrolyte and other factors \cite{brown1973use,shimizu1999novel,serna2006critical}, but the results have not been fully conclusive as to the causal mechanism.
While many previous density functional theory (DFT) studies have examined Al/Al$_2$O$_3$ interfaces at an atomic level \cite{kim2013nature,batyrev2001plane,koberidze2018structural,costa2014atomistic,siegel2002adhesion}, they have typically not included oxygen vacancies and chloride insertion.
For those that have examined vacancies \cite{weber2009point,carrasco2004theoretical,hine2009supercell,janetzko2004first}, they do not include Al metal in the calculation.
Previous DFT studies on chloride insertion into Al/Al$_2$O$_3$ interfaces that have included contributions from the metal, the oxide, vacancies, and chloride ions \cite{liu2019dft,liu2021density,leung2021first} have either implicitly limited their scope to defects with no charge, or tested only a small range of voltages.
This lack of study of charged ions in the full electrochemical environment stems from the difficulty of needing to include several different length scales in a simulation to get a realistic picture of corrosion behavior.
Modeling the realistic interaction of point defects with the surface requires a quantum mechanical treatment, with DFT providing a convenient compromise between accuracy and speed.
DFT is typically limited, however, to simulating $\sim$1000s of atoms as larger systems are computationally prohibitive.
Since a realistic model of an interface, including the surrounding solution, can be 10s to 100s of nm wide, DFT simulations must necessarily focus on a single region of interest, typically the surface.
The entire surrounding electrode helps determine the electric field (and therefore voltage) that the region of interest operates at.
Simulations of these isolated regions therefore do not provide information as to the voltage the electrode is under and, in effect, operate at ``flatband'' conditions where the electric field is flat across the system.
This necessarily limits the scope of traditional DFT calculations when examining these interfaces and defects.
Since the defects under examination are charged, it is crucial to incorporate a realistic model of voltage into DFT calculations.
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{fig1.png}
\caption{A schematic illustration of the Quantum Continuum Approximation applied to a metal-insulator interface immersed in solution.}
\label{fig:qca-schematic}
\end{figure*}
In this paper, we develop a Quantum Continuum Approximation (QCA) methodology that allows for the first-principles simulation of the metal-insulator interfaces at controlled voltages using limited DFT simulation cell sizes.
QCA self-consistently couples DFT calculations of the metal-insulator interface and the insulator-solution interface with Poisson-Boltzmann distribution of charges in the bulk insulator and solution to provide an equilibrium potential drop for the entire electrode, as schematically illustrated in Fig.~\ref{fig:qca-schematic}.
By reducing the number of atoms needed for DFT calculations to include voltage information by at least an order of magnitude, QCA enables voltage-dependent first-principles simulation of realistic electrochemical interfaces at reasonable computational cost.
We demonstrate this methodology for the simulation of chloride insertion into the aluminum oxide layer of an Al(111)/$\alpha$-Al$_2$O$_3$(0001) system as a function of voltage.
We predict that insertion is not favorable for Al/Al$_2$O$_3$ interfaces that do not include grain boundaries at any voltage.
We do see, however, that in systems with grain boundaries in the Al$_2$O$_3$, chloride insertion is favorable at $\Phi > -0.2$ V (SHE), in reasonable agreement with experiment and helping confirm the importance of grain boundaries in the corrosion process.
This work provides a template for attaching a potentiostat to electrochemical DFT calculations of interfaces.
\section{Methods}
\subsection{Chloride insertion energy calculations}
We calculate the formation energy of a Cl$^-$ ion inserting into an oxygen vacancy as
\begin{equation}
\Delta E_{\text{Cl-O}}[\Phi] = E_{\text{DFT}}(\text{Cl ins.})[\Phi] - E_{\text{DFT}}(\text{O vac.})[\Phi] - E_{\text{Cl}^-}[\Phi] ,
\label{eq:total-e}
\end{equation}
where $E_{\text{DFT}}(\text{O vac.})[\Phi]$ is the DFT energy of the slab with the oxygen vacancy as a function of the global potential $\Phi$ of the system, $E_{\text{DFT}}(\text{Cl ins.})[\Phi]$ is the DFT energy of the slab with a chlorine ion inserted into the previous location of the oxygen vacancy as a function of the global potential, and $E_{\text{Cl}^-}[\Phi]$ is the energy of the Cl$^-$ ion in solution as a function of voltage.
For the chloride ion, we follow a computational hydrogen electrode-like approach \cite{norskov2004origin,rossmeisl2007electrolysis,man2011universality}, where we determine the energy of the chloride ion as
\begin{equation}
E_{\text{Cl}^-}[\Phi] = \mu (\text{Cl}^-) + e\Phi^{\text{SHE}}.
\label{eq:chloride-e}
\end{equation}
Here, $\mu (\text{Cl}^-)$ is the equilibrium chemical potential of the chloride ion, and $\Phi^{\text{SHE}}$ is the potential that the electrode is operating at, referenced to the standard hydrogen electrode (SHE).
We determine the chemical potential by referencing the following chemical reaction
\begin{equation}
\text{Cl}_2 (\text{g}) + 2 e^- \rightarrow 2\text{Cl}^-(\text{aq}),
\end{equation}
which is at equilibrium at $\Phi_{\text{Cl}_2/\text{Cl}^-} = 1.36$ V (SHE) \cite{atkins2006physical}.
Using this reaction, we can then calculate the chemical potential of a chloride ion as
\begin{equation}
\mu (\text{Cl}^-) =\frac{1}{2} E_{\text{DFT}}(\text{Cl}_2(\text{g})) - e\Phi_{\text{Cl}_2/\text{Cl}^-}.
\end{equation}
With this energetic relationship established, we can now calculate the formation energy of chloride insertion into the Al/Al$_2$O$_3$ surface as a function of potential the electrode is under.
For this relationship to be strictly accurate, however, the calculated DFT energies of the slabs need to be taken at simulation conditions consistent with $\Phi$ for the electrode.
We outline how we can achieve this using the QCA methodology in the next section.
\subsection{Quantum Continuum Approximation (QCA) Methodology for Predicting Voltage at an interface}
\label{subsec:qca}
\subsubsection{Interface Energetics and Electrostatics}
At a fundamental level, our approach to calculations with metal-semiconductor-solution interfaces works by assigning a global potential to the system.
This potential drop can then be split up between several different voltage drop regions within the interface.
Finding the correct distribution of potential drops requires a self-consistent process for minimizing the free energy of the system while taking into account the electrostatic requirements of following the Poisson equation and maintaining electronic equilibrium throughout the interface.
To accomplish this task, we employ a new extension of the Quantum Continuum Approximation (QCA) methodology that was first developed by Campbell and Dabo for semiconductor-solution interfaces \cite{campbell2017quantum,campbell2019voltage}, which allowed for the voltage of the system to be additionally predicted by DFT slab calculations.
It was later extended to solid-state metal-semiconductor interfaces, providing predictions of Schottky barriers that corresponded well with nano-ARPES measurements \cite{subramanian2020photophysics}.
Here, we extend this to include metal-insulator or metal-semiconductor interfaces that are also immersed in solution, which are common in electrochemical applications.
The basic goal is to self-consistently couple DFT calculations of an interface with analytic Poisson-Boltzmann descriptions of the bulk semiconductor and solution regions.
As shown in Fig.~\ref{fig:qca-schematic}, we perform DFT calculations only on the crucial interface sections, such as the metal-metal oxide transition and the metal oxide-solution transition, or around atomic defects in the system.
We then apply Poisson-Boltzmann calculations to the oxide regions in between the DFT calculations and the surrounding solution.
The free energy of the system is
\begin{equation}
\mathcal{F} = \mathcal{F}^{\text{m}|\text{ox.}} + \mathcal{F}^{\text{ox.}} + \mathcal{F}^{\text{ox.}|\text{sol.}} + \mathcal{F}^{\text{sol.}} - \frac{1}{2}\int dr \epsilon_0 \epsilon(\mathbf{r}) |\nabla(\phi(\mathbf{r}))|^2 ,
\end{equation}
where $\epsilon_0$ is the vacuum permittivity, $\epsilon(\mathbf{r})$ is the dielectric constant as a function of position, $\phi$ is the electrostatic potential, and $\mathcal{F}^{\text{m}|\text{ox.}}$ and $\mathcal{F}^{\text{ox.}|\text{sol.}}$ are the free energies of the metal-oxide and oxide-solution interface, respectively, which can be calculated using the typical calculations for DFT energy.
Similarly, $\mathcal{F}^{\text{ox.}}$ and $\mathcal{F}^{\text{sol.}}$ are the contributions of the bulk oxide and solution regions, which can be calculated as analytic functions of the distribution of charge carriers within each system.
This decomposition of the free energy naturally leads to the following Poisson equation for the system:
\begin{equation}
\nabla(\epsilon_0\epsilon(\mathbf{r})\nabla \phi(\mathbf{r})) = p(\mathbf{r}) - n(\mathbf{r}) -p_d(\mathbf{r}) + n_d(\mathbf{r}) + c_+(\mathbf{r}) - c_-(\mathbf{r}) + \rho_+(\mathbf{r}) -\rho_-(\mathbf{r}),
\label{eq:electrostatics}
\end{equation}
where and $\rho_-(\mathbf{r})$ is the electron density obtained from the self-consistent Kohn-Sham equation for a given distribution $\rho_+(\mathbf{r})$ of atomic cores, and the source terms can be expressed as
\begin{equation}
n(\mathbf{r}) = n_d(\mathbf{r}) \left[1 + \exp\left(\frac{\phi(\mathbf{r}) + \epsilon_c - \epsilon_F}{k_{\text{B}}T}\right)\right]^{-1},
\end{equation}
\begin{equation}
p(\mathbf{r}) = p_d(\mathbf{r}) \left[1 + \exp\left(\frac{\epsilon_F-\epsilon_v-\phi(\mathbf{r})}{k_{\text{B}}T}\right)\right]^{-1},
\end{equation}
\begin{equation}
c_{\pm}(\mathbf{r}) = c^{\circ}(\mathbf{r}) \exp\left(\mp \frac{\phi(\mathbf{r})}{k_{\text{B}}T}\right).
\end{equation}
Here, $\epsilon_F$ denotes the Fermi energy, $\epsilon_v$ is the energy at the valence band maximum, $\epsilon_c$ is the energy at the conduction band minimum, $n_d(\mathbf{r})$ is the concentration of $n$-type carriers in the semiconducting/insulating region and is 0 outside this region, $p_d(\mathbf{r})$ is similarly the concentration of $p$-type carriers within the semiconducting regions, and $c^{\circ}(\mathbf{r})$ is the equilibrium ionic concentration inside the electrolyte region and is 0 outside of the electrolyte region.
In an $n$-type semiconducting region, the charge carriers can be solved for using the Boltzmann distribution as
\begin{equation}
p(\mathbf{r}) = p_d = 0,
\end{equation}
\begin{equation}
n(\mathbf{r}) = n_d\exp\left(\frac{\phi_0 - \phi(\mathbf{r})}{k_{\text{B}}T}\right),
\end{equation}
where $\phi_0$ is the asymptotic value of the potential in this region.
Analogous equations can be written for $p$-type semiconducting regions.
\subsubsection{Process for solving electrostatics}
To practically solve Eq.~\ref{eq:electrostatics} for a given potential, we start by dividing the system into two portions that we will simulate with DFT separately.
The first is the surface of the interface, i.e. where the passivating oxide layer meets the solution.
We will refer to this region with a superscript ox.$|$sol. in variables throughout this work.
The other DFT region is the interface of the metal with the passivating oxide, referred to as a superscript m$|$ox. within this work.
For both of these regions we start by doing DFT calculations on these systems when no charge has been applied.
This corresponds to the so-called `flatband' conditions.
We then measure both the planar averaged potential profile of these systems $\phi_{\text{fb}}^{\text{m}|\text{ox.}}(z)$ and $\phi_{\text{fb}}^{\text{ox.}|\text{sol.}}(z)$, and the Fermi level of these systems $\varepsilon_{\text{F,fb}}^{\text{m}|\text{ox.}}$ and $\varepsilon_{\text{F,fb}}^{\text{ox.}|\text{sol.}}$.
Next we assign a total charge $Q$ to the electrode.
The electrode is here defined as the region from the metal to the interface with the solution.
The distribution of this charge $Q$ within the interface mechanically determines the total potential drop via Gauss' law.
For a perfectly insulating/semiconducting system, the charge distribution can be well approximated by the following Poisson--Boltzmann relationship \cite{schmickler2010interfacial}:
\begin{equation}
\frac{d^2\phi}{dz^2} = \frac{n_{\text{d}}}{\epsilon_o \epsilon_{\text{sc}}} \left[ 1 - \exp \left( \frac{-\phi}{k_{\text{B}}T}\right)\right].
\label{eq:pb-ms-conditions}
\end{equation}
Here, $n_{\text{d}}$ is the concentration of charge carriers (dopants) within the semiconductor, $\epsilon_o$ is the vacuum permittivity, $\epsilon_{\text{sc}}$ is the dielectric constant of the semiconductor, and $k_{\text{B}}$ is the Boltzmann constant.
For most real-world systems, however, surface states will form at both the metal-insulator and insulator-solution interface \cite{bardeen1947surface,cowley1965surface,bard1980role}.
These surface states result from the abrupt nature of the interface disrupting the symmetry of the material, and from defects induced by interaction.
These surface states trap charge at the interface, thus reducing the charge throughout the remainder of the insulator and typically reducing the potential drop of the electrode.
Capturing these surface states requires a quantum mechanical treatment of the interfaces using, e.g., DFT.
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{fig-interface-potentials.pdf}
\caption{A schematic illustration of the potential profile of a metal-oxide-solution interface within QCA with important variables depicted. Distances and potentials shown are approximate.}
\label{fig:qca-potentials}
\end{figure*}
To simulate charge accumulation at these surface interfaces we place Helmholtz planar counter-charges on either side of a DFT slab, as illustrated in Fig.~\ref{fig:qca-potentials}.
We place the Helmholtz planes 4 \AA\ away from the surfaces of the slab to mimic the Debye length.
We obey charge neutrality for the DFT systems simulated, such that
\begin{equation}
q_{\text{h}}^{\text{ox.}|\text{sol.}} + q_{\text{DFT}}^{\text{ox.}|\text{sol.}} + q_{\text{h}}^{\text{sol.}} = 0
\label{eq:chg-neutrality-ox-sol}
\end{equation}
at the oxide-solution interface, where $q_{\text{h}}^{\text{ox.}|\text{sol.}}$ and $q_{\text{h}}^{\text{sol.}}$ is the charge on the Helmholtz planes on the oxide and solution side of the DFT slab, respectively.
Additionally, $q_{\text{DFT}}^{\text{ox.}|\text{sol.}}$ is the charge on the DFT slab representing the interface between the oxide and solution.
Since the electrode charge is defined as including the metal-oxide interface, the continuum oxide, and the oxide-solution interface, we can automatically assign the charge on the Helmholtz plane representing the solution as the opposite of this electrode charge, \textit{i.e.} $q_{\text{h}}^{\text{sol.}} = -Q$.
This arrangement of charge creates a potential gradient across the system.
We measure this as $\bar{\phi}^{\text{ox.}|\text{sol.}}(z) = \phi^{\text{ox.}|\text{sol.}}_{q} - \phi^{\text{ox.}|\text{sol.}}_{\text{fb}}$, where $\phi^{\text{ox.}|\text{sol.}}_{q}$ is the planar averaged potential of the charged system and $\bar{\phi}^{\text{ox.}|\text{sol.}}(z)$ is the difference in the planar averaged potential due to the charge.
We can finally couple this DFT potential to a bulk charge distribution by extracting the derivative of the subtracted potential $\frac{d\bar{\phi}}{dz}$ at a specific cutoff point $z_{\text{edge}}^{\text{ox.}|\text{sol.}}$.
In DFT, our oxide will necessarily only extend for a few layers, but in reality it should extend for at least several nanometers.
This nonphysical oxide surface creates spurious surface effects on the potential.
By setting the cutoff value for the potential $z_{\text{edge}}^{\text{ox.}|\text{sol.}}$ a few layers within the material, we avoid these surface impacts changing the potential away from the correct bulk-like potential.
We can now use this information to inform the potential distribution of the bulk oxide region using the Poisson-Boltzmann conditions outlined in Eq.~\ref{eq:pb-ms-conditions}.
The potential within the continuum oxide region can then be approximated as:
\begin{equation}
\bar{\phi}^{\text{ox.}}(z) = \bar{\phi}(z_{\text{edge}}^{\text{ox.}|\text{sol.}}) + \frac{\epsilon_o \epsilon_{\text{sc}}}{2n_{\text{d}}}\left[\left(\frac{d\bar{\phi}^{\text{ox.}}}{dz}(z)\right)^2 - \left(\frac{d\bar{\phi}}{dz}(z_{\text{edge}}^{\text{ox.}|\text{sol.}})\right)^2\right] - k_{\text{B}}T .
\label{eq:ox-pot}
\end{equation}
This relationship applies throughout the bulk oxide region, which lasts from $z_{\text{edge}}^{\text{ox.}|\text{sol.}}$ to $z_{\text{edge}}^{\text{m}|\text{ox.}}$.
We will specify the potential drop across this region as $\Phi_{\text{ox.}} = \bar{\phi}(z_{\text{edge}}^{\text{m}|\text{ox.}}) -\bar{\phi}(z_{\text{edge}}^{\text{ox.}|\text{sol.}}) $.
Similarly, for the metal-oxide interface we can say that
\begin{equation}
q_{\text{DFT}}^{\text{m}|\text{ox.}} + q_{\text{h}}^{\text{m}|\text{ox.}}= 0 ,
\label{eq:chg-neutrality-m-ox}
\end{equation}
where $q_{\text{h}}^{\text{m}|\text{ox.}}$ is the charge on the Helmholtz planes on the oxide sides of the DFT slab, and $q_{\text{DFT}}^{\text{m}|\text{ox.}}$ is the charge on the DFT metal-oxide interface.
We note that for metals, all excess charge can be expected to accumulate at the interface, leading to a flat potential within the bulk metal.
This means that it is unnecessary to include a symmetric Helmholtz plane representing charge within the rest of the metal electrode.
Here, the value of $q_{\text{h}}^{\text{m}|\text{ox.}}$ should be set equal to the amount of charge that has accumulated in the system up to that point.
Using Gauss' law, this would mean that
\begin{equation}
q_{\text{h}}^{\text{m}|\text{ox.}} = \epsilon_o \epsilon_{\text{sc}} A \frac{d\bar{\phi}^{\text{ox.}}}{dz}\left(z_{\text{edge}}^{\text{m}|\text{ox.}}\right),
\label{eq:int2-chg}
\end{equation}
where $A$ is the lateral surface area of the metal-oxide DFT interface.
Since this charge will be localized to the metal-oxide oxide interface in the subsequent DFT calculation, the area of the metal-oxide DFT interface is therefore the relevant one for Gauss' law.
While this setup ensures continuity between the two different DFT equations, it does not yet provide guidance on what the correct distribution of charges in Eq.~\ref{eq:chg-neutrality-ox-sol} should be.
The final condition to find the correct distribution of charges is to specify that a system at equilibrium must have the Fermi level be the same throughout the system:
\begin{equation}
\varepsilon_{\text{F}}^{\text{ox.}|\text{sol.}} = \varepsilon_{\text{F}}^{\text{m}|\text{ox.}} + \Phi_{\text{ox.}},
\end{equation}
where $\varepsilon_{\text{F}}^{\text{ox.}|\text{sol.}}$ and $\varepsilon_{\text{F}}^{\text{m}|\text{ox.}}$ are the Fermi level of the DFT oxide-solution interface and DFT metal-oxide interface, respectively.
By adding in the potential shift from the oxide $\Phi_{\text{ox.}}$, we make sure these are evaluated from the same reference point.
We can relate these DFT calculated Fermi levels to the vacuum energy referenced work function by adding on a potential shift from the solution $\Phi_{\text{sol.}}$:
\begin{equation}
W = \Phi_{\text{sol.}} + \varepsilon_{\text{F}}^{\text{ox.}|\text{sol.}} .
\end{equation}
This solution potential shift can be determined directly by another Poisson-Boltzmann shift \cite{fisicaro2016generalized,mathew2019implicit}, or by a constant solvation shift.
In real world scenarios, this potential drop would further be determined by the salt concentration and pH of the solution.
These effects can be partially included when calculating the formation energy of ions as in Eqn.~\ref{eq:chloride-e}, and can also be somewhat included in the implicit Poisson-Boltzmann solvers mentioned above.
For the sake of simplicity, within this work we will run our simulations in vacuum and do not add a solution contribution since we already including water molecules and Helmholtz planes as a first-order approximation for the voltage-charge response.
Future work should focus on integrating more detailed knowledge of the surrounding solution into the voltage drop within the solution.
To relate the calculated work function from DFT to the equilibrium voltage of a DFT calculation, we will use the Trasatti convention \cite{trasatti1986absolute}, where we define the potential of a given DFT slab as (versus SHE) as
\begin{equation}
\Phi = W/|e| - 4.44\text{ V},
\end{equation}
where $W$ is the work function of the system, defined as the opposite of the DFT calculated Fermi level, i.e. $W=-\varepsilon_{\text{F}}$.
(This relation is only strictly true for pure water, but we will accept the approximation for this system).
Thus, by using this methodology, we can specify a charge $Q$ on the electrode, and then test different distributions of $q_{\text{h}}^{\text{ox.}|\text{sol.}}$ and $q_{\text{DFT}}^{\text{ox.}|\text{sol.}}$ to find the equilibrium charge state and voltage of the system.
This allows standard DFT slab calculations of metal-insulator-solution interfaces to fully include voltage effects-- without the need to simulate 10s of nm of material.
By reducing the number of atoms needed to simulate within DFT by at least an order of magnitude, this work significantly speeds up calculations of realistic voltages for the system, to the point of making $>$ 5 nm systems computationally feasible.
Nevertheless, several smaller DFT calculations are needed to optimize the free energy within the system, finding the charge distribution which leads to a Fermi level being in equilibrium throughout the interface.
In this work, we manually test different combinations of charge to find the correct equilibrium distribution, which leads to roughly five to ten DFT slab calculations needed for each data point.
We have previously implemented an automated Newton-Raphson optimization algorithm for finding equilibrium free energy within semiconductor-solution systems \cite{campbell2019voltage}.
A similar algorithm could also be extended to these metal-insulator-solution systems for automated voltage processing.
\subsection{DFT Details}
We use structures based on the interfaces used in Leung's work on Al surfaces \cite{leung2021first}, which are illustrated in Fig.~\ref{fig:no-gb}.
Leung investigated chloride insertion into Al(111)/$\alpha$-Al$_2$O$_3$(0001) surfaces, finding that chloride insertion was only favorable in situations with grain boundaries.
By using these atomic structures, we can make more direct comparisons with the Leung work.
The $\alpha$ phase of Al$_2$O$_3$ is selected as an approximation of the amorphous alumina that would develop on a surface in contact with water molecules.
The Al/$\alpha$-Al$_2$O$_3$ structure is based on the ``FCC'' interface structure \cite{siegel2002adhesion}, but with double lattice cell dimensions.
We work with two interfaces throughout the work, one without a grain boundary, and one with a grain boundary.
For each interface, we create a structure both at the metal-oxide interface and the oxide-solution interface.
All of the oxide-solution structures in question include a sub monolayer of water on the surface representing a buildup from humidity.
The arrangement of water molecules in this work was taken from Leung's work \cite{leung2021first}, which was from a snapshot from \textit{ab-initio} molecular dynamics (AIMD).
The exact arrangement of water molecules has been shown to have significant effects on the total work function of DFT calculations \cite{cheng2012alignment,kharche2014first,hormann2019absolute}.
At non-zero temperatures, the water molecules will necessarily be dynamic, and this represents a reasonable snapshot.
By keeping the arrangement of water molecules consistent between the structures with and without a grain boundary, moreover, we minimize the impacts this has on the relative error between the two structures.
All electronic structure calculations are done using the {\sc quantum espresso} package~\cite{giannozzi2009quantum}.
We use norm-conserving pseudopotentials from the PseudoDojo repository~\cite{van2018pseudodojo} and the Perdew-Burke-Ernzerhof exchange-correlation functional~\cite{perdew1996generalized}.
We use kinetic energy cutoffs of 50 Ry and 200 Ry for the plane wave basis sets used to describe the Kohn-Sham orbitals and charge density, respectively.
We use a 2$\times$2$\times$1 Monkhorst-Pack grid~\cite{monkhorst1976special} to sample the Brillioun zone in our calculations.
We use the {\sc Environ} package \cite{andreussi2012revised} to calculate the parabolic corrections to the surface dipole.
For simplicity, all calculations are done in vacuum.
\section{Results}
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{fig_no_gb_structures.pdf}
\caption{The structures used for the Al(0001)/$\alpha$-Al$_2$O$_3$-water interface calculations for (a) a full interface, (b) an oxygen vacancy, highlighted by the red circle, (c) and Cl insertion into the interface.}
\label{fig:no-gb}
\end{figure*}
We first consider chloride ion insertion in an interface without any grain boundaries.
To start, we calculate the flatband potential of the system before any oxygen vacancies or chlorine atoms are inserted.
For the oxide-solution system we find a Fermi level of $ \varepsilon_{\text{F,fb}}^{\text{ox.}|\text{sol.}} = -6.2$ eV and for the metal-oxide system we find $\varepsilon_{\text{F,fb}}^{\text{m.}|\text{ox.}} = -6.1$ eV.
To find the flatband potential of the total system it is necessary to find the arrangement of charge that causes the Fermi level to be in equilibrium throughout the system, and the total charge on the electrode must be neutral, i.e. $Q=0$ .
This constrains the problem sufficiently that the flatband potential of the entire system can be calculated as $\Phi_{\text{fb}} \approx 1.7 $ V (SHE).
We then investigate placing an oxygen vacancy into the structure.
Following Leung's work, we place the oxygen vacancy 4.3 \AA\ away from the metal surface, where the zero $z$-location is here defined as the first oxygen atoms of the oxide structure moving away from the metal.
We allow the system to relax around the vacancy.
We then place a chlorine atom into this oxygen vacancy and again allow the system to relax.
We assume that oxygen vacancies are not mobile within the timeframe of chloride insertion.
For the sake of this demonstration, we do not examine the kinetics of chloride insertion, but the QCA techniques presented here do provide an opportunity for future work calculating kinetic barriers while the surface is under applied voltage.
Using Bader charge analysis \cite{henkelman2006fast}, we confirm that the Cl atom carries $\approx$ 8 electrons, and thus is, in fact, acting as a Cl$^{-}$ ion within the system.
\begin{figure*}
\centering
\includegraphics[width=0.8\textwidth]{figgb_structures.pdf}
\caption{The structures used for the Al(0001)/$\alpha$-Al$_2$O$_3$-water interface calculations including a grain boundary for (a) an oxygen vacancy, highlighted by the red circle, and (b) Cl insertion into the interface.}
\label{fig:gb}
\end{figure*}
We then apply QCA as outlined in section \ref{subsec:qca} over a variety of voltages, and calculate the resulting formation energy for chloride insertion into the oxygen vacancy.
Here, the DFT energies for the chloride insertion and oxygen vacancy, $E_{\text{DFT}}(\text{Cl ins.})$ and $E_{\text{DFT}}(\text{O vac.})$ in Eq.~\ref{eq:total-e} need to be calculated including the impact of voltage as found in the equilibrium QCA configuration.
As seen in Fig.~\ref{fig:stability-diagram}, the formation energy for chloride insertion is highly positive across a wide range of voltages, implying that chloride is highly unlikely to form at these potentials.
The chloride insertion energy only dips below 0 eV at $\gtrapprox$ 1.2 V (SHE).
This matches previous work by Leung who reports that chloride ions are not stable inserted into crystalline $\alpha$-Al$_2$O$_3$ \cite{leung2021first}.
We predict that the relationship between the applied voltage and the chloride insertion energy is largely linear.
This indicates that the dominant term in the energy is in fact the chloride ion energy introduced in Eqn.~\ref{eq:chloride-e}.
The can be attributed to a majority of the charge in the system accumulating at the oxide-solution interface, leaving only a small potential drop in the bulk oxide region, where the parabolic relationships for voltage would become more dominant.
This also highlights the importance of pairing QCA with a computational standard hydrogen electrode or similar approach which allows for inclusion of the voltage in the calculation of the ionic energy.
We next consider an interface that includes grain boundaries.
We again first measure the Fermi level of the oxide-solution system as $ \varepsilon_{\text{F,fb}}^{\text{ox.}|\text{sol.}} = -5.54$ eV and the metal-oxide system as $\varepsilon_{\text{F,fb}}^{\text{m.}|\text{ox.}} = -6.1$ eV.
Using the same procedure as before we can thus find a flatband potential for the grain boundary system of $\Phi_{\text{fb}} \approx 1.4 $ V (SHE).
We then insert oxygen vacancies or chlorine atoms into the structure and allow the system to relax.
It should be noted that due to periodic boundary conditions, the density of grain boundaries is much higher in this system than would be realistically expected.
We again confirm with Bader charge analysis that the Cl atom is acting as a Cl$^{-}$ ion within the system.
\begin{figure}
\centering
\includegraphics[width=0.6\textwidth]{al2o3-chloride-insetion-qca.pdf}
\caption{A stability diagram indicating the calculated insertion energy of a chloride ion as a function of voltage.
When the chloride ion is inserted into a grain boundary it becomes stable (i.e. $< 0$ eV, indicated by the horizontal black dashed line) at voltages $> -0.2$ V (SHE).
This compares favorably with the experimentally measured onset of chloride pitting at -0.5 V (SHE) indicated by the vertical red dashed line.
In contrast, in systems without grain boundaries, chloride ion insertion is not favorable until much higher voltages. }
\label{fig:stability-diagram}
\end{figure}
As seen in Fig.~\ref{fig:stability-diagram}, we then test a variety of voltages across the system with grain boundaries.
We predict that chloride ion insertion becomes stable at voltages $> -0.2$ V (SHE).
This correlates relatively well with the previous experiments showing pitting corrosion at voltages $> -0.5$ V (SHE) \cite{dibari1971electrochemical,bessone1992eis} (see Sec.~\ref{sec:discussion} for a discussion on comparing theoretical and experimental voltages).
It also matches previous results from Leung, who similarly finds that chloride ion insertion is only favorable at grain boundaries within the aluminum oxide \cite{leung2021first}.
This work emphasizes the role that grain boundaries likely play in pitting corrosion according to the point defect model, serving as locations for chloride insertion into the aluminum oxide and serving as corrosion centers.
This results also follows intuition: Cl$^-$ is a larger anion than oxygen and the extra room within the grain boundary allows for a more favorable insertion energy.
\begin{figure}
\centering
\includegraphics[width=0.6\textwidth]{al2o3-charge_density.pdf}
\caption{The equilibrium charge density at the interface with a Cl$^{-1}$ ion inserted both with and without grain boundaries.
Here, negative values are taken to indicate excessive electrons. }
\label{fig:chg-density}
\end{figure}
The voltage dependence of the charge density, shown in Fig.~\ref{fig:chg-density}, also matches the intuitive picture of chloride stability in grain boundaries we have established.
For interfaces with and without a grain boundary, the charge density is largely linear as a function of voltage.
This reflects the fact that, as previously noted, the majority of the charge in our simulations of the system is caught at the oxide/solution interface, causing the interface to behave in a more metallic manner, particularly away from the flatband potential.
Furthermore, the interface with grain boundaries consistently has lower charge densities at voltages where the Cl$^{-1}$ ion is predicted to be stable, somewhat matching the intuition that more charge is allowed in the system.
(In this notation, ``lower'' charge densities indicate a greater density of electrons induced at the interface.)
This is only violated at $\Phi \lessapprox$ -0.2 V (SHE), corresponding to the location where a Cl$^{-1}$ ion is no longer stable in the system.
Below this voltage, the charge density for the two systems becomes nearly equal.
\section{Discussion}
\label{sec:discussion}
While QCA provides a useful tool for attaching relative voltages to DFT calculations of interfaces, it does not solve a variety of outstanding problems in defining the electronic voltage of a neutral DFT system relative to experiment.
In particular, the arrangement of water molecules within the first monolayer of coverage has been shown to have a large impact on the estimated flatband potential, with previous studies on this system showing the impact of different H$_2$O configurations being 0.2-0.5 eV \cite{sakong2016structure,sakong2018electric}.
This is a difficult problem to overcome as there is no clearly ``correct'' configuration of the water molecules on the surface.
One possible solution would be to average over several possible H$_2$O configurations sampled from an AIMD calculation and attempt to form some averaging.
As indicated previously, another concern is the relative concentration of charged defects being examined.
The periodic nature of these DFT calculations ensures that the 2D density of any defects will be relatively high.
Furthermore, realistic systems likely have larger macrostructure, with grain boundaries and defects that are not actually periodic in nature.
This can be partially overcome by creating larger lateral system sizes and performing convergence tests.
Increasing the lateral size of the interfaces typically necessitates a a tradeoff of the length of the slab to keep calculations computationally feasible.
This shorter slab can still be accommodated to provide useful results with QCA provided that a few layers have been included around the defect and interface.
These twin uncertainties make exact comparison of our predicted chloride insertion onset voltage to macroscopic potentials extracted from experiment problematic.
Our QCA predictions of exact potentials should then be taken as semi-quantitative about the system, not exact.
For instance, the finding that grain boundaries within the aluminum oxide make chloride insertion much more favorable are consistent regardless of the exact voltage at which they become favorable.
Similarly, if our results predicted that chloride insertion was only favorable only at voltages $>$ 3 V (SHE), it would be likely that we were not correctly representing the interface within DFT (or that QCA itself was an incorrect framework for the problem).
When predicting relative voltage differences in the same system (i.e. the difference between two voltages predicted using the same surface termination), however, QCA should be more rigorous as the previous drawbacks to exact work function determination are held constant.
QCA nevertheless presents an advantage over other computational approaches to controlling the voltage of an electrode by coupling DFT to continuum and Poisson-Boltzmann descriptions of the solution.
This is namely the ability to include an extended insulating (oxide) region within the simulation.
Previous work has needed to focus on metals with a linear relationship between the charge and voltage of the system, consistent with a constant capacitance model \cite{schnur2011challenges,keilbart2017quantum,haruyama2018electrode,kano2021study}.
While these approaches can be quite useful, their utility can be limited in comparison to experiment, especially if large passivating layers are expected to form over the metal electrode and the capacitance can no longer be expected to be constant as a function of potential.
Given that the contact layer between metal and oxide has been shown to sometimes have barriers as high as 2-3 V \cite{yu2019combining,cornette2020relation}, the impact of the extended oxide layer is vital to fully understanding the voltage behavior of an electrode.
QCA thus presents a computationally scalable method for addressing voltages in systems with extended insulating regions.
\section{Conclusion and Outlook}
In this work, we have extended the QCA methodology to encompass metal-oxide interfaces that are also in contact with a solution.
This extension allows us to provide first principles calculations of atomic mechanisms behind corrosion and other electrochemical processes as a function of applied voltage, which has previously been difficult to manage within a DFT framework.
We demonstrate this methodology on an Al/$\alpha$-Al$_2$O$_3$ interface, calculating that chloride insertion is only favorable above -0.2 V (SHE) when a grain boundary is included at the interface.
This conclusion is broadly consistent with both previous first principles work, and experimental measurements.
This provides helpful context for parameterizing the point defect model with first-principles results.
While QCA provides a useful method for examining interfacial problems that includes voltage, it does not resolve some of the traditional difficulties of DFT.
Band gaps will be underestimated in semiconductors and insulators when using semi-local functionals.
As long as the potential drop across the electrode is less than the band gap of the DFT material, however, this should not affect the relative voltage results.
Similarly, the exact atomic structures of both the metal-oxide and oxide-solution interfaces exert a large influence on the assigned flatband potential of the interfacial structure.
Numerous DFT methods have been developed to address this issue, with high temperature ab-initio molecular dynamics providing one option for generating structures.
Larger surveys of possible structures should still be made for QCA analysis of any given system.
It nevertheless presents advantages over previous work in being able to examine systems with large passivation layers.
In this work, we demonstrated QCA in the context of corrosion of aluminum, but it can clearly also be used in numerous applications of interest.
In battery electrode research, for instance, the exact atomistic interaction of a solid-electrolyte interface (SEI) with Li ions shuttling through the structure is often not well understood as a function of voltage \cite{leung2020dft}.
QCA provides one potential route for exploring these systems while including a realistic voltage to the system.
Another application may be photocatalysis, where a semiconductor is placed in contact with solution to catalyze a reaction, often attached to a metal electrode to provide some level of voltage to the system \cite{wenderich2016methods}.
Beyond electrochemistry, a similar methodology can be applied to electronic interfaces of interest such as p-n junctions and Schottky barriers \cite{subramanian2020photophysics}.
Thus, QCA provides a new tool for computational electrochemistry across a broad variety of interfaces and applications, including voltage effects at reasonable computational cost.
\ack
We would like to thank Kevin Leung for providing the initial Al/Al$_2$O$_3$ system and a critical reading of the draft manuscript, as well as Andrew Baczewski for useful feedback.
This work was supported by the Laboratory Directed Research and Development program at Sandia National Laboratories under project 226354 and 229741.
Sandia National Laboratories is a multi-mission laboratory managed and operated by National Technology and Engineering Solutions of Sandia, LLC, a wholly owned subsidiary of Honeywell International, Inc., for DOE's National Nuclear Security Administration under contract DE-NA0003525.
This paper describes objective technical results and analysis.
Any subjective views or opinions that might be expressed in the paper do not necessarily represent the views of the U.S. Department of Energy or the United States Government.
\section*{References}
\bibliographystyle{iopart-num}
\section{Introduction}
Corrosion of metallic interfaces leads to weakening of their structural integrity and has been widely estimated to cost billions of dollars annually \cite{koch2002corrosion}.
While this has inspired numerous studies on corrosion and strategies to protect against it, the fundamental atomistic mechanisms behind corrosion are difficult to identify for any given system.
Corrosion in aluminum provides an excellent example of the issues.
In aluminum, it has been experimentally demonstrated that pitting corrosion begins at potentials $\Phi > -0.50$ V relative to the standard hydrogen electrode (SHE), dependent on the salt and pH concentration of the surrounding solution \cite{dibari1971electrochemical,bessone1992eis}.
The exact mechanism that causes this pitting corrosion is still under debate, however.
Pitting corrosion is often described using the point defect model \cite{lin1981point,chao1981point,engelhardt2004unification}, which states that corrosion is initially driven by the formation of charged defects in the surface of the material, of which chloride ions are particularly damaging \cite{frankel1998pitting,natishan20182017}.
On aluminum surfaces, a several nm thick aluminum oxide layer forms \cite{evertsson2015thickness} and the insertion of chloride into oxygen vacancies in this layer has been theorized as one of the main mechanisms leading to corrosion.
Experiments have investigated the dependence of pitting corrosion on chlorine concentration in the surrounding electrolyte and other factors \cite{brown1973use,shimizu1999novel,serna2006critical}, but the results have not been fully conclusive as to the causal mechanism.
While many previous density functional theory (DFT) studies have examined Al/Al$_2$O$_3$ interfaces at an atomic level \cite{kim2013nature,batyrev2001plane,koberidze2018structural,costa2014atomistic,siegel2002adhesion}, they have typically not included oxygen vacancies and chloride insertion.
For those that have examined vacancies \cite{weber2009point,carrasco2004theoretical,hine2009supercell,janetzko2004first}, they do not include Al metal in the calculation.
Previous DFT studies on chloride insertion into Al/Al$_2$O$_3$ interfaces that have included contributions from the metal, the oxide, vacancies, and chloride ions \cite{liu2019dft,liu2021density,leung2021first} have either implicitly limited their scope to defects with no charge, or tested only a small range of voltages.
This lack of study of charged ions in the full electrochemical environment stems from the difficulty of needing to include several different length scales in a simulation to get a realistic picture of corrosion behavior.
Modeling the realistic interaction of point defects with the surface requires a quantum mechanical treatment, with DFT providing a convenient compromise between accuracy and speed.
DFT is typically limited, however, to simulating $\sim$1000s of atoms as larger systems are computationally prohibitive.
Since a realistic model of an interface, including the surrounding solution, can be 10s to 100s of nm wide, DFT simulations must necessarily focus on a single region of interest, typically the surface.
The entire surrounding electrode helps determine the electric field (and therefore voltage) that the region of interest operates at.
Simulations of these isolated regions therefore do not provide information as to the voltage the electrode is under and, in effect, operate at ``flatband'' conditions where the electric field is flat across the system.
This necessarily limits the scope of traditional DFT calculations when examining these interfaces and defects.
Since the defects under examination are charged, it is crucial to incorporate a realistic model of voltage into DFT calculations.
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{fig1.png}
\caption{A schematic illustration of the Quantum Continuum Approximation applied to a metal-insulator interface immersed in solution.}
\label{fig:qca-schematic}
\end{figure*}
In this paper, we develop a Quantum Continuum Approximation (QCA) methodology that allows for the first-principles simulation of the metal-insulator interfaces at controlled voltages using limited DFT simulation cell sizes.
QCA self-consistently couples DFT calculations of the metal-insulator interface and the insulator-solution interface with Poisson-Boltzmann distribution of charges in the bulk insulator and solution to provide an equilibrium potential drop for the entire electrode, as schematically illustrated in Fig.~\ref{fig:qca-schematic}.
By reducing the number of atoms needed for DFT calculations to include voltage information by at least an order of magnitude, QCA enables voltage-dependent first-principles simulation of realistic electrochemical interfaces at reasonable computational cost.
We demonstrate this methodology for the simulation of chloride insertion into the aluminum oxide layer of an Al(111)/$\alpha$-Al$_2$O$_3$(0001) system as a function of voltage.
We predict that insertion is not favorable for Al/Al$_2$O$_3$ interfaces that do not include grain boundaries at any voltage.
We do see, however, that in systems with grain boundaries in the Al$_2$O$_3$, chloride insertion is favorable at $\Phi > -0.2$ V (SHE), in reasonable agreement with experiment and helping confirm the importance of grain boundaries in the corrosion process.
This work provides a template for attaching a potentiostat to electrochemical DFT calculations of interfaces.
\section{Methods}
\subsection{Chloride insertion energy calculations}
We calculate the formation energy of a Cl$^-$ ion inserting into an oxygen vacancy as
\begin{equation}
\Delta E_{\text{Cl-O}}[\Phi] = E_{\text{DFT}}(\text{Cl ins.})[\Phi] - E_{\text{DFT}}(\text{O vac.})[\Phi] - E_{\text{Cl}^-}[\Phi] ,
\label{eq:total-e}
\end{equation}
where $E_{\text{DFT}}(\text{O vac.})[\Phi]$ is the DFT energy of the slab with the oxygen vacancy as a function of the global potential $\Phi$ of the system, $E_{\text{DFT}}(\text{Cl ins.})[\Phi]$ is the DFT energy of the slab with a chlorine ion inserted into the previous location of the oxygen vacancy as a function of the global potential, and $E_{\text{Cl}^-}[\Phi]$ is the energy of the Cl$^-$ ion in solution as a function of voltage.
For the chloride ion, we follow a computational hydrogen electrode-like approach \cite{norskov2004origin,rossmeisl2007electrolysis,man2011universality}, where we determine the energy of the chloride ion as
\begin{equation}
E_{\text{Cl}^-}[\Phi] = \mu (\text{Cl}^-) + e\Phi^{\text{SHE}}.
\label{eq:chloride-e}
\end{equation}
Here, $\mu (\text{Cl}^-)$ is the equilibrium chemical potential of the chloride ion, and $\Phi^{\text{SHE}}$ is the potential that the electrode is operating at, referenced to the standard hydrogen electrode (SHE).
We determine the chemical potential by referencing the following chemical reaction
\begin{equation}
\text{Cl}_2 (\text{g}) + 2 e^- \rightarrow 2\text{Cl}^-(\text{aq}),
\end{equation}
which is at equilibrium at $\Phi_{\text{Cl}_2/\text{Cl}^-} = 1.36$ V (SHE) \cite{atkins2006physical}.
Using this reaction, we can then calculate the chemical potential of a chloride ion as
\begin{equation}
\mu (\text{Cl}^-) =\frac{1}{2} E_{\text{DFT}}(\text{Cl}_2(\text{g})) - e\Phi_{\text{Cl}_2/\text{Cl}^-}.
\end{equation}
With this energetic relationship established, we can now calculate the formation energy of chloride insertion into the Al/Al$_2$O$_3$ surface as a function of potential the electrode is under.
For this relationship to be strictly accurate, however, the calculated DFT energies of the slabs need to be taken at simulation conditions consistent with $\Phi$ for the electrode.
We outline how we can achieve this using the QCA methodology in the next section.
\subsection{Quantum Continuum Approximation (QCA) Methodology for Predicting Voltage at an interface}
\label{subsec:qca}
\subsubsection{Interface Energetics and Electrostatics}
At a fundamental level, our approach to calculations with metal-semiconductor-solution interfaces works by assigning a global potential to the system.
This potential drop can then be split up between several different voltage drop regions within the interface.
Finding the correct distribution of potential drops requires a self-consistent process for minimizing the free energy of the system while taking into account the electrostatic requirements of following the Poisson equation and maintaining electronic equilibrium throughout the interface.
To accomplish this task, we employ a new extension of the Quantum Continuum Approximation (QCA) methodology that was first developed by Campbell and Dabo for semiconductor-solution interfaces \cite{campbell2017quantum,campbell2019voltage}, which allowed for the voltage of the system to be additionally predicted by DFT slab calculations.
It was later extended to solid-state metal-semiconductor interfaces, providing predictions of Schottky barriers that corresponded well with nano-ARPES measurements \cite{subramanian2020photophysics}.
Here, we extend this to include metal-insulator or metal-semiconductor interfaces that are also immersed in solution, which are common in electrochemical applications.
The basic goal is to self-consistently couple DFT calculations of an interface with analytic Poisson-Boltzmann descriptions of the bulk semiconductor and solution regions.
As shown in Fig.~\ref{fig:qca-schematic}, we perform DFT calculations only on the crucial interface sections, such as the metal-metal oxide transition and the metal oxide-solution transition, or around atomic defects in the system.
We then apply Poisson-Boltzmann calculations to the oxide regions in between the DFT calculations and the surrounding solution.
The free energy of the system is
\begin{equation}
\mathcal{F} = \mathcal{F}^{\text{m}|\text{ox.}} + \mathcal{F}^{\text{ox.}} + \mathcal{F}^{\text{ox.}|\text{sol.}} + \mathcal{F}^{\text{sol.}} - \frac{1}{2}\int dr \epsilon_0 \epsilon(\mathbf{r}) |\nabla(\phi(\mathbf{r}))|^2 ,
\end{equation}
where $\epsilon_0$ is the vacuum permittivity, $\epsilon(\mathbf{r})$ is the dielectric constant as a function of position, $\phi$ is the electrostatic potential, and $\mathcal{F}^{\text{m}|\text{ox.}}$ and $\mathcal{F}^{\text{ox.}|\text{sol.}}$ are the free energies of the metal-oxide and oxide-solution interface, respectively, which can be calculated using the typical calculations for DFT energy.
Similarly, $\mathcal{F}^{\text{ox.}}$ and $\mathcal{F}^{\text{sol.}}$ are the contributions of the bulk oxide and solution regions, which can be calculated as analytic functions of the distribution of charge carriers within each system.
This decomposition of the free energy naturally leads to the following Poisson equation for the system:
\begin{equation}
\nabla(\epsilon_0\epsilon(\mathbf{r})\nabla \phi(\mathbf{r})) = p(\mathbf{r}) - n(\mathbf{r}) -p_d(\mathbf{r}) + n_d(\mathbf{r}) + c_+(\mathbf{r}) - c_-(\mathbf{r}) + \rho_+(\mathbf{r}) -\rho_-(\mathbf{r}),
\label{eq:electrostatics}
\end{equation}
where and $\rho_-(\mathbf{r})$ is the electron density obtained from the self-consistent Kohn-Sham equation for a given distribution $\rho_+(\mathbf{r})$ of atomic cores, and the source terms can be expressed as
\begin{equation}
n(\mathbf{r}) = n_d(\mathbf{r}) \left[1 + \exp\left(\frac{\phi(\mathbf{r}) + \epsilon_c - \epsilon_F}{k_{\text{B}}T}\right)\right]^{-1},
\end{equation}
\begin{equation}
p(\mathbf{r}) = p_d(\mathbf{r}) \left[1 + \exp\left(\frac{\epsilon_F-\epsilon_v-\phi(\mathbf{r})}{k_{\text{B}}T}\right)\right]^{-1},
\end{equation}
\begin{equation}
c_{\pm}(\mathbf{r}) = c^{\circ}(\mathbf{r}) \exp\left(\mp \frac{\phi(\mathbf{r})}{k_{\text{B}}T}\right).
\end{equation}
Here, $\epsilon_F$ denotes the Fermi energy, $\epsilon_v$ is the energy at the valence band maximum, $\epsilon_c$ is the energy at the conduction band minimum, $n_d(\mathbf{r})$ is the concentration of $n$-type carriers in the semiconducting/insulating region and is 0 outside this region, $p_d(\mathbf{r})$ is similarly the concentration of $p$-type carriers within the semiconducting regions, and $c^{\circ}(\mathbf{r})$ is the equilibrium ionic concentration inside the electrolyte region and is 0 outside of the electrolyte region.
In an $n$-type semiconducting region, the charge carriers can be solved for using the Boltzmann distribution as
\begin{equation}
p(\mathbf{r}) = p_d = 0,
\end{equation}
\begin{equation}
n(\mathbf{r}) = n_d\exp\left(\frac{\phi_0 - \phi(\mathbf{r})}{k_{\text{B}}T}\right),
\end{equation}
where $\phi_0$ is the asymptotic value of the potential in this region.
Analogous equations can be written for $p$-type semiconducting regions.
\subsubsection{Process for solving electrostatics}
To practically solve Eq.~\ref{eq:electrostatics} for a given potential, we start by dividing the system into two portions that we will simulate with DFT separately.
The first is the surface of the interface, i.e. where the passivating oxide layer meets the solution.
We will refer to this region with a superscript ox.$|$sol. in variables throughout this work.
The other DFT region is the interface of the metal with the passivating oxide, referred to as a superscript m$|$ox. within this work.
For both of these regions we start by doing DFT calculations on these systems when no charge has been applied.
This corresponds to the so-called `flatband' conditions.
We then measure both the planar averaged potential profile of these systems $\phi_{\text{fb}}^{\text{m}|\text{ox.}}(z)$ and $\phi_{\text{fb}}^{\text{ox.}|\text{sol.}}(z)$, and the Fermi level of these systems $\varepsilon_{\text{F,fb}}^{\text{m}|\text{ox.}}$ and $\varepsilon_{\text{F,fb}}^{\text{ox.}|\text{sol.}}$.
Next we assign a total charge $Q$ to the electrode.
The electrode is here defined as the region from the metal to the interface with the solution.
The distribution of this charge $Q$ within the interface mechanically determines the total potential drop via Gauss' law.
For a perfectly insulating/semiconducting system, the charge distribution can be well approximated by the following Poisson--Boltzmann relationship \cite{schmickler2010interfacial}:
\begin{equation}
\frac{d^2\phi}{dz^2} = \frac{n_{\text{d}}}{\epsilon_o \epsilon_{\text{sc}}} \left[ 1 - \exp \left( \frac{-\phi}{k_{\text{B}}T}\right)\right].
\label{eq:pb-ms-conditions}
\end{equation}
Here, $n_{\text{d}}$ is the concentration of charge carriers (dopants) within the semiconductor, $\epsilon_o$ is the vacuum permittivity, $\epsilon_{\text{sc}}$ is the dielectric constant of the semiconductor, and $k_{\text{B}}$ is the Boltzmann constant.
For most real-world systems, however, surface states will form at both the metal-insulator and insulator-solution interface \cite{bardeen1947surface,cowley1965surface,bard1980role}.
These surface states result from the abrupt nature of the interface disrupting the symmetry of the material, and from defects induced by interaction.
These surface states trap charge at the interface, thus reducing the charge throughout the remainder of the insulator and typically reducing the potential drop of the electrode.
Capturing these surface states requires a quantum mechanical treatment of the interfaces using, e.g., DFT.
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{fig-interface-potentials.pdf}
\caption{A schematic illustration of the potential profile of a metal-oxide-solution interface within QCA with important variables depicted. Distances and potentials shown are approximate.}
\label{fig:qca-potentials}
\end{figure*}
To simulate charge accumulation at these surface interfaces we place Helmholtz planar counter-charges on either side of a DFT slab, as illustrated in Fig.~\ref{fig:qca-potentials}.
We place the Helmholtz planes 4 \AA\ away from the surfaces of the slab to mimic the Debye length.
We obey charge neutrality for the DFT systems simulated, such that
\begin{equation}
q_{\text{h}}^{\text{ox.}|\text{sol.}} + q_{\text{DFT}}^{\text{ox.}|\text{sol.}} + q_{\text{h}}^{\text{sol.}} = 0
\label{eq:chg-neutrality-ox-sol}
\end{equation}
at the oxide-solution interface, where $q_{\text{h}}^{\text{ox.}|\text{sol.}}$ and $q_{\text{h}}^{\text{sol.}}$ is the charge on the Helmholtz planes on the oxide and solution side of the DFT slab, respectively.
Additionally, $q_{\text{DFT}}^{\text{ox.}|\text{sol.}}$ is the charge on the DFT slab representing the interface between the oxide and solution.
Since the electrode charge is defined as including the metal-oxide interface, the continuum oxide, and the oxide-solution interface, we can automatically assign the charge on the Helmholtz plane representing the solution as the opposite of this electrode charge, \textit{i.e.} $q_{\text{h}}^{\text{sol.}} = -Q$.
This arrangement of charge creates a potential gradient across the system.
We measure this as $\bar{\phi}^{\text{ox.}|\text{sol.}}(z) = \phi^{\text{ox.}|\text{sol.}}_{q} - \phi^{\text{ox.}|\text{sol.}}_{\text{fb}}$, where $\phi^{\text{ox.}|\text{sol.}}_{q}$ is the planar averaged potential of the charged system and $\bar{\phi}^{\text{ox.}|\text{sol.}}(z)$ is the difference in the planar averaged potential due to the charge.
We can finally couple this DFT potential to a bulk charge distribution by extracting the derivative of the subtracted potential $\frac{d\bar{\phi}}{dz}$ at a specific cutoff point $z_{\text{edge}}^{\text{ox.}|\text{sol.}}$.
In DFT, our oxide will necessarily only extend for a few layers, but in reality it should extend for at least several nanometers.
This nonphysical oxide surface creates spurious surface effects on the potential.
By setting the cutoff value for the potential $z_{\text{edge}}^{\text{ox.}|\text{sol.}}$ a few layers within the material, we avoid these surface impacts changing the potential away from the correct bulk-like potential.
We can now use this information to inform the potential distribution of the bulk oxide region using the Poisson-Boltzmann conditions outlined in Eq.~\ref{eq:pb-ms-conditions}.
The potential within the continuum oxide region can then be approximated as:
\begin{equation}
\bar{\phi}^{\text{ox.}}(z) = \bar{\phi}(z_{\text{edge}}^{\text{ox.}|\text{sol.}}) + \frac{\epsilon_o \epsilon_{\text{sc}}}{2n_{\text{d}}}\left[\left(\frac{d\bar{\phi}^{\text{ox.}}}{dz}(z)\right)^2 - \left(\frac{d\bar{\phi}}{dz}(z_{\text{edge}}^{\text{ox.}|\text{sol.}})\right)^2\right] - k_{\text{B}}T .
\label{eq:ox-pot}
\end{equation}
This relationship applies throughout the bulk oxide region, which lasts from $z_{\text{edge}}^{\text{ox.}|\text{sol.}}$ to $z_{\text{edge}}^{\text{m}|\text{ox.}}$.
We will specify the potential drop across this region as $\Phi_{\text{ox.}} = \bar{\phi}(z_{\text{edge}}^{\text{m}|\text{ox.}}) -\bar{\phi}(z_{\text{edge}}^{\text{ox.}|\text{sol.}}) $.
Similarly, for the metal-oxide interface we can say that
\begin{equation}
q_{\text{DFT}}^{\text{m}|\text{ox.}} + q_{\text{h}}^{\text{m}|\text{ox.}}= 0 ,
\label{eq:chg-neutrality-m-ox}
\end{equation}
where $q_{\text{h}}^{\text{m}|\text{ox.}}$ is the charge on the Helmholtz planes on the oxide sides of the DFT slab, and $q_{\text{DFT}}^{\text{m}|\text{ox.}}$ is the charge on the DFT metal-oxide interface.
We note that for metals, all excess charge can be expected to accumulate at the interface, leading to a flat potential within the bulk metal.
This means that it is unnecessary to include a symmetric Helmholtz plane representing charge within the rest of the metal electrode.
Here, the value of $q_{\text{h}}^{\text{m}|\text{ox.}}$ should be set equal to the amount of charge that has accumulated in the system up to that point.
Using Gauss' law, this would mean that
\begin{equation}
q_{\text{h}}^{\text{m}|\text{ox.}} = \epsilon_o \epsilon_{\text{sc}} A \frac{d\bar{\phi}^{\text{ox.}}}{dz}\left(z_{\text{edge}}^{\text{m}|\text{ox.}}\right),
\label{eq:int2-chg}
\end{equation}
where $A$ is the lateral surface area of the metal-oxide DFT interface.
Since this charge will be localized to the metal-oxide oxide interface in the subsequent DFT calculation, the area of the metal-oxide DFT interface is therefore the relevant one for Gauss' law.
While this setup ensures continuity between the two different DFT equations, it does not yet provide guidance on what the correct distribution of charges in Eq.~\ref{eq:chg-neutrality-ox-sol} should be.
The final condition to find the correct distribution of charges is to specify that a system at equilibrium must have the Fermi level be the same throughout the system:
\begin{equation}
\varepsilon_{\text{F}}^{\text{ox.}|\text{sol.}} = \varepsilon_{\text{F}}^{\text{m}|\text{ox.}} + \Phi_{\text{ox.}},
\end{equation}
where $\varepsilon_{\text{F}}^{\text{ox.}|\text{sol.}}$ and $\varepsilon_{\text{F}}^{\text{m}|\text{ox.}}$ are the Fermi level of the DFT oxide-solution interface and DFT metal-oxide interface, respectively.
By adding in the potential shift from the oxide $\Phi_{\text{ox.}}$, we make sure these are evaluated from the same reference point.
We can relate these DFT calculated Fermi levels to the vacuum energy referenced work function by adding on a potential shift from the solution $\Phi_{\text{sol.}}$:
\begin{equation}
W = \Phi_{\text{sol.}} + \varepsilon_{\text{F}}^{\text{ox.}|\text{sol.}} .
\end{equation}
This solution potential shift can be determined directly by another Poisson-Boltzmann shift \cite{fisicaro2016generalized,mathew2019implicit}, or by a constant solvation shift.
In real world scenarios, this potential drop would further be determined by the salt concentration and pH of the solution.
These effects can be partially included when calculating the formation energy of ions as in Eqn.~\ref{eq:chloride-e}, and can also be somewhat included in the implicit Poisson-Boltzmann solvers mentioned above.
For the sake of simplicity, within this work we will run our simulations in vacuum and do not add a solution contribution since we already including water molecules and Helmholtz planes as a first-order approximation for the voltage-charge response.
Future work should focus on integrating more detailed knowledge of the surrounding solution into the voltage drop within the solution.
To relate the calculated work function from DFT to the equilibrium voltage of a DFT calculation, we will use the Trasatti convention \cite{trasatti1986absolute}, where we define the potential of a given DFT slab as (versus SHE) as
\begin{equation}
\Phi = W/|e| - 4.44\text{ V},
\end{equation}
where $W$ is the work function of the system, defined as the opposite of the DFT calculated Fermi level, i.e. $W=-\varepsilon_{\text{F}}$.
(This relation is only strictly true for pure water, but we will accept the approximation for this system).
Thus, by using this methodology, we can specify a charge $Q$ on the electrode, and then test different distributions of $q_{\text{h}}^{\text{ox.}|\text{sol.}}$ and $q_{\text{DFT}}^{\text{ox.}|\text{sol.}}$ to find the equilibrium charge state and voltage of the system.
This allows standard DFT slab calculations of metal-insulator-solution interfaces to fully include voltage effects-- without the need to simulate 10s of nm of material.
By reducing the number of atoms needed to simulate within DFT by at least an order of magnitude, this work significantly speeds up calculations of realistic voltages for the system, to the point of making $>$ 5 nm systems computationally feasible.
Nevertheless, several smaller DFT calculations are needed to optimize the free energy within the system, finding the charge distribution which leads to a Fermi level being in equilibrium throughout the interface.
In this work, we manually test different combinations of charge to find the correct equilibrium distribution, which leads to roughly five to ten DFT slab calculations needed for each data point.
We have previously implemented an automated Newton-Raphson optimization algorithm for finding equilibrium free energy within semiconductor-solution systems \cite{campbell2019voltage}.
A similar algorithm could also be extended to these metal-insulator-solution systems for automated voltage processing.
\subsection{DFT Details}
We use structures based on the interfaces used in Leung's work on Al surfaces \cite{leung2021first}, which are illustrated in Fig.~\ref{fig:no-gb}.
Leung investigated chloride insertion into Al(111)/$\alpha$-Al$_2$O$_3$(0001) surfaces, finding that chloride insertion was only favorable in situations with grain boundaries.
By using these atomic structures, we can make more direct comparisons with the Leung work.
The $\alpha$ phase of Al$_2$O$_3$ is selected as an approximation of the amorphous alumina that would develop on a surface in contact with water molecules.
The Al/$\alpha$-Al$_2$O$_3$ structure is based on the ``FCC'' interface structure \cite{siegel2002adhesion}, but with double lattice cell dimensions.
We work with two interfaces throughout the work, one without a grain boundary, and one with a grain boundary.
For each interface, we create a structure both at the metal-oxide interface and the oxide-solution interface.
All of the oxide-solution structures in question include a sub monolayer of water on the surface representing a buildup from humidity.
The arrangement of water molecules in this work was taken from Leung's work \cite{leung2021first}, which was from a snapshot from \textit{ab-initio} molecular dynamics (AIMD).
The exact arrangement of water molecules has been shown to have significant effects on the total work function of DFT calculations \cite{cheng2012alignment,kharche2014first,hormann2019absolute}.
At non-zero temperatures, the water molecules will necessarily be dynamic, and this represents a reasonable snapshot.
By keeping the arrangement of water molecules consistent between the structures with and without a grain boundary, moreover, we minimize the impacts this has on the relative error between the two structures.
All electronic structure calculations are done using the {\sc quantum espresso} package~\cite{giannozzi2009quantum}.
We use norm-conserving pseudopotentials from the PseudoDojo repository~\cite{van2018pseudodojo} and the Perdew-Burke-Ernzerhof exchange-correlation functional~\cite{perdew1996generalized}.
We use kinetic energy cutoffs of 50 Ry and 200 Ry for the plane wave basis sets used to describe the Kohn-Sham orbitals and charge density, respectively.
We use a 2$\times$2$\times$1 Monkhorst-Pack grid~\cite{monkhorst1976special} to sample the Brillioun zone in our calculations.
We use the {\sc Environ} package \cite{andreussi2012revised} to calculate the parabolic corrections to the surface dipole.
For simplicity, all calculations are done in vacuum.
\section{Results}
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{fig_no_gb_structures.pdf}
\caption{The structures used for the Al(0001)/$\alpha$-Al$_2$O$_3$-water interface calculations for (a) a full interface, (b) an oxygen vacancy, highlighted by the red circle, (c) and Cl insertion into the interface.}
\label{fig:no-gb}
\end{figure*}
We first consider chloride ion insertion in an interface without any grain boundaries.
To start, we calculate the flatband potential of the system before any oxygen vacancies or chlorine atoms are inserted.
For the oxide-solution system we find a Fermi level of $ \varepsilon_{\text{F,fb}}^{\text{ox.}|\text{sol.}} = -6.2$ eV and for the metal-oxide system we find $\varepsilon_{\text{F,fb}}^{\text{m.}|\text{ox.}} = -6.1$ eV.
To find the flatband potential of the total system it is necessary to find the arrangement of charge that causes the Fermi level to be in equilibrium throughout the system, and the total charge on the electrode must be neutral, i.e. $Q=0$ .
This constrains the problem sufficiently that the flatband potential of the entire system can be calculated as $\Phi_{\text{fb}} \approx 1.7 $ V (SHE).
We then investigate placing an oxygen vacancy into the structure.
Following Leung's work, we place the oxygen vacancy 4.3 \AA\ away from the metal surface, where the zero $z$-location is here defined as the first oxygen atoms of the oxide structure moving away from the metal.
We allow the system to relax around the vacancy.
We then place a chlorine atom into this oxygen vacancy and again allow the system to relax.
We assume that oxygen vacancies are not mobile within the timeframe of chloride insertion.
For the sake of this demonstration, we do not examine the kinetics of chloride insertion, but the QCA techniques presented here do provide an opportunity for future work calculating kinetic barriers while the surface is under applied voltage.
Using Bader charge analysis \cite{henkelman2006fast}, we confirm that the Cl atom carries $\approx$ 8 electrons, and thus is, in fact, acting as a Cl$^{-}$ ion within the system.
\begin{figure*}
\centering
\includegraphics[width=0.8\textwidth]{figgb_structures.pdf}
\caption{The structures used for the Al(0001)/$\alpha$-Al$_2$O$_3$-water interface calculations including a grain boundary for (a) an oxygen vacancy, highlighted by the red circle, and (b) Cl insertion into the interface.}
\label{fig:gb}
\end{figure*}
We then apply QCA as outlined in section \ref{subsec:qca} over a variety of voltages, and calculate the resulting formation energy for chloride insertion into the oxygen vacancy.
Here, the DFT energies for the chloride insertion and oxygen vacancy, $E_{\text{DFT}}(\text{Cl ins.})$ and $E_{\text{DFT}}(\text{O vac.})$ in Eq.~\ref{eq:total-e} need to be calculated including the impact of voltage as found in the equilibrium QCA configuration.
As seen in Fig.~\ref{fig:stability-diagram}, the formation energy for chloride insertion is highly positive across a wide range of voltages, implying that chloride is highly unlikely to form at these potentials.
The chloride insertion energy only dips below 0 eV at $\gtrapprox$ 1.2 V (SHE).
This matches previous work by Leung who reports that chloride ions are not stable inserted into crystalline $\alpha$-Al$_2$O$_3$ \cite{leung2021first}.
We predict that the relationship between the applied voltage and the chloride insertion energy is largely linear.
This indicates that the dominant term in the energy is in fact the chloride ion energy introduced in Eqn.~\ref{eq:chloride-e}.
The can be attributed to a majority of the charge in the system accumulating at the oxide-solution interface, leaving only a small potential drop in the bulk oxide region, where the parabolic relationships for voltage would become more dominant.
This also highlights the importance of pairing QCA with a computational standard hydrogen electrode or similar approach which allows for inclusion of the voltage in the calculation of the ionic energy.
We next consider an interface that includes grain boundaries.
We again first measure the Fermi level of the oxide-solution system as $ \varepsilon_{\text{F,fb}}^{\text{ox.}|\text{sol.}} = -5.54$ eV and the metal-oxide system as $\varepsilon_{\text{F,fb}}^{\text{m.}|\text{ox.}} = -6.1$ eV.
Using the same procedure as before we can thus find a flatband potential for the grain boundary system of $\Phi_{\text{fb}} \approx 1.4 $ V (SHE).
We then insert oxygen vacancies or chlorine atoms into the structure and allow the system to relax.
It should be noted that due to periodic boundary conditions, the density of grain boundaries is much higher in this system than would be realistically expected.
We again confirm with Bader charge analysis that the Cl atom is acting as a Cl$^{-}$ ion within the system.
\begin{figure}
\centering
\includegraphics[width=0.6\textwidth]{al2o3-chloride-insetion-qca.pdf}
\caption{A stability diagram indicating the calculated insertion energy of a chloride ion as a function of voltage.
When the chloride ion is inserted into a grain boundary it becomes stable (i.e. $< 0$ eV, indicated by the horizontal black dashed line) at voltages $> -0.2$ V (SHE).
This compares favorably with the experimentally measured onset of chloride pitting at -0.5 V (SHE) indicated by the vertical red dashed line.
In contrast, in systems without grain boundaries, chloride ion insertion is not favorable until much higher voltages. }
\label{fig:stability-diagram}
\end{figure}
As seen in Fig.~\ref{fig:stability-diagram}, we then test a variety of voltages across the system with grain boundaries.
We predict that chloride ion insertion becomes stable at voltages $> -0.2$ V (SHE).
This correlates relatively well with the previous experiments showing pitting corrosion at voltages $> -0.5$ V (SHE) \cite{dibari1971electrochemical,bessone1992eis} (see Sec.~\ref{sec:discussion} for a discussion on comparing theoretical and experimental voltages).
It also matches previous results from Leung, who similarly finds that chloride ion insertion is only favorable at grain boundaries within the aluminum oxide \cite{leung2021first}.
This work emphasizes the role that grain boundaries likely play in pitting corrosion according to the point defect model, serving as locations for chloride insertion into the aluminum oxide and serving as corrosion centers.
This results also follows intuition: Cl$^-$ is a larger anion than oxygen and the extra room within the grain boundary allows for a more favorable insertion energy.
\begin{figure}
\centering
\includegraphics[width=0.6\textwidth]{al2o3-charge_density.pdf}
\caption{The equilibrium charge density at the interface with a Cl$^{-1}$ ion inserted both with and without grain boundaries.
Here, negative values are taken to indicate excessive electrons. }
\label{fig:chg-density}
\end{figure}
The voltage dependence of the charge density, shown in Fig.~\ref{fig:chg-density}, also matches the intuitive picture of chloride stability in grain boundaries we have established.
For interfaces with and without a grain boundary, the charge density is largely linear as a function of voltage.
This reflects the fact that, as previously noted, the majority of the charge in our simulations of the system is caught at the oxide/solution interface, causing the interface to behave in a more metallic manner, particularly away from the flatband potential.
Furthermore, the interface with grain boundaries consistently has lower charge densities at voltages where the Cl$^{-1}$ ion is predicted to be stable, somewhat matching the intuition that more charge is allowed in the system.
(In this notation, ``lower'' charge densities indicate a greater density of electrons induced at the interface.)
This is only violated at $\Phi \lessapprox$ -0.2 V (SHE), corresponding to the location where a Cl$^{-1}$ ion is no longer stable in the system.
Below this voltage, the charge density for the two systems becomes nearly equal.
\section{Discussion}
\label{sec:discussion}
While QCA provides a useful tool for attaching relative voltages to DFT calculations of interfaces, it does not solve a variety of outstanding problems in defining the electronic voltage of a neutral DFT system relative to experiment.
In particular, the arrangement of water molecules within the first monolayer of coverage has been shown to have a large impact on the estimated flatband potential, with previous studies on this system showing the impact of different H$_2$O configurations being 0.2-0.5 eV \cite{sakong2016structure,sakong2018electric}.
This is a difficult problem to overcome as there is no clearly ``correct'' configuration of the water molecules on the surface.
One possible solution would be to average over several possible H$_2$O configurations sampled from an AIMD calculation and attempt to form some averaging.
As indicated previously, another concern is the relative concentration of charged defects being examined.
The periodic nature of these DFT calculations ensures that the 2D density of any defects will be relatively high.
Furthermore, realistic systems likely have larger macrostructure, with grain boundaries and defects that are not actually periodic in nature.
This can be partially overcome by creating larger lateral system sizes and performing convergence tests.
Increasing the lateral size of the interfaces typically necessitates a a tradeoff of the length of the slab to keep calculations computationally feasible.
This shorter slab can still be accommodated to provide useful results with QCA provided that a few layers have been included around the defect and interface.
These twin uncertainties make exact comparison of our predicted chloride insertion onset voltage to macroscopic potentials extracted from experiment problematic.
Our QCA predictions of exact potentials should then be taken as semi-quantitative about the system, not exact.
For instance, the finding that grain boundaries within the aluminum oxide make chloride insertion much more favorable are consistent regardless of the exact voltage at which they become favorable.
Similarly, if our results predicted that chloride insertion was only favorable only at voltages $>$ 3 V (SHE), it would be likely that we were not correctly representing the interface within DFT (or that QCA itself was an incorrect framework for the problem).
When predicting relative voltage differences in the same system (i.e. the difference between two voltages predicted using the same surface termination), however, QCA should be more rigorous as the previous drawbacks to exact work function determination are held constant.
QCA nevertheless presents an advantage over other computational approaches to controlling the voltage of an electrode by coupling DFT to continuum and Poisson-Boltzmann descriptions of the solution.
This is namely the ability to include an extended insulating (oxide) region within the simulation.
Previous work has needed to focus on metals with a linear relationship between the charge and voltage of the system, consistent with a constant capacitance model \cite{schnur2011challenges,keilbart2017quantum,haruyama2018electrode,kano2021study}.
While these approaches can be quite useful, their utility can be limited in comparison to experiment, especially if large passivating layers are expected to form over the metal electrode and the capacitance can no longer be expected to be constant as a function of potential.
Given that the contact layer between metal and oxide has been shown to sometimes have barriers as high as 2-3 V \cite{yu2019combining,cornette2020relation}, the impact of the extended oxide layer is vital to fully understanding the voltage behavior of an electrode.
QCA thus presents a computationally scalable method for addressing voltages in systems with extended insulating regions.
\section{Conclusion and Outlook}
In this work, we have extended the QCA methodology to encompass metal-oxide interfaces that are also in contact with a solution.
This extension allows us to provide first principles calculations of atomic mechanisms behind corrosion and other electrochemical processes as a function of applied voltage, which has previously been difficult to manage within a DFT framework.
We demonstrate this methodology on an Al/$\alpha$-Al$_2$O$_3$ interface, calculating that chloride insertion is only favorable above -0.2 V (SHE) when a grain boundary is included at the interface.
This conclusion is broadly consistent with both previous first principles work, and experimental measurements.
This provides helpful context for parameterizing the point defect model with first-principles results.
While QCA provides a useful method for examining interfacial problems that includes voltage, it does not resolve some of the traditional difficulties of DFT.
Band gaps will be underestimated in semiconductors and insulators when using semi-local functionals.
As long as the potential drop across the electrode is less than the band gap of the DFT material, however, this should not affect the relative voltage results.
Similarly, the exact atomic structures of both the metal-oxide and oxide-solution interfaces exert a large influence on the assigned flatband potential of the interfacial structure.
Numerous DFT methods have been developed to address this issue, with high temperature ab-initio molecular dynamics providing one option for generating structures.
Larger surveys of possible structures should still be made for QCA analysis of any given system.
It nevertheless presents advantages over previous work in being able to examine systems with large passivation layers.
In this work, we demonstrated QCA in the context of corrosion of aluminum, but it can clearly also be used in numerous applications of interest.
In battery electrode research, for instance, the exact atomistic interaction of a solid-electrolyte interface (SEI) with Li ions shuttling through the structure is often not well understood as a function of voltage \cite{leung2020dft}.
QCA provides one potential route for exploring these systems while including a realistic voltage to the system.
Another application may be photocatalysis, where a semiconductor is placed in contact with solution to catalyze a reaction, often attached to a metal electrode to provide some level of voltage to the system \cite{wenderich2016methods}.
Beyond electrochemistry, a similar methodology can be applied to electronic interfaces of interest such as p-n junctions and Schottky barriers \cite{subramanian2020photophysics}.
Thus, QCA provides a new tool for computational electrochemistry across a broad variety of interfaces and applications, including voltage effects at reasonable computational cost.
\ack
We would like to thank Kevin Leung for providing the initial Al/Al$_2$O$_3$ system and a critical reading of the draft manuscript, as well as Andrew Baczewski for useful feedback.
This work was supported by the Laboratory Directed Research and Development program at Sandia National Laboratories under project 226354 and 229741.
Sandia National Laboratories is a multi-mission laboratory managed and operated by National Technology and Engineering Solutions of Sandia, LLC, a wholly owned subsidiary of Honeywell International, Inc., for DOE's National Nuclear Security Administration under contract DE-NA0003525.
This paper describes objective technical results and analysis.
Any subjective views or opinions that might be expressed in the paper do not necessarily represent the views of the U.S. Department of Energy or the United States Government.
\section*{References}
\bibliographystyle{iopart-num}
|
\section{Introduction}
The celebrated Brezis-Oswald theorem \cite{BO} states, among other things, that the problem
\begin{equation}
\label{bo}
\begin{cases}
-\Delta u=f(x, u) &\text{in $\Omega$}\\
u=0&\text{on $\partial\Omega$}\\
u\ge 0 &\text{in $\Omega$}
\end{cases}
\end{equation}
in a connected smooth bounded domain $\Omega\subseteq \mathbb{R}^N$ has a unique solution whenever $f(\cdot, t)\in L^\infty(\Omega)$ for all $t\ge 0$ and $t\mapsto f(x, t)/t$ is strictly decreasing in $t>0$, for a.\,e.\,$x\in \Omega$. This result has later been generalised to the quasi-linear setting in \cite{DS}, to the nonlocal one (see \cite{IM}) and even to mixed local-nonlocal problems in \cite{BMV}, producing countless applications in nonlinear analysis.
The general principle underlying Brezis-Oswald theorem appears to be the following. If $L$ denotes a (pseudo)-differential elliptic operator enjoying a strong minimum principle and which is {\em homogeneous} of degree $q$, then uniqueness of the solution
\begin{equation}
\label{Lq}
\begin{cases}
L(u)=f(x, u) &\text{in $\Omega$}\\
u=0&\text{on $\partial\Omega$}\\
u\ge 0 &\text{in $\Omega$}
\end{cases}
\end{equation}
should follow under the assumption that $t\mapsto f(x, t)/t^{q}$ is strictly decreasing. These kind of results, therefore, are called in the literature "Brezis-Oswald uniqueness theorems".
A number of additional assumptions, however, are generally required for such statements to hold, and we discuss here a few of them.
\begin{enumerate}
\item
{\em Regularity of the domain}.
Often, the domain $\Omega$ is assumed to have smooth (say, $C^{1, \alpha}$) boundary, so that a Hopf-type maximum principle holds true for the corresponding solutions. This usually requires rather fine regularity theory for solutions of the corresponding problems, and is usually avoided requiring additional technical assumptions on the reaction (see e.\,g.\,\cite{BMV}).
\item
{\em Growth conditions on $f$}.
The monotonicity assumption on $t\mapsto f(x, t)/t^q$ only grants an upper bound on the growth of $f$. Even if the operator $L$ is variational, namely $L= dE$ for a suitable differentiable convex functional $E:X\to \mathbb{R}$ on a Banach space $X\hookrightarrow L^p(\Omega)$, this poses delicate issues. Indeed, it is then natural to consider solutions of \eqref{Lq} as critical points of the functional
\[
J(v)=E(v)-\int_\Omega F(x, u)\, dx, \qquad F(x, t)=\int_0^t f(x, s)\, ds
\]
but without assuming any further lower growth control on $f$, $J$ is, typically, only lower semicontinuous. This problem is often side-stepped by either requiring a-priori boundedness in the notion of solution of \eqref{Lq} (so that actually $f$ can be truncated), or by directly requiring a two sided growth condition on $f$.
\item
{\em Monotonicity assumption}. It is clear that uniqueness fails even for \eqref{bo} without {\em strict} monotonicity, as seen by considering the case $f(x, t)=\lambda_1(\Omega)\, t$, where $\lambda_1(\Omega)$ is the first eigenfunction of the Laplacian with Dirichlet boundary conditions. In this case $t\mapsto f(x, t)/t$ is only non-increasing and the corresponding problem \eqref{bo} has infinitely many positive solutions, namely all first eigenfunctions of the Dirichlet Laplacian. However, at least for autonomous reactions independent of $x$, this is essentially the only obstruction to uniqueness. Indeed, in \cite[Proposition 3.8]{BMS1} it is proved that, for $L=-\Delta_{p} $ and $t\mapsto f(t)/t^{p-1}$ only {\em non-increasing} in $t>0$, either there is a unique solution to the corresponding problem \eqref{Lq}, or all solutions are first eigenfunctions for the Dirichlet $p$-Laplacian in $\Omega$. In particular, uniqueness holds true if $t\mapsto f(t)/t^{p-1}$ is non-increasing and strictly decreasing only on an arbitrarily small interval $]0, \delta[$, $\delta>0$.
\item
{\em Regularity of $L$}. As anticipated in point (2) above, the typical framework for a variational Brezis-Oswald uniqueness result is $L(u)=dE(u)$, for $E(u)=\|u\|_X^p$ and $\|\ \|_X$ is the norm of a Banach space $X\hookrightarrow L^p(\Omega)$, smooth away from the origin. Differentiability of the norm is typically used to have a single valued operator $L$ in equation \eqref{Lq}. Considering non-differentiable and only locally Lipschitz energies $E$ yields a differential inclusion formally of the form
\[
f(x, u)\in \partial E(u)
\]
for critical points of the relevant functional, where here $\partial$ denotes the multivalued convex subdifferential. In this more general setting, uniqueness results are clearly stronger statements.
\end{enumerate}
\subsection{Main results}
In this manuscript, we plan to address the previous themes, aiming to obtain a general version of the uniqueness part of Brezis-Oswald theorem. In particular, we seek for a classification of the critical points of functionals of the form
\begin{equation}
\label{J}
J(v)=\int_\Omega H(Du) - F(x, v)\, dx, \qquad v\in W^{1,p}_0(\Omega)
\end{equation}
under fairly general conditions on $\Omega$, $H$, and $f(x, t)=\partial_t F(x, t)$. Our main result reads as follows (for a more complete statement, see Theorem \ref{thfin}).
\begin{theorem}\label{Mth}
Let $\Omega\subseteq \mathbb{R}^N$ be open, connected and of finite measure.
For some $p>1$ suppose that
\begin{itemize}
\item[i)] $H:\mathbb{R}^N\to [0, \infty[$ is strictly convex and positively $p$-homogeneous
\item[ii)] $f:\Omega\times \mathbb{R}\to \mathbb{R}$ is Caratheodory and fulfils
\[
\sup_{x\in \Omega, |s|\le t} |f(x, s)|<\infty\qquad \forall t>0
\]
\item[iii)]
The map $ t\mapsto f(x, t)/t^{p-1}$ is non-increasing on $]0, \infty[$.
\end{itemize}
For $J$ as in \eqref{J} with
\[
F(x, t)=\int_0^t f(x, s)\, ds,
\]
let $C_+(J)$ be the set of its non-negative critical points $v\in W^{1,p}_0(\Omega)\setminus\{0\}$. Then
\begin{enumerate}
\item
All elements in $C_+(J)$ are multiple of each other and each of them minimise $J$ among non-negative functions in $ W^{1,p}_{0}(\Omega)$
\item
If $t\mapsto f(x, t)/t^{p-1}$ is strictly decreasing on $t>0$, $C_+(J)$ contains at most one element
\item
If $f$ does not depend on $x$ and $C_+(J)$ contains more than one element, each function in $C_+(J)$ minimises the Rayleigh quotient
\begin{equation}
\label{l0}
\frac{\displaystyle{\int_\Omega H(Dv)\, dx}}{\displaystyle{\int_\Omega v^p\, dx}}
\end{equation}
among all $v\in W^{1,p}_0(\Omega)\setminus \{0\}$, $v\ge 0$.
\end{enumerate}
\end{theorem}
\begin{remark}
Let us make some comments on the features of the previous theorem.
\begin{itemize}
\item
No regularity assumptions are made on the boundary of the open set $\Omega$. The finite measure assumption could be further weakened requiring additional conditions on the behaviour of $f(x, t)$ near $0$, but we chose not to analyse this case in order to have a cleaner statement. Notice that in this setting, there is no hope to have a Hopf maximum principle available.
\item
The stated assumption on $H$ enforces the two sided bound
\[
c\, |z|^p\le H(z)\le d\, |z|^p, \qquad \forall z\in \mathbb{R}^N
\]
for some $0<c\le d<\infty$. However, since $H$ is not supposed to be $C^1$, the first term of $J$ in \eqref{J} is only locally Lipschitz, with multi-valued differential. Notice also that $H$ is not supposed to be even. The strict convexity of $H$ is also not needed in assertion {\em (2)}, as well as in the second statement of assertion {\em (1)}.
\item
Conditions {\em ii)} and {\em iii)} on the reaction $f$ are equivalent to
\[
f(\cdot, t)\in L^\infty(\Omega), \qquad f(x, t)\le C\, (1+t^{p-1}), \qquad t\mapsto \frac{f(x, t)}{t^{p-1}} \ \text{non-increasing}
\]
for $t> 0$, which are more commonly found in Brezis-Oswald type results.
\item
Under the s\^ole assumptions {\em ii)} and {\em iii)}, even for smooth $H$, $J$ is only l.\,s.\,c., and the notion of critical point has to be dealt with care. We employ here the notion of {\em energy critical point} used in \cite{DGM}, meaning those $u\in W^{1,p}_0(\Omega)$ which, roughly speaking, minimise the corresponding semi-linearised convex functional
\[
v\mapsto \int_\Omega H(Dv) - f(x, u)\, v\, dx.
\]
For the sake of this introductory discussion, let us note that the usually encountered notions of critical point, when available, yield an energy critical point.
For more details, we refer to Definition \ref{defecp} and Remark \ref{remc}.
\item
Notice that we don't give any condition ensuring that $C_+(J)$ is actually not empty. One of the main features of the original Brezis-Oswald result (and of the related literature on the subject) was actually to provide necessary and sufficient condition for $J$ to have non-negative, non-trivial critical points, under a strict monotonicity assumption on $t\mapsto f(x, t)/t^{p-1}$. We won't investigate this theme here.
\item
Under assumption {\em iii)} alone, the structure of $C_+(J)$ is described in an optimal way by conclusion {\em (1)} of the theorem. Indeed, when $H(z)=|z|^2$, given a first positive eigenfunction $u$ of the Dirichlet Laplacian in the unit ball $B_1$ and an interval $[a, b]\subseteq [0, +\infty[$, any reaction $f$ obeying $f(x, t)=\lambda_1(B_1)\, t$ if $a\, u(x)\le t\le b\, u(x)$ and $f(x, t)/t$ strictly decreasing otherwise has $C_+(J)=\{t\, u: t\in [a, b]\}$.
\item
A similar statement holds true for non-positive critical points of $J$, with obvious modifications to {\em ii)} and {\em iii)}. Notice, however, that since we are not assuming that $H$ is even, the infimum of the ratios \eqref{l0} may change under the corresponding constrain $v\le 0$. This feature of possibly non-even, positively $p$-homogeneous energies $H$ and their corresponding first Dirichlet eigenvalues is discussed in section \ref{DE}.
\end{itemize}
\end{remark}
In order to prove Theorem \ref{Mth}, we need to develop basic regularity theory for energy critical points. The main step is contained in the following {\em a-priori} bound, which is of independent interest. In its statement, we let as usual $p^*=N\, p/(N-p)$ if $p<N$, $p^*=\infty$ if $p>N$ and $p^*$ be any finite number if $p=N$.
\begin{theorem}\label{Mth2}
Let $\Omega\subseteq \mathbb{R}^N$ have finite measure and, for given $p\ge 1$, suppose that
\begin{itemize}
\item[i)] $H:\mathbb{R}^N\to [0, \infty[$ is convex and fulfils $H(z)\ge c\, |z|^p$ for some $c>0$ and all $z\in \mathbb{R}^N$
\item[ii)] $f:\Omega\times \mathbb{R}\to \mathbb{R}$ is Caratheodory and fulfils
\[
\sup_{x\in \Omega, |s|\le t} |f(x, s)|<\infty\qquad \forall t>0
\]
\item[iii)]
There exist $\mu, \theta\ge 0$ and $1\le q\le p^*$ such that
\begin{equation}
\label{fgh}
{\rm sign}\,(t) f(x, t)\le \mu\, (|t|+\theta)^q\qquad \text{for a.\,e.\,$x\in \Omega$ and all $t\in \mathbb{R}$}.
\end{equation}
\end{itemize}
Then, any critical point $u\in W^{1,p}_0(\Omega)$ of $J$
is bounded. If additionally $q<p^*$, the estimate
\begin{equation}
\label{supboundi}
\|u\|_{L^\infty(\Omega)}\le
\max\left\{\theta, C\,\left((\mu/c)^{\frac{N}{p}}\int_\Omega |u|^q\, dx\right)^{\frac{p}{p\, q+ (p-q)\, N}}\right\}
\end{equation}
holds true for some $C=C(N, p, q)$ which, for any $N$, is positive and continuous in its other arguments.
\end{theorem}
\begin{remark}
Regarding this second result, the following should be noted.
\begin{itemize}
\item
Not only $H$ is only assumed to be locally Lipschitz, but no growth control from above is required in order to obtain boundedness of the critical points of $J$.
\item
Assumption {\em iii)} is understood to be empty in the case $q=\infty$, which forces $p>N$ thanks to $q\le p^*$. Clearly, in this case boundedness is trivial thanks to Morrey embedding. Anyway, \eqref{supboundi} only holds for $q<p^*$ and therefore for finite $q$'s.
\item
Assumption {\em ii)} and {\em iii)} are required in order for the functional $J$ in \eqref{J} to be well defined (possibly assuming the value $+\infty$) on $W^{1,p}_0(\Omega)$. In any case, under these assumptions $J$ turns out to be only l.\,s.\,c.\,on $W^{1,p}_0(\Omega)$ with the strong topology, and again we understand by critical point an energy one (see subsection \ref{nscp}).
\item
The form \eqref{supboundi} of the sup bound is optimal, as discussed in Remark \ref{remsup}. Moreover, it is invariant by scaling and scalar multiples, meaning the following. A function $u$ is a critical point for $J$ on $W^{1,p}_0(\Omega)$ if and only if for any $\lambda, r>0$ $u_{\lambda, r}(x)=\lambda\, u(x/r)$ is a critical point for a suitable functional $J_{\lambda, r}$ on $W^{1,p}_0(r\, \Omega)$ (where $r\, \Omega=\{r\, x: x\in \Omega\}$). More precisely, the functional $J_{\lambda, r}$ is of the form \eqref{J} for a reaction $\hat{f}$ obeying \eqref{fgh} with $\hat{\theta}=\lambda\, \theta$, $\hat{\mu}=\mu/\lambda^q$ and and $\hat{H}$ obeying $\hat{H}(z)\ge \hat{c}\, |z|^p$ with $\hat{c}=c\, \lambda/r$. The corresponding estimates \eqref{supboundi} turn out to be independent of $\lambda$ and $r$.
\item
Regarding the continuity statement for the constant $C$ in the last part of the theorem, it is to be understood in its domain $p\ge 1$, $q<p^*$. As is to be expected, the constant $C$ blows up as $q\uparrow p^*$.
\end{itemize}
\end{remark}
\subsection{Related results and motivations}
Applications and extensions of the Brezis-Oswald theorem are countless and it would be impossible to list all the contributions on the subject.
The principal part of the operator considered in the various generalisation can be classical quasi-linear operators, as in \cite{DS}, nonlocal operators as in \cite{IM} and even mixed local-nonlocal ones as in \cite{BMV}. In all these cases the operator is $(p-1)$-homogeneous and the corresponding monotonicity assumption on the reaction pertains $t\mapsto f(x, t)/t^{p-1}$. Non-homogeneous operators of Orlicz type are considered in \cite{CGSS}. The case when $t\mapsto f(x, t)/t^{p-1}$ is non-increasing but not strictly decreasing is naturally related to weighted nonlinear eigenvalue problems, where $f(x, t)=a(x)\, t^{p-1}$. Corresponding eigenfunctions are considered in \cite{BF,T, TTU} when the energy density $H$ is $p$-homogeneous and $C^1$, while the case where $H$ may be non-differentiable is studied in \cite{LS}. Various applications to nonlinear parabolic problems, encompassing the non-differentiable setting, are described in \cite{D}.
A complete Brezis-Oswald theorem usually consists of two statements. The first one is devoted to derive necessary and sufficient condition for the solvability of \eqref{Lq}, under the aforementioned monotonicity assumption of $t\mapsto f(x, t)/t^{p-1}$, as is done in \cite{BO, DS, IM}. The second one regards uniqueness of the solution, and is clearly the focus of this research.
As briefly outlined in the first section, our main intent with this investigation was to clarify how the various technical assumptions which usually appear in Brezis-Oswald type result could be weakened to have a cleaner picture, especially regarding its uniqueness part. It is clear, for instance, that the regularity of the domain or of $H$ is not necessary for the uniqueness statement to hold. Our interest in this part of Brezis-Oswald theorem, however, lies mainly in its applications to the qualitative study of solutions to PDE's.
Indeed, uniqueness is the basic step needed to effectively approximate the particular PDE problem under scrutiny of the form \eqref{Lq} with solutions of regularised problems
\begin{equation}
\label{Lqn}
L_n(u)=f_n(x, u),
\end{equation}
each having a smooth solution $u_n$ allowing computations can be more effectively performed. The sequence $(u_n)$ then shown to enjoy uniform bounds and therefore to converge up to subsequences to a solution of the limiting, original problem \eqref{Lq}. Then, uniqueness of solutions of \eqref{Lq} is the key point ensuring that the approximating solutions actually converge to the original one, which then (hopefully) inherits some properties of the $u_n$'s. This scheme of proof is nowadays standard in both regularity theory and qualitative properties of solution to PDE's.
As a typical example, let us consider positive solutions of the quasi-linear problem
\[
-\Delta_pu =f(u)
\]
with Dirichlet boundary conditions and a qualitative property of these solutions known as {\em quasi-concavity}, meaning that the super-level sets of $u$ are convex.
It is quite fortunate that the monotonicity assumption on $t\mapsto f(t)/t^{p-1}$ is strictly connected to the general conditions found in \cite{BMS1} ensuring quasi-concavity of the aforementioned solutions. Actually, the results in \cite{BMS1, BMS2} are proved via the smoothing technique outlined above under the assumption that the domain is convex with $C^{1,\alpha}$ boundary, but Theorem \ref{Mth} allow to extend the results to general convex domains.
As a motivation for Theorem \ref{Mth2}, let us also stress that, in order for the previously described approximation scheme to work, uniform bounds are also essential to ensure compactness. This is especially true when, in the family of problems \eqref{Lqn}, both the operators $L_n$ and the reactions $f_n$ may have a varying growth control, only qualitatively bounded. In this regard, Theorem \ref{Mth2} is inspired by \cite{BL}, where the stability of the available a-priori bounds had to be treated with extreme care for the employed contradiction argument to carry on. In particular, \eqref{supboundi} should be compared with \cite[Proposition 2.4]{BL}, corresponding to the case $\theta=0$ in \eqref{fgh}. In \cite{BL}, stability with respect to $p$ wasn't an issue, and the a-priori bound was derived via a Moser iteration technique involving the $p$-Sobolev inequality. Since the constant in the latter blows up as $p\uparrow N$, the obtained bounds in \cite{BL} are not stable in this limit. To be honest, we initially hoped for this apparent instability to be substantial, suggesting the possibility of interesting phenomena exhibited by solutions of, say, the $p$-Laplace equations as $p\uparrow N$. Unfortunately, a different and, we think, more streamlined proof of the $L^\infty$ bound in the spirit of De Giorgi method, shows that this is not the case.
Let us finally mention that our proof of Theorem \ref{Mth}, detailed in the next section, relies on a non-smooth version of the Picone inequality which relies only on the positive $p$-homogeneity of $H$. This should be compared to \cite{J} where, however, differentiable convex $H$ are considered. A natural alternative route to the Brezis-Oswald uniqueness result, as in \cite{D, DS, TTU}, is to consider the hidden convexity properties of the functional $J$ in \eqref{J}. Indeed, the core assumption that $t\mapsto f(x, t)/t^{p-1}$ be non-increasing is implied by the weaker hypothesis that $t\mapsto F(x, t^{1/p})$ is concave. Notice that this latter condition does not require $F$ to be differentiable in its second variable, nor any bound on its derivative is needed to define the functional $J$. Concavity of $t\mapsto F(x, t^{1/p})$ implies {\em hidden convexity} of
\[
J_F(v)=-\int_\Omega F(x, v)\, dx,
\]
on the cone of non-negative $v\in W^{1,p}_0(\Omega)$, which means convexity of $J_F$ along the curves
\[
t\mapsto \left(t \, u^{p}+(1-t)\, v^{p}\right)^{1/p} \qquad \forall u, v\in W^{1,p}_0(\Omega), \quad u, v\ge 0.
\]
On the other hand, by \cite{BF}, our non-smooth Picone inequality is equivalent to the hidden convexity of $J_H$. Hence the whole $J$ turns out to be hidden convex on the cone of nonnegative functions in $W^{1,p}_0(\Omega)$ and it is only natural to expect few critical points from such a functional. We chose not to adopt this approach since our notion of energy critical point, while very weak, requires at least that the reaction $f(x, t)=\partial_t F(x, t)$ is well defined and Caratheodory, but we point out that the interplay between hidden convexity and the notion of critical point for l.\,s.\,c.\,functionals in the sense of \cite{DG} has been initiated in \cite{DGM2} in the framework of nonlinear eigenvalue problems.
\subsection{Outline of the proof}
The proof of Theorem \ref{Mth} is split in two main parts: regularity and uniqueness. The first obstacle to regularity is the lack of a proper Euler-Lagrange equation to work with, in particular when dealing with critical points of the l.\,s.\,c.\,functional $J$. Here we take advantage of the variational characterisation of critical points $u$ as minima of the semi-linearised functional
\[
J_u(v)=\int_\Omega H(Dv)-f(x, u)\, v\, dx.
\]
This allows to test the minimality of $J_u(u)$ against $J_u(u_k)$, where $u_k:=\max\{k, u\}$ for $k\ge 0$, and $J_u(u)\le J_u(u_k)$ reads
\[
\int_\Omega H(D(u-k)_+)\, dx\le \int_\Omega f(x, u)\, (u-k)_+\, dx.
\]
This inequality and the one-sided growth control on $f$ suffice to derive a "fast decay" relation of the $L^q$ norms of $(u-k)_+$ much in the spirit of De Giorgi regularity proof. However, in order to ensure stability of the estimates in $p$, we avoid the natural Sobolev inequality and instead use an interpolation inequality {\em \'a la} Gagliardo-Nirenberg with a more tamed constant (see Proposition \ref{PGN}). Once the $L^\infty$ bound of Theorem \ref{Mth2} is proved, the reaction is bounded and Giaquinta-Giusti hole-filling method \cite{GG} allows to prove that the solutions belong to suitable De Giorgi classes, thus providing continuity and a strong minimum principle for non-negative critical points. In particular, an Euler-Lagrange equation (or rather, inclusion) is satisfied by critical points, namely
\begin{equation}
\label{eli}
\int_\Omega (\xi, D\varphi)=\int_\Omega f(x, u)\, \varphi\, dx\qquad \forall \varphi\in W^{1,p}_0(\Omega)
\end{equation}
where $\xi:\Omega\to \mathbb{R}^N$ is a measurable choice of the multifunction $x\mapsto \partial H(Du(x))$ and $(\ , \ )$ denotes the Euclidean scalar product.
With these regularity tools at our disposal we can prove the uniqueness part. First we derive a non-smooth version of Picone inequality in Lemma \ref{thpicone}. If $H$ positively $p$-homogeneous and convex, the point-wise inequality
\begin{equation}
\label{as}
p\, H(z)\ge (\xi, z), \qquad \forall \xi\in \partial H(z)
\end{equation}
holds true.
The non-smooth Picone inequality follows from the latter and reads as follows: if $u$ and $v$ are positive functions in $W^{1,1}_{\rm loc}(\Omega)$, then a.\,e.\,in $\Omega$ it holds
\[
p\, H(Dv)\ge \left( \xi, D \frac{v^p}{u^{p-1}}\right),
\]
for any measurable $\xi:\Omega\to \mathbb{R}^N$ such that $\xi(x) \in \partial H(Du(x))$ a.\,e.. Now, if $H$ is differentiable its convex subdifferential is single-valued, i.\,e.\,$\partial H=\{DH\}$ and actually \eqref{as} is an equality, so the smooth Picone inequality becomes
\[
\left( DH(Dv), Dv\right)\ge \left(DH(Du), D\frac{v^p}{u^{p-1}}\right).
\]
If both $u$ and $v$ are positive critical points, \eqref{eli} gives (cheating a little bit, since we are testing it for $u$ with $\varphi=v^p/u^{p-1}$)
\[
\int_\Omega f(x, v)\, v\, dx\ge \int_\Omega f(x, u)\, \frac{v^p}{u^{p-1}}\, dx.
\]
Exchanging the r\^ole of $u$ and $v$ and summing the corresponding inequalities, one arrives at
\begin{equation}
\label{pl}
\int_\Omega\left(\frac{f(x, u)}{u^{p-1}}-\frac{f(x, v)}{v^{p-1}}\right) (u-v)\, dx\ge 0
\end{equation}
which yields uniqueness when $t\mapsto f(x,t)/t^{p-1}$ is strictly decreasing.
The point, in the non-differentiable case, is that in general the inequality in \eqref{as} can be strict. However, it turns out that the measurable selection $\xi$ of $\partial H(Du)$ arising from \eqref{eli} actually attains equality in \eqref{as} almost everywhere. This property is related to our variational notion of critical point, and allows to carry on the proof as us usual.
Finally, the characterisation of the set of critical points described in Theorem \ref{Mth} under the non-increasing assumption on $t\mapsto f(x, t)/t^{p-1}$ is based on the analysis of the equality case in the non-smooth Picone inequality, derived from the identity
\[
\frac{f(x, u)}{u^{p-1}}=\frac{f(x, v)}{v^{p-1}}
\]
which must hold a.\,e.\,if $t\mapsto f(x, t)/t^{p-1}$ is non-increasing, thanks to \eqref{pl}.
\subsection{Structure of the paper}
Section \ref{Spre} contains preliminary material: we collect the notations used in the paper, provide some basic functional inequalities, derive basic properties of convex, positively $p$ homogeneous functions and describe the notion of energy critical points and its basic features. Section \ref{Sregularity} contains the proof of Theorem \ref{Mth2} and some of its consequences, such as continuity and minimum principles for energy critical points. In Section \ref{DE} we define and study the first Dirichlet eigenvalues associated to possibly non-even convex energies, aiming at the proof of point {\em (3)} of Theorem \ref{Mth}. In the final section we prove the non-smooth Picone inequality and Theorem \ref{Mth}.
\bigskip
\begin{center}{\sc Acknowledgements}
\end{center}
The author is member of {\em Gruppo Nazionale per l'Analisi Ma\-te\-ma\-ti\-ca, la Probabilit\`a e le loro Applicazioni} (GNAMPA),
is partially supported by project PIACERI - Linea 2 and 3 of the University of Catania and by GNAMPA's project "Equazioni alle derivate parziali di tipo ellittico o parabolico con termini singolari". Prof. M. Degiovanni is warmly thanked for outlining a mistake in a first version of the manuscript and for drawing our attention to paper \cite{DGM2}. Prof. J.\,I.\,D\'iaz is also thanked for kindly pointing out reference \cite{D}.
\section{Preliminaries}\label{Spre}
\subsection{Notations}
Throughout the paper we will make use of the following notations.
\medskip
Given a number $a\in \mathbb{R}$, we set $a_+=\max\{a, 0\}$ and $a_-=-\min\{0, a\}$ and similarly for real valued functions.
For every $v \in \mathbb{R}^N$ we denote by $|v|$ its Euclidean norm while $(v, w)$ stands for the scalar product, $v, w \in \mathbb{R}^N$. The symbol $B$ denotes a generic ball with unspecified center and radius while $B_r$ denotes a ball of radius $r$ centred at the origin. $S_1$ stands for the unit sphere $\partial B_1$.
\medskip
For a Lebesgue measurable $E\subseteq \mathbb{R}^N$, $|E|$ denotes its measure and we set $\omega_N=|B_1|$. We will omit the domain of integration when it is the whole $\mathbb{R}^N$, if this cause no confusion. For a measurable $E\subseteq \mathbb{R}^N$, ${\mathcal L}(E; \mathbb{R}^M)$ will denote the set of Lebesgue measurable functions from $E$ to $\mathbb{R}^M$. Given an open set $\Omega\subseteq \mathbb{R}^N$, $f$ will denote a Caratheodory function $f:\Omega\times \mathbb{R}\to \mathbb{R}$ and
\begin{equation}
\label{Ff}
F(x, t)=\int_0^tf(x, s)\, ds\qquad \forall t\in \mathbb{R}.
\end{equation}
Given $u\in {\mathcal L}(\Omega; \mathbb{R})$, we'll denote by $f(\cdot, u)$ the measurable function $x\mapsto f(x, u(x))$ and similarly for $F(\cdot, u)$. The Lebesgue spaces are denoted as usual by $L^p(E)$ for $p\ge 1$, with norm $\|\cdot \|_{L^p(E)}$, while $\sup_E u$ and $\inf_E u$ stand for the essential supremum and infimum respectively. The Lebesgue spaces of vector valued functions $u:E\to \mathbb{R}^M$, will be denoted by $\left(L^p(E)\right)^M$ with norms $\| |u| \|_{L^p(E)}$, where $|u|$ is the standard Euclidean norm of $\mathbb{R}^M$. Furthermore, we set for the sake of notational simplicity $\|f\|_p:= \|f\|_{L^p(\mathbb{R}^N)}$.
\medskip
Given $p\in [1, \infty]$, $p'$ denotes its conjugate, namely $p'=p/(p-1)$ if $p\in \, ]1, \infty[$, $1'=\infty$ and $\infty'=1$. Given an integer $N\ge 2$ and $p\in [1, \infty]$, we denote by
\[
p^*=
\begin{cases}
\dfrac{N\, p}{N-p}&\text{if $p<N$}\\
\text{any positive number}&\text{if $p=N$}\\
\infty&\text{if $p> N$}.
\end{cases}
\]
For an open $\Omega\subseteq \mathbb{R}^N$ with finite measure, the Sobolev space $W^{1,p}_0(\Omega)$ is the closure of $C^\infty_c(\Omega)$ w.\,r.\,t.\,the seminorm $\|Du\|_p$. In the finite measure case, $W^{1,p}_0(\Omega)$ is a Banach space, continuously embedded in $L^q(\Omega)$ for all $q\in [1, p^*]$. Its dual will be denoted by $W^{-1, p'}(\Omega)$, with duality pairing denoted by $\langle \Lambda, \varphi\rangle$ for $\Lambda\in W^{-1, p'}(\Omega)$ and $\varphi\in W^{1,p}_0(\Omega)$.
If $p<N$, the homogeneous Sobolev space $D^{1, p}(\mathbb{R}^N)$ is the space of functions in $L^1_{\rm loc}(\mathbb{R}^N)$ with distributional derivative in $L^p(\mathbb{R}^N)$ and which {\em vanish at infinity}, meaning that
\[
|\{|u|>k\}|<\infty\qquad \text{ for all $k>0$}.
\]
\subsection{Functional inequalities}
Functions in $D^{1,p}(\mathbb{R}^N)$ for $1\le p<N$ enjoy the Sobolev inequality
\begin{equation}
\label{Sob}
\|u\|_{p^*}\le C(N, p)\, \|Du\|_p.
\end{equation}
We are particularly interested in the $p=1$ case, for which $C(N, 1)=N^{-1}\, \omega_N^{-1/N}$ for all $N\ge 2$.
The following result is probably well known, but we couldn't find a precise references and thus provide its proof.
\begin{proposition}
For $1\le p, q<\infty$, let $X_{p, q}(\mathbb{R}^N)$ be the vector space of functions in $L^q (\mathbb{R}^N)$ with distributional derivative in $L^p(\mathbb{R}^N)$, equipped with the norm $\|u\|=\|Du\|_p+\|u\|_q$. Then $C^\infty_c(\mathbb{R}^N)$ is dense in $X_{p, q}(\mathbb{R}^N)$.
\end{proposition}
\begin{proof}
The set $C^\infty(\mathbb{R}^N)\cap X_{p, q}$ is dense in $X_{p, q}$ by known results (see \cite[Theorem 1, Section 1.1.5]{Mazya}), so it suffices to show that any $u\in C^\infty(\mathbb{R}^N)\cap X_{p, q}$ can be approximated by functions in $C^\infty_c(\mathbb{R}^N)$. Fix a cutoff $\eta\in C^\infty_c(B_2; [0, 1])$ such that $\eta\equiv 1$ in $B_1$ and, for any $u\in C^\infty(\mathbb{R}^N)\cap X_{p, q}$, define for $k\in \mathbb{N}$
\[
\eta_k(x)=\eta(x/k), \qquad u_k = u\, \eta_k,\qquad A_k=B_{2\, k}\setminus B_k,
\]
so that $u_k\in C^\infty_c(\mathbb{R}^N)$, ${\rm supp}\, (D\eta_k)\subseteq A_k$ and $\|D\eta_k\|_\infty=\|D\eta\|_\infty/k$. Clearly $\|u-u_k\|_q\le \|u\|_{L^q(\mathbb{R}^N\setminus B_k)}\to 0$ as $k\to \infty$, so we only have to estimate $\|D(u-u_k)\|_p$. It holds
\begin{equation}
\label{split}
\begin{split}
\int |Du-Du_k|^p\, dx&\le C_p\int |Du|^p\, (1-\eta_k)^p\, dx+ C_p\, \int |u|^p\, |D\eta|^p\, dx\\
&\le C_p\, \int_{\mathbb{R}^N\setminus B_k} |Du|^p\, dx+ \frac{C_p\, \|D\eta\|_\infty^p }{k^p}\, \int_{A_k} |u|^p\, dx.
\end{split}
\end{equation}
and the first term on the right vanishes for $k\to \infty$ since $|Du|\in L^p(\mathbb{R}^N)$, so we are left to estimate the second term.
The case $p<N$ is standard: by Chebichev inequality any $u\in L^q(\mathbb{R}^N)$ vanishes at infinity, hence $X_{p, q}\subseteq D^{1,p}(\mathbb{R}^N)\subseteq L^{p^*}(\mathbb{R}^N)$.
Therefore, by H\"older inequality
\[
\int_{A_k} |u|^p\, dx\le \left(\int_{A_k}|u|^{p^*}\, dx\right)^{\frac{p}{p^*}}\, |A_k|^{1-\frac{p}{p^*}}=C\, k^p \, \left(\int_{A_k}|u|^{p^*}\, dx\right)^{\frac{p}{p^*}},
\]
and the last factor on the right vanishes as $k\to \infty$ due to $u\in L^{p^*}(\mathbb{R}^N)$, hence also the last term in \eqref{split} does so. If $p\ge N$, we use the inequality
\[
\int_{B_2\setminus B_1} |v|^p\, dx\le C\, \int_{B_2\setminus B_1} |Dv|^p\, dx +C\, \left(\int_{B_2\setminus B_1} |v|^q\, dx\right)^{p/q}
\]
which is trivial by H\"older when $q\ge p$ and holds true thanks to Gagliardo-Nirenberg inequality when $q<p$. Applying it to $v(x)=u(k\, x)$ and changing variables, we get
\[
\frac{1 }{k^p}\, \int_{A_k} |u|^p\, dx\le C\, \int_{A_k} |Du|^p\, dx+ \frac{C\, k^{N\, \left(1-\frac{p}{q}\right)}}{k^p}\left(\int_{A_k} |u|^q\, dx\right)^{p/q}.
\]
As $k\to \infty$, the first term on the right vanishes again, as well as the second, since
\[
N\left(1-\frac{p}{q}\right)< p\qquad \forall p\ge N, q<\infty.
\]
\end{proof}
The next proposition is a particular case of the general Gagliardo-Nirenberg inequality. In order to get stable $L^\infty$ bounds in Section \ref{Sregularity}, however, we have to pay special attention to the dependance from the parameters and thus give a direct proof depending only on \eqref{Sob}.
\begin{proposition}[Gagliardo-Nirenberg]\label{PGN}
Let $N\ge 2$, $u\in W^{1,1}_{\rm loc}(\mathbb{R}^N)$ and $1\le p, q<\infty$. Then for any $u\in W^{1, 1}_{\rm loc}(\mathbb{R}^N)$ it holds
\begin{equation}
\label{GN}
\left(\int |u|^{\frac{N\, (q\, p+p-q)}{(N-1)\, p} }\, dx\right)^{\frac{N-1}{N}}\le G_{p, q, N}\left(\int |Du|^p\, dx\right)^{1/p}\left(\int |u|^q\, dx\right)^{1/p'},
\end{equation}
where
\[
G_{p, q, N}=\frac{1}{N\, \omega_N^{1/N}}\, \frac{q\, p+p-q}{p}.
\]
\end{proposition}
\begin{proof}
We can suppose that the right-hand side is finite, which amounts to $u\in X_{p, q}$. By the previous Proposition we can also suppose that $u\in C^\infty_c(\mathbb{R}^N)$.
For $p=1$ \eqref{GN} reduces to \eqref{Sob}, so we can suppose that $p>1$.
Let
\[
r=\frac{q}{p'}+1=\frac{q\, p+p-q}{p}
\]
and apply \eqref{Sob} to $v=|u|^r$ and then H\"older inequality, to get
\[
\begin{split}
\left(\int |u|^{\frac{q\, p+p-q}{p} 1^*}\, dx\right)^{1/1^*}&\le \frac{1}{N\, \omega_N^{1/N}}\, \int |D|u|^r|\, dx=\frac{1}{N\, \omega_N^{1/N}}\, r\, \int |u|^{r-1}\, |Du|\, dx\\
&\le \frac{1}{N\, \omega_N^{1/N}}\, r\, \left(\int |Du|^p\, dx\right)^{1/p}\left(\int |u|^{q}\, dx\right)^{1/p'}.
\end{split}
\]
\end{proof}
\subsection{Convex $p$-homogeneous functions}
Let $X$ be a Banach space in duality with $X^*$ via the duality pairing $\langle\cdot, \cdot\rangle$. Given a continuous convex function $C:X\to \mathbb{R}$ (we will alway restrict to this setting) its {\em subdifferential} at $x\in X$ is
\[
\partial C(x)=\left\{ x^*\in X^*: C(x+v)- C(x)\ge \langle x^*, v\rangle \ \forall v\in X\right\}.
\]
When $X=\mathbb{R}^N$ we'll identify $X$ with $X^*$ via the scalar product as a duality pairing, hence in this case $\partial C(x)\subseteq \mathbb{R}^N$.
Throughout the paper, $H:\mathbb{R}^N\to [0, \infty[$ will denote a convex function. Since its domain is $\mathbb{R}^N$, any such function is locally Lipschitz. The function $H$ is positively $p$-homogeneous for some $p\ge 1$ if
\[
H(\lambda\, z)=\lambda^p\, H(z)\qquad \text{ for all $\lambda>0$ and $z\in \mathbb{R}^N$}.
\]
Notice that we do not assume regularity of $H$, nor that $H$ is even.
In the positively $p$-homogeneous case the whole $H$ is completely determined by its {\em associated ball}, i.\,e.\,the convex closed set
\[
B_H=\{z\in \mathbb{R}^N: H(z)\le 1\},
\]
which in general is not symmetric around the origin but always contains $0$ in its interior. Vice-versa, given any convex $K$ containing $0$ in its interior, the function
\[
H_K(z)=\left(M_K(z)\right)^p
\]
where $M_K$ is the Minkowski functional of $K$
\[
M_K(z)=\inf \{ t>0: z/t\in K\},
\]
is a convex, positively $p$-homogeneous function with associated ball $K$. The set of symmetries of $H$ is the set of symmetries of its associated ball, namely
\[
{\rm Sym}\, (H)=\{T\in O(N): T(B_H)=B_H\}
\]
where $O(N)$ is the orthogonal group in dimension $N$. By construction, for any $T\in {\rm Sym}\, (H)$ it holds $H\circ T=H$.
\begin{proposition}\label{Lcoer}
Let $H:\mathbb{R}^N\to [0, \infty[$ be convex and positively $p$-homogeneous for some $p\ge 1$. Then for any $w\in \mathbb{R}^N$ and $\xi\in \partial H(w)$, it holds
\begin{equation}
\label{phom1}
p\, H(z)\ge (\xi, z)\qquad \forall \xi\in \partial H(z).
\end{equation}
\end{proposition}
\begin{proof}
Let $\xi \in \partial H(z)$ for some $z\in \mathbb{R}^N$. By $p$-homogeneity and convexity
\[
t^p\, H(z)=H(t\, z)\ge H(z)+(\xi, t\, z-z)
\]
for all $t\ge 1$, while the left hand side coincides with the right hand side when $t=1$. Therefore
\[
\left.\frac{d}{dt} \big[ t^p\, H(z)-H(z)-(\xi, t\, z- z)\big] \right|_{t=1}\ge 0,
\]
which gives \eqref{phom1}.
\end{proof}
\subsection{Energy critical points} \label{nscp}
In this section we consider functionals of the form \eqref{J} with more general structural conditions than those required in Theorem \ref{Mth}. More precisely, we assume that
\begin{equation}
\label{hypH}
\begin{split}
\text{ $H:\mathbb{R}^N\to [0, +\infty[$ is convex and for some $p\ge 1$ and $0\le c\le d\le \infty$ obeys}\\
c\, |z|^p\le H(z)\le d\, |z|^p \qquad \forall z\in \mathbb{R}^N,\qquad \qquad\qquad\qquad \,
\end{split}
\end{equation}
\begin{equation}
\label{hypf0}
\text{$f:\Omega\times \mathbb{R}\to \mathbb{R}$ is Caratheodory and}\quad
\sup_{|s|\le t}\|f(\cdot, s)\|_{L^\infty(\Omega)}<\infty \quad \forall t>0,
\end{equation}
\begin{equation}
\label{hypf}
\begin{split}
\text{ there exist $\theta, \mu\ge 0$, $1\le q\le p^*$ such that} \\
{\rm sign}\,(t)\, f(x, t)\le \mu \, ( |t|+\theta)^{q-1} \qquad \forall t\in \mathbb{R}. \quad \quad
\end{split}
\end{equation}
Regarding $F$ as per \eqref{Ff}, it follows from \eqref{hypf0} that when $p>N$ then $F(\cdot, v)\in L^\infty(\Omega)$ for any $v\in W^{1,p}_0(\Omega)$ by Morrey embedding, while, when $p\le N$, from \eqref{hypf} we readily obtain the one-sided estimate
\[
\sup_\Omega F(\cdot, t)\le \frac{\mu}{q}\,\left(|t|+ \theta\right)^q.
\]
Therefore, as long as $\Omega$ has finite measure (which will be assumed henceforth),
\begin{equation}
\label{g}
\left(H(Dv)-F(\cdot, v)\right)_-\in L^1(\Omega) \qquad \forall v\in W^{1,p}_0(\Omega).
\end{equation}
and the functional $J$ in \eqref{J} is well defined on $W^{1,p}_0(\Omega)$, eventually assuming the value $\infty$ if $F(\cdot, v)\notin L^1(\Omega)$. Fatou's lemma ensures that $J$ is l.\,s.\,c.\,on $W^{1,p}_0(\Omega)$ with the strong topology (regardless of $p\ge1$), hence the non-smooth critical point theory of \cite{DG} can be applied. Notice also that if \eqref{hypf} is strengthened to hold for both $f$ and $-f$,
then $J$ turns out to be locally Lipschitz, since its second term is actually $C^1$ on $W^{1,p}_0(\Omega)$, allowing to employ Clarke critical point theory \cite{Clarke}. We will not pursue these approaches here for sake of brevity and will use the following notion of critical point, taken from \cite{DGM}.
\begin{defn}[Energy critical point]\label{defecp}
Suppose $\Omega\subseteq \mathbb{R}^N$ has finite measure. Under assumptions \eqref{hypH}, \eqref{hypf0} and, if $p\le N$, \eqref{hypf}, we say that $u\in W^{1,p}_0(\Omega)$ is an {\em energy critical point} for $J$, if $J(u)<\infty$ and $u$ minimises the convex functional
\begin{equation}
\label{defJu}
J_u(v)=\int_\Omega H(Dv) - f(x, u)\, v\, dx
\end{equation}
defined on the space
\[
V_u=\left\{v\in W^{1,p}_0(\Omega): f(\cdot, u)\, v\in L^1(\Omega)\right\}.
\]
\end{defn}
A few remarks on this notion of critical point are in order.
\begin{remark}
\label{remc}
\begin{itemize}
\item
Notice that an energy critical point $u$ is {\em assumed} to lie in $V_u$, i.\,e.\,$f(\cdot, u)\, u\in L^1(\Omega)$. The r\^ole of $p$ in the previous definition is connected to the validity of \eqref{g} (granted, as noted, by \eqref{hypf}), so that $J$ is at least well defined. It should be remarked that, at this level of generality, it has nothing to do with \eqref{hypH}, where one could also have $c=0$ and $d=\infty$.
Regardless, from $H\ge 0$ and the assumption $J(u)<\infty$, we see that any energy critical point $u$ fulfils $H(Du)\in L^1(\Omega)$. Let us also remark that if $\in W^{1,p}_0(\Omega)$ is assumed, or proved, to be bounded, assumption \eqref{hypf} is irrelevant as $J$ is always well defined as long as $\Omega$ has finite measure and in this case $V_u=W^{1,p}_0(\Omega)$.
\item
We stress here that the notion of energy critical point is a very general one as it does not involve any topological structure on $W^{1,p}_0(\Omega)$. As such, it is useless if one wants to develop an effective critical point theory, where topological considerations are paramount.
Under conditions \eqref{hypH} with $c>0$ and $p>1$, as well as \eqref{hypf0} and \eqref{hypf} if $p\le N$, a critical point for $J$ in the sense of \cite{DG} (with the strong topology on $L^{p^*}(\Omega)$) is an energy critical point (see \cite[Theorem 5]{DGM}). Similarly, if $d<\infty$ and $p>1$ in \eqref{hypH} and \eqref{hypf} holds for both $f$ and $-f$, then any critical point in the sense of Clarke for $J$ is an energy critical point.
\item
In general, \eqref{hypf0} and the finite measure assumption on $\Omega$ ensure that $f(x, u)\in L^1(\Omega)$, hence $V_u$ contains $C^\infty_c(\Omega)$ and is dense in $W^{1,p}_0(\Omega)$. In the case $d<\infty$ of \eqref{hypH}, being
\begin{equation}
\label{defJH}
J_H(v):=\int_\Omega H(Dv)\, dx
\end{equation}
convex and finite on $W^{1,p}_0(\Omega)$, it is locally Lipschitz, hence for any energy critical point $u$ it holds
\[
\int_\Omega f(x, u)\, (v-u)\, dx\le J_H(v)-J_H(u)\le C_r\, \|D(v-u)\|_{p}
\]
for all $v\in V_u$ with $\|D(v-u)\|_{p}\le r$. Therefore the linear functional
\[
V_u\ni v\mapsto \int_\Omega f(x, u)\, v\, dx
\]
has a unique extension $T$ to $W^{1,p}_0(\Omega)$ as an element in $W^{-1, p}(\Omega)$. Notice that while $\langle T, v\rangle$ is finite for all $v\in W^{1,p}_0(\Omega)$, this does not imply that $f(\cdot, u)\, v\in L^1(\Omega)$, unless a one-sided bound on $f(x, u)\, v$ holds true, enabling to apply Brezis-Browder theorem.
\end{itemize}
\end{remark}
\begin{proposition}\label{propositionEL}
Let $\Omega\subseteq\mathbb{R}^N$ be open with finite measure and assume \eqref{hypH} with $d<\infty$, \eqref{hypf0} and, if $p\le N$, \eqref{hypf}.
If $u\in W^{1,p}_0(\Omega)$ is an energy critical point for $J$, there exists $\xi\in L^{p'}(\Omega; \mathbb{R}^N)$ such that
\begin{equation}
\label{EL}
\begin{cases}
\text{$\xi(x)\in \partial H(Du(x))$}&\text{ for a.\,e.\,$x\in \Omega$}\\[10pt]
\displaystyle{\int_\Omega \left(\xi, D\varphi\right) - f(x, u)\, \varphi\, dx=0}&\ \forall \varphi\in V_u.
\end{cases}
\end{equation}
\end{proposition}
\begin{proof}
Let us compute $\partial J_H$ (see \eqref{defJH}), following \cite{PP}. The operator
\[
W^{1,p}_0(\Omega)\ni v\mapsto A(v)=Dv \in \left(L^p(\Omega)\right)^N
\]
is linear and continuous with transpose
\[
\big(L^{p'}(\Omega)\big)^N\ni g\mapsto A^*(g)={\rm div}\, g\in W^{-1, p'}(\Omega)
\]
where, as usual, ${\rm div}\, g\in W^{-1, p'}(\Omega)$ acts as
\begin{equation}
\label{qwe}
\langle {\rm div}\, g, \varphi\rangle=\int_\Omega \left( g, D\varphi\right)\, dx\qquad \varphi\in W^{1, p}_0(\Omega).
\end{equation}
The convex functional
\[
\left(L^p(\Omega)\right)^N\ni g\mapsto I_H(g)=\int_\Omega H(g)\, dx
\]
is locally Lipschitz thanks to $d<\infty$ in \eqref{hypH} and has convex subdifferential (see \cite[Theorem 21]{R})
\[
\partial I_H(g)=\left\{\xi\in L^{p'}(\Omega; \mathbb{R}^N): \text{$\xi(x)\in \partial H(g(x))$ for a.\,e.\,$x\in \Omega$}\right\}.
\]
Clearly $J_H=I_H\circ A$ and by \cite[Theorem 19]{R} it holds
\[
\partial \left(I_H\circ A\right)(u)=A^*\left(\partial I_H (A(u))\right).
\]
It follows that at any $v\in W^{1,p}_0(\Omega)$
\[
\partial J_H(v)=\left\{{\rm div}\, \xi : \xi\in L^{p'}(\Omega; \mathbb{R}^N) \text{ and $\xi(x)\in \partial H(Dv(x))$ a.\,e.\,in $\Omega$}\right\},
\]
where an element of the set on the right acts as in \eqref{qwe}. To conclude, let $T\in W^{-1, p'}(\Omega)$ be the unique extension of
\[
v\mapsto \int_\Omega f(\cdot, u)\, v\, dx, \qquad v\in V_u
\]
granted by Remark \ref{remc}.
Then $u$ also minimises $v\mapsto J_H(v) - \langle T, v\rangle$ on $W^{1,p}_0(\Omega)$, by the continuity of $J_H-T$ and the density of $V_u$ in $W^{1,p}_0(\Omega)$. Therefore
\[
0\in \partial (J_H-T)(u)=\partial J_H(u)-T
\]
which implies \eqref{EL} when bracketed with $v\in V_u$.
\end{proof}
The following easy consequence will be useful later.
\begin{corollary}
Suppose that the assumptions of the previous Proposition hold true and additionally that $H$ is positively $p$-homogeneous.
Then any energy critical point $u$ for $J$ fulfils
\begin{equation}
\label{id2}
(\xi, Du)=p\, H(Du) \qquad \text{a.\,e.\,in $\Omega$}
\end{equation}
for any $\xi\in L^{p'}(\Omega; \mathbb{R}^N)$ obeying \eqref{EL}. In particular,
\begin{equation}
\label{id}
p\, \int_\Omega H(Du)\, dx=\int_\Omega f(x, u)\, u\, dx.
\end{equation}
\end{corollary}
\begin{proof}
Since $u$ minimises the convex functional $J_u$ defined in \eqref{defJu}, it does so on the half-line through $u$, namely on $\{t\, u: t>0\}\subseteq V_u$. The function
\[
h(t):=J_u(t\, u)=t^p\, \int_\Omega H(Du)\, dx - t\, \int_\Omega f(x, u)\, u\, dx
\]
is smooth and convex on $]0, \infty[$, therefore the previous statement forces $h'(1)=0$, which is \eqref{id}. Let now $\xi \in L^{p'}(\Omega; \mathbb{R}^N)$ fulfil \eqref{EL}, so that testing with $\varphi=u\in V_u$, we get
\[
\int_\Omega (\xi, Du)\, dx=\int_\Omega f(x, u)\, u\, dx.
\]
By using \eqref{id}, we deduce
\[
\int_\Omega p\, H(Du)-(\xi, Du)\, dx=0,
\]
but by \eqref{phom1} the integrand is non-negative, therefore it must vanish a.\,e.\,in $\Omega$, proving \eqref{id2}.
\end{proof}
\section{Regularity}\label{Sregularity}
The following lemma enucleates the De Giorgi method for proving $L^\infty$ bounds.
\begin{lemma}\label{propositionGiusti}
Let $\Omega\subseteq \mathbb{R}^N$, $q\ge 1$. If $u\in L^q(\Omega)$ satisfies
\begin{equation}
\label{it}
\int_\Omega (u-k)_+^q\, dx\le \frac{b}{(k-h)^{\beta}}\left(\frac{k}{k-h}\right)^\delta\left(\int_\Omega (u-h)_+^q\, dx\right)^{1+\gamma}
\end{equation}
for all $h_0\le h<k$ and positive parameters $b, \gamma, \beta, \delta$, then
\begin{equation}
\label{supest}
\sup_\Omega u\le \max\left\{2\, h_0, \, C\, \left( b^{1/\gamma} \int_\Omega u_+^q\, dx\right)^{\gamma/\beta}\right\}
\end{equation}
for $C=C(b, \gamma, \beta, \delta)>0$ smoothly depending on the data.
\end{lemma}
\begin{proof}
Let $M\ge 2\, h_0$ to be determined and set $k_n=M\, (1-2^{-1-n})$ for $n\ge 0$. Notice that $k_n\uparrow M$, $k_0=M/2\ge h_0$ and
\[
\frac{b}{(k_{n+1}-k_n)^{ \beta}}\left(\frac{k_{n+1}}{k_{n+1}-k_n}\right)^\delta\le \frac{b}{M^{ \beta}} \, 2^{(n+2)\, ( \beta+\delta)}.
\]
We can then apply \eqref{it} for $k=k_{n+1}$ and $h=k_{n}$ to get
\[
\int_\Omega (u-k_{n+1})_+^q\, dx\le \frac{A\, b\, A^n}{M^{\beta}}\left(\int_\Omega (u-k_n)_+^q\, dx\right)^{1+\gamma}.
\]
for $A=4^{\beta+\delta}$.
By \cite[Proposition 7.1]{Giusti}, if
\begin{equation}
\label{piccol}
\int_\Omega (u-k_{0})_+^q\, dx\le \frac{M^{\beta/\gamma}}{(A\, b)^{1/\gamma}\, A^{1/\gamma^2}}
\end{equation}
then
\[
\lim_n \int_\Omega (u-k_n)_+^q\, dx=0,
\]
which forces $u\le M$ in $\Omega$. Condition \eqref{piccol} is satisfied if $M\ge 2\, h_0$ and
\[
\int_\Omega u_+^q\, dx\le \frac{M^{\beta/\gamma}}{(A\, b)^{1/\gamma}\, A^{1/\gamma^2}},
\]
therefore $u\le M$ for all $M$ such that
\[
M\ge 2\, h_0\qquad \text{and} \qquad M\ge A^{\frac{\gamma+1}{\gamma\, \beta}}\, \left( b^{1/\gamma} \int_\Omega u_+^q\, dx\right)^{\gamma/\beta},
\]
proving \eqref{supest}.
\end{proof}
\begin{theorem}[$L^\infty$ bounds]\label{thbound}
Let $\Omega\subseteq \mathbb{R}^N$ be open with finite measure and suppose \eqref{hypH} holds with $c>0$, as well as \eqref{hypf0} and, if $p\le N$, \eqref{hypf}. Then, any energy critical point $u\in W^{1,p}_0(\Omega)$ of $J$
is bounded. Moreover, if \eqref{hypf} holds true (regardless of $p$) for some $q<p^*$, then
\begin{equation}
\label{supbound}
\|u\|_{L^\infty(\Omega)}\le
\max\left\{\theta, C\,\left((\mu/c)^{\frac{N}{p}}\int_\Omega |u|^q\, dx\right)^{\frac{p}{p\, q+ (p-q)\, N}}\right\}
\end{equation}
for some $C=C(N, p, q)$ independent of $d$, $\theta$, $\mu$ and $\Omega$. Moreover, for any fixed $N$, $C(N, \cdot, \cdot):\, [1, \infty[\times [1, p^*[\, \to \mathbb{R}_+$ is positive and continuous.
\end{theorem}
\begin{proof}
By considering $H/c$ and $f/c$, $u$ is still an energy critical point of the corresponding functional, so we can assume that $c=1$.
Without loss of generality, it suffices to prove that $u$ is bounded above (otherwise we look at $v(x)=-u(x)$, which is a critical point of a functional with $\check{H}(z)=H(-z)$ and $\check{f}(x, t)=-f(x, -t)$ obeying the same structural conditions as $J$). The boundedness of $u$ being trivial from Morrey embedding when $p>N$, we may suppose that $p\ge 1$ is arbitrary and that \eqref{hypf} holds true. Moreover, we can assume $q<\infty$ in \eqref{hypf}, since otherwise the condition $q\le p^*$ forces $p>N$ and again boundedness follows from Morrey embedding. Finally, by the definition of energy critical point and the discussion following Definition \ref{defecp}, we can also assume that $H(Du)$, $f(\cdot, u)\, u$ and $f(\cdot, u)$ all belong to $L^1(\Omega)$.
Given $k\ge 0$, we compare the minimality of $u$ with $\min\{k, u\}\in V_u$, namely
\[
J_u(u)\le J_u(\min\{k, u\})
\]
to get, by the locality of the weak gradient and \eqref{hypf},
\[
\int_{{\mathcal A}_k} H(Du)\, dx\le \int_{A_k} f(x, u)\, (u-k)\, dx\le \mu\, \int_{{\mathcal A}_k}(u+\theta)^{q-1}\, (u-k)\, dx
\]
where we set
\[
{\mathcal A}_k=\{x\in \Omega: u(x)\ge k\}.
\]
The inequality
\begin{equation}
\label{t2}
\mu\, (t+\theta)^{q-1}\, (t-k)\le \mu\,4^{q-1} (k^{q}+ (t-k)^q),
\end{equation}
holds true for any $t\ge k\ge \theta/2$\footnote{Indeed, if $\theta/2\le k\le t\le 2\, k$ then $(t+\theta)^{q-1}\, (t-k)\le k\, (2\,k +\theta)^{q-1}\le 4^{q-1}\, k^q$, while if $t\ge 2\, k$ then $t\le 2\, (t-k)$ and thus $(t+\theta)^{q-1}\, (t-k)\le 2^{q-1}\, (t-k)^q$.}.
Using \eqref{hypH}, the growth condition on $f$ and \eqref{t2}, we get
\begin{equation}
\label{t4}
\int_{{\mathcal A}_k}|Du|^p\, dx\le \mu\, 4^{q-1}\, \int_{{\mathcal A}_k} k^q+(u-k)^q\, dx
\end{equation}
which holds for any $k\ge \theta/2$.
We now aim at applying Lemma \ref{propositionGiusti}. First notice that, as long as $q<p^*$ (even in the case $p\ge N$), it always holds
\[
q\, p+(p-q)\, N>0
\]
which is equivalent, if $N\ge 2$, to
\[
\frac{N\, (q\, p+p-q)}{(N-1)\, p}>q.
\]
For $N\ge 2$, we can then chain H\"older, Gagliardo-Nirenberg inequality \eqref{GN} and \eqref{t4} to get
\begin{equation}
\label{t3}
\begin{split}
\int_\Omega& (u-k)_+^q\, dx\le \left(\int_\Omega (u-k)_+^{1^*\frac{q\, p+p-q}{p}}\, dx\right)^{\frac{q\, p}{1^*\, (q\, p+p-q)}}\, |{\mathcal A}_k|^{1-\frac{q\, p}{1^*\, (q\, p+p-q)}}\\
&\le C\, \left(\int_\Omega |D(u-k)_+|^p\, dx\right)^{\frac{q}{ q\, p+p-q}}\left(\int_\Omega (u-k)_+^q\, dx\right)^{\frac{q\, p}{p'(q\, p+p-q)}}\, |{\mathcal A}_k|^{\frac{q\, p +(p-q)\, N}{N\, (q\, p+p-q)}}\\
&\le C\, \left(\mu \, \int_{{\mathcal A}_k} k^q+(u-k)^q\, dx\right)^{\frac{q}{ q\, p+p-q}}\left(\int_\Omega (u-k)_+^q\, dx\right)^{\frac{q\, (p-1)}{q\, p+p-q}}\, |{\mathcal A}_k|^{\frac{q\, p+(p-q)\, N}{N\, (q\, p+p-q)}}
\end{split}
\end{equation}
with
\[
C=\left( 4^{q-1}\, G_{p, q, N}^p\right)^{\frac{q}{q\, p+p-q}}
\]
smoothly depending on the data.
Let now $\theta/2\le h<k$. By Chebishev inequality
\begin{equation}
\label{t5}
|{\mathcal A}_k|\le \frac{1}{(k-h)^q}\int_\Omega (u-h)_+^q\, dx,
\end{equation}
so that, since $(u-k)_+\le (u-h)_+$,
\[
\int_{{\mathcal A}_k} k^q+(u-k)^q\, dx=k^q\, |{\mathcal A}_k|+\int_\Omega (u-k)_+^q\, dx\le \left(\frac{k^q}{(k-h)^q}+1\right)\int_\Omega (u-h)_+^q\, dx,
\]
hence, since $k/(k-h)>1$,
\[
\int_{{\mathcal A}_k} k^q+(u-k)^q\, dx\le 2\left(\frac{k}{k-h}\right)^q\int_\Omega (u-h)_+^q\, dx.
\]
Inserting the latter into \eqref{t3} and using again \eqref{t5} on the last factor gives
\[
\int_\Omega (u-k)_+^q\, dx\le \frac{C\, \mu^{\frac{q}{ q\, p+p-q}}}{(k-h)^{q\, \frac{q\, p+(p-q)\, N}{N\, (q\, p+p-q)}}}\, \left(\frac{k}{k-h}\right)^{\frac{q^2}{ q\, p+p-q}}\, \left(\int_{\Omega}(u-k)_+^q\, dx\right)^{1+\frac{q\, p}{N\, ( q\, p+p-q)}}
\]
and Lemma \ref{propositionGiusti} with the choices
\[
b=C\, \mu^{\frac{q}{ q\, p+p-q}},\qquad \beta=q\, \frac{q\, p+(p-q)\, N}{N\, (q\, p+p-q)}, \qquad \gamma=\frac{q\, p}{N\, (q\, p+p-q)}, \qquad \delta=\frac{q^2}{q\, p+p-q}
\]
(which are all positive if $q<p^*$), yields the conclusion.
The case $N=1$ can be dealt similarly: instead of the Gagliardo-Nirenberg inequality \eqref{GN} we use
\[
\|v\|_{\infty}^{\frac{q\, p+p-q}{p}}\le \frac{q\, p+p-q}{p}\, \|Dv\|_p\, \|v\|_q^{q/p'},
\]
which can be proved by elementary means, and proceed in the same way. The details are omitted.
\end{proof}
\begin{remark}
\label{remsup}
It is worth analysing the optimality of the form \eqref{supbound} of the $L^\infty$ bound in the simple case of the Dirichlet energy, the main point being that one cannot expect the simpler bound
\begin{equation}
\label{ref2}
\|u\|_{L^\infty(\Omega)}\le C\,\left((\mu/c)^{\frac{N}{p}}\int_\Omega |u|^q\, dx\right)^{\frac{p}{p\, q+ (p-q)\, N}}
\end{equation}
to be true when $\theta>0$ with a constant $C=C(N, p, q)$ independent of $\theta$.
Let $\mu \in \ ]0, \lambda_1[$, where $\lambda_1$ is the first eigenvalue of the negative Dirichlet Laplacian in the ball $B_1$. For any $\theta>0$, the problem
\begin{equation}
\label{ref}
\begin{cases}
-\Delta u= \mu\, (u+\theta) &\text{in $B_1$},\\
u=0&\text{in $\partial B_1$}
\end{cases}
\end{equation}
has, by standard methods, a unique positive solution, which minimises a functional obeying the assumptions of the previous Theorem with $c=1$, $q=p=2$ and $\mu, \theta$ as in \eqref{hypf}. By comparison, $u$ is greater or equal than the solution of
\[
\begin{cases}
-\Delta v= \mu\, \theta &\text{in $B_1$},\\
v=0&\text{in $\partial B_1$},
\end{cases}
\]
which is $v(x)=\mu\, \theta \, (1-|x|^2)/(2\, N)$, therefore
\begin{equation}
\label{ref1}
\|u\|_{L^\infty(B_1)}\ge \frac{\mu\, \theta}{2\, N}.
\end{equation}
On the other hand, by testing the weak form of \eqref{ref} with $u$ and by the definition of $\lambda_1$, we get
\[
\mu\, \int_{B_1} u^2+\theta\, u\, dx=\int_{B_1} |Du|^2\, dx\ge \lambda_1\, \int_{B_1} u^2\, dx
\]
so that, rearranging and using H\"older inequality, we find
\[
(\lambda_1-\mu)\int_{B_1} u^2\, dx\le \mu\, \theta \, \int_{B_1} u\, dx\le \mu\, \theta\, \left(\int_{B_1} u^2\, dx\right)^{1/2} \omega_N^{1/2}.
\]
The latter gives the bound
\[
\left(\mu^{N/2}\, \int_{B_1} u^2\, dx\right)^{1/2}\le \omega_N^{1/2}\, \frac{\mu^{1+\frac{N}{4}}\, \theta}{\lambda_1-\mu},
\]
so that, if $\mu>0$ fulfils
\[
C(N, 2, 2)\, \omega_N^{1/2}\, \frac{\mu^{\frac{N}{4}}}{\lambda_1-\mu}<\frac{1}{2\, N},
\]
by \eqref{ref1} we have
\[
\|u\|_{L^\infty(B_1)}> C(N, 2, 2)\left(\mu^{N/2}\, \int_{B_1} u^2\, dx\right)^{1/2}
\]
and the bound \eqref{ref2} cannot hold.
\end{remark}
Once boundedness is proved, assumption \eqref{hypf0} ensures that $f(\cdot, u)\in L^\infty(\Omega)$ and therefore $V_u=W^{1,p}_0(\Omega)$. Even if \eqref{EL} is just a differential inclusion, basic regularity theory is nevertheless at hand, thanks to the theory of quasi-minima developed by Giaquinta and Giusti in \cite{GG}.
\begin{corollary}[Regularity]\label{reg}
Suppose \eqref{hypH} holds true with $0<c\le d<\infty$ and $p>1$, as well as \eqref{hypf0}. Let $u$ be a bounded energy critical point for $J$ on $W^{1,p}_0(\Omega)$ and $M_0>0$ be such that $\|u\|_{L^\infty(\Omega)}\le M_0$. If $\Omega$ is bounded and obeys
\begin{equation}
\label{L}
L=\inf_{x\in \partial\Omega, r>0} \frac{|B_r(x)\setminus \Omega|}{|B_r(x)|}>0,
\end{equation}
then
\[
\|u\|_{C^\beta(\overline\Omega)}\le C
\]
for $C>0$ and $\beta\in (0, 1)$ depending smoothly on $p$, $c$, $d$, $M_0$, $f$, $L$ and $|\Omega|$, continuously for fixed $N$.
If $\Omega$ has finite measure,
\[
\|u\|_{C^\beta(\Omega')}\le C
\]
for any $\Omega'\Subset \Omega$, with $C$ additionally depending on $\Omega'$ and $\Omega$.
\end{corollary}
\begin{proof}
To make the statement precise, we rewrite \eqref{hypf0} as
\[
\sup_{|s|\le t}\|f(\cdot, s)\|_{L^\infty(\Omega)}\le h(t)\qquad \forall t>0
\]
for a smooth, increasing $h$, so that the constants and exponents in the statement will depend smoothly just on $M_0$, through the function $h$ which in turn depends on $f$.
As $u$ is bounded, it minimises the convex functional $J_u$ defined in \eqref{defJu} over the whole $W^{1,p}_0(\Omega)$,
we can adapt the proof of \cite[Theorem 7.1]{Giusti} to our setting, focusing on the Dirichlet boundary condition.
We focus on the first statement, since the second one will follow from it by choosing smooth $\Omega'\Subset\Omega$ obeying \eqref{L}.
Consider the case $p\le N$.
For any ball $B_R$, not necessarily contained in $\Omega$ and $0<r<R$, we fix $\eta\in C^{\infty}_c(B_R)$ such that
\begin{equation}
\label{defeta}
{\rm supp}\, \eta\subseteq B_R, \qquad 0\le \eta\le 1, \qquad \left.\eta\right|_{B_r}\equiv 1, \qquad |D\eta|\le \frac{C(N)}{R-r},
\end{equation}
and compare the minimality of $u$ with the function $v=u-\eta\, (u-k)_+$ for $k\ge 0$. Notice that $v\in W^{1,p}_0(\Omega)$ regardless of the relative position of $B_R$ and $\Omega$, and
\[
{\rm supp}\, (v-u)\subseteq {\mathcal A}(k, R):=\{x\in \overline{\Omega}\cap B_R: u(x)\ge k\},
\]
\begin{equation}
\label{der}
Dv=(1-\eta)\, Du - (u-k)_+\, D\eta\qquad \text{in\ } {\mathcal A}(k, R).
\end{equation}
Therefore $J_u(u)\le J_u(v)$ entails
\[
\int_{{\mathcal A}(k, R)} H(Du)\, dx\le \int_{{\mathcal A}(k, R)} H(Dv)+f(x, u)\, \eta\, (u-k)_+\, dx.
\]
By using the growth conditions on $H$, \eqref{hypf0} and the boundedness of $u$, we infer
\[
c\, \int_{{\mathcal A}(k, R)}|Du|^p\, dx\le d\, \int_{{\mathcal A}(k, R)}|Dv|^p+ M_0\, \sup_{|s|\le M_0} \|f(\cdot, s)\|_{L^\infty(\Omega)}\, dx\\
\]
so that by \eqref{der}
\[
\begin{split}
\int_{{\mathcal A}(k, r)}|Du|^p\, dx&\le C\, \int_{{\mathcal A}(k, R)}\left[(1-\eta)^p\,|Du|^p +|D\eta|^p\, (u-k)^p\right]\, dx +C\, |{\mathcal A}(k, R)|\\
&\le C\, \int_{{\mathcal A}(k, R)\setminus {\mathcal A}(k, r)}|Du|^p\, dx +C\, \int_{{\mathcal A}(k, R)}\frac{(u-k)^p}{(R-r)^p}\, dx +C\, |{\mathcal A}(k, R)|
\end{split}
\]
for $C=C(N, p, c, d, M_0)>0$ and continuously depending of the data. By hole-filling the first term on the right, i.\,e.\,adding to the inequality $C$ times the left hand side, we get
\[
(C+1) \int_{{\mathcal A}(k, r)}|Du|^p\, dx\le C\int_{{\mathcal A}(k, R)}|Du|^p\, dx+\frac{C}{(R-r)^p}\, \int_{{\mathcal A}(k, R)}(u-k)^p\, dx +C\, |{\mathcal A}(k, R)|,
\]
or
\[
\int_{{\mathcal A}(k, r)}|Du|^p\, dx\le (1-\delta)\left[\int_{{\mathcal A}(k, R)}|Du|^p\, dx +\frac{1}{(R-r)^p}\int_{{\mathcal A}(k, R)} (u-k)^p\, dx +\, |{\mathcal A}(k, R)|\right]
\]
for $\delta=1/(C+1)$, so that \cite[Proposition 6.1]{Giusti} provides
\[
\int_{{\mathcal A}(k, r)}|Du|^p\, dx\le \frac{C}{(R-r)^p}\int_{{\mathcal A}(k, R)} (u-k)^p\, dx +C\, |{\mathcal A}(k, R)|.
\]
A similar conclusion holds true for $-u$, which is a critical point of a functional with the same structural conditions. Therefore $u$ belongs to the De Giorgi class as considered in \cite[p. 81 and p. 90]{LU}, namely ${\mathcal B}_p(\Omega\cup \partial\Omega, M_0, C, 1, 0)$, yielding the stated H\"older regularity up to the boundary. Luckily, the H\"older exponent found in \cite[Theorem 6.1 and Proposition 7.1]{LU} is very explicit in its dependance from the parameters, showing the stated dependance of $\beta$ and on the $C^\beta(\overline\Omega)$ norm of $u$.
If $p>N$, by comparing the minimality of $u$ with $\pm u_\pm$, we directly have
\[
\int_{\{\pm u\ge 0\}} H(Du)\, dx\le \int_{\{\pm u\ge 0\}} f(x, u)\, u_{\pm}\, dx\le C
\]
with a constant $C$ smoothly depending on $M_0$.
Extending $u$ as $0$ outside $\Omega$, using the growth condition \eqref{hypH} and Morrey embedding yields $u\in C^{1-N/p}(\mathbb{R}^N)$ with a uniform bound on the norm.
\end{proof}
\begin{corollary}[Minimum principles]\label{smp}
Let \eqref{hypH} hold with $0<c\le d<\infty$ and $p>1$, as well as \eqref{hypf0}, and let $u$ be a bounded energy critical point for $J$ on $W^{1,p}_0(\Omega)$, where $\Omega$ has finite measure.
\begin{enumerate}
\item
If $f(\cdot , t)\ge 0$ for $t\le 0$,
then $u\ge 0$ in $\Omega$.
\item
If $u\ge 0$ and there exist $A\ge 0$, $\delta>0$ such that for all $t\in\, ]0, \delta[$
\begin{equation}
\label{smph}
\inf_\Omega f(\cdot, t)\ge -A\, t^{p-1},
\end{equation}
then, in each connected component of $\Omega$, $u$ is either strictly positive or vanishes identically.
\end{enumerate}
\end{corollary}
\begin{proof}
We test the minimality with respect to $J_u$ of $u$ against $u_+$, to get
\[
\int_{\{u<0\}} H(Du)\, dx\le \int_\Omega f(x, u)\, (u-u_+)\, dx=-\int_{\{u\le 0\}} f(x, u)\, u_-\, dx
\]
and since $f(\cdot, u)\ge 0$ on $u\le 0$, the term on the left must vanish, implying that $u\ge 0$ a.\,e.. To prove the strict positivity of $u$ we further test the minimality of $u$ with $v=u+\eta\, (k-u)_+$ for $k\ge 0$ and $\eta$ as in \eqref{defeta} with $B_R\subseteq \Omega$. Since
\[
{\rm supp}\, (u-v)\subseteq \{x\in B_R: u(x)\le k\}=:{\mathcal B}(k, R)
\]
the minimality relation $J_u(u)\le J_u(v)$ reads
\[
\int_{{\mathcal B}(k, R)} H(Du)\, dx\le \int_{{\mathcal B}(k, R)} H(Dv)+ f(x, u)\, (u-v)\, dx.
\]
Since $u$ is bounded, we can assume \eqref{smph} holds true for all $t\ge 0$, with a bigger constant $A$ also depending on $\delta$ and $\|u\|_\infty$.
Since $u-v=-\eta\, (k-u)_+\le 0$, \eqref{smph} gives
\[
\begin{split}
\int_{{\mathcal B}(k, R)} H(Du)\, dx&\le \int_{{\mathcal B}(k, R)} H(Dv)+ A\, u^{p-1}\, \eta\, (k-u)_+\, dx.\\
&\le \int_{{\mathcal B}(k, R)} H(Dv)\, dx +A\, k^p\, |{\mathcal B}(k, R)|.
\end{split}
\]
Proceeding with the same calculations of the previous proof one arrives via hole-filling at
\[
\int_{{\mathcal B}(k, r)} |Du|^p\, dx\le \frac{C}{(R-r)^p}\int_{{\mathcal B}(k, R)}(k-u)_+^p\, dx +A\, k^p\, |{\mathcal B}(k, R)|
\]
for all $k\ge 0$. Therefore $u\ge 0$ belongs to the homogeneous De Giorgi class $DG^-_p(\Omega)$ as in \cite{DBT}. A weak Harnack inequality holds true in this setting (see \cite[Theorem 2]{DBT}), from which the strong minimum principle readily follows.
\end{proof}
\section{Dirichlet eigenvalues}
\label{DE}
For $H:\mathbb{R}^N\to [0, \infty[$ convex and obeying \eqref{hypH} we can define its {\em first Dirichlet eigenvalue} as
\begin{equation}
\label{lambda1}
\lambda_{1, H}(\Omega)=\inf\left\{\int_\Omega H(Dv)\, dx: v\in W^{1,p}_0(\Omega), \int_\Omega |v|^p\, dx=1\right\}.
\end{equation}
A {\em first normalized Dirichlet eigenfunction} for $\lambda_{1, H}(\Omega)$ is a function solving the minimum problem for $\lambda_{1, H}(\Omega)$ with its constraints. When $\Omega$ has finite measure, the embedding $W^{1,p}_0(\Omega)\hookrightarrow L^p(\Omega)$ is compact, therefore, if $c>0$ in \eqref{hypH}, problem \eqref{lambda1} is always solvable and its value is positive.
The value \eqref{lambda1} makes sense in general, but we focus here on the $p$-positively homogeneous setting. In this case, $\lambda_{1, H}(\Omega)$ is related to the best constant for the inequality
\begin{equation}
\label{bc}
\int_\Omega |v|^p\, dx\le C_{\rm opt}(H, \Omega)\, \int_\Omega H(Dv)\, dx\qquad \forall v\in W^{1, p}_0(\Omega)
\end{equation}
which, by positive $p$-homogeneity, satisfies $C_{\rm opt}(H, \Omega)=1/\lambda_{1, H}(\Omega)$. Moreover, if $u$ solves problem \eqref{lambda1}, any {\em positive} multiple of $u$ attains the best constant in \eqref{bc}, and will be called a {\em first Dirichlet eigenfunction}.
Notice that we are not assuming any regularity on $H$ except the natural local Lipschitz one and, more substantially, we do not assume that $H$ is even.
In the possibly non-even setting, it is convenient to define
\begin{equation}
\label{lambdapm}
\begin{split}
\lambda_{1, H}^+(\Omega)&=\inf\left\{\int_\Omega H(Dv)\, dx: v\in W^{1,p}_0(\Omega), \int_\Omega |v|^p\, dx=1, \ v\ge 0\right\}\\
\lambda_{1, H}^-(\Omega)&=\inf\left\{\int_\Omega H(Dv)\, dx: v\in W^{1,p}_0(\Omega), \int_\Omega |v|^p\, dx=1, \ v\le 0\right\},
\end{split}
\end{equation}
which may in principle be different. We say that $u\ge 0$ is a {\em first normalized Dirichlet positive eigenfunction} for $\lambda^{+}_{1, H}(\Omega)$ if $u$ realises the minimum for $\lambda_{1, H}^+(\Omega)$ (with its constraints) and that $u\le 0$ is a {\em first normailzed negative Dirichlet eigenfunction} if it does so for $\lambda_{1, H}^-(\Omega)$. Similar considerations as in the unconstrained eigenvalue $\lambda_{1, H}(\Omega)$ are valid for sign-constrained ones and we will drop the term normalized accordingly.
Notice that, without any evennes assumption on $H$, it is no longer true that if $u$ is a first positive eigenfunction then $-u$ is a first negative eigenfunction. Moreover, it holds
\begin{equation}
\label{checkH}
\lambda_{1, H}^-(\Omega)=\lambda_{1, \check{H}}^+(\Omega), \qquad \check{H}(z)=H(-z),
\end{equation}
since $v$ obeys the constraints involving $\lambda_{1, H}^-(\Omega)$ if and only if $-v$ does so for $\lambda_{1, H}^+(\Omega)$.
\begin{proposition}
Suppose $H:\mathbb{R}^N\to [0, \infty[$ is convex and positively $p$-homogeneoues, for $p\ge 1$. Then, for any open $\Omega\subseteq \mathbb{R}^N$
\begin{equation}
\label{minlambda}
\lambda_{1, H}(\Omega)=\min\left\{\lambda_{1, H}^+(\Omega), \lambda_{1, H}^-(\Omega)\right\}.
\end{equation}
\end{proposition}
\begin{proof}
Clearly
\[
\lambda_{1, H}(\Omega)\le \min\left\{\lambda_{1, H}^+(\Omega), \lambda_{1, H}^-(\Omega)\right\},
\]
so we prove the opposite inequality. To this end, notice that all the eigenvalues are finite so, assuming the contrary, let $\delta>0$ be such that
\begin{equation}
\label{t10}
\min\left\{\lambda_{1, H}^+(\Omega), \lambda_{1, H}^-(\Omega)\right\}-\lambda_{1, H}(\Omega)=\delta.
\end{equation}
Fix $0<\varepsilon<\delta$ and choose $v$ such that $\|v\|_p=1$ and
\[
\int_\Omega H(Dv)\, dx< \lambda_{1, H}(\Omega)+\varepsilon.
\]
If $v_-\equiv 0$ a.\,e., then $v\ge 0$ and
\[
\min\left\{\lambda_{1, H}^+(\Omega), \lambda_{1, H}^-(\Omega)\right\}\le \lambda_{1, H}^+(\Omega)\le \int_\Omega H(Dv)\, dx<\lambda_{1, H}(\Omega)+\varepsilon,
\]
contradicting $\varepsilon<\delta$.
A similar conclusion holds if $v_+=0$, so both $v_+$ and $v_-$ do not vanish identically. Thus $v_+/\|v_+\|_p$ and $-v_-/\|v_-\|_p$ are admissible competitors for the corresponding problems \eqref{lambdapm}, giving, by $p$-homgeneity
\[
\lambda_{1, H}^\pm(\Omega)\le \frac{1}{\|v_\pm\|_p^p}\, \int_\Omega H( D(\pm v_\pm))\, dx,
\]
respectively, thus by \eqref{t10}
\[
\left(\delta+\lambda_{1, H}(\Omega)\right)\, \int_\Omega v_\pm^p\, dx\le \int_\Omega H(D(\pm v_\pm))\, dx.
\]
The corresponding inequalities can be summed up to get, through the locality of the weak gradient,
\[
\begin{split}
\delta+\lambda_{1, H}(\Omega)&=\left(\delta+\lambda_{1, H}(\Omega)\right)\, \left(\int_\Omega v_+^p\, dx+\int_\Omega v_-^p\, dx\right)\\
&\le \int_\Omega H(Dv_+)\, dx+\int_\Omega H(D(-v_-))\, dx\\
&\le \int_\Omega H(Dv)\, dx<\lambda_{1, H}(\Omega)+\varepsilon
\end{split}
\]
against $\varepsilon<\delta$.
\end{proof}
In general, we do not expect that the two values in \eqref{lambdapm} coincide, although we have not been able to produce an explicit example. If equality is always true, however, it cannot follow from convexity, $p$-homogeneity and modularity of the energy alone, as the following toy example shows.
\begin{ex}
Let $H:\mathbb{R}^2\to [0, \infty[$ be
\[
H(x, y)= a\, x_+^2 + x_-^2+a\, y_+^2+ y_-^2
\]
for some $a>1$.
We look at $\mathbb{R}^3$ as the set of functions $u:\Omega\to \mathbb{R}$ with $\Omega=\{0, 1, 2\}$, whose "gradient" is $Du:=(u(0)-u(1), u(1)-u(2))\in \mathbb{R}^2$. The corresponding positive and negative Dirichlet eigenvalue problems are, respectively,
\[
\lambda_{1, H}^\pm(\Omega)=\inf\left\{\sum H(Du) :u(0)=u(2)=0, u(1)=\pm1\right\}.
\]
The function $H$ is convex, positively $2$-homogeneous and modular, meaning
\[
H(Du)+H(Dv)=H(D\max\{u, v\})+H(D\min\{u, v\}),
\]
due to \cite[Remark 2.2]{GM}, so the truncation arguments employed in the previous proof still stand. However
\[
\lambda_{1, H}^+(\Omega)=1<a= \lambda_{1, H}^-(\Omega).
\]
A similar construction can clearly be generalised to finite sets $\Omega$ containing an arbitrary number of points.
\end{ex}
A simple case where the sign-constrained eigenvalues are equal is when $\Omega$ enjoys a symmetry opposite with $H$.
In order to make a precise statement, let
\[
{\rm Sym}\, (\Omega)=\left\{T\in O(N): T(\Omega)=\Omega+x_0 \text{ for some $x_0\in \mathbb{R}^N$} \right\}.
\]
\begin{proposition}
Under the assumptions of the previous Proposition, suppose that
\begin{equation}
\label{sym}
{\rm Sym}\, (\Omega)\cap -{\rm Sym}\, (H)\ne \emptyset.
\end{equation}
Then
\begin{equation}
\label{propositionlambda}
\lambda_{1, H}^+(\Omega)=\lambda_{1, H}^-(\Omega)=\lambda_{1, H}(\Omega).
\end{equation}
\end{proposition}
\begin{proof}
Fix $T\in {\rm Sym}\, (\Omega)\cap -{\rm Sym}\, (H)$, so that it holds
\[
H(-T(z))=H(z)\qquad \forall z\in \mathbb{R}^N.
\]
Then, for any $v\ge 0$ in $W^{1,p}_0(\Omega)$ with $\|v\|_p=1$, the function $w=-v\circ T$ is non positive, belongs to $W^{1,p}_0(\Omega-x_0)$ for some $x_0\in \mathbb{R}^N$ and fulfils $\|w\|_p=1$. By translation invariance
\[
\lambda_{1, H}(\Omega)=\lambda_{1, H}(\Omega-x_0)
\]
and
\[
\int_{\Omega-x_0} H(Dw)\, dx=\int_{\Omega-x_0} H(-T Dv\circ T)\, dx=\int_\Omega H(Dv)\, dx.
\]
Therefore $\lambda_{1, H}^+(\Omega)\ge \lambda_{1, H}^-(\Omega)$. Proceeding symmetrically from $v\le 0$ we get the opposite inequality, while applying \eqref{minlambda} concludes the proof of \eqref{propositionlambda}.
\end{proof}
\begin{remark}
By the translation invariance of the energy, the same conclusion holds true if each connected component of $\Omega$ separately fulfils assumption \eqref{sym}.
In particular, \eqref{propositionlambda} holds true if each connected component has at least one axis of symmetry. Using this fact it is readily shown that \eqref{propositionlambda} always holds true for arbitrary open $\Omega\subseteq \mathbb{R}$. Indeed, open connected subset of $\mathbb{R}$ are intervals, half-lines or the whole $\mathbb{R}$. But if $\Omega$ has an unbounded connected component, a scaling argument shows that all eigenvalues vanish, while if $\Omega$ has only bounded connected components the previous Proposition applies, since open intervals are symmetric with respect to their midpoint.
\end{remark}
\begin{proposition}\label{regeigen}
Suppose that $H:\mathbb{R}^N\to [0, +\infty[$ is convex, positively $p$-homogeneous for some $p>1$, vanishing only at the origin and that $\Omega$ is connected with finite measure. If $u$ is a first Dirichlet eigenfunction (sign-constrained or not), then $u\in C^\alpha(\Omega)$ and never vanishes in $\Omega$.
\end{proposition}
\begin{proof}
We first observe that any eigenfunction is a critical point for a functional of the form \eqref{J} obeying \eqref{hypH}, \eqref{hypf0} and \eqref{hypf}. In particular, from Proposition \ref{Lcoer}, \eqref{hypH} holds for $H$ with $0<c\le d<\infty$.
In the unconstrained case, for the functional
\[
J(v)=\int_\Omega H(Dv)-\lambda_{1, H}(\Omega) |v|^p\, dx
\]
it holds $J\ge 0$ on $W^{1, p}_0(\Omega)$ by construction and $J(u)=0$, hence $u$ minimises $J$ and is therefore an energy critical point for it. In the sign-constrained case, by considering $\check{H}$ instead of $H$ and using \eqref{checkH}, it suffices to consider the case when $u$ is a positive eigenfunction.
The functional
\[
J_+(v)=\int_\Omega H(Dv)-\lambda_{1, H}^+(\Omega)\, v_+^p\, dx
\]
vanishes at $u$ and it holds $J_+(v)\ge J_+(v_+)$. But $J_+(v_+)\ge 0$ by the definition of $\lambda^+_{1, H}(\Omega)$, so again $u$ minimises $J_+$. Clearly in all cases \eqref{hypf} holds true with $q=p$ and Theorem \ref{thbound} implies boundedness, and thus continuity by Corollary \ref{reg}, of all eigenfunctions. Moreover, Corollary \ref{smp} applies for positive eigenfunction, since in this case $f(x, t)=t_+^{p-1}\ge 0$, yielding the first statement for positive eigenfunctions.
For an arbitrary eigenfunction $u$, we claim that either $u_+$ or $-u_-$ are eigenfunctions for $\lambda^{\pm}_{1, H}(\Omega)$ respectively, regardless of any assumption on $\Omega$. Indeed, it suffices to prove it for normalized eigenfunctions, so let $u$ be a first Dirichlet eigenfunction for $\lambda_{1, H}(\Omega)$ such that $\|u\|_p=1$. If $u$ is of constant sign, the claim is trivial. Otherwise, both $u_+$ and $u_-$ are non trivial and then $\pm u_\pm/\|u_{\pm}\|_p$ are admissible for the corresponding problems in \eqref{lambdapm}. By the positive $p$-homogeneity of $H$
\[
\lambda_{1, H}^\pm(\Omega)\le \frac{1}{\|u_\pm\|_p^p}\, \int_\Omega H(D(\pm u_\pm))\, dx
\]
which, summed up, imply
\[
\begin{split}
\lambda_{1, H}(\Omega)&=\min\left\{\lambda_{1, H}(\Omega)^+, \lambda_{1, H}^-(\Omega)\right\}\left(\|u_+\|_p^p+\|u_-\|_p^p\right)\\
&\le \int_\Omega H(Du_+)+ H(D(-u_-))\, dx=\lambda_{1, H}(\Omega),
\end{split}
\]
so that all inequalities above are actually equalities, proving the claim. It follows that if $u_+\ne 0$, then $u_+$ is a positive eigenfunction and therefore never vanishes by the previous part of the proof, so $u_-= 0$ and $u=u_+>0$ in $\Omega$. A similar conclusion holds true if $u_-\ne 0$, yielding $u=-u_-<0$ in $\Omega$ and concluding the proof.
\end{proof}
\section{Main result}
In order to prove a Brezis-Oswald comparison principle, we'll need the following generalisation to the non-smooth setting of the Picone inequality.
\begin{lemma}[Picone inequality]\label{thpicone}
Suppose $H$ is convex and positively $p$-homogeneous. Let $u, v\in W^{1,1}_{\rm loc}(\Omega)\cap L^\infty_{\rm loc}(\Omega)$ be such that $\inf_\Omega u>0$, $\inf_\Omega v\ge 0$ and let $\xi\in {\mathcal L}(\Omega; \mathbb{R}^N)$ be such that $\xi(x)\in \partial H(Du(x))$ a.\,e.. Then it holds
\begin{equation}
\label{picone}
\frac{1}{p}\left(\xi , D \frac{v^p}{u^{p-1}}\right)\le H(Dv)\qquad \text{a.\,e.\,in $\Omega$. }
\end{equation}
Suppose additionally that $H$ is strictly convex, $\Omega$ is connected and equality holds a.\,e.\,in \eqref{picone}. Then $v=k\, u$ for some $k>0$.
\end{lemma}
\begin{proof}
By considering $v+\varepsilon$ instead of $v$ and then passing to the limit for $\varepsilon\downarrow 0$, we can suppose that $\inf_\Omega v>0$.
First observe that by the assumptions $v^p/u^{p-1}\in W^{1,1}_{\rm loc}(\Omega)$ and Liebnitz rule for weak derivatives ensures
\begin{equation}
\label{crule}
D\frac{v^p}{u^{p-1}}=(1-p)\, \frac{v^p}{u^p}\, Du + p\, \frac{v^{p-1}}{u^{p-1}}\, Dv.
\end{equation}
Let $\xi\in \partial H(w)$ for some $w\in \mathbb{R}^N$. The following computation are henceforth made at a point $x\in \Omega$ such that $\xi=\xi(x)\in \partial H(Du(x))$. In the inequality
\[
H(z)\ge H(Du)+(\xi, z-Du)\qquad \forall z\in \mathbb{R}^N
\]
we choose $z=(u/v)\, Dv$ to get, by $p$-homogeneity,
\[
\left(\frac{u}{v}\right)^p\, H(Dv)=H\left(\frac{u}{v} \, Dv\right)\ge H(Du)+\frac{u}{v}\, (\xi, Dv)-(\xi, Du)
\]
and by \eqref{phom1} at $w=Du(x)$
\[
\left(\frac{u}{v}\right)^p\, H(Dv)\ge \left(\frac{1}{p}-1\right)(\xi, Du)+\frac{u}{v}\, (\xi, Dv).
\]
Multiplying both sides by $(v/u)^p$ and using \eqref{crule} we finally get
\[
H(Dv)\ge \frac{1}{p} \left(\xi, (1-p)\, \frac{v^p}{u^p}\, Du + p\, \frac{v^{p-1}}{u^{p-1}}\, Dv\right)=\frac{1}{p}\left(\xi, D \frac{v^p}{u^{p-1}}\right)
\]
and \eqref{picone} is proved.
To prove the second statement, suppose equality holds in \eqref{picone} a.\,e.\,in $\Omega$ for some choice of $\xi$ obeying $\xi\in \partial H(Du)$ a.\,e.. Then all the previous inequalities are equalities and in particular
\[
H\left(\frac{u}{v} \, Dv\right)= H(Du)+ \left(\xi, \frac{u}{v}\, Dv- Du\right)
\]
holds for a.\,e.\,$x\in\Omega$. Let, for any such $x$, $w=Du(x)$ and $z=(u(x)/v(x))\, Dv(x)$ and suppose that $w\ne z$. From $H(z)=H(w)+(\xi, z-w)$ and the convexity of $H$ we infer
\[
H\big(t\, z+(1-t)\, w\big)\le t\, H(z)+(1-t)\, H(w)=H(w)+\big(\xi, t\, (z-w)\big),
\]
while from $\xi\in \partial H(z)$ we deduce
\[
H(w)+\big(\xi, t\, (z-w)\big)=H(w)+\big(\xi, (t\, z+(1-t)\, w) -w\big)\le H\big(t\, z+(1-t)\, w\big).
\]
Therefore $H$ is linear on the segment from $z$ to $w$, contradicting the strict convexity of $H$. Therefore $z=w$, hence
\[
D\log u=D\log v
\]
a.\,e.\,in $\Omega$, which implies that $u$ and $v$ are proportional by the connectedness of $\Omega$.
\end{proof}
\begin{remark}
By inspecting the proof, it is readily seen that assumption $\inf_\Omega u>0$ can be weakened to
\[
\inf_{\Omega'} u>0, \qquad \forall \Omega'\Subset \Omega.
\]
\end{remark}
We can now prove our main result, Theorem \ref{Mth}, which we now restate in a form convenient for its proof.
\begin{theorem}\label{thfin}
Let $\Omega\subseteq \mathbb{R}^N$ be open, connected and of finite measure.
For $p>1$ suppose that
\begin{itemize}
\item[i)] $H:\mathbb{R}^N\to [0, \infty[$ is convex, positively $p$-homogeneous and vanishes only at the origin
\item[ii)] $f$ fulfils \eqref{hypf0}
\item[iii)]
The map $ t\mapsto f(x, t)/t^{p-1}$ is non-increasing on $\mathbb{R}_+$ for a.\,e.\,$x\in \Omega$.
\end{itemize}
Let $u\in W^{1, p}_0(\Omega)\setminus \{0\}$ be a non-negative critical point for $J$. The following assertions hold.
\begin{enumerate}
\item
$u$ minimises $J$ among non-negative functions in $W^{1,p}_{0}(\Omega)$.
\item
If $t\mapsto f(x, t)/t^{p-1}$ is strictly decreasing, $u$ is the unique positive energy critical point for $J$.
\item
If $H$ is strictly convex, any other non-negative energy critical point is a constant multiple of $u$.
\item
If $H$ is strictly convex and $f$ doesn't depend on $x$, either $u$ is the unique positive critical point for $J$ or $u$ is a first Dirichlet positive eigenfunction.
\end{enumerate}
\end{theorem}
\begin{proof}
Since $u\ge 0$ by assumption, we can assume $f(x, t)\equiv f(x, 0)$ for all $t\le 0$.
Let
\[
h(t)=\sup_{0\le s\le t}\|f(\cdot, s)\|_\infty.
\]
By the monotonicity assumption {\em iii)} and \eqref{hypf0}
\begin{equation}
\label{t22}
f(x, t)\le
\begin{cases}
h(1)\, t^{p-1}& \forall t\ge 1\\
h(1)&\forall t\le 1.
\end{cases}
\end{equation}
so that Theorem \ref{thbound} and Corollary \ref{reg} apply and $u\in C^0(\Omega)\cap L^\infty(\Omega)$. Moreover
\begin{equation}
\label{t20}
f(x, t)\ge\frac{ f(x,\|u\|_\infty)}{\|u\|_\infty^{p-1}}\, t^{p-1}\ge -\frac{h(\|u\|_\infty)}{\|u\|_\infty^{p-1}}\, t^{p-1} \qquad \forall t\in [0, \|u\|_\infty]
\end{equation}
so that the connectedness of $\Omega$ and Corollary \ref{smp} ensure that $u>0$ in $\Omega$.
Since $u$ is bounded, it holds $V_u=W^{1,p}_0(\Omega)$, hence its Euler-Lagrange equation \eqref{EL} holds true for any $\varphi\in W^{1,p}_0(\Omega)$. Let $\xi\in L^{p'}(\Omega; \mathbb{R}^N)$ be such that
\[
\xi(x)\in \partial H(Du(x))\quad \text{a.\,e.},\qquad \int_{\Omega}\left(\xi, D\varphi\right) -f(x,u)\, \varphi\, dx=0\quad \forall \varphi\in W^{1,p}_0(\Omega)
\]
and let $v\in W^{1,p}_0(\Omega)\cap L^\infty(\Omega)$ be non-negative. It is readily checked that for any $\varepsilon>0$ the function $\varphi=v^p/(u+\varepsilon)^{p-1}$ is an admissible test function in $W^{1,p}_0(\Omega)$, hence
\[
\int_\Omega \left(\xi, D \frac{v^p}{(u+\varepsilon)^{p-1}}\right)\, dx=\int_\Omega f(x, u)\, \frac{v^p}{(u+\varepsilon)^{p-1}}\, dx
\]
and by Picone inequality \eqref{picone} (notice that $\xi\in \partial H(Du)=\partial H(D(u+\varepsilon))$), we have
\begin{equation}
\label{t21}
p\, \int_\Omega H(Dv)\, dx\ge \int_\Omega \left(\xi, D \frac{v^p}{(u+\varepsilon)^{p-1}}\right)\, dx=\int_\Omega f(x, u)\, \frac{v^p}{(u+\varepsilon)^{p-1}}\, dx.
\end{equation}
We claim that
\begin{equation}
\label{t16}
\lim_{\varepsilon\downarrow 0} \int_\Omega \frac{f(x, u)}{(u+\varepsilon)^{p-1}}\, v^p\, dx=\int_\Omega \frac{f(x, u)}{u^{p-1}}\, v^p\, dx.
\end{equation}
Indeed, let
\[
\Omega_+=\big\{x\in \Omega: f(x, u(x))\ge 0\big\}, \qquad \Omega_-=\big\{x\in \Omega: f(x, u(x))<0\big\}.
\]
By Beppo-Levi monotone convergence theorem, \eqref{t16} holds true with $\Omega_+$ instead of $\Omega$, while on $\Omega_-$ it holds, by \eqref{t20},
\[
0\ge \frac{f(x, u)}{(u+\varepsilon)^{p-1}}\, v^p\ge -\frac{h(\|u\|_\infty)}{\|u\|_\infty^{p-1}}\, \frac{u^{p-1}}{(u+\varepsilon)^{p-1}}\, v^p\ge -\frac{h(\|u\|_\infty)}{\|u\|_\infty^{p-1}}\, \|v\|_\infty^p
\]
so that dominated convergence yields \eqref{t16} on $\Omega_-$ as well.
Therefore \eqref{t16} is proved, and passing to the limit as $\varepsilon\downarrow 0$ in \eqref{t21} yields
\begin{equation}
\label{t7}
p\, \int_\Omega H(Dv)\, dx\ge \int_\Omega \frac{f(x, u)}{u^{p-1}}\, v^p\, dx.
\end{equation}
Another monotone convergence argument (separately on $\Omega_+$ and $\Omega_-$) allows to remove the assumption $v\in L^\infty(\Omega)$, so that \eqref{t7} holds for all $v\in W^{1,p}_0(\Omega)$, $v\ge 0$. As a byproduct of this argument, let us notice that the integrand on the right of \eqref{t7} is always in $L^1(\Omega)$, for any $v\in W^{1, p}_0(\Omega)$, $v\ge 0$.
Let us prove assertion {\em (1)}. To this end, recall that $F(\cdot, v)_+\in L^1(\Omega)$ for all non-negative $v\in W^{1,p}_0(\Omega)$, thanks to the one-sided growth condition \eqref{t22}. Therefore it suffices to compare $J(u)$ with $J(v)$ for non-negative $v\in W^{1,p}_0(\Omega)$ such that $F(\cdot, v)\in L^1(\Omega)$.
We rewrite \eqref{t7} as
\[
J(v)\ge \int_\Omega \frac{f(x, u)}{u^{p-1}}\, \frac{v^p}{p}-F(x, v)\, dx
\]
and recall that, by \eqref{id},
\[
J(u)=\int_\Omega\frac{1}{p}\, f(x, u)\, u-F(x, u)\, dx,
\]
so we have
\[
J(v)-J(u)\ge \int_\Omega \frac{f(x, u)}{u^{p-1}}\, \left(\frac{v^p}{p}-\frac{u^p}{p}\right)+ F(x, u)-F(x, v)\, dx.
\]
We claim that the integrand is non-negative. Indeed, for a.\,e.\,$x\in \Omega$ it holds
\[
F(x, t)-F(x, s)=\int_s^t \frac{f(x, \sigma)}{\sigma^{p-1}}\, \sigma^{p-1}\, d\sigma
\]
so, considering separately the two cases $t\ge s\ge 0$ and $0\le t\le s$ and using the monotonicity of $\sigma\mapsto f(x, \sigma)/\sigma^{p-1}$, we get
\[
F(x, t)-F(x, s)\ge \frac{f(t)}{t^{p-1}}\, \left(\frac{t^p}{p}-\frac{s^p}{p}\right)
\]
for all $t, s\ge 0$. Setting $t=u(x)$ and $s=v(x)$ proves the claim, therefore $J(v)\ge J(u)$ for all $v\ge 0$, $v\in W^{1,p}_0(\Omega)$, concluding the proof of the first statement.
To prove statement {\em (2)}, suppose now that $v$ is another nontrivial, non-negative critical point for $J$, so that \eqref{t7} holds true with $u$ and $v$ exchanged. Recalling \eqref{id} we get
\[
\int_\Omega f(x, v)\, v\, dx\ge \int_\Omega \frac{f(x, u)}{u^{p-1}}\, v^p\, dx, \qquad \int_\Omega f(x, u)\, u\, dx\ge \int_\Omega \frac{f(x, v)}{v^{p-1}}\, u^p\, dx
\]
which, summed up, give
\[
\int_\Omega \left(\frac{f(x, u)}{u^{p-1}}-\frac{f(x, v)}{v^{p-1}}\right)\, (u^p-v^p)\, dx\ge 0.
\]
By assumption {\em ii)}, the latter forces
\begin{equation}
\label{t24}
\frac{f(x, u)}{u^{p-1}}=\frac{f(x, v)}{v^{p-1}}
\end{equation}
a.\,e.\,in $\Omega$, which implies that $u=v$ if $t\mapsto f(x, t)/t^{p-1}$ is strictly decreasing.
To prove assertion {\em (3)}, recall that by Picone inequality
\[
p\, H(Dv)-\left(\xi, D \frac{v^p}{(u+\varepsilon)^{p-1}}\right)\ge 0,
\]
for all $\varepsilon\ge 0$. Hence, chaining Fatou lemma, the weak form of the equation for $u$, \eqref{t16}, \eqref{t24} and \eqref{id}, we get
\[
\begin{split}
\int_\Omega p\, H(Dv)-\left(\xi, D \frac{v^p}{u^{p-1}}\right)\, dx&\le \lim_{\varepsilon\downarrow 0}\int_\Omega p\, H(Dv)-\left(\xi, D \frac{v^p}{(u+\varepsilon)^{p-1}}\right)\, dx\\
&= \lim_{\varepsilon\downarrow 0}\int_\Omega p\, H(Dv)-f(x, u)\, \frac{v^p}{(u+\varepsilon)^{p-1}}\, dx\\
&= \int_\Omega p\, H(Dv)-\frac{f(x, u)}{u^{p-1}}\, v^p\, dx\\
&= \int_\Omega p\, H(Dv)-f(x, v)\, v\, dx=0
\end{split}
\]
so that equality must occur a.\,e.\,in the Picone inequality. If $H$ is strictly convex, Lemma \ref{thpicone} ensures that
\begin{equation}\label{eq:prop}
u=k\, v \,,\qquad\mbox{for some $k>0$,}
\end{equation}
giving assertion {\em (3)}.
Finally, statement {\em (4)} amounts to prove that, in the case $f(x, t)=f(t)$, the validity of \eqref{eq:prop} for some $k\ne 1$ forces $u$ to be a first Dirichlet positive eigenfunction.
Let $g(t):=f(t)/t^{p-1}$ and assume that $k>1$ in \eqref{eq:prop}. For any fixed $\delta\in \ ]0, \|u\|_\infty[$, by continuity $u(\Omega)$ is connected and contains $]0, \|u\|_\infty-\delta]$. Hence for any $n\in \mathbb{N}$ there exists $x_n\in \Omega$ such that
\[
u(x_{n})=\frac{\| u \|_\infty-\delta}{k^n}
\]
and thus \eqref{t24} and \eqref{eq:prop} give
\[
\begin{split}
g(\|u\|_\infty-\delta)&=g(u(x_0))=g(v(x_0))=g(u(x_0)/k)=g((\|u\|_\infty-\delta)/k)=g(u(x_1))\\
&=g(v(x_1))=g(u(x_1)/k)=g((\|u\|_\infty-\delta)/k^2)=\dots=g((\|u\|_\infty-\delta)/k^n)
\end{split}
\]
for any $n\ge 0$. Therefore $g$, being continuous and non-increasing, is constantly equal to some $\lambda\ge 0$ on $]0, \|u\|_\infty-\delta]$. By letting $\delta\downarrow 0$, the same conclusion holds on $]0, \|u\|_\infty]$. Similarly, if $0<k<1$, we infer that $g\equiv \lambda\ge 0$ on
\[
]0, \|v\|_\infty]=\ ]0, \|u\|_\infty/k]\supseteq \ ]0, \|u\|_\infty].
\]
Notice that then $f(t)=\lambda\, t^{p-1}$ for $t\in \left[0, \|u\|_\infty\right]$, hence
\begin{equation}
\label{F}
F(t)=\frac{\lambda}{p}\, t^p\qquad \text{ for all $t\in [0, \|u\|_\infty ]$}.
\end{equation}
By \eqref{id} it holds
\[
J(u)=\int_\Omega H(Du)-\frac{\lambda}{p} \, u^p\, dx=0
\]
and by the first assertion of the Theorem
\[
J(v)\ge J(u)\qquad \forall v\in W^{1,p}_0(\Omega), \ v\ge 0.
\]
If $0\le v\le \|u\|_\infty$, from the last two displays and \eqref{F} we infer
\[
\frac{\lambda}{p} \int_\Omega v^p\, dx\le \int_\Omega H(Dv)\, dx.
\]
By positive $p$-homogeneity, the latter holds for any bounded $v\ge 0$, $v\ne 0$, (simply writing it for $w=v\, \|u\|_\infty/\|v\|_\infty$, which fulfils $0\le w\le \|u\|_\infty$) and thus for any $v\ge 0$ in $W^{1,p}_0(\Omega)$ by a monotone convergence argument. It follows that $\lambda/p\le \lambda_{1, H}^+(\Omega)$ and $J(u)=0$ implies
\[
\lambda_{1, H}^+(\Omega)\, \int_\Omega u^p\, dx\le \int_\Omega H(Du)\, dx=\frac{\lambda}{p}\, \int_\Omega u^p\, dx \le \lambda_{1, H}^+(\Omega)\, \int_\Omega u^p\, dx
\]
i.\,e., $u$ is a first Dirichlet positive eigenfunction, as claimed.
\end{proof}
|
\section{Introduction}
The current high precision era of lattice QCD introduces new questions on data generation, availability and access. With the increase in precision comes also an increase in numerical cost as well as the size and complexity of the data. Especially data sharing and community access become important as the process of transferring data can be a problem in itself and not having access can prohibit impactful physics analyses.
Simultaneously, the current level of precision is built on a wealth of previous data driven studies. In principle progress can be slowed by having to invest first to re-establish a comparable foundation for a new method.
In this situation it is important as community to re-evaluate the existing data sharing infrastructure, e.g. in form of the ILDG, and develop an expanded understanding of open science as a tool for scientific progress.
\section{The open lattice initiative, setup and features}
The open lattice initiative (OpenLat) was established as a response to some of these questions. Bringing together researchers across the globe, the goal is to generate state-of-the-art QCD gauge ensembles for physics applications and to share them with the community to strengthen open science.
As part of this effort the initiative aims to transparently define and uphold quality for the configurations provided (1), share and maintain an easily accessible repository (2), gather resources towards a common aim (3), and to grant access to any interested party freely and quickly (4).
The stabilised Wilson fermion (SWF) framework \cite{Francis:2019muy} includes a modified improvement term for the fermion action and also a number of advancements aimed at algorithmic stability\footnote{The SWF framework is also being used in the context of the master-field approach, see \cite{pfritzsch:lattice22} for an overview.}.
These measures are implemented in addition to the established techniques implemented within the open source software package \texttt{openQCD-2.0} and \texttt{2.4} \cite{mluscher:openqcd}, such as the Schwarz-alternating-procedure (SAP), local deflation, multi-grid, mass-preconditioning, multiple time-scale integrators and others.
All generated gauge ensembles are in $N_f=2+1$ QCD. To set the scale the gradient flow time criterion \cite{Luscher:2010iy} is used and results are converted to physical units via $\sqrt{8t_0}=0.414(5)\textrm{fm}$ \cite{Bruno:2016plf} .
\subsection{Stabilised Wilson Fermions}
\subsubsection{Exponentiated Clover action}
An important new component in OpenLat is the inclusion of a new fermion action resulting from a modification of the standard Wilson clover (WCF) action. The (traditional) $\mathcal{O}(a)$-improved Dirac operator takes the form
\begin{equation}
D = \frac{1}{2} \Big[\, \gamma_\mu\Big( \nabla_\mu^*+\nabla_\mu \Big)\,- \nabla_\mu^*\nabla_\mu \Big] + c_{SW} \frac{i}{4}\sigma_{\mu\nu}\hat F_{\mu\nu} + m_0~~,
\end{equation}
which, after even/odd preconditioning, may be written as $\hat D = D_{ee} - D_{eo} ( D_{oo} )^{-1} D_{oe}$, with a diagonal part ($M_0=4 +m_0$) given by
$D_{ee} + D_{oo} = M_0 + c_{SW} \frac{i}{4} \sigma_{\mu\nu} \hat F_{\mu\nu} $. Notice that arbitrarily small eigenvalues can occur as the clover term can saturate the bound
$\Vert \frac{i}{4} \sigma_{\mu\nu} \hat F_{\mu\nu} \Vert_2 \leq 3$
while the clover coefficient $c_{SW}$ is one at tree-level and then grows monotonically with $g_0^2$.
Furthermore, the positive and negative eigenvalues of the clover term are equally distributed.
Taken together this makes the above statement more precise as we see that $D_{oo}$ is not protected from arbitrarily small eigenvalues. The probability to encounter arbitrarily small eigenvalues of $D_{oo}$ is larger, the smaller the quark masses, and/or the coarser/larger the lattice.
This issue can be resolved by reformulating the clover term:
\begin{equation}
D_{ee} + D_{oo} = M_0 + c_{SW} \frac{i}{4} \sigma_{\mu\nu} \hat F_{\mu\nu} ~\rightarrow~~ M_0\exp\Big[ \frac{c_{SW}}{M_0}\frac{i}{4} \sigma_{\mu\nu} \hat F_{\mu\nu} \Big]~~,
\end{equation}
this is the so-called exponentiated clover \cite{Francis:2019muy}.
This formulation is still invertible, local and has cutoff effects described, as usual, by a Symanzik effective theory. While the lattice Dirac operator
is still not protected from exceptionally small eigenvalues stemming from the Wilson term, this new formulation eliminates the clover term as a potential
source of instabilities.
\subsubsection{Algorithmic features}
Part of the SWF framework is to include a number of algorithmic improvements, one of them is choosing the Stochastic Molecular Dynamics (SMD) algorithm \cite{Horowitz:1985kd,Horowitz:1986dt,HOROWITZ1991247,Jansen:1995gz} over the commonly used HMC. Benefits and features of the SMD were presented in \cite{Luscher:2011kk,Luscher:2017cjh}. Here we note that the algorithm leads to an increased stability and decreased autocorrelation times by reducing unbounded energy violations $|\delta H|\gg 1$.
The HMC and SMD are closely related, for example the HMC is a limiting case of the SMD for large step lengths and friction coefficients. In the opposite direction the SMD becomes related to a Langevin process.
As such, in an SMD update cycle the momentum fields are randomly rotated as opposed to randomly re-drawn and an update length a fraction of the typical trajectory length is chosen.
One special feature of the SMD is that once there is rejection the momentum tends to reverse and trajectories backtrack with a period $t_{acc}=\delta \tau P_{acc}/(1-P_{acc})$ \cite{Luscher:2011kk}. This makes a very a high acceptance rate necessary.
Further improvement in stability can be achieved by changing the solver stopping criterion a volume-independent norm $\Vert \eta \Vert_\infty=\textrm{sup}_x \Vert \eta \Vert_2$ that prevents precision losses from local effects.
Finally, to limit the effect of an accumulation of numerical errors on global sums quadruple precision is implemented and used.
\subsection{Tuning Setup}
The clover coefficient $c_{SW}$ is tuned according to the procedure followed in \cite{Luscher:1996ug}. It is performed in the range $\beta\geq 3.685$ \cite{Francis:2019muy,Francis:2022hyr} corresponding to a coarse lattice spacing of $a=0.12$~fm. The resulting interpolation formula is given by
\begin{equation}
c_{SW}(g_0^2) = \frac{1-0.325022g_0^2-0.0167274g_0^4}{1-0.489157g_0^2}, \qquad \beta=6/g_0^2\geq 3.685.
\end{equation}
Special attention needs to be given to choosing the bare mass parameters for fixing the trajectory to physical masses and the continuum limit with Wilson fermions. We adopt a strategy that aims at reducing mass-dependent cutoff effects by working at constant trace of the simulated quark mass matrix \cite{Bietenholz:2010jr, Bruno:2014jqa,Strassberger:2021tsu}.
Starting at the flavour-symmetric point, where all pseudoscalar meson masses are degenerate, we set
$
\textrm{tr}[M] = m_u + m_d + m_s = N_f \cdot m_\ell = \textrm{const} ~,
$
such that it depends on a single mass parameter. The value of $m_\ell$ amounts to matching the combination of the physical ground state masses of the pion and the kaon:
$ m_{\pi K}^2 = \frac{2}{3}\Big( m_K^2 + m_\pi^2/2\Big) \equiv m_{\pi K}^2|_{\textrm{phys}}$.
Following the procedure outlined in \cite{Aoki:2016frl} the input value is $m_{\pi K}|_{\textrm{phys}}=410.9(2)\,\rm{MeV}$.
Once the starting point is tuned and fixed, the quark masses are changed such that $\textrm{tr}[M]=\textrm{const}$.
\subsection{Quality control strategy}\label{sec:validation}
\subsubsection{Run stability and health}
During the generation of the gauge configurations a number of standard observables is continuously monitored and sanity checks are performed to test the health of the run. A few prominent validation observables are: the plaquette, topological charge, $\Delta H$ distribution, reweighting factors, distributions of the lowest eigenvalue $\lambda$ of $\sqrt{D^\dag D}$, and simple spectroscopic observables.
In particular, our checks are meant to ensure that\footnote{The last two items in the following complement the list we made in earlier references \cite{Francis:2022hyr}. In the future we plan to add the sign of the reweighting factor to the standard observables}:
\begin{itemize}
\itemsep0em
\item $\phi_4=8t_0( m_K^2 + m_\pi^2/2)=1.115$ within $0.5\%$, with an error of max. $1\sigma$.
\item The total reweighting factor fluctuations are mild, and ideally below $5\%$.
\item The SMD step distance $\delta \tau$ maximises the backtracking period.
\item The distribution of $\delta H$ matches the one set by the acceptance rate.
\item The distribution of the lowest $\sqrt{D^\dag D}$ eigenvalue is well-behaved and gapped.
\item\label{item:poles1} The distribution of the lower/upper bounds of the spectral gap for the strange quark are within the input ranges, and the degree of the Zolotarev is sufficiently high, $12 (V/2) \delta^2 < 10^{-4}$ \cite{mluscher:openqcd}.
\item\label{item:poles2} There is no significant loss of precision caused by unbalanced contributions to the total action that might drive instabilities in the evolution.
\item\label{item:topCharge} The distribution of the flowed topological charge is symmetric around zero with no signs of metastability.
\end{itemize}
\subsubsection{Distance between configurations and thermalisation}
Estimating the distance between two configurations that can be safely labelled as independent is based on evaluating the integrated autocorrelation times of several observables and choosing the slowest observed one. In most cases this is the topological charge $Q$, which is computed using the gradient flow \cite{Luscher:2010iy}. However, if another scale is seen to be even slower the distance is switched to this value instead.
The boundary conditions are periodic/anti-periodic by default. However, once a significant increase of the autocorrelation time of $Q$ is observed the boundaries are changed to open boundary conditions. The increase is interpreted as a signal of entering a regime of topology freezing and critical slowing down.
Currently this is only the case for $\beta=4.1$ and $a=0.055$~fm. All ensembles reported on in this proceedings contribution do not exhibit this feature and have periodic/anti-periodic boundary conditions.
The plaquette, $t^2\langle E\rangle$, where $t$ is the flow time and the density $E$ is defined in \cite{Luscher:2010iy}, pseudoscalar-pseudoscalar ($PP$) and pseudoscalar-axial ($PA$) meson correlators are checked for visible effects in order to estimate sufficient thermalisation. Nevertheless, a minimum of five integrated autocorrelation lengths, of the slowest scale, is imposed. This naturally relies on an accurate determination of the autocorrelation lengths which become available only after sufficiently long MC strings have been generated.
\subsubsection{Tuning and production levels}
\label{sec:tuningAndProd}
The gauge generation process is separated into two levels: The first is where thermalisation performed and the parameters of the algorithm tuned or adjusted. Once the algorithm is stable and thermalised, 100 independent configurations are generated. This is called the tuning level and once the health of the run is confirmed it is deemed production ready. At this point the initial 100 configurations become part of the full ensemble statistics.
At the production level configurations are produced with all run parameters fixed. Once the production reaches 500+ independent configurations in an ensemble it is labelled publication ready.
\section{Gauge generation and dissemination status}
\subsection{Production, publication and access plan}
\label{sec:plan}
The production of ensembles will proceed in three stages:
\begin{itemize}
\itemsep0ex
\item [{\bf 1.}] {\bf Stage:} Perform high precision tuning and generate ensembles with 3 dynamical flavors at the SU(3) flavor symmetric point. The minimum goal is to generate 500+ independent configurations on 4+ lattice spacings ($a=0.12, 0.094, 0.077$ and $0.064$~fm).
\item [{\bf 2.}] {\bf Stage:} Reduce the light quark masses with Tr$[M]=$const. The minimum goal is to have matched pion passes at $M_\pi=300, 200$~MeV for $a=0.12, 0.094, 0.077$ and $0.064$~fm. An additional lattice with $a=0.055$fm (open boundary conditions) will be added at this stage.
\item [{\bf 3.}] {\bf Stage:} Go towards the physical values of the pion mass $m_\pi=135$~MeV. The minimum goal is to supply at least one ensemble at the physical point with 500+ configurations and then to extend towards all lattice spacings.
\end{itemize}
Multiple lattice volumes are generated at all stages and are extended towards 500+ configurations, where reasonable (i.e. $L\gtrsim 3$fm, $m_\pi L\gtrsim 4$).
Each completed stage is accompanied by a reference publication. With this publication all configurations and metadata of the given stage are made openly available without further embargo time. The data that will be made public are a metadata catalogue, the gauge configurations themselves and additionally the standard, or validation, observables highlighted above.
Details on how and where to download the configurations and their corresponding auxiliary data will be given at the same time.
The metadata catalogue will be made compliant with further community standards, such as ILDG, to facilitate data sharing and a common data standard within the community.
Users may obtain access to the configurations of ongoing, i.e. unpublished, stages. This user-access is granted on a case-by-case basis\footnote{Users have reported results using OpenLat ensembles already at this and other conferences in \cite{fjoswig:lattice22,jgreen:lattice22,jkim:ectstar22}.}.
\subsection{Data management plan}
Currently OpenLat maintains a repository of 100TB of data. This includes $\sim20$k saved configurations and is projected to grow to 500+TB with completion of stage 2. The repository is mirrored at currently two separate locations to ensure redundancy and a tape option is used for long term storage. All metadata is preserved on disk and in the main online repository at \href{https://openlat1.gitlab.io}{https://openlat1.gitlab.io}. Currently this access is only internal, it will however be made open upon completion of stage 1. The internal metadata follows a detailed provenance policy that includes run and machine information along side the previously mentioned standard observables.
To maintain data integrity a two part strategy is followed: the configurations contain the plaquette in the header, as per openQCD format standard, and is augmented by a separately kept list of checksums of sufficient complexity (e.g. \texttt{sha512}).
\begin{table}[t!]
\centering
\begin{tabular}{ccccrrccccr}
\toprule
Stage & $m_\pi$[MeV] & $\beta$ & $a\rm{[fm]}$ & {$L$} & {$T$}& BC & {$N_{\rm{cfg}}$} & MDU \\\hline
1 & 412 & 3.685 & 0.12 & 24 & 96 & P & 1200 & 24800\\
& & & & 32 & 96 & P & 400 & 8300 \\
& & 3.8 & 0.094 & 24 & 96 & P & 1200 & 35000\\
& & & & 32 & 96 & P & 1300 & 38800\\
& & 3.9 & 0.077 & 48 & 96 & P & 300 & 11000 \\
& & 4.0 & 0.064 & 48 & 96 & P & 600 & 24500\\ \hline
2 & 300 & 3.685 & 0.12 & 24 & 96 & P & 600 & 13000\\
& & 3.8 & 0.094 & 32 & 96 & P & 300 & 8000 \\
& & 4.0 & 0.064 & 48 & 96 & P & 200 & 6000 \\ \hline
\toprule
\end{tabular}
\caption{Stage 1 and 2 production status overview: All ensembles are currently still gathering more statistics and numbers should be regarded as preliminary. See Sec.\ref{sec:plan} for details on planned ensembles and target statistics. Note that the numbers in MDU have been derived by converting from SMD trajectory length.
}
\label{tab:ensembles}
\end{table}
\section{Production status}
The initiative is currently in the process of completing stage 1 of the production plan while stage 2 is being pursued, see Sec.~\ref{sec:plan}. An overview of the current number of configurations and ensembles available at the production level is given in Tab.~\ref{tab:ensembles}.
The information on the MC chain length in MDU was obtained by multiplying the SMD step length with the total number of cycles
An overview of the full production landscape is given in Fig.~\ref{fig:overview}.
The figure shows the sets of pion masses and lattice spacings that are at the tuning level, i.e. have less than 100 independent configurations or are being otherwise thermalised or tuned, in shaded boxes. The solid boxes show those ensembles that are at the production level. Their height denotes the number of configurations they have.
The colored curved lines denote lines of constant $m_\pi L$, while the grey straight line denotes $L=3$~fm. The dashed floor lines are to guide the eyes as to the location of the target values in $m_\pi$[MeV]. The width and depth of the boxes have been chosen for legibility and do not reflect errors in $m_\pi$ or $L$.
Note that 500 configurations for all $SU(3)$ flavor points is threshold for the completion of stage 1 and also making the ensembles openly accessible.
The focus on performing a high precision tuning and providing high statistics ensembles at the flavor symmetric point is clearly visible. Furthermore the newest lattice spacing, $a=0.077$~fm, stands out for being the lowest in statistics. With respect to a previous version of this figure we have removed an older physical point ensemble at $a=0.094$~fm for having a too small volume with $m_\pi L\simeq 2.1$.
\begin{figure}
\centering
\includegraphics[width=0.69\textwidth]{./pics/mLs_swf2.png}
\caption{Overview of ensembles at the tuning (shaded boxes) and production (solid boxes) levels. The shaded or colored lines are to highlight: Lines of constant $m_\pi L$, values of $m_\pi$[MeV], lattice volumes $L$ and the target number of configurations.}
\label{fig:overview}
\end{figure}
\section{Summary}
In this contribution, we highlighted the recent progress made by the open lattice initiative. The initiative was founded with the expressed purpose of generating and providing ensembles for physics applications under open science. The ensembles are generated using the stabilised Wilson fermion framework, which is further researched and refined as part of the effort.
The three stage production plan has been laid out and an overview of the production status given. As stage 1 is nearing completion, once its reference publication is published the plan is that access will be granted to the configurations under the presented data management and access plans.
The procedures and results presented in this contribution should be regarded as preliminary and will be superseded by their final versions in the upcoming publication.
\section*{Acknowledgements}
The authors acknowledge support from the HPC computing centres and resources hpc-qcd (CERN),
HPE Apollo Hawk (HLRS) under the grant number stabwf/44185, Cori (NERSC), Frontera (TACC), Piz Daint (CSCS) as well as Occigen (CINES), Jean-Zay (IDRIS) and Ir\`ene-Joliot-Curie (TGCC) under projects (2020,2021,2022)-A0080511504 and (2020,2021,2022)-A0080502271 by GENCI as well as project 2021250098 by PRACE. AS acknowledges funding support under the National Science Foundation grant PHY-2209185. AF acknowledges support by the Ministry of Science and Technology Taiwan (MOST) under grant 111-2112-M-A49-018-MY2. We greatly acknowledge the leading role of Martin L\"uscher during the development and implementation of the SWF framework.
\bibliographystyle{JHEP}
|
\section{Introduction}\label{sec:introduction}
It is well-known that blockchain-based systems suffer from the information isolation problem~\cite{xie2019survey}, which prevents
message transports, value exchanges, and collaborative operations among blockchains, hindering the advantages of blockchain technologies in consensus, trust, and cooperations~\cite{xu2022spdl,chen2022survey, xu2022cloudchain,guo2022blockchain,liu2022blockchain}. Cross-chain technology has been considered to be one of the effective ways to solve the isolation problem, aiming to build a bridge for the communications and coordinations among isolated blockchain systems~\cite{belchior2021survey}.
Nevertheless, most existing cross-chain techniques such as notaries in notary schemes~\cite{belchior2021survey} and relay chains in sidechains/relays~\cite{wood2016polkadot}, rely on third parties that are assumed to be safe, which reduces their availability and makes them extremely vulnerable to the single point of failure problem.
Hashed TimeLock Contract (HTLC)~\cite{poon2016bitcoin} is a decentralized cross-chain scheme that employs smart contracts to ensure atomicity of transactions. But unfortunately, in HTLC, one cross-chain exchange requires multiple on-chain consensus, which would undoubtedly decrease the transaction rate and increase the transaction latency, further deepening the blockchain system's poor scalability.
Therefore, one can conclude that current cross-chain schemes cannot realize efficient cross-chain interoperability without relying on third parties.
In fact, researchers have put forward feasible schemes such as blockchain sharding, off-chain channels, and roll-up techniques (zk-rollup, optimistic rollup)~\cite{qi2022s, huang2022brokerchain, papadis2022payment, xie2019survey}, to enhance the scalability of single blockchain systems.
Among them, off-chain channels (see an example shown in Fig.~\ref{Fig:UAS}), which employ on-chain processes to establish and close a channel and off-chain operations to carry out tasks within the channel, provide faster transaction processing, need lower effort in hardware configuration, and have been successfully applied to Lightning Network~\cite{poon2016bitcoin}.
Therefore, it is of great significance to extend the current off-chain channel schemes for cross-chain services. Nevertheless, this is a nontrivial task. There exist two open challenges that should be addressed in order to take advantage of the high throughput of channels for cross-chain operations.
\begin{figure}[t!]
\centering
\includegraphics[width=0.8\textwidth]{figure/Intro-UAS.pdf}
\caption{A channel scheme example, in which a solid line represents an interaction process, and the shaded belt represents a payment channel, a state channel or a virtual channel. Note that step \textcircled{1} and step \textcircled{4} are on-chain processes that open and close the channel $\Omega$, respectively. In step \textcircled{2}, $\mathcal{S}$ sends transaction $Tx_1$ to $\mathcal{R}$. Step \textcircled{3} indicates that $\mathcal{S}$ and $\mathcal{R}$ send more transactions to each other via this channel. These transaction operations take place in the channel $\Omega$ and do not consume any on-chain resource.
}
\label{Fig:UAS}
\end{figure}
First, current channel schemes such as payment channels~\cite{poon2016bitcoin, hearn2013micro, decker2015fast, papadis2022payment}, state channels~\cite{dziembowski2018general,coleman2018counterfactual} and virtual channels~\cite{dziembowski2019perun, dziembowski2019multi} cannot support spending unsettled amounts. As shown in Fig.~\ref{Fig:UAS}, suppose sender $\mathcal{S}$ sends some amount $x$ in $\mathsf{Tx_1}$ to receiver $\mathcal{R}$ at time $\mathsf{T_1}$. $\mathcal{R}$ cannot use $x$ before both parties successfully close the channel at $\mathsf{T_2}$. In fact, the longer time the channel stays open, the longer the time ($\mathsf{T_2}$-$\mathsf{T_1}$) $\mathcal{R}$ needs to wait before using $x$. This latency becomes even bigger
when cross-chain operations are involved as the multi-chain heterogeneous design brings more dimensions of complexity.
We term this challenge \emph{Unsettled Amount Congestion (UAC) problem}. To address this issue, one needs to design a new channel architecture and a corresponding smart contract protocol to support flexible user joins while ensuring the correctness of settlement.
\begin{figure}[t]
\centering
\includegraphics[width=0.8\textwidth]{figure/Intro-UE.pdf}
\caption{An example of the Unfair Exchange (UE) problem. A solid line represents an interaction process. }
\label{Fig:UE}
\end{figure}
Second, the scenarios of cross-chain interactions are diversified, which requires a protocol to help channels ensure interaction fairness.
Current fair exchange schemes such as ZKCP~\cite{zkcp} and FairSwap~\cite{dziembowski2018fairswap} rely on cryptocurrencies for settlement, which assumes that one of the exchange objects can be made public (generally this object defaults to cryptocurrency).
Unfortunately, in a cross-chain scenario, affected by the risk of sharp currency price fluctuations, more users choose to interact in a barter way. This makes current fair exchange schemes fail to guarantee the fairness of the interaction process, as none of the two parties might be willing to be the first to disclose its secret that is employed to protect the exchange object (shown in Fig.~\ref{Fig:UE}). We define this to be the \emph{Unfair Exchange (UE) problem}, which severely limits the applications of cross-chain protocols. The difficulty in solving this problem lies in how to ensure the disclosure of both parties' secrets.
In this paper, we propose $\mathsf{Cross}$-$\mathsf{Channel}$ to effectively address the above two challenges.
First, in order to solve the UAC problem, we design a novel hierarchical channel architecture with a hierarchical settlement protocol, which allows channel initiators to establish sub-channels with other participants in order to use an unsettled amount, improve the throughput of the processed transactions, and ensure the correctness of the final settlement.
Moreover, we present a general fair exchange protocol based on zk-SNARK~\cite{groth2016size} and $(t,n)$-VSS (Verifiable Secret Sharing)~\cite{pedersen1991non} to address the UE problem. Specifically, this protocol adopts zk-SNARK to guarantee the authenticity and privacy of information, and employs $(t,n)$-VSS with a smart contract to ensure the fair disclosure of both parties' secrets.
Finally, we adopt the HTLC to enhance the atomicity of cross-chain interactions. Note that HTLC was originally proposed for synchronous networks, while in an asynchronous network, some nodes may be affected by high latency and cannot successfully receive or send messages, which makes blocked nodes unable to complete the protocol due to timeouts. Therefore, we develop an incentive mechanism to make HTLC, thereby $\mathsf{Cross\text{-}Channel}$, suitable for asynchronous networks.
For convenience, we highlight our contributions as follows:
\begin{enumerate}
\item $\mathsf{Cross}$-$\mathsf{Channel}$ is the first channel scheme to support cross-chain operations in both synchronous and asynchronous networks. To the best of our knowledge, current channel schemes only support intra-chain operations
\item We design a novel hierarchical channel architecture with a hierarchical settlement protocol, which can effectively solve the UAC problem. The proposed new architecture can further improve the throughput of $\mathsf{Cross}$-$\mathsf{Channel}$ and is well-suited for processing large-scale transactions.
%
\item A general fair exchange protocol is proposed in this paper to guarantee the disclosure of both parties' secrets, ensure the fairness of cross-chain interactions, and further solve the UE problem.
\item $\mathsf{Cross}$-$\mathsf{Channel}$ can support various cross-chain operations, especially for the exchange of encrypted information that does not rely on cryptocurrencies.
\item Extensive simulation experiments in AliCloud are conducted to validate the performance of $\mathsf{Cross}$-$\mathsf{Channel}$.
\end{enumerate}
The rest of the paper is organized as follows.
We first review the most related work in Section~\ref{sec:related work}.
Then, we present the models of our scheme,
and briefly introduce the necessary preliminary knowledge in Section~\ref{sec:Preliminaries}.
In Section~\ref{sec:Cross-Channel}, we detail our scheme $\mathsf{Cross}$-$\mathsf{Channel}$ considering different applications.
Section~\ref{sec:Implementation} reports the simulation experiments on AliCloud to evaluate the performance of $\mathsf{Cross}$-$\mathsf{Channel}$.
Finally, we provide concluding remarks in Section~\ref{sec:Conclusion}.
\section{Related Work and Motivation}\label{sec:related work}
In this section, we introduce a few well-known off-chain channel schemes related to our design, including payment channels, state channels and virtual channels.
\noindent\textbf{Payment channel schemes.}
A payment channel is a temporary off-chain trading channel for improving the transaction throughput of the entire system.
It was originally designed as a one-way channel~\cite{hearn2013micro}, and later evolved into a bi-directional channel so that one party can be both a sender and a receiver~\cite{decker2015fast}.
The most widely discussed recent projects are
Lightning Network~\cite{poon2016bitcoin} and Raiden~\cite{Raiden}, which establish payment channels in Bitcoin~\cite{nakamoto2008bitcoin} and Ethereum~\cite{wood2014ethereum}, respectively. In recent years, payment channel schemes with different features such as re-balancing, throughput maximization, attack resistance, and privacy protection,
have been constructed~\cite{li2020secure, khalil2017revive, papadis2022payment, green2017bolt, malavolta2017concurrency}.
\noindent\textbf{State channel schemes.}
A state channel enriches the functionality of
a payment channel. Concretely, the users of a state channel can,
besides payments, execute complex smart contracts in an off-chain
way (e.g., voting, auctions) and allow the exchange of states between two or more participants~\cite{dziembowski2018general, close2018forcemove}. The concept of state channel was proposed by Jeff Coleman~\cite{statechannel}. Later, Counterfactual~\cite{coleman2018counterfactual}
gave a detailed design and Dziembowski {\em et~al.}~\cite{dziembowski2018general} provided formal definitions and security proofs for the general state channel network. ForceMove~\cite{close2018forcemove} is a framework that can support $n$-party participation in a state channel. State channel schemes with
faster payment speeds were developed in~\cite{miller2019sprites, chakravarty2020hydra}.
\noindent\textbf{Virtual channel schemes.}
Virtual channels enable the creation,
progression, and closing of the channel without interacting with the underlying blockchain.
Dziembowski {\em et~al.}~proposed Perun~\cite{dziembowski2019perun}, the first virtual channel scheme in Ethereum. Later, they presented another scheme in~\cite{dziembowski2019multi}, discussing how to support virtual multi-party state channels. Aumayr {\em et~al.}~\cite{aumayr2021bitcoin} designed a virtual channel compatible with Bitcoin, proving that the establishment of a virtual channel can be independent of smart contracts.
\noindent\textbf{Summary and motivation.} According to the above analysis, one can see that
the emergence of state channels broadens the application of payment channels, enabling off-chain channels to provide more services. Virtual channels can effectively reduce the cost of channel network establishment and improve the efficiency of transaction processing.
Even though these channel schemes can successfully enhance the scalability of blockchain systems, they were originally proposed for operations within a single-chain, and cannot be directly extended to support cross-chain operations considering the challenges brought by the problems of UAC and UE. Furthermore, the design of the
current cross-chain solutions that do not rely on third parties, e.g., HTLC,
targets synchronous networks, while the unbounded latency in asynchronous networks may render them completely fail.
Motivated by these considerations,
we propose $\mathsf{Cross}$-$\mathsf{Channel}$ in this paper, which can effectively support efficient and fair atomic cross-chain operations under decentralized asynchronous networks.
\section{Models and Preliminaries}\label{sec:Preliminaries}
In this section, we first define our system model and threat model, then
provide preliminaries on zero-knowledge proof, fair exchange, hashed timelock contracts, and threshold key management
\subsection{Models}\label{model}
In this paper, we consider building a channel between heterogeneous blockchains. Such a channel involves three entities: sender ($\mathcal{S}$), receiver ($\mathcal{R}$), and blockchain miners ($\mathcal{M}$). $\mathcal{S}$ and $\mathcal{R}$ are the two parties of a channel interaction, being responsible for opening and closing the channel, uploading signature information, etc.
$\mathcal{M}$ is required to execute a smart contract to determine the legitimacy of the uploaded information, and honest miners would be accordingly rewarded by the blockchain incentive mechanism (just like the main chain of Ethereum).
\begin{itemize}
\item Sender $\mathcal{S}$. We assume that they can be arbitrarily malicious, and can act in their best interests.
\item Receiver $\mathcal{R}$. We assume that they can be arbitrarily malicious, and can act in their best interests.
\item Miner $\mathcal{M}$. Multiple miners follow a secure consensus algorithm to maintain the blockchain. Adversaries cannot compromise the majority of them to bring down the overall blockchain system.
\end{itemize}
All transactions can be divided into two categories, with one being the traditional on-chain transactions (or called transactions), which are confirmed and verified through the blockchain consensus mechanism, and the other being the off-chain transactions (or called receipts), which exist in channels and are verified by nodes within the channel. Some receipts would eventually be packaged into on-chain transactions and update the on-chain states of the nodes in the channel.
$\mathsf{Cross}$-$\mathsf{Channel}$ aims to realize scalability, fairness, and atomicity.
To achieve these goals, we next introduce the adopted key technologies.
\subsection{Zero-knowledge Proof: zk-SNARK} \label{zk-snark}
zk-SNARK (zero knowledge Succinct Non-interactive ARgument of Knowledge) is one type of the zero-knowledge proofs, which allows one party (the prover) to prove to another party (the verifier) that a statement is true, without revealing any information beyond the validity of the statement~\cite{groth2016size}.
\newtheorem{mydef}{Definition}
\begin{mydef}[zk-SNARK for an $\mathbb{F}$-arithmetic Circuit]
An $\mathbb{F}$-arithmetic circuit $C$ takes inputs (public inputs $\vec{x}$, private inputs $\vec{w}$ ) from a finite field such as ($\mathbb{F}^n$, $\mathbb{F}^h$), and outputs the result ( $\in \mathbb{F}^l$) based on the circuit logic. A zk-SNARK scheme essentially aims to ensure the satisfaction ($C(\vec{x}, \vec{w}) = 0^l$) of $C$, denoted as $R_C$. The whole process can be represented by a tuple of polynomial-time algorithms $\Pi$ $\overset{\text{def}}{=}$ (Setup, Prove, Verify):
\begin{itemize}
\item ${\mathtt{Setup}}(1^{\lambda}, C) \rightarrow (\mathsf{pk, vk})$.
The algorithm $\mathtt{Setup}$ takes a security parameter $1^{\lambda}$ and a circuit $C$ as inputs to obtain the key pair $(\mathsf{pk, vk})$, where $\mathsf{pk}$ is the proving key for proof generation and $\mathsf{vk}$ is the verification key for proof verification. The pair $(\mathsf{pk, vk})$ constitutes the common reference string $\mathsf{crs}$.
\item $\mathtt{Prove}(\mathsf{pk}, \vec{x}, \vec{w}) \rightarrow \pi $.
The algorithm takes as inputs the proving key $\mathsf{pk}$, the public inputs $\vec{x}$ and the private inputs $\vec{w}$ to generate a succinct zero-knowledge proof $\pi$.
\item $ \mathtt{Verify}(\mathsf{vk}, \vec{x}, \pi) \rightarrow \mathsf{1/0}$.
The algorithm verifies $\pi$ based on the verification key $\mathsf{vk}$ and public inputs $\vec{x}$. It returns 1 if the verification is successful and 0 otherwise.
\end{itemize}
\end{mydef}
Given a security parameter $\lambda$ and a circuit $C$ with a relation $R_C$, an honest $\mathcal{S}$ can generate a proof $\pi$ to convince $\mathcal{R}$ for every pair $(\vec{x}, \vec{w})$ $\in$ $R_C$. In the algorithm $\Pi.\mathsf{Setup}$, ${\mathsf{vk}}$ and $C$ are public, which means that anyone with $\vec{x}$ can verify a proof.
\subsection{Fair Exchange Based on zk-SNARK}\label{fair-exchange}
Fair exchange refers to the scenario where users exchange currency for digital commodities, e.g., digital assets and valuable information.
A fair exchange protocol was designed to guarantee that the exchange is executed in a fair way~\cite{dziembowski2018fairswap}.
zk-SNARK (see Sec.~\ref{zk-snark}) is one of the key technologies to realize fair exchange. It can help $\mathcal{S}$ protect the privacy of information content while proving its authenticity. Next, we give the definition of the circuit used for realizing fair exchange.
\newtheorem{mydef2}{Definition}
\begin{mydef}[The Circuit for Fair Exchange]\label{fair-exchange-circuit}
The whole process can be represented by a tuple of polynomial-time algorithms $\Upsilon$ $\overset{\text{def}}{=}$ ($\mathtt{DataAuth}$, $\mathtt{KeyAuth}$):
\begin{itemize}
\item $\mathtt{DataAuth}(m) \rightarrow h(m) $.
The algorithm takes the digital commodity (or plaintext) $\mathsf{m}$ as input, and computes the authenticator $h(m)$, which is the hash result of the digital commodity.
\item $ \mathtt{KeyAuth}(k, m) \rightarrow (\overline{m}, h(k))$.
The algorithm takes the encryption key $\mathsf{k}$ and the digital commodity $\mathsf{m}$ as inputs, and generates the encrypted digital commodity (or ciphertext) $\overline{m}$ as well as the hash result of the encryption key $h(k)$.
\end{itemize}
\end{mydef}
This circuit is illustrated in Fig.~\ref{Fig:fair-exchange-circuit}. $\mathcal{S}$ can use the circuit to generate a zero-knowledge proof based on the algorithm $\Pi.\mathtt{Prove}$, which proves the authenticity of the encrypted information and the encryption key.
\begin{figure}[htb]
\centering
\includegraphics[width=0.6\textwidth]{figure/fair-exchange-circuit.pdf}
\caption{The logic diagram of the fair exchange circuit. The parameters with gray background are private ones protected with zk-SNARK.}
\label{Fig:fair-exchange-circuit}
\vspace{-0.2cm}
\end{figure}
\subsection{Hashed Timelock Contracts}\label{htlc}
The Hashed Timelock Contract (HTLC) was first applied and implemented in Bitcoin's Lightning network~\cite{poon2016bitcoin}, which aims to ensure the atomicity of cross-chain asset exchanges.
HTLC requires that both sides of an interaction (e.g., $\mathcal{S}$ and $\mathcal{R}$) have accounts in each blockchain (i.e., accounts $\mathcal{S^\alpha}$ and $\mathcal{R^\alpha}$ in chain $\alpha$, and accounts $\mathcal{S^\beta}$ and $\mathcal{R^\beta}$ in chain $\beta$, for both $\mathcal{S}$ and $\mathcal{R}$).
The smart contracts\footnote{
Some blockchain systems such as Bitcoin do not support smart contracts~\cite{zou2019smart}. In such a case, HTLC is implemented with other mechanisms such as scripting~\cite{script}.
For convenience, we use smart contracts to represent true smart contracts as well as other techniques such as scripting when presenting HTLC in this study.} in blockchain $\alpha$ and $\beta$ are denoted by $\xi^\alpha$ and $\xi^\beta$, respectively.
\newtheorem{mydef4}{(Definition)}
\begin{mydef}[The Process of HTLC]
Assume that account $\mathcal{S^\alpha}$ and account $\mathcal{R^\beta}$ intend to exchange assets with each other. The whole interaction process in HTLC is divided into three steps: $\mathsf{Lock}$, $\mathsf{Update}$, and $\mathsf{Refund}$.
\begin{itemize}
\item $\mathsf{Lock:}$ First, $\mathcal{S^\alpha}$ selects a random 256-bit integer as the preimage pre and computes its hash value h(pre). Then in the smart contract $\xi^\alpha$, $\mathcal{S^\alpha}$ opens h(pre), employs h(pre) to lock its asset sent to $\mathcal{R^\alpha}$, and sets a timer $T_3$. Similarly, in the smart contract $\xi^\beta$, $\mathcal{R^\beta}$ locks its asset sent to $\mathcal{S^\beta}$ with the same h(pre) and sets a timer $T_4$, where $T_3>T_4$.
\item $\mathsf{Update:}$ $\mathcal{S^\beta}$ offers pre to $\xi^\beta$ within $T_4$ to unlock the asset sent by $\mathcal{R^\beta}$.
After $\mathcal{R^\beta}$ learns pre, $\mathcal{R^\alpha}$ provides pre to $\xi^\alpha$ within $T_3$ to unlock the asset sent by $\mathcal{S^\alpha}$.
\item $\mathsf{Refund:}$ If the time exceeds $T_4$ and $S^\beta$ does not provide pre, the locked asset in $\xi^\beta$ would be returned to $R^\beta$. In this case, since $\mathcal{R^\alpha}$ does not know pre (only $\mathcal{S}$ has pre), $\mathcal{R^\alpha}$ cannot provide pre within the specified time $T_3$ in blockchain $\alpha$. When $T_3$ times out, the locked asset would be returned to $S^\alpha$.
\end{itemize}
\end{mydef}
Note that the information in accounts $\mathcal{R^\alpha}$ ($\mathcal{S^\alpha}$) and $\mathcal{R^\beta}$ ($\mathcal{S^\beta}$) are shared because they both belong to the same entity $\mathcal{R}$ ($\mathcal{S}$). Therefore, when $\mathcal{R^\beta}$ learns $pre$ in blockchain $\beta$, $\mathcal{R^\alpha}$ can send $pre$ to the smart contract in blockchain $\alpha$.
Also note that $T_3>T_4$ is necessary in order to ensure atomicity.
Nevertheless, in asynchronous networks, we find that HTLC may not guarantee atomicity due to network delay. Therefore, we introduce an incentive mechanism to overcome this problem in Sec.~\ref{hierarchical-channel-scheme}.
\subsection{Pedersen’s Verifiable Secret Sharing}\label{vss}
Pedersen’s verifiable secret sharing scheme does not need any trusted third party, which enables $n$ participants to share a secret in a completely decentralized way~\cite{pedersen1991non}.
\newtheorem{mydef3}{(Definition)}
\begin{mydef}[Pedersen’s $(t,n)$-VSS]
Let $\mathbb{G}_q$ be a $q$-order subgroup of the prime $P$, with $g$ and $h$ being generators of $\mathbb{G}_q$. Let $s$ be the shared secret, $\mathcal{O}$ the owner of $s$, $n$ the number of participants, $t$ the threshold value, and $\mathsf{U_i}$ the $i$-th participant. Define Pedersen commitment as $E(a,b)=g^ah^b$. Then the whole process can be divided into three steps: $\mathsf{Share}$, $\mathsf{Verify}$ and $\mathsf{Recover}$.
\begin{itemize}
\item $\mathsf{Share}$: First, $\mathcal{O}$ selects a random number $r$, computes commitment $E(s,r)=g^sh^r$, and opens $E(s,r)$. Then, $\mathcal{O}$ selects $t-1$ random numbers $a_i, i\in[1,t-1]$, constructs a polynomial $f(x)=s + \sum_{i=1}^{t-1}a_ix^i$, and computes $s_i=f(i)$. Next, $\mathcal{O}$ selects another set of random numbers $b_i, i\in[1,t-1]$, calculates $E_{a_i}=g^{a_i}h^{b_i}$, and opens them. Finally, $\mathcal{O}$ constructs a polynomial $g(x)=r + \sum_{i=1}^{t-1}b_ix^i$, computes $r_i=g(i)$, and sends the $i$th secret share ($s_i,r_i$) to $\mathsf{U_i}$.
\item $\mathsf{Verify}$: When $\mathsf{U_i}$ receives ($s_i,r_i$), it computes $E(s_i,r_i)$ and $\prod_{j=0}^{t-1}E_j^{i^j}$, where $E_j^{i^j}=g^{a_ji^i}h^{b_ji^i}$. If the computed $E(s_i,r_i)$ and $\prod_{j=0}^{t-1}E_j^{i^j}$ are equal, the received $s_i$ is correct.
\item $\mathsf{Recover}$: When at least $t$ participants share the secret correctly and contribute their shares, the secret can be recovered by Lagrange polynomial interpolation, i.e. $s=\sum_{i=1}^ts_i\prod_{1 \leq j \leq t, j \neq i} \frac{i}{i-j} $.
\end{itemize}
\end{mydef}
The participants can verify the validity of the received shares in step $\mathsf{Verify}$, so as to detect the invalid messages sent by adversaries.
\section{The Cross-Channel}\label{sec:Cross-Channel}
In this section, we first provide an overview on $\mathsf{Cross}$-$\mathsf{Channel}$, then present the protocol in detail, and finally analyze its scalability, fairness and atomicity.
\subsection{Overview} \label{overview}
\begin{figure*}[htb]
\centering
\includegraphics[width=0.85\textwidth]{figure/overview.pdf}
\caption{The cross-chain currency exchange procedure with $\mathsf{Cross}$-$\mathsf{Channel}$. The whole process includes three protocols, i.e, the hierarchical interaction protocol, the general fair exchange protocol, and the hierarchical settlement protocol. The hierarchical interaction protocol contains a hierarchical channel structure that solves the UAC problem. The general fair exchange protocol solves the UE problem that occurs during the interaction in the (sub-)channel, especially for the encrypted information exchange scenario. The hierarchical settlement protocol adopts an improved HTLC protocol, overcoming the impact of high latency in asynchronous networks while ensuring the correctness of cross-chain settlements.}
\label{Fig:overview}
\end{figure*}
$\mathsf{Cross}$-$\mathsf{Channel}$ is an efficient channel scheme that supports complex services such as cross-chain.
For the sake of convenience, we use an example (currency exchange) to illustrate the general process of $\mathsf{Cross}$-$\mathsf{Channel}$ (shown in Fig.~\ref{Fig:overview}).
In this example, an entity $\mathcal{S}$ has an account $\mathcal{S^\alpha}$ in Bitcoin $\alpha$, and an entity $\mathcal{R}$ has an account $\mathcal{R^\beta}$ in Ethereum $\beta$.
$\mathcal{S}$ and $\mathcal{R}$
attempts to frequently exchange $\mathcal{S^\alpha}$'s Bitcoins (BTC) for $\mathcal{R^\beta}$'s Ether (ETH).
In order to achieve the above goal, $\mathcal{S}$ needs to creates an account $\mathcal{S}^\beta$ in $\beta$ to get $\mathcal{R^\beta}$'s ETH, and $\mathcal{R}$ also needs to create an account $\mathcal{R}^\alpha$ in $\alpha$ to get $\mathcal{S^\alpha}$'s BTC. The whole process can be summarized as follows.
First, $\mathcal{S}$ and $\mathcal{R}$ need to establish channels in $\alpha$ and $\beta$. Specifically, two accounts in the same blockchain ($\mathcal{S}^\alpha$ with $\mathcal{R}^\alpha$, or $\mathcal{S}^\beta$ with $\mathcal{R}^\beta$) execute the hierarchical interaction protocol $\Psi$ to establish a channel, then send currency in this channel. The channel can be a hierarchical one with multiple sub-channels in order to spend the unsettled amounts (e.g., accounts $D$ and $E$ in Fig.~\ref{Fig:overview}) based on $\Psi$. Currency exchanges within the channel follow the general fair exchange protocol $\Theta$, which also supports fair exchange (i.e., exchange currency with encrypted information) and encrypted information exchange.
Finally, when the channel needs to be closed, all involved accounts in the channel, including those for the sub-channels, upload their final states to the corresponding smart contracts, and the hierarchical settlement protocol $\Phi$ is executed to complete the settlement. Note that, we adopt an improved HTLC in $\Phi$ to ensure the correctness and atomicity of the cross-chain settlement.
In the following two subsections, we detail the hierarchical interaction protocol with settlement (Sec.~\ref{hierarchical-channel-scheme}) and the general fair exchange protocol (Sec.~\ref{general-fair-exchange}).
\subsection{Hierarchical Channel Design} \label{hierarchical-channel-scheme}
\begin{figure}[htb]
\centering
\includegraphics[width=0.8\textwidth]{figure/hibe-channel2.pdf}
\caption{An illustration example of a 3-level hierarchical channel.}
\label{Fig:hibe-channel}
\end{figure}
We present a hierarchical interaction protocol $\Psi$ with a settlement protocol $\Phi$ to solve the Unsettled Amount Congestion (UAC) problem.
The settlement protocol $\Phi$ can be written into smart contracts and verified by miners $\mathcal{M}$.
\noindent{\bf Hierarchical Interaction Protocol $\Psi$.} Fig.~\ref{Fig:hibe-channel} illustrates an example hierarchical channel. One can see that the whole structure has three levels, which are marked as $\mathsf{Level}$ 0--2.
At $\mathsf{Level}$ 0, $\mathcal{S}$ and $\mathcal{R}$ send $\mathsf{Tx}_{\text{Open}}$ to the smart contract $\xi$ to establish a channel $\Omega_0$ with initial state ($\mathsf{v_\mathcal{S}, v_\mathcal{R}}$), which can be denoted as $[\mathcal{S} \mapsto \mathsf{v_S}$, $\mathcal{R} \mapsto \mathsf{v_R}]_{\Omega_0}$, meaning that $\mathcal{S}$ has $\mathsf{v_\mathcal{S}}$ in $\Omega_0$, $\mathcal{R}$ has $\mathsf{v_\mathcal{R}}$ in $\Omega_0$, and the state of $\Omega_0$ is ($\mathsf{v_\mathcal{S}, v_\mathcal{R}}$).
\begin{equation}\notag
\mathsf{Tx}_{\text{Open}} \overset{\text{def}}{=} (\mathsf{From: \mathcal{S / R}; To: \xi; v_\mathcal{S}/v_\mathcal{R}}).
\end{equation}
For the sake of convenience, $\mathsf{v_S}$ and $\mathsf{v_R}$ can be the coins deposited by $\mathcal{S}$ and $\mathcal{R}$ (for some $\mathsf{v_S}$, $\mathsf{v_R}$ $\in$ $\mathbb{R}$ $\geq$ 0).
In $\Omega_0$, $\mathcal{S}$ and $\mathcal{R}$ can send transaction receipts \{$Tr$\} to each other, such as $Tr_{01\text{--}06}$ in Fig.~\ref{Fig:hibe-channel}. Let $Tr \overset{\text{def}}{=} (\mathsf{snd},\mathsf{rcv}, \mathsf{v}, \xi)_{\mathsf{\sigma_{snd}}}$, meaning that $\mathsf{snd}$ transfers amount $\mathsf{v}$ to $\mathsf{rcv}$ via smart contract $\xi$, where $\mathsf{\sigma_{snd}}$ is the message signature.
Note that $\mathsf{snd}$ and $\mathsf{rcv}$ can be omitted if clear from context.
To spend an unsettled amount in $\Omega_0$, the two parties of $\Omega_0$ can negotiate to open a sub-channel. For example, $\mathcal{S}$ can send a request to $\mathcal{R}$ asking for the permission to open a sub-channel $\Omega_1$ with $\mathcal{D}$ to spend the unsettled amount in $Tr_{06}$.
If $\mathcal{R}$ permits,
it would generate a sub-channel receipt $Sr_{06}$ and send it to $\mathcal{S}$, where $Sr_{06} \overset{\text{def}}{=} (\mathcal{D}, Tr_{06})_{\mathsf{Sig_{\mathcal{R}}}}$, with $\mathsf{Sig_{\mathcal{R}}}$ being $\mathcal{R}$'s signature for $Sr_{06}$, and $\mathcal{D}$ the address of the participant with which $\mathcal{S}$ would construct a sub-channel to spend the unsettled amount in $Tr_{06}$.
Then, $\mathcal{S}$ sends $Sr_{06}$ to $\mathcal{D}$, who needs to verify the legitimacy of $Sr_{06}$ based on $\mathsf{Sig_{\mathcal{R}}}$ and ensure the correctness of $\mathsf{v_{06}}$ according to the $\mathsf{\sigma_{\mathcal{R}}}$ carried by $Tr_{06}$.
When the verification is successful, $\mathcal{S}$ takes $\mathsf{v_{06}}$ as its initial balance to open sub-channel $\Omega_1$ with $\mathcal{D}$, i.e., $[\mathcal{S} \mapsto \mathsf{v_{06}}, \mathcal{D} \mapsto 0]_{\Omega_1}$. Note that $\Omega_1$ is a sub-channel that is constructed particularly for the spending of the unsettled amount in $Tr_{06}$ -- no other transactions between $\mathcal{S}$ and $\mathcal{D}$ are allowed. Following the same procedure, the parties ($\mathcal{S}$ and $\mathcal{R}$) in $\mathsf{Level}$ 0 can choose another $Tr$ to create another new sub-channel and the parties ($\mathcal{S}$ and $\mathcal{D}$) in $\mathsf{Level}$ 1 can also generate sub-channels.
For example, as shown in Fig.~\ref{Fig:hibe-channel}, $\mathcal{D}$ establishes a sub-channel $\Omega_2$ with user $\mathcal{Q}$ based on $Tr_{12}$. It is worthy of noting that all operations related to a sub-channel are off-chain, which means that smart contract is not involved thus conserving blockchain resources.
\noindent{\bf Hierarchical settlement protocol $\Phi$.} We design a new protocol $\Phi$
and implement it in smart contract $\xi$ to support settlement.
Not that, $\Phi$ can be adopted for both intra-chain and cross-chain channel settlement, which differ slightly.
We use the same example shown in Fig.~\ref{Fig:hibe-channel} to demonstrate the procedure for intra-chain settlement.
First, $\mathcal{S}$ and $\mathcal{R}$ send requests to $\xi$ to close the hierarchical channel. After receiving the channel closing requests, $\xi$ sets a timer $T_2$
and broadcasts this closing event to all blockchain participants and miners.
This message also requires the users ($\mathcal{S}$, $\mathcal{R}$, $\mathcal{D}$ and $\mathcal{Q}$) to compute their final states \{$f$\} based on the corresponding related receipts ($Tr$ and $Sr$).
For instance, $\mathcal{S}$ and $\mathcal{R}$ need to compute their final states based on $Tr_{01\text{--}06}$, $Sr_{06}$, $Tr_{11\text{--}13}$, and $Sr_{12}$.
Then, each participant packages its $f$ and the related receipts into a $\mathsf{Tx}_{\text{Close}}$ message and sends $\mathsf{Tx}_{\text{Close}}$ to $\xi$ within $T_2$.
\begin{equation}\notag
\mathsf{Tx}_{\text{Close}} \overset{\text{def}}{=} (\mathsf{From: Snd; To: \xi;}\; f, \{Sr\}).
\end{equation}
When $T_2$ times out, $\xi$ verifies the uploaded data from $\mathsf{Level}$ 0, i.e., the correctness of the signatures and account balances. Note that $\mathsf{Tx}_{\text{Close}}$ contains all sub-channel receipts agreed by $\mathsf{Snd}$, which are used by $\xi$ to check the correctness of account balances. If the verification is successful, $\xi$ would verify the data from the next sub-level, i.e., $\mathsf{Level}$ 1. If the verification succeeds, $\xi$ continues to verify the next higher level. If $\xi$ fails at any level, all sub-channels in that level and above would automatically fail. Such a failure drives $\xi$ to adjust the final state of each participant in all successful levels based on the received $\{Sr\}$'s. Finally, the miners update the corresponding on-chain states according to the settlement results.
To support cross-chain settlement, the hierarchical settlement protocol $\Phi$ adopts the HTLC protocol (introduced in Sec.~\ref{htlc}) to ensure the atomicity of the interaction.
As shown in Fig.~\ref{Fig:overview}, $\mathcal{S}$ and $\mathcal{R}$ need to have accounts in both blockchain $\alpha$ and $\beta$ for cross-chain operations. Based on the hierarchical interaction protocol and the intra-chain hierarchical settlement protocol mentioned above, accounts in the same chain ($\mathcal{S^\alpha}$ and $\mathcal{R^\alpha}$, $\mathcal{S^\beta}$ and $\mathcal{R^\beta}$) establish channels, send $Tr$, create $Sr$ to establish sub-channels, and upload the final state of each (sub-)channel when the channel needs to be closed.
Unlike the intra-chain settlement, which updates the on-chain states immediately, for cross-chain, when $\xi$ determines that the final states are valid, it runs the HTLC protocol to make $\mathcal{S^\alpha}$ and $\mathcal{R^\beta}$ lock their final states based on the step HTLC.$\mathsf{Lock}$, then $\mathcal{S}$ and $\mathcal{R}$ complete the final settlement according to the step HTLC.$\mathsf{Update}$ or HTLC.$\mathsf{Refund}$.
Particularly, in the step HTLC.$\mathsf{Refund}$ shown in Sec.~\ref{htlc}, if no one submits $\mathsf{pre}$ before $T_3$ or $T_4$ times out, the smart contract would not update the states. However, in an asynchronous network, affected by the high latency, after $\mathcal{S^\beta}$ provides $\mathsf{pre}$ to update the states in blockchain $\beta$, $\mathcal{R^\alpha}$ may fail to upload $\mathsf{pre}$ within $T_3$, which breaks the atomicity of HTLC.
Therefore, in order to make HTLC suitable for asynchronous networks, we set a timer $T_5$ in $\xi^\alpha$, during which any miner can help $\mathcal{R^\alpha}$ provide $\mathsf{pre}$ to get rewards from $\mathcal{R^\alpha}$.
Note that in more complex scenarios such as the encrypted information exchange, in addition to realizing the settlement mentioned above, the hierarchical settlement protocol needs to further accomplish the fair exchange of keys, which are detailed in Sec.~\ref{cross-channel and applications}.
\subsection{General Fair Exchange Protocol} \label{general-fair-exchange}
\begin{figure}[htb]
\centering
\includegraphics[width=\textwidth]{figure/general-fair-exchange.pdf}
\caption{ The sequence diagram of the general fair exchange protocol. Steps 1.1--1.9 demonstrate the $\Theta.$Share process, while steps 2.1--2.3 illustrating the $\Theta.$Exchange process and steps 3.1--3.7 standing for the $\Theta.$Recover process.
}
\label{Fig:fair-exchange}
\end{figure}
In this subsection, we propose a general fair exchange protocol $\Theta$ to solve the Unfair Exchange (UE) problem, which can guarantee the fairness of encrypted information exchange (EIE).
The whole protocol
involves four steps: {\bf Setup}, {\bf Share}, {\bf Exchange}, and {\bf Recover}, which are demonstrated in Fig.~\ref{Fig:fair-exchange}.
\noindent{\bf $\Theta.$Setup.} First, sender $\mathcal{S}$ builds a circuit $C_\Theta$ based on Fig.~\ref{Fig:fair-exchange-circuit2}.
Compared with the traditional fair exchange circuit (shown in Fig.~\ref{fair-exchange-circuit}), $C_\Theta$ appends $n$ key shares as private inputs and adds a key recover function (details shown in Sec.~\ref{vss}.$\mathsf{Recover}$) to recover the encryption key $k$.
The reason for designing this circuit lies in that, in $\Theta$, we adopt the $(t,n)$-VSS protocol (shown in Sec.\ref{vss}) to divide the encryption key into $n$ key shares. However, the input of traditional circuit (shown in Fig.~\ref{fair-exchange-circuit}) is the key itself, which cannot prove the correctness of the key shares.
Thus we propose a new circuit $C_\Theta$, which can prove the correctness of not only the key but also the key shares without exposing any key-related information.
Besides $C_\Theta$, $\mathcal{S}$ needs to generate a security parameter $1^{\lambda}$, and takes $C_\Theta$ and $1^{\lambda}$ as inputs of $\Pi.\mathsf{Setup}$ to construct the common reference string ($\mathsf{pk, vk}$).
\begin{figure}[htb]
\centering
\includegraphics[width=\textwidth]{figure/fair-exchange-circuit2.pdf}
\caption{The logic diagram of the circuit used for the general fair exchange protocol $\Theta$. The parameters with gray background are private ones protected with zk-SNARK.}
\label{Fig:fair-exchange-circuit2}
\vspace{-0.2cm}
\end{figure}
\noindent{\bf $\Theta.$Share.} This process is marked Step 1.1 to Step 1.9 in Fig.~\ref{Fig:fair-exchange}. First, $\mathcal{S}$ generates an encryption key $\mathsf{k_\mathcal{S}}$ and divides it into $\mathsf{n}$ shares based on $\mathsf{(t,n)\text{-}VSS.Share}$.
Then, $\mathcal{S}$ hashes $\mathsf{k_\mathcal{S}}$ and each individual key share $\mathsf{k_{\mathcal{S}i}(i \in [1,n])}$, packages them as well as the number of key shares $n$ and the key threshold $t$, into transaction $\mathsf{Tx}_{\text{Upload}}$.
Next, $\mathcal{S}$ sends $\mathsf{Tx}_{\text{Upload}}$ to the smart contract $\xi$.
Following the same procedure the receiver $\mathcal{R}$ packages $\mathsf{h(\mathsf{k_\mathcal{R}}})$, $n$, $t$, and $\mathsf{h({k_{\mathcal{R}{[1:n]}}})}$ into its $\mathsf{Tx}_{\text{Upload}}$ and sends it to $\xi$.
\begin{equation}\notag
\mathsf{Tx}_{\text{Upload}} \overset{\text{def}}{=} (\mathsf{From: \mathcal{S / R}; To: \xi; \; h(\mathsf{k_\mathcal{S / R}}), n ,t, \mathsf{h(k_{\mathcal{S / R}[1:n]}})}).
\end{equation}
When $\xi$ receives $\mathsf{Tx}_{\text{Upload}}$ from both parties, it would randomly select $\mathsf{n}$ miner addresses with a serial number $\mathsf{sn}$, bind each miner address with the hashes of two unique key shares (one from $\mathcal{S}$ and one from $\mathcal{R}$), and open these bindings.
Then, $\mathcal{S}$ and $\mathcal{R}$ sign their key shares with $\mathsf{sn}$, and distribute them to the selected miners based on the bindings.
After receiving the key shares, each miner verifies their legitimacy based on $\mathsf{(t,n)\text{-}VSS.Verify}$. Besides that, each miner recomputes the hashes of the received key shares and compares them with those in $\xi$ to detect possible errors.
If a miner detects a fake key share, it would report it to $\xi$ within $T_1$. Specifically, the miner packages the detected fake key share as well as the signature of the key share message with $\mathsf{sn}$ in $\mathsf{Tx}_{\text{Appeal}}$, then sends it to $\xi$.
When $\xi$ gets $\mathsf{Tx}_{\text{Appeal}}$, it verifies the signature of the key share owner with $\mathsf{sn}$, and recalculates the hash of the reported fake key share to check if the data in $\mathsf{Tx}_{\text{Appeal}}$ is legitimate.
If the verification succeeds, $\xi$ terminates the fair exchange protocol. Note that, $\mathsf{sn}$ can effectively prevent malicious behaviors of adversaries from destroying the execution of the protocol by providing the previous key share in $\mathsf{Tx}_{\text{Appeal}}$.
\begin{equation}\notag
\mathsf{Tx}_{\text{Appeal}} \overset{\text{def}}{=} (\mathsf{From: \mathcal{M}_i; To: \xi; Sig_{snd}, k'_{i}, sn), i \in [1,n]}.
\end{equation}
\noindent{\bf $\Theta.$Exchange.}
In this step,
$\mathcal{S}$ first computes the ciphertext $\overline{\mathsf{m}}_\mathcal{S}$ of the exchange object $\mathsf{m_\mathcal{S}}$ based on $\mathsf{k_\mathcal{S}}$. This can be done by some common encryption technologies, e.g., Elliptic Curve Cryptography (ECC)~\cite{liu2008tinyecc} and MIMC~\cite{albrecht2016mimc}.
Then, $\mathcal{S}$ takes $\mathsf{pk}$, $\mathsf{k_\mathcal{S}}$ and $\mathsf{m_\mathcal{S}}$ to generate $\pi_\mathcal{S}$ based on the algorithm $\Pi.\mathsf{Prove}$, and sends ($\pi_\mathcal{S}$, $\overline{\mathsf{m}}_\mathcal{S}$, $\mathsf{h({m_\mathcal{S}})}$, $\mathsf{h({k_\mathcal{S}})}$) to $\mathcal{R}$.
Similarly, $\mathcal{R}$ sends ($\pi_\mathcal{R}$, $\mathsf{\overline{m}_\mathcal{R}}$, $\mathsf{h({m_\mathcal{R}})}$, $\mathsf{h({k_\mathcal{R}})}$) to $\mathcal{S}$.
After that, $\mathcal{S}$ and $\mathcal{R}$ use $\mathsf{vk}$ and the received data
to respectively verify ($\pi_\mathcal{S}$, $\pi_\mathcal{R}$) based on the algorithm $\Pi.\mathsf{Verify}$.
The above process is marked Step 2.1 and Step 2.3 in Fig.~\ref{Fig:fair-exchange}.
\noindent{\bf $\Theta.$Recover.} If both parties verify successfully, $\mathcal{S}$ and $\mathcal{R}$ send requests to the smart contract $\xi$ for key recovery (marked Step 3.1 to Step 3.7 in Fig.~\ref{Fig:fair-exchange}).
When $\xi$ gets the requests from $\mathcal{S}$ and $\mathcal{R}$, it broadcasts this event to the selected miners, who then send their stored key shares to $\xi$ via $\mathsf{Tx}_{\text{Recover}}$.
\begin{equation}\notag
\mathsf{Tx}_{\text{Recover}} \overset{\text{def}}{=} (\mathsf{From: \mathcal{M}_i; To: \xi; k_{\mathcal{S}i}, k_{\mathcal{R}i}), i\in[1,n]}.
\end{equation}
$\xi$ verifies the legitimacy of each key share by comparing its hash result and checking the address of its sender. When the number of valid key shares is greater than the key threshold $t$, $\xi$ sends the collected key shares to the requester, who then employs $\mathsf{(t,n)\text{-}VSS.Recover}$ to recover the key, and further decrypts the message.
\subsection{Cross-Channel and Applications} \label{cross-channel and applications}
Based on the hierarchical channel and the general fair exchange protocol presented in the previous two subsections, we present $\mathsf{Cross}$-$\mathsf{Channel}$ to support various cross-chain services, e.g., currency exchange (CE), fair exchange (FE, i.e., exchange currency with encrypted information), and encrypted information exchange (EIE)), in this subsection. We adopt the same notations as before: ($\mathcal{S^\alpha}$, $\mathcal{R^\alpha}$, $\mathcal{M^\alpha}$, $\xi^\alpha$) and ($\mathcal{S^\beta}$, $\mathcal{R^\beta}$, $\mathcal{M^\beta}$, $\xi^\beta$).
Assume that $\mathcal{S}$ and $\mathcal{R}$ negotiate to exchange information ($\mathsf{m_{\mathcal{S}_i}, i \in \mathbb{Z^+}}$) in $\mathcal{S^\alpha}$ on blockchain $\alpha$ with information ($\mathsf{m_{\mathcal{R}_i}, i \in \mathbb{Z^+}}$) in $\mathcal{R^\beta}$ on blockchain $\beta$.
The whole scheme can be divided into four phases: {\bf Initialize}, {\bf Open}, {\bf Exchange}, and {\bf Close}, which are detailed in the following according to different application scenarios. Note that the first three phases are performed at each single chain while the last phase realizes the cross-chain operations via the cross-chain settlement protocol presented in Sec.~\ref{hierarchical-channel-scheme}. For better elaboration, we employ [ALL $\pmb{\Rightarrow}$] or [$\{\cdot\} \pmb{\Rightarrow}$] to denote that all or some of the three scenarios (CE, FE, EIE) need to execute the process that follows.
\noindent{\bf Initialize.} [ALL $\pmb{\Rightarrow}$] In $\alpha$ and $\beta$, each account is initialized with a unique address and a key pair ($\widetilde{\mathsf{pk}}$, $\widetilde{\mathsf{sk}}$).
[(FE, EIE) $\pmb{\Rightarrow}$] Each digital commodity owner generates a common reference string $\mathsf{(pk, vk)}$ based on
$\mathsf{\Theta.Setup}$ (introduced in Sec.~\ref{general-fair-exchange}).
\noindent{\bf Open.} [ALL $\pmb{\Rightarrow}$] According to the hierarchical interaction protocol $\Psi$, ($\mathcal{S^\alpha}$, $\mathcal{R^\alpha}$) and ($\mathcal{S^\beta}$, $\mathcal{R^\beta}$) respectively send $\mathsf{Tx_{open}}$ messages to call smart contracts $\xi^\alpha$ and $\xi^\beta$ to build channels $\Omega^\alpha_{0}$ and $\Omega^\beta_{0}$, and deposit their initial states, e.g., coins, into the channels. Note that the $\mathsf{Level}$ 0 of $\mathsf{Cross}$-$\mathsf{Channel}$ includes $\Omega^\alpha_0$ and $\Omega^\beta_0$, and the initial state is recorded as $[\mathcal{S}^\alpha \mapsto \mathsf{v}^\alpha_\mathcal{S}$, $\mathcal{R}^\alpha \mapsto \mathsf{v}^\alpha_\mathcal{R}$, $\mathcal{S}^\beta \mapsto \mathsf{v}^\beta_\mathcal{S}$, $\mathcal{R}^\beta \mapsto \mathsf{v}^\beta_\mathcal{R}]_{\Omega_0}$.
[EIE $\pmb{\Rightarrow}$] $\mathcal{S^\alpha}$ and $\mathcal{R^\beta}$ execute $\mathsf{\Theta.Share}$ to distribute their key shares
\noindent{\bf Exchange.} [CE $\pmb{\Rightarrow}$] The channel $\Omega_0$ allows two parties to instantaneously send payments between each other.
[(FE,EIE) $\pmb{\Rightarrow}$] The sender implements $\mathsf{\Theta.Exchange}$ to encrypt the exchanged information, generates the zero-knowledge proof, and verifies the proof sent by the receiver.
For example, $\mathcal{S^\alpha}$ can generate multiple encrypted information $\mathsf{\overline{m}_{\mathcal{S}_i}}$ based on $\mathsf{m_{\mathcal{S}_i}}$ and $\mathsf{k_\mathcal{S}, i \in \mathbb{Z^+}}$, and send them in $\Omega^\alpha_0$ to $\mathcal{R^\alpha}$.
Then, $\mathcal{S^\alpha}$ generates zero-knowledge proofs $\pi_{\mathcal{S}_i}$ to prove the authenticity of $\mathsf{k_\mathcal{S}}$ and $\mathsf{\overline{m}_{\mathcal{S}_i}}$ without exposing $\mathsf{k_\mathcal{S}}$ and $\mathsf{m_{\mathcal{S}_i}}$.
Next, it sends $\pi_{\mathcal{S}_i}$ and the public parameters shown in Fig.~\ref{Fig:fair-exchange-circuit2} to $\mathcal{R^\alpha}$.
$\mathcal{R^\alpha}$ can use $\mathsf{vk}$ and the received public parameters to verify $\pi_{\mathcal{S}_i}$.
Furthermore, with the consent of both parties in $\Omega_0$, one party can generate a sub-channel receipt $Sr$ to open a sub-channel and spend the unsettled amount based on the hierarchical interaction protocol $\Psi$ (details shown in Sec.~\ref{hierarchical-channel-scheme}).
\noindent{\bf Close.} [ALL $\pmb{\Rightarrow}$] Based on the hierarchical settlement protocol $\Phi$,
all participants in the hierarchical channel are required to upload their final states based on $\mathsf{Tx}_{\text{Upload}}$ within $T_2$.
[(CE) $\pmb{\Rightarrow}$] $\mathcal{S^\alpha}$ generates a preimage $\mathsf{pre}$ (a random 256-bits integer) and packages its hash result $\mathsf{h(pre)}$ in $\mathsf{Tx}_{\text{Lock}}$ to $\xi^\alpha$. %
\begin{equation}\notag
\mathsf{Tx}_{\text{Lock}} \overset{\text{def}}{=} (\mathsf{From: \mathcal{S / R}; To: \xi; h(pre)}).
\end{equation}
$\xi^\alpha$ opens $\mathsf{h(pre)}$ and uses $\mathsf{h(pre)}$ to lock the state of $\mathcal{S^\alpha}$ in the channel $\Omega^\alpha_0$. When $\mathcal{R^\alpha}$ learns $\mathsf{h(pre)}$ in the blockchain $\alpha$, $\mathcal{R^\beta}$ sends $\mathsf{Tx}_{\text{Lock}}$ to lock the state of $\Omega^\beta_0$.
We set timers $T_3$ and $T_4$ in blockchain $\alpha$ and $\beta$, respectively.
One needs to provide $\mathsf{pre}$ within $T_3$ and $T_4$ to update the state of $\alpha$ and $\beta$, respectively.
For example, $\mathcal{S^\beta}$ provides $\mathsf{pre}$ based on $\mathsf{Tx}_{\text{Update}}$ to update the states in blockchain $\beta$. Once $\mathsf{pre}$ is successfully verified by $\mathcal{\xi^\beta}$, $\mathcal{R^\beta}$ can learn $\mathsf{pre}$. Then, $\mathcal{R^\alpha}$ packages $\mathsf{pre}$ in $\mathsf{Tx}_{\text{Update}}$ to update the states of $\alpha$.
Note that, according to the HTLC protocol, $T_4$ should be less than $T_3$, which effectively guarantees the atomicity of the interaction process (the related discussion is shown in Sec.~\ref{discussions}).
Moreover, considering the high latency of asynchronous networks, we set a time threshold $T_5$. When miners observe that the $T_3$ times out and no one upload $\mathsf{pre}$ in $\alpha$, they can offer $\mathsf{pre}$ within $T_5$ to get rewards.
[(FE) $\pmb{\Rightarrow}$] Compared with the process in CE, $\mathcal{S^\alpha}$ needs to use $\mathsf{k_\mathcal{S}}$ as $\mathsf{pre}$ rather than regenerate a random 256-bit integer.
\begin{equation}\notag
\mathsf{Tx}_{\text{Update}} \overset{\text{def}}{=} (\mathsf{From: \mathcal{S / R / M}; To: \xi; pre}).
\end{equation}
[EIE $\pmb{\Rightarrow}$] In addition to $\mathsf{pre}$, $\mathcal{S^\alpha}$ and $\mathcal{R^\beta}$ also need to package the hash result of the key that needs to be recovered in $\mathsf{Tx}_{\text{Update-EIE}}$.
According to $\mathsf{\Theta.Recover}$, $\mathcal{S^\alpha}$ and $\mathcal{R^\beta}$ would get at least $t$ key shares from the smart contract, and they can execute $\mathsf{\Theta.Recover}$ to get each other's keys fairly ($\mathcal{S^\alpha}$ $\leftarrow$ $k_\mathcal{R}$, $\mathcal{R^\alpha}$ $\leftarrow$ $k_\mathcal{S}$).
\begin{equation}\notag
\mathsf{Tx}_{\text{Update-EIE}} \overset{\text{def}}{=} (\mathsf{From: \mathcal{S / R / M}; To: \xi; pre, h(k)}).
\end{equation}
For convenience, we summarize the logic of the smart contract $\xi^\alpha$ in Fig.~\ref{fig:contract}. Note that the smart contract $\xi^\beta$ is the same as $\xi^\alpha$ except that the timer $T_3$ is replaced with $T_4$. The entire protocol is outlined in Fig.~\ref{fig:Cross-Protocol}.
\begin{figure}[htb] \footnotesize
\caption{The $\mathsf{Cross}$-$\mathsf{Channel}$ Smart Contract $\xi^\alpha$. Suppose $\mathcal{S}$ and $\mathcal{R}$ are participants in the Level 0. Variables with brace represent a set (e.g., $Sr$ represents a sub-channel receipt, and $\{Sr\}$ is a set of multiple sub-channel receipts).
}\label{fig:contract}
\begin{framed}
\centering \textbf{$\mathsf{Cross}$-$\mathsf{Channel}$ Contract $\xi$}\vspace{10pt}\\
\begin{basedescript}{\desclabelwidth{30pt}}
\item[\hspace{7pt}Initialize:]$\pmb{\rightarrow}$ Set $\mathsf{state} := \mathtt{INIT}$. \\
\vspace{5pt}
\item[\hspace{7pt}Open:]$\pmb{\rightarrow}$ Upon receiving (``Open", $\mathsf{v}_\mathcal{S}, \mathsf{v}_\mathcal{R}$) from $\mathcal{S}$ and $ \mathcal{R}$:\\
{
\setlength{\leftskip}{17pt}Assert $\mathsf{state} = \mathtt{INIT}$;\\
\setlength{\leftskip}{17pt}Assert $\mathsf{v}_\mathcal{S} \leq \mathcal{S}'s$ balance and $\mathsf{v}_\mathcal{R} \leq \mathcal{R}'s$ balance;\\
\setlength{\leftskip}{17pt}Open a channel with states $(\mathsf{v}_\mathcal{S}, \mathsf{v}_\mathcal{R})$; \\
\setlength{\leftskip}{17pt}Set $\mathsf{state} := \mathtt{Open_{CE}}$. \\
\vspace{2.5pt}
\setlength{\leftskip}{0pt}$\pmb{\rightarrow}$ Upon receiving (``Upload", $\mathsf{h(k}), n, t, \mathsf{h(k_{[1:n]}})$: \\
\setlength{\leftskip}{17pt}Assert $\mathsf{state} = \mathtt{Open_{CE}}$;\\
\setlength{\leftskip}{17pt}Randomly select $n$ addresses of miners $\mathsf{\mathcal{M}_{[1:n]}}$;\\
\setlength{\leftskip}{17pt}Randomly select a serious number $\mathsf{sn}$;\\
\setlength{\leftskip}{17pt}Bind each $\mathsf{\mathcal{M}_{i}}$ with $\mathsf{h_{k_{i}}, i\in[1,n]}$; \\
\setlength{\leftskip}{17pt}Set a timer $T_1$; \\
\setlength{\leftskip}{17pt}{\bf While} current $T \le T_1$: \\
\setlength{\leftskip}{25pt}Collect (``Appeal", $\mathsf{Sig_{snd}}$, $\mathsf{{k_{i}}}$, $\mathsf{sn}$) from $\mathsf{\mathcal{M}_{i}, i \in [1,n]}$; \\
\setlength{\leftskip}{25pt}Require (``Appeal", $\mathsf{Sig_{snd}}$, $\mathsf{{k_{i}}}$, $\mathsf{sn}$) is illegal; \\
\setlength{\leftskip}{17pt}{\bf End while};\\
\setlength{\leftskip}{17pt}Set $\mathsf{state} := \mathtt{Open}$.\\
}
\vspace{5pt}
\item[\hspace{7pt}Close:]$\pmb{\rightarrow}$ Upon receiving (``Close", $f$, \{$Sr$\}) from $\mathcal{S}$ and $\mathcal{R}$:\\
{
\setlength{\leftskip}{17pt}Assert $\mathsf{state} = \mathtt{Open_{CE}}$ or $\mathtt{Open}$; \\
\setlength{\leftskip}{17pt}Set a timer $T_2$; \\
\setlength{\leftskip}{17pt}{\bf While} current $T \le T_2$: \\
\setlength{\leftskip}{25pt}Collect (``Close", $f$, \{$Sr$\}) from other participants in the channel; \\
\setlength{\leftskip}{17pt}{\bf End while};\\
\setlength{\leftskip}{17pt}Verify and ensure final states based on the protocol $\Phi$; \\
\setlength{\leftskip}{17pt}Set $\mathsf{state} := \mathtt{Close}$.\\
\vspace{2.5pt}
\setlength{\leftskip}{0pt}$\pmb{\rightarrow}$ Upon receiving (``Lock", $\mathsf{h(pre)}$)
from $\mathcal{S}$ and $\mathcal{R}$:\\
\setlength{\leftskip}{17pt}Assert $\mathsf{state} = \mathtt{Close}$; \\
\setlength{\leftskip}{17pt}Locks final states based on $\mathsf{h(pre)}$; \\
\setlength{\leftskip}{17pt}Set timer $T_{3}$, $T_5$; \\
\setlength{\leftskip}{17pt}Set $\mathsf{state} := \mathtt{Lock}$. \\
\vspace{2.5pt}
\setlength{\leftskip}{0pt}$\pmb{\rightarrow}$ Upon receiving (``Update", $\mathsf{pre'}$) or (``Update-EIE", $\mathsf{pre'}$, $\mathsf{h'(k)}$):\\
\setlength{\leftskip}{17pt}Assert $\mathsf{state} = \mathtt{Lock}$; \\
\setlength{\leftskip}{17pt}$\left \langle \mathcal{S}/ \mathcal{R} \right \rangle$ Assert current $T \le T_{3}$; \\
\setlength{\leftskip}{17pt}$\left \langle \mathcal{M} \right \rangle$ Assert current $T$: $T_{3} < T \le T_5$; \\
\setlength{\leftskip}{17pt}$\pmb{\rightarrow}$(``Update"): \\
\setlength{\leftskip}{17pt}{\bf If} (hash($\mathsf{pre'}$) == $\mathsf{h({pre})}$) {\bf then} \\
\setlength{\leftskip}{25pt}Update on-chain final states;\\
\setlength{\leftskip}{25pt}Set $\mathsf{state} := \mathtt{Success}$; \\
\setlength{\leftskip}{17pt}{\bf End if}. \\
\setlength{\leftskip}{17pt}$\pmb{\rightarrow}$(``Update-EIE"): \\
\setlength{\leftskip}{17pt}{\bf If} (hash($\mathsf{pre'}$) == $\mathsf{h({pre})}$ and $\mathsf{h'(k}) == \mathsf{h(k)}$) {\bf then} \\
\setlength{\leftskip}{25pt}Send key recover request to $\mathcal{M}_{[1:n]}$;\\
\setlength{\leftskip}{25pt}Collect key shares $\mathsf{k_{[1:t]}}$ from miners (``Recover", $\mathsf{k_{i}}$, $\mathsf{i\in[1,n]}$); \\
\setlength{\leftskip}{25pt}Update on-chain final states;\\
\setlength{\leftskip}{25pt}Set $\mathsf{state} := \mathtt{Success}$; \\
\setlength{\leftskip}{17pt}{\bf End if}. \\
\vspace{2.5pt}
}
\end{basedescript}
\end{framed}\vspace{-10pt}
\end{figure}
\begin{figure*}[htb] \footnotesiz
\caption{An example of the encrypted information exchange (EIE) based on $\mathsf{Cross}$-$\mathsf{Channel}$. Variables with brace represent a set (e.g., $Tr$ represents a transaction receipt, and $\{Tr\}$ is a set of multiple transaction receipts).} \label{fig:Cross-Protocol}
\begin{framed}
\begin{multicols}{2}
\centering \textbf{$\mathsf{Cross}$-$\mathsf{Channel}$ Protocol Details} \vspace{5pt}
\begin{basedescript}{\desclabelwidth{30pt}}
\item[{\bf The process in the blockchain $\alpha$ ($\mathcal{S^\alpha}$, $\mathcal{R^\alpha}$, $\mathcal{M^\alpha}$):}]
\item[\hspace{7pt}Initialize:]~\\
{
\setlength{\leftskip}{-16pt}All get ($\widetilde{\mathsf{pk}}$, $\widetilde{\mathsf{sk}}$) from blockchain; \\
\setlength{\leftskip}{-16pt}$\mathcal{S^\alpha}$ generates $C_\Theta$ according to Fig.~\ref{Fig:fair-exchange-circuit2}. \\
\setlength{\leftskip}{-16pt}$\mathcal{S^\alpha}$ generates and opens $\mathsf{(pk,vk)}$ = $\Pi$.${\mathsf{Setup}}(1^{\lambda}, C_\Theta)$.\\
}
%
\item[\hspace{7pt}Open:] ~\\
{
\setlength{\leftskip}{-16pt}$\mathcal{S^\alpha}$, $\mathcal{R^\alpha}$ send $\mathsf{v}$ to $\xi^\alpha$ based on $\mathsf{Tx}_{\text{Open}}$;\\
\setlength{\leftskip}{-16pt}$\mathcal{S^\alpha}$ invokes $\mathsf{\Theta.Share}$ to generate key shares $\mathsf{k_{\mathcal{S}[1:n]}}$;\\
\setlength{\leftskip}{-16pt}$\mathcal{S^\alpha}$ computes hash results $\mathsf{h(k_\mathcal{S})}$, $\mathsf{h(k_{\mathcal{S}[1:n]})}$;\\
\setlength{\leftskip}{-16pt}$\mathcal{S^\alpha}$ sends ($\mathsf{h(k_\mathcal{S})}, n, t, \mathsf{h(k_{\mathcal{S}[1:n]})}$) to $\xi^\alpha$ based on $\mathsf{Tx}_{\text{Upload}}$;\\
\setlength{\leftskip}{-26pt}$\pmb{\rightarrow}$ Upon getting the addresses of $\mathsf{\mathcal{M}^\alpha_{[1:n]}}$ and $\mathsf{sn}^\alpha$ from $\xi^\alpha$:\\
\setlength{\leftskip}{-16pt}$\mathcal{S^\alpha}$ signs ($\mathsf{k_{\mathcal{S}i}}$, $\mathsf{sn}^\alpha$) and sends them to $\mathsf{\mathcal{M}^\alpha_{i}}$;\\
\setlength{\leftskip}{-16pt}$\mathsf{\mathcal{M}^\alpha_{i}}$ checks $\mathsf{k_{\mathcal{S}i}}$ based on $(t,n)$-VSS.$\mathsf{Verify}$;\\
\setlength{\leftskip}{-16pt}$\mathsf{\mathcal{M}^\alpha_{i}}$ sends ($\mathsf{Sig_\mathcal{S}}$,$\mathsf{{k_{\mathcal{S}i}}}$, $\mathsf{sn}^\alpha$) based on $\mathsf{Tx}_{\text{Appeal}}$ (When $\mathsf{\mathcal{M}^\alpha_{i}}$ gets the fake key share).\\
}
%
\item[\hspace{7pt}Exchange:]$\pmb{\rightarrow}$ Upon detecting the state of $\xi^\alpha$ becomes $\mathtt{Open_{CE}}$ or $\mathtt{Open}$:\\
{
\setlength{\leftskip}{-16pt}$\mathcal{S^\alpha}$ generates the ciphertext $\mathsf{\overline{m}_{{\mathcal{S}}}}$ based on the plaintext $\mathsf{m_{\mathcal{S}}}$ and the key $\mathsf{k_\mathcal{S}}$;\\
\setlength{\leftskip}{-16pt}$\mathcal{S^\alpha}$ generates the proof $\pi_{\mathcal{S}}$ = $\Pi.\mathsf{Prove}(\mathsf{pk, m_{\mathcal{S}}, k_{\mathcal{S}[1:n]})}$; \\
\setlength{\leftskip}{-16pt}/* Note: $\mathcal{S^\alpha}$ can repeat the above two steps to generate multiple ciphertexts and proofs. */ \\
\setlength{\leftskip}{-16pt}$\mathcal{S^\alpha}$ send (\{$Tr, \pi_\mathcal{S}$\}) to $\mathcal{R}^\alpha$;\\
\setlength{\leftskip}{-16pt}$\mathcal{R^\alpha}$ verifies each $\pi_{\mathcal{S}}$ based on $\Pi.\mathsf{Verify(vk,\pi_{\mathcal{S}})}$;\\
\setlength{\leftskip}{-16pt}$\mathcal{R^\alpha}$ can generate a $Sr$ to open a sub-channel. \\
}
%
\item[\hspace{7pt}Close:]~\\
{
\setlength{\leftskip}{-16pt}Channel participants ($\mathcal{S^\alpha}$, $\mathcal{R^\alpha}$ and others in sub-channels) fetch related (\{$Tr$, $Sr$\}) to compute the final state $f$ in each channel;\\
\setlength{\leftskip}{-16pt}Channel participants send (\{$f, Sr$\}) to $\xi^\alpha$ based on $\mathsf{Tx}_{\text{Close}}$;\\
\setlength{\leftskip}{-16pt}$\mathcal{S^\alpha}$ creates a preimage $\mathsf{pre}$ and compute its hash digest $\mathsf{h({pre})}$;\\
\setlength{\leftskip}{-16pt}$\mathcal{S^\alpha}$ sends ($\mathsf{h({pre})}$) to $\xi^\alpha$ based on $\mathsf{Tx}_{\text{Lock}}$;\\
\setlength{\leftskip}{-16pt}$\mathcal{R^\alpha}$ learns $\mathsf{h({pre})}$ from $\xi^\alpha$.\\
\setlength{\leftskip}{-26pt}$\pmb{\rightarrow}$ Upon $\mathcal{R^\beta}$ learning $\mathsf{pre'}$ from the blockchain $\beta$: \\
\setlength{\leftskip}{-16pt}$\mathcal{R^\alpha}$ sends $\mathsf{pre'}$ and $\mathsf{h'(k_\mathcal{S})}$ to $\xi^\alpha$ based on $\mathsf{Tx}_{\text{Update-EIE}}$.\\
\setlength{\leftskip}{-16pt}$\mathcal{M^\alpha}$ sends $\mathsf{pre'}$ and $\mathsf{h'(k_\mathcal{S})}$ to $\xi^\alpha$ based on $\mathsf{Tx}_{\text{Update-EIE}}$ (When $\mathcal{R^\alpha}$ does not provides $\mathsf{pre'}$ within $T_3$).\\
\setlength{\leftskip}{-26pt}$\pmb{\rightarrow}$ Upon receiving the request for key recovery from $\xi^\alpha$:\\
\setlength{\leftskip}{-16pt}$\mathsf{\mathcal{M}_{i}^\alpha}$ sends $\mathsf{k_{\mathcal{S}i}}$ to $\xi^\alpha$ based on $\mathsf{Tx}_{\text{Recover}}$.\\
\setlength{\leftskip}{-26pt}$\pmb{\rightarrow}$ Upon receiving key shares from $\xi^\alpha$:\\
\setlength{\leftskip}{-16pt} $\mathcal{R^\alpha}$ recovers the key $\mathsf{k_\mathcal{S}}$ and uses it to decrypt the ciphertext (sent by $\mathcal{S}^\alpha$) based on $\mathsf{\Theta.Recover}$.\\
}
\end{basedescript}
\hrule
\vspace{5pt}
\begin{basedescript}{\desclabelwidth{30pt}}
\item[{\bf The process in the blockchain $\beta$ ($\mathcal{S^\beta}$, $\mathcal{R^\beta}$, $\mathcal{M^\beta}$):}]
\item[\hspace{7pt}Initialize:]~\\
{
\setlength{\leftskip}{-16pt}All get ($\widetilde{\mathsf{pk}}$, $\widetilde{\mathsf{sk}}$) from blockchain. \\
}
%
\item[\hspace{7pt}Open:]~\\
{
\setlength{\leftskip}{-16pt}$\mathcal{S^\beta}$, $\mathcal{R^\beta}$ send $\mathsf{v}$ to $\xi^\beta$ based on $\mathsf{Tx}_{\text{Open}}$;\\
\setlength{\leftskip}{-16pt}$\mathcal{R^\beta}$ invokes $\mathsf{\Theta.Share}$ to generate key shares $\mathsf{k_{\mathcal{R}[1:n]}}$;\\
\setlength{\leftskip}{-16pt}$\mathcal{R^\beta}$ computes hash results $\mathsf{h(k_\mathcal{R})}$, $\mathsf{h(k_{\mathcal{R}[1:n]})}$;\\
\setlength{\leftskip}{-16pt}$\mathcal{R^\beta}$ sends ($\mathsf{h(k_\mathcal{R})}, n, t, \mathsf{h(k_{\mathcal{R}[1:n]})}$) to $\xi^\beta$ based on $\mathsf{Tx}_{\text{Upload}}$;\\
\setlength{\leftskip}{-26pt}$\pmb{\rightarrow}$ Upon getting the addresses of $\mathsf{\mathcal{M}^\alpha_{[1:n]}}$ and $\mathsf{sn}^\beta$ from $\xi^\beta$:\\
\setlength{\leftskip}{-16pt}$\mathcal{R^\beta}$ signs ($\mathsf{k_{\mathcal{R}i}}$, $\mathsf{sn}^\beta$) and sends them to $\mathsf{\mathcal{M}^\beta_{i}}$;\\
\setlength{\leftskip}{-16pt}$\mathsf{\mathcal{M}^\alpha_{i}}$ checks $\mathsf{k_{\mathcal{R}i}}$ based on $(t,n)$-VSS.$\mathsf{Verify}$;\\
\setlength{\leftskip}{-16pt}$\mathsf{\mathcal{M}^\beta_{i}}$ sends ($\mathsf{Sig_\mathcal{R}}$,$\mathsf{{k_{\mathcal{R}i}}}$, $\mathsf{sn}^\beta$) based on $\mathsf{Tx}_{\text{Appeal}}$ (When $\mathsf{\mathcal{M}^\beta_{i}}$ gets the fake key share).\\
}
%
\item[\hspace{7pt}Exchange:]$\pmb{\rightarrow}$ Upon detecting the state of $\xi^\beta$ becomes $\mathtt{Open_{CE}}$ or $\mathtt{Open}$:\\
{
\setlength{\leftskip}{-16pt}$\mathcal{R^\beta}$ generates the ciphertext $\mathsf{\overline{m}_{{\mathcal{R}}}}$ based on the plaintext $\mathsf{m_{\mathcal{R}}}$ and the key $\mathsf{k_\mathcal{R}}$;\\
\setlength{\leftskip}{-16pt}$\mathcal{R^\beta}$ generates the proof $\pi_{\mathcal{R}}$ = $\Pi.\mathsf{Prove}(\mathsf{pk, m_{\mathcal{R}}, k_{\mathcal{R}[1:n]}})$; \\
\setlength{\leftskip}{-16pt}/* Note: $\mathcal{R^\beta}$can repeat the above steps to generate multiple ciphertexts and proofs. */ \\
\setlength{\leftskip}{-16pt}$\mathcal{R^\beta}$ send multiple (\{$Tr, \pi_\mathcal{R}$\}) to $\mathcal{S}^\beta$;\\
\setlength{\leftskip}{-16pt}$\mathcal{S^\beta}$ verifies each $\pi_{\mathcal{R}}$ based on $\Pi.\mathsf{Verify(vk, \pi_{\mathcal{R}})}$; \\
\setlength{\leftskip}{-16pt}$\mathcal{S^\beta}$ can generate a $Sr$ to open a sub-channel. \\
}
%
\item[\hspace{7pt}Close:]~\\
{
\setlength{\leftskip}{-16pt}Channel participants ($\mathcal{S^\beta}$, $\mathcal{R^\beta}$ and others in sub-channels) fetch related (\{$Tr$, $Sr$\}) to compute the final state $f$ in each channel;\\
\setlength{\leftskip}{-16pt}Channel participants send (\{$f, Sr$\}) to $\xi^\beta$ based on $\mathsf{Tx}_{\text{Close}}$;\\
\setlength{\leftskip}{-26pt}$\pmb{\rightarrow}$ Upon $\mathcal{R^\alpha}$ learning $\mathsf{h({pre})}$ from $\xi^\alpha$:\\
\setlength{\leftskip}{-16pt}$\mathcal{R^\beta}$ sends ($\mathsf{h({pre})}$) to $\xi^\beta$ based on $\mathsf{Tx}_{\text{Lock}}$;\\
\setlength{\leftskip}{-26pt}$\pmb{\rightarrow}$ Upon detecting the state of $\xi^\beta$ becomes $\mathtt{Lock}$: \\
\setlength{\leftskip}{-16pt}$\mathcal{S^\beta}$ sends $\mathsf{pre'}$ and $\mathsf{h'(k_\mathcal{R})}$ to $\xi^\beta$ based on $\mathsf{Tx}_{\text{Update-EIE}}$.\\
\setlength{\leftskip}{-16pt}$\mathcal{R^\beta}$ learns $\mathsf{{pre'}}$ from $\xi^\beta$.\\
\setlength{\leftskip}{-26pt}$\pmb{\rightarrow}$ Upon receiving the request for key recovery from $\xi^\beta$: \\
\setlength{\leftskip}{-16pt}$\mathsf{\mathcal{M}_{i}^\beta}$ sends $\mathsf{k_{\mathcal{R}i}}$ to $\xi^\beta$ based on $\mathsf{Tx}_{\text{Recover}}$.\\
\setlength{\leftskip}{-26pt}$\pmb{\rightarrow}$ Upon receiving key shares from $\xi^\beta$:\\
\setlength{\leftskip}{-16pt} $\mathcal{S^\beta}$ recovers the key $\mathsf{k_\mathcal{R}}$ and uses it to decrypt the ciphertext (sent by $\mathcal{R}^\beta$) based on $\mathsf{\Theta.Recover}$.\\
}
\end{basedescript}
\end{multicols}\vspace{-15pt}
\end{framed}\vspace{-10pt}
\end{figure*}
\subsection{Analysis} \label{discussions}
In this subsection, we
prove that $\mathsf{Cross}$-$\mathsf{Channel}$ possesses the properties of fairness and atomicity. Its scalability will be demonstrated through experiments in Sec.~\ref{Performance}.
\begin{assumption}
We assume that behaviors of the participants $\mathcal{S, R, M}$ within each chain follow the security model defined in Sec.~\ref{model}.
Consider Byzantine fault-tolerance, we also assume that an $N_{node}$-node
network can tolerate up to $\ell$ Byzantine nodes and support any reconstruction threshold within $ [\ell+1, N_{node}-\ell]$~\cite{das2022practical,huang2022workload}, where $N_{node}=3\ell+1$.
\label{assumption-1}
\end{assumption}
\begin{lemma}
zk-SNARK is a non-interactive zero-knowledge proof technique that satisfies completeness, soundness, and zero knowledge~\cite{groth2016size}.
\label{lem-1}
\end{lemma}
\begin{lemma}
Based on zk-SNARK, a sender (e.g., $\mathcal{S}$) can convince a receiver (e.g., $\mathcal{R}$) that the output ciphertext $\overline{m}$ is encrypted based on $m$ and $k_{[1:n]}$. Besides that, the receiver $\mathcal{R}$ cannot get any knowledge about $m$ and $k_{[1:n]}$ from public information.
\label{lem-2}
\end{lemma}
\begin{proof}
Based on our model defined in Sec.~\ref{model}, $\mathcal{S}$ and $\mathcal{R}$ can be arbitrarily malicious. So,
a malicious $\mathcal{S}$ would provide fake $m$ and $k_{[1:n]}$ to convince an honest $\mathcal{R}$, and a malicious $\mathcal{R}$ would try to obtain $m$ and $k_{[1:n]}$ from the public information provided by $\mathcal{S}$.
In order to prevent the above malicious behaviors, we adopt zk-SNARK (introduced in Sec.~\ref{zk-snark}), and design a new circuit (introduced in Fig.~\ref{Fig:fair-exchange-circuit2}) for it. Specifically, we take the encrypted information $m$ and key shares $k_{[1:n]}$ as private inputs, and implement the Encrypt, Recover, and Hash algorithms in the circuit.
Based on Lemma~\ref{lem-1}, the completeness of zk-SNARK ensures that an honest $\mathcal{S}$ with valid $m$ and $k_{[1:n]}$ can always convince $\mathcal{R}$.
When $\mathcal{S}$ is malicious, the soundness of zk-SNARK makes it impossible for $\mathcal{S}$ (with probabilistic polynomial-time witness extractor $\mathcal{E}$) to provide fake secrets (i.e., $m$ and $k_{[1:n]}$) to deceive $\mathcal{R}$ (shown in Equation~(\ref{eq:soundness})). In other words, $\mathcal{R}$ can determine whether $\mathcal{S}$ provides fake private inputs based on the public parameters, e.g., the common reference string $\mathsf{crs} (\mathsf{pk,vk})$, the proof $\pi$, and the public inputs.
\begin{equation}
\Pr{\left[
\begin{split}
& C(m, k) \neq R_C \\
& {\mathrm{Verify}}(\mathsf{vk}, \pi) = 1
\end{split}
\Biggm\vert
\begin{split}
& {\mathtt{Setup}}(1^{\lambda}, C) \rightarrow (\mathsf{crs})\\
& \mathcal{S}(\mathsf{pk, vk}) \rightarrow (\pi)\\
& \mathcal{E}(\mathsf{pk, vk}) \rightarrow (m, k_{[1:n]})
\end{split}
\right]} \le
\mathsf{negl}(\lambda).
\label{eq:soundness}
\end{equation}
The zero knowledge of zk-SNARK ensures that for every probabilistic polynomial-time (PPT) malicious $\mathcal{R}$, the probability that $\mathcal{R}$ can take private inputs from the proof can be ignored.
Thus one can conclude that zk-SNARK in our general fair exchange protocol ensures the authenticity and privacy of plaintexts and key shares.
\end{proof}
\begin{theorem}[Fairness]
In the scenario of encrypted information exchange,
$\mathcal{S}$ and $\mathcal{R}$ can be guaranteed that if $\mathcal{S}$ gets $\mathcal{R}$'s secret (i.e., the plaintext), $\mathcal{R}$ would also get $\mathcal{S}$'s secret, and vice versa.
\label{theorem-3}
\end{theorem}
\begin{proof}
We design the general fair exchange protocol $\Theta$ to achieve fairness in $\mathsf{Cross}$-$\mathsf{Channel}$.
In the process of {\bf $\Theta$.Exchange} (details in Sec.~\ref{general-fair-exchange}), based on Lemma~\ref{lem-2}, one can get that
the general fair exchange protocol adopts zk-SNARK to help $\mathcal{S}$ and $\mathcal{R}$ achieve the authenticity and privacy of the exchanged information.
So, in this step, neither side has access to the other's secrets.
In other steps,
the general fair exchange protocol adopts $(t,n)$-VSS. Let's consider the impact of Byzantine fault-tolerance on the fairness of the protocol.
Suppose that $n$ nodes have been chosen to receive the key shares and the key threshold is $t$. There are two possible Byzantine behaviors that can break the protocol.
First, when the number of Byzantine nodes receiving the key shares is greater than or equal to $t$, these nodes can collude to break the fairness and recover the key.
Second, when the key threshold $t$ is smaller than the number of honest nodes in the $n$ nodes receiving the key shares, the sender of {\bf $\Theta$.Recover} cannot recover the key when all the Byzantine nodes maliciously refuse to provide their key shares because the number of key shares is less than $t$. To overcome the first problem, we require that $t>\ell$, and to counter the second one, we set $n\ge t+\ell$, where $\ell$ is the maximum number of Byzantine nodes in the whole network. These two constraints can accommodate the worst case in which all the $\ell$ malicious nodes are unluckily selected to receive the key shares.
%
%
In summary, one can see that by carefully setting $n$ and $t$ it is impossible for the Byzantine nodes to collect all $t$ key shares even if all Byzantine nodes collude, and thus allow the key to be successfully recovered even if all Byzantine nodes do not follow the protocol to send $\mathsf{Tx}_{\text{Recover}}$. By this way one can guarantee that the general fair exchange protocol can effectively provide fairness for the interaction between two parties.
\end{proof}
\begin{theorem}[Atomicity]
Let objects $x_\mathcal{S}$ $\in$ $\mathcal{S}$ and $x_\mathcal{R}$ $\in$ $\mathcal{R}$ before a cross-chain exchange. The settlement result after $\mathsf{Cross}$-$\mathsf{Channel}$ can only be $(x_\mathcal{S}$ $\in$ $\mathcal{S}$ $\bigwedge$ $x_\mathcal{R}$ $\in$ $\mathcal{R})$ $\bigvee$ $(x_\mathcal{S}$ $\in$ $\mathcal{R}$ $\bigwedge$ $x_\mathcal{R}$ $\in$ $\mathcal{S})$
\label{theorem-4}
\end{theorem}
\begin{proof}
Before closing the channel,
the refusal of settlement by one or both parties would result in $(x_\mathcal{S}$ $\in$ $\mathcal{S}$ $\bigwedge$ $x_\mathcal{R}$ $\in$ $\mathcal{R})$, which does not break the atomicity of the protocol.
After both parties enter settlement (the {\bf Close} phase in Sec.~\ref{cross-channel and applications}), the hierarchical settlement protocol (proposed in Sec.\ref{hierarchical-channel-scheme}) is adopted during the interaction process.
Specifically, $\mathcal{S}$ and $\mathcal{R}$ use the same hash lock $\mathsf{h(pre)}$ ($\mathsf{pre}$ is known only by $\mathcal{S}$) to lock the exchanged information, and set a timer ($T_3$ or $T_4$) in each blockchain ($\alpha$ or $\beta$) to avoid the situation of information being deadlocked. There are two cases we need to consider.
\textit{First,} $\mathcal{S}$ uses $\mathsf{pre}$ to unlock the information in $\beta$. Then, $\mathcal{R}$ learns $\mathsf{pre}$ and uses it to unlock the information in $\alpha$ $(x_\mathcal{S}$ $\in$ $\mathcal{R}$ $\bigwedge$ $x_\mathcal{R}$ $\in$ $\mathcal{S})$.
\textit{Second,} if $\mathcal{S}$ does not provide $\mathsf{pre}$, $\mathcal{R}$ cannot get $\mathsf{pre}$; thus it cannot unlock the information in $\alpha$. When the timer expires, the smart contract returns the locked information $(x_\mathcal{S}$ $\in$ $\mathcal{S}$ $\bigwedge$ $x_\mathcal{R}$ $\in$ $\mathcal{R})$. Note that $T_3$ in $\alpha$ is longer than $T_4$, ensuring that a malicious $\mathcal{S}$ in $\beta$ cannot provide $\mathsf{pre}$ after $T_3$ in $\alpha$ times-out.
However, \textit{in an asynchronous network}, each account may not be able to receive or upload information within a certain time due to network latency. This implies that $\mathcal{R}$ may not be able to receive $\mathsf{pre}$ and upload it within $T3$ after $\mathcal{S}$ provides $\mathsf{pre}$ $(x_\mathcal{S}$ $\in$ $\mathcal{S}$ $\bigwedge$ $x_\mathcal{R}$ $\in$ $\mathcal{S})$. Therefore $\mathsf{Cross}$-$\mathsf{Channel}$ sets a timer $T_5$ in $\alpha$ to ensure that if $\mathcal{R}$ cannot provide $\mathsf{pre}$ within $T_3$, any honest miner who receives $\mathsf{pre}$ can help $\mathcal{R}$ to upload $\mathsf{pre}$ within $T_5$ $(x_\mathcal{S}$ $\in$ $\mathcal{R}$ $\bigwedge$ $x_\mathcal{R}$ $\in$ $\mathcal{S})$. Correspondingly, the honest miner would get rewards from the smart contract. As for the incentive mechanism, our scheme can be compared to a specific application of some blockchains such as the main chain of Ethereum, where miners can get rewards
for their work.
\end{proof}
\section{Implementation and Performance Evaluation}\label{sec:Implementation}
In this section, we present our concrete implementation of $\mathsf{Cross}$-$\mathsf{Channel}$ and test its performance.
\subsection{Implementation}\label{Implementation}
\noindent{\bf On-chain deployment: Ethereum and smart contract. } The Ethereum Geth\footnote{https://github.com/ethereum/go-ethereum} and Solidity\footnote{https://github.com/ethereum/solidity} come from Github. We use Geth to construct a test network for $\mathsf{Cross}$-$\mathsf{Channel}$ validation, and implement the smart contract $\xi$ based on Solidity.
In order to facilitate the interactions between smart contract and Ethereum, we adopt web3.py\footnote{https://pypi.org/project/web3} to deploy and call $\xi$.
\noindent{\bf Off-chain deployment: zk-SNARK and $(t,n)$-VSS. } We employ the zk-SNARK algorithm in Github\footnote{https://github.com/scipr-lab/libsnark} and $(t,n)$-VSS scheme in~\cite{pedersen1991non}. For the circuit in zk-SNARK (shown in Fig.~\ref{Fig:fair-exchange-circuit2}), we adopt MIMC to implement the $\mathsf{Encryption}$ and $\mathsf{Hash}$ algorithms because MIMC is encryption-friendly and can reduce circuit complexity and computational overhead.
We test the performance of $\mathsf{Cross}$-$\mathsf{Channel}$ on a local server and AliCloud.
The local server is equipped with an Intel$^\circledR$ Xeon(R) Silver 4214R CPU @ 2.40 GHz * 16 and 78.6 GB RAM running 64-bit Ubuntu 20.04.2 LTS.
In the experiment on AliCloud, we use 50 ecs.g6.2xlarge instances, with each running Ubuntu 20.04 system Intel Xeon (Cascade Lake) Platinum 8269CY processor and having 8 vCPUs of frequency 2.5/3.2 GHz and 16 GB RAM. We start 4 docker nodes in each instance to form two 100-node blockchains based on the Proof-of-Work (PoW) consensus algorithm, and the number of transactions at each blockchain reaches up to 1,000.
\subsection{Performance Evaluation}\label{Performance}
\noindent {\bf On-chain performance: smart contract.}
In this experiment, we build a 2-level channel (including a sub-channel) in a 20-node blockchain to test the basic performance of the smart contract, i.e., the execution time and gas consumption of each function.
As shown in TABLE~\ref{tab:sc_cost}, one can see that the execution time of each function is in the millisecond level, and $\mathsf{{Upload}}$ consumes the most gas (about 345,000).
The above results are reasonable because $\mathsf{Tx}_{\text{Upload}}$ involves more uploaded data, e.g., multiple signatures and keys (the definition of $\mathsf{Tx}_{\text{Upload}}$ is shown in Sec.~\ref{general-fair-exchange}).
TABLE~\ref{tab:comparation} displays the total smart contract costs in scenarios of $N$ currency exchanges (CE), $N$ fair exchanges (FE), and $N$ encrypted information exchanges (EIE), and compares HTLC and MAD-HTLC (the two most related cross-chain schemes) with our $\mathsf{Cross\text{-}Channel}$. Note that HTLC and MAD-HTLC do not support FE and EIE, and take one on-chain exchange for each CE operation, while our $\mathsf{Cross\text{-}Channel}$ takes only one on-chain exchange for $N$ operations, benefiting from the proposed channel scheme, where $N$ can be arbitrarily large. More specifically,
for CE, $\mathsf{Cross\text{-}Channel}$ consumes about 1,330,000 gas to process $N$ cross-chain exchanges and HTLC (MAD-HTLC) needs to take about 420,000$\times N$ (750,000$\times N$) gas to process the same volume of operations.
For FE, $\mathsf{Cross\text{-}Channel}$ consumes the same gas as that for currency exchange, while for EIE, $\mathsf{Cross\text{-}Channel}$ needs to consume more gas (around 2,700,000 gas) due to the adoption of the general fair exchange protocol (introduced in Sec.~\ref{general-fair-exchange}) to solve the UE problem.
Next we test the impact of the number of sub-channels on gas consumption and throughput (TPS). Our results indicate that whenever a new sub-channel is added,
the gas consumed by the entire protocol is increased by nearly 400,000, because both the number of $\mathsf{Tx}_{\text{Close}}$ and the amount of data in $\mathsf{Tx}_{\text{Close}}$, e.g., sub-channel receipts, are increased. The benefits obtained from this gas increase is the increased number of processed transaction receipts. For example, when the number of sub-channels is increased to $L$, $N \times L$ transaction receipts can be processed, assuming that each sub-channel can process $N$ receipts. In fact, given a quantitative resource budget, $\mathsf{Cross\text{-}Channel}$ can handle more operations, as $N$ and $L$ can be large, compared to HTLC and its variation, implying that the system throughput with $\mathsf{Cross\text{-}Channel}$ can be significantly enhanced.
\begin{table}[htb]
\centering
\begin{threeparttable} \scriptsize
\centering
\begin{tabular}{cccc}
\toprule
Contract functions & Execution time & Gas/ ETH/ USD\tnote{$\ast$} \\% & Applications\tnote{$\dagger$} \\
\midrule
$\mathsf{{Open}}$ & 13.663ms & 70,062/ 0.0000701/ 0.0948 \\
$\mathsf{{Upload}}$ & 19.102ms & 345,146/ 0.000345/ 0.467 \\
$\mathsf{{Appeal}}$ & 13.047ms & 57,542/ 0.0000575/ 0.0778 \\
$\mathsf{{Close}}$ & 18.905ms & 149,942/ 0.000150 / 0.203 \\
$\mathsf{{Recover}}$ & 10.883ms & 28,219/ 0.0000282/ 0.0382 \\
$\mathsf{{Lock}}$ & 14.843ms & 146,300/ 0.000146 / 0.198 \\
$\mathsf{{Update}}$ & 14.062ms & 79,121/ 0.0000791/ 0.107 \\
$\mathsf{{Update\text{-}EIE}}$ & 14.578ms & 108,791/ 0.000109/ 0.147 \\
\bottomrule
\end{tabular}
\begin{tablenotes}
\footnotesize
\item[$\ast$] Gasprice = 1 Gwei, 1 Ether = $10^9$ Gwei, and 1 Ether = 1353 USD.
\end{tablenotes}
\end{threeparttable}
\caption{Smart-contract experiments of $\mathsf{Cross}$-$\mathsf{Channel}$}\label{tab:sc_cost}
\end{table}
\begin{table}[htb]
\centering
\begin{threeparttable}\scriptsize
\centering
\begin{tabular}{cccccc}
\toprule
& Cross-Channel & Cross-Channel & HTLC\tnote{$\S$} & MAD-HTLC\tnote{$\S$} \\
& CE \& FE & EIE & CE & CE \\
\midrule
Gas & 1,330,858 & 2,701,308 & 429,532$\times N$ & 758,095$\times N$ \\
\bottomrule
\end{tabular}
\begin{tablenotes}
\footnotesize
\item[$\S$] The results are calculated based on the data provided in~\cite{tsabary2021mad}.
\end{tablenotes}
\end{threeparttable}
\caption{The comparison with other cross-chain protocols}\label{tab:comparation}
\end{table}
\begin{figure*}[htb]
\subfigure[]{\includegraphics[width=0.24\textwidth]{figure/CC_node_Latency3.pdf}}
\subfigure[]{\includegraphics[width=0.24\textwidth]{figure/CC_node_Latency4.pdf}}
\subfigure[]{\includegraphics[width=0.24\textwidth]{figure/CC_node_Latency1.pdf}}
\subfigure[]{\includegraphics[width=0.24\textwidth]{figure/CC_node_Latency2.pdf}}
\subfigure[]{\includegraphics[width=0.24\textwidth]{figure/CC_node_TPS3.pdf}}
\subfigure[]{\includegraphics[width=0.24\textwidth]{figure/CC-zksnark-time.pdf}}
\subfigure[]{\includegraphics[width=0.24\textwidth]{figure/CC-zksnark-size.pdf}}
\subfigure[]{\includegraphics[width=0.24\textwidth]{figure/CC-vss-time.pdf}}
\caption{Transaction delay tests (a) (b) (c) (d), TPS experiments (e), and the performance of zk-SNARK (f) (g) and verifiable secret sharing scheme (h)
.}
\label{fig:test}
\end{figure*}
\noindent {\bf On-chain performance: transaction delay.}
We simulate an EIE scenario to test the latency of each transaction. Specifically, we first build a $N_{node}$-blokchain network on AliCloud, where $N_{node}$ is the number of nodes in the blockchain. Then we let each node opens $m$ channels and each channel has two levels (i.e., including one sub-channel); each channel is open for about 100 seconds in average, during which users are allowed to interact (EIE) within the channel. The transmission rate is set to be roughly 390 MB/s.
Based on the above experimental setup, we test the transaction delay by changing $N_{node}$ and $m$, where $N_{node}$ varies from 10 to 100 and $m \in \{5,10\}$.
Note that the transaction delay refers to the time interval from when a transaction is sent to the blockchain until the corresponding block is confirmed by the miners.
The transaction delays are reported in Fig.~\ref{fig:test} (a) (b) (c) (d). One can see that when the number of nodes rises from 10 to 100, the transaction delay increases. The reason for this trend lies in that the more nodes in the network, the longer time the broadcast and consensus of transactions consume.
Besides that, the number of channels created in the network would also affect the transaction delay. For example, in a 100-node network, when $m = 5$, which means that 500 channels in total are constructed between nodes, the transaction latency is about 3.0-6.0 seconds (Fig.~\ref{fig:test} (a) (b)). When $m = 10$, the latency of various transactions for constructing 1,000 channels grows to about 3.2-7.4 seconds (Fig.~\ref{fig:test} (c) (d)).
The reason for this trend is that when the number of created channels increases, the number of transactions waiting in the queue increases, thus increasing the transaction delay.
\noindent {\bf On-chain performance: throughput (TPS) and scalability.}
In this experiment, we simulate a dynamic equilibrium state with a fixed number of channels (one-level), and discuss the throughput and scalability in three scenarios, i.e., CE, FE and EIE.
The unit for throughput is TPS, which refers to the number of transaction receipts our scheme can process per second.
Specifically, the numbers of channels opened and closed are dynamic variables, denoted as $v_{1}$ and $v_{2}$, respectively. We set $v_{1} = v_{2} = 10$ and $m = 5$, ensuring that in a $N_{node}$-blockchain, the total number of channels remains unchanged ($5 \times N_{node}$), but the number of newly opened channels and that of closed ones are both set to 10, maintaining a dynamic balance.
Besides that, we set the transmission rate of a channel to be roughly 390 MB/s and the test time lasts 100 seconds.
The transaction receipt is about 130 Bytes in CE (the definition of transaction receipt is described in Sec. \ref{hierarchical-channel-scheme}), and around 1.3 KB in FE and EIE (adding encrypted data blocks and information related to zero-knowledge proofs). Based on the above data, one can obtain that a channel can send $3 \times 10^6$ transaction receipts per second ($Tr$/s) in CE, and $3 \times 10^5$ $Tr$/s in FE and EIE.
Note that, the values of variables $v_{1}, v_{2}, m$ do not affect the trend of the experimental results; thus we make them fixed.
Fig.~\ref{fig:test} (e) demonstrates that the TPS of $\mathsf{Cross\text{-}Channel}$ is linear to the network size, showing good scalability. This implies that the more nodes (channels) in the network, the higher the system throughput.
In a $N_{node}$-blockchain, the number of channels in the network is proportional to that of the nodes, and channels can process transaction receipts in parallel; thus the overall transaction processing rate of the system is nearly $\mathcal{O}(N_{node})$,
and the time to process the above transaction receipts grows at rate $\mathcal{O}(1)$.
\noindent {\bf Off-chain performance: zk-SNARK and VSS.}
In this experiment, we first test the effect of the size of the exchanged encrypted objects on the performance of zk-SNARK. The encrypted object includes multiple data blocks, and each block has 100-bit data.
As shown in Fig.~\ref{fig:test} (e), one can see that as the number of data blocks goes from 10 to 10,000, the time consumption of $\Pi.\mathsf{Setup}$ and that of $\Pi.\mathsf{Prove}$ gradually increase. This does not cost extra on-chain resources, because zk-SNARK is run off-chain and is done before running $\mathsf{Cross}$-$\mathsf{Channel}$.
Additionally, as shown in Fig.~\ref{fig:test} (g),
zk-SNARK demonstrates great succinctness.
The proof size is kept at 1,019 bits and the running time of $\Pi.\mathsf{Verify}$ is at the millisecond level. Of course, one can further combine with other schemes, e.g. ZKCPlus~\cite{li2021zkcplus}, to optimize the performance of zero-knowledge proof based on specific scenarios.
Then, we test the three steps ($\mathsf{Share}$, $\mathsf{Verify}$ and $\mathsf{Recover}$) at different thresholds and set $(t,n)$ to be (11, 31), (21, 61), (31, 91), (41, 121), and (51, 151), which can effectively solve the Byzantine fault (details shown in Theorem~\ref{theorem-3}).
The time consumption of $(t,n)$-VSS is illustrated in Fig.~\ref{fig:test} (h), and one can get that the time for each step increases steadily as the threshold increases but it remains at the millisecond level.
\section{Conclusion}\label{sec:Conclusion}
In this paper, we propose $\mathsf{Cross}$-$\mathsf{Channel}$, a scalable channel that supports cross-chain services with high throughput.
Specifically,
we design a new hierarchical channel structure and propose a general fair exchange protocol to respectively solve the Unsettled Amount Congestion problem and the Unfair Exchange problem.
Additionally, we design a hierarchical settlement protocol based on HTLC and incentive mechanisms, which can help $\mathsf{Cross}$-$\mathsf{Channel}$ to ensure the correctness of the settlement and enhance the atomicity of the cross-chain interactions in asynchronous networks.
Finally, we implement $\mathsf{Cross}$-$\mathsf{Channel}$ in two 100-node blockchains on AliCloud, and conduct a test with up to 1,000 transactions.
Compared with the state-of-the-art cross-chain protocols, $\mathsf{Cross}$-$\mathsf{Channel}$ adds a small amount of on-chain resource overhead but can bring high throughput.
In our future research, we will extend $\mathsf{Cross}$-$\mathsf{Channel}$ to support multiparty channels, and consider more general operations such as digital asset transfers.
|
\section{Introduction}
The generation of quantum states of light has long been a milestone of quantum optics. An example of such states are squeezed states, in which the uncertainty in one observable reduces at the expense of the other, An example of such variables are number and phase. States of light squeezed in photon number promise numerous applications in quantum information and sensing, allowing, for example, signal detection below the shot noise limit \cite{andersen201630, lawrie2019quantum}. In the limit of infinite number squeezing, one approaches a so-called Fock state, an energy eigenstate of the electromagnetic field Hamiltonian with a perfectly defined photon number $n$. In addition to their importance as fundamental quantum states of light, Fock states can be used to realize numerous quantum computing protocols that require operations on photon number (e.g., photon number subtraction and boson sampling) \cite{zhong2020quantum}. Thus, many applications would be aided by the generation of Fock or other highly number squeezed states.
Despite these potential rewards, macroscopic Fock and near-Fock states are inherently challenging to produce. The main reason for this is that there is a general lack of nonlinear mechanisms which selects specifically for $n$ photons, especially where $n$ may be large. Moreover, most highly quantum states of light are difficult to maintain, due to the deleterious effects of loss.
To address these issues, the physical mechanism of nonlinear dispersive loss has been recently introduced as a potential new method to create ``Fock'' lasers, which possess steady state photon number distributions approaching macroscopic Fock states \cite{rivera2021macroscopic}. This phenomenon relies critically on the ability to simultaneously harness strong optical nonlinearities and low background loss (high $Q$ cavities). Although the fundamental phenomenology of this effect has been developed, there are still large outstanding questions about potential platforms for realizing these new lasers.
One platform of potential interest for lasers which create strongly intensity squeezed light are semiconductor lasers. Semiconductor lasers have become ubiquitous due to their compact form factor, ease of electrical pumping, and wide gain bandwidths. There has already been considerable work on the topic of reducing noise in semiconductor lasers. The earliest methods for generating intensity-squeezed light with these devices involved exploiting the sub-shot noise regime of space charge limited current in diodes \cite{machida1987observation, yamamoto1992photon}. In these demonstrations, reduced photon number noise and increased phase noise were obtained in constant current-driven semiconductor lasers. However, reducing pump noise was shown to reduce intensity noise by at most a few dB. Since these early works, attempts have been made to combine pump noise suppression with optical feedback using amplitude-phase decorrelation \cite{newkirk1991amplitude, kitching1995room, jeremie1999room}. Still, relative intensity noise suppression did not surpass 10 dB, was limited by the linewidth enhancement factor $\alpha_L$, and only appreciable for frequencies below the relaxation oscillation frequency. Because these methods achieve non-negligible squeezing in only part of the intensity noise spectrum, the total variance in photon number is weakly squeezed at best. To use semiconductor lasers in on-chip quantum sensing far below the shot noise limit and in quantum computing protocols, enhanced squeezing is warranted.
Here, we show how semiconductor lasers with a strong and frequency dependent outcoupling could be used to create strongly intensity squeezed light with broadband squeezing of 10 dB or more below the shot noise limit. Additionally, we show that semiconductor lasers with strong carrier nonlinearities in conjunction with dispersive loss can exhibit a new kind of carrier bistability. We show how strong carrier nonlinearity can drastically impact laser steady state behavior, and also show how strong noise reduction can still be achieved even in the presence of this effect. Semiconductor lasers are a strong candidate platform to realize these effects due to their ability to sustain wavelength modulation owing to wide gain bandwidth and also possibly due to strong nonlinearities owing to smaller active regions like quantum wells. They are also amenable to on-chip integration with optical elements such as low loss resonators, photonic crystals, and lithium niobate optics, making them suitable for potential applications in quantum information and sensing \cite{numai2015fundamentals}.
We structure the rest of this paper as follows. Section \ref{sec:theory} reviews the concept of a ``Fock laser,'' shows how this concept can be implemented in a semiconductor laser architecture, and presents analytical formulae for intensity noise in the presence and absence of nonlinearity. Section \ref{sec:results} shows the noise condensation effect for the semiconductor laser architecture, and also presents a new type of bistability based on carrier nonlinearity. Finally, Section \ref{sec:discussion} summarizes our results and presents outlooks and future extensions of this work for semiconductor laser systems.
\section{Theory}
\label{sec:theory}
\subsection{The concept of a ``Fock laser''}
In this section, we introduce and describe the operation of a ``Fock laser'' which can produce laser cavity states far below the shot noise limit, even approaching Fock states in extreme cases. The essential elements of a Fock laser are: (a) a pumped gain medium, (b) Kerr nonlinearity which gives the laser cavity an intensity dependent frequency shift, and (c) a frequency dependent outcoupling mechanism which gives the laser cavity a frequency dependent loss rate (Fig. \ref{fig:schema}a).
We first describe how under the right conditions, the combination of Kerr nonlinearity and frequency dependent loss lead to a laser cavity with an effective \emph{intensity dependent loss} that controls the quantum state of light produced by the laser. Consider the cavity architecture shown in Fig. \ref{fig:schema}a, but without any gain medium. Additionally, we focus on a single cavity mode, with annihilation operator $a$. As is well known, a cavity containing a Kerr nonlinearity develops an intensity dependent resonance frequency due to the intensity dependent index of the Kerr material. More specifically, the resonance frequency takes the form $\omega(n) = \omega_0(1 - \beta n),$ where $\omega_0$ is the bare cavity frequency, and $\beta$ is a dimensionless per-photon nonlinearity.
In addition to this Kerr nonlinearity, the cavity contains a frequency dependent outcoupling mirror characterized by the complex transmission amplitude coefficient $t(\omega)$ (associated with the intensity transmission coefficient $T(\omega) \equiv |t(\omega)|^2$). This frequency dependence can come from, for example, a Bragg reflector, or a Fano resonance which gives a Lorentzian reflectivity spectral profile. This frequency dependent outcoupling results in a corresponding frequency dependent damping rate in the cavity. Assuming $t \ll 1$, the frequency dependent damping rate of the cavity is given as $\kappa(\omega) \approx \text{FSR} \cdot T(\omega)$.
When combined, the intensity dependent resonance frequency and frequency dependent dissipation give the cavity mode an effective \emph{intensity dependent dissipation}. The one critical assumption for this description is that the temporal response of the frequency dependent mirror is fast compared to the round trip time of the cavity. This corresponds to an adiabatic limit in which the loss rate is able to near-instantaneously follow shifts in the cavity frequency caused by the Kerr nonlinearity. When these assumptions are fulfilled, the cavity field is subject to an effective intensity dependent damping rate $\kappa(n)$. Sharply frequency dependent reflectivity profiles enable the dissipation rate $\kappa(n)$ to take on forms which are highly nonperturbative in $n$, making this type of nonlinear dissipation fundamentally different than the types of nonlinear dissipation realized by multi-photon absorption. The fact that loss can vary sharply with photon number is fundamental to enabling the strong laser noise condensation we describe below.
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{laser_fig.pdf}
\caption{\textbf{The concept of a Fock laser.} (a) Essential components of a Fock laser (illustrated as a ring laser for concreteness) consisting of at least one mirror with frequency dependent transmission $t(\omega)$. (b) Comparison of conventional and sharp loss Fock laser steady state photon probability distribution $p(n)$. The steady state photon number is determined by the location of intersection between saturable gain and loss. The variance of the probability distribution is determined by the effective ``steepness'' of intersection of the gain and loss curves. While the conventional laser architecture with linear loss results in a near-coherent state far above threshold, the sharp loss architecture results in states with variance below the mean, which correspond to non-classical light. In the most extreme limit, this mechanism can enable the generation of near-Fock states inside the laser cavity.}
\label{fig:schema}
\end{figure}
We now describe how this intensity dependent loss can compress steady state photon statistics (Fig. \ref{fig:schema}b). The steady states of all lasers are characterized by a balance between saturable gain and loss. Gain saturation is itself a form of ``intensity dependent gain,'' in the sense that the net gain seen by the cavity field depends on the intensity. Quantum mechanically, this means that different photon number states experience different amounts of gain. In a conventional laser with intensity independent ``linear loss,'' the loss rate seen by the cavity field is the same for all photon numbers. For photon numbers where gain exceeds loss, an effective ``force'' encourages occupation of yet higher photon numbers; for photon numbers where loss dominates gain, an effective force encourages occupation of lower photon numbers. The intersection point where ``gain equals loss'' represents the equilibrium point between these two forces, and consequently determines the mean photon number of the cavity in the laser steady state. While the intersection point determines the mean photon number, the behavior of the photon number dependent gain and loss in the vicinity of this intersection dictates the variance of the photon number probability distribution $p(n)$. In conventional lasers which are far above threshold, the probability distribution approaches that of a coherent state, with Poissonian statistics (variance equal to the mean). It is this fact that leads to the common statement that ``lasers produce coherent states.''
This situation changes significantly if the number dependent gain or loss profiles deviate substantially from their usual forms. We will focus here on situations where the gain remains identical, but the ordinary linear loss is replaced with a strongly intensity dependent loss. If the loss rises sharply with photon number around its intersection with the saturable gain, then the steady state probability distribution becomes compressed compared to the case of linear loss. To understand this, we can return to the picture of ``forces'' on the probability distribution described for the normal laser. The magnitudes of these forces are essentially determined by the extent to which gain exceeds loss (or vice versa) below (or above) the mean photon number. If we compare the conventional and sharp loss lasers, it is clear that above the mean photon number, loss exceeds gain more aggressively for the sharp loss. Since loss dominates gain more strongly, the probability distribution is more strongly ``discouraged'' from entering this region. The result of this is that the probability distribution has the same mean, but can become strongly compressed, resulting in a much lower variance. Roughly speaking, the mean photon number is determined by the location of the intersection point, while the variance is determined by the ratio of the slopes of the gain and loss.
This mechanism enables the sharp loss laser to create steady states with variance lower than the mean, a feature only possible in non-classical light. In the most extreme limit, the loss rises so sharply that only a single number state (the mean) has a substantial probability of occupation, resulting in a cavity Fock state.
\subsection{Semiconductor Fock laser architecture}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{semiconductor_schematic_JS.pdf}
\caption{\textbf{Semiconductor Fock laser architecture.} (a) Basic semiconductor laser diode heterostructure design with nonlinear dispersive loss. Dispersive outcoupling is generated via the sharp frequency dependent transmission of a photonic crystal element. Coupling of combined Kerr and carrier nonlinearity with dispersive loss $R(\omega)$ creates sharp nonlinear loss $\kappa(n,N)$. Here, $\Delta$ denotes detuning from the dispersive (Lorentzian) resonance and $\gamma$ denotes the width of the dispersive resonance (related to its FWHM). (b) Semiconductor optical nonlinearities, including intensity dependent Kerr effect as well as carrier-dependent free carrier dispersion (FCD) and two photon absorption (TPA). These nonlinearities shift the real part of the active region's refractive index, in turn shifting the resonance frequency in the diode cavity. Weak nonlinear loss from shifting the imaginary part of the refractive index via the Kramers-Kronig relations is also generated, but in most cases is negligible compared to the nonlinear dispersive loss.}
\label{fig:schema2}
\end{figure}
In this section, we introduce the semiconductor ``Fock laser'' architecture that generates strongly noise condensed steady states inside the optical cavity. The device is a semiconductor heterostructure with an electrically pumped active region (Fig. \ref{fig:schema2}a). The cavity mode (with annihilation operator $a$) has a bare resonance frequency $\omega_0$. One of the end facets is a broadband reflector, while the other is a sharply dispersive element, such as a Fano resonance structure (main text) or a Bragg reflector (Supplementary Information (SI)), which equips the cavity with sharply frequency dependent dissipation. Both Kerr and carrier nonlinearity inside the cavity shift the index, and hence the resonance frequency (Fig. \ref{fig:schema2}b). However, as we show in Section \ref{sec:carrier}, the effect of these two nonlinearities on photon quantum statistics is very different.
To lowest order, the cavity frequency depends linearly on the photon and inverted carrier numbers $n$ and $N$ as
\begin{equation}
\omega(n,N)=\omega_0(1-\beta n-\sigma N),
\end{equation}
as analyzed previously using coupled mode theory and supported experimentally \cite{heuck2013heterodyne, yu2013switching, lunnemann2012nonlinear, said1992determination}. The dimensionless, per-photon Kerr nonlinear strength $\beta$ is related to the Kerr nonlinear coefficient $n_2$ (m\textsuperscript{2}/W) via $\beta=-\frac{\hbar\omega n_{\mathrm{eff}}^2 n_2 c}{8V_m}$, where $V_m$ denotes the modal volume, $n_{\mathrm{eff}}$ the cavity's effective refractive index, and $\omega$ the laser frequency. The carrier nonlinearity $\sigma$ is material-dependent, arising predominantly from two-photon absorption, band filling effects, and free carrier dispersion.
The Kerr nonlinearity $\beta$ may come from the gain medium itself (typical strengths for a modal volume equal to the cavity volume near GaAs bandgap are $\beta\sim -10^{-10}$), or from an additional Kerr medium incorporated into the cavity. For the purposes of our analysis, we will also assume that the index, and thus resonance frequency, responds instantaneously to changes in photon and carrier number, relative to other timescales in the system. Specifically, we ignore dispersive effects in the nonlinear strength near bandgap resonance. Materials and operating frequencies may need to be optimized so that this assumption is valid.
In particular, for the examples of Figs. \ref{fig:main} and \ref{fig:carrier}, we will take values typical of GaAs for both $\beta$ and $\sigma$. Note that the results here do not strictly hold for an all GaAs laser, because the third-order nonlinearity $\beta$ is assumed to be instantaneous, which does not pertain to GaAs above bandgap. The theory we obtain below would hold directly for an ``external cavity'' geometry, in which a thin slab of GaAs gain is embedded in a larger cavity (formed by two or more mirrors, as in Fig. \ref{fig:schema}), with the cavity containing a slab of a nonlinear medium which has no dispersion or two-photon absorption at the lasing wavelength (for example, silica). In that case, the effect of carrier nonlinearities becomes much smaller due to the small filling fraction of the gain slab, and the instantaneous Kerr nonlinearity becomes the dominant effect.
Additionally, in the SI we describe how small variations in steady state and noise behavior arise from other weak nonlinear effects such as two-photon absorption (TPA) and nonlinear gain saturation.
Following the same reasoning of the previous section, the photon and carrier number dependence of the resonance frequency can transfer to the loss, so long as the cavity resonance frequency responds instantaneously to changes in both photon and carrier number. In this case, the carrier and photon number dependent loss is given by
\begin{equation}
\kappa(n,N)\equiv \kappa(\omega(n,N)) = -\mathrm{FSR}\cdot \log{R(\omega(n,N))}
\label{eq:loss}
\end{equation}
Hence the sharp frequency dependent element with reflectivity $R(\omega)$ now controls the nonlinear loss. As we will show, these systems can provide new behaviors not just in their steady states, but also through new quantum noise behaviors.
\subsection{Semiconductor laser dynamics}
Semiconductors typically fall into the category of so-called ``class B'' lasers, in which the polarization dynamics decay quickly relative to the timescales associated with carrier recombination and cavity decay. In this case, the polarization dynamics are adiabatically eliminated, resulting in Heisenberg-Langevin equations for photon number, carrier number operators, and phase \cite{agrawal2012fiber}:
\begin{subequations}
\begin{align}
\dot n &=\left(G(n,N) - \kappa(n,N)\right)n+F_n \\
\dot N &=I - \left(nG(n,N) + \gamma_\parallel N\right)+F_N \\
\dot \phi &= \frac{\alpha_L}{2}G(n,N)+F_\phi
\label{eq:rates}
\end{align}
\end{subequations}
To be maximally general here, we allow the gain $G$ and loss $\kappa$ to depend on both the carrier density $N$ and photon number $n$ (the latter could account for gain saturation). In writing this form of the gain and loss, we have assumed that the gain and loss respond effectively instantaneously to changes in the photon and carrier number. Pumping is performed by carrier injection using current $I$ (in units of carrier density per unit time), and $\gamma_\parallel$ denotes the nonradiative decay rate of carriers. Furthermore, $\alpha_L$ denotes the linewidth enhancement factor and is associated with amplitude-phase coupling. The decay rates are associated with Langevin force terms $F_{n,N,\phi}$ with nonzero correlators given by the diffusion coefficients (including the effects of TPA, $\dot{n}=-\alpha_{\mathrm{TPA}}n^2$): $\langle F_n^\dagger F_n\rangle = \langle 2D_{nn}\rangle= 2R_{sp}n+\alpha_{\mathrm{TPA}}n^2, \langle F_N^\dagger F_N\rangle = \langle 2D_{NN} \rangle= 2(R_{sp}n+\gamma_\parallel N), \langle F_N^\dagger F_n\rangle = \langle 2D_{NN} \rangle = -Rn, \langle F_\phi^\dagger F_\phi\rangle = \langle 2D_{\phi\phi} \rangle = R_{sp}/2n$ where $R_{sp}\approx G(n,N)$ denotes the rate of spontaneous emission into the cavity mode, $R_{abs}\approx 0$ denotes the rate of absorption (negligible above threshold), and $R=R_{sp}+R_{abs}$. These correlators give rise to nonzero fluctuations in $n,N$ about their steady state values. A derivation of $\langle 2D_{nn}\rangle$ including the effects of TPA is provided in the SI.
In all examples presented in the main text, we consider linear gain neglecting saturation effects, $G(n,N)=G(N)=G_N(N-N_{\mathrm{trans}})$ with $N_{\mathrm{trans}}$ the transparency carrier density. We found no significant differences using logarithmic quantum well gain or including the effects of gain saturation \cite{agrawal1990effect}. Quantum well lasers may allow larger nonlinearities owing to smaller active region volumes as well as lower threshold currents, though their laser damage thresholds (in photon number) are also expected to be lower, potentially suppressing the maximum shifts in resonance frequency due to Kerr nonlinearity.
\subsection{Noise properties}
\label{sec:rin_exp}
The steady state noise properties of semiconductor lasers can be computed by considering operator valued fluctuations of the Heisenberg-Langevin equations from their mean solutions. In the steady state, this results in a pair of coupled linear equations for the operator values fluctuations $\delta n$ and $\delta N$, which are given as:
\[
\begin{bmatrix}
\delta \dot n \\
\delta\dot N
\end{bmatrix}
= M
\begin{bmatrix}
\delta n\\
\delta N
\end{bmatrix}
+
\begin{bmatrix}
F_n-n\kappa_{\dot\phi}F_\phi\\
F_N
\end{bmatrix}
\]
\noindent with the fluctuation matrix
\[
M =
\begin{bmatrix}
-n\kappa_n & n\left(G_N\left(1-\frac{\alpha_L}{2}\kappa_{\dot\phi} \right)-\kappa_N\right)\\
-G_0 & -(nG_N+\gamma_\parallel)
\end{bmatrix}
\]
\noindent Here, $\kappa_{\dot\phi}=\kappa_\omega\equiv d\kappa/d\omega$ denotes the sharpness of the dispersive loss, with the instantaneous laser frequency being given by $\omega(n,N)+\dot\phi$. As discussed in more detail in the SI, the terms involving $\kappa_{\dot\phi}$ can be neglected under the adiabatic approximation $\gamma \gg \mathrm{FSR}$ for the systems considered in the main text.
By solving this linear system in the frequency domain, one can construct the frequency resolved power spectrum of fluctuations for either of the operators. Of particular interest is the photon number variance, given by $(\Delta n)^2=\frac{1}{\pi}\int_0^\infty \langle \delta n^\dagger(\omega) \delta n(\omega) \rangle$. The fluctuation spectrum inside the integral is related to the frequently defined relative intensity noise (RIN) as
\begin{equation}
\mathrm{RIN}=\frac{1}{n^2}\langle \delta n^\dagger(\omega) \delta n(\omega) \rangle,
\label{eq:rin}
\end{equation}
A useful parameter to quantify the quantum nature of light is the Fano factor, defined as $F=(\Delta n)^2/n$. The Fano factor is 1 for Poissonian light, corresponding to the shot noise limit; values below one indicate sub-Poissonian light below the show noise level. We calculate the most general expression for $F$ in the presence of nonlinear dispersive loss in the SI.
Here, we summarize the results corresponding to a few important limiting cases where the effect of amplitude-phase coupling on intensity noise may be neglected. This holds when the sharpness of the dispersive loss satisfies $\kappa_{\dot\phi}=\kappa_\omega\ll 1$, which is true under the adiabatic approximation $\gamma\gg\mathrm{FSR}$ (systems considered in the main text). For completeness, we include the effect of two-photon absorption (TPA) when considering carrier nonlinearities (two-photon carrier generation is negligible compared to the pump current). However, for most systems, the magnitude of dispersive loss is at least an order of magnitude greater than TPA-induced loss; the effects of TPA are only seen at the weakest nonlinearities and lowest background losses. A more thorough discussion of TPA is provided in the SI. Important physical parameters to characterize intensity noise are the relaxation oscillation frequency and damping rate of relaxation oscillations. As derived in the SI, these are respectively given by
\begin{align*}
\Omega_R^2 &\approx (nG_N+\gamma_\parallel)(n\kappa_n)+n(G_N-\kappa_N)\kappa \\
\Gamma_1 &\approx n(G_N+\kappa_n)+\gamma_\parallel
\end{align*}
\noindent These measures provide an important way to understand the effect of nonlinear dispersive loss on intensity noise. The limiting Fano factors are thus:
\begin{enumerate}
\item For weak Kerr and carrier nonlinearities, $\kappa_n,\kappa_N\rightarrow 0$, we have $F\rightarrow 1+\kappa/(nG_N)$ when pumping far above threshold, recovering linear behavior. When $n$ becomes large far about threshold, the Fano factor approaches 1, resulting in Poissonian (coherent) statistics.
\item For strong Kerr nonlinearity but weak carrier nonlinearity, $n|\kappa_n| \gg \kappa_0,n|\kappa_N|,\gamma_\parallel,G_N$, the Fano factor $F\rightarrow \kappa/|n\kappa_n|$ for large $n$. Critically, the ratio $|\kappa_n|/\kappa$ is a measure of how sharply the loss varies with $n$ compared to the absolute loss rate at the steady state photon number, and thus dictates the dimensionless ``sharpness'' of the loss. The Fano factor is inversely proportional to this sharpness factor, and thus sharp losses lead to sub-Poissonian states.
\item For strong carrier nonlinearity but weak Kerr nonlinearity, $n|\kappa_N| \gg \kappa_0,n|\kappa_n|,\gamma_\parallel$, we have $F\rightarrow \kappa/(nG_N)+G_N/|G_N-\kappa_N|\rightarrow G_N/|G_N-\kappa_N|$ for large $n$. The carrier nonlinearity reduces dependence of the rate of change of intensity fluctuations on carrier fluctuations ($G_N\rightarrow G_N-\kappa_N$), lowering the relaxation oscillation frequency $\Omega_R^2$ while leaving the damping of these oscillations unchanged. This can amplify low-frequency intensity noise slightly.
\item For simultaneously strong Kerr and carrier nonlinearities, $n|\kappa_{n,N}| \gg \kappa_0,\gamma_\parallel$,
\begin{equation}
\begin{split}
F&\rightarrow \frac{\kappa}{n|G_N+\kappa_n|}\left(1 + \frac{nG_N^2}{n\kappa_n G_N+|G_N-\kappa_N|\kappa}\right).
\end{split}
\label{eq:ffactor}
\end{equation}
Roughly, this expression can be broken into Kerr nonlinearity (first term) and carrier nonlinearity (second term) contributions. The former describes squeezing via increased $\Omega_R^2$ and damping of relaxation oscillations due to ``sharp'' intensity dependent loss, while the latter reduces intensity noise-carrier noise coupling and thus $\Omega_R^2$. Kerr and carrier nonlinearities may therefore have competing effects, leading to interesting steady state and noise fluctuation behavior.
\end{enumerate}
\subsection{Setup parameters}
Unless otherwise noted, we use the following parameters for all setups, chosen to be similar to experimentally determined parameters for GaAs around $\lambda\approx 1$ micron: modal volume and active region volume $V=10^{-16}$ m\textsuperscript{3} (confinement factor $\Gamma=1$), bare cavity resonance frequency $\omega_0=2.16\times 10^{15}$ rad/s (873 nm), free spectral range $\mathrm{FSR}=17$ GHz, transparency density $N_{\mathrm{trans}}=2\times 10^{24}$ m\textsuperscript{-3}, nonradiative decay rate $\gamma_\parallel=3\times 10^8$ rad/s, and linear gain coefficient $G_N=1/V\cdot dG/dN=3694$ rad/s \cite{chow2012semiconductor}.
For sharp losses in the main text, we consider symmetric Fano loss profiles, with central frequency given by $\omega(n_c=10^6,N_c=N_{\mathrm{trans}})$ and resonance decay (FWHM) $\gamma=2\times 10^{12}$ rad/s.
\section{Results}
\label{sec:results}
\subsection{Noise condensation from nonlinear loss}
\label{sec:comparison}
We now discuss how this semiconductor laser architecture with sharp nonlinear loss can lead to new behaviors in both the mean field as well as the fluctuations, eventually leading to strongly noise-reduced states. We begin by considering systems in the absence of carrier nonlinearity. In Fig. \ref{fig:main}, we compare steady state and intensity noise plots for conventional linear loss and sharp (Lorentzian/symmetric Fano) loss. Linear loss presents an $n$-independent loss profile, and leads to the well known linear dependence of steady state photon number on pump current. In the presence of dispersive loss, moderate nonlinearity ($\beta = -10^{-10}$) begins to modify the steady state behavior. For pump currents just above threshold, the behavior is close to linear. However, as the pump current increases, so does the loss, pulling down the input-output curve to a sub-linear behavior. For even stronger nonlinearity ($\beta = -10^{-9}$), a bistable transition occurs that creates a range of photon numbers which have no stable steady state solution. In particular, this occurs because there is a nonzero photon number at which the cavity experiences minimum loss. This left bistable edge lies below threshold, and thus needs to be accessed "from above", by pumping to a high power (beyond the range of zero solutions), and slowly lowering the power (this is a hysteresis effect).
Beyond the point of bistability (roughly 30 mA in Fig. \ref{fig:main}b), $n$ stays approximately constant while the photon number variance $\Delta n^2$ can decrease sharply. We see the consequence of this in Fig. \ref{fig:main}c, which plots the Fano factor spectrum $n\cdot\mathrm{RIN}\equiv \Delta n^2(\omega)/n$. Here, the intensity noise fluctuations associated with relaxation oscillations are quenched closer to the bistable point. Due to the sharp loss, the relaxation oscillation peak is in general significantly suppressed compared to the case of linear loss. As a result of the bistability, the laser can exist in two states with very different photon numbers over a range of pump currents. The larger photon number branch corresponds to sharp loss ($\kappa_n>0$) in this scheme. Overall, nonlinear dispersive loss creates significant broadband intensity noise squeezing by orders of magnitude compared to analogous linear loss.
\begin{figure}[t]
\centering
\includegraphics[scale=0.8]{Fig1_221511.pdf}
\caption{Squeezing in semiconductor lasers with dispersive photonic loss and intensity dependent nonlinearity. (a) Intensity dependent loss profiles $\kappa(n)$ for three different (dimensionless) nonlinear strengths $\beta$ ($\beta=0$ corresponds to linear loss) coupled to symmetric Fano (Lorentzian) dissipation in the absence of carrier nonlinearity, as is a good approximation for an ``external cavity'' type geometry. (b) Steady state photon number $n$ as a function of pump current (S-curve) for three different linear background losses $\kappa_0$ expressed as a fraction of the free spectral range FSR. The indicated unstable region is bypassed by the bistable point and is not generally accessible during lasing. (c) Fano factor spectrum ($n\cdot\mathrm{RIN}\equiv \Delta n^2(\omega)/n$) as a function of noise frequency for three different pump powers $r\equiv I/I_\mathrm{thres}$. Here, $\kappa_0=\mathrm{FSR}$ for the linear loss (blue) and $\kappa_0=10^{-2}\cdot\mathrm{FSR}$ for the nonlinear loss (purple). (d) Fano factor as a function of pump current. Fano factors in (d) are plotted for the low noise (upper) branch in (b) when bistability is present. The Fano resonance decay rate is $\gamma=2\times 10^{12}$ rad/s.}
\label{fig:main}
\end{figure}
Finally, we calculate Fano factors as a function of pump current in Fig. \ref{fig:main}d by integrating the spectra in Fig. \ref{fig:main}c. For linear loss, as the intracavity intensity grows with pump current, the relaxation oscillations shift to higher noise frequencies but also undergo stronger damping, giving rise to a decaying Fano factor that approaches the shot noise limit far above threshold. Stronger linear losses amplify relaxation oscillations but do not affect the shot noise limit.
\begin{figure*}[t]
\centering
\includegraphics[scale=0.8]{carrier_with_squeezing.pdf}
\caption{Effects of carrier and Kerr nonlinearities composed with dispersive loss. (a) Schematically, in the presence of only carrier nonlinearity $\sigma$, the resonance frequency and thus loss depend ``directly'' on carrier density $N$. Thus, ``gain equals loss'' implies $G(N)=\kappa(N)$ (both plotted as a function of the ratio of mean carrier density to transparency density). For sufficiently strong $\sigma$ and low background loss $\kappa_0$, multiple steady state carrier densities $N$ can correspond to a given photon number $n$, resulting in different losses (detunings from the Fano resonance). The lowest loss solution (smallest detuning) is most likely to lase, though extra solutions may be accessible by dynamic pumping schemes. (b) The schematic effect of this ``carrier bistability'' is to create multiple branches in the S-curve of different slope/threshold current. Interesting features emerge when both carrier and Kerr nonlinearities are present, resulting in the behaviors shown in panels (c), (d), and (e). In particular, the carrier nonlinearity causes a deformation of the intensity-nonlinear Lorentzian loss profile, eventually pinching off the ``sharp loss'' from the linear loss for sufficiently strong carrier nonlinearity (purple curve). This stems from leftward motion of the carrier bistability boundaries and creates a demarcation between linear ($F\gg 1$) and nonlinear ($F<1$) loss regimes which may be separated by a region of lasing with no stable solution. System parameters used are the same as those in Fig. \ref{fig:main} with a nonlinear strength $\beta=-10^{-10}$, background loss $\kappa_0=10^{-2}\cdot\mathrm{FSR}$, and Fano resonance decay $\gamma=2\times 10^{12}$ rad/s. The magnitudes of Kerr and carrier nonlinearities taken here are comparable to what they might be in GaAs-based gain media: $\beta\sim -10^{-10}$ and $\sigma\sim -3\times 10^{-27}$ m\textsuperscript{3} (with the proviso of being taken as instantaneous and being evaluated at a single wavelength). Note also that the Kerr nonlinearity varies with mode volume (here, we consider a mode volume equal to the cavity volume).}
\label{fig:carrier}
\end{figure*}
The behavior of Fano factor for nonlinear dispersive loss is phenomenologically different. For simplicity, in Fig. \ref{fig:main}d, we only plot the sharp loss (upper) branch when bistability is present (purple curve) \footnote{The lower branch, accessible by normal pumping from threshold, resembles linear behavior apart from the bistable point, which creates a discontinuity in the Fano factor as a function of pump current.}. On the upper branch, linear behavior (shot noise) is restored as the detuning from the Fano resonance grows large: this corresponds to low reflectivity, high loss, and high pump currents. Approaching the left bistable edge, the cavity frequency approaches the Fano resonance and, for a certain $n$, the ratio $\kappa/(n\kappa_n)$ approaches a minimum, corresponding to minimum Fano factor. The Fano factor does not decrease indefinitely approaching the left bistable edge. Coupling of intensity and carrier density implies that even at zero loss, nonzero carrier noise and thus intensity noise exist due to the nonradiative decay $\gamma_\parallel$. This results in an increase of Fano factor at the left bistable edge. Nonetheless, nonlinear dispersive loss creates a range of pump currents over which intensity noise drops far below the shot noise limit. The effect is stronger for lower linear background losses and larger (Kerr) nonlinearities, with a maximum of 92\% (12 dB) reduction in Fano factor below the shot noise limit displayed in Fig. \ref{fig:main}d.
\subsection{Bistability from carrier nonlinearity}
\label{sec:carrier}
In this section, we describe how the interplay of carrier nonlinearity with dispersive loss can result in unexplored ``carrier bistability'' behavior. We consider carrier nonlinear strengths $\sigma$ comparable to what they might be in, for example, GaAs-based gain media \cite{boyd2020nonlinear, said1992determination}.
We first describe how multiple lasing steady states can exist when strong carrier nonlinearity and dispersive loss are simultaneously present. In an ordinary semiconductor laser, the ``gain equals loss'' requirement leads to a so-called ``gain clamping'' condition, wherein above threshold, the inverted carrier density is fixed at some value, regardless of the intensity (i.e. the carrier density $N$ such that $G_N(N-N_{\mathrm{trans}})=\kappa$). This is depicted in Fig. \ref{fig:carrier}a by the ``linear loss'' case which shows only a single intersection point of the carrier dependent gain and carrier independent loss. However, in the presence of strong carrier nonlinearity and sharply frequency dependent outcoupling (with a Fano mirror for example), the loss of the cavity mode can depend nonlinearly on the carrier density $N$, $\kappa(\omega(N))=\kappa(N)$. As the carrier density changes, so does the cavity frequency, and hence the damping rate via the frequency dependent mirror. The ``gain equals loss'' condition now reads $G_N(N-N_{\mathrm{trans}})=\kappa(N)$. As shown in Fig. \ref{fig:carrier}a, this leads to a situation where more than one carrier density $N$ can cause gain and loss to be equal, corresponding to multiple cavity resonance frequencies. In the case of the Fano resonance, we see that up to three different steady states are possible.
Fig. \ref{fig:carrier}b shows how this phenomenon manifests in the steady state laser behavior. The dependence of steady state intensity on the pump current is still linear, but there can be up to three independent branches, corresponding to different steady state $N$ and different lasing frequencies. For the Fano mirror example, one resonance frequency is always present such that the detuning from the Fano resonance $\Delta\approx 0$. Since this solution has lowest loss, and thus the lowest threshold, lasing will occur here by default. The other branches are also stable, but disconnected from the lowest branch. It may be possible to experimentally access these higher branches through dynamic pumping schemes which generate transients that can travel from one branch to another.
When Kerr nonlinearity is also introduced, additional phenomena appear due to the simultaneous nonlinear dependence of the damping rate on intensity and carrier number. It is important to note that the profile $\kappa(\omega)$ is unchanged, though $\kappa(n,N)$ will vary based on the nonlinear strengths. Furthermore, the gain (and thus loss) is monotonically increasing in $N$. For typical materials (and for the results presented in Fig. \ref{fig:carrier}), $\sigma<0$ increases the resonance frequency $\omega(n,N)$ and thus pushes lasing solutions rightward along $\kappa(\omega)$.
Consider first weak carrier nonlinearity (orange and green curves in Fig. \ref{fig:carrier}c). Then, the carrier nonlinearity can be treated as a perturbation to the initially symmetric Lorentzian loss $\kappa(n)$. On the $\kappa_\omega\equiv d\kappa/d\omega>0$ (right) branch of the dispersive loss, the carrier nonlinearity shifts the loss curve upward. On the other hand, the $\kappa_\omega<0$ (left) branch shifts downwards since an increase in $\omega$ corresponds to a point of lower loss (approaching detuning $\Delta=0$).
Suppose we increase the carrier nonlinearity further (red curve in Fig. \ref{fig:carrier}c). For $n$ near threshold, far below the ``magic'' photon number $n_c\approx 10^6$ of lowest loss, the carrier nonlinearity pushes solutions rightward along the Lorentzian. However, the laser still lies on the $\kappa_\omega<0$ branch - the carrier nonlinearity pushes the mode closer to $n_c$, which is near zero loss and thus $N\approx N_{\mathrm{trans}}$. This yields one steady state solution. For higher $n$, near but still below $n_c$, we eventually reach an $n$ at which two solutions are possible: $N\approx N_{\mathrm{trans}}$ (lower loss) or $N > N_{\mathrm{trans}}$ (higher loss). Immediately afterwards, a third solution is possible with still higher loss/higher carrier density, phenomenologically similar to the dashed curve in Fig. \ref{fig:carrier}a. Finally, as $\kappa_\omega$ drops past the inflection point of $\kappa(\omega)$, a point corresponding to two solutions marks the end of the carrier bistability and for the largest $n$ we again obtain only one solution (the Lorentzian loss looks approximately linear).
For even stronger carrier nonlinearity (purple curve in Fig. \ref{fig:carrier}c), the carrier bistability boundaries shift leftward in photon number. Comparing the red and purple curves in Fig. \ref{fig:carrier}c, the left boundary eventually crosses zero and becomes negative, at which point the loss curve detaches into two parts: a sharp part at low loss and linear part at higher loss, separated by a range of pump currents over which no stable lasing solution occurs. When the right bistability boundary also crosses $n=0$ the sharp loss vanishes and laser operation only occurs on the linear high-loss branch (with correspondingly larger threshold currents), as shown for the brown curve in Fig. \ref{fig:carrier}c.
We now examine the effects of this carrier bistability on intensity noise. As shown in Fig. \ref{fig:carrier}e, the minimum achievable Fano factor is relatively independent of the level of carrier nonlinearity. This can be seen by noting that the first term in Eq. \ref{eq:ffactor} dominates the Fano factor at these points. However, past the sharp loss region, the linear branch created by the carrier nonlinearity possesses a higher loss that pulls the Fano factor upward for larger pump currents. For large carrier nonlinearities, the system eventually hits bistability and a region of unstable lasing, transitioning to (approximately) linear behavior again. For carrier nonlinearities much stronger than the Kerr nonlinearity (brown curve), approximately linear loss is restored as described above and no intensity noise reduction is observed. Mathematically, Eq. \ref{eq:ffactor} essentially contains a combination of dominant Kerr and dominant carrier nonlinearity terms, demarcated by pump currents smaller and larger than the Fano factor minimum/sharp loss regime, respectively.
\section{Discussion}
\label{sec:discussion}
\subsection{Experimental platforms}
In this section, we describe some of the experimental platforms for realizing the effects described here. Because semiconductor platforms are conducive to integration with on-chip photonic crystal optical elements, many designs have already achieved the dispersive losses considered here and therefore could exhibit intensity noise reduction if quality factors and nonlinear strengths are within the tolerances required. For example, previous work has realized ``Fano lasers'' that exhibit self-pulsing due to the interplay between dispersive loss and carrier nonlinearity \cite{yu2017demonstration}. A Fano resonance is created by coupling between a waveguide and nanocavity (point defect) in a photonic crystal slab. By increasing the quality factor of the lasing waveguide mode in these types of structures, intensity noise reduction by nonlinear dispersive loss could be observable.
Photonic crystal surface-emitting lasers (PCSELs) are another platform that may be used for demonstrating the effects of nonlinear dispersive loss \cite{noda2017photonic}. PCSELs may present advantages such as single-mode operation and high output powers; in contrast to the Fano laser concept, lasing occurs transversely (and thus the Fano mirror is aligned transversely rather than longitudinally). However, because losses may be significant in both longitudinal and transverse directions, it is necessary to optimize quality factors in both directions. Nevertheless, advances in fabrication have achieved $Q > 5\times 10^6$ in photonic crystal nanocavities, well within the background losses considered here \cite{asano2017photonic}.
In addition to Fano-type dispersive losses, distributed feedback-based losses have been commonly exploited to enforce single-mode operation. Examples include distributed Bragg reflector (DBR) fiber lasers, vertical cavity surface-emitting lasers (VCSELs), and DBR diode lasers \cite{carroll1998distributed, iga2018forty}. All of these architectures include sharply frequency dependent elements that may be used to achieve strong noise condensation. High quality fabrication is necessary to minimize background losses (e.g., scattering) at interfaces in order to observe the intensity noise condensation described here.
Lastly, we note that even stronger nonlinearities are achievable in systems such as microdisk and quantum dot lasers due to enhanced confinement and ultralow mode volumes \cite{srinivasan2006cavity}. For mode volumes achieving $\lambda^3$, the dimensionless Kerr coefficient can be orders of magnitude larger than the values considered here.
However, damage threshold photon numbers may also be lower, limiting the maximum shifts to refractive index and resonance frequency from Kerr nonlinearity. Nevertheless, these systems can exhibit ultrahigh quality factors, and are thus of potential interest for studying the effects presented here.
\subsection{Outlook}
In this paper, we have shown how semiconductor lasers with sharply frequency dependent outcoupling and Kerr nonlinearity could be used to create lasers which produce high levels of intensity squeezing in the laser cavity. Additionally, we have demonstrated how strong carrier nonlinearities in these systems can lead to a new type of ``carrier bistability'' which leads to multiple possible laser steady states. We anticipate that many existing experimental platforms could realize the intensity noise condensation and bistable effects described here, especially systems employing a geometry that maximizes photonic (Kerr) nonlinearity over carrier nonlinearities, as in the external cavity-like geometry of Fig. \ref{fig:schema} (with the gain being a semiconductor gain chip).
This work naturally suggests additionally possibilities for using nonlinear disperisve dissipation to control the output state of semiconductor lasers. Examples of topics for additional investigation include the effect of nonlinear dispersive loss on phase noise and linewidth, as well as the effects of optical feedback on intensity and phase noise (e.g., in external cavity lasers). Semiconductor lasers are ubiquitous in many real-world applications and we envision that the use of nonlinear dispersive loss could render them important sources of low noise states for use in sensing and quantum computing applications.
\section{Acknowledgements} S.P. acknowledges the financial support of the Undergraduate Research Opportunities Program. J.S. is supported in part by a Mathworks Fellowship, and acknowledges previous support from a National Defense Science and Engineering Graduate (NDSEG) Fellowship (F-1730184536). N.R. acknowledges the support of a Junior Fellowship from the Harvard Society of Fellows, as well as earlier support from a Computational Science Graduate Fellowship of the Department of Energy (DE-FG02-97ER25308), and a Dean's Fellowship from the MIT School of Science. This material is based upon work supported in part by the Air Force Office of Scientific Research under the award number FA9550-20-1-0115; the work is also supported in part by the U. S. Army Research Office through the Institute for Soldier Nanotechnologies at MIT, under Collaborative Agreement Number W911NF-18-2-0048. We also acknowledge support of Parviz Tayebati.
\bibliographystyle{unsrt}
|
\section{Introduction}
\label{sec:intro}
\textit{Discriminative classification} and \textit{generative classification} are two different approaches to solving classification problems. In discriminative classification, we directly model the posterior distribution of the class labels $p({k}|\bm{x})$. This posterior distribution is typically obtained by applying the \textit{softmax} function to the logit outputs of the classifier.
\begin{equation}
p({k}|{\bm{x}}) = \frac{\exp (a_k)}{\sum_{j=1}^{K} \exp (a_j)},
\end{equation}
where $a_k$ is the model's $k$-th output.
Because the discriminative classifier makes use of the softmax function, it is also known as\textit{ softmax classifier}. The discriminative classifier is in fact modeling the decision boundary between different classes.
An alternative approach is to first model the class-conditional distributions $p(\bm{x}|{k})$ and then obtain the posterior distribution using Bayes' theorem
\begin{equation}
p({k}|{\bm{x}}) = \frac{p({\bm{x}}|{k})p({k})}{p({\bm{x}})},
\end{equation}
where the prior distribution $p({k})$ can be estimated from the fractions of the training set data points in each of the classes. This is the generative approach to solving classification problems. It seems that in order to solve the classification problem, the generative classifier needs to solve a more complex problem of modeling the class-conditional distributions. Indeed, the class-conditional distribution may contain a lot of structure that has little effect on $p({k}|{\bm{x}})$. But directly modeling $p(\bm{x}|{k})$ can be advantageous when we want to use $p(\bm{x}|{k})$ to detect out-of-distribution inputs.
Deep neural network-based discriminative classifier is the dominant approach to solving classification problems. A widely known issue of this approach is the existence of adversarial examples.
Various defense mechanisms have been proposed to address adversarial examples, with \textit{Adversarial training (AT)} being the most successful one. AT improves the classifier's robustness by training the classifier against adversarially perturbed inputs. One issue with AT is that adversarially robust classifiers tend to have lower accuracy on clean data than standard, non-robust classifiers. \cite{tsipras2018robustness} conjecture that the reduced accuracy on clean data is the consequence of robust classifiers learning fundamentally different features than standard classifiers.
Recent work on AT has been focus on using data augmentation (combined with model weight averaging~\cite{izmailov2018averaging})~\cite{rebuffi2021fixing}, synthetic training data produced by generative models~\cite{rebuffi2021fixing,gowal2021improving}, and high capacity models~\cite{alayrac2019labels, xie2019intriguing, gowal2020uncovering} to improve the standard and robust accuracies.
Discriminative classifier also has the issue of producing overconfident predictions on out-of-distribution inputs.
\cite{augustin2020adversarial} propose to address this issue by combining in- and out-distribution adversarial training to enforce low confidence on adversarial out-of-distribution samples.
Compared to discriminative robust classification, generative robust classification is a less-explored area.
Robust generative classification requires the density model $p_{\bm{\theta}}({\bm{x}}|{k})$ to have low likelihood outputs on regular as well as adversarial out-of-distribution samples, which is not achievable with existing density models.
In this work we explore modeling $p({\bm{x}}|{k})$ with the AT generative model proposed in \cite{yin2022learning}, and then using these conditional models to perform robust generative classification. We demonstrate that this robust generative classifier achieves comparable standard and robust classification accuracies to state-of-the-art softmax robust classifiers, and at the same time is more interpretable.
One issue with the proposed generative classifier is that it does not easily scale to problems with many classes. However, we note there are many practical problems with limited number of classes~\cite{Dua2019}, and our approach may find applications in this kind of problems.
We note that generative (robust) classification has also been investigated in \cite{Yin2020GAT}. In terms of the formulation of the generative classifier, both \cite{Yin2020GAT} and our work use AT to train a binary classifier to model the unnormalized density function of the class-conditional data. The key difference is that we do not assume that the normalizing constants of the density functions of different classes to be equal, and instead treat the normalizing constants as learnable parameters of the generative classifier. In terms of implementation, \cite{Yin2020GAT}'s generative classifier is ten times larger than the softmax robust classifier in terms of capacity~\cite{madry2017towards}, and yet its performance is much worse~\cite{tramer2020adaptive}.
In summary, our specific contributions are:
\begin{itemize}
\item We show that when the overall model capacities are similar, the generative classifier achieve comparable standard and robust accuracies to a baseline softmax robust classifier. In particular, the generative classifier is much more robust when the test perturbation is large.
\item We study the interpretability of the generative classifier and show that the generative classifier is more interpretable in terms of the qualities of generated samples and counterfactuals.
\item We propose to combine in- and out-distribution adversarial training to improve the model robustness.
\item We perform comprehensive ablation to study how calibration, model capacity, weight decay regularization, training perturbation size, data augmentation, and combining in- and out-distribution AT affect the model performance.
\end{itemize}
\section{Generative Robust Classification}
In a $K$ class classification problem, the proposed generative classifier consists of $K$ binary classifiers, with the $k$-th binary classifier trained to distinguish clean data of class $k$ from adversarially perturbed out-of-distribution data.
Following~\cite{yin2022learning}, denote the data distribution of class $k$ as $p_{k}$, the out-of-distribution dataset as $p_0$, then the $k$-th binary classifier $D_k:\mathcal{X} \subseteq \mathbb{R}^d \rightarrow [0,1]$ is trained by maximizing the following objective:
\begin{equation}
\label{eq:gat-obj-d}
\begin{split}
J({D_k}) = ~&\mathbb{E}_{{\bm{x}} \sim p_{k}}[\log {D_k}({\bm{x}})] \\ & + \mathbb{E}_{{\bm{x}} \sim p_{0}}[\min_{\bm{x}'\in\mathbb{B}(\bm{x},\epsilon)}\log (1 - {D_k}(\bm{x}')))],
\end{split}
\end{equation}
where $\mathbb{B}(\bm{x},\epsilon)$ is a neighborhood of $\bm{x}$: $\mathbb{B}(\bm{x},\epsilon)=\{\bm{x}' \in \mathcal{X}: \| \bm{x}'-\bm{x}\|_2 \leq \epsilon \}$.
$D_k$ is defined as $D_k(\bm{x}) = \sigma(d_k(\bm{x}))$, where $\sigma$ is the logistic sigmoid function, and $d_k:\mathbb{R}^d\rightarrow \mathbb{R}$ is a neural network with a single output node.
As is discussed in \cite{yin2022learning}, training with a larger perturbation $\epsilon$ causes the model to learn to generate more realistic and diverse samples of $p_k$.
Unfortunately, there is typically a trade-off between $D_k$'s discriminative capability and its generative capability. In this work we focus on optimizing the model's (standard and robust) discriminative performance by choosing an approximate $\epsilon$ to train the model. \cite{yin2022learning} also suggests that the $p_0$ should be diverse in order for $D_k$ to better capture $p_k$. However, in order to have a fair comparison with the discriminative classifier which does not use additional data, we use the the mixture distribution of other classes as the out-of-distribution dataset: $p_0=p_{\backslash k} =\frac{1}{K-1}\sum_{i=1,...,K, i\neq k} p_i$.
Following \cite{yin2022learning}, we interpret $d_k(\bm{x})$ as an unnormalized density model of $p(\bm{x}|k)$.
We can obtain the normalized density function by
\begin{equation}p(\bm{x}| k) = \frac{\exp(d_k(\bm{x}))}{Z_k},
\end{equation} with $Z_k$ being the partition function:
\begin{equation}
Z_k = \int\exp(d_k(\bm{x}))d\bm{x}.
\end{equation}
We can then apply the Bayes classification rule to obtain a \textit{generative classifier} $g(\bm{x}): \mathbb{R}^d \to \{1, ..., K\}$:
\begin{equation}
\label{eq:generative_classifier}
\begin{split}
g(\bm{x}) & = \argmax_k p(k|\bm{x}) \\
& = \argmax_k \frac{p(\bm{x}|k)p(k)}{p(\bm{x})} \\
& = \argmax_k \frac{\exp (d_k(\bm{x}))p(k)}{ Z_k},
\end{split}
\end{equation}
\subsection{Calibration}
\label{sec:calibration}
In general, $Z_k$ is intractable, and $Z_k$ of different classes are not equal, so $g(\bm{x})$ is also intractable.
While $Z_k$ cannot be computed directly, we can treat $Z_k$ as learnable parameters of the generative classifier, and then obtain their values by optimizing the generative classifier's performance on a validation set.
Absorbing $Z_k$ into the $ \exp $ and assuming that $p(k)$ of different classes are equal, then the generative classifier can be simplified as
\begin{equation}
\label{eq:generative_classifier_logit}
g(\bm{x})= \argmax_k d_k(\bm{x}) + c_k,
\end{equation}
where the calibration constants $\{c_1,...,c_K\}$ can be learned on a validation set.
\subsection{Combining in- and out-distribution adversarial training}
The robustness of the generative classifier can be further improved by combining in- and out-distribution adversarial training.
Consider an adversarially perturbed sample $ \bm{x}'=\bm{x}+\bm{\delta} $, where $\bm{x}$ is a clean sample of class $k$, and $ \bm{\delta} $ is an adversarial perturbation intended to cause misclassification.
In order to correctly classify $ \bm{x}' $, the generative classifier \cref{eq:generative_classifier_logit} needs satisfy
\begin{equation}\label{eq:inadv-req}
\forall i \in \{1, ..., K\} \backslash \{k\} \quad d_k(\bm{x}+\bm{\delta}) +c_k > d_i(\bm{x}+\bm{\delta})+c_i.
\end{equation}
However, in \cref{eq:gat-obj-d}, $d_k$ is only trained to have high outputs on clean samples of $p_k$, not perturbed samples of $p_k$. We can increase $d_k$'s outputs on perturbed samples of $p_k$ by also performing in-distribution adversarial training:
\begin{equation}
\label{eq:gat-obj-inclass}
\begin{split}
J({D_k}) =~ &\mathbb{E}_{\mathrm{\bm{x}} \sim p_{k}}[ \min_{\bm{x}'\in\mathbb{B}(\bm{x},\epsilon_1)}\log {D_k}({\bm{x}'})] \\& + \mathbb{E}_{\mathrm{\bm{x}} \sim p_{\backslash k}}[\min_{\bm{x}'\in\mathbb{B}(\bm{x},\epsilon_2)}\log (1 - {D_k}(\bm{x}')))].
\end{split}
\end{equation}
By explicitly training $d_k$ to have high outputs on \textit{adversarially perturbed samples} of class $k$, we make \cref{eq:inadv-req} easier to satisfy.
In \cref{sec:generative_classification_ablation} we provide an ablation on combining in- and out-distribution adversarial training.
\section{Experiments}
\subsection{Training setup}
We evaluate the generative classifier on the CIFAR-10 dataset (ten classes problem), a widely used benchmark for robust classification~\cite{croce2020robustbench}.
The proposed generative classifier consists of $K$ binary classifiers in a $K$ class classification problem. In order to have fair comparison with the discriminative approach which typically uses a single model, we limit the capacity of the binary classifier so that the overall capacities of these two approaches are similar.
Specifically, we use a customized model ``ResNet18Thinner'' (see also in \cref{fig:capacity_ablation}) to train the binary classifiers. The ResNet18Thinner model is an ResNet18 architecture with a width multiplier of 0.5, and has 10M parameters. The baseline discriminative robust classifier~\cite{robustness} is based on the ResNet50 architecture which has 90MB parameters.
All the binary classifier are trained with the SGD optimizer using a batch size of 128 and weight decay $10^{-4}$ for 2000 epochs. For classes 2, 3, 4, 5, 7, 8, 9 we use a starting learning rate of 0.1, and for classes 0, 1, 6 we find the 0.1 learning rate being too high and instead use a starting learning rate of 0.05. The learning rate is reduced to 0.01 after epoch 1500. The training perturbation size is 0.3 ($L_2$ norm), in contrast to the perturbation size of 0.5 used in training the softmax robust classifier. Following the common practice in the adversarial machine learning literature~\cite{rice2020overfitting}, we use early stopping on the test set to select the model.
\subsection{Evaluation}
We evaluate the generative classifier and softmax robust classifier on the clean test set (clean accuracy) and the adversarially perturbed test set (robust accuracy). The adversarial perturbations are computed by performing untargeted adversarial attack against the classifiers. Given a test sample $\bm{x}\in \mathbb{R}^d$ and its label $y$, for the softmax robust classifier $f$, the adversarial perturbation $\bm{\delta}^*$ is computed by solving
\begin{equation}
\bm{\delta}^* = \argmax_{\bm{\delta}\in \mathbb{R}^d, \|\bm{\delta}\|_p \le\epsilon } L (f(\bm{x}+\bm{\delta}), y),
\end{equation}
where $L$ is the cross-entropy loss. We solve this optimization using the PGD attack~\cite{madry2017towards}. For the generative classifier, it has been shown that directly attacking the cross-entropy loss is suboptimal~\cite{tramer2020adaptive}.
We instead use the attack proposed by \cite{tramer2020adaptive}. The attack first compute
\begin{equation}
\bm{\delta}_k^* = \argmax_{\bm{\delta}\in \mathbb{R}^d, \|\bm{\delta}\|_p \le\epsilon }d_k(\bm{x}+\bm{\delta})
\end{equation}
for $k = 1, ..., K, k\neq y$ and then compute $\bm{\delta}^*$ by
\begin{equation}
\bm{\delta}^* = \argmax_{k = 1, ..., K, k\neq y} d_k(\bm{x}+\bm{\delta}_k^*).
\end{equation}
\subsection{Results}
\subsubsection{Standard accuracy and robust accuracy}
\cref{tab:cifar10_acc} shows that the generative classifier has lower standard accuracy but higher robust accuracy compared to the softmax robust classifier.
To further investigate this phenomenon, we evaluate these two classifiers using the test set perturbed with different levels of adversarial noise. \cref{fig:cifar10_acc} shows that when there is no perturbation or when the test perturbation is small, the softmax robust classifier has better performance, and when the test perturbation is of moderate size or large size, the generative classifier outperforms the softmax robust classifier, and the larger the test perturbation, the larger the performance gap.
In adversarial machine learning, it has been observed that there is a trade-off between standard accuracy and robust accuracy~\cite{tsipras2018robustness} --- training with adversarially perturbed data allows the classifier to learn more robust and semantically meaningful features to achieve improved robust accuracy at the expense of the standard accuracy. ~\cite{robustness} further shows that larger training perturbations causes the model to perform better under large test perturbations but worse under small or no test perturbations.
As~\cref{fig:perturbation_ablation} suggests, this trade-off between standard performance and robust performance also exists when we use \cref{eq:gat-obj-d} to train the binary classifiers.
Note that the training perturbation of the generative classifier is already much smaller than that of the softmax robust classifier (0.3 vs. 0.5).
Based on these results, we conjecture that the generative classifier can learn more semantically meaningful features by capturing the class-conditional distributions and therefore has a better robust accuracy. Indeed, \cref{tab:fid_cifar10_gen}, \cref{fig:global_inte}, and \cref{fig:inter_counterfactuals} show that the generated samples of the generative classifier have better quality than that of the softmax robust classifier, suggesting that the generative classifier has better generative properties.
\begin{table}[h!]
\centering
\resizebox{\linewidth}{!}{%
\begin{tabular}{@{}lcc@{}}
\toprule
Model & Standard accuracy & Robust accuracy \\ \midrule
Generative classifier & 88.12\% & 72.27\% \\
Softmax robust classifier\cite{robustness} & 90.83\% & 70.17\% \\ \bottomrule
\end{tabular}%
}
\caption{CIFAR10 standard accuracy and robust accuracy ($\epsilon_\textrm{test}=0.5$, $L_2$ norm). }
\label{tab:cifar10_acc}
\end{table}
\begin{figure}[h!]
\centering
\begin{subfigure}[t]{0.49\linewidth}
\centering
\includegraphics[width=\linewidth]{figs/l2_sensitivity.pdf}
\end{subfigure}
\begin{subfigure}[t]{0.49\linewidth}
\centering
\includegraphics[width=\linewidth]{figs/linf_sensitivity.pdf}
\end{subfigure}
\caption{The performance of the generative classifier and the softmax robust classifier tested under different test perturbation sizes.}
\label{fig:cifar10_acc}
\end{figure}
\subsubsection{Interpretability}
We also study the interpretability the generative classifier. We use two approaches to evaluate the classifier's interpretability. First, we visualize the ``learned concepts'' of the classifier by generating synthetic samples that cause high-confidence predictions. This allows us to understand how the model make prediction in general and can be considered as a \textit{global interpretability} method~\cite{molnar2020interpretable}.
Second, we consider \textit{counterfactual explanation}~\cite{wachter2017counterfactual} which is also used by previous work~\cite{augustin2020adversarial} to evaluate the interpretability of adversarially robust classifiers.
To access the quality of generated samples (or counterfactuals), we compute the FID~\cite{heusel2017gans} between the training samples and the generated samples. FID is a distance metric that measures the similarity between two sets of data samples, and is widely used to evaluate the quality of images produced by generative models.
\noindent\textbf{Generated samples.} To generate samples that cause high-confidence predictions, we follow~\cite{santurkar2019image} and fit a multivariate normal distribution to the class-conditional data and then generate seed images by sampling from the this normal distribution. We then generate samples by performing targeted PGD attacks against the model with the seed images. We use a large perturbation limit so that the generated samples cause high-confidence predictions.
\cref{fig:global_inte} shows the seed images and generated images of the generative classifier and the robust discriminative classifier.
In both cases, the generated samples resemble the class-conditional data, suggesting the classifiers have captured the high-level features of the target classes. Qualitatively, generated samples of the generative classifier have less artifacts and the foreground objects are more recognizable.
To provide a quantitative assessment, we compute the FID of the generated samples. \cref{tab:fid_cifar10_gen} shows that the generated samples of the generative classifier has lower FID than that of the softmax robust classifier under different configurations of the PGD attack, suggesting that the generated samples of the generative classifier more closely resemble the training data.
We note that samples generated with different PGD attacks tend to have different FIDs, but samples produced by the generative classifier have consistently lower FIDs.
\noindent\textbf{Counterfactuals.} The idea of counterfactual explanation is to find the minimum change to an input sample such that the modified sample is classified to a predefined target class. To generate counterfactuals we take input samples and then perform targeted adversarial attacks against the classifier. For standard, non-robust models, this process would lead to adversarial examples~\cite{goodfellow2014explaining} which do not have visually meaningful changes. If on the other hand the classifier has captured the class-specific features, then the classifier's decisions change only when class-specific features appear in the attacked data.
\cref{fig:inter_counterfactuals} shows both the generative classifier and softmax robust classifier are able to generate plausible counterfactual explanations. The counterfactuals generated with the two approaches are qualitative similar in terms of the emergence of class specific features. We again use FID as a quantitative measure and \cref{tab:fid_cifar10_gen} shows that the counterfactuals of the generative classifier has better quality in terms of the FID score.
\begin{table}[h!]
\begin{center}
\resizebox{\linewidth}{!}{%
\begin{tabular}{@{}lcc@{}}
\toprule
& Generated samples & Counterfactuals \\ \midrule
\multicolumn{3}{l}{$L_2$ PGD attack of steps 7 and step-size 1.0} \\ \midrule
Generative classifier & 50.38 & 37.92 \\
Softmax robust classifier~\cite{robustness} & 61.84 & 43.65 \\ \midrule
\multicolumn{3}{l}{$L_2$ PGD attack of steps 10 and step-size 1.0} \\ \midrule
Generative classifier & 54.78 & 42.33 \\
Softmax robust classifier~\cite{robustness} & 66.15 & 50.85 \\ \bottomrule
\end{tabular}%
}
\end{center}
\caption{FID of generated samples and counterfactuals of the generative classifier and softmax robust classifier under different configurations of the PGD attack}
\label{tab:fid_cifar10_gen}
\end{table}
\begin{figure}[h!]
\centering
\begin{subfigure}[t]{0.32\linewidth}
\centering
\includegraphics[width=\linewidth]{figs/interpretability/sampling_input2}
\caption{Seed images}
\end{subfigure}
\begin{subfigure}[t]{0.32\linewidth}
\centering
\includegraphics[width=\linewidth]{figs/interpretability/sampling_generative2}
\caption{Generated samples (generative classifier)}
\end{subfigure}
\begin{subfigure}[t]{0.32\linewidth}
\centering
\includegraphics[width=\linewidth]{figs/interpretability/sampling_softmaxrobust2}
\caption{Generated samples (softmax classifier)}
\end{subfigure}
\caption{Seed images and generated images generated by performing targeted adversarial attack on the generative classifier and the robust discriminative classifier. The PGD attack is $L_2$-based attack of steps 10 and step-size 1.0.}
\label{fig:global_inte}
\end{figure}
\begin{figure}[h!]
\centering
\begin{subfigure}[t]{0.32\linewidth}
\centering
\includegraphics[width=\linewidth]{figs/interpretability/counterfactual_input}
\caption{Seed images}
\end{subfigure}
\begin{subfigure}[t]{0.32\linewidth}
\centering
\includegraphics[width=\linewidth]{figs/interpretability/counterfactual_generative}
\caption{Counterfactuals for the generative classifier}
\end{subfigure}
\begin{subfigure}[t]{0.32\linewidth}
\centering
\includegraphics[width=\linewidth]{figs/interpretability/counterfactual_softmaxrobust}
\caption{Counterfactuals for the softmax robust classifier}
\end{subfigure}
\caption{Seed images and counterfactuals generated by performing targeted adversarial attack on the generative classifier and the robust discriminative classifier. The PGD attack is $L_2$-based attack of steps 10 and step-size 1.0.}
\label{fig:inter_counterfactuals}
\end{figure}
\subsection{Ablation study}
\label{sec:generative_classification_ablation}
In this section we study how calibration, model capacity, weight decay regularization, training perturbation size, data augmentation, and combining in- and out-distribution adversarial training affect the model performance.
To make the analysis more concrete, we study how the above factors affect the performance of individual binary classifiers. We consider both the normal and adversarial scenarios, where in the normal scenario we evaluate the binary classifier on the task of separating clean in-distribution test samples from out-distribution test samples, and in the adversarial scenario we evaluate the binary classifier on the task of separating adversarially perturbed in-distribution and out-distribution test samples. We use the AUC (area under the ROC curve) score as the performance metric.
\subsubsection{Calibration}
The calibration parameters $\{c_1,...,c_K\}$ are learned by minimizing the cross-entropy loss of the generative classifier \cref{eq:generative_classifier_logit} on the training set of CIFAR10.
We use early stopping on the test set to select $\{c_1,...,c_K\}$.
The binary classifiers turn out to be well-calibrated, and the calibration does not have much effect on the generative classifier's performance (\cref{tab:calibration}). \cref{tab:calibrated_vals} shows the values of the learned $\{c_1,...,c_K\}$ (the initial values of $\{c_1,...,c_K\}$ are set to zeros).
\cref{fig:logit_hist} shows the histogram of $ d_k $'s outputs on samples of class $ k $ and samples of other classes. It can be seen that the learned $\{c_1,...,c_K\}$ are tiny compared to the outputs of $ d_k $, so they do not have much effect on \cref{eq:generative_classifier_logit}.
We note that although the calibration does not have much effect on the evaluated dataset (CIFAR-10), it does not by itself exclude the possibility that for certain datasets the calibration can be helpful.
\begin{table}[h!]
\begin{center}
\resizebox{\linewidth}{!}{%
\begin{tabular}{@{}llc@{}}
\toprule
& Training accuracy & Test accuracy\\ \midrule
Generative classifier & 95.25\% & 88.12\% \\
Generative classifier after calibration & 95.43\% & 88.15\% \\ \bottomrule
\end{tabular}}
\caption{Accuracies on the training set and test set before and after calibration.}
\label{tab:calibration}
\end{center}
\end{table}
\begin{table}[h!]
\centering
\resizebox{\linewidth}{!}{%
\begin{tabular}{@{}llllllllll@{}}
\toprule
$c_1$ & $c_2$ & $c_3$ & $c_4$ & $c_5$ & $c_6$ & $c_7$ & $c_8$ & $c_9$ & $c_{10}$ \\ \midrule
-0.068& -0.015& 0.066& 0.067& -0.037& 0.066& -0.068& -0.067& -0.058& 0.062 \\ \bottomrule
\end{tabular}%
}
\caption{The values of the learned $\{c_1,...,c_K\}$.}
\label{tab:calibrated_vals}
\end{table}
\begin{figure}[h!]
\centering
\resizebox{0.8\linewidth}{!}{%
\centering
\includegraphics[width=\textwidth]{figs/logit_hist}
}
\caption{Histogram of $ d_k $'s outputs on samples of class $ k $ and samples of other classes, $ k=1,2,3 $. (Note that the x axes of the subplots are shared.)}
\label{fig:logit_hist}
\end{figure}
\subsubsection{Model capacity and regularization}
Consistent with the previous findings~\cite{madry2017towards, alayrac2019labels, xie2019intriguing, gowal2020uncovering}, we find models with higher capacity tend to have better robustness
(\cref{fig:capacity_ablation}). Note that overfitting happened even with ResNet18Thinner, the model with the lowest capacity. To mitigate overfitting, we apply weight decay and \cref{fig:capacity_ablation} shows that the weight decay help the model achieve better robustness.
\begin{figure}[h!]
\centering
\resizebox{\linewidth}{!}{%
\centering
\includegraphics[width=\textwidth]{figs/generative_classification/capacity_ablation}
}
\caption{Adversarial AUROC score of the class 0 models of different capacities. ``ResNet18Thin'' and ``ResNet18Thinner'' are two ResNet18 models respectively with a width multiplier of 0.75 and 0.5.}
\label{fig:capacity_ablation}
\end{figure}
\subsubsection{Training perturbation size.}
Similar to \cite{tsipras2018robustness}, in \cref{fig:perturbation_ablation} we observe a decline in standard performance as the training perturbation size increases. Meanwhile, training with a larger perturbation helps the model achieve better adversarial robustness.
\begin{figure}[h!]
\centering
\begin{subfigure}[t]{\linewidth}
\centering
\includegraphics[width=\linewidth]{figs/generative_classification/perturbation_abalation_c0.pdf}
\end{subfigure}
\caption{Standard and adversarial performances of the class 0 model when trained with different perturbation sizes.}
\label{fig:perturbation_ablation}
\end{figure}
\subsubsection{Data augmentation}
Data augmentation is a widely used regularization technique for reducing overfitting in learning standard classification models. Unfortunately, the use of data augmentation in training robust classifiers is not as successful. On CIFAR-10, beyond the widely adopted random padding and cropping augmentation strategy, none of the more sophisticated augmentation techniques are beneficial for improving model robustness~\cite{rice2020overfitting,gowal2020uncovering}.
\cite{rebuffi2021fixing} studies this phenomenon and find that when combined with model weight averaging~\cite{izmailov2018averaging}, heuristics-driven augmentation techniques such as Cutout~\cite{devries2017improved}, CutMix~\cite{yun2019cutmix} and
MixUp~\cite{zhang2017mixup} can help improve robustness. However, data-driven data augmentation approaches such as AutoAugment~\cite{cubuk2019autoaugment} have not been found to be helpful. In contrast to these work, we find it straightforward to apply AutoAugment in our training to reduce overfitting and obtain better robustness (\cref{fig:aug_ablation}).
\begin{figure}[h!]
\centering
\resizebox{\linewidth}{!}{%
\centering
\includegraphics[width=\textwidth]{figs/generative_classification/aug_ablation}
}
\caption{Adversarial AUROC score of the class 0 model when trained with different data augmentation policy.
}
\label{fig:aug_ablation}
\end{figure}
\subsubsection{Combining in- and out-distribution adversarial training}
\cref{fig:inout_ablation} shows the effect of combining in- and out-distribution adversarial training. It can be seen that combining in- and out-distribution AT helps the model achieve better adversarial performance at the expense of decreased standard performance.
\begin{figure}[h!]
\centering
\resizebox{\linewidth}{!}{%
\centering
\includegraphics[width=\textwidth]{figs/generative_classification/inout_ablation}
}
\caption{Clean and adversarial AUROC scores of the class 0 model when trained with combined AT and out-distribution AT only.
}
\label{fig:inout_ablation}
\end{figure}
\subsection{Conclusion}
We have investigated applying the AT generative model to learn class-conditional distributions and then using the conditional models to perform generative classification. Compared to the baseline softmax robust classifier, the generative classifier achieves comparable performance when there is no adversarial perturbation or the perturbation is small, and much better performance when the perturbation is of moderate or large sizes. The generated samples and counterfactuals produced by the generative classifier are also more similar to the training data (as measured by FID) than that of the softmax robust classifier, suggesting that the generative classifier has better captured the class-conditional distributions. We note that the above results are obtained when we use a much smaller model for the binary classifiers so that the generative classifier and softmax classifier have similar capacities and inference speeds. We further show how model capacity and training perturbation size affect the model performance in a similar way as they would in standard adversarial training. Compared to the softmax robust classifier, we find it much easier to apply advanced augmentation such as AutoAugment in our training. We leave the evaluation of other data augmentation such as CutMix and MixUp to future work. The analysis in \cite{yin2022learning} also suggests that using a diverse out-distribution dataset improve the generative modeling performance of the binary classifier, but it is unclear whether using auxiliary OOD data helps improve the binary classifier's discriminative capability. Of related work is \cite{augustin2020adversarial} which show that incorporating auxiliary OOD data helps the robust classifier achieves better standard and robust accuracies on CIFAR-10. Further, some recent work show that generated in-distribution data produced by some generative models can be leveraged to achieve better robustness \cite{rebuffi2021fixing,gowal2021improving}. We leave the investigation of using auxiliary OOD data and/or generated in-distribution data in our approach to future work.
{\small
\bibliographystyle{ieee_fullname}
|
\section{Introduction}\label{sec:introduction}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{problemsetting_5.pdf}
\vspace{-20pt}
\caption{\textbf{Comparisons between UDA, OSUDA and OSDG.} UDA assumes there are a large number of unlabeled target real images available. Nevertheless, OSUDA and our \emph{proposed} OSDG only access one-shot target (target-relevant) domain image. }
\vspace{-15pt}
\label{fig:problem_setting}
\end{figure}
Semantic segmentation, which aims at assigning a semantic class label to each pixel of a given image, serves as an important and fundamental task in the field of computer vision. Deep leaning based semantic segmentation models~\cite{long2015fully,chen2017deeplab,cheng2021mask2former} have achieved great success recently, and are playing an important role in a slew of applications, {\it e.g.}, robot vision and autonomous driving. Training these models typically relies on collecting and labeling a large amount of diverse real-world data, which is notoriously costly. In order to circumvent the problem, leveraging the power of simulation to automatically synthesize abundant labeled images serves as a promising way~\cite{richter2016playing,ros2016synthia,dosovitskiy2017carla}. However, the models trained on the simulated images often do not perform well on the real images, because of the distribution difference between the simulated and real images. To this end, the unsupervised sim-to-real domain adaptation (UDA) methods~\cite{Tsai_adaptseg_2018,tranheden2021dacs,zou2018unsupervised,hoffman2018cycada,vu2019advent} are developed recently to mitigate the domain gap between the well-labeled source domain, \emph{simulated images}, and the unlabeled target domain, \emph{real images}. Although much progress has been made, traditional UDA methods typically assume the access to a large number of real images, constituting the target domain. However, such an assumption does not always hold in real-world applications, since the collection of real-world images is not only costly but can also be very difficult due to the data scarcity, {\it e.g.}, rare adverse weather and diseases. Thus, more recent works~\cite{choi2021robustnet,kim2021wedge,pan2018IBN-Net,yue2019domain,peng2022semantic,tobin2017domain,prakash2019structured} start to explore the more challenging problems of domain generalization (DG) and the domain randomization (DR), where the target domain images are assumed unavailable during training. While some progress is made, the DG/DR for semantic segmentation performance still falls far behind the UDA performance, suffering from the inaccessible target domain knowledge and limiting the sim-to-real applications.
Considering both the difficulty in data collection and the benefits from the guidance of the target domain, in this paper, we investigate the one-shot scenario, {\it i.e.}, there is only one image available as the target domain, named as one-shot unsupervised sim-to-real domain adaptation (OSUDA). The setup aims to bridge the domain gap with only one-shot target guidance, which is generally available in real-world applications. Existing methods on OSUDA~\cite{luo2020adversarial,wu2021style} typically fall into two main categories, 1) by exploiting auxiliary style images, {\it e.g.}, a separate set of images with diverse styles from ImageNet~\cite{deng2009imagenet}, to stylize the source domain images and search for the harder stylized samples around the one-shot target sample to improve the generality of the trained model, 2) by generating mixed-style images/ feature, {\it i.e.}, intermediate domain, between the source and target domain, and utilize the consistency regularization between them to improve the robustness of the model when adapting to the target domain.
Moreover, the domain gap between the source and target domain arises from the \emph{style} and \emph{spatial structure} differences~\cite{devaranjan2020meta,kar2019meta}. However, both the type 1) and 2) methods focus on reducing the domain gap on the style difference, but ignore the \emph{spatial structure} difference ({\it e.g.}, the spatial location of the bus in Fig.~\ref{fig:crossformer}\textcolor{red}{a}).
We propose to improve the generalization ability on the target domain by taking both style and spatial structure differences into account. As assumed by~\cite{luo2020adversarial,wu2021style}, we first assume that the one-shot target sample contains style information which does not change drastically within the target dataset.
To make full use of the one-shot target sample, we propose to firstly construct a \emph{pseudo-target domain} by stylizing the source domain images with the one-shot target domain image. This step bridges the style difference between the two domains. As the one-shot target sample does not necessarily provide the full picture of the spatial structure information in the entire target dataset, we propose to focus on the spatial structure information provided by the source domain. More specifically, we randomize the spatial structures from the source domain by class-mixed sampling strategy and synthesize a new domain called \emph{intermediate domain} . The intermediate domain contains randomized and augmented spatial structure information which can better generalized to the traget domain. We then design the \emph{class-aware cross-domain transformer} structure to facilitate the global alignment between the pseudo-target domain, source domain and intermediate domain. Compared to previous methods, our approach brings a few appealing benefits, (i) it does not require access to the large number of auxiliary data in type 1) methods, but still significantly outperforms the type 1) methods; (ii) it randomizes the intermediate domain to better mitigate the domain gap to the target domain on both the \emph{style} and \emph{spatial structure} level; (iii) it extracts the global domain-invariant knowledge among the source, intermediate and pseudo-target domain through the class-aware cross-domain transformer structure, leading to effective generalization to the target domain. Moreover, we extend the OSUDA task and relieve its constraint, by allowing the one-shot target image to be from a target-relevant domain instead of exactly the target domain, named as one-shot sim-to-real domain generalization (OSDG) problem. For example, in the GTA$\rightarrow$ Foggy Cityscapes setting, the proposed OSDG setup allows the one-shot target image, {\it i.e.}, a foggy image, to be collected by randomly downloading from the internet, instead of exactly from Foggy Cityscapes dataset, further reducing the data collection difficulty.
Remarkably, extensive experiments on different benchmarks demonstrate that our method significantly outperforms the previous state-of-the-art (SOTA) methods on OSUDA and OSDG by a large margin, 10.87, 9.59, 13.05 and 15.91 mIoU on GTA, SYNTHIA$\rightarrow$ Cityscapes, Foggy Cityscapes, respectively.
\section{Related Work}
\textbf{Sim-to-Real Domain Adaptation/ Randomization/ Generalization.} \emph{Previous UDA works}~\cite{hoffman2018cycada,Tsai_adaptseg_2018,tranheden2021dacs,wulfmeier2017addressing,palazzo2020domain,keser2021content,gogoll2020unsupervised,zhang2019vr,bousmalis2018using,fang2018multi,wulfmeier2018incremental,messikommer2022bridging,yun2021target,sakuma2021geometry} assume there are abundant unlabeled real images in the target domain, which might not be the case in practice due to the collection difficulty, {\it e.g.}, the rare adverse weather and diseases images. Instead, OSUDA and our proposed OSDG only require one-shot unlabeled image on the target/ target-relevant domain, which is more flexible and practical. From the method aspect, different from the \emph{semantic segmentation transformer based UDA method}~\cite{hoyer2022daformer} that only relies on the self-attention, our method proposes the class-aware cross-domain attention to further extract the domain-invariant knowledge. Besides, \emph{previous OSUDA/DR/DG methods}~\cite{luo2020adversarial,yue2019domain,kim2021wedge} typically need the access to auxiliary real images from other datasets, {\it e.g.}, ImageNet~\cite{deng2009imagenet} and WikiArt~\cite{huang2017arbitrary}, increasing the GPU and storage memory requirement. However, our method only need one-shot target/relevant image, without relying on any other auxiliary data.
\textbf{Vision Transformers.} Transformer is originally proposed to model the sequence-to-sequence data in the filed of Natural Language Processing (NLP)~\cite{vaswani2017attention}. Recently, transformer models and their variants have demonstrated their effectiveness for different computer vision tasks, {\it e.g.}, image classification~\cite{dosovitskiy2020vit,liu2021swin}, object detection~\cite{carion2020end,zhu2021deformable}, and semantic segmentation~\cite{zheng2021rethinking,strudel2021segmenter}. Among those works, the most related is~\cite{xu2022cdtrans} in terms of general methodology. The similarity is that, both \cite{xu2022cdtrans} and our method aim at improving the sim-to-real domain adaptation, with the help of cross-domain transformers. However, we have significant differences in the following aspects, 1) \cite{xu2022cdtrans} is designed for image classification, while we focus on semantic segmentation, yielding totally different framework structure and methods design; 2) The cross-domain attention in \cite{xu2022cdtrans} attends all image regions, while our class-aware cross-domain attention only attends regions of interest, {\it i.e.}, non-sampled class regions.
\section{Method}
In OSUDA, we are given the well-labeled source domain dataset $\cD_s = \{\x_i^s, \y_i^s\}_{i=1}^{N_s}$, where $\x_i^s\in \mathbb{R}^{H\times W\times 3}$ is the simulated RGB color image, and $\y_i^s\in \mathbb{R}^{H\times W}$ is the corresponding semantic label map. $N$ is the number of semantic classes. In the target domain, we are given the one-shot real image, $\cD_t = \{\x^t\}$. $\cD_s$ and $\cD_t$ are from different distributions, {\it i.e.}, $\x^s\sim P_s, \x^t\sim P_t$. The domain gap between $\cD_s$ and $\cD_t$, $P_s\neq P_t$, originates from the \emph{style} and \emph{spatial structure} difference between the simulated and real image. In OSDG, the one-shot target domain image is replaced with the one-shot target-relevant domain image, $\cD_r = \{\x^r\}$ (see Fig.~\ref{fig:problem_setting}), which is easy to be obtained. For example, the target-relevant domain image for Foggy Cityscapes dataset (target domain) can be easily obtained by, the user downloads one foggy image from the common search engine.
\textbf{Method Overview.} The main difference between the one-shot setup and the general UDA setup is the further limited information on target domain, therefore it is critical to make full use of the limited information. While a single unlabeled target sample does not provide a lot of information about the data distribution of the entire target dataset, we assume that its style information is particularly useful and does not change drastically within the target dataset. To mitigate the domain gap on the \emph{style} level, we firstly make use of the style information from the one-shot target sample, and construct the pseudo-target domain $\hat{\cD}_s = \{\hat{\x}_i^s\}_{i=1}^{N_s}$ (cf. Sec.~\ref{sec:pseudo_target}). Then, to compensate on the limited spatial structure information on the target dataset, we propose to randomize the source spatial structure to improve the generalization performance on spatial structure. To this end, we develop the class mixed sampling based intermediate domain randomization (IDR) training strategy to further reduce the domain gap on both \emph{style} level and \emph{spatial structure} level (cf. Sec.~\ref{sec:intermediate_dr}), where $\hat{\x}_i^s$, $\x_j^s$ and corresponding semantic label $\y_i^s$ and pseudo-label $\tilde{\y}_j^s$ are randomly mixed, respectively. Moreover, we propose the class-aware cross-domain transformers to extract the domain-invariant knowledge to facilitate the adaptation to the target domain (cf. Sec.~\ref{sec:mask_aware_trans}), by introducing the class-aware global cross-attention into the intermediate domain randomization training. The whole framework is shown in Fig.~\ref{fig:framework}.
\label{sec:train_strategy}\textbf{Training Strategy.} Following~\cite{tranheden2021dacs,olsson2021classmix}, we adopt the pseudo-label based self-training strategy, where the mean-teacher framework~\cite{tarvainen2017mean} (cf. Fig.~\ref{fig:framework}) is introduced. Both the teacher and the student model in the mean-teacher framework are semantic segmentation networks with the same structure. The teacher model, $\cF_{\theta^\prime}$, is used to output the final semantic prediction map. The student model, $\cF_{\theta}$, is used to backpropagate gradients and update weights $\theta$ based on all the relevant training loss. $\theta^\prime$ is an exponential moving average of $\theta$ throughout the optimization. Besides, the teacher model is also used to generate the pseudo-label $\tilde{\y}_j^s$ by feeding the source domain image sample $\x_j^s$, {\it i.e.}, $\tilde{\y}_j^s = \cF_{\theta^\prime}(\x_j^s)$. During the inference stage, the final semantic segmentation map is obtained by $\cF_{\theta^\prime}(\x^t_{ts})$, where $\x^t_{ts}$ is the testing image.
\begin{figure*}[t]
\centering
\includegraphics[width=\linewidth]{network_structure_2.pdf}
\vspace{-20pt}
\caption{\textbf{Framework Overview and Network Structure.} (a) The whole mean-teacher framework adopts the pseudo-label based self-training strategy, where the pseudo-target domain construction, class mixed sampling based IDR, and global cross attention modules are developed. (b) shows the semantic segmentation network architecture, {\it i.e.}, the teacher model $\cF_{\theta^{\prime}}$ and the student model $\cF_{\theta}$, which is composed of backbone, pixel decoder and transformer decoder.}
\label{fig:framework}
\vspace{-10pt}
\end{figure*}
\subsection{Pseudo-Target Domain for Style Alignment}\label{sec:pseudo_target}
In order to bridge the style difference between the source and the target domain, instead of directly aligning between the source and target domains, we propose to first construct a pseudo-target (PT) domain by augmenting the source with target style. For each source domain image ${\x}_i^s$, a pseudo-target sample $\hat{\x}_i^s$ will be generated according to the style of the one-shot target sample $\hat{\x}_i^s = \cS(\x_i^s|\x^t)$, with the guidance of the one-shot target image $\x^t$, wehre $\cS(\cdot|\x^t)$ represents the image translation mapping conditioned on the style of $\x^t$. In order to prevent the overfitting to the one-shot target image and preserve the content, for $\cS(\cdot|\x^t)$, we adopt an off-the-shelf image translation framework~\cite{huang2018multimodal} with the weighted perceptual loss~\cite{johnson2016perceptual} to generate the pseudo-target domain. By introducing the pseudo-target domain, we alleviate the style discrepancy between the domains and focus on the structural difference in other components.
With the generated pseudo-target domain, we employ the standard cross-entropy loss to learn the model, {\it i.e.}, $\cL_{pt} = \sum_i CE(\cF_\theta(\hat{\x}_i^s), \y_i^s)$.
\subsection{Class Mixed Intermediate Domain Randomization for Spatial Structural Generalization}\label{sec:intermediate_dr}
While the pseudo target domain mitigates style differences, spatial structure differences can still exist between the domains. In Fig.~\ref{fig:crossformer}\textcolor{red}{a}, we show an example of this phenomenon. Vehicles are often in an open field and wide roads in the source GTA domain, but are surrounded by crowded city street buildings in the Cityscapes target domain. City planning differences between virtual US cities and real European cities lead to this significant spatial structural difference. The target domain contains scene layouts that were never seen before in the source/ pseudo-target domain during training, therefore model performance can largely decrease if we do not mitigate the spatial structure differences.
To reduce the spatial structural domain gap between the source and unseen target images, we propose to randomize the spatial structure between source and pseudo-target using class-mixed sampling. To this end, we propose a class mixed sampling strategy between the source and pseudo-target domain (cf. Fig.~\ref{fig:crossformer}\textcolor{red}{a}). The sampling strategy randomizes the layouts between the two domains and aims at improving the generalization performance on the unseen target images. This is largely motivated by existing works on domain randomization~\cite{yue2019domain}. Instead of randomizing the textures of the objects, we randomly copy a class in an pseudo-target image to the source to create a new spatial layout.
\textbf{Class Mixed Sampling for IDR (CIDR).} More formally, given the pseudo-target domain image $\hat{\x}_i^s$, the corresponding semantic label map $\y_i^s$, and the source domain image $\x_j^s$, the mixed sampling mask $\m^s$ is defined as,
\begin{eqnarray}
\m^s(h, w) = \begin{cases}
1, \text{if } \y_i^s(h, w) = \c, \\
0, \text{otherwise},
\end{cases} \label{eq:mask}
\end{eqnarray}
where $(h, w)$ represents the (row, column) index, and $\c$ is the sampling class, which is randomly taken from the available classes in $\y_i^s$. Following~\cite{olsson2021classmix,tranheden2021dacs}, half of the available classes in $\y_i^s$ are randomly selected in each training iteration. Then, the intermediate domain sample $\x_i^m$ and the corresponding semantic label map $\y_i^m$ are written as,
\begin{eqnarray}
\x_i^m = \m^s \odot \hat{\x}_i^s + (1-\m^s) \odot \x_j^s, \\
\y_i^m = \m^s \odot \y_i^s + (1-\m^s) \odot \tilde{\y}_j^s,
\end{eqnarray}
where the pseudo-label $\tilde{\y}_j^s$ corresponding to $\x_j^s$, instead of the ground truth label $\y_j^s$, is used to prevent the model from overfitting to the source domain. Then the intermediate domain randomization training loss is written as, $\cL_{idr} = CE(\cF_\theta(\x_i^m), \y_i^m)$.
\begin{figure*}
\centering
\includegraphics[width=\linewidth]{cross_atten_intermediate_6.pdf}
\vspace{-20pt}
\caption{\textbf{Class Mixed Sampling based Intermediate Domain and Cross-Domain Transformers.} (a) (i) shows that the domain gap originates from both the \emph{style} ({\it e.g.}, the color and texture are different) and \emph{spatial structure} level ({\it e.g.}, the bus is always located between the crowded city street buildings in the target domain instead of the open field far from buildings in the source domain). (ii) indicates that the style based intermediate domain methods only focus on the \emph{style} differences, {\it e.g.}, the bus texture is changed, however, is still located in open fields far from buildings as in source domain. (iii) proves that our mixed sampling based intermediate domain can mitigate the domain gap on both the \emph{style} and \emph{spatial structure} level, {\it e.g.}, the bus is located in the crowded city street buildings in the intermediate domain image, similar to the scene in the target domain image. (b) describes the class-aware cross-domain attention mechanism (cf. purple dashed box) in transformer decoder $\cT$.}
\vspace{-10pt}
\label{fig:crossformer}
\end{figure*}
\subsection{Class-Aware Cross-Domain Transformers for Domain Invariant Features}\label{sec:mask_aware_trans}
Both the style alignment and the spatial structural alignment are based on the input space. To perform well in the target domain, it is also critical to learn a feature extractor that can provide domain-invariant knowledge.
Existing methods focus on employing the local consistency, {\it e.g.}, pixel-wise L1 loss in~\cite{yue2019domain} and the local patch-wise prototypical matching in~\cite{wu2021style} between different domains to learn domain-invariant knowledge.
However, as noticed in recent works~\cite{xie2021segformer,strudel2021segmenter}, the local knowledge mining operations are biased towards the local interactions, {\it e.g.}, the boundary and shape of road and building, leading to insufficient processing of global image context and sub-optimal adaptation. In contrast, without the built-in inductive prior, the transformers by design can capture the global context and interactions between different scene elements, {\it e.g.}, the scene structure layout of road, sidewalk and person. To this end, we propose to incorporate the class-aware cross-domain transformers into the IDR to extract the \emph{global} domain-invariant knowledge between the intermediate domain image and the pseudo-target domain image.
\textbf{Cross-Domain Transformers.} Following~\cite{cheng2021mask2former,cheng2021maskformer}, our semantic segmentation network $\cF_\theta, \cF_{\theta^\prime}$ is typically composed of the backbone $\cB$, pixel decoder $\cP$ and transformer decoder $\cT$ (cf. Fig.~\ref{fig:framework}). The backbone is used to extract the image feature $\f_{img}$. The pixel decoder gradually up-scales and maps the image feature to the per-pixel embedding, $\e_{pixel}\in \mathbb{R}^{C_\e\times H \times W}$ (cf. Fig.~\ref{fig:framework}\textcolor{red}{b}). The transformer decoder $\cT$ attends to the image feature $\f_{img}$ (cf. Fig.~\ref{fig:framework}\textcolor{red}{b}) and the learnable class query features $\f_{q}$ (cf. Fig.~\ref{fig:framework}\textcolor{red}{b}) to generate the per-class embedding $\e_{class}\in \mathbb{R}^{C_\e\times N}$ of dimension $C_\e$ (cf. Fig.~\ref{fig:framework}\textcolor{red}{b}). The final prediction of $\cF_\theta, \cF_{\theta^\prime}$ is produced by the product $\e_{class}^{T}\times \e_{pixel}$ (cf. Fig.~\ref{fig:framework}\textcolor{red}{b}). The standard self-attention in transformer decoder $\cT$ computes,
\begin{eqnarray}
Attention(\Q, \K, \V) = softmax(\Q\K^T)\V,
\end{eqnarray}
where $\Q, \K, \V$ are the query, key and value vectors linearly projected from the input of the self-attention module (cf. Fig.~\ref{fig:crossformer}\textcolor{red}{b}). On this basis, the cross-domain attention is derived by incorporating the query vectors $\Q_{pt}$ from the pseudo-target domain image $\hat{\x}_i^s$ and the key, value vectors $\K_{m}, \V_{m}$ from the intermediate domain $\x_i^m$ into the self-attention module (cf. Fig.~\ref{fig:crossformer}\textcolor{red}{b}), formulated as,
\begin{eqnarray}
CrossAttention(\Q_{pt}, \K_{m}, \V_{m}) =& \nonumber\\ &\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!softmax(\Q_{pt}\K_{m}^T)\V_{m}.\label{eq:cross_atten}
\end{eqnarray}
\textbf{Class-Aware Cross-Domain Transformers.} The area corresponding to the sampling class $\c$ (cf. Eq.~\eqref{eq:mask} and `road, bus' in Fig.~\ref{fig:crossformer}\textcolor{red}{a}) in the intermediate domain image is taken from pseudo-target domain image, sharing the same style and spatial structure knowledge. To promote the domain-invariant knowledge excavation, based on the cross-attention in Eq.~\eqref{eq:cross_atten}, the class-aware cross-domain attention (CACDA) is developed (cf. Fig.~\ref{fig:crossformer}\textcolor{red}{b}) to attend in non-sampled classes regions (cf. `sky, tree, building' in Fig.~\ref{fig:crossformer}\textcolor{red}{a}) where style and scene structure are different, written as,
\begin{eqnarray}
ClassCrossAtten(\Q_{pt}, \K_{m}, \V_{m}, \M_{c}) = \nonumber\\ softmax(\M_{c} + \Q_{pt}\K_{m}^T)\V_{m}, \label{eq:class_cross_attn}
\end{eqnarray}
where $\M_{c}\in \mathbb{R}^{N\times N}$ is the class-modulating matrix to tailor cross-attention operation, written as,
\begin{eqnarray}
\M_{c} (x, y) = \begin{cases}
0, \text{ if } x \neq \c \text{ and } y \neq \c, \\
-\infty, \text{ otherwise}.
\end{cases}
\end{eqnarray}
Then, integrating the cross-domain transformers or the class-aware cross-domain transformers into the intermediate domain randomization training leads to the training loss, $\cL_{cd} = CE(\cF_\theta(\x_i^m|\hat{\x}_i^s), \y_i^m)$. $\cF_\theta(\x_i^m|\hat{\x}_i^s)$ represents that the transformer decoder $\cT$ in $\cF_\theta$ utilizes the cross-attention mechanism in Eq.~\eqref{eq:cross_atten} or the class-aware cross-attention mechanism in Eq.~\eqref{eq:class_cross_attn}.
\subsection{Joint Training}
With the above pseudo-target domain, class mixed sampling based IDR, class-aware cross-domain transformers and the corresponding training losses, the total loss is,
\begin{eqnarray}
\cL_{total} = \cL_{pt} + \cL_{idr} + \lambda\cL_{cd},
\end{eqnarray}
where $\lambda$ is the hyper-parameter to balance the cross-domain transformer loss and other loss terms, which is set as 0.01 in our work. With the pseudo-label based self-training strategy in Sec.~\ref{sec:train_strategy}, the model $\cF_{\theta}, \cF_{\theta^\prime}$ are trained end-to-end with the loss $\cL_{total}$. Our proposed method for OSUDA is easy to be applied to OSDG, by replacing the one-shot target image $\x^t$ with the target-relevant image $\x^r$.
\section{Experiments}
In order to construct the pseudo-target domain, we stylize the source images with the one-shot target image using the MUNIT method~\cite{huang2018multimodal} and the frequency based Fourier Transform~\cite{yang2020fda} for OSUDA and OSDG, respectively. The perceptual loss weight in~\cite{huang2018multimodal} is set as 2.0. $\beta$ in~\cite{yang2020fda} is set as 0.05. The involved datasets description on, Cityscapes~\cite{cordts2016cityscapes}, GTA~\cite{richter2016playing}, SYNTHIA~\cite{ros2016synthia}, Foggy Cityscapes~\cite{sakaridis2018semantic,sakaridis2018model}, and training details are put in the supplementary.
\subsection{One-Shot Unsupervised Domain Adaptation}
\begin{table*}[t]
\centering
\resizebox{\linewidth}{!}{
\begin{tabular}{c|c|cccccccccccccc|cc}
\toprule
Method
& \rotatebox{60}{Source} & \rotatebox{60}{AdaNet~\cite{Tsai_adaptseg_2018}} & \rotatebox{60}{CLAN~\cite{luo2019taking}} & \rotatebox{60}{AENT~\cite{vu2019advent}} & \rotatebox{60}{CBST~\cite{zou2018unsupervised}} & \rotatebox{60}{CGAN~\cite{CycleGAN2017}} & \rotatebox{60}{OST~\cite{benaim2018one}} & \rotatebox{60}{FSDR~\cite{huang2021fsdr}} & \rotatebox{60}{DRPC\cite{yue2019domain}} & \rotatebox{60}{SADG\cite{peng2022semantic}} & \rotatebox{60}{RoNet~\cite{choi2021robustnet}} & \rotatebox{60}{WDGE~\cite{kim2021wedge}} & \rotatebox{60}{IBN~\cite{pan2018IBN-Net}} & \rotatebox{60}{ASM~\cite{luo2020adversarial}} & \rotatebox{60}{SPPM\cite{wu2021style}} & \rotatebox{60}{Ours(R)} & \rotatebox{60}{Ours(M)} \\
\midrule
Num$^\#$ & 0 & 1 & 1 & 1& 1& 1 & 1 & $\geq$15 & $\geq$15 & 0 & 0 & 1000 & 0 & $\geq$10 & 1 & 1 & 1 \\
\midrule
\multicolumn{18}{c}{GTA$\rightarrow$ Cityscapes} \\
\midrule
mIoU & 36.6 & 35.2 & 37.7 & 36.1 & 37.1 & 39.6 & 42.3 & 44.8 & 42.5 & 45.33 & 42.87 & 43.60 & 37.42 & 44.5 & 42.8 & \textbf{49.46} & \textbf{55.37} \\
\midrule
\multicolumn{18}{c}{SYNTHIA$\rightarrow$ Cityscapes} \\
\midrule
mIoU$^*$ & 33.65 & 39.1 & 40.4 & 39.9 & 38.5 & 42.1 & 42.8 & 47.3 & - & - & - & - & - & 40.7 & 47.3 & \textbf{51.72} & \textbf{57.80} \\
mIoU & 29.45 & - & - & - & - & - & - & 40.8 & 37.6 & 40.87 & 37.21 & 40.31 & 34.18 & 34.6 & 41.4 & \textbf{44.99} & \textbf{50.99} \\
\bottomrule
\end{tabular}
}
\caption{\textbf{OSUDA Performance.} Num$^{\#}$ is the number of \textbf{real} images used for training, which are from the target domain or other auxiliary datasets, {\it e.g.}, ImageNet. Baselines and ours (R) adopt ResNet-101 backbone. Ours (M) utilizes MiT-B5 backbone. mIoU$^*$ represents the 13 classes mIoU performance when removing the ``wall, fence, pole" classes, as the common practice in~\cite{Tsai_adaptseg_2018,tranheden2021dacs,vu2019advent}.}
\vspace{-10pt}
\label{tab:OSUDA}
\end{table*}
\textbf{Comparison to other SOTA methods.} To demonstrate the effectiveness of our method for OSUDA, we compare our methods with other SOTA methods on the popular benchmarks, GTA, SYNTHIA$\rightarrow$Cityscapes, respectively. The baseline methods include 1) the OSUDA methods, ASM~\cite{luo2020adversarial} and SPPM~\cite{wu2021style}; 2) the UDA methods, AdaNet~\cite{Tsai_adaptseg_2018}, CLAN~\cite{luo2019taking}, AENT~\cite{vu2019advent}, CBST~\cite{zou2018unsupervised}, CGAN~\cite{CycleGAN2017} and OST~\cite{benaim2018one}; 3) the DG/DR methods, FSDR~\cite{huang2021fsdr}, DRPC~\cite{yue2019domain}, RoNet~\cite{choi2021robustnet}, IBN~\cite{pan2018IBN-Net}, WDGE~\cite{kim2021wedge} and SADG~\cite{peng2022semantic}. From Table~\ref{tab:OSUDA}, it is shown that our proposed method outperforms the previous SOTA OSUDA methods with the same ResNet-101~\cite{he2016deep} backbone, 49.46\% {\it vs. } 44.5\%, 42.8\% for GTA, and 44.99\% {\it vs. } 34.6\%, 41.4\% for SYNTHIA. By replacing the ResNet-101 with the transformer based backbone MiT-B5~\cite{xie2021segformer}, the performance can be further improved to 55.37\% and 50.99\%, benefiting from the stronger knowledge extraction ability of transformers compared to convolutional neural networks (CNNs). Besides, the previous OSUDA method, ASM~\cite{luo2020adversarial}, and DG/DR methods, FSDR~\cite{huang2021fsdr}, WEDGE~\cite{kim2021wedge} and DRPC~\cite{yue2019domain}, typically utilize a large number of auxiliary real images ({\it e.g.}, $\geq$1000 for WEDGE) of diverse styles from ImageNet~\cite{deng2009imagenet} and other datasets for training to improve the sim-to-real domain generalization ability. Nevertheless, our method does not require these auxiliary real images, while still strongly outperforming the aforementioned methods. In Fig.~\ref{fig:idr_quali}, we show the qualitative comparison results of our proposed method and other methods.
\textbf{Pseudo-target domain construction.} In order to prove the effectiveness of our pseudo-target domain construction for transferring the target domain style and preserving the content of the source domain, we show the qualitative pseudo-target domain construction results in Fig.~\ref{fig:pt_example}. It is shown that our proposed framework is flexible, and different types of pseudo-target domain construction methods, learning/non-learning based methods, can be used.
\emph{Learning based approach:} In order to construct the pseudo-target domain, for OSUDA experiments, we stylize the images with the one-shot target image by using the deep learning based MUNIT method~\cite{huang2018multimodal}, setting the perceptual loss weight $w_p$ as 2.0. As shown in Fig.~\ref{fig:pt_example}\textcolor{red}{a}, by adding the higher weight $w_p$ to the perceptual loss, the example guided image translation method can translate the source-domain image to the pseudo-target domain image, without overfitting to the one-shot target image. In other words, the style of the target image is applied to the source domain image, while the content from the source domain image is preserved. It provides a simple yet effective strategy for the one-shot/ few-shot image translation, {\it i.e.}, adding higher weighted perceptual loss. \emph{Non-learning based approach:} Moreover, in order to prove the flexibility of our proposed whole framework for using different ways of pseudo-target domain construction, we adopt the frequency based Fourier Transform method~\cite{yang2020fda} for the OSDG experiments, which is non-learning based approach. The advantage of Fourier Transform is that it does not require the training process of the deep learning based approach ({\it e.g.}, the image translation method in Fig.~\ref{fig:pt_example}\textcolor{red}{a}), and is easy to be used and applied. The disadvantage is that it introduces more artifacts than the deep learning based image translation method, which is also shown in \cite{yang2020fda}. However, as shown in Fig.~\ref{fig:pt_example}\textcolor{red}{b}, though the artifacts exist, the style of the target relevant image is still transferred to the source domain and the content in the source domain image is still preserved. The OSDG experiments in Table~\ref{tab:OSDG} prove that, our framework can effectively realize the one-shot sim-to-real domain generalization with the Fourier Transform based image, verifying our framework is compatible with different learning/ non-learning based pseudo-target domain construction methods.
\begin{figure*}
\centering
\includegraphics[width=0.9\linewidth]{generated.pdf}
\vspace{-15pt}
\caption{\textbf{Pseudo-Target Domain Construction}. (a) is the deep learning based image translation method. (b) is the non-learning based Fourier Transform method. It proves the effectiveness and flexibility of our framework for utilizing different types of pseudo-target domain construction methods, learning/non-learning based approach.}
\label{fig:pt_example}
\vspace{-5pt}
\end{figure*}
\begin{table}
\centering
\resizebox{0.8\linewidth}{!}{
\begin{tabular}{c|cc|cc}
\toprule
\multirow{2}{*}{Setting}& \multicolumn{2}{c}{Res101} & \multicolumn{2}{c}{Mit-B5}\\
\cmidrule{2-5}
&DAFormer & Ours & DAFormer & Ours\\
\midrule
\midrule
G$\rightarrow$C & 48.53 & \textbf{49.46} &54.01 & \textbf{55.37}\\
S$\rightarrow$C & 43.79 & \textbf{44.99} & 48.90 & \textbf{50.99} \\
\midrule
G$\rightarrow$FC & 38.57 & \textbf{39.05} & 47.43 & \textbf{49.04}\\
S$\rightarrow$FC & 37.75 & \textbf{41.09} & 43.36 & \textbf{44.28}\\
\bottomrule
\end{tabular}
}
\vspace{-5pt}
\caption{\textbf{Comparison to DAFormer~\cite{hoyer2022daformer} for OSUDA and OSDG.} \cite{hoyer2022daformer} is combined with PT and CIDR for fair comparison.}
\label{tab:daformer_comp}
\vspace{-15pt}
\end{table}
\begin{table}
\centering
\resizebox{0.8\linewidth}{!}{
\begin{tabular}{cccc|c}
\toprule
PT & CIDR & CrossAtten & Class-Aware & mIoU \\
\midrule
\midrule
&&&&42.59\\
\checkmark& & & & 46.12\\
\checkmark& \checkmark & & & 53.45\\
\checkmark& \checkmark & \checkmark & & 54.62 \\
\checkmark& \checkmark & \checkmark & \checkmark & \textbf{55.37}\\
\bottomrule
\end{tabular}
}
\caption{\textbf{Ablation Study.} Experiments are conducted under OSUDA setting, GTA $\rightarrow$ Cityscapes, with the backbone as MiT-B5.}
\label{tab:ablation}
\vspace{-10pt}
\end{table}
\begin{table*}
\centering
\resizebox{\linewidth}{!}{
\begin{tabular}{c|c|cccccccc|cc}
\toprule
Method
& Source & BDL~\cite{li2019bidirectional} & FDA~\cite{yang2020fda} & GASF\cite{Kundu_2021_ICCV} & BDL~\cite{li2019bidirectional} & CBST~\cite{zou2018unsupervised} & MLSL~\cite{iqbal2020mlsl} & FAdapt~\cite{iqbal2022fogadapt} & ASM~\cite{luo2020adversarial} & Ours(R) & Ours(M) \\
\midrule
Num$^\#$ & 0 & 0 & 0 & 0 & 2975 & 2975 & 2975 & 2975 & 0 & 0 & 0 \\
\midrule
\multicolumn{12}{c}{GTA$\rightarrow$Foggy Cityscapes} \\
\midrule
mIoU & 33.4 & 30.3 & 35.3 & 38.3 & 36.3 & 37.7 & 39.1 & 41.0 & 35.99 & \textbf{39.05} & \textbf{49.04} \\
\midrule
\multicolumn{12}{c}{SYNTHIA$\rightarrow$Foggy Cityscapes} \\
\midrule
mIoU$^*$ & 24.4 & - & - & - & 38.1 & 38.4 & 40.8 & 41.4 & 33.12 & \textbf{47.28} & \textbf{50.17} \\
mIoU & 20.9 & - & - & - & 32.3 & 33.3 & 35.9 & 36.4 & 28.37 & \textbf{41.09} & \textbf{44.28} \\
\bottomrule
\end{tabular}
}
\caption{\textbf{OSDG Performance.} Num$^{\#}$ represents the number of \textbf{Foggy Cityscapes} (target domain) images used for training. Baselines and ours (R) adopt ResNet-101 backbone. Ours (M) utilizes MiT-B5 backbone. mIoU$^*$ represents the 13 classes mIoU performance when removing the ``wall, fence, pole" classes, as the common practice in~\cite{Tsai_adaptseg_2018,tranheden2021dacs,vu2019advent}.}
\label{tab:OSDG}
\vspace{-5pt}
\end{table*}
\begin{figure*}
\centering
\includegraphics[width=\linewidth]{combine_pt_quat.pdf}
\vspace{-20pt}
\caption{\textbf{Left:} Our class mixed sampling based IDR (46.36 mIoU) {\it vs. } other IDR methods, DLOW~\cite{gong2019dlow} (42.3 mIoU) and SMPPM~\cite{wu2021style} (42.8 mIoU), with ResNet-101 backbone under GTA$\rightarrow$Cityscapes. cIoU (solid line) is the each single class IoU performance, while the mIoU (dashed line) is the mean IoU performance over cIoU. \textbf{Right:} Qualitative semantic segmentation results comparison on the target domain between our method and the baseline method ASM~\cite{luo2020adversarial}, where GT is the ground truth semantic label map.}
\label{fig:idr_quali}
\vspace{-10pt}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=\linewidth]{cross_atten_supp.pdf}
\vspace{-20pt}
\caption{\textbf{Different Cross-Domain Transformer Variants.} (a) is the variant-ST, where the supervised training is conducted on the source domain and pseudo-target domain, and then the cross-attention is utilized between the source and pseudo-target domain. (b) is the vairiant-S, where the supervised training is conducted on the source domain, and then the cross-attention is utilized between the source and intermediate domain. (c) is our cross-domain transformer module structure in Sec.~\ref{sec:mask_aware_trans}, where the supervised training is conducted on the pseudo-target domain, and then the cross-attention is utilized between the pseudo-target domain and intermediate domain. $\cL_{src}$ is the supervised cross-entropy training loss on the source domain.}
\label{fig:cross_att_supp}
\vspace{-10pt}
\end{figure*}
\begin{table}
\centering
\begin{tabular}{c|ccc}
\toprule
Structure& Variant-ST & Variant-S & Ours\\
\midrule
mIoU & 52.54 & 54.05 & \textbf{55.37}\\
\bottomrule
\end{tabular}
\caption{\textbf{Different Cross-Domain Transformer Variants.} The experiments are conducted under GTA$\rightarrow$Cityscapes for OSUDA, with the MiT-B5 backbone.}
\vspace{-10pt}
\label{tab:cross_att_supp}
\end{table}
\textbf{Ablation Study.} We investigate the individual contributions of different modules towards the overall performance quantitatively in Table~\ref{tab:ablation}. The experiments are conducted under GTA$\rightarrow$Cityscapes, with the MiT-B5 backbone. It is shown that all modules, including pseudo-target domain, class mixed sampling based IDR, and class-aware cross-domain transformers, are able to contribute to the final performance and help boost the adaptation ability of the model. The pseudo-target domain improves the performance from 42.59\% to 46.12\%, proving the effectiveness of the \emph{style} knowledge from the one-shot target domain. The class mixed sampling based intermediate domain randomization further narrows the domain gap on \emph{style} and \emph{spatial structure}, reaching 53.45\%. Moreover, the class-aware cross-domain transformers extract the global domain-invariant knowledge, leading to 55.37\% performance and boosting other local knowledge extraction methods such as DRPC~\cite{yue2019domain} and FSDR~\cite{huang2021fsdr}. Besides, the class-aware cross-domain attention performs better than pure cross-domain attention, 55.37\% {\it vs. } 54.62\%. This is because the class-aware cross-domain attention attends to the non-sampled class regions, where the style and the spatial structure are different. Thus, it encourages and promotes the domain-invariant knowledge extraction. In Table~\ref{tab:daformer_comp}, we compare to the SOTA transformer based domain adaptation method, DAFormer~\cite{hoyer2022daformer}, and it is observed that our method outperforms DAFormer under various settings.
\textbf{Class Mixed Sampling based IDR {\it vs. } other IDR methods.} As shown in Fig.~\ref{fig:idr_quali}, the image-level and feature-level stylization based intermediate domain methods achieve similar performance, 42.3\% and 42.8\% under GTA$\rightarrow$Cityscapes. Nevertheless, our class mixed sampling based IDR method reaches 46.36\%. That is because, compared to the previous IDR methods such as DLOW~\cite{gong2019dlow,gong2021dlow} and SMPPM~\cite{wu2021style} that only focus on \emph{style}, our class mixed sampling based IDR can bridge the domain gap on both \emph{style} and \emph{spatial structure} level (cf. Fig.~\ref{fig:crossformer}\textcolor{red}{a}).
\textbf{Comparison to Cross-Domain Transformer Variants.} In Sec.~\ref{sec:mask_aware_trans}, we propose the class-aware cross-domain transformers, where the class-aware cross-domain attention is conducted between the pseudo-target domain and the intermediate domain. In order to demonstrate the effectiveness of our proposed cross-domain transformer module structure in Sec.~\ref{sec:mask_aware_trans}, we here compare it with other cross-domain variants, where different cross-attention strategies between source domain, intermediate domain and pseudo-target domain are conducted as shown in Fig.~\ref{fig:cross_att_supp}. In Table~\ref{tab:cross_att_supp}, we show the quantitative comparison results between our cross-domain transformer structure and other cross-domain transformer variants, variant-ST (Fig.~\ref{fig:cross_att_supp}\textcolor{red}{a}) and variant-S (Fig.~\ref{fig:cross_att_supp}\textcolor{red}{b}). It is observed that both our structure in Sec.~\ref{sec:mask_aware_trans} and variant-S outperform the variant-ST, 55.37\%, 54.05\% {\it vs. } 52.54\%. It proves the effectiveness and necessity of conducting the cross-domain attention on the intermediate domain, which bridges the source and the target domain and eases the domain-invariant knowledge extraction through the cross-domain attention. Besides, our structure in Sec.~\ref{sec:mask_aware_trans} performs better than the variant-S, 55.37\% {\it vs. } 54.05\%. It is because that the pseudo-target domain is closer to the target domain compared to the source domain, and eases the adaptation to the target domain, verifying the effectiveness of the pseudo-target domain construction.
\subsection{One-Shot Sim-to-Real Domain Generalization}
Similar to the experimental study on OSUDA, we compare our proposed method with other SOTA methods under OSDG setting in Table~\ref{tab:OSDG}. Remarkably, our proposed method not only performs better than the SOTA OSUDA method ASM, but also performs at par or even outperforms the SOTA UDA methods under GTA, SYNTHIA $\rightarrow$ Foggy Cityscapes. It is notable that the UDA methods utilize the whole 2975 target domain foggy cityscapes images for training, while our method only exploits one relevant foggy image from the website~\cite{sakaridis2018semantic}.
\section{Conclusion}
In this paper, we aim at addressing the one-shot unsupervised sim-to-real domain adaptation (OSUDA) and generalization (OSDG) problem. To this end, we propose a transformer based approach, where 1) the pseudo-target domain and class-mixed sampling based intermediate domain is proposed to bridge the domain gap on both the style and spatial structure level; 2) the class-aware cross-domain transformers further extract the domain-invariant knowledge, facilitating the adaptation to the target domain. Extensive experiments on different benchmarks prove the effectiveness of our proposed method for OSUDA and OSDG, improving the SOTA performance by a large margin.
{\small
\bibliographystyle{ieee_fullname}
\section*{Supplementary}
In this supplementary material, we provide the additional information for,
\begin{itemize}
\item[\textbf{S1}] discussion about limitations,
\item[\textbf{S2}] detailed datasets description involved in our experiments,
\item[\textbf{S3}] detailed implementation and training details of our proposed framework,
\item[\textbf{S4}] more experimental results and analysis.
\end{itemize}
\section{Limitations Analysis}
\noindent\textbf{Limitations.} Though our proposed method improves the OSUDA and OSDG performance by a large margin, it is still not saturated and yet to achieve the performance of fully supervised learning on the target domain. It is mainly constrained by the limited available knowledge on the target domain. One possible strategy to further improve the performance is to utilize other modal knowledge in the target domain, {\it e.g.} the text information.
More specifically, in OSDG problem, GTA$\rightarrow$ Foggy Cityscapes, the ``foggy" text can provide the guidance for target domain knowledge extraction and the adaptation to the target domain.
\section{Datasets Description}
\noindent\textbf{Cityscapes.} Cityscapes~\cite{cordts2016cityscapes} dataset is a real-world urban street scene dataset, which is collected from some European cities. Following~\cite{luo2020adversarial,wu2021style}, one randomly selected image from the training set is taken as the target domain for training in the one-shot unsupervised sim-to-real domain adaptation (OSUDA) setting, and 500 validation set images are used for testing. The original images are of resolution 2048$\times$1024, and are finely annotated with 19 semantic classes.
\noindent\textbf{GTA.} GTA~\cite{richter2016playing} dataset is a synthetic urban scene dataset, rendered from an open-world game engine. The scenes in GTAV are based on the Los Angeles city. It contains 24,966 images with the resolution of 1914$\times$1052, which are densely labeled with 19 semantic classes. The semantic annotations are compatible with Cityscapes.
\noindent\textbf{SYNTHIA.} SYNTHIA~\cite{ros2016synthia} dataset is a synthetic dataset rendered from a virtual city. It consists of 9,400 photo-realistic images with resolution 1280$\times$760, coming with 16-class pixel-level semantic annotations, which is a subset of Cityscapes.
\noindent\textbf{Foggy Cityscapes.} Foggy Cityscapes~\cite{sakaridis2018semantic,sakaridis2018model} derives from the Cityscapes dataset, by simulating the fog on the real Cityscapes images. Thus, Foggy Cityscapes inherits the semantic annotations of the real, clear counterparts from Cityscapes. Under the one-shot sim-to-real domain generalization (OSDG) setting, 500 validation set foggy cityscapes images are used for testing, and one randomly selected web foggy image from Foggy Driving~\cite{sakaridis2018semantic} is used as the relevant domain image.
\section{Framework Implementation and Training Details}
In order to construct the pseudo-target domain, 1) for OSUDA experiments, we stylize the images with MUNIT~\cite{huang2018multimodal}, setting the perceptual loss weight as 2.0, 2) for OSDG experiments, we adopt the frequency based Fourier Transform~\cite{yang2020fda}, with the $\beta$ as 0.05. For the semantic segmentation model training, in accordance with other DA methods~\cite{Tsai_adaptseg_2018,vu2019advent,hoyer2022daformer}, the Cityscapes, GTA images are resized to 1024$\times$512, 1280$\times$720, respectively. We adopt the AdamW~\cite{loshchilov2019decoupled} optimizer to train the model, with the learning rate as 6$\times$10$^{-5}$. Taking two random crops 512$\times$512 into each batch, the total training iteration is set as 80000. We implement the whole framework with PyTorch~\cite{NEURIPS2019_9015}.
\begin{table}[]
\centering
\resizebox{\linewidth}{!}{
\begin{tabular}{c|cc}
\toprule
Setting & Ground Truth $\y_j^s$ & Pseudo-Label $\tilde{\y}_j^s$ \\
\midrule
GTA$\rightarrow$Cityscapes & 53.22 & 55.37\\
SYNTHIA$\rightarrow$Cityscapes & 49.52 & 50.99\\
\bottomrule
\end{tabular}
}
\caption{\textbf{Ground Truth Label $\y_j^s$ {\it vs. } Pseudo-Label $\tilde{\y}_j^s$,} for intermediate domain randomization in Eq.~(\textcolor{red}{3}) of the main paper.}
\label{tab:gtvspseudo}
\end{table}
\section{Additional Experimental Results}
In Sec.~\textcolor{red}{4} of the main paper, the quantitative and qualitative experimental results on GTA, SYNTHIA$\rightarrow$Cityscapes benchmarks for OSUDA, and GTA, SYNTHIA$\rightarrow$Foggy Cityscapes benchmarks for OSDG are shown. Here we provide additional experimental results to further demonstrate the effectiveness of our proposed approach.
\noindent\textbf{Comparison to Other SOTA Methods.} In Table~\textcolor{red}{1} and Table \textcolor{red}{4} of the main paper, the comparisons to other state-of-the-art (SOTA) methods for OSUDA and OSDG are provided. Corresponding to Table~\textcolor{red}{1} and Table~\textcolor{red}{4}, we show the detailed per class IoU performance in Table~\ref{tab:gta_city}, Table~\ref{tab:synthia_city}, Table~\ref{tab:gta_fog} and Table~\ref{tab:synthia_fog}. It is shown that our proposed approach significantly outperforms other SOTA methods, further verifying the effectiveness of our method for both OSUDA and OSDG.
\noindent\textbf{Ground Truth Label $\y_j^s$ {\it vs. } Pseudo-Label $\tilde{\y}_j^s$.} In Eq.~(\textcolor{red}{3}) of the main paper, we utilize the pseudo-label $\tilde{\y}_j^s$ instead of the ground truth label $\y_j^s$ corresponding to $\x_j^{s}$ for the intermediate domain randomization, to prevent overfitting to the source domain. In order to prove the validity of using the pseudo-label $\tilde{\y}_j^s$, we conduct the experiments under the OSUDA benchmarks, GTA$\rightarrow$ Cityscapes, SYNTHIA$\rightarrow$ Cityscapes, respectively. As shown in Table~\ref{tab:gtvspseudo}, the strategy using the pseudo-label $\tilde{\y}_j^s$ outperforms the one using the ground truth label $\y_j^s$, 55.37\%, 50.99\% {\it vs. } 53.22\%, 49.52\%, demonstrating the effectiveness of Eq.~(\textcolor{red}{3}) in the main paper.
\begin{table*}[]
\centering%
\resizebox{\linewidth}{!}{%
\begin{tabular}{l|cccccccccccccccccccc|c}
\toprule
Method & Num$^{\#}$ & Road&SW&Build&Wall&Fence&Pole&TL&TS&Veg&Terrain&Sky&Person&Rider&Car&Truck&Bus&Train&MC&Bike& mIoU\\
\midrule
\midrule
Source only & 0 & 75.8 & 16.8 & 77.2 & 12.5 & 21.0 & 25.5 & 30.1 & 20.1 & 81.3 & 24.6 & 70.3 & 53.8 & 26.4 & 49.9 & 17.2 & 25.9 & 6.5 & 25.3 & 36.0 & 36.6 \\
\midrule
AdaptSegNet~\cite{Tsai_adaptseg_2018}& 1 &77.7 & 19.2 & 75.5 & 11.7 & 6.4 & 16.8 & 18.2 & 15.4 & 77.1 & 34.0 & 68.5 & 55.3 & 30.9 & 74.5 & 23.7 & 28.3 & 2.9 & 14.4 & 18.9 & 35.2 \\
CLAN~\cite{luo2019taking} & 1 & 77.1 & 22.7 & 78.6 & 17.0 & 14.8 & 20.5 & 23.8 & 12.0 & 80.2 & 39.5 & 74.3 & 56.6 & 25.2 & 78.1 & 29.3 & 31.2 & 0.0 & 19.4 & 16.7 & 37.7 \\
ADVENT~\cite{vu2019advent}&1 & 76.1 & 15.1 & 76.6 & 14.4 & 10.8 & 17.5 & 19.8 & 12.0 & 79.2 & 39.5 & 71.3 & 55.7 & 25.2 & 76.7 & 28.3 & 30.5 & 0.0 & 23.6 & 14.4 & 36.1\\
CBST~\cite{zou2018unsupervised}& 1& 76.1 & 22.2 & 73.5 & 13.8 & 18.8 & 19.1 & 20.7 & 18.6 & 79.5 & 41.3 & 74.8 & 57.4 & 19.9 & 78.7 & 21.3 & 28.5 & 0.0 & 28.0 & 13.2 & 37.1 \\
CycleGAN~\cite{CycleGAN2017}&1 & 80.3 & 23.8 & 76.7 & 17.3 & 18.2 & 18.1 & 21.3 & 17.5 & 81.5 & 40.1 & 74.0 & 56.2 & 38.3 & 77.1 & 30.3 & 27.6 & 1.7 & 30.0 & 22.2 & 39.6 \\
OST~\cite{benaim2018one}&1 & 84.3 & 27.6 & 80.9 & 24.1 & 23.4 & 26.7 & 23.2 & 19.4 & 80.2 & 42.0 & 80.7 & 59.2 & 20.3 & 84.1 & 35.1 & 39.6 & 1.0 & 29.1 & 23.2 & 42.3 \\
FSDR~\cite{huang2021fsdr}& $\geq$15 & 89.3 & 40.5 & 79.1 & 26.3 & 27.8 & 29.3 & 33.7 & 29.0 & 83.0 & 27.7 & 76.0 & 57.8 & 27.5 & 81.0 & 32.3 & 42.4 & 16.8 & 21.0 & 30.2 & 44.8 \\
DRPC\cite{yue2019domain}& $\geq$15 & - & - & - & - & - & - & - & - & - & - & - & - & - & - & - & - & - & - & - & 42.5\\
SADG\cite{peng2022semantic}&0 & - & - & - & - & - & - & - & - & - & - & - & - & - & - & - & - & - & - & - & 45.33\\
RobNet~\cite{choi2021robustnet} & 0 & - & - & - & - & - & - & - & - & - & - & - & - & - & - & - & - & - & - & - & 42.87\\
WEDGE~\cite{kim2021wedge} & 1000 & - & - & - & - & - & - & - & - & - & - & - & - & - & - & - & - & - & - & - & 43.60 \\
IBN~\cite{pan2018IBN-Net} & 0 & - & - & - & - & - & - & - & - & - & - & - & - & - & - & - & - & - & - & - & 37.42 \\
ASM~\cite{luo2020adversarial}& $\geq$10 & 86.2 & 35.2 & 81.4 & 24.2 & 25.5 & 31.5 & 31.5 & 21.9 & 82.9 & 30.5 & 80.1 & 57.3 & 22.9 & 85.3 & 43.7 & 44.9 & 0.0 & 26.5 & 34.9 & 44.5\\
SMPPM\cite{wu2021style}& 1 & 85.0 & 23.2 & 80.4 & 21.3 & 24.5 & 30.0 & 32.0 & 26.7 & 83.2 & 34.8 & 74.0 & 57.3 & 29.0 & 77.7 & 27.3 & 36.5 & 5.0 & 28.2 & 39.4 & 42.8 \\
\midrule
Ours (R101)& 1 & 80.88 & 32.62 & 85.82 & 36.11 & 30.68 & 40.7 & 43.66 & 41.71 & 84.07 & 30.72 & 84.48 & 65.38 & 27.56 & 85.98 & 36.47 & 51.36 & 24.13 & 26.68 & 30.67 & \textbf{49.46} \\
Ours (MiT-B5)& 1 & 83.41 & 35.30 & 87.11 & 44.79 & 32.27 & 42.53 & 50.19 & 52.47 & 87.99 & 46.09 & 90.43 & 66.71 & 25.55 & 88.64 & 50.32 & 50.77 & 44.54 & 34.36 & 38.58 & \textbf{55.37}\\
\bottomrule
\end{tabular}
}
\caption{\textbf{GTA$\rightarrow$Cityscapes: One-Shot Domain Adaptation.} Num$^{\#}$ represents the number of \textbf{real} images used for training, which are from the target domain or other auxiliary datasets, {\it e.g.}, ImageNet~\cite{deng2009imagenet} and WikiArt~\cite{huang2017arbitrary}. The baseline methods adopt the ResNet-101 backbone.}
\label{tab:gta_city}
\end{table*}
\begin{table*}[]
\centering%
\resizebox{\linewidth}{!}{%
\begin{tabular}{l|ccccccccccccccccc|cc}
\toprule
Method& Num$^{\#}$ & Road&SW&Build&Wall$^*$&Fence$^*$&Pole$^*$&TL&TS&Veg&Sky&Person&Rider&Car&Bus&MC&Bike& mIoU$^*$ & mIoU\\
\midrule
\midrule
Source only & 0 & 36.30 & 14.64 & 68.78 & 9.17 & 0.20 & 24.39 & 5.59 & 9.05 & 68.96 & 79.38 & 52.45 & 11.34 & 49.77 & 9.53 & 11.03 & 20.66 & 33.65 & 29.45\\
\midrule
AdaptSegNet~\cite{Tsai_adaptseg_2018}& 1 &64.1 & 25.6 & 75.3 & - & - & - & 4.7 & 2.7 & 77.0 & 70.0 & 52.2 & 20.6 & 51.3 & 22.4 & 19.9 & 22.3 & 39.1 & - \\
CLAN~\cite{luo2019taking}& 1 & 68.3 & 26.9 & 72.2 & - & - & - & 5.1 & 5.3 & 75.9 & 71.4 & 54.8 & 18.4 & 65.3 & 19.2 & 22.1 & 20.7 & 40.4 & - \\
ADVENT~\cite{vu2019advent}& 1 & 65.7 & 22.3 & 69.2 & - & - & - & 2.9 & 3.3 & 76.9 & 69.2 & 55.4 & 21.4 & 77.3 & 17.4 & 21.4 & 16.7 & 39.9 & - \\
CBST~\cite{zou2018unsupervised}& 1 & 59.6 & 24.1 & 72.9 & - & - & - & 5.5 & 13.8 & 72.2 & 69.8 & 55.3 & 21.1 & 57.1 & 17.4 & 13.8 & 18.5 & 38.5 & - \\
OST~\cite{benaim2018one}& 1 & 75.3 & 31.6 & 72.1 & - & - & - & 12.3 & 9.3 & 76.1 & 71.1 & 51.1 & 17.7 & 68.9 & 19.0 & 26.3 & 25.4 & 42.8 & - \\
FSDR~\cite{huang2021fsdr}& $\geq$15& 69.3 & 34.9 & 77.6 & 7.9 & 0.2 & 29.4 & 16.3 & 19.2 & 72.3 & 76.3 & 56.7 & 22.1 & 80.6 & 41.5 & 19.1 & 29.3 & 47.3 & 40.8 \\
DRPC~\cite{yue2019domain}& $\geq$15& - & - & - & - & - & - & - & - & - & - & - & - & - & - & - & - & - & 37.6\\
SADG~\cite{peng2022semantic}& 0 & - & - & - & - & - & - & - & - & - & - & - & - & - & - & - & - & - & 40.87\\
RobNet~\cite{choi2021robustnet} & 0 & - & - & - & - & - & - & - & - & - & - & - & - & - & - & - & - & - & 37.21\\
WEDGE~\cite{kim2021wedge} & 1000 & - & - & - & - & - & - & - & - & - & - & - & - & - & - & - & - & - & 40.31\\
IBN~\cite{pan2018IBN-Net} & 0 & - & - & - & - & - & - & - & - & - & - & - & - & - & - & - & - & - & 34.18 \\
ASM~\cite{luo2020adversarial}& $\geq$10& 85.7 & 39.7 & 77.1 & 1.1 & 0.0 & 24.2 & 2.1 & 9.2 & 76.9 & 81.7 & 43.4 & 11.4 & 63.9 & 15.8 & 1.6 & 20.3 & 40.7 & 34.6\\
SMPPM~\cite{wu2021style}& 1 & 79.3 & 35.3 & 75.9 & 5.6 & 16.6 & 29.8 & 25.4 & 22.7 & 79.9 & 76.8 & 54.6 & 23.5 & 60.2 & 23.9 & 21.2 & 36.6 & 47.3 & 41.4\\
\midrule
Ours(R101)& 1 & 82.53 & 33.83 & 77.75 & 12.61 & 0.78 & 34.18 & 30.8 & 34.42 & 79.75 & 82.43 & 55.42 & 30.71 & 72.5 & 28.44 & 15.89 & 47.76 & \textbf{51.72} & \textbf{44.99} \\
Ours(MiT-B5)& 1 & 81.35 & 37.34 & 84.76 & 19.54 & 1.24 & 43.67 & 43.03 & 34.37 & 86.49 & 90.03 & 63.84 & 32.76 & 79.62 & 42.68 & 27.96 & 47.24 & \textbf{57.80} & \textbf{50.99}\\
\bottomrule
\end{tabular}
}
\caption{\textbf{SYNTHIA$\rightarrow$Cityscapes: One-Shot Domain Adaptation.} Num$^{\#}$ represents the number of \textbf{real} images used for training, which are from the target domain or other auxiliary datasets, {\it e.g.}, ImageNet~\cite{deng2009imagenet} and WikiArt~\cite{huang2017arbitrary}. The baseline methods adopt the ResNet-101 backbone. mIoU$^*$ represents the 13 classes mIoU performance when removing the 3 classes denoted by $*$, as the common practice in~\cite{Tsai_adaptseg_2018,tranheden2021dacs,vu2019advent}.}
\label{tab:synthia_city}
\end{table*}
\begin{table*}[t]
\centering%
\resizebox{\linewidth}{!}{%
\begin{tabular}{l|cccccccccccccccccccc|c}
\toprule
Method& Num$^{\#}$ & Road&SW&Build&Wall&Fence&Pole&TL&TS&Veg&Terrain&Sky&Person&Rider&Car&Truck&Bus&Train&MC&Bike& mIoU\\
\midrule
\midrule
Source only & 0 & 69.5 & 12.9 & 65.6 & 10.5 & 6.8 & 39.5 & 41.7 & 20.4 & 62.7 & 7.5 & 63.5 & 58.5 & 31.1 & 62.3 & 16.3 & 31.9 & 1.4 & 22.0 & 10.8 & 33.4 \\
\midrule
BDL~\cite{li2019bidirectional} & 0 & - & - & - & - & - & - & - & - & - & - & - & - & - & - & - & - & - & - & - & 30.3 \\
FDA~\cite{yang2020fda} & 0 & - & - & - & - & - & - & - & - & - & - & - & - & - & - & - & - & - & - & - & 35.3 \\
GASF\cite{Kundu_2021_ICCV} & 0 & - & - & - & - & - & - & - & - & - & - & - & - & - & - & - & - & - & - & - & 38.3 \\
BDL~\cite{li2019bidirectional} & 2975 & 89.6 & 37.6 & 65.4 & 19.5 & 14.4 & 23.2 & 22.7 & 25.5 & 48.9 & 35.7 & 39.7 & 50.7 & 29.8 & 79.1 & 27.9 & 32.8 & 0.2 & 18.5 & 30.4 & 36.3 \\
CBST~\cite{zou2018unsupervised} & 2975 & 74.6 & 30.3 & 73.2 & 7.0 & 20.0 & 40.7 & 47.4 & 35.4 & 53.0 & 5.8 & 65.4 & 47.7 & 21.7 & 75.4 & 21.7 & 39.1 & 5.5 & 18.5 & 33.5 & 37.7 \\
MLSL~\cite{iqbal2020mlsl} & 2975 & 81.5 & 33.6 & 76.6 & 7.9 & 23.1 & 41.1 & 47.5 & 35.9 & 52.0 & 6.1 & 64.9 & 54.1 & 27.8 & 81.2 & 16.5 & 37.7 & 1.5 & 17.4 & 36.7 & 39.1 \\
FogAdapt~\cite{iqbal2022fogadapt} & 2975 & 85.1 & 31.7 & 76.7 & 16.5 & 20.3 & 41.2 & 46.2 & 34.9 & 70.8 & 9.1 & 63.8 & 53.9 & 26.2 & 81.5 & 22.0 & 38.0 & 5.9 & 19.0 & 36.3 & 41.0 \\
ASM~\cite{luo2020adversarial} & 0 & 83.15 & 32.83 & 69.41 & 13.3 & 23.47 & 29.24 & 22.24 & 22.81 & 66.35 & 23.32 & 64.7 & 53.38 & 22.07 & 75.49 & 18.41 & 22.25 & 0.00 & 18.85 & 22.46 & 35.99 \\
\midrule
Ours (R101) & 0 & 79.58 & 21.68 & 78.68 & 18.18 & 22.17 & 30.53 & 35.31 & 17.33 & 77.50 & 22.82 & 67.80 & 61.87 & 32.39 & 80.57 & 24.21 & 25.90 & 0.00 & 23.17 & 22.15 & \textbf{39.05}\\
Ours (MiT-B5) & 0 & 84.98 & 33.09 & 80.87 & 37.6 & 28.57 & 34.38 & 44.66 & 27.12 & 81.5 & 33.51 & 73.2 & 64.42 & 36.42 & 84.37 & 34.85 & 45.25 & 41.95 & 30.76 & 34.28 & \textbf{49.04}\\
\bottomrule
\end{tabular}
}
\caption{\textbf{GTA$\rightarrow$Foggy Cityscapes: One-Shot Domain Generalization.} Num$^{\#}$ represents the number of \textbf{Foggy Cityscapes} (target domain) images used for training. The baseline methods adopt the ResNet-101 backbone.}
\label{tab:gta_fog}
\end{table*}
\begin{table*}[]
\centering%
\resizebox{\linewidth}{!}{%
\begin{tabular}{l|ccccccccccccccccc|cc}
\toprule
Method& Num$^{\#}$ & Road&SW&Build&Wall$^*$&Fence$^*$&Pole$^*$&TL&TS&Veg&Sky&Person&Rider&Car&Bus&MC&Bike& mIoU$^*$ & mIoU\\
\midrule
\midrule
Source only & 0 & 29.3 & 21.3 & 34.5 & 0.8 & 0.0 & 17.5 & 15.8 & 8.2 & 17.1 & 33.5 & 57.1 & 4.7 & 71.2 & 12.2 & 2.9 & 8.9 & 24.4 & 20.9 \\
\midrule
BDL~\cite{li2019bidirectional} & 2975 & 83.2 & 43.2 & 63.6 & 2.38 & 0.1 & 19.1 & 6.8 & 5.3 & 35.4 & 19.9 & 55.4 & 31.8 & 65.2 & 21.0 & 27.6 & 37.1 & 38.1 & 32.3 \\
CBST~\cite{zou2018unsupervised} & 2975 & 70.5 & 31.2 & 57.62 & 2.9 & 0.02 & 31.7 & 29.1 & 23.1 & 38.5 & 41.1 & 61.3 & 18.9 & 75.0 & 8.3 & 11.9 & 32.1 & 38.4 & 33.3 \\
MLSL~\cite{iqbal2020mlsl} & 2975 & 48.9 & 27.2 & 53.4 & 11.4 & 0.4 & 31.9 & 32.4 & 21.0 & 49.2 & 40.1 & 65.8 & 24.4 & 77.8 & 20.9 & 19.0 & 50.5 & 40.8 & 35.9\\
FogAdapt~\cite{iqbal2022fogadapt} & 2975 & 62.2 & 28.0 & 56.4 & 13.1 & 0.7 & 30.3 & 30.1 & 27.4 & 61.7 & 61.8 & 54.9 & 30.0 & 66.1 & 2.6 & 12.1 & 44.8 & 41.4 & 36.4\\
ASM~\cite{luo2020adversarial} & 0 & 75.39 & 29.04 & 67.85 & 1.79 & 0.05 & 21.54 & 4.88 & 8.51 & 59.78 & 58.46 & 37.83 & 11.91 & 49.47 & 7.58 & 3.17 & 16.68 & 33.12 & 28.37 \\
\midrule
Ours(R101) & 0 & 79.91 & 32.49 & 69.53 & 11.97 & 0.22 & 30.52 & 29.89 & 22.56 & 65.98 & 59.91 & 59.57 & 26.23 & 78.34 & 28.87 & 25.37 & 36.04 & \textbf{47.28} & \textbf{41.09}\\
Ours(MiT-B5) & 0 & 74.5 & 30.2 & 78.21 & 16.48 & 1.89 & 37.93 & 35.17 & 24.87 & 77.47 & 69.03 & 61.14 & 25.45 & 81.96 & 37.97 & 24.69 & 31.56 & \textbf{50.17} & \textbf{44.28} \\
\bottomrule
\end{tabular}
}
\caption{\textbf{SYNTHIA$\rightarrow$Foggy Cityscapes: One-Shot Domain Generalization.} Num$^{\#}$ represents the number of \textbf{Foggy Cityscapes} (target domain) images used for training. The baseline methods adopt the ResNet-101 backbone. mIoU$^*$ represents the 13 classes mIoU performance when removing the 3 classes denoted by $*$, as the common practice in~\cite{Tsai_adaptseg_2018,tranheden2021dacs,vu2019advent}.}
\label{tab:synthia_fog}
\end{table*}
|
\section*{Table of Contents}
\tableofcontents
\newpage
\section{Introduction}
\label{sec:intro}
In the last decade, with the observation of gravitational wave events by the LIGO Scientific Collaboration \cite{LIGOScientific:2016aoc,LIGOScientific:2016sjg,LIGOScientific:2017bnn,LIGOScientific:2017vox,LIGOScientific:2017ycc,LIGOScientific:2017vwq}, and interferometry measurements of the centre of M87 and the Milky Way by the Event Horizon Telescope Collaboration \cite{EventHorizonTelescope:2019dse,EventHorizonTelescope:2022xnr,EventHorizonTelescope:2022xqj}, we have entered a new era of testing gravity, probing the nature of black holes and Einstein's theory of general relativity (GR) in the previously inaccessible strong field regime.
\par In GR, mathematical theorems guarantee that in (electro-)vacuum the gravitational field of stationary black holes is described \textit{uniquely} by the Kerr(-Newman) metric \cite{Chrusciel:2012jk}. As eloquently put by Subrahmanijan Chandrasekhar, the uniqueness
theorems along with a set of other results dubbed \textit{no-hair theorems} (see \cite{Herdeiro:2015waa} for a review) assert that
the Kerr metric provides ``the absolute exact representation of untold numbers of massive black holes that populate the universe''. While all strong regime observations are so far compatible with this ``Kerr hypothesis", any eventual deviation would provide a much sought after smoking-gun for new physics.
\par Indeed, once we go beyond GR and delve onto the realm of modified theories of gravity,
stationary vacuum spacetimes need not to be described by the Kerr metric. Popular examples of black hole spacetimes defying the Kerr hypothesis include gravity coupled with new (complex) bosonic degrees of freedom \cite{Doneva:2022ewd,Herdeiro:2014goa,Herdeiro:2015gia,Herdeiro:2016tmi}, scalar-Gauss-Bonnet gravity \cite{Sotiriou:2013qea,Sotiriou:2014pfa,Doneva:2017bvd,Silva:2017uqg,Antoniou:2017acq,Cunha:2019dwb,Dima:2020yac,Herdeiro:2020wei,Berti:2020kgk,Kanti:1995vq,Kleihaus:2011tg,Cunha:2016wzk,Delgado:2020rev}, 4D-Einstein-Gauss-Bonnet gravity \cite{Glavan:2019inb,Fernandes:2022zrq,Lu:2020iav,Kobayashi:2020wqy,Fernandes:2020nbq,Hennigar:2020lsl,Fernandes:2021dsb,Aoki:2020lig,Fernandes:2021ysi,Clifton:2020xhc}, and dynamical Chern-Simons gravity \cite{Alexander:2009tp,Jackiw:2003pm,Yagi:2012ya,Cano:2021rey,Cano:2019ore}.
Modification of the field equations describing gravity, however, naturally leads to an increase in their complexity such that analytic analysis becomes intractable. With closed-form solutions not available, one is forced to resort either to perturbation theory or numerical methods. In the strong-field regime, perturbative approximations may not be well-justified, leaving numerical studies as the most promising way forward. In this arena, the ever-increasing precision of our observations and measurements necessitates increasingly accurate solutions.
\par
In this paper, we will describe a numerical method capable of solving with high accuracy a system of non-linear elliptic partial differential equations (PDEs), such as those that appear when analyzing stationary and axially symmetric spacetimes, and implement this in a publicly available code.
A first version of our numerical implementation is available in the GitHub repository in Ref. \cite{gitlink}. The code is written in Julia and can be run with ease on laptop-class computers, with solutions being found in a matter of seconds. The Julia language is fast, memory efficient, and easy to manipulate, ensuring that implementing different modified gravity theories is not a difficult task.
Our code follows similar previous numerical solvers for these spacetimes, in particular the non-publicly-available FIDISOL/CADSOL solver \cite{Solver1,Solver2,Solver3} (which has been extensively used in the literature, see e.g. \cite{Herdeiro:2014goa, Herdeiro:2015gia, Herdeiro:2016tmi,Delgado:2020rev, Kleihaus:2015aje,Kleihaus:2011tg,Herdeiro:2020wei,Berti:2020kgk,Cunha:2019dwb}) and a recent publicly available solver developed in Refs.~ \cite{Sullivan:2019vyi,Sullivan:2020zpf}. We have several motivations for writing another code. First, in this work we show that pseudospectral methods are ideally suited to solving the type of equations at hand. In our tailor-made implementation we therefore make use of such methods\footnote{See also \textit{Kadath} \cite{Grandclement:2009ju}, which implements a spectral methods library for theoretical physics in C++.}, while both former codes utilise finite difference methods. In contrast to the first code mentioned above, our implementation is also open source, and moreover in our bench-marking we find our code to be far more accurate as detailed further below. The code of Refs. \cite{Sullivan:2019vyi,Sullivan:2020zpf} is also significantly more accurate than that of
the FIDISOL/CADSOL solver (though the documented accuracy is still less than our own when bench-marked on the Kerr solution) and is publicly available. This code is, however, written in C, and our use of Julia leads to simple code that can easily be adapted. Our code is also considerably faster. Our overall aim is a publicly available, accurate, well documented code that is transparent and easy to use code. Furthermore, our code provides a toolbox to explore several properties of the obtained black hole solutions, rather than being only a PDE solver.
This paper is organised as follows. In section \ref{sec:spectral} we introduce the reader to pseudospectral methods and the technical machinery that will be necessary to apply them in the context of black hole physics. Next, in section \ref{sec:bhs} we will describe how we can use the aforementioned methods to solve the stationary and axisymmetric field equations for gravity, discussing the boundary conditions, coordinate compactifications, and our numerical approach. We further discuss many of the properties that can be extracted from a spinning black hole solution. Finally, in section \ref{sec:numbhs} we start by validating our methods and code against the Kerr black hole, which is known in closed form, and later use our machinery to obtain stationary and axisymmetric black holes in Einstein-scalar-Gauss-Bonnet gravity for linear and exponential couplings. We also discuss
the accuracy of our code, and further compare with results from other codes in published literature. We work with units such that $G=c=1$.
\section{Spectral Methods}
\label{sec:spectral}
The idea behind spectral methods is to approximate a smooth solution to a system of differential or integral equations by a sum over a finite number of basis functions. In this section, we review how this works. Given that our aim is a clear and adaptable code, the presentation is relatively complete, and summarises that given in John P. Boyd's book on spectral methods \cite{Boyd}, to which the reader can turn for full details.
For simplicity, we begin with the one dimensional case and illustrate how the method finds an approximation to the smooth solution to a differential equation, $u(x)$, with the differential equation written in the form
\begin{equation}
\mathcal{R}\left(x,u\right) = 0,
\label{eq:residual}
\end{equation}
where $\mathcal{R}$ is called \textit{the residual} of the system. The solution, $u(x)$, can be approximated by a finite truncated series solution $u_N(x)$ such that
\begin{equation}
u(x) \approx u_N(x) = \sum_{n=0}^{N-1} \alpha_n \phi_n(x),
\end{equation}
where $\{\phi_n(x)\}_{n=0}^{\infty}$ is a set of \textit{global} and \textit{orthogonal} basis functions, $\{\alpha_n\}_{n=0}^{\infty}$ is the set of \textit{spectral coefficients}, and $N$ is the \textit{resolution}. In this setup, $u_N(x)$ can be said to be a numerical solution of the system \eqref{eq:residual} if spectral coefficients are found such that the residual is below a certain prescribed tolerance. The method is therefore global rather than local, with an exponential convergence with $N$ for problems with smooth solutions. Since black hole solutions are smooth, we expect exponential convergence when come to find such solutions using spectral methods.
This is in contrast to the polynomial convergence rate of most other numerical methods, such as finite element or finite difference schemes.
Furthermore, numerical solutions obtained via a spectral method provide an analytical approximation to the problem at hand (rather than a set of approximate numerical values at a discrete number of points).
\par As noted the basis functions must be orthogonal, which implies that
\begin{equation}
(\phi_n, \phi_m) = c_n \delta_{m n},
\end{equation}
where the brackets represent the inner product of two functions
$f(x)$ and $g(x)$ with respect to the weight function, $\omega(x)>0$, on the interval $[a,b]$ as
\begin{equation}
\left(f,g\right) \equiv \int_a^b f(x) g(x) \omega(x) dx.
\label{eq:innerprod}
\end{equation}
\par The set of basis functions used should have a number of further properties: i) they should be easy to compute (e.g. trigonometric functions or polynomials); ii) the approximations built out of the basis functions should converge rapidly to the true solution as the resolution is increased; iii) they should be complete, which means that any solution can be represented to arbitrarily high accuracy by taking the resolution to be sufficiently high. Two commonly used sets of basis functions that obey these requirements are sines and cosines, as used in a Fourier series, and a special class of polynomials dubbed Chebyshev polynomials.
\subsection{Chebyshev Polynomials}
\par For non-periodic problems, Chebyshev polynomials are the most natural choice as the spectral series is guaranteed to converge exponentially fast (provided our domain is restricted to the interval $x \in [-1,1]$). The $n$th Chebyshev polynomial (of the first kind) is defined as
\begin{equation}
T_n\left(x\right) = \cos \left(n \theta\right), \qquad \theta = \arccos x,
\end{equation}
or equivalently by the three-term recurrence relation
\begin{equation}
\begin{aligned}
&T_0(x)\equiv 1, \quad T_1(x) \equiv x, \\&
T_{n}(x) = 2xT_{n-1}(x) - T_{n-2}(x), \quad n\geq 2.
\end{aligned}
\end{equation}
The first six Chebyshev polynomials are shown in Fig. \ref{fig:chebyshevpolynomials} in the domain $x \in [-1,1]$.
\begin{figure}[h!]
\centering
\includegraphics[width=0.7\textwidth]{images/ChebyshevPolys.pdf}
\caption{First six Chebyshev polynomials in the domain $x \in [-1,1]$.}
\label{fig:chebyshevpolynomials}
\end{figure}
Chebyshev polynomials obey the orthogonality relation in the domain $x \in [-1,1]$
\begin{equation}
\int_{-1}^{1} \frac{T_m(x) T_n(x)}{\sqrt{1-x^2}} dx = \frac{\pi}{2}\left(1+\delta_{0n}\right) \delta_{mn},
\end{equation}
and hence form an orthogonal basis.
Their derivatives are given by
\begin{equation}
\frac{d}{dx}T_n(x) = n U_{n-1}(x),
\end{equation}
where $U_n(x)$ denotes the $n$th Chebyshev polynomial of the second kind, defined by the recurrence relation
\begin{equation}
\begin{aligned}
&U_0(x)\equiv 1, \quad U_1(x) \equiv 2x,\\&
U_{n}(x) = 2xU_{n-1}(x) - U_{n-2}(x), \quad n\geq 2,
\end{aligned}
\end{equation}
and with derivative
\begin{equation}
\frac{d}{dx}U_n(x) = \frac{(n+1)T_{n+1}(x)-x U_n(x)}{x^2-1}.
\end{equation}
Note that some derivatives require special care at the boundaries $x=\pm 1$, and must be computed as a well-defined limit, namely
\begin{equation}
\left. \frac{d^2 T_n}{dx^2} \right|_{x=-1} = \left(-1\right)^n\frac{n^4-n^2}{3}, \qquad \left. \frac{d^2 T_n}{dx^2} \right|_{x=1} = \frac{n^4-n^2}{3}.
\end{equation}
\subsection{Interpolation}
Interpolation is the process by which a function is approximated by a finite sum of suitable basis functions. The idea is that the sum is constructed such that the
aproximation agrees with the true function at the chosen set of interpolation points (also called collocation points). The objective is that the interpolant provides a good approximation to the true function also between those points. By virtue of the minimal amplitude theorem \cite{Boyd}, Chebyshev polynomials are widely used in interpolations. The reason is twofold. First, when using the so called Chebyshev nodes (or Gauss-Chebyshev points), $x_n$, as collocation points, the effect of the Runge phenomenon (numerical instabilities near the boundaries in the form of uncontrolled oscillations) is minimized. These points are the roots of the $N$th Chebyshev polynomial, and are given by
\begin{equation}
x_n = \cos \left( \frac{\left(2n+1\right)\pi}{2N} \right), \qquad n=0,\ldots, N-1.
\label{eq:chebnodes}
\end{equation}
Secondly, when Chebyshev polynomials are used as the basis for the interpolation, the interpolation error is distributed uniformly over the whole range.
The algorithm to interpolate a smooth function $u(x)$ using a truncated Chebyshev series written as\footnote{The prime in the sum denotes that the first coefficient is halved. We chose to halve the first coefficient in the sum in order to simplify some relations below, such as Eq. \eqref{eq:chebinterpolation}.}
\begin{equation}
u_N(x) = \frac{1}{2}\alpha_0 + \sum_{n=1}^{N-1} {\vphantom{\sum}} \alpha_n T_n(x)\equiv \sum_{n=0}^{N-1} {\vphantom{\sum}}' \alpha_n T_n(x),
\label{eq:chebseries}
\end{equation}
relies on finding the optimal spectral coefficients $\{\alpha_n\}$, and uses the discrete orthogonality relation of Chebyshev polynomials:
\begin{equation}
\sum_{j=0}^{N-1} T_n(x_j) T_m(x_j) = \frac{N}{2}\left(1+\delta_{0n}\right) \delta_{mn},
\end{equation}
where the $x_j$ are the points given in Eq. \eqref{eq:chebnodes}. These discrete relations imply that
\begin{equation}
\alpha_n = \frac{2}{N} \sum_{j=0}^{N-1} u(x_j) T_n(x_j).
\label{eq:chebinterpolation}
\end{equation}
We present in Fig. \ref{fig:interpolation} an illustrative example of a Chebyshev interpolation performed for several resolutions using the above expressions.
\begin{figure}[h!]
\centering
\includegraphics[width=\textwidth]{images/FunctionInterpolation.pdf}
\caption{Interpolation of the function $u(x) = x^2 e^{-2x}$ on a Gauss-Chebyshev grid, for resolutions ranging from $N=1$ to $N=6$. Using Eqs. \eqref{eq:chebseries} and \eqref{eq:chebinterpolation} we find that for $N=6$ the spectral coefficients of the approximation $u_6(x)$ are $\alpha_0 \approx 1.48427$, $\alpha_1 \approx -2.49232$, $\alpha_2 \approx 1.85409$, $\alpha_3 \approx -1.01286$, $\alpha_4 \approx 0.395175$, and $\alpha_5 \approx -0.111169$.}
\label{fig:interpolation}
\end{figure}
\subsection{Trigonometric functions}
For periodic problems, sines and cosines are the most suitable basis functions for a spectral series. These obey well known orthogonality relations, and form the basis for the Fourier series representation of a periodic function. As we will see, a finite sum of these functions can be used to generate a trigonometric interpolation to a periodic function.
Moreover, we can often simplify further by
taking into account symmetries. For example, considering the core problem considered in this paper, we note that
stationary and axisymmetric black holes are solutions to a system of two-dimensional elliptic PDEs that depend on the radial coordinate and the zenith angle $\theta \in [0,\pi]$. These solutions also often possess definite parity with respect to $\theta=\pi/2$ (i.e. in most cases they are symmetric about $\theta = \pi/2$), and therefore we need only to consider the range $\theta \in [0,\pi/2]$. In this range, the following discrete orthogonality relations hold
\begin{equation}
\begin{aligned}
&\sum_{j=0}^{N-1} \cos\left(2n \theta_j\right) \cos\left(2m \theta_j\right) = \frac{N}{2} \left(1+\delta_{0n}\right) \delta_{mn},\\&
\sum_{j=0}^{N-1} \cos\left(\left[2n+1\right] \theta_j\right) \cos\left(\left[2m+1\right] \theta_j\right) = \frac{N}{2} \delta_{mn},\\&
\sum_{j=0}^{N-1} \sin\left(2n \theta_j\right) \sin\left(2m \theta_j\right) = \frac{N}{2} \left(1-\delta_{0n}\right) \delta_{mn},\\&
\sum_{j=0}^{N-1} \sin\left(\left[2n+1\right] \theta_j\right) \sin\left(\left[2m+1\right] \theta_j\right) = \frac{N}{2} \delta_{mn},
\end{aligned}
\end{equation}
where
\begin{equation}
\theta_n=\frac{(2n+1)\pi}{4N}, \quad n=0,\dots, N-1.
\label{eq:thetapoints}
\end{equation}
Table \ref{tab:trig} summarizes the parity properties of the functions appearing in the relations above, and
together with these orthogonality relations we see that a function, $u(\theta)$, symmetric about $\theta=0,\pi/2$ can be interpolated using only even cosines such that
\begin{equation}
u_N(x) = \sum_{n=0}^{N-1} {\vphantom{\sum}}' \alpha_n \cos(2 n \theta),
\end{equation}
with the spectral coefficients
\begin{equation}
\alpha_n = \frac{2}{N} \sum_{j=0}^{N-1} u(\theta_j) \cos(2 n \theta_j).
\label{eq:triginterpolation}
\end{equation}
\renewcommand{\tabcolsep}{1pt}
\begin{table}[h!]
\centering
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline
Fourier series & Parity w.r.t. $\theta = 0$ & Parity w.r.t. $\theta = \pi/2$ & $u(0)$ & $u(\frac{\pi}{2})$ & $\partial_\theta u(0)$ & $\partial_\theta u(\frac{\pi}{2})$ \\ \hline
$\cos(\left[2n\right]\theta)$ & Even & Even & $\neq 0$ & $\neq 0$ & $=0$ & $=0$ \\ \hline
$\cos(\left[2n+1\right]\theta)$ & Even & Odd & $\neq 0$ & $=0$ & $= 0$ & $\neq 0$ \\ \hline
$\sin(\left[2n\right]\theta)$ & Odd & Odd & $=0$ & $=0$ & $\neq 0$ & $\neq 0$ \\ \hline
$\sin(\left[2n+1\right]\theta)$ & Odd & Even & $=0$ & $\neq 0$ & $\neq 0$ & $=0$ \\ \hline
\end{tabular}
\caption{Properties of the elements of a Fourier series of a function $u(\theta)$, depending on the parity symmetries, along with a scheme of its boundary values. Here, $n \in \mathbb{N}^0$. The entries on this table for $\theta = \pi$ would be equivalent to those of $\theta=0$.}
\label{tab:trig}
\end{table}
\subsection{Solving an ODE with a spectral method -- a first example}
\par So far we have seen how a known function can be approximated by a finite sum of suitable basis functions using interpolation. Now we turn to the problem of how to find such an approximation to an unknown function that is the solution to a given differential equation.
To understand how to solve differential equations using a spectral method, we will first consider a simple ordinary differential equation (ODE) example. Consider the one dimensional non-linear boundary value problem
\begin{equation}
\mathcal{R} = u_{xx}-u_{x}^2 = 0, \quad u(-1) - 2 = 0, \quad u(1) - 1 = 0.
\label{eq:odeproblem}
\end{equation}
We will find an approximate solution to this boundary value problem in the form of a Chebyshev spectral series, and later compare our results with the known exact solution, given by
\begin{equation}
u(x) = \log \left(\frac{2 e^2}{(e-1) x+e+1}\right).
\label{eq:exactODE1}
\end{equation}
To illustrate the calculations analytically, we will first consider a (very) low resolution approximate solution with $N=3$, where
\begin{equation}
u \approx u_3 = \frac{\alpha_0}{2} T_0(x) + \alpha_1 T_1(x) + \alpha_2 T_2(x) = \frac{\alpha_0}{2} + \alpha_1 x + \alpha_2 \left(2x^2-1\right).
\label{eq:odeexample1}
\end{equation}
Here there are $N=3$ unknowns ($\alpha_0$, $\alpha_1$, and $\alpha_2$), and $N_{BC} = 2$ boundary conditions. Once we substitute our ans\"atz of Eq. \eqref{eq:odeexample1} onto the residual given in Eq. \eqref{eq:odeproblem} we obtain
\begin{equation}
\mathcal{R} \approx 4 \alpha_2-(4 \alpha_2 x+\alpha_1)^2 = 0,
\label{eq:odeResidual}
\end{equation}
together with the boundary conditions
\begin{equation}
\frac{\alpha_0}{2} - \alpha_1 + \alpha_2 - 2 = 0, \qquad \frac{\alpha_0}{2} + \alpha_1 + \alpha_2 - 1 = 0.
\label{eq:odeBC}
\end{equation}
\begin{figure}[]
\centering
\includegraphics[width=.85\textwidth]{images/SolveODE.pdf}\hfill
\caption{Approximations to the solution of the boundary value problem given in Eq. \eqref{eq:odeproblem} (top) together with their absolute errors with respect to the exact solution (bottom) for resolutions $N=3$ (left) and $N=24$ (right).}
\label{fig:odesolution}
\end{figure}
To find the approximate solution we simply need to determine values for the three unknowns. Given that we have only $N=3$ degrees of freedom, and the two boundary conditions provide two constraints, we need only one further equation to find the values. To get this constraint the idea is to evaluate the residual at the $N-N_{BC}=1$ collocation point given by Eq.~\eqref{eq:chebnodes} (with the $N$ in that expression given by $1$), which gives the point $x=0$. With our resolution of $N=3$, finding an approximate solution to the boundary value problem then reduces to solving three non-linear coupled algebraic equations for the spectral coefficients, given by the two boundary conditions of Eq. \eqref{eq:odeBC} together with the residual of Eq. \eqref{eq:odeResidual} evaluated at $x=0$. The solution to the system is
\begin{equation*}
\alpha_0 = \frac{23}{8}, \quad \alpha_1 = -\frac{1}{2}, \quad \alpha_2 = \frac{1}{16}.
\end{equation*}
By construction this is an interpolation to the exact solution \eqref{eq:exactODE1}.
We note that had we chosen the resolution $N=4$, the number of collocation points where we would have to evaluate the residual would be $N-N_{BC}=2$, and would be given by Eq. \eqref{eq:chebnodes} as $x=\pm \sin \pi/8 \approx \pm 0.382683$. This would give $N=4$ coupled equations to find the four unknowns in this case. This then generalises to arbitrary $N$.
On Fig. \ref{fig:odesolution} we plot the exact solution against the approximation obtained with $N=3$, together with the absolute error, $|1-u_N/u|$, whose maximum can be seen to be $\mathcal{O}\left(10^{-2}\right)$ already for a very low resolution $N=3$. We also plot the approximation to the solution of the boundary value problem, but for a resolution $N=24$, where we observe that errors become of order machine precision ($\mathcal{O}\left(10^{-16}\right)$). In Fig. \ref{fig:odeerror} we plot the behaviour of the maximum absolute error as a function of the resolution, where exponential convergence is observed. As a rule of thumb, the truncation error is typically the same order-of-magnitude as the last coefficient retained in the truncation series.
\begin{figure}[]
\centering
\includegraphics[width=0.7\textwidth]{images/ODE_error.pdf}
\caption{Logarithmic plot of the maximum absolute error in the approximation to the solution of the boundary value problem as a function of the resolution. Spectral convergence is observed, together with a roundoff plateau.}
\label{fig:odeerror}
\end{figure}
\par An important point to make here is that even though for $N=3$ the approximation system has a closed-form analytical solution for the spectral coefficients, once higher resolutions are considered, a numerical root-finding method (such as Newton-Raphson) has to be employed. To successfully employ a Newton-Raphson method, a good initial guess for the spectral coefficients is of the utmost importance. We will come back to this in the next section.
\par To conclude, spectral collocation methods, also known as pseudospectral methods, are powerful tools that can be used to find high accuracy numerical solutions to differential equations. They provide global analytical approximations for the solution, and handling any kind of boundary condition is straightforward.
\subsection{Root-finding methods -- Newton-Raphson}
\par To numerically solve the system of algebraic equations for the spectral coefficients
a root-finding method must, in general, be employed. In particular, we will utilise the well known Newton-Raphson method.
In
the one-dimensional case, the method attempts to solve the equation $f(x)=0$ iteratively, starting with a initial guess, $x_0$. Successive values of $x$ are then generated
until a value, $x_*$, is reached at which the equation is approximately solved to a certain prescribed tolerance. The series of iterations takes the form
\begin{equation}
x_{n+1} = x_n - \frac{f(x_n)}{f'(x_n)}.
\label{eq:newton1d}
\end{equation}
For example, assume we want to find the root of the function $f(x) = x^3+x-1$, known to be $x_* \approx 0.6823278$ to eight decimal places. Starting with $x_0=1$ as our initial guess, using Eq. \eqref{eq:newton1d} we obtain
\begin{equation*}
x_1 = 0.75, \quad x_2 = 0.68604651, \quad x_3 = 0.68233958, \quad x_4 = 0.6823278,
\end{equation*}
thus converging to $x_*$ in four iterations to the prescribed tolerance of eight decimal places. Convergence, is however, not guaranteed, and particularly in more complicated settings an appropriate choice of starting point is extremely important, and must be chosen carefully.
\par The generalization of the method to $N$ variables with $N$ equations finds the root of a vector-valued function $F:\mathbb{R}^N \to \mathbb{R}^N$, and amounts to solving the linear system
\begin{equation}
\mathcal{J}(\mathbf{x}_n) \left(\mathbf{x}_{n+1}-\mathbf{x}_{n}\right) = -F\left(\mathbf{x}_{n}\right),
\end{equation}
at each iteration for the unknown $\mathbf{x}_{n+1}-\mathbf{x}_{n}$, where $\mathcal{J}$ is the $N \times N$ Jacobian matrix of the system, defined as
\begin{equation}
\mathcal{J}_{ij} = \frac{\partial F_i}{\partial x_j}.
\end{equation}
Constructing the Jacobian matrix of a given system is not always an easy task,
but is relatively
straightforward for the system of equations that arises when using the spectral method to solve an ODE. Another advantage of this method. As described, in this case, the system to be solved, $F$, will be composed of the residual $\mathcal{R}$ evaluated at the Gauss-Chebyshev points \eqref{eq:chebnodes}, and the boundary conditions, and will in general involve $u$, $u_x$ and $u_{xx}$. Our unknowns are the spectral coefficients $\alpha_j$. Thus, to facilitate the computation of the Jacobian, we may use the chain rule
\begin{equation}
\mathcal{J}_{ij} = \frac{\partial F_i}{\partial \alpha_j} = \frac{\partial F_i}{\partial u}\frac{\partial u}{\partial \alpha_j} + \frac{\partial F_i}{\partial u_x}\frac{\partial u_x}{\partial \alpha_j} + \frac{\partial F_i}{\partial u_{xx}}\frac{\partial u_{xx}}{\partial \alpha_j},
\end{equation}
and only then substitute the spectral expansions for the function $u$. This process is easily generalizable to a system of differential ODEs/PDEs (rather than a single one)\footnote{One must be careful when labelling the spectral coefficients of the different functions as it might be a source of errors.}.
\par As previously stated, when using a Newton-Raphson method, the choice of initial guess to the spectral coefficients is extremely important, because a non-appropriate choice will likely result in non-convergence of the algorithm. A good initial guess can sometimes be difficult to obtain, especially when dealing with systems of PDEs, where the number of coefficients is large (for our specific black holes problem, typically of $\mathcal{O}\left(10^{3}\right)$ coefficients). A good way of tackling this issue stems from a good understanding of the problem in question. For example, from an effective field theory point of view, a Kerr black hole is probably a good approximation to a black hole solution in modified theories. Therefore, since we have a closed-form expression for a Kerr black hole, an interpolation of this solution can used to generate an initial guess for the spectral coefficients in modified theories.
\section{Black Holes -- Metric Ans\"atz, The Kerr Solution, Boundary Conditions, and Connection with the Numerical Approach}
\label{sec:bhs}
\par We will now apply the methods described in the previous sections to black hole physics and approximately solve the coupled PDEs that arise when obtaining stationary solutions in a given theory of gravity.
We will focus on a particular ans\"atz for the black hole spacetime written in quasi-isotropic coordinates with line-element
\begin{equation}
ds^2 = -f \mathcal{N}^2 dt^2 + \frac{g}{f} \left[ h \left(dr^2 + r^2 d\theta^2\right) + r^2 \sin^2\theta \left(d\varphi - \frac{W}{r}\left(1-\mathcal{N}\right) dt\right)^2\right],
\label{eq:metric}
\end{equation}
which is stationary, axisymmetric, and circular.
Here $f$, $g$, $h$ and $W$ are dimensionless functions of the radial and angular coordinates $r$ and $\theta$, and
\begin{equation*}
\mathcal{N} \equiv \mathcal{N}(r) = 1-\frac{r_H}{r},
\end{equation*}
where $r_H$ is the (coordinate) location of the event horizon. The spatial coordinates range over the intervals
\begin{equation}
r \in [r_H,\infty[, \qquad \theta \in [0,\pi], \qquad \varphi \in [0,2\pi].
\end{equation}
In order for the line-element to be a solution to the theory of gravity at hand, the functions, $f$, $g$, $h$ and $W$ must satisfy a set of PDEs that result from the field equations of the theory.
The spacetime presented possesses two Killing vector fields, $k=\partial_t$ and $\Phi = \partial_\varphi$, and the linear combination
\begin{equation}
\xi = \partial_t + \Omega_H \partial_\varphi,
\label{eq:killingvector1}
\end{equation}
where $\Omega_H$ is the angular velocity of the horizon (to be defined below), is orthogonal to and null on the event horizon.
This Lewis-Papapetrou form for the metric is motivated by the discussion of Ref. \cite{Xie:2021bur}, which asserts that the above metric ans\"atz is consistent for a generic theory of gravity provided that its solutions can be obtained perturbatively from a solution in the general relativity limit. Note that our form of the metric functions on the line element of Eq. \eqref{eq:metric} differ somewhat from the standard form used in other works (see e.g. \cite{Herdeiro:2014goa, Herdeiro:2015gia, Herdeiro:2016tmi,Delgado:2020rev, Kleihaus:2015aje,Kleihaus:2011tg,Herdeiro:2020wei,Berti:2020kgk,Cunha:2019dwb}). The reasons for this will become clearer once we make a connection to our numerical approach, and are related to numerical accuracy issues.
\subsection{General Relativity -- The Kerr Black Hole}
To begin, let us consider the known Kerr black hole, which is the solution to the stationary and axisymmetric field equations of GR in vacuum. For completeness, we present its charged generalization, the Kerr-Newman solution of electrovacuum in Appendix~\ref{ap:KN}. The Kerr black hole solves the field equations
\begin{equation}
G_{\mu \nu} = 0,
\end{equation}
where $G_{\mu \nu}$ is the Einstein tensor, which follow from the Einstein-Hilbert action
\begin{equation}
\mathcal{S} = \frac{1}{16\pi} \int d^4x \sqrt{-g} R,
\end{equation}
where $R$ is the Ricci scalar of the metric $g_{\mu \nu}$. With the ans\"atz of Eq. \eqref{eq:metric} the Kerr black hole solution reads
\begin{equation}
\begin{aligned}
&f = \left(1+\frac{r_H}{r}\right)^2 \frac{\mathcal{A}}{\mathcal{B}},\\&
g = \left(1+\frac{r_H}{r}\right)^2,\\&
h = \frac{\mathcal{A}^2}{\mathcal{B}},\\&
W = \frac{2M\left(M r+r^2+r_H^2\right)}{r_H r^3 \mathcal{B}}\sqrt{M^2-4 r_H^2}
\end{aligned}
\label{eq:kerr}
\end{equation}
where
\begin{equation}
\begin{aligned}
&\mathcal{A} = \frac{2 M r \left(M r+\left(r^2+r_H^2\right)\right)+\left(r^2-r_H^2\right)^2}{r^4}-\frac{\left(M^2-4 r_H^2\right)}{r^2}\sin^2\theta ,\\&
\mathcal{B} = \left(\mathcal{A} + \frac{ \left(M^2-4 r_H^2\right) }{r^2} \sin^2\theta\right)^2 - \frac{\left(r^2-r_H^2\right)^2 \left(M^2-4 r_H^2\right)}{r^6} \sin^2\theta,
\end{aligned}
\end{equation}
and $M$ is the ADM mass of the black hole. The total angular momentum per unit mass, $a$, of the solution is related to $M$ and $r_H$ via
\begin{equation}
r_H = \frac{\sqrt{M^2-a^2}}{2} \equiv \frac{M}{2}\sqrt{1-\chi^2},
\label{eq:rh}
\end{equation}
where we have defined the dimensionless spin
\begin{equation}
\chi \equiv a/M = J/M^2.
\end{equation}
The mass $M$ and total angular momentum $J$ can be read
off from the metric components as $r \to \infty$, where
\begin{equation}
\begin{aligned}
&g_{tt} = -f \mathcal{N}^2 + \frac{g \left(1-\mathcal{N}\right)^2 W^2}{f} \sin^2\theta = -1+\frac{2M}{r} + \mathcal{O}\left(r^{-2}\right), \\&
g_{t\varphi} = -\frac{g r \left(1-\mathcal{N}\right) W}{f}\sin^2 \theta = -\frac{2J}{r}\sin^2 \theta + \mathcal{O}\left(r^{-2}\right),
\end{aligned}
\label{eq:asymptoticexpansionsmet}
\end{equation}
leading to
\begin{equation}
\begin{aligned}
&f = 1-\frac{2\left(M-r_H\right)}{r} + \mathcal{O}\left(r^{-2}\right), \\&
W = \frac{2J}{r_H r}+ \mathcal{O}\left(r^{-2}\right).
\end{aligned}
\label{eq:asymptoticexpansions}
\end{equation}
Note that the Kerr black hole in the quasi-isotropic coordinate system presented in Eq. \eqref{eq:metric} can be obtained from the standard textbook Boyer-Lindquist coordinates solution with the radial coordinate transformation
\begin{equation}
r_{BL} = r + M + \frac{M^2 - a^2}{4r} = r \left(1 + \frac{M}{r} + \frac{r_H^2}{r^2}\right).
\label{eq:coord_transf}
\end{equation}
The inverse transformation is given by
\begin{equation}
r = \frac{1}{2}\left(r_{BL} - M + \sqrt{\left(r_{BL}-M\right)^2 - 4 r_H^2}\right).
\label{eq:coord_transf_inv}
\end{equation}
\subsection{Boundary Conditions}
\par To solve the set of PDEs that result from the field equations in a particular theory of gravity, suitable boundary conditions should be imposed. These are obvious if an exact solution, such as the Kerr solution, is known by a trivial examination of the metric functions. However, in more intricate cases in modified gravity lacking an exact solution the boundary conditions must be found with a careful examination of the field equations and employing suitable expansions of the involved functions near the domain boundaries. For example if theories possess a GR limit when some parameter tends to zero, an expansion about the Kerr solution is possible. With this process, we find that in all cases to be discussed in this work within modified gravity theories, the metric functions must obey the same boundary conditions as the Kerr solution does.
These conditions are summarized next.
\par \textbf{(i) Axis boundary conditions:} Axial symmetry and regularity of the solutions on the symmetry axis $\theta=0, \pi$, imply the following boundary conditions
\begin{equation}
\partial_\theta f = \partial_\theta g = \partial_\theta h = \partial_\theta W = 0, \qquad \mathrm{for} \quad \theta=0, \pi.
\end{equation}
Moreover, the absence of conical singularities further imposes that on the symmetry axis
\begin{equation}
h = 1, \qquad \mathrm{for} \quad \theta=0, \pi.
\label{eq:conicalsing}
\end{equation}
All solutions to be discussed in this work
are are also symmetric with respect to a reflection on the equatorial plane $\theta=\pi/2$. Therefore, as was discussed above, it is enough to consider the range $\theta \in [0,\pi/2]$ and one of the boundary conditions becomes
\begin{equation}
\partial_\theta f = \partial_\theta g = \partial_\theta h = \partial_\theta W = 0, \qquad \mathrm{for} \quad \theta=\pi/2.
\end{equation}
\par \textbf{(ii) Event horizon boundary conditions:} The black hole solutions discussed here possess an event horizon located at a surface with constant radial variable $r=r_H$. The boundary conditions that the metric functions $f$, $g$ and $h$ obey at $r=r_H$ are
\begin{equation}
\begin{aligned}
&f-r_H \partial_r f = 0 \\&
g+r_H \partial_r g = 0, \\
\partial_r h = 0.
\end{aligned}
\end{equation}
The reason for the Robin-type boundary conditions that the functions $f$ and $g$ obey comes from the inclusion of the $\mathcal{N}^2$ factor in front of $f$ in the coefficient that multiplies $dt^2$ in the metric ans\"atz, Eq. \eqref{eq:metric}. This factor is chosen such that these functions do not contain a double-zero in a near-horizon expansion, allowing for more accurate solutions in this region, and therefore, a more accurate extraction of horizon physical quantities such as the area and temperature of the event horizon. We find that there are (at least) two possibilities for the condition that the function $W$ should obey at the horizon, one of which must be chosen appropriately such that the number of input parameters is kept at two\footnote{The ``input parameters" are the parameters needed to uniquely define a solution, this is discussed fully below in Sec. \ref{sec:connectio_numerical_approach}}
\begin{equation}
W = r_H \Omega_H
\label{eq:WBC_H_1}
\end{equation}
or
\begin{equation}
W - \frac{r_H}{2} \partial_r W = 0,
\end{equation}
where $\Omega_H$ is a constant interpreted as the angular velocity of the event horizon, which in the case of a Kerr black hole is given by
\begin{equation}
\Omega_H^{\mathrm{Kerr}} = \frac{\sqrt{M^2-4 r_H^2}}{2M \left(M+2r_H\right)} = \frac{\chi^2 - 1 + \sqrt{1-\chi^2}}{4r_H \chi}.
\label{eq:omh}
\end{equation}
\par \textbf{(iii) Asymptotic boundary conditions:} Requiring asymptotic flatness (i.e., that as $r \to \infty$, our solution approaches the Minkowski spacetime), the functions $f$, $g$, and $h$ obey
\begin{equation}
\lim_{r \to \infty} f = \lim_{r \to \infty} g = \lim_{r \to \infty} h = 1.
\end{equation}
Similarly to the boundary conditions at the event horizon, we find (at least) two suitable conditions for the function $W$
\begin{equation}
\lim_{r \to \infty} W = 0,
\end{equation}
or, from the asymptotic expansion of Eq. \eqref{eq:asymptoticexpansions}
\begin{equation}
\lim_{r \to \infty} r_H r^2 \partial_r W + 2 M^2 \chi = 0 \Leftrightarrow \lim_{r \to \infty} \frac{r^2}{2r_H} \partial_r W + \left(1+\frac{r^2}{2r_H} \partial_r f\right)^2 \chi = 0.
\end{equation}
\subsection{Connection with the numerical approach}
\label{sec:connectio_numerical_approach}
\par To recap, the field equations of a gravitational theory once applied to the line element of Eq. \eqref{eq:metric} will result in a set of non-linear coupled elliptic PDEs in $r$ and $\theta$ subject to the boundary conditions described above. Our objective is to solve this system of PDEs numerically using a spectral method. For this we introduce the compactified radial coordinate
\begin{equation}
x = 1-\frac{2r_H}{r},
\label{eq:x}
\end{equation}
mapping the range $r\in [r_H, \infty[$ to
\begin{equation}
x \in [-1,1].
\end{equation}
With the compactified coordinate, the radial boundary conditions change, and we proceed to give the new conditions next.
\noindent\textbf{Event horizon boundary conditions:} The boundary conditions that the metric functions $f$, $g$ and $h$ now obey are
\begin{equation}
\begin{aligned}
&f-2 \partial_x f = 0 \\&
g+2 \partial_x g = 0,\\
\partial_x h = 0,
\end{aligned}
\end{equation}
for $x=-1$.
For the function $W$, the first possibility (Eq. \eqref{eq:WBC_H_1}) remains unchanged ($W|_{x=1} = r_H \Omega_H$), whereas the second becomes
\begin{equation}
W - \partial_x W = 0,
\label{eq:W_BC_CHI_H}
\end{equation}
at $x=-1$.
\noindent\textbf{Asymptotic boundary conditions:} The asymptotic boundary conditions the functions $f$, $g$, and $h$ are now
\begin{equation}
f=g=h=1, \qquad \mathrm{for} \quad x=1
\end{equation}
Asymptotically, function $W$ now obeys either
\begin{equation}
W = 0,
\label{eq:W0_BC_x1}
\end{equation}
or
\begin{equation}
\partial_x W + \left(1 + \partial_x f\right)^2 \chi = 0,
\label{eq:W_BC_CHI_I}
\end{equation}
at $x=1$.
With our compactified radial coordinate, and given the symmetries of our problem\footnote{From now on we consider only the cases with even parity with respect to $\theta=\pi/2$.}, a suitable spectral expansion for the black hole metric functions (collectively denoted by $\mathcal{F}=\{f,g,h,W\}$) is given by
\begin{equation}
\mathcal{F}^{(k)} = \sum_{i=0}^{N_x-1} {\vphantom{\sum}}' \sum_{j=0}^{N_\theta-1} {\vphantom{\sum}}' \alpha_{ij}^{(k)} T_i(x) \cos \left(2j\theta\right),
\label{eq:spectralexpansion}
\end{equation}
where $N_x$ and $N_{\theta}$ are the resolutions in the radial and angular coordinates. Note that, as discussed above, the angular boundary conditions are automatically satisfied by this expansion (c.f. Table \ref{tab:trig}).
As mentioned previously, we will usually use the Kerr metric itself to set our initial guess when working with modified theories of gravity, and to do so we will need the expression for the spectral coefficients that follow from an interpolation of a two-dimensional function $u(x,\theta)$, which is given by
\begin{equation}
\alpha_{ij} = \frac{4}{N_x N_\theta} \sum_{k=0}^{N_x-1} \sum_{l=0}^{N_\theta-1} u(x_k,\theta_l) T_i(x_k) \cos\left(2j \theta_l\right),
\label{eq:interpolation}
\end{equation}
where $x_k$ and $\theta_l$ are given in Eqs. \eqref{eq:chebnodes} and \eqref{eq:thetapoints} respectively.
\par Each Kerr black hole is uniquely described by two input parameters. For example, in the presentation given in Eq. \eqref{eq:kerr}, these are the location of the event horizon $r_H$ and the ADM mass $M$. We have seen, however, in expressions \eqref{eq:rh} and \eqref{eq:omh} that they are related to the dimensionless spin $\chi$ and the horizon angular velocity $\Omega_H$. Therefore, using the correct parametrization, the Kerr solution can be described by any input pair chosen from $r_H$, $\Omega_H$, $\chi$, and $M$. In the numerical approach, in a theory agnostic setting, one input parameter that must be used is $r_H$ because it enters directly the metric ans\"atz and the definition of our compactified coordinate $x$. We have, however, freedom in the choice of the other input parameter in the numerics. To the best of our knowledge, so far in the literature for similar problems \cite{Herdeiro:2014goa, Herdeiro:2015gia, Herdeiro:2016tmi,Delgado:2020rev, Kleihaus:2015aje,Kleihaus:2011tg,Herdeiro:2020wei,Berti:2020kgk,Cunha:2019dwb}, the other input parameter has always been chosen as the event horizon angular velocity $\Omega_H$. Using this input pair $(r_H, \Omega_H)$, we find compatibility with the boundary conditions for the function $W$ if we choose Eqs. \eqref{eq:WBC_H_1} and \eqref{eq:W0_BC_x1} at the horizon and infinity, respectively. Then, in the case of a Kerr black hole, one finds that for a fixed value of $r_H$, two branches of solutions exist, as shown in Fig. \ref{fig:kerrbranches}. This follows from inverting the relation \eqref{eq:omh}.
\begin{figure}[]
\centering
\includegraphics[width=0.65\textwidth]{images/KerrBranches.pdf}
\caption{Choosing $(r_H, \Omega_H)$ as input parameters, for fixed $r_H$, two branches exist.}
\label{fig:kerrbranches}
\end{figure}
The first branch of solutions starts at a vanishing value of $\Omega_H$ (for fixed $r_H$) and exists until
\begin{equation}
r_H \Omega_H = \frac{\sqrt{5\sqrt{5}-11}}{4\sqrt{2}} \approx 0.0750708,
\end{equation}
at which point
\begin{equation}
\chi = \sqrt{\frac{\sqrt{5}-1}{2}} \approx 0.786151.
\end{equation}
Then, a second branch appears, and $\Omega_H$ tends backwards towards zero. As $\Omega_H \to 0$ on this second branch, extremal solutions are approached. The existence of two branches of solutions is not unique to Kerr, and is observed as well in the modified theories of gravity to be discussed in this work. We note that the numerical procedure gets rather difficult as near-extremal solutions are approached, as our metric ans\"atz with the described boundary conditions is not compatible with extremal solutions.
A novel approach that we can also adopt is to choose the pair $(r_H, \chi)$ as the input pair. This input pair is compatible with the $W$ boundary conditions of Eqs. \eqref{eq:W_BC_CHI_H} and \eqref{eq:W_BC_CHI_I} while maintaining the number of input parameters at two. We often find it very convenient to use the dimensionless spin as an input parameter, for example when exploring domains of existence, or simply when working on a single solution where a certain $\chi$ is wanted. Our numerical spectral method is not only powerful because high accuracy solutions are produced, but also because highly non-linear boundary conditions can be handled with ease (which is the case of the boundary condition of Eq. \eqref{eq:W_BC_CHI_I}).
\begin{figure*}[t!]
\centering
\includegraphics[width=\textwidth]{images/grid.png}
\caption{Fiducial grid with $N_x \times N_\theta = 11 \times 5$, highlighted in blue. The field equations (residuals) are evaluated in the blue region, with the boundary conditions imposed on the red region. The yellow highlight concerns the imposition of the condition of Eq. \eqref{eq:conicalsing}.}
\label{fig:grid}
\end{figure*}
To solve the system of field equations subject to the discussed boundary conditions we must construct a suitable grid. This is done as follows. We assume a resolution $N_x \times N_\theta$. The discrete grid points in the $x$ direction are chosen according to Eq. \eqref{eq:chebnodes}, where we take $N = N_x-2$, together with the boundary points $x=- 1$ and $x=1$, such that the total number of points in the $x$ direction is $N_x$\footnote{This approach is also called \textit{boundary-bordering} in the spectral methods' literature.}. In $\theta$, our points are chosen as in Eq. \eqref{eq:thetapoints}, where we take $N=N_\theta$. The $x$ and $\theta$ points together form the schematically shown in Fig. \eqref{fig:grid}, in blue.
Assuming there are a total number, $N_{funcs}$, of functions to solve for, there are $N_{funcs}\times N_x \times N_\theta$ degrees of freedom (spectral coefficients) in the problem, as seen in the spectral expansion of Eq. \eqref{eq:spectralexpansion}.
For each value of $\theta$ in the grid at the $x$ boundaries we impose for each function the horizon and asymptotic boundary conditions as discussed before. This gives us a total of $N_{funcs} \times 2 \times N_\theta$ equations (Fig. \ref{fig:grid}, in red). The remaining $N_{funcs} \times \left(N_x-2\right) \times N_\theta$ equations come from imposing the $N_{funcs}$ residuals resulting from the field equations at each non-boundary $x$ value, for each $\theta$. The number of degrees of freedom is then equal to the number of equations to solve, as it should. A small caveat -- the absence of conical singularities imposes that Eq. \eqref{eq:conicalsing} must be obeyed (i.e. for our coordinate range, $h=1$ at $\theta=0$). While we could leave this condition outside the numerical scheme and use it as another test to the code, we find that imposing it allows obtaining solutions with (much) higher accuracy. In our particular implementation, therefore, we have swapped the evaluation of one of the residuals at $\theta=0$ (for all interior values of $x$)\footnote{We empirically found that any of the field equations should equally valid to remove for this process, resulting in similar outcomes} with the condition of Eq. \eqref{eq:conicalsing}, see Fig. \ref{fig:grid} in yellow.
\subsubsection{Numerical Approach: A summary}
\par Here we summarize our numerical approach for clarity. To solve the field equations, some preliminary work must be done. First, we employ the metric ans\"atz of Eq. \eqref{eq:metric} which contains four unknown functions, $f$, $g$, $h$, and $W$. Plugging this metric ans\"atz onto the field equations of the theory, leads to a set of non-linear coupled PDEs that depend on the functions and their first and second derivatives $\left(\mathcal{F}, \partial_r \mathcal{F},\partial_r^2 \mathcal{F},\partial_\theta \mathcal{F},\partial_\theta^2 \mathcal{F}, \partial_{r\theta} \mathcal{F}\right)$. The set of field equations is then expressed in terms of the compactified coordinate $x$ defined in Eq. \eqref{eq:x} and put in residual form (i.e., $\mathcal{R}\left(x,\theta, \partial\mathcal{F}\right) = 0$). The same is done for the appropriate boundary conditions as discussed. This part of the process is usually done resorting to a computer algebra system such as \emph{Mathematica}, \emph{Maple} or \emph{SageMath}\footnote{In this work we have used \emph{Mathematica} along with the \emph{OGRe} package \cite{Shoshany:2021iuc} to obtain the explicit field equations of many theories.}. The residuals (and appropriate Jacobian) are then exported to a coding file in order to solve the problem using the developed numerical infrastructure. Each function is expanded in a spectral series given by Eq. \eqref{eq:spectralexpansion} and the input parameters are then specified (depending on the chosen boundary conditions for the function $W$). To successfully solve the field equations, a good initial guess must be provided to our Newton solver. For this, we interpolate the functions of the known Kerr solution using Eq. \eqref{eq:interpolation}, obtaining appropriate spectral coefficients to be provided as a good initial guess. If new fields are present, as is the case with modified theories, we typically take advantage of perturbative solutions and interpolate them as a guess. Convergence is assumed once the norm difference between the spectral coefficients of two successive iterations is less than a certain prescribed tolerance.
To speed up the solver, the values of our basis functions and their first and second derivatives are calculated at all the grid points and stored, such that no repeated evaluations are performed. Another optimization that we found particularly impactful was to store the values on the grid of the trigonometric functions that typically appear in the residuals, $\sin \theta$ and $\cos \theta$.
Once a solution is obtained, physical quantities can be extracted from it as we discuss in the next section, and the solution can be used for numerous investigations.
\subsection{Physical Properties of Stationary and Axisymmetric Black Holes}
\label{sec:physical_properties}
\par Once a numerical stationary and axisymmetric black hole solution has been found using our code, we can extract important quantities of physical relevance. In this section, we review many of the quantities that one can extract from a solution, some of which can be used to test the accuracy of our code. We have implemented additional code to extract all these quantities from a numerical solution.
\subsubsection{Quantities of interest}
Starting with the asymptotic quantities, we have seen that the mass $M$ and angular momentum $J$ can be extracted from the asymptotic expansion of Eq. \eqref{eq:asymptoticexpansionsmet} or Eq. \eqref{eq:asymptoticexpansions}. In terms of the coordinate $x$ these are given by
\begin{equation}
M = r_H \left(1+\partial_x f\right)|_{x=1}, \qquad J = -r_H^2 \partial_x W|_{x=1}.
\end{equation}
We remark that such a simple expression for the extraction of $J$ is the reason why we have defined the function $W$ in this way -- such that its decay is of the form $\sim 1/r$, allowing for more accurate results. In a circular spacetime, the zeroth law of black hole mechanics holds, which means that the surface gravity is constant on the horizon of the stationary black hole. The surface gravity is defined as $\kappa^2 = -1/2 (\nabla_\mu \xi_\nu)(\nabla^\mu \xi ^\nu)$, where $\xi$ was defined in Eq. \eqref{eq:killingvector1}. The Hawking temperature \cite{Hawking:1975vcx} can then be obtained from the surface gravity as
\begin{equation}
T_H = \frac{\kappa}{2\pi} = \left. \frac{1}{2\pi r_H} \frac{f}{\sqrt{g h}} \right|_{x=-1}.
\end{equation}
The induced metric on the horizon is
\begin{equation}
d\Sigma^2 = \mathfrak{h}_{ij} dx^i dx^j = r_H^2 \left.\frac{g}{f}\left[h d\theta^2 + \sin^2 \theta d\varphi^2\right] \right|_{x=-1},
\end{equation}
and from it we can compute several quantities of interest, the most important being the event horizon area
\begin{equation}
A_H = \int_H \sqrt{\mathfrak{h}} d\theta d\varphi = \left. 2\pi r_H^2 \int_{0}^{\pi} d\theta \sin \theta \frac{g \sqrt{h}}{f} \right|_{x=-1}.
\end{equation}
Also of importance is the entropy, which is given in the Iyer-Wald formalism by \cite{Iyer:1994ys}
\begin{equation}
S = \left. -2\pi \int_H \frac{\delta \mathcal{L}}{\delta R_{\mu \nu \alpha \beta}} \epsilon_{\mu \nu} \epsilon_{\alpha \beta} dA \right|_{\mathrm{on-shell}},
\label{eq:entropy_wald}
\end{equation}
where $\epsilon_{\mu \nu}$ is the binormal vector to the event horizon surface. In the case of a Kerr black hole the above expression reduces to the simple form $S = A_H/4$.
The horizon and asymptotic quantities are connected via the Smarr type relation \cite{PhysRevLett.30.71,Bardeen:1973gs,Iyer:1994ys,Liberati:2015xcp}
\begin{equation}
M = \left. 2T_H S + 2 \Omega_H J - 2 \int_{\Sigma} d^3x \sqrt{-g} \mathcal{L}\right|_{\mathrm{on-shell}}.
\label{eq:smarr}
\end{equation}
The Smarr relation is extremely important when studying numerical solutions as it provides a test to the code that relates physical quantities obtained on the horizon and asymptotic regions, allowing us to estimate the accuracy of the numerical method.
Also of interest is the perimetral radius $\mathfrak{R}$ which is a geometrically significant radial coordinate such that a circumference along the equatorial plane has perimeter $2\pi \mathfrak{R}$. It is related to the coordinate $r$ by
\begin{equation}
\mathfrak{R} = \left. \sqrt{g_{\phi \phi}} \right|_{\theta = \pi/2} = \left. \sqrt{\frac{g}{f}} r \right|_{\theta = \pi/2}.
\end{equation}
To explore the horizon geometry, it is useful to define the horizon circumference along the equator
\begin{equation}
L_e = 2\pi \mathfrak{R}_H,
\end{equation}
and along the poles
\begin{equation}
L_p = 2\int_{0}^{\pi} \sqrt{g_{\theta \theta}}|_{x=-1}d\theta = \left. 2r_H \int_{0}^{\pi} \sqrt{\frac{gh}{f}} \right|_{x=-1} d\theta.
\end{equation}
With these two quantities, we can define the sphericity
\begin{equation}
s = \frac{L_e}{L_p}.
\end{equation}
For a Kerr black hole $s \geq 1$, with $s$ increasing with spin. That means that spin deforms the horizon towards oblateness. The linear velocity of the horizon quantifies how fast the null geodesic generators of the horizon spin relative to a static observer at infinity, and is given by
\begin{equation}
v_H = \Omega_H \mathfrak{R}_H.
\end{equation}
For a Kerr black hole we have in terms of $M$ and $r_H$
\begin{equation}
\begin{aligned}
&J = M^2\sqrt{1-\left(\frac{2r_H}{M}\right)^2}, \\&
T_H = \frac{1}{4\pi M \left(1+\frac{M}{2r_H}\right)},\\&
A_H = 8\pi M^2\left(1+\frac{2r_H}{M}\right) \\&
L_e = 4\pi M, \\&
L_p = 4M\sqrt{2\left(1+\frac{2r_H}{M}\right)} \mathrm{EllipticE}\left(\frac{1}{2}\left[1-\frac{2r_H}{M}\right]\right), \\&
\mathfrak{R}_H = 2M,
\end{aligned}
\end{equation}
where $\mathrm{EllipticE}$ denotes the complete elliptic integral of the second kind, and we also note that $2r_H/M = \sqrt{1-\chi^2}$.
The Kerr solution is Ricci flat, and thus the Lagrangian of GR vanishes on-shell and therefore so does the last term in Eq. \eqref{eq:smarr}.
\subsubsection{Ergoregion}
\par The ergoregion is defined as the domain outside the event horizon wherein the norm of the asymptotically timelike Killing vector $k=\partial_t$ becomes positive, $g_{\mu \nu} k^\mu k^\nu > 0$. It is bounded by the event horizon and by the surface where
\begin{equation}
g_{tt} = -f \mathcal{N}^2 + \frac{g \left(1-\mathcal{N}\right)^2 W^2}{f} \sin^2\theta = 0.
\end{equation}
Within the ergoregion, an object cannot appear stationary with respect to a distant observer due to the intense frame-dragging.\footnote{This immediately follows from the fact that the 4-velocity of a massive particle must be timelike, $g_{\mu \nu} u^\mu u^\nu<0$. Indeed, the worldline of an object standing still at a fixed point implies that $u = \partial_t$, and if $g_{tt}\geq 0$, then $g_{\mu \nu} u^\mu u^\nu\geq 0$.} Furthermore, ergoregions raise the possibility of extracting energy from a black hole via the Penrose process, or superradiant scattering \cite{Brito:2015oca}. Starting from the well-known result for the ergosphere of a Kerr black hole in Boyer-Lindquist coordinates and inverting the relation of Eq. \eqref{eq:coord_transf} we obtain that in quasi-isotropic coordinates the ergosphere of a Kerr black hole is located at
\begin{equation}
r_E^{Kerr} = \frac{r_H}{\sqrt{1-\chi^2}} \left(\sqrt{1-\chi^2 \cos^2\theta}+\chi \sin \theta\right),
\end{equation}
where the subscript ``$E$" refers to ``ergoregion".
Due to the symmetries of our problem, we need only consider the range $\theta \in [0,\pi/2]$. To visualize the ergoregion, we introduce the coordinates
\begin{equation}
X = \frac{r}{r_H} \sin \theta, \qquad Z = \frac{r}{r_H} \cos \theta.
\end{equation}
In Fig. \ref{fig:ergokerr} we observe the ergoregion of a Kerr black holes in the $X-Z$ plane for several values of dimensionless spin.
\begin{figure}[]
\centering
\includegraphics[width=0.75\textwidth]{images/ergo_kerr.pdf}
\caption{Ergoregion of a Kerr black hole with $\chi = 0.3$ (orange), $\chi = 0.6$ (green), and $\chi = 0.8$ (red) visualized on the $X-Z$ plane. The event horizon is shown in blue.}
\label{fig:ergokerr}
\end{figure}
\subsubsection{Petrov type}
\par The Petrov classification allows for a kinematic characterization of the gravitational field in a coordinate independent manner using algebraic properties of the Weyl tensor $C_{\mu \nu \alpha \beta}$, namely its number of distinct principal null directions. This classification is useful, for example, when searching for exact solutions, or for a Carter-like constant \cite{Carter:1968}. Using the Newman-Penrose formalism, the information is contained in five complex scalars known as the Weyl scalars. With the null tetrad $\{l^\mu, n^\mu, m^\mu, \overline{m}^\mu\}$, where $l^{\mu}$ and $n^{\mu}$ are real, and $m^\mu, \overline{m}^\mu$ are complex conjugate satisfying the orthonormality conditions $l^\mu n_\mu=1$, $m^\mu \overline{m}_\mu=-1$ and all other products zero, the Weyl scalars are defined as
\begin{equation}
\begin{aligned}
&\psi_0 = -C_{\mu \nu \alpha \beta} l^\mu m^\nu l^\alpha m^\beta, \\&
\psi_1 = -C_{\mu \nu \alpha \beta} l^\mu n^\nu l^\alpha m^\beta, \\&
\psi_2 = -C_{\mu \nu \alpha \beta} l^\mu m^\nu \overline{m}^\alpha n^\beta, \\&
\psi_3 = -C_{\mu \nu \alpha \beta} l^\mu n^\nu \overline{m}^\alpha n^\beta, \\&
\psi_4 = -C_{\mu \nu \alpha \beta} n^\mu \overline{m}^\nu n^\alpha \overline{m}^\beta.
\end{aligned}
\end{equation}
With the above scalars, the following Lorentz invariant quantities can be constructed
\begin{equation}
\begin{aligned}
&I = \psi_0 \psi_4 - 4\psi_1 \psi_3 + 3 \psi_2^2, \\&
J = -\psi_2^3 + \psi_0 \psi_2 \psi_4 + 2\psi_1 \psi_2 \psi_3 - \psi_4 \psi_1^2 - \psi_0 \psi_3^2,\\&
D = I^3 - 27 J^2,\\&
K = \psi_4^2 \psi_1 - 3\psi_4 \psi_3 \psi_2 + 2\psi_3^3,\\&
L = \psi_4 \psi_2 - \psi_3 ^2,\\&
N = 12L^2-\psi_4^2 I.
\end{aligned}
\end{equation}
Given the above quantities, it is possible to determine the Petrov type of a given spacetime. The classification is summarized in Table \ref{tab:petrov} \cite{Achour:2021pla}. In particular, a spacetime is said to be \textit{algebraically special} if $D = 0$. The Kerr(-Newman) spacetime is Petrov type D. In a numerical setup, we also find useful to introduce the \textit{speciality index} defined as \cite{Berti:2004ny}
\begin{equation}
S = \frac{27 J^2}{I^3}.
\label{eq:specindex}
\end{equation}
\begin{table}[]
\centering
\begin{tabular}{|c|c|}
\hline
Type & Conditions \\ \hline
O & $\psi_0 = \psi_1 = \psi_2 = \psi_3 = \psi_4 = 0$ \\ \hline
I & $D \neq 0$ \\ \hline
II & $D = 0, I\neq 0, J \neq 0, K \neq 0, N \neq 0$ \\ \hline
III & $D = 0, I = J = 0, K \neq 0, L \neq 0$ \\ \hline
N & $D = 0, I = J = K = L = 0$ \\ \hline
D & $D = 0, I \neq 0, J \neq 0, K = N = 0$ \\ \hline
\end{tabular}
\caption{Summary of Petrov classification.}
\label{tab:petrov}
\end{table}
With an appropriate choice of tetrad, following Ref. \cite{Berti:2004ny}, it is possible to gauge away $\psi_1$ and $\psi_3$ to zero. Such a tetrad would be for example
\begin{equation}
\begin{aligned}
&l^\mu = \sqrt{\frac{g_{\varphi \varphi}}{2 \left(g_{t\varphi}^2 - g_{tt} g_{\varphi \varphi}\right)}} \left(1,0,0,-\frac{g_{t\varphi}+\sqrt{g_{t\varphi}^2 - g_{tt} g_{\varphi \varphi}}}{g_{\varphi \varphi}}\right),\\&
n^\mu = \sqrt{\frac{g_{\varphi \varphi}}{2 \left(g_{t\varphi}^2 - g_{tt} g_{\varphi \varphi}\right)}} \left(1,0,0,-\frac{g_{t\varphi}-\sqrt{g_{t\varphi}^2 - g_{tt} g_{\varphi \varphi}}}{g_{\varphi \varphi}}\right),\\&
m^\mu = \frac{1}{\sqrt{2}} \left(0,\frac{i}{\sqrt{g_{rr}}},\frac{1}{\sqrt{g_{\theta \theta}}},0\right).
\end{aligned}
\end{equation}
\subsubsection{Marginal Stable Circular Orbits: Light Rings and ISCO}
\par The study of marginal stable circular orbits is highly relevant for the observational properties of black holes. The innermost stable circular orbit (ISCO) of massive particles is the smallest possible radius for a stable circular orbit and is often taken to mark the inner edge of an accretion disk around a black hole. Accelerated charged particles orbiting the black hole emit synchroton radiation whose physical properties are connected with the frequency of geodesics at the ISCO. Therefore, physical properties of an astrophysical black hole can be inferred via measurements of the ISCO through accretion disks.
\par Light rings are circular null geodesics, typically unstable, allowing light to encircle a black hole before being scattered to infinity or falling into the event horizon. From an observational point of view, they are important for observations made with the Event Horizon Telescope as they are intimately connected with the shadow of the black hole \cite{Cunha:2018acu}.
To compute the ISCO and light rings we follow Ref. \cite{Sullivan:2020zpf}. We start by considering the line element of Eq. \eqref{eq:metric} in the form
\begin{equation}
ds^2 = g_{tt} dt^2 + g_{rr} dr^2 + g_{\theta \theta} d\theta^2 + g_{\varphi \varphi} d\varphi^2 + 2g_{t \varphi} dt d\varphi.
\end{equation}
The two independent killing vectors of the spacetime, $k^\mu = \partial_t$ and $\Phi^\mu = \partial_\varphi$, have the associated conserved reduced energy $E$ and angular momentum $L$
\begin{equation}
\begin{aligned}
&E = - k_\mu \frac{dx^\mu}{d\lambda} = - g_{tt} \dot t - g_{t \varphi} \dot \varphi,\\&
L = \Phi_\mu \frac{dx^\mu}{d\lambda} = g_{t \varphi} \dot t + g_{\varphi \varphi} \dot \varphi,
\end{aligned}
\end{equation}
where $\dot \equiv d/d\lambda$. The above expressions can be rearranged in terms of $\dot t$ and $\dot \varphi$
\begin{equation}
\begin{aligned}
&\dot t = \frac{E g_{\varphi \varphi} + L g_{t \varphi}}{g_{t \varphi}^2 - g_{tt}g_{\varphi \varphi}},\\&
\dot \varphi = - \frac{E g_{t \varphi} + L g_{tt}}{g_{t \varphi}^2 - g_{tt}g_{\varphi \varphi}}.
\end{aligned}
\label{eq:dottdotp}
\end{equation}
Considering orbits restricted to the equatorial plane, $\theta = \pi/2$, the condition associated with the normalization of the four-velocity of the particles becomes
\begin{equation}
-\epsilon = g_{tt} \dot t^2 + g_{rr} \dot r^2 + g_{\varphi \varphi} \varphi^2 +2 g_{t \varphi} \dot t \dot \varphi,
\end{equation}
with $\epsilon = \{0,1,-1\}$ for a massless, massive and tachyon particle, respectively. We disregard $\epsilon=-1$ from now on. Substituting the expressions of Eq. \eqref{eq:dottdotp} in the above condition, and solving for $\dot r^2$, we can define the effective potential
\begin{equation}
U_{eff} = \frac{1}{g_{rr}} \left(-\epsilon + \frac{E^2 g_{\varphi \varphi} + 2E L g_{t\varphi} + L^2 g_{tt}}{g_{t \varphi}^2 - g_{tt}g_{\varphi \varphi}}\right),
\end{equation}
such that
\begin{equation}
\dot r^2 = U_{eff}.
\end{equation}
The conditions for a circular orbit are $\dot r = 0$ and $\ddot r = 0$, from which follows that
\begin{equation}
U_{eff} = 0, \qquad \frac{dU_{eff}}{dr} \equiv U_{eff}' = 0,
\end{equation}
at the location of orbit. The dash denotes a derivative with respect to $r$. These conditions can further be rearranged into algebraic equations that must be satisfied simultaneously
\begin{equation}
\begin{aligned}
&E^2 g_{\varphi \varphi} + 2E L g_{t\varphi} + L^2 g_{tt} - \epsilon \left(g_{t \varphi}^2 - g_{tt}g_{\varphi \varphi}\right) = 0,\\&
E^2 g_{\varphi \varphi}' + 2E L g_{t\varphi}' + L^2 g_{tt}' - \epsilon \left(g_{t \varphi}^2 - g_{tt}g_{\varphi \varphi}\right)' = 0.
\end{aligned}
\label{eq:circular_orbits1}
\end{equation}
\noindent \textbf{Light Rings}\\
For a light particle, $\epsilon=0$. In this case, calculations are simpler than in the massive case. Solving the first equation for $L$ in \eqref{eq:circular_orbits1} and substituting in the second we obtain
\begin{equation}
g_{\varphi \varphi}' + 2 g_{t \varphi}' \left(\frac{g_{t\varphi} \pm \sqrt{g_{t \varphi}^2 - g_{tt}g_{\varphi \varphi}}}{g_{tt}}\right) + g_{tt}' \left(\frac{g_{t\varphi} \pm \sqrt{g_{t \varphi}^2 - g_{tt}g_{\varphi \varphi}}}{g_{tt}}\right)^2 = 0,
\end{equation}
which is to be evaluated on a radius $r$. The smallest root of the above equation is the location of the light ring.
In Boyer-Lindquist coordinates the location of the circular photon orbits of a Kerr black hole are given by \cite{Rezzolla:2016}
\begin{equation}
r_{BL}^{LR \pm} = 2M \left(1+\cos\left(\frac{2}{3}\arccos\left(\mp \chi\right)\right) \right),
\end{equation}
where the plus sign refers to co-rotating photons, and the minus sign to counter-rotating photons. In quasi-isotropic coordinates the location of the circular photon orbits can be obtained using the inverse transformation in Eq. \eqref{eq:coord_transf_inv}.
\noindent \textbf{ISCO}\\
For a massive particle, $\epsilon=1$. The ISCO is located at a saddle point of the effective potential, such that the condition $U_{eff}''=0$ should be imposed. This is equivalent to imposing
\begin{equation}
E^2 g_{\varphi \varphi}'' + 2E L g_{t\varphi}'' + L^2 g_{tt}'' - \epsilon \left(g_{t \varphi}^2 - g_{tt}g_{\varphi \varphi}\right)'' = 0,
\label{eq:circular_orbits2}
\end{equation}
in addition to Eq. \eqref{eq:circular_orbits1}. To find the location of the ISCO, we first solve Eq. \eqref{eq:circular_orbits1} for $E$ and $L$ as functions of the metric functions and their first derivatives, and later substitute these onto Eq. \eqref{eq:circular_orbits2}. Similarly to the light-ring case, we obtain a second order equation to be solved for $r$, the smallest root of which corresponds to the location of the ISCO.
In Boyer-Lindquist coordinates the location of the circular massive particle orbits of a Kerr black hole are given by \cite{Rezzolla:2016}
\begin{equation}
r_{BL}^{ISCO\pm} = M \left(3+Z_2 \mp \sqrt{(3-Z_1)(3+Z_1+2Z_2)}\right),
\end{equation}
where
\begin{equation*}
\begin{aligned}
&Z_1 = 1+\left(1-\chi^2\right)^{1/3} \left[(1+\chi)^{1/3} + (1-\chi)^{1/3}\right],\\&
Z_2 = \sqrt{3\chi^2+Z_1^2},
\end{aligned}
\end{equation*}
and the plus sign refers to co-rotating particles, and the minus sign to counter-rotating particles. In quasi-isotropic coordinates the location of the circular orbits can be obtained using the inverse transformation in Eq. \eqref{eq:coord_transf_inv}.
\noindent \textbf{Orbital frequencies at the ISCO and Light Ring}\\
The orbital angular frequency of particles both at the ISCO and light ring is given by
\begin{equation}
\omega_\pm = \frac{\dot \varphi}{\dot t} = \frac{-g_{t\varphi}' \pm \sqrt{g_{t\varphi}'^2 - g_{tt}' g_{\varphi \varphi}'}}{g_{\varphi \varphi}'},
\end{equation}
where the above expression is to be evaluated at the location of the ISCO/light ring, $\omega_+$ is the angular frequency of co-rotating particles and $\omega_-$ is the angular frequency of counter-rotating particles. In the case of a Kerr black hole we have
\begin{equation}
M \omega_\pm = \pm \frac{1}{\sqrt{48 \cos^4 \left(\frac{1}{3}\arccos{\left(\mp \chi\right)}\right)+\chi^2}},
\end{equation}
at the light ring, and
\begin{equation}
M \omega_\pm = \pm \frac{1}{\left(r_{BL}^{ISCO\pm}/M\right)^{3/2} \pm \chi},
\end{equation}
at the ISCO. The orbital frequency at the ISCO is associated with the cut-off frequency of the emitted synchrotron radiation generated from accelerated charges in accretion disks, and the angular frequency at the light ring is related to the time-scale of the response of the black hole when it is perturbed (real part of the frequency of the black hole quasi-normal modes) \cite{Cardoso:2008bp}.
\section{Numerical spinning black hole solutions}
\label{sec:numbhs}
In this section we first validate our numerical infrastructure against well-known results, namely the Kerr black hole, and then proceed to use it to obtain spinning black holes in a modified gravity theory, the Einstein-scalar-Gauss-Bonnet theory.
\subsection{Validating the code against the Kerr black hole}
To validate our numerical infrastructure we will solve the axisymmetric vacuum Einstein equations to numerically obtain the Kerr solution, and compare with analytical results. We choose to solve the the following combination of the field equations
which diagonalize the Einstein tensor with respect to the operator $\partial_r^2 + r^{-2}\partial_\theta^2$:
\begin{equation}
\begin{aligned}
&-\mathcal{E}^{\mu}_{\phantom{\mu} \mu} + 2 \mathcal{E}^{t}_{\phantom{t} t} + \frac{2W r_H}{r^2} \mathcal{E}^{\varphi}_{\phantom{\varphi} t} = 0,\\&
\mathcal{E}^{\varphi}_{\phantom{\varphi} t} = 0,\\&
\mathcal{E}^{r}_{\phantom{r} r} + \mathcal{E}^{\theta}_{\phantom{\theta} \theta} = 0,\\&
\mathcal{E}^{\varphi}_{\phantom{\varphi} \varphi} - \frac{W r_H}{r^2} \mathcal{E}^{\varphi}_{\phantom{\varphi} t} - \mathcal{E}^{r}_{\phantom{r} r} - \mathcal{E}^{\theta}_{\phantom{\theta} \theta} = 0.
\end{aligned}
\label{eq:feqs_combination}
\end{equation}
In Fig. \ref{fig:compare_kerr_sol} we present the results for the comparison of the metric functions obtained numerically with the analytically known ones for a Kerr black hole with $\chi=0.6$. Given that in this case the initial guess cannot be the Kerr metric itself, to obtain the results in Fig. \ref{fig:compare_kerr_sol} we used a Schwarzschild black hole with comparable $r_H$\footnote{We find the code to be robust against initial guesses, converging quickly even when these are somewhat (but not extremely) bad.}. The maximum observed error is of $\mathcal{O}\left(10^{-13}\right)$ for the metric function $h$, with all other metric functions being successfully obtained to machine precision.
\begin{figure}[]
\centering
\includegraphics[width=0.5\textwidth]{images/fKerrPlot.pdf}\hfill
\includegraphics[width=0.5\textwidth]{images/gKerrPlot.pdf}\vfill
\includegraphics[width=0.5\textwidth]{images/hKerrPlot.pdf}\hfill
\includegraphics[width=0.5\textwidth]{images/WKerrPlot.pdf}
\caption{Comparison between the numerical and analytical results for a Kerr black hole with $\chi = 0.6$, using $N_x=42$, $N_\theta = 8$. The maximum observed error is of order $\mathcal{O}\left(10^{-13}\right)$ for the function $h$, with all other functions being obtained to machine precision. A Schwarzschild black hole was used as an initial guess, and we have used $r_H=1$.}
\label{fig:compare_kerr_sol}
\end{figure}
We also explored the whole domain of existence of Kerr black holes, comparing numerically obtained physically relevant quantities with analytical ones, see Fig. \ref{fig:kerr_domain} below. These include the mass $M$, angular momentum $J$, horizon area $A_H$ and Hawking temperature $T_H$ of the black holes. Furthermore, we computed the (normalized) Smarr relation in Eq. \eqref{eq:smarr}. Overall, in all quantities we have found remarkable agreement between numerical and analytical results, with the Smarr relation providing accurate maximum error estimates. We also observe that errors are higher when the black holes approach the extremal case $(\chi \to 1)$. This is because in the extremal limit, our setup is not valid and another metric \"ansatz is needed (see e.g. Ref. \cite{Herdeiro:2015gia}).
\begin{figure}[]
\centering
\includegraphics[width=0.5\textwidth]{images/Kerr_Err_M_50_12.pdf}\hfill
\includegraphics[width=0.5\textwidth]{images/Kerr_Err_J_50_12.pdf}\vfill
\includegraphics[width=0.5\textwidth]{images/Kerr_Err_AH_50_12.pdf}\hfill
\includegraphics[width=0.5\textwidth]{images/Kerr_Err_TH_50_12.pdf}\vfill
\includegraphics[width=0.5\textwidth]{images/Kerr_Err_SMARR_50_12.pdf}\vfill
\caption{Comparison of numerical results for $M$, $J$, $A_H$ and $T_H$ with analytical ones, throughout the domain of existence of Kerr black holes. Each point represents a different black hole solution. Numerical results were obtained using $N_x=50$, $N_\theta = 12$. We observe remarkable agreement and small errors overall.}
\label{fig:kerr_domain}
\end{figure}
\subsection{Einstein-scalar-Gauss-Bonnet Gravity}
Einstein-scalar-Gauss-Bonnet (EsGB) theories of gravity are a popular set of scalar tensor theories of gravity that have been extensively studied \cite{Sotiriou:2013qea,Sotiriou:2014pfa,Doneva:2017bvd,Silva:2017uqg,Antoniou:2017acq,Cunha:2019dwb,Dima:2020yac,Herdeiro:2020wei,Berti:2020kgk,Kanti:1995vq,Kleihaus:2011tg,Cunha:2016wzk,Delgado:2020rev}, and which admit black hole solutions different to those of GR. Here we use this set of theories to test our methods and code in a non-trivial, but previously studied setting.
EsGB theories are described by the action
\begin{equation}
\mathcal{S} = \frac{1}{16\pi} \int d^4x \sqrt{-g}\left(R - \left(\nabla \phi\right)^2 + \frac{\alpha}{4}\xi\left(\phi\right) \mathcal{G} \right),
\label{eq:EsGB6}
\end{equation}
where $\phi$ is a real scalar field that couples non-minimally to the Gauss-Bonnet term via the coupling function $\xi(\phi)$, and where $\alpha$ is a coupling constant with dimensions of length squared. No closed-form black hole solutions are known in these models, even in the static case. One is therefore forced to resort to numerical methods to study black holes in these theories.
The field equations of the action \eqref{eq:EsGB6} are
\begin{equation}
\mathcal{E}_{\mu \nu} \equiv G_{\mu \nu} - T_{\mu \nu} = 0,
\end{equation}
where
\begin{equation*}
T_{\mu \nu} = \nabla_\mu \phi \nabla_\nu \phi - \frac{1}{2} g_{\mu \nu} \left(\nabla \phi \right)^2 + \alpha\, P_{\mu \alpha \nu \beta} \nabla^\alpha \nabla^\beta \xi\cbr{\phi},
\end{equation*}
and
\begin{equation*}
P_{\alpha \beta \mu \nu} \equiv \frac{1}{4} \epsilon_{\alpha \beta \gamma \delta} R^{\rho \sigma \gamma \delta} \epsilon_{\rho \sigma \mu \nu} = 2\, g_{\alpha [\mu}G_{\nu] \beta} + 2\, g_{\beta [\nu} R_{\mu] \alpha} -R_{\alpha \beta \mu \nu},
\end{equation*}
is the double-dual Riemann tensor (the square brackets denote anti-symmetrization). The scalar field equation is
\begin{equation}
\mathcal{E}_{\phi} \equiv \Box \phi + \frac{\alpha}{8} \dot \xi(\phi) \mathcal{G} = 0,
\label{eq:sfequation_num}
\end{equation}
where the dot denotes differentiation with respect to the scalar field $\phi$. In the stationary and axisymmetric setting, we find that the scalar field is subject to the boundary conditions \cite{Sullivan:2019vyi,Sullivan:2020zpf}
\begin{equation}
\begin{aligned}
&\partial_x \phi = 0, \qquad x=-1,\\&
\phi = 0, \qquad x=1,\\&
\partial_\theta \phi = 0, \qquad \theta = 0,\pi/2,
\end{aligned}
\end{equation}
while the boundary conditions for the metric functions remain those given above. We therefore choose the same spectral expansion for the scalar field as we did for the metric functions.
Black holes in the EsGB theory should obey the Smarr formula \eqref{eq:smarr}, which becomes
\begin{equation}
M + M_s = 2T_H S + 2 \Omega_H J,
\end{equation}
where\footnote{This relation can also be written as $$ M_s = \frac{1}{4\pi} \int d^3x \sqrt{-g} \left(\nabla \phi\right)^2 \frac{\partial}{\partial \phi} \left(\frac{\xi(\phi)}{\xi'(\phi)} \right),$$ provided the coupling does not obey $\xi(\phi) \propto \xi'(\phi)$ and the scalar field asymptotically vanishes. This is advantageous from a numerical point of view because no second derivatives of the scalar field are required, increasing the accuracy in computing $M_s$.}
\begin{equation}
M_s = -\frac{1}{4\pi} \int d^3x \sqrt{-g} \frac{\xi(\phi)}{\xi'(\phi)} \Box \phi,
\end{equation}
and the entropy is given by Eq. \eqref{eq:entropy_wald} that in the EsGB case becomes
\begin{equation}
S = \frac{A_H}{4} + \frac{\alpha}{8}\int_H d^2x \sqrt{\mathfrak{h}} \xi(\phi) R^{(2)},
\end{equation}
where $R^{(2)}$ is the Ricci scalar of the induced metric on the horizon.
We will focus on two coupling examples, the linear coupling
\begin{equation}
\xi(\phi) = \phi,
\end{equation}
and the exponential coupling
\begin{equation}
\xi(\phi) = e^{\gamma \phi}.
\end{equation}
We find that for the exponential coupling the Smarr relation takes a rather simple form
\begin{equation}
M + Q_s/\gamma = 2T_H S + 2 \Omega_H J,
\label{eq:smarr_esgb_exp}
\end{equation}
where $Q_s$ is the scalar charge of the solution, appearing in the asymptotic expansion of the scalar field
\begin{equation*}
\phi \approx \frac{Q_s}{r} + \mathcal{O}\left(r^{-2}\right).
\end{equation*}
It can also be proved that for the linear coupling the following relation holds \cite{Prabhu:2018aun}
\begin{equation}
Q_s = 2\pi \alpha T_H.
\label{eq:rel_esgb_lin}
\end{equation}
In what follows we use the relations in Eqs. \eqref{eq:smarr_esgb_exp} and \eqref{eq:rel_esgb_lin} to address the accuracy of our numerical solutions for the exponential and linear couplings respectively. This is necessary as closed-form solutions are unknown. We use the same combination of field equations as in the Kerr case (Eq. \eqref{eq:feqs_combination}), along with the scalar field equation \eqref{eq:sfequation_num} to solve the system. To solve the system we use a comparable Kerr black hole as an initial guess for the metric functions, and for the scalar field we use the perturbative solution \cite{Sullivan:2019vyi,Sullivan:2020zpf}
\begin{equation}
\phi \approx \frac{\alpha}{r_H^2} \frac{415-1047 x+942 x^2-358 x^3+51 x^4-3 x^5}{12 (-3+x)^6}.
\end{equation}
We present the accuracy estimate results (in a part of the domain of existence) using the Smarr relation for the exponential coupling and the relation in Eq. \eqref{eq:rel_esgb_lin} for the linear coupling in Fig. \ref{fig:EsGB_domain_example_exp}.
\begin{figure}[]
\centering
\includegraphics[width=0.6\textwidth]{images/SmarrExp.pdf}\vfill
\includegraphics[width=0.6\textwidth]{images/SmarrLin.pdf}
\caption{Smarr relation (top) and relation in Eq. \eqref{eq:rel_esgb_lin} (bottom) for numerical solutions in a part of the domain of existence for the theory with the exponential coupling with $\gamma=1$ and linear coupling, respectively, for different values of $\alpha/r_H^2$. Each point represents a different black hole solution. Numerical results were obtained using $N_x=50$, $N_\theta = 12$. We observe small errors, similarly to the Kerr case.}
\label{fig:EsGB_domain_example_exp}
\end{figure}
We observe that errors, as measured by the relations \eqref{eq:smarr_esgb_exp} and \eqref{eq:rel_esgb_lin}, are small and similar to those presented for the Kerr black hole in Fig. \ref{fig:kerr_domain}, despite a dramatic increase in the complexity and number of terms in the field equations. Our results also agree remarkably well with perturbative solutions, such as the ones obtained in Ref. \cite{Sullivan:2020zpf}.
\begin{figure}[]
\centering
\includegraphics[width=0.5\textwidth]{images/VaryNx.pdf}\hfill
\includegraphics[width=0.5\textwidth]{images/VaryNy.pdf}
\caption{Smarr relation for numerical solutions with a dilaton coupling ($\gamma=1$) as a function of the resolution in $x$ (left) and $\theta$ (right). We observe exponential convergence to as the resolution is increased.}
\label{fig:EsGB_VaryN}
\end{figure}
As another test to the code, in Figure \ref{fig:EsGB_VaryN} we plot the accuracy as estimated by the Smarr relation \eqref{eq:smarr_esgb_exp} as a function of both resolutions $N_x$ and $N_\theta$. We observe exponential convergence, similarly to the toy model presented in Fig. \ref{fig:odeerror}. Note that the Smarr relation provides only an estimate of maximum error -- recall the Kerr case, where most metric functions were actually obtained to a precision of $\sim \mathcal{O}\left(10^{-16}\right)$ but the Smarr relation attained errors on the order of $\sim \mathcal{O}\left(10^{-13}\right)$.
To further demonstrate the capabilities of our code, in the following we present some results for the physical properties of EsGB black holes. A plot of the ergoregion for a dilaton black hole with $\gamma=1$, $\chi=0.1$ and $\alpha/M^2=1.15$ can be found in Fig. \ref{fig:ergo_EsGB}.
\begin{figure}[]
\centering
\includegraphics[width=0.5\textwidth]{images/ergo_sGB.pdf}\hfill
\caption{Ergosphere for a EsGB dilaton black hole with $\gamma=1$, $\chi=0.1$ and $\alpha/M^2=1.15$ (red), together with the ergosphere of a Kerr black hole with the same $\chi$ (blue). The event horizon for both is presented in black.}
\label{fig:ergo_EsGB}
\end{figure}
In Fig. \ref{fig:petrov_EsGB} we plot $|1-S|$ as a function of $x$ and $\theta$, where S is the speciality index defined in Eq. \eqref{eq:specindex}, for the same EsGB black hole as before, where we can observe that the spacetime is not algebraically special, being Petrov type I. Spinning EsGB black holes were always observed to be Petrov type I.\footnote{With our numerical setup, a Kerr black hole typically yields values of $|1-S|$ on the order of $10^{-15}$ everywhere, in good agreement with the fact that it is Petrov type D.}
\begin{figure}[]
\centering
\includegraphics[width=0.7\textwidth]{images/petrov_esgb.pdf}
\caption{$|1-S|$ plotted as a function of $x$ and $\theta$, where S is the speciality index defined in Eq. \eqref{eq:specindex}, for a EsGB dilaton black hole with $\gamma=1$, $\chi=0.1$ and $\alpha/M^2=1.15$. The non-vanishing value of $|1-S|$ demonstrates that the spacetime is Petrov type I.}
\label{fig:petrov_EsGB}
\end{figure}
The perimetral location and angular frequencies at the ISCO and light rings of EsGB dilaton black holes ($\gamma=1$) are compared with those of a Kerr black hole (with the same $\chi$ and $M$) in Fig. \ref{fig:ISCO_EsGB}. Note that we have neglected any couplings between the dilaton and matter (see e.g. \cite{Kleihaus:2015aje,Pani:2009wy}). We have compared our results in the static and slowly rotating cases with those in Ref. \cite{Pani:2009wy}, observing remarkable agreement (in the appropriate setup). From Fig. \ref{fig:ISCO_EsGB} we observe differences of a few percent in most cases, with the most drastic differences occurring for the location of the co-rotating light ring due to its proximity to the horizon. The qualitative behaviour is as follows: the perimetral radius of both the ISCO and the light ring decreases with $\alpha/M^2$, and the opposite happens for the angular frequencies\footnote{We note that, similarly to Refs. \cite{Sullivan:2019vyi,Sullivan:2020zpf}, positive coordinate shifts in the location of the ISCO/light ring were observed. These are, however, not physically relevant and the perimetral radius should be used, where negative shifts are observed.}. Co-rotating orbits are most affected, and black hole spin enhances the differences of co-rotating orbits with respect to the Kerr case.
\begin{figure}[]
\centering
\includegraphics[width=0.5\textwidth]{images/Risco_sgb.pdf}\hfill
\includegraphics[width=0.5\textwidth]{images/wisco_sgb.pdf}\vfill
\includegraphics[width=0.5\textwidth]{images/RLR_sgb.pdf}\hfill
\includegraphics[width=0.5\textwidth]{images/wLR_sgb.pdf}
\caption{Comparison between EsGB dilaton ($\gamma=1$) and Kerr black holes with the same $\chi$ (and $M$) regarding the perimetral radius and angular frequencies at the ISCO (top) and light ring (bottom) as a function of $\alpha/M^2$, in a part of the domain of existence of solutions.}
\label{fig:ISCO_EsGB}
\end{figure}
\subsection{Comparison with other codes}
Similar codes to the one we have developed in this chapter are scarce. Indeed, most of the numerical studies of spinning black holes in modified theories of gravity make use of the non-publicly-available FIDISOL/CADSOL solver \cite{Solver1,Solver2,Solver3}, which implements a finite difference method together with the root finding Newton-Raphson method. The solver is written in Fortran and was first developed in the eighties. Works that use the FIDISOL/CADSOL solver can be found e.g. in Refs. \cite{Herdeiro:2014goa, Herdeiro:2015gia, Herdeiro:2016tmi,Delgado:2020rev, Kleihaus:2015aje,Kleihaus:2011tg,Herdeiro:2020wei,Berti:2020kgk,Cunha:2019dwb}. Some of these works have applied the FIDISOL/CADSOL solver in studies of EsGB gravity, much like we did here. However, they report an error of order $\mathcal{O}\left(10^{-3}\right)$, as estimated by the Smarr relation. In the appendix of Ref. \cite{Delgado:2022pwo}, the author gives a comprehensive overview of the FIDISOL/CADSOL solver, benchmarking it against the Kerr solution, with results again showing errors several orders of magnitude higher than those presented in Fig. \ref{fig:kerr_domain}.
More recently, in Ref.~\cite{Sullivan:2020zpf} the authors developed the \emph{eXtreme Partial Differential Equations Solver} (XPDES) code which is publicly available, to address similar problems. The code is written in C language, and implements a finite difference method to solve the field equations, similarly to the FIDISOL/CADSOL package. It makes use of the software \emph{Maple} to export the field equations to many large C programming files.
Ref. \cite{Sullivan:2020zpf} does not discuss errors as estimated by Smarr relations, instead, they (also) benchmark their code against the Kerr solution, and compare their EsGB results to perturbative solutions, finding good agreement. They report typical maximum errors on obtaining the Kerr solution of $\mathcal{O}\left(10^{-6}\right)$, which represents a good improvement when compared with the FIDISOL/CADSOL package, especially given that the XPDES code is open-source and publicly available.
Our code is written in the Julia programming language, which when compared with complied languages such as C code makes it logistically easier to use and adapt, and to implement
new models.
In our implementation the field equations and boundary conditions are written in a very simple way. For example, the boundary condition
\begin{equation*}
f-2\partial_x f = 0,
\end{equation*}
is written as a residual in code language as
\begin{equation*}
f - 2*dfdx.
\end{equation*}
The code is memory efficient and fast, making use of pseudospectral methods as explained above, with solutions to the field equations being obtained in the order of a few seconds in laptop-class computers. In our (limited) comparisons with the XPDES code, we found that where our code took only a few seconds the XPDES code would take minutes to achieve a lower accuracy.
The results of this section, for example in Figs. \ref{fig:kerr_domain} and \ref{fig:EsGB_domain_example_exp}, show that the accuracy of our code is many orders of magnitude better than the accuracy presented by either the FIDISOL/CADSOL package or even the XPDES code.
Once a solution to the field equations has been obtained, our code has built-in functions to compute all the physical properties of the black holes discussed in section \ref{sec:physical_properties}, therefore allowing for a simple and comprehensive study of different models.
\section{Conclusions}
In this paper we have reviewed the spectral method for solving differential equations and subsequently argued that such methods are ideal for finding stationary and axisymmetric black hole solutions in modified theories of gravity. In particular, they allow complicated field equations and boundary conditions to be implemented in a straightforward manner. We showed how this can be done, and have implemented the method in a new code. To show it in action, and to benchmark its performance against other codes, we applied the code in the GR setting, and verified that the solution found is extremely close to the known Kerr black hole. We then applied it to a popular set of modified theories of gravity, Einstein-scalar-Gauss-Bonnet gravity, where it is known that black hole solutions different from Kerr exist. In this latter setting we verified the accuracy using analytical expressions that should hold identically. We found that even in the Gauss-Bonnet setting our code took just seconds to find accurate spinning black hole solutions.
Within the code we have also implemented many built in functions to calculate black hole properties of physical interest. In the future, obtained solutions together with these functions could be used to study a huge range of phenomena observational interest. Other possible studies include the quasi-normal modes of black hole mergers (hence permitting realistic data analysis with Bayesian methods), the electromagnetic emission from accretion disks, black hole shadows, and our code's solutions could also be used as seed solutions for numerical evolutions. Given that the code has been completed only recently, we have, however, not yet applied it widely. Although a first application in research work to EsGB theories is contained in Ref.~\cite{Fernandes:2022kvg}. In the future we hope to apply the code to other theories, such as the so called regularized 4D-Einstein-Gauss-Bonnet gravity theory \cite{Glavan:2019inb,Fernandes:2022zrq,Lu:2020iav,Kobayashi:2020wqy,Fernandes:2020nbq,Hennigar:2020lsl,Fernandes:2021dsb,Aoki:2020lig,Fernandes:2021ysi} where thus far spinning black holes have not been found, and use it to further understand and constrain such theories using the physical properties described.
\section*{Acknowledgements}
P.F. acknowledges support by the Royal Society Grant No. RGF/EA/180022 and is supported by a Research Leadership Award from the Leverhulme Trust. D.J.M. is supported by a Royal Society University Research Fellowship.
\begin{appendices}
\section{The Kerr-Newman Black Hole}
\label{ap:KN}
The Kerr-Newman solution solves the Einstein-Maxwell field equations
\begin{equation}
G_{\mu \nu}=2\left(F_{\mu}^{\phantom{\mu} \alpha} F_{\nu \alpha} - \frac{1}{4} g_{\mu \nu} F_{\alpha \beta} F^{\alpha \beta}\right).
\end{equation}
The Einstein-Maxwell field equations can be obtained with the following action principle
\begin{equation}
\mathcal{S} = \frac{1}{16\pi} \int d^4x \sqrt{-g} \left(R - F_{\mu \nu} F^{\mu \nu}\right),
\end{equation}
where $F_{\mu \nu} = \nabla_\mu A_\nu - \nabla_\nu A_\mu$ is the Maxwell tensor. With the ans\"atz of Eq. \eqref{eq:metric} the Kerr-Newman black hole solution reads (in terms of $r_H$, $M$ and $Q$)
\begin{equation}
\begin{aligned}
&f_{\mathrm{KN}} = \left(1+\frac{r_H}{r}\right)^2 \frac{\mathcal{A}}{\mathcal{B}},\\&
g_{\mathrm{KN}} = \left(1+\frac{r_H}{r}\right)^2,\\&
h_{\mathrm{KN}} = \frac{\mathcal{A}^2}{\mathcal{B}},\\&
W_{\mathrm{KN}} = \frac{r \left(2 M^2-Q^2\right)+2 M \left(r^2+r_H^2\right)}{r_H r^3 \mathcal{B}}\sqrt{M^2-Q^2-4 r_H^2}
\end{aligned}
\label{eq:kerr-newman}
\end{equation}
where
\begin{equation}
\begin{aligned}
&\mathcal{A} = \frac{r^2 \left(2 M^2-Q^2\right)+2 M r \left(r^2+r_H^2\right)+\left(r^2-r_H^2\right)^2}{r^4}-\frac{\left(M^2-Q^2-4 r_H^2\right)}{r^2}\sin^2\theta ,\\&
\mathcal{B} = \left(\mathcal{A} + \frac{ \left(M^2-Q^2-4 r_H^2\right) }{r^2} \sin^2\theta\right)^2 - \frac{\left(r^2-r_H^2\right)^2 \left(M^2-Q^2-4 r_H^2\right)}{r^6} \sin^2\theta,
\end{aligned}
\end{equation}
together with the four-potential
\begin{equation}
A_\mu dx^\mu = \left(\tilde{A}_t - \frac{W_{KN}}{r}\left(1-\mathcal{N}\right) \tilde{A}_\varphi \sin^2 \theta \right) dt + \tilde{A}_\varphi \sin^2 \theta d\varphi,
\end{equation}
where
\begin{equation}
\tilde{A}_\varphi = \frac{Q r \left(1 + \frac{M}{r} + \frac{r_H^2}{r^2}\right) \sqrt{M^2-Q^2-4 r_H^2} }{r^2 \left(1 + \frac{M}{r} + \frac{r_H^2}{r^2}\right)^2+\left(M^2-Q^2-4 r_H^2\right)\cos^2\theta},
\end{equation}
and
\begin{equation}
\tilde{A}_t = \Phi - \frac{Q r \left(1 + \frac{M}{r} + \frac{r_H^2}{r^2}\right)}{r^2 \left(1 + \frac{M}{r} + \frac{r_H^2}{r^2}\right)^2+\left(M^2-Q^2-4 r_H^2\right)\cos^2\theta} + \frac{W_{KN}}{r}\left(1-\mathcal{N}\right) \tilde{A}_\varphi \sin^2 \theta,
\end{equation}
where $Q$ the electric charge and $\Phi$ the electrostatic potential (which can be chosen such that $\tilde{A}_t|_{r_H} = 0$). This particular choice of functions $\tilde{A}_t$ and $\tilde{A}_\varphi$ for the vector potential is such that they are optimised for a numerical setup such as ours.
The total angular momentum (per unit mass), $a$, of the solution is related to $M$, $Q$ and $r_H$ via
\begin{equation}
r_H = \frac{\sqrt{M^2-a^2-Q^2}}{2} \equiv \frac{M}{2}\sqrt{1-\chi^2-q^2},
\end{equation}
where we have defined the dimensionless charge
\begin{equation}
q \equiv Q/M.
\end{equation}
The electric charge can be read off the asymptotic decay of the temporal part of the four potential
\begin{equation}
\tilde{A}_t = \Phi -\frac{Q}{r} + \mathcal{O}\left(r^{-2}\right).
\end{equation}
The Kerr-Newman black hole obeys the well-known Smarr relation
\begin{equation}
M = 2 T S + 2\Omega_H J + \Phi Q.
\end{equation}
Note that the Kerr-Newman black hole in the quasi-isotropic coordinate system presented in Eq. \eqref{eq:metric} can be obtained from the standard textbook Boyer-Lindquist coordinates solution with the radial coordinate transformation
\begin{equation}
r_{BL} = r + M + \frac{M^2 - a^2 - Q^2}{4r} = r \left(1 + \frac{M}{r} + \frac{r_H^2}{r^2}\right).
\end{equation}
Details about marginal stable circular orbits in the Kerr-Newman case can be found in Refs. \cite{LRISCO_KN, Wang:2022ouq}.
\end{appendices}
\section*{Bibliography}
|
\section{Introduction}
The Large Hadron Collider (LHC) is currently the highest energy particle collider in the world. It accelerates beams of protons to almost the speed of light and then collides them at a centre-of-mass energy of 13.6 TeV at the centre of large, multi-purpose particle detectors that are designed to reconstruct the outcome of those collisions. Among the key physics objectives of the LHC are precise measurements of the properties of the Higgs boson, shedding light on the elusive particle(s) that may constitute dark matter, and searching for a wide breadth of new physics phenomena beyond the Standard Model (SM) via exotic decay signatures like long-lived particles.
To attain these physics goals, the LHC is preparing for an upgrade that will deliver an order of magnitude more data to the experiments by increasing the intensity of the proton beams, resulting in a higher instantaneous luminosity and thus many more collisions taking place every time the proton bunches cross \cite{https://doi.org/10.48550/arxiv.1902.10229}. At this upgraded High Luminosity LHC (HL-LHC) the number of concurrent, overlapping proton-proton interactions (pileup) is expected to reach up to 200, a significant increase from the current average pileup of 40. Such a step change in the running conditions of the collider will significantly increase our capabilities to fulfil the goals of the LHC programme. However, it also presents challenges. The significant increase in detector occupancy will impact the performance of the entire pipeline, including data acquisition, processing, and analysis, as well as simulating the collisions in the detector. This presents significant overhead on the computational resources, with some elements, such as reconstructing charged particle trajectories, becoming a major bottleneck.
To address these high demands on the computational resources, numerous approaches are being pursued, ranging from the development of more efficient algorithms and the application of state-of-the-art machine learning techniques to the use of graphics processing units (GPUs)~\cite{ATLAS:2802799,Contardo:2020886} to execute code that is parallelizable. One of the intriguing new avenues being pursued to tackle these challenges is quantum computing. This new paradigm offers a fundamentally new form of computing by leveraging the phenomena of quantum mechanics and opens the prospect of significantly speeding up our current algorithms and performing calculations that could only be done to some approximation with classical computers.
Particle physics has seen a surge of interest in ascertaining how quantum computers may impact the future of the field and establishing the scenarios in which they may be most advantageous. The current Noisy Intermediate Scale Quantum (NISQ) devices~\cite{Bharti_2022}, while a stepping stone on the way to universal, fault-tolerant quantum computers, have enabled many of these proof-of-principle studies to be performed. This exploratory phase of applying current NISQ era quantum computers to challenging problems in particle physics will pave the way for the emergence of new ideas and techniques needed to fully exploit quantum computation and identify the specific problems for which they are most suitable.
Quantum computing algorithms have been studied for a range of different scenarios in high energy physics. The calculation of simple scattering processes via the helicity spinor formalism and the simulation of a parton shower was demonstrated in~\cite{Bepari_2021}. A quantum walk framework was proposed in~\cite{Bepari:2021kwv}, demonstrating that the parton shower is more naturally and efficiently simulated using a quantum walk in two dimensions. Quantum computing has also been applied to jet clustering~\cite{Wei:2019rqy,Pires:2021fka,PhysRevD.106.036021}, classification of collisions of interest from those that are not~\cite{Terashi:2020wfi,dur34540,Wu:2021xsj}, and anomaly detection in searches for new physics~\cite{Ngairangbam:2021yma}.
The challenging task of connecting the hits left by charged particles in the tracking detector and associating them with the same particle has been studied from several different perspectives including: quantum associative memory to store all the different track patterns and subsequently employ Grover's search algorithm to search through the database and recall the right track pattern~\cite{Shapoval:2019txi}; quantum graph neural networks~\cite{Tuysuz:2020ocw,Tuysuz:2020gjh}; and quantum annealing devices to minimise an objective function~\cite{Zlokapa:2019tkn,refId0,bapst2020pattern}.
This paper approaches the problem of track reconstruction by proposing a hybrid quantum-classical algorithm that uses a support vector machine with a quantum-estimated kernel. The problem is decomposed into that of classifying short segments of tracks. Often such segments can form the `seeds' for extrapolating to the full trajectory of the track. This `seeding' step is expected to be a large consumer of CPU time at the HL-LHC~\cite{ATLAS:2802799}. Simplifications are implemented to fit the limitations of the presently available quantum simulators.
\section{Data preprocessing}
This study utilises the TrackML dataset~\cite{amrouche:2019wmx,amrouche2021tracking} which has been widely used for proof-of-principle studies of classical machine learning algorithms and quantum-based approaches. The dataset provides a simplified simulation of the detector geometry and conditions expected at the HL-LHC. It features a silicon tracking detector with 9 cylindrical layers in the central region and disk geometry in the forward regions, which is typically representative of the ATLAS~\cite{ATLAS:2008xda} and CMS detectors~\cite{CMS:2008xjf}. The detector is segmented into three sub-detectors differing in their spatial resolution, with the inner pixel detector comprising of 4 layers, followed by a short strip detector of 4 layers and then a 2-layer long strip detector. These tracking detectors are immersed in a strong magnetic field aligned with the direction of the proton-proton beam, so charged particles moving through these detectors will typically follow an approximately helical trajectory and show curved trajectories in the transverse $x-y$ plane, the plane perpendicular to the beam line. Figure~\ref{fig_detector} shows the layout of this virtual detector used to produce the TrackML dataset and the coverage of each sub-detector in the $r-z$ plane, where $r$ is the radial dimension and measures the distance from the beam line and $z$ is the distance along the beam line. For the analysis in this paper, only the hits in the barrel region of the detector are used to reduce the total number of hits to a level that can be processed within the current computational constraints.
\begin{figure}[h!]
\includegraphics[width=0.5\textwidth]{figs_for_transfer.003.jpeg}
\caption{A schematic of the virtual general-purpose detector simulated in the TrackML challenge and the coverage of each sub-detector in the r-z plane. Highlighted is the barrel region used in the analysis. The numbers indicate the various detector components and layers respectively. Original image is taken from~\cite{amrouche:2019wmx}.}
\label{fig_detector}
\end{figure}
The TrackML dataset contains 10,000 simulated events. The process of interest is top-antitop production and overlaid on this `signal' are 200 additional proton-proton collisions to simulate the conditions expected at the HL-LHC. This results in an average of 100,000 hits per event in the tracking detector which must be associated with approximately 10,000 tracks.
\begin{figure}[h!]
\includegraphics[width=4.0in]{paper_show_hits.png}
\includegraphics[width=4.0in]{paper_show_tracks.png}
\caption{The 6518 hits in an example event in the $x$--$y$ detector plane (top) and a fraction of the true tracks reconstructed from those hits (bottom). The hits come from 879 particles which produced triplets in the barrel region. The 10 layers of the detector for the barrel region are also shown. The blue, red, and green layers correspond to the pixel, short strip and long strip detectors, respectively.}
\label{fig_hits}
\end{figure}
The 3-dimensional spatial information for every hit in the detector is provided and this information is used to build the track candidates. The total number of possible combinations of those hits that can lead to a track is very large. Identifying the correct combination of hits that reconstruct the true trajectory of a particle is thus a challenging combinatorics problem. Figure~\ref{fig_hits} illustrates this by showing all hits for an event in the $x$--$y$ plane of the detector and a fraction of true reconstructed tracks formed from a combination of those hits.
To avoid unphysical hit sequences which would dominate the resulting dataset, selection criteria are applied to reduce the number of possible connections between hits in each event such that they can be processed without overburdening the computational resources. In addition, the problem is formulated as a classification task, with track segments consisting of a set of 3 hits in adjacent layers of the detector, called triplets, being classified as belonging to a single particle track or not.
The hits are described by three coordinates; $r,\phi$, and $z$, where $\phi$ is defined as the angle around the $z$ axis. A total of 300 events have been processed for classification. The first step in constructing the triplets is to make a dataset of doublets, which are defined as two consecutive hits in the detector. Selection criteria are applied to reduce the size of the doublet dataset and improve its quality. The following observables are used in the selection; the intercept from the extrapolation of the doublet to the $z$ axis, $z_{0}$, and the ratio $\frac{\Delta \phi}{\Delta r}$, as calculated from the difference in $\phi$ and $r$ between each hit forming the doublet. This selection is summarised in Table~\ref{tab:edge_cuts}, and was originally implemented in \cite{Farrell:2018cjr}.
The selection of triplets is based on the estimation of the transverse momentum ($p_{T}$) as determined from the three hits, the $\theta$-breaking angle and the $\phi$-breaking angle. The angle $\theta$ is defined in the $r$-$z$ plane and a breaking angle is that between the straight lines (connecting the two hits in a doublet) of two doublets that form a triplet. The triplet selection is summarised in Table~\ref{tab:triplet_cuts}.
\begin{table}[h!]
\begin{tabular}{c|c}
Variable & Selection\\ \hline
$\frac{\Delta \phi}{\Delta r}$ & $\le 6 \times 10^4 $ [$\frac{\text{rad}}{\text{mm}}$] \\
$|z_0|$ & $\le$ 100 [mm] \\
\end{tabular}
\caption{The selection criteria applied to select doublets, using the $z_{0}$ intercept from the extrapolation of the doublet to the $z$ axis and the ratio of the difference in $\phi$ and $r$ between each hit forming the doublet.}
\label{tab:edge_cuts}
\end{table}
\begin{table}[h!]
\begin{tabular}{c|c}
Variable & Selection\\ \hline
$\theta$-breaking angle & $\le 0.05-0.07$ [rad] \\
$\phi$-breaking angle & $\le 0.05 - 0.12$ [rad]\\
$p_T$ & $\ge$ 0.75 [GeV] \\
\end{tabular}
\caption{The selection criteria applied to select triplets based on the estimated $p_{T}$ and the $\theta$ and $\phi$ angles between two doublets that form a triplet. A range of values is given when the selection depends upon detector components traversed.}
\label{tab:triplet_cuts}
\end{table}
\section{Support vector machine}
The proposed algorithm utilises a support vector machine (SVM) \cite{cristianini_shawe-taylor_2000}, where a kernel function is calculated either on a (simulated) quantum or a classical computer. A support vector machine is a supervised machine learning algorithm that classifies data by drawing linear decision boundaries (hyperplanes) between different groups of data. This paper focuses on discriminating between two classes of data. It takes a training dataset of size $N$ of the form $(\mathbf{x}^{1},y^{1}),\ldots ,(\mathbf{x}^{N},y^{N})$, where $\mathbf{x}^{i}$ is an $M$-dimensional vector and $y^{i} =\pm 1$ for data that belongs to one of two classes. The hyperplane is defined by $\langle \mathbf{w} \cdot \mathbf{x} \rangle + b = 0$, where $\mathbf{w}$ is the normal vector to the hyperplane and $b$ is an offset. These parameters are determined during the learning process. For the simple case of linearly-separable data, the training points $\mathbf{x}^{i}$ of the two classes are placed on either side of the decision boundary, satisfying $f(\mathbf{x}^{i}) = \mathrm{sign}\left(\langle \mathbf{w} \cdot \mathbf{x}^{i} \rangle + b\right) = y_i$, where $f(\mathbf{x})$ is called the decision function.
The points closest to the hyperplane are called support vectors and the distance between them and the hyperplane is called the margin. The goal is to optimise the parameters of the hyperplane such that the margin is maximised. Figure~\ref{fig:svm} shows a visual representation of this. Once the hyperplane has been found, a previously unseen data point $\mathbf{z}$ can be classified using the decision function.
\begin{figure}[h!]
\includegraphics[width=3.5in]{figs_for_transfer.001.jpeg}
\caption{A visual representation of two classes of data in a 2-dimensional space, separated by a hyperplane $\langle \mathbf{w}\cdot\mathbf{x}\rangle+b$ (solid line). The highlighted points lying closest to the separation plane are called support vectors and the dotted lines passing through them define the margins.}
\label{fig:svm}
\end{figure}
The decision boundary is usually defined not in the original data space but in a higher-dimensional \textit{feature space} obtained with a feature map $\phi(\mathbf{x})$. This can introduce non-linearity whilst keeping the decision boundary linear. The goal of this operation is to achieve better separation of the two classes. Figure~\ref{fig:feature_map} shows a simple example of a feature map's functionality.
SVMs are an example of a kernel method, where the kernel $k(\mathbf{x}, \mathbf{y}) = \langle \phi(\mathbf{x}) \cdot \phi(\mathbf{y}) \rangle$ is a function with arguments in the original space of the data, defining a distance measure between two points in the feature space. The remarkable property of this function is that it returns the inner product in the feature space, sidestepping the explicit application of the feature map, which can become computationally expensive for sophisticated feature spaces. In support vector machines, this property can be utilised to find the separation hyperplane. This is possible because linear learning machines can be expressed in a dual representation, following the Karush-Kuhn-Tucker theory \cite{sundaram_1996}. During the optimisation of the dual problem, one needs to find a kernel matrix $K_{\mathbf{x,y}} = k(\mathbf{x},\mathbf{y})$ (an $N \times N$ symmetric matrix) from all pairs of training data points.
Expressed in its dual form, the decision function becomes:
\begin{equation}
y(\mathbf{x}) =\mathrm{sign}\left(\sum_{i=1}^{N}y^{i}\alpha^{i}k(\mathbf{x^{i}},\mathbf{x}) + b\right),
\end{equation}
where $\alpha^{i}$ are the coefficients which need to be optimised.
Just like quantum computing, kernel methods perform implicit computations in a possibly intractably-large Hilbert space through the efficient manipulation of data inputs.
\begin{figure}[h!]
\includegraphics[width = 3.5in]{figs_for_transfer.002.jpeg}
\caption{A visual representation of a simple feature map that takes an inseparable dataset in one-dimension to a two-dimensional feature space. Separation with a linear hyperplane is possible in the new feature space.}
\label{fig:feature_map}
\end{figure}
\section{Quantum kernel estimation}
Quantum computers can be utilised in kernel methods if one considers a quantum circuit $\mathcal{U}(\mathbf{x})$ whose gates are parametrised by the original features of some classical data. The result of such a circuit before measurement is a quantum state which exists in a higher-dimensional Hilbert space.
This is equivalent to a feature map. The quantum state is defined as~\cite{Schuld2021}:
\begin{equation}
\mathbf{x} \rightarrow \rho(\mathbf{x}) = |\psi(\mathbf{x})\rangle \langle \psi(\mathbf{x})|,
\end{equation}
where $| \cdot \rangle$ denotes the usual Dirac vector and $\rho(\mathbf{x})$ is obtained via
\begin{equation}
\rho(\mathbf{x}) = \mathcal{U}(\mathbf{x})\rho_{0}\mathcal{U}^{\dagger}(\mathbf{x}),
\end{equation}
with an initial state $\rho_0$. An all-zero initial state is used with $|\psi_0\rangle = |0^{\otimes M}\rangle$.
The kernel associated with such a feature map is obtained from~\cite{Havl_ek_2019}:
\begin{equation}
k(\mathbf{z}, \mathbf{x}) = \text{tr}\{\rho(\mathbf{z}) \rho(\mathbf{x})\} = \left|\langle\psi(\mathbf{z})|\psi(\mathbf{x})\rangle\right|^2.
\label{eq:quantum_kernel}
\end{equation}
This inner product can be calculated from the transition amplitude of two states;
\begin{equation}
\left|\langle\psi(\mathbf{z})|\psi(\mathbf{x})\rangle\right|^2 = \left| \langle0^{\otimes M}|\mathcal{U}^{\dagger}(\mathbf{z})\mathcal{U}(\mathbf{x})|0^{\otimes M} \rangle \right|^2.
\end{equation}
The circuit $\mathcal{U}^{\dagger}(\mathbf{z})\mathcal{U}(\mathbf{x})|0^{\otimes M}\rangle$ is run repeatedly over $R$ identical runs (shots). The fraction of measurements yielding an all-zero output gives an estimation of the kernel function for the two points $\mathbf{x}$ and $\mathbf{z}$, which forms an entry in the kernel matrix. Repeated evaluations for all combinations of the input dataset give the full kernel matrix.
Similar states have large kernel matrix entries while orthogonal points give $k(\mathbf{x}, \mathbf{z}) = 0$.
Feature maps of particular interest are those that are difficult to calculate using classical means whilst providing good classification of data. Ideally, a kernel matrix resulting from Eq. \ref{eq:quantum_kernel} would produce results better than any classical classifier and be calculated significantly faster on a quantum device. The kernel function proposed in~\cite{Havl_ek_2019} is based on the 3-fold forrelation (`Fourier correlation') problem~\cite{doi:10.1137/15M1050902}. The function is conjectured to have an exponential separation in complexity between its quantum and classical estimation. Further discussion of the potential for speedup is presented later.
The kernel-generating circuit is of the form $\mathcal{U}(\mathbf{x}) = U_{\phi(\mathbf{x})} H^{\otimes M} U_{\phi(\mathbf{x})} H^{\otimes M}$ where $H$ is the Hadamard gate and
\begin{equation}
U_{\phi(\mathbf{x})} = \exp{
\left(
i\sum_{S \subseteq [M]}{\phi_S(\mathbf{x})\prod_{i \in S}{Z_i}}
\right)}.
\label{eq:havlicek_map}
\end{equation}
$Z_i$ is a gate rotating the $i$-th qubit around the $Z$ axis on the Bloch sphere by an amount defined by $\phi_S(\mathbf{x})$.
$S$ denotes a subset of qubits. Only subsets with $|S|\leq 2$ are considered. The circuit for kernel estimation with $\mathcal{U}(\mathbf{x})$ in the case of 3-dimensional data is shown in Fig.~\ref{fig:circuit1}.
Ideas for generalising the circuit have been proposed in \cite{Shaydulin_2022} and \cite{https://doi.org/10.48550/arxiv.2012.07725}. Following the latter, we implement unitaries of the form:
\begin{equation}
U_{\phi(\mathbf{x})} = \exp{
\left(
i\alpha\sum_{S \subseteq [M]}{\phi_S(\mathbf{x})\prod_{i \in S}{\sigma^a_i}}
\right)},
\label{eq:upgraded_map}
\end{equation}
where $\sigma^{a} \in X,Y,Z$ and $\alpha$ is a constant factor to regulate the degree of rotation of the qubits. An example of $U_{\phi(\mathbf{x})}$ for 3-dimensional data can be found in Fig.~\ref{fig:circuit2}.
\begin{figure}[!h]
\includegraphics[width=3.6in]{figs_for_transfer.011.jpeg}
\caption{Quantum circuit diagram used to estimate the kernel and determine the inner product between two quantum states shown for data with three features. }
\label{fig:circuit1}
\end{figure}
\begin{figure}[!h]
\includegraphics[width=3.2in]{figs_for_transfer.010.jpeg}
\caption{Circuit diagram used to calculate $U_{\phi(\mathbf{x})}$ in the full circuit, shown for a datapoint with three features, which correspond to the spatial coordinates of a single hit in the tracking detector. The single-qubit gates are shown in pink and two-qubit gates in blue.}
\label{fig:circuit2}
\end{figure}
This quantum-estimated kernel is then used as input to a classical support vector machine which performs the training and classification. The full circuit thus takes events of dimension $M$ and projects them into an $2^{M}$-dimensional quantum space where the hyperplane separating the two classes of data is calculated.
\section{Results}
Results from the quantum-enhanced algorithm presented in this section were obtained with $\alpha = 0.1$, $\phi_{k}(\mathbf{x}) = x_k$, $\sigma^a_k = Z$ for single qubit rotations and $\phi_{l,m}(\mathbf{x}) = (\pi - x_l)(\pi - x_m)$, $\sigma^a_{l,m} = Y_lY_m$ for two-qubit rotations. These were compared to an RBF kernel~\cite{Musavi1992OnTT}, defined as $K_{RBF}(\mathbf{x}, \mathbf{z}) = \exp{\left(-\gamma\lVert \mathbf{x} - \mathbf{z} \rVert^2\right)}$ with $\gamma = 1$. To reduce generalisation error and model complexity, a regularisation term $C$ can be included into the optimisation loss function. The optimal coefficients and classical kernel type were found using a grid search~\cite{scikit-learn} with cross validation and a parameter scan optimising for validation score and training time.
Large values of $C$ result in a larger penalty for overfitting. The optimal value of $C = 10^6$ determined from this optimisation procedure was used in both the classical and quantum kernels.
The metrics used to quantify the performance of the classifiers are defined below, through the confusion matrix shown in Table ~\ref{tab:confusion}.
\begin{table}[h!]
\centering
\begin{tabular}{c|c|c}
& Predicted positive & Predicted negative \\ \hline
Actual positive & True positive (TP) & False negative (FN) \\
Actual negative & False positive (FP) & True negative (TN) \\
\end{tabular}
\caption{Confusion matrix used in defining the performance metrics for the classifiers used in triplet recognition.}
\label{tab:confusion}
\end{table}
\begin{equation}
\text{Accuracy} = \frac{\text{TP+TN}}{\text{TP+FP+TN+FN}},
\end{equation}
\begin{equation}
\text{Efficiency} = \frac{\text{TP}}{\text{TP+FN}},
\end{equation}
\begin{equation}
\text{Purity} = \frac{\text{TP}}{\text{TP+FP}}.
\end{equation}
A good model is expected to score high in all three metrics. Accuracy gives an overall percentage of correct guesses, efficiency is the fraction of actual true objects correctly recognised and purity measures how often the model mistakes a fake object for a true one.
\subsection{Full detector triplets}
The dataset used in the classification consists of triplets that passed the preprocessing step described above, which ensures a sample purity of 52\% and 80\% for doublets and triplets, respectively and a 99\% sample efficiency in both datasets. An average of 4,600 triplets remain per event.
The spatial coordinates of each hit in the triplet are used as the input data to the hybrid algorithm. This results in a 9-qubit circuit. To accommodate the dataset into our current computational constraints, it is further divided into 16 equal sections in the $\phi$ plane, with each section subtending $\frac{2\pi}{16}$ radians in $\phi$. A support vector machine is then defined for each of these regions and the relevant quantum kernel estimated.
The data is divided into 50 events for training and 15 events for testing, equivalent to a total of around 230,000 and 70,000 triplets, respectively.
The performance of both the classical algorithm and the quantum algorithm are evaluated using the three metrics defined above; accuracy, efficiency and purity. Furthermore, since the preprocessing step selects triplets that are more likely to form track candidates, a benchmark scenario is introduced to illustrate the performance of the classical and quantum algorithms on top of the preprocessed data. Triplets are randomly selected from the dataset and the classical and quantum classifiers are compared against this benchmark to demonstrate the improvements in classification accuracy.
\begin{figure}[!h]
\includegraphics[scale=0.55]{paper_good_err_prelim.png}
\caption{Track reconstruction efficiency and purity for triplets in the barrel detector as a function of $\phi$, $|\eta|$, $p_{T}$, particle multiplicity, and the number of hits associated with the track (track length). These are compared for the quantum-estimated kernel and the classical kernel.}
\label{fig:eff_pur_all}
\end{figure}
\begin{figure}[!h]
\includegraphics[scale=0.26]{paper_good_err_prelim_acc_w_size.png}
\caption{Accuracy to identify triplets in the barrel detector as a function of the size of the training dataset for the quantum-estimated kernel, classical kernel and selecting random triplets from the preprocessed dataset. }
\label{fig:acc_all}
\end{figure}
Figure~\ref{fig:eff_pur_all} shows the dependence of the efficiency and purity scores on relevant geometric and kinematic properties of the triplets; the angle $\phi$ of the first (innermost) hit of the triplet, the pseudorapidity $|\eta|$ of the triplet, the true $p_{T}$ of the triplet, the number of true particles in the event (particle multiplicity), and the number of hits corresponding to a true track (track length). For fake triplets, there is some ambiguity in determining the true $p_{T}$ and track length, as the three hits can come from three separate particles. In such cases, the choice was made to define these variables using the particle associated with the first hit.
The two classifiers show mostly comparable performance and a similar dependence on the observables defined above. Efficiencies close to 1.0 are achieved for most bins. Reduced values of the purity are observed in regions with reduced number of triplets for training, such as high-$\eta$ and high-$p_T$.
The accuracy scores of the classical and quantum algorithms as a function of the size of the training data are shown in Fig.~\ref{fig:acc_all}. The training size grows up to the computational limit imposed by the quantum simulator.
Whilst the overall performance of the two algorithms follows a similar trend, the classical algorithm performs slightly better at low training size and the quantum algorithm shows a small advantage for training size above 6,000 triplets. Both algorithms significantly outperform the benchmark scenario of randomly selecting triplets.
It is also instructive to study the performance of these algorithms for different layers of the detector, as the detector occupancy progressively decreases from the inner to the outer layers. Figure~\ref{fig:eff_pur_layers} shows the comparison between the quantum and classical algorithms for the efficiency and purity in layers 1-9 of the detector. While the efficiency and purity are similar between the two algorithms for the full detector, the largest difference in purity occurs for triplets identified in the first layer of the detector (the first hit is in the first layer). Since triplet formation is part of the seeding step used in many track reconstruction algorithms, we study the performance of our algorithms for triplets identified in the inner detector, with the first hit being in the first layer of the detector.
\begin{figure}[!h]
\includegraphics[scale=0.27]{paper_good_err_prelim_eff_pur_layer_id.png}
\caption{Efficiency and purity as a function of the detector layer, starting from the innermost layer nearest the centre of the detector. Layers 1$-$4 correspond to the pixel detector, layers 5$-$8 belong to the short strip detector and layers 9 and 10 are the long strip detector. Triplets are binned by their innermost hit.}
\label{fig:eff_pur_layers}
\end{figure}
\subsection{Innermost triplets}
This section presents results when restricting the classification of triplets to those in the innermost layers of the tracking detector, with the three hits of a triplet in the first three layers. The reduction in the total number of triplets per event allows more events to be processed before reaching the computational limit. The dataset is split into 240 events for training and 60 events for testing, equivalent to about 230,000 and 60,000 triplets, respectively.
The efficiency and purity as a function of triplet parameters are shown in Fig.~\ref{fig:eff_pur_inner} and the accuracy is shown as a function of the size of training data in Fig.~\ref{fig:acc_inner}.
The accuracy indicates a clearer separation between the quantum and classical performances. We see a continued trend of better purity for the quantum-enhanced classifier and a more comparable performance in terms of efficiency.
\begin{figure}[!h]
\includegraphics[scale=0.55]{paper_good_err_prelim_inner.png}
\caption{Track reconstruction efficiency and purity for triplets in the inner detector barrel region as a function of $\phi$, $|\eta|$, $p_{T}$, particle multiplicity, and the number of hits associated with the track (track length). These are compared for the quantum-estimated kernel and a classical kernel.}
\label{fig:eff_pur_inner}
\end{figure}
\begin{figure}[!h]
\includegraphics[scale=0.26]{paper_good_err_acc_w_size_inner.png}
\caption{Accuracy to identify triplets in the inner detector barrel region as a function of the size of the training data for the quantum-estimated kernel, classical kernel and selecting random triplets from the preprocessed data. }
\label{fig:acc_inner}
\end{figure}
\section{Potentials of quantum speedup}
\label{sec:q_speedup}
In general, we can assume that the evaluation of the matrix elements dominates the complexity of the SVM \cite{10.5555/1296233}. Thus, the complexity of the algorithm is $\mathcal{C} = \mathcal{O}(\beta N^2)$, where the $\beta$ factor depends on the kernel type and method used. For a single value of the kernel, the quantum complexity is $\beta_Q = \mathcal{O}(\epsilon^{-2})$ for some additive error $\epsilon$~\cite{Havl_ek_2019,Gentinetta:2022fia}. The current best classical algorithm proposed in \cite{Bravyi:2021ghk} has $\beta_C = \mathcal{O}(\epsilon^{-\frac{2}{3}}2^{\frac{2M}{3}})$. Demanding that $\beta_Q < \beta_C$ for $\epsilon = 10^{-3}$ results in a possible advantage for $M \approx 20$.
In \cite{Gentinetta:2022fia} it has been shown that in order for the full kernel matrix to approximate the ideal kernel, the propagation of the desired accuracy into the classification with an SVM causes $\beta_Q$ to pick up a non-negligible scaling with training size $N$; $\beta_Q \rightarrow \beta_Q' = \mathcal{O}(N^{\frac{8}{3}}\epsilon^{-2})$. It is possible that a similar analysis could introduce an $N$-scaling to $\beta_C$. Regardless, it appears that at the current stage, quantum kernel estimation could provide possible advantage for small datasets where the data has many features. For the triplet classification data presented in this paper, nine features were used. We envisage that possible speedup may be obtained if we extend the length of the track segments considered in the classification task.
Another point to consider for quantum kernel estimation is the noise on current quantum devices. In \cite{Thanasilp:2022agd}, the authors show that the presence of noise can cause kernel entries produced with Eq. \ref{eq:havlicek_map} evaluated over different input data to concentrate around some fixed value. The difference between any kernel entry and that value becomes exponentially small with the number of qubits. This results in an exponential number of shots necessary to resolve kernel entries for successful training. This dependence would have to be added into $\beta_Q$ in order for the required precision to be obtained.
Some proposals for different quantum kernel estimation methods can be found in \cite{Gentinetta:2022fia}, where a probabilistic algorithm calculates only a subset of the kernel entries and \cite{Haug:2021kom} where the estimation of the entire kernel matrix scales linearly with training size $N$.
Further studies could include empirical tests of classical and quantum complexities, study of noise effects in simulations and real quantum devices as well as implementation of other proposed quantum kernel estimation techniques in the context of high energy physics.
\section{Summary}
Reconstructing the trajectories of charged particles at particle colliders like the Large Hadron Collider is a challenging, computationally intensive problem. This is expected to become increasingly complex with the upgraded Large Hadron Collider (HL-LHC) where $\mathcal{O}(10^5)$ hits in the tracking detector must be quickly and accurately connected to form around 10,000 tracks.
This paper presents a hybrid quantum-classical algorithm with a support vector machine (SVM) using a quantum-estimated kernel to classify track segments or seeds for this challenging track reconstruction problem. Using a publicly available dataset that simulates a generic particle detector for the HL-LHC, we apply a selection criteria to select doublets (set of two consecutive hits) and subsequently triplets (set of three consecutive hits). The proposed algorithm classifies these triplets as either belonging to a particle track or not. A comparison is made between the performance of a quantum-estimated kernel, a classical kernel, and randomly selected triplets from the dataset. A similar level of performance is achieved for the quantum and classical algorithms. However, when only the triplets from the inner part of the tracking detector are considered, the quantum algorithm shows an improvement in accuracy scores against the classical algorithm. This is promising as the innermost layers are expected to be the most important for the seeding procedure at the HL-LHC. This is the first implementation of a quantum-kernel SVM approach to the track reconstruction problem.
\section{Acknowledgements}
Sarah Malik and Tim Scanlon are funded by grants from the Royal Society. S\'{e}bastien Rettie acknowledges support from the Banting Postdoctoral Fellowship program, and the Natural Sciences and Engineering Research Council of Canada.
We also acknowledge funding from the STFC.
We thank Simeon Hatzopoulos for helpful research assistance in visualising the TrackML dataset and Mohammad Hassanshahi for valuable discussions.
|
\section{Introduction}
In order to improve Language Models (LMs), the Natural Language Processing field has removed most of the system-induced biases in the last few years. For instance, practices that were once standard such as lemmatization, stemming and feature engineering have progressively disappeared in favor of general architectures trained on huge amounts of data, learning end-to-end which features may be leveraged to attain better performances. However, one essential part of LMs has seen little evolution: tokenization. Tokenizers convert sequences of characters into sequences of tokens (substrings of smaller length) which can then be embedded by the model. Subword tokenization algorithms~\cite{sennrich-etal-2016-neural,wu2016google,kudo-2018-subword} are a specific class of tokenizers designed in such a way that almost every string can be encoded and decoded with very few out-of-vocabulary tokens. They are used in the vast majority of recent LMs, but have been an essential part of NLP systems since much longer~\cite{mielke2021between}.
\begin{figure}
\centering
\includegraphics[width=\linewidth]{images/full_difftok_schema.pdf}
\caption{The differentiable tokenization scheme of MANTa-LM{}. Input bytes are first assigned a \textit{separation probability} using a Sliding Window Attention Transformer. These probabilities are used to compute the contribution of each byte embedding in the pooled representations of the \textit{blocks}. The block embeddings are fed to the Encoder-Decoder layers which predict the masked bytes. All the components are optimized with the LM objective.}
\label{fig:overview_diagram}
\end{figure}
The success of these algorithms can be attributed to several reasons. Firstly, they produce token sequences whose length is greatly reduced compared to the original character sequence. This characteristic is helpful because limitations in compute power and architectural constraints, such as the quadratic complexity with respect to sequence length of Transformers~\cite{vaswani2017attention}, prevent models from processing arbitrary long sequences. Secondly, they compress the corpus using occurrence statistics that may help LMs. For instance, if a word appears frequently in the training corpus, it will be encoded as a single token in the vocabulary and the model will be able to build a representation for that particular token more easily.
However, the induced biases of tokenizers are also harmful for modelization. One such limitation lies in their brittleness to character deformations which are commonly found in real world, noisy data. For instance, BERT's tokenizer~\cite{devlin-etal-2019-bert} encodes ``performance'' as \texttt{[``performance'']} but \mbox{``perfonmance''} as \texttt{[`per', `\#\#fo', `\#\#n', `\#\#man', `\#\#ce']}, which makes it hard for the model to behave similarly in both cases. Moreover, the tokenizer is fixed after its training and is therefore impossible to update, for instance to reflect new domains~\cite{el-boukkouri-etal-2020-characterbert} where tokenization might over-segment specific or technical terms. \citet{clark2022canine} list other issues emerging when using static subword tokenizers, especially when modeling languages with a more complex morphology than English.
To overcome these issues, \textit{tokenization-free} models~\cite{clark2022canine,xue2022byt5,tay2021charformer} produce character-based or byte-based embeddings for LMs instead of subword embeddings. These methods improve the robustness of LMs to naturally occurring noise as well as their expressiveness when dealing with out-of-domain or multilingual data. In order to cope with increased input lengths, some of these methods compress sequences with constant reduction rates obtained using specialized modules~\cite{clark2022canine,tay2021charformer}, subsequently removing any notion of subwords.
We argue that learning a subword tokenization together with input representations in an end-to-end fashion is beneficial for language modeling. In this work, we introduce MANTa{}, a gradient-based tokenizer and embedding module. It can easily be plugged-in to replace the classical combination of fixed tokenizers and trainable subword embedding matrices existing in most encoder-decoder models, without any increase in the total number of trainable parameters. We also introduce MANTa-LM{}, a Transformer encoder-decoder that incorporates MANTa{} and that is trained end-to-end. By learning a soft, adaptive segmentation of input sequences jointly with the LM pre-training objective, MANTa-LM{} produces byte-based representations with sequence lengths similar to those produced by static subword tokenizers. Additionally, by propagating gradients through our soft segmentation module during fine-tuning as well, we are able to adapt the segmentation to new domains, removing the limitations imposed by static subword tokenization.
We show that MANTa-LM{} is robust to noisy text data and able to adapt to new domains while being significantly faster than byte-level models. Interestingly, MANTa{} learns a simple but explainable segmentation using only the LM objective while effectively reducing the length of byte sequences.
In summary, the contributions of this paper are the following:
\begin{itemize}
\item We introduce MANTa{}, a gradient-based tokenization and pooling module that can learn jointly with an encoder-decoder LM;
\item We train MANTa-LM{} on English data and we evaluate its robustness to synthetic and natural variation and its ability to adapt to new domains compared to byte-level models.
\end{itemize}
\section{Related Work}
Non-neural subword-level tokenization methods have dominated in the last few years as the default way to encode textual data, the most used being BPE \citep{sennrich-etal-2016-neural}, WordPiece~\cite{wu2016google} and Unigram \citep{kudo-2018-subword}. However, they have inherent flaws that limit their multilingual performance~\citep{rust-etal-2021-good}, their adaptability to new languages and new domains after pre-training~\citep{el-boukkouri-etal-2020-characterbert,garcia-etal-2021-towards} and the downstream performance of language models in general \citep{bostrom-durrett-2020-byte}.
To alleviate these issues, tokenization-free (or character-level) models leverage characters instead of subwords to build text representations. Some of the first neural networks for sequence generation used characters directly as inputs \citep{sutskever2011generating,graves2013generating}, and following works modified the approach to create input word representations based on characters \citep{kim2016character,Jzefowicz2016ExploringTL,peters-etal-2018-deep}. Similar architectures were recently adapted to work with Transformers \citep{el-boukkouri-etal-2020-characterbert,ma-etal-2020-charbert}. Nevertheless, they still rely on fixed tokenization heuristics (for instance segmenting using whitespaces) which may not be suited to some languages or certain types of language variations. Recent works have tried to remove these induced biases by working purely with characters or bytes as input \citep{clark2022canine,tay2021charformer,xue2022byt5}. However, they either have to use various tricks to reduce the sequence lengths based on other induced biases like downsampling rates \citep{clark2022canine,tay2021charformer} or have extremely low training and inference speeds \citep{xue2022byt5}. \citet{chung2016hierarchical} create tokens in a differentiable manner by predicting frontiers and using the representations of each character inside a ``token'', but it remains unclear how their model could be adapted to be used with newer architectures such as Transformers. \citet{mofijul2022vocabulary} propose to segment tokens using a trained ``frontier predictor.'' Nevertheless, this differentiable tokenizer is not trained with the main language model objective but instead mimics a BPE subword tokenizer, carrying some of its flaws.
\section{\protectMANTa{}}
\subsection{Differentiable Tokenization}
\label{sec:differentiable_tokenization}
Our main contribution is the introduction of an end-to-end differentiable tokenization architecture that consists in softly aggregating input bytes into what we refer to as \textit{blocks}. As an analogy with hard tokenization schemes, blocks can be compared to tokens with smooth borders.
We decompose the tokenization process into several differentiable operations, ensuring that our model can be trained end-to-end. Our approach consists in predicting a segmentation, and then combining byte embeddings according to this segmentation. MANTa{} can be divided in three different parts:
\begin{itemize}
\item Predicting block frontiers using a parameterized layer to assign a probability $p_{F_i}$ to each input byte $b_i$ of being a frontier;\footnote{$F$ in $p_{F_i}$ stands for \textit{Frontier}.}
\item Building a byte-block unnormalized joint distribution using the frontier probabilities $(p_{F_i})_{i \in [1, L]}$ corresponding to a soft assignment from bytes to blocks;
\item Pooling byte representations for each block $B_j$ weighted by the probability of each byte to belong in the current block $P(b_i \in B_j)$.
\end{itemize}
This process results in a sequence of embeddings that can be given directly to the encoder-decoder model. We provide an overview of the entire model in Figure~\ref{fig:overview_diagram}. We also summarize the process of mapping byte embeddings to block embeddings in appendix~\ref{sec:appendix_detailled}.
\subsubsection{Predicting Subword Frontiers}
\label{sec:frontpred}
Our frontier predictor consists in a parameterized module mapping each byte $b_i$ to the probability of being a block frontier $p_{F_i}$. In a first part, we embed each byte $b_i$ to an embedding $e_{b_i}$. Working with bytes instead of characters allows modeling a larger array of symbols while having very small embedding matrices with $256\times hidden\ size$ parameters. Since the input sequences fed to the frontier predictor may be particularly long, we use a Transformer with sliding window attention~\citep{beltagy2020longformer}. This layer achieves a linear complexity with respect to sequence length by computing attention using only a local context. This reduced context forces the model to focus on local surface features rather than long-range dependencies which may be hard to model at the byte level.
We make the assumption that long-range dependencies are not relevant for segmentation and that this reduced context window should not harm the quality of the tokenization.
\subsubsection{Modeling the Byte-Block Assignment}
Once the frontier probabilities $(p_{F_i})_{i \in [1, L]}$ are predicted for the whole sequence, we use them to model an assignment between bytes and block slots. Each byte is given a probability distribution over the available block slots, and the expected block position of a byte in the block sequence increases along the byte sequence (i.e. the next byte is always more likely to be assigned to the next block).
Let us introduce $(B,\ b_i)$, the slot random variables for each byte $b_i$, describing the position of the block containing $b_i$ in the block sequence. In other words, the event $(B=k,\ b_i)$ describes the fact that the $i$-th byte belongs in the $k$-th block. These variables can only take values in $[1, L]$, as there cannot be more blocks than there are bytes. We can model the $(B,\ b_i)$ as a \textit{cumulative sum} of the random variables $F_i$: the position of the block in which a byte belongs is exactly the number of frontier bytes before this one.
Since $F_i \sim \mathcal{B}(p_{F_i})$, we can model the block index variables $B$ depending on the index of the bytes $b$ using the Poisson Binomial distribution $\mathcal{PB}$ which models the cumulative sum of Bernoulli variables: $\left(B,\ b_i\right) \sim \mathcal{PB}\!\left(\left(p_{F_k}\right)_{k \leq i}\right)$. There exists no closed form for this distribution's mass function, but some fast methods have been developed to avoid exploring the whole event tree \cite{BISCARRI201892, poibin_fft}. However, to reduce computational cost, we use a truncated Gaussian kernel $G$ with the same mean and variance to approximate the $(B,\ b_i)$ probability mass function:
$$
\forall k \in [1, L_B], P\!\left(B = k,\ b_i\right) \simeq P_{k,i} \triangleq \frac{1}{Z}G_{\mu_i, \sigma_i}(k)
$$
where $Z=\sum\limits_{1\leq k\leq L_B}G_{\mu_i, \sigma_i}(k)$ is a normalization term, and:
\begin{equation}
\begin{cases}
L_B = \min\left(L, \left(\mu_L + 3\sigma_L\right)\right)\\[11pt]
\mu_i = \sum_{k=1}^{i} p_{F_i} \\[11pt]
\sigma_i = \sqrt{\sum_{k=1}^{i} p_{F_i}\left(1-p_{F_i}\right)}
\end{cases}
\end{equation}
We denote by $P_{k,i}$ the approximation of the probability of membership of the byte $i$ to block $k$. We display an example of this map at different steps during training in Figure~\ref{fig:mapping_example}. We truncate the block sequences after $(\mu_L + 3\sigma_L)$ since all the probabilities beyond this position are negligible.
\iffalse
\begin{figure*}
\centering\small
\begin{subfigure}[b]{0.4\textwidth}
\includegraphics[width=\linewidth]{images/mapping_example_0.png}
\caption{Step 0}
\end{subfigure}\hfill
\begin{subfigure}[b]{0.4\textwidth}
\includegraphics[width=\linewidth]{images/mapping_example_3000.png}
\caption{Step 3,000}
\end{subfigure}
\begin{subfigure}[b]{0.90\textwidth}
~
\end{subfigure}
\begin{subfigure}[b]{0.4\textwidth}
\includegraphics[width=\linewidth]{images/mapping_example_7000.png}
\caption{Step 7,000}
\end{subfigure}\hfill
\begin{subfigure}[b]{0.4\textwidth}
\includegraphics[width=\linewidth]{images/mapping_example_13000.png}
\caption{Step 13,000}
\end{subfigure}
\caption{The block-byte soft assignment $P$ in the first pre-training steps. MANTa{} learns to downsample so that no information is lost through truncation, but also converges towards a sharp segmentation. }
\label{fig:mapping_example}
\end{figure*}
\fi
\begin{figure}[t]
\centering\small
\includegraphics[width=0.85\columnwidth]{images/mapping_example_0.png}\\
Step 0\\[2mm]
\includegraphics[width=0.85\columnwidth]{images/mapping_example_3000.png}\\
Step 3,000\\[2mm]
\includegraphics[width=0.85\columnwidth]{images/mapping_example_7000.png}\\
Step 7,000\\[2mm]
\includegraphics[width=0.85\columnwidth]{images/mapping_example_13000.png}\\
Step 13,000\\
\caption{The block-byte assignment $P$ during the first pre-training steps. MANTa{} learns to downsample input sequences so that no information is lost through truncation, but also converges towards a sharp segmentation.}
\label{fig:mapping_example}
\end{figure}
\subsubsection{Pooling Block Embeddings}
\label{sec:pooling_block}
At this point in the forward pass, we have estimated the position of the block in which each input byte belongs, along with the block sequence maximum plausible length $L_B$.
In order to provide block embeddings to the LM, we now focus on the contribution of each byte to the block given by the block-byte assignment map. For each block position $k \in [1, L_B]$, this map actually provides an unnormalized contribution $(P_{k,i})_{i \in [1, L]}$ of each byte in this block. We can then use the byte embeddings $e_b$ from the frontier predictor described in Section~\ref{sec:frontpred} and, for the $k$-th block, build a block embedding where each byte $b_i$ contributes based on its probability of being in this block $P_{k,i}$.
To build $e^B_k$, the embedding of block $B_k$ in $k\text{-th}$ position, we first compute the weighted byte embeddings ${\left(P_{k,i} \times e^b_i\right)_{i \in [1, L]}}\in \mathbb{R}^H$, with $H$ the hidden size of the byte embeddings. To make the block embeddings aware of the ordering of the bytes (so that \textit{ape} and \textit{pea} can have different representations), we proceed to a depthwise 1-D convolution along the dimension of the bytes after weighting. This convolution also improves the expressiveness of the block embeddings. We discuss our efficient implementation of these operations in Appendix~\ref{sec:appendix_cache}.
We finally apply a max-pooling operation on the contextualized weighted byte embeddings for each block. This yields one embedding per block, with the same dimension as the byte embeddings. We use a linear layer to map the block embeddings to the right input dimension for the encoder-decoder model, i.e. its hidden size.
The final step consists in truncating the block embedding sequence to a fixed length $\hat{L} = \min(L_B, L / K)$ with $K\in \mathbb{N^*}$ a fixed \textit{truncation factor}. This simple heuristic ensures that all sequences fed to the encoder-decoder have a length at least $K$ times shorter than the input byte sequence length. We choose $K=4$ throughout the paper which is in average the number of bytes in an English BPE token. Most importantly, this truncation incentivizes the frontier predictor to produce sufficiently long blocks. We discuss the influence of this mechanism in more depth in Section~\ref{sec:truncation}.
\subsection{Model Training}
We obtain from the differentiable tokenizer and pooling module a sequence of block embeddings that can be used exactly like subword embeddings. Thus, we use an encoder decoder architecture identical to T5~\citep{raffel2020t5}. Nevertheless, since we do not have a fixed subword vocabulary, our decoder operates at the byte level similarly to ByT5~\citep{xue2022byt5}.
\begin{table*}[]
\centering\small
\begin{tabular}{cccccccccc}
\toprule
Model & $|\theta|$ & MNLI & QNLI & MRPC & SST-2 & QQP & STSB & COLA & AVG \\ \midrule
$\text{T5}_{Small}$ & 60M & \textbf{79.7/79.7} & \textbf{85.7} & 80.2/86.2 & 89.0 & \textbf{90.2/86.6} & 80.0 & 30.3 & 76.6 \\[3pt]
$\text{MANTa-LM{}}_{Small}$ (ours) & 57M & 79.2/78.6 & 84.5 & \textbf{82.3/87.2} & \textbf{89.6} & 89.9/\textbf{86.5} & \textbf{81.4} & \textbf{32.0} & \textbf{77.1} \\ \bottomrule
\end{tabular}
\caption{Results on dev sets for the GLUE benchmark for small models following our pre-training procedure.}
\label{tab:glue_small}
\end{table*}
\begin{table*}[t]
\centering\small
\begin{tabular}{cccccccccc}
\toprule
Model & $|\theta|$ & MNLI & QNLI & MRPC & SST-2 & QQP & STSB & COLA & AVG \\ \midrule
$\text{BERT}_{Base}^\dagger$ & 110M & \textbf{84.4} / - & 88.4 & 86.7/- & \textbf{92.7} & - & - & - & - \\[3pt]
$\text{T5}_{Base}^\dagger$ & 220M & 84.2/\textbf{84.6} & 90.5 & \textbf{88.9/92.1} & \textbf{92.7} & \textbf{91.6/88.7} & \textbf{88.0} & 53.8 & 84.3 \\ \midrule
$\text{CharBERT}_{Base}^\mathsection$ & 125M & - & \textbf{91.7} & 87.8/- & - & 91/- & - & \textbf{59.1} & - \\[3pt]
$\text{Byte-level T5}_{Base}^\dagger$ & 200M & 82.5/82.7 & 88.7 & 87.3/91.0 & 91.6 & 90.9/87.7 & 84.3 & 45.1 & 81.5 \\[3pt]
$\text{Charformer}_{Base}^\dagger$ & 203M & 82.6/82.7 & 89.0 & 87.3/91.1 & 91.6 & 91.2/88.1 & 85.3 & 42.6 & 81.4 \\[3pt]
$\text{MANTa-LM{}}_{Base}$ (ours) & 200M & 77.5/78.8 & 88.2 & 82.4/88.2 & 91.3 & 90.8/87.7 & 79.2 & 51.0 & 80.3\\ [3pt] \bottomrule
\end{tabular}
\caption{Results on dev sets for the GLUE benchmark. $\dagger$ indicates results obtained by \citet{tay2021charformer}, which are very similar to our models in terms of compute, but use a smaller batch size which may enhance their performance. $\mathsection$ indicates results obtained by \citet{ma-etal-2020-charbert}. The top section concerns model trained using a subword tokenizer.}
\label{tab:glue}
\end{table*}
\subsubsection{Pre-Training Details}
\paragraph{Objective}
Our objective is identical to the one used in ByT5. We mask 15\% of bytes randomly and choose a number of spans such that each has an average length of 20 bytes. Each span is then replaced by an \texttt{<extra\_id\_i>} token with \texttt{i} identifying the order of the span in the sequence. On the decoder side, the model has to predict in an autoregressive way the span identifier and the masked bytes.
\paragraph{Data}
We pre-train our model on English text data using C4~\cite{raffel2020t5}, a large corpus scraped from the Internet. This corpus is particularly suited to our pre-training due to its diversity in terms of content and linguistic variations. In addition, it enables a better comparison with other tokenizer-free models trained using it such as Charformer. Since this dataset is not available publicly, we use the English split of the mC4 distributed by AllenAI. We filter long documents containing more than $2^{15}$ bytes, which is a simple proxy to remove important quantities of unwanted code data.
\paragraph{Hyperparameters}
We pre-train two versions of our model: $\text{MANTa-LM{}}_{Small}$ and $\text{MANTa-LM{}}_{Base}$. Each of them stacks a $\text{MANTa{}}_{Small}$ (resp. $\text{MANTa{}}_{Base}$) tokenizer and embedding module and a $\text{T5}_{Small}$ (resp. $\text{T5}_{Base}$) encoder-decoder model stripped of its tokenizer and subword embedding matrix. Details about $\text{MANTa{}}$ hyperparameters can be found in Appendix~\ref{sec:appendix_hp}.
Following T5 and ByT5, we use the Adafactor optimizer with a learning rate of $10^{-2}$ for the encoder-decoder model, parameter scaling for the whole system and no weight decay. However, to maintain stability of our differentiable tokenizer, we use a learning rate of $10^{-3}$ for the parameters of the byte embeddings, the frontier predictor, and the pooling module. We also use a triangular learning rate schedule with 1000 (resp. 5000) warm-up steps for batch size 1024 (resp. 64).
\paragraph{Training} We train $\text{T5}_{Small}$, $\text{MANTa-LM{}}_{Small}$, and $\text{MANTa-LM{}}_{Base}$ for 65k steps with a batch size of 1024. Sequence lengths are respectively 1024 for $Small$ models and 2048 for the $Base$ model. Thus, the models are trained on roughly the same amount of bytes as in \citet{tay2021charformer}, where a batch size of 64 is used for 1M steps.
We also train a $\text{ByT5}_{Small}$ model on the same data, using a batch size of 64 and a sequence length of 1024. We consider the ``Scaled'' architecture which provides the encoder with more layers than the decoder \cite{xue2022byt5}. To avoid prohibitive computation costs and ensure fairness in terms of available resources between models, we limit its training time to the one of $\text{MANTa-LM{}}_{Small}$. Hence, our $\text{ByT5}_{Small}$ is only trained for 200k steps.
\section{Experiments and Results}
\subsection{Evaluation on GLUE}
To ensure that our model is competitive with existing language models exploiting subword tokenization algorithms, we evaluate it on several English datasets and compare it with other baseline models.
\paragraph{Setup} We use GLUE~\cite{wang-etal-2018-glue}, a Natural Language Understanding benchmark consisting of 7 tasks, to evaluate our model. Similarly to T5, we cast the classification tasks as generation tasks where the model has to predict autoregressively the bytes forming the answer.
We compare our model to an encoder-decoder model with subword tokenization (pre-trained with the same denoising objective as T5) and a fully byte-level encoder-decoder, similar to ByT5. We compare $Small$ models with our pre-trained versions, and $Base$ models with results mentioned in \citet{tay2021charformer}. We report the number of parameters given in \citet{tay2021charformer} for $\text{Byte-level T5}_{Base}$, and gather from its low value that their implementation corresponds to a $\text{T5}_{Base}$ architecture trained on byte-level inputs.
\paragraph{Results} Results can be found on Tables~\ref{tab:glue_small} and~\ref{tab:glue}. Overall, MANTa-LM{} exhibits a performance slightly below Charformer but stays within a small margin on average (1.1 points below). Nonetheless, the main objective of our method is to balance decent performance with robustness and speed which we show in the following sections.
\subsection{Robustness to Domain Change}
Static subword tokenizers tend to show important limitations when used with texts originating from a domain unseen during training. For instance, \citet{el-boukkouri-etal-2020-characterbert} show that tokenizing medical texts with a tokenizer trained on Wikipedia data often results in an over-segmentation of technical terms which in turn affects the downstream performance. By removing this static bottleneck in MANTa-LM{}, we hope that it should be able to adapt more easily to new domains. To test this hypothesis, we finetune it on a medical Natural Language Inference dataset.
\begin{table}[t]
\centering\small
\begin{tabular}{lc}
\toprule
Model & Accuracy \\ \midrule
$\text{BERT}_{Base}^\ddagger$ & 77.7 \\
$\text{CharacterBERT}_{Base}^\ddagger$ & 77.9 \\
$\text{T5}_{Small}$ & 75.3 \\
$\text{MANTa-LM{}}_{Small}$ (ours) & 75.6 \\\bottomrule
\end{tabular}
\caption{Results on MedNLI. $\ddagger$ indicates results from \citet{el-boukkouri-etal-2020-characterbert}, who use a different pre-training corpus than C4. All other results are from models trained with our codebase.}
\label{tab:mednli}
\end{table}
\paragraph{Setup} We finetune MANTa-LM{} on \textsc{MedNLI}~\cite{romanov-shivade-2018-lessons}, a dataset consisting of 14,049 sentence pairs extracted from clinical notes. We follow the same finetuning setup than for the GLUE Benchmark i.e. use the same batch size and learning rate. We compare our results to the ones obtained by~\citet{el-boukkouri-etal-2020-characterbert} with models pretrained on the general domain.
\paragraph{Results} We present our results on Table~\ref{tab:mednli}. Although we notice a significant drop in performance compared to the encoder models trained by~\cite{el-boukkouri-etal-2020-characterbert}, we believe this drop may be due to the different pretraining data used---CharacterBERT uses splits of Wikipedia, which may be helpful to learn some technical terms related to the clinical domain---, and the different model sizes---CharacterBERT uses all of its parameters to encode example, while we keep half of the parameters in the decoder. Nonetheless, we note that MANTa-LM{} reaches a better performance than its subword tokenization counterpart T5.
\subsection{Robustness to Noisy Data}
Although LMs may learn complex patterns even from noisy input texts, this ability is conditioned by how the tokenizer segments character sequences. Since MANTa{} is not static and can be finetuned on non-standard data, we expect it should be able to learn to be more robust to variation/noise compared to a subword tokenizer paired with a LM. To evaluate this hypothesis, we study how MANTa-LM{} behaves on both naturally occurring text variation and multiple levels of synthetic noise.
\subsubsection{Naturally Occurring Noise}
\paragraph{Setup} Similarly to \citet{tay2021charformer}, we test our model on a toxicity detection task constructed with user generated data. We use the \textsc{ToxicComments}\iffalse\footnote{Other names for this dataset include ``Jigsaw Toxicity Prediction'' and ``Wiki Toxicity.''}\fi{} dataset~\cite{wulczyn2017ex} which contains 223,549 sentences annotated with a binary label indicating whether each sentence can be classified as toxic or not. We also use the same finetuning setup here as the one used for evaluating on the GLUE benchmark.
\paragraph{Results} We present our results in Table~\ref{tab:toxic} and compare them to the ones reported in \citet{tay2021charformer}. As expected, noisy user generated data is particularly harmful for models using subword tokenization. On the other hand, constructing sentence representations with byte-level information helps and our model is more accurate than Charformer. This gain may be due to a better segmentation of specific terms encountered in the data.
\subsubsection{Synthetic Noise}
\paragraph{Setup} We also compare T5 and ByT5 with our approach when facing different levels of noise. This study pictures how these models react to unseen noise at evaluation time (\textsc{Dev-Only} setup) and how they adapt to a given noise via fine-tuning (\textsc{Train-Dev} setup). We apply synthetic noise at different levels $\tau\in\{0.05, 0.10, 0.15\}$ by picking randomly $\tau \times L$ positions in the byte sequences and equiprobably deleting, replacing or inserting bytes at these positions.
\begin{table}[t]
\centering\small
\begin{tabular}{lc}
\toprule
Model & Accuracy \\ \midrule
$\text{T5}_{Base}^\dagger$ & 91.5 \\
$\text{Charformer}_{Base}^\dagger$ & 92.7 \\
$\text{MANTa-LM{}}_{Base}$ (ours) & {\bf 93.2} \\\bottomrule
\end{tabular}
\caption{Results on the \textsc{ToxicComments} dataset. Results indicated by $\dagger$ are from~\citet{tay2021charformer}.}
\label{tab:toxic}
\end{table}
\begin{figure}[t]
\centering\small
\begin{subfigure}[b]{\columnwidth}
\centering\small
\includegraphics[width=0.9\columnwidth]{images/noise_curve_dev.pdf}\vspace*{-2mm}
\caption{\textsc{Dev-Only}}
\end{subfigure}
\vspace{2mm}
\begin{subfigure}[b]{\columnwidth}
\centering\small
\includegraphics[width=0.9\columnwidth]{images/noise_curve.pdf}\vspace*{-2mm}
\caption{\textsc{Train-Dev}}
\end{subfigure}
\caption{Best accuracy on the SST-2 development set as the noise level increases. The \textsc{Train-Dev} setting corresponds to models finetuned on noisy data while models in the \textsc{Dev-Only} setting have been finetuned on clean data.}
\end{figure}
\paragraph{Results} We found that models performed similarly for the different noise levels in the \textsc{Dev-Only} setting. On the contrary, in the \textsc{Train-Dev} setting, MANTa-LM{} can be finetuned as well as ByT5 for all levels of noise, while the performance of T5 quickly degrades.
\begin{table}[t]
\centering\small
\begin{tabular}{lrr}
\toprule
Model & $|\theta|$ & Seconds/step \\ \midrule
$\text{Byte-level T5}_{Small}$ & 57M & 9.06 ($\times$ 8.0) \\[3pt]
$\text{MANTa-LM{}}_{Small}$ & 57M & 2.61 ($\times$ 2.3) \\[3pt]
$\text{T5}_{Small}$ & 60M & 1.13 ($\times$ 1) \\ \bottomrule
\end{tabular}
\caption{Comparison of training speeds. All the experiments were run on 16 NVIDIA V100 GPUs using a batch size of 1024 and a sequence length of 1024 bytes or 256 tokens}
\label{tab:speed}
\end{table}
\section{Training Speedups}
In terms of speed, we compare our model to $\text{MANTa-LM{}}_{Small}$ to $\text{T5}_{Small}$ counterparts: one that is trained at the classical subword-level, and one trained at byte-level, hence using sequences that are roughly 4 times longer. We also report the speed of the larger $\text{ByT5}_{Small}$ architecture as described in \citet{xue2022byt5}.
MANTa-LM{} is approximately 4 times faster than $\text{Byte-level T5}_{Small}$, and 5 times faster than $\text{ByT5}_{Small}$, which can be explained by the reduced sequence length we use in the encoder-decoder model. MANTa-LM{} is only 2.3 times slower than $\text{T5}_{Small}$ which furthermore benefits from already tokenized sequences at training time.
\section{Discussion}
\label{sec:discussion}
\begin{table*}[t]
\centering\small
\begin{tabular}{lc}
\toprule
\textbf{Original} & Oh, it's me vandalising?xD See here. Greetings, \\
\textbf{MANTa{}} & \texttt{O\nors{}h\rs{},\rs{} \nors{}i\nors{}t\rs{}'\nors{}s\rs{} \nors{}m\nors{}e\rs{} \nors{}v\nors{}a\nors{}n\nors{}d\nors{}a\nors{}l\nors{}i\nors{}s\nors{}i\nors{}n\nors{}g\nors{}?\rs{}x\nors{}D\rs{} \nors{}S\nors{}e\nors{}e\rs{} \nors{}h\nors{}e\nors{}r\nors{}e\rs{}.\rs{} \nors{}G\nors{}r\nors{}e\nors{}e\nors{}t\nors{}i\nors{}n\nors{}g\nors{}s\rs{},} \\
\textbf{T5 tokenizer} & \texttt{O\nors{}h\rs{},\rs{} \nors{}i\nors{}t\rs{}'\rs{}s\rs{} \nors{}m\nors{}e\rs{} \nors{}v\nors{}a\nors{}n\rs{}d\rs{}a\nors{}l\rs{}i\nors{}s\nors{}i\nors{}n\nors{}g\rs{}?\rs{}x\rs{}D\rs{} \nors{}S\nors{}e\nors{}e\rs{} \nors{}h\nors{}e\nors{}r\nors{}e\rs{}.\rs{} \rs{}G\nors{}r\nors{}e\nors{}e\nors{}t\nors{}i\nors{}n\nors{}g\rs{}s\rs{},}\\ \midrule
\textbf{Original} & The patient was started on Levophed at 0.01mcg/kg/min. \\
\textbf{MANTa{}} & \texttt{T\nors{}h\nors{}e\rs{} \nors{}p\nors{}a\nors{}t\nors{}i\nors{}e\nors{}n\nors{}t\rs{} \nors{}w\nors{}a\nors{}s\rs{} \nors{}s\nors{}t\nors{}a\nors{}r\nors{}t\nors{}e\nors{}d\rs{} \nors{}o\nors{}n\rs{} \nors{}L\nors{}e\nors{}v\nors{}o\nors{}p\nors{}h\nors{}e\nors{}d\rs{} \nors{}a\nors{}t\rs{} \nors{}0\rs{}.\nors{}0\nors{}1\nors{}m\nors{}c\nors{}g\rs{}/\nors{}k\nors{}g\rs{}/\nors{}m\nors{}i\nors{}n\rs{}.} \\
\textbf{T5 tokenizer} & \texttt{T\nors{}h\nors{}e\rs{} \nors{}p\nors{}a\nors{}t\nors{}i\nors{}e\nors{}n\nors{}t\rs{} \nors{}w\nors{}a\nors{}s\rs{} \rs{}s\nors{}t\nors{}a\nors{}r\nors{}t\nors{}e\nors{}d\rs{} \nors{}o\nors{}n\rs{} \nors{}L\nors{}e\rs{}v\nors{}o\rs{}p\rs{}h\nors{}e\rs{}d\rs{} \nors{}a\nors{}t\rs{} \nors{}0\nors{}.\rs{}0\nors{}1\rs{}m\rs{}c\rs{}g\rs{}/\rs{}k\nors{}g\rs{}/\rs{}m\nors{}i\nors{}n\rs{}.} \\\bottomrule
\end{tabular}
\caption{Examples of segmentations produced by our module (pre-trained only) and by T5's BPE tokenizer. The sentences are samples from \textsc{ToxicComments} and \textsc{MedNLI}.}
\label{tab:segmentation}
\end{table*}
\subsection{Truncating Embedding Sequences}
\label{sec:truncation}
Once we obtain block embeddings, the final step in MANTa{} consists in truncating sequences to a length 4 times smaller than the original byte sequence, as described in Section~\ref{sec:pooling_block}. This is essential to make MANTa-LM{} work.
First, it increases the control over the encoder-decoder's computation cost. Without this bottleneck, the Transformer can receive sequences varying from a single block containing the whole sequence ($L_B=1$) to one block per byte in the sequence ($L_B = L$). In the latter case, which mimics ByT5's input segmentation, the computation becomes extremely slow due to the quadratic cost of the attention with respect to the sequence length. Using the bottleneck ensures that we can control the worst case complexity of the encoder Transformer and keep it similar to that of a subword-based encoder model.
Second, it serves as a kind of regularization for the block segmentations. We noted that training our module without the bottleneck often led to block sequences as long as byte sequences ($L_B=L$). This may be due to the beginning of training where having very local information helps - for instance bytes to the left and right of masked spans. However, such a segmentation degrades the model speed and performance later in training. Truncating the sequence forces the model to construct larger blocks in order to ``fit'' all the information from the input sequence.
\subsection{Learnt Block Segmentation}
Segmentation examples can be found in Table \ref{tab:segmentation}. For each byte, we retrieve the expected block position produced by MANTa{} and approximate it with the closest integer to mimic hard tokenization. We found that MANTa{} is not keen to produce subword level segmentations. Most of the key symbols for word separation have been identified as block delimiters during pre-training. As expected, MANTa{} is less prone to over-segmentation of unknown words like named entities. We also found that a trained MANTa{} produced spiked separation probabilities, meaning that it converged towards a ``hard'' segmentation. This can also be observed by monitoring the value $\min(p_{F_i}, 1 - p_{F_i})$ which always converges towards values of magnitude $10^{-5}$.
\subsection{Gradient-Based Segmentation}
We employ a radically different downsampling approach compared to other gradient-based tokenization methods such as CANINE~\cite{clark2022canine} or Charformer~\cite{tay2021charformer}. While CANINE downsamples sequences using a fixed rate after byte contextualization and Charformer's GBST (Gradient Based Subword Tokenizer) pools representations created using various downsampling rates, MANTa{} only applies downsampling right before the LM to limit the length of block sequences. Hence, our model is able to build word-level representations of \textit{arbitrary length} as long as it divides the whole byte sequence length by a fixed factor.
We also argue that our method yields more explainable pooled representations as the segmentation can be explicitly derived from the outputs of MANTa{}. Indeed, contrary to CANINE and Charformer, MANTa{} disentangles the segmentation of blocks from their representations, allowing to study each part separately.
\subsection{Main hyperparameters}
We discuss here some of the major hyperparameters of our method. Constrained by limited computational resources, we were unable to assess their exact importance on MANTa's performance. We try to give some intuitions on their influence.
\paragraph{Frontier Predictor} We used a small Transformer network with sliding window attention for this module. A much larger network would be slower and may not bring significant improvements to the overall performance of the model, since it is only used for predicting the block byte assignment but does not ``expand'' the overall expressivity of the model.
\paragraph{Convolution kernel applied on byte embeddings} This kernel adds positional information to the byte embeddings and expressivity when constructing the block embeddings. Using a larger kernel or a concatenation of kernels might help for better block representations. However, our experiments did not show any significant difference in the pretraining performance.
\paragraph{Block embedding sequence truncation factor} Trimming block sequences was instrumental to produce meaningful input segmentations and blocks containing more than a single byte. We settled for a factor of 4 since other values led to minor degradations early in training. This factor roughly corresponds to the average number of bytes in a subword created by an English tokenizer.
We believe that a more thorough hyperparameter search could improve the performance of our model. We leave this for future work due to computational limitations.
\section{Conclusion}
In this work, we present MANTa{}, a fully differentiable module that learns to segment input byte sequences into blocks of arbitrary lengths, and constructs a robust representation for these blocks. We train this module jointly with an encoder-decoder LM on a span denoising objective to obtain MANTa-LM{}. We then show that MANTa-LM{} is more robust when applied to noisy or out-of-domain data than models using static subword tokenizers. At the same time, it performs on par with fully byte-level models on these setups while operating with a much reduced computational cost.
Beyond the noisy and out-of-domain settings, we believe that our approach could lead to interesting results for a number of languages, especially those whose writing system do not use whitespace separators, such as Chinese.
Finally, tokenizers are hypothesized to be an important limiting factor when segmenting multilingual data~\cite{rust-etal-2021-good}. We believe MANTa{} could be used in the multilingual setting to ensure a more balanced segmentation between languages.
\section*{Limitations}
Although MANTa{} can help alleviate some of the inherent issues accompanying subword tokenizers, it also suffers some flaws that we believe could be addressed in future work.
Contrary to encoder-decoder models that can decode long sequences efficiently, our model has to decode sequences byte-per-byte (similarly to~\citet{clark2022canine,xue2022byt5,tay2021charformer}) which adds an important computational overhead at generation time. Previous works have attempted to reduce this computational cost by decreasing the size of the decoder layers compared to the encoder~\cite{xue2022byt5} or by projecting embeddings to a smaller latent space~\cite{jaegle2021perceiver} for the decoding.
Finally, we presented in this work a proof of concept of adaptive segmentation algorithms on relatively small models, ranging from 50M to 200M parameters. Although we hypothesize that our model would scale relatively well since it keeps most of the encoder-decoder architecture untouched, this hypothesis should be tested in a future work.
\section*{Acknowledgements}
This work was funded by the last authors' chair in the PRAIRIE institute funded by the French national agency ANR as part of the ``Investissements d'avenir'' programme under the reference ANR-19-P3IA-0001.
This work was granted access by GENCI to the HPC resources of IDRIS under the allocation 2022-AD011012676R1.
|
\section{Introduction}
\label{sect:Introduction}
In spite of major efforts for more than 50 years, the understanding of radiation damage in silicon by hadrons remains a big challenge.
Reasons are the large number of radiation-induced states in the silicon band gap, the formation of defect clusters and the difficulties of their simulation.
It is well known that Si recoils from the interaction of neutrons and energetic electrons and hadrons produce locally dense cascades of Si vacancies, V, and interstitials, I, with densities as high as $10^{20}$\,cm$^{-3}$\,\cite{Fleming:2007}.
Inside these clusters the mobile vacancies can agglomerate to multi-vacancy states, in particular to the divacancy, V$_2$, which is immobile at room temperature.
In \cite{Lint:1972} it has been estimated that there are about 12 V$_2$\,states per cluster, which have typical dimensions of (5\,nm)$^3$.
The V$_2$ defect is known to come in four charge states: $+,\,0,\, -,\, --$.
The neutral V$_2^0$ state, which has an energy $E_c - 0.21$\,eV, can trap a hole or an electron resulting in states with energies $E_v+0.25$\,eV or $E_c - 0.42$\,eV, respectively, where $E_v$ denotes the energy of the valence and $E_c$ of the conduction band.
The negative V$_2^-$ state can trap electrons only, and the resulting state has the energy $E_c - 0.23$\, eV~\cite{Gill:1997}.
As discussed in \cite{Fan:1959,Cheng:1966,Pajot:2013}, the V$_2^0$ state gives rise to absorption bands at wavelengths of $1.8\,\upmu$m and $3.9\,\upmu$m, and the V$_2^-$ state at $3.6\,\upmu$m.
From the absence of photocurrent for illumination with $1.8\,\upmu$m light\,\cite{Fan:1959}, it is concluded that the photons induce transitions from the V$_2^0$ ground state to an excited state with energy $E_\mathit{ex}$.
A qualitative term scheme, deduced from the results of EPR experiments is shown in\,\cite{Cheng:1966}.
The NIR absorption data of\,\cite{Fan:1959} (Fig.\,2) also show that the overall intensity of the $1.8\,\upmu$m band is approximately independent of temperature.
However, the mean photon energy, $E_\mathit{ex}$, decreases from about 730\,meV ($\lambda = 1.70\,\upmu$m) at 100\,K to 690\,meV ($\lambda = 1.79\,\upmu$m) at 300\,K, and the full-width-half-maximum, $\Gamma _\mathit{ex}$, increases from 90\,meV to 110\,meV.
The annealing of the V$_2^0$ in silicon with different doping and impurities has been studied by several groups.
An example are the results from\,\cite{Cheng:1966} for 20\,minutes isochronal annealing and temperatures up to 330\,$^\circ $C:
Significant annealing starts around 150\,$^\circ $C, the 50\,\% point is around 240\,$^\circ $C, and complete annealing is reached at 280\,$^\circ $C.
From the publications it is however not clear if the overall intensity or the maximum of the $1.8\,\upmu$m band is shown, nor if $E_\mathit{ex}$ or $\Gamma _\mathit{ex}$ change during annealing.
In this work the light transmission for photon energies, $E_\gamma $, between 0.62 and 1.3\,eV has been measured for phosphorous-doped high-ohmic Si irradiated by reactor neutrons to fluences, $\Phi$, of (1, 5, 10$)\times 10^{16}$\,cm$^{-2}$ and 15 minutes isochronal annealing in $30\,^\circ $C steps up to $T_\mathit{ann} = 330\,^\circ $C.
From the transmission the absorbance $\alpha (E_\gamma , T_\mathit{ann}, \Phi) $ has been obtained.
For the radiation-induced absorbance, $\alpha _\mathit{irr}$, the absorption data for non-irradiated Si from \cite{Green:2021} were subtracted.
The $\alpha _\mathit{irr}$\,spectra for $0.62$\,eV$ \leq E_\gamma \leq 1.05$\,eV were fitted by a Breit-Wigner function for the V$_2^0$ excitation and a phenomenological background parametrisation to determine the values of the V$_2^0$ intensity, $I_{\mathrm{V}_2^0}$, and of the Breit-Wigner parameters $E_\mathit{\mathrm{V}_2^0}$ and $\Gamma _\mathit{\mathrm{V}_2^0}$, as a function of $\Phi $ and $T_\mathit{ann}$.
The next section gives an overview over the Si samples and the measurement techniques.
This is followed by a discussion of the analysis method.
Finally the results are presented and their relevance discussed.
\section{Samples and light-transmission measurements}
\label{sect:Samples}
For the study phosphorous-doped float-zone silicon crystals with approximately 3.5\,k$\Omega \cdot$cm resistivity and 3\,mm thickness were used.
They were irradiated by reactor neutrons \cite{IJS:Irrad} to fluences of $\Phi = (1, 5, 10) \times 10^{16}\,\mathrm{cm}^{-2}$.
The irradiations were performed approximately at room temperature, however the precise value is not known.
The estimated uncertainties of $\Phi $ are 10\,\%.
For the transmission measurements an Agilent CARY 5000 UV-VIS-NIR \cite{Agilent} was used.
The wavelength range was 0.95 to 2\,$\upmu$m and the temperature about 295\,K.
The first measurements were performed before annealing.
For the isochronal annealing the samples were heated for 15 min to temperatures of (80, 100, 120, 150, 180, 210, 240, 270, 300, 330) $^\circ $C.
The samples were stored at $-30\,^\circ $C, except for the times of the irradiation, transport, annealing and measurement.
Fig.\,\ref{fig:FigTrans} shows the measured transmission as a function of $E_\gamma $\,[eV] = 1.24$/\lambda \,[\upmu $m].
\begin{figure}[!ht]
\centering
\begin{subfigure}[a]{\textwidth}
\includegraphics[width=\textwidth]{Fig_Trans1E16.pdf}
\caption{ }
\label{fig:FigTransE16}
\end{subfigure}%
\newline
\begin{subfigure}[a]{\textwidth}
\includegraphics[width=\textwidth]{Fig_Trans5E16.pdf}
\caption{ }
\label{fig:FigTrans5E16}
\end{subfigure}%
\newline
\begin{subfigure}[a]{\textwidth}
\includegraphics[width=\textwidth]{Fig_Trans10E16.pdf}
\caption{ }
\label{fig:FigTrans10E16}
\end{subfigure}%
\caption{Measured transmission, \emph{Tr}, (left side) and radiation-induced absorbance $\alpha _\mathit{irr}$, (right side) for the different annealing steps as a function of $E_\gamma$ for
(a) $\Phi = 1 \times 10^{16}$\,cm$^{-2}$,
(b) $\Phi = 5 \times 10^{16}$\,cm$^{-2}$ and
(c) $\Phi = 10 \times 10^{16}$\,cm$^{-2}$.
In this and in other figures markers are used to distinguish the different data sets.
The distance in $E_\gamma $ of the individual data points is about 1\,meV.
For $\alpha _\mathit{irr}(E_\gamma$) also the results of the fits for $0.62\, \mathrm{eV} \leq E_\gamma \leq 1.05$\,eV described in the text, are shown as a thin black lines. }
\label{fig:FigTrans}
\end{figure}
The maximum value of \emph{Tr}, which is reached, approximately independent of $\Phi$ at $E_\gamma = 0.62$\,eV for $T_\mathit{ann} = 330\,^\circ$C, is about 0.52.
The value is close to the expectation of 0.53 from the reflection losses at the two Si-air boundaries of the samples calculated using the Fresnel formulae for normal incidence and the values of the refractive index of Si from \cite{Green:2021}.
The left side of Fig.\,\ref{fig:FigTrans} shows that \emph{Tr} decreases with increasing $\Phi$, and increases with increasing $T_\mathit{ann}$.
This is expected, as radiation damage increases with $\Phi$ and decreases with annealing.
The $E_\gamma $\,dependence of \emph{Tr} has a minimum at $E_\gamma \approx 0.7$\,eV with a full width of $\approx 0.1$\,eV, which is the absorption band of the V$_2^0$\,state.
For higher $E_\gamma$ values \emph{Tr} decreases slowly up to $\approx 1.05$\,eV, and rapidly above.
As discussed in \cite{Klanner:2022}, $E_\gamma = 1.040$\,eV is the threshold for the excitation of an electron from the valence to the conduction band by the absorption of an optical transverse phonon.
\section{Analysis of the absorption data and results}
\label{sectAnalysis}
From the measured transmission, $Tr(\lambda)$, the absorption coefficient, $\alpha (\lambda)$, is obtained using the standard formulae \cite{Scharf:2020}:
\begin{equation}\label{eq:a-data}
\alpha(\lambda) = \frac{1}{d} \cdot \ln\Bigg(\frac{\mathit{Tra}(\lambda) ^2+\sqrt{\mathit{Tra}(\lambda)^4 + 4 \cdot \mathit{Ref}(\lambda)^2 \cdot \mathit{Tr}(\lambda) ^2} } {2 \cdot \mathit{Tr}(\lambda)} \Bigg)
\end{equation}
with $\mathit{Ref}(\lambda) = \big(n(\lambda) - 1 \big)^2 / \big(n(\lambda) + 1 \big)^2 $ and
$\mathit{Tra}(\lambda) = 1 - \mathit{Ref}(\lambda)$.
The thickness of the sample is $d$, $\mathit{Ref}(\lambda)$ is the reflection of a single air-silicon interface, $\mathit{Tra}(\lambda)$ the corresponding transmission, and $n(\lambda)$ the index of refraction of silicon taken from \cite{Green:2021}.
When using Eq.\ref{eq:a-data} two important assumptions are made:
(1) The refractive index of silicon does not change with irradiation, and
(2) the transmission is dominated by absorption and the effects of scattering can be ignored.
To the authors' knowledge no investigation of the $\Phi $\,dependence of the refractive index has been reported in the literature.
An estimate can be made using the data presented in \cite{Klanner:2022a}:
The measured geometrical capacitance of irradiated pad diodes, which is reached either at high bias voltages or high frequencies, changes by less than 0.2\,\% for $\Phi$ between $3 \times 10^{15}$ and $13 \times 10^{15}$\,cm$^{-2}$.
From this observation it can be concluded that $\Delta \varepsilon_\mathit{Si}/\varepsilon_\mathit{Si} < 0.2\,\%$, and with $n = \sqrt{\varepsilon_\mathit{Si}}$, that \emph{n} changes by less than 0.1\,\% up to $\Phi = 13 \times 10^{15}$\,cm$^{-2}$.
Also for the $\Phi $\,dependence of the scattering no measurements are available.
The effect is that the scattered photons miss the sensor of the spectral-photometer and therefore reduce the measured transmission.
Thus, the absorbance $\alpha $ includes the contribution from scattering.
The following analysis uses $\alpha _\mathit{irr} (\Phi) = \alpha (\Phi) - \alpha (\Phi = 0)$.
For the absorbance of non-irradiated Si, $\alpha (\Phi = 0)$, the data from \cite{Green:2021} are used.
These data only extend to wavelengths $\lambda = 1.45\,\upmu$m, where $\alpha = 5 \times 10^{-8}$\,cm$^{-1}$ corresponding to an attenuation length of 20\,km, which can not be measured with a 0.3\,cm Si slab.
For $\lambda > 1.45\,\upmu$m the constant value of $5 \times 10^{-8}$\,cm$^{-1}$ is assumed.
It is noted that the subtraction of $\alpha (\Phi = 0)$ has no effect on the V$_2^0$ absorption band, which only covers the region $\lambda \gtrsim 1.6\,\upmu$m.
However, the subtraction of $\alpha (\Phi = 0)$ allows extending the fits to $\alpha $ to lower $\lambda $-, and thus to higher $E_\gamma $-values.
The right-hand side of Fig.\,\ref{eq:a-data} shows $\alpha _\mathit{irr} (E_\gamma)$ for the different annealing temperatures, $T_\mathit{ann}$, and the three $\Phi $\,values.
It can be seen that $\alpha _\mathit{irr}$ increases with $\Phi $, and decreases with $T_\mathit{ann}$.
At $E_\gamma \approx 0.7$\,eV, the centre of the V$_2^0$\,absorption band, a clear peak with a Breit-Wigner line shape is observed, whose amplitude decreases with $T_\mathit{ann}$ and which is practically zero for $T_\mathit{ann} = 330\,^\circ$C.
It is also observed that $\alpha _\mathit{irr}$ outside and under the V$_2^0$\,absorption band shows a similar decrease.
To quantitatively extract the parameters of the V$_2^0$-line shape and the background the following parametrisation is used:
\begin{equation}\label{eq:a-model}
\alpha (E_\gamma) = I_{\mathrm{V}_2^0} \cdot \left( \mathrm{BW}(E_\gamma ;E_{\mathrm{V}_2^0},\Gamma _{\mathrm{V}_2^0})+ \sqrt{a^2 + \left( b \cdot \mathrm{max} (0,E_\gamma - E_{BG}) \right)^2} \right),
\end{equation}
where BW is the is the normalized Breit-Wigner probability density with mean $E_{\mathrm{V}_2^0}$ and width $\Gamma _{\mathrm{V}_2^0}$, and
$I_{\mathrm{V}_2^0}$ is the contribution of the V$_2^0$\,band to $\alpha _\mathit{irr}$.
The background shape is parameterized by the constant $a$ for $E_\gamma $ below the background-transition energy, $E_\mathit{BG}$, and the square root of the sum of $a^2$ and the square of a straight line with slope $b$ for $E_\gamma > E_\mathit{BG}$.
Both, the Breit-Wigner and the background shape are multiplied with $I_{\mathrm{V}_2^0}$.
In this way the background parameters are constant if the signal-to-background ratio is constant.
Several ad-hoc parameterizations of the background have been tried, but Eq.\,\ref{eq:a-model} gave the best description of the data.
However, the results for the parameters of the Breit-Wigner function are hardly influenced by the background parametrizations which are compatible with the data.
Eq.\,\ref{eq:a-model} is fitted to the $\alpha _\mathit{irr}$\,data on the right-hand side of Fig.\,\ref{fig:FigTrans} in the range $0.62\,\mathrm{eV} \leq E_\gamma \leq 1.05$\,eV.
The fit results can be seen as black continuous lines.
As can be seen, the data are well described by the fit, with an \emph{rms} deviation of about $5 \times 10^{-4}$ of the mean $\alpha _\mathit{irr}$\,value of the individual data sets, which comprise 819 $E_\gamma$ values each.
\begin{figure}[!ht]
\centering
\begin{subfigure}[a]{0.5\textwidth}
\includegraphics[width=\textwidth]{Fig_alphaE17T20.pdf}
\caption{ }
\label{fig:alphaE17T20}
\end{subfigure}%
~
\begin{subfigure}[a]{0.5\textwidth}
\includegraphics[width=\textwidth]{Fig_alphaE17T150.pdf}
\caption{ }
\label{fig:alphaE17T150}
\end{subfigure}%
\newline
\begin{subfigure}[a]{0.5\textwidth}
\includegraphics[width=\textwidth]{Fig_alphaE17T210.pdf}
\caption{ }
\label{fig:alphaE17T20}
\end{subfigure}%
~
\begin{subfigure}[a]{0.5\textwidth}
\includegraphics[width=\textwidth]{Fig_alphaE17T300.pdf}
\caption{ }
\label{fig:alphaE17T150}
\end{subfigure}%
\caption{Measured $\alpha(E_\gamma )$, $\alpha_\mathit{irr}(E_\gamma )$ and $\alpha_\mathit{fit}(E_\gamma )$ for the data of $\Phi = 10^{17}$\,cm$^{-2}$ for
(a) no annealing,
(b) $T_\mathit{ann} = 150\,^\circ$C,
(c) $T_\mathit{ann} = 210\,^\circ$C, and
(d) $T_\mathit{ann} = 300\,^\circ$C.
In addition, the absorbance for the V$_2^0$ and the background, as determined by the fit in the range $0.62\,\mathrm{eV} \leq E_\gamma \leq 1.05$\,eV described in the text, are shown.
The symbols are markers and the spacing between the individual measurements is about 1\,meV.
The dashed lines represent the $\alpha(E_\gamma )$\,values determined from the transmission data. }
\label{fig:alphaE17}
\end{figure}
More details of the fits can be seen in Fig.\,\ref{fig:alphaE17} which shows the results for the $\Phi = 10^{17}$\,cm$^{-2}$ data before annealing, and for $T_\mathit{ann} = 150, 210$ and $ 300\,^\circ$C.
The V$_2^0$ absorption band, which peaks around 0.7\,eV, is followed by an approximately linear increase up to about 1.05\,eV, where $\alpha $ rapidly increases because the photons can excite electrons from the valence to the conduction band.
Above $E_\gamma \approx 1.15$\,eV the transmission is less than $10^{-4}$ and the measurements are dominated by noise.
For $\alpha_\mathit{irr}$, where the contribution of $\alpha $ for the non-irradiated silicon is subtracted, the fast rise above 1.05\,eV is essentially absent.
The open circles are the fits of Eq.\,\ref{eq:a-model} to $\alpha_\mathit{irr}(E_ \gamma)$, and the inclined and vertical crosses show the contributions of the V$_2^0$~Breit-Wigner and of the background, respectively.
It can be seen that the parameters of the V$_2^0$ do not change with $T_\mathit{ann}$, except for the intensity, which decreases by more than an order of magnitude.
The results for the $\Phi = 1$ and $5 \times 10^{16}$\,cm$^{-2}$ data are similar, only the V$_2^0$\,intensities are reduced approximately proportional to $\Phi$.
\begin{figure}[!ht]
\centering
\begin{subfigure}[a]{\textwidth}
\includegraphics[width=\textwidth]{Fig_aV2T.pdf}
\caption{ }
\label{fig:aV2T}
\end{subfigure}%
\newline
\begin{subfigure}[a]{\textwidth}
\includegraphics[width=\textwidth]{Fig_EV2T.pdf}
\caption{ }
\label{fig:EV2T}
\end{subfigure}%
\newline
\begin{subfigure}[a]{\textwidth}
\includegraphics[width=\textwidth]{Fig_GV2T.pdf}
\caption{ }
\label{fig:GV2T}
\end{subfigure}%
\caption{Parameters of the V$_2^0$\,absorption band as a function of $T_\mathit{ann}$ and $\Phi$ from the fit of Eq.\,\ref{eq:a-model} to $\alpha _\mathit{irr} (E_\gamma)$ for $0.62\,\mathrm{eV} \leq E_\gamma \leq 1.05$\,eV.
(a) Amplitude of the Breit-Wigner, $\alpha $-\emph{peak} in cm$^{-1}$,
(b) mean energy of the Breit-Wigner, $E _{\mathrm{V}_2^0}$, in eV, and
(c) full width of the Breit-Wigner, $\Gamma _{\mathrm{V}_2^0}$, in eV.
}
\label{fig:V2T}
\end{figure}
Fig.\,\ref{fig:V2T} shows the Breit-Wigner parameters as a function of $T_\mathit{ann}$ for the three $\Phi $\,values investigated from the fits of Eq.\,\ref{eq:a-model} to the $\alpha _\mathit{irr} (E_\gamma)$\,data for $0.62\,\mathrm{eV} \leq E_\gamma \leq 1.05$\,eV.
From Fig.\,\ref{fig:aV2T}, which shows the amplitude of the Breit-Wigner, $\alpha $-\emph{peak}, it is seen that $\alpha $-\emph{peak} before annealing (shown in the plot at $T_\mathit{ann} = 20\,^\circ$C) increases with $\Phi $.
As a function of $T_\mathit{ann}$ the V$_2^0$\,state anneals.
As expected, the annealing curves for the different $\Phi $\,values are similar.
At about $210\,^\circ$C, the amplitude is reduced by a factor 2.
The annealing curves are very similar to results reported in the literature\,\cite{Cheng:1966}.
Fig.\,\ref{fig:EV2T} shows the central energy of the Breit-Wigner peak, $E _{\mathrm{V}_2^0}$, as a function of $T_\mathit{ann}$.
For $T_\mathit{ann}$\,values up to $300\,^\circ$C the $\Phi = 1 \times 10^{16}$\,cm$^{-2}$ data show fluctuations of $\pm 4$\,meV, whereas for the higher $\Phi $\,values the fluctuations are about 1\,meV.
The larger fluctuations are ascribed to weak V$_2^0$\,signals for low $\Phi $ and for high $T_\mathit{ann}$.
The value of $699 \pm 1$\,meV is compatible with the value at room temperature which can be extracted from Fig.\,2 of Ref.\,\cite{Fan:1959}.
It is concluded that the position of the Breit-Wigner does not change within $\pm 1$\,meV.
Fig.\,\ref{fig:GV2T} shows the full width of the Breit-Wigner peak, $ \Gamma _{\mathrm{V}_2^0}$, as a function of $T_\mathit{ann}$.
Here the fluctuations are bigger than for $E _{\mathrm{V}_2^0}$ because of the strong anti-correlation between background and $\Gamma _{\mathrm{V}_2^0}$.
The value found for the two high-$\Phi $\,data sets and $T_\mathit{ann} < 300\,^\circ$C is $138 \pm 2$\,meV, without any trend of a $T_\mathit{ann} $\,dependence.
Again, this value is compatible with the value at room temperature which can be extracted from Fig.\,2 of Ref.\,\cite{Fan:1959}.
\begin{figure}[!ht]
\centering
\begin{subfigure}[a]{\textwidth}
\includegraphics[width=\textwidth]{Fig_aBGT.pdf}
\caption{ }
\label{fig:aBGT}
\end{subfigure}%
\newline
\begin{subfigure}[a]{\textwidth}
\includegraphics[width=\textwidth]{Fig_bBGT.pdf}
\caption{ }
\label{fig:bBGT}
\end{subfigure}%
\newline
\begin{subfigure}[a]{\textwidth}
\includegraphics[width=\textwidth]{Fig_EBGT.pdf}
\caption{ }
\label{fig:EBGT}
\end{subfigure}%
\caption{Parameters of the background shape as a function of $T_\mathit{ann}$ and $\Phi$ from the fit of Eq.\,\ref{eq:a-model} to $\alpha _\mathit{irr} (E_\gamma)$ for $0.62\,\mathrm{eV} \leq E_\gamma \leq 1.05$\,eV.
(a) Constant term, \emph{a},
(b) slope, \emph{b}, and
(c) background-transition energy, $E_{BG}$.
All parameters are relative to the intensity of the V$_2^0$\,band, and thus describe the shape of the background.
}
\label{fig:BGT}
\end{figure}
In Fig.\,\ref{fig:BGT} the shape parameters of the background are presented as a function of $T_\mathit{ann}$ for the three $\Phi$\,values.
It can be seen that, in particular the values for $\Phi = 1 \times 10^{16}$\,cm$^2$ show large fluctuations.
Fig.\,\ref{fig:aBGT}, shows that as a function of $T_\mathit{ann}$ the value of \emph{a}, which is the ratio of the background at the V$_2^0$\,peak to the V$_2^0$\,intensity, is approximately constant.
For the slope \emph{b}, which is shown in Fig.\,\ref{fig:bBGT}, a decrease from about 27\,eV$^{-1}$ to about 23\,eV$^{-1}$ is observed.
The background-transition energy, where a value $E_{BG} \approx 725$\,meV is determined, does not depend on $T_\mathit{ann}$.
It can be concluded that at the position of the V$_2^0$\,peak the annealing of the background is the same as the annealing of the V$_2^0$.
Above the V$_2^0$\,peak the decrease of \emph{b} with $T_\mathit{ann}$ indicates that the annealing of the background is stronger than the V$_2^0$ annealing.
The main results of this section are:
The NIR transmission, measured at room temperature, of silicon irradiated by reactor neutrons has been used to study the excitation of the V$_2^0$ state by photons for 15-minutes isochronal annealing for temperatures up to 330\,$^\circ$C.
It is found that at an annealing temperature of about 210\,$^\circ$C the V$_2^0$\,density is reduced by a factor 2, and that the energy and full width of the V$_2^0$\,absorption band do not change with annealing temperature.
It is also observed, that the annealing of the background under the V$_2^0$ band is similar to the annealing of the V$_2^0$.
\section{Comments to the results}
\label{sect:Results}
The results on the V$_2^0$\,absorption and annealing reported in this paper are certainly not new, and similar results have been reported since the late 1950-ies.
What is new, are the fits to the radiation-induced absorption coefficient, $\alpha _\mathit{irr} (E_\gamma)$, allowing to precisely determine the intensity, energy and width of the V$_2^0$\,absorption band and its background.
It is found, that for reactor-neutron fluences between (1 and 10)\,$\times 10^{16}$\,cm$^{-2}$ and subsequent isochronal annealing up to a temperature of 330\,$^\circ $C, the values of the energy and the width of the V$_2^0$\,absorption band do not change at the 1\,meV level, when the V$_2^0$\,intensity decreases by more than a factor 10 because of annealing.
As for neutron irradiation the V$_2^0$\,states are known to be mainly located in damage clusters, one can conclude that the cluster environment does not influence the V$_2^0$-excitation spectrum.
A probably related observation is reported in \cite{Klanner:2022}, where it is found that the silicon band-gap energy
changes by less than 1\,meV for neutron irradiation up to fluences of $10^{17}$\,cm$^{-2}$.
In Ref.\,\cite{Fan:1959} it is shown that the V$_2^0$\,absorption band changes with temperature, which shows that the V$_2^0$-excitation spectrum can be influenced by the environment.
It is also noted that in DLTS-\,\cite{Svensson:1991} and TSC-\,\cite{Donegani:2018}\,measurements a reduction of the electron-emission time constants of radiation-induced states in clusters, compared to the emission-time constants for isolated defects, is reported.
\section{Summary}
\label{sect:Conclusions}
Slabs of $n$-type silicon with about 3.5\,k$\Omega \cdot $cm resistivity and a thickness of 3\,mm have been irradiated by reactor neutrons to fluences of (1, 5, 10)$\times 10^{16}$\,cm$^{-2}$.
The transmission for light with wavelengths between 0.9 and 2.0\,$\upmu$m has been measured at room temperature of the samples, as irradiated and after 15\,min isochronal annealing with temperatures between $80\,^\circ$C and $330\,^\circ$C.
From the transmission, the absorption coefficient $\alpha (E_\gamma )$ has been determined.
By subtracting literature values of $\alpha (E_\gamma )$ for non-irradiated silicon, $\alpha _\mathit{irr}(E_\gamma )$, the radiation-induced contribution to $\alpha (E _\gamma)$, has been obtained.
The $E_\gamma $\,dependence of $\alpha _\mathit{irr}(E_\gamma )$ shows a peak for $E_\gamma \approx 700 $\,meV on a background.
The $\alpha _\mathit{irr} (E _\gamma)$\,spectrum has been fitted by a Breit-Wigner resonance curve, which is ascribed to the excitation of the neutral divacancy, V$_2^0$, on a parameterized absorption background.
The 6 parameters of the model have been determined as a function of irradiation fluence and annealing temperature.
The main results of the paper are:
\begin{itemize}
\item The model, which has 6 parameters, 3 for the V$_2^0$\,absorption band and 3 for the background, precisely describes $\alpha _\mathit{irr} (E_\gamma )$ for 0.62\,eV$\, \le E_\gamma \le$\,1.05\,eV.
\item The energy and the width of the V$_2^0$\,absorption band agree with literature values.
\item The energy and the width of the V$_2^0$\,absorption band do not depend on radiation fluence and annealing.
\item The intensity of the V$_2^0$\,absorption is approximately proportional to the fluence.
\item At an annealing temperature of 210\,$^\circ$C the V$_2^0$\,absorption is reduced by a factor of approximately 2.
\item The annealing of the background at the V$_2^0$\,peak is similar to the annealing of the V$_2^0$\,states. For photon energies above the V$_2^0$\,peak, the annealing is somewhat stronger.
\end{itemize}
The quantitative results of the study can be used to check models of the V$_2^0$ complex in hadron-irradiated silicon.
In addition, they can be used to correct for the radiation-induced change of photon absorption in experiments which characterise radiation-damaged silicon detectors using NIR light.
\section*{Acknowledgements}
\label{sect:Acknowledgement}
We thank Lukas Terkowski from the group of Roman Schnabel of the Institute of Laser Physics of Hamburg University for his help with the photo-spectrometer measurements, and Vladimir Cindro and Gregor Kramberger for the neutron irradiations performed at the TRIGA reactor of the Jozef Stefan Institute, Ljubljana.
We are also thankful to
Ioana Pintilie for information on the properties of radiation-induced defects.
The work was partially supported by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under Germany's Excellence Strategy -- EXC 2121 Quantum Universe - 390833306.
\input{bibliography}
\label{sect:Bibliography}
\end{document}
\section{List of References}
|
\section{Introduction}
This paper showcases \texttt{PSTMonitor}\xspace, a tool supporting the run-time monitoring of quantitative properties of
message-passing applications.
The underlying methodology of \texttt{PSTMonitor}\xspace, recently described in the
companion paper~\cite{DBLP:conf/coordination/BurloFSTT21}, relies on
\emph{(probabilistic) session types} (PST for short).
PSTs specify both \begin{enumerate*}[label=\textit{(\roman*)}] \item the
expected communication protocol; and \item quantitative constraints on the
branching behaviour of the protocol through probabilities, within one unified formalism. \end{enumerate*}
\begin{example}\label{eg:example-intro}
Consider a server hosting a guessing game. The server picks an integer $n$ between 1 and 100,
and the client is expected to either attempt to guess $n$, or ask for a hint.
In such setup, one might want to assess whether the server behaves fairly and gives the client a chance of guessing correctly.
On the other end, one might also want to assess whether the client is asking for too many hints without attempting to guess.
\qed
\end{example}
The methodology proposed in~\cite{DBLP:conf/coordination/BurloFSTT21} explores the post-deployment verification of systems such as the one in \Cref{eg:example-intro} against PSTs.
In particular, monitors are used to detect deviations of the current run-time execution of a system from the probabilistic behaviour specified by the PST.
Given a PST, our tool \texttt{PSTMonitor}\xspace
automates the synthesis of a
monitor that is able to both \begin{enumerate*}[label=\textit{(\roman*)}] \item report violations in case of miscommunications; and
\item issue warnings if the observed communications deviate substantially from the probabilistic
behaviour expressed in the PST.
\end{enumerate*}
One complication is that PSTs specify probabilistic behaviours over \emph{complete} executions,
whereas our synthesised monitors need to detect deviations in real-time,
hence they can only make decisions based on a \emph{partial} execution observed up to the current instant.
For this reason, \texttt{PSTMonitor}\xspace\ relies on \emph{confidence intervals} to
gauge whether the observed behaviour (up to the current execution point) is compatible with a given PST.
The rest of the paper is structured as follows.
\Cref{sec:workflow} illustrates key functionalities of our tool and
the workflow for deploying monitors derived from PSTs.
\Cref{s:evaluation} discusses the feasibility of our probabilistic
monitoring by considering a fragment of the Simple Mail Transfer
Protocol (SMTP)~\cite{SMTP}.
Finally, \Cref{sec:conc} draws some conclusions and sketches future
work.
\section{Workflow and Tool Functionality}\label{sec:workflow}
The workflow of \texttt{PSTMonitor}\xspace starts with a client-server communication protocol expressed as a \emph{probabilistic session type (PST)} \cite{DBLP:journals/corr/abs-2011-09037, DBLP:conf/concur/InversoMPTT20} --- \textit{i.e.,}\xspace a (binary) session type where the choice-points are augmented with quantitative information, namely probabilities describing the frequencies of the choices to be taken.
The PST is then used to automatically synthesise a monitor
which operates w.r.t.\xspace a given \emph{confidence level}, as detailed in \cite[Section 2.2]{DBLP:conf/coordination/BurloFSTT21}.
Intuitively, the synthesised monitor iteratively calculates \emph{confidence intervals} around all choice-points specified in the PST, and computes frequency-based estimates of the probabilities of the choices observed in the monitored system. The monitor continuously checks whether the run-time-observed choice probabilities fall within the corresponding PST confidence intervals: if not, the monitor issues a warning, meaning that
the observed behaviour is significantly deviating from the PST specification.
The monitor can later retract the warning, if the run-time-observed choice probabilities return within the PST confidence intervals.
\subsection{Specifying behaviour via Probabilistic Session Types (PSTs)}
Session types formalise communication protocols by specifying the order and choice of messages together with their corresponding payloads.
We take \emph{binary} session types (supporting client-server protocols) and augment the choice points with probability distributions that specify the frequency with which a choice should be taken by one of the components interacting in a session.
The syntax of our Probabilistic Session Types (PSTs) is thus:
\begin{align*}
\stS \;\bnfdef\; &\ \stBrai{i}{I} && \textit{(external choice)} \\
\bnfmid & \stSeli{i}{I} && \textit{(internal choice)} \\
\bnfmid & \stRec{X}.\stS && \textit{(recursion)} \\
\bnfmid & \stRecVar{X} && \textit{(recursion variable)} \\
\bnfmid & \stEnd && \textit{(termination)}
\end{align*}
In choice points ($\stBraOp$ and $\stSelOp$) the indexing set $I$ is finite and non-empty, the \emph{choice labels} $\texttt{l}_i$ are pairwise distinct, and the \emph{sorts} $\asort_i$ range over basic data types ($\typeInt$, $\typeStr$, $\typeBool$,\textit{etc.}\xspace) for typing \emph{variables} $x_i$.
We give a \emph{multinomial distribution} interpretation to each choice point ($\stBraOp$ and $\stSelOp$) in a PST:
we require that $\sum_{i \in I} p_i = 1$, where every $0 \leq p_i \leq 1$ is the probability of selecting the branch labelled by $\texttt{l}_i$.
The probabilities prescribed at a choice point represent a behavioural obligation on the interacting party that has control over the selection at that choice point.
As usual, we require that recursion is guarded, \textit{i.e.,}\xspace a recursion variable $X$ can only appear under an external or internal choice prefix.
\begin{example}
\label{eg:session-type}
We formalise the communication protocol outlined in \Cref{eg:example-intro} as the PST $\stS[game]$ in \Cref{fig:s-game}, written from the perspective of the server.
The type specifies that the server should wait for the client's choice (at the external branching point \stBraOp) to either \lab{Guess} a number, ask for \lab{Help}, or \lab{Quit}.
%
If the client asks for help, the server should reply with a \lab{Hint} message including a string, and the session loops.
The session should also loop after the outcome of a guess (\lab{Correct} or \lab{Incorrect}, at the internal choice \stSelOp) is communicated to the client.
The probability annotations in $\stS[game]$ specify the expected frequency of each choice, and rule out unwanted behaviours.
In particular, they require the server to reply with \lab{Correct} 1\% of the time, and the client to only request for help 20\% of the time.
\qed
\end{example}
\begin{figure}[h]
\small\begin{Verbatim}[commandchars=\\\{\},numbers=left,numbersep=1mm,frame=single,breaklines]
S_game = rec X.(+\{!Guess(num: Int)[0.75].
&\{?Correct()[0.01].X, ?Incorrect()[0.99].X\},
!Help()[0.2].?Hint(info: String)[1].X,
!Quit()[0.05].end\})
\end{Verbatim}
\vspace{-0.5cm}\cprotect\caption{Probabilistic session type \verb|S_game|.}\label{fig:s-game}
\end{figure}
\subsection{Monitor synthesis: behind the scenes}
\begin{figure}[htp]
\centering
\begin{subfigure}{0.45\textwidth}
\centering
\begin{tikzpicture}[decoration=penciline]
\node(monitor) [static-process]{\texttt{Monitor}};
\node(synth) [static-process, above=0.5cm of monitor, font=\footnotesize] {$\textsf{synth}$};
\node(cpspc) [draw, static-process, right=0.5cm of synth, font=\footnotesize] {$\textsf{CPSPc}$};
\node(scb) [decorate, draw, left=0.7cm of synth] {$\stS[game]$};
\draw[-angle 90,dashed] (scb) edge (synth);
\draw[-stealth, semithick] (synth) edge (monitor);
\draw[-stealth, semithick] (synth) edge (cpspc);
\end{tikzpicture}
\caption{\centering Monitor and CPSPc synthesis.}\label{fig:mon-cpspc-synth}
\end{subfigure}
\hspace{0.03\textwidth}
\hspace{0.03\textwidth}
\begin{subfigure}{0.45\textwidth}
\centering
\begin{tikzpicture}[decoration=penciline]
\node(monitor) [static-process]{\texttt{Monitor}};
\node(cm_c) [draw, left=0.2cm of monitor, font=\scriptsize] {$\textsf{CM}_c$};
\node(cm_m) [draw, right=0.2cm of monitor, font=\scriptsize] {$\textsf{CM}_s$};
\node(cpspc) [draw, static-process, above=0.5cm of monitor, font=\footnotesize] {$\textsf{CPSPc}$};
\node(scala-typechecker)[draw, right=1cm of cpspc, inner sep=2pt, font=\scriptsize, align=center]{\texttt{Scala}\\[-1mm]\texttt{type-checker}\\[-1mm]\texttt{+ lchannels}};
\draw[-angle 90, dashed] (cpspc) -- (scala-typechecker);
\draw[-, dashed] (cm_c) |- ($(cm_m.north)+(0,0.35)$);
\draw[-, dashed] (cm_m) |- (scala-typechecker);
\draw[-, dashed] (monitor) -- ($(monitor.north)+(0,0.35)$);
\node(satisfaction) [circle, draw, below=0.4cm of scala-typechecker, text=green!70!black, font=\scriptsize, inner sep=1pt, minimum size=4mm, semithick] {\ding{51}};
\draw[-stealth, semithick] (scala-typechecker) edge (satisfaction);
\end{tikzpicture}
\caption{\centering Type-checking implementations.}\label{fig:type-checking-implementations}
\end{subfigure}
\caption{Synthesising and compiling the monitor.}\label{fig:synth-compile}
\end{figure}
\texttt{PSTMonitor}\xspace generates executable session monitors in the Scala programming language.
However, session type constructs (internal or external choices) are not natively supported by Scala (nor other mainstream programming languages).
For this reason, \texttt{PSTMonitor}\xspace leverages the library \texttt{lchannels}\xspace \cite{DBLP:conf/ecoop/ScalasY16}, which encodes (binary) session types constructs in the Scala type system through \emph{Continuation-Passing Style Protocol classes} (\textsf{CPSPc}\xspace), capturing the order of the send and receive operations and choices in a session type.
\begin{example}
As depicted in \Cref{fig:mon-cpspc-synth}, our tool \texttt{PSTMonitor}\xspace generates the \textsf{CPSPc}\xspace from a session type.
\Cref{lst:cpspc} contains some of the \textsf{CPSPc}\xspace from the autogenerated file representing the type $\stS[game]$\xspace from \Cref{eg:session-type}.
\lstinputlisting[caption={Continuation-Passing Style Protocol classes for $\stS[game]$\xspace},label={lst:cpspc},captionpos=b,aboveskip=3mm]{cpspc.scala}
Intuitively, every class represents a message specified within the session type (lines 2, 4 and 5), containing the
\begin{enumerate*}[label=\emph{(\roman*)}]
\item type of the payload; and
\item the continuation type.
\end{enumerate*}
Every choice point is represented by an abstract class (lines 1 and 3) that is inherited by all the choices within the choice point.
The monitor uses these types to keep track of the current point of the interaction and verify that the components respect the session type. \qed
\end{example}
\begin{wrapfigure}[11]{l}{8.9cm}
\vspace{-0.6cm}
\lstinputlisting[caption={\vspace{1cm}\texttt{receive} and \texttt{send} methods of $\textsf{CM}_{c}$~in \Cref{fig:starting-monitor}.},label={lst:cm-receive},captionpos=b,aboveskip=3mm,frame=tbr,rulecolor=\color{black},firstnumber=1]{cm-receive.scala}
\end{wrapfigure}
\noindent The monitors generated by \texttt{PSTMonitor}\xspace are agnostic to the transport protocol in use:
they require user-supplied Connection Managers ($\textsf{CM}$s\xspace) that sit between them and the components under observation (as depicted in \Cref{fig:type-checking-implementations}).
A \cm acts as a translator and gatekeeper by transforming messages from the format used in the transport protocol to their respective \textsf{CPSPc}\xspace (and \emph{vice versa}), while governing the interaction with the component.
In order to do this, $\textsf{CM}$s\xspace must extend a provided abstract class (\verb|ConnectionManager|) and implement the methods \verb|setup|, \verb|close|, \verb|receive| and \verb|send| (as in \Cref{lst:cm-receive}).
As the name implies, the first two manage the connection with the component; the monitor invokes them when it sets up the connection and before it terminates
\begin{figure}[htp]
\centering
\begin{subfigure}{0.45\textwidth}
\centering
\begin{tikzpicture}
\node(monitor) [static-process]{\texttt{Monitor}};
\node(cm_c) [draw, left=0.2cm of monitor, font=\scriptsize] {$\textsf{CM}_c$};
\node(cm_m) [draw, right=0.2cm of monitor, font=\scriptsize] {$\textsf{CM}_s$};
\draw [thin, dotted, rounded corners] ($(cm_c.north west)+(-0.1cm,0.17cm)$) rectangle ($(cm_m.south east)+(0.1cm,-0.15cm)$);
\node(server) [process, fill=gray!20, right=1.3cm of monitor] {$s$};
\node(client) [process, fill=gray!20, left=1.3cm of monitor] {$c$};
\node(cpspc) [draw, above=0.5cm of monitor, static-process, font=\footnotesize] {$\textsf{CPSPc}$};
\draw[-] (cm_c) -- (monitor);
\draw[-] (cm_m) -- (monitor);
\draw[-] (cpspc) -- (monitor);
\draw[-] (cpspc) -| (cm_c);
\draw[-] (cpspc) -| (cm_m);
\draw[thin,double distance=2pt, fill=white] (cm_c) -- (client);
\draw[thin,double distance=2pt, fill=white] (cm_m) -- (server);
\node(vm) [below=0.8mm of cm_c, font=\tiny, draw=white, fill=white, inner sep=0pt] {\texttt{VM}};
\end{tikzpicture}
\caption{\centering Setup for black-box components.}\label{fig:wrapper-setup}
\end{subfigure}
\hspace{0.03\textwidth}
\hspace{0.03\textwidth}
\begin{subfigure}{0.45\textwidth}
\centering
\begin{tikzpicture}
\node(client) [process, fill=gray!20] {$c$};
\node(cm) [draw, right=0.25cm of client, font=\scriptsize] {$\textsf{CM}_c$};
\node(monitor) [process, right=0.2cm of cm]{\texttt{Monitor}};
\node(server) [process, right=1.7cm of monitor] {$s$};
\draw [thin, dotted, rounded corners] ($(cm.north west)+(-0.1cm,0.17cm)$) rectangle ($(server.south east)+(0.1cm,-0.15cm)$);
\draw[thin,double distance=2pt, fill=white] (cm) -- (client);
\draw[thin,double distance=2pt, fill=white] (monitor) -- node(lchannels)[above, font=\scriptsize, yshift=-1pt]{\texttt{lchannels}\xspace} (server);
\node(cpspc) [draw, above=0.5cm of monitor, static-process, font=\footnotesize] {$\textsf{CPSPc}$};
\draw[-] (cm) -- (monitor);
\draw[-] (cpspc) -- (monitor);
\draw[-] (cpspc) -| (cm);
\draw[-] (cpspc) -| (server);
\node(vm) [below=0.8mm of cm, font=\tiny, draw=white, fill=white, inner sep=0pt] {\texttt{VM}};
\end{tikzpicture}
\caption{\centering Grey-box setup.}\label{fig:hybrid-setup}
\end{subfigure}
\caption{Alternative monitoring setups for a client $c$ and a server $s$.}\label{fig:starting-monitor}
\end{figure}
\begin{example}
Listing \ref{lst:cm-receive} shows a code snippet for a \cm's \verb|receive| and \verb|send| methods.
In this case, the \cm is handling a TCP/IP socket, where the messages from the type $\stS[game]$ in \Cref{eg:example-intro} are serialised into a textual format, \textit{e.g.,}\xspace $\lab{Guess}(n)$ into \texttt{"GUESS $n$"}.
\noindent When invoked by the monitor, the \verb|receive| method checks the socket input buffer \texttt{inBuf}: if it finds a supported message (line 3, where the message matches the regex \texttt{guessR}), it returns an object of the corresponding CPSP class;
otherwise, it returns the unaltered message.
Therefore, when the client $c$ in \Cref{fig:starting-monitor} sends the message \texttt{"GUESS 23"}, the \cm[c] translates it to the CPSP class \texttt{Guess} (line 2 of \Cref{lst:cpspc}).
When the server sends a message to the monitor to be forwarded to the client, it invokes the method \verb|send| in Listing \ref{lst:cm-receive}: such a method translates messages from a CPSP class instance into the format accepted by the client, and sends them.
In this case, if the server replies with a message \texttt{Correct} (line 4 of Listing \ref{lst:cpspc}), the server translates it into the textual format \texttt{"CORRECT"} and writes it on the TCP/IP socket output buffer (line 7 of Listing \ref{lst:cm-receive}).
The catch-all case (line 10) is used for debugging purposes.
\qed
\end{example}
\subsection{Synthesising and using a monitor}
\label{sec:synthesise-monitor}
Our tool \texttt{PSTMonitor}\xspace automatically generates the monitor code and the \textsf{CPSPc}\xspace from a session type description, by running the following command:
\begin{lstlisting}[language=bash]
$ sbt "project monitor" "runMain monitor.Generate $DIR $ST $PRE"
\end{lstlisting}
\begin{enumerate}[nosep]
\item[\texttt{\$DIR}] is the directory where the source code of the monitor and classes will be generated;
\item[\texttt{\$ST}] is the file containing the probabilistic session type (as in \Cref{fig:s-game}); and
\item[\texttt{\$PRE}] (optional) is a file containing a preamble that will be added to the top of the generated files (\textit{e.g.,}\xspace containing package declarations and imports).
\end{enumerate}
Once completed, the auto-generated files \texttt{Monitor.scala} and \texttt{CPSPc.scala} will be saved in the directory \texttt{\$DIR}.
\subsection{Deploying the monitor}
The monitor generated in \Cref{sec:synthesise-monitor} can be used in a number of different setups.
We outline two such setups covering two common situations, depicted in \Cref{fig:starting-monitor}:
\begin{description}[nosep]
\item[Black-box monitoring setup (\Cref{fig:wrapper-setup}).] This is the most general way to deploy a monitor: both the client $c$ and server $s$ are treated as black boxes.
Here the monitor makes use of two connection managers, one for each end of the interaction.
To start the monitor itself, the user needs to write a simple proxy that sits between the monitored client and server, and starts the generated monitor whenever a new connection is established.
%
\item[Grey-box monitoring setup (\Cref{fig:hybrid-setup}).] This setup is possible when one of the components (\textit{e.g.,}\xspace the server \emph{s}\xspace) is implemented in Scala using \texttt{lchannels}\xspace and can be deployed together with the monitor.
This allows for a direct \texttt{lchannels}\xspace-based connection between the monitor and the component, without a connection manager.
Moreover, the monitor and component can be started on the same Java Virtual Machine (JVM),
thus improving the overall performance (as we illustrate in \Cref{s:evaluation}).\footnote{%
If a black-box component is supplied as a \texttt{.jar} file, it is also possible to deploy such a component and its monitor as different threads running on the same JVM; in this case, the monitor-component interaction would use a TCP/IP socket, which is less efficient than a direct \texttt{lchannels}\xspace-based connection. The resulting performance would be similar to the black-box scenario measured in \Cref{fig:wrapper-setup}.}
\end{description}
\subsection{Additional Features}\label{s:monitoring-procedure}
To further assist with the analysis of a monitored system, monitors generated by \texttt{PSTMonitor}\xspace can also log information about the current execution, thus enabling, \textit{e.g.,}\xspace the real-time visualisation of the monitor status.
The logs include \begin{enumerate*}[label=\emph{(\roman*)}]
\item the estimated probability of how often a choice within a branch was taken; and
\item the boundary of the confidence interval.
\end{enumerate*}
\begin{figure}
\hspace{2.5em}
\includegraphics[scale=0.11]{legend.png}
\smallskip
\centering
\newcommand{0.9}{0.9}
\begin{subfigure}{0.45\linewidth}
\scalebox{0.9}{
\input{Help_5_log_good.pgf}
}
\caption{\centering A compliant monitored client.}
\label{fig:compliant-client}
\end{subfigure}
\hfill
\begin{subfigure}{0.45\linewidth}
\scalebox{0.9}{
\input{Help_5_log_bad.pgf}
}
\caption{\centering A non-compliant monitored client.}
\label{fig:noncompliant-client}
\end{subfigure}
\caption{Evolution of the monitoring status for the \lab{Help} branch of $\stS[game]$ (\Cref{eg:example-intro}).}
\label{fig:live-logging}
\end{figure}
\begin{example}
\label{eg:monitor-plots}
The plots depicted in \Cref{fig:live-logging} are generated from monitor logs:
they show the executions of two different clients for the guessing game PST $\stS[game]$ (\Cref{eg:example-intro}). The two clients select the \lab{Help} choice with different patterns and frequencies.
The client in \Cref{fig:compliant-client} complies with the PST: it selects \lab{Help} with a frequency that always remains within the confidence interval of the probability specified in $\stS[game]$.
Instead, the client in \Cref{fig:noncompliant-client} diverges from the PST: it selects \lab{Help} too often, hence the estimated probability moves outside the confidence interval after a few iterations; when this happens, the monitor issues a warning.
\qed
\end{example}
\section{Evaluation}\label{s:evaluation}
We evaluate \texttt{PSTMonitor}\xspace by measuring the overheads induced by the monitors it synthesises.
As a benchmark, we consider a probabilistic fragment of the Simple Mail Transfer Protocol (SMTP) \cite{SMTP} (server-side) formalised as the session type $\stS[smtp]$ below:
{\small
\begin{align}
\stS[smtp] =\ & \stSnd{M220}{\typeStr}{}.\stBraOp\left\{\begin{array}{l}
\stRcv{Helo}{\typeStr}{}.\stSnd{M250}{\typeStr}{}.\stS[mail],\\
\stRcv{Quit}{}{}.\stSnd{M221}{\typeStr}{}
\end{array}\right\}\nonumber\\[2mm]
\label{s-mail-1}\stS[mail] =\ & \stRec{X}.\big(\stBra{\stRcv{MailFrom}{\typeStr}{0.5}.\stSnd{M250}{\typeStr}{}.\stRec{Y}.\\
&\ \big(\stBraOp\left\{\begin{array}{l}
\stRcv{RcptTo}{\typeStr}{0.6}.\stSnd{M250}{\typeStr}{}.\stRecVar{Y},\\
\stRcv{Data}{}{}.\stSnd{M354}{\typeStr}{}.\stRcv{Content}{\typeStr}{}.\stSnd{M250}{\typeStr}{}.\stRecVar{X},\\
\stRcv{Quit}{}{}.\stSnd{M221}{\typeStr}{}
\end{array}\right\}\big),\nonumber\\
&\quad \stRcv{Quit}{}{}.\stSnd{M221}{\typeStr}{}}\big)\nonumber
\end{align}
}%
When a client establishes a connection, the server sends a welcome message (\lab{M220}), and waits for the client to identify itself (\lab{Helo}).
Then, the client can recursively send emails by specifying the sender (\lab{MailFrom}) and recipient (\lab{RcptTo}) address(es), followed by the mail contents (\lab{Data}).
The client can send multiple emails by repeating the loop on ``$\stRecVar{X}$'' starting from line \eqref{s-mail-1}.
The purpose of the probability annotations in $\stS[smtp]$ and $\stS[mail]$ is to flag clients that appear to be sending spam, or using the server resources without a purpose.
Setting a confidence level of 95\%, such probability annotations result in a warning if a client \begin{enumerate*}[label=\emph{(\alph*)}]
\item sends three or more emails in a single connection (\lab{MailFrom}), or
\item includes six or more recipients (\lab{RcptTo}).
\end{enumerate*}
The type $\stS[smtp]$ above and the synthesised monitors are not tied to any specific message transport protocol; we implement them using TCP/IP (as per \cite{SMTP}) by providing suitable connection managers to the synthesised monitor (as explained in \Cref{sec:workflow}).
To conduct our experiments, we implement a client that sends emails to an SMTP server and measures the response time.
As a server, we use a default instance of Postfix\footnote{\url{http://www.postfix.org/}} (one of the most used SMTP servers \cite{SMTPSurvey}) configured to receive emails and discard them.
To study the monitoring overheads, we compare:
\begin{itemize}[nosep,noitemsep,label=-]
\item \emph{an unsafe setup}: the SMTP client and server interact directly;
\item \emph{a black-box monitored setup} where communication is mediated by a monitor instantiated separately,
as in \Cref{fig:wrapper-setup}; and
\item \emph{a grey-box monitored setup} with the client (written in Scala+\texttt{lchannels}\xspace) and monitor running on the same virtual machine, as in \Cref{fig:hybrid-setup}.
\end{itemize}
\begin{figure}
\newcommand{0.65}{0.65}
\hspace{2em}\includegraphics[scale=0.2]{legend2.png}
\smallskip
\scalebox{0.65}{
\input{smtp_cpu_consumption.pgf}
}
\scalebox{0.65}{
\input{smtp_mem_consumption.pgf}
}
\scalebox{0.65}{
\input{smtp_resp_time.pgf}
}
\caption{\centering Benchmark results for SMTP monitoring: CPU usage, maximum memory consumption, and response time. (Averages of 20 repetitions; 2 Xeon E5-2687W 8-core CPUs @ 3.10GHz, 128 GB RAM, GNU/Linux 5.10.27)}
\label{fig:smtp-benchmarks}
\end{figure}
We measure the monitor's response time, CPU utilisation, and maximum memory consumption,
by running experiments where the client sends an increasing number of emails to the server.
We measure the performance with logging enabled and disabled (as explained in \Cref{s:monitoring-procedure}).
The results are shown in \Cref{fig:smtp-benchmarks}\footnote{
Ideally, the results are also compared with other similar approaches. However, unfortunately, it is very hard to reliably compare overhead figures across different frameworks due to the discrepancies and peculiarities of each setup.}.
Overall, the plots in \Cref{fig:smtp-benchmarks} fall within the range of typical overheads introduced by run-time verification \cite{DBLP:series/lncs/BartocciFFR18, DBLP:journals/sttt/BartocciFBCDHJK19, DBLP:conf/fase/AcetoAFI21}.
They have two main origins:
\begin{enumerate}[nosep,label=\emph{(\arabic*)}]
\item the calculation and bookkeeping of probability estimates at choice points, and their checking against $\stS[smtp]$; and
\item the translation and duplication of messages being forwarded between the client and server (which is mitigated in the grey-box setup).
\end{enumerate}
The response time is arguably the most important measurement, since slower response times can be immediately perceived when interacting with a monitored system:
in the black-box monitored setup we measured an overhead of 25\% (or 30\% with logging enabled).
Such overhead can be reduced to 13\% (or 20\% with logging enabled) by adopting a grey-box setup which minimises network communication.
\section{Conclusions and Discussion}\label{sec:conc}
We have presented \texttt{PSTMonitor}\xspace, a tool aimed to analyse quantitative aspects of a system's interactions at runtime.
\texttt{PSTMonitor}\xspace is implemented as an extension of the monitoring framework \texttt{STMonitor}\xspace from \cite{DBLP:conf/ecoop/BurloFS21,DBLP:journals/darts/BurloFS21}.
Originally, \texttt{STMonitor}\xspace synthesises monitors from binary session types augmented with assertions (\textit{i.e.,}\xspace predicates over communicated values).
In order to support probabilistic session types (PSTs), we have replaced the type assertions with probabilities, and we have refactored the synthesis to generate monitors that conduct probabilistic analysis with the methodology introduced in \cite{DBLP:conf/coordination/BurloFSTT21} (and here summarised in \Cref{sec:workflow}).
We have also implemented the logging functionality outlined in \Cref{s:monitoring-procedure}, thus enabling further insight into the behaviour of a monitored system (as shown in \Cref{eg:monitor-plots}).
Notably, the proposed methodology and the tool instantiating it are independent of each other.
While the methodology can serve as the basis for other runtime analysis techniques for quantitative-based specifications, \texttt{PSTMonitor}\xspace is not limited, nor bound, to the current statistical technique.
Rather, the synthesis permits for interchangeable statistical inference paradigms.
For instance, we can improve our CI estimation by utilising
the Wilson score interval \cite{Wilson1927} which is more costly but also more reliable when the sample size (observed messages up to the current
point of execution) is small or the specified probability is close to 0 or 1.
\paragraph{Related work}
Other publications and tools (also discussed in \cite{DBLP:conf/coordination/BurloFSTT21})
propose related techniques or have objectives similar to our work.
Albeit addressing the problem from a different angle with different specification languages, these tools and their respective methodologies can complement the analysis conducted by \texttt{PSTMonitor}\xspace.
The work closest to ours is RT-MaC \cite{DBLP:conf/rtcsa/SammapunLS05}, a tool that generates monitors to determine whether a system satisfies a probabilistic property by analysing its behaviour at runtime and performing statistical hypothesis testing.
Similarly to us, they make use of confidence intervals to gauge the accuracy of the observed behaviour.
Unlike us, their tool sets up hypotheses from the specified (logic-based) properties and once the monitor has observed enough information to accept or reject the hypotheses, it decides whether the property is satisfied or not.
By adopting the technique from \cite{DBLP:conf/rtcsa/SammapunLS05}, our monitors can be made to reach \emph{irrevocable verdicts} on the observed behaviour, rather than issue retractable warnings.
Confidence intervals are also used in \cite{calinescu2015formal} for analysing quality of service properties (such as reliability, performance or cost) of a system.
Unlike our work, their tool-supported approach is applied in the post-deployment phase on models obtained from system logs or via runtime monitoring.
This is done with the aim of establishing unknown behavioural aspects of the system (\textit{e.g.,}\xspace how often information is requested from a cache), and since such properties cannot be definitively confirmed with the available information (\textit{i.e.,}\xspace a set of observations), confidence intervals are used to give an approximation of the system's behaviour.
Similarly, LogLens \cite{debnath2018loglens} analyses system logs to detect anomalies;
with no (or minimal) knowledge about the system, LogLens uses machine learning based techniques to discover patterns in its behaviour from previous system logs and then compare those in real-time logs to find inconsistencies.
On the contrary, our technique does not rely on \emph{any} previous information about the system; moreover, the focus of \cite{calinescu2015formal,debnath2018loglens} is not on the probabilistic properties of the system.
These tools can complement the methodology enabled by \texttt{PSTMonitor}\xspace:
given their capability of logging information on the system executions (\Cref{s:monitoring-procedure}), our monitors can be used to extract information about the system itself (even when it is a black-box) which can be passed on to tools such as \cite{calinescu2015formal,debnath2018loglens} for further analysis.
\bibliographystyle{elsarticle-num}
|
\section{Introduction}
The final task of the 3rd ACM-China IPCC is to optimize a program for graph spectral sparsification. In this competition, three test cases and source code of a program are provided for participants. These test cases are used for performance evaluation, which consist of 4K, 7K and 16K nodes respectively.
Participants of this competition need to craft a program that runs as fast as possible while outputs the same result as provided program on designated hardware platform.
For this task, we first proposed LGRASS, a graph spectral sparsification algorithm which runs in strictly linear time. Then we crafted a parallel processing scheme to optimize the hot-spot subroutines of LGRASS. The details of the two-step improvements is discussed in the following sections.
\section{Motivation}
\begin{figure}[htbp]
\centering
\subfloat[Baseline]{
\includegraphics[width=1.0in]{figures/Baseline.png}
\label{figure:baseline}
}
\subfloat[Basic LGRASS]{
\includegraphics[width=1.0in]{figures/LGRASS.png}
\label{figure:LGRASS}
}
\subfloat[Parallel LGRASS]{
\includegraphics[width=1.0in]{figures/pLGRASS.png}
\label{figure:pLGRASS}
}
\caption{Program Procedures}
\end{figure}
\begin{table}[ht]
\centering
\caption{Time breakdown of baseline program}
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline
& EFF & MST & INV & RES & MARK & ALL\\
\hline
Case 1 & 10.7ms & 205.3ms & 10.1s & 33.1ms & 21.7min & 22.9min \\
\hline
Case 2 & 5.6ms & 183.6ms & 52.4s & 13.4ms & 24.6min & 25.5min\\
\hline
\end{tabular}
\label{tab:baseline_breakdown}
\end{table}
To find hot-spot subroutines, we made a coarse-grained performance profiling for baseline program, whose procedure is demonstrated in Fig. \ref{figure:baseline}.
As is shown in Table \ref{tab:baseline_breakdown}, the running time of the baseline program, i.e. provided program, is dominated by its subroutine of marking edges. By the computational complexity analysis, there are three subroutines which costs super-linear time: marking edges, computing pseudo inverse and sorting edges. These subroutines costs over 99.9\% time of the overall running time, which means their super-linear running time has become the main performance bottleneck.
Thus, we propose LGRASS, to improve the time complexity of these three subroutines and reduce their running time to be \emph{linear}, and then optimize it for multiprocessor systems. The procedures of basic version and parallel version of our proposed method are demonstrated in Fig. \ref{figure:LGRASS} and Fig. \ref{figure:pLGRASS} respectively. The novel techniques used by basic LGRASS is presented in Section \ref{sec:BasicLGRASS} and the parallel processing scheme of parallel LGRASS is depicted in Section \ref{sec:ParallelLGRASS}.
\section{The Design of LGRASS}
\label{sec:BasicLGRASS}
\subsection{Marking Edges in Linear Time}
The pseudocode of marking edges in baseline program is described in Algorithm \ref{alg:baseline_marking}. The running time of Algorithm \ref{alg:baseline_marking} is dominated by a three-level nested loop with time complexity $O(N^2L)$, where $N$ is the number of nodes and $L$ is the number of edges. We observe that marking edges is arduous, while checking the edge mark only needs \emph{one} memory access, which is in constant time.
\begin{algorithm}
\caption{Marking Edges in Baseline Program}
\label{alg:baseline_marking}
\KwIn{Edge $(u, v)$}
$lca$ = the lowest common ancestor of $u$ and $v$ \\
$beta$ = max(min(depth[$u$], depth[$v$]) - depth[$lca$], 1) \\
$S_1$ = the nodes covered by $u$ with distance $beta$ \\
$S_2$ = the nodes covered by $v$ with distance $beta$ \\
\For{$x \in S_1$}{
\For{$y \in S_2$}{
\For{$e \in E$} {
\If{e = (x, y) or e = (y, x)}{
Mark edge $e$
}
}
}
}
\end{algorithm}
To mitigate the cost of marking edges, we start with the properties of \emph{marked edges}. An edge marked by $(u,v)$ must consist of two nodes covered by $u$ and $v$ respectively. By attaching edge marks to covered nodes, we can achieve $O(N)$ and $O(L)$ time complexity for marking edges and checking the edge mark respectively, which is described in Algorithm \ref{alg:linear_marking} and Algorithm \ref{alg:linear_checking}.
\begin{algorithm}
\caption{Marking Edges in Linear Time}
\label{alg:linear_marking}
\KwIn{Edge $e$ = $(u, v)$}
$lca$ = the lowest common ancestor of $u$ and $v$ \\
$beta$ = max(min(depth[$u$], depth[$v$]) - depth[$lca$], 1) \\
$S_1$ = the nodes covered by $u$ with distance $beta$ \\
$S_2$ = the nodes covered by $v$ with distance $beta$ \\
\For{$x \in S_1$}{
$M_x$ = $M_x \cup \{ e_1 \}$
}
\For{$y \in S_2$}{
$M_y$ = $M_y \cup \{ e_2 \}$
}
\end{algorithm}
\begin{algorithm}
\caption{Checking Edge Mark in Linear Time}
\label{alg:linear_checking}
\KwIn{Edge $(u, v)$}
\KwOut{Mark flag $flag$}
$flag$ = False \\
\For {$e \in E$ } {
\If{$e_1 \in M_u$ and $e_2 \in M_v$}{
$flag$ = True
}
\If{$e_1 \in M_v$ and $e_2 \in M_u$}{
$flag$ = True
}
}
\end{algorithm}
By the analysis of edges in test cases, we observe that the majority of off-tree edges are crossing edges, i.e. the edges that connect two nodes such that they do not have any ancestor and a descendant relationship between them.
We then focus on the properties of crossing edges, giving Lemma \ref{lemma:rlca} and Lemma \ref{lemma:lca}.
\begin{lemma}
\label{lemma:rlca}
For edge (x, y) and crossing edge (u, v), if edge (x, y) is covered by edge (u, v), then LCA(x, y) = LCA(u, v).
\end{lemma}
\begin{proof}
Without loss of generality, assume $x$ is covered by $u$ and $y$ is covered by $v$.
Let $w$ = LCA($u$, $v$). Note that $(u, v)$ is a crossing edge, which means $w \neq u$, then depth[$u$] $>$ depth[$w$]. So $beta \leq $ depth[$u$] - depth[$w$]. By $w \neq u$, $u$ is in a subtree of $w$, denoted by $T_1$. Note that $x$ is covered by $u$ with distance $beta \leq $ depth[$u$] - depth[$w$], then $x$ is in $T_1$ or $x$ is $w$.
Similarly, $y$ is in the subtree $T_2$ of $w$, or $y$ is $w$. Note that $T_1 \cap T_2$ = $\emptyset$, so LCA($x$, $y$) = $w$ = LCA($u$, $v$).
\end{proof}
\begin{lemma}
\label{lemma:lca}
For crossing edge (x, y) and crossing edge (u, v), if node x and node y are covered by node u or node v, and LCA(x, y) = LCA(u, v), then edge (x, y) is covered by edge (u, v).
\end{lemma}
\begin{proof}
We prove this lemma by the method of contradiction. Assume the edge $(x, y)$ is not covered by edge $(u, v)$. Without loss of generality, assume node $x$ and $y$ are both covered by node $u$.
Let $w$ = LCA($x$, $y$) = LCA($u$, $v$). Then $beta$ = min(depth[$u$], depth[$v$]) - depth[$w$] $\leq$ depth[$u$] - depth[$w$].
Note that $(x, y)$ is a crossing edge and LCA($x$, $y$) = $w$, then max(dist($x$, $u$), dist($y$, $u$)) $>$ dist($w$, $u$) = depth[$u$] - depth[$w$] $\geq beta$. So at least one node between $x$ and $y$ is not covered by $u$, which contradicts with the assumption.
\end{proof}
The converse proposition of Lemma \ref{lemma:lca} is also true, of which the proof is omitted. Then we have Algorithm \ref{alg:cross_marking} and Algorithm \ref{alg:cross_checking} to process crossing-edge marking and checking.
\begin{algorithm}
\caption{Crossing-edge Marking}
\label{alg:cross_marking}
\KwIn{Crossing edge $e$ = $(u, v)$}
$lca$ = the lowest common ancestor of $u$ and $v$ \\
$beta$ = min(depth[$u$], depth[$v$]) - depth[$lca$] \\
$S_1$ = the nodes covered by $u$ with distance $beta$ \\
$S_2$ = the nodes covered by $v$ with distance $beta$ \\
\For{$x \in S_1$}{
$M_{lca,x}$ = $M_{lca,x} \cup \{ e \}$
}
\For{$y \in S_2$}{
$M_{lca,y}$ = $M_{lca,y} \cup \{ e \}$
}
\end{algorithm}
\begin{algorithm}
\caption{Checking Crossing-edge Mark}
\label{alg:cross_checking}
\KwIn{Edge $(u, v)$}
\KwOut{Mark flag $flag$}
$lca$ = the lowest common ancestor of $u$ and $v$ \\
\uIf{$u \neq lca$ and $v \neq lca$}{
$flag$ = $M_{lca,u} \cap M_{lca,v}$ is not empty
}\Else{
$flag$ = False \\
\For{$e \in M_{lca,u} \cap M_{lca,v}$}{
\If{(u, v) is covered by $e$}{
$flag$ = True
}
}
}
\end{algorithm}
Although Algorithm \ref{alg:cross_checking} has no improvement on time complexity compared with Algorithm \ref{alg:linear_checking}, it can leverage the classic acceleration techniques for set operations, such as bitmap and SIMD, to reduce both running time and memory footprint.
As for non-crossing edges, the main obstacle is their aftereffects, which may withdraw marks of some crossing edges to effect more edges. So we designed an extra stage to recover non-crossing edges and deal with the complex aftereffects, which is described in Algorithm \ref{alg:noncross_fix}.
\begin{algorithm}
\caption{Recovering Non-crossing Edges}
\label{alg:noncross_fix}
\For{$e \in E$ }{
$flipped$ = False \\
\If{(e.isMarked and e.isEnforced) or \\
(not e.isMarked and e.isWithdrawn)}{
$flag$ = checking all-edge mark for $e$ \\
\If{$flag \neq e.isMarked$ }{
$flipped$ = True \\
$e.isMarked$ = $flag$
}
}
\If{(not e.isCrossing and not e.isMarked) \\ or $flipped$}{
\For{$e' \in $ edges covered by $e$}{
\uIf{$e.isMarked$}{
$e'.isEnforced$ = True
}\Else{
$e'.isWithdrawn$ = True
}
}
}
}
\end{algorithm}
\subsection{Computing Resistance in Linear Time}
The subroutine of computing resistance in baseline program is involved with operations on Laplacian matrix of minimal spanning tree. Considering the size of Laplacian matrix, the running time of operations on Laplacian matrix is at least quadratic. To achieve linear time complexity, we make use of the resistance computation subroutine of [1], which is proved to run in $O(L)$ time.
In our implementation, we made a slight change on LCA computation but greatly reduce its time consumption. For an off-tree edge $(u, v)$, we first check if $u$ and $v$ are in the same subtree of root. If so, compute the LCA of $u$ and $v$ as former procedure. Otherwise, the LCA of $u$ and $v$ must be root.
\subsection{Sorting Edges in Linear Time}
The baseline program utilizes stable sort function of STL library, whose time complexity is $O(L\log L)$. We notice that our data to sort is a bunch of non-negative double-precision floating-point number excluding NaN and infinity. By the IEEE standard [2], for 64-bit binary representations $a$ and $b$, if FP($a$) $<$ FP($b$), then INT($a$) $<$ INT($b$). That is, we can sort our data in an INT64 manner.
We choose radix sort as our basic sort algorithm for its stability, linearity and simplicity. Each 64-bit number is divided into eight unsigned 8-bit integer keys, because 256 int-type bucket counters can be put into exactly one page. With one round of scanning and eight rounds of relocation, the edges can be sorted in $O(L)$ time.
\section{Parallel Optimizations}
\label{sec:ParallelLGRASS}
\subsection{An Analysis of Hot-spot Subroutines}
Table \ref{tab:lgrass_breakdown} shows the time breakdown of basic LGRASS.
By our analysis, the subroutine of marking edges consumes up to 43.5\% of overall running time, while the other subroutines except LCA computation consumes 9.6\%-25.1\% of overall running time each.
\begin{table}[ht]
\centering
\caption{Time breakdown of basic LGRASS}
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline
& EFF & MST & LCA & RES & MARK & ALL\\
\hline
Case 1 & 2.8ms & 4.3ms & 2.1ms & 3.3ms & 4.6ms & 17.1ms \\
\hline
Case 2 & 1.5ms & 2.4ms & 1.1ms & 1.5ms & 5.0ms & 11.5ms\\
\hline
Case 3 & 14.9ms & 14.1ms & 9.5ms & 10.4ms & 33.5ms & 82.4ms \\
\hline
\end{tabular}
\label{tab:lgrass_breakdown}
\end{table}
Diving into these subroutines, we identified four parallel-optimization opportunities: edge marking, resistance computation, BFS (for computing effectiveness) and merge sort (for MST edge sort and resistance edge sort).
In the following subsections, we depict how we exploit these opportunities.
\subsection{Parallel Edge Marking}
By Lemma \ref{lemma:rlca}, the task of crossing-edge marking can be partitioned based on the LCA of off-tree edges. However, the majority of off-tree edges recognize root as their LCA, such that the workload of root-LCA subtask becomes much heavier than that of other subtasks. So we made a further partition based on which two subtrees those edges connect. The two-step partition is implemented with a carefully designed mapping function:
\begin{equation*} F(u, v) =
\begin{cases}
LCA(u, v) &\text{if $LCA(u, v) \neq root$} \\
N &\text{if $u = root$ or $v = root$} \\
N + 1 + \binom{S_1}{2} + S_2 &\text{otherwise}
\end{cases} \end{equation*}
where $N$ is the number of nodes, $S_1$ and $S_2$ are the maximal and minimal subtree index of edge $(u, v)$ respectively. And the node index and subtree index are both ranged from 0. After partition, each subtask can be independently processed with Algorithm \ref{alg:cross_marking} and Algorithm \ref{alg:cross_checking}. When all subtasks are done, Algorithm \ref{alg:noncross_fix} is still necessary to deal with the aftereffects of non-crossing edges.
The comparison of edge marking procedure between basic LGRASS and parallel LGRASS is demonstrated in Fig. \ref{figure:parallel}.
Besides, we implemented a greedy scheduling algorithm for dynamic task dispatching, which outperforms static task dispatching when using more than 4 threads.
\subsection{Parallel Resistance Computation }
\begin{figure}[htbp]
\centering
\subfloat[Basic LGRASS]{
\includegraphics[width=1.5in]{figures/v-a-c.pdf}
}
\subfloat[Parallel LGRASS]{
\includegraphics[width=1.5in]{figures/v-b.pdf}
}
\caption{Procedures of Edge Marking and Resistance Computation}
\label{figure:parallel}
\end{figure}
To make full use of multiprocessors, the resistance computation task is uniformly partitioned for each thread. For the same purpose, part of LCA computation is offloaded to resistance computation subroutine, and an efficient online LCA algorithm is implemented replace the former one.
Fig. \ref{figure:parallel} shows the difference of LCA and resistance computation between the basic version and parallel version of LGRASS.
\subsection{Parallel Breadth-first Search}
By our profiling result, the running time of computing effectiveness is dominated by BFS. To alleviate this, we implemented a level-synchronous parallel BFS algorithm. By our implementation, each thread first obtains vertices of current layer from a concurrent queue, then searches for the vertices of next layer and pushes them into the concurrent queue for next layer. We employ an atomic variable to record the number of elements in a concurrent queue, which ensures the linearizability of concurrent operations for multiple threads. Besides, we use atomic operations on distance array in relaxed memory order to prevent a vertex from being added to the queue too many times.
We also took several tricks for better performance: relaxed memory order for mutex variables, hand-crafted assembly code for compare-and-set operations and local-to-global merging for layer aggregation.
Fig. \ref{figure:pbfs} is an example of our parallel BFS scheme.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.8\linewidth]{figures/pbfs.PNG}
\caption{Parallel Breadth-first Search}
\label{figure:pbfs}
\end{figure}
\subsection{Parallel Merge Sort}
Considering the inherent intractability of sequential radix sort, we take merge sort as the main framework of our parallel sort and regard radix sort as its submodule.
By our design, each thread first sorts a block of the data by radix sort, then merge the blocks in parallel. If there are $P$ threads for merge sort, the thread with the longest running time need to perform up to $(2-\frac{1}{P})L$ comparisons.
We observe that only a small fraction of sorted data is accessed by later procedure, so we leave the final merger to later procedure
and reduce the maximum number of comparisons down to $(1-\frac{1}{P})L$ in merge sort.
Furthermore, when sorting edges by resistance, we can merge only top-$K$ edges, where $K$ depends on the number of edges for partition. Then each thread only needs to perform no more than $(\lfloor \log_2{P} \rfloor - 1)K$ times comparisons in merge sort. The procedure is shown in Fig. \ref{figure:pmergesort}.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.8\linewidth]{figures/mergesort.PNG}
\caption{Parallel Merge Sort of Edges with Resistance}
\label{figure:pmergesort}
\end{figure}
\section{Evaluation}
\begin{table}[ht]
\centering
\caption{Execution time comparison}
\begin{tabular}{|c|c|c|c|}
\hline
& Baseline & Basic LGRASS & Parallel LGRASS\\
\hline
Case 1 & 22.9min & 15.3ms & 6.4ms \\
\hline
Case 2 & 25.5min & 10.8ms & 5.0ms\\
\hline
Case 3 & 37.1h & 73.7ms & 24.0ms \\
\hline
\end{tabular}
\label{tab:comparison}
\end{table}
We evaluated our method on hardware platform provided by Beijing Super Cloud Computing Center, which is specified by organizers. Table \ref{tab:comparison} demonstrates the comparison among baseline program, basic LGRASS and parallel LGRASS. Compared with baseline program, basic LGRASS can reduce the execution time by several orders of magnitude with only \emph{one} processor. By utilizing multiprocessors, parallel LGRASS gains up to 3.1x speedup against basic LGRASS.
We also evaluated our proposed method on random test cases. As is shown in Fig. \ref{figure:linearity}, the using time of parallel LGRASS increases linearly as graph size scales up.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.7\linewidth]{figures/Linearity.png}
\caption{Using Time of Parallel LGRASS on Random Test Cases}
\label{figure:linearity}
\end{figure}
\section{Conclusion}
In this paper, we present our solution to the final task of the 3rd ACM-China IPCC. We proposed LGRASS, a linear graph spectral sparsification algorithm, and optimized it for parallel processing. Evaluation result shows that our proposed method can fulfils the task in 6.4ms, 5.0ms and 24.0ms respectively on official test cases and keep its linearity as graph size scales up on random test cases.
\ifCLASSOPTIONcompsoc
\section*{Acknowledgments}
\else
\section*{Acknowledgment}
\fi
We thank our shepherd, Jian Zhang, for his feedback and encouragement.
|
\section{Introduction} \label{sec:intro}
The wave instabilities generated by electron beams in various plasma setups are invoked in many applications in astrophysics and space plasmas. Magneto-genesis in galaxies and intergalactic medium can be explained by the Weibel-type instabilities involving interpenetrating electron beams, which can produce the primordial magnetic field seeds \citep{Schlickeiser2005, Lazar2009} to be amplified by cosmological dynamos \citep{Beck-2015}.
Nonthermal emissions from cosmological sources, e.g., gamma-ray bursts, active galactic nuclei, pulsars, etc., often invoke two-stream electrostatic (ES) or electromagnetic (EM) instabilities for the relaxation of the relativistic electron jets \citep{Stockem-etal-2007, Bret-2009, Schlickeiser-etal-2013}.
Radio emissions (e.g., type-II or type-III bursts) associated to solar flares and heliospheric shocks driven by coronal mass ejections (CMEs), are believed to be the result of an ES Langmuir relaxation of electron beams with energy in the range of a few hundred keV \citep{Ganse-etal-2012, Lee-etal-2019}.
Less energetic (up to a few keV) but more thermalized beams, known as strahls, are more recurrent in the solar wind, undergoing a continuous erosion with heliocentric distance \citep{Maksimovic-etal-2005, Anderson-etal-2012}, most probably due to the self-induced instabilities and wave fluctuations \citep{Verscharen-etal-2019, Che2019, Micera-etal-2020}, whose nature highly depend on the properties of electron strahls \citep{Lopez-etal-2020a}.
Recently, \cite{Lopez-etal-2020b} have revealed the existence of a firehose-like aperiodic instability driven by two symmetric populations of electrons, counter-moving along the regular magnetic field. It develops only at oblique angles of propagation with respect to the regular magnetic field, and remains aperiodic only if the electron counter-beams are perfectly symmetric, namely, with the same number densities, the same temperatures and the same beaming or drifting speeds.
These properties and the fact that the trigger is an excess of kinetic (free) energy of electrons in parallel direction, makes this instability similar to the aperiodic firehose instability driven by a temperature anisotropy $T_\parallel > T_\perp$ with $\parallel, \perp$ denoting directions relative to the magnetic field) \citep{Li2000, Gary2003, Camporeale2008, Shaaban2019, Lopez2019a, Moya-etal-2022}.
However, the ignition of this instability is conditioned by a relatively small drift ($v_d$) between counter-beaming electrons, that should not exceed the electron thermal speed ($\alpha_e$), otherwise (if, e.g., $v_d > \alpha_e$), it is highly competed by the ES instabilities. Indeed, recent results provided by \citet{Moya-etal-2022} have shown that in the $v_d/\alpha_e >1$ regime, quasi-parallel ES instabilities dominate with growth rates about one order of magnitude larger than the aperiodic (oblique) firehose instability \footnote{For a more fluid-like regime of electron (counter-)beams the instability wave spectra is dominated by the ES fluctuations (with the highest growth rates), while the EM or hybrid fluctuations may be predicted either for low drifts, i.e., $v_d < \alpha_e$, or for very high, relativistic beaming speeds corresponding to more than 100~keV \citep{Gary2003, Lazar2009, Bret-2009, Jao-Hau-2016}.}.
Focusing on solar plasma outflows, the observations have revealed electron counter-beams, parallel and anti-parallel to the local magnetic field (in the literature also known as couterstreaming or bi-directional electrons, or even double strahls), since the first satellite missions. e.g., VELA~5 and 6, and IMP~6 \citep{Montgomery-etal-1974}, ISEE~3 \citep{Bame-etal-1981,Gosling-etal-1987}, but also by heliospheric spacecraft, such as Helios 1 and 2 \citep{Pilipp-etal-1987,Macneil-etal-2020}, as well as Ulysses outside the ecliptic \citep{Hammond-etal-1996,Lazar-etal-2014}.
Electron counter-beams have also been reported by the more recent missions, e.g., ACE \citep{Skoug-etal-2000,Steinberg-etal-2005,Anderson-etal-2012}, Wind \citep{Larson-etal-1996,Fitzenreiter-etal-2003, Nieves-Chinchilla-2008}, STEREO A/B \citep{Lavraud-etal-2010,Kajdic-etal-2013,Carcaboso-etal-2020}.
In situ observations associate these electron counter-beams mainly with the foreshock or upstream regions of the bowshock \citep{Fitzenreiter-etal-2003} and interplanetary shocks driven by corotating interactions regions (CIRs) \citep{Lavraud-etal-2010} and coronal plasma ejection (CMEs) \citep{Lazar-etal-2014, Cremades-etal-2015}. More energetic beams are also associated with closed magnetic topology typical of CMEs with expanding coronal loops, more symmetric counter-beams at the apex, and less symmetric on flanks \citep{Skoug-etal-2000, Lazar-etal-2014}.
The properties of electron (counter-)beams can be determined either directly from the in-situ measurements of the velocity distributions \citep{Maksimovic-etal-2005, Nieves-Chinchilla-2008, Bercic-etal-2019}, or indirectly, either from the analysis of the enhanced fluctuations detected also in-situ \citep{Pulupa-Bale-2008}, or from the analysis of radio (or even harder) emissions of different nature, e.g., type-II bursts from upstream of the shock, or type-IV bursts related to the CME magnetic clouds \citep{Pick-Vilmer-2008}. The radiative or non-radiative relaxation of electron (counter-)beams may involve a wide palette of plasma wave instabilities, which depend on the properties of these beams including, as already mentioned above, instabilities of ES, EM or hybrid waves, periodic or aperiodic, and propagating parallel or obliquely to the magnetic field \citep{Verscharen-etal-2019, Lopez-etal-2020a,Moya-etal-2022}.
Here we investigate the aperiodic firehose instability driven by the counter-beaming electrons \citep{Lopez-etal-2020b,Moya-etal-2022}, taking into account that under common conditions in the heliosphere, the electron counter-beams are embedded in a background plasma of stationary electrons and ions, mainly protons \citep{Fitzenreiter-etal-2003, Lavraud-etal-2010, Anderson-etal-2012, Carcaboso-etal-2020}.
We will call it the beaming electron firehose instability, BEFI, for short.
In the next section we introduce the model for the electron distribution function, with two counter-beams and a stationary background population, as well as the analyzed parametric cases. The results from the linear kinetic theory of wave instabilities are presented in section 3. We analyze both the dispersion of the frequency and the growth rate, as a function of the wave number, but also the threshold conditions of BEFI as a function of the main parameters of the plasma. Moreover, Section 4 presents the results from the numerical simulations which, at this moment, seek to bring a confirmation of the predictions of the linear theory, and to qualitatively describe the temporal evolution of BEFI.
In the last section, we conclude our results, and discuss potential implications of this instability in heliospheric applications.
\section{Plasma model and parameters} \label{model}
We consider a plasma of electrons (subscript $e$) and protons (subscript $p$), the dominant plasma species in the solar outflows. Of interest for present study are events revealing counterbeaming electrons, e.g., associated with interplanetary shocks, CIRs and CME closed magnetic fields, and where the electron velocity distributions (VDs) exhibit three distinct populations
\begin{align}
f_e\left({ v_{\perp},v_{\parallel}}\right) = &
\frac{n_{0}}{n_e}~f_{0}\left({ v_{\perp},v_{\parallel}}\right)
+\frac{n_{1}}{n_e}~f_{1}\left({ v_{\perp},v_{\parallel}}\right) + \frac{n_{2}}{n_e}~f_{2}\left({ v_{\perp},v_{\parallel}}\right). \label{e1}
\end{align}
With numbering subscripts we indicate the stationary background component - subscript 0, and the more or less symmetric counterbeams - subscripts 1 and 2. Relative densities $n_j/n_e$ ($j=0, 1, 2$) are defined with respect to $n_e$, the total electron number density, equal to the proton density $n_e=n_p$.
For each individual beam the VD is assumed a temperature isotropic drifting Maxwellian of the form
\begin{equation}
f_j(v_\perp,v_\parallel)=\frac{1}{\pi^{3/2}\alpha_j^3}
\exp\left\{-\frac{v_\perp^2}{\alpha_j^2}
-\frac{(v_\parallel-U_j)^2}{\alpha_j^2}\right\}\,, \label{e2}
\end{equation}
where $\alpha_j=(2k_BT_j/m_e)^{1/2}$ is the thermal velocity and $U_j$ the drift velocity of the $j$-th beam. Using the zero-current condition, we find the drift velocities related by $n_2 U_2= -n_1 U_1$.
Stationary or non-drifting ($U_0 = 0$) background electrons are modelled by an isotropic Maxwellian distribution, and can be cooler or hotter than electron beams.
Protons are also considered a stationary (neutralizing) background, described by an isotropic non-drifting Maxwellian VD, and with the same temperature as the background electrons.
We also assume the electron counter-beams (subscript $b$ in the next) sufficiently symmetric, i.e., with the same relative drifts, $|U_1| = |U_2| = U_b$, the same number densities, $n_1 = n_2 = n_b = (n_e-n_0)/2$, and the same thermal velocities, $\alpha_1 = \alpha_2 = \alpha_b$.
Although in CIRs and interplanetary shocks asymmetric counter-beams are far more likely to occur, the symmetry considered here allows us to reduce the parameters space, and thus focus on the effects of the background electron population, for a series of new cases obtained by varying the main properties of electron populations, which contrast to the previous results reported by \cite{Lopez-etal-2020b} and \cite{Moya-etal-2022} for $n_0 = 0$.
The dispersion and stability properties are investigated for different parametrizations of electron populations, established by keeping constant the relative beaming speed $U_b = 0.065 c$ (where $c$ is the speed of light in vacuum), and varying thermal velocities $\alpha_0$ and $\alpha_b$, and also relative number densities, e.g., $n_b/n_e$.
Table~\ref{t1} shows the parameters for the most relevant plasma configurations that we have analyzed. These are classified in four cases corresponding to different thermal velocities $\alpha_0$ and $\alpha_b$ (in units of $c$), and for each of them a number of four subcases defined by different relative number densities of the beams $n_b/n_e$.
It should be noted that we also chose thermal velocities slightly lower than the beam or drift velocity, regimes for which previous qualitative estimates indicated a possible competition with electrostatic instabilities (parallel to the magnetic field). Here we will also discuss these regimes of transition from the dominance of BEFI to that of electrostatic instabilities, which are found to be sensitive not only to the properties of the counter-beams, but also to the presence of background electrons.
First cases with a lower relative density of the electron counter-beams are more relevant to the space plasma conditions, including conditions at CIRs, CMEs and interplanetary shocks, while the other cases with a lower density of the background electrons are more close to the configuration studied by \cite{Lopez-etal-2020b}.
For similar relative beaming speeds, with, e.g. $U_b/c = $ 0.06, 0.065, 0.07, and moderate values of plasma betas $\beta_e =$ 2, 4, but in the absence of the background population of electrons, linear theory predicts high enough maximum growth rates of beaming electron firehose instability (BEFI), and PIC simulations confirm that this unstable mode develops and can be faster than electrostatic instabilities \citep{Lopez-etal-2020b,Moya-etal-2022}.
\begin{table}[t!]
\caption{\label{t1} Four cases corresponding to different thermal velocities $\alpha_0$ and $\alpha_b$ (in units of $c$), and for each case other four distinct sub-cases as defined by different relative number densities of the electron background $n_0/n_e$ or beams $n_b/n_e$. For all cases $U_b/c = 0.065$ and $\beta_p = 1.96$.}
\centering
\begin{tabular}{l c c c c c c}
\hline\hline
Case & $\alpha_0/c$ & $\alpha_b/c$ & $n_0/n_e$ & $n_b/n_e$ & $\beta_0$& $\beta_b$ \\ [0.5ex]
\hline
1.a & 0.07 & 0.07 & 0.6 & 0.20 & 1.176 & 0.392 \\
1.b & 0.07 & 0.07 & 0.5 & 0.25 & 0.980 & 0.490 \\
1.c & 0.07 & 0.07 & 0.3 & 0.35 & 0.588 & 0.686 \\
1.d & 0.07 & 0.07 & 0.1 & 0.45 & 0.196 & 0.882 \\
\hline
2.a & 0.04 & 0.07 & 0.6 & 0.20 & 0.384 & 0.392 \\
2.b & 0.04 & 0.07 & 0.5 & 0.25 & 0.320 & 0.490 \\
2.c & 0.04 & 0.07 & 0.4 & 0.30 & 0.256 & 0.588 \\
2.d & 0.04 & 0.07 & 0.1 & 0.45 & 0.064 & 0.882 \\
\hline
3.a & 0.02 & 0.07 & 0.6 & 0.20 & 0.096 & 0.392 \\
3.b & 0.02 & 0.07 & 0.5 & 0.25& 0.08 & 0.490 \\
3.c & 0.02 & 0.07 & 0.3 & 0.35 & 0.048 & 0.686 \\
3.d & 0.02 & 0.07 & 0.1 & 0.45 & 0.016 & 0.882 \\
\hline
4.a & 0.07 & 0.04 & 0.6 & 0.20 & 1.176 & 0.128 \\
4.b & 0.07 & 0.04 & 0.5 & 0.25 & 0.980 & 0.160 \\
4.c & 0.07 & 0.04 & 0.4 & 0.30 & 0.784 & 0.192 \\
4.d & 0.07 & 0.04 & 0.3 & 0.35 & 0.588 & 0.224 \\
\hline
\end{tabular}
\end{table}
In the present analysis, the plasma beta parameter $\beta_j = 8 \pi n_j k_B T_j / B_0^2$ ($j= 0, b$) is calculated for each electron population using the corresponding number density $n_j$ and temperature $T_j$.
Our parameterization, see Table~\ref{t1}, attempts to cover conditions specific to interplanetary shocks triggered by the fast winds, e.g., in CIRs, where $\beta_e \gtrsim 1$, but also the low $\beta_e \lesssim 1$ conditions, more characteristic to CMEs.
Calculated with the total number density the plasma frequency $\omega_{p,e}^2 = (4 \pi n_e e^2/m_e)^{1/2}$ intervenes in the normalization of the wave number, while the electron gyrofrequency $\Omega_e = |e| B_0/(m_e c)$ in the normalization of wave frequency and growth rate.
For the plasma frequency to gyrofrequency ratio we consider $\omega_{p,e}/\Omega_e = 20$, which is relevant for solar wind conditions, and ensures a reasonable computational time in the numerical simulations (see section \ref{pic}). We analyze the full spectrum of unstable modes, triggered by the relative drift of the counter-beams (for all angles of propagation with respect to the magnetic field), by using the dispersion solver developed in \cite{Lopez2019b} and \cite{Lopez-etal-2021}.
Previous studies have shown that BEFI is mainly conditioned by the ratio between this drift, $U_b$, and the thermal velocity of the beaming electrons, $\alpha_b$.
Thus, if $\alpha_b$ is higher, and, implicitly, the corresponding $\beta_b$ is higher, then $U_b$ must also increase in order to excite the instability \citep{Lopez-etal-2020b, Moya-etal-2022}.
This condition shapes the drifting velocity thresholds of BEFI, see Fig.~3 in \cite{Lopez-etal-2020b}, and resembles that of the firehose heat-flux instability induced at (quasi-)parallel angles of propagation by the uni-directional electron strahls/beams (carrying the main heat-flux in the solar wind), see, e.g., Fig.~11 in \cite{Shaaban2018}.
\section{Linear theory} \label{befi}
\subsection{BEFI with background electrons: cases 1 and 2}
We first discuss the linear dispersion properties of BEFI, through a parametric analysis that allows the characterization of different regimes of this instability, when predicted to be dominant or in competition with other unstable modes.
Table~\ref{t1} presents plasma configurations which are found relevant for the existence of BEFI. These are classified into four cases, corresponding to different thermal velocities $\alpha_0$ and $\alpha_b$ (in units of $c$), and other four distinct sub-cases, as defined by various relative number densities of the electron background $n_0/n_e$ (or electron beams $n_b/n_e$). For all cases we consider the same relative drift $U_b/c = 0.065$, and for protons $\beta_p = 1.96$.
Displayed in Figs.~\ref{f1}--\ref{f7} are the results for all these configurations, that allow us to delimit the specific regimes of this instability: (i) Near the instability thresholds, e.g., cases 1.a, 2.a, 3.a and 4.a, the (maximum) growth rates are very small, approaching marginal stability, i.e., $\gamma \to 0$, but, luckily, there is no other instability predicted by the theory in competition with BEFI.
(ii) Regimes when additional instabilities can be identified in the wave spectra, but against which BEFI remains the dominant unstable mode, i.e., with the highest (maximum) growth rates. Overall, these results should show us how this instability is influenced by the background plasma, both by the relative density, $n_0/n_e$, and the thermal velocity, $\alpha_0/c$, of the background electrons. In Figs.~\ref{f1}, \ref{f4}, \ref{f5} and \ref{f6}, the white background signifies levels below the minimum level in the color bars (on the right) used to quantify the growth rate or wave frequency.
\begin{figure*}[t!]
\begin{center}
\includegraphics[width=0.9\textwidth]{case1}
\caption{\label{f1} Color coded growth rates $\gamma/\Omega_e$ (top panels) and wave frequencies $\omega/\Omega_e$ (bottom panels) for cases 1.a, 1.b, 1.c and 1.d (from left to right).}
\end{center}
\end{figure*}
\begin{figure*}[h!]
\begin{center}
\includegraphics[width=0.9\textwidth]{disp_1c}
\caption{\label{f2} The wave-number dispersion for the fastest growing unstable modes (with maximum growth rates) in case 1.c, corresponding to different angles of propagation ($\theta$) and different ranges of unstable wave-numbers: BEFI at $\theta = 65^{\rm o}$ (black), oblique HFI at $\theta = 45^{\rm o}$ (green) and FHFI at $\theta = 0^{\rm o}$ (orange). Top panels: wave frequency (left), growth rates (center), and polarization (right). Middle panels: longitudinal and transverse components of the wave electric components. Bottom panels: cartesian components of the wave magnetic field. See also details in the text.}
\end{center}
\end{figure*}
Fig.~\ref{f1} presents the four sub-cases of case 1, when thermal velocities of the beaming and background populations are comparable $\alpha_b/c \simeq \alpha_0/c = 0.07$, and are only slightly higher than beaming speed $U_b/c = 0.065$.
The unstable spectra show the wave number ($ck/\omega_{pe}$) dispersion of the growth-rate ($\gamma/\Omega_e$, upper panels) and the wave frequency ($\omega_r/\Omega_e$, lower panels), as a function of the propagation angle ($\theta$).
Both the growth rate and wave frequency are color coded on the right side of the respective panels.
Very low growth rates of BEFI are obtained in case 1.a (left panels), with, e.g., a maximum $\gamma_{\rm max}/\Omega_e = 0.007$, when background electrons have a major density $n_0/n_e = 0.6 > n_b/n_e = 0.2$.
With decreasing the density contrast between background and beaming electrons the peaking growth rate of BEFI increases with one order of magnitude for case 1.b, i.e., for $n_0/n_e =0.5$ and $n_b/n_e=0.25$, and may reach $\gamma_{\rm max} / \Omega_e \simeq 0.21$ for case 1.d, for even more dense electron beams with $n_b/n_e = 0.45 > n_0/n_e = 0.1$.
The results from case 1.d, are indeed very similar to those obtained by \cite{Lopez-etal-2020b} in the absence of background electrons.
The comparison of the four cases shows a clear and significant inhibition of BEFI under the influence of background electrons.
This instability remains aperiodic ($\omega_r = 0$) for all these plasma configurations, as a specific feature that may help to differentiate from other unstable modes.
\begin{figure*}[t!]
\begin{center}
\includegraphics[width=0.9\textwidth]{disp_1d}
\caption{\label{f3} The wave-number dispersion for the fastest growing unstable modes (with maximum growth rates) in case 1.d, corresponding to different angles of propagation ($\theta$) and different ranges of unstable wave-numbers: BEFI at $\theta = 68^{\rm o}$ (black), oblique HFI at $\theta = 40^{\rm o}$ (green), beaming ESI at $\theta = 0^{\rm o}$ (purple), and FHFI at $\theta = 0^{\rm o}$ (orange). Top panels: wave frequency (left), growth rates (center), and polarization (right). Middle panels: longitudinal and transverse components of the wave electric components. Bottom panels: cartesian components of the wave magnetic field. See also details in the text.}
\end{center}
\end{figure*}
In the unstable spectra for cases 1.c and 1.d, we can also distinguish other modes of different nature, in general with finite wave frequency $\omega_r \ne 0$, but (maximum) growth rates much lower than those of BEFI. Distinction can still be made between these two spectra. Thus, in panels 1.c, at small quasi-parallel angles $\theta < 20^{\rm o}$ and small wave-numbers, we may identify the firehose heat-flux instability (FHFI), and for more oblique angles and larger wave-numbers the oblique branches of heat-flux instabilities, which may combine FHFI and whistler heat-flux (WHF) instabilities, discussed to more detail in \cite{Lopez2019a}, see, e.g., their Fig.~3.
These heat-flux instabilities are triggered by an effective anisotropy in velocity space, as resulted from the asymmetry of thermal spread of the beaming and background populations. These instabilities may therefore not be very sensitive to the variation of (relative) number density.
For more dense counter-beams in case 1.d, the electrostatic beaming instabilities (i.e., with a major longitudinal electric field component $E_L = {\bf E} \cdot {\bf k}$) are also predicted at parallel and small $\theta$ propagation.
In this case the quasi-parallel unstable spectrum becomes already complex, showing a superposition of unstable modes with finite frequency, increasing with the wave number.
However, the maximum growth rates of all the other unstable modes remain much lower than those of BEFI. Clearly, in all these cases BEFI is not competed by other instability, and it is solely predicted to operate as the main radiative mechanism, with possible consequences on the relaxation of electron counter-beams. The corresponding plasma beta parameters, $\beta_0$ and $\beta_b$, see Table~\ref{t1}, take comparable values, around or slightly lower than 1, which means conditions near the equipartition of kinetic and magnetic energy, specific to the solar wind, in, e.g., CIRs and terrestrial bow-shock.
\begin{figure*}[t!]
\begin{center}
\includegraphics[width=0.9\textwidth]{case2}
\caption{\label{f4} Color coded growth rates $\gamma/\Omega_e$ (top panels) and wave frequencies $\omega/\Omega_e$ (bottom panels) for cases 2.a, 2.b, 2.c and 2.d, from left to right.}
\end{center}
\end{figure*}
Aiming to decipher the nature of the unstable modes, in Fig.~\ref{f2} we describe the main properties of the fastest growing modes, i.e., those with maximum growth rates, corresponding to each instability in case 1.c from Fig.~\ref{f1}.
Upper panels in Fig.~\ref{f2} show the wave frequency (left), the growth rate (middle), and the polarization (right) defined as $P = \text{Sign}(\omega_r) \text{Re} \left\{i(E_x/E_y)\right\}$.
This polarization is relevant for the electromagnetic modes, circular (or eliptically) polarized, $P> 0$ meaning right-handed (RH) polarization and $P < 0$ left-handed (LH) polarization. %
The maximum growth rate ($\gamma_{\rm max}/\Omega_e = 0.208$), associated with the fastest growing mode, is obtained for BEFI at $\theta \simeq 65^\circ$ (black lines), as an aperiodic mode ($\omega_r=0$) purely growing in time.
Lower panels in the middle row show components of the wave electric field, longitudinal (or parallel) and transverse to the direction of propagation (i.e., to ${\bf k}$), for three distinct modes: BEFI with maximum growth rate at $\theta = 65^{\rm o}$ (black lines in the left panel); the oblique branch of the HFIs, in this case, firehose-like modes, LH-polarized ($P < 0$), and with maximum growth rate at $\theta = 45^{\rm o}$ (green lines in the middle panel); and for parallel propagation ($\theta = 0$) a FHFI, circularly LH-polarized with $P = -1 < 0$ and only a transverse component ($E_T$) of the wave electric field (orange lines in the right panel).
Shown in the bottom panels are the corresponding cartesian components of the wave magnetic fields, which confirm the nature of these modes. Notice that for BEFI the major component is $B_y$, which is another common feature with the firehose instability driven by the temperature anisotropy of electrons \citep{Camporeale2008}.
Fig.~\ref{f3} shows the same details as in Fig.~\ref{f2}, but for the main properties of the fastest growing unstable modes in case 1.d, those corresponding to the peaking growth rates in Fig.~\ref{f1}.
In this case the relative density of the background electrons is only $n_0/n_e =0.1$, and the unstable wave spectra resemble those obtained by \cite{Lopez-etal-2020b}, for similar plasma parameters but in the absence of background electrons.
For BEFI (black lines) the value of maximum grow rate is higher, and is obtained at $\theta \simeq 68^\circ$. The maximum growth rate of the oblique HFI (green lines) remains lower than that of BEFI, and is obtained at $\theta \simeq 40^\circ$.
But in this case, in the oblique HFI one may observd that FHF (LH-polarized, with $P <0$, at lower wave-numbers) couples with WHF (RH-polarized, with $P > 0$, at higher wave-numbers), as also shown by \cite{Lopez-etal-2020a}.
It should also be remarked the similarity of the properties of these modes with those obtained for an asymmetric electron plasma-beam system \citep{Lopez-etal-2020a}.
For parallel propagation ($\theta = 0^\circ$) we find not only the FHFI (orange lines), but also the electrostatic (ES) electron beaming instability (EBI, with purple lines). For both of them maximum growth rates are less than that of BEFI.
This oscillatory ($\omega \ne 0$) beaming mode is most probably excited by the asymmetric counter-drifting beam and background populations of electrons, by contrast to previous studies in the absence of background electrons \citep{Lopez-etal-2020b}, where symmetric counter-beams were able to trigger an aperiodic ($\omega = 0$) two-stream instability.
This ES mode seems to couple to EM modes, FHF modes with a $B_x$ component at low wave-numbers, and the other oblique, BEFI or WHF modes, with a major $B_y$ transverse component of the wave magnetic field. Apparently with a hybrid nature, this mode is not of interest in our present analysis, but could be investigated in future studies.
\begin{figure*}[t!]
\begin{center}
\includegraphics[width=0.9\textwidth]{case3}
\caption{\label{f5} Color coded growth rates $\gamma/\Omega_e$ (top panels) and wave frequencies $\omega/\Omega_e$ (bottom panels) for cases 3.a, 3.b, 3.c and 3.d, from left to right.}
\end{center}
\end{figure*}
\begin{figure*}[t!]
\begin{center}
\includegraphics[width=0.9\textwidth]{case4}
\caption{\label{f6} Color coded growth rates $\gamma/\Omega_e$ (top panels) and wave frequencies $\omega/\Omega_e$ (bottom panels) for cases 4.a, 4.b, 4.c and 4.d, from left to right.}
\end{center}
\end{figure*}
\begin{figure*}[t!]
\begin{center}
\includegraphics[width=0.9\textwidth]{disp_4d}
\caption{\label{f7} The wave-number dispersion for the fastest growing unstable modes (with maximum growth rates) in case 4.c, corresponding to different angles of propagation ($\theta$) and different ranges of unstable wave-numbers: BEFI at $\theta = 63^{\rm o}$ (black), oblique HFI-1 at $\theta = 55^{\rm o}$ (green), oblique HFI-2 at $\theta = 83^{\rm o}$ (blue), beaming ESI at $\theta = 0^{\rm o}$ (purple), and FHFI at $\theta = 0^{\rm o}$ (orange). Top panels: wave frequency (left), growth rates (center), and polarization (right). Middle panels: longitudinal and transverse components of the wave electric components. Bottom panels: cartesian components of the wave magnetic field. See also details in the text.}
\end{center}
\end{figure*}
Fig.~\ref{f4} presents the unstable solutions obtained for cases 2.a - 2.d, similar to cases 1.a - 1.d, but for a cooler background population, this time with $\alpha_0/c = 0.04 < \alpha_b/c = 0.07$.
For the same relative densities, see Table~\ref{t1}, profiles of the unstable spectra are similar to those obtained in Fig.~\ref{f1}, showing a dominance of BEFI. The highest peaking (maximum) growth rates are obtained for BEFI, in general, at oblique angles, which increase with lowering the influence of the background electrons (from left to right). The growth rates increase the same way, but their maximum values, indicated in each panel in Fig.~\ref{f4}, are lower than those obtained in Fig.~\ref{f1}, meaning that BEFI is inhibited by a cooler background population of electrons. Note also, that unlike case 1.d., the ES instabilities are missing from the unstable spectra of case 2.d, despite the similarity between the plasma configurations.
\begin{figure*}[h!t!]
\begin{center}
\includegraphics[width=0.9\textwidth]{gmax}
\caption{\label{f8} Maximum growth rates $\gamma_\text{max}/\Omega_e$ (color coded) obtained for BEFI for $\alpha_0=\alpha_b$ (cases 1), and different density ratios $n_b/n_e=0.15$ (left), $n_b/n_e=0.25$ (center), and $n_b/n_e=0.35$ (right). Levels above contours at $\gamma_{\rm max}/\Omega_e=0.3$ correspond to the ESI (middle and right panels).}
\end{center}
\end{figure*}
\subsection{From BEFI to ES instabilities: cases 3 and 4}
Next let us see the properties of BEFI for cases 3.a - 3.d, when the electron background population is even cooler, i.e., $\alpha_0/c = 0.02$, and for the same set of relative number densities, see Table~\ref{t1}.
The unstable solutions are displayed in Fig.~\ref{f5}, i.e., color coded wave frequencies (top panels) and growth rates (bottom panels). In this case, BEFI remains fairly distinct and dominant, with growth rates higher (or even much higher) than all the other modes predicted by the linear theory.
The inhibiting effect of a cooler electron background is confirmed, by, e.g., the maximum growth rates, which are indicated in each panel, and are lower than those obtained for the corresponding cases in Fig.~\ref{f4}.
However, in case 3.d, when the relative density of the background population is very low, i.e., $n_0/n_e = 0.1$, the growth rates of ES modes with (quasi-)parallel wave-vectors become important, they are still less than those of BEFI, but are already markedly higher than those of ES modes obtained in case 1.d.
This is what we can call, as also suggested by \cite{Lopez-etal-2020b} and \cite{Moya-etal-2022}, the transition from the dominance of BEFI to the regime of ES instabilities, specific to much cooler electron populations, see, e.g., Fig.4 in \cite{Moya-etal-2022}.
Although these electrostatic instabilities are not the object of our study, we can explain these results by mentioning that in the velocity distributions (not shown here) case 3.d shows peaks of the counterbeams and corresponding slopes ($\delta f_b/\delta v \propto \gamma > 0$) more prominent than those for case 1.d. In case 2.d (and also case 1.c) the same peaks and corresponding slopes are much lower, below the threshold of these instabilities.
In Figs.~\ref{f6} and \ref{f7} we show that this transition can be even steeper when the electron beams are cooler than background population, i.e., for cases 4.a - 4.d., when $\alpha_b = 0.04 < \alpha_0 = 0.07$.
In Fig.~\ref{f6} growth rates of BEFI (top panels) show the same inhibition under the influence of background electrons, but contrary to that, BEFI keeps the highest growth rate only for sufficiently dense background population, for instance, in cases 4.a and 4.b.
Already in case 4.b, but especially in the other two cases, 4.c and 4.d, the spectra of instabilities become much more complicated due to new unstable solutions, both at oblique propagation angles and in directions parallel and quasi-parallel to the magnetic field.
By comparison with BEFI, these new instabilities are oscillatory (or periodic) in time, i.e. with $\omega_r \ne 0$, see the bottom panels in Fig.~\ref{f6}.
This property helps us to differentiate them, given that their growth rates become comparable (case 4.c) or even exceed (case 4.d) those of BEFI.
We should not forget that the two-stream aperiodic instability predicted in the absence of background population, is expected in this case as well. That seems to be identified in case 4.d at low angles and large wave-numbers, as the mode with highest (maximum) growing rates but with a very small frequency $\omega_r \to 0$.
For a better distinction, but also for a preliminary identification of the nature of the unstable modes, in Fig.~\ref{f7} we represent in detail the properties of the most unstable modes, associated with the maximum growth rates of different modes distinguished in case 4.c (as above in Figs.~\ref{f2} and \ref{f3}).
The maximum growth rate of BEFI (black lines) is obtained at $\theta = 63^{\rm o}$, and in this case is comparable with that of ES electron beaming instability (at $\theta = 0^{\rm o}$), indicated with purple lines.
At oblique angles, this time we can also identify two heat-flow (HF) instabilities with $\omega_r \ne 0$.
We know already the oblique HF found at intermediary oblique angles and larger wave-numbers, which is a whistler HF (WHF), with RH polarization and a maximum growth rate at $\theta = 55^{\rm o}$), indicated with green lines as in Fig.~\ref{f3} in Fig. 7.
A new unstable mode is predicted at very oblique angles (indicated with blue lines in Fig. 7), and combines a firehose HF (FHF) LH polarization, at low wave-numbers, with a WHF RH polarization, at large wave-numbers.
Even the profile of growth rates shows two humps, corresponding to two different modes.
The one obtained at large wave-numbers connects apparently to the WHF mode obtained at lower oblique angles, e.g., for all cases 4.b, 4.c and 4.d, but remains however distinct. For both cases 4.c and 4.d growth rates of these two modes remain lower than BEFI.
The influence of these RH-polarized modes extends to low angles of propagation, becoming visible at large enough wave-numbers and explaining the major magnetic field component $B_y$ obtained already in case 1.d for the ES mode (purple lines). Note, however, that in case 4.d, by far the highest growth rate is that of the aperiodic two-stream instability propagating in parallel direction, a robust and highly competitive instability of two symmetric highly dense counter-beams, as described by \cite{Lopez-etal-2020b}.
For the oblique instabilities, the wave-number dispersion of the electric and magnetic field components shows similar profiles, and all resemble those obtained in Fig.~\ref{f3}.
By comparison to case 1.d, BEFI has already a significant $E_L$ at low wave-numbers, where $E_L \sim E_T$.
However, the fastest growing mode, with maximum growth rate, has the same hybrid nature with a major $E_L >> E_T$, and a major $B_y$.
For $\theta = 0^{\rm o}$ we again find a purely electromagnetic FHF (orange lines) with LH polarization and a growth rate much lower than that of BEFI.
\subsection{Maximum growth rates (thresholds) of BEFI}
Linear theory can also offer a more comprehensive image of BEFI, if we compute the maximum growth rates and build maps of their contour levels as a function of the main plasma parameters, in this case, the (normalized) beam speed $U_b/c$, i.e., the main source of free energy, but also the plasma beta parameter, e.g., $\beta_b$ for the beam.
Fig.~\ref{f8} displays the normalized maximum growth rates $\gamma_{\rm max}/\Omega_e$, coded according to the color bar on the right side and obtained for BEFI for the situations specific to case 1, when the beam and background electrons have the same thermal spread $\alpha_b = \alpha_0$, and three different density ratios $n_b/n_e=0.15$ (left), $n_b/n_e=0.25$ (center), and $n_b/n_e=0.35$ (right).
The main features of BEFI are already known, i.e., for the same $\beta_b$ the growth rates are significantly enhanced with increasing the beam/drift speed.
Contour lines (black or white) can be fitted to various mathematical expressions of $U_b/c$ as a function of $\beta_b$, see e.g., in \cite{Shaaban-etal-2018-pop} and \cite{Moya-etal-2022}, to quantify the beam speed thresholds of this instability, though here we limit to a qualitative analysis.
The effect of background electrons becomes also obvious if we compare with the maximum growth rates of BEFI obtained in \cite{Lopez-etal-2020b} for $n_0/n_e = 0$, which are markedly higher than those derived here, e.g., in the right panel for the same beam/drift velocity and same plasma beta.
Moreover, the three panels in Fig.~\ref{f8} show a uniform effect of the background population, which tends to suppress the instability, markedly inhibiting (from right to left) the maximum growth rates, and increasing the instability thresholds, see, e.g., contours at $\gamma_{\rm max}/\Omega_e =$ 0.06 and 0.1 in the middle and right panels.
In the middle panel, for a relative density of the beam $n_b/n_e = 0.25$, and $n_0/n_e = 0.5$ for the background, one can observe that BEFI can still be triggered with a reasonable maximum growth rate $\gamma_{\rm max}/\Omega_e \simeq [0.10-0.15]$, if the beta parameter and beam speed are sufficiently high \footnote{Higher values of $U_b$ must be considered with caution to not exceed the non-relativistic limit of our approach $U_b/c \lesssim 0.4$ (where $c$ is the speed of light in vacuum), above which the electrons with energy $E > 100$ keV are weakly relativistic.}, respectively, $\beta_b > 0.2$ and $U_b/c > 0.065$.
If the background electrons are dominant, e.g., with a relative density $n_0/n_e = 0.7$, e.g., in the left panel, BEFI can be barely excited, with very low growth rates $\gamma_{\rm max}/\Omega_e \simeq 10^{-3}$ approaching and describing the plasma conditions of marginal stability ($\gamma \to 0$) against BEFI.
With decreasing the presence of background electrons, the beam speed characteristic to marginal stability is also markedly lowered, as already found for the instability thresholds.
On the other hand, in the middle and right panels, above the contour level around $\gamma_{\rm max}/\Omega_e \simeq 0.3$ we can identify the regime of ESI, whose maximum growth rates become much superior to BEFI.
The shape of these thresholds is very similar to the one obtained for the thresholds of firehose heat-flux instability (FHFI) induced in the direction parallel to the magnetic field by a single (asymmetric) strahl/beam in the solar wind \citep{Shaaban2018, Shaaban-etal-2018-pop}.
By virtue of these properties, we can treat BEFI as an instability triggered by a double heat-flux. But more than that, BEFI is from the category of the oblique heat-flux instabilities, that propagate/develop at highly oblique direction with respect to the magnetic field, as the oblique whistler heat-flux instability \citep{Verscharen-etal-2019,Lopez-etal-2020a}.
By contrast with the parallel heat-flux instabilities, the oblique ones can effectively contribute to the relaxation of the electron beams, through an efficient resonant scattering of beaming electrons (and do not require that the electrons and waves counter-propagate), as shown not only in numerical simulations \citep{Micera-etal-2020,Vo-Cattell-2022} but also in a series of recent observations \citep{Cattell-etal-2020}.
Therefore, we expect that BEFI-like instabilities play an effective role in the relaxation of double electron strahls/beams, those counterbeams with a sufficiently high thermal spread, as predicted by their linear proprieties discussed in this section.
This could be the case of electron counterbeams observed in CIRs, but also in the interplanetary shocks and CME foreshocks at sufficiently large heliocentric distances (e.g., 1~AU and beyond).
We do mention, however, that the observed electron counterbeams are not necessarily symmetrical, in which case the oblique instability can change its properties, becoming periodic ($\omega_r \ne 0$) and possibly whistler-like in nature.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=0.45\textwidth]{energy}
\caption{\label{f9} Temporal evolution of the magnetic (top) and electric (bottom) field energy, for cases 1.c, 1.d, 3.d and 4.d.}
\end{center}
\end{figure}
\begin{figure*}[h!]
\begin{center}
\includegraphics[width=0.7\textwidth]{kxky_1c}
\includegraphics[width=0.7\textwidth]{kxky_1d}
\includegraphics[width=0.7\textwidth]{kxky_3d}
\includegraphics[width=0.7\textwidth]{kxky_4d}
\caption{\label{f10} Time snapshots of the FFT (normalized) energy density computed for the out of plane (perpendicular) component of the magnetic field, for cases 1.c, 1.d, 3.d, and 4.d.}
\end{center}
\end{figure*}
\begin{figure*}[h!]
\begin{center}
\includegraphics[width=0.7\textwidth]{kxky_electric_4d}
\caption{\label{f11} Time snapshots of the FFT (normalized) energy density computed for the parallel electric field component in case 4.d, showing an early time fast growth and saturation of ES instability.}
\end{center}
\end{figure*}
\section{Particle-in-cell simulations}\label{pic}
In order to validate the predictions from linear theory examined in section \ref{befi}, here we present results from simulations which also describe the evolution of BEFI in time.
We have used a 2D explicit PIC code based on the KEMPO1 code from \cite{Matsumoto1993}.
Our simulation domain is composed by $512\times512$ grid cells, with $L_x=L_y=153.6\,c\omega_{pe}$ and 625 particles per grid per species. The mass ratio is $m_p/m_e=1836$, the plasma to gyro-frequency $\omega_{pe}/\Omega_{ce}=20$, the time step is $\Delta t=0.01/\omega_{pe}$ and the simulation runs until $t_\text{max}=2000/\omega_{pe}$.
We chose to present the simulation results for four cases, 1.c, 1.d, 3.d and 4.d, which confirm the excitation of BEFI for different plasma conditions, but also allow us to compare the wave fluctuations triggered by different initial conditions.
That is either for different relative densities of the electron beams, if we compare cases 1.c and 1.d, or for different thermal speeds of the electron populations, contrasting cases 1.d and 3.d, or 1.c and 4.d.
Fig.~\ref{f9} displays the evolution in time of the fluctuating magnetic field energy density $W_B = \int \delta B^2/B_0^2\, dx dy$ and the electric energy density $W_E = \int \delta E^2/B_0^2\, dx dy$, for the time interval of the simulations.
From the figure we can see that cases 1.c, 1.d and 3.d (all cases initially satisfying $U_b \lesssim \alpha_b$) are qualitatively similar.
In all three cases, in agreement with linear theory predictions, the fastest developing BEFI (i.e., with maximum growth rate) has a hybrid nature, with an electric field component (mainly contributing to $W_E$) that grows at the beginning faster than the electromagnetic (EM) transverse component ($W_B$).
However, as time advances the EM energy density $W_B$ arises and reaches levels of about one order of magnitude larger than the electric energy density $W_E$.
For cases 3.d and 4.d the increasing slopes of $W_B$ are indeed higher than case 1.d, as predicted by the maximum growth rates ($\gamma_{\rm max}$) obtained from linear theory \footnote{A direct correspondence of these slopes with $2\gamma_{\rm max}$ cannot be done because not only the mode with maximum growth rate develops with the increase of time.}.
The growth of BEFI fluctuations in time is confirmed in Fig.~\ref{f10} by the FFT spectra of the normalized energy density $|{\rm FFT}(\delta B/B_0)|^2$, computed for the out of plane (perpendicular) component of the fluctuating magnetic field. The levels of fluctuations are coded in the right-hand color bars.
Displayed are four time snapshots up to (or near) the saturation, for the same cases 1.c, 1.d, 3.d and 4.d, from top to bottom, respectively. The 2D dispersion at large propagation angles in the wave-vector space ($k_\parallel, k_\perp$) resemble those from linear theory, especially at early moments in time, when BEFI fluctuations do not yet reach very high amplitudes (intensities) to be affected by the nonlinear decays.
Additional spots that are visible later in time at different propagation angles, may indeed signify fluctuations of daughter waves generated nonlinearly via three- or four-waves nonlinear decays.
These results are very similar to Fig.~5 in \cite{Lopez-etal-2020b}, obtained for BEFI in the case with no electron background.
However, BEFI fluctuations are visibly inhibited by the presence of background electrons.
In this sense, the contrast between the levels of fluctuations in Figure 10 is also very relevant, such as those obtained for the same time snapshots in cases 1.c and 1.d.
In case 4.d (with $U_b > \alpha_b$), our BEFI is predicted by linear theory in close competition with the electrostatic (ES) instabilities.
(See also the results presented in Figs. 9, 10 and 11 in \citet{Lopez-etal-2020b}, where the initial conditions also considered beam speeds higher than thermal speeds, but in the absence of an electron background.)
Indeed, the green dotted-line in Fig.~\ref{f9} shows a quick increase and relaxation of the fluctuating electric energy density $W_E$, with a narrow and not very high peak, followed by a drop and then by a more robust growth of the magnetic energy density $W_B$ due to BEFI.
In this case primary excited is the ES instability, at much lower time scales, as already indicated in Fig.~\ref{f9}. BEFI develops as a secondary but more robust instability, and it is also confirmed in Fig.~\ref{f10} last row, for the same time scales of BEFI in cases 1.c, 1.d and 3.d.
However, for case 4.d, the oblique maxima of BEFI are more disperse or less compact, most probably, due to linear or nonlinear interactions with fluctuations of other nature.
The earlier ES excitations propagating at small angles with respect to the magnetic field are confirmed in Fig.~\ref{f11}, where we display earlier time snapshots of the FFT (normalized) energy density for the parallel electric field component in case 4.d.
The levels of fluctuations are color coded in the right-hand bars, and reach a maximum (saturation) at about $\omega_{pe} t = 61.44$, much earlier than the first time snapshot shown in Fig.~\ref{f10}.
\section{Conclusions} \label{conc}
Since space plasmas are weakly collisional (or even non-collisional), we expect wave instabilities to have multiple implications, especially by facilitating the conversion of free energy of plasma particles, as well as energy transfer between species.
\cite{Lopez-etal-2020b} and \cite{Moya-etal-2022} have recently shown that two symmetric electron counter-beams, aligned to the guiding magnetic field, can induce an electromagnetic (EM) firehose-like instability, aperiodic and propagating highly obliquely to the magnetic field.
In the present work we investigated this instability under typical conditions found in the heliosphere, calling it the beaming electron firehose instability (BEFI). Thus, we assumed a specific parameterization of the plasma system, including a background embedding plasma of electrons and ions (protons). Counter-beaming electrons penetrating the background solar wind are often reported by in-situ observations, in various contexts such as interplanetary shocks, corotating interaction regions (CIRs), and closed magnetic field topology specific to coronal mass ejections (CMEs).
We relied on such observations to define the plasma model introduced in section 2, and to identify the conditions found favorable to BEFI, see parametric cases in Table~\ref{t1}.
In section~\ref{befi} we described the linear spectra of unstable waves for the selected cases in Table~\ref{t1}, varying the relative densities and thermal speed of the electron components.
Particularly relevant for BEFI, are the regimes identified in Figs.~\ref{f1}-\ref{f5}, for cases 1.a-1.d, 2.a - 2.d and 3.a-3.c, when BEFI is either solely predicted, or has (maximum) growth rates much higher than all the other instabilities in the spectra.
The influence of background population can be quantified in terms of relative density and thermal spread.
For the cases studied here, BEFI growth rates are significantly reduced if relative beam densities are less than 20\% of the total density (implying background electrons with relative density exceeding 80\%), making the existence of this instability critical.
For a slightly cooler background population, compare for instance cases 1 with cases 2, the range of unstable wave-numbers increase.
Similar effect is obtained in cases 3 for a slightly cooler beam.
However, for even lower thermal speeds or higher densities of the beams, e.g., in cases 4, the (maximum) growth rates become dominated by the electrostatic (ES) instabilities at lower angles of propagation, as already shown in \cite{Lopez-etal-2020b} and \cite{Moya-etal-2022}.
Linear properties of dispersion and stability, including the instability thresholds led us to the conclusion that BEFI is analogous to heat-flux instabilities generated by unidirectional electron strahls/beams in the solar wind.
BEFI is however triggered by a double heat-flux, that of the counter-beams (or double strahl) of electrons, but for sufficiently low beaming speeds (or associated heat fluxes), in the range of thermal speed of electron beams.
However, in the present analysis with two electron counter-beams and background populations, the configuration of linear spectra of unstable modes becomes much more complicated.
Besides the ES instabilities (for higher beaming speeds) we have also identified periodic instabilities (with $\omega_r \ne 0$) that do not appear in the absence of the background electron population, when only symmetric counter-beams are present \citep{Lopez-etal-2020b}.
These unstable wave modes are specific to asymmetric electron beam-plasma configurations, which here result from the combination of each electron beam with the background population.
More details can be found in a recent parametric analysis of electron heat-flux instabilities in the solar wind conditions \citep{Lopez-etal-2020a}.
Future works should also investigate more complex plasma systems with asymmetric counter-beams embedded by background electrons, for which we expect BEFI to become a periodic mode as well, see, for instance, the case in Fig.~7 in \cite{Lopez-etal-2020b}.
In such a case, BEFI will probably blend more easily with other modes and make them difficult to distinguish.
From the analogy with the heat-flux instabilities, BEFI compares better with the oblique whistlers, which can contribute to the scattering and relaxation of unidirectional strahls in the solar wind \citep{Micera-etal-2020,Cattell-etal-2020}.
Particle-in-cell (PIC) simulations confirm the results of the linear kinetic theory (section~4), not only for the conditions in which the BEFI is predicted as primary excitation, with major growth rates, but also when it develops as a secondary instability.
Tested in the PIC simulations were those cases associated with high growth rates, in order to reduce the computational time and obtain results of increased confidence.
The BEFI fluctuations develop (aperiodically) at highly oblique propagation angles to the magnetic field, in agreement with the wavenumber and angular dispersion of the (initial) linear growth rates.
Moreover, levels reached by these fluctuations are diminished with increasing the presence of background electrons, also contrasting to the results in \cite{Lopez-etal-2020b} obtained in the absence of of background electrons.
In the regimes of competition with ES instabilities, BEFI still develops as a secondary but sufficiently robust instability to produce intense EM fluctuations, long lasting in time up to their saturation.
Therefore, we can expect BEFI to be involved in the regulation of electron counter-beams with properties similar to those investigated here.
Our present results should motivate future theoretical and observational studies, to model the evolution of such double electron strahls/beams under the consistent action of BEFI-like fluctuations, and to compare with in-situ observations in space plasmas.
\begin{acknowledgements}
The authors acknowledge support from the Ruhr-University Bochum and the Katholieke Universiteit Leuven, and Mansoura University. These results were also obtained in the framework of the projects C14/19/089 (C1 project Internal Funds KU Leuven), G.0D07.19N (FWO-Vlaanderen), SIDC Data Exploitation (ESA Prodex-12), Belspo project B2/191/P1/SWiM, and Fondecyt No. 1191351 (ANID, Chile). P.S. Moya is grateful for the support of KU Leuven BOF Network Fellowship NF/19/001 and ANID Chile through FONDECyT grant No. 119135. R.A.L. acknowledges the support of ANID Chile through FONDECyT grant No. 11201048. Powered@NLHPC: This research was partially supported by the supercomputing infrastructure of the NLHPC (ECM-02). We thank the anonymous reviewer for a careful reading of our paper, as well as for the pertinent observations.
\end{acknowledgements}
|
\section{Introduction}
The next-generation wireless communication systems aim for increased data rates, reduced lower latency and large-scale access. The mmWave technology has shown significant promise in meeting these requirements \cite{hemadeh2017millimeter, 8424038, shokri2015millimeter}. Communication at mmWave frequencies has recently attracted significant attention due to the fact that the sub-6 GHz band faces acute spectrum shortage \cite{8565897, 8770151}. In this context, the under-utilized mmWave band has the potential of satisfying the requirements of next generation wireless networks. Furthermore, the abundance of free spectrum in the mmWave band is capable of supporting the massive connectivity of devices required by the Internet of Things (IoT) \cite{sahoo2018enabling}.
However, as shown via practical measurements in \cite{heath2016overview}, \cite{sohrabi2016hybrid}, mmWave networks suffer from signal blockage and penetration losses, which degrades the received signal strength.
The dense deployment of access points (APs) for creating small cells has shown significant promise for overcoming these challenges. However, inter-cell interference (ICI) becomes a significant bottleneck with the ever-increasing network densification. Several cooperative signal processing concepts, such as cooperative MIMO networks, network MIMO, and coordinated multipoint (CoMP) transmission \cite{gesbert2010multi, venkatesan2007network, irmer2011coordinated}, are presented in the literature to address the issue of ICI. Note that the network architectures mentioned above are conventionally implemented in a network-centric fashion by dividing the APs into disjoint clusters. In \cite{ngo2017cell, nayebi2017precoding, ammar2021user}, a user-centric cell-free MIMO structure relying on a large number of distributed APs is considered, which are connected to a central processing unit (CPU). The network jointly serves multiple users in conjunction with no cell boundaries, similar to the cooperative MIMO structure. However, cooperative MIMO networks mitigate the ICI by grouping multiple cells into fixed cooperative clusters, which shifts the interference management from the cell level to the cluster level. The resultant inter-cluster interference imposes a fundamental limit on cooperation. Hence, users close to the cluster edge might not benefit from the cooperative MIMO clustering. By contrast, cell-free MIMO facilitates improved service for cell-edge users due to the elimination of the inter-cluster interference. Furthermore, the other significant aspect of cell-free MIMO is the deployment of a higher number of APs than that of the users in comparison to the cooperative MIMO. Thus, cell-free networks promise a further improvement in spectral efficiency by mitigating the interference through intelligent, cooperative beamforming.
Apart from the benefit of wider bandwidth availability, the short wavelength of mmWave frequencies allows us to pack large antenna array into a compact space for supporting highly directional beamforming, while simultaneously serving numerous users. Thus, the multiple-input-multiple-output (MIMO) technology can be exploited to obtain high beamforming gains for mitigating the effects of path loss, atmospheric absorption and penetration loss in the mmWave regime \cite{rangan2014millimeter}.
However, the high power consumption and associated cost/complexity of assigning a dedicated per-antenna RF chain prohibits the employment of the conventional fully-digital (FD) transceiver architecture in mmWave MIMO systems \cite{heath2016overview}.
This has motivated the conception of the novel hybrid MIMO transceiver architecture as a promising solution, which employs much fewer RF chains in comparison to the antenna elements, thus easing its practical implementation. The overall MIMO signal processing in such a hybrid architecture is divided into the RF and baseband domains, where RF transmit precoding (TPC) yields a directional beamforming gain, whereas spatial multiplexing and interference cancellation are achieved via suitable baseband TPC. It can be readily observed that the design of the baseband and RF precoders and combiners is the key to address the various challenges associated with the proposed cooperative cell-free multi-user (CFMU) mmWave MIMO system, and also to harness the significant gains promised by the mmWave band for 5G/B5G.
\subsection{Related Works}
In the recent on mmWave literature, sophisticated techniques have been proposed for designing hybrid TPCS and receive combiners (RCs) for mmWave MIMO systems \cite{el2014spatially, wang2019low, noh2016zero, alkhateeb2015limited}. In particular, the early contributions focused more on single-user (SU) scenarios.
Omar \textit{et al.,} in \cite{el2014spatially} considered SU mmWave hybrid MIMO systems and derived the hybrid transmit beamformers (HTBF) in which the sparse scattering nature of mmWave wireless channel has been exploited to obtain a two-stage solution via the popular orthogonal matching pursuit (OMP) algorithm. Wang \textit{et al.,} \cite{wang2019low}, proposed another hybrid TPC design that decomposes the original HTBF design problem into several sub-problems to render the optimization tractable.
In order to further improve the spectral efficiency and to mitigate the inter-user interference (IUI) in practical mmWave cellular systems a multi-user (MU) mmWave MIMO system has also been considered in the pioneering works \cite{noh2016zero}, \cite{alkhateeb2015limited}, wherein a BS simultaneously serves multiple users/devices.
In a mmWave MU MIMO system, the digital component of the hybrid architecture provides higher degrees of freedom than analog-only beamforming, and can hence be employed for mitigating the interference among the different users. Note that the design of HTBF is a key challenge in MU mmWave MIMO systems, since the resultant IUI degrades the overall system performance \cite{ding2015performance}, \cite{zhang2015downlink}. Therefore, there is an urgent need for minimizing the IUI, while the HTBF gain has to be simultaneously maximized for each user. Thus, it is of pivotal importance to design efficient HTBF for mmWave MU MIMO systems that have a low complexity. Hence, sophisticated HTBF techniques have been developed for optimizing their performance. In \cite{alkhateeb2015limited}, a low-complexity iterative algorithm was proposed, which exploited the sparse scattering nature of the mmWave wireless channel for designing the HTBF of the BS and analog RC of each user. The proposed solution was shown to be asymptotically optimal for the large-scale antenna regime, which also demonstrates the well-known fact that MU MIMO systems attain a high capacity than SU MIMO systems.
In \cite{hong2013joint, zhang2010cooperative, li2015joint, cheng2013joint, yang2017opportunistic}, the coordination of different base stations (BSs) has been proposed for conventional sub-6 GHz systems for improving their performance. However, the proposed schemes targeted exclusively fully-digital sub-6 GHz MU MIMO systems. Hence, they are unsuitable for mmWave MU MIMO cellular systems.
Michaloliakos \textit{et al.} \cite{michaloliakos2016joint}, proposed a novel scheme relying on jointly selecting the analog TPCS for maximizing the data rate of the users by harnessing a set of predefined beam patterns. Zhu \textit{et al.} \cite{zhu2017hybrid}, proposed a novel cooperative HTBF design for multi-cell multi-user (MCMU) scenarios based on the Kronecker decomposition. First, analog transmit beamformer (TBF) is obtained by decomposing the beamformer as well as the data and interference
path-vectors into the corresponding Kronecker products of unit-modulus phase-shift vectors. Next, the baseband TBF is designed by using the minimum mean square error (MMSE) TBF technique. Shu \textit{et al.} \cite{sun2018analytical}, presented both signal-to-leakage-plus-noise-ratio (SLNR) and regularized zero-forcing based HTBF techniques for multi-stream MCMU systems to study the benefits of interference coordination in terms of the spectral efficiency.
Bai \textit{et al.} \cite{bai2018cooperative}, presented a two-step hybrid cooperative TPC design for a distributed antenna scenario combined with user selection for energy-efficient transmissions. While their work has immense utility, it is restricted to a single-cell scenario, which does not exploit the additional advantages offered by multi-cell cooperation. Fang \textit{et al.} \cite{fang2021hybrid}, maximize the sum-power consumption of the APs supporting single-antenna users relying on cooperative HTBF.
While several authors have addressed the problem of hybrid beamformer design for classic cellular systems, the research of their cell-free counterparts is limited. In \cite{kim2019joint}, the authors proposed a hybrid beamformer scheme, where the analog and digital beamformers are alternately optimized based on the weighted minimum mean squared error (WMMSE) criterion. In \cite{hou2017joint}, the authors proposed a two-step procedure for their hybrid beamformer design. In the first step, the analog TPC is obtained from a codebook, followed by semidefinite relaxation (SDR) and convex approximation-aided baseband TPC design. In \cite{femenias2019cell}, the authors designed a hybrid beamforming scheme for CFMU massive MIMO systems toward maximizing the user rate. Chenghao \textit{et. al.} \cite{feng2021weighted} proposed a block coordinated descent (BCD) algorithm based hybrid transmit beamformer (HTBF) maximizing the weighted sum-rate of cell-free mmWave MIMO systems. However, the aforementioned treatises assume the availability of either perfect knowledge of CSI or that of its second-order statistics. The HTBF design of a cell-free network requires the full knowledge of AoAs/AoDs of the multipath components and their complex-valued path gains. However, in practice, acquiring accurate estimates of these quantities is challenging. In the recent studies of cell-free MIMO systems, sophisticated techniques have been proposed for channel estimation in CFMU mmWave MIMO systems \cite{jin2019channel, souza2021effective, guo2021joint, song2021joint, wang2022two}. Furthermore, the channel matrix in the mmWave regime is sparse, which can be jointly attributed to the reduced scattering and diffraction effects as well as to the focused beam emanating from the large antenna arrays. Exploiting this sparsity can lead to a significant reduction in the number of pilot transmissions required and also a substantial improvement in the estimation accuracy. In this context, the compressed sensing (CS) techniques proposed in \cite{gilbert2005applications, barbotin2012estimation} utilize a sparse signal recovery model for channel estimation. The authors of \cite{huang2018iterative, srivastava2021sparse} have demonstrated the applicability of some classical CS based schemes for sparse channel estimation at a reduced pilot overhead. One can readily employ these existing schemes for efficient sparse channel estimation in CFMU mmWave MIMO networks.
In wireless communication systems, there are several scenarios, where common data has to be sent to a group of users, for instance, during the transmission of content pertaining to a live event. In such scenarios, multicast is an efficient framework for improving both the spectral and power efficiency by simultaneously transmitting the common data to a group of users \cite{dartmann2013low, guo2018multi, dong2020multi}. This has huge potential in numerous applications, such as mobile TV, live video streaming, video conferencing etc.
As a further advance in \cite{dartmann2013low}, a low-complexity uplink-downlink duality based suboptimal algorithm is presented to jointly design the beamformers for multiple users in a single-cell wireless multicast network.
Chengjun \textit{et al.} \cite{guo2018multi} conceived a low-complexity multi-quality multicast beamforming schemes..
The authors of \cite{dong2020multi} proposed a weighted MMSE based multicast beamformer employing the popular successive convex approximation and Lagrangian duality based techniques.
Note that all the above contributions only consider conventional sub-6 GHz fully digital transceiver architectures, which cannot be used in mmWave CFMU MIMO systems. Multicast transmission, specifically designed for mmWave MIMO networks, was studied in \cite{choi2015iterative,dai2015hybrid,wang2018hybrid}.
In \cite{choi2015iterative}, a single group scenario is considered for HTBF-aided multicast transmission, wherein an iterative algorithm was designed based on the principle of alternating minimization. In the seminal work \cite{dai2015hybrid}, the author presented a single-cell single-group HTBF design relying on max-min fairness for multicasting using a limited number of RF chains. In \cite{wang2018hybrid}, a suboptimal low-complexity HTBF design was proposed for the downlink of unicast and multicast mmWave MIMO systems. However, all these papers only considered a single AP scenario for mutlicast transmission. To the best of our knowledge, none of the works in the existing literature have sophistically addressed the problem of cooperative HTBF design in CFMU mmWave MIMO systems. Thus, motivated by these limitations of the existing literature reviewed above, this paper presents a novel cooperative beamformer design conceived for efficiently mitigating the interference in CFMU mmWave MIMO networks. Table \ref{tab_lit_rev} boldly and explicitly contrasts to the literature discussed above. Next, we further detail the various contributions of our work.
\begin{table*}
\centering
\caption{Contrasting our contribution to the existing literature} \label{tab_lit_rev}
\begin{tabular}{|l|c|c|c|c|c|c|c|c|c|c|c|c|c|}
\hline
& \cite{el2014spatially} & \cite{wang2019low} & \cite{alkhateeb2015limited} & \cite{hong2013joint} & \cite{zhu2017hybrid} & \cite{sun2018analytical} & \cite{femenias2019cell} & \cite{feng2021weighted} & \cite{dartmann2013low} & \cite{choi2015iterative} & \cite{dai2015hybrid} & \cite{wang2018hybrid} & Proposed Work \\ [0.5ex]
\hline
mmWave communication & \checkmark & \checkmark & \checkmark & & \checkmark & \checkmark & \checkmark & \checkmark & & \checkmark & \checkmark & \checkmark & \checkmark\\
\hline
Hybrid architecture & \checkmark & \checkmark & \checkmark & & \checkmark & \checkmark & \checkmark & \checkmark & & \checkmark & \checkmark & \checkmark & \checkmark\\
\hline
Multi-user & & & \checkmark & \checkmark & \checkmark & \checkmark & \checkmark & \checkmark & \checkmark & \checkmark & \checkmark & \checkmark & \checkmark \\
\hline
Downlink & \checkmark & \checkmark & \checkmark & \checkmark & & \checkmark & \checkmark & \checkmark & \checkmark & \checkmark & \checkmark & \checkmark & \checkmark \\
\hline
Uplink & & & & & \checkmark & & & & \checkmark & & & & \checkmark \\
\hline
Cooperative beamforming & & & & \checkmark & \checkmark & \checkmark & \checkmark & \checkmark & \checkmark & & & & \checkmark\\
\hline
Multicast single-group & & & & & & & & & \checkmark & \checkmark & \checkmark & \checkmark & \checkmark\\
\hline
Multicast multi-group & & & & & & & & & \checkmark & & & & \checkmark\\
\hline
Broadcast & & & & & & & & & & \checkmark & & \checkmark & \checkmark\\
\hline
Unicast & & & & \checkmark & \checkmark & \checkmark & \checkmark & \checkmark & \checkmark & & & \checkmark & \checkmark\\
\hline
\end{tabular}
\end{table*}
\subsection{Contributions}
Again, we design HTBFs and RCs for broadcast, unicast and multicast scenarios in cooperative CFMU mmWave MIMO cellular systems.
\begin{itemize}
\item Initially, we design an optimal hybrid broadcast beamformer for CFMU mmWave MIMO systems. Furthermore, the optimal TBFs having per-AP power constraints are also derived for this scenario. At the users, a low-complexity maximal ratio combiner (MRC) is used in the baseband for maximizing the effective SNR.
\item Next, the design and analysis is extended to both unicast and multicast transmission scenarios. Toward this, the minimum variance distortionless response (MVDR) \cite{johnson1992array}, \cite{trees2002optimum} based optimal successive minimum variance hybrid beamforming (OSHB) schemes are designed for mitigating both the inter-user and inter-group interference. It has been demonstrated theoretically that the proposed schemes can support a higher number of users than popular block diagonalization (BD) method.
\item Subsequently, a Bayesian learning (BL)-based method is developed for jointly designing the RF and baseband TPCs/RCs, which is of significant utility, when perfect knowledge of the angles of arrival/departure (AoAs/AoDs) and path gains of the multipath components is not available.
\item Furthermore, in order to address the cooperative beamformer design problem of the uplink, a successive CFMU uplink hybrid beamforming (SCUHBF) scheme is developed that maximizes the SINR. It is shown that the proposed scheme achieves an improved data rate as a benefit of interference cancellation.
\item Simulation results are provided for characterizing the OSHB and SCUHBF schemes in cancelling the interuser/intergroup interference together with the improved spectral efficiency compared to a no-coordination scenario, the performance attained is close to that of the ideal fully-digital design.
\end{itemize}
\subsection{Organization of the paper}
The remainder of the paper is arranged as follows. Section \ref{sec 2} introduces the system model and channel model. In Section \ref{sec HBB}, we formulate the problem of HBBF for CFMU mmWave MIMO systems. Section IV and V describe the problem of cooperative TPC design of unicast and multicast scenarios, respectively, and develop the OSHB schemes for it. The SCUHBF scheme of an uplink scenario is derived in Section VI, followed by the BL-based joint precoder design in Section-VII. Finally, Section VIII presents our simulation results for illustrating the merits of the proposed work, followed by our conclusions in Section IX.
\subsection{Notation}
Vectors and matrices are denoted by boldfaced lower and upper case letters respectively, while scalars are represented by lower case letters. The transpose, Hermitian, conjugate, and inverse of a matrix $\mathbf{F}$ are denoted by $\mathbf{F}^T$, $\mathbf{F}^H$, $\mathbf{F}^*$, and $\mathbf{F}^{-1}$, respectively. Furthermore, $\text{Tr}(\mathbf{A})$ denotes the trace of a matrix $\mathbf{A}$, $\norm{\mathbf{x}}_2$ represents the $l_2$-norm of a vector $\mathbf{x}$, and $\mathop{\mathbb{E}}[\cdot]$ is the expectation operator. Furthermore, ${\mathbf{x}} \sim \mathcal{CN} (\mathbf{b}, \mathbf{A})$ denotes a complex Gaussian random vector having mean $\mathbf{b}$ and covariance matrix $\mathbf{A}$. Finally, $\mathbf{I}_N$ represents an $N \times N$ identity matrix and $\mathbf{A} \succcurlyeq 0$ indicates that $\mathbf{A}$ is a positive semi-definite (PSD) matrix.
\section{mmWave CFMU System and Channel Model}\label{sec 2}
Consider a downlink cell-free mmWave hybrid MIMO system with $M$ cooperative APs, where the $m$th AP has $N_T$ transmit antennas (TAs) and $N_{\mathrm{RF},m}$ RF chains so that $1 \leq N_{\mathrm{RF},m} << N_T$, $\forall m, 1 \leq m \leq M$. These $M$ APs are connected to a central processing unit (CPU) and serve $G$ user-groups, where the $g$th group is comprised of $U_g$ users, as shown in Fig. \ref{multicell}. Each user is equipped with $N_R$ receive antennas (RAs) and $N_{\mathrm{RF},u}$ RF chains, which satisfy $1 \leq u \leq U_g$, $\forall 1 \leq g \leq G$. We consider a next-generation multicast scenario, where several users request identical services relying on the same data. Such a group of users is referred to as a multicast group in the literature, and many such groups may coexist as shown in Fig. \ref{multicell}. For convenience, user $u$ in the $g$th multicast group is denoted as $U_u^{(g)}$. The signal, $\mathbf{r}^{(g)}_u \in \mathbb{C}^{N_{R} \times 1}$, received at the user $u$ in the $g$th multicast group can be formulated as
\begin{align}
\mathbf{r}^{(g)}_u =& \sum_{m=1}^{M}{\mathbf{H}^{(g)}_{u, m} \mathbf{F}_{\mathrm{RF},m} \mathbf{f}^{(g)}_{\mathrm{BB},m} x^{(g)}} \nonumber \\ +& \sum_{\substack{l=1\\ l \neq g}}^{G} \sum_{m=1}^{M}{\mathbf{H}^{(g)}_{u, m} \mathbf{F}_{\mathrm{RF},m} \mathbf{f}^{(l)}_{\mathrm{BB},m} x^{(l)}} + {\boldsymbol{\delta}}^{(g)}_{u},
\end{align}
where $\mathbf{H}^{(g)}_{u, m} \in \mathbb{C}^{N_{R} \times N_{T}}$ represent the mmWave MIMO channel matrix from the $m$th AP to the user $U_{u}^{(g)}$, $\mathbf{F}_{\mathrm{RF},m} \in \mathbb{C}^{N_{T} \times N_{\mathrm{RF},m}}$ is the RF TPC of the $m$th AP, and $\mathbf{f}^{(g)}_{\mathrm{BB}, m} \in \mathbb{C}^{N_{\mathrm{RF},m} \times 1}$ is the baseband TPC of the $m$th AP for the user-group $g$. Furthermore, ${\boldsymbol{\delta}}^{(g)}_u \in \mathbb{C}^{N_{R} \times 1}$ represents the complex additive Gaussian noise vector with distribution ${\boldsymbol{\delta}}^{(g)} \sim \mathcal{CN} (\mbox{0}, \sigma_{\delta}^2 \mathbf{I}_{N_R})$ and $x^{(g)}$ signifies the transmitted symbol that has unit power i.e., $\mathop{\mathbb{E}}\left[{|x^{(g)}|}^2\right] = \mbox{1}$. The received signal $\tilde{\mathbf{r}}_{u}^{(g)}\in\mathbb{C}^{N_{\mathrm{RF},u}\times 1} $ after RF combining at the user $u$ in the $g$th group is given by
\begin{align*}
\tilde{\mathbf{r}}_{u}^{(g)} =& \left(\mathbf{W}_{\mathrm{RF},u}^{(g)}\right)^H \sum_{m=1}^{M}{\mathbf{H}^{(g)}_{u, m} \mathbf{F}_{\mathrm{RF},m} \mathbf{f}^{(g)}_{\mathrm{BB},m} x^{(g)}} \nonumber \\ +& \left(\mathbf{W}_{\mathrm{RF},u}^{(g)}\right)^H \sum_{\substack{l=1\\ l \neq g}}^{G} \sum_{m=1}^{M}{\mathbf{H}^{(g)}_{u, m} \mathbf{F}_{\mathrm{RF},m} \mathbf{f}^{(l)}_{\mathrm{BB},m} x^{(l)}} + \tilde{{\boldsymbol{\delta}}}^{(g)}_{u},
\end{align*}
where $\mathbf{W}_{\mathrm{RF},u}^{(g)} \in \mathbb{C}^{N_{R} \times N_{\mathrm{RF}, u}}$ denotes the RF RC of the user $U_{u}^{(g)}$ and $ \tilde{{\boldsymbol{\delta}}}^{(g)}_{u} = \left(\mathbf{W}_{\mathrm{RF},u}^{(g)}\right)^H{\boldsymbol{\delta}}^{(g)}_{u}$. Let us define the effective channel matrix between the user $U_{u}^{(g)}$ and $m$th AP as
$\mathbf{H}^{(g)}_{\mathrm{eff},u,m} = \left(\mathbf{W}_{\mathrm{RF},u}^{(g)}\right)^H\mathbf{H}^{(g)}_{u, m}\mathbf{F}_{\mathrm{RF},m} \in \mathbb{C}^{N_{\mathrm{RF},u} \times N_{\mathrm{RF},m}}$, and also define the concatenated baseband TPC $\mathbf{f}_{\mathrm{BB}}^{(g)}$ of the $g$th group as $\mathbf{f}_{\mathrm{BB}}^{(g)} = \left[\left(\mathbf{f}_{\mathrm{BB},1}^{(g)}\right)^H, \left(\mathbf{f}_{\mathrm{BB},2}^{(g)}\right)^H, \ldots, \left(\mathbf{f}_{\mathrm{BB},M}^{(g)}\right)^H \right]^H$. The concatenated effective channel matrix between the user $U_{u}^{(g)}$ and all the cooperating APs is denoted as $\mathbf{H}^{(g)}_{\mathrm{eff},u} = [\mathbf{H}^{(g)}_{\mathrm{eff},u,1}, \mathbf{H}^{(g)}_{\mathrm{eff},u,2}, \ldots, \mathbf{H}^{(g)}_{\mathrm{eff},u,m}]\in \mathbb{C}^{N_{\mathrm{RF},u}\times \left(\sum_{m=1}^{M}N_{\mathrm{RF},m}\right)}$. Therefore, the effective received signal after baseband combining, $\hat{\mathbf{r}}_{u}^{(g)}$, at user $U_{u}^{(g)}$ can be written as
\begin{figure}
\begin{center}
\includegraphics[width = 5cm, height = 4cm]{Figure1.eps}
\caption{Cooperative beamforming in cell-free mmWave MIMO for mulitcast scenario. \vspace{-20pt}}
\label{multicell}
\end{center}
\end{figure}
\begin{align*}
\hat{\mathbf{r}}_{u}^{(g)} =&\, \left(\mathbf{w}_{\rm BB ,u}^{(g)}\right)^H\mathbf{H}^{(g)}_{\rm eff ,u}\mathbf{f}_{\rm BB}^{(g)} x^{(g)} \nonumber \\ +&\, \left(\mathbf{w}_{\rm BB ,u}^{(g)}\right)^H\sum_{\substack{l=1\\ l \neq m}}^{G}\mathbf{H}^{(g)}_{\rm eff ,k}\mathbf{f}_{\rm BB}^{(l)} x^{(l)} + \left(\mathbf{w}_{\rm BB ,u}^{(g)}\right)^H \tilde{{\boldsymbol{\delta}}}^{(g)}_{u},
\end{align*}
where $\mathbf{w}_{\rm BB ,u}^{(g)} \in \mathbb{C}^{N_{\rm RF, u} \times 1}$ denotes the baseband RC of the user $U_{u}^{(g)}$. The SINR of the user $U_{u}^{(g)}$ can be expressed as
\begin{align*}
\text{SINR}^{(g)}_u =&\, \frac{\left| \left(\mathbf{w}_{\rm BB ,u}^{(g)}\right)^H \mathbf{H}^{(g)}_{\rm eff ,u}\mathbf{f}_{\rm BB}^{(g)}\right|^2}{\sum_{l=1,l \neq m}^{G}{\left| \left(\mathbf{w}_{\rm BB ,u}^{(g)}\right)^H \mathbf{H}^{(g)}_{\rm eff ,u}\mathbf{f}_{\rm BB}^{(l)}\right|^2}+ \sigma_{\tilde{\delta}}^2}.
\end{align*}
Furthermore, the downlink capacity represented as $C_{\mathrm{DL}}$ is given by
\begin{align*}
C_{\mathrm{DL}}=&\, \sum_{g=1}^{G}\sum_{u=1}^{U_g} \log_2 \left(\mbox{1} + \text{SINR}^{(g)}_u\right).
\end{align*}
\subsection{mmWave MIMO channel model}
Using the geometric channel model described in \cite{el2014spatially}, \cite{rappaport2015millimeter} for the mmWave MIMO system, the channel matrix between the $m$th AP and $u$th user, denoted by $\mathbf{H}_{u,m} \in \mathbb{C}^{N_R \times N_T}$, can be expressed as
\begin{equation}
\mathbf{H}_{u,m}=\sqrt{\frac{N_{T}N_{R}}{L}}\sum_{l=1}^{L}\alpha_{u,m}^{(l)}\mathbf{a}_{R}(\phi_{l,u,m}^{R})\mathbf{a}_{T}^H(\phi_{l,u,m}^{T}),
\label{Channel}
\end{equation}
where $L$ represents the number of dominant multipath components between the $m$th AP and user $u$, $\alpha _{u,m}^l$ denotes the complex-valued path gain of the $l\text{th}$ multipath component. Furthermore, the terms $\mathbf{a}_{T}(\phi_{l}^{T})$ and $\mathbf{a}_{R}(\phi_{l}^{R})$ denote the transmit and receive array response vectors, respectively, with $\phi_{l}^{R}$ being the angle of arrival (AoA) and $\phi_{l}^{T}$ the angle of departure (AoD).
For uniformly spaced linear arrays (ULA), where $\mathbf{\lambda}$ and $d$ denote the signal wavelength and distance between the antenna elements, respectively, the array response vectors $\mathbf{a}_{R}(\phi_{l}^{R})$ and $\mathbf{a}_{T}(\phi_{l}^{T})$ are given by
\begin{align*}
& \mathbf{a}_{R}(\phi_{l}^{R}) \\ \nonumber &=\frac{1}{\sqrt{N_{R}}}[1,\ e^{j\frac{2\pi}{\lambda}d\sin(\phi_{l}^{R})},\ \ldots,\ e^{j\frac{2\pi}{\lambda}(N_{R}-1)d\sin(\phi_{l}^{R})}]^{T},
\end{align*}
\begin{align*}
& \mathbf{a}_{T}(\phi_{l}^{T}) \\ \nonumber &=\frac{1}{\sqrt{N_{T}}}[1,\ e^{j\frac{2\pi}{\lambda}d\sin(\phi_{l}^{T})},\ \ldots,\ e^{j\frac{2\pi}{\lambda}(N_{T}-1)d\sin(\phi_{l}^{t})}]^{T}.
\end{align*}
One can also represent this mmWave MIMO channel in a compact form as
\begin{equation*}
\mathbf{H}_{u,m} = \mathbf{A}_{R,u,m}\mathbf{H}_{u,m}^{\alpha}\mathbf{A}_{T,u,m}^{H},
\end{equation*}
where $\mathbf{A}_R = \left[\mathbf{a}_{R}(\phi_{1,u,m}^{R}),\dots,\mathbf{a}_{R}(\phi_{L,u,m}^{R})\right]\in \mathbb{C}^{N_R \times L}$ and $\mathbf{A}_T = \left[\mathbf{a}_{T}(\phi_{1,u,m}^{T}),\dots,\mathbf{a}_{T}(\phi_{L,u,m}^{T})\right]\in \mathbb{C}^{N_T \times L}$ contain the array response vectors of the AP and user, respectively, and $\mathbf{H}_{u,m}^{\alpha} = \mathrm{diag}\left(\alpha_{u,m}^{(1)},\alpha_{u,m}^{(2)},\dots,\alpha_{u,m}^{(L)}\right)$ denotes the diagonal matrix of path gains. Although this channel model is described for a uniform linear array (ULA), one can also readily extend the framework of this paper to uniform planar arrays (UPA). The algorithms and results derived in this work are also valid for a UPA, the corresponding array response vector can be written as
\begin{align}
\mathbf{a}_{x}\left(\phi,\theta\right)&= \frac{1}{\sqrt{N_x}}\bigg[1, e^{j\frac{2\pi}{\lambda}d\left(\sin\phi \sin\theta + \cos\theta \right)}, \ldots, \nonumber \\ & \ldots, e^{j\frac{2\pi}{\lambda}d\left(\left(N_{x}^{h}-1\right)\sin\phi \sin\theta + \left(N_{x}^{v}-1\right) \cos\theta \right)} \bigg],
\end{align}
where $x \in \lbrace R,T \rbrace$, while $N_{x}^{h}$ and $N_{x}^{v}$ denote the number of horizontal and vertical elements of the UPA in the two-dimension plane of the array. The UPA structure at each AP has a total of $N_T$ elements associated with $N_T = N_{T}^{h} \times N_{T}^{v}$.
Note that the system model developed in Section-II describes a general multicast scenario, where a group of users may request identical services. A special case of the multicast scenario is popularly termed as a unicast scenario in which each user requests different information. The next section designs HTBFs for a broadcast system, where all the users request the same information, i.e., there is only a single group. Thus, it can also be viewed as a special instance of the multicast scenario, where there is no inter-group or inter-user interference. This interference-free scenario simplifies the HTBF design problem, as described in the next section.
\section{Hybrid Broadcast Beamforming (HBBF)}\label{HBB}
\label{sec HBB}
In a broadcast scenario, the sum signal-to-noise ratio (SSNR) of all the users in the system is maximized subject to the total power constraint ($P_T$) in order to obtain the optimal HBBF. The motivation for SSNR maximization can be summarised as follows. Note that at low-SNRs, we have $\log_2 \left(1 + \mathrm{SNR}_u\right) \approx \frac{1}{\mathrm{ln}\left(2\right)}\mathrm{SNR}_u$. Therefore, one can deduce that the maximization of capacity is identical to SSNR maximization. Furthermore, it can observed from the following relationship, which exploits Jensen's inequality, that SSNR maximization provides an upper bound of the true capacity:
$$\frac{1}{U} \sum_{u=1}^U \log_2 \left(1 + \mathrm{SNR}_u\right) \leq \log_2 \left(1 + \frac{1}{U}\sum_{u=1}^U\right).$$
Considering a broadcast scenario to $U$ users, the signal $\mathbf{r}_u \in \mathbb{C}^{N_R \times 1} $ received at user $u$ can be expressed as
\begin{equation*}
\mathbf{r}_u = \sum_{m=1}^{M}\mathbf{H}_{u,m}\mathbf{F}_{\mathrm{RF},m}\mathbf{f}_{\mathrm{BB},m}x + \boldsymbol{\delta}_u,
\end{equation*}
where $\mathbf{H}_{u,m}\in \mathbb{C}^{N_{R} \times N_{T}} $ represents the channel matrix between the user $u$ and the $m$th AP, $\mathbf{F}_{\mathrm{RF},m} \in \mathbb{C}^{N_{T} \times N_{\mathrm{RF},m}}$ and $\mathbf{f}_\mathrm{BB,m} \in \mathbb{C}^{N_{\mathrm{RF},m} \times 1}$ denote the RF and baseband TPCs respectively, at the $m$th AP. Furthermore, $\boldsymbol{\delta}_u \in \mathbb{C}^{N_{R} \times 1}$ represents the AWGN at user $u$ and $x$ denotes the transmitted symbol corresponding to the broadcast scenario, which is normalized to have unit power i.e., $ \mathop{\mathbb{E}}\left[|x|^2 \right] = 1$. After RF combining using the combiner matrix $\mathbf{W}_{\mathrm{RF},u} \in \mathbb{C}^{N_R \times N_{\mathrm{RF},u}}$,
the received signal $\tilde{\mathbf{r}}_u$ can be written as
\begin{align}
\tilde{\mathbf{r}}_u =& \mathbf{W}_{\mathrm{RF},u}^H \left( \sum_{m=1}^{M}\mathbf{H}_{u,m}\mathbf{F}_{\mathrm{RF},m} \mathbf{f}_{\mathrm{BB},m}x \right) + \tilde{\boldsymbol{\delta}}_u,
\label{Equation 2}
\end{align}
where $\tilde{\boldsymbol{\delta}}_u = \mathbf{W}_{\mathrm{RF},u}^H \boldsymbol{\delta}_u$. Note that the columns of the RF TPC $\mathbf{F}_{\mathrm{RF},m}$ can be chosen as the dominant $N_{\mathrm{RF},m}$ transmit array steering vectors corresponding to each user. Similarly, the RF RC $\mathbf{W}_{\mathrm{RF},u}$ can be configured with the $N_{\mathrm{RF},u}$ dominant receive array steering vectors corresponding to each AP. Thus, to design the RF precoder for the $m$th AP, one can first find the index $\mathcal{K}_{u,m}$ of the multipath component corresponding to the $u$th user, which has the maximum magnitude path gain, i.e., $\mathcal{K}_{u,m} = \max\left\{\left|\alpha_{u,m}^{(l)}\right|, \left|\alpha_{u,m}^{(2)}\right|,\dots,\left|\alpha_{u,m}^{(L)}\right|\right\}$. Then, the RF TPC $\mathbf{F}_{\mathrm{RF},m}$ for the $m$th AP
can be obtained by extracting the $N_{\mathrm{RF},m}$ columns indexed by the set $\mathcal{K}_m = \left\{\mathcal{K}_{1,m},\mathcal{K}_{2,m},\dots,\mathcal{K}_{U,m}\right\}$ from the matrices $\mathbf{A}_{T,u,m}$. This can be mathematically represented as
\begin{equation}
\mathbf{F}_{\mathrm{RF},m} = \left[\mathbf{A}_{T,1,m}(:,\mathcal{K}_{1,m}),\dots, \mathbf{A}_{T,U,m}(:,\mathcal{K}_{U,m})\right].
\end{equation}
In a similar fashion, one can also obtain the RF RC $\mathbf{W}_{\mathrm{RF},u}$ for each user. Let the effective channel matrix $\mathbf{H}_{\mathrm{eff},u} \in \mathbb{C}^{N_{\mathrm{RF},u} \times N_{\mathrm{RF},m}}$ between the user $u$ and the $m$th AP be defined as
$\mathbf{H}_{\mathrm{eff},u,m} = \mathbf{W}_{\mathrm{RF},u}^H\mathbf{H}_{u,m}\mathbf{F}_{\mathrm{RF},m}.$
Furthermore, let the concatenated effective channel matrix $\mathbf{H}_{\mathrm{eff},u} \in \mathbb{C}^{N_{\mathrm{RF},u} \times \sum_{m=1}^{M}N_{\mathrm{RF},m}}$ for user $u$ be defined as
$$\mathbf{H}_{\mathrm{eff},u} = \left[\mathbf{H}_{\mathrm{eff},u,1} \mathbf{H}_{\mathrm{eff},u,2} \cdots \mathbf{H}_{\mathrm{eff},u,M} \right].$$
The received signal $\tilde{\mathbf{r}}_u$ of \eqref{Equation 2} can be completely rewritten as
\begin{align}\label{EqBHBRBroad}
&\tilde{\mathbf{r}}_u = \sum_{m=1}^{M}\mathbf{H}_{\mathrm{eff},u,m}\mathbf{f}_{\mathrm{BB},m}x + \tilde{\boldsymbol{\delta}}_ u = \mathbf{H}_{\mathrm{eff},u} \mathbf{f}_\mathrm{{BB}}x + \tilde{\boldsymbol{\delta}}_ u,
\end{align}
where $\mathbf{f}_\mathrm{{BB}} = \left[\mathbf{f}_{\mathrm{BB},1}^T, \mathbf{f}_{\mathrm{BB},2}^T, \cdots, \mathbf{f}_{\mathrm{BB},M}^T\right]^T \in \mathbb{C}^{\sum_{m=1}^{M} N_{\mathrm{RF},m} \times 1}$ denotes the stacked baseband TPs of all the APs. Note that the signal power received at user $u$ is given as $\mathbf{f}_\mathrm{BB}^H\mathbf{H}_{\mathrm{eff},u}^H\mathbf{H}_{\mathrm{eff},u}\mathbf{f}_\mathrm{BB} $, which implies that the total received power for all the users in the system can be written as $\sum_{u=1}^{U}\mathbf{f}_\mathrm{BB}^H\mathbf{H}_{\mathrm{eff},u}^H\mathbf{H}_{\mathrm{eff},u}\mathbf{f}_\mathrm{BB} = \mathbf{f}_\mathrm{BB}^H \mathbf{H}\mathbf{f}_\mathrm{BB}$,
where $\mathbf{H} = \sum_{u=1}^{U}\mathbf{H}_{\mathrm{eff},u}^H\mathbf{H}_{\mathrm{eff},u} $. Our objective is to find the optimal stacked baseband TPC $\mathbf{f}_\mathrm{BB}$ for the HBBF in order to maximize the SSNR. Toward this, the optimization problem can be formulated as
\begin{equation*}
\begin{aligned}
& \underset{\mathbf{f}_\mathrm{BB}}{\text{maximize}}
& & \mathbf{f}_\mathrm{BB}^H \mathbf{H}\mathbf{f}_\mathrm{BB} \\
& \text{subject to}
& & \left\lVert \mathbf{f}_\mathrm{BB} \right\rVert^2 \leq P_T.
\end{aligned}
\end{equation*}
The optimal baseband TPC $\mathbf{f}_\mathrm{BB}$ for the optimization problem above is given by $\mathbf{f}_\mathrm{BB}= \sqrt{P_T} \boldsymbol{\lambda}_{ \text{max}} \left(\mathbf{H}\right),$
where $\boldsymbol{\lambda}_{ \text{max}}(\mathbf{H})$ represents the eigenvector corresponding to the largest eigenvalue of the matrix $\mathbf{H}$. Interestingly, it can be observed that the SSNR of the mmWave broadcast system can be maximized by sending the information in the direction of the leading eigenmode of the effective channel matrix $\mathbf{H}$.
Substituting this optimal $\mathbf{f}_{\mathrm{BB}}$ into \eqref{EqBHBRBroad}, we have
\begin{equation}
\tilde{\mathbf{r}}_u = \sqrt{P_T} \mathbf{H}_{\mathrm{eff},u} \boldsymbol{\lambda}_{ \text{max}} \left(\mathbf{H}\right)x + \tilde{\boldsymbol{\delta}}_u .
\end{equation}
The covariance matrix $\mathbf{R}_{\tilde{\delta}_ u} \in \mathbb{C}^{N_{\mathrm{RF},u} \times N_{\mathrm{RF},u}}$ of the combined noise vector $\tilde{\boldsymbol{\delta}}_u $ can be expressed as
\begin{equation}
\mathbf{R}_{\tilde{\delta}_ u}= \mathop{\mathbb{E}}\left[\tilde{\boldsymbol{\delta}}_ u\tilde{\boldsymbol{\delta}}_ u^{H}\right] = \sigma^2_{\delta} \mathbf{W}_{\mathrm{RF},u}^H\mathbf{W}_{\mathrm{RF},u}.
\end{equation}
Therefore, the whitened system model can be written as
\begin{equation}
\mathbf{R}_{\tilde{\delta}_ u}^{-\frac{1}{2}} \tilde{\mathbf{r}}_u = \sqrt{P_T} \mathbf{R}_{\tilde{\delta}_ u}^{-\frac{1}{2}} \mathbf{H}_{\mathrm{eff},u} \boldsymbol{\lambda}_{ \text{max}} \left(\mathbf{H}\right)x + \mathbf{R}_{\tilde{\mathbf{\delta}}_ u}^{-\frac{1}{2}}\tilde{\boldsymbol{\delta}}_ u .
\end{equation}
The optimal baseband RC $\mathbf{w}_{\mathrm{BB},u}$ for user $u$ can now be readily found using the MRC criterion as
\begin{equation}
\mathbf{w}_{\mathrm{BB},u} = \frac{\mathbf{R}_{\tilde{\delta}_ u}^{-\frac{1}{2}}\mathbf{H}_{\mathrm{eff},u} \boldsymbol{\lambda}_{ \text{max}} \left(\mathbf{H}\right)}{\left\lVert \mathbf{R}_{\tilde{\delta}_ u}^{-\frac{1}{2}}\mathbf{H}_{\mathrm{eff},u} \boldsymbol{\lambda}_{ \text{max}} \left(\mathbf{H}\right) \right\rVert}.
\end{equation}
Next, we derive the cooperative HBBF for a broadcast system having per AP-power constraints. Furthermore, we also maximize the minimum SNR of the users in the system to incorporate fairness in the power allocation.
\subsection{HBBF-P}
This sub-section considers per-AP power constraints for performing cooperative HBBF for CFMU systems that maximizes the sum-SNR. Let us assume the power among the APs to be equally distributed as $\frac{P_T}{M}$ in the broadcast group. The associated optimization problem of the CFMU HBBF, which maximizes the sum-SNR, is given by \vspace{-10pt}
\begin{align*}
& \underset{\mathbf{F}_{\mathrm{BB},u}}{\text{maximize}}\quad \sum_{u=1}^{U} \text{Tr} \left(\mathbf{f}_\mathrm{BB}^H \mathbf{H}_{\mathrm{eff},u}^H\mathbf{H}_{\mathrm{eff},u}\mathbf{f}_\mathrm{BB}\right)\\ &\qquad\quad\ \equiv \sum_{u=1}^{U}\sum_{m=1}^{M} \text{Tr} \left(\mathbf{H}_{\mathrm{eff},u,m}^H\mathbf{H}_{\mathrm{eff},u,m}\mathbf{F}_{\mathrm{BB},m}\right)\\
& \text{subject to} \ \text{Tr}\left(\mathbf{F}_{\mathrm{BB},m}\right) \leq \frac{P_{T}}{M}, \, m=\mbox{1}, \mbox{2}, \ldots, M \nonumber \\ &\qquad\quad\ \ \ \ \ \mathbf{F}_{\mathrm{BB},m}\succcurlyeq 0 \ \ \text{and} \ \text{rank}\left(\mathbf{F}_{\mathrm{BB},m}\right)=1 \ \ \ \forall m,
\end{align*}
where $\mathbf{F}_{\mathrm{BB},m}=\mathbf{f}_{\mathrm{BB},m}\left(\mathbf{f}_{\mathrm{BB},m}\right)^H \in \mathbb{C}^{N_{\mathrm{RF},m} \times N_{\mathrm{RF},m}} $ denotes a rank-1 positive semi-definite (PSD) matrix. The above optimization problem is non-convex due to the rank-1 constraint, hence it is difficult to solve. Upon relaxing this rank-1 constraint, the equivalent semi-definite programme (SDP) can be expressed as
\begin{equation}\label{HBBP}
\begin{aligned}
& \underset{\mathbf{F}_{\mathrm{BB},m}}{\text{maximize}}
& & \sum_{u=1}^{U}\sum_{m=1}^{M} \text{Tr} \left(\mathbf{H}_{\mathrm{eff},u,m}^H\mathbf{H}_{\mathrm{eff},u,m}\mathbf{F}_{\mathrm{BB},m}\right)\\
& \text{subject to}
& & \text{Tr}\left(\mathbf{F}_{\mathrm{BB},m}\right) \leq \frac{P_{T}}{M}, \ \ \ m=1,2, \cdots, M, \\
& & & \ \mathbf{F}_{\mathrm{BB},m}\succcurlyeq 0 \ \ \ \forall m.
\end{aligned}
\end{equation}
\subsection{HBBF-P with a max-min per-AP power constraint}
In order to ensure fairness in the power allocation, the minimum SNR of the users can be maximized. Towards this, let the target SNR of each user be denoted by $\gamma$. To perform HBBF, the max-min fairness based optimization problem can be framed as
\begin{equation} \label{HBB max-min}
\begin{aligned}
\ \ \ \underset{\gamma, \mathbf{F}_\mathrm{BB,n}}{\text{maximize}}
\ \gamma \\
\hspace{-3cm} \text{s.t.}\ \
& \sum_{m=1}^{M} \text{Tr} \left(\mathbf{H}_{\mathrm{eff},u,m}^H\mathbf{H}_{\mathrm{eff},u,m}\mathbf{F}_{\mathrm{BB},m}\right) \geq \gamma, \ \forall u, \\
& \text{Tr}\left(\mathbf{F}_{\mathrm{BB},m}\right) \leq \mathrm{\frac{P_{T}}{M}}, \ \ \ \ m=1,2, \cdots, M, \\
& \mathbf{F}_{\mathrm{BB},m}\succcurlyeq 0, \ \ \ \forall m.
\end{aligned}
\end{equation}
It can be observed that the optimization problems \eqref{HBBP} and \eqref{HBB max-min} can be efficiently solved using widely available tools, such as CVX \cite{grant2014cvx}. The baseband beamformer $\mathbf{f}_{\mathrm{BB},m}$ can be finally determined as the eigenvector corresponding to the largest eigenvalue of $\mathbf{F}_{\mathrm{BB},m}^{*}$, i.e., $\mathbf{f}_{\mathrm{BB},m}^{\mathrm{opt}} = \sqrt{\lambda}_{\mathrm{max}}\hat{\mathbf{f}}_{\mathrm{BB},m}$, where $\hat{\mathbf{f}}_{\mathrm{BB},m}$ represents the unit-norm eigenvector corresponding to the eigenvalue $\lambda_{\mathrm{max}}$. Note that the relaxed SDP based beamformer obtained is an approximation of the optimal beamformer. It has been shown that the SDP relaxation results in a close approximation of the optimization problems involving non-convex quadratically constrained quadratic programs (QCQPs) \eqref{HBBP} and max–min-based quadratic programs \eqref{HBB max-min}.
Furthermore, it has been demonstrated that for SDP relaxation, the rank of an optimal solution is upper bounded by $\sqrt{M}$ \cite{palomar2010convex}. For practical values of $M$, the experimental studies in \cite{palomar2010convex} demonstrate that the average approximation ratio is close to unity. The computational complexity order of HBBF-P scheme is $\mathcal{O}\left[\left(MN_{\mathrm{RF},m}\right)^{3.5}\right]$ \cite{luo2010semidefinite, vandenberghe1996semidefinite}.
\section{Unicast Scenario: Optimal Successive Hybrid Transmit Beamforming}\label{Unicast}
This section jointly designs the optimal HTBFs and RCs for the mmWave unicast scenarios by maximizing the SINR in successive order. The algorithm developed herein will be extended to multicast systems in the next section.
In this setting, there are $G = U$ user groups with one user in each group, i.e., $U_g = 1$, where $U$ denotes the total number of users present in the system. The signal $\mathbf{r}_u \in \mathbb{C}^{N_R \times 1}$ received at the $u$th user can be written as
\begin{align}
\mathbf{r}_u =& \sum_{m=1}^{M}\mathbf{H}_{u,m}\mathbf{F}_{\mathrm{RF},m}\mathbf{f}_{\mathrm{BB},u,m}x_u \nonumber \\ +& \sum_{\substack{i=1\\ i \neq u}}^{U} \sum_{m=1}^{M}\mathbf{H}_{u,m}\mathbf{F}_{\mathrm{RF},m}\mathbf{f}_{\mathrm{BB},i,m}x_i + \boldsymbol{\delta}_{u},
\label{Equation u1}
\end{align}
where $\boldsymbol{\delta}_u \in \mathbb{C}^{N_{R} \times 1}$ represents the AWGN at user $u$ and $x_u$ denotes the symbol transmitted to user $u$, which is normalized to have unit power. Furthermore, $\mathbf{f}_{\mathrm{BB},u,m} \in \mathbb{C}^{N_{\mathrm{RF},m} \times 1}$ represents the baseband TPC at the $m$th AP for the user $u$.
After employing the RF RC $\mathbf{W}_{\mathrm{RF},u} \in \mathbb{C}^{N_{u} \times N_{\mathrm{RF},u}}$, the received signal, $\tilde{\mathbf{r}}_u = \mathbf{W}_{\mathrm{RF},u}^H \mathbf{r}_u$, is given as:
\begin{align}\label{EqBHBR}
\tilde{\mathbf{r}}_u =&\, \mathbf{W}_{\mathrm{RF},u}^H\sum_{m=1}^{M}\mathbf{H}_{u,m}\mathbf{F}_{\mathrm{RF},m}\mathbf{f}_{\mathrm{BB},u,m}x_u \nonumber \\ +& \mathbf{W}_{\mathrm{RF},u}^H\sum_{\substack{i=1\\ i \neq u}}^{U} \sum_{m=1}^{M}\mathbf{H}_{u,m}\mathbf{F}_{\mathrm{RF},m}\mathbf{f}_{\mathrm{BB},i,m}x_i + \tilde{\boldsymbol{\delta}}_u,
\end{align}
where $\tilde{\boldsymbol{\delta}}_u = \mathbf{W}_{\mathrm{RF},u}^H \boldsymbol{\delta}_u$. Let $\mathbf{H}_{\mathrm{eff},u,m} = \mathbf{W}_{\mathrm{RF},u}^H\mathbf{H}_{u,m}\mathbf{F}_{\mathrm{RF},m}$ denote the effective channel matrix between the $m$th AP and user $u$.
Let the concatenated baseband TPC and the effective channel matrix of the $u$th user be defined as
$$\mathbf{f}_{\mathrm{BB},u} = [\mathbf{f}_{\mathrm{BB},u,1}^{T}, \mathbf{f}_{\mathrm{BB},u,2}^{T}, \ldots, \mathbf{f}_{\mathrm{BB},u,M}^{T}]^{T},$$
$$\!\!\!\!\!\!\!\!\! \mathbf{H}_{\mathrm{eff},u} = \left[\mathbf{H}_{\mathrm{eff},u,1} \mathbf{H}_{\mathrm{eff},u,2} \cdots \mathbf{H}_{\mathrm{eff},u,M} \right],$$ respectively.
After baseband combining at user $u$ employing the RC $\mathbf{w}_{\mathrm{BB},u} \in \mathbb{C}^{N_{\mathrm{RF}, u} \times 1}$, the estimate $\hat{x}_u$ of the symbol $x_u$ is given by
\begin{align}\label{estUnicast}
\hat{x}_u =& \mathbf{w}_{\mathrm{BB},u}^H\mathbf{H}_{\mathrm{eff},u} \mathbf{f}_{\mathrm{BB},u} x_u + \sum_{i=1, i \neq u}^{U}\mathbf{w}_{\mathrm{BB},u}^H \mathbf{H}_{\mathrm{eff},u} \mathbf{f}_{\mathrm{BB},i} x_i \nonumber \\ +& \mathbf{w}_{\mathrm{BB},u}^H \boldsymbol{\delta}_{u}.
\end{align}
Note that the first and second terms in \eqref{estUnicast} correspond to the desired signal and inter-user interference, respectively. Next, we develop the OSHB-U scheme similar to the successive optimization method in multi-user MIMO systems \cite{spencer2004zero} that employs cooperative beamforming for iteratively multiplexing multiple users in the downlink, together with exploiting the MVDR beamformer of \cite{van2002modulation} for interference minimization in the mmWave cellular system. Furthermore, it maximizes the SINR of each user while cancelling the interferences imposed on the users, which have already been scheduled. Toward this, we now derive the proposed OSHB-U technique for a unicast scenario.
To begin with, it can be readily observed that while scheduling the $u$th user, the zero-forcing (ZF) interference constraint of $\mathbf{w}_{\mathrm{BB},v}^H \mathbf{H}_{\mathrm{eff},v} \mathbf{f}_{\mathrm{BB},u} = \mbox{0}$, $\mbox{1}\leq v \leq u-\mbox{1}$, successfully eliminates the interference imposed by the previously scheduled users. Hence, the baseband TPC is given by $\mathbf{f}_{\mathrm{BB},u}=\mathbf{M}_{u-1}^{\perp} \boldsymbol{\nu}_{u}$, where $\boldsymbol{\nu}_{u} \in \mathbb{C}^{\sum_{m=1}^{M}N_{\mathrm{RF},m}-u+1}$, and $\mathbf{M}_{u-1}^{\perp}\in\mathbb{C}^{\sum_{m=1}^{M}N_{\mathrm{RF},m}\times(\sum_{m=1}^{M}N_{\mathrm{RF},m}-u+1)}$ represents the orthonormal basis to the null space of the matrix $\mathbf{M}_{u-1} \in \mathbb{C}^{(u-1) \times \sum_{m=1}^{M} N_{\mathrm{RF},m}}$, which is given by
\begin{align}\label{MainMatrix2}
\mathbf{M}_{u-1}=\left[\left(\mathbf{w}_{\mathrm{BB},1}^H\mathbf{H}_{\mathrm{eff},1}\right)^T,\ldots, \left(\mathbf{w}_{\mathrm{BB},u-1}^H\mathbf{H}_{\mathrm{eff},u-1}\right)^T\right]^T.
\end{align}
The matrix $\mathbf{M}_{u-1}^{\perp}$ can be readily derived from the singular value decomposition (SVD) of the matrix $\mathbf{M}_{u-1}$, and it corresponds to the right singular vectors matrix composed of $\left(\sum_{m=1}^{M}N_{\mathrm{RF},m}-u+1\right)$. Furthermore, the transmissions of the successive users $u+1, u+2, \ldots, U$ are also orthogonal to those of the $u$th user, i.e., $\mathbf{w}_{\mathrm{BB},u}^H \mathbf{H}_{\mathrm{eff},u} \mathbf{f}_{\mathrm{BB},v}=\mbox{0}$, $u + 1 \leq v \leq U$. Therefore , the effective received signal after RF precoding and combining at user $u$ can be expressed as
\begin{align}
\tilde{\mathbf{r}}_u =&\, \mathbf{H}_{\mathrm{eff}, u} \mathbf{f}_{\mathrm{BB},u} x_u + \sum_{i=1}^{u-1}\mathbf{H}_{\mathrm{eff}, u} \mathbf{f}_{\mathrm{BB},i} x_i + \tilde{\boldsymbol{\delta}}_{u} \nonumber \\ =&\, \mathbf{H}_{\mathrm{eff},u} \mathbf{M}_{u-1}^{\perp} \boldsymbol{\nu}_{u} x_u + \sum_{i=1}^{u-1}\mathbf{H}_{\mathrm{eff},u} \mathbf{M}_{i-1}^{\perp} \boldsymbol{\nu}_{i} x_i + \tilde{\boldsymbol{\delta}}_{u}.
\label{eq 16}
\end{align}
Let the interference and noise terms in \eqref{eq 16} be collectively denoted by $\boldsymbol{\tau}_u = \sum_{i=1}^{u-1}\mathbf{H}_{\mathrm{eff},u} \mathbf{M}_{u-1}^{\perp} \boldsymbol{\nu}_{i} x_i + \tilde{\boldsymbol{\delta}}_{u} $. Its covariance matrix $\mathbf{C}_{\boldsymbol{\tau}_{u}} \in \mathbb{C}^{N_{\mathrm{RF},u} \times N_{\mathrm{RF},u}}$ is given by
\begin{align}\label{covar matrix}
\mathbf{C}_{\boldsymbol{\tau}_{u}} =&\, \sum_{i=1}^{u-1}\mathbf{H}_{\mathrm{eff},u} \mathbf{f}_{\mathrm{BB},i} \left(\mathop{\mathbb{E}}\left[|x_i|^2 \right]\right)\mathbf{f}_{\mathrm{BB},i}^{H} \mathbf{H}_{\mathrm{eff},u}^{H} + \mathop{\mathbb{E}}\left[\boldsymbol{\delta}_{u} \boldsymbol{\delta}_{u}^H \right]\nonumber \\ =&\, \sum_{i=1}^{u-1} \mathbf{H}_{\mathrm{eff},u} \left(\mathbf{M}_{i-1}^{\perp} \boldsymbol{\nu}_{i} \boldsymbol{\nu}_{i}^{H} \left(\mathbf{M}_{i-1}^{\perp}\right)^{H}\right) \mathbf{H}_{\mathrm{eff},u}^{H} \nonumber \\ +& \sigma_{\boldsymbol{\delta}}^{2}\mathbf{W}_{\mathrm{RF},u}^H\mathbf{W}_{\mathrm{RF},u}.
\end{align}
To obtain the optimal baseband TPC and RC, one can write the effective interference-whitened system model corresponding to (\ref{eq 16}) as
\begin{equation}
\mathbf{C}_{\boldsymbol{\tau}_{u}}^{-\frac{1}{2}}\tilde{\mathbf{r}}_{u}= \mathbf{C}_{\boldsymbol{\tau}_{u}}^{-\frac{1}{2}}\mathbf{H}_{\mathrm{eff},u}\mathbf{M}_{u-1}^{\perp} \boldsymbol{\nu}_{u} x_{u} +\mathbf{C}_{\boldsymbol{\tau}_{u}}^{-\frac{1}{2}}\boldsymbol{\tau}_{u}.
\end{equation}
Therefore, the optimal baseband TPC and RC for user $u$, which maximize the SINR are given by $\mathbf{f}_{\mathrm{BB},u} = \mathbf{M}_{u-1}^{\perp} \boldsymbol{\nu}_{u}$ and $\mathbf{w}_{\mathrm{BB},u} = {\left(\mathbf{C}_{\boldsymbol{\tau}_{u}}^{-(1/2)}\right)}^H \tilde{\mathbf{w}}_{\mathrm{BB},u}$, where $\tilde{\mathbf{w}}_{\mathrm{BB},u}$ and $\boldsymbol{\nu}_{u}$ are the principal left and right singular vectors, respectively, of the matrix $\mathbf{C}_{{\boldsymbol{\tau}}_{u}}^{-({1}/{2})}\mathbf{H}_{\mathrm{eff},u}\mathbf{M}_{u-1}^{\perp} \in \mathbb{C}^{N_{\mathrm{RF},u} \times (MN_{\mathrm{RF},m} - u+1)}$. The detailed proofs of the results obtained above are given in Appendix A.
The optimal vector $\mathbf{w}_{\mathrm{BB},u} = {\left(\mathbf{C}_{\boldsymbol{\tau}_{u}}^{-(1/2)}\right)}^H \tilde{\mathbf{w}}_{\mathrm{BB},u}$ is known as the MVDR beamformer that maximizes the SINR. This is termed as MVDR, because the output signal at user $u$ has the minimum variance and the desired signal is not distorted.
Therefore, the minimum variance distortionless response (MVDR) together with the optimal successive minimum variance hybrid beamforming (OSHB) results in the optimal TPC that maximizes the SINR, while also nulling the interference emanating from the previously scheduled $u-1$ users. Note that the proposed optimal hybrid beamformer is obtained in two steps. In the first step, we employ the CSI toward design of the optimal RF TPCs. In the second step, the proposed baseband TPC is derived based on the MVDR beamformer. Therefore, this results in an optimal hybrid beamformer that maximizes the SINR along with interference cancellation.
Algorithm \ref{algo_S_OMP} presents the step-by-step procedure of the OSHB algorithm for obtaining the baseband TPC and RC. The key features of the proposed OSHB technique for the mmWave unicast CFMU MIMO system are described next.
To begin with, note that the interference cancellation procedure of the proposed OSHB method is similar to the classic successive interference cancellation (SIC) methods, wherein once the signal of user $u$ is successfully recovered, one can remodulate it and then subtract it from the received composite signal for decontaminating it and for reducing the decoding complexity of the remaining users at the APs. Hence, the baseband TPC configured for the $u$th user is specifically chosen for ensuring that the interference emanating from the formerly selected $(u-1)$ users is zero. This can be realized by selecting the TPC for ensuring that it lies in the subspace orthogonal to the column space of the matrix $\mathbf{M}_{u-1}$.
Furthermore, one can observe that the maximum number of users supported by the OSHB-U scheme is $\sum_{m=1}^{M}N_{\mathrm{RF},m}$. In order to cancel the interference imposed on user $u$, the number of columns in the null space of the matrix $\mathbf{M}_{u-1}^{\perp}\in\mathbb{C}^{\sum_{m=1}^{M}N_{\mathrm{RF},m}\times(\sum_{m=1}^{M}N_{\mathrm{RF},m}-u+1)}$ must be larger than or equal to one, i.e. $\sum_{m=1}^{M}N_{\mathrm{RF},m}-u+1 \geq 1$, which yields $U \leq \sum_{m=1}^{M}N_{\mathrm{RF},m}$. This implies that the number of RF chains at each AP has to be increased for multiplexing a higher number of users. On the other hand, if one considers the overall effective channel matrix $\mathbf{H}_\mathrm{eff} = {[\mathbf{H}_{\mathrm{eff},1}^T,\mathbf{H}_{\mathrm{eff},u}^T, \ldots,\mathbf{H}_{\mathrm{eff},U}^T]}^T \in \mathbb{C}^{UN_{\mathrm{RF},u} \times M N_{\mathrm{RF},m}}$ to perform block diagonalization, then one has to satisfy the constraint $UN_{\mathrm{RF},u} \leq \sum_{m=1}^{M}N_{\mathrm{RF},m}$. Hence, the block diagonalization method specifically designed for perfect interference removal can only support $\lfloor \sum_{m=1}^{M} N_{\mathrm{RF},m}/N_{\mathrm{RF},u}\rfloor$ users. Therefore, it can be readily observed that a substantially higher number of users can be served using the proposed OSHB scheme than by the conventional block diagonalization approach, which demonstrates its practical suitability.
\begin{algorithm}[t!]
\caption{OSHB algorithm}\label{algo_S_OMP}
\textbf{Input:} Effective channel matrix $\mathbf{H}_{\mathrm{eff},u}$, $\mathbf{W}_{\mathrm{RF},u}$\\
\textbf{Initialization:} $\mathbf{M}_{0} = \mathbf{I}_{1 \times MN_{\mathrm{RF},m}}$, $\mathbf{C}_{\tau_{1}} = \mathbf{W}_{\mathrm{RF},1}^H\mathbf{W}_{\mathrm{RF},1}$ \\
\textbf{for} $u = 1:U$ \textbf{do}
\begin{enumerate}
\item $\left[\mathbf{U} \ \mathbf{S} \ \mathbf{V}\right] = \mathrm{svd}\left(\mathbf{M}_{u-1}\right)$
\item $\mathbf{M}_{u-1}^{\perp} = \mathbf{V}\left(:,u:MN_{\mathrm{RF},m}\right)$
\item $\left[\mathbf{U}^{'} \ \mathbf{S}^{'} \ \mathbf{V}^{'} \right] = \mathrm{svd}\left(\mathbf{C}_{\tau_{u}}^{\frac{-1}{2}}\mathbf{H}_{\mathrm{eff},u}\mathbf{M}_{u-1}^{\perp}\right)$
\item $\boldsymbol{\nu}_{u} = \mathbf{V}^{'}\left(:,1\right)$
\item $\mathbf{f}_{\mathrm{BB},u} = \mathbf{M}_{u-1}^{\perp}\boldsymbol{\nu}_{u}$
\item $\mathbf{w}_{\mathrm{BB},u} = \left(\mathbf{C}_{\tau_{u}}^{\frac{-1}{2}}\right)^H\mathbf{U}^{'}\left(:,1\right)$
\item $\mathbf{C}_{\tau_{u}} \rightarrow $ update according to \eqref{covar matrix}
\item $\mathbf{M}_{u-1} \rightarrow $ update according to \eqref{MainMatrix2}
\end{enumerate}
\textbf{end for}\\
\end{algorithm}
\section{Optimal Successive Hybrid Beamforming for Multicast Scenarios}
\label{Multicastsytem}
As described previously, in a multicast scenario, the users belonging to the same multicast group tend to request the same data. For such a system, one can develop the OSHB-M scheme, similar to the OSHB-U scheme derived for the unicast scenario, which successfully cancels the interference arising from the previously chosen groups $1, 2,\ldots, g-1$ contaminating the $g$th group. Furthermore, to minimize the interference generated by the previously chosen groups, the MVDR-based beamformer can be employed for the $g$th group.
As described in Section II, the signal $\tilde{\mathbf{r}}^{(g)}_u \in \mathbb{C}^{N_{\mathrm{RF},u} \times 1}$ received for the user $U_u^{(g)}$ at the output of the RF RC is given as
\begin{align}
\tilde{\mathbf{r}}^{(g)}_u = \mathbf{H}^{(g)}_{\mathrm{eff},u}\mathbf{f}_{\mathrm{BB}}^{(g)} x^{(g)} + \sum_{\substack{l=1\\ l \neq i}}^{G}\mathbf{H}^{(g)}_{\mathrm{eff},u}\mathbf{f}_{\mathrm{BB}}^{(l)} x^{(l)} + \check{{\boldsymbol{\delta}}}^{(g)}_{u}.
\end{align}
The received signal concatenated for all the users in group $g$ can be formulated as
\begin{align}
\underset{{\mathbf{r}^{(g)}}}{\underbrace{\begin{bmatrix}\mathbf{r}_{1}^{(g)} \\ \mathbf{r}_{2}^{(g)} \\ \vdots \\ \mathbf{r}_{U_{g}}^{(g)} \end{bmatrix}}}=\underset{\mathbf{H}^{(g)}_{\mathrm{eff}}}{\underbrace{\begin{bmatrix} \mathbf{H}^{(g)}_{\mathrm{eff},1} \\ \mathbf{H}^{(g)}_{\mathrm{eff},2} \\ \vdots \\ \mathbf{H}^{(g)}_{\mathrm{eff},U_{g}} \end{bmatrix}}}\mathbf{f}_{\mathrm{BB}}^{(g)} x^{(g)}+\sum_{\substack{l=1\\ l \neq g}}^{G} \mathbf{H}^{(g)}_{\mathrm{eff}}\mathbf{f}_{\mathrm{BB}}^{(l)} x^{(l)}+ \underset{\boldsymbol{\delta}^{(g)}}{\underbrace{\begin{bmatrix} \boldsymbol{\delta}_{1}^{(g)} \\ \boldsymbol{\delta}_{2}^{(g)} \\ \vdots \\ \boldsymbol{\delta}_{U_{g}}^{(g)} \end{bmatrix}}}.
\label{Main_eq}
\end{align}
Let $\mathbf{w}_{\mathrm{BB}}^{(g)} = \left[\left(\mathbf{w}_{\mathrm{BB},1}^{(g)}\right)^T, \left(\mathbf{w}_{\mathrm{BB},2}^{(g)}\right)^T,\ldots,\left(\mathbf{w}_{\mathrm{BB},U_g}^{(g)}\right)^T\right]^T$ denote the concatenated baseband combiner of all the users for the $g$th group. The signal $\check{r}_u^{(g)} = \left(\mathbf{w}_{\mathrm{BB},u}^{(g)}\right)^H \hat{\mathbf{r}}_{u}^{(g)}$ received by the $g$th group after baseband RC is given by
\begin{align}\label{McastInt}
\check{r}_u^{(g)} =& \left(\mathbf{w}_{\mathrm{BB}}^{(g)}\right)^H\mathbf{H}^{(g)}_{\mathrm{eff}}\mathbf{f}_{\mathrm{BB}}^{(g)} x^{(g)} + \left(\mathbf{w}_{\mathrm{BB}}^{(g)}\right)^H\sum_{\substack{l=1\\ l \neq i}}^{G}\mathbf{H}^{(g)}_{\mathrm{eff}}\mathbf{f}_{\mathrm{BB}}^{(l)} x^{(l)} \nonumber \\ +& \check{{\boldsymbol{\delta}}}^{(g)}.
\end{align}
The proposed OSHB-M algorithm employs the ZF constraint $\mathbf{G}^{(g)}\mathbf{f}_{\mathrm{BB}}^{(g)}=\mbox{0}$ for cancelling the interference denoted by the second term in \eqref{McastInt}, where the matrix $\mathbf{G}^{(g)} \in \mathbb{C}^{(\sum_{j=1}^{g-1}U_j) \times \sum_{m=1}^{M}N_{\mathrm{RF},m}}$ is defined as $\mathbf{G}^{(g)} = \left[\left(\mathbf{F}^{(1)}\right)^T, \left(\mathbf{F}^{(2)}\right)^T, \ldots, \left(\mathbf{F}^{(g-1)}\right)^T\right]^T$, with
\begin{align}
\mathbf{F}^{j}=\begin{bmatrix} \left(\mathbf{w}_{\mathrm{BB},1}^{(g)}\right)^{H} \mathbf{H}^{(j)}_{\mathrm{eff},1}\\ \left(\mathbf{w}_{\mathrm{BB},2}^{(g)}\right)^{H} \mathbf{H}^{(j)}_{\mathrm{eff},2}\\ \vdots\\ \left(\mathbf{w}_{\mathrm{BB},U_{j}}^{(g)}\right)^{H} \mathbf{H}^{(j)}_{\mathrm{eff},U_{j}} \end{bmatrix} \in \mathbb{C}^{U_j \times MN_\mathrm{RF},m}.
\end{align}
Hence, the baseband TPC constructed for the $g$th group is given by $\mathbf{f}_{\mathrm{BB}}^{(g)} = {(\mathbf{G}^{(g)})}^{\perp} \boldsymbol{\lambda}^{(g)}$, where $\boldsymbol{\lambda}^{(g)} \in \mathbb{C}^{\left(\sum_{m=1}^{M}N_{\mathrm{RF},m} - \sum_{j=1}^{g-1}U_j\right) \times 1}$, and the columns of the matrix ${(\mathbf{G}^{(g)})}^{\perp} \in \mathbb{C}^{\sum_{m=1}^{M}N_{\mathrm{RF},m} \times (MN_{\mathrm{RF},m} - \sum_{j=1}^{g-1}U_j)}$ constitute an orthonormal basis for the null-space of the matrix $\mathbf{G}^{(g)}$, and can be obtained as described in Section IV for the matrix $\mathbf{G}_{u-1}^{\perp}$. Next, we denote the interference-plus-noise part $\boldsymbol{\tau}^{(g)}$ in (\ref{Main_eq}) as $\boldsymbol{\tau}^{(g)} = \sum_{l=1}^{g-1} \mathbf{H}^{(g)}_{\mathrm{eff}} \mathbf{f}_{\mathrm{BB}}^{(l)} + \boldsymbol{\delta}^{(g)}$. The covariance matrix $\mathbf{C}_{\tau}^{(g)} = \mathrm{E}[\boldsymbol{\tau}^{(g)} {(\boldsymbol{\tau}^{(g)})}^H] \in \mathbb{C}^{U_{g}N_{\mathrm{RF},u}\times U_{g}N_{\mathrm{RF},u}}$ of $\boldsymbol{\tau}^{(g)}$ for the $g$th multicast group can be written as
\begin{equation*}
\mathbf{C}_{\tau}^{(g)} = \sum_{l=1}^{g-1}\mathbf{H}^{(g)}_{\mathrm{eff}}\left(\mathbf{f}_{\mathrm{BB}}^{(l)}\left({\mathbf{f}_{\mathrm{BB}}^{(l)}}\right)^{H}\right) \left(\mathbf{H}^{(g)}_{\mathrm{eff}}\right)^{H}+\sigma_\delta^2 \mathbf{I}_{U_{g}N_{\mathrm{RF},u}}.
\end{equation*}
Therefore, the optimal baseband TPC $\mathbf{f}_{\mathrm{BB}}^{(g)}$ and RC $\mathbf{w}_{\mathrm{BB}}^{(g)}$ conditioned for the $g$th multicast group are given as $\mathbf{f}_{\mathrm{BB}}^{(g)} = {(\mathbf{G}^{(g)})}^{\perp} \boldsymbol{\lambda}^{(g)}$ and $\mathbf{w}_{\mathrm{BB}}^{(g)} = {\left({\left(\mathbf{C}_{\tau}^{(g)}\right)}^{-({1}/{2})}\right)}^H \tilde{\mathbf{w}}_{\mathrm{BB}}^{(g)}$, where $\tilde{\mathbf{w}}_{\mathrm{BB}}^{(g)}$ and $\boldsymbol{\lambda}^{(g)}$ represent the principal left and right singular vectors, respectively, of the matrix $\tilde{\mathbf{H}}^{(g)}_{\mathrm{eff}} = {(\mathbf{C}_{\tau}^{(g)})}^{-({1}/{2})}\mathbf{H}^{(g)}_{\mathrm{eff}}{(\mathbf{G}^{(g)})}^{\perp} \in \mathbb{C}^{U_g N_{\mathrm{RF},u} \times (MN_{\mathrm{RF},m} - \sum_{j=1}^{g-1}U_j)}$. This result follows from our analysis similar to that provided for the unicast scenario described in Appendix A. Finally, it can be readily observed that the proposed OSHB-U and OSHB-M schemes can be implemented at the complexity order of $\mathcal{O}\left(\left(\sum_{m=1}^{M}N_{\mathrm{RF},m}\right)^3\right)$. This is due to the fact that these schemes are based on the singular value decomposition of the matrices $\mathbf{M}_{u-1}$ and $\mathbf{G}^{(g)}$. Note that while deriving the multicast beamforming solutions we assumed that the multicast users are in the same geographical location. However, if the users that request the same data are in different geographical regions, it is more efficient to transmit separately, i.e., unicast, to each user \cite{sadeghi2018joint}.
\section{Hybrid precoder design using the Bayesian learning (BL)-based approach }
One can note that in the previous sections, the designs of the RF TPCs and RCs require the knowledge of the dominant array response vectors of the APs and users, respectively, which eventually necessitates the full knowledge of the AoAs/AoDs of the multipath components and their complex-valued path gains. However, in practice, it is challenging to acquire accurate estimates of these quantities, which motivates us to develop a scheme that can efficiently design the hybrid TPCs and RCs without requiring the knowledge of the full channel state information (CSI).
To this end, in this section, we jointly design the baseband and RF TPCs and RCs for a unicast scenario. However, a similar approach can be readily developed for a general multicast scenario. The proposed joint design first obtains the optimal digital baseband TPC and RC $\mathbf{f}_{\mathrm{opt}}$ and $\mathbf{w}_{\mathrm{opt}}$, respectively, assuming a fully digital architecture. This step only requires the knowledge of the mmWave MIMO channel matrices $\mathbf{H}_{u,m}$, rather than their component AoAs/AoDs and path gains. Subsequently, a BL-based algorithm decomposes the fully digital precoders/combiners $\mathbf{f}_{\mathrm{opt}}$ and $\mathbf{w}_{\mathrm{opt}}$ into their RF and baseband components. The signal $\mathbf{r}_u \in \mathbb{C}^{N_R \times 1}$ received by the $u$th user can be expressed as
\begin{align}
\mathbf{r}_u = \sum_{m=1}^{M}\mathbf{H}_{u,m}\mathbf{f}_{u,m}x_u + \sum_{\substack{i=1\\ i \neq u}}^{U} \sum_{m=1}^{M}\mathbf{H}_{u,m}\mathbf{f}_{i,m}x_i + \boldsymbol{\delta}_{u}.
\end{align}
Let the concatenated fully digital TPC and channel matrix for the $u$th user be defined as $\mathbf{f}_{u} = [\mathbf{f}_{u,1}^{T}, \mathbf{f}_{u,2}^{T}, \ldots, \mathbf{f}_{u,M}^{T}]^{T}$, and
$\mathbf{H}_{u} = \left[\mathbf{H}_{u,1} \mathbf{H}_{u,2} \cdots \mathbf{H}_{,u,M}\right]$, respectively. Employing the fully digital RC $\mathbf{w}_{u} \in \mathbb{C}^{N_R \times 1}$, the signal $\tilde{r}_u = \mathbf{w}_{u}^H\mathbf{r}_u$ received by the $u$th user can be written as
\begin{align}
\tilde{r}_u = \mathbf{w}_{u}^H\mathbf{H}_{u}\mathbf{f}_{u}x_u + \sum_{\substack{i=1\\ i \neq u}}^{U} \mathbf{w}_{u}^H\mathbf{H}_{u}\mathbf{f}_{i}x_i + \underset{\tilde{\boldsymbol{\delta}}_u}{\underbrace{\mathbf{w}_{u}^H \boldsymbol{\delta}_u}}.
\end{align}
The SINR of the $u$th user is given by
\begin{align}
\text{SINR}_u = \frac{\left|\mathbf{w}_{u}^H\mathbf{H}_{u}\mathbf{f}_{u}\right|^2}{\sum_{i=1,i \neq u}^{U}{\left|\mathbf{w}_{u}^H\mathbf{H}_{u}\mathbf{f}_{i}\right|^2}+ \sigma_{\tilde{\delta}}^2}.
\end{align}
Upon employing the proposed OSHB-U algorithm, the optimal fully digital TPC and RC are obtained as $\mathbf{f}_{u}=\mathbf{A}_{u-1}^{\perp} \boldsymbol{\nu}_{u}$ and $\mathbf{w}_{u} = {(\mathbf{R}_{\tau_{u}}^{-(1/2)})}^H \tilde{\mathbf{w}}_{u}$ respectively, where $\boldsymbol{\nu}_{u}$ and $\tilde{\mathbf{w}}_{u}$ are the principal right and left singular vectors, respectively, of the matrix $\mathbf{R}_{\tau_{u}}^{-({1}/{2})}\mathbf{H}_{u}\mathbf{A}_{u-1}^{\perp} \in \mathbb{R}^{N_{R} \times (NN_{T} - u+1)}$. The columns of the matrix $\mathbf{A}_{u-1}^{\perp}\in\mathbb{C}^{MN_T \times \left(MN_T -u+1\right)}$ represent a basis orthonormal to the null-space of the matrix $\mathbf{A}_{u-1}=\left[\left(\mathbf{w}_{1}^H\mathbf{H}_{1}\right)^T,\ldots, \left(\mathbf{w}_{u-1}^H\mathbf{H}_{u-1}\right)^T\right]^T \in\mathbb{C}^{\left(u-1\right) \times MN_T}$. Furthermore, the concatenated fully digital TPC $\mathbf{F}_{\mathrm{opt}} \in \mathbb{C}^{MN_T \times U}$ corresponding to all users is given by $\mathbf{F}_{\mathrm{opt}} = \left[\mathbf{f}_{1}, \mathbf{f}_{2}, \ldots, \mathbf{f}_{U} \right]$. Now, the decomposition problem of jointly designing the baseband TPC $\mathbf{F}_{\mathrm{BB}} \in \mathbb{C}^{N_{\sum_{m=1}^{M}\mathrm{RF},m} \times U}$ and RF TPC $\mathbf{F}_{\mathrm{RF}} \in \mathbb{C}^{MN_{T} \times \sum_{m=1}^{M} N_{\mathrm{RF},m}}$ can be formulated as
\begin{align}\label{SBL1}
& \!\!\!\!\!\!\!\!\!\!\!\!\! \left(\mathbf{F}_{\mathrm{BB}}^{*}, \mathbf{F}_{\mathrm{RF}}^{*}\right)= \underset{\mathbf{F}_{\mathrm{BB}}, \mathbf{F}_{\mathrm{RF}}}{\arg \ \ \ \min}\left\|\mathbf{F}_{\mathrm{opt}} -\mathbf{F}_{\mathrm{RF}} \mathbf{F}_{\mathrm{BB}}\right\|_{\mathrm{F}}^2\nonumber \\
& \hspace{2cm} \text { s.t. } \ \ \ \ \vert \mathbf{F}_{\mathrm{RF}}(i,j)\vert = \frac{1}{\sqrt{N_T}},
\end{align}
where $\mathbf{F}_\mathrm{BB} = \left[\mathbf{f}_{\mathrm{BB},1}, \mathbf{f}_{\mathrm{BB},2},\ldots, \mathbf{f}_{\mathrm{BB},U}\right]$ denotes the concatenated baseband TPC and $\mathbf{F}_{\mathrm{RF}}=\mathrm{blkdiag}\left(\mathbf{F}_{\mathrm{RF},1},\ldots,\mathbf{F}_{\mathrm{RF},m}\right)$ represents the block-diagonal matrix of RF TPCs corresponding to all the APs. The constant magnitude constraint in \eqref{SBL1} is due to the analog phase shifter based architecture, which results in a non-convex optimization problem. Towards solving this, one can first define a dictionary matrix of the feasible transmit array response vectors as $\mathbf{G}_{T} = \left[\mathbf{a}_{T}\left(\phi_{1}\right), \mathbf{a}_{T}\left(\phi_{2}\right), \cdots, \mathbf{a}_{T}\left(\phi_{S}\right)\right] \in \mathbb{C}^{N_{T} \times S}$, where the set of AoDs $\{\phi_{s},\ \forall 1\leq s\leq S\}$ spans the angular range $[0,\ \pi]$, obeying $\displaystyle \cos(\phi_{s})=\frac{2}{S}(s-1)-1$, and the quantity $S$ represents the angular grid size \cite{heath2016overview}. Note that the elements of the dictionary matrix satisfy the constant magnitude constraint of \eqref{SBL1}. Hence, the columns of the RF TPC $\mathbf{F}_{\mathrm{RF}}$ can be readily chosen from the dictionary matrix $\mathbf{G}_T$.
The equivalent problem for hybrid precoder design in the CFMU mmWave MIMO system can be formulated as
\begin{equation}\label{SBLmain}
\begin{array}{c}
\underset{\tilde{\mathbf{F}}_{\mathrm{BB}}}{\arg \min}\left\|\mathbf{F}_{\mathrm{opt}}-\mathbf{G}_{T} \tilde{\mathbf{F}}_{\mathrm{BB}}\right\|_{\mathrm{F}} \\
\hspace{1.2cm}\mathrm{ s.t. } \hspace{0.5cm} \left\|\operatorname{diag}\left(\tilde{\mathbf{F}}_{\mathrm{BB}} \tilde{\mathbf{F}}_{\mathrm{BB}}^{H}\right)\right\|_{0} = \sum_{m=1}^{M} N_{\mathrm{RF},m},
\end{array}
\end{equation}
where $\tilde{\mathbf{F}}_{\mathrm{BB}} \in \mathbb{C}^{S \times U}$ denotes the intermediate baseband TPC. The constraint in \eqref{SBLmain} arises due to the fact that there can only be $\sum_{m=1}^{M} N_{\mathrm{RF},m}$ non-zero rows in the matrix $\tilde{\mathbf{F}}_{\mathrm{BB}}$, since there are only $N_{\mathrm{RF},m}$ RF chains at each AP. This leads to a simultaneous sparse structure in the matrix $\tilde{\mathbf{F}}_{\mathrm{BB}}$. The procedure of obtaining the proposed BL-based hybrid TPC corresponding to the optimization problem \eqref{SBLmain} is discussed next.
The BL method proposed for our hybrid TPC design assigns the parameterized Gaussian prior $p\left(\displaystyle \tilde{\mathbf{F}}_{\mathrm{BB}};{\boldsymbol{\Gamma}}\right)$ to the baseband TPC matrix $\tilde{\mathbf{F}}_{\mathrm{BB}}$, where we have
\begin{align}
p\left(\displaystyle \tilde{\mathbf{F}}_{\mathrm{BB}};{\boldsymbol{\Gamma}}\right)=&\, \prod_{i=1}^{S}p\left(\tilde{\mathbf{F}}_{\mathrm{BB}}(i,:);\gamma_{i}\right) \nonumber \\ =&\, \prod_{i=1}^{S} \frac{1}{\pi\gamma_i} \exp\left(- \frac{\norm{\tilde{\mathbf{F}}_{\mathrm{BB}}(i,:)}^2}{\gamma_i}\right),
\end{align}
and $\boldsymbol{\Gamma}=\mathrm{diag}\left(\gamma_{1},\ldots,\gamma_{S}\right) \in \mathbb{R}^{S \times S}$ denotes the diagonal matrix of hyper-parameters. One can observe that the $i$th row of the matrix $\tilde{\mathbf{F}}_{\mathrm{BB}}$ is assigned the hyperparameter $\gamma_i$, which enforces the row sparsity, as given in the constraint \eqref{SBLmain}. The MMSE estimator of the matrix $\tilde{\mathbf{F}}_{\mathrm{BB}}$, represented by the matrix $\boldsymbol{\Phi} \in \mathbb{C}^{S \times U}$ and the associated covariance matrix ${\boldsymbol{\Pi}}\in \mathbb{C}^{S \times S}$ are given as
\begin{equation}
\boldsymbol{\Phi} =\frac{1}{\sigma_{e}^{2}}\boldsymbol{\Pi}\mathbf{G}_{T}^{H}\mathbf{F}_{\mathrm{opt},n} \ \ \text{and} \ \ \displaystyle \mathbf{\Pi}=\left(\frac{1}{\sigma_{e}^{2}}\mathbf{G}_{T}^{H}\mathbf{G}_{T}+\mathbf{\Gamma}^{-1}\right)^{-1},
\label{eq. mu and sigma}
\end{equation}
where $\sigma_{e}^2$ denotes the variance of the approximation error. It can be observed that the MMSE estimate $\boldsymbol{\Phi}$ depends on the hyperparameter matrix $\mathbf{\Gamma}$. Furthermore, as $\gamma_i \rightarrow 0$, the $i$th row of the matrix $\tilde{\mathbf{F}}_{\mathrm{BB}}$ denoted by $\tilde{\mathbf{F}}_{\mathrm{BB}}(i,:)\rightarrow0$. Thus, the estimation of $\tilde{\mathbf{F}}_{\mathrm{BB}}$ is equivalent to the estimation of the associated hyperparameter vector $\boldsymbol{\gamma}=[\gamma_{1},\ldots,\gamma_{S}]^{T}$. To find the maximal likelihood estimate of $\boldsymbol{\gamma}$, the Bayesian evidence $p(\mathbf{F}_{\mathrm{opt}};\boldsymbol{\Gamma)}$ can be maximized. However, the pertinent optimization problem becomes non-convex, and it is hence difficult to solve. Therefore, the BL technique harnessed for hybrid TPC design in CFMU systems relies on a low-complexity iterative expectation-maximization (EM) framework, which guarantees likelihood maximization in each iteration, and convergence to a local optimum.
To begin with, let $\hat{\boldsymbol{\Gamma}}^{(k-1)} $ denote the estimate of the hyperparameter matrix $\boldsymbol{\Gamma}$ obtained from the $(k-1)$st iteration. The EM procedure involves two key steps. In the first step, known as the expectation step (E-step), we evaluate the average log-likelihood function $\mathcal{L}\left(\boldsymbol{\Gamma}\mid\hat{\boldsymbol{\Gamma}}^{(k-1)}\right)$ of the hyper-parameters as follows
\begin{align*}
\mathcal{L}\left(\boldsymbol{\Gamma}\mid\hat{\boldsymbol{\Gamma}}^{(k-1)}\right) = \mathbb{E}_{\tilde{{\bf F}}_{\mathrm{BB}}\mid{\bf F}_{\mathrm{opt}};\hat{\boldsymbol{\Gamma}}^{(k-1)}}\Bigg \lbrace\log p\left({\bf F}_{\mathrm{opt}},\tilde{{\bf F}}_{\mathrm{BB}};\boldsymbol{\Gamma}\right)\Bigg \rbrace.
\end{align*}
The second step, which is known as the maximization step (M-step), maximizes the average log-likelihood computed above with respect to the hyperparameter vector $\boldsymbol{\gamma}$ as
\begin{align}
\hat{\boldsymbol{\gamma}}^{(k)}=\arg \max _{\boldsymbol{\gamma}} \mathbb{E} \Bigg \lbrace\log p\left(\mathbf{F}_{\mathrm{opt}} \mid \tilde{\mathbf{F}}_{\mathrm{BB}}\right) + \log p\left(\tilde{\mathbf{F}}_{\mathrm{BB}} ; \boldsymbol{\Gamma}\right)\Bigg \rbrace. \label{eqHyp}
\end{align}
It can be readily observed from \eqref{eqHyp} that the expression $p\left(\mathbf{F}_{\mathrm{opt}} \mid \tilde{\mathbf{F}}_{\mathrm{BB}}\right)$ is independent of the hyperparameter $\boldsymbol{\gamma}$ and can be disregarded in the successive M-step. Therefore, the equivalent optimization problem is formulated as
\begin{align}\label{estimate_gamma}
\hat{\boldsymbol{\gamma}}^{(k)} =&\, \arg \max _{\boldsymbol{\gamma}} \mathbb{E}_{\tilde{\mathbf{F}}_{\mathrm{BB}} \mid \mathbf{F}_{\mathrm{opt}},\hat{\boldsymbol{\Gamma}}^{(k-1)}}\left\{\log p\left(\tilde{\mathbf{F}}_{\mathrm{BB}} ; \boldsymbol{\Gamma}\right)\right\}\nonumber \\
\equiv &\, \arg \max _{\boldsymbol{\gamma}} \sum_{i=1}^{S}\left[ -\log \left( \gamma_{i}\right)-\frac{\mathbb{E}\left(\norm{\tilde{\mathbf{F}}_{\mathrm{BB}}(i,:)}^{2}_{2}\right)}{\gamma_{i}}\right], \nonumber \\
\equiv &\, \arg \max _{\boldsymbol{\gamma}} \sum_{i=1}^{S} -\log \left( \gamma_{i}\right)-\frac{\norm{\boldsymbol{\Phi}^{(k)}(i,:)}^{2} + U \boldsymbol{\Pi}_{(i, i)}^{(k)}}{\gamma_{i}},
\end{align}
where $\boldsymbol{\Phi}^{(k)}$ and $\boldsymbol{\Pi}^{(k)}$ are obtained from (\ref{eq. mu and sigma}) by setting $\boldsymbol{\Gamma} = \hat{\boldsymbol{\Gamma}}^{(k-1)}$. The optimal value of $\hat{\gamma}_{i}^{(k)}$ can be evaluated by equating the gradient of the objective function in \eqref{estimate_gamma} with respect to $\boldsymbol{\gamma}$ to zero, which yields the update equation for each hyperparameter as
\begin{equation*}
\displaystyle \hat{\gamma}_{i}^{(k)}= \frac{1}{U} \norm{\boldsymbol{\Phi}^{(k)}(:,i)}^2+\boldsymbol{\Pi}^{(k)}_{(i,i)}.
\end{equation*}
Upon convergence of the EM procedure, the RF and baseband TPCs are obtained as follows.
Let $\mathcal{S}$ denote the set of indices for the $\left(\sum_{m=1}^{M}N_{\mathrm{RF},m}\right)$ hyper-parameters having the largest magnitudes. The concatenated optimal baseband TPC matrix $\mathbf{F}_{\mathrm{BB}}^{*}$ can be extracted from $\tilde{\mathbf{F}}_{\mathrm{BB}}$ as
\begin{align}\label{FbbBL}
\mathbf{F}_{\mathrm{BB}}^{*} = \tilde{\mathbf{F}}_{\mathrm{BB}}\left(\mathcal{S},:\right).
\end{align}
Similarly, the optimal RF TPC $\mathbf{F}_{\mathrm{RF}}^{*}$ can be obtained from the dictionary matrix of the transmit array response $\mathbf{G}_T$ as
\begin{align}\label{FrfBL}
\mathbf{F}_{\mathrm{RF}}^{*} = \mathbf{G}_{T}\left(:,\mathcal{S} \right).
\end{align}
Algorithm 2 presents the step-by-step procedure of the BL-based technique for designing the hybrid TPC. One can follow a similar procedure to obtain the digital baseband and RF RCs from the fully digital design. The complexity analysis of the proposed two-stage distributed hybrid beamformer design is summarized next. Due to lack of space, the detailed derivations for the computational complexities of Algorithms 1 and 2 have been relegated to our technical report \cite{TechReport}. It can be observed that the complexity of Algorithm-1, which designs the fully-digital transmit precoder (TPC), is $\mathcal{O}\left(\left(U-1 \right)^2 MN_T\right)$. Next, the fully-digital TPC is decomposed into its constituent radio frequency (RF) and baseband (BB) precoders using Algorithm-2. This step incurs a complexity of order $\mathcal{O}\left(S^3\right)$, which can be attributed to the matrix inversion of size $\left[S \times S \right]$ in Eq. 29. Since, $S >> \left(U-1 \right)^2MN_T$, the overall complexity of the coordinated hybrid TPC can be closely approximated by $\mathcal{O}\left(S^3\right)$. Furthermore, the global convergence of the BL algorithm's cost function is obtained at the sparsest solution, which guarantees the sparsest representation of the digital precoder \cite{wipf2004sparse}. Further, as a benefit of the EM-algorithm, the proposed BL technique is guaranteed to converge to a fixed point of the log-likelihood function, regardless of its initialization. This leads to the robust performance of the BL algorithm and makes it ideally suited for hybrid precoder/ combiner design in mmWave MIMO networks.
\begin{algorithm}\label{BLtechnique}
\SetAlgoLined
\textbf{Input:} Concatenated optimal digital precoder matrix $\mathbf{F}_{\mathrm{opt}}$, dictionary matrix $\mathbf{G}_T$, RF chains $\sum_{m=1}^{M}N_{\mathrm{RF},m}$, variance of approximation error $\sigma_e^2$, stopping parameters $\epsilon$ and $k_{\mathrm{max}}$\;
\textbf{Initialization:} $\hat{\gamma}_i^{(0)} = 1$, $\forall 1 \leq i \leq S \rightarrow \hat{\boldsymbol{\Gamma}}^{(0)} = \mathbf{I}_S$. Set counter $k = 0$ and $\boldsymbol{\Gamma}^{(-1)} = \mathbf{0}$\;
\While{$\left(\norm{\hat{\boldsymbol{\gamma}}^{(k)} - \hat{\boldsymbol{\gamma}}^{(k-1)}}^2 > \epsilon \ \mathrm{and} \ k < k_{\mathrm{max}}\right)$}{
\textbf{E-step:} Evaluate the \textit{a posteriori} covariance and mean $$\boldsymbol{\Gamma}^{(k)} = \left(\frac{1}{\sigma^2_e}\mathbf{G}_T^H \mathbf{G}_T + \left(\hat{\boldsymbol{\Gamma}}^{(k-1)}\right)^{-1}\right)^{-1}$$
$$\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \tilde{\mathbf{F}}_{\mathrm{BB}}^{(k)} = \frac{1}{\sigma^2_e}\boldsymbol{\Pi}^{(k)}\mathbf{G}_T^H\mathbf{F}_{\mathrm{opt}}$$ \
\textbf{M-step:} Evaluate the hyperparameter estimate\
$$\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \textbf{for} \ i = 1,\ldots, S$$
$$\hspace{1.2cm} \hat{\gamma}_{i}^{(k)} = \boldsymbol{\Pi}^{(k)}_{(i,i)} + \frac{1}{U} \sum_{j=1}^{U}\left|\tilde{\mathbf{F}}_{\mathrm{BB},n}(i,j)\right|^2 $$
\ \ \ \ \ \ \ \textbf{end}\
} \textbf{Output:} Obtain $\mathbf{F}_{\mathrm{BB}}^{*}$ and $\mathbf{F}_{\mathrm{RF}}^{*}$ using \eqref{FbbBL} and \eqref{FrfBL}.
\caption{BL-based hybrid TPC design}
\end{algorithm}
\section{Successive Multiuser Uplink Hybrid Beamforming (SCUHBF)}
Consider now the cooperative CFMU mmWave MIMO uplink system with $U$ users and $M$ cooperative APs. The $u$th user has $N_R$ TAs and $N_{\mathrm{RF},u}$ RF chains, whereas the $m$th AP is equipped with $N_T$ RAs and $N_{\mathrm{RF},m}$ RF chains. The symbol transmitted by user $u$ can be expressed as $ \mathbf{s}_u = \mathbf{F}_{\mathrm{RF},u}\mathbf{f}_{\mathrm{BB},u}x_u.$
The signal received at the $m$th AP can be expressed as
\begin{align*}
\mathbf{r}_m = \sum_{u=1}^{U}\mathbf{H}_{u,m}\mathbf{s}_u+ \boldsymbol{\delta}_m = \sum_{u=1}^{U}\mathbf{H}_{u,m}\mathbf{F}_{\mathrm{RF},u}\mathbf{f}_{\mathrm{BB},u}x_u + \boldsymbol{\delta}_m,
\end{align*}
where $\mathbf{F}_{\mathrm{RF},u}\in \mathbb{C}^{N_R \times N_{\mathrm{RF},u}}$ represents the RF TPC and $\mathbf{f}_{\mathrm{BB},u}\in \mathbb{C}^{N_{\mathrm{RF},u} \times 1}$ denotes the baseband TPC of the user $u$. Furthermore, $\mathbf{H}_{u,m} \in \mathbb{C}^{N_T \times N_R}$ denotes the channel matrix between the $m$th AP and the user $u$. Therefore, the stacked signal $\mathbf{r} = \left[\mathbf{r}_1^H, \mathbf{r}_2^H,\ldots, \mathbf{r}_M^H\right]^H \in \mathbb{C}^{MN_T \times 1} $ received at all the cooperating APs can be written as
\begin{align}
\mathbf{r} = \sum_{u=1}^{U}\mathbf{H}_u\mathbf{F}_\mathrm{{RF},u}\mathbf{f}_\mathrm{{BB},u}x_u + \boldsymbol{\delta},
\end{align}
where $\mathbf{H}_u = {[\mathbf{H}_{u,1}^H, \mathbf{H}_{u,2}^H, \ldots, \mathbf{H}_{u,M}^H]}^H \in \mathbb{C}^{MN_T \times N_R}$ denotes the aggregate channel matrix between all the APs and the $u$th user. Also, ${\boldsymbol{\delta}} \in \mathbb{C}^{MN_T \times 1}$ denotes the complex AWGN with covariance matrix $\mathop{\mathbb{E}}[\boldsymbol{\delta} \boldsymbol{\delta}^H] = \sigma_{\delta}^2 \mathbf{I}_{MN_T}$. The estimated received symbol corresponding to user $u$ after employing the hybrid RC $\mathbf{w}_u = \mathbf{W}_{\rm RF}\mathbf{w}_{\mathrm{BB},u}$ at the cooperative APs is given by
\begin{align}\label{UplinkEstimated}
\hat{x}_u =& \mathbf{w}^H_u \mathbf{r} = \mathbf{w}^H_u \mathbf{H}_u \mathbf{s}_u + \sum_{i=1,i \neq u}^{U} \mathbf{w}^H_u \mathbf{H}_i \mathbf{s}_i + \underset{\tilde{\boldsymbol{\delta}}}{\underbrace{\mathbf{w}_{u}^H \boldsymbol{\delta}}}\nonumber\\
=& \mathbf{w}^H_{\mathrm{BB},u} \mathbf{W}^H_{\mathrm{RF}} \mathbf{H}_u \mathbf{F}_{\mathrm{RF},u}\mathbf{f}_{\mathrm{BB},u}x_u \nonumber \\ +& \sum_{i=1,i \neq u}^{U} \mathbf{w}^H_{\mathrm{BB},u} \mathbf{W}^H_{\rm RF} \mathbf{H}_i \mathbf{F}_{\mathrm{RF},i}\mathbf{f}_{\mathrm{BB},i}x_i + \tilde{{\boldsymbol{\delta}}}.
\end{align}
Let us define the effective channel matrix $\mathbf{H}_{\mathrm{eff},u}$ between the user $u$ and the APs as $\mathbf{H}_{\mathrm{eff},u} = \mathbf{W}^H_{\rm RF} \mathbf{H}_u \mathbf{F}_{\mathrm{RF},u} \in \mathbb{C}^{MN_{\mathrm{RF},m} \times N_{\mathrm{RF},u}}$. From \eqref{UplinkEstimated}, the effective system model for the estimated symbol $\hat{x}_u$ is given by
\begin{equation}
\hat{x}_u = \mathbf{w}^H_{\mathrm{BB},u} \mathbf{H}_{{\rm eff},u}\mathbf{f}_\mathrm{{BB},u}x_u + \sum_{i=1,i \neq u}^{U}\mathbf{w}^H_{\mathrm{BB},u} \mathbf{H}_{{\rm eff},i}\mathbf{f}_{\mathrm{BB},i}x_i + \tilde{{\boldsymbol{\delta}}}.
\label{Eq7MRC}
\end{equation}
\begin{figure*}[t]
\centering
\subfloat[]{\includegraphics[width=0.45\linewidth]{Figure2a}\label{Broadcast}}
\subfloat[]{\includegraphics[width=0.45\linewidth]{Figure2b}\label{BroadcastLargeAntenna}}
\caption{Capacity comparison of various schemes for a broadcast scenario with(a) $M = 4$, $N_T = 16$, $N_R = 8$, $N_{\mathrm{RF},m} = 4$, and $U = 10$, (b) $M = 4$, $N_T = 64$, $N_R = 8$, $N_{\mathrm{RF},m} = 4$, and $U = 15$. \vspace{-10pt}}\label{Broadcast_Main_Fig}
\end{figure*}
The baseband RC designed for the $u$th user is given by $\mathbf{w}_{\mathrm{BB},u} = \mathbf{H}_{{\rm eff},u}\mathbf{f}_{\mathrm{BB},u}$. It can be observed from \eqref{Eq7MRC} that the term $\mathbf{f}_{\mathrm{BB},u}^H \left(\mathbf{H}_{\mathrm{eff},u}^H \mathbf{H}_{\mathrm{eff},u}\right) \mathbf{f}_{\mathrm{BB},u}$ represents the signal power of the $u$th user in the estimate $\hat{x}_u$ and the term $\sum_{i=1,i \neq u}^{U} | \mathbf{f}_{\mathrm{BB},u}^H\mathbf{H}_{\mathrm{eff},u}^H \mathbf{H}_{\mathrm{eff},i} \mathbf{f}_{\mathrm{BB},i}|^2$ denotes the interference part. In the cooperative uplink scenario, the objective is to design the uplink hybrid TPCs for ensuring that they maximize the SINR of each user in the system. Hence, the pertinent optimization problem is given by
\begin{align}
& \underset{\mathbf{f}_{\mathrm{BB},u}}{\mathrm{maximize}}\quad \mathbf{f}_{\mathrm{BB},u}^H \left(\mathbf{H}_{\mathrm{eff},u}^H \mathbf{H}_{\mathrm{eff},u}\right) \mathbf{f}_{\mathrm{BB},u} \nonumber \\ & \text{s.t.}\quad\ \left(\mathbf{f}_{\mathrm{BB},i}^H \mathbf{H}_{\mathrm{eff},i}^H \mathbf{H}_{\mathrm{eff},u}\right) \mathbf{f}_{\mathrm{BB},u} =\mbox{0},\ i = \mbox{1},\mbox{2},\ldots, u-\mbox{1}, \nonumber \\ &\qquad\quad \|\mathbf{W}_{\rm RF}\mathbf{w}_{\mathrm{BB},u} \|_2^2 \leq P.
\label{UplinkOpt}
\end{align}
From the first constraint in \eqref{UplinkOpt}, it can be readily observed that if the baseband TPC $\mathbf{f}_{\mathrm{BB},u}$ is designed so that any user $i$, $i < u$, does not face interference due to transmission of user $u$, then the same holds vice versa, i.e., the transmission of user $i$ does not create any interference for user $u$, $u > i$. Hence, one can deduce that the interference cancellation process is symmetric in nature. Furthermore, it can also be deduced that in the successive CFMU mmWave uplink system there is no interference among the scheduled users, which is in contrast to the successive downlink scenario. The solution to the optimization problem in \eqref{UplinkOpt} can be achieved by solving the following problem:
\begin{align}
& \underset{\mathbf{b}_u}{\mathrm{maximize}}\quad \mathbf{b}_u^H \left[\left(\mathbf{M}_{u-1}^{\perp}\right)^H \mathbf{H}_{{\rm eff},u}^H \mathbf{H}_{{\rm eff},u} \mathbf{M}_{u-1}^{\perp} \right] \mathbf{b}_u\nonumber \\ & \ \ \ \ \ \ \text{s.t.} \ \ \qquad \| \mathbf{b}_u \|_2^2 \leq P,
\label{Eq 9}
\end{align}
where $\mathbf{b}_u \in \mathbb{C}^{(N_{\mathrm{RF},u}-u+1)\times 1}$, and $\mathbf{M}_{u-1}^{\perp} \in \mathbb{C}^{N_{\mathrm{RF},u} \times (N_{\mathrm{RF},u}-u+1)}$ denotes a basis for the null space of the matrix $\mathbf{M}_{u-1} \in \mathbb{C}^{(u-1) \times N_{\mathrm{RF},u}}$, which is given as
\begin{equation*}
\mathbf{M}_{u-1} = \left[\begin{array}{c}\mathbf{f}_{{\rm BB},1}^H \mathbf{H}_{{\rm eff},1}^H \\ \mathbf{f}_{{\rm BB},2}^H \mathbf{H}_{{\rm eff},2}^H \\ \vdots \\ \mathbf{f}_{{\rm BB},u-1}^H \mathbf{H}_{{\rm eff},u-1}^H\end{array} \right] \mathbf{H}_{{\rm eff},u},
\end{equation*}
and can be obtained from the SVD of $\mathbf{M}_{u-1}$. Therefore, the uplink baseband TPC designed for the $u$th user is given as $\mathbf{f}_{\mathrm{BB},u} = \sqrt{P}\mathbf{M}_{u-1}^{\perp}\mathbf{b}_u^{*}$, where $\mathbf{b}_u^{*}$ denotes the dominant eigenvector of the matrix $(\mathbf{M}_{u-1}^{\perp})^H \mathbf{H}_{{\rm eff},u}^H \mathbf{H}_{{\rm eff},u} \mathbf{M}_{u-1}^{\perp}$. Furthermore, the rate $R_u$ of the $u$th user is given as $R_u = \log_2 (\mbox{1} + {\| \mathbf{H}_{{\rm eff},u} \mathbf{M}_{u-1}^{\perp} \mathbf{b}_{u}^{*} \|}_2^2)$, and the capacity $C_{\mathrm{sum}}$ in the uplink scenario is given as $C_{\mathrm{sum}} = \sum_{u=1}^{U}{R_u}$. Hence, one can notice that the proposed SCUHBF scheme maximizes each user's SINR, while successfully cancelling the IUI.
\begin{figure*}[t]
\centering
\subfloat[]{\includegraphics[width=0.45\linewidth]{Figure3a}\label{4Ua}}
\subfloat[]{\includegraphics[width=0.45\linewidth]{Figure3b}\label{4Ub}}
\caption{Unicast Scenario: (a) Capacity performance of the OSHB-U with various schemes for $N_T = 16$, $N_R = 8$, $N_{\mathrm{RF},m} = 4$, and $U = 4$, (b) capacity comparison of the fully-digital design and BL-based hybrid beamformer design for $N_T = 32$ and $N_R = 8$ antennas.\vspace{-15pt}}\label{UnicastFig}
\end{figure*}
\section{Simulation Results}
In this section, we provide simulation results for illustrating the performance of the proposed HTBFs in various scenarios. In the cell-free multi-user network considered, the $M$ APs and $U$ users are considered to be uniformly distributed at random within a square coverage area of size $200 \times 200$ $m^2$. The mmWave MIMO channel $\mathbf{H}_{u,m}$ between each user and the AP has been modelled using \eqref{Channel} with $L = 6$ multipath components. The AoA/AoDs of the channels between the APs and users are generated with uniform distribution over $\left[0 \ \ 2\pi\right]$. Furthermore, the multipath gains $\alpha_{u,m}^l$ are assumed to be complex Gaussian distributed as $\mathcal{C}\mathcal{N}\left(0,1\right)$. For the BL-based precoder and combiner design, the initial value of all the hyper-parameters is set to $\hat{\gamma}_i^{(0)} = 1, \forall 1 \leq i \leq S$, with the maximum number of EM iterations and the stopping parameter set as $k_{\mathrm{max}} = 50$ and $\epsilon = 10^{-6}$, respectively. We consider $3000$ such channel realizations for plotting the capacity $R_{\mathrm{sum}}$ versus the total power $P_T$ (dB) available at the APs. \vspace{-10pt}
\subsection{Beamforming for broadcast scenarios: HBBF}
For a broadcast scenario, a $\left(16 \times 8\right)$ elements mmWave MIMO system is considered, where we have $N_T = 16$ TAs at AP and $N_R = 8$ RAs at user, $M = 4$ APs and $U = 8$ users. For HTBF, we assume that the number of RF chains at each AP is equal to the number of users i.e. $N_{\mathrm{RF},m} = U$, and number of RF chains at each user is set equal to the total number of cooperative APs, i.e. $N_{\mathrm{RF},m} = M$. Figure \ref{Broadcast} illustrates the performance of the HBBF and HBBF-P schemes described in Section \ref{HBB}. It can be observed that the hybrid design of the proposed HBBF schemes closely approaches the performance of an ideal fully-digital beamformer (FDBF), where the number of RF chains is equal to the number of antennas. Furthermore, it can also be observed that the fairness-based HBBF-P scheme associated with per-AP power constraints closely follows the performance of the HBBF, given a pooled power constraint. Thus, the proposed HBBF-aided hybrid TPC design results in the optimal solution, and it is also well-suited for transmission in broadcast CFMU cooperative mmWave MIMO scenarios. Fig. \ref{BroadcastLargeAntenna} illustrates the performance of the HBBF scheme in a large-scale antenna array regime. It can be observed that the capacity increases upon increasing the number of TAs at the AP, thanks to the array gain of the large antenna arrays. This further establishes the importance of having a large number of antennas in the mmWave regime for improved power efficiency.\vspace{-10pt}
\begin{figure*}
\centering
\subfloat[]{\includegraphics[height=5cm, width=8cm]{Figure4a}\label{UPA_ULAFig}}
\subfloat[]{\includegraphics[height=5cm, width=8cm]{Figure4b}\label{UplinkFig}}
\caption{Unicast scenario: (a) Capacity versus number of users with $M = 4$, (b) Uplink scenario: capacity versus total power with $M = 4$, $N_T = 16$, $N_R = 8$, $G = 2$ and $U = 4$. \vspace{-10pt}}
\end{figure*}
\begin{figure*}[t]
\centering
\subfloat[]{\includegraphics[width=0.45\linewidth]{Figure5a}\label{MulticastFig}}
\subfloat[]{\includegraphics[width=0.45\linewidth]{Figure5b}\label{MulticastLargeAntenna}}
\caption{Unicast Scenario: (a) Capacity versus total power for a multicast scenario with $M = 4$, $N_T = 16$, $N_R = 8$, $G = 2$ and $U_g = 4$, (b) capacity versus total power for a multicast scenario $M = 4$, $N_R = 8$, $G = 2$ and $U_g = 4$. \vspace{-10pt}}
\end{figure*}
\subsection{Beamforming schemes for downlink and uplink unicast scenarios: OSHB-U and SCUHBF}
Next, we consider both downlink and uplink CFMU mmWave MIMO unicast scenarios in the face of interuser interference and compare the capacity performance of various schemes. Fig. \ref{4Ua} shows the capacity performance of the OSHB-U in a downlink system along with those of the other schemes for $M = 4$ APs and $U = 4$ users with $N_T = 16$, $N_R = 8$, $N_{\mathrm{RF},m} = 4$ and $N_{\mathrm{RF},u} = M$. It can be observed that the OSHB-U scheme performs similarly to the ideal fully-digital beamformer. The proposed OSHB-U scheme achieves an improved performance in terms of capacity as a benefit of the SIC and the maximization of the SINR by the MVDR-based beamformer. For multi-cell cooperative scenarios, one can also employ the sub-optimal successive constrained eigen hybrid beamforming (SCEHBF) \cite{6607449} scheme based on MRC, which was originally derived for a sub-6 GHz system and has been suitably extended for mmWave for fair comparison. Its performance is poor, since it disregards the interference inflicted by the previously scheduled users. Furthermore, the simplistic white power allocation (WPA) \cite{6607449} scheme using MRC results in a significant performance degradation because it fails to cancel the interference.
Fig. \ref{4Ub} compares the capacity of the BL-based HTBF design with that of the ideal fully-digital beamformer. The BL-based HBF design does not require the full knowledge of AoAs/AoDs and their path gains for designing the baseband and RF precoders. It can be observed that the BL-based design attains a performance close to that of the fully-digital design with fewer RF chains. This is attributed to the fact that the mmWave MIMO channel is comprised of fewer multipath components, hence resulting in a low-rank channel. Therefore, the ideal FDBF can be approximated using a linear combination of only a few transmit array response vectors.
Fig. \ref{UPA_ULAFig} plots the capacity versus the number of users considering both the uniform planar array (UPA) and the uniform linear array (ULA) structures. Interestingly, it can be seen that the capacity increases upon increasing the number of AP antennas, which can be attributed to the fact that the CFMU mmWave MIMO system and the proposed design technique is capable of efficiently exploiting the resultant increased array gain. Furthermore, one can also observe that the proposed OSHBF-U algorithm outperforms the SCEHBF, when serving different number of users. To be specific, as the number of served users increases, the gap between the proposed and benchmark algorithms progressively increases, which further confirms the stability and effectiveness of our proposed algorithm.
Fig. \ref{UplinkFig} plots the capacity versus transmit power for the uplink scenario. For this analysis, $U=4$ active users are equipped with $N_R = 8$ TAs, and assumed to be distributed in $M = 4$ cooperating cells with each AP having $N_T = 16$ RAs. It can be observed that the proposed SCUHBF scheme performs close to the capacity of $U$ interference-free point-to-point mmWave MIMO systems having dimensions of $\sum_{m=}^{M}N_{\mathrm{RF},m} \times N_{\mathrm{RF},u}$, and it also closely matches the performance of the ideal FD design. Furthermore, it is also observed that the capacity achieved by the proposed SCUHBF scheme for a cooperative uplink scenario is much higher than that of employing maximum ratio transmit (MRT) beamforming at the users and MRC at the APs.
\subsection{Beamforming for multicast scenarios: OSHB-M}
Fig. \ref{MulticastFig} compares the capacity performance of the OSHB-M scheme in the face of inter-group interference for different values of the number of RF chains, $M = 4$ APs and $U_g = 4$ users having $N_T = 16$, $N_R = 8$ antennas, and $G = 2$ user groups. It can be observed that the performance of the proposed OSHB-M scheme closely follows that of the ideal fully-digital TBF design, but with a much lower number of RF chains at the APs. Next, Fig. \ref{MulticastLargeAntenna} demonstrates the large-scale antenna benefits of the mmWave regime. It can be observed that the performance of the OSHB-M schemes improve significantly upon increasing the number of antennas at the APs. \vspace{-10pt}
\section{Conclusion \vspace{-8pt}}
Optimal hybrid beamformers for broadcast, unicast and multicast CFMU mmWave MIMO systems are designed in the face of interuser/intergroup interference. Initially, for a broadcast scenario, the optimal beamforming solutions were derived by considering a total power constraint as well as per-AP power constraints. Subsequently, OSHB-based techniques relying on the optimal MVDR principle were proposed for unicast and multicast scenarios, which efficiently cancel the resultant interuser and intergroup interference. Next, BL-based joint hybrid transceiver was also design which did not require the knowledge of the AoAs/AoDs and the gains of the various multipath components of the mmWave MIMO channels. Finally, we also derived the SCUHBF-based cooperative uplink HBF, which maximizes the SINR, hence achieving high data rates. Our simulation results demonstrated the improved spectral efficiency as well as interference cancellation capability of the proposed HBBF, OSHB and SCUHBF techniques. The proposed hybrid designs have a low-complexity, and can simultaneously support an increased number of users in comparison to the existing methods, which makes them eminently suitable for practical mmWave CFMU MIMO systems. \vspace{-10pt}
\begin{appendices}
\section{MVDR Beamformer}
Considering the system model of \eqref{eq 16}, the signal $\tilde{\mathbf{y}}_u$ received at user $u$ after RF combining can be written as
\begin{align*}
\tilde{\mathbf{y}}_u = \mathbf{H}_{\mathrm{eff},u} \mathbf{M}_{u-1}^{\perp} \boldsymbol{\nu}_{u} x_u + \boldsymbol{\tau}_u,
\end{align*}
where $\boldsymbol{\tau}_u$ represents the interference-plus-noise component associated with the covariance matrix $\mathbf{C}_{\boldsymbol{\tau}_u}$. After baseband combining at user $u$, the signal estimate $\hat{x}_u$ is given as
\begin{align*}
\hat{x}_u = \mathbf{w}_{\text{BB},u}^H \mathbf{H}_{\mathrm{eff},u} \mathbf{M}_{u-1}^{\perp} \boldsymbol{\nu}_{u} x_u + \mathbf{w}_{\text{BB},u}^H \boldsymbol{\tau}_u.
\end{align*}
Therefore, the SINR at user $u$ is given by
\begin{align*}
\mathrm{SINR}_u = \frac{\vert \mathbf{w}_{\text{BB},u}^H \mathbf{H}_{\mathrm{eff},u} \mathbf{M}_{u-1}^{\perp} \boldsymbol{\nu}_{u} \vert^2}{\mathbf{w}_{\text{BB},u}^H \mathbf{C}_{\boldsymbol{\tau}_u}\mathbf{w}_{\text{BB},u}}.
\end{align*}
The MVDR beamformer is the optimal beamformer that maximizes the SINR \cite{van2002modulation}, and it is given as $\mathbf{w}_{\text{BB},u} = \mathbf{C}^{-1}_{\boldsymbol{\tau}_u} \mathbf{H}_{\mathrm{eff},u} \mathbf{M}_{u-1}^{\perp} \boldsymbol{\nu}_{u}$. Furthermore, the maximum value of the SINR is determined as
\begin{align*}
\mathrm{SINR}_{\mathrm{max},u}=&\, \frac{\vert \boldsymbol{\nu}_{u}^H \mathbf{M}_{u-1}^{\perp H} \mathbf{H}_{\mathrm{eff},u}^H \mathbf{C}^{-1}_{\boldsymbol{\tau}_u} \mathbf{H}_{\mathrm{eff},u} \mathbf{M}_{u-1}^{\perp} \boldsymbol{\nu}_{u} \vert ^2}{\boldsymbol{\nu}_{u}^H \mathbf{M}_{u-1}^{\perp H} \mathbf{H}_{\mathrm{eff},u}^H \mathbf{C}^{-1}_{\boldsymbol{\tau}_u} \mathbf{H}_{\mathrm{eff},u} \mathbf{M}_{u-1}^{\perp} \boldsymbol{\nu}_{u}} \\ =&\, \boldsymbol{\nu}_{u}^H \mathbf{M}_{k-1}^{\perp H} \mathbf{H}_{\mathrm{eff},u}^H \mathbf{C}^{-1}_{\boldsymbol{\tau}_u} \mathbf{H}_{\mathrm{eff},u} \mathbf{M}_{u-1}^{\perp} \boldsymbol{\nu}_{u}.
\end{align*}
Observe from the above expression that the optimal vector $\boldsymbol{\nu}_{u}$ that maximizes the $\mathrm{SINR}_{\mathrm{max}, u}$ is the principal right singular vector of $\mathbf{C}^{-(1/2)}_{\boldsymbol{\tau}_u} \mathbf{H}_{\mathrm{eff},u} \mathbf{M}_{u-1}^{\perp}$. Therefore, the optimal MVDR receive beamformer $\mathbf{w}_{\mathrm{BB},u}$ is given by $\mathbf{w}_{\mathrm{BB},u} = {(\mathbf{C}_{\boldsymbol{\tau}_{u}}^{-(1/2)})}^H \tilde{\mathbf{w}}_{\mathrm{BB},u}$, where $\tilde{\mathbf{w}}_{\mathrm{BB},u}$ is the dominant left singular vector
of the matrix $\mathbf{C}^{-(1/2)}_{\boldsymbol{\tau}_u} \mathbf{H}_{\mathrm{eff},u} \mathbf{M}_{u-1}^{\perp}$. This completes the proof. \vspace{-12pt}
\end{appendices}
\bibliographystyle{ieeetran}
|
\section{Implementation details} \label{sec: implementation}
\subsection{Training Procedure} \label{sec: training_details}
The overall training procedure is shown in algorithm \ref{alg:train}. We use the weighted version of our nonparametric representation during training, as displayed in lines 3 - 9. Another option is to use standard-sampling representation (see Sec. \ref{sec:model}) and utilize the Gumbel-Softmax trick~\citep{jang2016categorical} to enable backpropagation when drawing expert id $k$ from the categorical distribution $\mathcal{G}(\boldsymbol{\pi})$.
\begin{algorithm}
\caption{Training Procedure}\label{alg:train}
\begin{algorithmic}[1]
\Require{Training data $\mathcal{D}$ with $M$ ground truth labels per image, number of experts $K$, number of samples per expert $S$, batch size $B$, leaning rate $\eta$, loss hyperparameters $\beta, \gamma$.}
\Require{initial expert network parameters $\theta_s$, initial gating network parameters $\theta_{\pi}$. }
\While{$\theta_s, \theta_{\pi}$ not converged}
\State Sample $\{x_n,y^{(j)}_n, v^{(j)}_n\}_{j=1,\cdots,M}^{n=1,\cdots,B} $ a batch from the training dataset $\mathcal{D}$.
\For{$k = 1,\cdots, K$}
\For{$t = 1,\cdots, S$}
\State Sample $\{z_n^{(k,t)}\}^{n=1,\cdots,B} \sim P_k$ a batch of expert prior codes.
\State $s_n^{(k,t)}\leftarrow F(x,z_n^{(k,t)} ; \theta_s)$ with ${n=1,\cdots,B}$.
\State $u_n^{(k,t)}\leftarrow \frac{1}{S}\pi_k(x ; \theta_{\pi})$ with ${n=1,\cdots,B}$.
\EndFor
\EndFor
\State Rearrange all outputs as $\{s_n^{(i)}, u_n^{(i)}\}_{i=1,\cdots,N}$, where $N = K\times S$.
\State Compute pairwise cost matrix $\mathbf{C}_n$, with $\mathbf{C}_n[i,j] = c({s_n^{(i)}},{y_n^{(j)}})$, ${n=1,\cdots,B}$.
\State Solve the optimal coupling matrix $\mathbf{P}^*_n$ = LP\_SOLVER($\mathbf{C}_n, \mathbf{u}_n, \mathbf{v}_n, \gamma$), ${n=1,\cdots,B}$.
\State $\ell \leftarrow\frac{1}{B}\sum_{n=1}^B\left(\langle\mathbf{P}^*_{n},\mathbf{C}_n(\theta_s)\rangle + \beta KL(\mathbf{P}^*_{n}\mathbbm{1}_M || \mathbf{u}_n(\theta_{\pi})) \right)$
\State {$\theta_s, \theta_{\pi} \leftarrow \text{Adam}({\nabla}^{\text{soft}}\ell , \eta)$}
\EndWhile
\end{algorithmic}
\end{algorithm}
\paragraph{Details on solving the coupling matrix}
In each iteration of training, we need to solve the best coupling matrix $\mathbf{P^*}$. This corresponds to the LP\_SOLVER function module in algorithm \ref{alg:train}~ and is formally defined as:
\begin{equation}
\mathbf{P^*} =\mathop{\arg\min}\limits_{\mathbf{P} \in \mathbf {U'}} \sum_{i,j}\mathbf{P}_{ij}\mathbf{C}_{ij} \quad \mathbf {U'} \coloneqq \{\mathbf{P} \in \mathbbm{R}_{+}^{N\times M}: \mathbf{P}\mathbbm{1}_M\le \gamma,\mathbf{P}^T\mathbbm{1}_N = \mathbf{v}_n\}
\end{equation}
Especially, when $\gamma = 1$, the corresponding inequality constraint term vanished and $\mathbf {U'} \coloneqq \{\mathbf{P} \in \mathbbm{R}_{+}^{N\times M}: \mathbf{P}^T\mathbbm{1}_N = \mathbf{v}_n\}$.
{We can then solve the problem by moving the mass of each ground truth label to its nearest prediction.} The time complexity is $O(N)$.
When $\gamma < 1$, the problem can be solved by a standard linear programming solver like simplex or interior-point~\citep{ignizio1994linear}.
We adopt a greedy strategy here, which empirically provides similar performance but with lower time complexity. Specifically, we first sort the ground truth labels according to their minimum cost to the predicted sample in ascending order. For each ground truth label, we also sort the predicted samples in terms of the corresponding cost in ascending order.
Then, we go through the sorted ground truth labels and transfer their mass to the predicted samples in the sorted order given the constraint $\mathbf{P}\mathbbm{1}_M\le \gamma$ is satisfied. This algorithm has a time complexity of $O(NMlog(N))$.
\paragraph{Details on computing Soft Gradient}
{We observe that optimizing the loss (cf. Eq.~\ref{eq:relax}) via a direct gradient descent could sometimes suffer from slow convergence and low performance due to large variance in the target signal $\sum_j\mathbf{P}^*_{ij}$ in the second term. This is partially caused by the random samples generated by the stochastic experts and unstable $\mathbf{P}^*$ inferred during the learning process.}
{To address this, we adopt a simple smoothing technique to produce a stable gradient. Specifically, we group those experts with similar outputs during training, and average the gradients of the experts within each group. This prevents the gradients from changing dramatically in consecutive steps and resembles the Straight-Through method~\citep{jang2016categorical, bengio2013estimating}. }
We show the details for computing soft gradients in Algorithm \ref{alg:gradient} . {Groups $g_1, \cdots, g_G$ are sets of sample indices, and can be computed by a clustering method.} We here take a simple strategy and merge experts with similar outputs. Specifically, we initialize each group as an expert, and merge two experts $k, k'$ if IoU$(s^k, s^{k'}) \ge G_0$, where $s^k \coloneqq \frac{1}{S}\sum_t s^{k,t}$ is the mean of samples for expert $ k\in[K]$, $G_0$ is a hyper-parameter where we initialize as 1 and linearly decreased during training. Notice that we mainly apply this soft gradient operation on the Cityscapes dataset, where multiple experts generate similar predictions.
\begin{algorithm}
\caption{Compute Soft Gradient}\label{alg:gradient}
\hspace*{\algorithmicindent} \textbf{Input:} loss function $\ell$, model outputs $\{s ^{(i)}, u ^{(i)}\}_{i=1}^{N}$, coupling matrix $\mathbf{P}^* $, hyperparameter $G_0$.\\
\hspace*{\algorithmicindent} \textbf{Output:} soft gradients ${\nabla}^{\text{soft}}_{\mathbf{u} }\ell$.
\begin{algorithmic}[1]
\Procedure{SoftGradient}{}
\State $g_1, \cdots, g_G \leftarrow \text{Clustering}(\mathbf{s} , G_0)$
\State $\mathbf{o} \leftarrow \mathbf{P} ^*\mathbbm{1}_M$
\For{$m \in 1 \cdots G $}
\State ${\bar{u}}^{(m)} \leftarrow \frac{1}{|g_m|}\sum_{i\in {g_m}}u^{(i)} $
\State ${\bar{o}}^{(m)} \leftarrow \frac{1}{|g_m|}\sum_{i\in g_m}o^{(i)} $
\State {${\nabla}^{\text{soft}}_{u ^{(i)}}\ell(u ^{(i)},o ^{(i)}) \leftarrow {\nabla}_{{\bar{u}}^{(m)}}\ell({\bar{u}}^{(m)},{\bar{o}}^{(m)}) \quad (i\in g_m)$}
\EndFor
\State\Return ${\nabla}^{\text{soft}}_{\mathbf{u} }\ell$
\EndProcedure
\end{algorithmic}
\end{algorithm}
\paragraph{Details on Network Architechtures}
For the backbone, we use the same encoder-decoder architecture as in U-net~\citep{ronneberger2015u} and follow the model design as~\citep{baumgartner2019phiseg}. Specifically, there are 6 down- and upsampling blocks, with each block consisting of three convolutional layers, each followed by a batch normalization layer and a ReLU-activation\footnote{Other details on backbone architecture can be referred to~\cite{baumgartner2019phiseg} and their code link \url{https://github.com/baumgach/PHiSeg-code}.}.
The global feature map $u_g\in R^{F_1\times H'\times W'}$ has feature dimension $F_1 = 192$. The dense feature map $u_d \in R^{F_2\times H\times W}$ has feature dimension $F_2 = 32.$ The gating network contains three-layer MLP, with the first two layers having 192 neurons and the last layer having K neurons (K is the expert number). The three subsequent 1 x 1 convolution blocks after the decoder have 32 channels for the first two blocks and C channels for the last block ( C is the number of classes).
\paragraph{Other Training Details}
On the LIDC dataset, we apply random flip and random rotation augmentations same as~\cite{baumgartner2019phiseg, NEURIPS2020_95f8d990}.
We use the latent code with dimension 1, which is empirically found to perform well and can be further refined.
The models are trained for 200 epochs with batch size 16.
On the Cityscapes dataset, we apply random horizontal flips, cropping, and scaling to size 128×128 same as~\cite{kassapis2021calibrated}. We use a batch size of 12, latent code with dimension 5, and train the models for 1000 epochs. {The ground truth marginal probabilities $\mathbf{v}$ (cf. \ref{eq:relax}) are initialized in uniform and annealing to the original values after some warm-up epochs.}
The additional prediction inputs are from the pretrained model xception71\_dpc\_cityscapes\_trainfine provided by DeepLab Model Zoo \footnote{Other details on this additional input can be referred to in~\cite{kassapis2021calibrated} and their public code: \url{https://github.com/EliasKassapis/CARSSS}.}.
As commonly practiced~\citep{kassapis2021calibrated, NEURIPS2018_473447ac}, we ignore the unlabeled pixels when calculating loss. On both datasets, we use weight-decay with weight $1e^{-5}$ and Adam optimizer with an initial learning rate of $1e^{-3}$. {The learning rates and loss slack factor $\gamma$ are linearly updated with scheduled updates.}
We do the model selection and tune hyperparameters according to the GED score on the validation set.
We use two NVIDIA TITAN RTX GPUs on the LIDC dataset and four NVIDIA TITAN RTX GPUs on the Cityscapes dataset.
\subsection{Evaluation Metrics} \label{sec: evaluation_metrics}
{We aim to measure how well the model captures the multimodal aleatoric uncertainty in segmentation. The aleatoric uncertainty refers to the irreducible randomness in the data-generating process and can be represented by the ground truth label distribution~\citep{NEURIPS2020_95f8d990, hullermeier2021aleatoric}. Following the literature~\citep{NEURIPS2018_473447ac,kohl2019hierarchical,baumgartner2019phiseg,NEURIPS2020_95f8d990,kassapis2021calibrated}, we use two distributional metrics: \textit{generalized energy distance} (GED) and \textit{matched-IoU} (M-IoU) to evaluate how the model fits the ground truth label distribution in the segmentation output space, which also demonstrates the quality of aleatoric uncertainty quantification. Due to the complexity of the high-dimensional segmentation distribution, both metrics use the sampling strategy to approximate the distributional distance via two sets of samples. In complementary {to the sample-wise metrics}, we also adopt the \textit{expected calibration error} (ECE)~\citep{guo2017calibration} as an additional metric to measure pixel-wise uncertainty calibration. We note that while the ECE measures the quality of more general predictive uncertainty, our problem setting is dominated by the aleatoric uncertainty (cf. \ref{sec: uncertainty_type_lidc_results} and \ref{sec: uncertainty_type_city_results}). Besides, as it is mainly used in the calibration of classification problems, the ECE can not measure the important multimodal and structural characteristics of segmentation uncertainty, and thus is adopted as a secondary metric.
{Below, we summarize the details of these three metrics. For clarity, we denote $\hat{y}$ as the hard prediction (after 'argmax' operation) to distinguish the soft prediction $s$ (after 'softmax' operation) . Other notations are kept consistent with previous content.}
\textbf{{Generalized energy distance (GED)}}~\citep{szekely2013energy} measures the distance between two distributions by calculating the pair-wise distance $d(\cdot)$ = 1- IoU of samples from predictive distribution and samples from ground truth distribution. The expected value is then subtracted with the sample diversity of each distribution to get the final score:
\begin{equation}\label{eq: ged}
\mathcal{D}^2_{GED}(\mu,\nu) \coloneqq 2E_{\hat{y}\sim \mu, y\sim \nu}[d(\hat{y},y)] - E_{y,y'\sim \nu}[d(y,y')] - E_{\hat{y},\hat{y}'\sim {\mu}}[d(\hat{y},\hat{y}')].
\end{equation}
This metric is widely adopted by the previous literature~\citep{NEURIPS2018_473447ac,kohl2019hierarchical,baumgartner2019phiseg,NEURIPS2020_95f8d990,kassapis2021calibrated}.
{While effective in most situations, this metric potentially rewards sample diversity and sometimes gives a biased estimation of distance~\citep{kohl2019hierarchical}.}
\textbf{Matched-IoU (M-IoU)}~\citep{kuhn1955hungarian, cunningham1976network} measures the distributional distance by solving an optimal matching matrix $\mathbf{P}$ between samples from the predictive distribution and ground truth distribution:
\begin{equation}\label{eq: m-iou}
\mathcal{W}_{IoU}(\mu,\nu) \coloneqq \max_{\mathbf{P}}\sum_{i,j}\mathbf{P}_{i,j}\text{IoU}(\hat{y}_i,y_j) \quad \text{s.t. }\mathbf{P}\mathbbm{1}_M = \mathbf{u}, \mathbf{P}^T\mathbbm{1}_N = \mathbf{v}.
\end{equation}
It is previously adopted by~\cite{kohl2019hierarchical,kassapis2021calibrated} as compensation for GED measure {since this metric does not favor sample diversity.}
We here use an extended version to enable a measure when one of the marginal distributions is not uniform.
Specifically, on the LIDC dataset and evaluated with standard-sampling representation, we adopt the Hungarian matching algorithm~\citep{kuhn1955hungarian} to solve matching matrix $\mathbf{P}$ same as~\cite{kohl2019hierarchical,kassapis2021calibrated}. When evaluated with our compact representation, we adopt the network simplex algorithm~\citep{cunningham1976network} to solve $\mathbf{P}$. On the Cityscapes dataset where ground truth labels are aligned with non-uniform probabilities, we also adopt the network simplex algorithm.
{\textbf{{Expected calibration error (ECE)}}~\citep{guo2017calibration} measures the difference between probability prediction and the real accuracy. It is defined as:
\begin{equation}\label{eq: m-iou}
\text{ECE} \coloneqq E_{\hat{P}}\left[\left|P(\hat{Y} = Y | \hat{P} = p) - p\right|\right]
\end{equation}
Here, $Y$ is the random variable for the label, $\hat{Y}$ is the class prediction, and $\hat{P}$ is the associated probability.
In our case, we get the pixel-wise label distribution and predictive distribution by marginalization and treat each pixel i.i.d.
For the binary case in the LIDC dataset, we measure how the foreground class is calibrated. For the multi-class case in the Cityscapes dataset, we measure how the most-likely class is calibrated (aka confidence calibration)~\citep{guo2017calibration, kull2019beyond}.}
{On the Cityscapes dataset, we follow the convention~\citep{kohl2019hierarchical,kassapis2021calibrated} to only calculate metrics of the ten switchable classes.}
\section{Additional experiments on LIDC}
\subsection{Complementary Results}\label{sec: compl_results}
In complementary to the results shown in the main text, we show the performance of our method with M-IoU and ECE under sample-number of 50, 100 in Table \ref{tab:lidc_ours}, and the complementary ablation study under the sample number of 50, 100 in Table \ref{tab:abla_50_100}. For short, we name each experiment by its replaced term and refer to the comprehensive explanation in Sec. \ref{sec: ablation}.
\begin{table}[h]
\centering
\caption{Other Quantitative results on LIDC-IDRI dataset. {The ECE is in \%.}}
\label{tab:lidc_ours}
\resizebox{\textwidth}{!}{%
\setlength{\tabcolsep}{6pt}
\renewcommand{\arraystretch}{1.1}
\begin{tabular}{llllll}
\toprule[1pt]\midrule[0.3pt]
Method & {\# labels} & {M-IoU $\uparrow$(50)} & {M-IoU $\uparrow$(100)} & {ECE$\downarrow$(50)} & {ECE$\downarrow$(100)} \\
\midrule
{Ours} & \multirow{2}{*}{All} & {0.631 $\pm$ 0.002} & {0.636 $\pm$ 0.002} & {0.053 $\pm$ 0.018} & {0.052 $\pm$ 0.017} \\
{Ours - compact} & & {0.637 $\pm$ 0.003} & {0.638 $\pm$ 0.002} & {0.049 $\pm$ 0.014} & {0.049 $\pm$ 0.014} \\
\midrule
{Ours} & \multirow{2}{*}{One} & {0.604 $\pm$ 0.003} & {0.606 $\pm$ 0.003} & {0.089 $\pm$ 0.010} & {0.085 $\pm$ 0.012} \\
{Ours - compact} & & {0.607 $\pm$ 0.003} & {0.607 $\pm$ 0.003} & {0.084 $\pm$ 0.011} & {0.083 $\pm$ 0.012} \\
\midrule[0.3pt]\bottomrule[1pt]
\end{tabular}%
}
\end{table}
\begin{table}[h]
\centering
\caption{Complementary results of the ablation study with sample number 50, 100. The ECE is in \%.}
\label{tab:abla_50_100}
\resizebox{\textwidth}{!}{%
\setlength{\tabcolsep}{6pt}
\renewcommand{\arraystretch}{1.1}
\begin{tabular}{lllllll}
\toprule[1pt]\midrule[0.3pt]
Method & GED $\downarrow$ (50) & GED $\downarrow$ (100) & M-IoU $\uparrow$ (50) & M-IoU $\uparrow$ (100) & ECE $\downarrow$ (50) & ECE $\downarrow$ (100) \\ \midrule
IoU loss & 0.533 $\pm$ 0.001 & 0.533 $\pm$ 0.001 & 0.533 $\pm$ 0.001 & 0.533 $\pm$ 0.001 & 0.277 $\pm$ 0.016 & 0.277 $\pm$ 0.016 \\
Uniform expert weights & 0.256 $\pm$ 0.003 & 0.249 $\pm$ 0.003 & 0.554 $\pm$ 0.006 & 0.556 $\pm$ 0.007 & 0.209 $\pm$ 0.008 & 0.209 $\pm$ 0.009 \\
Deterministic Expert & 0.221 $\pm$ 0.004 & 0.216 $\pm$ 0.004 & 0.601 $\pm$ 0.008 & 0.603 $\pm$ 0.009 & 0.127 $\pm$ 0.005 & 0.121 $\pm$ 0.003 \\\midrule[0.3pt]
Ours & 0.195 $\pm$ 0.002 & 0.189 $\pm$ 0.002 & 0.631 $\pm$ 0.002 & 0.636 $\pm$ 0.002 & 0.053 $\pm$ 0.018 & 0.052 $\pm$ 0.017 \\
\midrule[0.3pt]\bottomrule[1pt]
\end{tabular}
}
\end{table}
\subsection{Reliability plots} \label{sec: reliability_plot_lidc}
In Figure \ref{fig:reliability_diagram_lidc}, we show the reliability plot of our model (left) and compare with the plot of~\cite{kassapis2021calibrated} (right). We generate the result by running the pre-trained model provided by~\cite{kassapis2021calibrated}. For both methods, 16 samples are drawn for evaluation and we use the standard-sampling representation. This corresponds to the results on the last column of Table \ref{tab:lidc_comp}. We see that the confidence of our model calibrates well with the accuracy, resulting in a small gap in the curve and achieving a better pixel-wise calibration result than~\cite{kassapis2021calibrated}.
\begin{figure}[htbp]
\centering
\begin{minipage}[t]{0.49\textwidth}
\centering
\includegraphics[width=\linewidth]{img/lidc_reliab_ours_16.pdf}
\end{minipage}
\begin{minipage}[t]{0.49\textwidth}
\centering
\includegraphics[width=\linewidth]{img/lidc_reliab_carsss_16.pdf}
\end{minipage}
\caption{We show the reliability plot of our model (left) and compare with the plot of~\cite{kassapis2021calibrated} (right) on the LIDC dataset. For both methods, 16 samples are drawn for evaluation and we use the standard-sampling representation.
We achieves 0.041\% ECE (one of the three random runs in Table \ref{tab:lidc_comp}).}
\label{fig:reliability_diagram_lidc}
\end{figure}
\subsection{Analysis on the type of uncertainty} \label{sec: uncertainty_type_lidc_results}
While specifically designed for modeling the data-inherent uncertainty (aleatoric), the outputs of our model may also contain the model uncertainty (epistemic). We note that our major evaluation metrics GED and M-IoU only measure how the model learns the aleatoric uncertainty. Therefore, the epistemic component in the output could lead to an additional error in these metrics. To explore how much the epistemic uncertainty accounts for the total predictive uncertainty in our model and how it affects the final results, we conduct the following experiments.
Specifically, we follow~\cite{kendall2017uncertainties} to utilize the property that epistemic uncertainty varies under different training data sizes while aleatoric part does not. By observing how the predictive uncertainty changes with different sizes of training data, we can infer how much the epistemic uncertainty accounts for in the predictive uncertainty. Besides, we also show how the metrics GED, M-IoU and ECE change, in order to analyze how the variation of epistemic uncertainty affects the final metrics. Experiments are done using all the labels and evaluated in 16 samples with standard-sampling representation.
As shown in Table \ref{tab:uncertainty_type_lidc}, we change the size of the original LIDC training set to 1/2, 1/4 and summarize the uncertainty score by using pixel-wise entropy (same as~\cite{kendall2017uncertainties}) and sample diversity. The sample diversity is calculated as the sample-wise IoU distance, which is the same as the diversity term in calculating GED (cf. Eq.\ref{eq: ged}). We find this score provides a better measure of the output variation under the class-imbalance situation. We see that both uncertainty scores increase slightly with the size of the training set decreasing. However, the magnitude of this change is in 1e-3, which is negligible when compared to the standard derivation of the repeated experiments. Therefore, we conclude that the predictive uncertainty contains only a small amount of epistemic uncertainty, which is nearly negligible. Besides, the changes in GED, M-IoU and ECE metrics are all in the same magnitude as the standard derivation of the repeated experiments and are at least one order of magnitude smaller than the difference in comparison results among methods in Table \ref{tab:lidc_comp}.
\begin{table}[h]
\centering
\caption{We conduct experiments on the LIDC dataset with different training data sizes. Here, entropy and sample diversity are used to measure the uncertainty level of output in both pixel-wise and sample-wise. Experiments are done using all the labels and evaluated in 16 samples with standard-sampling representation.}
\label{tab:uncertainty_type_lidc}
\setlength{\tabcolsep}{6pt}
\renewcommand{\arraystretch}{1.1}
\resizebox{\textwidth}{!}{%
\begin{tabular}{l|ll|lll}
\toprule
Train dataset & Entropy (\%) & Sample Diversity & GED $\downarrow$ & M-IOU $\uparrow$ & ECE $\downarrow$ (\%) \\ \midrule
LIDC / 4 & 0.561 $\pm$ 0.006 & 0.509 $\pm$ 0.010 & 0.226 $\pm$ 0.002 & 0.615 $\pm$ 0.006 & 0.097 $\pm$ 0.016 \\
LIDC / 2 & 0.559 $\pm$ 0.007 & 0.505 $\pm$ 0.011 & 0.225 $\pm$ 0.002 & 0.619 $\pm$ 0.004 & 0.076 $\pm$ 0.010 \\
LIDC & 0.554 $\pm$ 0.007 & 0.501 $\pm$ 0.007 & 0.218 $\pm$ 0.003 & 0.624 $\pm$ 0.004 & 0.064 $\pm$ 0.015 \\ \bottomrule
\end{tabular}%
}
\end{table}
\subsection{Analysis on the number of experts} \label{sec: num_experts}
We analyze how expert number influences the model performance. To do this, we conduct experiments on the LIDC full label dataset under expert number = 1,2,4,6,8,10, where 4 is our default expert number that is used in other experiments. We show the GED score evaluated with 16 samples. As shown in Figure \ref{fig:lidc_expert_number}, with the increasing number of experts, the model achieves a better GED score in general. However, the improvement is much slower when the expert number is larger than 4. Intuitively, more experts correspond to a more powerful model capacity. Though with too many experts, there could be some duplicate ones predicting roughly similar predictions. {By using 4 experts, we keep a balance between the model capacity and the expert compactness.}
\begin{figure}[h]
\centering
\includegraphics[width=\linewidth]{img/lidc_expert_number.pdf}
\caption{The GED score changes with the number of experts, evaluated with 16 samples in standard-sampling representation. Experiments are conducted on the LIDC dataset with the full label set. }
\label{fig:lidc_expert_number}
\end{figure}
\subsection{Analysis of the model architecture}\label{sec: model_arch}
{Following, we explore the flexibility of our model design by conducting analysis experiments on three kinds of model variations: 1. different segmentation backbone; 2. the gating network takes input from different layers of the encoder; 3. the noise injects into the different layers of the decoder (cf. Sec. \ref{sec:model}). Experiments are conducted on the LIDC dataset with full labels and evaluated with 16 samples in the standard-sampling representation. Results are shown in Table \ref{tab:model_arch} and we summarize that with all variations explored, the model shows satisfactory results and consistently performs the best compared with previous works (cf. Table \ref{tab:lidc_comp}), which demonstrates the flexibility of our MoSE design. Following, we analyze the experimental results for each variation in detail.}
{On row \# 1, we change the currently used Unet architecture to DeepLabv3+~\citep{Chen_2018_ECCV} with ResNet101\footnote{Details can be found in \url{https://github.com/jfzhang95/pytorch-deeplab-xception}.}. The gating module takes feature from the ASPP output and noise injection are added in the last layer of the encoder. Other architecture designs and hyperparameters the kept same as default model. We see that, without further tuning, the experiment using DeepLabv3+ as the backbone has achieved satisfactory results on all metrics. This validates the feasibility of using other segmentation network as the backbone.
On row \# 2 and \#3, we make the gating network take the input features from the 2nd and 4th layers from the overall 6-layer encoder.
On row \# 4 and \#5, we conduct experiments with the latent noise fused with the 2nd and 4th layers from the overall 6-layer decoder.
In comparison, we list the results of our default setting in the last row, which takes features from the last layer of the encoder and fuse latent noises in the last layer of the decoder.
All other experimental settings are kept the same in these experiments. We observe that all these model variations show good performance on the final metrics, which validates the efficacy of using different levels of image features from the encoder network for the gating network input and injecting expert-specific noises in the different layers of decoder. For a new dataset, we suggest designing the gating network, its connection with the segmentation backbone, and noise injection way via prior knowledge or by architecture tuning.
}
\begin{table}[h]
\centering
\caption{We analyze different model variations by replacing the backbone and changing the noise injection layer and the gating input layer. Experiments are done on the full-label LIDC dataset and evaluated with 16 samples in the standard-sampling representation. Below we denote L as the full layer of encoder and Q as the full layer of decoder. In our default Unet architecture, L = Q = 6.}
\label{tab:model_arch}
\resizebox{\textwidth}{!}{%
\renewcommand{\arraystretch}{1.1}
\begin{tabular}{lll|lll}
Backbone & Noise Injection Layer & Gating Input Layer & GED $\downarrow$ & M-IOU $\uparrow$ & ECE $\downarrow$ (\%) \\ \midrule
Deeplab v3+ & - & - & 0.237 $\pm$ 0.005 & 0.609 $\pm$ 0.005 & 0.104 $\pm$ 0.018 \\\midrule
Unet & 1/3 L & Q & 0.222 $\pm$ 0.005 & 0.620 $\pm$ 0.004 & 0.080 $\pm$ 0.037 \\
Unet & 2/3 L & Q & 0.218 $\pm$ 0.001 & 0.626 $\pm$ 0.002 & 0.082 $\pm$ 0.009 \\\midrule
Unet & L & 1/3 Q & 0.224 $\pm$ 0.003 & 0.625 $\pm$ 0.001 & 0.086 $\pm$ 0.006 \\
Unet & L & 2/3 Q & 0.220 $\pm$ 0.001 & 0.627 $\pm$ 0.002 & 0.082 $\pm$ 0.011 \\\midrule
Unet & L & Q & 0.218 $\pm$ 0.003 & 0.624 $\pm$ 0.004 & 0.064 $\pm$ 0.015 \\\bottomrule
\end{tabular}%
}
\end{table}
\section{Additional experiments on Cityscapes} \label{sec: city_additional}
\subsection{Reliability plots}\label{sec: reliability_plot_city}
On the left of Figure \ref{fig:reliability_diagram_city}, we show the reliability plot of our model with the sample number $S = 35$ under the compact representation. For comparison, we generate the reliability plot of the previous state of the art~\citep{kassapis2021calibrated} by running the pre-trained model provided by~\cite{kassapis2021calibrated} with 35 samples and show the result on the right. We see that our model achieves a better calibration result.\footnote{{Notice that our model has all confidence scores less than 0.8. This is reasonable since the maximum occurrence probability of the evaluated ground truth class is $(1-\frac{4}{17}) \approx 0.76$ (cf. Sec. \ref{sec:datasets}).}}
\begin{figure}[htbp]
\centering
\begin{minipage}[t]{0.49\textwidth}
\centering
\includegraphics[width=\linewidth]{img/city_reliab_ours_weighted_1.pdf}
\end{minipage}
\begin{minipage}[t]{0.49\textwidth}
\centering
\includegraphics[width=\linewidth]{img/city_reliab_carsss_35.pdf}
\end{minipage}
\caption{We show the reliability plots of our model (left) and~\cite{kassapis2021calibrated} (right) on the Cityscapes dataset evaluated with 35 samples. Our result is taken from one of the three random runs in Table \ref{tab:cityscape2} last column, last row. }
\label{fig:reliability_diagram_city}
\end{figure}
\begin{figure}[htbp]
\centering
\begin{minipage}[t]{\textwidth}
\centering
\includegraphics[height=0.48\textheight]{img/cityscape_prob_calibration.pdf}
\end{minipage}
\begin{minipage}[t]{\textwidth}
\centering
\includegraphics[height=0.48\textheight]{img/cityscape_prob_calibration_carsss35.pdf}
\end{minipage}
\caption{We compare the calibration of our method and Kassapis et al. (2021) in terms of mode-wise probabilities and pixel-wise class flip ratios. In the table below, we use `0' to demonstrate the class is not flipped, and use `1' to demonstrate the class is flipped. }
\label{fig:mode_prob}
\end{figure}
\subsection{Calibration in mode ratio and pixel flip ratio}\label{sec: city_mode_ratio}
To have a further analysis of the predictive distribution and see how well it learns the characteristic of the multimodal ground truth distribution, we follow~\cite{NEURIPS2018_473447ac,kassapis2021calibrated} to calculate the mode ratio and pixel flip ratio of model outputs. To calculate the mode ratio, we match each output sample to the nearest ground truth label to determine its mode, where the distance is computed as 1-IoU. For the class flip ratio, we count the relative proportions between the five flipped classes and the corresponding original classes for all pixels across all samples. In Figure \ref{fig:mode_prob}, we show the mode ratio and pixel-wise class flip ratio of our method, and compare with ~\cite{kassapis2021calibrated}. Specifically, we use the weighted output representation with 35 samples (last row in Table \ref*{tab:cityscape2}) and generate the output of \cite{kassapis2021calibrated} by running the pre-trained model provided, and evaluated with 35 samples (row \#4 in Table \ref*{tab:cityscape2}). We see that our method has better calibration with the ground truth probabilities at both mode-level and pixel-level.
\subsection{Analysis on the type of uncertainty} \label{sec: uncertainty_type_city_results}
{We conduct experiments to analyze the type of predictive uncertainty on the Cityscapes dataset similar to the experiments conducted on the LIDC dataset. We refer to Sec.~\ref{sec: uncertainty_type_lidc_results} for the detail explanation on experimental settings and do not repeat here. Results are evaluated in 35 samples with standard-sampling representation. As shown in Table \ref{tab:uncertainty_type_cityscapes}, we get similar observations as in the LIDC where both the uncertainty scores and the evaluation metrics do not change much with the varies in training data size. Therefore, the epistemic uncertainty has less account in the predictive uncertainty of our model.}
\begin{table}[h]
\centering
\caption{We conduct experiments on the Cityscapes dataset with different training data sizes.
Results are evaluated in 35 samples with standard-sampling representation.}
\label{tab:uncertainty_type_cityscapes}
\setlength{\tabcolsep}{6pt}
\renewcommand{\arraystretch}{1.1}
\resizebox{\textwidth}{!}{%
\begin{tabular}{l|ll|lll}
\toprule
Train dataset & Entropy & Sample Diversity & GED $\downarrow$ & M-IOU $\uparrow$ & ECE $\downarrow$ (\%) \\ \midrule
Cityscapes / 4 & 0.823 $\pm$ 0.006 & 0.598 $\pm$ 0.002 & 0.157 $\pm$ 0.002 & 0.686 $\pm$ 0.004 & 1.878 $\pm$ 0.009 \\
Cityscapes / 2 & 0.814 $\pm$ 0.004 & 0.597 $\pm$ 0.002 & 0.156 $\pm$ 0.001 & 0.687 $\pm$ 0.005 & 1.858 $\pm$ 0.013 \\
Cityscapes & 0.817 $\pm$ 0.010 &0.595 $\pm$ 0.004 & 0.155 $\pm$ 0.002 & 0.690 $\pm$ 0.002 & 1.861 $\pm$ 0.015 \\\bottomrule
\end{tabular}%
}\end{table}
\subsection{Analysis on the number of experts} \label{sec: city_num_experts}
For the main experiments on the Cityscapes dataset, we assume the ground truth mode number (32) is known and hence choose a slightly larger expert number K = 35. This allows us to validate how the model deals with the redundant experts. If we do not have any prior information on the mode number, we can also use a validation set to draw the curve on how performance changes with the increase of expert numbers and select the K number at the inflection point of the curve.
{Below, we show the experimental analysis of the expert number K for model training on the Cityscapes dataset. Specifically, we evaluate the model performance with expert-number K = 25, 30, 32, 35, 40, where 35 is our default expert number. As shown in Figure \ref{fig:cityscape_expert_number}, we see that with the increasing number of experts, the model achieves a better GED score in general. Besides, K = 32 is the inflection point and can be chosen when no prior information is known. Results are evaluated with 35 samples in standard-sampling representation.}
\begin{figure}[h]
\centering
\includegraphics[width=\linewidth]{img/cityscape_expert_number.pdf}
\caption{The GED score changes with the number of experts, evaluated with 35 samples in standard-sampling representation. Experiments are conducted on the Cityscapes dataset. }
\label{fig:cityscape_expert_number}
\end{figure}
\subsection{Complementary Qualitative Analysis}\label{sec: city_qualitative}
In Figure \ref{fig:cityscape_sup}, we show the predictions from all 35 experts. We matched 32 of them with 32 ground truth modes for easy comparison. We see that the expert networks learn the variety of modes well and the gating network gives well-calibrated mode probabilities. Also, we notice that our model learns relatively compact experts, with the extra experts shown at the bottom of the predictions having relatively small weights.
\begin{figure}[h]
\centering
\includegraphics[width=\linewidth]{img/city_vis_sup.pdf}
\caption{The extra visualization results on the Cityscapes. On the first line of each case, we show the input image masked with a ground truth label. Rest lines show samples from 35 experts and 32 ground truth modes. {For clarity, we show matched pairs for ground truth masks on the top and predicted masks on the bottom and divide them into 5 subgroups labeled as (1) - (5).}
The predictions are generated by sampling once from each expert, we denote the corresponding predicted probabilities on the left head of each prediction. }
\label{fig:cityscape_sup}
\end{figure}
\section{Discussion} \label{sec: limitations}
We discuss some of the limitations of our method in the following:
Firstly, when the expert number is larger than the real mode number in the dataset, it is possible to have two expert networks giving similar predictions. As discussed in Sec.~\ref{sec: training_details}, this makes the matching process unstable and we
adopt the soft gradient trick to handle this scenario. In the future, efforts can be done by either stabilizing the matching process or automatically pruning/merging those duplicated experts. Secondly, the matching is based on the cost matrix, however, some existed pairwise loss functions may not provide a good measure for comparison. For example, on the LIDC dataset where the class imbalance scenario exists,
using the cross-entropy loss makes the measure dominated by the background pixels, which leads to less satisfactory matching results and further disturbs the training.
\section{Conclusion}
In this work, we tackle the problem of modeling aleatoric uncertainty in semantic segmentation. We propose a novel mixture of stochastic experts model to capture the multi-modal segmentation distribution, which contains several expert networks and a gating network. This yields an efficient two-level uncertainty representation. To learn the model, we develop a Wasserstein-like loss that computes an optimal coupling matrix between the model outputs and the ground truth label distribution. Also, it can easily integrate traditional segmentation metrics and be efficiently optimized via constraint relaxation. We validate our method on the multi-grader LIDC-IDRI dataset and the muti-modal Cityscapes dataset. Results demonstrate that our method achieves the state-of-the-art or competitive performance on all metrics. {We discuss the limitations of our method in Appendix \ref{sec: limitations}.}
\section{Experiment} \label{sec:exp}
We evaluate our method on two public benchmarks, including LIDC-IDRI dataset~\citep{armato2011lung} and {multimodal Cityscapes dataset~\citep{Cordts2016Cityscapes}}.Below we first introduce dataset information in Sec. \ref{sec:datasets} and experiment setup in Sec. \ref{sec:Experiment Setup} . Then we present our experimental results on the LIDC dataset in Sec. \ref{sec:lidc} and the Cityscapes dataset in Sec. \ref{sec:cityscapes}.
\subsection{Datasets} \label{sec:datasets}
\paragraph{LIDC-IDRI Dataset}
The LIDC dataset~\citep{armato2011lung} consists of 1018 3D thorax CT scans, with each annotated by four anonymized radiologists.\footnote{A total of twelve radiologists participated in the image annotation and can revise their own annotations by reviewing the anonymized annotations of other radiologists. Therefore, the final disagreement reached among radiologists indicates the true ambiguities and typically represents the aleatoric uncertainty~\citep{armato2011lung}. } For fair comparison, we use a pre-processed 2D dataset provided by~\cite{NEURIPS2018_473447ac} with 15096 slices each cropped to $128\times 128$ patches
and adopt the 60-20-20 dataset split manner same as~\cite{baumgartner2019phiseg, NEURIPS2020_95f8d990}. We conduct experiments in two settings, which train the model using either full labels or one randomly-sampled label per image as in the literature.
\paragraph{Multimodal Cityscapes Dataset}
The Cityscapes dataset~\citep{Cordts2016Cityscapes} contains images of street scenes with 2975 training images and 500 validation images, labeled with 19 different semantic classes. To inject ambiguity,
we follow~\cite{NEURIPS2018_473447ac} to randomly flip five original semantic classes into five new classes with certain probabilities. Specifically, the original classes of `sidewalk', `person', `car' , `vegetation', `road' are randomly transformed to sidewalk2', `person2', `car2' , `vegetation2', `road2' with corresponding probabilities of 8/17, 7/17, 6/17, 5/17, 4/17. This brings a total of 24 semantic classes in the dataset, with $2^5 = 32$ modes of ground truth segmentation labels for each image (See Figure \ref{fig:city_vis}). For fair comparison, we follow the setting in~\citep{NEURIPS2018_473447ac, kassapis2021calibrated} and report results on the validation set.
\subsection{Experiment Setup} \label{sec:Experiment Setup}
\paragraph{Performance Measures}
To measure how well the model captures the aleatoric uncertainty in segmentation, we follow~\cite{NEURIPS2018_473447ac,kohl2019hierarchical,baumgartner2019phiseg,NEURIPS2020_95f8d990,kassapis2021calibrated} to adopt two main metrics for evaluation: the intersection-over-union (IoU) based \textit{Generalized Energy Distance} (GED)~\citep{szekely2013energy} and the \textit{Hungarian-matched IoU}~\citep{kuhn1955hungarian}. We extend the latter to \textit{Matched IoU} (M-IoU) by adopting network simplex algorithm~\citep{cunningham1976network}
for the setting that the marginal distribution of predictions or ground truth is not uniform. Furthermore, we use the \textit{expected calibration error} (ECE)~\citep{guo2017calibration} with 10 bins as a secondary metric to quantify the pixel-wise uncertainty calibration.
For details on these metrics please refer to Appendix \ref{sec: evaluation_metrics}. Below, we conduct all experiments in triplicate and report the results with mean and standard derivation.
\paragraph{Implementation Details}
We employ the same encoder-decoder architecture as in U-net~\citep{ronneberger2015u} and follow the model design as~\cite{baumgartner2019phiseg}. On the LIDC dataset, we use $K = 4$ experts each with $S = 4$ samples. In the loss function, we use the IoU~\citep{milletari2016v} as the pair-wise cost function and set the hyperparameters $\gamma_0 = 1/2$, $\beta = 1$ for full-annotation case and $\beta = 10$ for one-annotation case. On the Cityscapes dataset, we use a slightly larger number $K = 35$ of experts each with $S = 2$ samples. We use the CE as the pair-wise cost function and set the hyperparameters $\gamma_0 = 1/32$, $\beta = 1$ for the loss. To stabilize the training, we adopt a gradient-smoothing trick in some cases and refer the reader to Appendix \ref{sec: training_details} for details.
\subsection{Experiments on the LIDC-IDRI Dataset}\label{sec:lidc}
\begin{table}[t]
\centering
\caption{Quantitative results on the LIDC dataset. The number of sampled outputs is shown in the parentheses aside the metrics. Our method achieves SOTA on all evaluation metrics for all settings.}
\label{tab:lidc_comp}
\resizebox{\textwidth}{!}{%
\setlength{\tabcolsep}{6pt}
\renewcommand{\arraystretch}{1.1}
\begin{tabular}{llllllll}
\toprule[1pt]\midrule[0.3pt]
Method & {\# labels} & {GED $\downarrow$ (16)} & {GED $\downarrow$ (50)} & {GED $\downarrow$ (100)} & {M-IoU $\uparrow$ (16)} & {ECE $\downarrow$ (\%) (16)} \\
\midrule
\cite{NEURIPS2018_473447ac} & \multirow{8}{*}{All} & 0.320 $\pm$ 0.030 & - & 0.252 $\pm$ N/A $^{\dag}$ & 0.500 $\pm$ 0.030 & -\\
\cite{kohl2019hierarchical} & & 0.270 $\pm$ 0.010 & - & - & 0.530 $\pm$ 0.010 & -\\
\cite{hu2019supervised} & & - & 0.267 $\pm$ 0.012 & - & - & -\\
\cite{baumgartner2019phiseg} & & - & - & 0.224 $\pm$ N/A & - & -\\
\cite{NEURIPS2020_95f8d990} & & - & - & 0.225 $\pm$ 0.002 & - & -\\
\cite{kassapis2021calibrated} & & 0.264 $\pm$ 0.002 & 0.248 $\pm$ 0.004 & 0.243 $\pm$ 0.004 & 0.592 $\pm$ 0.005 & 0.214 $^{*}$ \\
\textbf{Ours} & & \textbf{0.218 $\pm$ 0.003} & \textbf{0.195 $\pm$ 0.002} & \textbf{0.189 $\pm$ 0.002} & \textbf{0.624 $\pm$ 0.004} & \textbf{0.064 $\pm$ 0.015}\\
\textbf{Ours - compact} & & \textbf{0.195 $\pm$ 0.005} & \textbf{0.187 $\pm$ 0.003} & \textbf{0.186 $\pm$ 0.002} & \textbf{0.635 $\pm$ 0.003} & \textbf{0.054 $\pm$ 0.015}\\
\midrule
\cite{NEURIPS2018_473447ac} & \multirow{5}{*}{One} & - & - & 0.445 $\pm$ N/A & - & -\\
\cite{baumgartner2019phiseg} & & - & - & 0.323 $\pm$ N/A & - & -\\
\cite{NEURIPS2020_95f8d990} & & - & - & 0.365 $\pm$ 0.005 & - & -\\
\textbf{Ours} & & \textbf{0.252 $\pm$ 0.004} & \textbf{0.229 $\pm$ 0.005} & \textbf{0.223 $\pm$ 0.005} & \textbf{0.596 $\pm$ 0.003} & \textbf{0.105 $\pm$ 0.009}\\
\textbf{Ours - compact} & & \textbf{0.228 $\pm$ 0.004} & \textbf{0.221 $\pm$ 0.004} & \textbf{0.220 $\pm$ 0.005} & \textbf{0.605 $\pm$ 0.003} & \textbf{0.090 $\pm$ 0.011}\\
\midrule[0.3pt]\bottomrule[1pt]
\end{tabular}%
\begin{tablenotes}
\small
\item $\dag$ This score is taken from~\cite{baumgartner2019phiseg}.
\item $*$ {This score is generated by running the pre-trained model provided by~\cite{kassapis2021calibrated}. }
\end{tablenotes}
\end{table}
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{img/lidc_vis.pdf}
\caption{Qualitative results on the LIDC dataset. Each row of predictions represents a sampling sequence from an expert, with the last line showing the entropy map. We denote the mode probabilities on the left head of each row.
}
\label{fig:lidc_vis}
\end{figure}
We compare our method with the previous state of the arts and summarize the results under sample number N = 16, 50, 100 in the Table \ref{tab:lidc_comp} and Appendix Table \ref{tab:lidc_ours}. We validate our method with two kinds of output representation (cf. Sec. \ref{sec:representation}): the standard-sampling-based representation denoted as `Ours' and the weighted form as `Ours - compact'. For the latter, we report the results by taking $S=$ 4, 12, and 25 samples from each expert in order to generate the comparable number of total samples (i.e., 16, 48, and 100) as the standard-sampling form.
As shown in Table \ref{tab:lidc_comp}, our method is consistently superior to others with the standard-sampling form in all evaluation settings. Moreover, the compact weighted representation improves the performance further, especially in the small sample (16) case, and is more stable across different sample sizes. For the challenging one label setting, our method is less affected and also outperforms the others by a large margin, demonstrating its robustness to the number of annotations per image.
{We show the reliability plots in Appendix \ref{sec: reliability_plot_lidc}. To validate that our MoSE model does capture the aleatoric uncertainty, we further conduct experiments with varying training data sizes in Appendix \ref{sec: uncertainty_type_lidc_results}.}
Figure \ref{fig:lidc_vis} visualizes our compact uncertainty representation on two cases. We can see that each expert predicts a distinct mode of aleatoric uncertainty with a shared style across images. The image-specific weighting of experts shows a coarse-level uncertainty, while the variations in samples from each expert demonstrate a fine-level uncertainty.
\subsubsection{Ablation Study}\label{sec: ablation}
\begin{table}[t]
\caption{Ablation study on our model components, evaluated with sample number 16.}
\label{tab:ablation}
\centering
\resizebox{0.97\textwidth}{!}{
\setlength{\tabcolsep}{10pt}
\renewcommand{\arraystretch}{1.1}
\begin{tabular}[t]{cccccc}
\toprule[1pt]\midrule[0.3pt]
Expert type & Expert weights & loss & {GED $\downarrow$} & {M-IoU $\uparrow$} & {ECE $\downarrow$ {(\%)} } \\
\midrule
stochastic & learnable & IoU loss & 0.533 $\pm$ 0.001 & 0.533 $\pm$ 0.001 & 0.277 $\pm$ 0.017\\
stochastic & uniform & OT-based & 0.282 $\pm$ 0.002 & 0.545 $\pm$ 0.007 & 0.215 $\pm$ 0.006\\
deterministic & learnable & OT-based & 0.246 $\pm$ 0.006 & 0.591 $\pm$ 0.001 & 0.142 $\pm$ 0.003\\
\midrule
stochastic & learnable & OT-based & \textbf{0.218 $\pm$ 0.003} & \textbf{0.624 $\pm$ 0.004} & \textbf{0.064 $\pm$ 0.015}\\
\midrule[0.3pt]\bottomrule[1pt]
\end{tabular}%
}
\end{table}
We conduct an ablation study of the model components on the LIDC dataset under the training setting with all (four) labels. We show the results with the sample number $S=16$ under the standard-sampling form in Table \ref{tab:ablation}, and leave the other settings to Appendix Table \ref{tab:abla_50_100}. In Row \#1, we first analyze our OT-based loss by replacing it with the IoU loss~\citep{milletari2016v} calculated on every possible pair of predictions and labels. This leads to a performance drop of more than 0.3 in the GED score and the resulting model tends to predict an averaged label mask. Moreover, we modify our MoSE design by adopting a uniform weighting of experts in Row \#2, and replacing the stochastic experts with deterministic ones in Row \#3. For both cases, compared with our full model in the last row, we see a clear performance drop. Those comparison results validate the efficacy of our design. We also conduct more analysis experiments by varying different aspects of the model architecture. Please refer to Appendix \ref{sec: num_experts} and \ref{sec: model_arch} for details.
\subsection{Experiments on the multimodal Cityscapes Dataset}\label{sec:cityscapes}
On this dataset, we compare our method with two previous works~\citep{NEURIPS2018_473447ac,kassapis2021calibrated}.
{For fair comparison, we follow~\cite{kassapis2021calibrated} to use predictions~\citep{deeplabmodelzo1} from a pretrained model as an additional input.}
We evaluate our model with the standard-sampling representation under 16 and 35 samples, as well as the compact representation under 35 samples (1 sample for each expert). As the previous works only provide their results on GED with 16 samples, we generate other results from the pre-trained model provided by~\cite{kassapis2021calibrated} and summarize the results in Table \ref{tab:cityscape2}.
Overall, our method achieves the state-of-the-art M-IoU and ECE under both sample number settings, as well as comparable GED scores with~\cite{kassapis2021calibrated}, demonstrating the superiority of our approach. The discrepancy in the GED performance is likely caused by the evaluation bias of the GED on sample diversity~\citep{kohl2019hierarchical}, as~\cite{kassapis2021calibrated} typically produces more diverse outputs due to its adversarial loss. Moreover, we observe that our compact representation achieves better performance than the standard-sampling form under the same sample number (35), which demonstrates the effectiveness of such compact representation in multimodal uncertainty estimation.
We show some visualization results in Figure \ref{fig:city_vis}. For clarity, we pair each ground truth label with the closest match in model predictions. We can see that our model can predict multimodal outputs with calibrated mode probabilities well.
Furthermore, we include additional reliability plots, experimental analysis on the calibration of mode ratio, detailed ablation on model architecture, and more visualization results in the Appendix \ref{sec: city_additional}.
\begin{table}[t]
\centering
\caption{Quantitative results on the multimodal Cityscapes dataset. }
\label{tab:cityscape2}
\resizebox{\textwidth}{!}{
\setlength{\tabcolsep}{10pt}
\renewcommand{\arraystretch}{1.1}
\begin{tabular}{lllll}
\toprule[1pt]\midrule[0.3pt]
Method & Sample number & {GED $\downarrow$} & {M-IoU $\uparrow$} & {ECE $\downarrow$ {(\%)}} \\
\midrule
\cite{NEURIPS2018_473447ac} & \multirow{3}{*}{16} & 0.206 $\pm$ N/A & - & - \\
\cite{kassapis2021calibrated} & & \textbf{0.164 $\pm$ 0.01} & 0.633 & 3.274 \\
Ours & & 0.176 $\pm$ 0.002 & \textbf{0.638 $\pm$ 0.002} & \textbf{3.030 $\pm$ 0.021} \\
\midrule
\cite{kassapis2021calibrated} & \multirow{3}{*}{35} & \textbf{0.139} & 0.672 & 2.196 \\
Ours & & 0.155 $\pm$ 0.002 & \textbf{0.690 $\pm$ 0.002} & \textbf{1.861 $\pm$ 0.015} \\
Ours - compact & & 0.142 $\pm$ 0.001 & \textbf{0.761 $\pm$ 0.002} & \textbf{0.352 $\pm$ 0.025} \\
\midrule[0.3pt]\bottomrule[1pt]
\end{tabular}%
}
\end{table}
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{img/city_vis.pdf}
\caption{ Qualitative results on the multimodal Cityscapes dataset. Within each case, (a) shows the image and a part of ground truth labels with corresponding probabilities, and (b) shows the entropy map, predictive samples and predicted probabilities.
}
\label{fig:city_vis}
\end{figure}
\section{Introduction}
Semantic segmentation, a core task in computer vision, has made significant progress thanks to the powerful representations learned by deep neural networks. The majority of existing work focus on generating a single (or sometimes a fixed number of) segmentation output(s) to achieve high pixel-wise accuracy~\citep{minaee2021image}. Such a prediction strategy, while useful in many scenarios, typically disregards the predictive uncertainty in the segmentation outputs, even when the input image may contain ambiguous regions. Equipping predicted segmentation with calibrated uncertainty, however, is essential for many safety-critical applications such as medical diagnostics and autonomous driving to prevent problematic low-confidence decisions~\citep{amodei2016concrete,armato2011lung,levinson2011towards}.
An important problem of modeling predictive uncertainty in semantic segmentation is to capture \textit{aleatoric uncertainty}, which aims to predict multiple possible segmentation outcomes with calibrated probabilities when there exist ambiguities in input images~\citep{NEURIPS2020_95f8d990}. In lung nodule segmentation, for example, an ambiguous image can either be annotated with a large nodule mask or non-nodule with different probabilities~\citep{armato2011lung}.
Such a problem can be naturally formulated as label distribution learning~\citep{geng2013label}, of which the goal is to estimate the conditional distribution of segmentation given input image. Nonetheless, due to the high-dimensional output space, typical multimodal characteristic of the distributions and limited annotations, it remains challenging to predict well-calibrated uncertainty for segmentation.
Most previous works~\citep{NEURIPS2018_473447ac,kohl2019hierarchical,baumgartner2019phiseg,hu2019supervised} adopt the conditional variational autoencoder (cVAE) framework~\citep{NIPS2015_8d55a249}
to learn the predictive distribution of segmentation outputs, which has a limited capability to capture the multimodal distribution due to the over-regularization from the Gaussian prior
or posterior collapse~\citep{razavi2018preventing, qiu2020modal}.
Recently, ~\cite{NEURIPS2020_95f8d990} propose a multivariate Gaussian model in the logit space of pixel-wise classifiers. However, it has to use a low-rank covariance matrix for computational efficiency, which imposes a restriction on its modeling capacity. To alleviate this problem, ~\cite{kassapis2021calibrated} employ the adversarial learning strategy to learn an implicit density model, which requires additional loss terms to improve training stability~\citep{luc2016semantic,samson2019bet}.
Moreover, all those methods have to sample a large number of segmentation outputs to represent the predictive distribution, which can be inefficient in practice.
In this work, we aim to tackle the aforementioned limitations by explicitly modeling the multimodal characteristic of the segmentation distribution. To this end, we propose a novel \textit{mixture of stochastic experts} (MoSE)~\citep{masoudnia2014mixture}
framework, in which each expert network encodes a distinct mode of aleatoric uncertainty in the dataset and its weight represents the probability of an input image being annotated by a segmentation sampled from that mode. This enables us to decompose the output distribution into two granularity levels and hence provides an efficient and more interpretable representation for the uncertainty. Moreover, we formulate the model learning as an \textit{Optimal Transport} (OT) problem~\citep{villani2009optimal}, and design a \textit{Wasserstein}-like loss that directly minimizes the distribution distance between the MoSE and ground truth annotations.
Specifically, our MoSE model comprises a set of stochastic networks and a gating network. Given an image, each expert computes a semantic representation via a shared segmentation network, which is fused with a latent Gaussian variable to generate segmentation samples of an individual mode. The gating network takes a semantic feature embedding of the image and predicts the mode probabilities. As such, the output distribution can be efficiently represented by a set of samples from the experts and their corresponding probabilities. To learn the model, we relax the original OT formulation and develop an efficient bi-level optimization procedure to minimize the loss.
We validate our method on the LIDC-IDRI dataset~\citep{armato2011lung} and a modified multimodal Cityscapes dataset~\citep{Cordts2016Cityscapes, NEURIPS2018_473447ac}. Results demonstrate that our method achieves the state-of-the-art or competitive performance on all metrics.
To summarize, our main contribution is three-folds: (i) We propose a novel mixture of stochastic experts model to capture the multimodal aleatoric uncertainty in segmentation; (ii) We develop a Wasserstein-like loss for learning the uncertainty, which can be efficiently optimized via constraint relaxation; (iii) Our method achieves the state-of-the-art or competitive performance on two challenging segmentation benchmarks.
\section{Method}
In this section, we introduce our method for capturing the multimodal aleatoric uncertainty in semantic segmentation. Specifically, we first present the problem formulation in Sec.~\ref{sec:overview}. We then describe the detailed design of our mixture of stochastic experts framework in Sec.~\ref{sec:model}, followed by our Optimal Transport formulation of the model training in Sec.~\ref{sec:train}.
\begin{figure*}[t!]
\centering
\includegraphics[width=1.0\textwidth]{img/overview_0929.pdf}
\caption{Given an ambiguous image as input, our model consists of several expert networks and a gating network and produces segmentation samples with corresponding weights. It represents the aleatoric uncertainty at two granularity levels, (e.g. the input image can be annotated as either ``2" or ``3" and also has variations near boundaries.) and is trained by an Optimal-Transport-based loss function which minimizes the distance between the MoSE outputs and ground truth annotations.}
\label{fig:model_overview}
\end{figure*}
\subsection{Problem Setup and Formulation}\label{sec:overview}
We aim to build a flexible model for capturing the multimodal aleatoric uncertainty in semantic segmentation, where each input image could be annotated in multiple ways due to its ambiguity. To this end, we formulate it as the problem of estimating the probability distribution of segmentation outputs given an input image. Formally, we assume an annotated image dataset {$\mathcal{D} = \{x_n, {y}_n^{(i)}, v_n^{(i)}\}_{n=1,\cdots,D}^{i =1,\cdots,M}$} is available for training,
where ${x}_n\in \mathbbm{R}^{H\times W\times 3}$ is an image and has $M$ {($\ge 1$ )}
distinct segmentation annotations ${y}_n^{(1)}, \cdots, {y}_n^{(M)}$ with corresponding relative frequencies $v_n^{(1)},\cdots,v_n^{(M)}\in [0,1]$, where $\sum_i v_n^{(i)} = 1$.
Each annotation ${y}_n^{(i)}\in \mathcal{Y} \coloneqq \{1,\cdots,C\}^{H\times W}$, where $C$ is the number of classes and $H,W$ are the height and width of the image, respectively. Here $\{ {y}_n^{(i)}, v_n^{(i)}\}$ represent random samples from the underlying label distribution of $x_n$.
To represent the aleatoric uncertainty, we introduce a conditional probabilistic model on segmentation, denoted as $\mu_\theta(y\mid x)\in P_\mathcal{Y}$, where $P_\mathcal{Y}$ is the space of segmentation distributions. The model learning problem corresponds to minimizing a loss function $\ell: P_\mathcal{Y}\times P_\mathcal{Y}\to\mathbbm{R}$ between our model output $\mu_{\theta}(\cdot|x_n)$ and the empirical annotation distribution of the image $x_n$, denoted as $\nu_n$, over the dataset $\mathcal{D}$. Namely, our learning task can be written as,
\begin{equation}
\label{eq:problem_setup}
\min_{\theta}\frac{1}{D}\sum_{n=1}^{D}\ell(\mu_{\theta}(y\mid x_n), \nu_n) \quad \text{where } \nu_n \coloneqq \sum_{i=1}^M v_n^{(i)}\delta({y_{n}^{(i)}})
\end{equation}
In this work, we focus on developing an effective and easy-to-interpret uncertainty representation for semantic segmentation. In particular, we propose a novel \textit{multimodal representation} in our model $\mu_\theta(y\mid x)$ and a \textit{Wasserstein}-like distribution distance as our loss $\ell$, which admits an efficient minimization strategy via partial constraint relaxation. Below we will present our model design and learning strategy in detail.
\subsection{The Mixture of Stochastic Experts (MoSE) framework}\label{sec:model}
In order to capture the multimodal structure of annotations, we develop a \textit{mixture of stochastic experts} (MoSE) framework, which decomposes the output uncertainty into two granularity levels.
At the fine-level, we employ an expert network to estimate a distinct mode of the aleatoric uncertainty in the dataset, which typically corresponds to certain local variations in segmentation. At the coarse-level, we use a gating network to predict the probabilities of an input image being segmented by those stochastic experts, which depicts the major variations in the segmentation outputs.
Specifically, we assume there exist $K$ stochastic experts $\{\mathcal{E}_k\}_{k=1}^K$, each of which defines a conditional distribution of the segmentation $y$ given an input image $x$, denoted as $P_k(y|x)$. Each expert $\mathcal{E}_k$ is associated with a weight $\pi_k(x)$ predicted by a gating network, which represents the probability of an input being annotated in the mode corresponding to $\mathcal{E}_k$.
We instantiate $P_k(y|x)$ as an implicit distribution that maps an expert-specific prior distribution $P_k(z_k)$ (where $z_k$ is a latent variable) and the input $x$ to the output distribution via a deep network $F(x,z_k)$.
As such, given the input $x$, we are able to generate a segmentation sample $s\in\mathcal{Y}$ as follows: we first sample an expert id $k$ from the categorical distribution $\mathcal{G}(\boldsymbol{\pi}(x))$ where $\boldsymbol{\pi}(x)=[\pi_1(x),\cdots,\pi_K(x)]$. Then we sample a latent code $z_k$ from $P_k(z_k)$ and compute the prediction $s$ by the network $F(\cdot)$:
\begin{equation}
\label{eq:moe}
s = F(x, z_k)\quad \text{where } k\sim \mathcal{G}(\boldsymbol{\pi}(x)), z_k \sim P_k(z_k).
\end{equation}
An overview of our MoSE framework is illustrated in Figure~\ref{fig:model_overview}.
Below we introduce our network design in Sec.~\ref{sec:network} and present an efficient output representation in Sec.~\ref{sec:representation}.
\subsubsection{Network Design} \label{sec:network}
We adopt an encoder-decoder architecture for our MoSE model, which consists of an encoder network that extracts an image representation, $K$ expert-specific prior models and a decoder network that generates the segmentation outputs, as well as a gating module for predicting the expert probabilities. We detail the design of the experts and the gating module below.
\paragraph{Expert Network}
The stochastic expert network $\mathcal{E}_k$ comprises an encoder $f_{e}(\cdot)$, a decoder $f_{d}(\cdot)$ and latent prior models $\{P_k(z_k)\}$.
We expect each expert to generate segmentation outputs for different images with a consistent uncertainty mode. To achieve this, we introduce a set of input-agnostic Gaussian distributions $\{P_k(z_k)=N(m_k, \sigma_kI), k = 1, \cdots, K\}$ as the expert-specific priors, where $m_k\in R^{L}$ and $I$ is an $L\times L$ identity matrix.
Given an input image $x$, we use a Conv-Net as the encoder $f_{e}(x)$ to compute an image representation $u_g$, followed by a Deconv-Net as the decoder $f_{d}(u_g,z_k)$ which takes the feature $u_g$ and a sampled noise $z_k$ to generate a segmentation output for the $k$-th expert. Concretely, assume the global representation $u_g\in R^{F_1\times H'\times W'}$ has $F_1$ channels and a spatial size of $ H'\times W'$. The decoder first generates a dense feature map $u_d \in R^{F_2\times H\times W}$ and then fuses with the noise $z_k$ to produce the output. Here we adopt a similar strategy as in~\cite{kassapis2021calibrated}, which projects $z_k$ to two vectors $\tilde{z}_k^{b} = w_1z_k \in R^{F_2}$ and $\tilde{z}_k^{w} = w_2z_k \in R^{F_2}$, and generates an expert-specific dense feature map as $\bar{u}_d^k[:,i,j] = (u_d[:,i,j] + \tilde{z}_k^{b}) \cdot\tilde{z}_k^{w}$, $(i,j) \in [1,H]\times [1,W]$. Finally, we use three $1\times 1$ convolution layers to generate pixel-wise predictions. For notation clarity, we denote all the parameters in the expert networks as $\theta_s$.
\paragraph{Gating Module}
The gating network shares the encoder $f_e(\cdot)$ and uses the global representation $u_g$ to predict the probabilities of experts. Specifically,
we first perform the average pooling on $u_g$ and pass the resulting vector $\bar{u}_g = \text{AvgPool}(u_g)\in R ^{F_1}$ through a three-layer MLP to generate the probabilities $\boldsymbol{\pi}(x)$. We denote the parameters in the gating module as $\theta_\pi$.
We note that our MoSE framework is flexible, which can be integrated into most modern segmentation networks with a small amount of additional parameters. Moreover, it can be easily extended to other model variants where the gating module takes the input features from other convolution layers in the encoder or the latent noise is fused with other deconvolutional layers in the decoder {(cf. \ref{sec: model_arch})}.
\subsubsection{Uncertainty Representation}\label{sec:representation}
Due to the implicit distributions used in our MoSE framework, we can only generate samples from expert networks.
In this work, we consider two types of nonparametric representations for the predictive uncertainty $\mu_{\theta}(y\mid x)$. The first representation utilizes the standard sampling strategy to generate a set of $N$ samples $\{s^{(i)}\}_{i=1}^N$ for an input image (cf. Eq.\ref{eq:moe}), and approximates the uncertainty by the resulting empirical distribution {$\frac{1}{N}\sum_{i=1}^{N}\delta (s^{(i)})$}. The second method utilizes the MoSE framework and approximates the uncertainty distribution by a set of weighted samples. Specifically, we first generate $S$ samples from each expert, denoting as $\{s_k^{(i)}\}_{i=1}^S$ for the $k$-th expert, and then approximate the predictive distribution $\mu_{\theta}$ with the following $\hat{\mu}_{\theta}$:
\begin{equation}
\label{moe_emp}
\hat{\mu}_{\theta}(y\mid x) = \sum_{k=1}^{K}\frac{\pi_{k}(x)}{S}\sum_{i=1}^{S}\delta (s_k^{(i)})
\end{equation}
where $s_k^{(i)}= F(x, z_{k}^{(i)})$, and $z_k^{(i)}\sim N(m_k, \sigma_kI)$.
It is worth noting that we can sample from multiple experts simultaneously as above, which leads to an efficient and compact representation of the segmentation uncertainty.
For clarity, we denote the nonparametric representation of both forms as $\sum_{i=1}^{N}u^{(i)}\delta (s^{(i)})$ in the following, where $u^{(i)}$ is the sample weight and $N$ is the sample number.
\subsection{Optimal Transport-based Calibration Loss}\label{sec:train}
The traditional segmentation loss functions typically encourage the model prediction to align with unique ground truth (e.g. Dice) or require an explicit likelihood function (e.g. Cross-Entropy).
To fully capture the label distribution with our implicit model, we formulate the model learning as an \textit{Optimal Transport} (OT) problem, which minimizes a \textit{Wasserstein}-like loss between the model distribution $\mu_{\theta}(y\mid x_n)$ and the observed annotations $\nu_n$ of $x_n$ on the training set $\mathcal{D}$.
Specifically, our loss measures the best way to transport the mass of model distribution $\mu_\theta(y\mid x_n)$ to match that of the ground truth distribution $\nu_n$, with a given cost function $c: \mathcal{Y}\times\mathcal{Y}\to \mathbbm{R}$. Based on this loss, our learning problem can be written as follows,
\begin{equation}
\label{eq:ot}
\min_{\theta}\sum_{n=1}^{D}W_{c}({\mu}_{\theta}(y\mid x_n), \nu_n) \quad \text{where } W_c(\mu, \nu)\coloneqq \inf_{\gamma\in \Pi (\mu, \nu)}\int_{ \mathcal{Y}\times\mathcal{Y}} c(s,y)d\gamma(s,y).
\end{equation}
Here $\Pi(\mu,\nu)$ denotes the set of joint distributions on $ \mathcal{Y}\times\mathcal{Y}$ with $\mu$ and $\nu$ as marginals. Substituting in our discrete empirical distributions $ \hat{\mu}_{\theta}(y\mid x) = \sum_{i=1}^{N}u^{(i)}\delta (s^{(i)})$ and $\nu= \sum_{j=1}^M v^{(j)}\delta({y^{(j)}})$, we derive the following objective function:
\begin{equation}\label{eq:learn}
\begin{aligned}
&\min_{\theta_s,\theta_{\pi}}\sum_{n}\sum_{i,j}\mathbf{P}_{ij}^*c(s_n^{(i)}(\theta_s),y_n^{(j)}) \quad \text{where }\mathbf{P^*} =\mathop{\arg\min}\limits_{\mathbf{P} \in \mathbf {U}} \sum_{i,j}\mathbf{P}_{ij}\mathbf{C}_{ij}\\
&\mathbf {U} \coloneqq \{\mathbf{P} \in \mathbbm{R}_{+}^{N\times M}: \mathbf{P}\mathbbm{1}_M = \mathbf{u}_n(\theta_{\pi}), \mathbf{P}^T\mathbbm{1}_N = \mathbf{v}_n\}
\end{aligned}
\end{equation}
where $\mathbf{C}\in \mathbbm{R}^{N\times M}$ is the ground cost matrix with each element $\mathbf{C}_{ij} \coloneqq c(s^{(i)},y^{(j)})$ being the pairwise cost. In our problem, we can use a commonly-adopted segmentation quality measure to compute the cost matrix. $\mathbf{P}$ is the coupling matrix that describes the transportation path, and $\mathbf {U} $ is the set for all possible values of $\mathbf{P}$ which satisfy the marginal constraints. This formulation can be regarded as a bi-level optimization problem, where the inner problem finds an optimal coupling matrix $\mathbf{P}$ and the outer problem optimizes the model output distribution under this coupling.
\paragraph{Constraint Relaxation}
We note that the marginal probabilities, $\mathbf{u}(x;{\theta_{\pi}})$, occurs in the constraint of the inner problem in Eq~\ref{eq:learn}. As such, we can not directly use backpropagation to optimize the gating network parameters $\theta_{\pi}$~\citep{peyre2019computational}.
To tackle the challenge, we propose a novel imbalanced constraint relaxation. {Our main idea is to have fewer constraints when optimizing the inner problem, and use the optimally solved matrix $\mathbf{P}$ as supervision for the gating network in the outer problem.} Formally, we introduce the following relaxation:
\begin{equation}\label{eq:relax}
\begin{aligned}
&\min_{\theta_s,\theta_{\pi}}\sum_{n}\sum_{i,j}\mathbf{P}_{ij}^*c(s_n^{(i)}(\theta_s),y_n^{(j)}) + \beta KL(\sum_j\mathbf{P}^*_{ij} || \mathbf{u}_n(\theta_{\pi})) \\
&\text{where }\mathbf{P^*} =\mathop{\arg\min}\limits_{\mathbf{P} \in \mathbf {U'}} \sum_{i,j}\mathbf{P}_{ij}\mathbf{C}_{ij} \quad \mathbf {U'} \coloneqq \{\mathbf{P} \in \mathbbm{R}_{+}^{N\times M}: \mathbf{P}\mathbbm{1}_M\le \gamma,\mathbf{P}^T\mathbbm{1}_N = \mathbf{v}_n\}
\end{aligned}
\end{equation}
where $\beta$ is used to balance the loss terms, and $\gamma\in[\frac{1}{N},1]$ is designed for preventing sub-optimal solutions in the early training stage. Empirically, we start with a value $\gamma_0< 1$ and annealing it to 1 (meaning no constraint).
It is worth noting that in the upper-level problem, the first term can be regarded as a normal segmentation loss for matched pairs of the predictions and ground truth labels, while the second term can be considered as optimizing the Cross-Entropy loss between the predicted probabilities and a pseudo label generated from the coupling matrix.
Our constraint relaxation has several interesting properties for model learning. Firstly, by moving the constraint of the predictive marginal distribution from the inner problem to the upper problem, we find it helps deal with the gradient bias~\citep{bellemare2017cramer, fatras2020learning} when the ground truth annotations are limited for each image, which is common in real scenarios. In addition, our relaxed problem can be regarded as minimizing an approximation of the original problem, and achieves the same value when the KL term being 0. Finally, our relaxation leads to a more efficient alternating learning procedure. In particular, when $\gamma = 1$, we have only one constraint in the inner problem and this can be solved in linear time~\citep{ignizio1994linear}. We show the overall training procedure and details on solving the inner problem in Appendix \ref{sec: implementation}.
\section{Related work}
\paragraph{Aleatoric uncertainty in semantic segmentation}
Aleatoric uncertainty refers to the uncertainty inherent in the observations and can not be explained away with more data~\citep{hullermeier2021aleatoric}. Early works mainly focus on the classification task, modeling such uncertainty by exploiting Dirichlet prior~\citep{malinin2018predictive, malinin2019reverse} or post-training with calibrated loss~\citep{sensoy2018evidential, guo2017calibration, kull2019beyond}. In semantic segmentation, due to the high-dimensional structured output space, aleatoric uncertainty estimation typically employs two kinds of strategies. The first kind~\citep{kendall2017uncertainties,wang2019aleatoric,jungo2020analyzing, ji2021learning} ignore the structure information and may suffer from inconsistent estimation. In the second line of research, \cite{NEURIPS2018_473447ac} first propose to learn a conditional generative model via the cVAE framework~\citep{NIPS2015_8d55a249, Kingma2014}, which can produce a joint distribution of the structured outputs. It is then improved by~\cite{baumgartner2019phiseg, kohl2019hierarchical} with hierarchical latent code injection. Other progresses based on the cVAE framework attempt to introduce an additional calibration loss on the sample diversity~\citep{hu2019supervised}, change the latent code to discrete~\citep{qiu2020modal}, or augment the posterior density with Normalizing Flows~\citep{valiuddin2021improving}.
Recently,~\cite{NEURIPS2020_95f8d990} directly model the joint distribution among pixels in the logit space with a multivariate Gaussian distribution with low-rank parameterization, while~\cite{kassapis2021calibrated} adopts a cascaded framework and an adversarial loss for model learning. Different from those methods, we adopt an explicit multimodal framework training with a Wasserstein-like loss, which enables us to capture the multimodal characteristic of the label space effectively.
\paragraph{Optimal Transport and Wasserstein distance}
The optimal transport problem studies how to transfer the mass from one distribution to another in the most efficient way~\citep{villani2009optimal}. The corresponding minimum distance is known as Wasserstein distance or Earth Mover's distance in some special cases~\citep{ruschendorf1985wasserstein}, which has attracted much attention in the machine learning community~\citep{torres2021survey}.
We apply the Optimal Transport theory to semantic segmentation for calibrating the aleatoric uncertainty, where a Wasserstein-like loss between the predicted distribution and ground truth label distribution is minimized. Unlike most work dealing with fixed marginal probabilities in the Optimal Transport problem, we use parameterized marginal probabilities and jointly optimize marginal probabilities and segmentation predictions. Furthermore, we design a novel constraint relaxation to efficiently compute and optimize the loss.
\paragraph{Mixture of Experts}
The Mixture of Experts model~\citep{masoudnia2014mixture} typically contains several expert models and a gating function. It divides a complex prediction task into different sub-tasks, each modeled by a distinct expert network. We adopt this framework in our task by designing a stochastic expert model with shared segmentation parameters and expert-specific latent prior, enabling us to model the multimodal distribution explicitly.
\paragraph{Other multiple output problems}
There have been many works in the literature on predicting \textit{multiple structured outputs}~\citep{guzman2012multiple,rupprecht2017learning,ilg2018uncertainty}. Unlike ours, those models are not necessarily calibrated and can only produce a limited number of predictions. Another similar task is \textit{image-to-image translation} ~\citep{isola2017image}, which considers transforming images from one style to another and is addressed by learning a conditional generative model. However, these methods~\citep{isola2017image, zhu2017toward, esser2018variational} seek more diverse images and place less emphasis on calibration. In contrast, we aim to model a well-calibrated conditional distribution on segmentation that can generate varying segments during inference.
|
\section*{Introduction}
Improving access to research is a broad and inclusive effort, championed and moved forward by individuals and organizations around the world. In scientific publishing, arXiv has played an important role in open access for over 30 years by removing financial, institutional, and geographic barriers to research.
Truly open access, though, means more than free and available. When we say that arXiv is open we also must ask: open to whom? Accessibility is the practice of \emph{ensuring access regardless of disability}. It is the next frontier in the open science movement.
\section{Barriers to access}
Barriers to accessing research are broad and can be related to a permanent or temporary disability, or situational factors. Vision impairments and learning disabilities can impede access to written material. If content display is not flexible then using a mobile device can be a barrier. Language barriers or lack of internet access are broadly experienced impediments to accessing research.
Countless people around the globe encounter such barriers. People with disabilities are the world’s largest minority (United Nations). More than a quarter of the world’s population has a diagnosed vision impairment (World Health Organization), and vision loss is among the top ten disabilities in the United States (CDC). For much of the world, including english speaking deaf, written English is not their first language (Hrastinski and Wilbur). 20\% of people in the United States have dyslexia (Yale Center for Dyslexia and Creativity), and 26\% of people in the United States self-report living with at least one disability (Elavsky et al.). A wide array of temporary and situational impairments (such as bright or low light conditions, or a temporary loss of vision) affect many more who do not have a permanent disability diagnosis.
\textbf{“Is there really open and immediate access for everyone, if scholars and students with disabilities cannot access and use research articles?”}(Wentz et al.)
In the United States, 14.5\% of K-12 students have a recognized disability (National Center for Education Statistics). Ever increasing numbers of high school students are participating in Computer Science curriculums (Vegas and Fowler) and other STEM fields, so naturally the percent of students with disabilities in these courses is proportionally increasing as well (Code.org). These efforts will, over time, lead to more students with disabilities in higher education who will fully expect and demand equal access to the research output in their field.
Furthermore, in the United States there are clear mandates for equivalent access to websites and federally funded research. The need to make research more accessible to all is evident and pressing.
\quotetext{We can only do so much to welcome people, but if they can’t participate fully it's all just nice talk.}
\quoteattr{—Dr. Kimberly Arcand, data visualizer and science communicator}
Feedback from arXiv users confirms the barriers to current participation:
\textbf{“There are not that many blind mathematicians and scientists because of all the stumbling blocks.”}
\textbf{“Most people veer away from STEM subjects towards text subjects because until a lot more content is accessible they need to work really hard to find it.”}
Despite the wealth of data around needs, and a web of regulations mandating equal access, the vast majority of research papers have low levels of accessibility, creating significant barriers for a large number of people.
\section{Mandates}
Access is a critical effort in academia. In the USA, the White House recently released a policy \href{https://www.whitehouse.gov/ostp/news-updates/2022/08/25/ostp-issues-guidance-to-make-federally-funded-research-freely-available-without-delay/}{mandating open access for all federally funded research}, while \href{https://www.coalition-s.org/}{Plan S} in Europe aims for no science to be locked behind paywalls. Internationally, we have the \href{https://www.go-fair.org/fair-principles/}{FAIR principles} (of which arXiv is a signatory), and the \href{https://en.unesco.org/science-sustainable-future/open-science/recommendation}{UNESCO recommendations on open science}, among others. The movement towards open access is making clear and steady progress.
The movement towards accessibility of research, however, has stalled. Accessibility should be part of every conversation about access. Accessibility is a requirement for most federally funded research in many countries, but in practice only a very low percentage of research papers pass accessibility criteria (Wang et al.). And of these, even fewer pass a more careful human review (Elavsky et al.).
There are formidable obstacles for researchers with disabilities to enter and participate in STEM fields. Foremost among them is equal access to research and data.
\quotetext{The biggest barrier for me is the time you need to spend tracking down formats you can access in order to gain the same information available to everybody else.}
\quoteattr{—Robin Williams, statistician}
Existing legal mandates include making websites and federally funded content accessible to all users. Further regulations continue to gain momentum at the Federal level in the USA (ADA Title III). Furthermore, educational experts have been hard at work for decades transforming primary education to be more inclusive, which will put pressure on all areas of higher education as more diverse students enter STEM, including on repositories of scientific research like arXiv.
\pagebreak
\quotetext{There is work going on earlier in the pipeline, for example making mathematical language accessible to high schoolers. As more people with disabilities enter STEM fields, students will have the expectation of accessing information.}
\quoteattr{—Joe Zesski, Assistant Director, Northeast ADA Center}
Most importantly, arXiv has a mandate from our global community to improve access. Scientists need this and the research community is asking for it. One way of meeting these mandates is through well formatted HTML.
HTML is also the foundation of machine readability with major repercussions for the future of scientific discovery. The science that researchers share on arXiv is important, and automated research flows using machine learning are increasingly relied upon for accelerating discovery.(Policy and Global Affairs et al.) Semantic markup and accessibility for assistive technology is a good first step towards ensuring the arXiv repository better supports emerging approaches to conducting science.
arXiv’s long-term mission is simply to serve the needs of the scientific community. Everyone should be able to participate in the wealth of scientific knowledge contributed to arXiv by researchers from all over the world. Accessibility is the next logical step.
\quotetext{Research should be accessible for everyone in the broadest possible way.}
\quoteattr{—Dr. Shiri Azenkot, Associate Professor and Accessibility researcher}
\section{Current state of research accessibility}
Scientific research is primarily shared in PDF format, which mimics the printed page. It has been the primary format for decades, so any analysis of the accessibility of research inevitably becomes—in part—an analysis of the PDF format itself.
For authors, PDF offers a reliable visual format for sharing and disseminating their work. For publishers, it offers an archival format that won’t change over time. The accessibility of PDFs can be improved by tagging and adding descriptions.
However, the format has serious limitations that have had a profound impact on the accessibility of research. PDF is not a suitable format for the web and has low native accessibility. PDFs are challenging for people with a variety of reading disabilities, including blindness, low vision, dyslexia, and more. It is time consuming to improve the accessibility of PDFs, and even if this work is done it has no effect on its poor mobile performance.
Research from the Allen AI institute reports a miserably low accessibility rate for research papers:
\textbf{“PDF accessibility adherence is low across all fields of study. Of the five accessibility criteria we assess, only 2.4\% of the PDFs we assess demonstrate full compliance.”}(Wang et al.)
\quotetext{The best PDF will ever achieve is what HTML delivers. All it can do is catch up.}
\quoteattr{—Dr. Jonathan Godfrey, Senior Lecturer in Statistics}
Screen readers rely on semantic markup (for example headers, images, formulas, and so on) to correctly interpret content. PDFs do not natively include semantic properties, and it must be tagged with this information after the fact to make it accessible. Tagging is time consuming, takes specialized knowledge, and requires proprietary tools. These tools (and the expertise to use them) are not free, nor are they intuitive. Adobe’s own manual for making PDFs accessible is 94 pages long! (Elavsky et al.) In addition, if you find you need to regenerate the PDF again for any reason, you must tag the document again from scratch.
It is perhaps not surprising then that efforts to promote tagging PDFs for accessibility have not become the norm in academic publishing.
\textbf{“In CHI 2014, a year in which considerable effort was spent giving author’s feedback on the accessibility of their documents, only 26.8\% included any document tags at all.”}(Bingham et al.)
\textbf{“The criterion with the lowest rate of compliance is Alt-text, which has remained stable between 5–10\% and has been lower in recent years. Since Alt-text is the only criterion of the five which always necessitates author intervention, we believe this is a sign that authors have not become more attuned to accessibility needs.”}(Wang et al.)
Even when this manual work is done the resulting PDF is still only partially accessible: Two column layouts often confuse screen reader software; Text and graphics don’t reflow in mobile devices or with magnification; There is limited parsability for third party tools; And if math and data visualizations did have rich markup in the original source—such as a TeX formula or a SVG graph—the data are often flattened when the PDF is generated.
The typesetting software used to generate PDF results in differing levels of accessibility, with Microsoft Word producing the highest level of compliance, and LaTeX producing the lowest (Wang et al.). In the scientific fields that arXiv covers, LaTeX is widely used, and 90\% of arXiv’s corpus in recent years was submitted as TeX source. The LaTeX core team is working on generating more accessible PDFs, and we plan to incorporate this work into arXiv when it is released.
On mobile devices, PDFs are far behind standard and don’t meet basic user expectations. Research done by Adobe found that 65\% of Americans find consuming content on mobile frustrating; 45\% stopped reading or didn’t even try; and 72\% say they would work on their mobile devices more if it were easier to read documents (Adobe). Mobile needs in academia vary widely, from a researcher catching up on a paper while traveling, to one with limited access to technology using their mobile phone as their only option for reading papers. Flexible reflow of content is a must to expand access and efficiency.
ACM conference have also found low accessibility rates of PDF, despite significant effort and have chosen HTML5 format as a goal to \textbf{“ultimately make accessibility easier and more standardized.”} (Mankoff, et al.)
\quotetext{You can make PDF accessible, but screen readers are much more efficient when working with HTML. To tag PDF you need specialized skills and tools. For HTML, all these things are comparatively easier.}
\quoteattr{—Avneesh Singh, chair, accessibility task force - W3C EPUB 3 Working Group and Publishing Community Group}
PDF is the current standard, but our user research tells us that providing HTML—in addition to PDF and TeX source—will substantially increase accessibility to research posted to arXiv. Well formatted HTML will support and empower the many different ways that scientists consume research data.
\subsection{Existing tools}
A great deal of work has been done by sung and unsung heroes in the accessibility space, and as arXiv explores our role in making research more accessible we know that we are standing on the shoulders of giants. Critical and transformative work has been achieved in the standardizing and formatting of math on the web; converting TeX to HTML (and PDF, Word and other formats to HTML as well); guiding researchers in writing alternative content for figures and images; making data in charts and graphs parseable by screen readers; translation services (including for Braille); identifying and measuring accessibility on the web; and the incredibly important and advanced work that has gone into screen readers and haptic displays.
We also want to take the opportunity to acknowledge the countless scientists with disabilities who have invented and contributed to tools, standards, and community understanding over their careers. We have spoken to so many scientists who had to first invent or build the tools they use just to participate in their field. They have paved the way for countless others, as well as for the assistive software in use today. They donate their time to open source tools, guide new scientists in navigating the confusing landscape of access, generously share their expertise (including with arXiv), and endlessly advocate for equal access. All the while overcoming the daily obstacles they face in their own work. Thank you.
To the professors and teachers helping students overcome the formidable barriers built into how we share and publish research, thank you. To the researchers who make their work accessible, thank you.
We can make things easier and more equal for our colleagues. Making research available to everyone regardless of disability is the next stage of Open Science.
\subsection{The tool we need}
You might ask, with so many tools already built, why do we need anything else? After thorough analysis we found that none of the existing tools, as they are, can provide a smooth experience for authors and readers. Authors should be able to submit their work with the software they currently use and without being an accessibility expert. To achieve that will require new technological solutions as well as cultural changes beyond arXiv, but we have an important role to play.
Our goal is to be able to say to the arXiv community: you bring your expertise in your field; we will help close the gap on all the rest. Thanks to the work of so many in the broad and complex field of accessibility, we believe this is achievable today, though still not easy.
\section{Researching accessibility in the arXiv ecosystem}
We undertook to research the experience of people with a variety of disabilities and other barriers as it relates to accessing research articles, on arXiv and beyond. We also gathered input from experts in different fields: accessibility researchers; writers of web standards for accessibility, Math, and PDF; TeX and LaTeX experts; developers of screen reader and other assistive technologies; scholars of accessibility law; and science communicators.
Our research took two forms, a quantitative survey and a series of qualitative interviews.
\subsection{Survey}
We developed a survey to investigate behavior and preferences related to accessing content on arXiv.
\subsubsection{Process}
We invited two groups of people to take this survey: researchers who directly rely on assistive technology, and professors and colleagues who assist researchers in accessing the research and data they need.
Out of a pool of 275 volunteers we had a response rate of 18\% and received a total of 53 individual responses. Though this number appears small, it is above the industry expected ~6\% when taking into account the long length and detail of the survey. The arXiv community again proved its generosity of spirit.
\subsubsection{Demographics}
Our respondents were primarily frequent arXiv users. 58\% use arXiv daily and 25\% weekly. 25\% of respondents directly use assistive technology, while 75\% are not direct users, but are involved in assisting people who are.
Respondents represent a variety of fields, with the highest number of respondents (*42\%) from Physics, followed by Math and Computer Science.
\begin{table}[!ht]
\centering
\begin{tabular}{ll}
\hline
Field
& Percentage
\\ \hline
Physics
& 31\%
\\ \hline
Mathematics
& 21\%
\\ \hline
Computer Science
& 13\%
\\ \hline
Other
& 13\%
\\ \hline
Astrophysics
& 12\%
\\ \hline
Engineering
& 6\%
\\ \hline
Biology
& 4\%
\\ \hline
\end{tabular}
\end{table}
Respondents come from various, but not all, geographic regions. Europe had the highest representation at 44\%, followed by Asia at 23\%, North America at 21\%, South America at 8\%, and Africa at 4\%. We continue to work towards greater global representation in our surveys.
The assistive technology reported in use by our respondents are:
\begin{table}[!ht]
\centering
\begin{tabular}{ll}
\hline
Assistive technology
& Percentage
\\ \hline
Screen reader
& 43\%
\\ \hline
Adjust screen color/contrast
& 24\%
\\ \hline
Magnifier
& 24\%
\\ \hline
Voice command
& 10\%
\\ \hline
\end{tabular}
\end{table}
\subsubsection{Survey Results}
\textsc{Access to research output}
\textbf{Summary:} Our respondents heavily depend on access to research, but users of assistive technology report they only have access to 38\% of the research they need without assistance. Overall, participants reported that access has improved somewhat over the last five years.
Our respondents heavily depend on access to research, with 89\% saying that research is completely or somewhat essential to their professional work:
\begin{table}[!ht]
\centering
\begin{tabular}{lll}
\hline
Dependence level
& Yes, use assistive technology
& No, don’t use assistive technology
\\ \hline
Completely Essential
& 80\%
& 78\%
\\ \hline
Somewhat Essential
& 10\%
& 11\%
\\ \hline
Somewhat Optional
& 10\%
& 8\%
\\ \hline
Completely Optional
& 0\%
& 2\%
\\ \hline
\end{tabular}
\end{table}
We next asked what level of access respondents have today, without requiring assistance from others. Overall the numbers were high: 89\% report having access to all or most of the research they need without assistance. However, those numbers look quite different if the respondent uses assistive technology, with only 30\% reporting access to all papers without assistance:
\begin{table}[!ht]
\centering
\begin{tabular}{lll}
\hline
Current access
& Yes, use assistive technology
& No, don’t use assistive technology
\\ \hline
All research is accessible
& 30\%
& 56\%
\\ \hline
Most research is accessible
& 40\%
& 38\%
\\ \hline
About half is accessible
& 10\%
& 6\%
\\ \hline
Most is not accessible
& 20\%
&
\\ \hline
\end{tabular}
\end{table}
We also asked whether access to research had improved in the last five years, and for the majority of users, including those who use assistive technology, it has improved at least somewhat:
\begin{table}[!ht]
\centering
\begin{tabular}{lll}
\hline
Improvement level
& Yes, use assistive technology
& No, don’t use assistive technology
\\ \hline
It has improved a lot
& 40\%
& 47\%
\\ \hline
It has improved a little
& 20\%
& 24\%
\\ \hline
It is about the same
& 30\%
& 26\%
\\ \hline
It is worse
& 10\%
& 3\%
\\ \hline
\end{tabular}
\end{table}
\rule{480pt}{.5pt}
\textsc{Barriers to access}
\textbf{Summary:} Survey respondents agree that PDF formatting is the biggest barrier. The main reason reported for not submitting accessible papers is a lack of understanding around requirements.
When asked what the biggest barriers were to accessing papers, PDF formatting topped the list:
\begin{table}[!ht]
\centering
\begin{tabular}{lll}
\hline
Barriers
& Yes, use assistive technology
& No, don’t use assistive technology
\\ \hline
PDF formatting
& 22\%
& 25\%
\\ \hline
Images
& 15\%
& 10\%
\\ \hline
Math
& 12\%
& 12\%
\\ \hline
Graphs and charts
& 10\%
& 11\%
\\ \hline
TeX macros
& 7\%
& 10\%
\\ \hline
Language
& 7\%
& 5\%
\\ \hline
Font size
& 7\%
& 6\%
\\ \hline
Colors
& 7\%
& 2\%
\\ \hline
Other
& 5\%
& 4\%
\\ \hline
Contrast
& 5\%
& 4\%
\\ \hline
arxiv.org website
& 3\%
& 11\%
\\ \hline
\end{tabular}
\end{table}
The arxiv.org website is not a significant barrier to users of assistive technology, which our interviews also highlighted for us and is good news. But we do not ignore that the website has barriers unrelated to accessibility (these include limited coverage of related fields and a poor search function), though beyond the scope of this report.
We asked respondents the reasons that stopped them from submitting accessible papers.
\pagebreak
\begin{table}[!ht]
\centering
\begin{tabular}{lll}
\hline
Reason
& Yes, use assistive technology
& No, don’t use assistive technology
\\ \hline
Their own papers are \\
accessible already
& 50\%
& 21\%
\\ \hline
Not knowing what an \\
accessible paper requires
& 25\%
& 43\%
\\ \hline
arXiv requires submitting \\
TeX, even if you have created \\
an accessible PDF
& 25\%
& 25\%
\\ \hline
Lack of accessibility mandates
& 25\%
& 11\%
\\ \hline
Conference or journal \\
guidelines are not accessible
& 0
& 11\%
\\ \hline
Not knowing how to make \\
TeX accessible
& 0
& 25\%
\\ \hline
No access to authoring \\
tools to create accessible \\
content
& 0
& 11\%
\\ \hline
\end{tabular}
\end{table}
For users who do not use assistive technology the top reason was not knowing what an accessible paper requires, while for users of assistive technology it was that they consider their papers to be accessible already.
\rule{480pt}{.5pt}
\textsc{Preferred format}
\textbf{Summary:} Respondents who use assistive technology preferred HTML, while those who didn't preferred PDF. Use of specific types of assistive technology, such as screen magnification and color and contrast remediation, correlated with a strong preference for HTML.
We asked survey respondents what their preferred format is for reading papers. Interestingly, the survey responses around format contradict the results from our interviews, where HTML was very strongly indicated as the preferred format for accessibility. We conjecture that this difference in result is due to several factors:
\begin{enumerate}
\item Our interview participants include a high percentage of accessibility and standards professionals, while the survey respondents are predominantly researchers who are less familiar with accessibility and the pros and cons of different formats.
\item It is difficult to predict future behaviour, even our own. The dominance of the PDF format in scientific publishing—and its resultant workflows—makes it difficult to imagine alternatives.
\item Many researchers like PDF. Based on user feedback it is clear that new formats for publishing must appear alongside existing options so researchers can interact with papers in their preferred way.
\end{enumerate}
When asked what their preferred format would be \emph{if well formatted HTML was available}, 67\% still indicated PDF would be preferred; among assistive technology users a small majority of 55\% would prefer HTML:
\pagebreak
\begin{table}[!ht]
\centering
\begin{tabular}{lll}
\hline
Preferred format
& Yes, use assistive technology
& No, don’t use assistive technology
\\ \hline
HTML
& 55\%
& 29\%
\\ \hline
PDF
& 45\%
& 67\%
\\ \hline
TeX
& 0\%
& 4\%
\\ \hline
\end{tabular}
\end{table}
When respondents were asked which format would be \emph{most useful} to them, there was a strong split between those who use assistive technology, and those who do not:
\begin{table}[!ht]
\centering
\begin{tabular}{lll}
\hline
Useful format
& Yes, use assistive technology
& No, don’t use assistive technology
\\ \hline
Well formatted HTML
& 45\%
& 16\%
\\ \hline
Well formatted PDF
& 45\%
& 69\%
\\ \hline
\end{tabular}
\end{table}
Interestingly, respondents with specific barriers indicated a preference for HTML: those who need to adjust font size, color, or contrast; all barriers which indicate a vision impairment.
Also preferring HTML are professors who help students with translation or by describing images and charts. One professor described how HTML would help him: \emph{“I translate material into braille for one user, which is highly specialized, but starting from html is much better than pdf.”}
\rule{480pt}{.5pt}
\textsc{Additional suggestions}
Respondents were asked to select from a list of potential site improvements which they would find useful. The following two changes had the highest positive response rate:
\begin{enumerate}
\item The ability to build a customizable arXiv feed (73\%), and
\item A quick way to get to a paper’s conclusions and references (58\%).
\end{enumerate}
Other suggestions included:
\begin{itemize}
\item Coverage of more scientific fields
\item More use of AI and machine learning for categorization and discovery
\item The option to enlarge fonts
\item Improve TeX upload function
\item Add dark mode, and
\item To keep arXiv going:
\end{itemize}
\textbf{“I REALLY appreciate the arXiv service. I can't afford to subscribe to a plethora of physics journals. Hats off to the arXiv team!”}
\pagebreak
\subsection{Interviews}
To better understand the ecosystem arXiv is operating in, we interviewed a wide range of researchers from the larger arXiv community.
\subsubsection{Process}
We interviewed a total of 44 individuals. They include researchers with reading disabilities or other access barriers who use arXiv, professors who help students with disabilities, researchers whose focus is on various fields of accessibility, experts on standards for web content, and leaders in LaTeX, MathML, and other languages critical to the success of this project. Some participants fit into more than one category.
Our participants were recruited through our accessibility survey, two accessibility mailing lists, direct invitations, and word of mouth. The mode for interview length was 30 minutes with some going longer.
We conducted semi-structured interviews with individuals or small groups using video conferencing tools, and in two cases in person. An interview guide was developed and used as a loose guiding tool, but prioritizing the participants' conversational lead: we wanted to learn what they most wanted us to know. All participants were interviewed by the main interviewer, assisted by other members of arXiv staff and student assistants.
All personally identifying data relating to interviewees is omitted in this report in order to preserve anonymity.
\subsubsection{Demographics}
Our participants were diverse in terms of their career stage and included PhD students, professors, and researchers working in industry. Participants come from multiple fields of research including physics, math, statistics, computer science, legal, and regulatory. We interviewed a number of participants who serve on various W3C boards related to accessibility and web standards.
Participants were not asked to disclose their disability, but were asked to describe if and how their disability affected their access to research. In responding to this question participants disclosed the following: 7 participants disclosed blindness, 1 participant disclosed dyslexia, 1 participant disclosed ADHD, and 2 participants disclosed a movement disability.
\subsubsection{Analysis}
All interviews were transcribed by either the main interviewer, or by student assistants and then reviewed by the main interviewer. Transcripts were broken down into observations, then documented following Atomic Research principles by mapping each observation to a semantic layer to facilitate discovery.
All observations were anonymized prior to thematic analysis to protect the privacy of our interviewees. During analysis we looked for similarities, grouped them into themes, and compared to data from the survey and looked for disparities and correlations.
\subsubsection{Interview Results}
{\Large\textbf The Accessible User Journey}
To evaluate the user journey we sorted feedback into five primary user goals: Find Research, Read Research, Participate in Scholarship, Prepare my Document, and Submit:
\begin{figure}[htp]
\centering
\includegraphics[scale=0.5]{./images/pos-neg-stacked.png}
\caption{This graphic displays a stacked bar chart for each step of the user journey showing the percent of positive vs. negative experiences. Find research: 38\% negative, 62\% positive. Read papers: 83\% negative, 17\% positive. Participate in scholarship: 67\% negative, 33\% positive. Write papers: 89\% negative, 11\% positive. Submit papers: 57\% negative, 43\% positive.}
\label{fig:stacked bar chart of each step of the user journey}
\end{figure}
When analyzing each of these steps we asked: “when it comes to accessing research, was the experience of the participant positive or negative?” Most steps were dominated by negative experiences, with only Find Research being mostly positive. An accessible table display of the user journey, as well as the anonymized qualitative data behind it, is available in our \href{https://docs.google.com/spreadsheets/u/0/d/1Uvb-A1ePpYyWETxAizVI7ExHobCCSbuLczwxkFN6Ss4/edit}{User Journey Table}.
\emph{Read Research} elicited the highest number of comments, followed by \emph{Prepare my Document}:
\begin{figure}[htp]
\centering
\includegraphics[scale=0.4]{./images/pos-neg.png}
\caption{This graphic displays each of the five steps as an individual positive/negative bar chart. The number of bars at each step shows the number of experiences, while the height of each bar represent the impact score. Varying numbers of experiences were shared for each step, with Read eliciting the most by far. Find research: 12. Read papers: 57. Participate in scholarship: 17. Write papers: 27. Submit: 14.}
\label{fig:positive/negative bar charts for each of the five steps in the user journey}
\end{figure}
\pagebreak
{\Large\textbf Themes}
Our thematic analysis of interview feedback identified 5 themes:
\begin{enumerate}
\item The PDF format as a barrier.
\item The benefits of HTML as a format.
\item Skepticism on the potential for real change.
\item arXiv has a role to play in improving the accessibility of research papers.
\item HTML is just the beginning.
\end{enumerate}
\bigskip
\rule{480pt}{.5pt}
\textsc{Theme 1: researchers who rely on assistive technology point to PDF as the biggest stumbling block to accessible research.}
Participants who rely on assistive technology are particularly aware of—and frustrated with—its shortcomings:
\textbf{“Forget PDF!”}
\textbf{“I would prefer HTML over PDF.”}
\textbf{“You can make a pdf accessible but it is still a PDF (poor mobile use, no magnification).”}
\textbf{“PDF is not the best thing from an accessibility point of view.”}
\textbf{“[with PDF] I need to borrow someone and say ‘hey can you help me look at this data set.’”}
\textbf{“Screen reader does not do equations. Figures are a crapshoot.”}
\textbf{“Even accessible PDF, that just means they are tagged. But how do you read math in PDF?”}
\textbf{“I have walked away from PDF. Why are you wasting your time and mine?”}
For researchers who \emph{do not} rely on assistive technology, and for whom PDF already is sufficient, is is challenging to visualize accessing research in new ways that might break their current PDF-dependent workflow:
\textbf{“For scientific documents PDF is still excellent, in my opinion. I would not be that interested in HTML papers.”}
\textbf{“It is universal, and can keep things quite simple.”}
\textbf{“I generally prefer PDF… it’s useful to have it as a single file to add it to Zotero or wherever.”}
Some interviewees who initially dismissed HTML point out its benefits as they thought about it more:
\textbf{“I want the PDF so I can add it to Paper Pile… Although, [HTML] would make it easier to quote parts of a paper.”}
Making PDF accessible is challenging:
\textbf{“For PDF you need specialized people who understand the PDF standard. For tagging, you need tools from Acrobat and skilled people.”}
\bigskip
\rule{480pt}{.5pt}
\textsc{Theme 2: HTML mitigates a wide range of access issues, positively addressing many disabilities and improving the experience for most users.}
HTML has a significant edge for researchers with disabilities:
\textbf{“I prefer HTML versions. As an assistive tech user I find it much faster to navigate.”}
\textbf{“HTML also works for the deaf community. English is a second language to deaf readers… they would rather get ASL. There are tools for converting HTML to ASL.”}
\textbf{“What I hear from colleagues in astronomy who are blind or low vision is that HTML is the preferred delivery mechanism. It is the most accessible.”}
\textbf{“HTML gives you a lot more freedom. You can use software to change colors or typefaces.”}
\textbf{“I don’t see how an accessible PDF can be better than accessible HTML. I don’t see the upside to it.”}
\textbf{“With the HTML version, you can throw MathJax on your website as well, then be able to render that MathML and have it spoken correctly to AT users.”}
\textbf{“Screen readers are more efficient while working with HTML.”}
HTML offers benefits beyond disability access, such as for machine learning and mobile phones:
\textbf{“The parsability and adaptability of HTML is better than PDF.”}
\textbf{“With HTML, [researchers] can adjust the content to their needs. I'm all for HTML.”}
\textbf{“The PDF format is not the best suited to mobile phones.”}
\textbf{“Reading pdf articles on a smartphone is anyway difficult.”}
\textbf{“HTML5 is the right basis for longevity.”}
HTML also offers greater reduction of legal risks. Though not yet specified, the consensus of the legal scholars we spoke to was that legal mandates in the USA for user-uploaded content are inevitable and already underway:
\textbf{“Long term, user uploaded content will need to be accessible. But the standards aren't there yet.”}
\textbf{“Given how large your scope is it’s good that you are doing this now… you have such a massive volume of material it could be a headache.”}
\textbf{“arXiv is covered by the ADA. There is no pass for third-party submitted content.”}
\bigskip
\rule{480pt}{.5pt}
\textsc{Theme 3: There is deep skepticism, and even a sense of despair, around the likelihood of real change.}
Researchers with disabilities feel their advocacy has gone unheard:
\textbf{“I have waited 20 years for PDF to do anything with accessibility. That is half a working lifetime.”}
\textbf{“Progress is not moving anywhere on accessible papers.”}
\textbf{“For a long time now accessibility has been viewed as a 'nice to have.' Pushing accessibility up the agenda is the biggest thing that can happen.”}
Making all research accessible is an enormous challenge that extends beyond arXiv:
\textbf{“arXiv can’t do it without the involvement of ordinary people.”}
\textbf{“Encouraging authors to tag will only be successful if the main players apply the new standards together.”}
\textbf{“It's an upward struggle because of how the blind reader space works. Screen readers want users to use their tool and don't provide APIs.”}
\textbf{“Even with much bigger technological steps you will still need buy-in from authors to do this.”}
We repeatedly heard we should not let the perfect be the enemy of good. Some improvement in accessibility is much better than no improvement:
\textbf{“There is no silver bullet. No point in waiting or looking for one, when you have tools that will do 80\% of the work now.”}
\textbf{“70\% great and another 15\% good is not a terrible success rate.”}
\textbf{“Even a small stride is a great leap for the people it helps. Don't be discouraged.”}
Participants called out the need for more authoring tools and education if this effort is to succeed:
\textbf{“I think one of the biggest problems that we have is the paucity of tools that make it easy to make HTML.”}
\textbf{“The biggest problems [in making research papers accessible] is communicating the importance of this to authors. I hope awareness will grow over time.”}
\textbf{“There needs to be more [accessibility] guidance, especially if you want authors to do some of the work.”}
\bigskip
\rule{480pt}{.5pt}
\textsc{Theme 4: We heard from our community that arXiv can have a powerful and immediate impact.}
We are part of an ecosystem, and we have a role to play in making research accessible:
\textbf{“arXiv is well set up [to impact accessibility] in a way that a single publisher isn't.”}
\textbf{“Your job at arXiv is just to put out proper HTML.”}
\textbf{“In a dreamworld every technical document will just work [with screen readers]. In a smaller dream world, it just works on arXiv.”}
\textbf{“My area in physics is stuck in its ways. Not many people want to go against the norm. And arXiv sets the norm.”}
Because arXiv has direct control over the submission pipeline, it opens up opportunities:
\textbf{“The good news is that the potential for significant impact is just right there. The fact that you are interested in this is very exciting to me. We could shift how we author papers in a way that doesn't add burden but makes life easier. It is a value-add for everyone in authoring and output.”}
\textbf{“If someone is in control of the pipeline there is a chance accessibility will get built in.”}
\textbf{“There is no polished product you can buy somewhere but all the pieces are coming together. And if we can show the user what they need to do on their end, this could work very beautifully. ArXiv has the content, and you are compiling your own content, and you have the traffic.”}
\textbf{“Your push on the pipeline-based tool is the right way because you cannot scale up on it in any other way for it to be automated.”}
\bigskip
\rule{480pt}{.5pt}
\textsc{Theme 5: HTML is just one step towards a greater vision of accessibility.}
Research is consumed in different ways. The ultimate goal is flexibility and giving readers the ability to access the content in the way that works best for them:
\textbf{“In a perfect world I could be given the source files and reproduce and be in total control of the method I choose to consume.”}
\textbf{“I have a right to the information that the author was trying to communicate… There is an ethos in the blind community in [Country] that everyone consumes PDF, so we have the right to consume it as well. I believe I have the right to consume the content.”}
\textbf{“It really helps when people include the source.”}
\textbf{“I have written an add-on that can sonify a series... higher pitch for higher values and so on.”}
\textbf{“The ideal is that I get to choose how much information I skip over in the same way a sighted person does.”}
Current accessibility standards are only a starting point:
\textbf{“Accessibility has limitations; it is the bare minimum ground in its standards and is too legally driven. But the work of real access, making things work for people with disabilities, requires more.”}
Many interviewees with disabilities pointed out that a paper is just a portal to the knowledge behind it. They don’t necessarily need access to the paper if they can absorb the research itself:
\textbf{“Anything like tables are super helpful for me. Source data, or code, is much easier to understand. Then you don’t need to read the paper. If it is well written code then I would prefer that to reading the paper.”}
\textbf{“I would prefer if there was raw data available in an excel spreadsheet or similar. Then I can find how to make sense of the data. I would compile the statistics.”}
\section{arXiv's Plan for accessible research papers}
HTML is an even older and more established standard than PDF. In fact, HTML was invented to facilitate the sharing of scientific knowledge (CERN). It has well defined criteria for achieving accessibility on the web and does not require proprietary tools to author or consume. HTML also provides a better foundation for machine readability, and can help usher in the next generation of tools that will help us all find and access research more efficiently.
Of course, HTML is not automatically accessible. When we refer to HTML in our accessibility plan we mean well-formatted, semantic HTML with necessary ARIA tags. There will be some limitations to what we can do based on the richness of the original uploaded TeX, but our plan is to achieve the most accessible HTML possible within those constraints.
PDF can theoretically be tagged for accessibility, too. But as presented in this paper, the reality is that only about 2.4\% of PDFs are accessible and there are significant barriers to improving that number. One promising effort is that the LaTeX core working group are addressing accessibility now, and we plan on incorporating their work as soon as it is available so that we can also provide more accessible PDF documents on arXiv.
\pagebreak
Based on provided feedback, we have rated how well we expect the PDFs that arXiv generates now vs the well-formatted HTML that our plan will provide to see how they score on a number of criteria in the table below.
\textbf{arXiv assessment of HTML and PDF}
\small \emph{Scale: 0=non-functional, 1=OK, 2=good, 3=excellent.}
\begin{table}[!ht]
\centering
\begin{tabular}{lll}
\hline
Criteria
& PDF
& Well-formatted HTML
\\ \hline
Screen reader legible text
& 2
& 3
\\ \hline
Screen reader legible math
& 1
& 3
\\ \hline
Screen reader legible charts
& 0
& 0-1
\\ \hline
Screen reader legible images
& 0
& 0-1
\\ \hline
Screen magnifier compatibility
& 0
& 3
\\ \hline
Colors and contrast adjustment
& 0
& 3
\\ \hline
Mobile friendly
& 0
& 3
\\ \hline
Machine readability
& 1
& 3
\\ \hline
Portability
& 3
& 2
\\ \hline
Archival nature
& 3
& 3
\\ \hline
Ability to make accessible
& 1
& 2
\\ \hline
Established use in academia
& 3
& 0
\\ \hline
Open source
& 1
& 3
\\ \hline
Adjustable content*
& 0
& 3
\\ \hline
Legal risk mitigation
& 1
& 3
\\ \hline
Total score
& PDF: 16
& HTML: 35
\\ \hline
\end{tabular}
\end{table}
\emph{*examples: to meet publishers' requests to hide author names while the paper is under double blind review, or promoting best practices by displaying the license the author chose during submission on the work itself}
To offer the flexibility of well formatted HTML downstream requires, ironically, restrictions upstream during content creation. Well structured, parseable content that follows established standards must be either provided or generated during submission.
And this is the difficulty. 90\% of arXiv submissions are provided as TeX (mainly LaTeX), and converting to HTML is not easy due to its extensibility.
\quotetext{On the one hand it's great that LaTeX is so extensible. On the other hand it is such a pain that it is so extensible.}
\quoteattr{—Frank Mittelbach, Head of Development, LaTeX group}
Incorporating the conversion into arXiv’s submission process will mean substantial changes to the pipeline behind the scenes, not visible to the author or affecting their submission experience. However we will need to lean on author engagement in two ways: to add alternative text for images and other content that can’t be parsed, and to view and approve their HTML output before submitting, just as they do their PDF output today. arXiv will then make this content available directly on the website alongside the PDF and TeX source.
Because the relationship between authors and the arXiv platform is direct—with no third party typesetting the documents before publishing—we have a tremendous opportunity to make small changes in the submission pipeline that have profound results on accessibility.
\quotetext{arXiv has a closeness to the practitioner that is exciting for accessibility. A lot of remediations require a human touch.}
\quoteattr{—Frank Elavsky, Data Visualization and Accessibility expert}
\section*{Conclusions}
The level of accessibility of research papers is low, and we cannot claim to have achieved truly open science while those with disabilities are barred from equivalent access. Based on our user research, the step our community wants arXiv to take is clear: offer well formatted, accessible HTML alongside existing sources.
Adding HTML will allow all researchers to experience its benefits, try new workflows, and adjust how papers are authored over time. It will support existing and emerging assistive technologies that work most efficiently with HTML, and normalize the format across more fields.
It is not an easy goal, but it is an achievable one. And because of arXiv’s reach across many fields and control over the submission pipeline we are positioned to leverage HTML in an impactful way.
\section*{References}
Adobe. “The case for reimagining reading.” Adobe Blog, 5 December 2018, \sloppy
\url{https://blog.adobe.com/en/fpost/2020/reimagining-reading-infographic}.
Bingham, Jeffrey P., et al. “An Uninteresting Tour Through Why Our Research Papers Aren’t Accessible.” ACM CHI, 2016. \sloppy
\url{http://dx.doi.org/10.1145/2851581.2892588}.
Centers for Disease Control and Prevention. “Burden of Vision Loss.” Centers for Disease Control and Prevention, 2020, \sloppy
\url{https://www.cdc.gov/visionhealth/risk/burden.htm}.
CERN. “The birth of the Web.” CERN, \sloppy
\url{https://home.cern/science/computing/birth-web}.
Code.org Advocacy Coalition, et al. “2022 State of Computer Science Education.” Code.org Advocacy Coalition, 2022, \sloppy
\url{https://advocacy.code.org/2022\_state\_of\_cs.pdf}.
Elavsky, Frank, et al. “How accessible is my visualization? Evaluating visualization accessibility with Chartability.” Computer Graphics Forum, vol. 41, no. 3, 2022, \sloppy
\url{https://www.frank.computer/chartability/}.
Hrastinski, Iva, and Ronnie B. Wilbur. “Academic Achievement of Deaf and Hard-of-Hearing Students in an ASL/English Bilingual Program.” Journal of Deaf Studies and Deaf Education, vol. 21, no. 2, 2016. \sloppy
\url{https://www.jstor.org/stable/26172441}.
Mankoff, Jennifer, et al. “2019 Access SIGCHI report.” ACM SIGACCESS Accessibility and Computing, no. 126, 2020. \sloppy
\url{https://doi.org/10.1145/3386280.3386287}.
National Center for Education Statistics. “Children 3 to 21 years old served under Individuals with Disabilities Education Act (IDEA), Part B, by age group and sex, race/ethnicity, and type of disability: 2020-21.” National Center for Education Statistics, 2021, \sloppy
\url{https://nces.ed.gov/programs/digest/d21/tables/dt21\_204.50.asp}.
Policy and Global Affairs, et al. Automated Research Workflows for Accelerated Discovery: Closing the Knowledge Discovery Loop. National Academies Press, 2022, \sloppy
\url{https://nap.nationalacademies.org/catalog/26532/automated-research-workflows-for-accelerated-discovery-closing-the-knowledge-discovery}. Accessed 14 October 2022.
United Nations. “Factsheet on Persons with Disabilities.” United Nations, Department of Economic and Social Affairs, \sloppy
\url{https://www.un.org/development/desa/disabilities/resources/factsheet-on-persons-with-disabilities.html}. Accessed 28 September 2022.
Vegas, Emiliana, and Brian Fowler. “What do we know about the expansion of K-12 computer science education?” Brookings Institution, 4 August 2020, \sloppy
\url{https://www.brookings.edu/research/what-do-we-know-about-the-expansion-of-k-12-computer-science-education/}.
Wang, Lucy, et al. “Improving the accessibility of scientific documents.” arXiv, 2021. arxiv.org, \sloppy
\url{https://arxiv.org/pdf/2105.00076}.
“Website Accessibility Regulations On The Horizon: DOJ To Start Title II Rulemaking For State and Local Governments Next Year.” ADA Title III, Seyfarth, 29 July 2022, \sloppy
\url{https://www.adatitleiii.com/2022/07/website-accessibility-regulations-on-the-horizon-doj-to-start-title-ii-rulemaking-for-state-and-local-governments-next-year/}.
Wentz, Brian, et al. “A Socio-Legal Framework for Improving the Accessibility of Research Articles for People With Disabilities.” Journal of Business \& Technology Law, vol. 16, no. 223, 2021. DigitalCommons@UM Carey Law, \sloppy
\url{https://digitalcommons.law.umaryland.edu/cgi/viewcontent.cgi?article=1333\&context=jbtl}
World Health Organization. World Report on Vision. 2019, \sloppy
\url{https://www.who.int/publications-detail-redirect/9789241516570}.
Yale Center for Dyslexia and Creativity. “Dyslexia FAQ.” Yale Center for Dyslexia, \sloppy
\url{https://dyslexia.yale.edu/dyslexia/dyslexia-faq/}.
\end{document}
|
\section{Introduction}
\label{sec:introduction}
It is well established that galaxy mergers play a fundamental role in the growth of galaxies in the Universe. In addition to the assembly of the main matter components, the merger process is predicted to have short-lived, but dramatic, effects on numerous galactic properties.
The most obvious visible impact of a galaxy-galaxy interaction is the re-arrangement of stellar material, an effect which was seen in the earliest N-body simulations \citep[][]{Toomre1972,White1978,Roos1979,Villumsen1982}. Strong tidal forces result in disturbed morphologies including shells and tidal features that are dependent on the properties of both the initial galaxy members and the orbital properties of the interaction \citep[][]{Casteels2014,Nevin2019,Blumenthal2020,Patton2016}. The advent of hydrodynamical simulations further demonstrated that the generation of internal asymmetric structures produced torques that could drain angular momentum from the gas, causing it to flow to the centre and resulting in central starbursts \citep[][]{Hernquist1989a,Barnes1991,Mihos1996,DiMatteo2007,Capelo2016,Blumenthal2018}, a feature that is widely observed in both the pre- \citep[][]{Barton2000,Woods2006,Woods2007,Ellison2008,Woods2010,Scudder2012,Patton2013,Knapen2015,Cao2016} and post- \citep[][]{Ellison2008,Ellison2013,Thorp2019,Bickley2022} merger regimes in galaxy surveys.
Merger driven gas inflows have also been predicted to potentially fuel black hole growth \citep[][]{Sanders1988,DiMatteo2005,Springel2005,Hopkins2008,Capelo2015}. However, whilst there is broad agreement in the observational literature supporting merger-enhanced star formation, there continues to be controversy over the role of mergers in triggering active galactic nuclei (AGN). Although numerous observational studies do not find evidence for a merger-AGN connection \citep[][]{Cisternas2011,Schawinski2011,Kocevski2012,Bohm2013,Shah2020,Lambrides2021}, many others have demonstrated a statistically significant excess of AGN in interacting pairs and post-mergers \citep[][]{Alonso2007,Ellison2011,Bessiere2012,Ellison2013,Hong2015,Kocevski2015,Hewlett2017,Marian2020,Bickley2022Submitted}.
Recently, \cite{Ellison2019} explored the possibility that contradictory results may result from different experimental approaches; namely that some studies assess the excess of AGN in mergers, whereas others quantify the excess of mergers amongst AGN. Using a sample of low redshift AGN and high quality imaging from the Canada France Imaging Survey, \cite{Ellison2019} demonstrated that there is \textit{both} an excess of AGN in mergers \textit{and} an excess of mergers amongst AGN. At least at low redshift (and for the optical and mid-IR selection used by \citealt{Ellison2019}), experimental approach does not seem to explain the conflicting results.
There are numerous other factors that may affect our assessment of the merger-AGN connection. AGN are a multi-wavelength phenomenon that manifest across the electromagnetic spectrum and different selection methods identify different objects (see \citealt{HickoxAlexander2018} for a recent review). The extent of a connection between mergers and AGN may depend on the wavelengths at which the AGN is selected. For example, several studies have shown that the AGN excess in mergers is greater for mid-IR selected AGN, compared to those identified through their optical emission lines \citep[][]{Satyapal2014,Weston2017,Goulding2018,Ellison2019,Gao2020}, with no X-ray AGN enhancement found in low redshift post-mergers \citep[][]{Secrest2020}. However, even for selection in a given frequency regime disagreement can exist. For example, some studies of radio-selected AGN find no merger connection \citep[][]{Dicken2012,Ellison2015} whilst others do \citep[][]{RamosAlmeida2011,RamosAlmeida2012,Gao2020,Pierce2022}, a tension that could be linked to further distinctions within the radio population into high and low excitation sources \citep[][]{Chiaberge2015,Bernhard2022}. Disagreement also exists as to whether the dominance of merger triggering is linked to AGN luminosity \citep[][]{Schawinski2012,Treister2012,Villforth2014,Glikman2015,Mechtley2016,Villforth2017,Marian2019,Pierce2022}.
The above summary of the observational literature demonstrates the considerable diversity amongst recent results. However, there are a few broad statements regarding the merger-AGN connection that are relatively uncontroversial. First, that the majority of AGN are \textit{not} merger induced, a statement that is also supported by simulations \citep[][]{Steinborn2018,McAlpine2020}. Second, that \textit{some} AGN are probably triggered by interactions, although the fraction likely depends on subtleties such as redshift \citep[e.g.][]{Rosario2015} and selection method \citep[e.g.][]{Gao2020}. Third, it is the most obscured \citep[][]{Satyapal2014,Fan2016,Weston2017,Ellison2019,Gao2020} and most luminous \citep[][]{Schawinski2012,Treister2012,Glikman2015,Donley2018,Goulding2018,UrbanoMayorgas2019,Pierce2022} AGN that seem to be most likely to be linked to mergers.
Beyond the simple understanding of which mechanisms (merging versus various secular processes) lead to black hole growth, the potential for AGN triggering during the interaction is arguably most relevant for assessing the end point of the merger sequence. Since AGN have been widely implicated as a way to quench star formation \citep[][]{DiMatteo2005,Springel2005}, if most mergers experience an accretion+feedback event during the interaction, then we might expect most post-coalescence galaxies to rapidly shut-down their star formation \citep[][]{Hopkins2008}.
An additional complication to the proposed connections between mergers, star formation and AGN triggering, followed by feedback driven quenching, is the different timescales for these processes. High resolution simulations have shown that merger induced central starbursts and high SMBH accretion rates are for the most part not temporally correlated \citep[][]{Volonteri20151}. Furthermore, most observational studies do not find a connection between central starburst and AGN activity, except in the highest luminosity systems, an effect that is largely attributed to short timescales for AGN variability \citep[][]{RowanRobinson1995,Schweitzer2006,Lutz2010,Shao2010,Santini2012,Rosario20131,Rosario20132}.
In order to assess the impact of mergers on star formation, black hole accretion and quenching in a statistical and holistic way, we have been undertaking a series of studies using a large, state-of-the-art hydrodynamical simulation suite from IllustrisTNG (hereafter TNG; \citealt[][]{Nelson2017,Naiman2018,Mariancci2018,Pillepich20171,Springel2017}). Although large box simulations are lacking in resolution compared with idealized binary merger runs or cosmological zoom in simulations (e.g. \citealt{Hopkins2013,Moreno2015,Moreno2019}), they offer several advantages. First, many thousands of interactions can be drawn from the large sample of galaxies in a cosmological simulation, in contrast to the suites of binary mergers which typically contain only tens of interactions. Second, no a priori decisions are required to define either galaxy properties (morphologies, gas fractions etc.) or interaction properties (e.g. mass ratios, orbits etc.). Finally, the full cosmological setting provides a more realistic interaction with both the circumgalactic and intergalactic environments.
In the first paper in our series, \cite{Patton2020} demonstrated that enhancements in the star formation rate (SFR) of interacting galaxies in TNG are experienced in the pre-coalescence phase out to separations of a few hundred kpc. Although the exact number varies depending on the simulation used, triggered star formation out to large separations was found to be a common feature of all of simulations in the TNG suite. \cite{Hani2020} continued the assessment of SFR enhancement into the post-coalescence regime and showed that the elevated levels persist for typically 500 Myrs post-merger. \cite{Quai2021} demonstrated that although most post-mergers do not lead to rapid shut down of star formation, there is nonetheless an excess of quenched galaxies amongst simulated post-mergers compared with controls.
In the work presented here, we investigate the role of mergers in triggering AGN in TNG, specifically whether there is an enhancement of supermassive black hole accretion rates in the post-merger phase. In Section \ref{sec:methods}, we summarize the salient details of the TNG simulation, our merger sample selection, and our algorithm for matching post-merger galaxies to control comparisons. In Section \ref{sec:Results} we quantify the enhancement of black hole accretion rates in post-mergers and investigate whether these enhancements depend on galaxy and merger properties (Section \ref{subsec:Enhancements}). Furthermore, we quantify the timescale over which accretion rate enhancements persist (Section \ref{subsec:timescale}) and compare this with the triggered star formation in order to assess synchronicity between these processes (Section \ref{subsec:correlations}). Finally, we address whether mergers are likely to produce AGN activity and quantify the contribution of recent mergers to the total AGN population, i.e. what fraction of mergers produce high accretion rates (Section \ref{subsec:fractionAGN}) and what fraction of high accretion rate systems have undergone a recent merger (Section \ref{subsec:fractionMergers}).
\section{Methods}
\label{sec:methods}
\subsection{IllustrisTNG}
\label{subsec:Illustris}
We use the IllustrisTNG galaxy formation simulation, a state-of-the-art cosmological magneto-hydrodynamic simulation including galactic scale stellar feedback, stellar population evolution and chemical enrichment, primordial and metal-line gas cooling and heating, and multi-mode blackhole feedback \citep[][]{Nelson2017,Naiman2018,Mariancci2018,Pillepich20171,Springel2017}. In the work presented here, we use the TNG100-1 simulation run, the intermediate volume and resolution run of the fiducial TNG galaxy formation model, which has a $(110.7 \, \mathrm{Mpc})^3$ volume, a baryonic resolution of $1.4 \times 10^6 \, \mathrm{M_{\odot}}$, and a dark matter resolution of $7.5 \times 10^6 \, \mathrm{M_{\odot}}$.
We briefly describe the physical models for supermassive black hole seeding and feedback in TNG, with the full details available in \cite{Weinberger2017} and \cite{Pillepich2018}. Supermassive black holes, hereafter SMBHs, are seeded at $M_{BH} = 8 \times 10^5 \, \mathrm{h^{-1} M_{\odot}}$ at the centre of halos that meet a threshold mass of $5 \times 10^{10} \, \mathrm{h^{-1} M_{\odot}}$. The black holes can then grow by either accreting gas from the region surrounding the black hole, or by merging with other black holes.
SMBH accretion is calculated using a Bondi-Hoyle-Lyttleton subgrid model,
\begin{equation}
\dot M_{Bondi} = \frac{4 \pi G^2 M_{BH}^2 \rho}{c_s^3}
\label{eq:Bondi}
\end{equation}
where $G$ is the gravitational constant, $M_{BH}$ is the black hole mass, and $\rho$ and $c_s$ are the density and sound speed sampled in a kernel-weighted sphere centred on the SMBH, labelled the accretion region. The SMBH accretion rate is capped by the Eddington rate,
\begin{equation}
\dot M_{Edd} = \frac{4 \pi G M_{BH} m_p}{\epsilon_r \sigma_T c}
\label{eq:Edd}
\end{equation}
where $m_p$ is the proton mass, $\epsilon_r$ is the radiative accretion efficiency, $\sigma_T$ is the Thompson cross-section, and $c$ is the vacuum speed of light.
SMBHs are merged when the black holes are within the accretion regions of one-another. Furthermore, in order to prevent the wandering of SMBHs away from the halo centre, SMBHs are fixed to the local gravitational potential minima, which has the added effect of promptly merging SMBHs when their host subhalos coalesce.
TNG uses a dual feedback mode model, where SMBHs use a different feedback prescription depending on the SMBH accretion rate and SMBH mass. At high accretion rates, feedback is implemented according to a radiative mode model, and at low accretion rates the feedback is implemented according to a kinetic mode model. The classification of high vs. low accretion is made by calculating the ratio of the Bondi accretion rate to the Eddington accretion rate, where a high accretion rate is defined as $\dot M_{Bondi} \geq \chi \dot M_{Edd}$ and $\chi$ is defined as,
\begin{equation}
\chi = min\bigg[\chi_0 \bigg(\frac{M_{BH}}{10^8 M_{\odot}}\bigg)^{\beta},0.1\bigg].
\label{eq:chi}
\end{equation}
The maximum $\chi=0.1$ follows observational constraints set by X-ray binaries \cite{Dunn2010}, whereas the parameters $\chi_0$ and $\beta$ are simulation parameters tuned to 0.002 and 2 respectively.
For SMBHs with high accretion rates, the simulation uses a radiative feedback prescription, where a fraction of the accreted mass energy is injected as thermal energy into the region surrounding the black hole. At low accretion rates, once again a fraction of the accreted mass energy is injected into the surrouding BH region, however the energy injected is in the form of kinetic energy in randomized directions away from the black hole. For our work, focused on SMBH accretion rates, we do not comment in detail on the feedback prescriptions, and refer the reader to \cite{Weinberger2017}.
\subsection{Identifying galaxies and galaxy mergers in IllustrisTNG}
\label{subsec: ID mergers}
Before we can identify the population of galaxy mergers in the TNG simulation, we first apply selection constraints to our entire galaxy sample.
First, we exclude galaxies at redshift$>$1 from our sample because the increased frequency of mergers and interactions exacerbate issues such as numerical stripping and subhalo switching \citep[][]{Rodriguez-Gomez2015}, which makes robust merger identification more difficult. We then apply a stellar mass requirement of $10^9 M_{\odot}$, equivalent to $\sim 1000$ particles, to ensure our selected galaxies are properly resolved. We combine our minimum stellar mass requirement with our minimum merger mass ratio of 1:10, i.e. requiring post-mergers have progenitors of at least $10^9 M_{\odot}$, by requiring a minimum stellar mass of $10^{10} M_{\odot}$ in the post-merger sample.
We also apply an additional environmental constraint to remove galaxies susceptible to numerical stripping from ongoing interactions with neighbouring subhalos. Following the procedure of \cite{Patton2020}, we calculate $r_{sep}$,
\begin{equation}
r_{sep} = \frac{r}{R^{host}_{1/2} + R^{comp}_{1/2}}
\label{eq:rsep}
\end{equation}
where r is the 3D separation of the galaxies, and $R^{host}_{1/2}$ and $R^{comp}_{1/2}$ are the stellar half mass radii of the host and companion. \cite{Patton2020} demonstrated that TNG galaxies with $r_{sep} \lesssim 2$ had stripped stellar mass and we therefore excluded them from our galaxy sample.
Once we have identified appropriate candidate galaxies from TNG, we identify galaxy mergers using the merger-trees created by the \textsc{SUBLINK} algorithm \citep[][]{Rodriguez-Gomez2015}, which associate each galaxy in TNG with progenitor and/or and descendant galaxies. We identify galaxy mergers by selecting nodes within the merger trees. Nodes occur when particles which are assigned into distinct subhalos at a given snapshot are subsequently assigned into the same subhalo in the following snapshots.
In the work presented here, we calculate the mass ratio of the merger using a different methodology from previous works in this series. We apply additional steps, once again, as a precaution against numerical stripping and subhalo switching. For each merger, the mass ratio of the merger is calculated using the stellar mass within twice the stellar half-mass radius, in the snapshots leading up to coalescence of the subhalos, up to a maximum of 10 snapshots. From the mass ratio estimates, the maximum and minimum are excluded, which removes abnormally large or small values that are due to subhalo switching or significant numerical stripping. Finally, we take the average and standard deviation of the remaining mass ratio measurements to calculate the merger stellar mass ratio, $\mu$, and assign a merger mass ratio error, $\sigma_{\mu}$. For the work presented here, we define mergers as coalescence of subhalos with a merger mass ratio within $0.1 < \mu \leq 1$. We do not limit the overall post-merger sample using the mass ratio error except when specified.
\subsection{Post-merger and non-merger galaxy samples}
\label{subsec: PM and NM}
Hydrodynamical simulations have suggested that SMBH accretion rates peak post-coalescence \citep[][]{Hopkins2008}, a framework that is corroborated by some observational studies \citep[][]{Ellison2013,Satyapal2014,Bickley2022Submitted}. In the work presented here we focus on post-merger galaxies, and construct our post-merger sample using only galaxy mergers that have coalesced in the time between the previous and current snapshot, or within $\sim 160$ Myrs. Note that we discuss the long term effects on the post-merger sample in Section \ref{subsec:timescale}. Selecting galaxies immediately after coalescence, and applying the galaxy selection criteria outlined above, we identify a sample of 1973 post-merger galaxies from TNG100-1. The stellar mass, SMBH mass, redshift and mass ratio distributions of this complete post-merger sample are shown in the dotted teal line in Figure \ref{fig:PMdist}.
In addition to the post-mergers, we define a sample of non-mergers whose properties are representative of the post-mergers. We construct a non-merger sample in order to generate a comparative population that has the same underlying distribution as the post-merger sample, and in order to contextualize how the underlying stochasticity of SMBH accretion rates affect our calculation of accretion rate enhancements. The representative non-merger sample is constructed as follows. For each of the 1973 post-mergers, we identify one non-merging galaxy that is selected to be the single best match in redshift, stellar mass, gas mass, environment, and feedback mode (see Section \ref{subsec:matching} for details on the necessity of feedback mode matching) that has not undergone a merger of mass ratio $\mu > 0.1$ within the last two Gyrs. The environment is quantified using two parameters taken from \cite{Patton2020}: $r_1$ and $N_2$. $r_1$ is the 3D distance to the nearest neighbour, within two Mpc, which has a mass above 10\% of the target galaxy mass. $N_2$ is the number of neighbours within two Mpc of the galaxy centre, with a minimum mass of $10^{9} M_{\odot}$. The single best match is selected as an exact match in redshift and feedback mode, and the best simultaneous match of stellar mass, gas mass, $r_{1}$ and $N_{2}$ following the statistical weighting scheme of \cite{Patton2016}. We therefore begin with a sample
of 1973 post-mergers and 1973 non-mergers.
\begin{figure}
\includegraphics[width=\columnwidth]{PMsampleDist_legend.pdf}
\caption{Distribution of post-merger and non-merger sample (from top to bottom) in stellar mass, SMBH mass, redshift, and mass ratio. The dotted teal line shows the distribution of all the post-mergers, following our criteria, in TNG100-1. The solid teal line shows the distribution of the 1563 successfully control matched post-mergers, and the dashed yellow line shows the 1522 successfully control matched non-mergers.}
\label{fig:PMdist}
\end{figure}
\subsection{Control Matching Post-merger and Non-merger Galaxies}
\label{subsec:matching}
In order to quantify the difference between a given physical property (e.g. SMBH accretion rate) of an individual (post-merger or non-merger) galaxy and the expected value of a `normal' galaxy, we use a control matching algorithm to identify a sample of comparative controls for each individual galaxy in the post-merger and non-merger sample. The pool of possible control galaxies consists of all galaxies, meeting the criteria outlined in Section \ref{subsec: ID mergers}, that have not undergone a merger of mass ratio $\mu > 0.1$ within the last two Gyrs. We note that when control matching the non-mergers, we exclude galaxies that have already been assigned into the non-merger sample from the control pool, such that no galaxy can be matched to itself.
We begin by down-selecting the pool of potential control galaxies by imposing two cuts (which are effectively broad matching constraints). First, we require that the controls for a given galaxy are drawn from the same simulation snapshot, which corresponds to matching in redshift. Second, we require that the controls for a given galaxy are in the same instantaneous feedback mode as the post-merger (or non-merger), using the feedback mode classification outlined in Section \ref{subsec:Illustris}. The requirement of feedback mode matching diverges from the methods applied by \cite{Patton2020}, \cite{Hani2020}, and \cite{Quai2021}, but is necessary for the work presented here to avoid spurious features in the results. The issue is illustrated in Figure \ref{fig:Mdot_vs_Mstar} which shows the distribution of accretion rates as a function of stellar mass for all galaxies in TNG100-1, colour-coded to distinguish those in radiative feedback mode (dark blue) from those in kinetic feedback mode (red). Figure \ref{fig:Mdot_vs_Mstar} demonstrates that (at a fixed stellar mass) the accretion rate distributions are bimodal. Without feedback mode matching, post-merger galaxies may (artificially) exhibit extremely enhanced or suppressed accretion rates if they are matched to controls in the other feedback mode.
\begin{figure}
\includegraphics[width=\columnwidth]{Mdot_vs_Mstar_legend.pdf}
\caption{SMBH accretion rates as a function of stellar mass for all the TNG100-1 galaxies meeting the selection criteria of Section \ref{subsec: ID mergers}. The dark blue contours represent the sample in radiative mode feedback and the orange contours represent the sample in kinetic mode feedback.}
\label{fig:Mdot_vs_Mstar}
\end{figure}
In addition to snapshot (redshift) and feedback mode, the control galaxies for each post-merger and non-merger are matched on several further properties. First, the control sample is limited to galaxies whose stellar mass is matched to within $\pm 0.05$ dex and whose gas mass is matched to within $\pm 0.1$ dex. Next, the matched control sample is limited to galaxies within $\pm 10\%$ of $r_1$ and $N_2$. We allow all of the error tolerances to grow, 0.01 dex in stellar mass, 0.1 dex in gas mass, and $10\%$ in $r_1$ and $N_2$, up to four times until at least five matched control galaxies are found. If fewer than five control galaxies are found, then the post-merger is excluded from the sample. On average, galaxies have to grow their error tolerances twice to meet the required number of controls, and 410 post-merger galaxies (and 451 non-merger galaxies) did not meet the control requirements within permitted error tolerances after the maximum number of grows and were hence rejected from our sample. Therefore, our final matched sample consists of 1563 matched post-merger galaxies and 1522 non-merger galaxies, with an average of 10 matched controls each. The properties of the final post-merger and non-merger samples are shown in the solid teal and dashed yellow lines in Figure \ref{fig:PMdist}. We note that the top panel of Figure \ref{fig:PMdist} reveals that the control matching algorithm preferentially fails to match higher stellar mass post-mergers. It was determined that the strict error tolerance on stellar mass matching combined with gas mass matching resulted in higher mass post-mergers being less likely to find a sufficient number of controls. However, we note that both matching criteria are necessary in order to select an unbiased set of controls. Without the strict error tolerance, high mass post-mergers are preferentially matched to low mass controls. We therefore prioritize the selection of unbiased controls over the completeness of the matched post-merger sample.
Figure \ref{fig:ControlsMatch} shows the distribution of stellar mass, gas mass, $r_1$, and $N_2$ for the post-merger galaxies, shown in the teal line, and the matched control galaxies, shown in dashed pink line. In the figure, the distributions trace one another closely, demonstrating the success of the control matching methodology in identifying galaxies with similar properties to the post-merger sample. Although not shown in Figure \ref{fig:ControlsMatch}, we note that the non-merger sample and its controls exhibit similarly closely matched properties.
\begin{figure}
\includegraphics[width=\columnwidth]{MatchingHist_legend.pdf}
\caption{Distribution of parameters matched in control matching process: (from top to bottom) stellar mass and gas mass measured within twice the stellar half mass radius, distance to nearest neighbour within two Mpc, and number of neighbours within two Mpc. Post-mergers are represented with the teal line, and the controls are represented with the dashed pink line.}
\label{fig:ControlsMatch}
\end{figure}
Once the control matching procedure is complete, we then calculate a relative SMBH accretion rate enhancement, $\Delta \dot M_{BH}$, for each post-merger and non-merger galaxy as
\begin{equation}
\Delta \dot M_{BH} = log_{10}(\dot M_{BH}) - median[log_{10}(\dot M_{BH Controls})].
\label{eq:enhancement}
\end{equation}
\section{Results}
\label{sec:Results}
\subsection{Enhancement of accretion rates in post-merger galaxies}
\label{subsec:Enhancements}
Figure \ref{fig:DMdot_hist_all} shows the distribution of $\Delta \dot M_{BH}$ for the post-merger sample, shown in the teal line, and non-merger sample, shown in the yellow dashed line. The non-merger population peaks at 0 dex, consistent with no enhancement in the SMBH accretion rate of non-mergers relative to their controls. A lack of statistical offset may be expected for the non-merger sample, as it is a subset of the control population. However, since the non-merger sample is not randomly selected from the control pool, and is selected to match the population characteristics of the post-merger sample, we have therefore demonstrated that the population characteristics of the non-merger sample do not give rise to enhanced SMBH accretion rates relative to their controls. Nonetheless, the non-merger distribution provides a useful reference. For example, both populations show a similar distribution of $\Delta \dot M_{BH}$, ranging from 100 times enhanced to 100 times suppressed accretion rates relative to controls, with 30 non-mergers (2\%) achieving accretion rates at least 10 times higher than controls. The comparable spread demonstrates the large variability of accretion rate in both samples. In contrast to the non-mergers, the post-merger sample peaks at a positive accretion rate enhancement, with a median enhancement of 0.23 dex, corresponding to an enhancement of $\sim 70 \%$. In the post-merger sample, 102 post-mergers (6.5\%) have accretion rates at least 10 times higher than controls. Figure \ref{fig:DMdot_hist_all} therefore demonstrates our first main result, that mergers in TNG, on average, have an elevated accretion rate in the post-merger phase.
\begin{figure}
\includegraphics[width=\columnwidth]{DMdot_hist_all_legend_axis.pdf}
\caption{Histogram of accretion rate enhancements for the post-merger population and non-merger population. Post-mergers are represented with the teal line, and non-mergers are represented with the dashed yellow line. The median enhancement of the post-merger sample is 0.23 dex.}
\label{fig:DMdot_hist_all}
\end{figure}
Figure \ref{fig:DMdot_vs_z} shows the SMBH accretion rate enhancement of the post-merger and non-merger samples as a function of redshift. The full distribution of the post-merger sample is shown in the background density plot, and the foreground points are the median accretion enhancement for the post-merger sample (teal circles) and non-merger sample (yellow squares) within equally spaced bins of redshift. The errorbars represent the standard error on the median within each bin. The median accretion rate enhancement of the post-merger sample is consistently above 0 across the redshift range. We present both the median enhancement points and the background distribution to emphasize that while there is a positive median $\Delta \dot M_{BH}$ in post-mergers, not all post-mergers show accretion rate enhancements (as expected from Figure \ref{fig:DMdot_hist_all}). Overall, neither sample (post-mergers nor non-mergers) show any significant dependence on redshift, indicating that the merger process elevates accretion rates out to at least $z=1$ (the limit of our sample selection).
\begin{figure}
\includegraphics[width=\columnwidth]{DMdot_vs_z.pdf}
\caption{Accretion rate enhancements of post-merger galaxies for different redshifts. The background density plot shows the distribution of $\Delta \dot M_{BH}$ for the post-merger sample, and the foreground points represent the median value at that redshift. The teal circles are the median post-merger $\Delta \dot M_{BH}$ and the yellow squares are the median non-merger $\Delta \dot M_{BH}$. The error on the x-axis represents the bin width for each data point, and the error on the y-axis is the standard error on the median for that redshift bin.}
\label{fig:DMdot_vs_z}
\end{figure}
We find a similarly consistent result when investigating the dependence of SMBH accretion rate enhancements on stellar mass, shown in Figure \ref{fig:DMdot_vs_Mstar}. As expected from its construction, we find that the non-merger sample has accretion rate enhancements consistent with zero at all stellar masses. In contrast the teal points demonstrate that the SMBH accretion rates in post-mergers are consistently enhanced, on average, at all stellar masses within our sample.
\begin{figure}
\includegraphics[width=\columnwidth]{DMdot_vs_Mstar.pdf}
\caption{Accretion rate enhancements of post-merger galaxies for different stellar masses. The background density plot shows the distribution of $\Delta \dot M_{BH}$ for the post-merger sample, and the foreground points represent the median value at that stellar mass. The teal circles are the median post-merger $\Delta \dot M_{BH}$ and the yellow squares are the median non-merger $\Delta \dot M_{BH}$. The error on the x-axis represents the bin width for each data point, and the error on the y-axis is the standard error on the median for that stellar mass bin.}
\label{fig:DMdot_vs_Mstar}
\end{figure}
Figure \ref{fig:DMdot_vs_Mgas} shows the accretion rate enhancements $\Delta \dot M_{BH}$ as a function of the gas mass (top panel) and gas fraction (bottom panel) of the post-merger or non-merger galaxy, where the gas mass is the sum of all gas particles within twice the stellar half-mass radius and the gas fraction is the ratio of the gas mass to the baryon mass (gas + stars). In the top panel, we see that for a gas mass less than $\sim 10^9 \mathrm{M_{\odot }}$, the median enhancements of the post-merger sample are not significantly distinct from the non-merger sample, with the exception of one point at $\sim 10^{8.3-8.4} \mathrm{M_{\odot}}$. However in the regime of $> 10^{9} \mathrm{M_{\odot}}$, post-mergers consistently have on average enhanced accretion rates. Therefore, Figure \ref{fig:DMdot_vs_Mgas} suggests that post-mergers of a lower gas mass are less likely to have enhanced accretion rates. In addition, at gas masses above $10^9 \mathrm{M_{\odot}}$, post-merger galaxies consistently show accretion rate enhancements, where non-merger galaxies of this mass range do not, suggesting that the presence of a large amount of gas does not guarantee higher than average accretion rates.
We find a qualitatively consistent result looking at the gas fraction (bottom panel of Figure \ref{fig:DMdot_vs_Mgas}). We see that for post-mergers with a lower gas fraction, there is a trend towards low to no relative accretion enhancement. Once again, there is some evidence for an exception to this trend at the lowest gas fraction, however we caution that the poor statistics at lower gas mass and gas fraction make the median data point more susceptible to the variability on a galaxy by galaxy basis. Therefore, we find that the presence of a significant amount of gas is an essential, but insufficient criterion to produce accretion rate enhancements, and that the merger itself is important for driving gas into the region surrounding the SMBH whence it can accrete.
\begin{figure}
\includegraphics[width=\columnwidth]{DMdot_vs_gasMass_gasFrac.pdf}
\caption{Accretion rate enhancements of post-merger galaxies for different $M_{gas }$ (shown in top panel) and gas fractions (shown in bottom panel). The gas mass is the sum of all gas particles within twice the stellar half-mass radius of the galaxy. The gas fraction is the ratio of the gas mass to the baryon mass (gas mass + stellar mass). The background density plot shows the distribution of $\Delta \dot M_{BH}$ for the post-merger sample, and the foreground points represent the median value at that $M_{gas}$ or gas fraction. The teal circles are the median post-merger $\Delta \dot M_{BH}$ and the yellow squares are the median non-merger $\Delta \dot M_{BH}$. The error on the x-axis represents the bin width for each data point, and the error on the y-axis is the standard error on the median for that $M_{gas}$ or gas fraction bin.}
\label{fig:DMdot_vs_Mgas}
\end{figure}
Finally, Figure \ref{fig:DMdot_vs_MassRatio} shows $\Delta \dot M_{BH}$ for post-merger galaxies of different merger mass ratios. We note that in this plot, we only include galaxies with a merger mass ratio error of $\sigma_{\mu} \leq 0.1$, which excludes 99 galaxies from the sample of 1563 post-mergers. By visual inspection, we find no significant relationship between the strength of the accretion rate enhancement and the merger mass ratio. In fact, we see that the majority of enhancements are consistent, within error, with the overall sample accretion rate enhancement of 0.23 dex. We further confirm the lack of correlation with a statistical Pearson correlation test, which yields a correlation coefficient of $\sim$0.1.
The lack of correlation between $\Delta \dot M_{BH}$ and mass ratio is somewhat contradictory to previous simulation results. For example, \cite{Capelo2015} find mass ratio to be the most important factor influencing SMBH accretion rate in a suite of high resolution binary merger simulations. However, there are notable differences between the experiments of \cite{Capelo2015} and the work presented here. Specifically, \cite{Capelo2015} look at the effect of varying the merger mass ratio while keeping the orbital geometry and gas fraction constant, whereas our result looks at a population averaged enhancement and is therefore subject to the variable conditions of every merger. In addition, \cite{Capelo2015} comment on the role of resolution in their result, and that some of the torques generated in the galaxy interaction require high resolution simulations in order to be resolved. Finally, our result should not be interpreted as mass ratio having zero role in regulating gas inflows. Instead, Figure \ref{fig:DMdot_vs_MassRatio} demonstrates that, once the full demographic of merger properties is sampled, mass ratio is not a dominant factor in predicting the strength of a SMBH accretion rate enhancement.
\begin{figure}
\includegraphics[width=\columnwidth]{DMdot_vs_MassRatio.pdf}
\caption{Accretion rate enhancements of post-merger galaxies for different merger mass ratio. The background density plot shows the distribution of $\Delta \dot M_{BH}$ for the post-merger sample, and the foreground points represent the median value at that mass ratio. The error on the x-axis represents the bin width for each data point, and the error on the y-axis is the standard error on the median for that mass ratio bin.}
\label{fig:DMdot_vs_MassRatio}
\end{figure}
\subsection{Timescale of accretion rate enhancements}
\label{subsec:timescale}
We have thus far demonstrated that post-merger galaxies show, on average, enhanced accretion rates when compared to matched controls. This result is in agreement with previous theoretical studies that demonstrate merger driven gas inflows \citep[][]{DiMatteo2005,Springel2005,Hopkins2008}, which would increase the local gas density surrounding the black hole. However, the timescales of enhanced SMBH accretion are crucial to understanding the influence of the merger beyond the immediate post-merger phase, and the role of merger induced accretion rate enhancements on a galaxy's evolution. In addition, studying the average temporal behaviour of SMBH accretion rate enhancements can help alleviate the challenges associated with studying the highly stochastic individual SMBH accretion rates.
In the following section, we re-perform an experiment from a previous paper in this series. Following the procedure of \cite{Hani2020} (who investigated enhancements in SFR), we calculate a $\Delta \dot M_{BH}$ for post-merger galaxies in the snapshots following coalescence. Specifically, for each post-merger in our sample, we follow the descendants for as long as the merger tree allows or until the next merger event (with a mass ratio greater than 1:10). We control match the post-merger descendants to identify the enhancement to SMBH accretion rate as a function of the time since the most recent merger. We re-select the control galaxies at each subsequent snapshot following the same matching criteria as Section \ref{subsec:matching}. Therefore, a galaxy at snapshot N will not be matched to the same controls in the subsequent snapshot N+1 if the controls fall out of the acceptable matching criteria and tolerances specified in Section \ref{subsec:matching}.
Figure \ref{fig:DMdot_DSFR_TPM} shows the SMBH accretion rate enhancement as a function of time post-merger. The median SMBH accretion rate enhancement is shown in the teal line, where the shaded region represents the 25th and 75th percentiles. We find that, on average, SMBH accretion rate enhancements persist for up to two Gyr post-merger. A long-lived accretion rate enhancement may seem to be in contradiction with the short lived (tens-hundreds of Myrs) high accretion rate events observed in previous simulation studies \citep[][]{DiMatteo2005,Springel2005,Hopkins2008}. However we emphasise our relative enhancement variable $\Delta \dot M_{BH}$ is distinct from an enhancement in absolute accretion rate in a given galaxy. Our result demonstrates that post-mergers sustain a population-averaged enhanced accretion rate relative to matched controls, which does not necessarily correspond to a long lived high accretion rate event in an individual galaxy. A long-lived population average suggests that the dynamical disturbance that leads to higher than normal feeding of the SMBH persists for up to $\sim 2$ Gyrs in TNG100-1.
We are also interested in investigating the temporal correlation, or lack there-of, between star formation rate enhancements and SMBH accretion rate enhancements. Studying the population averaged enhancements can smooth out the large temporal variability of SMBH accretion rates, and allow us to look at the timescales over which connections between star formation and SMBH accretion rate are present in the overall post-merger population. Therefore, we also compute the star formation rate enhancements of our post-merger sample,
\begin{equation}
\Delta SFR = log_{10}(SFR) - median[log_{10}(SFR_{Controls})],
\label{eq:enhancement2}
\end{equation}
where the star formation rate is the sum of star formation rates for all cells within twice the stellar half-mass radius. When calculating $\Delta SFR$, we introduce an additional matching criterion from \cite{Hani2020}. We require that galaxies be matched within a classification of star forming or passive in order to avoid (spuriously) large/small values of $\Delta SFR$. Passive galaxies are defined as galaxies which lie more than 2$\sigma$ below the star forming main sequence, where we calculate the star forming main sequence by applying a linear fit to TNG100-1 galaxies with stellar mass between $10^{9-10.2}M_{\odot}$ and between redshift 0 to 1, and extrapolate the linear fit to higher stellar masses following the procedure of \citep[][]{Donnari2019}. In this way, post-merger star-forming galaxies are matched to star-forming controls, in addition to the fiducial matching criteria outlined in Section \ref{subsec:matching}. Likewise, passive post-mergers are matched to passive controls. We therefore identify a different set of controls when calculating $\Delta \dot M_{BH}$ and $\Delta SFR$ for the experiments presented in Section \ref{subsec:timescale}.
Figure \ref{fig:DMdot_DSFR_TPM} shows the star formation rate enhancement in the purple line, once again the shaded region represents the 25th and 75th percentiles. We qualitatively recover the result of \cite{Hani2020}, who find that, on average, post-merger galaxies demonstrated star formation rate enhancements\footnote{We note that the magnitude of the star formation rate enhancement in post-merger galaxies is lower in the work presented compared with the main result of \cite{Hani2020}. There exist a number of subtle matching scheme differences, specifically the inclusion of gas mass matching (which is demonstrated in \cite{Hani2020} to reduce the magnitude of the SFR enhancement).}, and that the enhancements persisted for up to $\sim 500$ Myrs after the merger, consistent with previous simulation studies \citep[][]{DiMatteo2008} and observational estimates \citep[][]{Wild2010}. We therefore demonstrate that, on average, SMBH accretion rate enhancements are significantly longer lived than star formation rate enhancements within our post-merger sample. We compare the result of Figure \ref{fig:DMdot_DSFR_TPM} with \cite{Volonteri20151}, who investigate the temporal correlation between SFR and SMBH accretion rate in 10 high resolution binary merger simulations. \cite{Volonteri20151} find that, on an individual merger basis, properties are only temporally correlated within the 200-300 Myrs of the final coalesence event of the merger, a timescale that is consistent with our population averaged result. Our result that enhancements of SMBH accretion rates can be sustained up to 2 Gyrs is also consistent with the result of \cite{Volonteri20152}, who find that the SMBH accretion rates of merger remnants can remain sufficiently high such that the luminosity of the AGN is dominant over the stellar luminosity up to 1.5 Gyrs after the merger.
As a possible explanation for the difference in the timescale of SFR and accretion rate enhancements, we comment that in TNG, both the star formation and SMBH accretion depend on local gas density, where stars form in gas following the empirical Kennicutt-Schmidt relation \citep[][]{Schmidt1959,Kennicutt1998} and the gas dependence of SMBH accretion is shown in Eq. \ref{eq:Bondi}. However, star formation in TNG additionally requires a threshold density of $n_{H} \gtrsim 0.1 cm^{-3}$ \citep[][]{Pillepich2018}. We suggest a possible scenario where a merger event increases the central gas density significantly within $\sim 500$ Myrs of coalesence, resulting in enhanced star formation and SMBH accretion rates. However, the increase of material to the central gas reservoir is insufficient to sustain star formation past 500 Myrs yet sufficient to sustain a long lived low accretion rate enhancements.
We considered whether the above average accretion rates in the post-merger sample, which persist $\sim$ two Gyrs after the merger event, may be a feature of the matching methodology, as we define controls as galaxies that have not undergone a merger of mass ratio $>1:10$ within the last two Gyrs. We test whether the timescale of averaged accretion rate enhancements is sensitive to the minimum elapsed time post-merger that we allow for controls. We regenerate our control sample, now requiring that control galaxies to have had at least three Gyrs since their most recent 1:10 merger, and find that the timescale of two Gyrs is robust.
\begin{figure}
\includegraphics[width=\columnwidth]{DMdot_DSFR_vs_TPM_line_2Gyr_legend.pdf}
\caption{Running median of the accretion rate enhancement (teal) and star formation rate enhancement (purple) over the time post-merger. The error of the shaded region represents the 25th and 75th percentiles associated with the median at that time post-merger.}
\label{fig:DMdot_DSFR_TPM}
\end{figure}
Figure \ref{fig:DMdot_DSFR_TPM} demonstrates the averaged behaviour of $\Delta \dot M_{BH}$ and $\Delta SFR$ in the total post-merger sample. However, not all galaxies show positive SMBH accretion rate or SFR enhancements (e.g. Figure \ref{fig:DMdot_hist_all}). In particular, the results of Section \ref{subsec:Enhancements} suggest that gas poorer galaxies do not exhibit, on average, accretion rate enhancements. We therefore separate the post-merger population based on the instantaneous feedback mode, which separates the post-mergers with high absolute accretion rates (radiative mode) from low absolute accretion rates (kinetic mode). The distinction in feedback mode also broadly separates lower mass, gas rich galaxies (radiative mode) from higher mass, gas poor galaxies (kinetic mode).
Figure \ref{fig:DMdot_DSFR_TPM_split} shows SMBH accretion rate enhancements and SFR enhancements, separated into the radiative mode, shown in blue, and kinetic mode, shown in red. In the top panel, we see that both populations show, on average, positive accretion rate enhancements within two Gyrs of coalescence. The radiative mode feedback population has a higher peak accretion rate enhancement of $\sim$ 0.3 dex compared with a relative enhancement of $\sim$0.2 dex in the kinetic mode feedback population. The lower median accretion rate enhancements in kinetic mode feedback galaxies is consistent with Section \ref{subsec:Enhancements}, demonstrating that accretion rate enhancements are lower in galaxies with lower gas mass or gas fraction.
The bottom panel of Figure \ref{fig:DMdot_DSFR_TPM_split} shows the star formation rate enhancement separated by feedback mode, where the solid lines show the star formation rate enhancement calculated within twice the stellar half-mass radius and dashed lines within one stellar half-mass radius. We include a second radius in order to compare the global and central star formation rate enhancements. Beginning with the solid lines in Figure \ref{fig:DMdot_DSFR_TPM_split}, we see that the two feedback modes show different behaviour in the first 500 Myrs post-merger. Radiative mode feedback galaxies show star formation rate enhancements for $\sim$ 500 Myrs after the merger. We therefore find that galaxies in radiative mode feedback, on average, have both accretion rate and star formation rate enhancements within the first few hundred Myrs. In contrast, kinetic mode feedback galaxies show a relative star formation \textit{suppression} for 500 Myrs after the merger. Therefore, Figure \ref{fig:DMdot_DSFR_TPM_split} demonstrates that despite an increased supply of gas, resulting in enhanced SMBH accretion rates, kinetic mode post-mergers display lower than average rates of star formation within $\sim $ 500 Myrs of coalescence. Our result that, on average, kinetic mode feedback galaxies demonstrate both relatively enhanced SMBH accretion rates and relatively suppressed of star formation rates may be explained considering the relationship between star formation quenching and kinetic mode feedback in TNG \citep[][]{Davies2020,Luo2020,Terrazas2020,Nelson2021,Piotrowska2021}, which we will discuss in further detail in Section \ref{subsec:feedback}.
Another possible explanation for the star formation rate suppression as well as the difference in timescales between $\Delta \dot M_{BH}$ and $\Delta SFR$ is that the enhancement in accretion rate is measured in the accretion region for the simulation, i.e. within the cells in the immediate vicinity of the black hole particle, whereas the star formation rate is measured over a much larger spatial scale, within twice the stellar half-mass radius. We therefore investigate whether central SFR enhancements are present in the PM sample. In the bottom panel of Figure \ref{fig:DMdot_DSFR_TPM_split}, the dashed lines represent the star formation rate enhancement calculated within one stellar half-mass radius, $\Delta SFR$\textsubscript{half}, once again split by feedback mode into the blue and red lines. Comparing the solid lines, $\Delta SFR$, with the dashed lines, $\Delta SFR$\textsubscript{half}, we can see how a more centralized aperture, although still much larger in spatial extent than the BH accretion region, affects the timescales of the star formation rate enhancements. For the radiative feedback mode galaxies, we see that the SFR enhancement peak is higher, reaching almost 0.4 dex. We also see that the enhancement is slightly longer lived. In the kinetic mode galaxies, we see that the star formation rate is slightly enhanced, 0.1 dex, within the $\sim$500 Myr window, demonstrating relative star formation rate enhancements despite suppressed star formation rates on a larger spatial scale.
\begin{figure}
\includegraphics[width=\columnwidth]{DMdot_DSFR_vs_TPM_sep_line_2Gyr_legend.pdf}
\caption{Running median of the accretion rate enhancement (top panel) and the star formation rate enhancement (bottom panel) over the time post-merger. The dark blue line represents the population in radiative mode feedback, and the orange line represents the population in kinetic mode feedback. The solid lines show the star formation rate measured within twice the stellar half-mass radius. The dashed lines represent the star formation rate enhancement calculated in a smaller aperture of one stellar half-mass radius. The error on the y-axis represents the 25th and 75th percentiles associated with the median $\Delta \dot M_{BH}$ and $\Delta SFR$ at that time post-merger.}
\label{fig:DMdot_DSFR_TPM_split}
\end{figure}
\subsection{Correlation of star formation rate and accretion rate enhancements}
\label{subsec:correlations}
In the previous section, Figure \ref{fig:DMdot_DSFR_TPM} demonstrated that within $\sim $ 500 Myrs of the merger, on average, galaxies have both star formation rate enhancements and accretion rate enhancements. However, Figure \ref{fig:DMdot_DSFR_TPM} does not assess whether the two enhancements are temporally correlated in a given galaxy. Although the merger process triggers gas inflows \citep[][]{Hernquist1989a,Barnes1991,Mihos1996,DiMatteo2007,Capelo2016,Blumenthal2018} that might naturally lead to quasi-simultaneous enhancements in both nuclear star formation and BH accretion \citep[][]{Sanders1988,DiMatteo2005,Springel2005,Hopkins2008}, observations show little evidence for a starburst-AGN connection outside of the most luminous systems \citep[][]{RowanRobinson1995,Schweitzer2006,Lutz2010,Shao2010,Santini2012,Rosario2015}. High resolution binary merger simulations have also demonstrated the lack of temporal correlations for the majority of the duration of the merger, with \cite[][]{Volonteri20151} finding a correlation only during the 200-300 Myrs over which the galaxies coalesce. The challenge is the different timescales of the two processes; BH accretion is stochastic on very rapid timescales, whereas star formation is more sustained. Observations that take a snapshot of a single point in time can not capture any potential extended connection between these processes. Simulation snapshots suffer from the same effect. Nonetheless, a merger event pinpoints the time of major gas inflow and may therefore be expected to demonstrate a connection between star formation and accretion.
In order to test for a correlation between the SFR enhancement and SMBH accretion rate enhancements in post-mergr galaxies, and determine whether selecting the post-mergers within different windows of time post-merger may affect a measured correlation, we consider the relationship between $\Delta \dot M_{BH}$ and $\Delta SFR$ on a galaxy by galaxy basis. We note that for the remainder of this section, where we compare $\Delta \dot M_{BH}$ and $\Delta SFR$ for each individual galaxy, we use a single matching criteria to calculate both $\Delta \dot M_{BH}$ and $\Delta SFR$, where we include the star forming vs passive classification discussed in Section \ref{subsec:timescale} to our fiducial matching scheme outline in Section \ref{subsec:matching}. That is, for each post-merger we identify a set of suitable controls and calculate $\Delta \dot M_{BH}$ and $\Delta SFR$ relative to the same set of controls.
Figure \ref{fig:DMdot_vs_DSFR} shows the distribution of $\Delta \dot M_{BH}$ vs $\Delta SFR$ on a galaxy by galaxy basis, divided into three samples based on time post-merger. The percentages of galaxies occupying each quadrant are shown, where a high occupation fraction in the top right quadrant would be indicative of an excess of simultaneous enhancements. We find that within 200 Myrs of the merger, 42\% of post-mergers have both accretion rate and SFR enhancements. However, the majority of post-mergers occupy each of the other quadrants, demonstrating that the star formation and accretion rate processes are not generally synchronized in post-mergers. We perform a Pearson correlation test for each window of time post-merger, with the correlation coefficient quoted in the textbox of each panel. The correlation coefficient is strongest within 200 Myrs of the merger and decreases with each subsequent time bin. However, even in the shortest time-since-merger interval the correlation between SFR and SMBH accretion rate enhancements is modest (correlation coefficient = 0.29). Our result therefore demonstrates that the synchronicity or correlation between SFR and SMBH accretion rate enhancements is affected by the timescale on which the galaxy is observed. We also demonstrate that a significant correlation between $\Delta \dot M_{BH}$ and $\Delta SFR$ is only present in the first few hundred Myrs post-merger, consistent with the timescale of temporally correlated SFR and SMBH accretion rates in \cite{Volonteri20151}. Finally, that even within 200 Myrs of coalescence, the majority of post-mergers do not have synchronized $\Delta SFR$ and $\Delta \dot M_{BH}$.
\begin{figure*}
\includegraphics[width=2\columnwidth]{DMdot_vs_DSFR_TPM.pdf}
\caption{Accretion rate enhancements vs star formation rate enhancements, where each panel represents bins of time post-merger (TPM in Gyrs), since the most recent 1:10 mass ratio merger. The percentages represent the number of galaxies within each quadrant relative to the total number of galaxies in the panel. The Pearson correlation coefficient is shown in the light blue text box.}
\label{fig:DMdot_vs_DSFR}
\end{figure*}
Returning to the left-most panel of Figure \ref{fig:DMdot_vs_DSFR}, we note that despite a bias for co-incident positive $\Delta \dot M_{BH}$ and $\Delta SFR$, the majority of post-mergers do not demonstrate synchronicity in SFR and SMBH accretion rate enhancements. We explore the diversity of the post-mergers within 200 Myrs of the merger in Figure \ref{fig:DMdot_vs_DSFR_3x3}, which shows the distribution of the post-mergers in $\Delta \dot M_{BH}$ and $\Delta SFR$ space, separated into bins of mass ratio and gas mass. The panels are organized such that gas mass is increasing from top to bottom and mass ratio is increasing from left to right.
The post-mergers in the bottom row of panels of Figure \ref{fig:DMdot_vs_DSFR_3x3}, corresponding to a gas mass $10^{9.75-11} M_{\odot}$, demonstrate that the majority of gas-rich post-mergers have both positive $\Delta \dot M_{BH}$ and $\Delta SFR$ (i.e. over 50\% are in the top right quadrant). In fact, comparing the bottom left panel (gas-rich `minor' mergers) to the bottom right panel (gas-rich major mergers), we see that gas-rich major mergers are more likely to have synchronicity (51\% compared with 76\%). We also find that overall, gas rich major mergers are more likely to produce SMBH accretion rate enhancements (68\% positive $\Delta \dot M_{BH}$ compared with 89\%). Our result suggests that gas rich major mergers more reliably produce accretion rate enhancements compared with minor mergers, though not necessarily stronger enhancements (result of Figure \ref{fig:DMdot_vs_MassRatio}).
The panels of the top row of Figure \ref{fig:DMdot_vs_DSFR_3x3}, corresponding to gas mass $10^{7-9.75} M_{\odot}$, show a slight bias for galaxies to occupy the left-side quadrants, corresponding to relatively lower star formation rates. A star formation rate suppression in lower gas mass post-mergers is consistent with the results of Sections \ref{subsec:Enhancements} and \ref{subsec:timescale}. We also emphasize that we are considering the gas mass of the galaxy post-coalescence, an important distinction if comparing our results with previous work such as \cite{DiMatteo2007} and \cite{Scudder2015}. Specifically, \cite{DiMatteo2007} do not find a strong dependence of star formation rate enhancement on the initial gas mass (i.e. the amount of gas available just before coalescence), while \cite{Scudder2015} find that galaxies with the lowest initial gas fraction have the highest SFR enhancements due to interactions. Our results demonstrate that star formation rate enhancements are less likely in post-mergers with a low post-coalescence gas mass, but do not comment on the relationship with initial gas mass. Our dependence of $\Delta SFR$ on gas mass may be explained, in part, by the correlation between SFR and gas fraction (as was demonstrated in \citealt[][]{Scudder2015} and \citealt[][]{Hani2020}). However, we note that \cite{Hani2020} still find a dependence of $\Delta SFR$ on gas fraction even when explicitly control matching on gas mass. If we focus on the top right panel of Figure \ref{fig:DMdot_vs_DSFR_3x3}, we also find that the bias for negative $\Delta SFR$ is present even in the highest mass ratio mergers. Therefore, our results suggest that even major mergers are unlikely to produce star formation rate enhancements if they have a low gas mass.
Overall, Figure \ref{fig:DMdot_vs_DSFR_3x3} demonstrates that the strongest bias for synchronicity and the strongest correlation (coefficient=0.47) occurs in gas-rich major mergers. We note that for the results presented in Figure \ref{fig:DMdot_vs_DSFR_3x3}, we find the same qualitative and similar quantitative results when separating the post-mergers by gas fraction rather than gas mass.
\begin{figure*}
\includegraphics[width=1.5\columnwidth]{DMdot_vs_DSFR_MR_2x2.pdf}
\caption{Accretion rate enhancement vs star formation rate enhancement for galaxies within 200 Myrs of a 1:10 mass ratio merger. The plots are organized by increasing gas mass, in units $\mathrm{log}_{10}\mathrm{M_{\odot}}$, from bottom to top and increasing mass ratio from left to right. The percentages represent the number of galaxies within each quadrant relative to the total number of galaxies in each panel. The Pearson correlation coefficient is shown in the light blue text box for each quadrant}
\label{fig:DMdot_vs_DSFR_3x3}
\end{figure*}
\section{Discussion}
\label{sec:Discussion}
\subsection{Effect of Resolution and Numerical Considerations}
\label{subsec:ResolutionAndEtc}
In the work presented here, we use the intermediate volume and resolution run of TNG, TNG100-1. We can repeat our experiment using the large volume and low resolution run, TNG300-1, which has a $(302.6 \, \mathrm{Mpc})^3$ volume, a baryonic resolution of $1.1\times 10^7 \mathrm{M_{\odot}}$, and a dark matter resolution of $5.9 \times 10^7 \mathrm{M_{\odot}}$ \citep[][]{Weinberger2017,Pillepich2018}. We obtain a sample of 25576 successfully matched post-merger galaxies from TNG300-1. Figure \ref{fig:DMdot_hist_res} shows the accretion rate enhancement parameter for the post-merger sample from both resolutions, TNG100 shown in teal and TNG300 shown in the dashed line. The median $\Delta \dot M_{BH}$ of the TNG300 post-merger population is $\sim$0.27 dex, corresponding to an accretion rate enhancement of roughly 90\%, slightly higher but consistent with the median enhancement for TNG100. Therefore, we demonstrate the comparability of our results between the two resolutions and volumes.
\begin{figure}
\includegraphics[width=\columnwidth]{DMdot_hist_res_compare.pdf}
\caption{Histogram of accretion rate enhancements for the post-merger populations from the intermediate resolution, TNG100-1, and low resolution, TNG300, simulations. TNG100 is represented with the teal line, and TNG300 is represented with the dashed line.}
\label{fig:DMdot_hist_res}
\end{figure}
A limitation of any work looking at SMBH accretion rates using cosmological scale simulations is the subgrid model required to calculate $\dot M_{BH}$. Subgrid models are an essential part of cosmological simulations, as they include important physical processes, such as stellar and SMBH feedback, that occur on scales below the resolution of the simulation. As mentioned, the TNG physics model uses a Bondi-Hoyle-Lyttleton accretion model \citep[][]{Weinberger2017,Pillepich2018}. However, the accretion model includes simplifying assumptions such as spherical symmetry, which would not hold for an accretion disk of material surrounding a SMBH. \cite{NegriVolonteri2017} demonstrated the variability of accretion rates calculated by different Bondi models and the dependance on simulation choices such as how the local density and sound speed are sampled. They find that Bondi models often overestimate the accretion rate, particularly when AGN feedback is inefficient at evacuating gas from the central region of the galaxy. However, they also find that in some simulations the accretion rate can be underestimated, which occurs when simulation cells of hot gas are over-represented in calculation of local parameters. Therefore, while we are unable to comment on the effects of mergers on the subparsec scales around the SMBH, the work presented here provides insight on how a merger event can impact the supply of material to the nuclear region, which may go on to truly accrete onto the SMBH.
An additional limitation of this work concerns the implementation of SMBH relocation in cosmological simulations, and the subsequent effect on SMBH accretion and SMBH mergers. \cite{Bahe2022} investigate the effects of SMBH repositioning in simulations using the EAGLE galaxy physics model, and demonstrate that the repositioning of SMBHs to the potential minima can result in significant boosts to SMBH accretion rate due to the increased density at the gravitational potential minima and the reduced relative velocity of the SMBH after repositioning. In addition, SMBH repositioning promotes early SMBH merging as it places the SMBHs of the merging galaxies in close proximity. A higher SMBH mass, due to ‘premature’ merging, will promote higher accretion rates due to the dependence on SMBH mass in the Bondi model. Therefore, we may expect the TNG model to overestimate accretion rates in the PM sample since SMBHs in merging galaxies do not wander and may merge prematurely. In addition, SMBH mergers may further complicate the timescale and delay between processes like starbursts and AGN activity. For example, \cite{Ni2022} demonstrate using the ASTRID cosmological simulations, which do not anchor SMBH particles to the gravitational minima and instead include a subgrid recipe to compensate for unresolved dynamical friction, that there is a delay of $\sim $200 Myrs between first close encounter of SMBH pairs and the SMBH merger. The above limitations highlight the imperative to investigate the galaxy merger and AGN connection in other cosmological simulations with varying subgrid implementations for SMBH physics.
\subsection{Implications for feedback and quenching}
\label{subsec:feedback}
In Section \ref{subsec:timescale}, we suggested that the enhancement of SMBH accretion rates in post-mergers with kinetic mode AGN feedback may have implications for galaxy quenching in TNG. Numerous studies have demonstrated that in TNG effective kinetic-mode AGN feedback is responsible for galaxy quenching \citep[][]{Davies2020,Luo2020,Terrazas2020,Nelson2021,Piotrowska2021}. Here, we briefly comment on how merger induced accretion enhancements may influence galaxy quenching in TNG. For simplicity, we can consider idealistic scenarios through which the enhanced accretion rates in post-merger galaxies may affect the quenching process.
First, increased accretion rates could result in post-merger galaxies that are in a high accretion state and are therefore using the radiative mode feedback model, which is ineffective at quenching in the TNG model. Therefore, mergers in TNG could inhibit or delay the quenching process, as galaxies may be in an inefficient feedback mode (in terms of quenching star formation). Since the median enhancement of accretion rates is subtle, $\sim1.7$ times higher than controls, a delayed quenching effect would be most applicable in galaxies that have accretion rates close to the transition between high and low accretion states, as defined in Section \ref{subsec:Illustris}.
Alternatively, increased SMBH accretion rates may enhance quenching in TNG. First, accretion rate enhancements in galaxies with low accretion rates could increase the energy input via effective kinetic mode feedback and contribute to the quenching of galaxies. In addition, enhancement in accretion rate can promote growth of the SMBH. According to the TNG model, the transition between radiative and kinetic mode feedback is strongly dependent on the SMBH mass, such that high mass SMBHs are more likely to be have kinetic mode feedback. Therefore, post-mergers may transition towards effective kinetic mode feedback sooner, promoting an excess of quenching.
\cite{Quai2021} explicitly investigate quenching in post-merger galaxies in TNG and conclude that quenching is rare in post-merger galaxies, but that there is an excess of quenched post-mergers when compared with matched controls. Specifically, \cite{Quai2021} find that quenching in post-mergers occurs only in galaxies that were already predisposed to quenching. The rarity of quenched post-mergers would be consistent with the first scenario outlined above. Post-mergers with strongly enhanced accretion rates will likely be in radiative feedback mode, where the SMBH feedback would not interact with the host galaxy in a way that promotes quenching, leading to a rarity of quenched post-merger galaxies. Furthermore, the merger process may introduce more gas into the nuclear region of the host galaxy, prolonging the lifetime of the high accretion state and preventing quenching.
The second scenario we have proposed is consistent with the slight excess of quenched post-mergers found in \cite{Quai2021}, where increased effective kinetic mode feedback speeds up quenching in the post-merger galaxies that are predisposed to quench. Furthermore, a truncation of star formation in the massive post-merger galaxies following a period of enhanced AGN activity would be consistent with the findings of \cite{Dubois2016} who observe a decrease of star formation in massive galaxies due to merger induced AGN activity. In addition, we discuss how our results relate to observational studies in Section \ref{subsec:observations}.
\subsection{What fraction of mergers are AGN?}
\label{subsec:fractionAGN}
So far we have demonstrated that post-merger galaxies have (on average) enhanced accretion rates. We have also demonstrated that not all post-mergers show enhancements in their SMBH accretion rate. Next we address the question of how frequently mergers actually trigger an AGN. Specifically, in this experiment, we will quantify what fraction of all post-mergers will have a high accretion event within 500 Myrs of coalescence.
To begin, we define an enhancement fraction, $f_{\mathrm{enhanced}}$, as the number of galaxies that have an SMBH accretion rate at or above a cutoff, $\dot M_{BH}^{\mathrm{cutoff}}$, divided by the total number of galaxies in the sample. We calculate $f_{\mathrm{enhanced}}$ for both the post-merger and non-merger sample, defined in Section \ref{subsec: PM and NM}, following the post-merger for 500 Myrs after the merger and the non-merger sample for 500 Myrs of secular evolution. The result of the experiment is shown in Figure \ref{fig:fEnhancedAcc}.
The top panel of Figure \ref{fig:fEnhancedAcc} shows $f_{\mathrm{enhanced}}$ as a function of $\dot M_{BH}^{\mathrm{cutoff}}$, where the post-merger sample is shown in the solid teal line and the non-merger sample is shown in the dashed yellow line. The top axis shows the associated bolometric luminosity for the SMBH accretion rate cutoff, calculated as 10\% of the accretion mass energy, or $0.1 \dot M_{BH}c^2$. We demonstrate that $\sim$ 60\% of post-merger galaxies have accretion rates exceeding $L_{bol} > 10^{43} \mathrm{erg/s}$ within 500 Myrs of the merger. However, we find that 50\% of non-merger galaxies also have an AGN phase of $L_{bol} > 10^{43} \mathrm{erg/s}$ within the same time period. Our result demonstrates that accretion rate events exceeding $L_{bol} > 10^{43} \mathrm{erg/s}$ are common in both samples, though slightly more common in the post-mergers, as expected.
In the bottom panel of Figure \ref{fig:fEnhancedAcc} we show the ratio of the post-merger to non-merger $f_{\mathrm{enhanced}}$ (the ratio of the blue solid line to the yellow dashed line from the top panel), or the fractional excess of AGN in the post-merger. We demonstrate that even though the fraction of galaxies that have an AGN phase decreases with increasing luminosity for both samples (as expected), the fractional excess of post-mergers with an AGN phase increases with luminosity. Therefore while less than 10\% of post-mergers will achieve accretion rates exceeding $L_{bol} > 10^{45} \mathrm{erg/s}$, four times more AGN appear in the post-merger sample than the non-merger sample.
Overall, we find that only a small fraction of mergers have high accretion rate events but that post-mergers are more likely to have an AGN event than a non-merger as a function of luminosity, consistent with results from the Magneticum Pathfinder simulation \citep[][]{Steinborn2018} and the EAGLE simulation \citep[][]{McAlpine2020}, as well as complimentary to \cite{Bhowmick2020} who find that high luminosity AGN are more likely to appear in environments with a higher density of SMBHs in TNG.
\begin{figure}
\includegraphics[width=\columnwidth]{fracEnhancedDotMbh_legend.pdf}
\caption{The top panel shows the fraction of galaxies that have a SMBH accretion rate at or higher than $\dot M_{BH}^{\mathrm{cutoff}}$ within a period of 500 Myrs, relative to the total number of galaxies. The teal line corresponds to the post-merger sample and the yellow dashed line is the non-merger sample, as defined in Section \ref{subsec: PM and NM}. The top x-axis corresponds to the AGN luminosity, calculated as $L_{\mathrm{bol}}=0.1\dot M_{BH}c^2$. The bottom panel shows the ratio of the fraction of post-merger galaxies to non-merger galaxies, or the fractional excess of AGN in the post-merger sample. The error in the shaded region is the Poisson error reflecting the number of galaxies.}
\label{fig:fEnhancedAcc}
\end{figure}
\subsection{What fraction of AGN are mergers?}
\label{subsec:fractionMergers}
A complementary question to the one posed in Section \ref{subsec:fractionAGN} is what fraction of galaxies with high accretion rates are mergers? In Section \ref{subsec:fractionAGN}, we demonstrated that the post-merger sample galaxies are more likely to be AGN than the non-merger sample of galaxies. However, post-mergers are rare in both TNG and the observed universe and thus may not be the major pathway to AGN triggering. Therefore, we measure the merger fraction of AGN in TNG100-1 in order to quantify the contribution of mergers to the total AGN population in the simulation.
To investigate the merger fraction of AGN, we first select an `AGN sample' by selecting all galaxies that have a SMBH accretion rate at or above $\dot M_{BH}^{\mathrm{cutoff}}$. In this experiment, the merger fraction is defined as the fraction of galaxies that have had a merger in the last 500 Myrs. Figure \ref{fig:fMerger} shows the merger fraction as a function of the cutoff accretion rate. The x-axis defines the AGN sample used to calculate the merger fraction, where the AGN sample will consist of all galaxies with accretion rates at or exceeding $\dot M_{BH}^{\mathrm{cutoff}}$ (and meeting the selection criteria outlined in Section \ref{subsec: ID mergers}, i.e $M_{\star} > 10^{10} \mathrm{M_{\odot}}$ and $z<1$). Once again, we show the equivalent bolometric luminosity along the top axis, calculated as $L_{bol} = 0.1 \dot M_{BH}c^2$. Figure \ref{fig:fMerger} shows that the merger fraction increases as a function of accretion rate (or, equivalently, luminosity). The horizontal dashed line represents the total merger fraction for the entire sample of galaxies, $f_{\mathrm{merger}}\sim$ 3\%. We see that the merger fraction of the AGN sample exceeds the total $f_{\mathrm{merger}}$ around $L_{bol}\sim 10^{44} \mathrm{erg/s}$. Therefore, there is no significant luminosity dependence of the merger fraction for $L_{bol}\lesssim 10^{44} \mathrm{erg/s}$. Beyond $L_{bol}\sim 10^{44} \mathrm{erg/s}$ the merger fraction increases with AGN luminosity, with a peak merger fraction of $\sim$13\%.
Although there is a luminosity dependence on the merger fraction for $L_{bol}\gtrsim 10^{44} \mathrm{erg/s}$, we find that recent mergers never dominate the AGN sample, even at the highest luminosities. This result is once again consistent with the results of \cite{Steinborn2018} and \cite{McAlpine2020} for the Magneticum Pathfinder and EAGLE simulations. We note that the merger fraction that we have defined does not include galaxies in the pre-merger phase, which may contribute to the remaining $\sim$90\% of galaxies at or above $10^{45}$ erg/s. There may also be minor mergers that are not accounted for in the merger fraction. However, the result of Figure \ref{fig:fMerger} also supports a scenario where secular processes play a significant role in fueling the highest luminosity AGN (at least in the redshift regime 0-1). This experiment concludes that post-mergers contribute more to the AGN population at the highest accretion rates, but they are non-dominant over the entire accretion rate range.
\begin{figure}
\includegraphics[width=\columnwidth]{AGNMergerFrac_legend.pdf}
\caption{The teal line shows the fraction of galaxies that have had a merger of mass ratio greater than 1:10 within 500 the past Myrs. The x-axis defines the sample used to calculate the merger fraction from left to right, where only galaxies with a black hole accretion rate of at least $\dot M_{BH}^{\mathrm{cutoff}}$ are included in the calculation of the merger fraction. The top x-axis shows the corresponding AGN luminosity, calculated as $L_{\mathrm{bol}}=0.1\dot M_{BH}c^2$. The error in the shaded region is the Poisson error reflecting the number of galaxies. The horizontal dashed line represents the overall merger fraction of the sample, $\sim$3\%, for all galaxies in TNG with stellar mass of at least $10^{10} M_{\odot}$ and a redshift $< 1$.}
\label{fig:fMerger}
\end{figure}
\subsection{Comparisons to observations}
\label{subsec:observations}
In the previous sections, we have demonstrated that there is a population averaged enhancement in SMBH accretion rate in post-mergers and that post-mergers are more likely to host highly accreting SMBHs. Our findings are in agreement with the observational studies that find that AGN are more likely to appear in a post-merger or interacting galaxy sample when compared with undisturbed controls \citep{Woods2007,Alonso2007,Koss2010,Ellison2011,RamosAlmeida2012,Ellison2013,Satyapal2014,Hong2015,Kocevski2015,Rosario2015,Weston2017,Hewlett2017,Goulding2018,Ellison2019,Gao2020,Marian2020,Pierce2022}. We find that at most, AGN occur four times more commonly in the post-mergers compared with the non-mergers. The maximum excess of AGN is quantitatively consistent with numerous observational studies finding an excess of approximately four or lower \citep[][]{Ellison2013,Goulding2018,Gao2020}, though some studies find a significantly larger excess of AGN in merging galaxies \citep[][]{Satyapal2014,Weston2017}. Such variations in the observed excess may be, in part, due to the different AGN selection techniques, which have been demonstrated to result in different AGN fractions \citep[][]{Secrest2020,Bickley2022Submitted}. Variations may also exist within the same AGN selection techniques, as demonstrated in \cite{Bickley2022Submitted} who find the AGN excess may also depend on the strength of visual disturbances in the PM sample..
In addition, we demonstrate that the majority of high luminosity AGN are not associated to recent mergers. Our results are in agreement with observations which find the most luminous AGN are not mergers \citep{Cisternas2011,Schawinski2011,Schawinski2012,Kocevski2012,Bohm2013,Villforth2014,Mechtley2016,Hewlett2017,Villforth2017,Marian2019,Lambrides2021}, but are in contrast to observational studies that find the majority of quasars have disturbed morphology \citep[][]{Treister2012,Glikman2015,Fan2016,Goulding2018,UrbanoMayorgas2019}. Therefore, our results support a scenario where mergers can, but do not always, trigger AGN, and where mergers do not play a major role is triggering even the most luminous AGN.
Despite an increased likelihood to host highly accreting SMBHs, we demonstrate that the majority of mergers do not host a highly accreting SMBH, suggesting the majority of mergers will not undergo a strong AGN feedback event. In fact, observations have demonstrated that mergers have normal (or even enhanced) gas fractions (e.g. \citealt{Ellison20152,Ellison2018,Violino2018,Pan2018}), consistent with a lack of strong instantaneous feedback. However, observations do find that AGN can have effects on the small (kpc or below) scale interstellar medium (e.g. \citealt[][]{Oosterloo2017,Izumi2020,Ellison2021,GarciaBurillo2021,RamosAlmeida2022,Saito2022}).
In Section \ref{subsec:correlations}, we demonstrated that the strength of the star formation rate and SMBH accretion rate enhancements, on a galaxy by galaxy basis, are not generally correlated, in agreement with simulations such as \cite{Hickox2014,Volonteri20151}. However, we do find a connection between star formation rate and SMBH accretion rate in post-mergers with both a high mass ratio and significant amount of gas. Such a connection may reflect the general preference for AGN to reside in gas rich and star forming galaxies, as demonstrated in simulations \citep[][]{Ward2022} and observations \citep[][]{Rosario20131,Bernhard2016,Jarvis2020,Ellison20192,Shangguan2020,Xie2021,Koss2021}. In addition, we demonstrate in Figure \ref{fig:DMdot_vs_DSFR_3x3} numerous individual cases where a post-merger may have a strongly enhanced SMBH accretion rate and simultaneously suppressed or normal star formation rate relative to controls. Our results are therefore consistent with observations which show that most AGN (not strictly post-mergers) have typical rates of star formation \citep[][]{Rosario20131,Rosario2015}, although some observations demonstrate higher than normal star formation rates in the highest luminosity AGN \citep[][]{Schweitzer2006,Lutz2010,Shao2010,Santini2012}.
Finally, in Section \ref{subsec:ResolutionAndEtc}, we comment the consistency between our results and the rarity of rapid quenching in TNG post-mergers found in \cite{Quai2021}. However, \cite{Ellison2022} demonstrate that in observations, there exists a significant excess of rapidly/recently quenched post starburst galaxies (PSBs) in galaxy mergers. There is still uncertainty about the quenching mechanisms within PSBs, as observations show they do not lack molecular gas (e.g. \citealt{Rowlands2015,French2015}). Instead, it might be that PSBs lack dense star-forming gas \citep[][]{French20181}, perhaps due to enhanced turbulence (e.g. \citealt[][]{Smercina2022}). Such mechanisms may not be captured within TNG due to the resolution constraints and effective equation of state treatment of the interstellar medium, warranting an investigation of AGN and rapid quenching in post-merger galaxies using higher resolution simulations or simulations with varied physical models.
\section{Conclusions}
\label{sec:conclusion}
In the work presented here, we study the effect of galaxy mergers on SMBH accretion rates in a collection of 1563 post-merger galaxies from the IllustrisTNG simulation. Our post-merger sample reflects a diverse collection of stellar masses, star formation rates, gas masses, environments, and mass ratios. For each post-merger galaxy in our sample, we identify control galaxies for comparison, and are able to isolate the effect of the merger on the instantaneous SMBH accretion rates. Our results are summarized in the following points:
\begin{itemize}
\item On average, post-merger black hole accretion rates are enhanced by a factor of 0.23 dex, corresponding to a $\sim$70\% increase in accretion rate (Figure \ref{fig:DMdot_hist_all}). We find that there is significant variation in the accretion rate enhancement on a galaxy by galaxy basis, where $\sim$ 30\% of post-mergers have accretion rates consistent with no enhancement or lower than controls.
\medskip
\item SMBH accretion rate enhancements persist for up to two Gyrs after coalescence, and are significantly longer lived than SFR enhancements, which only persist for $\sim$500 Myrs post-merger (Figure \ref{fig:DMdot_DSFR_TPM}).
\medskip
\item We find that the co-incidence of accretion rate and SFR enhancements is most pronounced within the first few hundred Myrs post-merger, and that the correlation strength of the enhancements decreases with time post-merger, shown in Figure \ref{fig:DMdot_vs_DSFR}. However, even within 200 Myrs of coalescence, the majority of post-mergers do not demonstrate synchronicity in SMBH acretion rate enhancements and SFR enhancements.
\medskip
\item In Figure \ref{fig:DMdot_vs_DSFR_3x3}, we find that gas rich major mergers demonstrate the strongest correlation coefficient between $\Delta SFR$ and $\Delta \dot M_{BH}$, and the majority have enhancements in both. However, we find that the presence of gas alone is an insufficient criteria for enhancing SMBH accretion rates, and that gas rich non-mergers, on average, do not have positive $\Delta \dot M_{BH}$ (Figure \ref{fig:DMdot_vs_Mgas}).
\medskip
\item Post-mergers are more likely to host high luminosity AGN than non-merger galaxies (Figure \ref{fig:fEnhancedAcc}), however high luminosity AGN are rare in the post-merger sample. For AGN luminosity in excess of $L_{\mathrm{bol}} > 10^{45} \mathrm{erg/s}$, corresponding to $\sim$10\% of post-mergers, there are four times more AGN in the post-merger sample than the matched non-merger sample.
\medskip
\item We find a luminosity dependence in the merger fraction of AGN for $L_{\mathrm{bol}} > 10^{44} \mathrm{erg/s}$ (Figure \ref{fig:fMerger}). However, we find that the majority of high luminosity AGN are not recent mergers, which at most only contribute $\sim$13\% to AGN with $L_{\mathrm{bol}} > 10^{45} \mathrm{erg/s}$.
\end{itemize}
\section*{Acknowledgements}
We thank the IllustrisTNG collaboration for making their data accessible. We thank the anonymous referee for their constructive comments. SBM acknowledges the receipt of a British Columbia Graduate Scholarship and the Dr. Margaret Perkins Hess Research Fellowship from the University of Victoria. This research was enabled in part by support provided by WestGrid (www.westgrid.ca) and Compute Canada (www.computecanada.ca). DRP gratefully acknowledges NSERC of Canada for a Discovery Grant which helped to fund this research.
\section*{Data Availability}
The data used in this work are publicly available at \hyperlink{}{https://www.tng-project.org}.
\bibliographystyle{mnras}
\section{Introduction}
\label{sec:introduction}
It is well established that galaxy mergers play a fundamental role in the growth of galaxies in the Universe. In addition to the assembly of the main matter components, the merger process is predicted to have short-lived, but dramatic, effects on numerous galactic properties.
The most obvious visible impact of a galaxy-galaxy interaction is the re-arrangement of stellar material, an effect which was seen in the earliest N-body simulations \citep[][]{Toomre1972,White1978,Roos1979,Villumsen1982}. Strong tidal forces result in disturbed morphologies including shells and tidal features that are dependent on the properties of both the initial galaxy members and the orbital properties of the interaction \citep[][]{Casteels2014,Nevin2019,Blumenthal2020,Patton2016}. The advent of hydrodynamical simulations further demonstrated that the generation of internal asymmetric structures produced torques that could drain angular momentum from the gas, causing it to flow to the centre and resulting in central starbursts \citep[][]{Hernquist1989a,Barnes1991,Mihos1996,DiMatteo2007,Capelo2016,Blumenthal2018}, a feature that is widely observed in both the pre- \citep[][]{Barton2000,Woods2006,Woods2007,Ellison2008,Woods2010,Scudder2012,Patton2013,Knapen2015,Cao2016} and post- \citep[][]{Ellison2008,Ellison2013,Thorp2019,Bickley2022} merger regimes in galaxy surveys.
Merger driven gas inflows have also been predicted to potentially fuel black hole growth \citep[][]{Sanders1988,DiMatteo2005,Springel2005,Hopkins2008,Capelo2015}. However, whilst there is broad agreement in the observational literature supporting merger-enhanced star formation, there continues to be controversy over the role of mergers in triggering active galactic nuclei (AGN). Although numerous observational studies do not find evidence for a merger-AGN connection \citep[][]{Cisternas2011,Schawinski2011,Kocevski2012,Bohm2013,Shah2020,Lambrides2021}, many others have demonstrated a statistically significant excess of AGN in interacting pairs and post-mergers \citep[][]{Alonso2007,Ellison2011,Bessiere2012,Ellison2013,Hong2015,Kocevski2015,Hewlett2017,Marian2020,Bickley2022Submitted}.
Recently, \cite{Ellison2019} explored the possibility that contradictory results may result from different experimental approaches; namely that some studies assess the excess of AGN in mergers, whereas others quantify the excess of mergers amongst AGN. Using a sample of low redshift AGN and high quality imaging from the Canada France Imaging Survey, \cite{Ellison2019} demonstrated that there is \textit{both} an excess of AGN in mergers \textit{and} an excess of mergers amongst AGN. At least at low redshift (and for the optical and mid-IR selection used by \citealt{Ellison2019}), experimental approach does not seem to explain the conflicting results.
There are numerous other factors that may affect our assessment of the merger-AGN connection. AGN are a multi-wavelength phenomenon that manifest across the electromagnetic spectrum and different selection methods identify different objects (see \citealt{HickoxAlexander2018} for a recent review). The extent of a connection between mergers and AGN may depend on the wavelengths at which the AGN is selected. For example, several studies have shown that the AGN excess in mergers is greater for mid-IR selected AGN, compared to those identified through their optical emission lines \citep[][]{Satyapal2014,Weston2017,Goulding2018,Ellison2019,Gao2020}, with no X-ray AGN enhancement found in low redshift post-mergers \citep[][]{Secrest2020}. However, even for selection in a given frequency regime disagreement can exist. For example, some studies of radio-selected AGN find no merger connection \citep[][]{Dicken2012,Ellison2015} whilst others do \citep[][]{RamosAlmeida2011,RamosAlmeida2012,Gao2020,Pierce2022}, a tension that could be linked to further distinctions within the radio population into high and low excitation sources \citep[][]{Chiaberge2015,Bernhard2022}. Disagreement also exists as to whether the dominance of merger triggering is linked to AGN luminosity \citep[][]{Schawinski2012,Treister2012,Villforth2014,Glikman2015,Mechtley2016,Villforth2017,Marian2019,Pierce2022}.
The above summary of the observational literature demonstrates the considerable diversity amongst recent results. However, there are a few broad statements regarding the merger-AGN connection that are relatively uncontroversial. First, that the majority of AGN are \textit{not} merger induced, a statement that is also supported by simulations \citep[][]{Steinborn2018,McAlpine2020}. Second, that \textit{some} AGN are probably triggered by interactions, although the fraction likely depends on subtleties such as redshift \citep[e.g.][]{Rosario2015} and selection method \citep[e.g.][]{Gao2020}. Third, it is the most obscured \citep[][]{Satyapal2014,Fan2016,Weston2017,Ellison2019,Gao2020} and most luminous \citep[][]{Schawinski2012,Treister2012,Glikman2015,Donley2018,Goulding2018,UrbanoMayorgas2019,Pierce2022} AGN that seem to be most likely to be linked to mergers.
Beyond the simple understanding of which mechanisms (merging versus various secular processes) lead to black hole growth, the potential for AGN triggering during the interaction is arguably most relevant for assessing the end point of the merger sequence. Since AGN have been widely implicated as a way to quench star formation \citep[][]{DiMatteo2005,Springel2005}, if most mergers experience an accretion+feedback event during the interaction, then we might expect most post-coalescence galaxies to rapidly shut-down their star formation \citep[][]{Hopkins2008}.
An additional complication to the proposed connections between mergers, star formation and AGN triggering, followed by feedback driven quenching, is the different timescales for these processes. High resolution simulations have shown that merger induced central starbursts and high SMBH accretion rates are for the most part not temporally correlated \citep[][]{Volonteri20151}. Furthermore, most observational studies do not find a connection between central starburst and AGN activity, except in the highest luminosity systems, an effect that is largely attributed to short timescales for AGN variability \citep[][]{RowanRobinson1995,Schweitzer2006,Lutz2010,Shao2010,Santini2012,Rosario20131,Rosario20132}.
In order to assess the impact of mergers on star formation, black hole accretion and quenching in a statistical and holistic way, we have been undertaking a series of studies using a large, state-of-the-art hydrodynamical simulation suite from IllustrisTNG (hereafter TNG; \citealt[][]{Nelson2017,Naiman2018,Mariancci2018,Pillepich20171,Springel2017}). Although large box simulations are lacking in resolution compared with idealized binary merger runs or cosmological zoom in simulations (e.g. \citealt{Hopkins2013,Moreno2015,Moreno2019}), they offer several advantages. First, many thousands of interactions can be drawn from the large sample of galaxies in a cosmological simulation, in contrast to the suites of binary mergers which typically contain only tens of interactions. Second, no a priori decisions are required to define either galaxy properties (morphologies, gas fractions etc.) or interaction properties (e.g. mass ratios, orbits etc.). Finally, the full cosmological setting provides a more realistic interaction with both the circumgalactic and intergalactic environments.
In the first paper in our series, \cite{Patton2020} demonstrated that enhancements in the star formation rate (SFR) of interacting galaxies in TNG are experienced in the pre-coalescence phase out to separations of a few hundred kpc. Although the exact number varies depending on the simulation used, triggered star formation out to large separations was found to be a common feature of all of simulations in the TNG suite. \cite{Hani2020} continued the assessment of SFR enhancement into the post-coalescence regime and showed that the elevated levels persist for typically 500 Myrs post-merger. \cite{Quai2021} demonstrated that although most post-mergers do not lead to rapid shut down of star formation, there is nonetheless an excess of quenched galaxies amongst simulated post-mergers compared with controls.
In the work presented here, we investigate the role of mergers in triggering AGN in TNG, specifically whether there is an enhancement of supermassive black hole accretion rates in the post-merger phase. In Section \ref{sec:methods}, we summarize the salient details of the TNG simulation, our merger sample selection, and our algorithm for matching post-merger galaxies to control comparisons. In Section \ref{sec:Results} we quantify the enhancement of black hole accretion rates in post-mergers and investigate whether these enhancements depend on galaxy and merger properties (Section \ref{subsec:Enhancements}). Furthermore, we quantify the timescale over which accretion rate enhancements persist (Section \ref{subsec:timescale}) and compare this with the triggered star formation in order to assess synchronicity between these processes (Section \ref{subsec:correlations}). Finally, we address whether mergers are likely to produce AGN activity and quantify the contribution of recent mergers to the total AGN population, i.e. what fraction of mergers produce high accretion rates (Section \ref{subsec:fractionAGN}) and what fraction of high accretion rate systems have undergone a recent merger (Section \ref{subsec:fractionMergers}).
\section{Methods}
\label{sec:methods}
\subsection{IllustrisTNG}
\label{subsec:Illustris}
We use the IllustrisTNG galaxy formation simulation, a state-of-the-art cosmological magneto-hydrodynamic simulation including galactic scale stellar feedback, stellar population evolution and chemical enrichment, primordial and metal-line gas cooling and heating, and multi-mode blackhole feedback \citep[][]{Nelson2017,Naiman2018,Mariancci2018,Pillepich20171,Springel2017}. In the work presented here, we use the TNG100-1 simulation run, the intermediate volume and resolution run of the fiducial TNG galaxy formation model, which has a $(110.7 \, \mathrm{Mpc})^3$ volume, a baryonic resolution of $1.4 \times 10^6 \, \mathrm{M_{\odot}}$, and a dark matter resolution of $7.5 \times 10^6 \, \mathrm{M_{\odot}}$.
We briefly describe the physical models for supermassive black hole seeding and feedback in TNG, with the full details available in \cite{Weinberger2017} and \cite{Pillepich2018}. Supermassive black holes, hereafter SMBHs, are seeded at $M_{BH} = 8 \times 10^5 \, \mathrm{h^{-1} M_{\odot}}$ at the centre of halos that meet a threshold mass of $5 \times 10^{10} \, \mathrm{h^{-1} M_{\odot}}$. The black holes can then grow by either accreting gas from the region surrounding the black hole, or by merging with other black holes.
SMBH accretion is calculated using a Bondi-Hoyle-Lyttleton subgrid model,
\begin{equation}
\dot M_{Bondi} = \frac{4 \pi G^2 M_{BH}^2 \rho}{c_s^3}
\label{eq:Bondi}
\end{equation}
where $G$ is the gravitational constant, $M_{BH}$ is the black hole mass, and $\rho$ and $c_s$ are the density and sound speed sampled in a kernel-weighted sphere centred on the SMBH, labelled the accretion region. The SMBH accretion rate is capped by the Eddington rate,
\begin{equation}
\dot M_{Edd} = \frac{4 \pi G M_{BH} m_p}{\epsilon_r \sigma_T c}
\label{eq:Edd}
\end{equation}
where $m_p$ is the proton mass, $\epsilon_r$ is the radiative accretion efficiency, $\sigma_T$ is the Thompson cross-section, and $c$ is the vacuum speed of light.
SMBHs are merged when the black holes are within the accretion regions of one-another. Furthermore, in order to prevent the wandering of SMBHs away from the halo centre, SMBHs are fixed to the local gravitational potential minima, which has the added effect of promptly merging SMBHs when their host subhalos coalesce.
TNG uses a dual feedback mode model, where SMBHs use a different feedback prescription depending on the SMBH accretion rate and SMBH mass. At high accretion rates, feedback is implemented according to a radiative mode model, and at low accretion rates the feedback is implemented according to a kinetic mode model. The classification of high vs. low accretion is made by calculating the ratio of the Bondi accretion rate to the Eddington accretion rate, where a high accretion rate is defined as $\dot M_{Bondi} \geq \chi \dot M_{Edd}$ and $\chi$ is defined as,
\begin{equation}
\chi = min\bigg[\chi_0 \bigg(\frac{M_{BH}}{10^8 M_{\odot}}\bigg)^{\beta},0.1\bigg].
\label{eq:chi}
\end{equation}
The maximum $\chi=0.1$ follows observational constraints set by X-ray binaries \cite{Dunn2010}, whereas the parameters $\chi_0$ and $\beta$ are simulation parameters tuned to 0.002 and 2 respectively.
For SMBHs with high accretion rates, the simulation uses a radiative feedback prescription, where a fraction of the accreted mass energy is injected as thermal energy into the region surrounding the black hole. At low accretion rates, once again a fraction of the accreted mass energy is injected into the surrouding BH region, however the energy injected is in the form of kinetic energy in randomized directions away from the black hole. For our work, focused on SMBH accretion rates, we do not comment in detail on the feedback prescriptions, and refer the reader to \cite{Weinberger2017}.
\subsection{Identifying galaxies and galaxy mergers in IllustrisTNG}
\label{subsec: ID mergers}
Before we can identify the population of galaxy mergers in the TNG simulation, we first apply selection constraints to our entire galaxy sample.
First, we exclude galaxies at redshift$>$1 from our sample because the increased frequency of mergers and interactions exacerbate issues such as numerical stripping and subhalo switching \citep[][]{Rodriguez-Gomez2015}, which makes robust merger identification more difficult. We then apply a stellar mass requirement of $10^9 M_{\odot}$, equivalent to $\sim 1000$ particles, to ensure our selected galaxies are properly resolved. We combine our minimum stellar mass requirement with our minimum merger mass ratio of 1:10, i.e. requiring post-mergers have progenitors of at least $10^9 M_{\odot}$, by requiring a minimum stellar mass of $10^{10} M_{\odot}$ in the post-merger sample.
We also apply an additional environmental constraint to remove galaxies susceptible to numerical stripping from ongoing interactions with neighbouring subhalos. Following the procedure of \cite{Patton2020}, we calculate $r_{sep}$,
\begin{equation}
r_{sep} = \frac{r}{R^{host}_{1/2} + R^{comp}_{1/2}}
\label{eq:rsep}
\end{equation}
where r is the 3D separation of the galaxies, and $R^{host}_{1/2}$ and $R^{comp}_{1/2}$ are the stellar half mass radii of the host and companion. \cite{Patton2020} demonstrated that TNG galaxies with $r_{sep} \lesssim 2$ had stripped stellar mass and we therefore excluded them from our galaxy sample.
Once we have identified appropriate candidate galaxies from TNG, we identify galaxy mergers using the merger-trees created by the \textsc{SUBLINK} algorithm \citep[][]{Rodriguez-Gomez2015}, which associate each galaxy in TNG with progenitor and/or and descendant galaxies. We identify galaxy mergers by selecting nodes within the merger trees. Nodes occur when particles which are assigned into distinct subhalos at a given snapshot are subsequently assigned into the same subhalo in the following snapshots.
In the work presented here, we calculate the mass ratio of the merger using a different methodology from previous works in this series. We apply additional steps, once again, as a precaution against numerical stripping and subhalo switching. For each merger, the mass ratio of the merger is calculated using the stellar mass within twice the stellar half-mass radius, in the snapshots leading up to coalescence of the subhalos, up to a maximum of 10 snapshots. From the mass ratio estimates, the maximum and minimum are excluded, which removes abnormally large or small values that are due to subhalo switching or significant numerical stripping. Finally, we take the average and standard deviation of the remaining mass ratio measurements to calculate the merger stellar mass ratio, $\mu$, and assign a merger mass ratio error, $\sigma_{\mu}$. For the work presented here, we define mergers as coalescence of subhalos with a merger mass ratio within $0.1 < \mu \leq 1$. We do not limit the overall post-merger sample using the mass ratio error except when specified.
\subsection{Post-merger and non-merger galaxy samples}
\label{subsec: PM and NM}
Hydrodynamical simulations have suggested that SMBH accretion rates peak post-coalescence \citep[][]{Hopkins2008}, a framework that is corroborated by some observational studies \citep[][]{Ellison2013,Satyapal2014,Bickley2022Submitted}. In the work presented here we focus on post-merger galaxies, and construct our post-merger sample using only galaxy mergers that have coalesced in the time between the previous and current snapshot, or within $\sim 160$ Myrs. Note that we discuss the long term effects on the post-merger sample in Section \ref{subsec:timescale}. Selecting galaxies immediately after coalescence, and applying the galaxy selection criteria outlined above, we identify a sample of 1973 post-merger galaxies from TNG100-1. The stellar mass, SMBH mass, redshift and mass ratio distributions of this complete post-merger sample are shown in the dotted teal line in Figure \ref{fig:PMdist}.
In addition to the post-mergers, we define a sample of non-mergers whose properties are representative of the post-mergers. We construct a non-merger sample in order to generate a comparative population that has the same underlying distribution as the post-merger sample, and in order to contextualize how the underlying stochasticity of SMBH accretion rates affect our calculation of accretion rate enhancements. The representative non-merger sample is constructed as follows. For each of the 1973 post-mergers, we identify one non-merging galaxy that is selected to be the single best match in redshift, stellar mass, gas mass, environment, and feedback mode (see Section \ref{subsec:matching} for details on the necessity of feedback mode matching) that has not undergone a merger of mass ratio $\mu > 0.1$ within the last two Gyrs. The environment is quantified using two parameters taken from \cite{Patton2020}: $r_1$ and $N_2$. $r_1$ is the 3D distance to the nearest neighbour, within two Mpc, which has a mass above 10\% of the target galaxy mass. $N_2$ is the number of neighbours within two Mpc of the galaxy centre, with a minimum mass of $10^{9} M_{\odot}$. The single best match is selected as an exact match in redshift and feedback mode, and the best simultaneous match of stellar mass, gas mass, $r_{1}$ and $N_{2}$ following the statistical weighting scheme of \cite{Patton2016}. We therefore begin with a sample
of 1973 post-mergers and 1973 non-mergers.
\begin{figure}
\includegraphics[width=\columnwidth]{PMsampleDist_legend.pdf}
\caption{Distribution of post-merger and non-merger sample (from top to bottom) in stellar mass, SMBH mass, redshift, and mass ratio. The dotted teal line shows the distribution of all the post-mergers, following our criteria, in TNG100-1. The solid teal line shows the distribution of the 1563 successfully control matched post-mergers, and the dashed yellow line shows the 1522 successfully control matched non-mergers.}
\label{fig:PMdist}
\end{figure}
\subsection{Control Matching Post-merger and Non-merger Galaxies}
\label{subsec:matching}
In order to quantify the difference between a given physical property (e.g. SMBH accretion rate) of an individual (post-merger or non-merger) galaxy and the expected value of a `normal' galaxy, we use a control matching algorithm to identify a sample of comparative controls for each individual galaxy in the post-merger and non-merger sample. The pool of possible control galaxies consists of all galaxies, meeting the criteria outlined in Section \ref{subsec: ID mergers}, that have not undergone a merger of mass ratio $\mu > 0.1$ within the last two Gyrs. We note that when control matching the non-mergers, we exclude galaxies that have already been assigned into the non-merger sample from the control pool, such that no galaxy can be matched to itself.
We begin by down-selecting the pool of potential control galaxies by imposing two cuts (which are effectively broad matching constraints). First, we require that the controls for a given galaxy are drawn from the same simulation snapshot, which corresponds to matching in redshift. Second, we require that the controls for a given galaxy are in the same instantaneous feedback mode as the post-merger (or non-merger), using the feedback mode classification outlined in Section \ref{subsec:Illustris}. The requirement of feedback mode matching diverges from the methods applied by \cite{Patton2020}, \cite{Hani2020}, and \cite{Quai2021}, but is necessary for the work presented here to avoid spurious features in the results. The issue is illustrated in Figure \ref{fig:Mdot_vs_Mstar} which shows the distribution of accretion rates as a function of stellar mass for all galaxies in TNG100-1, colour-coded to distinguish those in radiative feedback mode (dark blue) from those in kinetic feedback mode (red). Figure \ref{fig:Mdot_vs_Mstar} demonstrates that (at a fixed stellar mass) the accretion rate distributions are bimodal. Without feedback mode matching, post-merger galaxies may (artificially) exhibit extremely enhanced or suppressed accretion rates if they are matched to controls in the other feedback mode.
\begin{figure}
\includegraphics[width=\columnwidth]{Mdot_vs_Mstar_legend.pdf}
\caption{SMBH accretion rates as a function of stellar mass for all the TNG100-1 galaxies meeting the selection criteria of Section \ref{subsec: ID mergers}. The dark blue contours represent the sample in radiative mode feedback and the orange contours represent the sample in kinetic mode feedback.}
\label{fig:Mdot_vs_Mstar}
\end{figure}
In addition to snapshot (redshift) and feedback mode, the control galaxies for each post-merger and non-merger are matched on several further properties. First, the control sample is limited to galaxies whose stellar mass is matched to within $\pm 0.05$ dex and whose gas mass is matched to within $\pm 0.1$ dex. Next, the matched control sample is limited to galaxies within $\pm 10\%$ of $r_1$ and $N_2$. We allow all of the error tolerances to grow, 0.01 dex in stellar mass, 0.1 dex in gas mass, and $10\%$ in $r_1$ and $N_2$, up to four times until at least five matched control galaxies are found. If fewer than five control galaxies are found, then the post-merger is excluded from the sample. On average, galaxies have to grow their error tolerances twice to meet the required number of controls, and 410 post-merger galaxies (and 451 non-merger galaxies) did not meet the control requirements within permitted error tolerances after the maximum number of grows and were hence rejected from our sample. Therefore, our final matched sample consists of 1563 matched post-merger galaxies and 1522 non-merger galaxies, with an average of 10 matched controls each. The properties of the final post-merger and non-merger samples are shown in the solid teal and dashed yellow lines in Figure \ref{fig:PMdist}. We note that the top panel of Figure \ref{fig:PMdist} reveals that the control matching algorithm preferentially fails to match higher stellar mass post-mergers. It was determined that the strict error tolerance on stellar mass matching combined with gas mass matching resulted in higher mass post-mergers being less likely to find a sufficient number of controls. However, we note that both matching criteria are necessary in order to select an unbiased set of controls. Without the strict error tolerance, high mass post-mergers are preferentially matched to low mass controls. We therefore prioritize the selection of unbiased controls over the completeness of the matched post-merger sample.
Figure \ref{fig:ControlsMatch} shows the distribution of stellar mass, gas mass, $r_1$, and $N_2$ for the post-merger galaxies, shown in the teal line, and the matched control galaxies, shown in dashed pink line. In the figure, the distributions trace one another closely, demonstrating the success of the control matching methodology in identifying galaxies with similar properties to the post-merger sample. Although not shown in Figure \ref{fig:ControlsMatch}, we note that the non-merger sample and its controls exhibit similarly closely matched properties.
\begin{figure}
\includegraphics[width=\columnwidth]{MatchingHist_legend.pdf}
\caption{Distribution of parameters matched in control matching process: (from top to bottom) stellar mass and gas mass measured within twice the stellar half mass radius, distance to nearest neighbour within two Mpc, and number of neighbours within two Mpc. Post-mergers are represented with the teal line, and the controls are represented with the dashed pink line.}
\label{fig:ControlsMatch}
\end{figure}
Once the control matching procedure is complete, we then calculate a relative SMBH accretion rate enhancement, $\Delta \dot M_{BH}$, for each post-merger and non-merger galaxy as
\begin{equation}
\Delta \dot M_{BH} = log_{10}(\dot M_{BH}) - median[log_{10}(\dot M_{BH Controls})].
\label{eq:enhancement}
\end{equation}
\section{Results}
\label{sec:Results}
\subsection{Enhancement of accretion rates in post-merger galaxies}
\label{subsec:Enhancements}
Figure \ref{fig:DMdot_hist_all} shows the distribution of $\Delta \dot M_{BH}$ for the post-merger sample, shown in the teal line, and non-merger sample, shown in the yellow dashed line. The non-merger population peaks at 0 dex, consistent with no enhancement in the SMBH accretion rate of non-mergers relative to their controls. A lack of statistical offset may be expected for the non-merger sample, as it is a subset of the control population. However, since the non-merger sample is not randomly selected from the control pool, and is selected to match the population characteristics of the post-merger sample, we have therefore demonstrated that the population characteristics of the non-merger sample do not give rise to enhanced SMBH accretion rates relative to their controls. Nonetheless, the non-merger distribution provides a useful reference. For example, both populations show a similar distribution of $\Delta \dot M_{BH}$, ranging from 100 times enhanced to 100 times suppressed accretion rates relative to controls, with 30 non-mergers (2\%) achieving accretion rates at least 10 times higher than controls. The comparable spread demonstrates the large variability of accretion rate in both samples. In contrast to the non-mergers, the post-merger sample peaks at a positive accretion rate enhancement, with a median enhancement of 0.23 dex, corresponding to an enhancement of $\sim 70 \%$. In the post-merger sample, 102 post-mergers (6.5\%) have accretion rates at least 10 times higher than controls. Figure \ref{fig:DMdot_hist_all} therefore demonstrates our first main result, that mergers in TNG, on average, have an elevated accretion rate in the post-merger phase.
\begin{figure}
\includegraphics[width=\columnwidth]{DMdot_hist_all_legend_axis.pdf}
\caption{Histogram of accretion rate enhancements for the post-merger population and non-merger population. Post-mergers are represented with the teal line, and non-mergers are represented with the dashed yellow line. The median enhancement of the post-merger sample is 0.23 dex.}
\label{fig:DMdot_hist_all}
\end{figure}
Figure \ref{fig:DMdot_vs_z} shows the SMBH accretion rate enhancement of the post-merger and non-merger samples as a function of redshift. The full distribution of the post-merger sample is shown in the background density plot, and the foreground points are the median accretion enhancement for the post-merger sample (teal circles) and non-merger sample (yellow squares) within equally spaced bins of redshift. The errorbars represent the standard error on the median within each bin. The median accretion rate enhancement of the post-merger sample is consistently above 0 across the redshift range. We present both the median enhancement points and the background distribution to emphasize that while there is a positive median $\Delta \dot M_{BH}$ in post-mergers, not all post-mergers show accretion rate enhancements (as expected from Figure \ref{fig:DMdot_hist_all}). Overall, neither sample (post-mergers nor non-mergers) show any significant dependence on redshift, indicating that the merger process elevates accretion rates out to at least $z=1$ (the limit of our sample selection).
\begin{figure}
\includegraphics[width=\columnwidth]{DMdot_vs_z.pdf}
\caption{Accretion rate enhancements of post-merger galaxies for different redshifts. The background density plot shows the distribution of $\Delta \dot M_{BH}$ for the post-merger sample, and the foreground points represent the median value at that redshift. The teal circles are the median post-merger $\Delta \dot M_{BH}$ and the yellow squares are the median non-merger $\Delta \dot M_{BH}$. The error on the x-axis represents the bin width for each data point, and the error on the y-axis is the standard error on the median for that redshift bin.}
\label{fig:DMdot_vs_z}
\end{figure}
We find a similarly consistent result when investigating the dependence of SMBH accretion rate enhancements on stellar mass, shown in Figure \ref{fig:DMdot_vs_Mstar}. As expected from its construction, we find that the non-merger sample has accretion rate enhancements consistent with zero at all stellar masses. In contrast the teal points demonstrate that the SMBH accretion rates in post-mergers are consistently enhanced, on average, at all stellar masses within our sample.
\begin{figure}
\includegraphics[width=\columnwidth]{DMdot_vs_Mstar.pdf}
\caption{Accretion rate enhancements of post-merger galaxies for different stellar masses. The background density plot shows the distribution of $\Delta \dot M_{BH}$ for the post-merger sample, and the foreground points represent the median value at that stellar mass. The teal circles are the median post-merger $\Delta \dot M_{BH}$ and the yellow squares are the median non-merger $\Delta \dot M_{BH}$. The error on the x-axis represents the bin width for each data point, and the error on the y-axis is the standard error on the median for that stellar mass bin.}
\label{fig:DMdot_vs_Mstar}
\end{figure}
Figure \ref{fig:DMdot_vs_Mgas} shows the accretion rate enhancements $\Delta \dot M_{BH}$ as a function of the gas mass (top panel) and gas fraction (bottom panel) of the post-merger or non-merger galaxy, where the gas mass is the sum of all gas particles within twice the stellar half-mass radius and the gas fraction is the ratio of the gas mass to the baryon mass (gas + stars). In the top panel, we see that for a gas mass less than $\sim 10^9 \mathrm{M_{\odot }}$, the median enhancements of the post-merger sample are not significantly distinct from the non-merger sample, with the exception of one point at $\sim 10^{8.3-8.4} \mathrm{M_{\odot}}$. However in the regime of $> 10^{9} \mathrm{M_{\odot}}$, post-mergers consistently have on average enhanced accretion rates. Therefore, Figure \ref{fig:DMdot_vs_Mgas} suggests that post-mergers of a lower gas mass are less likely to have enhanced accretion rates. In addition, at gas masses above $10^9 \mathrm{M_{\odot}}$, post-merger galaxies consistently show accretion rate enhancements, where non-merger galaxies of this mass range do not, suggesting that the presence of a large amount of gas does not guarantee higher than average accretion rates.
We find a qualitatively consistent result looking at the gas fraction (bottom panel of Figure \ref{fig:DMdot_vs_Mgas}). We see that for post-mergers with a lower gas fraction, there is a trend towards low to no relative accretion enhancement. Once again, there is some evidence for an exception to this trend at the lowest gas fraction, however we caution that the poor statistics at lower gas mass and gas fraction make the median data point more susceptible to the variability on a galaxy by galaxy basis. Therefore, we find that the presence of a significant amount of gas is an essential, but insufficient criterion to produce accretion rate enhancements, and that the merger itself is important for driving gas into the region surrounding the SMBH whence it can accrete.
\begin{figure}
\includegraphics[width=\columnwidth]{DMdot_vs_gasMass_gasFrac.pdf}
\caption{Accretion rate enhancements of post-merger galaxies for different $M_{gas }$ (shown in top panel) and gas fractions (shown in bottom panel). The gas mass is the sum of all gas particles within twice the stellar half-mass radius of the galaxy. The gas fraction is the ratio of the gas mass to the baryon mass (gas mass + stellar mass). The background density plot shows the distribution of $\Delta \dot M_{BH}$ for the post-merger sample, and the foreground points represent the median value at that $M_{gas}$ or gas fraction. The teal circles are the median post-merger $\Delta \dot M_{BH}$ and the yellow squares are the median non-merger $\Delta \dot M_{BH}$. The error on the x-axis represents the bin width for each data point, and the error on the y-axis is the standard error on the median for that $M_{gas}$ or gas fraction bin.}
\label{fig:DMdot_vs_Mgas}
\end{figure}
Finally, Figure \ref{fig:DMdot_vs_MassRatio} shows $\Delta \dot M_{BH}$ for post-merger galaxies of different merger mass ratios. We note that in this plot, we only include galaxies with a merger mass ratio error of $\sigma_{\mu} \leq 0.1$, which excludes 99 galaxies from the sample of 1563 post-mergers. By visual inspection, we find no significant relationship between the strength of the accretion rate enhancement and the merger mass ratio. In fact, we see that the majority of enhancements are consistent, within error, with the overall sample accretion rate enhancement of 0.23 dex. We further confirm the lack of correlation with a statistical Pearson correlation test, which yields a correlation coefficient of $\sim$0.1.
The lack of correlation between $\Delta \dot M_{BH}$ and mass ratio is somewhat contradictory to previous simulation results. For example, \cite{Capelo2015} find mass ratio to be the most important factor influencing SMBH accretion rate in a suite of high resolution binary merger simulations. However, there are notable differences between the experiments of \cite{Capelo2015} and the work presented here. Specifically, \cite{Capelo2015} look at the effect of varying the merger mass ratio while keeping the orbital geometry and gas fraction constant, whereas our result looks at a population averaged enhancement and is therefore subject to the variable conditions of every merger. In addition, \cite{Capelo2015} comment on the role of resolution in their result, and that some of the torques generated in the galaxy interaction require high resolution simulations in order to be resolved. Finally, our result should not be interpreted as mass ratio having zero role in regulating gas inflows. Instead, Figure \ref{fig:DMdot_vs_MassRatio} demonstrates that, once the full demographic of merger properties is sampled, mass ratio is not a dominant factor in predicting the strength of a SMBH accretion rate enhancement.
\begin{figure}
\includegraphics[width=\columnwidth]{DMdot_vs_MassRatio.pdf}
\caption{Accretion rate enhancements of post-merger galaxies for different merger mass ratio. The background density plot shows the distribution of $\Delta \dot M_{BH}$ for the post-merger sample, and the foreground points represent the median value at that mass ratio. The error on the x-axis represents the bin width for each data point, and the error on the y-axis is the standard error on the median for that mass ratio bin.}
\label{fig:DMdot_vs_MassRatio}
\end{figure}
\subsection{Timescale of accretion rate enhancements}
\label{subsec:timescale}
We have thus far demonstrated that post-merger galaxies show, on average, enhanced accretion rates when compared to matched controls. This result is in agreement with previous theoretical studies that demonstrate merger driven gas inflows \citep[][]{DiMatteo2005,Springel2005,Hopkins2008}, which would increase the local gas density surrounding the black hole. However, the timescales of enhanced SMBH accretion are crucial to understanding the influence of the merger beyond the immediate post-merger phase, and the role of merger induced accretion rate enhancements on a galaxy's evolution. In addition, studying the average temporal behaviour of SMBH accretion rate enhancements can help alleviate the challenges associated with studying the highly stochastic individual SMBH accretion rates.
In the following section, we re-perform an experiment from a previous paper in this series. Following the procedure of \cite{Hani2020} (who investigated enhancements in SFR), we calculate a $\Delta \dot M_{BH}$ for post-merger galaxies in the snapshots following coalescence. Specifically, for each post-merger in our sample, we follow the descendants for as long as the merger tree allows or until the next merger event (with a mass ratio greater than 1:10). We control match the post-merger descendants to identify the enhancement to SMBH accretion rate as a function of the time since the most recent merger. We re-select the control galaxies at each subsequent snapshot following the same matching criteria as Section \ref{subsec:matching}. Therefore, a galaxy at snapshot N will not be matched to the same controls in the subsequent snapshot N+1 if the controls fall out of the acceptable matching criteria and tolerances specified in Section \ref{subsec:matching}.
Figure \ref{fig:DMdot_DSFR_TPM} shows the SMBH accretion rate enhancement as a function of time post-merger. The median SMBH accretion rate enhancement is shown in the teal line, where the shaded region represents the 25th and 75th percentiles. We find that, on average, SMBH accretion rate enhancements persist for up to two Gyr post-merger. A long-lived accretion rate enhancement may seem to be in contradiction with the short lived (tens-hundreds of Myrs) high accretion rate events observed in previous simulation studies \citep[][]{DiMatteo2005,Springel2005,Hopkins2008}. However we emphasise our relative enhancement variable $\Delta \dot M_{BH}$ is distinct from an enhancement in absolute accretion rate in a given galaxy. Our result demonstrates that post-mergers sustain a population-averaged enhanced accretion rate relative to matched controls, which does not necessarily correspond to a long lived high accretion rate event in an individual galaxy. A long-lived population average suggests that the dynamical disturbance that leads to higher than normal feeding of the SMBH persists for up to $\sim 2$ Gyrs in TNG100-1.
We are also interested in investigating the temporal correlation, or lack there-of, between star formation rate enhancements and SMBH accretion rate enhancements. Studying the population averaged enhancements can smooth out the large temporal variability of SMBH accretion rates, and allow us to look at the timescales over which connections between star formation and SMBH accretion rate are present in the overall post-merger population. Therefore, we also compute the star formation rate enhancements of our post-merger sample,
\begin{equation}
\Delta SFR = log_{10}(SFR) - median[log_{10}(SFR_{Controls})],
\label{eq:enhancement2}
\end{equation}
where the star formation rate is the sum of star formation rates for all cells within twice the stellar half-mass radius. When calculating $\Delta SFR$, we introduce an additional matching criterion from \cite{Hani2020}. We require that galaxies be matched within a classification of star forming or passive in order to avoid (spuriously) large/small values of $\Delta SFR$. Passive galaxies are defined as galaxies which lie more than 2$\sigma$ below the star forming main sequence, where we calculate the star forming main sequence by applying a linear fit to TNG100-1 galaxies with stellar mass between $10^{9-10.2}M_{\odot}$ and between redshift 0 to 1, and extrapolate the linear fit to higher stellar masses following the procedure of \citep[][]{Donnari2019}. In this way, post-merger star-forming galaxies are matched to star-forming controls, in addition to the fiducial matching criteria outlined in Section \ref{subsec:matching}. Likewise, passive post-mergers are matched to passive controls. We therefore identify a different set of controls when calculating $\Delta \dot M_{BH}$ and $\Delta SFR$ for the experiments presented in Section \ref{subsec:timescale}.
Figure \ref{fig:DMdot_DSFR_TPM} shows the star formation rate enhancement in the purple line, once again the shaded region represents the 25th and 75th percentiles. We qualitatively recover the result of \cite{Hani2020}, who find that, on average, post-merger galaxies demonstrated star formation rate enhancements\footnote{We note that the magnitude of the star formation rate enhancement in post-merger galaxies is lower in the work presented compared with the main result of \cite{Hani2020}. There exist a number of subtle matching scheme differences, specifically the inclusion of gas mass matching (which is demonstrated in \cite{Hani2020} to reduce the magnitude of the SFR enhancement).}, and that the enhancements persisted for up to $\sim 500$ Myrs after the merger, consistent with previous simulation studies \citep[][]{DiMatteo2008} and observational estimates \citep[][]{Wild2010}. We therefore demonstrate that, on average, SMBH accretion rate enhancements are significantly longer lived than star formation rate enhancements within our post-merger sample. We compare the result of Figure \ref{fig:DMdot_DSFR_TPM} with \cite{Volonteri20151}, who investigate the temporal correlation between SFR and SMBH accretion rate in 10 high resolution binary merger simulations. \cite{Volonteri20151} find that, on an individual merger basis, properties are only temporally correlated within the 200-300 Myrs of the final coalesence event of the merger, a timescale that is consistent with our population averaged result. Our result that enhancements of SMBH accretion rates can be sustained up to 2 Gyrs is also consistent with the result of \cite{Volonteri20152}, who find that the SMBH accretion rates of merger remnants can remain sufficiently high such that the luminosity of the AGN is dominant over the stellar luminosity up to 1.5 Gyrs after the merger.
As a possible explanation for the difference in the timescale of SFR and accretion rate enhancements, we comment that in TNG, both the star formation and SMBH accretion depend on local gas density, where stars form in gas following the empirical Kennicutt-Schmidt relation \citep[][]{Schmidt1959,Kennicutt1998} and the gas dependence of SMBH accretion is shown in Eq. \ref{eq:Bondi}. However, star formation in TNG additionally requires a threshold density of $n_{H} \gtrsim 0.1 cm^{-3}$ \citep[][]{Pillepich2018}. We suggest a possible scenario where a merger event increases the central gas density significantly within $\sim 500$ Myrs of coalesence, resulting in enhanced star formation and SMBH accretion rates. However, the increase of material to the central gas reservoir is insufficient to sustain star formation past 500 Myrs yet sufficient to sustain a long lived low accretion rate enhancements.
We considered whether the above average accretion rates in the post-merger sample, which persist $\sim$ two Gyrs after the merger event, may be a feature of the matching methodology, as we define controls as galaxies that have not undergone a merger of mass ratio $>1:10$ within the last two Gyrs. We test whether the timescale of averaged accretion rate enhancements is sensitive to the minimum elapsed time post-merger that we allow for controls. We regenerate our control sample, now requiring that control galaxies to have had at least three Gyrs since their most recent 1:10 merger, and find that the timescale of two Gyrs is robust.
\begin{figure}
\includegraphics[width=\columnwidth]{DMdot_DSFR_vs_TPM_line_2Gyr_legend.pdf}
\caption{Running median of the accretion rate enhancement (teal) and star formation rate enhancement (purple) over the time post-merger. The error of the shaded region represents the 25th and 75th percentiles associated with the median at that time post-merger.}
\label{fig:DMdot_DSFR_TPM}
\end{figure}
Figure \ref{fig:DMdot_DSFR_TPM} demonstrates the averaged behaviour of $\Delta \dot M_{BH}$ and $\Delta SFR$ in the total post-merger sample. However, not all galaxies show positive SMBH accretion rate or SFR enhancements (e.g. Figure \ref{fig:DMdot_hist_all}). In particular, the results of Section \ref{subsec:Enhancements} suggest that gas poorer galaxies do not exhibit, on average, accretion rate enhancements. We therefore separate the post-merger population based on the instantaneous feedback mode, which separates the post-mergers with high absolute accretion rates (radiative mode) from low absolute accretion rates (kinetic mode). The distinction in feedback mode also broadly separates lower mass, gas rich galaxies (radiative mode) from higher mass, gas poor galaxies (kinetic mode).
Figure \ref{fig:DMdot_DSFR_TPM_split} shows SMBH accretion rate enhancements and SFR enhancements, separated into the radiative mode, shown in blue, and kinetic mode, shown in red. In the top panel, we see that both populations show, on average, positive accretion rate enhancements within two Gyrs of coalescence. The radiative mode feedback population has a higher peak accretion rate enhancement of $\sim$ 0.3 dex compared with a relative enhancement of $\sim$0.2 dex in the kinetic mode feedback population. The lower median accretion rate enhancements in kinetic mode feedback galaxies is consistent with Section \ref{subsec:Enhancements}, demonstrating that accretion rate enhancements are lower in galaxies with lower gas mass or gas fraction.
The bottom panel of Figure \ref{fig:DMdot_DSFR_TPM_split} shows the star formation rate enhancement separated by feedback mode, where the solid lines show the star formation rate enhancement calculated within twice the stellar half-mass radius and dashed lines within one stellar half-mass radius. We include a second radius in order to compare the global and central star formation rate enhancements. Beginning with the solid lines in Figure \ref{fig:DMdot_DSFR_TPM_split}, we see that the two feedback modes show different behaviour in the first 500 Myrs post-merger. Radiative mode feedback galaxies show star formation rate enhancements for $\sim$ 500 Myrs after the merger. We therefore find that galaxies in radiative mode feedback, on average, have both accretion rate and star formation rate enhancements within the first few hundred Myrs. In contrast, kinetic mode feedback galaxies show a relative star formation \textit{suppression} for 500 Myrs after the merger. Therefore, Figure \ref{fig:DMdot_DSFR_TPM_split} demonstrates that despite an increased supply of gas, resulting in enhanced SMBH accretion rates, kinetic mode post-mergers display lower than average rates of star formation within $\sim $ 500 Myrs of coalescence. Our result that, on average, kinetic mode feedback galaxies demonstrate both relatively enhanced SMBH accretion rates and relatively suppressed of star formation rates may be explained considering the relationship between star formation quenching and kinetic mode feedback in TNG \citep[][]{Davies2020,Luo2020,Terrazas2020,Nelson2021,Piotrowska2021}, which we will discuss in further detail in Section \ref{subsec:feedback}.
Another possible explanation for the star formation rate suppression as well as the difference in timescales between $\Delta \dot M_{BH}$ and $\Delta SFR$ is that the enhancement in accretion rate is measured in the accretion region for the simulation, i.e. within the cells in the immediate vicinity of the black hole particle, whereas the star formation rate is measured over a much larger spatial scale, within twice the stellar half-mass radius. We therefore investigate whether central SFR enhancements are present in the PM sample. In the bottom panel of Figure \ref{fig:DMdot_DSFR_TPM_split}, the dashed lines represent the star formation rate enhancement calculated within one stellar half-mass radius, $\Delta SFR$\textsubscript{half}, once again split by feedback mode into the blue and red lines. Comparing the solid lines, $\Delta SFR$, with the dashed lines, $\Delta SFR$\textsubscript{half}, we can see how a more centralized aperture, although still much larger in spatial extent than the BH accretion region, affects the timescales of the star formation rate enhancements. For the radiative feedback mode galaxies, we see that the SFR enhancement peak is higher, reaching almost 0.4 dex. We also see that the enhancement is slightly longer lived. In the kinetic mode galaxies, we see that the star formation rate is slightly enhanced, 0.1 dex, within the $\sim$500 Myr window, demonstrating relative star formation rate enhancements despite suppressed star formation rates on a larger spatial scale.
\begin{figure}
\includegraphics[width=\columnwidth]{DMdot_DSFR_vs_TPM_sep_line_2Gyr_legend.pdf}
\caption{Running median of the accretion rate enhancement (top panel) and the star formation rate enhancement (bottom panel) over the time post-merger. The dark blue line represents the population in radiative mode feedback, and the orange line represents the population in kinetic mode feedback. The solid lines show the star formation rate measured within twice the stellar half-mass radius. The dashed lines represent the star formation rate enhancement calculated in a smaller aperture of one stellar half-mass radius. The error on the y-axis represents the 25th and 75th percentiles associated with the median $\Delta \dot M_{BH}$ and $\Delta SFR$ at that time post-merger.}
\label{fig:DMdot_DSFR_TPM_split}
\end{figure}
\subsection{Correlation of star formation rate and accretion rate enhancements}
\label{subsec:correlations}
In the previous section, Figure \ref{fig:DMdot_DSFR_TPM} demonstrated that within $\sim $ 500 Myrs of the merger, on average, galaxies have both star formation rate enhancements and accretion rate enhancements. However, Figure \ref{fig:DMdot_DSFR_TPM} does not assess whether the two enhancements are temporally correlated in a given galaxy. Although the merger process triggers gas inflows \citep[][]{Hernquist1989a,Barnes1991,Mihos1996,DiMatteo2007,Capelo2016,Blumenthal2018} that might naturally lead to quasi-simultaneous enhancements in both nuclear star formation and BH accretion \citep[][]{Sanders1988,DiMatteo2005,Springel2005,Hopkins2008}, observations show little evidence for a starburst-AGN connection outside of the most luminous systems \citep[][]{RowanRobinson1995,Schweitzer2006,Lutz2010,Shao2010,Santini2012,Rosario2015}. High resolution binary merger simulations have also demonstrated the lack of temporal correlations for the majority of the duration of the merger, with \cite[][]{Volonteri20151} finding a correlation only during the 200-300 Myrs over which the galaxies coalesce. The challenge is the different timescales of the two processes; BH accretion is stochastic on very rapid timescales, whereas star formation is more sustained. Observations that take a snapshot of a single point in time can not capture any potential extended connection between these processes. Simulation snapshots suffer from the same effect. Nonetheless, a merger event pinpoints the time of major gas inflow and may therefore be expected to demonstrate a connection between star formation and accretion.
In order to test for a correlation between the SFR enhancement and SMBH accretion rate enhancements in post-mergr galaxies, and determine whether selecting the post-mergers within different windows of time post-merger may affect a measured correlation, we consider the relationship between $\Delta \dot M_{BH}$ and $\Delta SFR$ on a galaxy by galaxy basis. We note that for the remainder of this section, where we compare $\Delta \dot M_{BH}$ and $\Delta SFR$ for each individual galaxy, we use a single matching criteria to calculate both $\Delta \dot M_{BH}$ and $\Delta SFR$, where we include the star forming vs passive classification discussed in Section \ref{subsec:timescale} to our fiducial matching scheme outline in Section \ref{subsec:matching}. That is, for each post-merger we identify a set of suitable controls and calculate $\Delta \dot M_{BH}$ and $\Delta SFR$ relative to the same set of controls.
Figure \ref{fig:DMdot_vs_DSFR} shows the distribution of $\Delta \dot M_{BH}$ vs $\Delta SFR$ on a galaxy by galaxy basis, divided into three samples based on time post-merger. The percentages of galaxies occupying each quadrant are shown, where a high occupation fraction in the top right quadrant would be indicative of an excess of simultaneous enhancements. We find that within 200 Myrs of the merger, 42\% of post-mergers have both accretion rate and SFR enhancements. However, the majority of post-mergers occupy each of the other quadrants, demonstrating that the star formation and accretion rate processes are not generally synchronized in post-mergers. We perform a Pearson correlation test for each window of time post-merger, with the correlation coefficient quoted in the textbox of each panel. The correlation coefficient is strongest within 200 Myrs of the merger and decreases with each subsequent time bin. However, even in the shortest time-since-merger interval the correlation between SFR and SMBH accretion rate enhancements is modest (correlation coefficient = 0.29). Our result therefore demonstrates that the synchronicity or correlation between SFR and SMBH accretion rate enhancements is affected by the timescale on which the galaxy is observed. We also demonstrate that a significant correlation between $\Delta \dot M_{BH}$ and $\Delta SFR$ is only present in the first few hundred Myrs post-merger, consistent with the timescale of temporally correlated SFR and SMBH accretion rates in \cite{Volonteri20151}. Finally, that even within 200 Myrs of coalescence, the majority of post-mergers do not have synchronized $\Delta SFR$ and $\Delta \dot M_{BH}$.
\begin{figure*}
\includegraphics[width=2\columnwidth]{DMdot_vs_DSFR_TPM.pdf}
\caption{Accretion rate enhancements vs star formation rate enhancements, where each panel represents bins of time post-merger (TPM in Gyrs), since the most recent 1:10 mass ratio merger. The percentages represent the number of galaxies within each quadrant relative to the total number of galaxies in the panel. The Pearson correlation coefficient is shown in the light blue text box.}
\label{fig:DMdot_vs_DSFR}
\end{figure*}
Returning to the left-most panel of Figure \ref{fig:DMdot_vs_DSFR}, we note that despite a bias for co-incident positive $\Delta \dot M_{BH}$ and $\Delta SFR$, the majority of post-mergers do not demonstrate synchronicity in SFR and SMBH accretion rate enhancements. We explore the diversity of the post-mergers within 200 Myrs of the merger in Figure \ref{fig:DMdot_vs_DSFR_3x3}, which shows the distribution of the post-mergers in $\Delta \dot M_{BH}$ and $\Delta SFR$ space, separated into bins of mass ratio and gas mass. The panels are organized such that gas mass is increasing from top to bottom and mass ratio is increasing from left to right.
The post-mergers in the bottom row of panels of Figure \ref{fig:DMdot_vs_DSFR_3x3}, corresponding to a gas mass $10^{9.75-11} M_{\odot}$, demonstrate that the majority of gas-rich post-mergers have both positive $\Delta \dot M_{BH}$ and $\Delta SFR$ (i.e. over 50\% are in the top right quadrant). In fact, comparing the bottom left panel (gas-rich `minor' mergers) to the bottom right panel (gas-rich major mergers), we see that gas-rich major mergers are more likely to have synchronicity (51\% compared with 76\%). We also find that overall, gas rich major mergers are more likely to produce SMBH accretion rate enhancements (68\% positive $\Delta \dot M_{BH}$ compared with 89\%). Our result suggests that gas rich major mergers more reliably produce accretion rate enhancements compared with minor mergers, though not necessarily stronger enhancements (result of Figure \ref{fig:DMdot_vs_MassRatio}).
The panels of the top row of Figure \ref{fig:DMdot_vs_DSFR_3x3}, corresponding to gas mass $10^{7-9.75} M_{\odot}$, show a slight bias for galaxies to occupy the left-side quadrants, corresponding to relatively lower star formation rates. A star formation rate suppression in lower gas mass post-mergers is consistent with the results of Sections \ref{subsec:Enhancements} and \ref{subsec:timescale}. We also emphasize that we are considering the gas mass of the galaxy post-coalescence, an important distinction if comparing our results with previous work such as \cite{DiMatteo2007} and \cite{Scudder2015}. Specifically, \cite{DiMatteo2007} do not find a strong dependence of star formation rate enhancement on the initial gas mass (i.e. the amount of gas available just before coalescence), while \cite{Scudder2015} find that galaxies with the lowest initial gas fraction have the highest SFR enhancements due to interactions. Our results demonstrate that star formation rate enhancements are less likely in post-mergers with a low post-coalescence gas mass, but do not comment on the relationship with initial gas mass. Our dependence of $\Delta SFR$ on gas mass may be explained, in part, by the correlation between SFR and gas fraction (as was demonstrated in \citealt[][]{Scudder2015} and \citealt[][]{Hani2020}). However, we note that \cite{Hani2020} still find a dependence of $\Delta SFR$ on gas fraction even when explicitly control matching on gas mass. If we focus on the top right panel of Figure \ref{fig:DMdot_vs_DSFR_3x3}, we also find that the bias for negative $\Delta SFR$ is present even in the highest mass ratio mergers. Therefore, our results suggest that even major mergers are unlikely to produce star formation rate enhancements if they have a low gas mass.
Overall, Figure \ref{fig:DMdot_vs_DSFR_3x3} demonstrates that the strongest bias for synchronicity and the strongest correlation (coefficient=0.47) occurs in gas-rich major mergers. We note that for the results presented in Figure \ref{fig:DMdot_vs_DSFR_3x3}, we find the same qualitative and similar quantitative results when separating the post-mergers by gas fraction rather than gas mass.
\begin{figure*}
\includegraphics[width=1.5\columnwidth]{DMdot_vs_DSFR_MR_2x2.pdf}
\caption{Accretion rate enhancement vs star formation rate enhancement for galaxies within 200 Myrs of a 1:10 mass ratio merger. The plots are organized by increasing gas mass, in units $\mathrm{log}_{10}\mathrm{M_{\odot}}$, from bottom to top and increasing mass ratio from left to right. The percentages represent the number of galaxies within each quadrant relative to the total number of galaxies in each panel. The Pearson correlation coefficient is shown in the light blue text box for each quadrant}
\label{fig:DMdot_vs_DSFR_3x3}
\end{figure*}
\section{Discussion}
\label{sec:Discussion}
\subsection{Effect of Resolution and Numerical Considerations}
\label{subsec:ResolutionAndEtc}
In the work presented here, we use the intermediate volume and resolution run of TNG, TNG100-1. We can repeat our experiment using the large volume and low resolution run, TNG300-1, which has a $(302.6 \, \mathrm{Mpc})^3$ volume, a baryonic resolution of $1.1\times 10^7 \mathrm{M_{\odot}}$, and a dark matter resolution of $5.9 \times 10^7 \mathrm{M_{\odot}}$ \citep[][]{Weinberger2017,Pillepich2018}. We obtain a sample of 25576 successfully matched post-merger galaxies from TNG300-1. Figure \ref{fig:DMdot_hist_res} shows the accretion rate enhancement parameter for the post-merger sample from both resolutions, TNG100 shown in teal and TNG300 shown in the dashed line. The median $\Delta \dot M_{BH}$ of the TNG300 post-merger population is $\sim$0.27 dex, corresponding to an accretion rate enhancement of roughly 90\%, slightly higher but consistent with the median enhancement for TNG100. Therefore, we demonstrate the comparability of our results between the two resolutions and volumes.
\begin{figure}
\includegraphics[width=\columnwidth]{DMdot_hist_res_compare.pdf}
\caption{Histogram of accretion rate enhancements for the post-merger populations from the intermediate resolution, TNG100-1, and low resolution, TNG300, simulations. TNG100 is represented with the teal line, and TNG300 is represented with the dashed line.}
\label{fig:DMdot_hist_res}
\end{figure}
A limitation of any work looking at SMBH accretion rates using cosmological scale simulations is the subgrid model required to calculate $\dot M_{BH}$. Subgrid models are an essential part of cosmological simulations, as they include important physical processes, such as stellar and SMBH feedback, that occur on scales below the resolution of the simulation. As mentioned, the TNG physics model uses a Bondi-Hoyle-Lyttleton accretion model \citep[][]{Weinberger2017,Pillepich2018}. However, the accretion model includes simplifying assumptions such as spherical symmetry, which would not hold for an accretion disk of material surrounding a SMBH. \cite{NegriVolonteri2017} demonstrated the variability of accretion rates calculated by different Bondi models and the dependance on simulation choices such as how the local density and sound speed are sampled. They find that Bondi models often overestimate the accretion rate, particularly when AGN feedback is inefficient at evacuating gas from the central region of the galaxy. However, they also find that in some simulations the accretion rate can be underestimated, which occurs when simulation cells of hot gas are over-represented in calculation of local parameters. Therefore, while we are unable to comment on the effects of mergers on the subparsec scales around the SMBH, the work presented here provides insight on how a merger event can impact the supply of material to the nuclear region, which may go on to truly accrete onto the SMBH.
An additional limitation of this work concerns the implementation of SMBH relocation in cosmological simulations, and the subsequent effect on SMBH accretion and SMBH mergers. \cite{Bahe2022} investigate the effects of SMBH repositioning in simulations using the EAGLE galaxy physics model, and demonstrate that the repositioning of SMBHs to the potential minima can result in significant boosts to SMBH accretion rate due to the increased density at the gravitational potential minima and the reduced relative velocity of the SMBH after repositioning. In addition, SMBH repositioning promotes early SMBH merging as it places the SMBHs of the merging galaxies in close proximity. A higher SMBH mass, due to ‘premature’ merging, will promote higher accretion rates due to the dependence on SMBH mass in the Bondi model. Therefore, we may expect the TNG model to overestimate accretion rates in the PM sample since SMBHs in merging galaxies do not wander and may merge prematurely. In addition, SMBH mergers may further complicate the timescale and delay between processes like starbursts and AGN activity. For example, \cite{Ni2022} demonstrate using the ASTRID cosmological simulations, which do not anchor SMBH particles to the gravitational minima and instead include a subgrid recipe to compensate for unresolved dynamical friction, that there is a delay of $\sim $200 Myrs between first close encounter of SMBH pairs and the SMBH merger. The above limitations highlight the imperative to investigate the galaxy merger and AGN connection in other cosmological simulations with varying subgrid implementations for SMBH physics.
\subsection{Implications for feedback and quenching}
\label{subsec:feedback}
In Section \ref{subsec:timescale}, we suggested that the enhancement of SMBH accretion rates in post-mergers with kinetic mode AGN feedback may have implications for galaxy quenching in TNG. Numerous studies have demonstrated that in TNG effective kinetic-mode AGN feedback is responsible for galaxy quenching \citep[][]{Davies2020,Luo2020,Terrazas2020,Nelson2021,Piotrowska2021}. Here, we briefly comment on how merger induced accretion enhancements may influence galaxy quenching in TNG. For simplicity, we can consider idealistic scenarios through which the enhanced accretion rates in post-merger galaxies may affect the quenching process.
First, increased accretion rates could result in post-merger galaxies that are in a high accretion state and are therefore using the radiative mode feedback model, which is ineffective at quenching in the TNG model. Therefore, mergers in TNG could inhibit or delay the quenching process, as galaxies may be in an inefficient feedback mode (in terms of quenching star formation). Since the median enhancement of accretion rates is subtle, $\sim1.7$ times higher than controls, a delayed quenching effect would be most applicable in galaxies that have accretion rates close to the transition between high and low accretion states, as defined in Section \ref{subsec:Illustris}.
Alternatively, increased SMBH accretion rates may enhance quenching in TNG. First, accretion rate enhancements in galaxies with low accretion rates could increase the energy input via effective kinetic mode feedback and contribute to the quenching of galaxies. In addition, enhancement in accretion rate can promote growth of the SMBH. According to the TNG model, the transition between radiative and kinetic mode feedback is strongly dependent on the SMBH mass, such that high mass SMBHs are more likely to be have kinetic mode feedback. Therefore, post-mergers may transition towards effective kinetic mode feedback sooner, promoting an excess of quenching.
\cite{Quai2021} explicitly investigate quenching in post-merger galaxies in TNG and conclude that quenching is rare in post-merger galaxies, but that there is an excess of quenched post-mergers when compared with matched controls. Specifically, \cite{Quai2021} find that quenching in post-mergers occurs only in galaxies that were already predisposed to quenching. The rarity of quenched post-mergers would be consistent with the first scenario outlined above. Post-mergers with strongly enhanced accretion rates will likely be in radiative feedback mode, where the SMBH feedback would not interact with the host galaxy in a way that promotes quenching, leading to a rarity of quenched post-merger galaxies. Furthermore, the merger process may introduce more gas into the nuclear region of the host galaxy, prolonging the lifetime of the high accretion state and preventing quenching.
The second scenario we have proposed is consistent with the slight excess of quenched post-mergers found in \cite{Quai2021}, where increased effective kinetic mode feedback speeds up quenching in the post-merger galaxies that are predisposed to quench. Furthermore, a truncation of star formation in the massive post-merger galaxies following a period of enhanced AGN activity would be consistent with the findings of \cite{Dubois2016} who observe a decrease of star formation in massive galaxies due to merger induced AGN activity. In addition, we discuss how our results relate to observational studies in Section \ref{subsec:observations}.
\subsection{What fraction of mergers are AGN?}
\label{subsec:fractionAGN}
So far we have demonstrated that post-merger galaxies have (on average) enhanced accretion rates. We have also demonstrated that not all post-mergers show enhancements in their SMBH accretion rate. Next we address the question of how frequently mergers actually trigger an AGN. Specifically, in this experiment, we will quantify what fraction of all post-mergers will have a high accretion event within 500 Myrs of coalescence.
To begin, we define an enhancement fraction, $f_{\mathrm{enhanced}}$, as the number of galaxies that have an SMBH accretion rate at or above a cutoff, $\dot M_{BH}^{\mathrm{cutoff}}$, divided by the total number of galaxies in the sample. We calculate $f_{\mathrm{enhanced}}$ for both the post-merger and non-merger sample, defined in Section \ref{subsec: PM and NM}, following the post-merger for 500 Myrs after the merger and the non-merger sample for 500 Myrs of secular evolution. The result of the experiment is shown in Figure \ref{fig:fEnhancedAcc}.
The top panel of Figure \ref{fig:fEnhancedAcc} shows $f_{\mathrm{enhanced}}$ as a function of $\dot M_{BH}^{\mathrm{cutoff}}$, where the post-merger sample is shown in the solid teal line and the non-merger sample is shown in the dashed yellow line. The top axis shows the associated bolometric luminosity for the SMBH accretion rate cutoff, calculated as 10\% of the accretion mass energy, or $0.1 \dot M_{BH}c^2$. We demonstrate that $\sim$ 60\% of post-merger galaxies have accretion rates exceeding $L_{bol} > 10^{43} \mathrm{erg/s}$ within 500 Myrs of the merger. However, we find that 50\% of non-merger galaxies also have an AGN phase of $L_{bol} > 10^{43} \mathrm{erg/s}$ within the same time period. Our result demonstrates that accretion rate events exceeding $L_{bol} > 10^{43} \mathrm{erg/s}$ are common in both samples, though slightly more common in the post-mergers, as expected.
In the bottom panel of Figure \ref{fig:fEnhancedAcc} we show the ratio of the post-merger to non-merger $f_{\mathrm{enhanced}}$ (the ratio of the blue solid line to the yellow dashed line from the top panel), or the fractional excess of AGN in the post-merger. We demonstrate that even though the fraction of galaxies that have an AGN phase decreases with increasing luminosity for both samples (as expected), the fractional excess of post-mergers with an AGN phase increases with luminosity. Therefore while less than 10\% of post-mergers will achieve accretion rates exceeding $L_{bol} > 10^{45} \mathrm{erg/s}$, four times more AGN appear in the post-merger sample than the non-merger sample.
Overall, we find that only a small fraction of mergers have high accretion rate events but that post-mergers are more likely to have an AGN event than a non-merger as a function of luminosity, consistent with results from the Magneticum Pathfinder simulation \citep[][]{Steinborn2018} and the EAGLE simulation \citep[][]{McAlpine2020}, as well as complimentary to \cite{Bhowmick2020} who find that high luminosity AGN are more likely to appear in environments with a higher density of SMBHs in TNG.
\begin{figure}
\includegraphics[width=\columnwidth]{fracEnhancedDotMbh_legend.pdf}
\caption{The top panel shows the fraction of galaxies that have a SMBH accretion rate at or higher than $\dot M_{BH}^{\mathrm{cutoff}}$ within a period of 500 Myrs, relative to the total number of galaxies. The teal line corresponds to the post-merger sample and the yellow dashed line is the non-merger sample, as defined in Section \ref{subsec: PM and NM}. The top x-axis corresponds to the AGN luminosity, calculated as $L_{\mathrm{bol}}=0.1\dot M_{BH}c^2$. The bottom panel shows the ratio of the fraction of post-merger galaxies to non-merger galaxies, or the fractional excess of AGN in the post-merger sample. The error in the shaded region is the Poisson error reflecting the number of galaxies.}
\label{fig:fEnhancedAcc}
\end{figure}
\subsection{What fraction of AGN are mergers?}
\label{subsec:fractionMergers}
A complementary question to the one posed in Section \ref{subsec:fractionAGN} is what fraction of galaxies with high accretion rates are mergers? In Section \ref{subsec:fractionAGN}, we demonstrated that the post-merger sample galaxies are more likely to be AGN than the non-merger sample of galaxies. However, post-mergers are rare in both TNG and the observed universe and thus may not be the major pathway to AGN triggering. Therefore, we measure the merger fraction of AGN in TNG100-1 in order to quantify the contribution of mergers to the total AGN population in the simulation.
To investigate the merger fraction of AGN, we first select an `AGN sample' by selecting all galaxies that have a SMBH accretion rate at or above $\dot M_{BH}^{\mathrm{cutoff}}$. In this experiment, the merger fraction is defined as the fraction of galaxies that have had a merger in the last 500 Myrs. Figure \ref{fig:fMerger} shows the merger fraction as a function of the cutoff accretion rate. The x-axis defines the AGN sample used to calculate the merger fraction, where the AGN sample will consist of all galaxies with accretion rates at or exceeding $\dot M_{BH}^{\mathrm{cutoff}}$ (and meeting the selection criteria outlined in Section \ref{subsec: ID mergers}, i.e $M_{\star} > 10^{10} \mathrm{M_{\odot}}$ and $z<1$). Once again, we show the equivalent bolometric luminosity along the top axis, calculated as $L_{bol} = 0.1 \dot M_{BH}c^2$. Figure \ref{fig:fMerger} shows that the merger fraction increases as a function of accretion rate (or, equivalently, luminosity). The horizontal dashed line represents the total merger fraction for the entire sample of galaxies, $f_{\mathrm{merger}}\sim$ 3\%. We see that the merger fraction of the AGN sample exceeds the total $f_{\mathrm{merger}}$ around $L_{bol}\sim 10^{44} \mathrm{erg/s}$. Therefore, there is no significant luminosity dependence of the merger fraction for $L_{bol}\lesssim 10^{44} \mathrm{erg/s}$. Beyond $L_{bol}\sim 10^{44} \mathrm{erg/s}$ the merger fraction increases with AGN luminosity, with a peak merger fraction of $\sim$13\%.
Although there is a luminosity dependence on the merger fraction for $L_{bol}\gtrsim 10^{44} \mathrm{erg/s}$, we find that recent mergers never dominate the AGN sample, even at the highest luminosities. This result is once again consistent with the results of \cite{Steinborn2018} and \cite{McAlpine2020} for the Magneticum Pathfinder and EAGLE simulations. We note that the merger fraction that we have defined does not include galaxies in the pre-merger phase, which may contribute to the remaining $\sim$90\% of galaxies at or above $10^{45}$ erg/s. There may also be minor mergers that are not accounted for in the merger fraction. However, the result of Figure \ref{fig:fMerger} also supports a scenario where secular processes play a significant role in fueling the highest luminosity AGN (at least in the redshift regime 0-1). This experiment concludes that post-mergers contribute more to the AGN population at the highest accretion rates, but they are non-dominant over the entire accretion rate range.
\begin{figure}
\includegraphics[width=\columnwidth]{AGNMergerFrac_legend.pdf}
\caption{The teal line shows the fraction of galaxies that have had a merger of mass ratio greater than 1:10 within 500 the past Myrs. The x-axis defines the sample used to calculate the merger fraction from left to right, where only galaxies with a black hole accretion rate of at least $\dot M_{BH}^{\mathrm{cutoff}}$ are included in the calculation of the merger fraction. The top x-axis shows the corresponding AGN luminosity, calculated as $L_{\mathrm{bol}}=0.1\dot M_{BH}c^2$. The error in the shaded region is the Poisson error reflecting the number of galaxies. The horizontal dashed line represents the overall merger fraction of the sample, $\sim$3\%, for all galaxies in TNG with stellar mass of at least $10^{10} M_{\odot}$ and a redshift $< 1$.}
\label{fig:fMerger}
\end{figure}
\subsection{Comparisons to observations}
\label{subsec:observations}
In the previous sections, we have demonstrated that there is a population averaged enhancement in SMBH accretion rate in post-mergers and that post-mergers are more likely to host highly accreting SMBHs. Our findings are in agreement with the observational studies that find that AGN are more likely to appear in a post-merger or interacting galaxy sample when compared with undisturbed controls \citep{Woods2007,Alonso2007,Koss2010,Ellison2011,RamosAlmeida2012,Ellison2013,Satyapal2014,Hong2015,Kocevski2015,Rosario2015,Weston2017,Hewlett2017,Goulding2018,Ellison2019,Gao2020,Marian2020,Pierce2022}. We find that at most, AGN occur four times more commonly in the post-mergers compared with the non-mergers. The maximum excess of AGN is quantitatively consistent with numerous observational studies finding an excess of approximately four or lower \citep[][]{Ellison2013,Goulding2018,Gao2020}, though some studies find a significantly larger excess of AGN in merging galaxies \citep[][]{Satyapal2014,Weston2017}. Such variations in the observed excess may be, in part, due to the different AGN selection techniques, which have been demonstrated to result in different AGN fractions \citep[][]{Secrest2020,Bickley2022Submitted}. Variations may also exist within the same AGN selection techniques, as demonstrated in \cite{Bickley2022Submitted} who find the AGN excess may also depend on the strength of visual disturbances in the PM sample..
In addition, we demonstrate that the majority of high luminosity AGN are not associated to recent mergers. Our results are in agreement with observations which find the most luminous AGN are not mergers \citep{Cisternas2011,Schawinski2011,Schawinski2012,Kocevski2012,Bohm2013,Villforth2014,Mechtley2016,Hewlett2017,Villforth2017,Marian2019,Lambrides2021}, but are in contrast to observational studies that find the majority of quasars have disturbed morphology \citep[][]{Treister2012,Glikman2015,Fan2016,Goulding2018,UrbanoMayorgas2019}. Therefore, our results support a scenario where mergers can, but do not always, trigger AGN, and where mergers do not play a major role is triggering even the most luminous AGN.
Despite an increased likelihood to host highly accreting SMBHs, we demonstrate that the majority of mergers do not host a highly accreting SMBH, suggesting the majority of mergers will not undergo a strong AGN feedback event. In fact, observations have demonstrated that mergers have normal (or even enhanced) gas fractions (e.g. \citealt{Ellison20152,Ellison2018,Violino2018,Pan2018}), consistent with a lack of strong instantaneous feedback. However, observations do find that AGN can have effects on the small (kpc or below) scale interstellar medium (e.g. \citealt[][]{Oosterloo2017,Izumi2020,Ellison2021,GarciaBurillo2021,RamosAlmeida2022,Saito2022}).
In Section \ref{subsec:correlations}, we demonstrated that the strength of the star formation rate and SMBH accretion rate enhancements, on a galaxy by galaxy basis, are not generally correlated, in agreement with simulations such as \cite{Hickox2014,Volonteri20151}. However, we do find a connection between star formation rate and SMBH accretion rate in post-mergers with both a high mass ratio and significant amount of gas. Such a connection may reflect the general preference for AGN to reside in gas rich and star forming galaxies, as demonstrated in simulations \citep[][]{Ward2022} and observations \citep[][]{Rosario20131,Bernhard2016,Jarvis2020,Ellison20192,Shangguan2020,Xie2021,Koss2021}. In addition, we demonstrate in Figure \ref{fig:DMdot_vs_DSFR_3x3} numerous individual cases where a post-merger may have a strongly enhanced SMBH accretion rate and simultaneously suppressed or normal star formation rate relative to controls. Our results are therefore consistent with observations which show that most AGN (not strictly post-mergers) have typical rates of star formation \citep[][]{Rosario20131,Rosario2015}, although some observations demonstrate higher than normal star formation rates in the highest luminosity AGN \citep[][]{Schweitzer2006,Lutz2010,Shao2010,Santini2012}.
Finally, in Section \ref{subsec:ResolutionAndEtc}, we comment the consistency between our results and the rarity of rapid quenching in TNG post-mergers found in \cite{Quai2021}. However, \cite{Ellison2022} demonstrate that in observations, there exists a significant excess of rapidly/recently quenched post starburst galaxies (PSBs) in galaxy mergers. There is still uncertainty about the quenching mechanisms within PSBs, as observations show they do not lack molecular gas (e.g. \citealt{Rowlands2015,French2015}). Instead, it might be that PSBs lack dense star-forming gas \citep[][]{French20181}, perhaps due to enhanced turbulence (e.g. \citealt[][]{Smercina2022}). Such mechanisms may not be captured within TNG due to the resolution constraints and effective equation of state treatment of the interstellar medium, warranting an investigation of AGN and rapid quenching in post-merger galaxies using higher resolution simulations or simulations with varied physical models.
\section{Conclusions}
\label{sec:conclusion}
In the work presented here, we study the effect of galaxy mergers on SMBH accretion rates in a collection of 1563 post-merger galaxies from the IllustrisTNG simulation. Our post-merger sample reflects a diverse collection of stellar masses, star formation rates, gas masses, environments, and mass ratios. For each post-merger galaxy in our sample, we identify control galaxies for comparison, and are able to isolate the effect of the merger on the instantaneous SMBH accretion rates. Our results are summarized in the following points:
\begin{itemize}
\item On average, post-merger black hole accretion rates are enhanced by a factor of 0.23 dex, corresponding to a $\sim$70\% increase in accretion rate (Figure \ref{fig:DMdot_hist_all}). We find that there is significant variation in the accretion rate enhancement on a galaxy by galaxy basis, where $\sim$ 30\% of post-mergers have accretion rates consistent with no enhancement or lower than controls.
\medskip
\item SMBH accretion rate enhancements persist for up to two Gyrs after coalescence, and are significantly longer lived than SFR enhancements, which only persist for $\sim$500 Myrs post-merger (Figure \ref{fig:DMdot_DSFR_TPM}).
\medskip
\item We find that the co-incidence of accretion rate and SFR enhancements is most pronounced within the first few hundred Myrs post-merger, and that the correlation strength of the enhancements decreases with time post-merger, shown in Figure \ref{fig:DMdot_vs_DSFR}. However, even within 200 Myrs of coalescence, the majority of post-mergers do not demonstrate synchronicity in SMBH acretion rate enhancements and SFR enhancements.
\medskip
\item In Figure \ref{fig:DMdot_vs_DSFR_3x3}, we find that gas rich major mergers demonstrate the strongest correlation coefficient between $\Delta SFR$ and $\Delta \dot M_{BH}$, and the majority have enhancements in both. However, we find that the presence of gas alone is an insufficient criteria for enhancing SMBH accretion rates, and that gas rich non-mergers, on average, do not have positive $\Delta \dot M_{BH}$ (Figure \ref{fig:DMdot_vs_Mgas}).
\medskip
\item Post-mergers are more likely to host high luminosity AGN than non-merger galaxies (Figure \ref{fig:fEnhancedAcc}), however high luminosity AGN are rare in the post-merger sample. For AGN luminosity in excess of $L_{\mathrm{bol}} > 10^{45} \mathrm{erg/s}$, corresponding to $\sim$10\% of post-mergers, there are four times more AGN in the post-merger sample than the matched non-merger sample.
\medskip
\item We find a luminosity dependence in the merger fraction of AGN for $L_{\mathrm{bol}} > 10^{44} \mathrm{erg/s}$ (Figure \ref{fig:fMerger}). However, we find that the majority of high luminosity AGN are not recent mergers, which at most only contribute $\sim$13\% to AGN with $L_{\mathrm{bol}} > 10^{45} \mathrm{erg/s}$.
\end{itemize}
\section*{Acknowledgements}
We thank the IllustrisTNG collaboration for making their data accessible. We thank the anonymous referee for their constructive comments. SBM acknowledges the receipt of a British Columbia Graduate Scholarship and the Dr. Margaret Perkins Hess Research Fellowship from the University of Victoria. This research was enabled in part by support provided by WestGrid (www.westgrid.ca) and Compute Canada (www.computecanada.ca). DRP gratefully acknowledges NSERC of Canada for a Discovery Grant which helped to fund this research.
\section*{Data Availability}
The data used in this work are publicly available at \hyperlink{}{https://www.tng-project.org}.
\bibliographystyle{mnras}
|
\section{Introduction}
In his $1940$ book \emph{The Theory of Group Characters and Matrix
Representations of Groups}, D.~E.~Littlewood devotes a section to the
evaluation of the Schur function $s_\lambda$ at a set of variables ``twisted''
(not his term) by a primitive $t$-th root of unity $\zeta$
\cite[\S7.3]{Littlewood40}.
In modern terminology,
Littlewood's theorem asserts that $s_\lambda$ evaluated at the
variables $\zeta^jx_i$ for $1\leq i\leq n$ and $0\leq j\leq t-1$ is
zero unless the $t$-core of $\lambda$ is empty.
Moreover, when it is nonzero, it factors as a product of Schur functions
indexed by the elements of the $t$-quotient of $\lambda$, each with the
variables $x_1^t,\dots,x_n^t$.
The Schur functions are characters of the irreducible
polynomial representations of
the general linear group $\mathrm{GL}(n,\mathbb{C})$.
Ayyer and Kumari \cite{AK22} have recently generalised
Littlewood's theorem to the characters of the other classical groups
$\mathrm{O}(2n,\mathbb{C})$, $\mathrm{Sp}(2n,\mathbb{C})$ and
$\mathrm{SO}(2n+1,\mathbb{C})$ indexed by partitions.
While their factorisations are still indexed by the $t$-quotient of the
corresponding partition, the vanishing is governed by the $t$-core having a
particular form.
More precisely, $t\textup{-core}(\lambda)$ is of the form $(a\mid a+z)$ in Frobenius
notation, where $z=-1,1,0$, for
$\mathrm{O}(2n,\mathbb{C})$, $\mathrm{Sp}(2n,\mathbb{C})$ and
$\mathrm{SO}(2n+1,\mathbb{C})$ respectively.
Note that these are the same partitions occurring in Littlewood's Schur
expansion of the Weyl denominators for types $\mathrm{B}_n$, $\mathrm{C}_n$
and $\mathrm{D}_n$ \cite[p.~238]{Littlewood40}
(see also \cite[p.~79]{Macdonald95}).
Littlewood's proof, and the proofs of Ayyer and Kumari, use the Weyl-type
expressions for the characters as ratios of alternants.
In the Schur case, Chen, Garsia and Remmel \cite{CGR84} and independently
Lascoux \cite[Theorem~5.8.2]{Lascoux?} gave an alternate
proof based on the Jacobi--Trudi formula \eqref{Eq_JT}.
Macdonald later applied the same technique to give an extension of
Littlewood's original theorem to skew Schur functions, again using the
Jacobi--Trudi formula \cite[p.~91]{Macdonald95}; see Theorem~\ref{Thm_skew}
below.
In this article we lift the results of Ayyer and Kumari to the
much more general universal characters of the groups
$\mathrm{O}(2n,\mathbb{C})$, $\mathrm{Sp}(2n,\mathbb{C})$ and
$\mathrm{SO}(2n+1,\mathbb{C})$ as defined by Koike and Terada \cite{KT87}.
These are symmetric functions indexed
by partitions which, under appropriate specialisation of the variables,
become actual characters of their respective groups.
In fact, these generalise the Jacobi--Trudi-type
formulas for the characters of these groups, which were first written down
by Weyl \cite[Theorems~7.8.E \& 7.9.A]{Weyl39}.
For the universal characters we generalise the notion of ``twisting''
a set of variables by introducing operators
$\varphi_t:\Lambda\longrightarrow\Lambda$ for each integer $t\geq 2$
which act on the complete homogeneous symmetric functions as
\begin{equation}\label{Eq_phi-def}
\varphi_t h_r=\begin{cases}h_{r/t} &\text{if $t$ divides $r$}, \\
0 &\text{otherwise}.\end{cases}
\end{equation}
It is not at all hard to show that the image of $\varphi_t$ when acting on
symmetric functions at the variables $x_1^t,\dots,x_n^t$ agrees with the result
of twisting the variables $x_1,\dots,x_n$ by $\zeta$.
The advantages of this framework for such factorisations are that the
proofs are much simpler, and the structure of the factorisations is made
transparent.
Moreover, we are able to discuss dualities between these objects
which are only present at the universal level.
A particularly important tool for our purposes is Koike's universal character
$\mathrm{rs}_{\lambda,\mu}$ \eqref{Eq_rs-def} associated with
a rational representation of $\mathrm{GL}(n,\mathbb{C})$.
This object, which is used later in Subsection~\ref{Sec_CK} to prove
other character factorisations,
appears to be the correct universal character analogue
of the Schur function with variables $(x_1,1/x_1,\dots,x_n,1/x_n)$.
The remainder of the paper reads as follows.
In the next section we outline the preliminaries on partitions and symmetric
functions needed to state our main results, which follow in
Section~\ref{Sec_results}.
In the following Section~\ref{Sec_lemmas} we prepare for the proofs of these
results by giving a series of lemmas regarding cores and quotients and their
associated signs.
The factorisations are then proved in Section~\ref{Sec_proofs},
including a detailed proof of the Schur case, following Macdonald.
The final Section~\ref{Sec_stuff} concerns other factorisation results
relating to Schur functions and other characters.
This includes universal extensions of factorisations very different
from those already discussed originally due to Ciucu and Krattenthaler, later
generalised by Ayyer and Behrend.
\section{Preliminaries}
\subsection{Partitions}
A \emph{partition} $\lambda=(\lambda_1,\lambda_2,\lambda_3,\dots)$ is a weakly
decreasing sequence of nonnegative integers such that only finitely many of the
$\lambda_i$ are nonzero.
The nonzero $\lambda_i$ are called \emph{parts} and the number of parts the
\emph{length}, written $l(\lambda)$.
We say $\lambda$ is \emph{a partition of $n$} if
$\abs{\lambda}:=\lambda_1+\lambda_2+\lambda_3+\cdots=n$.
Two partitions are regarded as the same if they agree up to trailing zeroes,
and the set of all partitions is written $\mathscr{P}$.
A partition is identified with its \emph{Young diagram},
which is the left-justified array of squares consisting of $\lambda_i$ squares in
row $i$ with $i$ increasing downward.
For example
\smallskip
\begin{center}
\begin{tikzpicture}[scale=0.4]
\foreach \i [count=\ii] in {6,4,3,2}
\foreach \j in {1,...,\i}{\draw (\j,1-\ii) rectangle (\j+1,-\ii);}
\end{tikzpicture}
\end{center}
is the Young diagram of $(6,4,3,2)$.
We define the \emph{conjugate} partition $\lambda'$ by reflecting the diagram of
$\lambda$ in the main diagonal $x=y$, so that the conjugate of $(6,4,3,2)$ above is
$(4,4,3,2,1,1)$.
If $\lambda=\lambda'$ then $\lambda$ is called \emph{self-conjugate}.
For a square at coordinate $(i,j)$ where $1\leq i\leq l(\lambda)$ and
$1\leq j\leq \lambda_i$ the \emph{hook length} is $h(i,j)=\lambda_i+\lambda_j'-i-j+1$.
For example the square $(1,2)$ in
\smallskip
\begin{center}
\begin{tikzpicture}[scale=0.4]
\filldraw[color=green!40] (2,0) rectangle (7,-1);
\filldraw[color=green!40] (2,-1) rectangle (3,-4);
\foreach \i [count=\ii] in {6,4,3,2}
\foreach \j in {1,...,\i}{\draw (\j,1-\ii) rectangle (\j+1,-\ii);}
\end{tikzpicture}
\end{center}
has hook length $8$, with its hook shaded.
A partition $\lambda$ \emph{is a $t$-core} if it contains no squares of
hook length $t$, the set of which is denoted $\mathscr{C}_t$.
For a pair of partitions $\lambda,\mu$ we write $\mu\subseteq\lambda$ if the diagram
of $\mu$ can be drawn inside the diagram of $\lambda$, i.e. if
$\mu_i\leq \lambda_i$ for all $i\geq 1$.
In this case we can form the \emph{skew shape} $\lambda/\mu$ by removing
the digram of $\mu$ from that of $\lambda$.
For example $(3,2,1,1)\subseteq(6,4,3,2)$ and
the diagram of $(6,4,3,2)/(3,2,1,1)$ is given by the non-shaded squares of
\smallskip
\begin{center}
\begin{tikzpicture}[scale=0.4]
\foreach \i [count=\ii] in {6,4,3,2}
\foreach \j in {1,...,\i}{\draw (\j,1-\ii) rectangle (\j+1,-\ii);}
\foreach \i [count=\ii] in {3,2,1,1}
\foreach \j in {1,...,\i}{\draw[fill=gray] (\j,1-\ii) rectangle (\j+1,-\ii);}
\end{tikzpicture}
\end{center}
A skew shape is called a \emph{ribbon} (or \emph{border strip}, \emph{rim hook},
\emph{skew hook}) if its diagram is connected
and contains no $2\times 2$ square.
A $t$-ribbon is a ribbon with $t$ boxes.
The \emph{height} of a $t$-ribbon $R$, written $\mathrm{ht}(R)$,
is one less than the number of rows it occupies.
In our example above $R=(6,4,3,2)/(3,2,1,1)$ is an $8$-ribbon with
height $\mathrm{ht}(R)=3$.
We say a skew shape is \emph{tileable by $t$-ribbons} or \emph{$t$-tileable} if
there exists a sequence of partitions
\begin{equation}\label{Eq_RD}
\mu=\nur{0}\subseteq\nur{1}\subseteq\cdots\subseteq\nur{k-1}\subseteq\nur{k}
=\lambda
\end{equation}
such that $\nur{i}/\nur{i-1}$ is a $t$-ribbon for $1\leq i\leq k$.
A sequence $D=(\nur{0},\dots,\nur{k})$ (not to be confused with the
$t$-quotient of $\nu$ below, for which we use the same notation)
satisfying \eqref{Eq_RD} is called a
\emph{ribbon decomposition} (or \emph{border strip decomposition}) of $\lambda/\mu$.
We define the height of a ribbon decomposition to be the sum of the heights
of the individual ribbons:
$\mathrm{ht}(D):=\sum_{i=1}^k\mathrm{ht}(\nur{i}/\nur{i-1})$.
As shown by van Leeuwen \cite[Proposition~3.3.1]{vanLeeuwen99} and
Pak \cite[Lemma~4.1]{Pak00} (also in \cite[\S6]{APRU21}),
the quantity $(-1)^{\mathrm{ht}(D)}$ is the same for every ribbon decomposition
of $\lambda/\mu$.
We therefore define the \emph{sign} of a $t$-tileable skew shape $\lambda/\mu$ as
\begin{equation}\label{Eq_sgn-def}
\sgn_t(\lambda/\mu):=(-1)^{\mathrm{ht}(D)}.
\end{equation}
Let $d(\lambda)$ be the greatest integer such that $d(\lambda)\geq \lambda_{d(\lambda)}$.
Equivalently, $d(\lambda)$ is the side length of the largest square which fits
inside the diagram of $\lambda$ (the \emph{Durfee square}).
A partition can alternatively be written in \emph{Frobenius notation} as
\[
\lambda=\big(\lambda_1-1,\dots,\lambda_{d(\lambda)}-d(\lambda)\mid
\lambda'_1-1,\dots,\lambda'_{d(\lambda)}-d(\lambda)\big).
\]
Any pair of integer sequences
$a_1>\cdots>a_k\geq0$ and $b_1>\cdots>b_k\geq0$ thus
determines a partition $\lambda=(a\mid b)$ with $d(\lambda)=k$.
For $z\in\mathbb{Z}$ and an integer sequence of predetermined length
$a=(a_1,\dots,a_k)$ we write $a+z:=(a_1+z,\dots,a_k+z)$.
Following Ayyer and Kumari \cite[Definition~2.9]{AK22},
$\lambda$ is called \emph{$z$-asymmetric} if $\lambda=(a\mid a+z)$
for some integer sequence $a$ and integer $z$.
Clearly a $0$-asymmetric partition is self-conjugate.
Partitions which are $-1$- and $1$-asymmetric are called
\emph{orthogonal} and \emph{symplectic} respectively.
\subsection{Cores and quotients}
We now describe the $t$-core and $t$-quotient of $\lambda$ arithmetically
following \cite[p.~12]{Macdonald95}.
There are many equivalent descriptions, see for instance
\cite{GKS90,HJ11,JK81,WW20}.
We begin with the \emph{beta set} of a partition,
which, is simply the set of $n$ integers
\[
\beta(\lambda;n):=\{\lambda_1+n-1,\lambda_2+n-2,\dots,\lambda_{n-1}+1,\lambda_n\},
\]
where $n\geq l(\lambda)$ is fixed.
The number of elements in this set congruent to $r$ modulo
$t$ is denoted by $m_r(\lambda;n)=m_r$.
Each element which falls into residue class $r$ for $0\leq r\leq t-1$ can
be written as $\xir{r}_kt+r$ for some integers
$\xir{r}_1>\cdots>\xir{r}_{m_r}\geq 0$.
These integers are used to define a partition with parts
$\lar{r}_k=\xir{r}_k-m_r(\lambda;n)+k$ where $1\leq k\leq m_r(\lambda;n)$,
and the ordered sequence $(\lar{0},\dots,\lar{t-1})$ of these partitions
is called the \emph{$t$-quotient}.
The precise order of the constituents of the $t$-quotient depends on the
residue class of $n$ modulo $t$. However, the orders only differ by cyclic
permutations, and Macdonald comments that it is best to think of the
quotient as a sort-of ``necklace'' of partitions.
To simplify things somewhat, we adopt the convention that the $t$-quotient is
always computed with $n$ a multiple of $t$, so that the order of its
constituents is fixed.
To define the $t$-core, one writes down the $n$ distinct integers $kt+r$ where
$0\leq k\leq m_r(\lambda;n)-1$ and $0\leq r\leq t-1$ in descending order,
say as $\tilde\xi_1>\dots>\tilde\xi_n$. Then $t\textup{-core}(\lambda)_i:=\tilde\xi_i-n+i$.
If $t\textup{-core}(\lambda)$ is empty then we say $\lambda$ \emph{has empty $t$-core}.
It will prove useful later on to work with the
\emph{bead configurations} (or \emph{bead diagram}, \emph{abacus model})
of James and Kerber \cite[\S2.7]{JK81}, which give a different model for
$t$-cores and $t$-quotients.
The ``board'' for a bead configuration is the set of nonnegative integers,
arranged in downward-increasing columns
$0\leq r\leq t-1$, called \emph{runners}, according to their
residues modulo $t$.
A bead is then placed at the space corresponding to each element of
$\beta(\lambda;n)$.
For example, $\lambda=(4,4,3,2,1)$ with $t=3$ and $n=6$ has configuration
\begin{center}
\begin{tikzpicture}[scale=0.6]
\foreach \i in {0,...,2}{
\draw[dashed,opacity=0.5] (\i,0) -- (\i,-3);}
\foreach \j in {0,...,3}{
\foreach \i in {0,...,2}{
\draw[fill=gray] (\i,-\j) circle (2pt);}}
\filldraw[black] (2,0) circle (4pt); \filldraw[black] (0,0) circle (4pt);
\filldraw[black] (1,-1) circle (4pt); \filldraw[black] (0,-2) circle (4pt);
\filldraw[black] (2,-2) circle (4pt);\filldraw[black] (0,-3) circle (4pt);
\end{tikzpicture}
\end{center}
Moving a bead up one space is equivalent to reducing one of the elements of
$\beta(\lambda;n)$ by $t$.
This is, in turn, equivalent to removing a $t$-ribbon
from $\lambda$ such that what remains is still a Young diagram
(see for instance \cite[p.~12]{Macdonald95}).
Pushing all beads to the top will give the bead configuration of
$t\textup{-core}(\lambda)$, and this is clearly independent of the order in which the beads
are pushed. It follows that $t\textup{-core}(\lambda)$ is the unique partition obtained
by removing $t$-ribbons (in a valid way) from the diagram of $\lambda$ until
it is no longer possible to do so.
We note that if removing a ribbon $R$ corresponds to moving a bead from position
$b$ to $b-t$, then $\mathrm{ht}(R)$ is equal to the number of beads lying
at the positions strictly between $b-t$ and $b$.
The $t$-quotient can be obtained from the bead configuration by reading the
$r$-th runner, bottom-to-top, as a bead configuration with $m_r(\lambda;n)$ beads.
For our example, this means that $\beta(t\textup{-core}(\lambda);6)=\{6,5,3,2,1,0\}$,
so that $t\textup{-core}(\lambda)=(1,1)$, with the quotient $((1,1),(1),(1))$ computed
similarly.
The above procedure of computing the $t$-core and $t$-quotient actually
encodes a bijection
\begin{align*}
\phi_t:\mathscr{P}&\longrightarrow\mathscr{C}_t\times\mathscr{P}^t \\
\lambda &\longmapsto\big(t\textup{-core}(\lambda),(\lar{0},\dots,\lar{t-1})\big),
\end{align*}
such that $\abs{\lambda}=\abs{t\textup{-core}(\lambda)}+t(\abs{\lar{0}}+\cdots+\abs{\lar{t-1}})$.
The arithmetic description of this correspondence was first written down by
Littlewood \cite{Littlewood51}.
The idea of removing ribbons from a partition until a unique core is obtained
goes back to Nakayama \cite{Nakayama40}.
The $t$-quotient of a partition has its origin in the \emph{star diagrams}
of Nakayama, Osima, Robinson and Staal \cite{NO51,Robinson48,Staal50},
which were shown to be equivalent to Littlewood's $t$-quotient by Farahat
\cite{Farahat53}.
Let $w_t(\lambda;n)$ be the permutation of $\beta(\lambda;n)$ which sorts the elements
of the beta set so that their residues modulo $t$ are increasing, and the
elements within each residue class decrease.
The sign of $w_t(\lambda;n)$ will be denoted $\sgn(w_t(\lambda;n))$.
The permutation $w_t(\lambda;n)$ can also be read off the bead configuration
by first labelling the beads ``backwards'': label the bead with largest place
$1$, second-largest $2$,
and so on. Reading the labels column-wise from bottom-to-top gives
$w_t(\lambda;n)$ in one-line notation.
An inversion in this permutation corresponds to a pair of beads $b_1,b_2$ such
that $b_2$ lies weakly below and strictly to the right of $b_2$.
In the above example $w_3(\lambda;6)=136425$ and
the bead at position $0$ generates three inversions, as it ``sees''
the beads $2, 4$ and $8$.
As follows from the above, a partition $\lambda$ has empty $t$-core if and only
if it is $t$-tileable.
In our results below we will need the following characterisation of when
a skew shape is $t$-tileable, generalising the notion of ``empty $t$-core''
to this case.
We briefly recall our convention that $t$-quotients are always computed
with the number of beads in the bead configuration a multiple of $t$.
\begin{lemma}\label{Lem_tilable}
A skew shape $\lambda/\mu$ is tileable by $t$-ribbons if and only if
$t\textup{-core}(\lambda)=t\textup{-core}(\mu)$ and $\mur{r}\subseteq\lar{r}$ for each
$0\leq r\leq t-1$.
\end{lemma}
\begin{proof}
The skew shape being $t$-tileable is equivalent to the diagram of
$\mu$ being obtainable from the diagram of $\lambda$ by removing $t$-ribbons.
In other words, we can obtain the bead configuration of $\mu$ from that of
$\lambda$, where both have $nt$ beads, by moving beads upwards.
Assume that this is the case.
Then $m_r(\lambda;nt)=m_r(\mu;nt)$ for each $0\leq r\leq t-1$, so that the $r$-th
runner has the same number of beads in each diagram.
This implies that $t\textup{-core}(\lambda)=t\textup{-core}(\mu)$.
It also follows that the $i$-th bead in each runner of $\lambda$'s
bead configuration must lie weakly below the $i$-th bead in the same
runner of $\mu$'s bead configuration.
Equivalently, $\mur{r}_i\leq \lar{r}_i$ for all $0\leq r\leq t-1$ and
$1\leq i\leq m_r(\lambda;nt)$, which in turn is equivalent to
$\mur{r}\subseteq\lar{r}$.
The reverse direction is now clear.
\end{proof}
Note that the lemma is also true when the $t$-quotients of $\lambda$ and
$\mu$ are computed using the same integer $n$ of any residue class modulo $t$.
If $\lambda/\mu$ is $t$-tileable, then we think of
$\lar{0}/\mur{0},\dots,\lar{t-1}/\mur{t-1}$ as its $t$-quotient.
When $\lambda/\mu$ is not $t$-tileable, it is not so clear how to define
the $t$-quotient.
\subsection{Symmetric functions and universal characters}
Here we discuss some basics of the theory of symmetric functions,
following \cite{Macdonald95}.
Let $\Lambda$ denote the \emph{ring of symmetric functions} in an
arbitrary countable set of variables $X$, called an \emph{alphabet}.
Where possible, we write elements of $\Lambda$ without reference to an alphabet if
the expression is independent of the chosen alphabet.
If we specialise to $X_n=(x_1,\dots,x_n)$ then the elements of
$\Lambda_{X_n}:=\mathbb{Z}[X_n]^{\mathfrak{S}_n}$ become
symmetric polynomials in the variables $x_1,\dots,x_n$.
Another specialisation is $X_n^{\pm}=(x_1,x_1^{-1},\dots,x_n,x_n^{-1})$,
which gives Laurent polynomials in the $x_i$ invariant under permutation and
inversion of the variables (i.e., $\mathrm{BC}_n$-symmetric functions).
We will later write $(x_1^\pm,\dots,x_n^\pm)$ for this alphabet.
Two fundamental algebraic bases for $\Lambda$ are the \emph{complete
homogeneous symmetric functions} and the \emph{elementary symmetric functions},
defined for any positive integer $r$ by
\[
h_r(X):=\sum_{1\leq i_1\leq\cdots\leq i_r}x_{i_1}\cdots x_{i_r}
\quad\text{and}\quad
e_r(X):=\sum_{1\leq i_1<\cdots<i_r}x_{i_1}\cdots x_{i_r},
\]
respectively. We further set $h_0=e_0:=1$ and $h_{-r}=e_{-r}=0$ for
positive $r$.
These admit the generating functions
\begin{align*}
H_z(X)&:=\sum_{r\geq0}z^rh_r(X)=\prod_{i\geq1}\frac{1}{1-zx_i}\\
E_z(X)&:=\sum_{r\geq0}z^re_r(X)=\prod_{i\geq1}(1+zx_i).
\end{align*}
The $h_r$ and $e_r$ for $r\geq 1$ are algebraically independent
over $\mathbb{Z}$ and generate $\Lambda$.
In view of this, we can define a homomorphism $\omega:\Lambda\longrightarrow\Lambda$
by $\omega h_r=e_r$.
It then follows from the relation $H_z(X)E_{-z}(X)=1$ that
$\omega e_r=h_r$, so that $\omega$ is an involution.
The most important family of symmetric functions are the \emph{Schur functions}.
These have several definitions, but
for our purposes it is best to define them, already for skew shapes, by the
\emph{Jacobi--Trudi formula}.
If $\lambda/\mu$ is a skew shape and $n$ an integer such that
$n\geq l(\lambda)$ we define
\begin{equation}\label{Eq_JT}
s_{\lambda/\mu}:=\det_{1\leq i,j\leq n}(h_{\lambda_i-\mu_j-i+j}).
\end{equation}
This is independent of $n$ as long as $n\geq l(\lambda)$.
If $\mu\not\subseteq\lambda$ then we set $s_{\lambda/\mu}:=0$.
There is also an equivalent formula in terms of the $e_r$,
called the \emph{dual Jacobi--Trudi formula} (rarely also the
\emph{N\"agelsbach--Kostka identity})
\[
s_{\lambda/\mu}=\det_{1\leq i,j\leq m}(e_{\lambda_i'-\mu_j'-i+j}).
\]
Restricting to the $\mu$ empty case,
we have $s_{(r)}=h_r$ and $s_{(1^r)}=e_r$.
Moreover, it is clear that $\omega s_{\lambda/\mu}=s_{\lambda'/\mu'}$.
If the set of variables $(x_1,\dots,x_n)$ is finite then the
Schur function for $\mu=0$ admits another definition as a ratio of alternants
\begin{equation}\label{Eq_s-alt}
s_\lambda(x_1,\dots,x_n)=\frac{\det_{1\leq i,j\leq n}(x_i^{\lambda_j+n-j})}
{\det_{1\leq i,j\leq n}(x_i^{n-j})}.
\end{equation}
The denominator is the Vandermonde determinant
$\det_{1\leq i,j\leq n}(x_i^{n-j})=\prod_{1\leq i<j\leq n}(x_i-x_j)$.
In this case we also define $s_\lambda(x_1,\dots,x_n)=0$ if $l(\lambda)>n$.
If $\lambda$ is a partition of length at most $n$, then
\begin{equation}\label{Eq_det-rep}
s_{(\lambda_1+1,\dots,\lambda_n+1)}(x_1,\dots,x_n)
=(x_1\cdots x_n)s_{(\lambda_1,\dots,\lambda_n)}(x_1,\dots,x_n).
\end{equation}
This allows for Schur functions on a finite set of $n$ variables to be extended
to weakly decreasing sequences of integers of length exactly $n$.
Following Koike and Terada we define the \emph{universal characters} for
$\mathrm{O}(2n,\mathbb{C})$ and $\mathrm{Sp}(2n,\mathbb{C})$ as
the symmetric functions \cite[Definition~2.1.1]{KT87}
\begin{align}\label{Eq_oh}
\o_\lambda&:=\det_{1\leq i,j\leq n}\big(h_{\lambda_i-i+j}-h_{\lambda_i-i-j}\big)\\
\sp_\lambda&:=\frac{1}{2}
\det_{1\leq i,j\leq n}\big(h_{\lambda_i-i+j}+h_{\lambda_i-i-j+2}\big),\label{Eq_sph}
\end{align}
where $n\geq l(\lambda)$.
Like the Schur functions, these determinants also have dual versions
\begin{align}
\o_\lambda&=\frac{1}{2}\notag
\det_{1\leq i,j\leq m}\big(e_{\lambda_i'-i+j}+e_{\lambda_i'-i-j+2}\big)\\
\sp_\lambda&=
\det_{1\leq i,j\leq m}\big(e_{\lambda_i'-i+j}-e_{\lambda_i'-i-j}\big), \label{Eq_sp-e}
\end{align}
where $m\geq \lambda_1$.
From this it is clear that $\omega \o_\lambda=\sp_{\lambda'}$.
Koike alone added a third universal character for the
group $\mathrm{SO}(2n+1,\mathbb{C})$ \cite[Definition~6.4]{Koike97}
(see also \cite[Equation~(3.8)]{LRW20})
\[
\mathrm{so}_\lambda:=\det_{1\leq i,j\leq n}\big(h_{\lambda_i-i+j}+h_{\lambda_i-i-j+1}\big)
=\det_{1\leq i,j\leq m}\big(e_{\lambda_i'-i+j}+e_{\lambda_i'-i-j+1}\big).
\]
This universal character is self-dual under $\omega$, so
$\omega \mathrm{so}_\lambda=\mathrm{so}_{\lambda'}$.
For later convenience we also define a variant of the above as
\[
\mathrm{so}_\lambda^{-}:=\det_{1\leq i,j\leq n}\big(h_{\lambda_i-i+j}-h_{\lambda_i-i-j+1}\big)
=\det_{1\leq i,j\leq m}\big(e_{\lambda_i'-i+j}-e_{\lambda_i'-i-j+1}\big).
\]
If $X=(x_1,x_2,x_3,\dots)$ is a set of variables (which may be finite or
countable)
and $-X:=(-x_1,-x_2,-x_3,\dots)$, then
\[
\mathrm{so}_\lambda^{-}(X)=(-1)^{\abs{\lambda}}\mathrm{so}_\lambda(-X),
\]
since the $h_r$ and $e_r$ are homogeneous of degree $r$.
For $l(\lambda)\leq n$, each of the three above universal characters become actual
characters of irreducible representations of their associated groups
when specialised to $(x_1^\pm,\dots,x_n^\pm)$ (hence the name
universal characters).
The irreducible polynomial representations of $\mathrm{GL}(n,\mathbb{C})$
are indexed by partitions of length at most $n$.
On the other hand, the irreducible rational representations are indexed by
weakly decreasing sequences of integers of length $n$,
which are called \emph{staircases} by Stembridge \cite{Stembridge87}.
Such sequences are equivalent to pairs of partitions $\lambda,\mu$
such that $l(\lambda)+l(\mu)\leq n$.
Given such a pair, one defines the associated staircase $[\lambda,\mu]$ by
$[\lambda,\mu]_i:=\lambda_i-\mu_{n-i+1}$ for $1\leq i\leq n$.
The characters of the rational representations of $\mathrm{GL}(n,\mathbb{C})$
are then given by $s_{[\lambda,\mu]}(x_1,\dots,x_n)$ for all staircases with
$n$ entries.
Note that \eqref{Eq_det-rep} implies that this object is just a Schur function
up to a power of $x_1\cdots x_n$.
In \cite{Littlewood44}, Littlewood gave the expansion
\begin{equation}\label{Eq_Littlewood-def}
s_{[\lambda,\mu]}(x_1,\dots,x_n)
=\sum_{\nu}(-1)^{\abs{\nu}}s_{\lambda/\nu}(x_1,\dots,x_n)
s_{\mu/\nu'}(1/x_1,\dots,1/x_n).
\end{equation}
For a pair of partitions $\lambda,\mu$ and sets of indeterminates $X$, $Y$,
this may be used to define the universal character associated to a rational
representation of $\mathrm{GL}(n,\mathbb{C})$ as
\begin{equation}\label{Eq_rs-def}
\mathrm{rs}_{\lambda,\mu}(X;Y)
:=\sum_{\nu}(-1)^{\abs{\nu}}s_{\lambda/\nu}(X)s_{\mu/\nu'}(Y).
\end{equation}
Note that the only terms which contribute are those with $\nu\subseteq\lambda$
and $\nu'\subseteq\mu$.
If we let $\omega_X$ and $\omega_Y$ denote the involution $\omega$ acting on
the set of variables in its subscript, then
\begin{align*}
\omega_X\omega_Y\mathrm{rs}_{\lambda,\mu}(X;Y)
&=\sum_{\nu}(-1)^{\abs{\nu}}s_{\lambda'/\nu'}(X)s_{\mu'/\nu}(Y)\\
&=\sum_{\nu'}(-1)^{\abs{\nu}}s_{\lambda'/\nu}(X)s_{\mu'/\nu'}(Y)\\
&=\mathrm{rs}_{\lambda',\mu'}(X;Y).
\end{align*}
As shown by Koike \cite{Koike89}, this object has a Jacobi--Trudi-type
expression as a block matrix
\begin{equation}\label{Eq_rs-h}
\mathrm{rs}_{\lambda,\mu}(X;Y)=
\det_{1\leq i,j\leq n+m}
\begin{pmatrix}
\big(h_{\lambda_i-i+j}(X)\big)_{1\leq i,j\leq n} &
\big(h_{\lambda_i-i-j+1}(X)\big)_{\substack{1\leq i\leq n\\ 1\leq j\leq m}} \\
\big(h_{\mu_i-i-j+1}(Y)\big)_{\substack{1\leq i\leq m\\ 1\leq j\leq n}} &
\big(h_{\mu_i-i+j}(Y)\big)_{1\leq i,j\leq m}
\end{pmatrix},
\end{equation}
where $n\geq l(\lambda)$ and $m\geq l(\mu)$.
As for the other determinants, this is independent of $n$ and $m$ as long as
$n\geq l(\lambda)$ and $m\geq l(\mu)$.
The relation under $\omega_X\omega_Y$ implies we have the dual
form \cite[Definition~2.1]{Koike89}
\begin{equation}\label{Eq_rs-e}
\mathrm{rs}_{\lambda,\mu}(X;Y)=
\det_{1\leq i,j\leq n+m}
\begin{pmatrix}
\big(e_{\lambda_i'-i+j}(X)\big)_{1\leq i,j\leq n} &
\big(e_{\lambda_i'-i-j+1}(X)\big)_{\substack{1\leq i\leq n\\ 1\leq j\leq m}} \\
\big(e_{\mu_i'-i-j+1}(Y)\big)_{\substack{1\leq i\leq m\\ 1\leq j\leq n}} &
\big(e_{\mu_i'-i+j}(Y)\big)_{1\leq i,j\leq m}
\end{pmatrix},
\end{equation}
where $n\geq \lambda_1$ and $m\geq\mu_1$.
The definition \eqref{Eq_rs-def} and the determinants \eqref{Eq_rs-h}
and \eqref{Eq_rs-e} are related by taking the Laplace expansion of
each determinant according to its presented block structure;
see \cite[Equation~(2.1)]{Koike89}.
Also, by the definition \eqref{Eq_rs-def} and \eqref{Eq_Littlewood-def} it is
immediate that, for $l(\lambda)+l(\mu)\leq n$,
\[
\mathrm{rs}_{\lambda,\mu}(x_1,\dots,x_n;1/x_1,\dots,1/x_n)=s_{[\lambda,\mu]}(x_1,\dots,x_n).
\]
We will always take $X=Y$ in $\mathrm{rs}_{\lambda,\mu}(X;Y)$, which we write as
$\mathrm{rs}_{\lambda,\mu}$ in the rest of the paper.
In particular we note that $\mathrm{rs}_{\lambda,\mu}=\mathrm{rs}_{\mu,\lambda}$.
As mentioned in the introduction, the notion of twisting the set of variables
$x_1,\dots,x_n$ by a primitive $t$-th root of unity $\zeta$ is replaced by
the operator $\varphi_t$ \eqref{Eq_phi-def}, which has been considered
by Macdonald \cite[p.~91]{Macdonald95} and, for $t=2$, by Baik and Rains
\cite[p.~25]{BR01}.
Let $X^t:=(x_1^t,x_2^t,x_3^t\dots)$ and denote by $\psi_t$ the
homomorphism
\begin{align*}
\psi_t: \Lambda&\longrightarrow \Lambda_{X^t}\\
f&\longmapsto f(X,\zeta X,\dots,\zeta^{t-1}X).
\end{align*}
Since $\psi_t H_z(X)=H_{z^t}(X^t)$, both $\varphi_t$ and $\psi_t$ act
on the $h_r$ in the same way, i.e., the diagram
\begin{equation}\label{Eq_diagram}
\begin{tikzcd}
\Lambda\arrow{r}{\varphi_t} \arrow{dr}{\psi_t} & \Lambda \arrow{d}{X^t}\\
& \Lambda_{X^t}
\end{tikzcd}
\end{equation}
commutes, where the arrow labelled $X^t$ is the substitution map.
This implies the claim of the introduction that the action of $\varphi_t$ is
equivalent to twisting the alphabet $X$ by a primitive $t$-th root of unity
$\zeta$.
If one wishes to think about this as a map $\Lambda_X\longrightarrow\Lambda_X$
where $X$ is some concrete alphabet, then substitute each $x\in X$ by its set
of $t$-th roots $x^{1/t},\zeta x^{1/t},\dots,\zeta^{t-1}x^{1/t}$ and
evaluate this expression.
By the action of this map on the $h_r$, such a map gives a symmetric function
again in the variables $X$.
Using the generating function $E_z(X)$ one may also show that
\cite[\S5.8]{Lascoux?}
\[
\varphi_t e_r=\begin{cases}
(-1)^{r(t-1)/t}e_{r/t} & \text{if $t$ divides $r$}, \\
0 & \text{otherwise}.
\end{cases}
\]
Therefore $\omega$ and $\varphi_t$ commute if $t$ is odd, but not in general.
Proposition~\ref{Prop_signs} shows that, in some cases, the maps commute up
to a computable sign.
Different, but closely related operators are discussed at the end of this paper.
\section{Summary of results}\label{Sec_results}
With the preliminary material of the previous section under our belts, we
are now ready to state our main results regarding factorisations of
universal characters under $\varphi_t$.
The first of these is the action of the map on the skew Schur functions.
\begin{theorem}\label{Thm_skew}
We have that $\varphi_t s_{\lambda/\mu}=0$ unless $\lambda/\mu$ is tileable by
$t$-ribbons, in which case
\[
\varphi_t s_{\lambda/\mu}=\sgn_t(\lambda/\mu)
\prod_{r=0}^{t-1}s_{\lar{r}/\mur{r}}.
\]
\end{theorem}
For $\mu$ empty, this result is due to Littlewood
\cite[p.~131]{Littlewood40}, who proves it by direct manipulation of the
ratio of alternants \eqref{Eq_s-alt}.
By \eqref{Eq_diagram} with $X=(x_1,\dots,x_n)$, one can recover Littlewood's
result by simply evaluating the right-hand side of the equation
at $(x_1^t,\dots,x_n^t)$.
He also states the $\mu$ empty case of the theorem
in the language of symmetric group characters:
see both \cite[p.~144]{Littlewood40} and \cite[p.~340]{Littlewood51}.
The generalisation to skew characters was
discovered by Farahat \cite{Farahat54} (see also \cite[Theorem~3.3]{EPW14}).
The form we state here is precisely that of Macdonald
\cite[p.~91]{Macdonald95}.
Curiously, Prasad recently rediscovered the $\mu$ empty case
independently, with a proof identical to Littlewood's, but
in a more representation-theoretic context \cite{Prasad16}.
A version of the result for Schur's $P$-and $Q$-functions has been given by
Mizukawa \cite[Theorem~5.1]{Mizukawa02}.
Theorem~\ref{Thm_skew} has been rediscovered many times for
both skew and straight shapes, and often only in special cases.
We make no attempt to give a complete history, but it appears to us that
the theorem deserves to be better known.
The interested reader can consult \cite{Wildon22} for some exposition on
the character theory side of this story.
On the symmetric functions side, such an exposition is lacking in the
literature.
We now state, in sequence, the three factorisations lifting
\cite[Theorems~2.11, 2.15 \& 2.17]{AK22} to the level of universal characters,
beginning with the universal orthogonal character.
\begin{theorem}\label{Thm_AK_o}
Let $\lambda$ be a partition of length at most $nt$.
Then $\varphi_t\o_\lambda=0$ unless $t\textup{-core}(\lambda)$ is orthogonal, in which case
\[
\varphi_t\o_\lambda=
(-1)^{\varepsilon^{\o}_{\lambda;nt}} \sgn(w_t(\lambda;nt))
\o_{\lar{0}}\prod_{r=1}^{\lfloor(t-1)/2\rfloor}\mathrm{rs}_{\lar{r},\lar{t-r}}
\times\begin{cases}
\mathrm{so}_{\lar{t/2}}^- & \text{if $t$ is even}, \\
1 & \text{if $t$ is odd},
\end{cases}
\]
where
\[
\varepsilon^{\o}_{\lambda;nt}
=\sum_{r=\lfloor(t+2)/2\rfloor}^{t-1}\binom{m_r(\lambda;nt)+1}{2}
+d(t\textup{-core}(\lambda))
+\begin{cases}
\binom{n+1}{2}+nd(t\textup{-core}(\lambda)) & \text{if $t$ is even}, \\
0 & \text{if $t$ is odd}.
\end{cases}
\]
\end{theorem}
Our next result is the same factorisation for the symplectic character.
\begin{theorem}\label{Thm_AK_sp}
Let $\lambda$ be a partition of length at most $nt$.
Then $\varphi_t\sp_\lambda=0$ unless $t\textup{-core}(\lambda)$ is symplectic, in which case
\[
\varphi_t\sp_\lambda=
(-1)^{\varepsilon^{\sp}_{\lambda;nt}} \sgn(w_t(\lambda;nt))\sp_{\lar{t-1}}
\prod_{r=0}^{\lfloor(t-3)/2\rfloor}\mathrm{rs}_{\lar{r},\lar{t-r-2}}
\times\begin{cases}
\mathrm{so}_{\lar{(t-2)/2}} & \text{if $t$ is even}, \\
1 & \text{if $t$ is odd}
\end{cases}
\]
where
\[
\varepsilon^{\sp}_{\lambda;nt}
=\sum_{r=\lfloor t/2\rfloor}^{t-2}\binom{m_r(\lambda;nt)+1}{2}
+\begin{cases} \binom{n+1}{2}+
nd(t\textup{-core}(\lambda)) & \text{if $t$ is even}, \\
0 & \text{if $t$ is odd}.
\end{cases}
\]
\end{theorem}
Finally, we can claim a similar factorisation for $\mathrm{so}_\lambda$.
\begin{theorem}\label{Thm_AK_so}
Let $\lambda$ be a partition of length at most $nt$.
Then $\varphi_t\mathrm{so}_\lambda=0$ unless $t\textup{-core}(\lambda)$ is self-conjugate, in which case
\[
\varphi_t\mathrm{so}_\lambda=
(-1)^{\varepsilon^{\mathrm{so}}_{\lambda;nt}} \sgn(w_t(\lambda;nt))
\prod_{r=0}^{\lfloor(t-2)/2\rfloor}\mathrm{rs}_{\lar{r},\lar{t-r-1}}
\times\begin{cases}
1 & \text{if $t$ is even},\\
\mathrm{so}_{\lar{(t-1)/2}} & \text{if $t$ is odd}.
\end{cases}
\]
where\footnote{We have corrected the lower bound in the sum defining
$\varepsilon^{\mathrm{so}}_{\lambda;nt}$ from $\lfloor t/2\rfloor$ in
\cite[Theorem~2.17]{AK22} (there denoted $\epsilon$) to
$\lfloor (t+1)/2\rfloor$.}
\[
\varepsilon^{\mathrm{so}}_{\lambda;nt}=
\sum_{r=\lfloor (t+1)/2\rfloor}^{t-1}\binom{m_r(\lambda;nt)+1}{2}+
\begin{cases}
0 & \text{if $t$ is even}, \\
nd(t\textup{-core}(\lambda)) & \text{if $t$ is odd}.
\end{cases}
\]
\end{theorem}
Some remarks are in order.
Firstly, the three signs $\sgn(w_t(\lambda;nt))(-1)^{\varepsilon_{\lambda;nt}^\bullet}$
are actually independent of $n$ as long as $nt\geq l(\lambda)$, a fact which
we prove in Lemma~\ref{Lem_sign-invariance} below.
As remarked by Ayyer and Kumari \cite[Remark~2.19]{AK22},
the order of the quotient is unchanged upon replacing $n\mapsto n+1$,
so the product in the evaluation is independent of $n$.
It is in principle possible to carry out our proof technique below under
the assumption that $l(\lambda)$ is bounded by an arbitrary integer, say $k$,
where $k$ is not necessarily a multiple of $t$.
In this case the evaluation is of course the same, however the sign will be
expressed differently and the $t$-quotients in the evaluations will be
a cyclic permutation of the ones presented.
Since the proof is simplest when this $k$ is a multiple of $t$,
we stick to this case.
To obtain the theorems of Ayyer and Kumari one evaluates the right-hand side
of each identity at the set of variables $(x_1^{\pm t},\dots,x_n^{\pm t})$.
Using \eqref{Eq_det-rep} and the definition of $\mathrm{rs}_{\lambda,\mu}$ it follows
that in this case the rational universal characters occurring in each evaluation
agree with the Schur functions
$s_{\mur{k}_i}(x_1^{\pm t},\dots,x_n^{\pm t})$ in the notation of
\cite{AK22}.
As we have already seen the maps $\omega$ and $\varphi_t$ do not commute in
general.
However, when the symmetric function is one of $s_{\lambda/\mu},\o_\lambda,\sp_\lambda,
\mathrm{so}_\lambda$, the maps commute up to a sign which can be computed.
\begin{proposition}\label{Prop_signs}
There holds
\[
\omega\varphi_t s_{\lambda/\mu}=
(-1)^{(t-1)(\abs{\lar{0}}+\cdots+\abs{\lar{t-1}}-
\abs{\mur{0}}-\cdots-\abs{\mur{t-1}}
)}
\varphi_t\omega s_{\lambda/\mu}.
\]
Moreover,
\[
\omega\varphi_t \bullet_\lambda =
(-1)^{(t-1)(\abs{\lar{0}}+\cdots+\abs{\lar{t-1}})}\varphi_t\omega \bullet_\lambda
\]
for $\bullet\in\{\o,\sp,\mathrm{so}\}$.
\end{proposition}
\section{Auxiliary results}\label{Sec_lemmas}
The purpose of this section is to collect all the small facts about
beta sets and the signs \eqref{Eq_sgn-def} which we need to prove our
main results.
To begin, we relate the bead configurations of a partition and its conjugate.
\begin{lemma}\label{Lem_reverse}
Let $\lambda$ be a partition of length at most $nt$ such that $\lambda_1\leq mt$.
Then the bead configuration for $\beta(\lambda';mt)$ can be obtained from
the bead configuration for $\beta(\lambda;nt)$ with $n+m$ rows by rotating
the picture by $180^\circ$ and then interchanging beads and spaces.
\end{lemma}
\begin{proof}
This is a consequence of the fact \cite[p.~3]{Macdonald95} that
for $l(\lambda)\leq n$ and $\lambda_1\leq m$,
\[
\{0,1,\dots,m+n-1\}=\{\lambda_i+n-i:1\leq i\leq n\}\sqcup
\{m+n-1-(\lambda_j'+m-j):1\leq j\leq m\},
\]
where $\sqcup$ denotes a disjoint union.
\end{proof}
This lemma immediately implies the following relationship between the
$t$-core and $t$-quotient of $\lambda$ and $\lambda'$.
\begin{corollary}\label{Cor_reverse}
Let $\lambda$ be a partition with $t$-quotient $(\lar{0},\dots,\lar{t-1})$.
Then $t\textup{-core}(\lambda')=t\textup{-core}(\lambda)'$ and
$\big((\lar{t-1})',\dots,(\lar{0})'\big)$ is the $t$-quotient of $\lambda'$.
\end{corollary}
The next pair of lemmas are due to Ayyer and Kumari, the first of which
characterises partitions with $z$-asymmetric $t$-cores in terms of their
beta sets \cite[Lemma~3.6]{AK22}.
\begin{lemma}\label{Lem_AK_orth}
For a partition $\lambda$ of length at most $nt$, $t\textup{-core}(\lambda)$ is of the form
$(a\mid a+z)$ for some integer $z$ if and only if
\begin{subequations}\label{Eq_m_conditions}
\begin{align}
m_i(\lambda,nt)+m_{t-i-z-1}(\lambda,nt)&=2n \quad \text{for $0\leq i\leq t-z-1$}, \\
m_i(\lambda,nt)&=n \quad \text{for $t-z\leq i\leq t-1$}.
\end{align}
\end{subequations}
\end{lemma}
The second lemma of Ayyer and Kumari we need is \cite[Lemma~3.13]{AK22},
which is used later on to simplify signs.
\begin{lemma}
Let $\lambda$ be a partition of length at most $nt$.
If $t\textup{-core}(\lambda)$ is orthogonal, then
\begin{equation}\label{Eq_o-rk}
d(t\textup{-core}(\lambda))=\sum_{r=1}^{\lfloor(t-1)/2\rfloor}\abs{m_r(\lambda;nt)-n}
=\sum_{r=\lfloor(t+2)/2\rfloor}^{t-1}\abs{m_r(\lambda;nt)-n}
\end{equation}
If $t\textup{-core}(\lambda)$ is symplectic, then
\begin{equation}\label{Eq_sp-rk}
d(t\textup{-core}(\lambda))=\sum_{r=0}^{\lfloor(t-3)/2\rfloor}\abs{m_r(\lambda;nt)-n}
=\sum_{r=\lfloor t/2\rfloor}^{t-2}\abs{m_r(\lambda;nt)-n}
\end{equation}
If $t\textup{-core}(\lambda)$ is self-conjugate, then
\begin{equation}\label{Eq_so-rk}
d(t\textup{-core}(\lambda))=\sum_{r=0}^{\lfloor(t-2)/2\rfloor}\abs{m_r(\lambda;nt)-n}
=\sum_{r=\lfloor(t+1)/2\rfloor}^{t-1}\abs{m_r(\lambda;nt)-n}
\end{equation}
\end{lemma}
Next, we show that the sign of a tileable skew shape can be expressed in
terms of the signs of the permutations $w_t(\lambda;n)$.
\begin{lemma}\label{Lem_signs0}
For $\lambda/\mu$ $t$-tileable and any integer $n$ such that $n\geq l(\lambda)$,
\[
\sgn_t(\lambda/\mu)=\sgn(w_t(\lambda;n))\sgn(w_t(\mu;n)).
\]
\end{lemma}
\begin{proof}
Since $\lambda/\mu$ is $t$-tileable, it has a ribbon decomposition
$D=(\nur{0},\dots,\nur{k})$ where $\nur{0}=\mu$ and $\nur{k}=\lambda$.
Also, $\nur{k-1}$ can be obtained from $\lambda$ by moving one bead at some
position upward one space.
By our characterisation of the inversions in the permutation $w_t(\lambda;n)$,
we see that moving a bead at position $\ell$ up one space changes the sign by
$(-1)^{b_k}$ where $b_k$ is the number of beads at positions between
$\ell-t$ and $\ell$.
In other words, $\sgn(w_t(\lambda;n))=(-1)^{b_k}\sgn(w_t(\nur{k-1};n))$.
Moreover, $b_k=\mathrm{ht}(\nur{k}/\nur{k-1})$, so that
\[
\sgn(w_t(\lambda;n))\sgn(w_t(\mu;n))=(-1)^{\sum_{i=1}^k b_i}=
(-1)^{\mathrm{ht}(D)}=\sgn_t(\lambda/\mu). \qedhere
\]
\end{proof}
We also have the following useful relationship between the sign of $\lambda/\mu$
and $\lambda'/\mu'$.
\begin{lemma}\label{Lem_signs1}
For $\lambda/\mu$ $t$-tileable,
\[
\sgn_t(\lambda/\mu)\sgn_t(\lambda'/\mu')=
(-1)^{(t-1)(\abs{\lar{0}}+\cdots+\abs{\lar{t-1}}-
\abs{\mur{0}}-\cdots-\abs{\mur{t-1}})}.
\]
\end{lemma}
\begin{proof}
To prove the claim of the lemma we will proceed by induction on $\abs{\lambda/\mu}$.
If $\abs{\lambda/\mu}=0$ then $\lambda=\mu$ and the equation is trivial.
Now fix $\mu$ and assume the result holds for $\lambda/\mu$ being $t$-tileable.
Adding a $t$-ribbon to $\lambda/\mu$ moves one of the beads, say at position $b$,
in the bead configuration for $\lambda$ down a single space.
The change in the number of inversions in $w_t(\lambda;nt)$ is the number of beads
$b'$ such that $b<b'<b+1$.
A consequence of Lemma~\ref{Lem_reverse} is that $w_t(\lambda';mt)$ will change
by the number of empty spaces between $b$ and $b+1$.
There are $t-1$ spaces and beads between $b$ and $b+1$, so the left-hand
side changes by $(-1)^{t-1}$ when adding a $t$-ribbon.
But adding a $t$-ribbon to $\lambda/\mu$ changes some element of the $t$-quotient
of $\lambda$ by a single box, also corresponding to a change in sign of
$(-1)^{t-1}$.
\end{proof}
There is another sign relation between orthogonal and symplectic $t$-cores,
but this time using the permutations $w_t$.
\begin{lemma}\label{Lem_signs2}
Let $\lambda$ be an orthogonal or symplectic $t$-core whose diagram is contained
in an $nt\times nt$ square. Then
\[
\sgn(w_t(\lambda;nt))\sgn(w_t(\lambda';nt))=(-1)^{d(\lambda)}.
\]
\end{lemma}
\begin{proof}
Assume that $\lambda$ is a non-empty, orthogonal $t$-core (if $\lambda$ is empty
the result is trivial) and fix $n$ so that the condition of the theorem holds.
The key observation is that for an orthogonal $t$-core, the bead configuration
of $\lambda'$ with $nt$ beads can be obtained from the bead configuration of $\lambda$
with $nt$ beads by reducing the labels by $1$ modulo $t$.
For example if $\lambda=(12,7,5,3,2,2,1,1,1,1,1)$ then
$\lambda'=(11,6,4,3,3,2,2,1,1,1,1,1)$ and their bead configurations are
\begin{center}
\begin{tikzpicture}[scale=0.5]
\foreach \i in {0,...,5}{\draw[dashed,opacity=0.5] (\i,0) -- (\i,-3);}
\foreach \i in {0,...,5}{\draw[dashed,opacity=0.5] (\i+10,0) -- (\i+10,-3);}
\foreach \j in {0,...,3}{
\foreach \i in {0,...,5}{
\draw[fill=gray] (\i,-\j) circle (2pt);}}
\filldraw[black] (0,0) circle (4pt); \filldraw[black] (0,-1) circle (4pt);
\filldraw[black] (2,0) circle (4pt); \filldraw[black] (2,-1) circle (4pt);
\filldraw[black] (2,-2) circle (4pt);\filldraw[black] (3,0) circle (4pt);
\filldraw[black] (3,-1) circle (4pt);\filldraw[black] (4,0) circle (4pt);
\filldraw[black] (5,0) circle (4pt);\filldraw[black] (5,-1) circle (4pt);
\filldraw[black] (5,-2) circle (4pt);\filldraw[black] (5,-3) circle (4pt);
\foreach \j in {0,...,3}{
\foreach \i in {0,...,5}{
\draw[fill=gray] (\i+10,-\j) circle (2pt);}}
\filldraw[black] (15,0) circle (4pt); \filldraw[black] (15,-1) circle (4pt);
\filldraw[black] (11,0) circle (4pt); \filldraw[black] (11,-1) circle (4pt);
\filldraw[black] (11,-2) circle (4pt);
\filldraw[black] (12,0) circle (4pt);\filldraw[black] (12,-1) circle (4pt);
\filldraw[black] (13,0) circle (4pt);\filldraw[black] (14,0) circle (4pt);
\filldraw[black] (14,-1) circle (4pt);\filldraw[black] (14,-2) circle (4pt);
\filldraw[black] (14,-3) circle (4pt);
\end{tikzpicture}
\end{center}
respectively.
This is a consequence of Lemma~\ref{Lem_AK_orth} with $z=\pm1$ and
Lemma~\ref{Lem_reverse}.
When passing from $\lambda$ to $\lambda'$, the inversions contributed by the beads
in the first runner are removed and replaced by additional inversions associated
to the remaining beads in the first $n$ rows.
Modulo $2$, this is equivalent to each bead in the zeroeth runner now seeing
all of the beads in the same row twice, plus all other beads in the other
runners once.
Let $b$ be the number of beads in the first $n$ rows of the runners
from $1$ to $t-1$ in the bead configuration of $\lambda$.
Then the sign change is
\[
\sgn(w_t(\lambda;nt))=\sgn(w_t(\lambda';nt))(-1)^{n^2(t-1)+b}.
\]
Since $\lambda$ is an orthogonal $t$-core, $n^2(t-1)+b\equiv d(\lambda)\Mod{2}$ by
\eqref{Eq_o-rk} and Lemma~\ref{Lem_AK_orth} with $z=-1$.
\end{proof}
The next lemma proves the claim made after
Theorems~\ref{Thm_AK_o}--\ref{Thm_AK_so} that the signs occurring in those
factorisations are independent of $n$.
\begin{lemma}\label{Lem_sign-invariance}
The signs $(-1)^{\varepsilon_{\lambda;nt}^{\bullet}}\sgn(w_t(\lambda;nt))$ for
$\bullet\in\{\o,\sp,\mathrm{so}\}$ are independent of $n$ as long as $nt\geq l(\lambda)$.
\end{lemma}
\begin{proof}
Assume that $nt\geq l(\lambda)$. Incrementing $n$ by one adds a row of beads to
the top of the bead configuration of $\lambda$, and so
$m_r(\lambda;(n+1)t)=m_r(\lambda;nt)+1$.
In the inversion count, the $r$th bead in the new first row sees
\[
\sum_{k=r+1}^{t-1}(m_{k}(\lambda;nt)+1)
\]
other beads.
Summing over $k=0,\dots,t-1$ we see that
\[
\sgn(w_t(\lambda;(n+1)t))=\sgn(w_t(\lambda;nt))
(-1)^{\sum_{r=1}^{\lfloor t/2\rfloor}(m_{2r-1}(\lambda;nt)+1)}.
\]
Now assume that $\lambda$ has an orthogonal $t$-core.
Then by Lemma~\ref{Lem_AK_orth} with $z=-1$ the above has the same parity as
\[
\sum_{r=1}^{\lfloor t/2\rfloor}(m_{2r-1}(\lambda;nt)+1)
\equiv\begin{cases}\frac{(n+1)t}{2} & \text{if $t$ is even,}
\\[2mm]
\frac{t-1}{2}+\sum_{r=1}^{(t-1)/2}m_r(\lambda;nt) & \text{if $t$ is odd},
\end{cases}
\Mod{2}.
\]
A short calculation shows that
\begin{align*}
\varepsilon_{\lambda;(n+1)t}^{\o}
&=\varepsilon_{\lambda;nt}^{\o}+\sum_{r=1}^{\lfloor (t-1)/2\rfloor}
m_r(\lambda;nt)+\begin{cases}
n+\frac{t}{2}+d(t\textup{-core}(\lambda)) & \text{if $t$ is even},\\[1mm]
\frac{t-1}{2} & \text{if $t$ is odd},
\end{cases} \\
&\equiv\varepsilon_{\lambda;nt}^{\o}+\begin{cases}
\frac{(n+1)t}{2} & \text{if $t$ is even},\\[1mm]
\frac{t-1}{2} +
\sum_{r=1}^{(t-1)/2}m_r(\lambda;nt)
& \text{if $t$ is odd}
\end{cases}
\Mod{2}.
\end{align*}
where the last equality uses \eqref{Eq_o-rk}.
The remaining two cases follow similarly.
\end{proof}
We conclude this section with a small lemma relating the indices in the
Jacobi--Trudi determinants with partition quotients.
\begin{lemma}\label{Lem_phi}
Let $\lambda,\mu$ be partitions of length at most $nt$ and
assume that for $0\leq r,s\leq t-1$ we have
$\lambda_i-i\equiv r\Mod{t}$, $\mu_j-j\equiv s\Mod{t}$ for $1\leq i,j\leq nt$.
If $r-s+z\equiv 0\Mod{t}$ for some $z\in\mathbb{Z}$, then
\[
\frac{\lambda_i-\mu_j+j-i+z}{t}=\lar{r}_k-\mur{s}_\ell-k+\ell
+m_r(\lambda;nt)-m_s(\mu;nt)+(r-s+z)/t,
\]
for some $k,\ell$ such that
$1\leq k\leq m_r(\lambda;nt)$ and $1\leq\ell\leq m_s(\mu;nt)$.
Alternatively, if $r+s+z\equiv 0\Mod{t}$ then
\[
\frac{\lambda_i+\mu_j-i-j+z}{t}
=\lar{r}_k+\mur{s}_\ell-k-\ell-2n+1+m_r(\lambda;nt)+m_s(\mu;nt)+(r+s+z)/t
\]
for some $k, \ell$ such that
$1\leq k\leq m_r(\lambda;nt)$ and $1\leq\ell\leq m_{s}(\mu;nt)$.
\end{lemma}
\begin{proof}
We first write $\lambda_i+nt-i=\xir{r}_kt+r$ and $\mu_j+nt-j=\pi_\ell^{(s)}t+s$
for $1\leq k\leq m_r(\lambda;nt)$ and $1\leq \ell\leq m_s(\mu;nt)$.
Then
\begin{align*}
\frac{\lambda_i-\mu_j-i+j+z}{t}
&=\xir{r}_k+\pi_\ell^{(s)}+(r-s+z)/t\\
&=\lar{r}_k+\mur{s}_\ell-k+\ell+m_r(\lambda;nt)-m_s(\mu;nt)+(r-s+z)/t,
\end{align*}
by the definition of the $t$-quotient. The second claim is analogous.
\end{proof}
\section{Proofs of theorems}\label{Sec_proofs}
In this section we provide proofs of Theorems~\ref{Thm_AK_o},~\ref{Thm_AK_sp}
and \ref{Thm_AK_so}. Since our proof strategy follows that of Macdonald's proof
of the skew Schur case \cite[p.~91]{Macdonald95} (Theorem~\ref{Thm_skew} above),
we reproduce this proof in detail as preparation for what remains.
\subsection{Proof of Theorem~\ref{Thm_skew}}
Let $n$ be a nonnegative integer and $\mu\subseteq\lambda$ be a pair of partitions
such that $l(\lambda)\leq nt$.
Consider the Jacobi--Trudi determinant
\[
s_{\lambda/\mu}=\det_{1\leq i,j\leq nt}(h_{\lambda_i-\mu_j-i+j}).
\]
Before applying the map $\varphi_t$, we rearrange the rows and columns of
this determinant by the permutations $w_t(\lambda;nt)$ and $w_t(\mu;nt)$
respectively.
By Lemma~\ref{Lem_signs0} this introduces a sign of $\sgn_t(\lambda/\mu)$.
The rows and columns are now arranged in such a way that the residue
classes of $\lambda_i-i$ and $\mu_j-j$ are grouped in ascending order, and
the values within each class are decreasing.
From this vantage point it is easy to apply the map $\varphi_t$ since
$\varphi_t h_{\lambda_i-\mu_j-i+j}$ vanishes unless $\lambda_i-i\equiv\mu_j-j\Mod{t}$.
Therefore, $\varphi_t s_{\lambda/\mu}$ has a block-diagonal structure, with
each block having size $m_r(\lambda;nt)\times m_r(\mu;nt)$ for $0\leq r\leq t-1$.
We conclude that $\varphi_t s_{\lambda/\mu}=0$ unless $m_r(\lambda;nt)=m_r(\mu;nt)$
for all $0\leq r\leq t-1$.
Assuming this is the case, then the entries of the of the minor corresponding
to the residue class $r$ are given by Lemma~\ref{Lem_phi}, and are
\[
h_{(\lambda_i-\mu_j-i+j)/t}=h_{\lar{r}_k-\mur{r}_\ell-k+\ell}
\]
for some $k$ and $\ell$ with $1\leq k,\ell\leq n$.
Note that the rows and columns are in the desired order (i.e., in each
$n\times n$ minor the indices increase from $1$ to $n$) thanks to the
permutations we applied at the beginning of the proof.
We have therefore shown that if $m_r(\lambda;nt)=m_r(\mu;nt)$ for all
$0\leq r\leq t-1$, then
\[
\varphi_t s_{\lambda/\mu}=\sgn_t(\lambda/\mu)\prod_{r=0}^{t-1}s_{\lar{r}/\mur{r}}.
\]
Now, if $\mur{r}\not\subseteq\lar{r}$ for any $r$ such that
$0\leq r\leq t-1$ this expression will give zero,
from which we conclude, by Lemma~\ref{Lem_tilable},
that $\varphi_t s_{\lambda/\mu}=0$ unless $\lambda/\mu$ is $t$-tileable.
\subsection{Proof of Theorem~\ref{Thm_AK_o}}
Let $\lambda$ be a partition such that $l(\lambda)\leq nt$ and consider the
definition \eqref{Eq_oh} of $\o_\lambda$
\[
\o_\lambda=\det_{1\leq i,j\leq nt}\big(h_{\lambda_i-i+j}-h_{\lambda_i-i-j}\big).
\]
We permute the rows and columns by $w_t(\lambda;nt)$ and $w_t(0;nt)$ respectively,
which introduces a sign of
\begin{equation}\label{Eq_first-sign}
(-1)^{\binom{n+1}{2}\binom{t}{2}}\sgn(w_t(\lambda;nt)).
\end{equation}
The modular behaviour of the indices of each row is now known.
There are three possibilities for the entries of $\varphi_t\o_\lambda$:
both $h$'s may survive, one $h$ may survive, or the entry is necessarily zero.
For both to survive, we see that
$h_{\lambda_i-i+j}$ and $h_{\lambda_i-i-j}$ are nonzero under $\varphi_t$
if and only if $\lambda_i-i\equiv-j\equiv0\Mod{t}$ or, if $t$ is even,
$\lambda_i-i\equiv-j\equiv t/2\Mod{t}$.
In the first instance, by Lemma~\ref{Lem_phi},
\[
\varphi_t \big(h_{\lambda_i-i+j}-h_{\lambda_i-i-j}\big)
=h_{\lar{0}_k-k+\ell+m_0(\lambda;nt)-n}-h_{\lar{0}_k-k-\ell+m_0(\lambda;nt)-n},
\]
where $1\leq k\leq m_0(\lambda;nt)$ and $1\leq \ell\leq n$.
Moreover, all other entries in the first $m_0(\lambda;nt)$ rows and $n$ columns
are zero.
If $t$ is even then we also find a submatrix of size $m_{t/2}(\lambda;nt)\times n$
in the rows $1+\sum_{r=0}^{(t-2)/2}m_r(\lambda;nt)$ to
$\sum_{r=0}^{t/2}m_r(\lambda;nt)$ and columns $1+nt/2$ to $n(t+2)/2$.
The entries of this submatrix are
\[
\varphi_t\big(h_{\lambda_i-i+j}-h_{\lambda_i-i-j}\big)
=h_{\lar{t/2}_k-k+\ell+m_{t/2}(\lambda;nt)-n}-
h_{\lar{t/2}_k-k-\ell+m_{t/2}(\lambda;nt)-n+1},
\]
where $1\leq k\leq m_{t/2}(\lambda;nt)$ and $1\leq\ell\leq n$.
Again, all other entries in these rows and columns are necessarily zero
under $\varphi_t$.
Given a row corresponding to the residue class $r$ where
$1\leq r\leq\lfloor(t-1)/2\rfloor$, there are two possibilities for the entry to
potentially survive: the column corresponds to the residue class $r$ or $t-r$.
Again, by Lemma~\ref{Lem_phi},
\[
\varphi_t\big(h_{\lambda_i-i+j}-h_{\lambda_i-i-j}\big)
=\begin{cases}
h_{\lar{r}_k-k+\ell+m_r(\lambda;nt)-n} & \text{if $j\equiv -r\Mod{t}$}, \\
-h_{\lar{r}_k-k-\ell+m_r(\lambda;nt)-n+1} & \text{if $j\equiv r\Mod{t}$}. \\
\end{cases}
\]
The set of indices of the complete homogeneous symmetric functions in
such a row are
\begin{subequations}\label{Eq_index-sets}
\begin{align}\label{Eq_index-set1}
&\big\{\lar{r}_k-k-\ell+m_r(\lambda,nt)+1\mid 1\leq \ell\leq 2n\big\}\\
&\,=\big\{\lar{r}_k-k+\ell\mid 1\leq \ell\leq m_r(\lambda,nt)\big\}
\sqcup\big\{\lar{r}_k-k-\ell+1\mid 1\leq\ell\leq m_{t-r}(\lambda,nt)\big\}.\notag
\intertext{If we look at the complementary row corresponding to $t-r$,
then a similar computation shows that the indices are}
&\big\{\lar{t-r}_k-k-\ell+m_{t-r}+1\mid 1\leq\ell\leq 2n\big\}
\label{Eq_index-set2}\\
&\,=\big\{\lar{t-r}_k-k+\ell\mid 1\leq \ell\leq m_{t-r}(\lambda,nt)\big\}
\sqcup\big\{\lar{t-r}_k-k-\ell+1\mid 1\leq\ell\leq m_r(\lambda,nt)\big\}.\notag
\end{align}
\end{subequations}
We have now identified the entries which do not necessarily vanish under
$\varphi_t$.
These can be rearranged into a block-diagonal matrix.
If $t$ is even, we move the submatrix corresponding to $t/2$ to the
bottom-right $m_{t-1}(\lambda;nt)$ rows and $n$ columns, which picks up a sign of
\[
(-1)^{m_{t/2}(\lambda;nt)\sum_{r=(t+2)/2}^{t-1}m_r(\lambda;nt)+n^2(t-2)/2}.
\]
We then group the rows and columns corresponding to the residue classes
$r$ and $t-r$ together with $0\leq r \leq \lfloor(t-1)/2\rfloor$ increasing.
The determinant is now block-diagonal and the blocks have dimension
$m_0(\lambda;nt)\times n$, $(m_r(\lambda;nt)+m_{t-r}(\lambda;nt))\times 2n$
for $1\leq r\leq \lfloor (t-1)/2\rfloor$ and, if $t$ is even, $m_{t/2}(\lambda;nt)
\times n$.
Since the determinant of a block-diagonal matrix vanishes if one of the
blocks is not a square, we can therefore conclude that $\varphi_t\o_\lambda$
vanishes unless the conditions \eqref{Eq_m_conditions} with $z=-1$ hold in
Lemma~\ref{Lem_AK_orth}, i.e., unless $t\textup{-core}(\lambda)$ is orthogonal.
In this case the top-left $n\times n$ minor is equal to
$\o_{\lar{0}}$ and if $t$ is even the bottom-right minor corresponds to
$\mathrm{so}_{\lar{t/2}}^-$.
Note that in this case the grouping of the $2n\times 2n$ minors does not
change the sign of the determinant since each row and column is pushed
past an even number of rows or columns.
For each $1\leq r\leq \lfloor(t-1)/2\rfloor$ these final minors are of the form
\[
\begin{pmatrix}
h_{\lar{r}_1+m_r-n} & \cdots & h_{\lar{r}_1+m_r-1}
& -h_{\lar{r}_1+m_r-n-1} & \cdots & -h_{\lar{r}_1+m_r-2n} \\
\vdots & & \vdots & \vdots & & \vdots \\
h_{\lar{r}_{m_r}+1-n} & \cdots & h_{\lar{r}_{m_r}}
& -h_{\lar{r}_{m_r}-n} & \cdots & -h_{\lar{r}_{m_r}-2n+1} \\
-h_{\lar{t-r}_1+m_{t-r}-n-1} & \cdots & -h_{\lar{t-r}_1+m_{t-r}-2n}
& h_{\lar{t-r}_1+m_{t-r}-n} & \cdots & h_{\lar{t-r}_1+m_{t-r}-1} \\
\vdots & & \vdots & \vdots & & \vdots \\
-h_{\lar{t-r}_{m_{t-r}}-n} & \cdots & -h_{\lar{t-r}_{m_{t-r}}-2n+1}
& h_{\lar{t-r}_{m_{t-r}}+1-n} & \cdots & h_{\lar{t-r}_{m_{t-r}}}
\end{pmatrix},
\]
where we write $m_r=m_r(\lambda;nt)$.
Clearing the negatives in this minor produces the sign
$(-1)^{m_r(\lambda;nt)+n}$.
If $m_r(\lambda;nt)=m_{t-r}(\lambda;nt)=n$ then we are done.
If $m_r(\lambda;nt)>n$ then we need to move the
columns $n+1$ to $m_r(\lambda;nt)$ so they are the first
$m_r(\lambda;nt)-n$ columns, and then
reverse the order. This gives a sign of
\[
(-1)^{n(m_r(\lambda;nt)-n)+\binom{m_r(\lambda;nt)-n}{2}}
=(-1)^{\binom{m_r(\lambda;nt)}{2}-\binom{n}{2}}.
\]
If $m_r(\lambda;nt)<n$ then we need to push the $m_{t-r}(\lambda;nt)-n$ missing rows
past the $n-m_r$ rows to their right and then reverse again, giving the same
sign
\[
(-1)^{(m_{t-r}(\lambda;nt)-n)m_r(\lambda;nt)+\binom{m_{t-r}(\lambda;nt)-n}{2}}
=(-1)^{\binom{m_r(\lambda;nt)}{2}-\binom{n}{2}}
\]
since $m_{t-r}(\lambda;nt)-n=n-m_r(\lambda;nt)$.
In each of the three cases the resulting determinant is equal to
$\mathrm{rs}_{\lar{r},\lar{t-r}}$.
Collecting all of the above determinant manipulations, the value of
$\varepsilon^{\o}_{\lambda;nt}$ is
\[
\frac{(n+1)nt(t-1)}{4}+
\sum_{r=1}^{\lfloor(t-1)/2\rfloor}
\bigg(\binom{m_r+1}{2}+\binom{n+1}{2}\bigg)
+\begin{cases}
n\sum_{r=1}^{(t-2)/2} m_{t-r}+\frac{n^2(t-2)}{2} &
\text{$t$ even}, \\
0 & \text{$t$ odd}.
\end{cases}
\]
To see that this agrees with the sign of Ayyer and Kumari, we use
\eqref{Eq_o-rk} together with the fact
that for odd $t$, $(t+1)(t-1)n(n+1)/4$ is even and for even $t$,
$n(n+1)(t^2-2)/4$ has the same parity as $n(n+1)/2$.
The above exponent therefore has the same parity as
\[
\varepsilon^{\o}_{\lambda;nt}
=\sum_{r=\lfloor(t+2)/2\rfloor}^{t-1}\binom{m_r(\lambda;nt)+1}{2}
+d(t\textup{-core}(\lambda))+\begin{cases}
\binom{n+1}{2}+nd(t\textup{-core}(\lambda)) & \text{if $t$ is even}, \\
0 & \text{if $t$ is odd}.
\end{cases}
\]
This completes the proof.
\subsection{Proof of Theorem~\ref{Thm_AK_sp}}\label{Sec_Proof3}
It is of course possible to prove Theorem~\ref{Thm_AK_sp} by direct
manipulation of the $h$ Jacobi--Trudi-type formula for $\sp_\lambda$
\eqref{Eq_sph}.
However, it will be more insightful to
begin with the $e$ Jacobi--Trudi-type formula \eqref{Eq_sp-e}
\[
\sp_{\lambda}=\det_{1\leq i,j\leq nt}\big(e_{\lambda_i'-i+j}-e_{\lambda_i'-i-j}\big),
\]
where we assume that $n$ is an integer such that $nt\geq \lambda_1$.
We further assume that $nt\geq l(\lambda)$, since, in the end, our sign
will be independent of $n$.
The values of $\varphi_t h_r$ and $\varphi_t e_r$ differ by a sign of
$(-1)^{(t-1)r/t}$, and the indices of the $e$'s in this formula are the same
as the $h$'s in the formula for $\o_{\lambda'}$ (\eqref{Eq_oh} with
$\lambda\mapsto\lambda'$), so we can simply replace each $h$
by a signed $e$ in the previous proof.
Moreover, by Corollary~\ref{Cor_reverse}, we know that the $t$-quotient of
$\lambda'$ is simply the reverse of the $t$-quotient of $\lambda$.
We can therefore already claim that $\varphi_t \sp_\lambda$ vanishes unless
$t\textup{-core}(\lambda)$ is symplectic, in which case
\[
\varphi_t\sp_\lambda
=(-1)^{\delta} \sgn(w_t(\lambda';nt))
\sp_{\lar{t-1}}\prod_{i=0}^{\lfloor(t-3)/2\rfloor}\mathrm{rs}_{\lar{i},\lar{t-i-2}}
\times\begin{cases}
\mathrm{so}_{\lar{(t-2)/2}} & \text{if $t$ is even}, \\
1 & \text{if $t$ is odd}
\end{cases}
\]
where $\delta$ has the value
\[
(t-1)\sum_{r=0}^{t-1}\abs{\lar{r}}+
\sum_{r=\lfloor(t+1)/2\rfloor}^{t-1}\binom{m_r(\lambda';nt)}{2}
+d(t\textup{-core}(\lambda))+\begin{cases}
\binom{n+1}{2}+nd(t\textup{-core}(\lambda)) & \text{if $t$ is even}, \\
0& \text{if $t$ is odd}.
\end{cases}
\]
All that remains now is to show that this sign agrees with that of
Theorem~\ref{Thm_AK_sp}.
By a combination of Lemmas~\ref{Lem_signs1} and \ref{Lem_signs2} we may replace
$\sgn(w_t(\lambda';nt))$ by $\sgn(w_t(\lambda;nt))$, which cancels
$d(t\textup{-core}(\lambda))+\sum_{r=0}^{t-1}\abs{\lar{r}}$ in $\delta$.
If we call this new exponent $\delta'$, then we
also have by Lemma~\ref{Lem_reverse} that $m_r(\lambda';nt)=m_{r-1}(\lambda;nt)$
for $\lfloor(t+1)/2\rfloor\leq r\leq t-1$, which implies
$\delta'=\varepsilon_{\lambda;nt}^{\sp}$.
\subsection{Proof of Theorem~\ref{Thm_AK_so}}
The final proof closely follows the first.
Let $\lambda$ be a partition of length at most $nt$ and consider
\[
\mathrm{so}_\lambda=\det_{1\leq i,j\leq nt}\big(h_{\lambda_i-i+j}+h_{\lambda_i-i-j+1}\big).
\]
As before we apply the permutations $w_t(\lambda;nt)$ and $w_t(0;nt)$ to the
rows and columns of this determinant, introducing the sign
\eqref{Eq_first-sign}.
Unlike before, there is only one case in which both $h$'s may survive.
If $t$ is odd and $\lambda_i-i\equiv-j\equiv (t-1)/2\Mod{t}$ then we have
\[
\varphi_t\big(h_{\lambda_i-i+j}+h_{\lambda_i-i-j+1}\big)
=h_{\lar{(t-1)/2}_k-k+\ell+m_{(t-1)/2}(\lambda;nt)-n}
+h_{\lar{(t-1)/2}_k-k-\ell+1+m_{(t-1)/2}(\lambda;nt)-n}.
\]
where $1\leq k\leq m_{(t-1)/2}(\lambda;nt)$ and $1\leq\ell\leq n$.
These entries lie in the rows $1+\sum_{r=0}^{(t-3)/2}m_r(\lambda;nt)$ to
$\sum_{r=0}^{(t-1)/2}m_r(\lambda;nt)$ and columns
$1+n(t-1)/2$ to $n(t+1)/2$, and outside of their intersection, all other entries
in these rows and columns are zero.
Now consider a row corresponding to the residue class $r$ for
$0\leq r\leq \lfloor(t-2)/2\rfloor$.
Then the column must fall into the residue class $r$ or $t-r-1$ in order for
the entry to not necessarily vanish.
In this case we now have
\[
\varphi_t\big(h_{\lambda_i-i+j}+h_{\lambda_i-i-j+1}\big)
=\begin{cases}
h_{\lar{r}_k-k+\ell_1+m_r(\lambda;nt)-n} & \text{if $j\equiv -r\Mod{t}$}, \\
h_{\lar{r}_k-k-\ell_1+m_r(\lambda;nt)-n+1} & \text{if $j\equiv r+1\Mod{t}$}. \\
\end{cases}
\]
Again, a similar computation holds for the row corresponding to $t-r-1$,
and the sets of indices agree with \eqref{Eq_index-sets} but with
$t-r\mapsto t-r-1$ in \eqref{Eq_index-set2}.
If $t$ is odd we move the central submatrix corresponding to $(t-1)/2$ to
the top-left, picking up a sign of
\[
(-1)^{m_{(t-1)/2}(\lambda;nt)\sum_{r=0}^{(t-3)/2}m_r(\lambda;nt)+n^2(t-1)/2}.
\]
The grouping and rearrangement of the remaining minors is the same as in the
first proof above. We only remark that the result is the determinant of a
block-diagonal matrix with blocks of dimensions
$(m_r(\lambda;nt)+m_{t-r-1}(\lambda;nt))\times 2n$ for
$0\leq r\leq \lfloor(t-2)/2\rfloor$ plus one of size
$m_{(t-1)/2}(\lambda;nt)\times n$ if $t$ is odd.
Thus the determinant vanishes unless
\eqref{Eq_m_conditions} holds with $z=0$, i.e., unless $t\textup{-core}(\lambda)$ is
self-conjugate.
Accounting for the sign of $(-1)^{\binom{m_r(\lambda;nt)}{2}+\binom{n}{2}}$
from reordering the columns in the copies of $\mathrm{rs}_{\lar{r},\lar{t-r-1}}$,
the exponent $\varepsilon^{\mathrm{so}}_{\lambda;nt}$ has the value
\[
\frac{(n+1)nt(t-1)}{4}
+\sum_{r=0}^{\lfloor(t-2)/2\rfloor}\bigg(\binom{m_r}{2}+\binom{n}{2}\bigg)
+\begin{cases}
0 & \text{if $t$ is even},\\
n\sum_{r=0}^{(t-3)/2}m_r+n^2(t-1)/2 & \text{if $t$ is odd}.
\end{cases}
\]
By \eqref{Eq_so-rk} this has the same parity as
\[
\varepsilon^{\mathrm{so}}_{\lambda;nt}=
\sum_{r=\lfloor (t+1)/2\rfloor}^{t-1}\binom{m_r(\lambda;nt)+1}{2}+
\begin{cases}
0 & \text{if $t$ is even}, \\
nd(t\textup{-core}(\lambda)) & \text{if $t$ is odd}.
\end{cases}
\]
\subsection{Proof of Proposition~\ref{Prop_signs}}
To close out this section, we sketch the proof of Proposition~\ref{Prop_signs}.
In the Schur case, by Corollary~\ref{Cor_reverse} and the fact that
$\lambda/\mu$ is $t$-tileable if and only if $\lambda'/\mu'$ is, we already have
$\omega\varphi_t s_{\lambda/\mu}=\pm\varphi_t\omega s_{\lambda/\mu}$.
The precise difference in sign is then provided by Lemma~\ref{Lem_signs1}.
In the orthogonal case, again using Corollary~\ref{Cor_reverse}, we have
\begin{align*}
\omega\varphi_t\o_\lambda
&=(-1)^{\varepsilon^{\o}_{\lambda;nt}} \sgn(w_t(\lambda;nt))
\sp_{(\lar{0})'}\prod_{i=1}^{\lfloor(t-1)/2\rfloor}\mathrm{rs}_{(\lar{i})',(\lar{t-i})'}
\times\begin{cases}
\mathrm{so}_{(\lar{t/2})'}^- & \text{if $t$ is even}, \\
1 & \text{if $t$ is odd},
\end{cases}\\
&=(-1)^{\varepsilon^{\o}_{\lambda;nt}+\varepsilon^{\sp}_{\lambda';nt}}
\sgn(w_t(\lambda;nt))\sgn(w_t(\lambda';nt))
\varphi_t\sp_{\lambda'},
\end{align*}
where $n$ should be large enough so that $\lambda$ is contained in an $nt\times nt$
box.
Combining Lemmas~\ref{Lem_signs0}--\ref{Lem_signs2} shows that, in this case,
\[
\sgn(w_t(\lambda;nt))\sgn(w_t(\lambda';nt))=(-1)^{d(t\textup{-core}(\lambda))
+(t-1)(\abs{\lar{0}}+\cdots+\abs{\lar{t-1}})}.
\]
A sign computation as in Section~\ref{Sec_Proof3} above shows that the
total sign agrees with the proposition, and is equal to
$(-1)^{(t-1)(\abs{\lar{0}}+\cdots+\abs{\lar{t-1}})}$.
The case of $\mathrm{so}_\lambda$ is nearly identical, and so it is omitted.
\section{Other factorisations}\label{Sec_stuff}
\subsection{Littlewood-type factorisations}
In \cite[\S7.3]{Littlewood40}, Littlewood proves a factorisation slightly
more general than the one contained in Theorem~\ref{Thm_skew} for $\mu$
empty; see also \cite[Theorem~2.7]{AK22}.
Here, and below, we let $\lar{r}=\lar{k}$ if $k\equiv r\Mod{t}$.
\begin{theorem}
Let $\lambda$ be a partition of length at most $nt+1$ and
$X=(x_1,\dots,x_n)$ a set of variables. Then for another variable $q$,
\[
s_\lambda(X,\zeta X,\dots,\zeta^{t-1}X,q)=0
\]
unless $t\textup{-core}(\lambda)=(c)$ for some $0\leq c\leq t-1$, in which case
\[
s_\lambda(X,\zeta X,\dots,\zeta^{t-1}X,q)=
\sgn_t(\lambda/(c))q^cs_{\lar{c-1}}(X^t,q^t)\prod_{\substack{r=0\\r\neq c-1}}^{t-1}
s_{\lar{r}}(X^t).
\]
\end{theorem}
This theorem can also be placed in our framework, however in a somewhat less
elegant manner than our other results.
The operator
$\varphi_t^q:\Lambda\longrightarrow \Lambda\otimes_{\mathbb{Z}}\mathbb{Z}[q]$
which gives the above may be defined by
\[
\varphi_t^q h_{at+b}:=q^{b}\sum_{k\geq 0}q^{kt} h_{a-k}
=\sum_{k\geq0}q^k\varphi_t h_{at+b-k}.
\]
Note that the sums are finite since $h_r$ vanishes for negative $r$, and that
for $q=0$ this reduces to the operator $\varphi_t$ from the earlier sections.
Alternatively, since
\[
h_r(X,q)=\sum_{k\geq0}q^k h_{r-k}(X),
\]
the image of $\varphi_t^q$ acting on any symmetric function $f$ is the same as
the image of $\varphi_t$ acting on $f(X,q)$, where $\varphi_t$ acts only
on the $X$ variables.
Using $\varphi_t^q$,
Littlewood's above theorem may be phrased as follows.
After the statement we provide a short proof which relies only
on Theorem~\ref{Thm_skew}.
\begin{proposition}
There holds $\varphi_t^q s_\lambda=0$ unless $t\textup{-core}(\lambda)=(c)$ for some
$c$ such that $0\leq c\leq t-1$, in which case
\[
\varphi_t^q s_\lambda=
\sgn_t(\lambda/(c))q^c\prod_{\substack{r=0\\r\neq c-1}}^{t-1}s_{\lar{r}}
\sum_{k\geq0}q^{kt}s_{\lar{c-1}/(k)}.
\]
\end{proposition}
\begin{proof}
The first observation is that
\[
\varphi_t^qs_\lambda=\sum_{k\geq 0} q^k\varphi_ts_{\lambda/(k)},
\]
which is a simple consequence of the branching rule for Schur functions
\cite[p.~72]{Macdonald95}.
In the case that $l(t\textup{-core}(\lambda))>1$
then each term in the sum on the right-hand side vanishes
by Theorem~\ref{Thm_skew} as the $t$-cores of the inner and outer shape
can never be equal.
Now assume $t\textup{-core}(\lambda)=(c)$ for some $0\leq c\leq t-1$, which is a complete
set of $t$-cores with length one.
Then the nonzero terms in the sum on the right-hand side are those for
which $k$ is of the form $\ell t+c$ with $\ell\geq0$ and $\ell t+c\leq
\lambda_1$.
Therefore
\[
\varphi_t^qs_\lambda=q^{c}\sum_{\ell\geq0}q^{\ell t}
\varphi_ts_{\lambda/(\ell t+c)}
=q^c\sum_{\ell\geq0}\sgn_t(\lambda/(\ell t+c))q^{\ell t}s_{\lar{c-1}/(\ell)}
\prod_{\substack{r=0\\r\neq c-1}}^{t-1}s_{\lar{r}},
\]
again by Theorem~\ref{Thm_skew}. By our convention we always
compute the $t$-quotient using a beta set with number of elements a multiple
of $t$. This means that the single row $(\ell t+c)$ has one non-empty
element in its $t$-quotient, $\lar{c-1}$.
Moreover, since the partitions $(\ell t+c)$ all differ by a ribbon of
height zero, the sign of each term in the sum is the same and equal to
$\sgn_t(\lambda/(c))$.
Putting all of this together, we arrive at
\[
\varphi_t^qs_\lambda=
\sgn_t(\lambda/(c))q^c\prod_{\substack{r=0\\r\neq c-1}}^{t-1}s_{\lar{r}}
\sum_{\ell\geq0}q^{\ell t}s_{\lar{c-1}/(\ell)}.\qedhere
\]
\end{proof}
We do not see, at this stage, whether it is possible to extend the previous
result to skew Schur functions.
If we expand $\varphi_t^q s_{\lambda/\mu}$ as in the proof above, we find that
\[
\varphi_t^q s_{\lambda/\mu}
=\sum_{\nu\succ\mu}q^{\abs{\nu}-\abs{\mu}}\varphi_t s_{\lambda/\nu},
\]
where $\nu\succ\mu$ means that $\nu/\mu$ is a \emph{horizontal strip},
i.e., $\nu\supseteq\mu$ and $\nu/\mu$ contains at most one box in each column
of its Young diagram.
Of course, this implies that $\varphi_t^qs_{\lambda/\mu}=0$ if there does not
exist a $\nu$ such that $\nu\succ\mu$ and $\lambda/\nu$ is $t$-tileable.
However, the sum may vanish even if such a $\nu$ exists.
For example,
\begin{align*}
\varphi_2^qs_{(4,4)/(1)}&=q\varphi_t\big(s_{(4,4)/(2)}+s_{(4,4)/(1,1)}\big)
+q^3\varphi_t\big(s_{(4,4)/(4)}+s_{(4,4)/(3,1)}\big) \\
&=q(s_{(2)}s_{(2)/(1)}-s_{(2)}s_{(2)/(1)})
+q^3(s_{(2)}-s_{(2)}) \\
&=0.
\end{align*}
In a similar direction Pfannerer \cite[Theorem~4.4]{Pfannerer22} has shown
that, if $\lambda$ has empty $t$-core and $m=\ell t+k$ is any integer, then the
Schur function $s_\lambda(1,\zeta,\dots,\zeta^{m-1})$ always factors as a product
of Schur functions with variables all one indexed by the $t$-quotient of $\lambda$.
When $m$ is a multiple of $t$ this becomes a special case of Littlewood's
theorem (\ref{Thm_skew} with $\mu$ empty) noted by Reiner, Stanton and White
\cite[Theorem~4.3]{RSW04}.
Pfannerer's result does not include a statement about vanishing and, indeed,
the Schur function he considers does not appear to behave nicely when
$\lambda$ has nonempty $t$-core.
It is an open problem to see how this factorisation fits into our story.
\subsection{Factorisations of supersymmetric Schur functions}
Recently, Kumari has given a version of Theorem~\ref{Thm_skew} for the
so-called \emph{skew hook Schur functions} (or
\emph{supersymmetric skew Schur functions}) \cite[Theorem~3.2]{Kumari22}.
For two independent sets of variables (\emph{alphabets}), we denote their
plethystic difference by $X-Y$; see, e.g.,
\cite{Haglund08,Lascoux03} for the necessary background on plethystic
notation. We also note that for an alphabet $X$, we let $\varepsilon X$ be
the alphabet with all variables negated.
The \emph{complete homogeneous supersymmetric function} used in
\cite{Kumari22} may be defined as\[
\sum_{j=0}^r h_j(X)e_{r-j}(Y)=h_r[X-\varepsilon Y].
\]
The hook Schur function is then the Jacobi--Trudi determinant of these
functions, so that
\[
s_{\lambda/\mu}[X-\varepsilon Y]=\det_{1\leq i,j\leq n}\big(
h_{\lambda_i-\mu_j-i+j}[X-\varepsilon Y]\big).
\]
From this, it follows readily that Kumari's factorisation for the
hook Schur functions is contained in Theorem~\ref{Thm_skew} above at the
alphabet $X-\varepsilon Y$.
\subsection{Factorisations of $\mathrm{rs}_{\lambda,\mu}$}\label{Sec_CK}
To close, we point out that the universal character $\mathrm{rs}_{\lambda,\mu}$ can be
used to lift some factorisation results, discovered by Ciucu and Krattenthaler
\cite[Theorems~3.1--3.2]{CK09} and subsequently generalised by
Ayyer and Behrend \cite[Theorems~1--2]{AB19}, to the universal character level.
In the next result we write $\lambda+1^n=(\lambda_1+1,\dots,\lambda_n+1)$
where $n\geq l(\lambda)$.
\begin{theorem}\label{Thm_CK}
For $\lambda$ a partition of length at most $n$, there holds
\begin{subequations}\label{Eq_CK}
\begin{align}\label{Eq_CK1}
\mathrm{rs}_{\lambda,\lambda}&=\mathrm{so}_\lambda\mathrm{so}_\lambda^-,
\intertext{and}
\mathrm{rs}_{\lambda+1^n,\lambda}&=\o_{\lambda+1^n}\sp_\lambda.\label{Eq_CK2}
\end{align}
\end{subequations}
Moreover, for $\lambda$ a partition of length at most $n+1$,
\begin{subequations}\label{Eq_CK'}
\begin{align}
\mathrm{rs}_{(\lambda_1,\dots,\lambda_n),(\lambda_2,\dots,\lambda_{n+1})}
+\mathrm{rs}_{(\lambda_1-1,\dots,\lambda_{n+1}-1),(\lambda_2+1,\dots,\lambda_n+1)}
&=\sp_{(\lambda_1,\dots,\lambda_n)}\o_{(\lambda_2,\dots,\lambda_{n+1})},\label{Eq_CK3}
\intertext{and}
\mathrm{rs}_{(\lambda_1+1,\dots,\lambda_n+1),(\lambda_2,\dots,\lambda_{n+1})}
+\mathrm{rs}_{(\lambda_1,\dots,\lambda_{n+1}),(\lambda_2+1,\dots,\lambda_n+1)}
&=\mathrm{so}_{(\lambda_1+1,\dots,\lambda_n+1)}\mathrm{so}_{(\lambda_2,\dots,\lambda_{n+1})}^-.\label{Eq_CK4}
\end{align}
\end{subequations}
\end{theorem}
To get back to the results of Ayyer and Behrend one simply evaluates
both sides of each equation at the alphabet $(x_1^\pm,\dots,x_n^\pm)$.
The precise forms present in \cite[Equations~(18)--(21)]{AB19} then follow
from \eqref{Eq_det-rep}.\footnote{The factor of $(1+\delta_{0,\lambda_{n+1}})$
in \cite[Equation~(20)]{AB19} is not present in our generalisation
\eqref{Eq_CK3} since the second character vanishes if $\lambda_{n+1}=0$.}
As identities for Laruent polynomials the pairs of identities
\eqref{Eq_CK} and \eqref{Eq_CK'} admit uniform statements.
However no such uniform statement will exist for the above generalisation,
since this requires characters indexed by half-partitions, which cannot be
handled by the universal characters.
\begin{proof}[Proof of Theorem~\ref{Thm_CK}]
First up is \eqref{Eq_CK1}, which is the simplest of the four.
In the determinant
\[
\mathrm{rs}_{\lambda,\lambda}=
\det_{1\leq i,j\leq 2n}
\begin{pmatrix}
\big(h_{\lambda_i-i+j}\big)_{1\leq i,j\leq n} &
\big(h_{\lambda_i-i-j+1}\big)_{\substack{1\leq i,j\leq n}} \\[2mm]
\big(h_{\lambda_i-i-j+1}\big)_{\substack{1\leq i,j\leq n}} &
\big(h_{\lambda_i-i+j}\big)_{1\leq i,j\leq n}
\end{pmatrix},
\]
add the blocks on the right to the blocks on the left, and then subtract the
blocks on the top from the blocks on the bottom, giving
\begin{align*}
\mathrm{rs}_{\lambda,\lambda}&=
\det_{1\leq i,j\leq 2n}
\begin{pmatrix}
\big(h_{\lambda_i-i+j}+h_{\lambda_i-i-j+1}\big)_{1\leq i,j\leq n} &
\big(h_{\lambda_i-i-j+1}\big)_{\substack{1\leq i,j\leq n}} \\[2mm]
0 &
\big(h_{\lambda_i-i+j}-h_{\lambda_i-i+j}\big)_{1\leq i,j\leq n}
\end{pmatrix} \\
&=\mathrm{so}_\lambda\mathrm{so}_\lambda^-.
\end{align*}
For the second identity \eqref{Eq_CK2},
\[
\mathrm{rs}_{\lambda+1^n,\lambda}=
\det_{1\leq i,j\leq 2n}
\begin{pmatrix}
\big(h_{\lambda_i-i+j+1}\big)_{1\leq i,j\leq n} &
\big(h_{\lambda_i-i-j+2}\big)_{\substack{1\leq i,j\leq n}} \\[2mm]
\big(h_{\lambda-i-j+1}\big)_{\substack{1\leq i,j\leq n}} &
\big(h_{\lambda_i-i+j}\big)_{1\leq i,j\leq n}
\end{pmatrix},
\]
and we add columns $1, \dots, n-1$ to the columns $n+2,\dots,2n$
and then subtract the bottom two blocks from the top two, resulting in
\begin{align*}
\mathrm{rs}_{\lambda+1^n,\lambda}&=
\frac{1}{2}\det_{1\leq i,j\leq 2n}
\begin{pmatrix}
\big(h_{\lambda_i-i+j+1}-h_{\lambda_i-i-j+1}\big)_{1\leq i,j\leq n} &
0 \\
\big(h_{\lambda-i-j+1}\big)_{\substack{1\leq i,j\leq n}} &
\big(h_{\lambda_i-i+j}+h_{\lambda_i-i-j+2}\big)_{1\leq i,j\leq n}
\end{pmatrix} \\[2mm]
&=\o_{\lambda+1^n}\sp_\lambda.
\end{align*}
In the third identity, we consider the second determinant in the sum
in \eqref{Eq_CK3}
\begin{align*}
\det_{1\leq i,j\leq 2n}
\begin{pmatrix}
\big(h_{\lambda_i-i+j-1}\big)_{1\leq i,j\leq n+1} &
\big(h_{\lambda_i-i-j}\big)_{\substack{1\leq i\leq n+1\\ 1\leq j\leq n-1}} \\[3mm]
\big(h_{\lambda_{i+1}-i-j+2}\big)_{\substack{1\leq i\leq n-1\\1\leq j\leq n+1}} &
\big(h_{\lambda_{i+1}-i+j+1}\big)_{1\leq i,j\leq n-1}
\end{pmatrix}.
\end{align*}
Push the first column so it becomes the $(n+1)$-st, and then push
the $(n+1)$-st row to the final row, which picks up a minus sign.
The resulting determinant differs from that of
$\mathrm{rs}_{(\lambda_1,\dots,\lambda_n),(\lambda_2,\dots,\lambda_{n+1})}$
in only the last row, so we can take the sum of the two, giving
\begin{align*}
\det_{1\leq i,j\leq 2n}
\begin{pmatrix}
\big(h_{\lambda_i-i+j}\big)_{1\leq i,j\leq n} &
\big(h_{\lambda_i-i-j+1}\big)_{\substack{1\leq i\leq n}} \\[2mm]
\big(h_{\lambda_{i+1}-i-j+1}\big)_{\substack{1\leq i\leq n-1\\1\leq j\leq n}} &
\big(h_{\lambda_{i+1}-i+j}\big)_{\substack{1\leq i\leq n-1\\1\leq j\leq n}}\\[3mm]
(h_{\lambda_{n+1}-n-j+1}-h_{\lambda_{n+1}-n+j-1})_{1\leq j\leq n} &
(h_{\lambda_{n+1}-n+j}-h_{\lambda_{n+1}-n-j})_{1\leq j\leq n}
\end{pmatrix}.
\end{align*}
In this new determinant, add columns $n+1,\dots,2n-1$ to columns $2,\dots,n$,
and then subtract rows $2,\dots,n$ from rows $n+1,\dots,2n-1$, which gives
\begin{align*}
\frac{1}{2}
\det_{1\leq i,j\leq 2n}
\begin{pmatrix}
\big(h_{\lambda_i-i+j}+h_{\lambda_i-i-j+2}\big)_{1\leq i,j\leq n} &
\big(h_{\lambda_i-i-j+1}\big)_{\substack{1\leq i\leq n}} \\[2mm]
0 &
\big(h_{\lambda_{i+1}-i+j}-h_{\lambda_{i+1}-i-j}\big)_{1\leq i,j\leq n}\\[3mm]
\end{pmatrix},
\end{align*}
which equals
$\sp_{(\lambda_1,\dots,\lambda_n)}\o_{(\lambda_2,\dots,\lambda_{n+1})}$.
The final factorisation \eqref{Eq_CK4} follows almost identically.
\end{proof}
\subsection*{Acknowledgements}
I would like to thank Christian Krattenthaler for many useful comments,
corrections and suggestions.
This work also benefitted from discussions with
Ilse Fischer, Hans H\"ongesberg, Josef K\"ustner, Florian Schreier-Aigner
and Ole Warnaar.
|
\section{Lossless Compression}
\label{app:lossless}
The goal of lossless compression is to encode samples from a discrete probably distribution $x\sim p_d(x)$ $x\in \mathcal{X}$ as bit strings $m=\texttt{enc}(x) \in \{0,1\}^*$ of shortest possible length $\ell(m)$ such that $x$ can be decoded without loss of information $x = \texttt{dec}(m)$.
Block codes assign a unique binary code of equal length to each event in $\mathcal{X}$. The expected length per symbol of such codes is $l(m)=\lceil \log_2 |\mathcal{X}| \rceil$ bits.
We can improve this result if we consider to give shorter code words to more frequently occurring symbols and longer codes to less frequently occurring ones. In fact building on this idea, it can be shown that the smallest expected achievable average code length per point using an approximation to the true data distribution $p(x)$, is given by the \emph{cross-entropy}:
\begin{align}
H(p_d, p) := \mathbb{E}_{x \sim p_d} \left[ -\log p(x)\right].
\end{align}
This quantity is bounded from below (seen via Jensen's inequality) by the \emph{entropy} of the ``true'' probability distribution $p_d$, i.e.\ when $p=p_d$ \citep{mackay2003information, cover2012elements}.
Entropy coders reach this bound up to a small constant $\epsilon$. An entropy codec is a tuple of an inverse pair of functions, $\texttt{enc}_{p_{X}}$ and $\texttt{dec}_{p_{X}}$ that achieve near optimal compression rates on sequences of symbols.
\begin{equation}
\begin{aligned}
&\texttt{enc}_{p}:m, x\mapsto m'\\
&\texttt{dec}_{p}:m'\mapsto (m, x),
\end{aligned}
\end{equation}
such that $\ell(m')=\ell(m)+\log_2 p(x) + \epsilon $.
Specifically in this paper, we rely on \emph{Arithmetic coding}, a type of stream coder that represents data by means of a subinterval of the unit interval $[0,1]$ of the real line.
\clearpage
\section{Qualitative Samples}
\label{app:qual_samples}
\begin{figure}[h!]
\begin{subfigure}[t]{0.5\linewidth}
\centering
\includegraphics[width=\linewidth, height=0.35\paperheight]{figs/qual_samples/orig_38.png}
\caption{Original}
\end{subfigure}
\hfill
\begin{subfigure}[t]{0.5\linewidth}
\centering
\includegraphics[width=\linewidth, height=0.35\paperheight]{figs/qual_samples/38_0.1653120219707489.jpg}
\caption{\textsc{PQ-MIM}\xspace (bpp=0.165)}
\end{subfigure}
\end{figure}
\begin{figure}[h!]
\begin{subfigure}[t]{0.5\linewidth}
\centering
\includegraphics[width=\linewidth, height=0.35\paperheight]{figs/qual_samples/orig_6.png}
\caption{Original}
\end{subfigure}
\hfill
\begin{subfigure}[t]{0.5\linewidth}
\centering
\includegraphics[width=\linewidth, height=0.35\paperheight]{figs/qual_samples/6_0.12431618571281433.png}
\caption{\textsc{PQ-MIM}\xspace (bpp=0.124)}
\end{subfigure}
\end{figure}
\begin{figure}[h!]
\begin{subfigure}[t]{0.5\linewidth}
\centering
\includegraphics[width=\linewidth, height=0.35\paperheight]{figs/qual_samples/orig_78.png}
\caption{Original}
\end{subfigure}
\hfill
\begin{subfigure}[t]{0.5\linewidth}
\centering
\includegraphics[width=\linewidth, height=0.35\paperheight]{figs/qual_samples/78_0.15514478087425232.png}
\caption{\textsc{PQ-MIM}\xspace (bpp=0.155)}
\end{subfigure}
\end{figure}
\begin{figure}
\begin{subfigure}[t]{0.5\linewidth}
\centering
\includegraphics[width=\linewidth, height=0.35\paperheight]{figs/qual_samples/orig_82.png}
\caption{Original}
\end{subfigure}
\hfill
\begin{subfigure}[t]{0.5\linewidth}
\centering
\includegraphics[width=\linewidth, height=0.35\paperheight]{figs/qual_samples/82_0.17619465291500092.png}
\caption{\textsc{PQ-MIM}\xspace (bpp=0.176)}
\end{subfigure}
\end{figure}
\section{Conclusion}
\label{sec:conc}
In this paper, we have revisited vector quantization for neural image compression. We introduced a product-quantization variants of VQ-VAE and shown that it has a better scaling properties in terms of bit-rate. Additionally, we introduced a novel conditional entropy model based on masked image modeling. We have shown that combined with the qunicunx partitioning pattern, \textsc{PQ-MIM}\xspace provides strong reduction in bit-rate. \textsc{PQ-MIM}\xspace exhibits a competitive performance for PSNR and MS-SSIM metric compared to strong neural compression baselines. Furthermore, when we train \textsc{PQ-MIM}\xspace using perceptual losses, it provides a strong performance on multiple metrics (e.g. FID, KID and MS-SSIM) compared the strong baseline of HiFiC. Finally, we have shown that \textsc{PQ-MIM}\xspace can operate in a hybrid compression/generation mode where it can fill the gaps for non-transmitted patches.
\section{Experiments}
\label{sec:exp}
\begin{figure*}
\vspace{-5mm}
\begin{subfigure}[t]{0.32\linewidth}
\caption{FID}
\includegraphics[trim={0.0cm, 2.0cm, 1.2cm, 0.0cm}, clip, width=1.0\linewidth]{figs/rd_plots/fid.pdf}
\end{subfigure}
\hfill
\begin{subfigure}[t]{0.32\linewidth}
\centering
\caption{KID}
\includegraphics[trim={0.0cm, 2.0cm, 1.2cm, 0.0cm}, clip, width=1.04\linewidth]{figs/rd_plots/kid.pdf}
\end{subfigure}
\hfill
\begin{subfigure}[t]{0.32\linewidth}
\centering
\caption{MS-SSIM}
\includegraphics[trim={0.0cm, 2.0cm, 1.2cm, 0.0cm}, clip, width=1.0\linewidth]{figs/rd_plots/msssim.pdf}
\end{subfigure}
\caption{\textbf{Perceptual training and evaluation using \clic2020 test-set.} Performance of our adversarially trained \textsc{PQ-MIM}\xspace w.r.t percpetual metrics compared to HiFiC \cite{mentzer20nips}. \textsc{PQ-MIM}\xspace provides a stronger performance on FID, KID and MS-SSIM across all reported operating points.}
\label{fig:fid}
\end{figure*}
We first present our experimental setup in \sect{setup} and the present our results in \sect{results}.
We provide ablation studies in Section \ref{sec:ablations}.
We will share our code and models.
\subsection{Experimental setup}
\label{sec:setup}
\noindent \textbf{Rate-distortion control.} For all our experiments we fix the codebook size $V=256$ and only vary the number of sub-vectors $M \in \{2, 4, 6\}$ for two different down-sampling factors $f \in \{8,16\}$.
\mypar{PQ-VAE implementation details.}
Our PQ-VAE training uses the straight-through estimator \citep{bengio2013estimating} to propagate gradients through the quantization bottleneck. As for the quantization, the elements of the latent representation $\mathbf{z}$ are first linearly projected to a low dimensional look-up vector (dim=8 per sub-vector) followed by $\ell_{2}$ normalization, following~\citet{yu21iclr}.
We train our model using ImageNet \citep{deng09cvpr} for 50 epochs with a batch size of 256. We use an AdamW \citep{loshchilov2017decoupled} optimizer with a peak learning rate of $1.10^{-3}$, weight decay of $0.02$ and $\beta_{2}=0.95$. We apply a linear warmup for the first 5 epochs of training followed by a cosine decay schedule for the remaining 45 epochs to a minimum learning rate of $5.10^{-5}$. Unless mentioned otherwise, for all experiments, the encoder and decoder use an XCiT-L6 with 6 layers and hidden dimension of 768. We use sinusoidal positional embedding \citep{vaswani17nips} such that our model can flexibly operate on variable sized images.
For the results reported in Figure~\ref{fig:results_psnr_msssim}, the models are trained using solely MSE ($\eta=0.5$) or MS-SSIM ($\eta=10.0$) distortion losses for their corresponding plots. As for the models trained with perceptual objectives (Figure~\ref{fig:fid} and Table~\ref{tab:discriminator}), they are trained with a weighted sum of MSE, LPIPS ($\alpha=1$) and adversarial loss ($\gamma=0.1$). We use a ProjectedGAN Discriminator~\citep{sauer2021projected} architecture. The perceptual training is initialized with an MSE only trained checkpoint and trained for 50 epochs using ImageNet with a learning rate of $10^{-4}$ and weight decay of $5.10^{-5}$. Similar to HiFiC~\citep{mentzer20nips}, we freeze the encoder during the perceptual training. Additionally, we find that clipping the gradient norm to a maximum value of $4.0$ improves the training stability. Our discriminator takes only the decoded image as an input and does not rely on any other conditional signal.
\begin{figure}
\centering
\vspace{-5mm}
\includegraphics[width=\linewidth]{figs/qual_samples/extreme_compression.pdf}
\caption{\textbf{Extreme Image Compression.} \textsc{PQ-MIM}\xspace exhibits non-trivial compression performance at the extreme compression regime (e.g. 0.03 bpp), leading to compressed image codes that can fit in a short tweet (280 characters).}
\label{fig:extreme_compression}
\end{figure}
\mypar{MIM implementation details.} Our MIM module is an XCiT-L12 with 12 layers and embedding dimension of 768. MIM is trained simultaneously with the PQ-VAE, but the gradients are not backpropagated to the encoder or the quantizer parameters.
The PQ indicies $q$ are split into inputs and targets for the MIM model as defined by the quincunx partitioning pattern.
By default we use $S=5$ stages. All stages are processed with the same MIM model.
The masked patches are replaced by a learnable ``mask'' token embedding.
The loss is computed only for the masked patches. Since every token is assigned $M$ PQ indices, the output of the MIM transformer is passed to $M$ separate linear heads to predict $M$ softmax normalized distributions over their corresponding codebooks.
The marginal distribution of the codebook is computed as a normalized histogram over the ImageNet training set. For entropy coding, we use the implementation of the \texttt{torchac}\footnote{\scriptsize \url{https://github.com/fab-jul/torchac}} arithmetic coder.
\mypar{Datasets.} We train our models using ImageNet~\citep{deng09cvpr}. For data augmentation, we apply random resized cropping to \myres{256} images and horizontal flipping.
For evaluation and comparison to prior work, we use \textsc{Kodak}\xspace~\citep{kodak1993kodak} and \textsc{Tecnick}\xspace~\citep{tecnick} datasets for PSNR and MS-SSIM. Moreover, we compute the perceptual metrics (FID~\citep{heusel17nips}, KID~\citep{binkowski2018demystifying}) for perceptually trained models using the \clic2020 test-set~\citep{CLIC2020} (428 images) using the same patch cropping scheme detailed by \cite{mentzer20nips}.
\mypar{Baselines.}
We compare to several existing neural compression baselines: the scale hyperprior model~\citep{balle18iclr}, mean \& scale hyperprior~\citep{minnen18nips}, GMM hyperprior~\citep{cheng20cvpr}, and HiFiC~\citep{mentzer20nips}.
Among the non-neural codecs, we compare to the popular BPG \citep{ballard_bpg}, WebP, and JPEG2000.
\subsection{Main experimental results}
\label{sec:results}
\noindent \textbf{Comparison to existing (neural) codecs.}
We compare \textsc{PQ-MIM}\xspace to other approaches across a wide range of bitrates in Figure~\ref{fig:results_psnr_msssim}\footnote{All results for the baselines are reported using the authors official repositories}.
Note that in our evaluation we consider bitrates that are an order of magnitude lower than what is typically studied in the literature: most previous studies were limited to 0.1 bpp and above, see \eg~\citep{balle18iclr,cheng20cvpr,minnen18nips}.
The extremely low bitrates we consider make it possible to transmit a \myres{256} image in an SMS or a tweet (280 characters)\footnote{For example, a bitrate of 0.03 yields $256\times 256 \times 0.03 / 8\approx 246$ bytes for a \myres{256} image.} as shown in Figure ~\ref{fig:extreme_compression}. \textsc{PQ-MIM}\xspace achieves a strong and competitive performance for both \textsc{Kodak}\xspace and \textsc{Tecnick}\xspace datasets, outperforming all prior neural and standard codecs with the exception of GMM hyperprior~\citep{cheng20cvpr}. We observe that \textsc{PQ-MIM}\xspace is particularly strong for low rates, making it a good fit for extreme compression scenarios. Moreover, \textsc{PQ-MIM}\xspace exhibits a particularly strong MS-SSIM performance which was designed to model the human visual contrast perception \citep{wang2004image, wang2003multiscale}.
\mypar{Perceptual metrics comparison.} In Figure~\ref{fig:fid}, we compare \textsc{PQ-MIM}\xspace to HiFiC~\citep{mentzer20nips} in perceptual quality measures like FID and KID as well as MS-SSIM.
HiFiC is based on the mean \& scale hyperprior model~\citep{minnen18nips}, but adds adversarially trained discriminator model to improve the perceptual quality of the image reconstructions. \textsc{PQ-MIM}\xspace, with perceptual training, achieves a strong performance for all reported metrics, outperforming HiFiC for all operating points.
\begin{table*}[t!]
\smallskip
\begin{minipage}[b]{0.48\linewidth}
\centering
\caption{\textbf{Discriminator Architecture}.
We investigate multiple discriminator including StyleGAN \citep{stylegan}, ProjectedGAN \citep{sauer2021projected} and UNet \citep{schonfeld2020u}. \label{tab:heads}}
\scalebox{0.85}{
\begin{tabular}{l c c c }
\toprule
Discriminator & FID $\downarrow$ & KID $\downarrow$ & MS-SSIM $\uparrow$ \\
\midrule
None & 26.1 & 1.2$\times{-2}$ & 15.3 \\
\midrule
StyleGAN & \textbf{3.57} & \textbf{4.8$\times10{-4}$} & 13.3\\
ProjectedGAN & 3.69 & 6.6$\times10{-4}$ & \textbf{14.1} \\
UNet & 3.87 & 6.7$\times10{-4}$ & 13.9 \\
\bottomrule
\label{tab:discriminator}
\end{tabular}}
\end{minipage}%
\hfill
\begin{minipage}[b]{0.5\linewidth}
\centering
\caption{\textbf{Different MIM masking policies.} \textsc{PQ-MIM}\xspace with quincunx pattern with 5 steps reduces the bpp significantly (27\%). Additionally, \textsc{PQ-MIM}\xspace is orders of magnitude cheaper in terms of FLOPs compared to an autoregressive raster order masking pattern. \label{tab:masking_policy}}
\scalebox{0.9}{
\begin{tabular}{l c c c}
\toprule
Masking policy & \#steps & bpp & MACs/Pixel (M)\\
\midrule
Marginal baseline & 1 & 0.512 & 0.69 \\
Raster order & $T$ & \texttt{OOM} & 24.4$\times10^3$ \\
Quincunx & 5 & 0.373 & 6.66 \\
\bottomrule
\end{tabular}}
\end{minipage}%
\end{table*}
\subsection{Analysis and Ablations}
\label{sec:ablations}
\noindent \textbf{Model size and architecture.}
In Figure~\ref{fig:autoencoder_capacity}, we analyze the effect of using XCiT of different capacities for the autoencoder with respect to rate-distortion trade-off. We observe that the performance improves with higher capacity autoencoders, but their is a diminishing return with further increase in capacity. For all our experiments we use an XCiT-L6 since it achieves the best performance.
\mypar{Masking patterns.}
In Table~\ref{tab:masking_policy}, we compare to predicting tokens one-by-one autoregressively in a raster-scan order, the same pattern used in VQ-VAE based generative image models such as DALL-E~\citep{ramesh21arxiv} and VQ-GAN~\citep{esser21cvpr}. In contrast to \textsc{PQ-MIM}\xspace, raster-scan
models require causal attention, which makes XCiT not a good fit. We use a standard ViT model instead. However, due to the quadratic complexity of ViT and the high resolution of images typically used for evaluation of compression method (e.g. \textsc{Tecnick}\xspace), our autoregressive variant consistently exceeded the memory limits, even when using \texttt{A100} GPUs with 40GB memory. Moreover, raster scan fully-autoregressive models results in extremely expensive FLOP count since it needs $T$ separate forward passes per image.
On the other hand, our stage-wise MIM with quincunx pattern requires 4 evaluations, and does not scale with the image resolution as is the case for raster, making it a more practical solution.
\mypar{Number of prediction stages.}
We compare the quincunx masking pattern with masking based on confidence score following MaskGiT~\citep{chang22arxiv} patch selection procedure for image generation.
In the latter case, at a given step, we pick the patches to transmit dynamically based on their confidence score.
The confidence is defined as the maximum across the probabilities assigned over the vocabulary by the model.\footnote{
Note that for decoding the same confidence score can be used to identify the group of tokens to decode.}
For the quincunx and confidence based masking policy we use the same 5-stage scheme, in which the number of tokens in subsequent groups doubles in size. We ablate the number of prediction stages $S$ for the quincunx and confidence-based sampling.
For the latter we consider two options: (i) a linear scheme where each group of tokens contains (approximately) the same number of tokens; and (ii) a doubling scheme where each subsequent group of tokens is double the size of the previous group, as is also used for the quincunx pattern.
Every point on the curves in \fig{mim_steps} corresponds to the bitrate when encoding/decoding with a given number of steps.
For example, $S=2$ steps means we have only two steps each encoding/decoding 50\% of the patches. For the 3-steps doubling schedule the groups have sizes of 1/4, 1/4, 1/2, and so on.
For all three patterns the bitrate monotonically decreases with the number of steps, as more tokens can be predicted from larger contexts.
On the one hand, using confidence-based masking patterns, doubling and linear schemes lead to mostly comparable bitrates for the same number of steps, with further improvement for linear with more steps however with diminishing return. On the other hand, quincunx provides a stronger reduction ratio with strictly 5 steps, even when compared with confidence-based masking with higher number of steps.
\begin{figure*}[t!]
\vspace{-2mm}
\begin{minipage}[t]{0.48\linewidth}
\centering
\includegraphics[width=1.0\linewidth]{figs/ablations/ae_capacity.pdf}
\caption{\textbf{Autoencoder capacity.} The RD performance of different encoder/decoder capacities. We use the same model trunk for (i) the encoder; (ii) the PQ-MIM entropy model ; and (iii) the decoder. Increasing the model size from a XCiT-T6 model (3.5M params) to XCiT-L6 (47M params) increases the performance by typically +0.8dB.} %
\label{fig:autoencoder_capacity}
\end{minipage}
\hfill
\begin{minipage}[t]{0.5\linewidth}
\centering
\includegraphics[width=1.0\linewidth]{figs/ablations/mim_steps.pdf}
\caption{\textbf{MIM number of steps.} In addition to quincunx, we explore using the prediction confidence as the patch masking policy following MaskGiT~\citep{chang22arxiv}. We test a linear and doubling schedules. Quincunx provides a higher rate saving, even compared to confidence policy with a longer schedule.}
\label{fig:mim_steps}
\end{minipage}
\end{figure*}
\mypar{Predicting missing patches.}
When encoding/decoding the discrete image representation $\mathbf{q}$, we reduce the bitrate with our models by their ability to predict the remaining tokens given those of preceding stages.
To illustrate the predictive ability of our model, we consider an experiment where we remove a subset of the tokens (sampled randomly), and use our model to fill in the missing patches by conditioning on the observed ones.
We then use the PQ-VAE decoder to decode the discrete latent codes including the filled in ones.
In \fig{inpaiting_rand_bitstream} we show the results obtained when removing 10\%, 20\%, 30\% and 50\% of the patches.
Our MIM is able to model redundancies among the patches, which corroborates our findings of its compression ability.
\subsection{Limitations}
Image compression, and compression of visual data in general, is an important technology to scale the distribution of visual data. This is ever more important to cope with the growing quantity of visual data that is streamed in the form of video and for augmented and virtual reality applications.
Compression is also critical to allow users with low-bandwidth connections to benefit from applications relying on sharing of image, video, or virtual reality data.
\mypar{Caveats of learned neural compression models: biases and performance. }
As with any machine learning model, potential biases in the training data may be transferred into the model via training.
In our setting, this can affect the autoencoding reconstruction abilities for content under-represented in the training data, as well as the compression abilities of the model for such data. Such biases should be assessed before deployment of the model.
Beyond rate-distortion trade-offs, important evaluation dimensions include the energy and latency performance of compression models.
Current neural compression methods, including ours, need to be further optimized to be competitive with existing codecs on these criteria.
\mypar{Specific limitations of our approach. }
In our work, we specifically focus on a high compression regime, with compression rate lower than 0.6 bit per pixel. This is a favorable case for our approach: it benefits from the generative model capability inherited from the VQ-VAE. Compared to VQ, PQ allows for higher rates, but it becomes comparatively less effective when increasing the number of subquantizers, as one can deduce from the slopes of the rate-distortion curves in Fig.~\ref{fig:results_psnr_msssim}. It is also computationally more expensive.
\begin{figure*}
\setlength{\tabcolsep}{1pt}
\centering
\vspace{-5mm}
\begin{tabular}{ccccc}
\texttt{No Patch Dropping} & \texttt{10\% Drop Rate} & \texttt{20\% Drop Rate} & \texttt{30\% Drop Rate} & \texttt{50\% Drop Rate} \\
\includegraphics[width=0.19\linewidth]{figs/hybrid_gen/img_0_drop0.0.jpg} & \includegraphics[width=0.19\linewidth]{figs/hybrid_gen/img_0_drop0.1.jpg} & \includegraphics[width=0.19\linewidth]{figs/hybrid_gen/img_0_drop0.2.jpg} & \includegraphics[width=0.19\linewidth]{figs/hybrid_gen/img_0_drop0.3.jpg} & \includegraphics[width=0.19\linewidth]{figs/hybrid_gen/img_0_drop0.5.jpg} \\
\includegraphics[width=0.19\linewidth]{figs/hybrid_gen/img_1_drop0.0.jpg} & \includegraphics[width=0.19\linewidth]{figs/hybrid_gen/img_1_drop0.1.jpg} & \includegraphics[width=0.19\linewidth]{figs/hybrid_gen/img_1_drop0.2.jpg} & \includegraphics[width=0.19\linewidth]{figs/hybrid_gen/img_1_drop0.3.jpg} & \includegraphics[width=0.19\linewidth]{figs/hybrid_gen/img_1_drop0.5.jpg} \\
\end{tabular}
\caption{\textbf{\textsc{PQ-MIM}\xspace can operate on a partial set of transmitted tokens.} Since \textsc{PQ-MIM}\xspace is compatible with generation, the conditional entropy model can be repurposed to predict the PQ codes for missing parts in an image. We show results for different dropping rates of transmitted tokens. \textsc{PQ-MIM}\xspace exhibits strong inpainting abilities. Even for the extreme case where half of the image patches are dropped, \textsc{PQ-MIM}\xspace can still retain a large percentage of the original image structure and details.
\label{fig:inpaiting_rand_bitstream}}
\vspace{-2mm}
\end{figure*}
\section{Introduction}
\label{sec:intro}
Image compression played a crucial role in accelerating multiple events in history: it was a major component of the Voyager mission~\citep{ludwig2016voyager}. Efficient image codecs have accelerated the rapid growth of the internet by enabling the transmission of images in a few dozens of kilobytes, thanks to the emergence of effective lossy methods. This democratization was accompanied by standardization efforts to facilitate the interoperability, which led to the emergence of standards such as the Joint Photographic Experts Groups (JPEG). Subsequent formats have leveraged scientific advances on all components of source coding, ranging from transforms~\citep{antonini1992image}, and quantization~\citep{gray98trinf}, to entropy coding~\citep{witten1987arithmetic,taubman2000high}, eventually leading to modern video compression codecs enabling streaming and video-conferencing applications.
Neural methods have recently become increasingly popular for image compression as well as other image processing tasks, such as denoising~\citep{tian20nn}, super-resolution~\citep{bruna2015super,dong2015image,ledig2017photo,wang21pami} or image reconstruction~\citep{wang2020deep,knoll2020deep}.
In typical scenarios, neural image compression is not necessarily mature enough to take over standard techniques like the BPG format inherited from the High-Efficiency Video Coding standard~\citep{sullivan2012overview}.
This is because they do not offer a significant quantitative advantage over prior works that would justify the higher complexity, which depends on the context and operational constraints.
A key advantage of neural compression methods is their enhanced qualitative reconstruction when
incorporating an adversarial loss or likewise psycho-visual objectives favoring visually appealing reconstructions~\citep{agustsson19iccv, mentzer20nips}.
From this perspective, neural compression is related to image generation. The two subfields, however, are currently dominated by different approaches, noticeably they employ different discretization procedures. Indeed, while earlier neural compression methods utilized vector quantization \citep[VQ]{agustsson17nips}, recent methods mostly employ scalar quantization (SQ).
In contrast, the recent literature on image generation \citep{chang22arxiv, yu21iclr, esser21cvpr, rombach2021highresolution} relies on Vector Quantization jointly with a distortion criterion akin to those used in compression.
In this work we aim to reduce the methodological gap and to make a step towards unification of neural image compression and image generation, and allowing image compression to more directly benefit from the rapid advances in image generation methods. Patch-based masking methods for self-supervised learning~\citep{bao21arxiv, he21mae, el2021large} have recently demonstrated their potential for image generation~\citep{chang22arxiv}. Inspired by this work we propose a compression approach built upon Vector Quantized Variational Auto-Encoders~\citep{oord17nips,razavi19nips}. In this context, we focus on two intertwined questions: (1) How to define a vector quantizer offering a range of rate-distortion operating points? (2) How to define an entropy model minimizing the cost of storing the quantization indexes, while avoiding the prohibitive complexity of an auto-regressive model?
To address the challenges above, we revisit vector quantization in image compression, and investigate product quantization~\cite{jegou10pami} (PQ) in a compression system derived from VQ-VAE~\cite{oord16nips}. We show that PQ offers a strong and scalable rate-distortion trade-off.
We then we focus on the spatial entropy modeling and coding of the quantization indexes in the VQ or PQ latent layer, hence we name our method as (Vector/Product)-Quantized Masked Image Modeling (VQ-MIM and PQ-MIM). To this end, we introduce a multi-stage vector-quantized image model: we gradually reduce the conditional entropy of the patch latent codes by increasing the number of observed patches we condition on for each stage. The conditional distribution over patches is estimated by a transformer model and provided to an entropy coder, symmetrically on the emitter and receiver sides.
In summary, we make the following contributions:
\begin{itemize}
\item We introduce a novel Masked Image Modeling conditional entropy model that significantly reduces the rates by leveraging the spatial inter-dependencies between latent codes.
\item We introduce product quantization for VQ-VAE. This simple PQ-VAE variant offers a strong and scalable rate-distortion trade-off.
\item When trained with adversarial and perpetual losses, \textsc{PQ-MIM}\xspace exhibits a strong performance in terms of perceptual metrics like FID and KID, outperforming HiFiC \citep{mentzer20nips}.
\item We qualitatively show that \textsc{PQ-MIM}\xspace is capable of operating in a hybrid mode, between generative and compression, without requiring further training and finetuning. This allows for higher resilience to corrupted or missing signal where our model can fill-in the missing information.
\end{itemize}
\begin{figure*}
\centering
\vspace{-9mm}
\includegraphics[width=\linewidth]{figs/teaser.pdf}
\caption{\textbf{Qualitative example of \textsc{PQ-MIM}\xspace compression.} \textsc{PQ-MIM}\xspace provides a strong compression performance. We retain many of the details present original image with minimal blurring effect even with comrpession rate as low as 0.196 bpp. Moreover, compared to HiFiC we achieve a lower rate for the same image. \textsc{PQ-MIM}\xspace provides colors that are more faithful to the original image, while HiFiC has a darkening effect and some high frequency artifacts. On the other hand, \textsc{PQ-MIM}\xspace can have some smoothing effect that can cause loss of detail for particular regions (e.g. some of the Arabic letters in the example above). More qualitative examples are provided in Appendix~\ref{app:qual_samples}.
\label{fig:teaser}}
\vspace{-4mm}
\end{figure*}
\section{Product Quantized Masked Image Modeling}
\label{sec:method}
\begin{figure*}[t!]
\centering
\vspace{-9mm}
\includegraphics[width=0.95\linewidth]{figs/overview_new.pdf}
\caption{\textbf{\textsc{PQ-MIM}\xspace overview.} Our model consists of (i) a a transformer based encoder and decoder, (ii) a masked image model (MIM) for conditional entropy modeling, and (iii) an entropy coder, \eg an arithmetic coder (AE/AD).
The input image $\mathbf{x}$ is projected to a set of latent features, followed by product
quantization to yield quantization indices $\mathbf{q}$. The arithmetic coder encodes (and decodes) $\mathbf{q}$ into a bitstream $\mathbf{b}$ in a lossless manner.
The elements in $\mathbf{q}$ are spatially split into groups, as detailed in Figure~\ref{fig:spatial_model}.
\textbf{Conditional Entropy Modeling.} Our model estimates the conditional probabilities of the discrete indices in $\mathcal{S}$ steps. Every step, a subset of the tokens $q_i$ is selected using the quincunx pattern. Our MIM transformer estimates $p(q_i | q_{s < i})$ and passes it to the Arithmetic encoder as a CDF, effectively reducing the lossless compression cost.
}
\label{fig:overview}
\vspace{-3mm}
\end{figure*}
This work takes a step towards to closing the gap between neural image compression and image generation methodologies. We revisit vector quantization for image compression and propose an entropy model inspired by masked image modelling. Our compression pipeline, depicted in \fig{overview}, relies on three neural networks:
\begin{enumerate}
\item \textbf{The Encoder network} $\hat{E}: \mathcal{X} \rightarrow \mathcal{Z} $ maps input images $\mathbf{x} \in \mathcal{X}$ to a quantized representation $\mathbf{z} \in \mathcal{Z}$.
\item \textbf{The Mask Image Model} (MIM) compresses the quantized representations without loss of information. This network is involved both on the compression and decompression side.
\item \textbf{The Decoder network} $G: \mathcal{Z} \rightarrow \mathcal{X}$ produces an estimate $\hat{\mathbf{x}}=G(\hat{E}(\mathbf{x}))$ of the original image $\mathbf{x}$.
\end{enumerate}
We now detail the architecture, in particular our PQ proposal, of the image model that we employ in the statistical lossless coding, as well as the training scheme.
\subsection{High-level architecture: PQ-VAE}
\paragraph{High-level architecture.} We follow recent work on discrete generative image models for the design of image encoder and decoder~\citep{chang22arxiv,esser21cvpr,oord17nips,razavi19nips,yu21iclr}.
The encoder $E: \mathcal{X}\rightarrow \mathbb{R}^{w\times h\times d}$ takes an RGB image $\mathbf{x}$ of resolution $W\!\times\!H$ as input and maps it to a latent representation $E(\mathbf{x})$ with $d$ feature channels and a reduced spatial resolution $w\!\times\!h$, downsampling the input resolution by a factor $f={W}/{w}={H}/{h}$.
The $T=w\!\times\!h$ elements of the latent representation $E(\mathbf{x})$ are quantized with a vector quantizer $Q(\cdot)$ to produce the quantized latent representation $\mathbf{z}=Q(E(\mathbf{x}))=\hat{E}(\mathbf{x})$, where each element in $E(\mathbf{x})$ is replaced with its nearest cluster center.
The decoder $G$ uses the quantized latents $\mathbf{z}$ to reconstruct the image.
\mypar{Product Quantization.}
In VQ-VAE, the quantizer $Q$ is simply an online k-means quantizer that produces quantization indices from real-valued vectors.
We denote by $\mathbf{q}\in\{1,\dots,V\}^T$ the map of quantization indices indicating for each element of $\mathbf{z}$ which of the $V$ centroids is selected.
The higher $V$ the more precise is the approximation $\mathbf{z}$, leading to higher bit rates. For instance, assuming that indices are coded with a naive coding scheme (see next section), the bit rate is doubled when moving to $V=256$ to $V=65536$ centroids.
\begin{wrapfigure}{r}{0.45\textwidth}
\begin{center}
\vspace{-1mm}
\includegraphics[trim={0.0cm, 1.0cm, 1.0cm, 0.0cm}, clip, width=1\linewidth]{figs/ablations/pq_vs_vq.pdf}
\caption{\textbf{PQ and VQ comparison}. While VQ provides a comparable performance to PQ for extremely low rates where the codebook size is small, PQ exhibits better scaling behaviour for higher rates. }
\label{fig:pq_vs_vq}
\end{center}
\vspace{-5mm}
\end{wrapfigure}
However, scaling the number of centroids $K$ is possible up to thousands of centroids, but beyond that it is computationally prohibitive. Additionally, it is challenging to train large codebooks where each centroid has a very low probability of being updated. %
To address this problem, we replace the online k-means quantizer by a product quantizer~\citep{jegou10pami} (PQ): the latent vector $\mathbf{z}$ is split into $M$ subvectors as $\mathbf{z}=[\mathbf{z}^1,\dots,\mathbf{z}^j,\dots,\mathbf{z}^M]$ of dimension $M/d$. Each subvector is quantized by a distinct quantizer having $V_\mathrm{s}$ quantization value. The set of quantizers implicitly defines a vector quantizer in the latent space with $V=V_\mathrm{s}^M$ distinct centroids. Hence, we can easily define very large codebooks without the computational and optimization problems mentioned above, because both the assignment and learning are marginalized over the different subspaces. Empirically, we observe in Figure~\ref{fig:pq_vs_vq} that PQ provides a better scaling behaviour for higher rates compared to VQ whose codebook size needs to grow exponentially to achieve the same rates.
\mypar{Neural network.}
Without loss of generality, we choose all neural network models to be identical.
This is not a requirement but this offers the property that the encoder and decoder have identical complexities, and that the memory and compute peaks are identical.
More specifically we choose a cross-variance transformer~\citep{elnouby21nips} (XCiT), whose complexity is linear with respect to image resolution. In contrast, standard vision transformers~\citep{dosovitskiy21iclr} (ViT) are quadratic in the image surface, which is prohibitive for high resolution images that can typically require strong compression. We point out that recent work has shown that Swin-Transformers~\citep{zhu2021transformer} could be a compelling choice as well in the context of image compression. Formally they also have a quadratic complexity, but this is amortized by the hierarchical structure of this architecture.
\subsection{Image entropy model}
In this section we present PQ-MIM.
The objective is to compress the discrete representations $\bm q$ without loss of information, producing a bitstream of the compressed representation that can be transmitted or stored.
During the decoding stage, we invert the aforementioned lossless compression. %
This model can be regarded as the VQ/PQ-VAE counterpart of adaptive contextual arithmetic coders, like EBCOT~\citep{taubman2000high} or CABAC~\citep{richardson2004h}, proposed in early compression standards, in that it couples a conditional probabilistic model with an arithmetic coder.
\begin{figure*}
\centering
\vspace{-2mm}
\includegraphics[width=\linewidth]{ppt/MIM_stages_2.pdf}
\caption{\textbf{Illustration of PQ-MIM with a quincunx pattern.} We employe the quincunx pattern both on the encoder and decoder side.
Each panel represents one of five stages in which we en/de-code a set of tokens in parallel using a probability model $p_s$ implemented by a transformer with parameters $\theta_s$.
The transformer predicts the tokens in $\mathbf{q}_s$, displayed in grayscale and marked by ``\textbf{?}'', and takes as input the preceding groups of tokens $\mathbf{q}_1,\dots,\mathbf{q}_{s-1}$ that are displayed in color. %
The distribution provided by this neural network is fed to an arithmetic en/de-coder. %
\label{fig:spatial_model}
}
\vspace{-3mm}
\end{figure*}
\mypar{Lossless compression.}
A naive manner for lossless compression of the discrete image codes $\mathbf{q}=\{q_t\}_{t=1}^T$ is to use fixed-length codes.
In that case, each code word is assigned to a unique binary representation of equal length, resulting in $\lceil{\log_2 V}\rceil$ bits per element $q_t$.
This approach is computationally very efficient as fixed-length codes are not model-based, and as such do not require likelihood estimation, and because all codes are of equal length by construction, the computation is perfectly parallelizable.
However, theoretically this coding scheme could be Shannon optimal only if codewords are uniformly and independently distributed.
Those assumptions are not met in practice due to the architecture choices we have made previously: k-means does not produce uniformly distributed indices except in singular cases~\citep{gray98trinf}. More details about losseless compression are covered in Appendix~\ref{app:lossless}.
\paragraph{Entropy model.}
To improve the bitrate, we hence rely on an entropy coder, which provides an inverse pair of functions, $\texttt{enc}_{p}$ and $\texttt{dec}_{p}$, achieving near optimal compression rates on sequences of %
symbols for any distribution $p$.
The better $p$ matches the (unknown) underlying data distribution, the better the compression rate~\citep{cover2012elements}. Generally, more powerful generative models will ensure better compression performance. %
Fully autoregressive generative models $p(\mathbf{q})=\prod_{t=1}^T p(q_t|\mathbf{q}_{<t})$ are powerful ~\citep{ding21arxiv,esser21cvpr,gafni22arxiv,ramesh21arxiv,yu21iclr},
however, they are inconvenient in that the likelihood estimation for this type of model is not trivially parallelizable: each patch index must be processed sequentially as it is used to condition subsequent patch indices.
Thus, similar to prior works~\citep{chang22arxiv,reed17icml} we propose a masked image model,
which we use to predict the image patch indices in several stages. %
Specifically, we partition $\mathbf{q}$ into $S$ subsets $\mathbf{q}_1, \mathbf{q}_2, \dots, \mathbf{q}_S$ of patch indices, that we refer to as \emph{tokens} by analogy to language modelling:
\begin{equation}
\mathbf{q} = \bigcup_{s=1}^S \mathbf{q}_s.
\end{equation}
We model the elements in each subset conditionally independent given all preceding groups:
\begin{align}
p(\mathbf{q}) & = \ \ \prod_{s=1}^{S} p\left(\mathbf{q}_s|\mathbf{q}_{<s};\theta_s\right),\\
p\left(\mathbf{q}_s|\mathbf{q}_{<s};\theta_s\right) & = \prod_{q_t\in \mathbf{q}_s} p\left(q_t|\mathbf{q}_1, \mathbf{q}_2, \dots, \mathbf{q}_{s-1}; \theta_s\right).
\end{align}
Since $p(\mathbf{q}_1)$ is not conditioned on any previous elements, it fully factorizes over $q_t\in\mathbf{q}_1$, and we model it as the marginal distribution over the vocabulary observed on the training data.
The non-trivial conditional distributions $p\left(\mathbf{q}_s|\mathbf{q}_{<s};\theta_s\right)$ for $s\geq 2$ are modeled using transformer networks, which have few inductive biases and have been successful across many tasks, including image generation. An overview of the MIM entropy model is illustrated in Figure~\ref{fig:overview}.
\mypar{Amortized encoding and decoding.}
From a computational point of view, our proposal allows for the compression (resp. decompression) to proceed in $S$ stages. In each stage $s$ we encode/decode the set $\mathbf{q}_s$ of tokens conditioned on the groups $\mathbf{q}_1, \dots, \mathbf{q}_{s-1}$ encoded/decoded in preceding stages, but independently among the tokens in the set $\mathbf{q}_s$.
This allows for paralellization among the elements in each subset $\mathrm{q}_s$, and requires strictly $S$ forward passes through the model independent of the image size, rather than $T$ sequential forwards passes for fully autoregressive models.
\mypar{Quincunx partitioning.}
In practice, we typically use $S=5$ stages.
We have explored different patterns to partition the $T$ tokens over the $S$ stages.
In particular we consider the ``quincunx'' regular grid pattern, where in each stage we double the number of tokens to predict, see \fig{spatial_model} for an illustration. This multi-level refinement was previously explored for image compression in the context of lifting schemes designed with oriented wavelets~\citep{chappelier2006oriented}.
In our experiments we contrast this partitioning with alternative ones with other patterns and subset cardinalities (Figure~\ref{fig:mim_steps}).
\begin{figure}[t!]
\vspace{-5mm}
\begin{minipage}{\linewidth}
\begin{subfigure}{0.48\linewidth}
\centering
\caption{\textsc{Tecnick}\xspace PSNR}
\includegraphics[trim={0.0cm, 1.0cm, 1.0cm, 0.0cm}, clip, width=1.0\linewidth]{figs/rd_plots/tecnick_mse.pdf}
\end{subfigure}%
\hfill
\begin{subfigure}{0.48\linewidth}
\centering
\caption{\textsc{Tecnick}\xspace MS-SSIM}
\includegraphics[trim={0.0cm, 1.0cm, 1.0cm, 0.0cm}, clip, width=1.0\linewidth]{figs/rd_plots/tecnick_msssim.pdf}
\end{subfigure}
\end{minipage}%
\hfill
\begin{minipage}{\linewidth}
\begin{subfigure}{0.48\linewidth}
\centering
\caption{\textsc{Kodak}\xspace PSNR}
\includegraphics[trim={0.0cm, 1.0cm, 1.0cm, 0.0cm}, clip, width=1.0\linewidth]{figs/rd_plots/kodak_mse.pdf}
\end{subfigure}%
\hfill
\begin{subfigure}{0.48\linewidth}
\centering
\caption{\textsc{Kodak}\xspace MS-SSIM}
\includegraphics[trim={0.0cm, 1.0cm, 1.0cm, 0.0cm}, clip, width=1.0\linewidth]{figs/rd_plots/kodak_msssim.pdf}
\end{subfigure}
\caption{\textbf{Rate-Distortion performance for \textsc{Tecnick}\xspace and \textsc{Kodak}\xspace datasets.} We report \textsc{PQ-MIM}\xspace PSNR and MS-SSIM performance for various operating points. \textsc{PQ-MIM}\xspace provides a competitive performance, particularly for MS-SSIM, compared to standard codecs such as JPEG 2000 \citep{Taubman2012} and BPG \citep{ballard_bpg} as well as recent neural methods \citep{cheng20cvpr, minnen18nips, balle18iclr}.
}
\label{fig:results_psnr_msssim}
\end{minipage}
\end{figure}
\subsection{Training the \textsc{PQ-MIM}\xspace}
\mypar{Reconstruction objective and training.}
The goal of lossy image compression is to match the image and its reconstruction as closely as possible according to some distortion metric. In this paper, we train our model to reduce the image distortion and the quantization loss using the following objective:
\begin{equation}
L = L_\textrm{Rec} + \eta \cdot L_\textrm{PQ}
\end{equation}
Following VQ-VAE, our quantization objective $L_\textrm{PQ}$ consists of an embedding and commitment losses, averaged over the $M$ different PQ sub-vectors. For the distortion loss $L_\textrm{Rec}$, we present two setups where we use different types of distortion measures:
\begin{itemize}[leftmargin=*]
\item \textbf{MSE \& MS-SSIM.} Typical distortion measures used in the majority of the neural compression literature such as mean squared error (MSE) or multi-scale structural similarity~\citep{wang2003multiscale} (MS-SSIM). For this setup, the model is trained solely using one distortion measure at a time.
\begin{equation}
L_\textrm{Rec}(\mathbf{x},\hat{E},G) = L_\textrm{MSE/MS-SSIM}(\mathbf{x},\hat{\mathbf{x}})
\end{equation}
\item \textbf{Perceptual measures.} Alternatively, we report a setup where we utilize perceptual objectives such as LPIPS~\citep{zhang18cvpr} and adversarial training~\citep{goodfellow14nips} to enhance psycho-visual image quality. The distortion loss is defined as:
\begin{equation}
L_\textrm{Rec}(\mathbf{x},\hat{E},G) = L_\textrm{MSE}(\mathbf{x},\hat{\mathbf{x}}) + \alpha \cdot L_\textrm{Perc}(\mathbf{x},\hat{\mathbf{x}}) + \gamma \cdot L_\textrm{Adv}(\hat{E},G,D),
\end{equation}
where $\alpha$ and $\gamma$ are weighing coefficients and the adversarial loss $L_\textrm{Adv}$ is defined as:
\begin{equation}
L_\textrm{Adv}(\hat{E},G,D) = \mathbb{E}_{\mathbf{x}}[ \ln D(\mathbf{x}) ] + \mathbb{E}_{\hat{\mathbf{x}}}[ \ln(1- D(\hat{\mathbf{x}})) ],
\label{eq:Ladv}
\end{equation}
where $D(\cdot)$ is the discriminator and $\mathbb{E}_\mathbf{x}$ denotes the expectation over $\mathbf{x}$ sampled uniformly from the training set. Similarly $\mathbb{E}_{\hat{\mathbf{x}}}$ denotes the expectation over reconstructed training images.
Note that, unlike the MSE and perceptual losses, the adversarial loss does not compare individual images and their reconstructions, but aims to match the \emph{distributions} of original images and their reconstructions.
\end{itemize}
\mypar{Training the entropy model.}
Our MIM module is an XCiT transformer that accepts $T$ tokens as input representing the image patches. During training, we randomly mask a set of tokens by sampling from a uniform distribution $U(0, 1)$. The masked tokens are replaced with a mask embedding vector, while the observed token indices are mapped to their corresponding continuous representation using an embedding look-up table. The MIM module outputs a context vector for every masked token which in turn is passed to $M$ linear heads, representing the different PQ sub-vector indices, followed by a softmax to yield a distribution $p({\bm q}_s| {\bm q}_{<s})$.
The module is trained using a standard cross-entropy objective. While we train the autoencoder and the MIM modules simultaneously, we do not backpropagate gradients from the MIM module to the encoder $E$ or the quantization parameters, so the two components can be trained separately in sequence.
\section{Related work}
\label{sec:related}
\noindent \textbf{Neural image compression}
Early approaches to neural image compression reach back to the late 1980s \citep{sonehara1989image,sicuranza1990artificial,bottou1998high}.
Recent rapid advances in explicit and implicit density modelling~\citep{goodfellow14nips,kingma2013auto,larochelle2011neural,van2016wavenet,rezende2015variational,salimans17iclr} have renewed interest in posing image compression as a learning problem.
Due to the connection to variational learning~\citep{gregor2016towards,frey1998bayesian,alemi2018fixing}, variational auto-encoders have been the primary choice for lossy image compression.
In contrast to standard variational models, the evaluation of neural compression models focuses on achievable bitrates, multi-scale applicability and computational complexity.
The initial works in the field used fully convolutional architectures for encoding/decoding~\citep{balle17iclr, theis2017lossy, mentzer18cvpr}.
The resulting encoded latent image representations are quantized and compressed via an entropy coder with learned explicit density model or ``entropy bottleneck''.
Initial variational approaches directly modeled a single level of code densities.
Ball\'e extended these models by introducing a second ``hyperprior'' that yielded improved performance~\citep{balle18iclr}.
Hyperprior models have been the basis for several subsequent advances with further improvements for density modeling, such as joint autoregressive models \citep{minnen18nips}, Gaussian mixture/attention \citep{cheng20cvpr}, and channel-wise auto-regressive models \citep{minnen2020channel}.
Another line of work has proposed to use vector quantization with histogram-based probabilities for image compression \citep{agustsson17nips,lu2019learning}.
Contrary to VQ-VAE models, these models typically optimize the rate (or a surrogate of the rate) directly and may include a spatial component for the quantized vectors.
Yang et al. \citep{yang2020improving} showed multiple ways to improve the encoding process, including the fine-tuning of the discretization process and employing bits-back coding in the entropy bottleneck.
Finally, \citep{santurkar2018generative,mentzer20nips, rippel17icml, agustsson19iccv} showed that altering the distortion metric to include an additional adversarial loss can make a large difference for compression rate.
Another interesting line of work considers image compression by training image-specific networks, or network adapters, that map image coordinates to RGB values, and compressing the image-specific parameters~\citep{dupont21arxiv,dupont22arxiv,strupler22eccv}.
\mypar{VQ-models for image generation.}
There has been significant interest in generative models based in discrete image representations, as introduced by VQ-VAE~\citep{oord17nips,razavi19nips}.
A discrete representation of reduced spatial resolution is learned by means of an autoencoder which quantizes the latent representations.
This discrete representation is coupled with a strong prior, for example implemented as an autoregressive pixel-CNN model.
VQ-GAN~\citep{esser21cvpr} replaces the prior architecture with a transformer model~\citep{vaswani17nips}, and introduces an adversarial loss term to learn an autoencoder with more visually pleasing reconstructions and improved sample quality.
This approach has been extended to text-based generative image models by extending the prior to model a longer sequence that combines the discrete image representation with a prefix that encoding the conditioning text.
This has yielded impressive results by scaling the model capacity and training data to tens or hundreds of million text-image pairs~\citep{ding21arxiv,gafni22arxiv,ramesh21arxiv}.
A fundamental limitation of autoregressive generative models is that they sample data sequentially, requiring separate non-parallel evaluation of the predictive (transformer) model to sample each token.
To alleviate this, several data items can be sampled independently in parallel, conditioning on all previously sampled tokens.
This has been leveraged to speed-up pixel-CNNs for small images and video by two to three orders of magnitude~\citep{reed17icml}.
More recently, MaskGIT~\citep{chang22arxiv} and follow-up work~\citep{lezama22eccv} explored this for generative models of VQ-VAE representations, and find that similar or better sample quality is obtained by parallel sampling of image patch subsets in few steps, reducing the generation time significantly.
Despite their success for image synthesis we are not aware of the use of such models for image compression in earlier work.
|
\section{Introduction}
\par
We study a~combinatorial property of a~pair ${\mathcal I},{\mathcal J}$ of ideals on~$\omega$. Namely, we consider the~following variation of a~P-ideal:
\begin{defi}\label{definiciaPJ}
We say that an~ideal ${\mathcal I}$ is a~$\mathrm{P}({\mathcal J})$-ideal if for any family $\{I_n:n\in\omega \}\subseteq{\mathcal I}$ there is $I\in{\mathcal I}$ such that $I_n\subseteq^{\mathcal J} I$ for each $n\in\omega$.
\end{defi}
The~property was introduced by M.~Mačaj and M.~Sleziak~\cite{macslez}, and applied later in~\cite{filipow, stanfil, stan, rep1, rep2}. In addition to its equivalent formulation via the Boolean algebra $\mathcal{P}(\omega)/{\mathcal J}$, M.~Mačaj and M.~Sleziak found the~property to be a~useful tool in characterizing certain properties of ideal convergence we describe below. They define a~sequence $\seqn{x_n}{n}$ in a~topological space~$X$ to be ${\mathcal I}^{\mathcal J}$-convergent to a~point $x\in X$ if there is $F\in{\mathcal I}^\ast$ such that the~sequence $\gseqn{x_n}{n\in F}$ is $({\mathcal J}\restriction F)$-convergent to~$x$. They have shown
\begin{theorem*}[M. Mačaj -- M. Sleziak]
Let $X$ be a~non-discrete {\rm T}$_1$ first countable topological space. The~following are equivalent:
\begin{enumerate}[(1)]
\item ${\mathcal I}$ is a~$\mathrm{P}({\mathcal J})$-ideal.
\item The family of all equivalence classes corresponding to sets in ${\mathcal I}$ is $\sigma$-directed in the quotient $\mathcal{P}(\omega)/{\mathcal J}$.
\item For any sequence $\seqn{x_n}{n}$ in~$X$, if $\seqn{x_n}{n}$ is ${\mathcal I}$-convergent to~$x$ then $\seqn{x_n}{n}$ is ${\mathcal I}^{\mathcal J}$-convergent to~$x$.
\end{enumerate}
\end{theorem*}
In~\cite{filipow} R.~Filipów and M.~Staniszewski study ideal version of equal convergence~\cite{CsaLa}, called quasi-normal convergence in~\cite{Buk91}. In fact, various versions of ideal equal convergence have been introduced in \cite{das, szuca} but the most general one has appeared in this very paper. A~sequence $\seqn{f_n}{n}$ of real-valued functions defined on a~set~$X$ is $({\mathcal I},{\mathcal J})$-equally convergent to a~function $f$ if there is ${\mathcal J}$-converging sequence of positive reals $\seqn{\varepsilon_n}{n}$ such that $\set{n}{|f_n(x)-f(x)|\geq\varepsilon_n}\in{\mathcal J}$ for every $x\in X$. They have shown
\begin{theorem*}[R.~Filipów -- M.~Staniszewski]
Let $X$ be a~non-empty set. The~following are equivalent:
\begin{enumerate}[(1)]
\item ${\mathcal I}$ is a~$\mathrm{P}({\mathcal J})$-ideal.
\item For any sequence $\seqn{f_n}{n}$ of real-valued functions on~$X$, if $\seqn{f_n}{n}$ is $({\mathcal I},{\mathcal I})$-equally convergent to~$f$ then $\seqn{f_n}{n}$ is $({\mathcal I},{\mathcal J})$-equally convergent to~$f$.
\item For any sequence $\seqn{f_n}{n}$ of real-valued functions on~$X$, if $\seqn{f_n}{n}$ is ${\mathcal I}$-uniformly convergent to~$f$ then $\seqn{f_n}{n}$ is $({\mathcal I},{\mathcal J})$-equally convergent to~$f$.
\item For any sequence $\seqn{f_n}{n}$ of real-valued functions on~$X$, if $\seqn{f_n}{n}$ is $\sigma$-${\mathcal I}$-uniformly convergent to~$f$ then $\seqn{f_n}{n}$ is $({\mathcal I},{\mathcal J})$-equally convergent to~$f$.
\end{enumerate}
\end{theorem*}
This is an interesting result since in the classical (non-ideal) case, uniform convergence is stronger than equal convergence and $\sigma$-uniform convergence is even equivalent to equal convergence.
\par
We shall study the~introduced variation of P-ideal for ideals which are defined via a~given partition of natural numbers. Four such critical ideals, which obtained standard notation $\fin\times\emptyset$, ${\mathcal{ED}}$, $\emptyset\times\fin$, $\fin\times\fin$, have been isolated in the literature (see below for the~definition and~\cite{BrFl17,BrFaVe,Hr,hrusakborel} for details). We consider an additional ideal -- ${\mathcal Sel}$, the~ideal generated by the family of graphs of functions from ${^\omega\omega}$. We have found all pairs ${\mathcal I},{\mathcal J}$ such that ${\mathcal I}$ is a~$\mathrm{P}({\mathcal J})$-ideal, see the~table below or Table~\ref{omxom}.
\begin{center}
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline
& $\mathrm{P}$ & $\mathrm{P}(\text{\rm Fin}\times\emptyset)$& $\mathrm{P}({\mathcal Sel})$ & $\mathrm{P}({\mathcal E}{\mathcal D})$ & $\mathrm{P}(\emptyset\times\text{\rm Fin})$ & $\mathrm{P}(\text{\rm Fin}\times\text{\rm Fin})$\\
\hline\hline
$\text{\rm Fin}\times\emptyset$ & \xmark & \cmark & \xmark &\cmark & \xmark & \cmark \\
\hline
${\mathcal Sel}$ & \xmark & \xmark & \cmark &\cmark & \cmark & \cmark\\
\hline
${\mathcal E}{\mathcal D}$ & \xmark & \xmark & \xmark &\cmark & \xmark &\cmark\\
\hline
$\emptyset\times\text{\rm Fin}$& \cmark & \cmark & \cmark &\cmark& \cmark & \cmark\\
\hline
$\text{\rm Fin}\times\text{\rm Fin}$ & \xmark & \cmark & \xmark & \cmark & \xmark &\cmark \\
\hline
\end{tabular}
\end{center}
Furthermore, we study a~more general situation when the~partition~${\mathcal A}$ inducing an ideal ${\mathcal I}$ is potentially different than the~partition~${\mathcal B}$ inducing an ideal ${\mathcal J}$. Therefore we introduce the~following notation. We begin with ideals which are not tall. Let ${\mathcal A}$ be an infinite partition of $\omega\times\omega$ into infinite sets.
\begin{center}
\begin{tabular}{ll}
$\finop{{\mathcal A}}$& the~ideal generated by~${\mathcal A}$,\\[0.2cm]
$\selp{{\mathcal A}}$& the~ideal generated by the~family of all selectors of~${\mathcal A}$,\footnotemark\\[0.2cm]
$\ofinp{{\mathcal A}}$& the~family of all sets with finite intersection with each element of~${\mathcal A}$.
\end{tabular}
\footnotetext{We say that a set $S\in{\mathcal P}(X)$ is a selector (semiselector) of an infinite family ${\mathcal A}\subseteq{\mathcal P}(X)$, if $\abs{S\cap A}=1$ ($\abs{S\cap A}\leq 1$ and $S$ is infinite) for any $A\in{\mathcal A}$.}
\end{center}
We sometimes write shortly $\langle{\mathcal A}\rangle$ instead of~$\finop{{\mathcal A}}$. A~supremum of pairs of previous ideals leads to the~following tall ideals.
\begin{center}
\begin{tabular}{ll}
$\EDp{{\mathcal A}}$& the~ideal generated by~$\finop{{\mathcal A}}\cup\selp{{\mathcal A}}$,\\[0.2cm]
$\finfinp{{\mathcal A}}$& the~ideal generated by~$\finop{{\mathcal A}}\cup\ofinp{{\mathcal A}}$.
\end{tabular}
\end{center}
One can see that the~introduced ideals are representations of all isomorphic copies of ideals $\fin\times\emptyset$, ${\mathcal Sel}$, ${\mathcal{ED}}$, $\emptyset\times\fin$, $\fin\times\fin$, respectively. Considering vertical sets $V_n=\{n\}\times\omega$,\footnote{We will use this notation for the sets $\{n \}\times \omega$ throughout the paper.} and corresponding partition ${\mathcal V}=\set{V_n}{n\in\omega}$, we obtain $\fin\times\emptyset=\finop{{\mathcal V}}$, ${\mathcal Sel}=\selp{{\mathcal V}}$, etc.
\par
The~paper is organized as follows. In the~next section, we continue recalling the results on P-property for two arbitrary ideals obtained by other authors. Our own observations are added as well. The~most important result here is the~introduction of the~relation~$\rsub{}{}$ between two ideals. It is a~natural sufficient condition for a~P-property for two ideals to be satisfied, and can be seen as a generalization of inclusion and orthogonality, see Proposition~\ref{ortaink}. Surprisingly, the~relation~$\rsub{}{}$ is a~characterization of a~P-property for many pairs of ideals we consider. Indeed, we say that $\rsub{{\mathcal I}}{{\mathcal J}}$ if the~whole ideal ${\mathcal I}$ has a~${\mathcal J}$-union\footnote{A~set~$E$ such that $I\subseteq^{\mathcal J} E$ for all $I\in{\mathcal I}$.} in~${\mathcal I}$, and by Corollary~\ref{prehlcor}, Theorems~\ref{ref1}, \ref{veze}, \ref{ed_ofin}, we obtain
\begin{theorem*}
If ${\mathcal I}$ is one of $\finop{{\mathcal A}}$, $\selp{{\mathcal A}}$, $\EDp{{\mathcal A}}$, and ${\mathcal J}$ one of $\finop{{\mathcal B}}$, $\ofinp{{\mathcal B}}$, then
\begin{center}
${\mathcal I}$ is a $\mathrm{P}({\mathcal J})$-ideal if and only if $\rsub{{\mathcal I}}{{\mathcal J}}$.
\end{center}
\end{theorem*}
\noindent Previous theorem is not true for ${\mathcal I}=\ofinp{{\mathcal A}}$ since $\ofinp{{\mathcal A}}$ is a~P-ideal.
\par
The~reasoning for Table~\ref{omxom} is included in Section~\ref{natural_finfin}. Copies of~$\fin\times\emptyset$ and~$\fin\times\fin$ are being examined in Section~\ref{natural_finfin} as well, and in Corollary~\ref{prehlcor} we prove the~following.
\begin{theorem*}
$\finop{{\mathcal A}}$ is a $\mathrm{P}({\mathcal J})$-ideal if and only if $\finfinp{{\mathcal A}}$ is a $\mathrm{P}({\mathcal J})$-ideal.
\end{theorem*}
\par
Considering copies of the~ideal ${\mathcal Sel}$ in Section~\ref{S-towers} we found a~schema of combinatorial properties of partitions describing necessary criteria for $\selp{{\mathcal A}}$ to be a~$\pid{{\mathcal J}}$-ideal for all mentioned copies~${\mathcal J}$ of ideals induced by a~partition, see Proposition~\ref{sufficient}. Moreover, the~conditions are sufficient in some cases as well. We say that a~family of $k$ disjoint semiselectors of~${\mathcal A}$ is an~($\omega,k$)-tower of ${\mathcal B}$-monochromatic semiselectors of~${\mathcal A}$ if all semiselectors from this family are selectors of a~single infinite subfamily of ${\mathcal A}$,
and each semiselector is covered by a~set in~${\mathcal B}$. In Theorem~\ref{veze} we show the~following.
\begin{theorem*}
$\selp{{\mathcal A}}$ is a $\pid{\ofinp{{\mathcal B}}}$-ideal if and only if there is no \emph{(}$\omega,k$\emph{)}-tower of ${\mathcal B}$-monochromatic semi\-se\-lectors of~${\mathcal A}$ for some~$k$.\footnote{However, to simplify the~exposition we treat just ${\mathcal A}={\mathcal V}$ in the~paper, which describes just when ${\mathcal Sel}$ is a $\pid{\ofinp{{\mathcal B}}}$-ideal. In such a~case, semiselectors are just partial functions on~$\omega$. One can easily modify the~definition of the~notion to obtain a~more general situation.
}
\end{theorem*}
\par
In Section~\ref{towers_equiv}, we add copies of the~ideal ${\mathcal{ED}}$ into consideration. We show that in two cases, the~orthogonality is equivalent to the P-property for two ideals. Thus, in Theorems~\ref{ref1} and~\ref{ed_ofin} we prove the~following.
\begin{theorem*}
\
\begin{enumerate}[\rm (a)]
\item $\EDp{{\mathcal A}}$ is a $\mathrm{P}(\finop{{\mathcal B}})$-ideal if and only if $\EDp{{\mathcal A}}\perp\finop{{\mathcal B}}$.
\item $\EDp{{\mathcal A}}$ is a $\pid{\ofinp{{\mathcal B}}}$-ideal if and only if $\EDp{{\mathcal A}}\perp\ofinp{{\mathcal B}}$.
\end{enumerate}
\end{theorem*}
\par
Finally, the~last section is devoted to the~construction of a~partition~${\mathcal E}$, and to the~non-trivial proof of ${\mathcal Sel}$ being not a~$\pid{\selp{{\mathcal E}}}$-ideal, see Theorem~\ref{tazkaThm}. However, the~partition~${\mathcal E}$ is constructed in a~way that restricts the presence of monochromatic towers to extremely small cases.
Thus, our convenient combinatorial property is not enough to characterize all cases in which the~ideal ${\mathcal Sel}$ satisfies P-property for two ideals.
\section{Introducing P-property for two ideals}
The~aim of the~section is to recall basic terminology and outline the basic behavior of P-property for two arbitrary ideals. All the~other sections are devoted to the study of P-property for two ideals induced by partition. Here, we shall summarize the~results appearing in the~literature, and we shall extend the list with our own observations.
\par
Let us begin with recalling basic terminology and notation on ideals. We use $(\forall^\infty n\in\omega)$ to abbreviate ``for all but finitely many $n$'s'' and $(\exists^\infty n\in\omega)$ is an abbreviation for ``there exist infinitely many $n$'s''. Let $M$ be a non-empty infinite set. An \textbf{ideal} on $M$ is a family ${\mathcal I}\subseteq {\mathcal P}(M)$ closed under taking subsets and finite unions, i.e., ${\mathcal I}$ is an ideal if $B\in{\mathcal I}$ for any $B\subseteq A\in{\mathcal I}$ and $A\cup B\in{\mathcal I}$ for any $A,B\in{\mathcal I}$. Moreover, we assume that all ideals are proper ($M\notin {\mathcal I}$) and contain all finite subsets of $M$. We are interested mostly in ideals on $\omega$. However, we study these ideals also on other countable sets $M$ by identifying $M$ with $\omega$ via fixed bijection. The ideal of all finite subsets of $M$ is denoted by $\text{\rm Fin}$, where the underlying set $M$ is usualy evident from the context. Calligraphic ${\mathcal I},{\mathcal J}$ are used exclusively to denote ideals on a countable set. A~family ${\mathcal B}\subseteq{\mathcal I}$ is a \textbf{base} of ${\mathcal I}$ if for any $A\in{\mathcal I}$ there is $B\in{\mathcal B}$ such that $A\subseteq B$. ${\mathcal B}$ is a~\textbf{subbase} for ${\mathcal I}$ if the family of all finite unions of members of ${\mathcal B}$ is a base for ${\mathcal I}$.
\par
For any $A, B\subseteq M$, we write $A\subseteq^{\mathcal J} B$ if $A\setminus B\in{\mathcal J}$, and $A=^{\mathcal J} B$ if $A\subseteq^{\mathcal J} B$ and $B\subseteq^{\mathcal J} A$. We use the notation $A\subseteq^*B$ instead of $A\subseteq^\text{\rm Fin} B$. We recall from Definition~\ref{definiciaPJ} that ${\mathcal I}$ is a~$\boldsymbol{\mathrm{P}({\mathcal J})}$\textbf{-ideal}\footnote{We use the notation from \cite{filipow}, although the notion was introduced in \cite{macslez} as $\mathrm{AP}({\mathcal I},{\mathcal J})$ property.} if for any family $\{I_n:n\in\omega \}\subseteq{\mathcal I}$ there is $I\in{\mathcal I}$ such that $I_n\subseteq^{\mathcal J} I$ for each $n\in\omega$. Note that for a~base~${\mathcal B}$ of ${\mathcal I}$ we have that ${\mathcal I}$ is $\mathrm{P}({\mathcal J})$ if and only if for every family $\{B_n:n\in\omega \}\subseteq{\mathcal B}$ there is $B\in{\mathcal B}$ such that $B_n\subseteq^{\mathcal J} B$ for every~$n$.\footnote{Let us consider the following two-player game: in the stage $n$ Player \rom{1} chooses a set $I_n\in{\mathcal I}$. Player \rom{2} responds by playing $J_n\in{\mathcal J}$ a subset of $I_n$. Player \rom{2} wins if $\bigcup_{n\in\omega}I_n\setminus J_n\in{\mathcal I}$. It is easy to see that in the language of games ${\mathcal I}$ is a $\pid{{\mathcal J}}$ if and only if Player \rom{2} has a winning strategy in this game.}
\par
We include a~list of reformulations of the~definition by M.~Mačaj and M.~Sleziak~\cite{macslez}. Let us recall that ${\mathcal I}^*$ is the~\textbf{dual filter} to ${\mathcal I}$, i.e., ${\mathcal I}^*=\{M\setminus I:I\in {\mathcal I} \}$.
\par
\begin{lema}[M. Mačaj, M. Sleziak, \cite{macslez} Lemma~3.9.]
Let ${\mathcal I}$ and ${\mathcal J}$ be ideals on the same set $M$. The following conditions are equivalent:
\begin{enumerate}[(1)]
\item ${\mathcal I}$ is a $\pid{{\mathcal J}}$-ideal.\footnote{We use the notation from \cite{filipow}, although the notion was introduced in \cite{macslez} as $\mathrm{AP}({\mathcal I},{\mathcal J})$ property.}
\item Any family $\{F_n:n\in\omega \}\subseteq{\mathcal I}^*$ has ${\mathcal J}$-intersection\footnote{We say that a set $X$ is ${\mathcal K}$-intersection of a family $\{X_n:n\in\omega \}$ if $X\subseteq^{\mathcal K} X_n$ holds for each $n\in\omega$. ${\mathcal K}$-intersection is "${\mathcal K}$-pseudointersection" in \cite{macslez}. } in ${\mathcal I}^*$.
\item For every family $\{I_n:n\in\omega \}\subseteq{\mathcal I}$ there is a family $\{J_n:n\in\omega\} \subseteq{\mathcal I}$ such that $I_n=^{\mathcal J} J_n$ for $n\in\omega$ and $\bigcup_{n\in\omega}J_n\in{\mathcal I}$.
\item For every family of mutually disjoint sets $\{I_n:n\in\omega \}\subseteq{\mathcal I}$ there is a family $\{J_n:n\in\omega\} \subseteq{\mathcal I}$ such that $I_n=^{\mathcal J} J_n$ for $n\in\omega$ and $\bigcup_{n\in\omega}J_n\in{\mathcal I}$.
\item For every family $\{I_n:n\in\omega \}\subseteq{\mathcal I}$ such that $I_n\subseteq I_{n+1}$ for each $n\in\omega$ there is a family $\{J_n:n\in\omega\} \subseteq{\mathcal I}$ such that $I_n=^{\mathcal J} J_n$ for $n\in\omega$ and $\bigcup_{n\in\omega}J
_n\in{\mathcal I}$.
\item In the Boolean algebra $\mathcal{P}(M)/{\mathcal J}$ the ideal ${\mathcal I}$ corresponds to a $\sigma$-directed subset\footnote{i.e., it contains an upper bound of each countable subset.}.
\end{enumerate}
\end{lema}
\noindent It is common to use the term $\mathrm{P}$-ideal instead of $\mathrm{P}(\text{\rm Fin})$-ideal.
\par
Some of the basic examples of pairs of ideals such that ${\mathcal I}$ is $\mathrm{P}({\mathcal J})$-ideal were pointed out by R.~Filipów and M.~Staniszewski in \cite{stanfil}. Ideals ${\mathcal I}$, ${\mathcal J}$ on $M$ are said to be \textbf{orthogonal}, written ${\mathcal I}\perp{\mathcal J}$, if there is a set $A\subseteq M$ such that $A\in{\mathcal I}$ and $M\setminus A\in {\mathcal J}$. An ideal ${\mathcal I}$ is \textbf{maximal}, if either $A\in {\mathcal I}$ or $M\setminus A\in {\mathcal I}$ for every $A\subseteq M$ (or equivalently, ${\mathcal I}$ is maximal with respect to ordering by inclusion).
\begin{lema}[R.~Filipów -- M.~Staniszewski]\label{ort}
\
\begin{enumerate}[(a)]
\item If ${\mathcal I}$ is a $\mathrm{P}$-ideal, then ${\mathcal I}$ is a $\mathrm{P}({\mathcal J})$-ideal for every ${\mathcal J}$.
\item If ${\mathcal I},{\mathcal J}$ are orthogonal then ${\mathcal I}$ is $\mathrm{P}({\mathcal J})$ and ${\mathcal J}$ is $\mathrm{P}({\mathcal I})$.
\item If ${\mathcal I},{\mathcal J}$ are maximal then ${\mathcal I}$ is $\mathrm{P}({\mathcal J})$ and ${\mathcal J}$ is $\mathrm{P}({\mathcal I})$.
\item If ${\mathcal I}\subseteq {\mathcal J}$, then ${\mathcal I}$ is a~$\mathrm{P}({\mathcal J})$-ideal. In particular, ${\mathcal I}$ is a~$\mathrm{P}({\mathcal I})$-ideal for every~${\mathcal I}$.
\item If ${\mathcal J}_1\subseteq{\mathcal J}_2$ and ${\mathcal I}$ is a~$\mathrm{P}({\mathcal J}_1)$-ideal, then ${\mathcal I}$ is a~$\mathrm{P}({\mathcal J}_2)$-ideal.
\end{enumerate}
\end{lema}
\noindent Taking into account Lemma~\ref{ort}(e), and assuming ${\mathcal I}$ is a~$\mathrm{P}({\mathcal J})$-ideal, it would be tempting to conjecture that either ${\mathcal I}'\subseteq{\mathcal I}$ or ${\mathcal I}'\supseteq{\mathcal I}$ leads to ${\mathcal I}'$ being a~$\mathrm{P}({\mathcal J})$-ideal. However, it does not necessarily need to be true, see Table~\ref{omxom}.
\par
In the~following proposition, we generalize parts~(b) and~(d) of Lemma~\ref{ort}, introducing a~unifying element, the~relation~$\rsub{}{}$. Given an~ideal ${\mathcal I}$ on $M$ we denote by ${\mathcal I}^+$ the family of $\boldsymbol{{\mathcal I}}$\textbf{-positive sets} (subsets of $M$ which are not in ${\mathcal I}$), i.e., ${\mathcal I}^+={\mathcal P}(M)\setminus{\mathcal I}$. If $X\in{\mathcal I}^+$, we denote by ${\mathcal I}\restriction X$ the ideal $\{I\cap X: I\in{\mathcal I} \}$ on~$X$. Finally, by $\boldsymbol{\rsub{{\mathcal I}}{{\mathcal J}}}$ we mean the condition
\[
(\exists E\in{\mathcal I}^*)\ {\mathcal I}\restriction E\subseteq{\mathcal J}.
\]
Note that if ${\mathcal I}\subseteq{\mathcal J}$ then $\rsub{{\mathcal I}}{{\mathcal J}}$.
\begin{prop}\label{ortaink}
\
\begin{enumerate}[(a)]
\item If ${\mathcal I},{\mathcal J}$ are orthogonal then $\rsub{{\mathcal I}}{{\mathcal J}}$.
\item If $\rsub{{\mathcal I}}{{\mathcal J}}$, then ${\mathcal I}$ is a $\mathrm{P}({\mathcal J})$-ideal.
\end{enumerate}
\end{prop}
\begin{proof}
(a) If ${\mathcal I}\perp {\mathcal J}$ then there is a set $E'\in{\mathcal I}$ such that $E=M\setminus E'\in{\mathcal J}\cap{\mathcal I}^*$.
(b) Let ${\mathcal I}$, ${\mathcal J}$ be ideals on $M$. Let $\{I_n:n\in\omega \}\subseteq {\mathcal I}$ and $I=M\setminus E\in {\mathcal I}$. Then $I_n\setminus I= I_n\cap (M\setminus I) =I_n\cap E\in {\mathcal I}\restriction E\subseteq {\mathcal J} $.
\end{proof}
Note that $\rsub{{\mathcal I}}{{\mathcal J}}$ is not necessarily equivalent to ${\mathcal I}\subseteq{\mathcal J}\ \lor\ {\mathcal I}\perp{\mathcal J}$. Before we give a~counterexample, let us recall that
if there is not a finite subfamily $\mathcal{C}$ of a~family ${\mathcal E}$ with finite $M\setminus \bigcup {\mathcal C}$ then \textbf{ideal generated by} $\boldsymbol{{\mathcal E}}$ is the smallest ideal containing ${\mathcal E}$ and $\text{\rm Fin}$, and we denote this ideal by $\gen{{\mathcal E}}$. So
$$\gen{{\mathcal E}}=\left\{E\in{\mathcal P}(M): E\subseteq^* \bigcup_{i<n} E_n\text{ for some } E_0,\dotso,E_{n-1}\in {\mathcal E} \right\}.$$
Thus $\gen{{\mathcal A}}=\finop{{\mathcal A}}$. Note that $M=\bigcup {\mathcal E}$ if and only if ${\mathcal E}$ is a subbase for $\gen{{\mathcal E}}$. The family ${\mathcal E}\cup\text{\rm Fin}$ is always a subbase for $\gen{{\mathcal E}}$.
Returning back to the~counterexample, let ${\mathcal I}=\gen{\{K\}}$ for some infinite and coinfinite $K\subseteq \omega$. Since $\omega\setminus K\in {\mathcal I}^*$ and ${\mathcal I}\restriction (\omega\setminus K)\subseteq\text{\rm Fin}$, we have $\rsub{{\mathcal I}}{\text{\rm Fin}}$. Evidently, ${\mathcal I}\not\subseteq \text{\rm Fin}$ and $\text{\rm Fin}\not\perp{\mathcal I}$ since ${\mathcal I}$ does not contain any cofinite set. Clearly, the condition $\rsub{{\mathcal I}}{{\mathcal J}}$ is equivalent to
\[
(\forall\kappa\leq\abs{{\mathcal I}})(\forall \mathcal{E}\in[{\mathcal I}]^\kappa)(\exists E\in {\mathcal I}^*)\ \mathcal{E}\restriction E\subseteq {\mathcal J}.
\]
Actually, using the definition and notation from~\cite{stan}, the~condition $\rsub{{\mathcal I}}{{\mathcal J}}$ is equivalent to any of the~following conditions:\footnote{A notion of a $\kappa\text{-}\mathrm{P}({\mathcal J},{\mathcal I})$-ideal in \cite{stan} is a~modification of the $\pid{{\mathcal J}}$-ideal notion. Let ${\mathcal I},{\mathcal J},{\mathcal K}$ be ideals on $M$ and $\kappa$ be a cardinal number. We say that an ideal ${\mathcal K}$ is a $\kappa$-$\mathrm{P}({\mathcal J},{\mathcal I})$-ideal, if for every family $\{K_\alpha\subseteq M:\alpha<\kappa \}\subseteq {\mathcal K}$ there exists $I\in {\mathcal I}$ such that $K_\alpha\setminus I\in {\mathcal J}$ for every $\alpha<\kappa$.} ${\mathcal I}$ is a $\abs{{\mathcal I}}\text{-}\mathrm{P}({\mathcal J},{\mathcal I})$-ideal, ${\mathcal I}$ is a $\mathtt{cof}({\mathcal I})\text{-}\mathrm{P}({\mathcal J},{\mathcal I})$-ideal,\footnote{Let us recall that $\mathtt{cof}({\mathcal I})=\min\{\abs{{\mathcal E}}:{\mathcal E}\subseteq{\mathcal I}\wedge{\mathcal E}\text{ is a base of }{\mathcal I} \}$.} ${\mathcal I}$ is a~$\kappa\text{-}\mathrm{P}({\mathcal J},{\mathcal I})$-ideal for every~$\kappa$.
\par
We say that an ideal ${\mathcal I}$ on $M$ is \textbf{tall} if for each $X\in[M]^\omega$ there exists $I\in{\mathcal I}$ such that $\abs{X\cap I}=\omega$. An~ideal ${\mathcal I}$ is \textbf{nowhere tall}, if
\[
(\forall A\in{\mathcal I}^+)(\exists B\in[A]^\omega)\ {\mathcal I}\restriction B=[B]^{<\omega},
\]
for more see \cite{filtry}. Clearly, $\fin\times\emptyset$ and $\emptyset\times\fin$ are nowhere tall ideals. If an ideal ${\mathcal I}$ is tall and an ideal ${\mathcal J}$ is nowhere tall, then~$\rsub{}{}$ is equivalent to orthogonality:
\begin{lema}\label{nowheretall}
Let ${\mathcal I},{\mathcal J}$ be ideals on $M$ such that ${\mathcal I}$ is tall and ${\mathcal J}$ is nowhere tall. The following statements are equivalent.
\begin{enumerate}
\item[\rm (1)] ${\mathcal I}\perp{\mathcal J}$.
\item[\rm (2)] $\rsub{{\mathcal I}}{{\mathcal J}}$.
\end{enumerate}
\end{lema}
\proof Part (1)$\to$(2) is trivial.
$\neg$(1)$\to\neg$(2). If $E\in {\mathcal I}^*$ then $E\notin {\mathcal J}$ by the assumptions. Ideal ${\mathcal J}$ is nowhere tall and therefore there is infinite $E'\subseteq E$ such that ${\mathcal J}\restriction E'=[E']^{<\omega}$. Since ${\mathcal I}$ is tall, there is an infinite $ D\subseteq E'$ such that $D\in {\mathcal I}$. Then $D\in {\mathcal I}\restriction E$ and $D\notin {\mathcal J}$, i.e., ${\mathcal I}\restriction E\not\subseteq{\mathcal J}$.
\qed
It may seem counterintuitive that a tall ideal can be ``almost'' a subset of an obviously less ``dense'' nowhere tall ideal but this situation may occur. Consider, e.g., bijective function $f\colon\omega\times\omega\to\omega\times\omega$ such that $f({\mathcal{ED}})\perp\fin\times\emptyset$. In this case $\rsub{f({\mathcal{ED}})}{\fin\times\emptyset}$.
Let ${\mathcal I},{\mathcal J}$ be ideals on $M$. One can easily see that if ideals ${\mathcal I},{\mathcal J}$ are not orthogonal then the set
$$
{\mathcal I}\lor {\mathcal J}=\{I\cup J: I\in{\mathcal I}, J\in{\mathcal J} \}
$$
is an ideal on $M$. In fact, $\lor$ is supremum of two non-orthogonal ideals in the family of all ideals on $M$ ordered by inclusion.
We have already commented that if ${\mathcal I}_1$ is a $\pid{{\mathcal J}}$-ideal and ${\mathcal K}\supseteq {\mathcal I}_1$ then it is not necessarily true that ${\mathcal K}$ is a $\pid{{\mathcal J}}$-ideal, see Table~\ref{omxom}. However, we show that if ${\mathcal K}={\mathcal I}_1\lor{\mathcal I}_2$ and ${\mathcal I}_2$ is also a $\pid{{\mathcal J}}$-ideal
then it is the case. Note that the~assumption about ${\mathcal I}_1,{\mathcal I}_2$ not being orthogonal should be read as ${\mathcal I}_1\lor{\mathcal I}_2$ is an ideal. Finally, note that if ${\mathcal K}\supseteq {\mathcal I}_1$ then ${\mathcal K}={\mathcal I}_1\lor{\mathcal K}$, ${\mathcal I}_1,{\mathcal K}$ are non-orthogonal, but it is not necessarily true that ${\mathcal K}\cap {\mathcal I}_1\subseteq{\mathcal J}$.
\begin{prop}\label{differJ}
Let ${\mathcal J}$ be an ideal on $M$ and let ${\mathcal I}_1,{\mathcal I}_2$ be non-orthogonal ideals on $M$ such that ${\mathcal I}_1\cap {\mathcal I}_2\subseteq{\mathcal J}$. The following statements are equivalent.
\begin{enumerate}
\item[\rm (1)] ${\mathcal I}_1$ and ${\mathcal I}_2$ are $\mathrm{P}({\mathcal J})$-ideals.
\item[\rm (2)] ${\mathcal I}_1\lor {\mathcal I}_2$ is a~$\mathrm{P}({\mathcal J})$-ideal.
\end{enumerate}
\end{prop}
\proof (1)$\to$(2) For any family $\{K_n:n\in\omega \}\subseteq {\mathcal I}_1\lor {\mathcal I}_2$ there are families $\{I^1_n\in{\mathcal I}_1:n\in\omega \}\subseteq{\mathcal I}_1$ and $\{I^2_n\in{\mathcal I}_2:n\in\omega \}\subseteq{\mathcal I}_2$, such that $K_n=I^1_n\cup I^2_n$ for each $n$. There are $I^1\in{\mathcal I}_1$, $I^2\in{\mathcal I}_2$ such that $I^1_n\setminus I^1\in{\mathcal J}$ and $I^2_n\setminus I^2\in{\mathcal J}$ for each $n$. Then $K_n\setminus K\in {\mathcal J}$ for each $n$, where $K=I^1\cup I^2$.
(2)$\to$(1) Let $\{I_n:n\in\omega \}\subseteq {\mathcal I}_1\subseteq {\mathcal I}_1\lor{\mathcal I}_2$. There is $I^1\in{\mathcal I}_1$ and $I^2\in{\mathcal I}_2$, such that $I_n\setminus \left(I^1\cup I^2\right)\in {\mathcal J}$ for each $n$. For every $n\in\omega$
$$I_n \setminus I^1\subseteq \left( I_n\setminus \left(I^1\cup I^2\right) \right)\cup \left(I_n\cap I^2\right).$$
$I_n\setminus \left(I^1\cup I^2\right)\in{\mathcal J}$ because we assume (2) and $\left(I_n\cap I^2\right)\in{\mathcal J}$ from the assumption ${\mathcal I}_1\cap {\mathcal I}_2\subseteq {\mathcal J}$. Thus, $I_n\setminus I^1\in{\mathcal J}$ for every $n\in\omega$. Proof for the ideal ${\mathcal I}_2$ is completely analogous.
\qed
Note that the implication (1)$\to$(2) of Proposition~\ref{differJ} does not depend on the assumption ${\mathcal I}_1\cap{\mathcal I}_2\subseteq{\mathcal J}$. On the other hand, it is crucial for the implication (2)$\to$(1). The counterexample are ideals ${\mathcal Sel}$ and $\emptyset\times\fin$ studied in the next section.
By Proposition~\ref{differJ} we obtain the~following.
\begin{corol}\label{finfin}
Let ${\mathcal I}_1,{\mathcal I}_2$ be non-orthogonal ideals on a set $M$ such that ${\mathcal I}_2$ is a $\mathrm{P}({\mathcal I}_1)$-ideal. Then ${\mathcal J}={\mathcal I}_1\lor{\mathcal I}_2$ is a $\mathrm{P}({\mathcal I}_1)$-ideal.
\end{corol}
\par
We conclude the~section with the~connections to cardinal invariants. It can be easily seen that ${\mathcal I}$ is a $\pid{{\mathcal J}}$-ideal if and only if $\mathfrak{b}({\mathcal I},\subseteq^{\mathcal J})\geq\omega_1$\footnote{For any preordered set without maximal elements $\langle X,\leq \rangle$, let
$$
\mathfrak{b}(X,\leq)=\min\{\abs{A}:A\subseteq X, A\text{ is unbounded from above in }X \}.
$$
}
(or ${\mathcal I}$ is not a $\pid{{\mathcal J}}$-ideal if and only if $\mathfrak{b}({\mathcal I},\subseteq^{\mathcal J})=\omega$). Note that for cardinal invariant $$\mathtt{add}^*({\mathcal I})=\min\{\abs{{\mathcal E}}:{\mathcal E}\subseteq{\mathcal I} \land (\forall X\in {\mathcal I})(\exists E\in {\mathcal E})\ E\not\subseteq^* X \}$$ studied in \cite{inv}, we have $\mathfrak{b}({\mathcal I},\subseteq^\text{\rm Fin})=\texttt{add}^*({\mathcal I})$, thus the invariant $\mathfrak{b}({\mathcal I},\subseteq^{\mathcal J})$ can be seen as $``\mathtt{add}^{\mathcal J}({\mathcal I})"$ -- generalization of the invariant $\mathtt{add}^*({\mathcal I})$.
\section{Natural partition and copies of product ideals}\label{natural_finfin}
For each pair among ideals $\text{\rm Fin}\times\emptyset$, ${\mathcal Sel}$, ${\mathcal{ED}}$, $\emptyset\times\text{\rm Fin}$, $\fin\times\fin$, we shall decide whether it satisfies P-property for two ideals or not, see Table~\ref{omxom}. The~second part of the~section is an~initiation of the~study of the~same question for their isomorphic copies. Although all five ideals have been already discussed in the~introduction, including their definition, we include a~more detailed treatment before we reach the~main result of the~present section. Therefore, let us recall another viewpoint on their definition not included in the~introduction.
\par
By $\emptyset$ we denote the~ideal $\{ \emptyset\}$. Here we violate our assumption, that $\text{\rm Fin}$ is a subset of every ideal, but this notation comes in useful, especially when dealing with Fubini products of ideals. Given two ideals ${\mathcal I},{\mathcal J}$ on~$\omega$, the \textbf{Fubini product} ${\mathcal I}\times{\mathcal J}$ is defined by
$${\mathcal I}\times{\mathcal J}=\{A\subseteq\omega\times\omega: \{n:\{m:\langle n,m\rangle\in A \}\notin{\mathcal J} \} \in{\mathcal I}\}.$$
Thus, $\text{\rm Fin}\times\emptyset$, $\emptyset\times\text{\rm Fin}$, $\fin\times\fin$ may be viewed also as Fubini products of ideals in $\{\text{\rm Fin}, \emptyset\}$. That is the~reason for the~choice of their abbreviation. A~direct definition of ${\mathcal Sel}$ and ${\mathcal{ED}}$ can be expressed as follows:
\begin{align*}
& {\mathcal Sel}=\{X\subseteq \omega\times\omega: (\exists m\in\omega)(\forall k\in\omega)\ \abs{\{l:\ev{k,l}\in X\}}\leq m \},\\
& {\mathcal{ED}}=\{X\subseteq \omega\times\omega: (\exists k\in\omega)(\forall^\infty n\in\omega)\ \abs{\{m:\ev{n,m}\in X\}}\leq k \}\\
& \hspace{0.6cm} = \{X\subseteq\omega\times\omega: \limsup_{n\to\infty}\abs{(\{ n\}\times\omega)\cap X}<\infty \}.
\end{align*}
Ideal ${\mathcal E}{\mathcal D}$ can be thought as the ideal on $\omega$ generated by infinite partition of $\omega$ into infinite sets and selectors of this partition. Considering the ideal ${\mathcal Sel}$, if we identify functions with their graphs, we get another equivalent definition ${\mathcal Sel}=\gen{{^\omega\omega}}$.
The~following diagram shows relationships between ideals with respect to inclusion relation.
\begin{center}
\begin{figure}[H]
\centering
\begin{tikzpicture}
\matrix (m) [matrix of math nodes, row sep=3em,
column sep=3em]{
\fin\times\emptyset&{\mathcal{ED}}&\fin\times\fin&\\
\text{\rm Fin} & {\mathcal Sel} &\emptyset\times\fin&\\};
\path[-stealth]
(m-1-1) edge (m-1-2)
(m-2-1) edge (m-1-1) edge (m-2-2)
(m-2-2) edge (m-1-2) edge (m-2-3)
(m-1-2) edge (m-1-3)
(m-2-3) edge (m-1-3);
\end{tikzpicture}
\caption{Inclusion between critical ideals on $\omega\times\omega$.}
\label{inclusion}
\end{figure}
\end{center}
Moreover, recall that
\begin{center}
${\mathcal Sel}={\mathcal{ED}}\cap(\emptyset\times\fin)$,\hspace{0.5cm} ${\mathcal{ED}}=(\text{\rm Fin}\times\emptyset)\lor{\mathcal Sel}$,\hspace{0.5cm} $\text{\rm Fin}\times\text{\rm Fin}=(\text{\rm Fin}\times\emptyset)\lor (\emptyset\times\text{\rm Fin})$.
\end{center}
The~following table shows which pairs~${\mathcal I},{\mathcal J}$ among $\text{\rm Fin}\times\emptyset$, $\emptyset\times\text{\rm Fin}$, $\fin\times\fin$, ${\mathcal Sel}$, and ${\mathcal{ED}}$ satisfy that ${\mathcal I}$ is a~$\pid{{\mathcal J}}$-ideal. The~reasoning can be found below Table~\ref{omxom}. In fact, many of the results come from the~fact that ${\mathcal I}\subseteq {\mathcal J}$ yields ${\mathcal I}$ is a $\pid{{\mathcal J}}$-ideal.
\begin{center}
\begin{table}[H]
\centering
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline
& $\mathrm{P}$ & $\mathrm{P}(\text{\rm Fin}\times\emptyset)$& $\mathrm{P}({\mathcal Sel})$ & $\mathrm{P}({\mathcal E}{\mathcal D})$ & $\mathrm{P}(\emptyset\times\text{\rm Fin})$ & $\mathrm{P}(\text{\rm Fin}\times\text{\rm Fin})$\\
\hline\hline
$\text{\rm Fin}\times\emptyset$ & \xmark & \cmark & \xmark &\cmark & \xmark & \cmark \\
\hline
${\mathcal Sel}$ & \xmark & \xmark & \cmark &\cmark & \cmark & \cmark\\
\hline
${\mathcal E}{\mathcal D}$ & \xmark & \xmark & \xmark &\cmark & \xmark &\cmark\\
\hline
$\emptyset\times\text{\rm Fin}$& \cmark & \cmark & \cmark &\cmark& \cmark & \cmark\\
\hline
$\text{\rm Fin}\times\text{\rm Fin}$ & \xmark & \cmark & \xmark & \cmark & \xmark &\cmark \\
\hline
\end{tabular}
\caption{$\mathrm{P}({\mathcal J})$ interactions between critical ideals on $\omega\times\omega$ }
\label{omxom}
\end{table}
\end{center}
Below we comment row by row in Table~\ref{omxom}. In fact, all the~positive statements are based on either of three following reasons, and their combinations: inclusion between ideals (Proposition~\ref{ort}(d), and Diagram~\ref{inclusion}), the~fact that $\emptyset\times\fin$ is a $\mathrm{P}$-ideal (see, e.g., \cite{Hr}), and finally, $\text{\rm Fin}\times\text{\rm Fin}$ is an ideal $(\text{\rm Fin}\times\emptyset)\lor(\emptyset\times\text{\rm Fin})$ (Lemma~\ref{finfin}). In particular, the~first among the~latter reasons yields that ${\mathcal I}$ is a~$\pid{{\mathcal I}}$-ideal, and therefore these pairs are excluded from the~further considerations (a~diagonal in Table~\ref{omxom}). In case of negative statement, the~corresponding reasoning is included.
\begin{itemize}
\item Since $\emptyset\times\fin$ is a $\mathrm{P}$-ideal, it is also $\pid{{\mathcal I}}$ for all ideals ${\mathcal I}$ on $\omega\times\omega$ by Proposition~\ref{ort}(d).
\item $\text{\rm Fin}\times\emptyset$ is a subset of $\text{\rm Fin}\times\text{\rm Fin}$ and also ${\mathcal{ED}}$, therefore it is a $\pid{\fin\times\fin}$-ideal and $\pid{{\mathcal{ED}}}$-ideal. A sequence $A_n=\{n \}\times\omega$ is a witness for $\text{\rm Fin}\times\emptyset$ being not a $\mathrm{P}(\emptyset\times\text{\rm Fin})$-ideal and since ${\mathcal Sel}\subseteq\emptyset\times\text{\rm Fin}$, we have that $\text{\rm Fin}\times\emptyset$ is not a $\pid{{\mathcal Sel}}$-ideal either.
\item $\text{\rm Fin}\times\text{\rm Fin}$ is an ideal $(\text{\rm Fin}\times\emptyset)\lor(\emptyset\times\text{\rm Fin})$. It follows from Lemma~\ref{finfin} that $\text{\rm Fin}\times\text{\rm Fin}$ is a $\mathrm{P}(\text{\rm Fin}\times\emptyset)$-ideal and since $\text{\rm Fin}\times\emptyset\subseteq{\mathcal E}{\mathcal D}$ we have that $\text{\rm Fin}\times\text{\rm Fin}$ is also a $\mathrm{P}({\mathcal E}{\mathcal D})$-ideal. A sequence $A_n=\{n \}\times\omega$ is a witness for $\text{\rm Fin}\times\text{\rm Fin}$ being not a $\mathrm{P}(\emptyset\times\text{\rm Fin})$-ideal and hence $\text{\rm Fin}\times\text{\rm Fin}$ is not a $\pid{{\mathcal Sel}}$-ideal.
\item ${\mathcal Sel}$ is a subset of $\emptyset\times\fin$, $\fin\times\fin$ and ${\mathcal{ED}}$ and hence is a $\pid{\emptyset\times\fin}$, $\pid{\fin\times\fin}$ and $\pid{{\mathcal{ED}}}$-ideal. A sequence $\omega\times\{n\}$ is a witness for ${\mathcal Sel}$ not being a $\pid{\text{\rm Fin}}$ and $\pid{\fin\times\emptyset}$-ideal.
\item ${\mathcal{ED}}$ is a subset of the ideal $\text{\rm Fin}\times\text{\rm Fin}$, so ${\mathcal{ED}}$ is a $\mathrm{P}(\text{\rm Fin}\times\text{\rm Fin})$-ideal. A sequence $A_n=\{ n \}\times\omega$ is a witness to ${\mathcal E}{\mathcal D}$ being a $\mathrm{P}(\emptyset\times\text{\rm Fin})$-ideal and hence ${\mathcal E}{\mathcal D}$ is not a $\pid{{\mathcal Sel}}$-ideal. ${\mathcal E}{\mathcal D}$ is not a $\mathrm{P}(\text{\rm Fin}\times\emptyset)$-ideal and the counterexample is a sequence $A_n=\omega\times\{n \}$.
\end{itemize}
The~rest of the~paper is devoted to the~investigations when the~isomorphic copies of ideals $\text{\rm Fin}\times\emptyset$, $\emptyset\times\text{\rm Fin}$, $\fin\times\fin$, ${\mathcal Sel}$, and ${\mathcal{ED}}$ satisfy P-property for two ideals. Let us recall that ideals ${\mathcal I}$ and ${\mathcal J}$ on $X$ and $Y$, respectively, are \textbf{isomorphic}, written ${\mathcal I}\simeq{\mathcal J}$, if there exists a bijection $f\colon X\to Y$, such that $I\in{\mathcal I}\equiv f[I]\in{\mathcal J}$ for every $I\in{\mathcal I}$ (or equivalently ${\mathcal J}=f({\mathcal I})=\{f[I]:I\in{\mathcal I} \}$) and we call such a bijection an~\textbf{isomorphism}.
\par
The~situation rapidly changes when it comes to isomorphic copies in comparison with original ideals $\text{\rm Fin}\times\emptyset$, $\emptyset\times\text{\rm Fin}$, $\fin\times\fin$, ${\mathcal Sel}$, and ${\mathcal{ED}}$. Now, before the examination of the situation, we shall make an important remark -- we state our results for ideals defined via arbitrary partition ${\mathcal A}$ (or ${\mathcal B}$), and a particular vertical partition ${\mathcal V}$. However, this is not a setback since the results are true for two arbitrary partitions, i.e., for the general formulation one can think of ${\mathcal V}$ being arbitrary partition.
Let us begin with a~basic observation.
\begin{lema}\label{existujeInkIzo}
For each pair of ideals ${\mathcal I},{\mathcal J}\neq \text{\rm Fin}$ on countable sets such that ${\mathcal I}$ is not tall there is an isomoprhism $f$ such that $f({\mathcal I})\subseteq {\mathcal J}$.
\end{lema}
\proof Since ${\mathcal I}$ is not tall there is an infinite set $X\subseteq M$ such that ${\mathcal I}\restriction X=[X]^{<\omega}$. Consider arbitrary $f$ sending $M\setminus X$ to a set belonging to ${\mathcal J}$ and $X$ to the rest. \qed
Moreover, note that for each pair of ideals ${\mathcal I},{\mathcal J}\neq\text{\rm Fin}$ on $M$ there is an isomorphism $f$ such that $f({\mathcal I})$ is a $\mathrm{P}({\mathcal J})$-ideal. Indeed, let $I\in{\mathcal I}$ be an infinite set and $J\in{\mathcal J}$ be a coinfinite set. It suffices to consider a mapping $f$ such that $f[I]=M\setminus J$ to see that $f({\mathcal I})$ and ${\mathcal J}$ are orthogonal and thus ${\mathcal I}$ is a $\mathrm{P}({\mathcal J})$-ideal. Consequently, for every ${\mathcal J}\neq\text{\rm Fin}$ there are partitions ${\mathcal A},{\mathcal B},\mathcal{C},\mathcal{D}$ of $\omega\times\omega$ such that $\selp{{\mathcal A}}$, $\EDp{{\mathcal A}}, \finop{{\mathcal B}},\ofinp{\mathcal{C}}$ and $\finfinp{\mathcal{\mathcal{D}}}$ are $\mathrm{P}({\mathcal J})$-ideals.
\par
Furthermore, the ideal $\emptyset\times\fin$ is a $\mathrm{P}$-ideal and therefore $\emptyset\times\fin$\footnote{as well as every isomorphic copy of $\emptyset\times\fin$, since isomorphisms preserve $\mathrm{P}$-property} is a $\pid{{\mathcal J}}$-ideal for each ideal ${\mathcal J}$ on $\omega\times\omega$. Thus, $\ofinp{{\mathcal A}}$ is a~$\pid{{\mathcal J}}$-ideal for each ideal ${\mathcal J}$ on $\omega\times\omega$. For this reason, the ideal $\ofinp{{\mathcal A}}$ is omitted from further considerations.
For the next part we will need the following notions. A family ${\mathcal A}\subseteq{\mathcal J}^+$ is~${\mathcal J}$-almost disjoint (${\mathcal J}$-AD) if $A\cap B\in{\mathcal J}$ for each $A,B \in {\mathcal A}$, $A\neq B$. An ${\mathcal J}$-AD family is an ${\mathcal J}$-MAD family if for each $X\in{\mathcal J}^+$ there is an $A\in{\mathcal A}$ such that $X\cap A\in{\mathcal J}^+$ (i.\,e.\ ${\mathcal A}$ is $\subseteq$-maximal among the ${\mathcal J}$-AD families). It is common to use terms AD-family and MAD-family instead of $\text{\rm Fin}$-AD family and $\text{\rm Fin}$-MAD family. It is easy to see that every ideal ${\mathcal I}$ with countable base ${\mathcal B}=\{B_n:n\in\omega \}$ has a countable subbase ${\mathcal S}=\{S_n:n\in\omega \}$ with $S_i\cap S_j=\emptyset$ as $i\neq j$. Indeed, let
$S_0=B_0$ and $S_k=B_k\setminus \bigcup_{l<k} B_l$ for $k>0$. Hence, every ideal with countable base is generated by an AD family. Note that an~ideal $\gen{{\mathcal C}}$ generated by a~family ${\mathcal C}$ is included in ideal~${\mathcal J}$ if and only if ${\mathcal C}\subseteq{\mathcal J}$.
\begin{prop}\label{generovaneAD}
Let ${\mathcal J}$ be an ideal on $M$ and ${\mathcal C}$ be an AD family. The following statements are equivalent.
\begin{enumerate}
\item[\rm (1)] $\gen{{\mathcal C}}$ is a $\mathrm{P}({\mathcal J})$-ideal.
\item[\rm (2)] ${\mathcal C}\setminus{\mathcal J}$ is finite.
\item[\rm (3)] $\rsub{\gen{C}}{{\mathcal J}}$.
\end{enumerate}
\end{prop}
\proof
$\neg$(2)$\to\neg$(1) Let $\abs{{\mathcal C}}=\kappa$ and ${\mathcal C}=\{C_\alpha:\alpha<\kappa \}$ be an enumeration of ${\mathcal C}$. Let $\abs{\{\alpha\in\kappa:C_\alpha\notin{\mathcal J} \}}\geq\omega$. Choose a strictly increasing sequence $\alpha_k$ such that $C_{\alpha_k}\notin{\mathcal J}$ for all $k$. For any $I\in\gen{C}$ pick a $k$ such that $\abs{I\cap C_{\alpha_k}}<\omega$.
(2)$\to$(3) Let $\gen{C}^*\ni I=\omega\setminus \bigcup ({\mathcal C}\setminus{\mathcal J})$.
(3)$\to$(1) Consequence of Proposition~\ref{ortaink}.
\qed
If we combine the~Proposition~\ref{differJ} and~\ref{generovaneAD} we get equivalence of the statements in Corollary~\ref{prehlcor}. Note that $\finop{{\mathcal A}}$ is included in an ideal~${\mathcal J}$ as a subset if and only if $A_n\in {\mathcal J}$ for all $n\in\omega$, compare condition (3) in Corollary~\ref{prehlcor}. On the other hand $\finfinp{{\mathcal A}}$ is not a subset of ${\mathcal J}$ for ${\mathcal J}=\fin\times\emptyset,{\mathcal{ED}},\emptyset\times\fin$ for any~${\mathcal A}$. Indeed, the statement for $\fin\times\emptyset$ and $\emptyset\times\fin$ follows from tallnes of $\finfinp{{\mathcal A}}$. In fact, if ${\mathcal I}$ is an ideal on $M$ and ${\mathcal J}\neq\text{\rm Fin}$ is not a tall ideal, then ${\mathcal I}$ is tall if and only if $f({\mathcal I})\not\subseteq {\mathcal J}$ holds for all bijective mappings $f$. ``Only if" part is trivial. The contrapositive of ``if" part follows from Lemma~\ref{existujeInkIzo}. The fact that $\finfinp{{\mathcal A}}$ is not below ${\mathcal{ED}}$ follows from $\fin\times\fin$ not being Katětov below ${\mathcal{ED}}$, see \cite{hrusakborel}. Finally, the~statement $\rsub{\finfinp{{\mathcal A}}}{{\mathcal J}}$ cannot be added to the~list of equivalent conditions in Corollary~\ref{prehlcor} since taking $A\in(\fin\times\fin)^\ast$, the~ideal $(\fin\times\fin)\restriction A$ is tall, but $(\fin\times\emptyset)\restriction A$ is not the~case.
\begin{corol}\label{prehlcor}
Let ${\mathcal J}$ be ideal on $\omega\times\omega$. The following statements are equivalent.
\begin{itemize}
\item[\rm (1)] $\finop{{\mathcal A}}$ is a $\mathrm{P}({\mathcal J})$-ideal.
\item[\rm (2)] $\finfinp{{\mathcal A}}$ is a $\mathrm{P}({\mathcal J})$-ideal.
\item[\rm (3)] $(\forall^\infty n\in\omega)\ A_n\in {\mathcal J}$.
\item[\rm (4)] $\rsub{\finop{{\mathcal A}}}{{\mathcal J}}$.
\end{itemize}
\end{corol}
\proof
(1)$\equiv$(2) Particular case of Proposition~\ref{differJ} for ${\mathcal I}_1=\finop{{\mathcal A}}$ and ${\mathcal I}_2=\ofinp{{\mathcal A}}$ since $\ofinp{{\mathcal A}}$ is a~P-ideal.
\par
(1)$\equiv$(3) Particular case of Proposition~\ref{generovaneAD} for ${\mathcal I}=\finop{{\mathcal A}}$.
\par
(3)$\to$(4) Straightforward.
\par
(4)$\to$(1) Follows from Proposition~\ref{ortaink}.
\qed
Let us focus on a particular case of Corollary~\ref{prehlcor} for ${\mathcal J}=\fin\times\emptyset$. We offer two more equivalent conditions to $\finop{{\mathcal A}}$ being a~$\mathrm{P}(\fin\times\emptyset)$-ideal in Corollary~\ref{ekvivfub}. Note that $\finop{{\mathcal A}}$ is included in ideal~$\fin\times\emptyset$ if and only if each~$A_n$ has nonempty intersection with finitely many $V_m$. Compare with part~(2) of the~following.
\begin{corol}\label{ekvivfub}
The following statements are equivalent.
\begin{enumerate}[(1)]
\item $\finop{{\mathcal A}}$ is a $\mathrm{P}(\fin\times\emptyset)$-ideal.
\item $(\forall^\infty n)(\forall^\infty m)\ A_n\cap V_m=\emptyset$.
\item $(\exists F\in\text{\rm Fin}^*)\ \finop{{\mathcal A}}\restriction\bigcup_{k\in F}A_k\subseteq\fin\times\emptyset$.
\item $(\forall^\infty n)\ A_n\in\fin\times\emptyset$.
\end{enumerate}
\end{corol}
\proof Straightforward.\qed
Table~\ref{omxom1} lists particular cases of Corollary~\ref{prehlcor} for ideal ${\mathcal J}$ being one of the studied ideals $\text{\rm Fin}\times\emptyset$, ${\mathcal Sel}$, ${\mathcal{ED}}$, $\emptyset\times\text{\rm Fin}$, $\fin\times\fin$. Note that Corollary~\ref{prehlcor} states that $\finop{{\mathcal A}}$ is a $\mathrm{P}({\mathcal J})$-ideal if and only if $\finfinp{{\mathcal A}}$ is a $\mathrm{P}({\mathcal J})$-ideal. Therefore the~conditions in Table~\ref{omxom1} are valid for $\finop{{\mathcal A}}$ as well as for $\finfinp{{\mathcal A}}$. For typographical reasons we omit ``for all but finitely many $n\in\omega$" from conditions in Table~\ref{omxom1}. Note that the~statement with ``for all $n$" is equivalent to $\finop{{\mathcal A}}\subseteq{\mathcal J}$ for ideal ${\mathcal J}$ being one of the ideals in the~table.
\begin{center}
\begin{table}[H]
\centering
\begin{tabular}{|c|c|c|c|c|c|}
\hline
& $\mathrm{P(\finop{{\mathcal B}})}$ & $\mathrm{P}(\selp{{\mathcal B}})$ & $\mathrm{P}(\EDp{{\mathcal B}})$ & $\mathrm{P}(\ofinp{{\mathcal B}})$ & $\mathrm{P}(\finfinp{{\mathcal B}})$ \\
\hline\hline
$\finop{{\mathcal A}}$& {\footnotesize $(\forall^\infty m)$}& {\footnotesize $(\exists k)(\forall m)$ } &{\footnotesize $(\exists k)(\forall^\infty m)$ } & {\footnotesize $(\forall m)$} & {\footnotesize $(\forall^\infty m)$} \\
$\finfinp{{\mathcal A}}$ & {\footnotesize $A_n\cap B_m=\emptyset$} & {\footnotesize $\abs{A_n\cap B_m}<k$} & {\footnotesize $\abs{A_n\cap B_m}<k$} & {\footnotesize $\abs{A_n\cap B_m}<\omega$} & {\footnotesize $\abs{A_n\cap B_m}<\omega$} \\
\hline
\end{tabular}
\caption{The statements hold for all but finitely many $n\in\omega$.}
\label{omxom1}
\end{table}
\end{center}
\section{Selector ideal and towers of functions}\label{S-towers}
In this section we introduce a somewhat instinctive set-theoretic object, which, as we shall see, is tightly interrelated with the studied P-like property. The goal is to find a description of relationships between ideals using natural notions such as partial functions.
It turned out that we are able to formulate necessary conditions for $\selp{{\mathcal A}}$ to be a~$\pid{{\mathcal J}}$-ideal for ${\mathcal J}$ being one of the~ideals among $\fin\times\emptyset$, ${\mathcal Sel}$, ${\mathcal{ED}}$, $\emptyset\times\fin$, $\fin\times\fin$ using these objects called towers of functions. In the~following sections we shall show that towers of functions may be or may be not a~sufficient tool for characterizing the studied P-notion.
If ${\mathcal B}$ is an infinite partition of $\omega\times\omega$ into infinite sets, we call a partial function $f\subseteq\omega\times\omega$ $\boldsymbol{{\mathcal B}}$\textbf{-monochromatic} if there is $B\in{\mathcal B}$ such that $f\subseteq B$.
\begin{defi}
Let $\kappa,\lambda\leq \omega$ be cardinal numbers and ${\mathcal B}$ be an~infinite partition of $\omega\times\omega$ into infinite sets. A~set ${\mathcal T}$ of nonempty partial funcions from $\omega$ to $\omega$ is called $\boldsymbol{(\kappa,\lambda)}$\textbf{-tower of functions} ($\boldsymbol{(\kappa,\lambda)}$\textbf{-TF} for short) if
\begin{itemize}
\item $(\exists a\in[\omega]^\kappa)(\forall g\in {\mathcal T})\ \dom{g}=a$,
\item $\abs{{\mathcal T}}=\lambda$,
\item $g\neq g'\in{\mathcal T}\ \to\ g\cap g'=\emptyset$.
\end{itemize}
If, in addition, each $g\in{\mathcal T}$ is ${\mathcal B}$-monochromatic, we call such a $\tower{\kappa}{\lambda}$ a $\boldsymbol{(\kappa,\lambda)}$\textbf{-tower of} $\boldsymbol{{\mathcal B}}$\textbf{-mo\-no\-chroma\-tic functions} ($\boldsymbol{\tow{\kappa}{\lambda}}$ \textbf{w.r.t. }$\boldsymbol{{\mathcal B}}$ for short).
\end{defi}
Briefly speaking, $\kappa$ is the~size of the~domain of functions in~${\mathcal T}$, and $\lambda$ is the~size of~${\mathcal T}$ itself. Sometimes we do not stress parameters~$\kappa,\lambda,{\mathcal B}$, and we say just towers of monochromatic functions or even just towers. Note that $\emptyset$ is the only $\tow{\kappa}{0}$ for every cardinal number~$\kappa$. On the other hand, there is no $\tow{0}{\lambda}$ for any~$\lambda>0$.
If ${\mathcal T}$ is a $\tow{\kappa}{\lambda}$, we use the following notation:
\begin{itemize}
\item If ${\mathcal T}\neq\emptyset$, then $C_{\mathcal T}=\{B\in{\mathcal B} : (\exists g\in{\mathcal T})\ g\subseteq B\}$ and
$\dom{{\mathcal T}}=\dom{g}$, where $g\in{\mathcal T}$.
\item If ${\mathcal T}=\emptyset$, then $C_{\mathcal T}=\emptyset$ and $\dom{{\mathcal T}}=\emptyset$.
\end{itemize}
We use the term ${\mathcal B}$\textit{-monochromatic} (or just monochromatic) because the partition ${\mathcal B}$ can be seen as a coloring of $\omega\times\omega$. Note that existence of $\tow{\kappa}{\lambda}$ ${\mathcal T}$ implies existence of $\tow{k}{l}$ ${\mathcal T}'$ with $\bigcup {\mathcal T}'\subseteq \bigcup {\mathcal T}$ for all $k<\kappa,l<\lambda$.
\par
We shall be interested in sequences of towers. Thus, we say that a~sequence $\seqn{{\mathcal T}_k}{k}$ of towers is a~sequence of \textbf{essentially different} towers if $C_{{\mathcal T}_i}\cap C_{{\mathcal T}_j}=\emptyset$ and $\dom{{\mathcal T}_i}\cap\dom{{\mathcal T}_j}=\emptyset$ for each $i\neq j$.
\par
Before we show how towers of functions and the P-property for two ideals are interconnected, we will need the following simple lemmas.
\begin{lema}\label{omegaveze}
For any finite ${\mathcal F}\subseteq{^\omega\omega}$ and any $\tower{\omega}{\abs{{\mathcal F}}+1}$ ${\mathcal T}$ there is $g\in {\mathcal T}$ such that $\abs{g\setminus\bigcup{\mathcal F}}=\omega$.
\end{lema}
\begin{proof}
By contradiction. Let ${\mathcal F}\subseteq {^\omega\omega}$ be finite and let ${\mathcal T}$ be an~$\tower{\omega}{\abs{{\mathcal F}}+1}$ such that $\abs{g\setminus \bigcup {\mathcal F}}<\omega$ for every $g\in {\mathcal T}$. Then $g\subseteq^* \bigcup {\mathcal F}$ for each $g\in {\mathcal T}$. So there is $k\in\omega$ such that $\bigcup_{g\in {\mathcal T}} g\restriction\{i>k:i\in\dom{{\mathcal T}} \}\subseteq \bigcup {\mathcal F}$. This is, of course, a contradiction, since $\bigcup_{g\in {\mathcal T}} g\restriction\{i>k:i\in\dom{{\mathcal T}} \}\subseteq \bigcup {\mathcal F}$ is a union of $\abs{{\mathcal F}}+1$ disjoint partial functions sharing a common domain.
\end{proof}
Consequently, for any $A\in {\mathcal Sel}$ such that $A\subseteq\bigcup {\mathcal F}$ there is $g\in {\mathcal T}$ such that $\abs{g\setminus A}=\omega$ as well.
\begin{lema}\label{kkveze}
Let ${\mathcal T}_k=\{g_i^k:i<k \}$ be a $\tower{k}{k}$ for every $k\in\omega$, such that $\dom{{\mathcal T}_i}\cap\dom{{\mathcal T}_j}=\emptyset$ for $i\neq j$ and let ${\mathcal F}\subseteq{^\omega\omega}$ be finite. Then there is $i\leq \abs{{\mathcal F}}$ such that
$$(\forall m\in\omega)(\exists n\geq m)\ \abs{g_i^n\setminus \bigcup{\mathcal F}}\geq m.$$
\end{lema}
\begin{proof}
Let wlog $m\geq 1$ and consider the tower ${\mathcal T}_{m\cdot (\abs{{\mathcal F}}+1)}$. Take $G=\{g_i^{m\cdot (\abs{{\mathcal F}}+1)}:i\leq \abs{{\mathcal F}} \}\subseteq{{\mathcal T}_{m\cdot (\abs{{\mathcal F}}+1)}}$, i.e., a set of the first $\abs{{\mathcal F}}+1$ partial functions of ${\mathcal T}_{m\cdot (\abs{{\mathcal F}}+1)}$. Now, we use the standard pigeonhole principle. We have $m\cdot(\abs{{\mathcal F}}+1)$ points in the domain of $G$ and for each $l\in\dom{G}$ the set $\{g_i^{m\cdot(\abs{{\mathcal F}}+1)}(l):i\leq \abs{{\mathcal F}}\}$ is of cardinality $\abs{{\mathcal F}}+1$, so there is $i(l)\leq \abs{{\mathcal F}}$ such that $\ev{l,g_{i(l)}^{m\cdot(\abs{{\mathcal F}}+1)}(l)}\notin\bigcup{\mathcal F}$. The set $\{\ev{l,g_{i(l)}^{m\cdot(\abs{{\mathcal F}}+1)}(l)}: l\in\dom{G} \}$ contains $m\cdot(\abs{{\mathcal F}}+1)$ points, but we have only $\abs{{\mathcal F}}+1$ possible values of $i(l)$. By pigeon principle, one of these numbers $i(l)$ must have been repeated at least $m$ times. Denote this number by $i_m$. Thus, for any $m\geq 1$ there is $i_m\leq\abs{{\mathcal F}}$ such that
$\abs{g_{i_m}^{m\cdot(\abs{{\mathcal F}}+1)}\setminus\bigcup{\mathcal F}}\geq m$.
A sequence $\ev{i_m:m\in\omega}$ takes only finitely many values, thus, at least one of these values, say $i'$, must be repeated infinitely many times. That is, $(\exists^\infty m\in\omega)\ \abs{g_{i'}^{m\cdot(\abs{{\mathcal F}}+1)}\setminus \bigcup {\mathcal F}}\geq m$.
\end{proof}
Consequently, for any $A\in{\mathcal Sel}$ such that $A\subseteq\bigcup{\mathcal F}$ we have that $(\forall m\in\omega)(\exists n\geq m)\ \abs{g_i^n\setminus A}\geq m$
holds as well.
There are ways how to express the ideals $\selp{{\mathcal A}}$ and $\EDp{{\mathcal A}}$ similar to those in Section~\ref{natural_finfin}. These definitions are probably better in expressing the intuition behind these ideals.
\begin{center}
\begin{tabular}{lcl}
$\selp{{\mathcal A}}$&$=$ &$\{X\subseteq\omega\times\omega: (\exists m\in\omega)(\forall A\in {\mathcal A})\ \abs{X\cap A}\leq m \}$,\\[0.2cm]
$\EDp{{\mathcal A}}$&$=$ & $\{X\subseteq\omega\times\omega: (\exists m\in\omega)\ \{A\in{\mathcal A}: \abs{X\cap A }> m \}\text{ is finite}\}$.
\end{tabular}
\end{center}
\begin{prop}\label{sufficient}
\
\begin{enumerate}[\rm (A)]
\item If ${\mathcal Sel}$ is a $\pid{\finop{{\mathcal B}}}$-ideal then there is no sequence $\seqn{{\mathcal T}_k}{k}$ of essentially different towers with each~${\mathcal T}_k$ being a~$\tow{1}{k}$ w.r.t.~${\mathcal B}$.\footnote{In fact, we shall prove the~equivalence, see Theorem~\ref{ref1}.}
\item If ${\mathcal Sel}$ is a $\pid{\selp{{\mathcal B}}}$-ideal then there is no $\tow{k}{k}$ w.r.t. ${\mathcal B}$ for some $k\in\omega$.
\item If ${\mathcal Sel}$ is a $\pid{\EDp{{\mathcal B}}}$-ideal then there is no sequence $\langle {\mathcal T}_k:k\in\omega \rangle$ of essentially different towers with each~${\mathcal T}_k$ being a~$\tow{k}{k}$ w.r.t. ${\mathcal B}$.
\item If ${\mathcal Sel}$ is a $\pid{\ofinp{{\mathcal B}}}$-ideal then there is no $\tow{\omega}{k}$ w.r.t. ${\mathcal B}$ for some $k\in\omega$.\footnote{In fact, we shall prove the~equivalence, see Theorem~\ref{veze}.}
\item If ${\mathcal Sel}$ is a $\pid{\finfinp{{\mathcal B}}}$-ideal then there is no sequence $\langle {\mathcal T}_k:k\in\omega \rangle$ of essentially different towers with each~${\mathcal T}_k$ being an~$\tow{\omega}{k}$ w.r.t. ${\mathcal B}$.
\end{enumerate}
\end{prop}
\proof
(A) Let $\ev{{\mathcal T}_k:k\in\omega}$ be a sequence of essentially different towers with each~${\mathcal T}_k=\{g^k_i:i<k \}$ being a~$\tow{1}{k}$ w.r.t.~${\mathcal B}$. Define a partial function $f_i=\bigcup_{k>i}g^k_i$ for each $i\in\omega$. We will show that ${\mathcal W}=\{f_i:i\in\omega \}\subseteq{\mathcal Sel}$ is a~witness for ${\mathcal Sel}$ not being a~$\pid{\finop{{\mathcal B}}}$-ideal. First, note that $f_i\cap f_j=\emptyset$ for $i\neq j$ and $\dom{f_i}=^*\dom{f_j}$ for each $i,j<\omega$. Let wlog $X=\bigcup {\mathcal F}\in{\mathcal Sel}$ where ${\mathcal F}$ is a finite subset of ${^\omega\omega}$ consisting of functions. Pick a~set ${\mathcal W}'$ of $\abs{{\mathcal F}}+1$ distinct functions from ${\mathcal W}$. Since we have $\abs{{\mathcal F}}+1$ disjoint partial functions from ${\mathcal Sel}$ on $\bigcap_{f\in{\mathcal W}'}\dom{f}$, there exists $f'\in{\mathcal W}'$ such that $\abs{f'\setminus X}=\omega$. However, ${\mathcal W}$ was chosen in such a~way that infinite parts of its elements do not belong to~$\finop{{\mathcal B}}$, so $f'\setminus X\notin\finop{{\mathcal B}}$. Thus, ${\mathcal Sel}$ is not a $\pid{\finop{{\mathcal B}}}$-ideal.
\par
(B) Assume that there is a $\tow{k}{k}$ w.r.t. ${\mathcal B}$ for all $k\in\omega$. Note that if there is a $\tow{k}{k}$ for each $k\in\omega$, then there is a sequence $\langle {\mathcal T}_k\subseteq\mathcal{P}(\omega\times\omega):k\in\omega\rangle$ such that
\begin{itemize}
\item[\rm (a)] ${\mathcal T}_k=\{g^k_i:i<k\}$ is a $\tow{k}{k}$ for each $k<\omega$,
\item[\rm (b)] $\dom{{\mathcal T}_i}\cap\dom{{\mathcal T}_j}=\emptyset$ for each $i\neq j$.
\end{itemize}
Let us define $f_i=\bigcup_{k>i}g^k_i$ for every $i<\omega$ and ${\mathcal W}=\{f_i:i<\omega \}$. Let $X\in{\mathcal Sel}$, wlog $X=\bigcup{\mathcal F}$ for some finite ${\mathcal F}\subseteq {^\omega\omega}$. Applying Lemma~\ref{kkveze}, we get $i\in\omega$ such that $(\forall m\in\omega)(\exists n\geq m)\ \abs{g_i^n\setminus\bigcup{\mathcal F}}\geq m$. Since $g_i^n$ is a monochromatic part of $f_i$ for all but finitely many $n$, by the definition of $\selp{{\mathcal B}}$ we get $f_i\setminus X\notin\selp{{\mathcal B}}$. Note that it does not matter if the color changes with $n$.
\par
(C) From the assumptions we have that there is a sequence ${\mathcal T}_k=\{g^k_i:i<k \}$ of esentially different towers, each ${\mathcal T}_k$ being a $\tow{k}{k}$ w.r.t. ${\mathcal B}$. Let us define $f_i$'s and ${\mathcal W}$ as in the previous cases and let, again, wlog $X=\bigcup{\mathcal F}$ for some finite $F\subseteq {^\omega\omega}$. The rest of the argumentation follows the lines of the previous case with a single exception -- $g_i^n$ and $g_i^m$ are of different colors for any $n\neq m$. This fact is crucial for $f_i\setminus X\neq \EDp{{\mathcal B}}$.
\par
(D) Assume that there is an~$\tow{\omega}{k}$, ${\mathcal T}_k=\{g_i^k: i<k \}$ w.r.t. ${\mathcal B}$, for any $k\in\omega$ and consider a countable family $\{g_i^k:i,k\in\omega \}$. Let $X=\bigcup{\mathcal F}\in{\mathcal Sel}$ for some finite ${\mathcal F}\subseteq{^\omega\omega}$. Applying Lemma~\ref{omegaveze} we get $i\leq \abs{{\mathcal F}}$ such that $\abs{g_i^{\abs{{\mathcal F}}+1}\setminus\bigcup{\mathcal F}}=\omega$.
\par
(E)
From the assumptions we have that there is a sequence ${\mathcal T}_k=\{g^k_i:i<k \}$ of essentially different towers, each ${\mathcal T}_k$ being an~$\tow{\omega}{k}$ w.r.t. ${\mathcal B}$. Define $f_i$'s as in the cases (A), (B), (C) and let $X=\bigcup{\mathcal F}\in{\mathcal Sel}$ for some finite ${\mathcal F}\subseteq{^\omega\omega}$. Let us denote by $\overline{{\mathcal T}}_k=\{g_i^k:i<\abs{{\mathcal F}}+1\}\subseteq {\mathcal T}_k$ for any $k>\abs{{\mathcal F}}$. Clearly, $\overline{{\mathcal T}}_k$ is an~$\tow{\omega}{\abs{{\mathcal F}}+1}$ for any $k>\abs{{\mathcal F}}$. Hence, by Lemma~\ref{omegaveze} there is a sequence $\ev{i_k:k\in\omega}\in{^\omega\{0,\dotso,\abs{{\mathcal F}} \} }$ such that $\abs{g^k_{i_k}\setminus\bigcup {\mathcal F}}=\omega$ for every $k\in\omega$. Clearly, one of $i_k$'s, say $i'$, must have been repeated infinitely many times, i.e., $(\exists^\infty k\in\omega)\ \abs{g_{i'}^k\setminus\bigcup{\mathcal F}}=\omega$.Thus, $f_{i'}\setminus\bigcup{\mathcal F}\notin\finfinp{{\mathcal B}}$.
\qed
\section{Eventually different ideal and towers are sometimes enough}\label{towers_equiv}
As we have already mentioned, towers of functions in some cases describe not only a necessary condition, but also a sufficient condition for the~ideal ${\mathcal Sel}$ to be a~$\mathrm{P}({\mathcal J})$-ideal. We shall try to point out these cases and show that sometimes they describe that the~ideal ${\mathcal{ED}}$ is a~$\mathrm{P}({\mathcal J})$-ideal as well. We begin with an~auxiliary assertion.
Let us recall the fact that $\EDp{{\mathcal A}}\not\subseteq\finop{{\mathcal B}}$ for any partition ${\mathcal A}$ because of the tallness of $\EDp{{\mathcal A}}$. However, by Lemma~\ref{existujeInkIzo} there are partitions~${\mathcal A}$ such that $\selp{{\mathcal A}}\subseteq\finop{{\mathcal B}}$.
\begin{theorem}\label{ref1}
The following statements are equivalent.
\begin{itemize}
\item[\rm (1)] ${\mathcal Sel}$ is a $\pid{\finop{{\mathcal B}}}$-ideal.
\item[\rm (2)] ${\mathcal{ED}}$ is a $\pid{\finop{{\mathcal B}}}$-ideal.
\item[\rm (3)] ${\mathcal{ED}}\perp\finop{{\mathcal B}}$.
\item[\rm (4)] $\rsub{{\mathcal{ED}}}{\finop{{\mathcal B}}}$.
\item[\rm (5)] $\rsub{{\mathcal Sel}}{\finop{{\mathcal B}}}$.
\item[\rm (6)] ${\mathcal Sel},\fin\times\emptyset$ are $\pid{\finop{{\mathcal B}}}$-ideal.
\item[\rm (7)] $\omega\times\omega$ can be covered by finitely many verticals, ${B_n}$'s, and functions in~${}^{\omega}{\omega}$.
\item[\rm (8)] There is no sequence $\seqn{{\mathcal T}_k}{k}$ of essentially different towers with each~${\mathcal T}_k$ being a~$\tow{1}{k}$ w.r.t.~${\mathcal B}$.
\end{itemize}
\end{theorem}
\proof (2)$\equiv$(6), and (2)$\to$(1) follow directly from Proposition~\ref{differJ}. Part (7) is just (3) restated in terms of generating families. Implications (3)$\to$(4) and (5)$\to$(1) follow directly from Proposition~\ref{ortaink}. Part (3)$\to$(2) follows from Lemma~\ref{ort}. The~implication (1)$\to$(8) is shown in Proposition~\ref{sufficient}.
\par
(4)$\to$(5) Let $F\in{\mathcal{ED}}^*$ be a witness for (4). If $F\in{\mathcal Sel}^*$ we are done. If $(\omega\times\omega)\setminus F$ contains infinite sections of ${V_n}$'s it is enough to add this sections to $F$ and form $F'$. It is not difficult to to see that $F'\in{\mathcal Sel}^*$ and since ${\mathcal Sel}\restriction V_n=[V_n]^{<\omega}$ for each $V_n$, if $E\in{\mathcal Sel}\restriction F'$ then $E$ is just union of a set belonging to ${\mathcal Sel}\restriction F\subseteq\finop{{\mathcal B}}$, and some finite set from ${\mathcal Sel}\restriction \bigcup_{n\in G} V_n$ for some finite $G\subseteq\omega$.
\par
$\neg$(3)$\to\neg$(8) Let us assume that ${\mathcal{ED}}$ and $\finop{{\mathcal B}}$ are not orthogonal. We inductively construct a~sequence $\seqn{{\mathcal T}_k}{k}$ of essentially different towers with each~${\mathcal T}_k$ being a~$\tow{1}{k}$ w.r.t.~${\mathcal B}$. Pick ${\mathcal T}_1$ arbitrarily contained in $V_0\cap B_{k_1}$ for some $k_1\in\omega$. Set $m_0=0$. For each $n\geq 2$, let
$$
X_n=\bigcup_{k>k_{n-1}}B_k.
$$
We have $X_n\notin{\mathcal{ED}}$ since ${\mathcal{ED}}$ and $\finop{{\mathcal B}}$ are not orthogonal. Consequently, $\limsup_{m\to\infty}\abs{V_m \cap X_n}=\infty$, so there is an~$m_n>m_{n-1}$ such that $\abs{V_{m_n}\cap X_n}\geq n+1$. Pick ${\mathcal T}_n\subseteq V_{m_n}\cap X_n$ of size~$n+1$ arbitrarily, and set $k_n>k_{n-1}$ to be a~number with ${\mathcal T}_n\subseteq\bigcup_{k\leq k_n}B_k$. It is straightforward that the~construction leads to the~required sequence $\seqn{{\mathcal T}_k}{k}$ of essentially different towers.
\qed
One can notice that the~condition ${\mathcal Sel}\perp\finop{{\mathcal B}}$ is not listed in~Theorem~\ref{ref1}. In the following we shall provide the reason why it cannot be added to the list.
\begin{lema}\label{selroz}
If ${\mathcal Sel}\perp\finop{{\mathcal B}}$ then $V_n\in\finop{{\mathcal B}}$ for all but finitely many $n\in\omega$.
\end{lema}
\proof It suffices to notice that every $E\in{\mathcal Sel}$ has only finite intersection with each $V_n$. If $E\in{\mathcal Sel}$ is such that $(\omega\times\omega)\setminus E\in \finop{{\mathcal B}}$ then $V_n\setminus E\subseteq(\omega\times\omega)\setminus E\in \finop{{\mathcal B}}$. The rest of the proof is implied by the fact $V_n\setminus E=^* V_n$.\qed
\begin{prop}
If ${\mathcal Sel}\perp \finop{{\mathcal B}}$ then ${\mathcal{ED}}$ is a $\pid{\finop{{\mathcal B}}}$-ideal. The converse is not generally true.
\end{prop}
\proof
It follows from Lemma~\ref{selroz} that $\fin\times\emptyset$ is a $\pid{\finop{{\mathcal B}}}$-ideal. Ideal ${\mathcal Sel}$ is also a $\pid{\finop{{\mathcal B}}}$-ideal because of the orthogonality. Now it suffices to apply Proposition~\ref{differJ}. For the second part let us define ${\mathcal B}$ in the following way: $B_0=\bigcup_{n>0}(\{n\}\times\omega)$, $\bigcup_{n>0}B_n=\{0\}\times \omega$. It is easy to see that ${\mathcal{ED}}\perp \finop{{\mathcal B}}$ but ${\mathcal Sel}\subseteq \finop{{\mathcal B}}$. \qed
Another consequence of Theorem~\ref{ref1} states that if ${\mathcal{ED}}$ is a $\pid{\finop{{\mathcal B}}}$-ideal then $\finop{{\mathcal B}}$ is a $\pid{{\mathcal{ED}}}$-ideal. Note that the~opposite implication is not true in general, since $\fin\times\emptyset\subseteq{\mathcal{ED}}$.
\par
Towers of functions describe that ${\mathcal Sel}$ is a~$\pid{\ofinp{{\mathcal B}}}$-ideal.
\begin{theorem}\label{veze}
The following statements are equivalent.
\begin{itemize}
\item[\rm (1)] ${\mathcal Sel}$ is a $\pid{\ofinp{{\mathcal B}}}$.
\item[\rm (2)] $\rsub{{\mathcal Sel}}{\ofinp{{\mathcal B}}}$.
\item[\rm (3)] There is $k\in\omega$ such that there is no $\tow{\omega}{m}$ w.r.t. ${\mathcal B}$ for every $m>k$.
\item[\rm (4)] $(\forall\mathcal{E}\in[{^\omega\omega}]^\omega)(\exists E\in[{^\omega\omega}]^{<\omega})(\forall f\in\mathcal{E})(\forall B\in{\mathcal B})\ \abs{(f\cap B)\setminus \bigcup E}<\omega$.
\end{itemize}
\end{theorem}
\proof
$\neg$(3)$\to\neg$(1) It is Proposition~\ref{sufficient}.
(3)$\to$(2) For each (color) $B\in {\mathcal B}$ let $$X_B^i=\{n:\abs{V_n\cap B}=i\},\ \ i\in\omega+1\setminus\{ 0\},$$
and $I_B=\{i\leq k: \abs{X_B^i}=\omega \}$. Note that we may omit the case $i>k$ since $X^i_B$ would be always finite by the assumption. If $i\in I_B$ then there is an $\tow{\omega}{i}$ $g^1_{B,i},\dotso, g^i_{B,i}$ such that $\{\langle n,m \rangle\in B: n\in X_B^i \}=\bigcup_{j=1}^ig^j_{B,i}$. Let us define a countable family ${\mathcal G}=\{g^j_{B,i}:B\in{\mathcal B}, i\in I_B, 1\leq j\leq i \}$ and consider an arbitrary bijective enumeration of ${\mathcal G}=\{g_j:j<\omega \}$. The family ${\mathcal G}$ consists of mutually disjoint monochromatic functions. We will show that there is a set of at most $k$ functions $f_j\in{^\omega\omega}$ such that $g_i\setminus\bigcup_{j<k}f_j$ is finite for each $i$.
Define \begin{alignat*}{2}
& h_i=g_i, \quad && \text{ for }i<k,\\
& h_i=g_i \setminus \left(g_i\restriction\bigcup_{\{m_1,\dotso, m_k \}\in[i]^k}\left(\bigcap_{j=1}^k \dom{g_{m_j}}\cap\dom{g_i} \right)\right),\quad && \text{ for } i\geq k.
\end{alignat*}
Notice that the set $\bigcap_{j=1}^k \dom{g_{m_j}}\cap\dom{g_i}$ -- an intersection of $g_i$ with a possible $\tow{\omega}{k}$ is always finite, since there does not exist an~$\tow{\omega}{k+1}$. As a consequence a finite union $$\bigcup_{\{m_1,\dotso, m_k \}\in[i]^k}\left(\bigcap_{j=1}^k \dom{g_{m_j}}\cap\dom{g_i} \right)$$ is also a finite set, thus $h_i=^*g_i$. We show by contradiction that the set $H=\bigcup_{i<\omega} h_i$ is equal to union of at most $k$ partial functions $f_j$, i.e., $\abs{V_n\cap H}\leq k$ for all $n\in\omega$. If $\abs{V_n\cap H}>k$ for some $n$, then there are distinct $i_1,\dotso, i_{k+1}$ such that $n\in \dom{h_{i_1}},\dotso,\dom{h_{i_{k+1}}}$ and consequently $n\in \dom{g_{i_1}},\dotso,\dom{g_{i_{k+1}}}$. Let wlog $i_{k+1}=\max\{i_1,\dotso,i_{k+1} \}$. Then $\{i_1,\dotso,i_k \}\in[i_{k+1}]^k$ and $n\in\bigcap_{j=1}^k\dom{g_{i_j}}\cap \dom{g_{i_{k+1}}}$, thus $n\notin\dom{h_{i_{k+1}}}$ by the definition of $h_{i_{k+1}}$, a contradiction. Moreover, $g_i\setminus H$ is finite since $g_i=^*h_i$ for each $i\in\omega$.
Now, it suffices to show that $H\in{\mathcal Sel}$ is a witness for $\rsub{{\mathcal Sel}}{\ofinp{{\mathcal B}}}$. We will show that $f\setminus H\in\ofinp{{\mathcal B}}$ for every function $f\in{^\omega\omega}$, which is a subbase of ${\mathcal Sel}$, consequently the relation $\rsub{{\mathcal Sel}}{\ofinp{{\mathcal B}}}$ holds true. By the definition, $f\setminus H\in\ofinp{{\mathcal B}}$ if and only if $(f\cap B)\setminus H$ is finite for every $B\in{\mathcal B}$. First, note that not only $X_B^i$ is finite for every $i>k$ but also $\bigcup_{i\notin I_B}X_B^i=\bigcup_{i\in \omega+1\setminus (I_B\cup\{0 \})}X_B^i\in \text{\rm Fin}$, otherwise there would exist an~$\tow{\omega}{k+1}$ (there would be infinitely many columns with at least $k+1$ points in~$B$). Denote by
$$C_B=\bigcup_{i\in I_B}\{\langle n,m \rangle\in B:n\in X_B^i \},\ \ D_B=B\setminus C_B.$$
Then $B=C_B\cup D_B$.
We know that $\{\langle n,m \rangle\in B: n\in X_B^i \}=\bigcup_{j=1}^ig^j_{B,i}$ and $I_B$ is finite, thus $C_B$ is finite union of monochromatic towers, i.e., there are $m<\omega$ and $i_0,\dotso, i_{m-1}<\omega$ such that $C_B= \bigcup_{j<m} g_j$, so $C_B\setminus H\subseteq \bigcup_{j<m} g_j\setminus H$ is a finite set. Clearly, $\mathrm{proj}_1 D_B=\bigcup_{i\in \omega+1\setminus (I_B\cup\{0 \})}X_B^i$, which is a finite set, then $f\cap D_B$ is finite because $\dom{f}\cap \mathrm{proj}_1 D_B\in\text{\rm Fin}$. Then
$$(f\cap B)\setminus H\subseteq (C_B\setminus H) \cup (f\cap D_B)\in \text{\rm Fin}.$$
is finite.
(1)$\equiv$(4) Implication (2)$\to$(4) is trivial. For the other implication let $\{I_n:n\in\omega \}\subseteq {\mathcal Sel}$. By the definition of ${\mathcal Sel}$, for each $n$ there is $m_n$ and $f^n_1,\dotso,f^n_{m_n}$ such that $I_n=\bigcup_{j<m_n}f^n_j$. Set $\mathcal{E}=\{f^n_j:n,j<\omega \}$ and apply (4). Then $I_n\setminus\bigcup E=\bigcup_{j<m_n}\left(f_j\setminus\bigcup E\right)$. For $j<m_n$ we know that $(f_j\cap A)\setminus\bigcup E$ is finite for each $A\in {\mathcal A}$, i.e., $f_j\setminus\bigcup E\in\finop{{\mathcal A}}$. Thus, $I_n\setminus\bigcup E$ is finite union of sets from $\finop{{\mathcal A}}$, thus, a set from $\finop{{\mathcal A}}$.
\qed
Although ideal ${\mathcal{ED}}$ was added to the~list of equivalent conditions in Theorem~\ref{ref1}, Theorem~\ref{veze} is not the case as one can see in Table~\ref{omxom}. However, ${\mathcal{ED}}$ is a $\pid{\ofinp{{\mathcal B}}}$-ideal if and only if both ideals are orthogonal, as one can see below.
\begin{theorem}\label{ed_ofin}
The following statements are equivalent.
\begin{itemize}
\item [\rm (1)] ${\mathcal{ED}}$ is a $\pid{\ofinp{{\mathcal B}}}$-ideal.
\item[\rm (2)] $\rsub{{\mathcal{ED}}}{\ofinp{{\mathcal B}}}$.
\item[\rm (3)] ${\mathcal{ED}}\perp\ofinp{{\mathcal B}}$.
\item[\rm (4)] $\fin\times\emptyset$ and ${\mathcal Sel}$ are $\pid{\ofinp{{\mathcal B}}}$-ideals.
\item[\rm (5)] $\rsub{\fin\times\emptyset}{\ofinp{{\mathcal B}}}$ and $\rsub{{\mathcal Sel}}{\ofinp{{\mathcal B}}}$.
\end{itemize}
\end{theorem}
\proof
(1)$\to$(4) follows from Proposition~\ref{differJ}. (4)$\to$(5) follows from Theorem~\ref{veze} and Corollary~\ref{prehlcor}.
For (5)$\to$(2) it is sufficient to note the following -- it can be easily proved that if $\rsub{{\mathcal I}_1}{{\mathcal J}}$ and $\rsub{{\mathcal I}_2}{{\mathcal J}}$ where ${\mathcal I}_1\not\perp{\mathcal I}_2$, then $\rsub{{\mathcal I}_1\lor{\mathcal I}_2}{{\mathcal J}}$.
(2)$\equiv$(3) is a consequence of Lemma~\ref{nowheretall}.
(3)$\to$(1) Follows from Lemma~\ref{ort}.
\qed
At this point it may seem that $\rsub{{\mathcal I}}{{\mathcal J}}$ and ${\mathcal I}$ is a $\pid{{\mathcal J}}$-ideal are always equivalent but this, in fact, is not the case. Consider, e.g., $\emptyset\times\fin$ and $\fin\times\emptyset$. Clearly, $\emptyset\times\fin\not\subseteq^\upharpoonright \fin\times\emptyset$ but $\emptyset\times\fin$ is a $\pid{\fin\times\emptyset}$-ideal.
\section{Towers of functions are not enough}
In contrary to Section~\ref{towers_equiv}, towers of functions are not always sufficient to fully describe the~$\pid{{\mathcal J}}$-property for ${\mathcal Sel}$. The~whole section is devoted to the~counterexample, its~construction, and a~proof that it possesses the~required properties.
\par
We shall define a~partition ${\mathcal E}=\set{A_{j,i}}{i,j\in\omega}\cup \{B_i:i\in \omega \}$ of~$\omega\times\omega$ into infinite sets.
Let us fix an~auxiliary partition~${\mathcal D}=\set{D_k}{k\in\omega}$ of~$\omega\setminus \{0 \}$ into infinite sets.
\begin{itemize}
\item $\langle m,i\rangle \in A_{j,i}$ if and only if $m$ is the~$(j+ki)$-th element of~$D_k$,
\item $\bigcup \{B_i:i\in\omega\} \supseteq \{0 \}\times\omega$.
\item The~elements of the~remaining subset of~$\omega\times\omega$, say $\{w_i:i\in\omega \}$, are distributed one-by-one into sets~$B_i$, i.e., $w_i\in B_i$ for all $i\in\omega$.
\end{itemize}
The last condition implies that if $m$ is $(j+ki)$-th element of $D_k$ for $-ki\leq j<0$ then $\ev{m,i}\in \bigcup \{B_i:i\in\omega \}$.
If a point $\ev{x,y}$ is colored with $A_{j,i}$, then $y=i$ because in the definition of ${\mathcal E}$ the color $A_{j,i}$ is assigned only to points whose second coordinate equals to $i$. The parameter $i$ in the notation $A_{j,i}$ thus represents the number of the row in which the color $A_{j,i}$ occurs. Consequently, $A_{j,i}\cap A_{j',i'}=\emptyset$ whenever $i\neq i'$. Also note that $\abs{A_{j,i}\cap (D_k\times\{ i\})}=1$ for all $i,j.k\in\omega$, that is, there is only one point of $\omega\times\omega$ ``above'' $D_k$ colored with $A_{j,i}$.
\par
The~construction of coloring~${\mathcal E}$ is done in a~way to avoid towers of functions.
\par\vspace{0.3cm}
\noindent \textbf{Observation.} There is no $\tow{k}{k}$ for $k\geq 3$ w.r.t. ${\mathcal E}$.
\proof
By the definition of ${\mathcal E}$ for any $i\in\omega$ there is exactly one element $\ev{x,y}$ of $\omega\times\omega$ with $x>0$ such that $\ev{x,y}\in B_i$. Hence, if $f$ is a function with $f\subseteq B_i$ for some $i\in\omega$, then $\abs{\dom{f}}\leq 2$. Thus, functions colored with colors from $\{B_i:i\in\omega \}$ cannot be elements of $\tow{3}{3}$s.
As a result, it is enough to consider the colors $\{A_{j,i}:i,j,\in\omega \}$ and the set $(\omega\times\omega)\setminus (\{ 0\}\times\omega)$. With this setup, we will show that there is not even $\tow{2}{2}$, which implies non-existence of $\tow{3}{3}$. Let $m_0, m_1\in\omega\setminus\{ 0\}, m_0\neq m_1,i_0,i_1\in\omega$ be such that $\ev{m_0,i_0},\ev{m_1,i_1}\in\omega\times\omega$ have the same color, i.e., there are $j,i\in\omega$ such that $\ev{m_0,i_0},\ev{m_1,i_1}\in A_{j,i}$. As mentioned previously, $A_{j,i}$ occurs only in the row $\omega\times \{ i\}$, thus $i_0=i_1=i$. So we have $\ev{m_0,i},\ev{m_1,i}\in A_{j,i}$.
Now, note that there are $k_0\neq k_1$ such that $m_0\in D_{k_0}$ and $m_1\in D_{k_1}$. If we assumed, by way of contradiction, that there is $k\in\omega$ such that $m_0, m_1\in D_k$, then $m_0$ is $(j+ki)$-th element of $D_k$ and so is $m_1$, a contradiction because $m_0\neq m_1$.
By the definition $m_0$ is $(j+k_0i)$-th element of $D_{k_0}$ and $m_1$ is $(j+k_1i)$-th element of $D_{k_1}$. Consider now $\ev{m_0,i'},\ev{m_1,i'}$ for some $i'\neq i$ and assume they are colored with the same color $A_{j',i'}$. Then
$$m_0\text{ is }(j'+k_0i')\text{-th element of }D_{k_0},$$
$$m_1\text{ is }(j'+k_1i')\text{-th element of }D_{k_1}.$$
So we have that $m_0$ is $(j+k_0i)$-th and also $(j'+k_0i')$-th element of $D_{k_0}$, $m_1$ is $(j+k_1i)$-th and also $(j'+k_1i')$-th element of $D_{k_1}$, i.e.,
$$j'+k_0i'=j+k_0i\ \land\ j'+k_1i'=j+k_1i$$
which is equivalent to
$$i'(k_0-k_1)=i(k_0-k_1)\ \equiv\ i=i',$$
a contradiction. \qed
We will show that ${\mathcal Sel}$ is not a~$\pid{\selp{{\mathcal E}}}$-ideal, i.e., the necessary condition stated in Proposition~\ref{sufficient} in terms of existence of towers of monochromatic functions is not a sufficient condition.
\begin{theorem}\label{tazkaThm}
${\mathcal Sel}$ is not a $\pid{\selp{{\mathcal E}}}$-ideal.
\end{theorem}
\proof By contradiction. Assume ${\mathcal Sel}$ is $\pid{\selp{{\mathcal E}}}$-ideal. Set $f_i=\omega\times\{ i\}$. We will show that ${\mathcal W}=\{f_i:i\in\omega \}$ is a witness for ${\mathcal Sel}$ being not a $\pid{\selp{{\mathcal E}}}$-ideal.
Since ${\mathcal Sel}$ is a $\pid{\selp{{\mathcal E}}}$-ideal, there is $X\in{\mathcal Sel}$, $X\subseteq \bigcup{\mathcal F}$ for some ${\mathcal F}\subseteq {^\omega\omega},\abs{{\mathcal F}}<\omega$ such that $f_i\setminus X\in\selp{{\mathcal E}}$ for all $i\in\omega$, i.e., there is $\langle k_i\in\omega:i<\omega \rangle$ with $\abs{(f_i\setminus X)\cap A}\leq k_i$ for all $A\in{\mathcal E}$ and $i\in\omega$.
\vspace{0.3cm}
Now, we introduce some new notation and definitions for the~sake of intelligibility of the~proof. We define a~function assigning a~color lying below a~selected color on set~$D_k$, i.e., function $(\cdot\downarrow\cdot)$, by setting
$$(A_{j,i}\downarrow D_k)=A_{j',i-1}$$
if $\proj{1}{A_{j,i}}\cap D_k=\proj{1}{A_{j',i-1}}\cap D_k$.
\begin{claim}\label{claim1}
$(A_{j,i+1}\downarrow D_k) = A_{j+k,i}$ for all $i,j,k\in\omega$.
\end{claim}
\proofofclaim Let $\ev{m,i+1}\in A_{j,i+1}$ with $m\in D_k$, i.e., $m$ is $(j+k(i+1))$-th element of $D_k$. One can easily see there is $j'\in\omega$ such that $(A_{j,i+1}\downarrow D_k)=A_{j',i}$. By the definition, $m$ is also $(j'+ki)$-th element of $D_k$. We have that $$j+k(i+1)=j'+ki\ \equiv\ j'=j+k.$$ \qed
An~integer interval $\{x\in\omega: a\leq x\leq b \}$ with endpoints $a,b\in\omega$ is denoted by $\boldsymbol{[a,b]}$. If $s\leq t,u\leq v$ then the~set
$$\bigcup_{j=s}^t A_{j,i}\cap \bigcup_{k=u}^v\left(D_k\times\{ i\}\right)$$
is called an~$\boldsymbol{\chain{i}{s,t}{u,v}}$. Note that an~$\chain{i}{s,t}{u,v}$ is a~finite function, which can be partitioned into $d(a)=\abs{[u,v]}=v-u+1$ many functions $g_i$, $i<d(a)$, such that $g_i$ is defined on $D_{u+i}$ and each $g_i$ contains $l(a)=\abs{[s,t]}=t-s+1$ many elements.
\begin{claim}\label{claim2}
For every $\chain{i+1}{s,t}{u,v}$ $a$ there is $\chain{i}{s',t'}{u,v}$ $b$ with $l(b)=l(a)-d(a)+1$ and $\dom{b}\subseteq \dom{a}$.
\end{claim}
\proofofclaim Denote by $L=l(a)-d(a)+1$. Consider $[(t-L+1)-n,t-n]$ and $D_{u+n}$ for $n<d(a)$.
Note that for any $n<d(a)$
$$a_n\coloneqq\bigcup_{\substack{j=(t-L+1)-n \\ =t-n-(L-1)}}^{t-n}A_{j,i+1}\cap (D_{u+n}\times\{i+1\})\subseteq a,$$
because $t-n\leq t$ and $(t-L+1)-n>(t-L+1)-d(a)=(t-(l(a)-d(a)+1)+1)-d(a)=t-l(a)=t-(t-s+1)=s-1$, so $(t-L+1)-n\geq s$. Also, Claim~1 gives us $$(A_{t-n-o,i+1}\downarrow D_{u+n})=A_{t+u-o,i}\ \text{for all }o<L.$$
We have that
$$\bigcup_{j=t-n-(L-1)}^{t-n} (A_{j,i+1}\downarrow D_{u+n})\cap (D_{u+n}\times\{ i\})=\bigcup_{o=0}^{L-1}(A_{t-n-o,i+1}\downarrow D_{u+n})\cap (D_{u+n}\times\{ i\})=$$
$$=\bigcup_{o=0}^{L-1}A_{t+u-o,i}\cap (D_{u+n}\times\{i \}).$$
By the definition of $(\cdot\downarrow\cdot)$ we have
\begin{equation}\tag{*}
\proj{1}{a_n}=\proj{1}{\bigcup_{o=0}^{L-1}A_{t+u-o,i}\cap (D_{u+n}\times\{i \})}.
\end{equation}
Set $$b=\bigcup_{n<d(a)}\left( \bigcup_{o=0}^{L-1}A_{t+u-o,i}\cap (D_{u+n}\times\{i \})\right)=\bigcup_{o=0}^{L-1}A_{t+u-o,i}\cap\bigcup_{n=0}^{d(a)-1}(D_{u+n}\times\{i \})=$$
$$=\bigcup_{o=0}^{L-1}A_{t+u-o,i}\cap \bigcup_{k=u}^v(D_k\times\{ i\}).$$
The set $b$ is $\chain{i}{s',t'}{u,v}$ with $s'=t+u-(L-1), t'=t+u$, so $l(b)=t'-s'+1=L=l(a)-d(a)+1$.
Using disjointness of $D_k$'s, fact that $a_n\subseteq a$ for $n<d(a)$ and (*) we have that
$$\dom{b}=\proj{1}{\bigcup_{n<d(a)}\left( \bigcup_{o=0}^{L-1}A_{t+u-o,i}\cap (D_{u+n}\times\{i \})\right)}\subseteq \proj{1}{a}=\dom{a}.$$ \qed
The $\chain{i}{s,t}{u,v}$ $a$ is \textbf{$\boldsymbol{X}$-covered} if $a\subseteq X$. The color $A_{j,i}$ is \textbf{$\boldsymbol{X}$-covered on $\boldsymbol{[u,v]}$ (on $\boldsymbol{D_k})$} if $$A_{j,i}\cap\bigcup_{k=u}^v(D_k\times\{ i\})\subseteq X\ \ \ (A_{j,i}\cap (D_k\times\{ i\})\subseteq X \text{ respectively}).$$
\begin{claim}\label{claim3}
For every $\chain{i}{s',t'}{u,v}$ $b$ with $d(b)\geq (k_i+1)^{l(b)}$ there is an $X$-covered $\chain{i}{s',t'}{u',v'}$ $c$ such that $d(c)=\floor*{\frac{d(b)}{(k_i+1)^{l(b)}}}$ and $c\subseteq b$.
\end{claim}
Before we prove this claim, let us point out that $l(c)=l(b)$. Also, we will need the following lemmas.
\begin{lema}\label{pravdivost}
Let $[u,v]$ be an integer interval with $d\coloneqq \abs{[u,v]}$. If $\varphi(x)$ is true for all but (at most) $k$ points in $[u,v]$, then there is an interval $[u',v']\subseteq [u,v]$ with $\abs{[u',v']}=\floor*{\frac{d}{k+1}}$, such that $\varphi(x)$ is true on $[u',v']$.
\end{lema}
\proof There exist $(k+1)$ disjoint subintervals of $[u,v]$ with cardinality $\floor*{\frac{d}{k+1}}$. Since $\varphi(x)$ is not true for (at most) $k$ points in $[u,v]$, $\varphi(x)$ must be true on (at least) one of these $(k+1)$ intervals by the pigeonhole principle.\qed
\begin{lema}[folklore]\label{celacast}
If $a\in\mathbb{Z},b,c,\in\omega\setminus\{ 0\}$, then $\floor*{\frac{\floor*{\frac{a}{b}}}{c}}=\floor*{\frac{a}{bc}}$.\qed
\end{lema}
\proofofclaim By the assumptions $\abs{(f_i\cap A_{j,i})\setminus X}=\abs{((\omega\times\{ i\})\cap A_{j,i})\setminus X}\leq k_i$ for all $j\in\omega$. We will inductively show that there are intervals $[u_j, v_j], s'\leq j\leq t'$ such that
\begin{itemize}
\item $[u_{j+1},v_{j+1}]\subseteq [u_j,v_j]$ for $s'\leq j<t$,
\item $A_{j,i}$ is $X$-covered on $[u_j,v_j]$ (hence, $A_{j,i}$ is $X$-covered also on $[u_{j'},v_{j'}]$ for $j'>j$ and $A_{j,i}$ is $X$-covered on $[u_{t'},v_{t'}]$ for every $j\in[s',t']$),
\item $\abs{[u_j, v_j]}=v_j-u_j+1=\floor*{\frac{d(b)}{(k_i+1)^{(j+1)-s'}}}$ for $j\in [s',t']$.
\end{itemize}
Induction:
\begin{itemize}
\item $j=s':$ Let $\varphi(k)$ be ``$A_{j,i}$ is $X$-covered on $D_k$''. By the assumptions $\varphi(k)$ is true for all but (wlog) $k_i$ points in $[u,v]$. After application of Lemma~\ref{pravdivost} we get $[u_{s'},v_{s'}]\subseteq [u,v]$ with $\abs{[u_{s'},v_{s'}]}=v_{s'}-u_{s'}+1=\floor*{\frac{d(b)}{k_i+1}}$.
\item $s'\leq j<t':$ Let $[u_j, v_j]$ be such that $A_{j,i}$ is $X$-covered on $[u_j,v_j]$ and $\abs{[u_j,v_j]}=\floor*{\frac{d(b)}{(k_i+1)^{(j+1)-s'}}}$. Let $\varphi(k)$ be ``$A_{j+1,i}$ is $X$-covered on $D_k$''. By the assumptions $\varphi(k)$ is true for all but (wlog) $k_i$ points in $[u_j,v_j]$. By Lemma~\ref{pravdivost} there is $[u_{j+1},v_{j+1}]\subseteq[u_j,v_j]$ such that $A_{j+1,i}$ is $X$-covered on $[u_{j+1},_{j+1}]$ and
$$\abs{[u_{j+1},_{j+1}]}=\floor*{\frac{\abs{[u_j,v_j]}}{k_i+1}}=\floor*{\frac{\floor*{\frac{d(b)}{(k_i+1)^{(j+1)-s'}}}}{k_i+1}}=\floor*{\frac{d(b)}{(k_i+1)^{((j+1)+1)-s'}}}.$$
The last equation follows from Lemma~\ref{celacast}.
\end{itemize}
Consequently, we get $[u_{t'},v_{t'}]$ such that $A_{t',i}$ is $X$-covered on $[u_{t'},v_{t'}]$ and
$$\abs{[u_{t'},v_{t'}]}=\floor*{\frac{d(b)}{(k_i+1)^{t'-s'+1}}}=\floor*{\frac{d(b)}{(k_i+1)^{l(b)}}}>0.$$
Since the sequence $[u_j,v_j]$ is $\subseteq$-decreasing, $[u_{t'},v_{t'}]\subseteq [u_j, v_j]$ for $ j\in[s',t']$.
If we denote by $u'=u_{t'}, v'=v_{t'}$ and define $$c=\bigcup_{j=s'}^{t'}A_{j,i}\cap \bigcup_{k=u'}^{v'}(D_k\times\{i\})\subseteq X,$$ we can write
$$d(c)=\abs{[u',v']}=\floor*{\frac{d(b)}{(k_i+1)^{t'-s'+1}}}=\floor*{\frac{d(b)}{(k_i+1)^{l(b)}}}.$$
One can easily check that $c\subseteq b$. This is implied by the fact that $[u',v']\subseteq [u,v]$ and $[s',t']$ remained the same.\qed
Now we will try to finish the proof of the Theorem.
Define the following sequences:
\begin{flalign*}
&q_0=1, p_0=1,\\
&q_l=q_{l-1}(k_{l-1}+1)^{p_{l-1}}, \\
&p_l=q_l+p_{l-1}-1=q_{l-1}(k_{l-1}+1)^{p_{l-1}}+p_{l-1}-1.
\end{flalign*}
We will show that there is an $X$-covered $\chain{i}{s_i,t_i}{u_i,v_i}$ $a_i$ with $l(a_i)=p_i$, $d(a_i)=q_i$, for each $\abs{{\mathcal F}}\geq i\geq 0$ such that $\dom{a_{i+1}}\subseteq\dom{a_i}$ for $0\leq i < \abs{{\mathcal F}}-1$.
Induction:
\begin{itemize}
\item Find an $X$-covered $\chain{\abs{{\mathcal F}}}{s_{\abs{{\mathcal F}}}, t_{\abs{{\mathcal F}}}}{u_{\abs{{\mathcal F}}}, v_{\abs{{\mathcal F}}}}$ $a_{\abs{{\mathcal F}}}$ with $s_{\abs{{\mathcal F}}}, t_{\abs{{\mathcal F}}}, u_{\abs{{\mathcal F}}}, v_{\abs{{\mathcal F}}}$ such that
$l(a_{\abs{{\mathcal F}}})=p_{\abs{{\mathcal F}}}$, $d(a_{\abs{{\mathcal F}}})=q_{\abs{{\mathcal F}}}$. Note that in every row $\omega\times\{ i\}$ there is always such a chain.\footnote{Take, e.g., $s_{\abs{{\mathcal F}}}=0, t_{\abs{{\mathcal F}}}=p_{\abs{{\mathcal F}}}-1$. For any $j\in[0,p_{\abs{{\mathcal F}}}-1]$ there is $m_j$ such that $A_{j,\abs{{\mathcal F}}}$ is $X$-covered on $D_m$ for $m>m_j$. Set $u_{\abs{{\mathcal F}}}=\max\{m_j:j\in[0,p_{\abs{{\mathcal F}}}-1]\}+1$ and $v_{\abs{{\mathcal F}}}=u_{\abs{{\mathcal F}}}+q_{\abs{{\mathcal F}}}-1$.}
\item $0<i\leq \abs{{\mathcal F}}:$ let $a_i$ be an $X$-covered $\chain{i}{s_i,t_i}{u_i,v_i}$ with $l(a_i)=p_i, d(a_i)=q_i$. Then by Claim~\ref{claim2} there is $\chain{i-1}{s_{i-1},t_{i-1}}{u,v}$ $a'_{i-1}$ with $$l(a'_{i-1})=l(a_i)-d(a_i)+1=p_i-q_i+1=q_i+p_{i-1}-1-q_i+1=p_{i-1},$$
$$d(a'_{i-1})=d(a_i),\ \ \dom{a'_{i-1}}\subseteq\dom{a_i}.$$
Claim~\ref{claim3} implies existence of $\chain{i-1}{s_{i-1},t_{i-1}}{u_{i-1},v_{i-1}}$ $a_{i-1}$ such that
$$d(a_{i-1})=\floor*{\frac{d(a'_{i-1})}{(k_{i-1}+1)^{l(a'_{i-1})}}}=\floor*{\frac{d(a_{i})}{(k_{i-1}+1)^{p_{i-1}}}}=\floor*{\frac{q_i}{(k_{i-1}+1)^{p_{i-1}}}}=$$
$$=\floor*{\frac{q_{i-1}(k_{i-1}+1)^{p_{i-1}}}{(k_{i-1}+1)^{p_{i-1}}}}=q_{i-1},$$
$$l(a_{i-1})=l(a'_{i-1})=p_{i-1}\ \ \text{and}\ \ a_{i-1}\subseteq a'_{i-1},$$
consequently $\dom{a_{i-1}}\subseteq \dom{a_i}$.
\end{itemize}
Lastly, we get $X$-covered $\chain{0}{s_0,t_0}{u_0,v_0}$ $a_0$ with $d(a_0)=q_0=1$, $l(a_0)=p_0=1$ and $\dom{a_0}\subseteq \dom{a_j}$ for $0\leq j\leq \abs{{\mathcal F}}$, i.e., we have got $m\in\dom{a_0}$ such that $\ev{m,i}\in X$ for $0\leq i\leq \abs{{\mathcal F}}$ and this cannot be done using $\abs{{\mathcal F}}$ functions, which X consists of, a contradiction.
\qed
It is possible to prove a~stronger result. However, the~proof is a~more complex copy of the~proof of Theorem~\ref{tazkaThm}. We do not proceed with a~complete proof but rather point out which parts of the~proof of Theorem~\ref{tazkaThm} should be modified.
\begin{prop}\label{selnotp-ex_towers}
${\mathcal Sel}$ is not a $\pid{\EDp{{\mathcal E}}}$-ideal.
\end{prop}
\proof By contradiction. Suppose ${\mathcal Sel}$ is a $\pid{\EDp{{\mathcal E}}}$-ideal. Set again $f_i=\omega\times\{i\}$. We will show, that the family ${\mathcal W}=\{f_i:i\in\omega \}$ is a witness for ${\mathcal Sel}$ not being a $
\pid{\EDp{{\mathcal E}}}$. By assumptions there is $X\in{\mathcal Sel}$, $X\subseteq\bigcup{\mathcal F}$ for some ${\mathcal F}\subseteq {^\omega\omega}$, $\abs{{\mathcal F}}<\omega$ such that $f_i\setminus X\in\EDp{{\mathcal E}}$ for each $i\in\omega$, i.e.,
$$(\exists\langle k_i\in\omega:i<\omega \rangle\in{^\omega\omega})(\forall i\in\omega)(\forall^\infty A\in A_{\mathrm{A}})\ \abs{(f_i\cap A)\setminus X}\leq k_i.$$
Since $\abs{f_i\cap B_j}\leq 2$ for any $i,j$ and $f_i\cap A_{j,i'}\neq \emptyset$ only for $i'=i$, it is equivalent to $$(\exists\langle k_i\in\omega:i<\omega \rangle\in{^\omega\omega})(\forall i\in\omega)(\forall^\infty j\in\omega)\ \abs{(f_i\cap A_{j,i})\setminus X}\leq k_i.$$ By the definition, $A_{j,i}\subseteq \omega\times\{ i\}=f_i$, hence $f_i\cap A_{j,i}=A_{j,i}$ for any $i,j\in\omega$. Thus, we get equivalent formula
\begin{equation}\tag{*}
(\exists\langle k_i\in\omega:i<\omega \rangle\in{^\omega\omega})(\forall i \in \omega)(\forall^\infty j\in\omega)\ \abs{A_{j,i}\setminus X}\leq k_i.
\end{equation}
Let $C=\{j\in\omega: (\exists i\leq \abs{{\mathcal F}})\ \abs{A_{j,i}\setminus X}>k_i \}$. Clearly, $C\in\text{\rm Fin}$ because for any $i\leq \abs{{\mathcal F}}$ we have that $\{j\in\omega: \abs{A_{j,i}\setminus X}>k_i \}\in\text{\rm Fin}$ by (*).
We will use most of the apparatus built in the proof of Theorem~\ref{tazkaThm}, particularly Claim~\ref{claim1}, Claim~\ref{claim2}, modified Claim~\ref{claim3}, sequences $p_l$ and $q_l$ and notation $a_i, a_i'$.
We will follow the lines of the proof of Theorem~\ref{tazkaThm}. In fact, the proof is identical. The only difference is in the modified version of Claim~\ref{claim3}, in which we have to carefully pick the first $X$-covered chain in the induction so as to avoid the ``bad'' colors (i.e., the colors $A_{j,i}$ such that $\abs{A_{j,i}\setminus X}>k_i$).
Take an $X$-covered $\chain{\abs{{\mathcal F}}}{s_{\abs{{\mathcal F}}},t_{\abs{{\mathcal F}}}}{u_{\abs{{\mathcal F}}},v_{\abs{{\mathcal F}}}}$ $a_{\abs{{\mathcal F}}}$ with $s_{\abs{{\mathcal F}}},t_{\abs{{\mathcal F}}},u_{\abs{{\mathcal F}}},v_{\abs{{\mathcal F}}}$ being such that $l(a_{\abs{{\mathcal F}}})=p_{\abs{{\mathcal F}}}, d(a_{\abs{{\mathcal F}}})=q_{\abs{{\mathcal F}}}$ and $s_{\abs{{\mathcal F}}}>\max C$. It is easy to see that there is such a chain.\footnote{Set $s_{\abs{{\mathcal F}}}=\max C+1 >\max\{j\in\omega: \abs{A_{j,\abs{{\mathcal F}}}\setminus X}>k_{\abs{{\mathcal F}}} \}$ and $t_{\abs{{\mathcal F}}}=s_{\abs{{\mathcal F}}}+p_{\abs{{\mathcal F}}}-1$. Then for $j\in[s_{\abs{{\mathcal F}}},t_{\abs{{\mathcal F}}}]$ there is $m_j$ such that $A_{j,\abs{{\mathcal F}}}$ is $X$-covered on $D_m$ for $m>m_j$. Set $u_{\abs{{\mathcal F}}}=\max\{m_j: j\in [s_{\abs{{\mathcal F}}},t_{\abs{{\mathcal F}}}] \}+1$ and $v_{\abs{{\mathcal F}}}=u_{\abs{{\mathcal F}}}+q_{\abs{{\mathcal F}}}-1$.}
We modify the Claim~\ref{claim3} by adding the assumption $s'>\max C$, thus, the new formulation is as follows.
\begin{clm}{3*}\label{claim3*}
For every $\chain{i}{s',t'}{u,v}$ $b$ with $d(b)\geq (k_i+1)^{l(b)}$ and $s'>\max C$ there is an $X$-covered $\chain{i}{s',t'}{u',v'}$ $c$ such that $d(c)=\floor*{\frac{d(b)}{(k_i+1)^{l(b)}}}$ and $c\subseteq b$.
\end{clm}
The sole purpose of the condition $s'>\max C$ is to ensure that $\abs{A_{j,i}\setminus X}\leq k_i$ for $j\in[s',t']$. The rest of the proof of this claim is identical.
In fact, the assumption $s_{\abs{{\mathcal F}}}>\max C$ ensures $s_i>\max C$ for $i\leq\abs{{\mathcal F}}$. Hence, it is perfectly safe to apply Claim~\ref{claim3*} throughout the induction to outputs of Claim~\ref{claim2}.
The rest of the proof is completely identical.
\qed
\section{Conclusion and open problems}
The~aim of the paper was to study the~notion of a~P-property for two ideals induced by a~partition of natural numbers, see Definition~\ref{definiciaPJ}. We chose standard critical ideals appearing in the~literature~\cite{BrFl17,BrFaVe,Hr,hrusakborel}, namely $\text{\rm Fin}\times\emptyset$, ${\mathcal Sel}$, ${\mathcal{ED}}$, $\emptyset\times\text{\rm Fin}$, $\fin\times\fin$. In the~case of ideals induced by the~same partition, we easily answered the~question which of such pairs satisfy the~P-property for two ideals. The~result is presented in~Table~\ref{omxom}. However, once considering pairs of ideals induced by potentially different partitions, the~answer appeared to be much harder. In such cases we seek a~combinatorial property of two partitions guaranteeing the~corresponding induced ideals possess the~P-property for two ideals.
\par
Considering the~partition~${\mathcal A}$, the~first ideal in the~list induced by~${\mathcal A}$ is $\ofinp{{\mathcal A}}$. However, the~ideal is a~P-ideal, and hence a~$\pid{{\mathcal J}}$-ideal for arbitrary~${\mathcal J}$. We have found easy conditions characterizing the~relation of partitions~${\mathcal A},{\mathcal B}$ such that $\finop{{\mathcal A}}$ and $\finfinp{{\mathcal A}}$ are a~$\pid{{\mathcal J}}$-ideal for ${\mathcal J}$ induced by~${\mathcal B}$, see Table~\ref{omxom1}.
\par
For the~selector ideal $\selp{{\mathcal A}}$ and the~eventually different ideal $\EDp{{\mathcal A}}$, we were able to reach our goal just partially. Namely, we were able to find full characterizations in four cases displayed in Table~\ref{overview}, see Theorems~\ref{ref1}, \ref{veze}, and~\ref{ed_ofin}. We were still able to find a~useful necessary condition on partitions~${\mathcal A},{\mathcal B}$ such that $\selp{{\mathcal A}}$ is a~$\pid{{\mathcal J}}$-ideal for ${\mathcal J}$ induced by~${\mathcal B}$, see Proposition~\ref{sufficient}. By Theorems~\ref{ref1}, \ref{ed_ofin}, our necessary conditions are also sufficient ones in two cases. However, as our Theorem~\ref{tazkaThm}, and Proposition~\ref{selnotp-ex_towers} show, it is not the~case for all considered ideals.
\begin{center}
\begin{table}[H]
\centering
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline
& $\mathrm{P}$ & $\mathrm{P}(\finop{{\mathcal B}})$ & $\mathrm{P}(\selp{{\mathcal B}})$ & $\mathrm{P}(\EDp{{\mathcal B}})$ & $\mathrm{P}(\ofinp{{\mathcal B}})$ & $\mathrm{P}(\finfinp{{\mathcal B}})$ \\
\hline\hline
$\ofinp{{\mathcal A}}$& \cmark & \cmark & \cmark&\cmark &\cmark &\cmark \\
\hline
$\finop{{\mathcal A}}$ & \xmark & $\rsub{}{}$ & $\rsub{}{}$ &$\rsub{}{}$ &$\rsub{}{}$ &$\rsub{}{}$ \\
\hline
$\finfinp{{\mathcal A}}$ & \xmark & $\uparrow$ & $\uparrow$ & $\uparrow$ & $\uparrow$ & $\uparrow$ \\
\hline
$\selp{{\mathcal A}}$ & \xmark & $\rsub{}{}$ & ? & ? & $\rsub{}{}$ & ? \\
\hline
$\EDp{{\mathcal A}}$ & \xmark & $\perp,\rsub{}{},\uparrow$ & ? & ? & $\perp,\rsub{}{}$ & ? \\
\hline
\end{tabular}
\caption{Characterizations of $\pid{{\mathcal J}}$-relationships between all combinations of considered ideals.}
\label{overview}
\end{table}
\end{center}
The~upward pointing arrow~$\uparrow$ in the table means that the relationship is characterized by the relationship provided above the corresponding cell, e.g., $\finfinp{{\mathcal A}}$ is a $\pid{\finop{{\mathcal B}}}$ if and only if $\rsub{\finop{{\mathcal A}}}{\finop{{\mathcal B}}}$.
Thus, we leave the~question when are the~ideals $\selp{{\mathcal A}}$ and $\EDp{{\mathcal A}}$ a~$\pid{{\mathcal J}}$-ideal for ${\mathcal J}$ induced by~${\mathcal B}$ for cases displayed in Table~\ref{overview} open. Among them, we believe the~following case is crucial:
\begin{quest}
Is there a~reasonable combinatorial characterization of partitions~${\mathcal A},{\mathcal B}$ such that $\selp{{\mathcal A}}$ is a~$\pid{\selp{B}}$-ideal?
\end{quest}
Finally, there is one more ideal induced by a~partition of natural numbers and appering very often in the~literature, namely Kat\v etov power ideal $\text{\rm Fin}^\alpha$. We have not yet touched it, and therefore the~same question about this ideal seems to be fully open.
\par
\vspace{0.5cm}
\input{biblio.tex}
\end{document}
|
\section{Introduction}
\label{intro}
Let $\Omega \subset \mathbb{R}^{n+1}$ be open; let $E_\ell\subset \mathbb{R}^{n+1}$, $\ell\in\mathbb{N}$, and $E$ be Caccioppoli sets in $\Omega$ (sets with locally finite perimeter); let $g_\ell:\Omega \to \mathbb{R}$, $\ell\in\mathbb{N}$, and $g:\Omega\to \mathbb{R}$ be functions in $C^0(\Omega)$. Assume that
\begin{equation}
\label{eq:bounds_assumptions_continuous}
g_\ell \to g \text{ in $C^0_{\text{loc}}(\Omega)$}; \,\,\sup_\ell \int_{U} |\nabla \chi_{E_\ell}| <\infty \text{ for any } U \subset \subset \Omega;\,\, \chi_{E_\ell} \to \chi_{E} \text{ in $L^1_{\text{loc}}(\Omega)$}.
\end{equation}
Note that, given the second condition in (\ref{eq:bounds_assumptions_continuous}), the existence of a Caccioppoli set $E$ for which the third condition holds is automatically true for a subsequence of $\{E_\ell\}$, by $BV$ compactness. (Also note that locally uniform $C^{0,\alpha}$-bounds on $g_\ell$ would guarantee the existence of $g$ for which the first condition holds for a subsequence of $g_\ell$.)
We denote by $\partial^* F$ the reduced boundary of a Caccioppoli set $F$, that is, the set of points where the measure-theoretic unit normal exists. This set is $n$-rectifiable. The inner measure-theoretic unit normal, well-defined at points in $\partial^* F$, is denoted by $\nu_{F}$. The perimeter measure $|\nabla \chi_{F}|$ (here $\nabla$ is distributional and $\chi_{F} \in \text{BV}_{loc}(\Omega)$ by definition of Caccioppoli sets) coincides with $\mathcal{H}^n \res \partial^*F$. These are consequences of De Giorgi's fundamental theorem. In particular, the second condition in (\ref{eq:bounds_assumptions_continuous}) can equivalently be written as $\sup_\ell \mathcal{H}^n(\partial^*E_\ell \cap U) <\infty$. We refer to \cite{Mag} for details.
\medskip
The hypersurfaces $\partial^* E_\ell$ will be assumed to have mean curvature prescribed by the ambient function $g_\ell$, in a weak sense that will be specified below. (In the case in which $\partial E_\ell$ and $g_\ell$ are smooth, then we will just be saying that the classical mean curvature vector of the hypersurface $\partial E_\ell$ is $g_\ell \nu_{E_\ell}$, where $\nu_{E_\ell}$ is the inward unit normal.) We will be interested in the mean curvature properties of the ``limit hypersurface'' obtained from the sequence $\partial^* E_\ell$; the notion of limit will be specified below. In particular we will be interested in how the mean curvature of the limit hypersurface relates to $g$.
This problem was successfully addressed by R. Sch\"atzle in \cite{Sch}. We start with a discussion that shows why the notion of ``limit hypersurface'' and its mean curvature properties are delicate matters. An example constructed by K. Gro\ss e-Brauckmann in \cite{GB}, and reported in \cite{Sch}, gives a sequence of CMC embedded surfaces in $\mathbb{R}^3$ with mean curvature $1$ that converge in the Hausdorff distance to a plane $P$. The convergence can be expressed effectively in the language of varifolds, in which case the limit is the integral varifold corresponding to the plane endowed with multiplicity $2$. Varifold convergence makes area continuous (under the limit operation) and keeps track of the nature of the CMC surfaces, that are ``double-layered'' on the plane, with necks that connect the two layers and become dense as $\ell \to \infty$, see the figures in \cite[p. 376]{Sch}. Note that in this example, with reference to the notation used earlier, $g_\ell \equiv g \equiv 1$ and the set $E$ is empty ($E_\ell$ is the set ``between the two layers''). In particular, if we take limits only in the sense of boundaries, or in the sense of currents, we would just get $\partial \llbracket E_\ell \rrbracket \to 0$ (as customary, $\llbracket E_\ell \rrbracket$ denotes the $(n+1)$-current of integration on $E_\ell$ and $\partial$ the boundary operator on currents). This convergence does not see the accumulation of mass (area) onto the double plane. Varifold convergence, on the other hand, in particular implies that $\mathcal{H}^n \res \partial^*E_\ell \rightharpoonup 2 \mathcal{H}^n \res P$ (as Radon measures). Clearly the mean curvature of $P$ is $0$, hence it is not prescribed by the ambient function $g$: the ``ambient-prescribed'' feature is lost in the limit. This is essentially due to a cancellation phenomenon: the two layers of the CMC boundaries approach onto $P$ with mean curvature vectors that are close to being opposite on a large set; this set gets larger and larger as $\ell \to \infty$, and on it the two vectors get closer and closer to being opposite.
\medskip
When we look at the convergence of $\mathcal{H}^n \res \partial^*E_\ell \rightharpoonup \|V\|$, where $\|V\|$ denotes the weight measure of the limit varifold, it is immediate to see that $\Sigma=\text{spt}(\|V\|)$ contains $\overline{\partial^* E}$. As shown by the previous example, where $\partial^* E = \emptyset$ and $\|V\|=2 \mathcal{H}^n \res P$, the strict inclusion is possible. It is also fairly straightforward to check that on $\text{spt}(\mu_V) \setminus \partial^* E$ the mean curvature is almost everywhere $0$ and that the multiplicity is almost everywhere even on $\text{spt}(\|V\|) \setminus \partial^* E$. A priori it could also happen that the multiplicity of the limit varifold on (a set of positive measure in) $\partial^* E$ is odd and higher than $1$. This is however shown to be impossible in \cite{Sch} if $g\neq 0$. In other words, if $g$ does not vanish then the only difference between the limit in the sense of boundaries and the varifold limit is given by ``even-multiplicity minimal pieces'', that appear only in the varifold limit. Any such portion is known in the literature as ``hidden boundary''.
\medskip
As a relevant special instance of Theorem \ref{thm:main} below, we have the following compactness result, valid for a class of boundaries with mean curvature prescribed by an ambient function that does not vanish (or only vanishes on a sufficiently small set). Given the above example, an additional assumption is to be expected, to avoid hidden boundaries. It is known that hidden boundaries do not appear if the Morse index of the hypersurfaces is uniformly bounded; this follows from the (more general) results in \cite{BW1} \cite{BW2}, \cite{SchSim}\footnote{Digressing a bit, the `cancellation of first variation' under study arises in other situations as well, see \cite[Section 1.2]{Sch}. For example, in the context (\cite{HT}) of the Allen--Cahn inhomogeneous PDE, a sequence of solutions with first variation equal to a given positive constant can give rise (as the Allen--Cahn parameter $\mathop{\varepsilon} \to 0$) to a hyperplane with multiplicity $2$ (vanishing first variation). In this case one wants to relate first variations with respect to different functionals (Allen--Cahn for the sequence, area for the limit) and the phenomenon is likely to be more complex. Even assuming uniform Morse index bounds on the sequence, it is at present unknown whether minimal limits can arise (the optimal regularity result for the limit is on the other hand available under index bounds, see \cite[Theorem 1.2]{BW3}).}. The scope here is to identify a weaker assumption that guarantees compactness (in particular, the preservation of the 'prescribed mean curvature condition'). We identify it in an $L^q$ bound on the curvatures, for $q>1$.
\begin{thm}
\label{thm:C^2_case}
Let $g_\ell, g$ ($\ell\in \mathbb{N}$) be functions and $E_\ell, E$ be sets of locally finite perimeter in $\mathbb{R}^{n+1}$ that satisfy (\ref{eq:bounds_assumptions_continuous}). Assume that $\mathcal{H}^n(\{g=0\})=0$ (in particular, $g \geq 0$ or $g \leq 0$) and that $\partial E_\ell = \partial^* E_\ell$ are $C^2$-embedded for all $\ell$. Denoting by $\nu_\ell$ the inner unit normal to $\partial E_\ell$, assume that the mean curvature vector of $\partial E_\ell$ is given by $g_\ell \nu_\ell$ for all $\ell\in \mathbb{N}$.
\noindent Denote by $B^{\ell}$ the second fundamental form of the hypersurface $\partial E_\ell$. Assume that for $q>1$ we have, for every open set $U \subset \subset \Omega$,
$$\sup_{\ell\in \mathbb{N}} \int_{\partial E_\ell \cap U} |B^{\ell}|^q \,d\mathcal{H}^n<\infty.$$
Then the set of locally finite perimeter $E$ in (\ref{eq:bounds_assumptions_continuous}) satisfies:
\begin{itemize}
\item $|\partial^* E_\ell|$ converges to $|\partial^* E|$ in the sense of varifolds (as customary, $|\partial^* E|$ is the multiplicity-$1$ varifold associated to the reduced boundary of $E$);
\item the first variation of $|\partial^* E|$ (as a functional on $C^1_c$-vector fields) is represented by integration (with respect to $\mathcal{H}^n \res \partial^* E$) of $g \nu$, where $\nu$ is the ($\mathcal{H}^n$-a.e.~defined) inward pointing unit normal to $\partial^* E$.
\end{itemize}
\end{thm}
\begin{oss}
It follows from the more general Theorem \ref{thm:main} that $C^2$-embeddedness can be replaced by $C^2$-quasi-embeddedness, in the sense of \cite{BW2}.
\end{oss}
The two conclusions of Theorem \ref{thm:C^2_case} state respectively that 'hidden boundaries' do not arise and that the 'prescribed mean curvature condition' is preserved in the limit. In fact, it is known from the main result in \cite{Sch} (which makes no curvature hypotheses) that when $g\neq 0$ the loss of the 'prescribed mean curvature condition' for a sequence of boundaries arises exactly where 'hidden boundaries' appear (and viceversa). The proof of Theorem \ref{thm:C^2_case} given here is independent of \cite{Sch}.
We treat the hypersurfaces in question neither as integral varifolds nor as boundaries, but rather as \textit{oriented} integral varifolds. This notion was introduced by J. E. Hutchinson in \cite{Hut}, which also provides the basic compactness result for this class (details in Section \ref{oriented_var}). We informally sketch some key ideas here: first we analyse the example of CMC boundaries given earlier from this different point of view.
\medskip
Viewing the smooth boundary $\partial E_\ell$ as an oriented integral varifold means that we ``lift'' it to the oriented Grassmannian $\Omega \times \mathbb{S}^n$ (as opposed to the unoriented Grassmannian, which corresponds to the more customary notion of (unoriented) varifold). The point $x \in \partial E_\ell$ is lifted to $(x, \nu_{E_\ell}(x))$. More precisely, the oriented integral varifold associated to the smooth oriented hypersurface $\partial E_\ell$ is the Radon measure $V^\ell$ in $\Omega \times \mathbb{S}^n \subset \Omega \times \mathbb{R}^{n+1}$ that is supported on the $n$-dimensional submanifold $\{(x, \nu_{E_\ell}(x)): x\in \partial E_\ell\}$ and is characterised by the fact that its pushforward via the projection onto the first factor $(x,v)\in \Omega \times \mathbb{R}^{n+1} \to x \in \Omega$ is the measure $\mathcal{H}^n \res \partial E_\ell$. In the example in question, assuming that $P$ is the plane spanned by $\textbf{e}_1, \textbf{e}_2$, taking the limit in the sense of oriented varifold means that we pass to the limit (as measures in $\mathbb{R}^{n+1} \times \mathbb{R}^{n+1}$) the lifts considered; we obtain, as limit, $V=\mathcal{H}^2 \res P_+ + \mathcal{H}^2 \res P_-$, where $P_+=\{(x,v):x\in P, v=\textbf{e}_3\}$ and $P_-=(x,v):x\in P, v=-\textbf{e}_3\}$. Taking the limit as oriented varifolds keeps track of the fact that the double plane arises as the union of two copies of $P$ with opposite orientations (respectively $\textbf{e}_3$ and $-\textbf{e}_3$). In the notation that will be introduced in Section \ref{prelim}, $V^\ell = (\partial E_\ell, \nu_{E_{\ell}}, 1, 0)$ and $V=(P, \textbf{e}_3, 1, 1)$.
The key advantage of this is that, not only we can speak of the unit normal in the limit, but crucially the unit normal is (the restriction of) a \textit{fixed} \textit{ambient} function: it is the vector valued function $\vec{v}$, where $v=(v_1, \ldots, v_{n+1})$ is the variable in the second factor of $\Omega \times \mathbb{R}^{n+1}$, in which we embed the oriented Grassmannian bundle $\Omega \times \mathbb{S}^{n}$.
Given the prescribed nature of the mean curvature for each $\ell$, the mean curvature vector can also be lifted to $\Omega \times \mathbb{R}^{n+1}$ as the \textit{vector-valued} function $\vec{\mathscr{g}}_\ell(x,v)=g_\ell(x)v$. This is a (continuous) ambient function, generally depending on $\ell$. Under hypothesis (\ref{eq:bounds_assumptions_continuous}) the sequence $\vec{\mathscr{g}}_\ell$ converges to $\vec{\mathscr{g}}(x,v)=g(x) v$ in $C^0_\text{loc}(\Omega \times \mathbb{R}^{n+1})$. (This is independent of the sequence $V^\ell$.) In the specific example, $g_\ell=g\equiv 1$. The issue of ``loss of prescribed mean curvature condition'' corresponds to the fact that the candidate vector field $\vec{\mathscr{g}}$ does not yield the mean curvature vector on the limit $V$.
In the example in question, the candidate is $v$; its restriction gives $\textbf{e}_3$ on $P^+$ and $-\textbf{e}_3$ on $P^-$, which are clearly not the mean curvature vectors of $P$ with orientation (respectively) $\textbf{e}_3$ and $-\textbf{e}_3$. One can note that the average of the candidate mean curvature vectors is $0$, which is the correct mean curvature of $P$, with either orientation; this is a general fact that we will observe in due course, and it corresponds to the cancellation effect of the mean curvature vectors that happens when hidden boundaries form, as discussed earlier.
This suggests that writing the usual first variation formula (in which one uses a vector field in $\Omega$) for $V^\ell$, as an identity in the oriented Grassmannian bundle, and passing it to the limit yields an identity that is not separately meaningful for portions that have opposite orientations, like $P^+$ and $P^-$. For $V$, $\vec{\mathscr{g}}$ itself has no variational meaning, only its average (in $v$) does; in fact, one can write an analogous (valid) identity for $V$, replacing $\vec{\mathscr{g}}$ with any vector-valued function that has the same average values (see Section \ref{lift_mean}). To obtain separate first variation information on $P^+$ and $P^-$, one would like to evaluate the first variation on a vector field that deforms $P^+$ and $P^-$ independently, which is not possible with a vector field in $\Omega$: one needs to employ a vector field that \textit{depends on} the oriented unit normal (a function of $(x,v)$). This suggests that full curvatures should be relevant, in order to argue that the first variation operator for the sequence gives rise to the first variation operator for the limit.
Again in Hutchinson's work \cite{Hut}, the notion of 'curvature integral varifolds' is introduced; this amounts to the fact that the (unoriented) integral varifold admits a weak notion of second fundamental form, or curvature coefficients, defined via the validity of a certain identity in the unoriented Grassmannian bundle. (See also C. Mantegazza \cite{Man}, where the notion is extended to include the possible presence of boundary.) In the absence of any regularity information in the limit, it is natural to employ a weak notion of mean curvature, and, more generally, a weak notion of curvature coefficients.
We will revisit the definition of curvature varifolds given in \cite{Hut}, formulating one that is more suited (for our purposes) to the \textit{oriented} class (details in Section \ref{curv_oriented_var}). The identity that defines the weak notion of curvature coefficients $W_{ab}^{\ell}$ associated to $\partial^* E_{\ell}$ holds in $\Omega \times \mathbb{R}^{n+1}$ (in which we embed the oriented Grassmannian bundle $\Omega \times \mathbb{S}^n$). This identity can be passed to the limit thanks to the $L^q$ bounds assumed. In view of the discussion above on the mean curvature, it is not a priori clear that the limit coefficients $W_{ab}$ appearing in this identity are ``meaningful curvature coefficients'' for the limit oriented integral varifold $V$.
However, it turns out that the identity for weak second fundamental form of oriented integral varifolds is rigid (unlike the one obtained from the first variation formula alone): the definition is well-posed, more precisely, if curvature coefficients exist, there is a unique choice and they must be odd in $v$.
We check this by relating the coefficients $W_{ab}$, that are in principle just functions in $L^q(V)$ satisfying an identity, to a ``geometric'' notion of curvature coefficients (Section \ref{geometric_coeff}). To make sense of this, we exploit the fundamental $C^2$-rectifiability of $\|V\|$ given by U. Menne \cite{Men}, which provides countably many $C^2$ hypersurfaces that cover $\|V\|$-almost every point. We use these $C^2$ hypersurfaces to build appropriate test functions in $\Omega \times \mathbb{R}^{n+1}$, to be used in the identity.
The parity property removes the issue that was arising earlier
and implies that the candidate mean curvature $\vec{\mathscr{g}}$ is the correct one (Section \ref{proofs}).
Having established that the prescribed mean curvature condition holds in the limit, the fact that the limit is a boundary follows by fairly straighforward arguments.
\medskip
In view of the ideas described, it is natural to prove a more general result that employs the framework of ``oriented integral varifolds with curvature'' in the assumptions as well. This allows in particular to remove the $C^2$ assumption in Theorem \ref{thm:C^2_case} (which was made to make sense classically of the second fundamental form). The result is a general compactness statement with preservation of the 'prescribed mean curvature condition'; we will give the most general form in Theorem \ref{thm:overall}, once all the relevant notions have been introduced. For the moment, we record the following fairly general instance, which immediately implies Theorem \ref{thm:C^2_case}. Note that $g\in C^0$ is unrestricted, and the limit is not necessarily a boundary (as easy examples show), it is just an oriented integral varifold $V$, and the multiplicity of $\|V\|$ is a.e.~equal to $1$ where $g\neq 0$.
\begin{thm}
\label{thm:main}
Let $g_\ell, g$ ($\ell\in \mathbb{N}$) be functions and $E_\ell, E$ be sets of locally finite perimeter in $\mathbb{R}^{n+1}$ that satisfy (\ref{eq:bounds_assumptions_continuous}). Denoting by $\nu_\ell$ the inner unit normal to $\partial^* E_\ell$ (in the sense of De Giorgi), assume that the first variation of $|\partial^* E_\ell|$ is given by integration (with respect to $\mathcal{H}^n \res \partial^* E_\ell$) of $g_\ell \nu_\ell$ for all $\ell\in \mathbb{N}$.
Assume that, for every $\ell$, the oriented integral varifold $V^{\ell}=(\partial^* E_\ell, \nu_\ell, 1, 0)$ has curvature in the sense of Definition \ref{Dfi:oriented_curvature_varifold} below, with curvature coefficients $W^{\ell}_{ia}$, for $a, i \in \{1, \ldots, n+1\}$. Moreover, let $q>1$ and assume that, for every open set $U \subset \subset \Omega$,
$$\sup_{\ell\in \mathbb{N}} \int_U |W^{\ell}_{ia}|^q \,dV^{\ell}<\infty.$$
Then there exists an oriented integral varifold $V$, to which $V^\ell$ converge (in the sense of oriented varifolds) such that
\begin{itemize}
\item
$V=(\partial^*E, \nu_E, 1, 0) + (\mathcal{R}, \textbf{n}, \theta, \theta)$, where $\mathcal{R}$ is $n$-rectifiable, $\textbf{n}$ is any measurable determination of unit normal a.e.~on $\mathcal{R}$, $\theta: \mathcal{R} \to \mathbb{N}\setminus \{0\}$ is $L^1_\text{loc}(\mathcal{H}^n \res \mathcal{R})$ and $\mathcal{R}\subset \{x:g(x)=0\}$;
\item the generalised mean curvature of $\|V\|$ is given ($\|V\|$-a.e.) by $g \nu_E$ on $\partial^* E$ and by $g \textbf{n} = -g \textbf{n} =0$ on $\mathcal{R}$ (as $\|V\|$ has automatically first variation that is an absolutely continuous measure with respect to $\|V\|$, the generalised mean curvature represents the first variation of $n$-area of $\|V\|$).
\end{itemize}
\end{thm}
The first conclusion implies that $\Theta(\|V\|, x)=1$ for $\|V\|$-a.e.~$x\in\mathbb{R}^{n+1}\setminus \{x:g(x)=0\}$. Note that $\mathcal{H}^n \res(\partial^*E \cap \{g\neq 0\}) = \|V\|\res \{g\neq 0\}$.
\begin{oss}
Part of the conclusions in Theorem \ref{thm:main} follows from \cite{Sch}, at least if we assume $g_\ell$ to satisfy locally uniform $C^{1}$ bounds. Namely \cite{Sch} gives (even without curvature assumptions) that the (unoriented) varifold obtained in the limit is of the type $|\partial^* E| + (\mathcal{R}, 2\theta)$ and that the generalised mean curvature of this varifold is (a.e.) given by $g \nu_E$ on $\partial^* E$ and $0$ on $\mathcal{R}$ (in particular, $\mathcal{R}\cap \partial^* E\subset \{g=0\}$, up to an $\mathcal{H}^n$-negligeable set). The relevant additional conclusion in Theorem \ref{thm:main} is that $\mathcal{R}$ must be contained in $\{g=0\}$, thereby preserving the ``prescribed mean curvature'' feature. (This is not true without the curvature assumption.) We will not rely on \cite{Sch} to obtain the partial conclusions just recalled, as our proof will give all the conclusions at once.
\end{oss}
\medskip
Under suitable restrictions on the nodal set $\{g=0\}$ of $g$, Theorem \ref{thm:main} becomes a compactness result in the class of boundaries. This was the case in Theorem \ref{thm:C^2_case}, by dimensional considerations. We provide another instance in the following corollary (whose proof is immediate, once Theorem \ref{thm:main} is proved).
\begin{cor}
\label{cor:boundaries}
Under the assumptions of Theorem \ref{thm:main}, further assume that $g$ is a smooth function such that $S=\{g=0\}$ is a smooth hypersurface with the property that its mean curvature $h_S$ can only vanish (on $S$) to finite order. Then $V=(\partial^* E, \nu_E, 1, 0)$, equivalently, $|\partial^* E_\ell| \to |\partial^* E|$.
\end{cor}
Theorems \ref{thm:C^2_case}, \ref{thm:main} and Corollary \ref{cor:boundaries} continue to hold if $\Omega$ is a Riemannian manifold; we prove this in Section \ref{Riemannian}, Theorem \ref{thm:Riemannian}, after giving the relevant definitions. We note that the class of functions considered in Corollary \ref{cor:boundaries} is generic in the set of smooth functions (in the sense of Baire category) on a compact Riemannian manifold, see \cite[Proposition 3.8]{ZZ}.
\medskip
\textit{\textbf{A more general class}}. As mentioned earlier, the arguments employed in this work lead to a compactness result, Theorem \ref{thm:overall} in Section \ref{Riemannian}, that permits to pass to the limit the prescribed-mean-curvature condition, regardless of whether the oriented varifolds in question arise as boundaries. (As a special instance we address the case of a sequence of two-sided immersions with mean curvature prescribed by ambient functions, see Theorem \ref{thm:immersions}.) Looking at the problem in this generality has the effect of decoupling the `loss of mean curvature' from the `formation of multiplicity'. Indeed, multiplicity may become higher than $1$ in the region $\{g\neq 0\}$ (unlike in the case of boundaries); the prescribed-mean-curvature condition is nonetheless preserved (with respect to a measurable unit normal on the limit).
\medskip
\textit{\textbf{Regularity aspects}}. We do not address here the question of optimal regularity of the limit varifold. Some regularity conclusions in Theorem \ref{thm:C^2_case} (and, restricting to $\{g\neq 0\}$, in Theorem \ref{thm:main}) are immediate from De Giorgi's and Allard's theorems. Indeed, $\|V\|=\mathcal{H}^n\res \partial^* E$ and at every $x\in \partial^* E$ the density satisfies $\Theta(\|V\|, x)=1$; as the first variation of $V$ is given by integration of $g \nu_E \in L^\infty(\|V\|;\mathbb{R}^{n+1})$, for every $x\in \partial^* E$ there exists an open neighbourhood of $x$ in which $\spt{V}$ is a $C^{1,\alpha}$ (embedded) hypersurface (for every $\alpha\in (0,1)$), and in which $\spt{V}=\partial^* E$. As $\mathcal{H}^n(\spt{V}\setminus \partial^* E)=0$ we conclude that there exists a closed set $N$ with $\mathcal{H}^n(N)=0$ such that $\spt{V}\setminus N$ is $C^{1,\alpha}$, for every $\alpha \in (0,1)$. (If $g\in C^{k,a}$, the regularity of $\spt{V}\setminus N$ is improved to $C^{k+2,a}$ by standard Schauder theory.)
This conclusion does not really make use of the curvature bounds and finer regularity results can probably be obtained. To give an explicit idea, the following appears to be a plausible conjectural statement (not the most ambitious one), to which we plan to return to in future work. In Theorem 1 assume further that $q=2$, $g_\ell \to g$ in $C^{0,1}_{\text{loc}}(\Omega)$, and $g>0$. Then $V=|\partial^* E|$ is the varifold of integration over a $C^{2,\alpha}$ embedded hypersurface $M$, and $\text{dim}\left(\spt{V}\setminus M\right)\leq n-2$.
A simple example of this situation can be obtained in $U\subset \mathbb{R}^3$, for $g_\ell\equiv g \equiv 1$, by considering a sequence of (portions of) unduloids $M_\ell$, with constant mean curvature $1$, that converge to the union of two spherical hemispheres $S_1, S_2$ (with constant mean curvature $1$) that intersect tangentially at a point $p=S_1\cap S_2$. We note that, in this case, $M_\ell$ have uniformly bounded areas and Morse index equal to $1$ (for the functional $\mathcal{H}^2(\partial^* F \cap U) - \mathcal{H}^3(F \cap U)$). This implies uniform $L^2$ bounds on the second fundamental forms.
(Uniform mass bounds, mean curvature prescribed by an ambient function and uniform Morse index bounds are the setting for the regularity theory of \cite{BW1}, \cite{BW2}.)
On the other hand, we may consider the hypersurfaces $M_\ell \times \mathbb{R}^k$ in $\mathbb{R}^{3+k}=\mathbb{R}^{n+1}$. They still satisfy that the masses are locally uniformly bounded, their mean curvatures are constant and equal to $1$, and the second fundamental forms are locally uniformly bounded in $L^2$. These hypersurfaces converge to $(S_1 \times \mathbb{R}^k)\cup (S_2 \times \mathbb{R}^k)$. In this example, the Morse index of $M_\ell \times \mathbb{R}^k$ tends to infinity (with $\ell$) in any bounded open set that intersects $\{p\}\times \mathbb{R}^k$.
\section{Preliminaries}
\label{prelim}
\subsection{Oriented integral varifolds}
\label{oriented_var}
We recall the notion of oriented integral varifold from \cite{Hut}, limiting ourselves to the codimension-$1$ case that will be of interest here. (For the moment, in Euclidean space --- see Section \ref{Riemannian} for the Riemannian setting.) We will identify $\mathbb{S}^n$ with the oriented Grassmannian of $n$-planes in $\mathbb{R}^{n+1}$. The identification is given by the Hodge star operator $\star$; for $v\in \mathbb{S}^n$, $\star v$ is the unit $n$-vector that spans the orthogonal to $v$ and whose orientation is such that $\star v \wedge v$ agrees with the orientation of $\mathbb{R}^{n+1}$.
Let $U\subset \mathbb{R}^{n+1}$ be open. A Radon measure $V$ in the oriented Grassmannian bundle $U\times \mathbb{S}^n$, that we will always view as embedded in $U\times \mathbb{R}^{n+1}$, is called an \textit{oriented $n$-varifold}. Such a $V$ is an oriented \textit{integral} $n$-varifold if there exist an $n$-rectifiable set $R$ in $U$, a couple $(\theta_1, \theta_2)$ of locally integrable $\mathbb{N}$-valued functions on $R$, with $(\theta_1, \theta_2) \neq (0,0)$, and a measurable choice of orientation $\xi:R \to \mathbb{R}^{n+1}$ (the latter means that, for $\mathcal{H}^n$-a.e.~$x\in R$, $\xi(x)\in\mathbb{S}^n \subset \mathbb{R}^{n+1}$ is one of the two choices of unit normal to the approximate tangent at $R$, and the map $\xi$ is measurable) such that $V$ acts on $\varphi\in C^0_c(U\times \mathbb{R}^{n+1})$ as follows:
\begin{equation}
\label{eq:action_oriented_var}
V(\varphi) = \int_R \theta_1(x) \varphi(x,\xi(x))) + \theta_2(x) \varphi(x,-\xi(x))) \,\, d\mathcal{H}^n(x).
\end{equation}
We write $V=(R, \xi, \theta_1, \theta_2)$ in the above situation. The class of oriented integral $n$-varifolds just introduced is denoted by $IV^o_n(U)$ (following the notation in \cite{Hut}).
\medskip
The $n$-current $\textbf{c}(V)$ associated to an oriented integral $n$-varifold $V$ is defined as follows by its action on an arbitrary $n$-form $\eta$ having compact support in $U$:
$$\textbf{c}(V)(\eta)=\int \langle \star v, \eta(x) \rangle d\,V(x,v).$$
It follows that, if $V$ is an oriented integral varifold, the current $\textbf{c}(V)$ is integral as well and, with the above notation, $\textbf{c}(V)$ is the current of integration on $R$ with orientation $\xi$ and multiplicity function $\theta_1-\theta_2$.
Finally, recall that given an oriented varifold $V$, we consider $\textbf{q}:U\times \mathbb{S}^n \to U\times \mathbb{RP}^n$, the projection from the the oriented to the unoriented Grassmannian bundle. The pushforward measure $\textbf{q}_\sharp V$ is the (unoriented) varifold associated to $V$. With slight abuse of notation, we will identify $\mathbb{RP}^n$ with the Grassmannian $G_n(n+1)$ of unoriented $n$-planes in $\mathbb{R}^{n+1}$, which is in turn identified with the $(n+1)\times (n+1)$-matrices of orthogonal projection onto said plane; the latter is a submanifold of $\mathbb{R}^{(n+1)^2}$. Concretely, for $v \in \mathbb{S}^n \subset \mathbb{R}^{n+1}$ we have $\textbf{q}(v)= \delta_{ij}-v_i v_j$, for $i,j\in \{1, \ldots, n+1\}$; the $(n+1)\times (n+1)$-matrix $\delta_{ij}-v_i v_j$ can be identified with a point in $\mathbb{R}^{(n+1)^2}$ using its entries as coordinates. We will denote by $P_{ij}=\delta_{ij} - v_i v_j$ the projection matrix.
Given an integral oriented $n$-varifold $V$ as above then $\textbf{q}_\sharp V$ is the varifold of integration on $R$ with multiplicity $\theta_1+\theta_2$. The weight measure $\|\textbf{q}_\sharp V\|$ is $(\theta_1 + \theta_2) \mathcal{H}^n \res R$ and is the same as the weight measure $\|V\|$ associated to $V$. (Recall that $\|V\|$ is the pushforward of $V$ under the projection map $U\times \mathbb{R}^{n+1}\to U$ and $\|\textbf{q}_\sharp V\|$ is the pushforward of $\textbf{q}_\sharp V$ under the projection map $U\times \mathbb{R}^{(n+1)^2} \to U$.) The first variation of $\textbf{q}_\sharp V$ is denoted by $\delta (\textbf{q}_\sharp V)$ and is defined as a linear functional on $C^1_c(U; \mathbb{R}^{n+1})$. If the action of $\delta (\textbf{q}_\sharp V)$ is bounded by the $C^0$-norm of the vector field, then $\delta (\textbf{q}_\sharp V)$ is a measure in $U$, denoted by $\|\delta (\textbf{q}_\sharp V)\|$.
Convergence as oriented varifolds means (weak*) convergence as Radon measures in $U\times \mathbb{S}^n\subset U\times \mathbb{R}^{n+1}$. The following compactness result is proved by Hutchinson.
\begin{theorem*}[compactness\footnote{The fact that there exist (subsequentially) limits that are oriented rectifiable varifolds with multiplicity taking values a.e.~in the half-integers is immediate from Allard's compactness for integral (unoriented) varifolds and from Federer-Fleming's compacntess for integral currents; in \cite{Hut} it is proved that the multiplicity must be a.e.~integer-valued.} for oriented integral varifold, \cite{Hut}]
Let $V_\ell$ be oriented integral varifolds on $U$ such that $\limsup_{\ell \to \infty}\|\textbf{q}_\sharp V\|(U)<\infty$, $\limsup_{\ell \to \infty}\|\delta (\textbf{q}_\sharp V)\|(U)<\infty$ and $\limsup_{\ell \to \infty}M_U(\partial \textbf{c}(V_\ell))<\infty$ (where $M_U$ denotes the mass of a current in $U$ and $\partial$ is the boundary operator for currents). Then $V_\ell$ subsequentially converges, in the sense of oriented varifolds, to an oriented \emph{integral} varifold $V$.
\end{theorem*}
We recall that, given an (unoriented) integral varifold $W$, the first variation (of area) evaluated on $X\in C^1_c(U)$ is $(\delta W)(X) = \int \text{div}_{T} X \,dW(x,T)$, where $\text{div}_T X$ is the divergence with respect to the (unoriented) $n$-plane $T$. When $\delta W$ is a $\mathbb{R}^{n+1}$-valued Radon measure (in $U$), denoted by $\|\delta W\|$, its absolutely continuous part with respect to $\|W\|$ is called generalised mean curvature of $W$, denoted by $\vec{H}_W$, see \cite{SimonNotes}. (We note explicitly, in view of forthcoming sections, that this is a function defined $\|W\|$-a.e., hence in $U$, not in the Grassmannian bundle.)
We finally notice that if $V^\ell \to V$ as oriented varifolds, then $\textbf{q}_\sharp V^\ell \to \textbf{q}_\sharp V$ as (unoriented) varifolds and $\textbf{c}(V^\ell) \to \textbf{c}(V)$ as currents. The former follows by definition of pushforward measure; the latter follows upon noticing that for any $n$-form $\eta$ with compact support in $U$, the function $\langle \star v, \eta(x) \rangle$ is in $C^0_c(U \times \mathbb{R}^{n+1})$.
\subsection{Lifting the first variation formula to $U \times \mathbb{R}^{n+1}$}
\label{lift_mean}
Let $V$ be an oriented integral $n$-varifold, $V\in IV^o_n(U)$ for $U\subset \mathbb{R}^{n+1}$, recall that $\textbf{q}_\sharp V$ denotes the associated (unoriented) integral $n$-varifold. Assume that the first variation $(\delta \textbf{q}_\sharp V)(X) = \int \text{div}_T X d\|V\|$ is represented by a (vector-valued) function $\vec{H}\in L^p(\|V\|)$ in the sense that, for any $X\in C^1_c(U)$, $(\delta \textbf{q}_\sharp V)(X) =-\int \vec{H} \cdot X \,d\|V\|$. Then the function $\vec{M}(x,v)=\vec{H}(x)$ is defined $V$-a.e. and automatically even, in the sense that $V$-a.e.~we have $\vec{M}(x,v)=\vec{M}(x,-v)$. Moreover, using the fact that $\|V\|$ is the pushforward of $V$, the following identity holds for every $i\in\{1, \ldots, n+1\}$ and for any $\varphi \in C^1_c(U)$, implicitly extended to $U\times \mathbb{R}^{n+1}$ independently of the $v$-variables (here $D_j$ denotes the partial derivative in $x_j$):
\begin{equation}
\label{eq:lift_first_var}
\int \left((\delta_{ij}-v_i v_j) D_j \varphi + M_i \varphi \right)\,d\,V =0;
\end{equation}
this is checked by choosing the vector field $X = \varphi(x) \textbf{e}_i$ in the first variation formula and computing that $\text{div}_T (\varphi(x) \textbf{e}_i) = P_{ij}D_j \varphi=(\delta_{ij}-v_i v_j) D_j \varphi$ (with implicit summation over repeated indices). Identity (\ref{eq:lift_first_var}) is a lift to the oriented Grassmannian bundle of the first variation formula; in the situation of interest in Theorems \ref{thm:C^2_case} and \ref{thm:main}, where the generalised mean curvature of $V^{\ell}=(\partial^*E_\ell, \nu_\ell, 1,0)$ is induced by an ambient function $g_\ell$, (\ref{eq:lift_first_var}) holds with $\vec{M^\ell}(x,v)=g_\ell(x) v$ in place of $\vec{M}$ and $V^{\ell}$ in place of $V$.
We have, in Theorems \ref{thm:C^2_case} and \ref{thm:main}, a uniform mass bound and a uniform $L^\infty$ bound for the mean curvatures.
Then we can pass to the limit, both in the sense of (unoriented) varifolds for $\textbf{q}_\sharp V^{\ell}$, by Allard's theorem, or in the sense of oriented varifolds by the result of \cite{Hut} recalled earlier (note that $\textbf{c}(V^\ell)=\partial\llbracket E_\ell\rrbracket$ are cycles). We let $V$ be the oriented integral varifold obtained in the (subsequential) limit; then $\textbf{q}_\sharp V$ is the limit of $\textbf{q}_\sharp V^{\ell}$. The interest in considering (\ref{eq:lift_first_var}) is that (since we have uniform bounds on $\int |\vec{M}^\ell|^p dV$, from the $L^\infty$-bounds on $\vec{H}$), the identity can be passed to the limit, obtaining $\int \left((\delta_{ij}-v_i v_j) D_j \varphi + g v_i \varphi \right)\,d\,V =0.$
However, this does not mean that $g \vec{v}$ (appearing in the identity) is the generalised mean curvature of $\textbf{q}_\sharp V$, as shown by the example of CMC surfaces converging to a double plane (see Section \ref{intro}). In that example (with $g_\ell\equiv 1$ for every $\ell$) the limit in the sense of oriented varifold is $(P, \nu, 1, 1)$, where $\nu$ is either of the two choices of unit normal on the plane $P$. The identity (\ref{eq:lift_first_var}) holds with $M_i= v_i$ for the limit. In other words, $\vec{M}=\nu$ on the lift of $(P, \nu)$ and $\vec{M}=-\nu$ on the lift of $(P, -\nu)$. It is only the average of $\vec{M}$ that is is $0$, i.e.~the mean curvature of $P$: as we are about to see, this is not accidental.
\begin{lem}
\label{lem:mean_curv}
Let $V$ be an oriented integral varifold in $U$, $V=(R, \xi, \theta_1, \theta_2)$ with notation as in Section \ref{oriented_var}. Assume that there exists for $i\in \{1,\ldots, n+1\}$ a function $M_i$ in $L^p(V)$ such that the following identity is valid for every $\varphi \in C^1_c(U)$ ($\varphi$ is thought of as a function on $U \times \mathbb{R}^{n+1}$ that is independent of the variables in the second factor\footnote{We do not worry about the fact that $\varphi$ is formally not compactly supported in $U\times \mathbb{R}^{n+1}$, as the support of $V$ is in $U\times \mathbb{S}^n$, hence one can always cut-off $\varphi$ in the $v$-variables.}):
\begin{equation}
\label{eq:mean_curv_coefficients_bundle}
\int \left((\delta_{ij}-v_i v_j) D_j \varphi + M_i \varphi \right)\,d\,V =0.
\end{equation}
The first variation $\delta (\textbf{q}_\sharp V)$ is a Radon measure in $U$ that is absolutely continuous with respect to $\|V\|$; the generalised mean curvature of $\textbf{q}_\sharp V$ (the Radon-Nikodym derivative of $\delta (\textbf{q}_\sharp V)$ with respect to $\|V\|$) is given at a.e.~point in $R$ by the weighted average of the vector-valued functions $\vec{M}$ at the points $(x, \xi)$ and $(x,-\xi)$:
$$\vec{H}(x) = \frac{\theta_1(x) \vec{M}(x,\xi(x)) + \theta_2(x) \vec{M}(x, -\xi(x))}{(\theta_1 + \theta_2)(x)}.$$
\end{lem}
\begin{proof}
For $i$ fixed, consider the measure $M_i V$ and denote by $\mu$ its pushforward by $\pi \circ \textbf{q}$. Then $\mu$ is absolutely continuous with respect to $\|V\|$. Indeed, if $A$ is such that $\|V\|(A)=0$ we also have $V(A\times \mathbb{R}^{n+1})=0$ and hence $(M_i V)(A\times \mathbb{R}^{n+1})=0$, which shows $\mu(A)=0$. We denote by $H_i$ the density of $\mu$ with respect to $\|V\|$, which gives (Radon--Nikodym theorem) $\mu = H_i \|V\|$. We claim that the first variation of $\textbf{q}_\sharp V$ is given by the Radon measure that is absolutely continuous with respect to $\|V\|$ and whose generalised mean curvature (the density of said Radon measure with respect to $\|V\|$) is the function $\vec{H}=(H_1, \ldots, H_{n+1})$. Indeed we have, for every $\varphi \in C^1_c(U)$, that the first variation of $\textbf{q}_\sharp V$ evaluated on the vector field $\varphi \textbf{e}_i$ is given by
$$\int \text{div}_V (\varphi \textbf{e}_i) d\|V\| = \int P_{ij} D_j \varphi d\|V\|$$
and $P_{ij}=\delta_{ij}-v_i v_j$, so the last term is $\int (\delta_{ij}-v_i v_j) D_j \varphi d\,V$ and this first variation coincides with (by assumption)
$$-\int M_i \varphi \,d\,V = -\int \varphi d(\pi\circ \textbf{q})_\sharp (M_i V) = -\int H_i \varphi d\|V\| = -\int \vec{H} \cdot (\varphi \textbf{e}_i) \,d\,\|V\| . $$
By linearity of the first variation, we conclude that for any vector field $X\in C^1_c(U)$ the first variation of $V$ is given by $ -\int \vec{H} \cdot X \,d\,\|V\|$. Note that $\vec{H}\in L^p(\|V\|)$.
To complete the proof of the lemma, note that the measure $M_i V$ is defined (by its action on any $\varphi \in C^0_c(U \times \mathbb{R}^{n+1})$ --- recall that the action of $V$ extends to functions in $L^p(V)$, and $M_i \varphi$ is such a function whenever $\varphi \in C^0_c$) by $(M_i V) (\varphi) = V(M_i \varphi) = \int_E \theta_1(x) M_i(x,\xi(x)))\varphi(x,\xi(x))) + \theta_2(x) (x,-\xi(x)))\varphi(x,-\xi(x))) \,\, d\mathcal{H}^n(x)$. The pushforward of $M_i V$ via $\pi\circ \textbf{q}$ is the measure $\mu$ defined (by its action on $\phi \in C^0_c(U)$) by
$$\mu(\phi)=\int_E \left(\theta_1(x) M_i(x,\xi(x)) + \theta_2(x)M_i(x,-\xi(x))\right) \phi(x) \,\, d\mathcal{H}^n(x).$$
Since $\|V\|=(\theta_1 + \theta_2) \mathcal{H}^n \res E$ and $(\theta_1 + \theta_2)\neq 0$ a.e., we conclude that, for $\|V\|$-a.e.~$x$, $H_i(x) = \frac{\theta_1(x) M_i(x,\xi(x)) + \theta_2(x) M_i(x, -\xi(x))}{(\theta_1 + \theta_2)(x)}$.
\end{proof}
We note that, if $\theta_1\neq 0, \theta_2\neq 0$ for the limit $V$ of $V^\ell$, we have that the function $gv$ is odd on $V$, while $\vec{M}$ in (\ref{eq:lift_first_var}) is naturally even; hence $gv$ cannot restrict to the ``lift of the mean curvature'', unless the limit $\|V\|$ has a.e.~multiplicity $1$. (This is in agreement with \cite[p. 375, case (ii)]{Sch}.)
\subsection{Oriented integral varifolds with curvature}
\label{curv_oriented_var}
\begin{Dfi}
\label{Dfi:oriented_curvature_varifold}
An oriented integral $n$-varifold $V$ in $U\subset \mathbb{R}^{n+1}$ (recall that $V$ is a Radon measure in $U \times \mathbb{S}^{n} \subset U \times \mathbb{R}^{n+1}$) has curvature if there exist real-valued $V$-measurable functions $W_{i a}$, for $i, a \in \{1, \ldots, n+1\}$, defined $V$-a.e., such that the following identity holds for any $\phi \in C^1_c(U \times \mathbb{R}^{n+1})$ (with summation over a repeated index):
\begin{equation}
\label{eq:oriented_curvature_varifold}
\int \left((\delta_{ij}-v_i v_j) \, D_j \phi + W_{i a} \, D^*_{a} \phi - (W_{rr} v_i + W_{ri} v_r) \phi\right) d\, V = 0.
\end{equation}
The notation $D^*_a$ denotes the partial derivative with respect to the variable $v_a$ in the second factor $\mathbb{R}^{n+1}$, while $D_j$ denotes the partial derivative with respect to the variable $x_j$ in $U$. We say that $V$ has curvature in $L^q$ if $W_{i a} \in L^q(V)$ for all $i,a$ and we refer to the collection $W_{ia}$ ($i, a \in \{1, \ldots, n+1\}$) as curvature coefficients.
\end{Dfi}
\begin{oss}
\label{oss:L1_firstvar}
If $V$ is as in Definition \ref{Dfi:oriented_curvature_varifold} with curvature in $L^1_{\text{loc}}(V)$, then the first variation of $\textbf{q}_\sharp V$ is representable by integration of an $L^1_{\text{loc}}(\|V\|)$ generalised mean curvature. To see this, test (\ref{eq:oriented_curvature_varifold}) with $\varphi\in C^1_c(U)$ (extended independently of the $v$-variables to $U\times \mathbb{R}^{n+1}$); we obtain $\int P_{is} (D_s \varphi) - (W_{rr} v_i + W_{ri} v_r) \varphi d\, V = 0$. Then Lemma \ref{lem:mean_curv}, used with $M_i=W_{rr} v_i + W_{ri} v_r \in L^1_{\text{loc}}(V)$, implies that the first variation of $\textbf{q}_\sharp V$ is represented by integration of the average of $M_i$, which is then in $L^1_{\text{loc}}(\|V\|)$.
\end{oss}
The following argument justifies the requirement of (\ref{eq:oriented_curvature_varifold}) and shows that a $C^2$ hypersurface oriented by a choice $\nu$ of unit normal is naturally an oriented $n$-varifold that satisfies Definition \ref{Dfi:oriented_curvature_varifold}. We use the notation (as in \cite{Hut}) $\delta_i=P_{is}D_s$. Equivalently, for any function $f$, $\delta_i f$ is the $i$-th component of $\nabla^T f$, the projection of $\nabla f$ onto the tangent space, or also $\delta_i f = \nabla_{\textbf{e}_i^T} f$, where $\textbf{e}_i^T$ is the projection of $\textbf{e}_i$ onto the tangent space.
\medskip
\noindent \textit{Claim}. Let $M$ be an oriented $C^2$-embedded hypersurface in $U$, with a chosen unit normal $\nu$. Let $V$ be the oriented integral varifold associated to $(M, \nu, \theta_1, \theta_2)$. Then $V$ has curvature in the sense of Definition \ref{Dfi:oriented_curvature_varifold}, with $W_{i a}(x,v)=\delta_i (\nu_a)$ for $(x,v)$ such that $v=\nu(x)$ and $W_{i a}(x,v)=-\delta_i (\nu_a)$ for $(x,v)$ such that $v=-\nu(x)$, where $\nu_a=\nu\cdot \textbf{e}_a$. (This defines $W_{i a}$ for $V$-a.e.~$(x,v)$.)
\medskip
\noindent The verification of the claim involves only the use of the divergence theorem, (just as in the case of (unoriented) curvature varifolds in \cite{Hut}) for the vector field $X = \phi(x,\nu(x)) \textbf{e}_i$ (where $\phi$ is as in Definition \ref{Dfi:oriented_curvature_varifold}). Denote the tangential part of $X$ by $X^T = \phi P_{ri} \textbf{e}_r$. Then
$\nabla^{M}(X^T)^r = P_{si}\frac{\partial}{\partial x_s} (\phi P_{ri}) \textbf{e}_i$, so that $\text{div}_{M} X^T = \textbf{e}_r \cdot \nabla^{M}(X^T)^r = P_{rs} \frac{\partial}{\partial x_s} (\phi P_{ri})$. Then the divergence theorem $\int_{M} \text{div}_{M} X^T d \mathcal{H}^n = 0$ gives (using $P_{rs}P_{ri}=P_{si}$ and $P_{rs}=P_{sr}$)
$$\int_{M} P_{rs} \frac{\partial}{\partial x_s} (\phi P_{ri}) d \mathcal{H}^n=\int_{M} P_{rs} \left(\frac{\partial}{\partial x_s} \phi(x,\nu(x))\right) P_{ri} + \phi \delta_r P_{ri} =$$
$$= \int_{M} P_{is} (D_s \phi) + (D^*_{\ell} \phi ) \delta_{i} \nu_\ell + \phi \delta_r P_{ri}=0.$$
Note that $\delta_i P_{jk} = \delta_i ( \delta_{jk}-v_j v_k) = -(\delta_i v_j) v_k - (\delta_i v_k) v_j$. The functions $P_{jk}$, $v_j, v_k$ are evaluated at $(x,\nu(x))$, hence $\delta_i (v_\ell(x,\nu(x))=\delta_i \nu_\ell$ (where $\nu_\ell = \nu \cdot \textbf{e}_\ell$). Therefore we can define $W_{i \ell}(x,v)=\delta_i \nu_\ell$ for $(x,v)$ such that $v=\nu(x)$. The previous identity becomes
$$\int \left((\delta_{is}-v_i v_s) \,D_s \phi + W_{i \ell}\, D^*_{\ell} \phi - \phi (W_{rr} v_i + W_{ri} v_r)\right)\,d \mathcal{H}^n=0,$$
with the integrand evaluated at $(x, \nu(x))$. Similarly we set $W_{i \ell}(x,v)=-\delta_i \nu_\ell$ for $(x,v)$ such that $v=-\nu(x)$ and we get
$$\int \left((\delta_{is}-v_i v_s) \,D_s \phi + W_{i \ell}\, D^*_{\ell} \phi - \phi (W_{rr} v_i + W_{ri} v_r)\right)\,d \mathcal{H}^n=0,$$
with the integrand evaluated at $(x, -\nu(x))$. Note that we have defined $W_{i\ell}$ $V$-a.e. A linear combination of the two identities with coefficients respectively $\theta_1, \theta_2$ gives (\ref{eq:oriented_curvature_varifold}).
\medskip
We note that, in the above example, $W_{i \ell}$ is odd in $v$, in the sense that $V$-a.e.~we have $W_{i\ell}(x,v)=-W_{i\ell}(x,-v)$. It does not seem a priori clear that the coefficients $W_{ia}$ in (\ref{eq:oriented_curvature_varifold}) are necessarily odd in $v$. A related issue is whether there could be distinct choices of $W_{ia}$ for which (\ref{eq:oriented_curvature_varifold}) holds. (As seen in Section \ref{lift_mean}, if we require the validity of (\ref{eq:oriented_curvature_varifold}) only for test functions $\varphi$ independent of $v$, then there are indeed multiple choices for $M_i$, since only the weighted average is uniquely determined.) The following result is key for our arguments and settles this issue.
\begin{Prop}[Curvature coefficients are unique and odd]
\label{Prop:odd_unique}
Let $V$ be an oriented integral varifold with curvature, as in Definition \ref{Dfi:oriented_curvature_varifold}, and assume that $W_{ia}\in L^1_{\text{loc}}(V)$ (for all $i,a$). Then $W_{ia}$ are odd functions, that is, $V$-a.e.~we have $W_{i a}(x,v)=-W_{i a}(x,-v)$. Moreover, any two choices of curvature coefficients in (\ref{eq:oriented_curvature_varifold}) must agree $V$-a.e.
\end{Prop}
\begin{oss}
If we were to add the requirement that $W_{ia}$ are odd in Definition \ref{Dfi:oriented_curvature_varifold}, we would not gain any advantage in our forthcoming arguments, since the odd condition does not pass in any standard way to the limit under oriented varifold convergence.
\end{oss}
We will prove Proposition \ref{Prop:odd_unique} in Section \ref{geometric_coeff}. We further have the following properties for curvature coefficients. We note that (a) is a consequence of Proposition \ref{Prop:odd_unique}, while (c) follows from the stronger statement estalished in Section \ref{geometric_coeff}, so we postpone its proof. We will make use of (a) in the proof of Theorem \ref{thm:main} in Section \ref{proofs}; (b) and (c) are, strictly speaking, not used in the proof of Theorem \ref{thm:main} and are given here for completeness.
\begin{Prop}
\label{Prop:further_prop_curv_coeff}
Let $V$ be an oriented integral varifold with curvature, as in Definition \ref{Dfi:oriented_curvature_varifold}, and assume that $W_{ia}\in L^1_{\text{loc}}(V)$ (for all $i,a$). Then
\begin{enumerate}
\item[(a)] the function $-(W_{rr} v_i + W_{ri} v_r)$ appearing in (\ref{eq:oriented_curvature_varifold}) is $V$-a.e.~the (even) lift of $\vec{H}\cdot \textbf{e}_i$, where $\vec{H}$ is the generalised mean curvature vector of $\textbf{q}_\sharp V$;
\item[(b)] $W_{ij} v_j=0$ $V$-a.e.~;
\item[(c)] $V$-a.e. we have $W_{ij}=W_{ji}$, and $W_{ri} v_r=0$ (so that the mean curvature term in (\ref{eq:oriented_curvature_varifold}) can be written as $-W_{rr} v_i$).
\end{enumerate}
\end{Prop}
\begin{proof}
(a) As in Remark \ref{oss:L1_firstvar}, for any $\varphi\in C^1_c(U)$ (extended independently of the $v$-variables to $U\times \mathbb{R}^{n+1}$) we have $\int P_{is} (D_s \varphi) - (W_{rr} v_i + W_{ri} v_r) \varphi d\, V = 0$. Proposition \ref{Prop:odd_unique} implies that $-(W_{rr} v_i + W_{ri} v_r)$ is even in $v$. By Lemma \ref{lem:mean_curv} it must then the (even) lift of $\vec{H}\cdot \textbf{e}_i$.
(b) For any $\psi\in C^1_c(U \times \mathbb{R}^{n+1})$ we test in (\ref{eq:oriented_curvature_varifold}) with $\phi=(1-v_p v_p) \psi$ and obtain $\int (1-v_p v_p) P_{is} (D_s \psi) + (1-v_p v_p) W_{ia} D^*_a \psi + 2 v_a W_{ia} \psi - (W_{rr} v_i + W_{ri} v_r) (1-v_p v_p) \psi d\, V = 0$. Note that $(1-v_p v_p)$ vanishes identically on $U\times \mathbb{S}^n$ (hence on $\text{spt}(V)$), therefore $\int v_a W_{ia} \psi d\, V = 0$. Since $\psi$ is arbitrary, we conclude.
\end{proof}
\begin{oss}
We will prove (c), that is, $W_{ij}=W_{ji}$ $V$-a.e.,~in Section \ref{geometric_coeff}. This, combined with (b), gives also that $W_{ri} v_r$ vanishes $V$-a.e. We note that the condition $W_{ri} v_r=0$ is equivalent to the fact that the generalised mean curvature $\vec{H}$ of $\textbf{q}_\sharp V$ is a.e.~orthogonal to the approximate tangent to $\|\textbf{q}_\sharp V\|$ (which is known to be a.e.~true, \cite[Chapter 5]{Brak}). Indeed, using (a), we have $-(P \vec{H}) \cdot \textbf{e}_a=P_{ai} (W_{rr} v_i + W_{ri} v_r) = W_{rr}(P_{ai}v_i) + W_{ri}\delta_{ai} v_r - W_{ri} v_a v_i v_r = W_{ra} v_r$ (using (b) and $P_{ai}v_i=0$ $V$-a.e.~).
\end{oss}
\subsection{Relation with curvature integral varifolds in \cite{Hut}}
\label{relation_with_Hut}
While not essential for the forthcoming sections, we compare here Definition \ref{Dfi:oriented_curvature_varifold} to the notion of curvature integral varifold in \cite{Hut} (see also \cite{Man}). The validity of (\ref{eq:oriented_curvature_varifold}) for $V$ with $W_{ia} \in L^1_{\text{loc}}(V)$ implies that $\textbf{q}_\sharp V$ is a curvature integral varifold in the sense of \cite{Hut}. Indeed, let $\varphi\in C^1_c(U \times \mathbb{R}^{(n+1)^2})$, with variables $P_{ij}$ in the second factor, $i, j \in \{1, \ldots, n+1\}$. Then $\phi(x,v) = \varphi(x, P(v))$ with $P_{ij}(v)=\delta_{ij}-v_i v_j$ is admissible in (\ref{eq:oriented_curvature_varifold}). The chain rule gives $D^*_\ell \phi = - D^*_{jp} \varphi \,(v_j \delta_{\ell p} + v_p \delta_{\ell j})$. We now substitute in (\ref{eq:oriented_curvature_varifold}):
$$\int \left((\delta_{ij}-v_i v_j) D_j \varphi - (W_{ip} v_j + W_{ij} v_p) D^*_{jp} \varphi - (W_{rr} v_i + W_{ri} v_r) \varphi\right) d\, V = 0,$$
where $\varphi$ and its derivatives are evaluated at $(x,P(v))$. As the function $-(W_{i p}v_j + W_{ij} v_p)$ is even in $v$ (by Proposition \ref{Prop:odd_unique}), it descends to a well-defined ($\textbf{q}_\sharp V$-a.e.) function in $U \times \mathbb{R}^{(n+1)^2}$ that we denote by $A_{ijp}$. By definition of push-forward measure then
\begin{equation}
\label{eq:from_orient_to_unorient}
\int \left(P_{ij} D_j \varphi + A_{ijp} D^*_{jp} \varphi + A_{rir} \varphi\right) d\, \textbf{q}_\sharp V =0.
\end{equation}
The existence of real-valued functions $A_{ijk}$, for $i, j, k \in \{1, \ldots, n+1\}$, defined $\textbf{q}_\sharp V$-a.e.~(in $U \times \mathbb{R}^{(n+1)^2}$) such that this identity holds expresses the fact that $\textbf{q}_\sharp V$ is a curvature integral varifold in the sense of \cite{Hut}. Note that (as expected from \cite[Proposition 5.2.4]{Hut}) $A_{ijk}=A_{ikj}$ and (using Proposition \ref{Prop:further_prop_curv_coeff} (b)) $A_{ijj}=0$.
In the above situation ($V$ is an oriented integral varifold with curvature in the sense of Definition \ref{Dfi:oriented_curvature_varifold}) the curvature coefficients $W_{ia}$ are determined by the curvature coefficients $A_{ijk}$ of $\textbf{q}_\sharp V$ by the following relation (we also include the coefficients $B_{ij}^k$ of the generalised second fundamental form of $\textbf{q}_\sharp V$ as in \cite{Hut}):
$$W_{ia}(x,v)=-P_{ad}(v)A_{idp}(x,\textbf{q}(v)) v_p = - B_{ia}^p(x,\textbf{q}(v)) v_p.$$
To see this, we observe that $A_{ijp} v_p = -W_{i p}v_j v_p - W_{ij}$, hence $P_{aj} A_{ijp} v_p = -P_{aj} W_{ij}$. Moreover, $W_{ia} = \delta_{aj}W_{ij}=P_{aj}W_{ij} + W_{ij} v_j v_a=P_{aj}W_{ij}$ (using Proposition \ref{Prop:further_prop_curv_coeff} (b)).
\medskip
While the idea (illustrated in Section \ref{curv_oriented_var}) behind Definition \ref{Dfi:oriented_curvature_varifold} is the same as for the notion of curvature (unoriented) varifolds in \cite{Hut}, to avoid confusion we remark that in \cite{Hut} the notion of curvature varifolds $\mathcal{C}V_n(U)$ is essentially given only for (unoriented) integral varifolds. In the case of an oriented integral varifold $V \in IV^o_n(U)$, \cite{Hut} defines $V\in \mathcal{C}V^o_n(U)$ (oriented curvature varifolds) by the requirement that $\textbf{q}_\sharp V \in \mathcal{C}V_n(U)$. That is, in \cite{Hut} an oriented integral varifold is said to be a curvature varifold when the associated (unoriented) varifold is a curvature varifold. In Definition \ref{Dfi:oriented_curvature_varifold} instead we require the defining identity (\ref{eq:oriented_curvature_varifold}) directly on the oriented varifold $V$, in order to say that it has curvature; the functions $W_{ij}$ are defined $V$-a.e. (thus in the oriented Grassmannian bundle $U \times \mathbb{S}^{n}\subset U\times \mathbb{R}^{n+1}$).
Our notion may a priori be more restrictive than membership to $\mathcal{C}V^o_n(U)$ in the sense of \cite{Hut}.
(The two notions coincide if $V$ is sufficiently regular. For example, if $V$ is associated to a $C^2$-embedded hypersurface $M$, the condition that $V$ is an oriented curvature varifold in the sense of Definition \ref{Dfi:oriented_curvature_varifold} is equivalent to $V \in \mathcal{C}V^o_n(U)$ in the sense of \cite{Hut}.)
In the forthcoming sections we will work with oriented integral varifolds associated to reduced boundaries (without any regularity assumption). We will use the terminology ``$V$ is an oriented varifold with curvature'', or ``the oriented varifold $V$ has curvature'', to mean the condition in Definition \ref{Dfi:oriented_curvature_varifold}, without reference to the class $\mathcal{C}V^o_n(U)$ of \cite{Hut}.
\section{Curvature coefficients and their geometric counterpart}
\label{geometric_coeff}
The aim of this section (see Proposition \ref{Prop:C2_coeff} below) is to relate the coefficients $W_{ia}$ appearing in (\ref{eq:oriented_curvature_varifold}), which are just functions on $V$, to ``geometric'' curvature coefficients associated to the weight measure $\|V\|$. We assume that $W_{ia}\in L^1_{\text{loc}}(V)$ for all $i,a$. As a byproduct, we will also establish Proposition \ref{Prop:odd_unique}.
\medskip
We begin by recalling that $\textbf{q}_\sharp V$ is an integral varifold with first variation represented by a function in $L^1_{\text{loc}}(\|\textbf{q}_\sharp (V)\|)$, as observed in Remark \ref{oss:L1_firstvar}. The weight measures of $V$ and $\textbf{q}_\sharp (V)$ are the same, $\|V\|=\|\textbf{q}_\sharp (V)\|$. We denote by $R$ the $n$-rectifiable set in $\mathbb{R}^{n+1}$ such that $V=(R, \xi, \theta_1, \theta_2)$, and $\textbf{q}_\sharp (V)=(R, \theta_1+\theta_2)$.
The result of Menne \cite{Men} (see also \cite{San}, which proves a weaker version that is sufficient for the proof of Theorem \ref{thm:main}, since an $L^\infty$ bound on the generalised mean curvatures is available) guarantees the $C^2$-rectifiability of $\|V\|$, namely there exists a collection $\{M_j\}_{j=1}^\infty$ of $C^2$-hypersurfaces in $U$ such that $\|V\|(U\setminus (\cup_{j=1}^\infty M_j))=0$.
Upon possibly redefining the hypersurfaces $M_j$, we may assume that each $M_j$ admits a continuous choice of unit normal $\nu_{M_j}$.\footnote{This can be ensured by considering, for each $j$ of the initial family $\{M_j\}$, the hypersurfaces $M_j\cap B^{n+1}_r(x)$ with $x\in U$ having rational coordinates, and with $r\in \mathbb{Q}$ sufficiently small to ensure that $M_j\cap B^{n+1}_r(x)$ is orientable. There are only countably many choices for $x$ and $r$, so the claim follows.}
Consider, for each $j$, the lift of $M_j$ to $U \times \mathbb{S}^{n} \subset U\times \mathbb{R}^{n+1}$, that is the $C^1$ hypersurface $\tilde{M}_j=\{(x,v)\in U\times \mathbb{R}^{n+1}: |v|=1, x\in M_j, v=\pm \nu_{M_j}(x)\}$. (Note that if $M_j$ is connected then $\tilde{M}_j$ has two connected components.)
\begin{oss}
The measure $V$ is absolutely continuous with respect to $\mathcal{H}^n\res (\cup_{j=1}^\infty \tilde{M}_j)$.
Assume that $A\subset U\times \mathbb{R}^{n+1}$ satisfies $\mathcal{H}^n(A\cap (\cup_{j=1}^\infty \tilde{M}_j))=0$. Denote by $\varpi: U\times \mathbb{R}^{n+1} \to U$ the standard projection and by $v^{\pm}(R)$ the two sets defined respectively by
$$v^{\pm}(R)=\{(x,v)\in U\times \mathbb{R}^{n+1}: x\in R, \text{ the approximate tangent $T_x R$ exists, } v=\pm \xi(x)\}.$$
Then by definition of $V$ we have
$$V(A)=(\theta_1 \mathcal{H}^n)(\varpi(A\cap v^+(R))) + (\theta_2 \mathcal{H}^n)(\varpi(A\cap v^-(R))).$$
By ($C^1$-)rectifiability, for any $j$ we have that at $\|V\|$-a.e.~$x\in M_j$ the approximate tangent to $R$ at $x$ exists and agrees with $T_x M_j$, and $\xi(x)=\pm \nu_{M_j}(x)$. In other words, for every $j$ there exists $N_j\subset M_j$ such that $\|V\|(N_j)=0$ and for every $x\in M_j \setminus N_j$ we have that $x\in R$, the approximate tangent to $R$ at $x$ exists and $(x,\pm\xi(x))\in \tilde{M}_j$. Hence $v^{\pm}(R) \setminus (N\times \mathbb{R}^{n+1}) \subset (\cup_{j=1}^\infty \tilde{M}_j)$, where $N=N_0 \cup (\cup_{j=1}^\infty N_j)$ and $N_0=R\setminus (\cup_{j=1}^\infty M_j)$; note that $\|V\|(N)=0$. We then have that $\varpi(A\cap v^\pm(R)) \subset \varpi(A \cap (\cup_{j=1}^\infty \tilde{M}_j)) \cup N$. This inclusion gives $V(A) \leq ((\theta_1+\theta_2) \mathcal{H}^n) \varpi(A \cap (\cup_{j=1}^\infty \tilde{M}_j))$. Recalling that the projection decreases the $\mathcal{H}^n$-measure, $\mathcal{H}^n\left(\varpi(A\cap (\cup_{j=1}^\infty \tilde{M}_j)) \right)\leq \mathcal{H}^n\left(A\cap (\cup_{j=1}^\infty \tilde{M}_j)\right)=0$, we conclude $V(A)=0$.
\end{oss}
By the previous remark and by Radon-Nikodym's theorem there exists a function $\sigma \in L^1_{\text{loc}}\left(\mathcal{H}^n\res (\cup_{j=1}^\infty \tilde{M}_j)\right)$ such that $V= \sigma\, \mathcal{H}^n \res (\cup_{j=1}^\infty \tilde{M}_j)$. This is (by definition) a rectifiable measure. By the characterisation of rectifiability through approximate tangents (e.g.~\cite[Chapter 3]{SimonNotes}), for $V$-a.e.~$(x,v)$ the approximate tangent measure to $V$ at $(x,v)$ exists and is $\sigma(x,v)\mathcal{H}^n\res \Pi_{x,v}$ for an $n$-dimensional subspace $\Pi_{x,v}$ in $\mathbb{R}^{n+1} \times \mathbb{R}^{n+1}$. Explicitly, denoting by $\eta_{(x,v),r}(\cdot)$ the dilation $\frac{\cdot - (x,v)}{r}$, this means that for $r\to 0$ the measures $\frac{1}{r^n}(\eta_{(x,v),r})_\sharp V$ converge to $\sigma(x,v)\mathcal{H}^n\res \Pi_{x,v}$, that is, for every function $\phi \in C^0_c(\eta_{(x,v),r}(U)\times \mathbb{R}^{n+1})$,
$$\int \frac{1}{r^n} \phi\, d({\eta_{(x,v)}}_\sharp V) \to \sigma(x,v)\int \phi \,d\mathcal{H}^n\res \Pi_{x,v}\,\,\,\, \text {as } r\to 0.$$
The above also implies (choosing $\phi$ to be a smooth, radially symmetric non-negative bump function in $\mathbb{R}^{2n+2}$ such that its value on the unit ball is $1$, it is decreasing in the radial variable, and its support is the ball of radius $1+\delta$, with $\delta>0$ arbitrarily small) that $\liminf_{r\to 0} \frac{1}{r^n}V(B_{(1+\delta)r})\geq \sigma(x,v) \omega_n$ and $\limsup_{r\to 0} \frac{1}{r^n}V(B_{r})\leq \sigma(x,v) (1+\delta)^n \omega_n$, hence
\begin{equation}
\label{eq:V_r^n}
\frac{V(B_r(x,v))}{r^n}\to \sigma(x,v) \omega_n.
\end{equation}
\medskip
We next check that for $V$-a.e.~$(x,v)$ the following holds (note that $V$-a.e.~$\sigma > 0$):
\begin{equation}
\label{eq:Leb_cont_modified}
\left|\int \frac{\chi_r}{\sigma(x,v)} W_{ia} d\,V - W_{ia}(x,v)\right| \to 0 \,\,\,\, \text {as } r\to 0, \end{equation}
where $\chi$ is an any smooth, radially symmetric bump function in $\mathbb{R}^{2n+2}$ such that its integral on any $n$-plane through the origin is equal to $1$ and we use the notation $\chi_r(\cdot)=\frac{1}{r^n} \chi\left(\frac{\cdot - (x,v)}{r}\right)$.
\noindent
Recall that $W_{ia}\in L^1_{\text{loc}}(V)$, therefore $V$-a.e.~point is of Lebesgue continuity for $W_{ia}$, that is, for $V$-a.e.~$(x,v)$
\begin{equation}
\label{eq:Leb_cont_W}
\frac{1}{V(B_r((x,v))}\int_{B_r((x,v))} |W_{ia}-W_{ia}(x,v)| \, dV \to 0.
\end{equation}
Then
$$\int \frac{\chi_r}{\sigma(x,v)} W_{ia} d\,V - W_{ia}(x,v) = \int \frac{\chi_r}{\sigma(x,v)} W_{ia} d\,V - W_{ia}(x,v) \int_{\Pi_{x,v}} \chi d\mathcal{H}^n =$$
$$= \int \frac{\chi_r}{\sigma(x,v)} \left(W_{ia}-W_{ia}(x,v)\right) d\,V+W_{ia}(x,v)\left(\underbrace{\int \frac{\chi_r}{\sigma(x,v)} d\,V - \int_{\Pi_{x,v}} \chi d\mathcal{H}^n}_{\to 0 \text{ as } r\to 0}\right),$$
where the last term tends to $0$ by the characterisation of tangent measure above, used with $\chi$ in place of $\phi$ (and using the definition of push-forward measure). Denoting by $C>0$ the radius of a ball that contains the support of $\chi$ we also find
$$\left|\int \frac{\chi_r}{\sigma(x,v)} \left(W_{ia}-W_{ia}(x,v)\right) d\,V \right|\leq \frac{\|\chi\|_{L^\infty}}{\sigma(x,v)}\frac{C^n}{(Cr)^n}\int_{B_{Cr}((x,v))}|W_{ia}-W_{ia}(x,v)|d\,V.$$
As $\frac{1}{r^n}\int_{B_r((x,v))}|W_{ia}-W_{ia}(x,v)|d\,V=\frac{V(B_r((x,v))}{r^n}\frac{1}{V(B_r((x,v))}\int_{B_r((x,v))}|W_{ia}-W_{ia}(x,v)|d\,V$ tends to $0$. By (\ref{eq:V_r^n}) and (\ref{eq:Leb_cont_W}), we conclude (\ref{eq:Leb_cont_modified}).
\medskip
Next, we point out that, for every given $j$, for $V$-a.e.~$(x,v)\in\tilde{M}_j$ it holds:
\begin{equation}
\label{eq:char_Mj}
\frac{(\chi_{\tilde{M}_j}V)(B_r(x,v))}{r^n}=\frac{(\sigma\,\mathcal{H}^n \res \tilde{M}_j)(B_r(x,v))}{r^n}\to \sigma(x,v) \omega_n\,\,\,\text{ as } r\to 0.
\end{equation}
Indeed $\chi_{\tilde{M}_j}$ is $V$-measurable and in $L^1_{\text{loc}}(V)$, so by the Lebesgue continuity property, for $V$-a.e.~$(x,v)$ it holds $\frac{1}{V(B_r((x,v))}\int_{B_r((x,v))}\chi_{\tilde{M}_j} dV \to \chi_{\tilde{M}_j}(x,v)$ as $r\to 0$. As observed in (\ref{eq:V_r^n}), $\frac{V(B_r((x,v))}{r^n}\to \omega_n \sigma(x,v)$ for $V$-a.e.~$(x,v)$, so (\ref{eq:char_Mj}) follows.
\noindent Similarly, for every given $j$, and for every $(i,a)$, for $V$-a.e.~$(x,v)\in\tilde{M}_j$ it holds:
\begin{equation}
\label{eq:char_Mj_W}
\frac{1}{ \sigma(x,v) \omega_n r^n}\int_{B_r(x,v)} |W_{ia}\,\chi_{\tilde{M}_j} - W_{ia}(x,v)| \,d\,V\to 0\,\,\,\text{ as } r\to 0.
\end{equation}
\medskip
We are now ready to prove the main result of this section, from which Proposition \ref{Prop:odd_unique} follows immediately.
\begin{Prop}
\label{Prop:C2_coeff}
Let $V$ be an oriented integral varifold with curvature in $L^1_{\text{loc}}(V)$ (as in Definition \ref{Dfi:oriented_curvature_varifold}) and let $\{M_j\}_{j=1}^\infty$, $\{\tilde{M}_j\}_{j=1}^\infty$ be as above. For any $j$, for $V$-a.e.~$(x,v)\in \tilde{M}_j$ the curvature coefficients $W_{ia}$ appearing in (\ref{eq:oriented_curvature_varifold}) are given ($i, a \in \{1, \ldots, n+1\}$) by $W_{ia}((x,v))=\delta_i (\nu_{M_j})_a$ for $v=\nu_{M_j}(x)$ and $W_{ia}((x,v))=-\delta_i (\nu_{M_j})_a$ for $v=-\nu_{M_j}(x)$.
\end{Prop}
\begin{oss}
\label{oss:well_posed_curv_coeff}
As shown earlier, $V$-a.e.~point is in at least one of the $\tilde{M}_j$'s, so the conclusion covers indeed $V$-a.e.~point. It also follows from the proof that the conclusion is $V$-a.e.~well-posed (even though a point may belong to more than one of the $\tilde{M}_j$'s).
\end{oss}
\begin{proof}
We are going to prove the statement at an arbitrary $(x_0,v_0)\in \tilde{M}_j$ such that $T_{x_0} R$ exists, $T_{x_0} M_j= T_{x_0} R$, $\pm v_0$ is orthogonal to $T_{x_0} M_j$ at $x_0$, and such that (\ref{eq:V_r^n}), (\ref{eq:Leb_cont_modified}), (\ref{eq:Leb_cont_W}), (\ref{eq:char_Mj}) and (\ref{eq:char_Mj_W}) hold with $x_0$, $v_0$ in place respectively of $x$, $v$. These conditions are all verified at $V$-a.e.~point in $\tilde{M}_j$, as checked above.
Since we have either $v_0=\nu_{M_j}(x_0)$ or $v_0=-\nu_{M_j}(x_0)$, we begin with the first of these two instances. Let $f$ be a $C^2$ function defined in a neighbourhood $O\subset U$ of $x_0$, chosen such that $M_j\cap O=\{f=0\}$, $\nabla f \neq 0$ in $O$ and $\frac{\nabla f}{|\nabla f|}=\nu_{M_j}$. Define $\phi_a:O\times \mathbb{R}^{n+1}\to \mathbb{R}$ by
$$\phi_a(x,v) = v_a - \frac{D_a f}{|\nabla f|}(x).$$
Let $\tilde{\psi}:\mathbb{R}^{n+1}\to [0,1]$ be a smooth function (of $v$) that is identically $1$ in an open neighbourhood of the set $\{v: v=\nu_{M_j}(p),\, p\in M_j\cap O\}$ and identically $0$ in an open neighbourhood of the set $\{v: v=-\nu_{M_j}(p),\, p\in M_j \cap O\}$. The fact that these two sets are a positive distance apart can be ensured by taking $O$ sufficently small. We define $\psi(x,v)=\tilde{\psi}(v)$ on $O\times \mathbb{R}^{n+1}$. We then check that
\begin{equation}
\label{eq:vanishing_integral_f}
\frac{1}{r^{n+1}}\int_{B_r((x_0,v_0))}|\psi \phi_a| d\,V \to 0 \text{ as }r\to 0.
\end{equation}
Indeed, $\psi \phi_a$ vanishes on $\tilde{M}_j$ by construction ($\phi_a$ vanishes on $\tilde{M}_j\cap \{\psi\neq 0\}$). Moreover (by $C^1$-regularity of $\psi \phi_a$), there exists $C>0$ such that $|\psi \phi_a|\leq C r$ on $B_r((x_0,v_0))$ for all sufficently small $r$. Therefore (\ref{eq:vanishing_integral_f}) follows from the fact that $\lim_{r\to 0}\frac{1}{r^{n}} V(B_r((x_0,v_0)) \setminus \tilde{M}_j) =0$, which is given by the assumed validity of (\ref{eq:V_r^n}) and (\ref{eq:char_Mj}) at $(x_0, v_0)$. Similarly, we have
\begin{equation}
\label{eq:vanishing_integral_f_2}
\frac{1}{r^{n+1}}\int_{B_r((x_0,v_0))}|D_\ell^* \psi|\, |\phi_a| d\,V \to 0 \text{ as }r\to 0\,\, \text{ for every }\ell.
\end{equation}
Testing (\ref{eq:oriented_curvature_varifold}) with $\psi \beta \phi_a$ (in place of $\varphi$), for an arbitrary $\beta\in C^1_c(O\times \mathbb{R}^{n+1})$, we find
\begin{eqnarray*}
\int (\delta_{is}-v_i v_s) \beta \psi \,\frac{\partial}{\partial x_s} \left(\frac{-D_a f}{|\nabla f|}\right) \,d\,V+ \int (\delta_{is}-v_i v_s)\,(D_s\beta) \psi\,\phi_a \,d\,V + \int W_{i a}\,\beta \psi\,d\,V\\
+\int W_{i\ell}\phi_a D^*_\ell (\beta \psi) \,d\,V- \int \beta \psi \phi_a (W_{rr} v_i + W_{ri} v_r)d\,V=0.
\end{eqnarray*}
We let $\beta=\frac{\chi_r}{\sigma(x_0,v_0)}$ in the above identity, where $\chi$ and $\chi_r$ are as described for (\ref{eq:Leb_cont_modified}), and analyse the five terms. Since $|D \beta|\leq \frac{K}{r^{n+1}}$ (for a constant $K$ depending on $(x_0,v_0)$), by (\ref{eq:vanishing_integral_f}) and (\ref{eq:vanishing_integral_f_2}) the second term tends to $0$ as $r\to 0$. The fourth term, as the integrand vanishes on $\tilde{M}_j$, can be replaced by $\int W_{i\ell}\phi_a D^*_\ell (\beta \psi) (1-\chi_{\tilde{M}_j})\,d\,V$, which is bounded in modulus by $\frac{C}{r^n}\int_{B_r((x_0,v_0))} |W_{i\ell}| (1-\chi_{\tilde{M}_j})\,d\,V$, which tends to $0$ as $r\to 0$ by (\ref{eq:V_r^n}), (\ref{eq:Leb_cont_W}), (\ref{eq:char_Mj_W}). The fifth term is bounded in modulus by $\frac{Cr}{r^n}\int |W_{rr} v_i + W_{ri} v_r|d\,V$, which also tends to $0$ as $r\to 0$ by (\ref{eq:V_r^n}), (\ref{eq:Leb_cont_W}).
Since $\psi\equiv 1$ on the support of $\beta$ for all sufficiently small $r$, we obtain that the third term tends to $W_{ia}(x_0,v_0)$ by (\ref{eq:Leb_cont_modified}). The first term tends to $(\delta_{is}-(v_0)_i (v_0)_s) \frac{\partial}{\partial x_s} \left(\frac{D_a f}{|\nabla f|}\right)(x_0)$ (because $(x_0, v_0)$ is a point of continuity and thus of Lebesgue continuity of $(\delta_{is}-v_i v_s) \frac{\partial}{\partial x_s} \left(\frac{D_a f}{|\nabla f|}\right)(x)$).
In conclusion
\begin{equation}
\label{eq:coefficient_correct_1}
W_{ia}(x_0,v_0) = (\delta_{is}-(v_0)_i (v_0)_s) \frac{\partial}{\partial x_s} \left(\frac{D_a f}{|\nabla f|}\right)(x_0).
\end{equation}
The second instance, that is the case in which $(x_0,v_0)$ satisfies $v_0=-\nu_{M_j}(x_0)$, is addressed in the same way, noting that $-f$ can be used in place of $f$, and it leads to
\begin{equation}
\label{eq:coefficient_correct_2}
W_{ia}(x_0,v_0) = -(\delta_{is}-(v_0)_i (v_0)_s) \frac{\partial}{\partial x_s} \left(\frac{D_a f}{|\nabla f|}\right)(x_0).
\end{equation}
In other words the value of $W_{ia}$ agrees with the curvature coefficient of $M_j$ computed with respect to $\nu_{M_j}$ if $v=\nu_{M_j}$ and with respect to $-\nu_{M_j}$ if $v=-\nu_{M_j}$.
\end{proof}
\begin{oss}
Given functions $W_{ia}\in L^1_{\text{loc}}(V)$ for which (\ref{eq:oriented_curvature_varifold}) holds, $W_{ia}$ have a well-defined Lebesgue value $V$-a.e., and this is the value $W_{ia}(x_0, v_0)$ for which we have the validity of (\ref{eq:Leb_cont_modified}) in the proof just given, regardless of the chosen $j$. Therefore the relations (\ref{eq:coefficient_correct_1}) and (\ref{eq:coefficient_correct_2}) are valid with the same left-hand-side regardless of $j$. In particular this guarantees that, except possibly for a set of vanishing $V$-measure, if $(x,v) \in \tilde{M}_j \cap \tilde{M}_k$ then the curvature coefficients of $M_j$ and $M_k$ are the same at $x$. This confirms Remark \ref{oss:well_posed_curv_coeff} and is not surprising: $V$ is a rectifiable measure and so $V$-a.e.~the approximate tangent $\Pi_{(x,v)}$ exists and agrees with the approximate tangent to $T_{(x,v)} \tilde{M}_j$ independently of $j$: the tangent $T_{(x_0,v_0)} \tilde{M}_j$ determines the right-hand-sides of (\ref{eq:coefficient_correct_1}) and (\ref{eq:coefficient_correct_2}).
\end{oss}
\begin{proof}[proof of Proposition \ref{Prop:further_prop_curv_coeff} (c)]
For each $C^2$ hypersurface $M$ oriented by a unit normal $\nu_M$, we have by direct computation $\delta_i (\nu_{M} \cdot \textbf{e}_a) =- (\nabla_{\textbf{e}_i^T} \textbf{e}_{a}^T)^N = -B(\textbf{e}_i^T, \textbf{e}_{a}^T)$, where the tangential and normal components of a vector are denoted respectively using upper indeces $T$ and $N$, and $B$ is the classical second fundamental form, whose symmetry gives $\delta_i (\nu_{M} \cdot \textbf{e}_a)=\delta_a (\nu_{M} \cdot \textbf{e}_i)$. Then the conclusion follows from Proposition \ref{Prop:C2_coeff}.
\end{proof}
\section{Proof of Theorems \ref{thm:C^2_case} and \ref{thm:main}}
\label{proofs}
For each $\ell$ we have the validity of (\ref{eq:oriented_curvature_varifold}), with $V^\ell=(\partial^* E_\ell, \nu_\ell, 1, 0)$ in place of $V$ and $W^\ell_{ia}$ in place of $W_{ia}$ (as assumed in Theorem \ref{thm:main}):
\begin{equation}
\label{eq:cuvature_identity_Vell}
\int \left((\delta_{is}-v_i v_s) \,D_s \phi + W^\ell_{i a}\, D^*_{a} \phi - \phi (W^\ell_{rr} v_i + W^\ell_{ri} v_r)\right)d\,V^\ell=0.
\end{equation}
\begin{oss}
Note that Theorem \ref{thm:C^2_case} is indeed a special case of Theorem \ref{thm:main}. This follows from the argument in Section \ref{curv_oriented_var}, which shows that $V^\ell=(\partial^* E_\ell, \nu_\ell, 1, 0)$ is an oriented integral varifold with curvature, with curvature coefficients $W^\ell_{ia}=\delta_i (\nu_{E_\ell})_a$. By direct computation, $\delta_i (\nu_{E_\ell} \cdot \textbf{e}_a) =- (\nabla_{\textbf{e}_i^T} \textbf{e}_{a}^T)^N$, where $T$ and $N$ stand respectively for tangential and normal components of a vector with respect to $\partial E_{\ell}$. Therefore $\int \sum_{i,j} |W^\ell_{ij}|^q d V^\ell\leq \int_{\partial E_\ell} |B^\ell|^q d\mathcal{H}^n$, where $B^{\ell}$ is the (classical) second fundamental form of $\partial E_{\ell}$.
\end{oss}
Our first observation is that the class of oriented $n$-varifolds in $\Omega\subset \mathbb{R}^{n+1}$ with curvature in $L^q$ is compact (in the oriented varifold topology) under locally uniform bounds on the masses and on the $L^q$-norms of the curvature coefficients, for $q>1$. This can be deduced as immediate consequence of the (more general) convergence of measure-function pairs in \cite{Hut} (using \cite[Proposition 4.2.4 and Proposition 4.4.2 (ii)]{Hut}). We prove it here directly in terms of convergence of (vector-valued) measures. (Here we work on the sequence $V^\ell$ under consideration; the general statement of this compactness property will be given in Proposition \ref{Prop:compactness_curvature_bounds_Riem}.)
\medskip
Consider the sequence of vector-valued measures $\vec{\mu}_\ell=\vec{W}^\ell V^\ell$, where $\vec{W}^\ell$ is the vector with $(n+1)^2$ entries $\{W^ \ell_{ia}\}$ for $(a,i)\in \{1, \ldots , n+1\} \times \{1, \ldots , n+1\}$ (for a fixed order on this set of indeces). By assumption, $\int_{U\times \mathbb{R}^{n+1}} |W_{ia}^\ell|^q d V^{\ell} \leq K$ for all $\ell$ and for all $i,a$, for the given $q>1$.
Then the measures $\vec{\mu}_\ell$ have locally uniformly bounded total variations, $|\vec{\mu}_\ell|(U\times \mathbb{R}^{n+1})=\int_{U\times \mathbb{R}^{n+1}} |\vec{W}^\ell| dV^\ell \leq K^{1/q} C^{1/q'}$, where $C$ is the upper bound for the perimeters of $E_\ell$ in $U$ (provided by (\ref{eq:bounds_assumptions_continuous})) and $q'$ is the conjugate of $q$. By standard weak* compactness for Radon measures we can extract a (subsequential) limit $\vec{\mu}$. (We will use below the fact that the total variation is lower semi-continuous.) Extracting possibly a further subsequence, we also have the existence of an oriented integral varifold $V$ to which $V^{\ell}$ converge, thanks to the compactness theorem of Hutchinson, recalled in Section \ref{oriented_var}. Let $A\subset U\times \mathbb{R}^{n+1}$ be open and bounded; for $t>0$ let $A_t$ be the set of points in $A$ at distance $\geq t$ from $\partial A$; let $u\in C^0_c(A)$ with $u\equiv 1$ on $A_t$ and $u\geq 0$. Then:
$$|\vec{\mu}|(A_t)\leq \liminf_{\ell\to \infty} |\vec{\mu_\ell}|(A_t) = \liminf_{\ell\to \infty}\int_{A_t} |\vec{W}^\ell| dV^\ell\leq \liminf_{\ell\to \infty}\int_{U\times \mathbb{R}^{n+1}} u |\vec{W}^\ell| dV^\ell\leq$$
$$\liminf_{\ell\to \infty} \left(\int_{U\times \mathbb{R}^{n+1}} |\vec{W}^\ell|^q dV^\ell\right)^{1/q} \left(\int u^{q'}dV^\ell\right)^{1/q'}\leq K^{1/q} \liminf_{\ell\to \infty} \left(\int u^{q'}dV^\ell\right)^{1/q'}$$
$$=K^{1/q} \left(\int u^{q'}dV \right)^{1/q'} \leq K^{1/q}\, V(A)^{1/q'}. $$
Then (since these are Radon measures) we get $|\vec{\mu}|(A)\leq K^{1/q}\, V(A)^{1/q'}$ for any Borel set $A$. In particular, $\vec{\mu}$ is absolutely continuous with respect to $V$ and is thus of the type $\vec{W} V$, with $\vec{W}\in L^1_{\text{loc}}(V)$. We will denote by $W_{ia}$ the components of $\vec{W}$.
The convergence $\vec{W}^\ell V^\ell \rightharpoonup \vec{W} V$ (weak* as vector-valued Radon measures) is equivalent to convergence as measure-function pairs of $(W^\ell_{i a}, V^\ell)$ to $(W_{ia}, V)$, in the sense of \cite[Proposition 4.2.4]{Hut}. Both notions amount to the fact that for every $(i,a)$
$$\int W_{ia}^{\ell} \phi dV^{\ell} \to \int W_{ia} \phi dV \text{ for every }\phi \in C^0_c(U\times \mathbb{R}^{n+1}).$$
Thus, for every $\phi\in C^1_c(U\times \mathbb{R}^{n+1})$, passing (\ref{eq:cuvature_identity_Vell}) to the limit, we obtain
\begin{equation}
\label{eq:cuvature_identity_limit}
\int \left((\delta_{is}-v_i v_s) \,D_s \phi + W_{i a}\, D^*_{a} \phi - \phi (W_{rr} v_i + W_{ri} v_r)\right)d\,V=0.
\end{equation}
The oriented integral varifold $V$ thus has curvature in the sense of Definition \ref{Dfi:oriented_curvature_varifold}, with curvature coefficients $W_{ia} \in L^1_{\text{loc}}(V)$. In fact, we have $W_{ia} \in L^q_{\text{loc}}(V)$. (For the rest of our arguments in this section, we will only require that $W_{ia} \in L^1_{\text{loc}}(V)$.) Being convex, $y\in \mathbb{R}^{(n+1)^2} \to |y|^q$ is a supremum of countably many affine functions, $|y|^q = \sup_{m\in \mathbb{N}} (a_m + \vec{b}_m\cdot y)$ for a suitable choice of $a_m \in \mathbb{R}$, $\vec{b}_m\in \mathbb{R}^{(n+1)^2}$. Then we have, for each $m$, and for any $\psi\in C^0_c(U \times \mathbb{R}^{n+1})$,
$$\int (a_m + \vec{b}_m \cdot \vec{W}) \psi \, dV =\lim_{\ell \to \infty}\int (a_m + \vec{b}_m \cdot \vec{W^\ell}) \psi \, dV^\ell \leq \liminf_{\ell \to \infty}\int |\vec{W}^\ell|^q \psi dV^\ell.$$ One can verify (this holds whenever a function is defined as a supremum) that $\int_A |\vec{W}|^q dV = \sup\left\{\sum_{m\in \mathbb{N}} \int (a_m + \vec{b}_m \cdot \vec{W})\phi_m \,dV\right\}$, where the supremum is taken over all possible choices of $\{\phi_m\}_{m\in\mathbb{N}}$ with $\phi_m\in C^0_c(A)$, $0\leq \phi_m\leq 1$, with $\text{spt}(\phi_m)$ pairwise disjoint and $A$ having compact closure. It then follows that $\int_A |\vec{W}|^q dV \leq \liminf_{\ell \to \infty}\int_A |\vec{W}^\ell|^q dV^\ell$.
\medskip
\begin{proof}[proof of Theorem \ref{thm:main}]
From Proposition \ref{Prop:further_prop_curv_coeff} (a), $-(W^\ell_{rr} v_i + W^\ell_{ri} v_r)$ in (\ref{eq:cuvature_identity_Vell}) is the lift ($V^{\ell}$-a.e.) of $\vec{H}^\ell\cdot \textbf{e}_i$, where $\vec{H}^\ell$ is the generalised mean curvature of $|\partial^* E_\ell|$, which is $g_\ell \nu_\ell$ by assumption. Hence $V^{\ell}$-a.e.~$(x,v)$ we have $-(W^\ell_{rr} v_i + W^\ell_{ri} v_r)(x,v) = g_{\ell}(x) v_i$, the $i$-th component of $\vec{\mathscr{g}}_\ell(x,v)$. (Recall that $V^{\ell}=(\partial^* E_\ell, \nu_{E_\ell}, 1,0)$, so the measurable choice of orientation is $\nu_{E_\ell}$.) Then (\ref{eq:cuvature_identity_Vell}) can be equivalently written as
$$\int \left((\delta_{is}-v_i v_s) \,D_s \phi + W^\ell_{i a}\, D^*_{a} \phi + \phi g_\ell v_i \right)d\,V^\ell=0.$$
Passing to the limit, since $g_\ell v_i \to gv_i$ in $C^0(U\times \mathbb{R}^{n+1})$ (and the total masses of $V^\ell$ are locally uniformly bounded), we obtain
\begin{equation}
\label{eq:cuvature_identity_limit_with_g}
\int \left((\delta_{is}-v_i v_s) \,D_s \phi + W_{i a}\, D^*_{a} \phi + \phi g v_i\right)d\,V=0.
\end{equation}
Comparing (\ref{eq:cuvature_identity_limit_with_g}) and (\ref{eq:cuvature_identity_limit}), the arbitraryness of $\phi\in C^1_c(U\times \mathbb{R}^{n+1})$ implies that $V$-a.e.~it holds $-(W_{rr} v_i + W_{ri} v_r) = g v_i$, the $i$-th component of $\vec{\mathscr{g}}(x,v)$. Again by Proposition \ref{Prop:further_prop_curv_coeff} (a), $-(W_{rr} v_i + W_{ri} v_r)$ is the lift of $\vec{H}\cdot \textbf{e}_i$, where $\vec{H}$ is the generalised mean curvature of $\|V\|$. Hence $\vec{\mathscr{g}}(x,v)=g(x)\vec{v}=\vec{H}(x)$ (that is, $\vec{\mathscr{g}}$ descends to $\vec{H}$ and, in particular, the prescribed mean curvature feature holds for $V$).
\medskip
Since $\textbf{c}(V^\ell)=\partial\llbracket E_\ell \rrbracket$ and $\partial\llbracket E_\ell \rrbracket \to \partial \llbracket E\rrbracket$ (as currents), we conclude that $\textbf{c}(V)=\partial \llbracket E \rrbracket$ and therefore $V=(\partial^* E, \nu_E, 1, 0)+(\mathcal{R}, \textbf{n}, \theta, \theta)$, for an $n$-rectifiable set $\mathcal{R}$ and $\theta:\mathcal{R}\to \mathbb{N}\setminus \{0\}$ in $L^1_{\text{loc}}(\mathcal{H}^n\res \mathcal{R})$, and $\textbf{n}$ a measurable choice of unit normal on $\mathcal{R}$. Defining $R=\partial^* E \cup \mathcal{R}$ we then write $V=(R, \xi, \theta_1, \theta_2)$, where $\xi$ is the measurable choice of orientation on $R$ taken to agree with $\nu_E$ on $\partial^*E$ and with $\textbf{n}$ on $\mathcal{R}\setminus \partial^* E$, and $(\theta_1, \theta_2)=(\theta+1,\theta)$ on $\partial^* E \cap \mathcal{R}$, $(\theta_1, \theta_2)=(1,0)$ on $\partial^* E \setminus \mathcal{R}$, $(\theta_1, \theta_2)=(\theta,\theta)$ on $\mathcal{R}\setminus \partial^* E$.
By Lemma \ref{lem:mean_curv} (which can be applied since we can test in (\ref{eq:cuvature_identity_limit_with_g}) with $\phi$ independent of $v$) we find $\vec{H}(x)=\frac{\theta_1(x) g(x)\xi(x)-\theta_2(x) g(x)\xi(x)}{\theta_1(x)+\theta_2(x)}$, so that necessarily we must have $(\theta_1(x), \theta_2(x))=(1,0)$ for $\|V\|$-a.e.~$x\in \{g(x) \neq 0\}$ . In particular, possibly redefining $\mathcal{R}$ by removing a set of $\mathcal{H}^n$-measure $0$, $\mathcal{R}\subset \{g=0\}$.
Theorems \ref{thm:C^2_case} and \ref{thm:main} are proved.
\end{proof}
\begin{proof}[proof of Corollary \ref{cor:boundaries}]
We note that we can use $\{g=0\}$ as one of the $M_j$'s that give the $C^2$-rectifiability property for the rectifiable set $R$ such that $V=(R, \xi, \theta_1, \theta_2)$ (see Section \ref{geometric_coeff}). More precisely, given the cover $\{M_j\}_{j=1}^\infty$ as in Section \ref{geometric_coeff}, we can redefine $C_{1} = \{g=0\}$, $C_j = M_{j-1} \setminus \{g=0\}$ for $j\geq 2$ to yield a new countable $C^2$-cover $\{C_j\}_{j=1}^\infty$. We can then work with the new cover to obtain (fixing $j=1$) that $V$-a.e.~$(x,v) \in \tilde{C}_{1}$ the curvature coefficients $W_{ia}(x,v)$ of $V$ agree with the ``geometric'' curvature coefficients of $C_{1}$ and, in particular, the generalised mean curvature of $\textbf{q}_\sharp V$ agrees with $h_S$ almost everywhere on $S=\{g=0\}$. On the other hand, by Theorem \ref{thm:main}, almost everywhere on $\mathcal{R}$ we have that the generalised mean curvature of $\textbf{q}_\sharp V$ vanishes and that $\mathcal{R}\subset S$. Therefore $h_S = 0$ almost everywhere on $\mathcal{R}$. If $\mathcal{H}^n(\mathcal{R})>0$, there is a $\mathcal{H}^n$-measurable set $Z \subset S$ with $\mathcal{H}^n(Z)>0$ and such that $h_S=0$ on $Z$. Then there would be a point (of Lebesgue density $1$ for $Z$) where $h_S$ vanishes at infinite order, contradiction. So $\mathcal{H}^n(\mathcal{R})=0$, that is, $V=(\partial^* E, \nu_E, 1, 0)$.
\end{proof}
\section{Riemannian setting and more general hypersurfaces}
\label{Riemannian}
Assume that $\Omega$ is an oriented Riemannian manifold of dimension $n+1$ (as we are addressing a local problem, orientability can always be assumed without loss of generality). We will embed $\Omega$ isometrically in and open set $U\subset \mathbb{R}^N$, for $N$ sufficiently large, and denote by $S(x)$ the matrix of orthogonal projection from $\mathbb{R}^N$ onto $T_x \Omega\subset \mathbb{R}^N$.
Following \cite{Hut}, $V$ is an oriented integral $n$-varifold in $\Omega$ if it is an oriented integral $n$-varifold in $U$ and $\spt{V}\subset \Omega$. This means that $V=(R, \vec{N}, \theta_1, \theta_2)$ for an $n$-rectifiable set $R\subset U$, a measurable choice of orientation $\vec{N}(x)$, with $\vec{N}(x)$ a unit simple $(N-n)$-vector whose span is orthogonal to $T_x R$ in $\mathbb{R}^N$, and two locally summable integer-valued functions $\theta_1, \theta_2$; moreover we have imposed $R\subset \Omega$. Recall that, in particular, $V$ is a measure in $U\times G^o(n,N)$, where the second factor is the Grassmannian of oriented $n$-planes in $\mathbb{R}^N$, which naturally embeds in $\Lambda^n(\mathbb{R}^N) \equiv \mathbb{R}^{N \choose {n}}$.
To exploit the codimension-$1$ property of $V$ in $\Omega$, we note that $\vec{N}$ determines uniquely a measurable function $\xi: R \to \mathbb{S}\Omega$, where $ \mathbb{S}\Omega$ is the unit sphere bundle of $\Omega$, by the relation $\xi \wedge \vec{N}_{\Omega} = \vec{N}$, where $\vec{N}_{\Omega}$ is the smooth determination of unit simple $(N-(n+1))$-vector that orients the normal bundle of $\Omega$ in $\mathbb{R}^N$. Note that $\mathbb{S}\Omega \subset T\Omega$ is realised as a smooth subbundle of $\Omega \times \mathbb{R}^N$, with fiber at $x\in \Omega$ given by $\mathbb{S}^{N-1} \cap T_x \Omega$ (an $n$-sphere). This allows us to treat $V$ as a Radon measure in $U \times \mathbb{R}^N$, whose defining action on $\varphi\in C^0_c(U \times \mathbb{R}^N)$ is
$$V(\varphi)=\int_R \left(\theta_1(x) \varphi(x, \xi(x)) + \theta_2(x) \varphi(x, -\xi(x))\right) d\mathcal{H}^n(x).$$
In fact, $V$ is supported in $\mathbb{S}\Omega$. We write $V=(R, \xi, \theta_1, \theta_2)$.
\begin{Dfi}
\label{Dfi:Riemannin_oriented_curvature}
We say that the oriented integral $n$-varifold $V=(R, \xi, \theta_1, \theta_2)$ in $\Omega$ has curvature in $L^q_{\text{loc}}$ if there exist functions $W_{ia} \in L^q_{\text{loc}}(V)$, for $i,a \in \{1, \ldots, N\}$ such that the following identity holds for all $\varphi\in C^1_c(U \times \mathbb{R}^N)$:
$$\int \left((\delta_{sk} - v_s v_k) S_{kb} (D_s \varphi) + (D^*_a \varphi) W_{ba} + (\delta_{ik} - v_i v_k) S_{kr} \overline{A}_{irb} \varphi +\right.$$
$$\left.-S_{rb} (W_{jr} v_k+ W_{jk}v_r) S_{kj} \varphi \right)\, dV=0,$$
where $\overline{A}_{ijk}=S_{ir}D_r S_{jk}$ are the curvature coefficients of $\Omega$ in $\mathbb{R}^N$ (these are smooth functions of $x$) and $D_j$, $D^*_a$ denote the partial derivatives respectively with respect to the variable $x_j$ (in $U\subset \mathbb{R}^N$) and $v_a$ (in $\mathbb{R}^N$).
\end{Dfi}
Define, for $i,j\in \{1, \ldots, N\}$, the function $P_{ij}(x,v) =(\delta_{ik} - v_i v_k) S_{kj}$ (which appears in the above identity). Note that when $v\in T_x \Omega$ is unit, the matrix $P(x,v)$ represents the matrix of orthogonal projection from $\mathbb{R}^N$ onto the $n$-dimensional subspace orthogonal to $v$ in $T_x \Omega$. When $v$ is a unit normal to a hypersurface $M\subset \Omega$, then $P$ is the orthogonal projection from $\mathbb{R}^N$ onto $T_x M \subset \mathbb{R}^N$.
We point out the following relations, that involve the (smooth) curvature coefficients of $\Omega$. Recall that (see \cite{Hut}) $\overline{A}_{irb}= \overline{A}_{ibr}= \overline{B}_{ir}^b+ \overline{B}_{ib}^r$, where $\overline{B}_{ij}^k = \langle \overline{\nabla}_{{\textbf{e}_i}^T} \textbf{e}_j^T, \textbf{e}_k^N\rangle$ stands for the second fundamental form coefficients of $\Omega$ in $\mathbb{R}^N$, for $i,j,k\in\{1,\ldots, N\}$, with $\overline{\nabla}$ denoting differentiation in $\mathbb{R}^N$ and $T$ and $N$ denoting respectively tangential and normal components with respect to $\Omega$. We have:
\begin{eqnarray*}
S_{jr} \overline{B}_{db}^j = S_{jr} \langle \nabla_{\textbf{e}_d^T} \textbf{e}_b^T, \textbf{e}_j^\perp\rangle = S_{jr} \langle \overline{B}(\textbf{e}_d^T, \textbf{e}_b^T), \textbf{e}_j \rangle= S\left(\overline{B}(\textbf{e}_d^T, \textbf{e}_b^T)\right) \cdot \textbf{e}_r = 0 \cdot \textbf{e}_r =0; \\
P_{ir} \overline{A}_{irb}=P_{ir} \overline{B}_{ir}^b+P_{ir} \overline{B}_{ib}^r = P_{ir} \overline{B}_{ir}^b, \text{ using } P_{ir}=P_{iy} S_{yr};\\
S_{ab} S_{yr} \overline{A}_{irb}= S_{ab} S_{yr} \overline{B}_{ir}^b= S_{yr} S_{ab} \overline{B}_{ir}^b = 0, \text{ hence also }P_{kb}P_{jd} \overline{A}_{dkj} =0.
\end{eqnarray*}
The defining identity for $W_{ia}$ in Definition \ref{Dfi:Riemannin_oriented_curvature} is then equivalently written as
\begin{equation}
\label{eq:Riemannin_oriented_curvature_B}
\int \left(P_{sb} (D_s \varphi) + (D^*_a \varphi) W_{ba} + P_{ir} \overline{B}_{ir}^{b} \varphi -S_{rb} (W_{jr} v_k+ W_{jk}v_r) S_{kj} \varphi \right)\, dV=0.
\end{equation}
\noindent \textit{Claim}. Let $M$ be a $C^2$ hypersurface in $\Omega$, oriented by a choice of unit normal $\nu$. The associated oriented integral varifold $V_M=(M, \nu, 1,0)$ satisfies the conditions in Definition \ref{Dfi:Riemannin_oriented_curvature}.
\begin{proof}[proof of the claim]
Consider the vector field $\phi \textbf{e}_b$, for $b\in \{1, \ldots, N\}$ and $\phi\in C^1_c(\Omega)$. Then the component tangential to $\Omega$ is given by $\phi S_{rb} \textbf{e}_r$, and the component tangential to $M$ is given by $(\phi S_{rb} \textbf{e}_r)^t=\phi P_{rj}S_{rb} \textbf{e}_j$, so that $\text{div}_M (\phi S_{rb} \textbf{e}_r)^t = P_{sj}D_s(\phi P_{rj} S_{rb})$. Then the divergence theorem gives
$$\int P_{sr}S_{rb}(D_s \phi) + P_{sr}(D_s S_{rb}) \phi + P_{sj} S_{rb} (D_s P_{rj}) \phi=0.$$
Since $P\circ S = P$ (and with the usual notation $\delta_j=P_{is}D_s$) we rewrite the identity as
$$\int P_{sb}(D_s \phi) + (\delta_r S_{rb}) \phi + S_{rb} (\delta_j P_{rj}) \phi =0.$$
Substituting $P_{rj}=(\delta_{rk} - v_r v_k) S_{kj}$ and setting $W_{ik}=\delta_i(\nu \cdot \textbf{e}_k)=P_{si}D_s(\nu \cdot \textbf{e}_k)$,
$$\delta_j P_{rj}=(\delta_{rk} - v_r v_k)\delta_j S_{kj}-(\delta_j v_r)v_k S_{kj} - (\delta_j v_k)v_r S_{kj} = (\delta_{rk} - v_r v_k)P_{jd}S_{dr}D_r S_{kj}$$ $$-(\delta_j v_r)v_k S_{kj} - (\delta_j v_k)v_r S_{kj} =(\delta_{rk} - v_r v_k) P_{jd}\overline{A}_{dkj}-W_{jr} v_k S_{kj} - W_{jk}v_r S_{kj}.$$
Moreover, $\delta_r S_{rb}=P_{rd}D_d S_{rb} = P_{ir}S_{id} D_d S_{rb} = P_{ir}(\overline{\delta}_i S_{rb}) = P_{ir}\overline{A}_{irb}$. (We use the notation $\overline{\delta}_i=S_{id} D_d$.)
The identity becomes then
$$\int P_{sb}(D_s \phi) +P_{ir}\overline{A}_{irb} \phi + \underbrace{S_{rb} (\delta_{rk} - v_r v_k)}_{P_{bk}} P_{jd} \overline{A}_{dkj}\phi-S_{rb} (W_{jr} v_k+ W_{jk}v_r) S_{kj} \phi =0,$$
and the third term vanishes (by the relations given above).
Taking $\phi(x)=\varphi(x,\nu(x))$ for $\varphi(x,v)$ with $x\in \mathbb{R}^N$ and $v\in \mathbb{R}^N$, we obtain:
$$\int P_{sb}(D_s \varphi) + (D^*_a \varphi)\underbrace{P_{sb}D_s(\nu(x) \cdot \textbf{e}_a)}_{W_{ba}} + P_{ir}\overline{A}_{irb} \varphi-S_{rb} (W_{jr} v_k+ W_{jk}v_r) S_{kj} \varphi =0,$$
with $\varphi$ evaluated at $(x,\nu(x))$, which concludes the proof that $(M, \nu, 1,0)$ is an oriented integral varifold with curvature as in Definition \ref{Dfi:Riemannin_oriented_curvature}.
\end{proof}
Next, we note that, testing with $\varphi(x,v)=\phi(x)$ in (\ref{eq:Riemannin_oriented_curvature_B}), for $\phi\in C^1_c(\Omega)$, we obtain
$$\int \left(P_{sb} (D_s \varphi) +P_{ir} \overline{B}_{ir}^{b} \varphi -S_{rb} (W_{jr} v_k+ W_{jk}v_r) S_{kj} \varphi \right)\, dV=0.$$
All functions in the integrand, except $W_{ia}$ and $v_j$, are even in $v$ (some only depend on $x$). Set $\tilde{W}_b(x)$ to be the weighted average (as in Lemma \ref{lem:mean_curv}) of $S_{rb} (W_{jr} v_k+ W_{jk}v_r) S_{kj}$ at $\xi(x)$ and $-\xi(x)$, which is $L^1_{\text{loc}} (\textbf{q}_\sharp V)$. We then obtain
$$\int \left(P_{sb} (D_s \phi) + P_{ir} \overline{A}_{irb} \phi -\tilde{W}_b \phi \right)\, d (\textbf{q}_\sharp V)=0.$$
The first term in the integrand is $\text{div}_{\textbf{q}_\sharp V} (\phi \textbf{e}_b)$. Therefore the identity just obtained amounts to the first variation formula evaluated on the vector field $\varphi \textbf{e}_b$ and implies that $\textbf{q}_\sharp V$ has first variation in $L^1_{\text{loc}} (\textbf{q}_\sharp V)$. Explicitly, $\int \text{div}_{\textbf{q}_\sharp V} (\varphi \textbf{e}_b) +\varphi \textbf{e}_b \cdot \vec{H}_{\textbf{q}_\sharp V}^{\mathbb{R}^N}=0$, with $\vec{H}_{\textbf{q}_\sharp V}^{\mathbb{R}^N}=(P_{ir} \overline{A}_{irb} -\tilde{W}_b) \textbf{e}_b$ (the generalised mean curvature of $\textbf{q}_\sharp V$ in $\mathbb{R}^N$). On the other hand, testing with the vector field $\phi(x)S_{ab}(x)\textbf{e}_a = (\varphi \textbf{e}_b)^T$ gives
$$\int \left(P_{sa} (D_s (\phi S_{ab})) + \underbrace{P_{ir} S_{ab} \overline{A}_{ira}}_{=0} \phi - \underbrace{S_{ab} \tilde{W}_a}_{\tilde{W}_b} \phi \right)\, d (\textbf{q}_\sharp V)=0,$$
where we used respectively the relations given earlier and the definition of $\tilde{W}_{ab}$ for the two braced terms.
This identity is $\int \text{div}_{\textbf{q}_\sharp V} (\varphi \textbf{e}_b)^T +(\varphi \textbf{e}_b)^T \cdot \vec{H}_{\textbf{q}_\sharp V}^{\Omega}=0$, where the generalised mean curvature $\vec{H}_{\textbf{q}_\sharp V}^{\Omega}$ of $\textbf{q}_\sharp V$ in $\Omega$ is given by $-\tilde{W}_b \textbf{e}_b$. Note that $\vec{H}_{\textbf{q}_\sharp V}^{\Omega}$ is the projection onto $\Omega$ of $\vec{H}_{\textbf{q}_\sharp V}^{\mathbb{R}^N}$.
\medskip
We have established in particular the following: if $V$ is as in Definition \ref{Dfi:Riemannin_oriented_curvature}, the first variation of $\textbf{q}_\sharp V$ in $\Omega$ is represented by a function in $L^q_{\text{loc}}(\textbf{q}_\sharp V)$. It follows, thanks to \cite{Men}, that for $V$ as in Definition \ref{Dfi:Riemannin_oriented_curvature}, with $q\geq 1$, we have the existence of $\{M_j\}_{j=1}^\infty$ such that $\|V\|(U \setminus (\cup_{j=1}^\infty M_j))=0$ and each $M_j$ is a $C^2$-hypersurface in $\Omega$. By redefining the $M_j$'s we can assume that each of them admits an orientation, with unit normal $\nu_{M_j}$. Then we can prove the following analogue of Propositions \ref{Prop:C2_coeff} and \ref{Prop:odd_unique}.
\begin{Prop}
\label{Prop:C^2_coeff_Riem}
Let $V$ and $W_{ia}$ (for $i,a\in\{1,\ldots, N\}$) be as in Definition \ref{Dfi:Riemannin_oriented_curvature}, with $q\geq 1$, and $\{M_j\}_{j=1}^\infty$ as above. Then for each $j$, for $V$-a.e.~$(x,v) \in M_j$ we have $W_{ia}(x,v)=\delta_i (\nu_{M_j} \cdot \textbf{e}_a)$ when $v=\nu_{M_j}(x)$, and $W_{ia}(x,v)=-\delta_i (\nu_{M_j} \cdot \textbf{e}_a)$ when $v=-\nu_{M_j}(x)$. This is well-defined $V$-a.e. In particular, $W_{ia}$ is a.e. uniquely defined, and odd in $v$.
\end{Prop}
The proof of Proposition \ref{Prop:C^2_coeff_Riem} proceeds as in Section \ref{geometric_coeff}, using test functions
$$\phi_a(x,v)= v_a - \frac{D_a f}{|\nabla f|}(x),$$
where $f:U \to \mathbb{R}$ is chosen so that $\{f=0\} \cap \Omega = M_j$, $\nabla f = \nu_{M_j}$ on $M_j$. Such $f$ can be defined on $\Omega$ and then extended to $U$ (since $\Omega$ is a smooth submanifold and we can use tubular neighbourhood coordinates, possibly making $U$ smaller).
\medskip
We moreover have, arguing either directly as in Section \ref{proofs}, or by invoking \cite[Proposition 4.2.4 and 4.4.2(ii)]{Hut}, that the class of oriented $n$-varifolds in $\Omega$ with curvature in $L^q_{\text{loc}}$ has the following compactness property.
\begin{Prop}
\label{Prop:compactness_curvature_bounds_Riem}
Let $\{V^\ell\}_{\ell\in \mathbb{N}}$ be a family of oriented integral $n$-varifolds in $\Omega$ that satisfy the assumptions of Hutchinson's compactness theorem (see \cite{Hut}, or Section \ref{prelim}) and that moreover have curvature coefficients $W_{ia}^\ell\in L^q_{\text{loc}}(V^\ell)$ for $i,a\in \{1, \ldots, N\}$ (in the sense of Definition \ref{Dfi:Riemannin_oriented_curvature}) such that, for every compact $K$, $\sup_{\ell}\int_K |W^\ell_{ia}|^q\,dV^\ell<\infty$ for $q>1$. Then any oriented integral varifold $V$ in $\Omega$ that is a limit point of $\{V^\ell\}$ has curvature $W_{ia}\in L^q_{\text{loc}}(V)$ and we have (subsequentially) $\vec{W}^{\ell_k} V^{\ell_k} \to \vec{W} V$ (as vector-valued measures --- here $\vec{W}$ denotes the vector with entries $W_{ia}$ for a fixed ordering of $\{(i,a)\}_{i,a\in \{1, \ldots, N\}}$, and $\vec{W^\ell}$ is analogously defined; the same convergence can be expressed in terms of measure-function pairs $(W_{ia}^\ell, V^\ell)$ and $(W_{ia}, V)$ as in \cite{Hut}).
\end{Prop}
Since $W_{ia}$ are odd in $v$ (Proposition \ref{Prop:C^2_coeff_Riem}), $S_{rb} (W_{jr} v_k+ W_{jk}v_r) S_{kj}$ is the even lift of $\tilde{W}_b$, which, as we saw, is $\vec{H}_{\textbf{q}_\sharp V}^\Omega \cdot \textbf{e}_b$. When the mean curvature of $V^\ell$ in $\Omega$ is prescribed by $g^\ell\in C^0(\Omega)$, we have $S_{rb} (W^\ell_{jr} v_k+ W^\ell_{jk}v_r) S_{kj} =- g^\ell v_b$. This is a continuous ambient function (in $U \times \mathbb{R}^N$, extending to a tubular neighbourhood of $\Omega$). Repeating the arguments of Section \ref{proofs} we establish the following general compactness result:
\begin{thm}
\label{thm:overall}
Let $V^{\ell}=(R_\ell, \xi_\ell, \theta_1^\ell, \theta_2^\ell)$ be a sequence of integral oriented $n$-varifolds in an $(n+1)$-dimensional oriented Riemannian manifold $\Omega$, with $\sup_{\ell\in\mathbb{N}}\|\textbf{q}_\sharp V^\ell\|(U)<\infty$ and $\sup_{\ell\in\mathbb{N}}M_U(\partial \textbf{c}(V^\ell))<\infty$ for every $U\subset \subset \Omega$. Assume that there is $q>1$ such that $V^{\ell}$ have curvatures in $L^q_{\text{loc}}$ (in the sense of Definition \ref{Dfi:Riemannin_oriented_curvature}), and that the curvature coefficients are locally uniformly bounded in the $L^q(V^\ell)$-norm.
By the curvature assumption, $\delta(\textbf{q}_\sharp V^\ell)$ is represented by integration on $\|\textbf{q}_\sharp V^\ell\|$ of $\vec{H}^\ell \in L^q_{\text{loc}}(\|V^\ell\|)$. We assume that there exists a measurable choice of orientation $\nu_\ell$ on $R_\ell$ such that $\vec{H}^\ell = g_\ell \nu_\ell$ a.e.
Then every varifold limit is of the type $\textbf{q}_\sharp V$ for an oriented integral varifold $V=(R, \xi, \theta_1, \theta_2)$ that has curvature in $L^q_{\text{loc}}$; moreover, there exists a measurable choice of orientation $\nu$ on $R$ such that $\vec{H} = g \nu$ a.e. (By Allard's theorem, $\delta(\textbf{q}_\sharp V)$ is represented by integration on $\|\textbf{q}_\sharp V\|$ of $\vec{H}\in L^q_{\text{loc}}(\|V\|)$, the generalised mean curvature.)
\end{thm}
Indeed, an analysis of the proof of Theorem \ref{thm:main} shows that the condition that the varifolds in question arise as boundaries of Caccioppoli sets is only used in the following steps. It permits to give an orientation to said varifolds, so that we can treat them as oriented integral varifolds. It permits the use of Hutchinson's compactness theorem, as it implies that the boundary in the sense of currents is $0$, in particular the boundary masses of the associated currents are locally uniformly bounded. Finally, once the prescribed mean curvature condition is established for the limit (in the paragraph following (\ref{eq:cuvature_identity_limit_with_g})), it permits the multiplicity-$1$ conclusion on $\{g\neq 0\}$. In Theorem \ref{thm:overall} the first two facts are already in the hypotheses and we do not aim to conclude the third, we only need to establish the preservation of the prescribed mean curvature condition: this follows from the arguments of Section \ref{proofs} up to the paragraph that follows (\ref{eq:cuvature_identity_limit_with_g}).
If, on the other hand, we add the hypothesis that $V^\ell$ are associated to boundaries of Caccioppoli sets, we follow the arguments of Section \ref{proofs} to the end and establish:
\begin{thm}
\label{thm:Riemannian}
Theorems \ref{thm:C^2_case} and \ref{thm:main}, and Corollary \ref{cor:boundaries}, hold with $\Omega$ replaced by a Riemannian manifold (of dimension $n+1$).
\end{thm}
\begin{oss}
For completeness, we point out that if $V$ is as in Definition \ref{Dfi:Riemannin_oriented_curvature} then:
\begin{itemize}
\item[(1)] (as in Proposition \ref{Prop:further_prop_curv_coeff}) we have that $W_{ij}v_j=0$ and $W_{ij}=W_{ji}$, and in particular that $S_{rb} (W_{jr} v_k+ W_{jk}v_r) S_{kj}= W_{jk}S_{kj} v_b $.
\item[(2)] (as in Section \ref{relation_with_Hut}) $\textbf{q}_\sharp V$ is a varifold with curvature in the sense of \cite{Hut}, where we view $\textbf{q}_\sharp V$ as measure in $U\times \mathbb{R}^{N^2}$, and $\mathbb{R}^{N^2}$ has coordinates $P_{ij}$, $i, j\in \{1, \ldots, N\}$. To see this, we take $\varphi\in C^1_c(U \times \mathbb{R}^{N^2})$ and use the chain rule in (\ref{eq:Riemannin_oriented_curvature_B}) with $\phi(x,v)=\varphi(x, P(x,v))$ with $P_{ij}(x,v)=(\delta_{ik}-v_iv_k) S_{kj}(x)$:
\begin{eqnarray*}
D_s \phi = D_s \varphi + D^*_{cd}\varphi D_s P_{cd} = D_s \varphi + (D^*_{cd}\varphi) (\delta_{ck}-v_c v_k) (D_s S_{kd}),\\
D^*_a \phi = - (D^*_{cd}\varphi) S_{kd} (\underbrace{(D^*_a v_c)}_{\delta_{ac}} v_k + \underbrace{(D^*_a v_k)}_{\delta_{ak}} v_c) = - (D^*_{ad}\varphi) S_{kd} v_k - (D^*_{cd}\varphi) S_{ad} v_c,
\end{eqnarray*}
so substituting we obtain
\begin{eqnarray*}
\int P_{sb} D_s \varphi+ (D^*_{cd}\varphi) (\delta_{ck}-v_c v_k)\underbrace{\overbrace{P_{jb}S_{js}}^{P_{bs}} (D_s S_{kd})}_{P_{jb}\overline{A}_{jkd}}-\\
-(D^*_{ad}\varphi) \underbrace{S_{kd} v_k}_{v_d} W_{ba}- (D^*_{cd}\varphi) S_{ad} v_c W_{ba} + \left(P_{ir} \overline{A}_{irb} - S_{rb} (W_{jr} v_k+ W_{jk}v_r) S_{kj} \right) \phi=0.
\end{eqnarray*}
Define $A_{bcd}=(\delta_{ck}-v_c v_k)P_{jb}\overline{A}_{jkd} - S_{kd}v_k W_{bc}-S_{ad}v_c W_{ba}.$ Then we only need
$$A_{dbd}=(\delta_{bk}-v_b v_k)P_{jd}\overline{A}_{jkd} - S_{kd} v_k W_{db}-S_{ad}v_b W_{da}$$
to coincide with the expression in parenthesis, that is,
$$P_{ir} \overline{A}_{irb} -S_{rb} W_{dr}v_k S_{kd} - \underbrace{S_{rb} v_r}_{v_b} W_{jk} S_{kj}.$$
When $W_{ia}$ are associated to a $C^2$ hypersurface oriented by $\nu$ we have $S_{rb} W_{dr} = S_{rb} \delta_d (\nu \cdot \textbf{e}_r)=S_{rb}P_{da} D_a (\nu \cdot \textbf{e}_r)=P_{da} D_a (\nu \cdot S_{rb}\textbf{e}_r)=\delta_d (\nu \cdot (S \textbf{e}_b))=\delta_d (\nu \cdot\textbf{e}_b)=W_{db}$. By Proposition \ref{Prop:C^2_coeff_Riem} we then have $S_{rb} W_{dr}=W_{db}$ holds $V$-a.e.
Finally, notice that $S_{bk}=P_{bk}+v_b v_k$ $V$-a.e., therefore $(\delta_{bk}-v_b v_k)P_{jd}\overline{A}_{jkd}=P_{jd}\overline{A}_{jbd}+(P_{bk}-S_{bk})P_{jd}\overline{A}_{jkd}=P_{jd}\overline{A}_{jbd}$. The proof of the claim is complete.
\end{itemize}
\end{oss}
A concrete instance of Theorem \ref{thm:overall} for hypersurfaces that do not arise as boundaries is the following:
\begin{thm}
\label{thm:immersions}
Let $g_\ell, g$ ($\ell\in \mathbb{N}$) be continuous functions in an $(n+1)$-dimensional Riemannian manifold $\Omega$ such that $g_\ell \to g$ in $C^0_{\text{loc}}(\Omega)$. Let $M_\ell=\iota_\ell(\Sigma_\ell)$ be two-sided properly $C^2$-immersed closed hypersurfaces in $\Omega$ such that, for every open set $U\subset \subset \Omega$, $\sup_{\ell\in \mathbb{N}} \mathcal{H}^n(M_\ell \cap U) <\infty$. (We denote by $\Sigma_\ell$ abstract $n$-manifolds and by $\iota_\ell$ the immersion.) Denote by $\nu_\ell$ a determination of unit normal for $M_\ell$ and assume that the mean curvature vector of $M_\ell$ is given by $g_\ell \nu_\ell$ for all $\ell\in \mathbb{N}$. (This is well-defined on $\Sigma_\ell$.) Denote by $B^{\ell}$ the second fundamental form of the immersion $M_\ell$ and assume that for every open set $U \subset \subset \Omega$ we have
$$\sup_{\ell\in \mathbb{N}} \int_{\Sigma_\ell \cap \iota_\ell^{-1}(U)} |B^{\ell}|^q \,d\mathcal{H}^n<\infty \text{ for some } q>1.$$
Then every varifold limit of $|M_\ell|$ is $\textbf{q}_\sharp V$ for an oriented integral varifold $V$ and the first variation of $\textbf{q}_\sharp V$ (as a functional on $C^1_c$-vector fields in $\Omega$) is represented by integration (with respect to $\textbf{q}_\sharp V$) of $g \nu$, where $\nu$ is a measurable unit normal field on $\textbf{q}_\sharp V$ (a.e.~defined and orthogonal to $\textbf{q}_\sharp V$).
\end{thm}
For example, let $g_\ell =g \equiv 2$ and $\Omega = B^n_1(0) \times (-1,1)$, and consider $S = \{(x,y):x\in B_1^n(0), y =-\sqrt{1-|x|^2}\}$ and $M_\ell = (S + \frac{1}{\ell}\textbf{e}_{n+1}) \cup (S - \frac{1}{\ell}\textbf{e}_{n+1})$, where the choice of orientation $\nu_\ell$ is given by the unit vector whose scalar product with $\textbf{e}_{n+1}$ is positive. Then the varifold limit is $2|S|$ and $\nu$ is the upwards unit normal to $S$. Note that, with the given orientation, $M_\ell$ are not boundaries.
\begin{proof}
The two-sidedness assumption permits to consider the sequence of oriented integral varifolds $V^\ell=(M_\ell, \nu_\ell, 1,0)$. As the hypersurfaces $M_\ell$ are closed in $\Omega$, the associated currents are cycles in $\Omega$, hence Hutchinson's compacntess theorem applies to yield a (subsequential) limit $V$ in the sense of oriented integral varifolds. Then the proof of Theorem \ref{thm:main} can be repeated up to the point where we establish that $V$ satisfies (\ref{eq:cuvature_identity_limit_with_g}), from which the prescribed mean curvature condition is seen to hold for the limit and the conclusion of the theorem follows.
\end{proof}
\begin{oss}
In view of the hypotheses in Hutchinson's compactness theorem in \cite{Hut} (recalled in Section \ref{prelim}) one could adapt the notion of curvature varifolds with boundary, following Mantegazza's work \cite{Man}, to the oriented case and obtain a compactness result for $M_\ell$ immersed hypersurfaces-with-boundary with masses of the boundaries locally uniformly bounded in $\ell$.
\end{oss}
\begin{small}
|
\section{Introduction}
We begin with the following definitions that play a pivotal role in this paper. In addition, by the capital letter $N$ we denote a standard Gaussian random variable $N \sim N(0,1)$ with density function $p_N (x)=(2\pi)^{-1/2}e^{-x^2/2} $, $x\in {\rm I\kern-0.16em R}$. All random variables throughout the paper are real-valued. The first definition concerning polynomial Stein operators is taken from our recent paper \cite{a-g-g-algebraic-stein-operators}.
\begin{mydef}\label{def:PSO}
Let $X$ be a (continuous) target random variable. Let $T \ge1$. We say that a linear differential operator $S = \sum_{t=0}^{T} p_t \partial^t $ acting on a suitable class $\mathcal{F}$ of functions is a \textbf{polynomial Stein operator} for $X$ if :
\begin{itemize}
\item[(a)] $Sf\in L^1(X)$,
\item[(b)] ${\rm I\kern-0.16em E} \left[ S f(X) \right] =0$ for all $f \in \mathcal{F}$, and
\item[(c)] the coefficients $p_t$, $t=0,1,\ldots,T$, are polynomial.
\end{itemize}
By $\mathrm{PSO}_{\mathcal{F}}(X)$ we denote the set of all polynomial Stein operators, acting on a class $\mathcal{F}$ of functions, for the target random variable $X$.
Hereafter, for the ease of notation, we write $\mathrm{PSO}_{\mathcal{F}}(X) = \mathrm{PSO}(X)$.
\end{mydef}
\begin{mydef}[\textbf{Characterising polynomial Stein operators}]\label{def:Charactrization}
We say that an operator $S \in \mathrm{PSO}(Y)$ is a \textbf{characterising} Stein operator for the target random variable $Y$ if for any given random variable $X$ such that $Sf \in L^1({\rm I\kern-0.16em P}_X)$ for all $f \in \mathcal{F}$, the condition
\begin{equation*}
{\rm I\kern-0.16em E} \left[ Sf (X)\right]=0, \quad \forall \, f \in \mathcal{F}
\end{equation*}
implies $X =_d Y$ (equality in distribution).
\end{mydef}
\begin{rem}
{\rm Observe that $0\in\mathrm{PSO}(X)$ is a trivial (and non-characterising) Stein operator for the target random variable $X$.}
\end{rem}
Polynomial Stein operators lie at the heart of Stein's method \cite{stein} and the Nourdin-Peccati Malliavin-Stein method \cite{np09,n-p-book}, powerful techniques for proving quantitative probabilistic limit theorems. Indeed, the starting point of Stein's method for Gaussian approximation \cite{stein,chen,n-p-book} is the following characterisation of the standard Gaussian distribution, known as \emph{Stein's Lemma}. Consider the divergence operator $G=\partial-x$, where $\partial$ denotes the usual differential operator. Then, a real-valued random variable $X$ is equal in distribution to $N\sim N(0,1)$ if and only if ${\rm I\kern-0.16em E}[Gf(X)]=0$ for all absolutely continuous $f:{\rm I\kern-0.16em R}\rightarrow{\rm I\kern-0.16em R}$ such that ${\rm I\kern-0.16em E}|f'(N)|<\infty$.
The operator $G$ is not the only non-trivial element of $\mathrm{PSO}(N)$. Letting $m\geq1$, consider the polynomial differential operators $S_m = H_{m-1}\partial - H_m $ and $ L_m= \partial^m - H_m $, where $H_m$ stands for the probabilist's Hermite polynomial of degree $m$. It is easily proved by integration by parts that $S_m, L_m \in \mathrm{PSO}(N)$. These operators were introduced as natural generalisations of the operator $G$ (reducing to $G$ when $m=1$) in the delightful paper \cite{gr05}.
More generally, if we let $k, t\geq1$, then $S(k,t)=H_{k}\partial^t-H_{k+t}\in\text{PSO}(N)$ (see \cite[p.\ 294]{b86}). The fact that $S(k,t)\in\text{PSO}(N)$ yields a useful integration by parts formula, which \cite{b86} exploited to derive asymptotic expansions for smooth functions of sums of independent random variables.
Two natural questions arise.
\begin{description}
\item[(Q1)] Are there other elements of $\text{PSO}(N)$ beyond the operators $S(k,t)$, $k,t\geq1$, and if so can a description of a typical element be given?
\item[(Q2)] Are the operators in the class $\text{PSO}(N)$ characterising in the sense of Definition \ref{def:Charactrization}?
\end{description}
In this paper, we provide answers to both questions (Q1) and (Q2). We do so by making a connection between Stein's method and noncommutative algebra. In particular, we view polynomial Stein operators as elements of the first Weyl algebra, the ring of linear differential operators with polynomial coefficients (in one variable). Weyl algebras enjoy a rich algebraic structure and their properties are well-understood; we refer the reader to \cite{Bjork-Book, Coutinho-Book, gw-Book, mr-Book87, Lam-Book, cf-Book} and references therein for introductions and comprehensive accounts of the theory. For example, it is well-known that Weyl algebras are simple integral domains, Noetherian (left/right) and Ore extensions. Two-sided ideals of Weyl algebras have plain structure; however, the structure of one-sided ideals is less obvious. Forasmuch as being Noetherian implies all one-sided ideals are finitely generated; however, more delicately, a famous theorem by Stafford \cite{Stafford-2generators} says that every one-sided (left/right) ideal of Weyl algebras can be generated by two elements. With this powerful theory at hand, we are able to address questions (Q1) and (Q2)
and also gain further insights on Stein's method that are perhaps surprising from a probabilistic perspective.
For (Q1), we provide a complete algebraic description of the class $\text{PSO}(N)$ (Theorem \ref{prop:PSO=Ideal}). We show that $\text{PSO}(N)$ is a vector space over ${\rm I\kern-0.16em R}$ with the operators $S(k,t)$, $k,t\geq1$, in fact forming a basis. In addition, $\text{PSO}(N)$ is identified as a principle right ideal of the first Weyl algebra generated by Stein's classical operator $G=\partial-x$. Various interpretations have been given as to why $G$ is a natural object in Gaussian analysis via Stein's method, such as the the generator approach to Stein's method of \cite{b88,barbour2,gotze} in which $(\partial-x)\partial$ is recognised as the generator of an Ornstein-Uhlenbeck process with standard Gaussian stationary distribution, and the canoncial Stein operator viewpoint of \cite{ley}. Here we provide another such interpretation. Also, whilst methods exist for obtaining families of Stein operators for a given target distribution (see the flexible approach to Stein's density method of \cite{ley,mrs21} and our algorithmic approach \cite{a-g-g-algebraic-stein-operators}), this is the first time in the literature that a complete description has been given for the class of polynomial Stein operators for a given target distribution.
The characterising problem (Q2) is the subject of Section \ref{sec4}. In Corollary \ref{cor:Hermite-Coeff-Is-Characterizing}, we show that the polynomial Stein operators $S_m=H_{m-1}\partial-H_m$, $m\geq1$, are characterising in the sense of Definition \ref{def:Charactrization}. This is a consequence of a more general result that all polynomial Stein operators for the standard Gaussian distribution that are first order linear differential operators are characterising (see again Corollary \ref{cor:Hermite-Coeff-Is-Characterizing}). We remark that the Stein operators $(\partial-x)\partial^n$, $n\geq0$, that is the Stein operators for the Gaussian distribution with coefficients with polynomial degree one, are seen to be characterising as a direct consequence of Proposition 2.1 of our recent paper \cite{a-g-g-Stein-charactrization}. We also show that the Stein operators $S(k,t)=H_{k}\partial^t-H_{k+t}\in\text{PSO}(N)$, $k,t\geq1$, are characterising amongst the class of symmetric and infinitely divisible distributions (Corollary \ref{barops}). We conjecture that the additional assumptions of symmetric and infinitely divisible distributions are artefacts of our proof, and that these assumptions can be disposed of. However, it turns out that there are non-trivial elements of $\mathrm{PSO}(N)$ that are not characterising unless additional distributional assumptions are made; in part (d) of Example \ref{ex:Moment_Assumption_Crucial}, we give a general recipe for constructing such non-characterising polynomial Stein operators for the Gaussian distribution. This non-characterising property of elements of the class $\mathrm{PSO}(N)$ is in fact a special case of the result that if the characteristic functions of $X_1$ and $X_2$ are holonomic (satisfy a linear homogeneous differential equation with polynomial coefficients), then $\mathrm{PSO}(X_1)\cap\mathrm{PSO}(X_2)\not=0$ (see Theorem \ref{thm:PSO_Intersection_NonTrivial}). From a probabilistic point of view this result may seem surprising, but it is in fact a consequence of basic properties of Weyl algebras. This demonstrates the power of the connection between Stein's method and noncommutative algebra, and in this paper we have likely only scratched the surface of what can be achieved through this connection. We hope this paper inspires both probabilists and algebraists to further explore this seemingly fruitful new window on Stein's method.
\begin{rem}[\textbf{Caution on the class of functions $\mathcal{F}$}]\label{rem_cau}
{\rm We require that the class $\mathcal{F}$ is \textit{invariant} under any polynomial differential operator, that is to say $S \mathcal{F} \subseteq \mathcal{F}$ for any differential operator $S$ with polynomial coefficients. For probabilistic purposes, the following natural choices for function space $\mathcal{F}$ are available: (a) $\mathcal{F}={\rm I\kern-0.16em R}[x]$ the ring of all polynomials with real coefficients; (b) $\mathcal{F} = C^\infty_c ({\rm I\kern-0.16em R})$ the space of all infinitely differentiable functions with compact support; (c) the class of Schwartz functions $\mathcal{F} = \mathcal{S} ({\rm I\kern-0.16em R}): = \{ f \in C^\infty ({\rm I\kern-0.16em R}) \, : \,
\sup_{x\in {\rm I\kern-0.16em R}} \vert x^\alpha \partial^\beta f (x) \vert < \infty, \, \forall \, \alpha,\beta \in {\rm I\kern-0.16em N}_0 \}$, which includes the class $C^\infty_c({\rm I\kern-0.16em R})$; and (d) $\mathcal{F}=P({\rm I\kern-0.16em R}):=\{ f \in C^\infty ({\rm I\kern-0.16em R}) \, : \,
\sup_{x\in {\rm I\kern-0.16em R}} \vert (1+x^\alpha)^{-1} \partial^\beta f (x) \vert < \infty, \, \forall \, \alpha,\beta \in {\rm I\kern-0.16em N}_0 \}$ the space of infinitely differentiable functions with polynomial growth. Note that these families of functions are invariant under any polynomial differential operator.
The Schwartz class $\mathcal{F} = \mathcal{S} ({\rm I\kern-0.16em R})$ is applicable to all target distributions. When target distributions have all moments (such as distributions in Wiener chaoses or probability distributions with bounded support), the classes $\mathcal{F}={\rm I\kern-0.16em R}[x]$ and $\mathcal{F}=P({\rm I\kern-0.16em R})$ are also available.
The class $\mathcal{F}={\rm I\kern-0.16em R}[x]$ is of particular interest for target distributions that are moment-determined. It is classical that the standard Gaussian distribution is moment-determined; see e.g. \cite[Lemma 3.1.]{n-p-book}. The class $\mathcal{F}=P({\rm I\kern-0.16em R})$ is useful in practical implementations of Stein's method, containing the real and and imaginary parts of the function $f(x)=e^{\mathrm{i}tx}$, $t\in\mathbb{R}$, and ensuring that, for $f$ in this class, ${\rm I\kern-0.16em E}|X^\alpha\partial^\beta f(X)|<\infty$ for all $\alpha,\beta\in{\rm I\kern-0.16em N}_0$, if $X$ has all moments. To keep the statements of our results concise, when considering an arbitrary target random variable, the shorthand notation $\text{PSO}(X)$ can be safely considered to represent $\text{PSO}_{\mathcal{F}}(X)$, where $\mathcal{F} = \mathcal{S}({\rm I\kern-0.16em R})$.
However, if all moments of the target random variable(s) exist (as is the case for the Gaussian distribution), then the results also apply for $\mathcal{F}={\rm I\kern-0.16em R}[x]$ and $\mathcal{F}=P({\rm I\kern-0.16em R})$.
}
\end{rem}
\section{Auxiliary background} \label{se:Background}
\subsection{Elements of Weyl algebra $A_n(\mathbb{K})$}\label{sec:Weyl_Algebra}
Throughout this section, $\mathbb{K}$ denotes a commutative field of characteristic zero. For our purposes, either $\mathbb{K}={\rm I\kern-0.16em R}$, or $\mathbb{K}=\mathbb{C}$ (the field of complex numbers). Next, we collect some basic definitions and results from noncommutative algebra. For a comprehensive treatment of the subject, the reader can consult the classical textbooks \cite{mr-Book87, cf-Book, gw-Book, Lam-Book}.
\begin{mydef}[\textbf{Ring}]\label{def:Ring}
A ring is a triplet $(R, +, \cdot)$ in which $R$ is a non-empty set equipped with two binary operations $+$ and $\cdot$ called addition and multiplication, respectively, such that:
\begin{itemize}
\item[(i)] $(R,+)$ is an abelian group.
\item[(ii)] $(R,\cdot)$ is a monoid.
\item[(iii)] Addition and multiplication are related to each other via distributive laws: $(a+b)\cdot c = a \cdot b + b\cdot c$ and $ a \cdot (b+c) = a \cdot b + a \cdot c$.
\end{itemize}
The neutral element of $R$ under the addition operator is called zero and denoted by $0_R=0$, whereas the neutral element under the multiplication operator is called one or the unit-element and is denoted by $1_R=1$. The inverse of an element $a \in R$ under the addition operator is denoted by $-a$, i.e., $a+ (-a)=0$. The ring $R$ is called commutative if $a\cdot b = b\cdot a$ for all $a,b \in R$, and otherwise we say $R$ is noncommutative. We often omit the multiplication sign and write $ab$ instead of $a \cdot b$.
\end{mydef}
\begin{mydef}[\textbf{Ideal}]\label{def:Ideal}
A left (right) ideal $I$ of a ring $R$ is a subgroup of $(R,+)$ such that $RI \subseteq I$ $(IR \subseteq I)$. We also write $I \lhd_l R $ and $I \lhd_r R$ to denote a left/right ideal. A two-sided ideal $I$ -- often referred to simply as an ideal -- is a left and right ideal simultaneously, that is $RI \subseteq I $ and $IR \subseteq I$. We denote it by $I \lhd R$.
\end{mydef}
\begin{mydef}[\textbf{Module}]\label{def:Module}
Let $R$ be a ring (not necessarily commutative). An abelian group $(M,+)$ is called a left $R$-module if there is a mapping $\bullet : R \times M \to M$ such that for all $a,b \in R$ and for all $m,m' \in M$:
\begin{itemize}
\item[(i)] $a \bullet (m+m') = a\bullet m + a \bullet m'$.
\item[(ii)] $ (a+b)\bullet m = a \bullet m + b \bullet m$.
\item[(iii)] $(ab) \bullet m = a \bullet (b\bullet m)$.
\item[(iv)] $1\bullet m = m$, where $1 = 1_R$.
\end{itemize}
In brief, we write ${}_R M$ to say $M$ is a left $R$-module. A right $R$-module is defined similarly, and denoted by $M_R$. A left $R$-module ${}_R M$ is said to be finitely generated if there is a finite set $X=\{ x_i \, : \, i \in I\} \subseteq M$, say of the cardinality $ \vert I \vert =n$, so that for every element $m \in M$ there are $a_1,\ldots,a_n \in R$ such that $m = a_1\bullet x_1 +\ldots+a_n \bullet x_n$. In this case we write ${}_R M = {}_R \langle X \rangle = {}_R \langle x_1,\ldots,x_n \rangle$. If $n=1$, that is when ${}_R M$ is generated by only one element, it is called a cyclic module.
\end{mydef}
\begin{ex}
(a) Let $R$ be a ring. Every left ideal $I \lhd_l R $ with the ring multiplication is a left $R$-module. Similarly, every right ideal $I \lhd_r R $ with the ring multiplication is a right $R$-module. (b) Let $a_1,\ldots,a_n \in R$. Then $ {}_R \langle a_1,\ldots,a_n \rangle = \sum_{i=1}^{n} Ra_i = \{ \sum_{i=1}^{n} r_i a_i \, : \, r_i \in R, \, 1 \le i \le n \}$ is a finitely generated left $R$-module.
\end{ex}
\begin{mydef}[\textbf{Annihilator}]\label{def:Annihilator}
Let ${}_R M$ be a left $R$-module. Let $X \subseteq M$ be a subset of $M$. The annihilator of $X$, denoted by $\mathrm{Ann}_{R}(X)$, consists of those elements $r \in R$ such that for all elements $x \in X$, $r\bullet x =0$. In set notation,
\begin{equation*}
\mathrm{Ann}_R (X) = \{ r \in R \, : \, r \bullet x =0 \,\, \forall \, x \in X \}.
\end{equation*}
\end{mydef}
\begin{mydef}\label{def:Basic_Ring_Theory}
Let $R$ be a ring (not necessarily commutative).
\begin{itemize}
\item[(a)] The ring $R$ is an integral domain if the product of two non-zero elements is non-zero, i.e., $ab\neq0$ for every $0 \neq a ,b \in R$.
\item[(b)] The ring $R$ is simple if $0$ and $R$ are the only two-sided ideals of $R$.
\item[(c)] The ring $R$ is a principal left (right) ideal ring if every left (right) ideal is principal (or generated by only one element).
\item[(d)] A left (right) ideal $I \lhd_l R$ ($I \lhd_r R$) is called maximal if there is no left (right) proper ideal $J$ containing $I$.
\item[(e)] The ring $R$ is prime if for every $0 \neq a,b \in R$ then $a R b \neq 0$. A left (right) ideal $I \lhd_l R$ ($I \lhd_r R$) is called a prime ideal if the quotient $R/I$ is a prime ring.
\item[(f)] The ring $R$ is called left (right) Noetherian if it satisfies the left (right) ascending chain condition,
namely that every chain of left (right) ideals $ I_1 \subseteq I_2\subseteq I_3\subseteq \cdots $ must terminate after a finite step, i.e., there exists $n\in{\rm I\kern-0.16em N}$ such that $I_n=I_{n+1}=\cdots$.
\end{itemize}
\end{mydef}
\begin{mydef}[\textbf{Weyl algebras}]\label{def:Weyl_Algebra}
Let $n \in {\rm I\kern-0.16em N}$ be a natural number. The $n$th Weyl algebra $A_n (\mathbb{K})$ is an associative $\mathbb{K}$-algebra (a vector space over the field $\mathbb{K}$ equipped with a bilinear operator which is associative, that is to say $r(a\cdot b) = (ra)\cdot b = a \cdot (rb) $ for every $r \in \mathbb{K}$ and every $a,b \in A_n (\mathbb{K})$) with generators $x_n,\ldots ,x_n$, $y_1,\ldots ,y_n$ subject to the relations
\begin{equation*}
[x_i,x_j]=[y_i,y_j]=0, \quad [x_i,y_j]=\delta_{ij}, \quad \forall \, 1 \le i,j\le n,
\end{equation*}
where $[a,b]=ab-ba$ is the commutator of the elements $a$ and $b$, and $\delta_{ij}$ denotes the Kronecker delta. Let $\mathbb{K}[x_1,\ldots ,x_n]$ be the ring of polynomials in $n$ variables $x_1,\ldots ,x_n$ with coefficients in $\mathbb{K}$. It is classical that the $n$th Weyl algebra $A_n (\mathbb{K})$ is isomorphic (as a $\mathbb{K}$-algebra) to the ring of differential operators with polynomial coefficients in $n$ variables by identifying each $y_i$ with $\partial_i = \partial_{x_i}$. Clearly, $A_n(\mathbb{K})$ is a noncommutative ring.
\end{mydef}
The following theorem collects some algebraic properties of Weyl algebras.
\begin{thm}\label{thm:Weyl_Algebra_Property}
\begin{itemize}
\item[(a)] The Weyl algebra $A_n (\mathbb{K})$ is a simple Noetherian (left/right) integral domain, and hence is a prime ring.
\item[(b)] The Weyl algebra $A_n (\mathbb{K})$ satisfies the left/right Ore property, namely that $A_n (\mathbb{K}) L \cap A_n (\mathbb{K}) S \neq 0$ for every two non-zero elements $L, S \in A_n (\mathbb{K})$.
\end{itemize}
\end{thm}
\begin{mydef}\label{def:Holonomic-Module}
\begin{itemize}
\item[(a)] {\rm{\textbf{(Bernstein filtration)}}}. Let $\nu \in {\rm I\kern-0.16em N}_0$. Let $\mathscr{F}_\nu$ be a $\mathbb{K}$-vector space which is generated by elements $\{ x^\alpha \partial^\beta \, :\, \vert \alpha \vert + \vert \beta \vert \le \nu \}$, where for a multi-index $\alpha=(\alpha_1,\ldots,\alpha_n)$ we let $\vert \alpha \vert = \sum_{i=1}^{n} \alpha_i$. Clearly, each $\mathscr{F}_\nu$ is a finite dimensional $\mathbb{K}$-vector space, $\mathscr{F}_0 \subset \mathscr{F}_1 \subset \cdots$, $\cup_{\nu=0}^{\infty} \mathscr{F}_\nu = A_n (\mathbb{K})$ and $\mathscr{F}_\nu \mathscr{F}_k \subset \mathscr{F}_{\nu+k}$. The family $\mathscr{F} = \{ \mathscr{F}_\nu \, : \, \nu \ge 0 \}$ is called the Bernstein filtration for the Weyl algebra $A_n(\mathbb{K})$. The direct sum $\mathrm{gr} \left( A_n(\mathbb{K}) \right) : = \mathscr{F}_0 \oplus \frac{\mathscr{F}_1}{\mathscr{F}_0} \oplus \frac{\mathscr{F}_2}{\mathscr{F}_1} \oplus \cdots $ is called the associated graded ring.
\item[(b)] {\rm{\textbf{(Good filtration)}}}. Let $M$ be a left $A_n(\mathbb{K})$-module. A filtration $\Gamma = (\Gamma_\nu \, : \, \nu \ge 0)$ is an increasing sequence of finite dimensional vector subspaces on $\mathbb{K}$ such that $M = \cup_{\nu=0}^{\infty} \Gamma_\nu$ and $ \mathscr{F}_k \Gamma_\nu \subset \Gamma_{\nu+k}$ for every pair $k$ and $\nu$. For a given filtration $\Gamma$, the associated graded $\mathrm{gr} \left( A_n(\mathbb{K})\right)$-module is given as $\mathrm{gr}_{\Gamma} M : = \Gamma_0 \oplus \frac{ \Gamma_1}{\Gamma_0} \oplus \frac{\Gamma_2}{\Gamma_1} \oplus \cdots$. A filtration $\Gamma$ is called a good filtration if the associated graded $\mathrm{gr}_{\Gamma} M$ is a finitely generated graded $\mathrm{gr} \left( A_n(\mathbb{K})\right)$-module.
\item[(c)] {\rm{\textbf{(Bernstein dimension)}}}. Let $\Gamma$ be a good filtration on a finitely generated left $A_n(\mathbb{K})$-module $M$. Then, it is well-known that there is $d \ge 0$ and rational numbers $a_0,\ldots,a_d$ so that $\mathrm{dim}_K \Gamma_\nu = a_d \nu^d + \cdots+a_0$ for $\nu \gg 0$.
The integer $d=: d_\Gamma(M) = d(M)$ is called the Bernstein dimension, and it is well-known that $d$ is independent of the choice of good filtration.
\item[(d)] {\rm{\textbf{(Holonomic module)}}}. A finitely generated left $A_n(\mathbb{K})$-module $M$ is called a holonomic module if $d(M)=n$. It is well-known that for every non-zero finitely generated left $A_n(\mathbb{K})$-module $M$ it holds $d(M)\ge n$. This is known as Bernstein's inequality.
\end{itemize}
\end{mydef}
\begin{thm}[\textbf{Stafford's Theorems}]\label{thm:Stafford}
\begin{itemize}
\item[(a)] Every left (right) ideal in the $n$th Weyl algebra $A_n(\mathbb{K})$ may be generated by two elements -- see e.g.
\cite[Theorem 3.1]{Stafford-2generators}.
\item[(b)] Every holonomic $A_n$-module is cyclic, i.e. generated by one element -- see e.g. \cite[Corollary 2.6]{Coutinho-Book}.
\end{itemize}
\end{thm}
\begin{ex}
Let $M=C^\infty(\mathbb{K})$ be the vector space of infinitely many differentiable functions over the field $\mathbb{K}$. Define the mapping $\bullet : A_1(\mathbb{K}) \times M \to M$, where the action of $L=\sum_{t=0}^{T} p_t \partial^t$ on $f$ is defined by $L \bullet f = \sum_{t=0}^{T} p_t \partial^t f$. Recall that the coefficients $p_t$, $t=0,1,\ldots,T$, are polynomials in one variable. It is straightforward to see that ${}_{A_1(\mathbb{K})} M$ is a left $A_1(\mathbb{K})$-module.
\end{ex}
\begin{mydef}[\textbf{Holonomic function}]\label{def:Holonomic-Function}
A function $f: {\rm{Dom}} (f) \subseteq \mathbb{K} \to \mathbb{K}$ is called holonomic (or $\partial$-finite, or $D$-finite) if it satisfies one of the following equivalent conditions:
\begin{itemize}
\item[(a)] $\mathrm{Ann}_{A_1(\mathbb{K})} (f) : = \{ L \in A_1(\mathbb{K}) \, : \, L \bullet f =0 \} \neq 0$.
\item[(b)] The vector space spanned by $(\partial^t f : t =0, 1, \ldots )$ is finite dimensional over the field of rational functions $\mathbb{K}(x) = \big\{ \frac{p(x)}{q(x)} \, : \, p, q \in \mathbb{K}[x] \big\}$ with the usual addition and multiplication operations on polynomials.
\end{itemize}
\end{mydef}
For an introduction to holonomic function theory the reader is referred to the excellent textbooks \cite{Stanley-Book, Flajolet-Book}. The class of holonomic functions enjoys rich closure properties. The following result is borrowed from \cite[Theorem B2]{Flajolet-Book} and \cite[Theorem 1]{Non-Holonomic-Character}.
\begin{thm}\label{thm:Holonomic-Closure-Properties}
\noindent{(a)} The class of univariate holonomic functions is closed under the following mathematical operations: (i) sum, (ii) product, (iii) differentiation, (iv) indefinite integration, and (v) algebraic substitution.
\vspace{1mm}
\noindent{(b)} A holonomic function has only finitely many singularities.
\vspace{1mm}
\noindent{(c)} If the function $f(x_1,\ldots,x_n)$ is holonomic, then, for any $a$ and $b$, the function given by $ \int_{a}^{b} f(x_1,\ldots,x_{n-1},x_n) \,dx_n$ is also holonomic, if it is defined (\cite[Proposition 3.5]{Zeilberger90}).
\end{thm}
\subsection{A gentle introduction to Malliavin operators}\label{sec:Malliavin}
For the scope of our paper, it is enough to define Malliavin operators in dimension $d=1$ algebraically, without discussing their functional analytic extensions.
For a state-of-the-art exposition of Malliavin calculus in full generality see \cite{n-p-book,n-n-book}.
\begin{mydef} Let $\mathscr{S}$ denote the space of smooth functions consisting of $C^\infty$-functions $f:{\rm I\kern-0.16em R} \to {\rm I\kern-0.16em R}$. In the univariate case, the Malliavin derivative $D$ and the divergence operator $\delta$ are defined as linear mappings acting on the space of smooth functions $\mathscr{S}$, with
\begin{align*}
&D f= \partial f = f', \quad
\delta f =(x-\partial) f = xf - f'.
\end{align*}
Let $p\ge 2$. We define the $p$th Malliavin derivative and divergence operators as follows: $D^p f = \partial^p f = f^{(p)}$, and $\delta^p f = \delta \left( \delta^{p-1} f\right) = \delta^{p-1} (\delta f)$.
\end{mydef}
\begin{prop}[\textbf{Gaussian integration by parts}] \label{prop:GIBP-d=1}
Let $N \sim N(0,1)$. For $f,g\in \mathscr{S}$,
\begin{align}\label{lem:GIBP-d=1}
{\rm I\kern-0.16em E}\bigl[ f(N) D g(N) \bigr]= {\rm I\kern-0.16em E}\bigl[ f(N) \partial g(N) \bigr]= {\rm I\kern-0.16em E}\bigl[ g(N) \delta f( N) \bigr].
\end{align}
\end{prop}
\begin{mydef}[\textbf{Hermite polynomials}]\label{def:Hermite_Polynomials}
We define the univariate Hermite polynomials as $H_0(x)=1$, $H_1(x)=x$, and
\begin{equation*}
H_n(x) =\delta H_{n-1}(x) = \bigl (\delta^n 1 \bigr) (x), \quad n \ge 2.
\end{equation*}
\end{mydef}
\begin{prop}[\textbf{Properties of Hermite polynomials}]\label{prop:Hermite_Polynomials_Properties}
\begin{enumerate}
\item[(a)]
$H_0(x),\dots, H_n(x)$ are monic polynomials spanning $\mathbb{K}_n[x]$ (the ring of polynomials of maximum degree $n$).
\item[(b)] For $n \ge 1$, $D H_n(x)= \partial H_n (x)= n H_{n-1}(x)$.
\item[(c)] Hermite polynomials are orthogonal in the sense that for $N \sim N(0,1)$ it holds that
\begin{equation*}
{\rm I\kern-0.16em E}\bigl[ H_m(N) H_n(N) \bigr]=
{\rm I\kern-0.16em E}\bigl[ H_m(N) (\delta^{n} 1 ) (N) \bigr]={\rm I\kern-0.16em E}\bigl[ 1 \; \partial_n H_m(N) \bigr ] = \left \{
\begin{matrix} n!, & n=m, \\
0, & \mbox{ otherwise. } \\
\end{matrix} \right .
\end{equation*}
\end{enumerate}
\end{prop}
\section{Algebraic structure of PSO($N$) }\label{sec:Aplgebraic_Structures}
We open this section with a general result on the algebraic structure of $\text{PSO}(X)$, the class of polynomial Stein operators of a given continuous target random variable $X$.
\begin{prop}\label{prop:PSO_General_Targets}
\begin{itemize}
\item[(a)] Let $\varphi_X$ denote the characteristic function of the random variable $X$. Then $\mathrm{Ann}_{A_1(\mathbb{C})} (\varphi_X) \neq 0$ if and only if $\varphi_X$ is holonomic. In addition, $\mathrm{PSO}(X) \neq 0$ if and only if $\mathrm{Ann}_{A_1(\mathbb{C})} (\varphi_X) \neq 0$.
\item[(b)] Let $p_X$ denote the probability density function of the continuous random variable $X$ having an integrable characteristic function. Then, $\mathrm{PSO}(X) \neq 0$ if and only if $p_X$ is holonomic.
\item[(c)] $\mathrm{PSO}(X) \lhd_r A_1({\rm I\kern-0.16em R})$ is a \textbf{right} ideal of the first Weyl algebra $A_1({\rm I\kern-0.16em R})$. Moreover, $\mathrm{PSO}(X)$ can be generated by two elements, i.e. $\mathrm{PSO}(X)= \langle G_1, G_2 \rangle_{A_1({\rm I\kern-0.16em R})}$ for some $G_1, G_2 \in A_1({\rm I\kern-0.16em R})$.
\item[(d)] $\mathrm{PSO}(X)$ is a right principal ideal if and only if $\mathrm{Ann}_{A_1(\mathbb{C})} (\varphi_X) \lhd_l A_1(\mathbb{C})$ is a \textbf{left} principal ideal in the first Weyl algebra $A_1(\mathbb{C})$.
\end{itemize}
\end{prop}
We will need the following lemma in our proof of Proposition \ref{prop:PSO_General_Targets}.
\begin{lem}\label{lem:Anti-Isomorphism}
The Fourier operator $\Psi : A_1(\mathbb{C}) \to A_1(\mathbb{C})$, given by
\begin{align*}
x^n \partial_x^k \mapsto
\Psi\bigl( x^n \partial_x^k )=
\mathrm{i}^{k-n} t^k \partial_t^n, \quad k,n\in {\rm I\kern-0.16em N}_0,
\end{align*}
extended by linearity is an anti-isomorphism of the Weyl algebra $A_1(\mathbb{C})$, that is to say, a bijection so that for every $S_1,S_2 \in A_1(\mathbb{C})$,
\begin{align*}
\Psi( S_1 S_2 ) =\Psi( S_2 ) \Psi(S_1).
\end{align*}
\end{lem}
\begin{proof}
Clearly, $\Psi$ is a bijective map. On the other hand, the set $\{ x^n \partial^k_x \, :\, n,k \in {\rm I\kern-0.16em N}_0\}$ provides a basis (as a vector space) over the field $\mathbb{C}$; see \cite[Proposition 1.2]{Bjork-Book}. Therefore, since
\begin{align*}
\Psi\bigl( x^n \partial^k_x\bigr)
\Psi\bigl( x^m \partial^{\ell}_x\bigr)
= \mathrm{i}^{k+\ell-n-m} t^k \partial^n_t t^{\ell}
\partial^m_t
= \mathrm{i}^{k+\ell-n-m} \sum_{r=0}^{n\wedge \ell}
\frac{n! \ell! }{r! (n-r)!(\ell-r)!}
t^{k+\ell-r} \partial_t^{m+n-r}
\end{align*}
and
\begin{align*}
\Psi\bigl( x^m \partial^{\ell}_x x^n \partial^k_x \bigr)
&=\Psi\biggl(
\sum_{r=0}^{\ell\wedge n} \frac{ \ell!}{r! (\ell-r)!}
\frac{n!}{(n-r)!} x^{n+m-r}\partial_x^{k+\ell-r}
\biggr)\\
& = \sum_{r=0}^{\ell\wedge n} \frac{ \ell!n!}{r! (\ell-r)!(n-r)!}
\Psi\bigl( x^{n+m-r}\partial_x^{k+\ell-r} \bigr)\\
&= \mathrm{i}^{k+\ell-n-m} \sum_{r=0}^{n\wedge \ell}
\frac{ \ell!n! }{r! (\ell-r)!(n-r)!}
t^{k+\ell-r} \partial_t^{m+n-r},
&\end{align*}
the result follows.
\end{proof}
\noindent{\emph{Proof of Proposition \ref{prop:PSO_General_Targets}.}} (a) The first claim is a direct application of Definition \ref{def:Holonomic-Function}. For the second claim, first assume $0 \neq S \in \text{PSO}(X)$. Let $\mathcal{L} = \Psi(S)$ be the image of $S$ under the Fourier mapping $\Psi$ in Lemma \ref{lem:Anti-Isomorphism}. Then $ 0 \neq \mathcal{L} \in \text{Ann}_{A_1(\mathbb{C})} (\varphi_X)$ because $\mathcal{L} \varphi_X (t) = {\rm I\kern-0.16em E}[S e^{\mathrm{i} t X}] =0$ for every real $t$, where we applied the Stein operator $S$ to the real and imaginary parts of $e^{\mathrm{i} t X}$.
For the other direction,
assume that $\text{Ann}_{A_1(\mathbb{C})}(\varphi_X)$ is non-trivial. In other words, the characteristic function $\varphi_X$ is holonomic. Consider the algebraic function $g(t)=-t$ for every real $t$. Then Theorem \ref{thm:Holonomic-Closure-Properties} implies that the function $\varphi_X \circ g$ is holonomic. Take $0 \neq \mathcal{L} \in \text{Ann}_{A_1(\mathbb{C})} (\varphi_X \circ g)$ and let $S = \Psi^{-1}(L)$. Then $0 \neq S \in A_1({\rm I\kern-0.16em R})$, and ${\rm I\kern-0.16em E}[S e^{-\mathrm{i} tX}]= \mathcal{L} (\varphi_X \circ g)(t) =0$ for every real $t$. In addition, for an arbitrary $f \in \mathcal{S}({\rm I\kern-0.16em R})$, the Schwartz class, by Fourier inversion formula, we can write
\begin{align*} f(x) = \frac{1}{2\pi} \int_{{\rm I\kern-0.16em R}} e^{-\mathrm{i} t x} \widehat{f}(t) \,dt.
\end{align*}
Therefore,
\begin{align*}
{\rm I\kern-0.16em E}[S f(X)] &= {\rm I\kern-0.16em E} \left[ S \biggl( \frac{1}{2\pi} \int_{{\rm I\kern-0.16em R}} e^{-\mathrm{i} t X} \widehat{f}(t)\, dt \biggr)\right] \\
&= \frac{1}{2\pi}{\rm I\kern-0.16em E} \left[ \int_{{\rm I\kern-0.16em R}} S (e^{-\mathrm{i} t X} ) \widehat{f}(t)\, dt \right] \\
& =\frac{1}{2\pi} \int_{{\rm I\kern-0.16em R}} \widehat{f}(t) {\rm I\kern-0.16em E} \big[ S (e^{-\mathrm{i} t X} ) \big] \,dt \quad (\text{using Fubini})\\
& = \frac{1}{2\pi} \int_{{\rm I\kern-0.16em R}} \widehat{f}(t) \mathcal{L} (\varphi_X \circ g)(t) \, dt\\
& =0.
\end{align*}
Note that in obtaining the second equality, we may interchange the integral and the operator $S$ because $\widehat{f}$ is also in $\mathcal{S}({\rm I\kern-0.16em R})$ (we emphasis that the operator $S$ acts on the variable $x$ and not on $t$). Hence, $S\in \text{PSO}(X)$.
\vspace{2mm}
\noindent{(b)} By part (a), the assumption $\text{PSO}(X) \neq 0$ implies that characteristic function $\varphi_X$ is holonomic and hence, by using the inversion formula (note that $\varphi_X$ is integrable) together with Theorem \ref{thm:Holonomic-Closure-Properties} part (c), the density function $p_X$ is holonomic. The other direction is just a direct application of part (c) in Theorem \ref{thm:Holonomic-Closure-Properties}.
\vspace{2mm}
\noindent{(c)} Clearly $(\text{PSO}(X), +)$ is a subgroup of $(A_1({\rm I\kern-0.16em R}),+)$. Now the claim follows at once due to the fact that the class $\mathcal{F}$ is invariant under elements of the first Weyl algebra $A_1({\rm I\kern-0.16em R})$. Finally, that $\text{PSO}(X)$ can be generated by two elements is Stafford's Theorem \ref{thm:Stafford}.
\vspace{2mm}
\noindent{(d)} It is easily seen that $\text{Ann}_{A_1(\mathbb{C})} (\varphi_X) \lhd_l A_1 (\mathbb{C})$ is a left ideal of the first Weyl algebra $A_1(\mathbb{C})$. Moreover, the anti-isomorphism $\Psi$ given in Lemma \ref{lem:Anti-Isomorphism} sends back and forth a generator to a generator. \hfill $\Box$
\begin{rem}\label{ex:(non)Holonomic_CF}{ \rm
(a) Many continuous probability distributions admit holonomic characteristic functions, including the Gaussian, exponential, gamma, beta,
variance-gamma,
Pearson distributions, and semicircular just to mention a few.
Therefore, according to Proposition \ref{prop:PSO_General_Targets} part (a), the associated polynomial Stein operator classes are non-trivial; indeed, polynomial Stein operators are known for all these distributions as well as many others; we refer the reader to \cite{a-g-g-algebraic-stein-operators,gms19,ley} and references therein for an overview of this literature.
\vspace{2mm}
\noindent{(b)} Let $X$ follow the extreme value Gumbel distribution with density function $p_X(x)= e^{-x} e^{-e^{-x}}$, $x \in {\rm I\kern-0.16em R}$. All moments of the Gumbel distribution exist. It is known that the characteristic function is given by $\varphi_X(t) =\Gamma(1 - \mathrm{i} t)$, where the gamma function is defined by $\Gamma(z)=\int_{0}^{\infty} x^{z-1} e^{-x}\,dx$, for $\text{Re}(z) >0$. The gamma function $\Gamma(z)$ is not a holonomic function, see, e.g., \cite[p.\ 4]{Holonomic-Toolkit}, and therefore the characteristic function $\varphi_X$ is not holonomic. Therefore, $\text{PSO}(X) =0$ is the trivial ideal. For Gumbel approximation via Stein's method the reader can consult \cite{ley, Gumber_Stein}.
\vspace{2mm}
\noindent{(c)} Let $X$ follow the log-normal distribution with density function given by $p_X(x)= (2\pi)^{-1/2}x^{-1}\exp(-(\log(x))^2/2)$, $x>0$. All moments of the log-normal distribution exist. It is readily seen that $p_X(x)$ is not a holonomic function \cite{Non-Holonomic-Character}, and therefore $\mathrm{PSO}(X)=0$ is the trivial ideal. A brief account of the Malliavin-Stein method for log-normal approximation is given by \cite{kusuotud}. We also remark that it is a well-known open problem to find a closed-form expression for the characteristic function of the log-normal distribution. Proposition \ref{prop:PSO_General_Targets} tells us that the characteristic function of the log-normal distribution is not holonomic and thus does not satisfy a homogeneous, linear ordinary differential equation (ODE) with rational coefficients. This partially explains the difficulty in finding a closed-form formula for the characteristic function, because many of the standard special functions are solutions of homoegenous, linear ODEs with rational coefficients.
}
\end{rem}
\begin{lem}\label{lem:delta-operate-on-basis}
Let $G= \partial-x \in A_1({\rm I\kern-0.16em R})$. Then, for every $k,l\in {\rm I\kern-0.16em N}_0$, it holds that
\begin{equation*}
G ( H_k \partial^l ) = H_k \partial^{l+1}-H_{k+1} \partial^l.
\end{equation*}
\end{lem}
\begin{proof}
By definition of the operator $G$, we can write
\begin{align*}
G ( H_k \partial^l ) & = \partial ( H_k \partial^l )-x H_k \partial^l= \partial (H_k) \partial^l + H_k \partial^{l+1}-x H_k \partial^l\\
& = H_k \partial^{l+1}-\left( xH_k - \partial H_k \right) \partial^l
=H_k \partial^{l+1}- \delta(H_k) \partial^l \\
&= H_k \partial^{l+1}-H_{k+1} \partial^l,
\end{align*}
where we used Definition \ref{def:Hermite_Polynomials} of Hermite polynomials.
\end{proof}
\begin{thm}\label{prop:PSO=Ideal}
Consider $\mathrm{PSO}(N)$, the class of all polynomial Stein operators associated to the standard Gaussian random variable $N$. Then the following statements are in order:
\begin{itemize}
\item[(a)] $\mathrm{PSO}(N)$ is a vector space over ${\rm I\kern-0.16em R}$ with basis
\begin{align}\label{eq:PSO-basis}
\mathbf{\mathfrak{B}}:= \Big\{ S(k,t):= H_k\partial^t- H_{k+t},\quad k,t\in {\rm I\kern-0.16em N}_0 \Big\}.
\end{align}
\item[(b)] $\mathrm{PSO}(N)$ is a principal maximal right ideal, and
\begin{equation}\label{eq:Generator}
\mathrm{PSO}(N) = \langle G \rangle _{A_1({\rm I\kern-0.16em R})}= G A_1({\rm I\kern-0.16em R}) : = \big\{ G L \, : \, L \in A_1({\rm I\kern-0.16em R}) \big \},
\end{equation}
where the generator $G=\partial-x$. (Equivalently, the ideal $\mathrm{PSO}(N)$, as a module on the first Weyl algebra $A_1({\rm I\kern-0.16em R})$, is a cyclic module.)
\item[(c)] The ideal $\mathrm{PSO}(N)$ (as a $A_1({\rm I\kern-0.16em R})$ module) is \textbf{not} holonomic, although the quotient $A_1({\rm I\kern-0.16em R}) / \mathrm{PSO}(N)$ is a holonomic module. In addition, $d ( A_1({\rm I\kern-0.16em R}) / \mathrm{PSO}(N))=1$ and $d(\mathrm{PSO}(N))=2$.
\end{itemize}
\end{thm}
\begin{rem}{ \rm
\begin{itemize}
\item[(a)] The first Weyl algebra $A_1({\rm I\kern-0.16em R})$ is not a principal ideal domain. In fact, the ideal
generated by elements $x^2$ and $1+x\partial$ is not principal, see \cite[Example 7.11.8]{mr-Book87}.
\item[(b)]Items (b)-(c) of Theorem \ref{prop:PSO=Ideal} reveal that $\text{PSO}(N)$ is a cyclic non-holonomic module on the first Weyl algebra $A_1({\rm I\kern-0.16em R})$. It is a well-known fact in $D$-module theory that every holonomic module is cyclic; see Theorem \ref{thm:Stafford} item (b).
\end{itemize}
}
\end{rem}
\begin{rem}\label{rem:PSO(N)-Generators}
{\rm Consider the right ideal $I=\langle G \rangle \lhd_r A_1({\rm I\kern-0.16em R})$, where $G= \partial-x$. Clearly, $I \subseteq \text{PSO}(N)$. The quotient ideal $A_1({\rm I\kern-0.16em R})/I$ is a holonomic module (see e.g. \cite[p.\ 86]{Coutinho-Book}) and by part (b) of Stafford's Theorem \ref{thm:Stafford} is cyclic meaning that there is $G' \in A_1({\rm I\kern-0.16em R})$ so that $A_1({\rm I\kern-0.16em R})/I = \langle G' + I \rangle$. Therefore, $\text{PSO}(N) = \langle G, G' \rangle_{A_1({\rm I\kern-0.16em R})}$. However, Theorem \ref{prop:PSO=Ideal} part (b) yields that the extra generator $G'$ is immaterial.
}
\end{rem}
\begin{proof}
(a) It is clear that $\text{PSO}(N)$ is a linear space. Now let $S=\sum_{t=0}^{T} p_t \partial^t \in A_1({\rm I\kern-0.16em R})$. Observe that by using the Gaussian integration by parts formula \eqref{lem:GIBP-d=1}, $S \in \text{PSO}(N)$ if and only if
\begin{equation*}
p_0=-\sum_{t=1}^{T} \delta^t p_t.
\end{equation*}
On the other hand, every polynomial can be written as a finite sum of Hermite polynomials; see Proposition \ref{prop:Hermite_Polynomials_Properties}. Hence, a basis of the linear space of Stein operators for the standard Gaussian distribution is given by the elements
\[ S(k,t)= H_k \partial^t - \delta^t H_k = H_k \partial^t - H_{k+t}, \quad t, k \in {\rm I\kern-0.16em N}_0. \]
(b) In order to show that $\text{PSO}(N)$ is a principal ideal, it is enough to show that for every element $S(k,t) \in \mathfrak{B}$ we have $S(k,t) = G L$ for some element $L \in A_1({\rm I\kern-0.16em R})$. By taking telescoping sums and applying Lemma \ref{lem:delta-operate-on-basis} in the second step,
\begin{align*}
S(k, t) &=- \sum_{r=1}^t
\biggl(H_{k+t+1-r} \partial^{r-1} -H_{k+t-r} \partial^r
\biggr)=G \biggl( \sum_{r=1}^ t H_{k+t+1-r} \partial^{r-1} \biggr).
\end{align*}
Maximality of $\text{PSO}(N)$ follows from \cite[Exercise 18, p.\ 50]{Lam-Book}. Maximality of principal ideals $xA_1({\rm I\kern-0.16em R})$ and $A_1({\rm I\kern-0.16em R})x$ is also discussed in \cite[p.\ 21]{mr-Book87}.
\vspace{1mm}
\noindent{(c)} According to \cite{Coutinho-Book}, Corollary 1.2 and Proposition 1.3, Chapter 10, torsion and holonomic finitely generated $A_1({\rm I\kern-0.16em R})$ modules are the same thing. Now, if $\text{PSO}(N)$ were a holonomic module then this would imply that the generator $G=\partial-x$ would be a torsion element, i.e., $G L =0$ for some non-zero element $L \in A_1({\rm I\kern-0.16em R})$, which is a contradiction with $A_1({\rm I\kern-0.16em R})$ being an integral domain. Lastly, the claim $d(\text{PSO}(N))=2$ follows directly from the Bernstein's inequality.
\end{proof}
\section{Characterising feature of \text{PSO}($N$)}\label{sec4}
\subsection{The annihilator ideal structure}
Denote by $\varphi_N (t)= e^{-t^2/2}$ the characteristic function of a standard Gaussian random variable $N$. Consider the left annihilator ideal $\text{Ann}_{A_1(\mathbb{C})} (\varphi_N) : = \big\{ \mathcal{S} \in A_1 (\mathbb{C}) \, : \, \mathcal{S}\varphi_N =0 \big \}$. In virtue of Theorem \ref{prop:PSO=Ideal} item (b),
$\text{Ann}_{A_1(\mathbb{C})}(\varphi_N)$ is a left principal ideal generated by the operator $\mathcal{G}= \Psi (G)= \Psi (\partial -x) = \partial +t$, where $\Psi$ is the Fourier map as introduced in Lemma \ref{lem:Anti-Isomorphism}, i.e.,
\begin{equation*}
\text{Ann}_{A_1(\mathbb{C})}(\varphi_N) = {}_{A_1(\mathbb{C})} \langle \mathcal{G} \rangle:= \{ \mathcal {A} \mathcal{G} \, : \, \mathcal{A} \in A_1 (\mathbb{C})\}.
\end{equation*}
In fact, following Proposition \ref{prop:PSO_General_Targets} part (d), it is generally true that, for a real-valued random variable $X$, $\text{PSO}(X)$ is a principal right ideal if and only if the annihilator $\text{Ann}_{A_1 (\mathbb{C})} (\varphi_X)$ is a left principal ideal in which the generators are linked via the Fourier map $\Psi$. On the other hand, one has to note that, providing some easy verifiable conditions that guarantee that either of ideals is principal is not a straightforward task. For example, with $X =H_2(N)= N^2-1$, one can show that $\text{PSO}(X)$ is a principal right ideal generated by the operator $2(x+1)\partial - x$ whilst the aforementioned question becomes remarkably non-trivial for any other generic element in the second Wiener chaos. This is under further investigation by the authors.
The annihilator ideal $\text{Ann}_{A_1(\mathbb{C})}(\varphi_N) $ plays a major role in the characterising Stein operators. Obviously, every Stein operator $L \in \text{PSO}(N)$ corresponds to a unique element $\mathcal{L} \in \text{Ann}_{A_1(\mathbb{C})}(\varphi_N) $ and vice versa via the Fourier map $\Psi$. Let us continue with the following simple observation.
\begin{prop}\label{prop:General_Characterization}
Let $ 0 \neq L \in \mathrm{PSO}(N)$ and $\mathcal{L} = \Psi (L)$ be the image of the operator $L$ under the Fourier map $\Psi$ (introduced in Lemma \ref{lem:Anti-Isomorphism}). Then $L$ is a characterising Stein operator for the standard Gaussian distribution in the sense of Definition \ref{def:Charactrization} if and only if
\begin{equation*}
\mathrm{Ker} (\mathcal{L}) \cap \widehat{\mathcal{P}}({\rm I\kern-0.16em R}) = \{ e^{-t^2/2}\},
\end{equation*}
where $\widehat{\mathcal{P}}({\rm I\kern-0.16em R})$ is the class of characteristic functions of probability distributions on the real line.
\end{prop}
\begin{proof}
It is immediate that $\varphi_N (t) = e^{-t^2/2} \in \text{Ker} (\mathcal{L})$ because $\mathcal{L} = \mathcal{A}(t+\partial)$ for some $\mathcal{A} \in A_1 (\mathbb{C})$ and $\mathcal{G} \varphi_N(t)= (t+\partial)e^{-t^2/2}=0$. In addition, $e^{-t^2/2} \neq \varphi \in \text{Ker} (\mathcal{L})$ if and only if ${\rm I\kern-0.16em E}[L f (X)]={\rm I\kern-0.16em E}[L f (N)]=0$ for all $f \in \mathcal{F}$, where $X$ is a random variable whose characteristic function is $\varphi$. Hence, the claim follows.
\end{proof}
We continue with some illustrating examples.
\begin{ex}\label{ex:Moment_Assumption_Crucial}\label{extocor}{\rm
(a) Let $G=x-\partial \in \mathrm{PSO}(N)$. Then, obviously, $G$ is a characterising Stein operator for the standard Gaussian distribution. This goes all the way back to Stein's seminal paper \cite{stein}. In order to apply Proposition \ref{prop:General_Characterization}, note that $\Psi (G) = \mathcal{G} = t + \partial$ and, clearly $\text{Ker} (\mathcal{G}) \cap \widehat{\mathcal{P}}({\rm I\kern-0.16em R}) = \{ e^{-t^2/2}\}$.
\vspace{2mm}
\noindent{(b)} Let $ x_0 \in {\rm I\kern-0.16em R}\setminus\{0\}$, and consider the operator $\mathcal{L} = \mathcal{A} (t+\partial) \in \mathrm{Ann}_{A_1(\mathbb{C})}(\varphi_N)$, where $\mathcal{A} = (t - \mathrm{i} x_0) +\partial \in A_1 (\mathbb{C})$. One can easily check that both characteristic functions
\begin{equation*}
e^{-t^2/2}, \, e^{-t^2/2} e^{\mathrm{i} t x_0} \in \text{Ker} (\mathcal{L}) \cap \widehat{\mathcal{P}}({\rm I\kern-0.16em R}).
\end{equation*}
These characteristic functions correspond to the $N(0,1)$ and $N(x_0,1)$ distributions, respectively. Hence, Proposition \ref{prop:General_Characterization} entails that the operator $L = \Psi^{-1}(\mathcal{L})$ is not a characterising Stein operator for the standard Gaussian distribution without assuming a further condition on the first derivative (evaluated at $t=0$) of characteristic functions $\varphi \in \widehat{\mathcal{P}}({\rm I\kern-0.16em R}) $ (note that $-\mathrm{i} \, \frac{\partial}{\partial t} ( e^{-t^2/2} e^{\mathrm{i} t x_0} )\big \vert_{t=0}= x_0 \neq 0$). For example, with the additional assumption ${\rm I\kern-0.16em E}[X]=0$ (that is $\varphi'(0)=0$), the operator $L$ is a characterising polynomial Stein operator for the $N(0,1)$ distribution.
\vspace{2mm}
\noindent{(c)} This time, let $ {\mathcal A}=( t \partial + 2 t^2 -1 ) $ and $ {\mathcal L}={\mathcal A}(t+\partial)
=(t \partial^2 +3 t^2 \partial - \partial + 2 t^3 )$.
Then, one can easily check that the characteristic functions
\begin{equation*}
e^{-t^2/2}, \, e^{-t^2} \in \text{Ker} (\mathcal{L}) \cap \widehat{\mathcal{P}}({\rm I\kern-0.16em R})
\end{equation*}
corresponding to the Gaussian distributions $N(0,1)$ and $N(0,2)$, respectively. Hence, operator $L=\Psi^{-1}(\mathcal{L})$ is not a characterising Stein operator for the standard Gaussian distribution without assuming a further condition on the second derivative (evaluated at $t=0$) of characteristic functions $\varphi \in \widehat{\mathcal{P}}({\rm I\kern-0.16em R}) $ (note that $-\frac{\partial^2}{\partial t^2} ( e^{-t^2})\big \vert_{t=0}=2$). For example, with the additional assumption ${\rm I\kern-0.16em E}[X^2]=1$ (that is $\varphi''(0)=-1$), the operator $L$ is a characterising polynomial Stein operator for the $N(0,1)$ distribution.
\vspace{2mm}
\noindent{(d)} More generally, there exist $T$-th order ($T\geq2$) differential operators in the class $\mathrm{PSO}(N)$ for which $T-1$ moment conditions must be specified for the operator $S$ to be characterising. Consider a mixture of Gaussian distributions $Y$ with density $p_Y(x)=\sum_{j=1}^T w_jp_j(x)$, where $w_1,\ldots,w_T>0$, $\sum_{j=1}^Tw_j=1$ and, for $j=1,\ldots,T$, the function $p_j$ is the density of a $N(0,\sigma_j^2)$ random variable, with $\sigma_1^2,\ldots,\sigma_T^2>0$ distinct and also $\sigma_1^2=1.$ Since $p_1,\ldots,p_T$ are linearly independent holonomic functions that satisfy first order homogeneous differential equations with polynomial coefficients, it follows that $p_Y(x)=\sum_{j=1}^T w_jp_j(x)$ is a holonomic function and satisfies a $T$-th order homogeneous differential equation of the form $Lp=0$, where $Lp(x)=\sum_{j=0}^m\sum_{k=0}^T a_{jk}x^jp^{(k)}(x)$, with real-valued coefficients $a_{jk}$ that do not depend on $w_1,\ldots,w_T$ (they do not depend on $w_1,\ldots,w_T$ because any linear combination of $p_1,\ldots,p_T$ will satisfy $Lp=0$). By a standard integration by parts argument (see, for example, the proof of Lemma 4.1 of \cite{gms19}) it follows that a Stein operator for $Y$ is given by $Sf(x)=\sum_{j=0}^m\sum_{k=0}^T (-1)^ka_{jk}\partial^k(x^jf(x))$, where $f\in P({\rm I\kern-0.16em R})$. Since the operator $S$ has no dependence on $w_1,\ldots,w_T$, for $S$ to be a characterising Stein operator for $X\sim N(0,1)$, additional conditions must be specified to ensure that $w_1=1$ and $w_2=\cdots=w_T=0$. For a characterisation based on moment conditions, $T-1$ moments of $X$ would need to be specified; we do not need to specify $T$ moments because $w_1=1-\sum_{j=2}^Tw_j$.
}
\end{ex}
The next proposition studies the structure of the annihilator ideal $\text{Ann}_{A_1(\mathbb{C})}(\varphi_N) $. It turns out that the null space of the annihilator ideal $\text{Ann}_{A_1(\mathbb{C})}(\varphi_N) $ is extremely large and in fact it includes any holonomic characteristic function.
\begin{prop}\label{prop:Ideal-Structure-In-Fourier-Side}
Let $$\mathrm{Ker} \big( \mathrm{Ann}_{A_1(\mathbb{C})}(\varphi_N) \big) : = \{ \mathrm{Ker} (\mathcal{L}) \, :\, \mathcal{L} \in \mathrm{Ann}_{A_1(\mathbb{C})}(\varphi_N) \} = \bigcup_{\mathcal{A} \in A_1 (\mathbb{C})} \mathrm{Ker} \left( \mathcal{A} (t+\partial) \right).$$ Then
\begin{equation*}
\mathrm{Ker} \big( \mathrm{Ann}_{A_1(\mathbb{C})}(\varphi_N) \big) \cap \widehat{\mathcal{P}}({\rm I\kern-0.16em R}) = \widehat{\mathcal{P}}({\rm I\kern-0.16em R}) ^{\mathrm{Hol}},
\end{equation*}
where $\widehat{\mathcal{P}}({\rm I\kern-0.16em R}) ^{\mathrm{Hol}}$ stands for the class of all holonomic characteristic functions (\textbf{possibly discrete distributions, e.g. the Bernoulli distribution}).
\end{prop}
\begin{proof}
It is clear that $ \text{Ker} ( \text{Ann}_{A_1(\mathbb{C})}(\varphi_N) ) \cap \widehat{\mathcal{P}}({\rm I\kern-0.16em R}) \subseteq \widehat{\mathcal{P}}({\rm I\kern-0.16em R}) ^{\mathrm{Hol}}$. For the other inclusion, let $\varphi \in \widehat{\mathcal{P}}({\rm I\kern-0.16em R}) ^{\mathrm{Hol}}$. Then, there is $\mathcal{L} \in A_1 (\mathbb{C})$ such that $\mathcal{L} \varphi=0$. On the other hand, by the left Ore property, one can write $ \mathcal{S} \mathcal{L} = \mathcal{A}(t+\partial)$ for some elements $\mathcal{S}, \mathcal{A} \in A_1 (\mathbb{C})$. Hence, $\mathcal{A}(t+\partial) \in \text{Ann}_{A_1(\mathbb{C})} (\varphi_N)$ and $ \mathcal{A}(t+\partial) \varphi =0$. Therefore, $\varphi \in \text{Ker} ( \mathcal{A}(t+\partial) )$ and we are done.
\end{proof}
\begin{cor}\label{cor:Shape-CFs-InTheIdeal}
If $\varphi \in \widehat{\mathcal{P}}({\rm I\kern-0.16em R}) ^{\mathrm{Hol}}$, then
\begin{equation}\label{eq:structural-form-Ideal-Fourier-side}
\varphi (t) = e^{-t^2/2} \left( 1+ \int_{0}^{t} e^{s^2/2} \psi (s)\, ds\right)
\end{equation}
for some holonomic function $\psi$. Conversely, if the function $\varphi$ (not necessarily being a characteristic function) takes the form \eqref{eq:structural-form-Ideal-Fourier-side} for some holonomic function $\psi$, then $\varphi \in \mathrm{Ker} (\mathrm{Ann}_{A_1 (\mathbb{C})} (\varphi_N))$.
\end{cor}
\begin{proof}
Using Proposition \ref{prop:Ideal-Structure-In-Fourier-Side} one can infer that $ \varphi \in \text{Ker} (\mathcal{A} (t+\partial))$ for some $\mathcal{A} \in A_1 (\mathbb{C})$. Hence, $ (t+\partial) \varphi = \psi (t)$ for some function $\psi \in \text{Ker} (\mathcal{A})$. This means that $\psi$ must be a holonomic function, and in addition for some constant $c$ we have
\begin{equation*}
\varphi (t) = e^{-t^2/ 2} \biggl( c + \int_0^t e^{s^2/2} \psi(s) \,ds \biggr).
\end{equation*}
Since $\varphi$ is a characteristic function, we have $\varphi(0)=c=1$. Conversely, assume that $\varphi$ takes the form \eqref{eq:structural-form-Ideal-Fourier-side} for some function $\psi$ so that $\mathcal{A}\psi = 0$ for some $\mathcal{A} \in A_1 (\mathbb{C})$. Then, $\mathcal{G}\varphi (t) = \psi (t)$. Therefore, $\mathcal{A}\mathcal{G}\varphi = \mathcal{A}\psi =0$.
\end{proof}
\begin{rem}\rm{
(a) In equation \eqref{eq:structural-form-Ideal-Fourier-side}, the second factor $1+ \int_{0}^{t} e^{s^2/2} \psi (s)\, ds$ is not necessarily the characteristic function of a probability distribution. Otherwise, $\varphi$ would be a characteristic function of a random variable $X =_d N + Y$ (equality in distribution) a convolution with the standard Gaussian distribution random variable that contradicts with Corollary \ref{cor:PSO_Intersection_NonTrivial}. As is mentioned below in part (a) of Remark \ref{rem4.2}, for example, one can take $X$ to be a semi-circular distribution. Then, $\text{PSO}(N) \cap \text{PSO}(X) \neq 0$ and $X$ cannot be written in the form of convolution with the standard Gaussian distribution because it has a bounded support.
\vspace{2mm}
\noindent{(b)} In equation \eqref{eq:structural-form-Ideal-Fourier-side}, the second factor $1+ \int_{0}^{t} e^{s^2/2} \psi (s)\, ds$ being a characteristic function of a probability measure is equivalent with whether the ratio of two characteristic functions is necessarily a characteristic function. However, it is a classical and known problem in probability theory that the ratio of two characteristic functions is not in general a characteristic function; see \cite{lukas} for further details. }
\end{rem}
\begin{thm}\label{thm:PSO_Intersection_NonTrivial}
Fix $n \ge 1$. Let $X_1, \ldots, X_n$ be continuous random variables whose characteristic functions are holonomic. Then
\begin{equation}\label{eq:POS-Intersection-General}
\mathrm{PSO}(X_1) \cap \cdots \cap \mathrm{PSO}(X_n) \neq 0.
\end{equation}
\end{thm}
\begin{proof}
First note that the assumption that the characteristic functions of $X_1,\ldots,X_n$ are holonomic implies that $\text{PSO}(X_k) \neq 0$ for each $k=1,\ldots,n$, by Proposition \ref{prop:PSO_General_Targets}, item (a). On the other hand, by Stafford's Theorem \ref{thm:Stafford} each right ideal $\text{PSO}(X_k) = \langle G^k_1, G^k_2 \rangle$ for two generators $G^k_1$ and $G^k_2$. Since Weyl algebras are right Noetherian integral domains they are right Ore domains \cite[Theorem 1.15]{mr-Book87}. Therefore, letting $R = A_1({\rm I\kern-0.16em R})$, we have
\begin{equation*}
\bigcap_{k=1}^{n} G^k_1 R \cap G^k_2 R \neq 0.
\end{equation*}
On the other hand, $ G^k_1 R \cap G^k_2 R \subseteq \text{PSO}(X_k)$ for each $k=1,\ldots,n$, and whence the claim follows.
\end{proof}
\begin{cor}\label{cor:PSO_Intersection_NonTrivial}
Let $N \sim N(0,1)$ and $X$ be a continuous random variable whose characteristic function is holonomic. Then
\begin{equation*}
\mathrm{PSO}(N) \cap \mathrm{PSO}(X) \neq 0.
\end{equation*}
\end{cor}
\begin{rem}\label{rem4.2}{ \rm (a) In the statement of Theorem \ref{thm:PSO_Intersection_NonTrivial}, the supports of the random variables $X_1,\ldots,X_n$ do not need to be the same. In fact, one can even consider random variables having disjoint supports. Let $X$ and $Y$ be two continuous random variables whose characteristic functions are holonomic with disjoint supports $\mathbb{D}_X$ and $\mathbb{D}_Y$, respectively. Let $\mathbb{U}_X$ and $\mathbb{U}_Y$ be two disjoint open domains so that $\mathbb{U}_X \supseteq \mathbb{D}_X$, $ \mathbb{U}_Y \supseteq \mathbb{D}_Y$. It is evident that ${\rm I\kern-0.16em E}[S_X f(Y)]=0$ for all $f\in \mathcal{S}(\mathbb{U}_X)$, where $S_X \in \text{PSO}(X)$. In addition, clearly $ \mathcal{S}(\mathbb{U}_X) \subseteq \mathcal{S}({\rm I\kern-0.16em R})$; however, the requirement ${\rm I\kern-0.16em E}[S_X f(Y)]=0$ for all $f \in \mathcal{S}({\rm I\kern-0.16em R})$ is a non-trivial task. According to Theorem \ref{thm:PSO_Intersection_NonTrivial} there is at least one non-trivial polynomial differential operator $ S \neq 0$ such that
\begin{equation*}
{\rm I\kern-0.16em E}[S f(X_k)]=0, \quad \forall \, f \in \mathcal{S}({\rm I\kern-0.16em R}) \supseteq \mathcal{S} (\mathbb{U}), \quad k=1,\ldots,n,
\end{equation*}
where each $\mathbb{D}_{X_k}$ is the support of random variable $X_k$, and the open domain $\mathbb{U} \supseteq \mathbb{D}_{X_1} \cup \cdots \cup \mathbb{D}_{X_n}$. An example of an operator $0 \neq L \in \text{PSO} (N) \cap \text{PSO}(X)$ is given in \cite[Section 3.1]{a-g-g-Stein-charactrization} when $X$ is distributed as a semi-circular distribution with probability density function $p_X(x)= (2/\pi) \sqrt{1-x^2} \textbf{1}_{[-1,1]}$ supported on the interval $[-1,1]$.
\vspace{2mm}
\noindent{(b)
Theorem \ref{thm:PSO_Intersection_NonTrivial} is also related to characterising Stein operators; see Definition \ref{def:Charactrization}. Firstly, in order that a polynomial Stein operator would be useful from a probabilistic point of view, the class $\mathcal{F}$ of test functions must contain functions being non-zero on the support of the underlying target random variable. Secondly, for random variables $X$ and $Y$ as in part (a) of this remark, Theorem \ref{thm:PSO_Intersection_NonTrivial} yields the existence of a polynomial Stein operator $S\not=0$ so that ${\rm I\kern-0.16em E}[Sf(X)]={\rm I\kern-0.16em E}[Sf(Y)]=0$ for all $f \in \mathcal{S}({\rm I\kern-0.16em R})$. Therefore, the Stein operator $S$ cannot be a characterising operator for neither $X$ nor $Y$ without any extra assumption; see also Example \ref{ex:Moment_Assumption_Crucial}.
}
\end{rem}
\subsection{First order characterising Stein operators }
For a suitable (typically belonging to $L^1({\rm I\kern-0.16em R})$) function $\varphi$ (playing the role of a characteristic function in the sequel), the Fourier transform $\mathbf{\mathcal{F}} (\varphi)$ is defined by
\begin{equation}\label{eq:Fourier_Transform}
\mathbf{\mathcal{F}} (\varphi) (\xi) := \int_{{\rm I\kern-0.16em R}} \varphi(t) e^{- \mathrm{i} t \xi} \, dt, \quad \forall \, \xi \in {\rm I\kern-0.16em R}.
\end{equation}
Let $P$ denote a polynomial. With the notation $P(\partial)$ we mean the differential operator in which the variable $x$ is replaced with derivative operator $\partial$. The following properties are immediate.
\begin{lem}\label{lem:FourierTransform_Basic_Properties}
Let $P$ be a polynomial. Then, for a suitable function $\varphi$, we have
\begin{itemize}
\item[(a)]
\begin{equation*}
P(\partial) \mathbf{\mathcal{F}}(\varphi)(\xi) = \mathbf{\mathcal{F}} \left( P(-\mathrm{i}t) \varphi \right)(\xi).
\end{equation*}
\item[(b)]
\begin{equation*}
\mathbf{\mathcal{F}} \left( P(\partial) \varphi \right) (\xi) = P (\mathrm{i} \xi) \mathbf{\mathcal{F}}(\varphi)(\xi).
\end{equation*}
\item[(c)] Combining the two previous cases, for any differential operator $P=\sum_k p_k \partial^k$ with polynomial coefficients, we have
\begin{equation*}
\mathbf{\mathcal{F}} \left( P (\cdot,\partial) \varphi\right) (\xi) = P (\mathrm{i} \partial, \mathrm{i} \xi) \mathbf{\mathcal{F}} (\varphi)(\xi).
\end{equation*}
\end{itemize}
\end{lem}
\begin{lem}\label{lem:G_commute_FourierTransform}
Consider the operator $\mathcal{G}=\partial + t $. Then $$\mathbf{\mathcal{F}} \mathcal{G} = \mathrm{i} \, \mathcal{G} \mathbf{\mathcal{F}}$$ (in other words, the operator $\mathcal{G}$ and the Fourier transform $\mathcal{F}$ are commutating up to the complex imaginary number $\mathrm{i}$).
\end{lem}
\begin{proof}
This a direct application of the Lemma \ref{lem:FourierTransform_Basic_Properties} and the linearity of Fourier transform.
\end{proof}
\begin{prop}\label{prop:L_nOperators}
Let $n \in {\rm I\kern-0.16em N}_0$. Consider the linear operator $\mathcal{L}_n: = \partial^n \mathcal{G} = \partial^n (t+\partial) \in \mathrm{Ann}_{A_1 (\mathbb{C})} (\varphi_N)$ (by the convention $\mathcal{L}_0=\mathcal{G}$). Then
\begin{equation*}
\mathrm{Ker}(\mathcal{L}_n ) \cap \widehat{\mathcal{P}}({\rm I\kern-0.16em R}) = \{ e^{-t^2/2} \}.
\end{equation*}
Consequently, for every $n\ge 0$, the operator $L_n := x^n \partial + (nx^{n-1} - x^{n+1}) \in \mathrm{PSO} (N)$ is a characterising Stein operator for the standard Gaussian distribution among those random variables whose characteristic functions belongs to the domain $\mathbb{D}(\mathcal{L}_n)$ of the operator $\mathcal {L}_n$ (one has to note that the case $n=0$ boils down to that of the classical Stein operator for the standard Gaussian distribution).
\end{prop}
\begin{proof}
Let $\varphi \in \text{Ker}(\mathcal{L}_n)$, and using Lemma \ref{lem:FourierTransform_Basic_Properties}, one can infer that
\begin{align*}
\left(\mathbf{\mathcal{F}} \mathcal{L}_n \varphi \right) (\xi) = \mathbf{\mathcal{F}} \left( \partial^n \mathcal{G} \varphi \right) (\xi)= (\mathrm{i} \xi)^n \left( \mathbf{\mathcal{F}} \mathcal{G} \varphi \right)(\xi)=0, \quad \forall \, \xi \in {\rm I\kern-0.16em R}.
\end{align*}
This yields that $\mathbf{\mathcal{F}} (\mathcal{G}\varphi)=0$. Now, the injectivity of the Fourier transform implies that $\varphi \in \text{Ker} (\mathcal{G})$. Finally, $\varphi$ being a characteristic function yields the result.
\end{proof}
\begin{prop}\label{prop:L_nOperators_Combiniation}
Let $\mathcal{L}=\sum_{n} c_n \mathcal{L}_n \in \mathrm{Ann}_{A_1 (\mathbb{C})}(\varphi_N)$ be an arbitrary finite linear combination of the operators $\mathcal{L}_n$ introduced in Proposition \ref{prop:L_nOperators}, with $c_n \in \mathbb{C}$ being non-zero simultaneously. Then
\begin{equation*}
\mathrm{Ker}(\mathcal{L}) \cap \widehat{\mathcal{P}}({\rm I\kern-0.16em R})= \{ e^{-t^2/2} \}.
\end{equation*}
Consequently, the operator $L : = \Psi^{-1}(\mathcal{L})\in \mathrm{PSO}(N)$ (possibly after a suitable adjustment to get real coefficients) is a characterising Stein operator for the standard Gaussian distribution among those random variables whose characteristic functions belong to domain $\mathbb{D}(\mathcal{L})$ of the operator $\mathcal{L}$ .
\end{prop}
\begin{proof}
Let $\varphi \in \text{Ker} (\mathcal{L})$ and using proof of Proposition \ref{prop:L_nOperators} we can write
\begin{align*}
\left(\mathbf{\mathcal{F}} \mathcal{L} \varphi \right)(\xi) =\sum_{n} c_n (\mathrm{i} \xi)^n \mathbf{\mathcal{F}} \mathcal{G} \varphi (\xi) = \mathbf{\mathcal{F}} \mathcal{G} \varphi (\xi) P (\mathrm{i} \xi)=0, \quad \forall \, \xi \in {\rm I\kern-0.16em R},
\end{align*}
where the polynomial $P(z) = \sum_{n} c_n z^n$, $z \in \mathbb{C}$. Then, this implies that $ \mathbf{\mathcal{F}} \mathcal{G} \varphi (\xi) P (\mathrm{i} \xi)=0$ for every $\xi \in {\rm I\kern-0.16em R}$. Since any polynomial can have finitely many isolated roots, and the Fourier transform is uniformly continuous, it follows that $ \mathbf{\mathcal{F}} \mathcal{G} \varphi (\xi)=0$ for all $\xi \in {\rm I\kern-0.16em R}$. Now injectivity of the Fourier transform yields that $\varphi \in \text{Ker} (\mathcal{G})$. Hence, the claim follows.
\end{proof}
\begin{cor}\label{cor:Hermite-Coeff-Is-Characterizing}
Let $p \in {\rm I\kern-0.16em R}[x]$ be a non-zero polynomial with real coefficients and $\delta$ stands for the adjoint operator. Then, the operator $$L = p(x) \partial - \delta p(x) \in \mathrm{PSO} (N)$$ is a characterising polynomial Stein operator for the standard Gaussian distribution among those random variables whose characteristic functions belong to domain $\mathbb{D}(\mathcal{L})$ of the operator $\mathcal{L} = \Psi (L)$. In particular, let $m\ge1$ and let $H_m$ denote the Hermite polynomial of degree $m$. Then, the operator $S_m : = H_{m-1} \partial - H_{m} \in \mathrm{PSO} (N)$ is a characterising polynomial Stein operator for the standard Gaussian distribution among all those random variables whose characteristic functions belong to domain $\mathbb{D}(\mathcal{S}_m)$ of the operator ${\mathcal{S}}_m = \Psi (S_m)$.
\end{cor}
\begin{proof}
The fact that the operator $L$ is a polynomial Stein operator for the standard Gaussian distribution follows directly from the Gaussian integration by parts formula in Proposition \ref{prop:GIBP-d=1}. In order to show that $L$ is in fact a characterising Stein operator note that it holds that $p (x) = \sum_{n \le \text{deg}(p)} a_n x^{n}$, for some real coefficients $a_n$, and the fact that the adjoint operator $\delta$ is linear. Hence, the operator $\mathcal{L} = \Psi(L)$ can be written as a finite linear sum of operators $\mathcal{L}_n$'s. Now, the claims follows at once from Proposition \ref{prop:L_nOperators_Combiniation}.
\end{proof}
\begin{rem}\label{rem11}\rm{We can use an alternative argument to prove that the operator $S_m : = H_{m-1} \partial - H_{m} \in \mathrm{PSO} (N)$, $m\geq 2$, is characterising without the need for the assumption that the characteristic function of the random variable $X$ belongs to the domain $\mathbb{D}(\mathcal{S}_m)$ of the operator $\mathcal{S}_m = \Psi (S_m)$.
We first note the following simple connection between the Stein operator $H_{m-1}\partial - H_m$ and the classical Stein operator $\partial - x$. Consider the Stein operator $f'(x) - xf(x)$. Let $f(x)=H_{m-1}(x)g(x)$. Then using basic properties of Hermite polynomials yields that $f'(x) - xf(x) = H_{m-1}(x) g'(x) - H_m(x) g(x)$. Now, let $X$ be a real-valued random variable with density $p_X(x)$ and suppose that ${\rm I\kern-0.16em E}[H_{m-1}(X) g'(X) - H_m(X) g(X)]=0$ for all $g\in \mathcal{G}$ for some suitable class of functions $\mathcal{G}$. Then, by an integration by parts argument, we have that $\int_{\rm I\kern-0.16em R} (p_X'(x)+xp_X(x)) H_{m-1}(x)g(x) \,dx=0$ for all $g\in \mathcal{G}$. We thus deduce that $p_X(x)$ satisfies the ODE $p_X'(x)+xp_X(x)=0$. On solving this ODE subject to the condition that $p_X(x)$ is the probability density function of a real-valued random variable (so that $\int_{\rm I\kern-0.16em R} p_X(x)\,dx=1)$ we deduce that $p_X(x)=(2\pi)^{-1/2}e^{-x^2/2}$, and therefore $X\sim N(0,1)$.}
\end{rem}
\subsection{Higher order characterising Stein operators}
In this section, we consider the characterising property of higher order Stein operators. In the light of part (d) of Example \ref{ex:Moment_Assumption_Crucial}, and more generally Theorem \ref{thm:PSO_Intersection_NonTrivial}, we know that there exist many higher order Stein operators that are not characterising. The purpose of this section is to attempt to give some examples of general classes of higher order polynomial Stein operators for the Gaussian distribution that are characterising. The most natural operator to start with is perhaps the higher order counterpart of the Stein operator $S_m$ given in Corollary \ref{cor:Hermite-Coeff-Is-Characterizing}, namely the Stein operator $L_m=\partial^m - H_m (x)$, $m\geq1$. As noted in the introduction, the fact that $L_m\in\mathrm{PSO}(N)$, $m\geq1$, was known to \cite{gr05}. The proof of the following proposition is technical and is given in Appendix \ref{appa}.
\begin{prop}\label{prop4.5}
Let $ m\ge 1$ be fixed. Then, the operator $L_m = \partial^m - H_m \in \mathrm{PSO}(N) $
is a characterising polynomial Stein operator for the standard Gaussian distribution among those random variables that are symmetric (that is $X=_d-X$) and have an infinitely divisible distribution. When $m=1,2,3$, the characterisation holds without the need for the symmetric and infinitely divisible distribution assumptions.
\end{prop}
\begin{rem}\rm{In Proposition \ref{prop4.5}, our characterisation of the standard Gaussian distribution has additional symmetry and infinite divisibility conditions when $m\geq4$.
We conjecture that these additional conditions are not necessary to ensure characterisation, but are rather artefacts resulting from our proof. We leave it as an interesting open problem to verify or disprove this conjecture. Similar comments apply for Corollary \ref{barops} and Proposition \ref{prop4.6} below.}
\end{rem}
We can generalise Proposition \ref{prop4.5} to deduce a characterisation result for the Stein operators $S(k,t)=H_{k}\partial^t-H_{k+t}\in\text{PSO}(N)$, $k,t\geq1$. Recall that the fact that $S(k,t)\in\mathrm{PSO}(N)$, $k,t\geq1$, was known to \cite{b86}.
\begin{cor}\label{barops}
Let $k, t\geq1$. Then, the operator $H_k \partial^t - H_{k+t} \in \mathrm{PSO}(N)$ is a characterising Stein operator for the standard Gaussian distributions among those random variables that are symmetric and have an infinitely divisible distribution. When $t=1,2,3$, the characterisation holds without the need for the symmetric and infinitely divisible distribution assumptions.
\end{cor}
\begin{proof}
We deduce the result from Proposition \ref{prop4.5} by applying the technique used in Remark \ref{rem11}. Consider the Stein operator $\partial^t f(x) - H_t(x)f(x)$. Now, let $X$ be a real-valued random variable with density $p_X(x)$ and suppose that ${\rm I\kern-0.16em E}[ f^{(t)}(X) - H_{t}(X) f(X)]=0$ for all $f\in \mathcal{F}$ for some suitable class of functions $\mathcal{F}$. Then, by an integration by parts argument, we have that $\int_{\rm I\kern-0.16em R} f(x) Tp_X(x) \,dx=0$ for all $f\in \mathcal{F}$, for some linear differential operator $T$ (the particular form that the operator $T$ takes is not important to our argument). We thus deduce that $p_X(x)$ satisfies the ODE $Tp_X(x)=0$. By proposition \ref{prop4.5}, we can infer that the only solution to this ODE that satisfies the properties of a probability density function is $p_X(x)=(2\pi)^{-1/2}e^{-x^2/2}$ if $t=1,2,3$, and again $p_X(x)=(2\pi)^{-1/2}e^{-x^2/2}$ if we impose the additional assumptions that $p_X(x)$ must be the density of a symmetric and infinitely divisible random variable when $t\geq4$.
Now, let $f(x)=H_k(x)g(x)$. Then we get that $\partial^t f(x) - H_t(x)f(x) = H_k(x)\partial^t g(x) - H_{k+t}(x)g(x)$.
Let $Y$ be a real-valued random variable with density $p_Y(x)$ and suppose that ${\rm I\kern-0.16em E}[H_{k}(Y) g^{(t)}(Y) - H_{k+t}(Y) g(Y)]=0$ for all $g\in \mathcal{G}$ for some suitable class of functions $\mathcal{G}$. Then, by the same integration by parts argument as before, we have that $\int_{\rm I\kern-0.16em R} H_{k}(x)g(x)Tp_Y(x) \,dx=0$ for all $g\in \mathcal{G}$, where $T$ is the same operator as introduced earlier in this proof. We thus deduce that $p_Y(x)$ satisfies the ODE $Tp_Y(x)=0$, the same ODE as satisfied by $p_X(x)$. Thus, under the relevant assumptions for the cases $t=1,2,3$ and $t\geq4$, the only solution that satisfies the properties of a probability density function is $p_Y(x)=(2\pi)^{-1/2}e^{-x^2/2}$, and therefore $Y\sim N(0,1)$.
\end{proof}
Another interesting class of higher order Stein operators for the standard Gaussian distribution can be derived via the so-called Rodriguez formula for Hermite polynomials. In fact, the Rodriguez formula implies that the characteristic function $\varphi_N (t) = e^{-t^2/2}$ satisfies the ODE
\begin{equation}\label{eq:Rodriguez-formula-Hermite}
\partial^m \varphi_N ( t) = (-1)^m H_m (t) \varphi_N(t), \quad \forall \, m \ge 1.
\end{equation}
Let $\mathcal{R}_m = \partial^m - (-1)^m H_m \in \mathrm{Ann}_{A_1 (\mathbb{C})} (\varphi_N)$ and hence $R_m = \Psi^{-1}(\mathcal{R}_m)=(-1)^mH_m(-\mathrm{i}\partial)-\mathrm{i}^mx^m \in \text{PSO}(N)$ (possibly after a multiplication by a suitable power of $\mathrm{i}$ to get real coefficients). The polynomial Stein operators $R_m$ are particularly interesting
whenever the sole $m$-th moment matching ${\rm I\kern-0.16em E}[X^m] = {\rm I\kern-0.16em E}[N^m]$ is applicable. The proof of the following proposition is given in Appendix \ref{appa}.
\begin{prop}\label{prop4.6}Let $ m\ge 1$ be fixed.
Then, the operator $R_m=(-1)^mH_m(-\mathrm{i}\partial)-\mathrm{i}^mx^m \in \mathrm{PSO}(N) $ is a characterising polynomial Stein operator for the standard Gaussian distribution among those random variables that are symmetric and have an infinitely divisible distribution. When $m=1,2,3$, the characterisation holds without the need for the symmetric and infinitely divisible distribution assumptions.
\end{prop}
|
\section{{\rm REFERENCES}}
\sloppy \hyphenpenalty10000
\begin{list}{}{\leftmargin1cm\listparindent-1cm
\itemindent\listparindent\parsep0pt\itemsep0pt}
{\end{list}\vspace{2mm}}
\def~{~}
\newlength{\DW}
\settowidth{\DW}{0}
\newcommand{\hspace{\DW}}{\hspace{\DW}}
\newcommand{\refitem}[5]{\item[]{#1} #
\def#5}\ifx\REFARG\TYLDA\else, {#5}\fi.{#3}\ifx#5}\ifx\REFARG\TYLDA\else, {#5}\fi.~\else, {\it#3}\fi
\def#5}\ifx\REFARG\TYLDA\else, {#5}\fi.{#4}\ifx#5}\ifx\REFARG\TYLDA\else, {#5}\fi.~\else, {\bf#4}\fi
\def#5}\ifx\REFARG\TYLDA\else, {#5}\fi.{#5}\ifx#5}\ifx\REFARG\TYLDA\else, {#5}\fi.~\else, {#5}\fi.}
\newcommand{\Section}[1]{\section{#1}}
\newcommand{\Subsection}[1]{\subsection{#1}}
\newcommand{\Acknow}[1]{\par\vspace{5mm}{\bf Acknowledgments.} #1}
\pagestyle{myheadings}
\newfont{\bb}{ptmbi8t at 12pt}
\newcommand{\rule{0pt}{2.5ex}}{\rule{0pt}{2.5ex}}
\newcommand{\rule[-1.8ex]{0pt}{4.5ex}}{\rule[-1.8ex]{0pt}{4.5ex}}
\def\fnsymbol{footnote}{\fnsymbol{footnote}}
\begin{center}
{\Large\bf New Short-Period Delta Scuti Stars in OGLE-IV Fields toward
the Galactic Bulge\footnote{Based on photometric observations obtained with
the 1.3-m Warsaw telescope at Las Campanas Observatory of the Carnegie
Institution for Science and spectroscopic data collected with the European Southern
Observatory (ESO) NTT telescope at La Silla Observatory under ESO programme 105.20EF.001.}}
\vskip1cm
{\bf
P.~~P~i~e~t~r~u~k~o~w~i~c~z$^1$,~~M.~~R~a~t~a~j~c~z~a~k$^1$,~~I.~~S~o~s~z~y~\'n~s~k~i$^1$,\\
A.~~U~d~a~l~s~k~i$^1$,~~M.~K.~~S~z~y~m~a~\'n~s~k~i$^1$,~~K.~~U~l~a~c~z~y~k$^2$,\\
R.~~P~o~l~e~s~k~i$^1$,~~S.~~K~o~z~{\l}~o~w~s~k~i$^1$,~~J.~~S~k~o~w~r~o~n$^1$,\\
D.~M.~~S~k~o~w~r~o~n$^1$,~~P.~~M~r~\'o~z$^1$,~~K.~~R~y~b~i~c~k~i$^{1,3}$,~~P.~~I~w~a~n~e~k$^1$,\\
M.~~W~r~o~n~a$^1$,~~and~~M.~~G~r~o~m~a~d~z~k~i$^1$\\}
\vskip3mm
{
$^1$ Astronomical Observatory, University of Warsaw, Al. Ujazdowskie 4, 00-478 Warszawa, Poland\\
$^2$ Department of Physics, University of Warwick, Coventry CV4 7AL, UK\\
$^3$ Department of Particle Physics and Astrophysics, Weizmann Institute of Science, Rehovot 76100, Israel\\
}
\end{center}
\Abstract{We report the classification of 24 puzzling short-period variable stars
located in OGLE-IV Galactic bulge fields. The stars are low-amplitude ($<0.05$ mag)
multi-periodic objects with dominant periods between 22 and 54 min whose type could
not have been unambiguously established based on photometry only. A low-resolution
spectroscopic follow-up has shown that all the objects are main sequence A/F-type
stars. Thus, all the variables are $\delta$ Sct-type pulsators.
We have added them to the OGLE-IV Collection of Variable Stars.}
{Catalogs -- Stars: variables: $\delta$~Scuti -- Galaxy: bulge -- Galaxy: disk}
\Section{Introduction}
$\delta$~Sct-type variable stars are pulsating stars with periods
below 0.3~d and $V$-band ($I$-band) amplitudes up to 0.9~mag (0.6~mag).
Most of them are multiple-mode pulsators (Netzel {\it et al.\ } 2022). They have
spectral types from A0 to F6 and luminosity classes from III (giants),
through IV (subgiants), to V (dwarfs). In the Hertzsprung-Russell
diagram, $\delta$~Sct stars reside in the part where the classical
instability strip intersects the main sequence (MS). They can also be found
in the pre-MS and post-MS stages. Space-based observations from the Kepler
satellite revealed that no more than 70 per cent of all stars in that part
pulsate (Murphy {\it et al.\ } 2019). The Kepler data also showed that the majority
of $\delta$~Sct stars are low-amplitude pulsators ($<0.1$~mag)
and only about half of them have amplitudes $>0.001$~mag and appear to be
variable in ground-based observations (Balona and Dziembowski 2011).
The upper bound of 0.3~d (7.2~h) for fundamental-mode $\delta$~Sct stars
was arbitrarily set due to scarcity of Population I classical pulsators with periods
in the range 0.3--1~d. Stars with periods longer than 0.3~d are classified
as classical Cepheids (Soszy\'nski {\it et al.\ } 2008). Until the early 2010s,
shortest periods of the known $\delta$~Sct-type stars were of about 25--30 min
({\it e.g.},\, 25.54 min in $\omega$~Cen V294, 26.25 min in $\omega$~Cen V295 -- Kaluzny {\it et al.\ } 2004).
Holdsworth {\it et al.\ } (2014) searched for high-frequency signals in SuperWASP data
and reported on the detection of over 350 candidates for short-period pulsators
including $\delta$~Sct stars with periods in the range 6--29 min.
Classification of variable stars with periods below 1~h ($\approx0.042$~d)
purely based on photometry is often very difficult due to low amplitudes
and sinusoidal shapes of the light curves. A position in the color-magnitude
diagram is usually insufficient to indicate the true variability type,
particularly in the case of heavily reddened objects located close to the Galactic plane.
There are many types of pulsating stars with periods below 1~h. Among them there
are various types of white dwarfs ({\it e.g.},\, ZZ Cet, V777 Her -- Winget and Kepler 2008),
pre-white dwarfs ({\it e.g.},\, EL CVn -- Maxted {\it et al.\ } 2013), hot subdwarfs
({\it e.g.},\, V361 Hya, V1093 Her -- Heber {\it et al.\ } 2016), Blue Large-Amplitude Pulsators
(Pietrukowicz {\it et al.\ } 2017, Kupfer {\it et al.\ } 2019), objects from the GW Vir family
including planetary nebulae nuclei variables (PNNV, Quirion {\it et al.\ } 2007),
$\delta$~Sct stars (Rodr\'iguez {\it et al.\ } 2000), and their Population II
analogues, SX Phe stars (Rodr\'iguez and L\'opez-Gonz\'alez 2000).
There is also a group of rapidly oscillating Ap (roAp) variables showing light
variations in the range 4.7--25.8 min and amplitudes of millimagnitudes or
smaller (Kurtz 2022). Recent analysis of TESS observations indicates that these
variables should be treated as normal $\delta$~Sct stars and the term ``roAp''
should be avoided (Balona 2022). Definitive classification of short-period
variables often requires a spectroscopic confirmation. This is the
case of objects presented in the work here.
Searches for short-period variables in 121 OGLE-IV fields covering the inner
Galactic bulge and central part of the Sagittarius Dwarf Spheroidal Galaxy brought
the identification of a total of 10~140 $\delta$~Sct stars with periods in the range
0.033--0.3~d (47.7~min--7.2~h) (Pietrukowicz {\it et al.\ } 2020, Soszy\'nski {\it et al.\ } 2021).
The main criteria used in the classification process were the period value and asymmetry
in the light curve shape. In addition to this set, 157 extremely short variables
with periods in the range 0.01--0.04~d (14.4--57.6~min) were found but could not have
been unambiguously classified. From this sample, 24 stars with sinusoidal variations
stable over years and located in the upper part of the MS in the
color-magnitude diagrams were subject of our spectroscopic follow-up. In Section~2
we provide details on the photometric and spectroscopic observations. Information
on the 24 targets selected for the spectroscopic campaign can be found in Section~3,
while the results are presented in Section~4.
\Section{Observations}
Photometric observations were obtained in the framework of a long-term variability
survey, the Optical Gravitational Lensing Experiment (OGLE), conducted on the 1.3-m
Warsaw telescope at Las Campanas Observatory, Chile. The analyzed in this work data
were collected during the fourth phase of the survey (OGLE-IV) in years 2010--2019.
OGLE monitors the Milky Way's bulge and disk and Magellanic Clouds in searches
for periodic and irregular brightness variations as well as for transient events.
The entire monitored area covers about 3600 deg$^2$. The OGLE-IV camera is a
mosaic of 32 CCD detectors covering a field of 1.4 deg$^2$ at a scale of
0.26 arcsec/pixel. Most of the observations were taken through the Cousins $I$ filter.
The remaining data were taken through the Johnson $V$ filter to secure color information.
Over the whole decade in the area of the inner Galactic bulge, OGLE collected about
181~500 $I$-band frames and 6400 $V$-band frames. The number of $I$-band measurements
per object varies from 60 up to 16~800, depending on the field. The exposure times
were of 100~s in $I$ and 150~s in $V$. In the case of three most crowded fields,
BLG501, BLG505, and BLG512, the cadence of $I$-band observations was as short as
19 min. Time-series photometry was extracted using the Difference Image Analysis
(DIA -- Wo\'zniak 2000), a technique developed especially for dense stellar fields
(Alard and Lupton 1998). Observations of the inner bulge cover a magnitude range
$12.5<I<21.5$. Technical details on the OGLE survey can be found in Udalski {\it et al.\ } (2015).
The OGLE-IV Collection of Variable Stars (OCVS) contains over one million
objects of various variability types: pulsating stars ({\it e.g.},\, Soszy\'nski {\it et al.\ } 2019),
spotted stars ({\it e.g.},\, Iwanek {\it et al.\ } 2019), eclipsing and ellipsoidal binaries
({\it e.g.},\, Pawlak {\it et al.\ } 2016, Wrona {\it et al.\ } 2022), cataclysmic systems ({\it e.g.},\, Mr\'oz {\it et al.\ } 2015).
Despite the huge number of known variable stars, some of newly detected objects
are puzzling and require spectroscopic investigation to properly classify them
and to understand their nature.
Low-resolution spectroscopic observations were conducted under ESO programme 105.20EF.001
in remote visitor mode (due to Covid-19 pandemic restrictions) over four consecutive
nights in gray time, from 2021 June 30/July~1 to July 3/4. The spectra were
collected with the ESO Faint Object Spectrograph and Camera 2 (EFOSC2) mounted
at the Nasmyth focus of the 3.58-m New Technology Telescope (NTT) located at
La Silla Observatory, Chile. Information about the characteristics of this instrument
are provided in Buzzoni {\it et al.\ } (1984). Thin cirrus clouds were present during the
whole observing run but the first half of the third night when they were thick.
All the spectra were taken with grism \#4 covering wavelengths 4085--7520~$\rm \AA$ at
the slit width of $1\makebox[0pt][l]{.}\arcs0$ and $2\times2$ binning readout, which provided a spectral
resolution of about 11~$\rm \AA$ at 5000~$\rm \AA$. For most of the objects the slit
was aligned to the parallactic angle, but in some cases the slit was oriented properly
to avoid light from neighboring stars. Exposure times were calculated with the help
of the ESO Exposure Time Calculator\footnote{https://www.eso.org/observing/etc/}
to obtain spectra with a signal-to-noise ratio of 50. For wavelength calibrations,
the He-Ar lamp was used. Three spectrophotometric standard stars were selected
for the flux calibrations. Bias and dome flat-field images were taken at dawn.
All spectra were reduced using the IRAF package\footnote{IRAF was distributed
by the National Optical Astronomy Observatory, USA, which is operated by the
Association of Universities for Research in Astronomy, Inc., under a cooperative
agreement with the National Science Foundation.}. Debiasing, flat-fielding,
wavelength and flux calibrations were performed in the standard way.
\Section{Target objects for the spectroscopic follow-up}
In Table~1, we provide information on the 24 stars selected for our spectroscopic
run. The stars are located toward the Galactic bulge. They have
magnitudes in the range $13.1<I<15.6$ and $14.2<V<17.5$. All the stars are
multi-periodic variables with amplitudes below 0.05 mag in $I$.
We list up to five detected periodicities. The periods were determined
with the help of the TATRY code (Schwarzenberg-Czerny 1996) after subsequent
pre-whitenings.
\begin{sidewaystable}
\centering \caption{\small Photometric parameters of the 24 target stars}
\medskip
{\footnotesize
\begin{tabular}{ccccclllll}
\hline
Name & RA(2000) & Dec(2000) & $\langle I \rangle$ & $\langle V \rangle$ & $P_1$ & $P_2$ & $P_3$ & $P_4$ & $P_5$ \\
OGLE- & & & [mag] & [mag] & [d] & [d] & [d] & [d] & [d] \\
BLG-DSCT- & & & & & & & & & \\
\hline
16832 & $17\uph22\upm51\makebox[0pt][l]{.}\ups74$ & $-30\arcd16\arcm23\makebox[0pt][l]{.}\arcs1$ & 15.53 & 17.49 & 0.033531956(24) & 0.036224500(35) & 0.042320454(45) & - & - \\
16833 & $17\uph32\upm20\makebox[0pt][l]{.}\ups04$ & $-22\arcd05\arcm49\makebox[0pt][l]{.}\arcs1$ & 13.10 & 14.20 & 0.034233043(20) & 0.039879897(52) & 0.016609671(15) & - & - \\
16834 & $17\uph35\upm57\makebox[0pt][l]{.}\ups82$ & $-26\arcd54\arcm50\makebox[0pt][l]{.}\arcs6$ & 14.34 & 15.72 & 0.035165296(11) & 0.038645168(18) & 0.040651018(27) & 0.045402275(52) & 0.035165937(64) \\
16835 & $17\uph38\upm33\makebox[0pt][l]{.}\ups75$ & $-34\arcd09\arcm26\makebox[0pt][l]{.}\arcs4$ & 14.85 & 16.65 & 0.035719287(23) & 0.039362439(60) & 0.08913910(39) & 0.043655649(97) & 0.04184227(10) \\
16836 & $17\uph44\upm34\makebox[0pt][l]{.}\ups25$ & $-33\arcd17\arcm56\makebox[0pt][l]{.}\arcs1$ & 15.28 & 16.76 & 0.035541111(20) & 0.09749517(25) & 0.051245588(88) & - & - \\
16837 & $17\uph47\upm50\makebox[0pt][l]{.}\ups67$ & $-20\arcd48\arcm46\makebox[0pt][l]{.}\arcs4$ & 15.17 & 16.53 & 0.037076495(22) & 0.046284106(35) & 0.04357615(11) & - & - \\
16838 & $17\uph50\upm02\makebox[0pt][l]{.}\ups34$ & $-33\arcd18\arcm59\makebox[0pt][l]{.}\arcs5$ & 14.48 & 15.71 & 0.033961025(13) & 0.037164189(44) & 0.036201125(56) & - & - \\
16839 & $17\uph50\upm31\makebox[0pt][l]{.}\ups18$ & $-29\arcd28\arcm06\makebox[0pt][l]{.}\arcs3$ & 15.06 & 16.30 & 0.025409872(10) & 0.027650646(25) & 0.029687043(47) & 0.032681617(62) & 0.029680207(58) \\
16840 & $17\uph51\upm35\makebox[0pt][l]{.}\ups11$ & $-28\arcd44\arcm39\makebox[0pt][l]{.}\arcs2$ & 15.11 & 16.10 & 0.0278339039(50) & 0.032812117(21) & 0.06660964(12) & 0.027834329(29) & 0.047779888(96) \\
16841 & $17\uph51\upm46\makebox[0pt][l]{.}\ups46$ & $-28\arcd26\arcm05\makebox[0pt][l]{.}\arcs3$ & 15.08 & 15.93 & 0.0190090983(40) & 0.0201545952(54) & 0.0201278558(72) & 0.0201484003(80) & 0.0213994311(96) \\
16842 & $17\uph51\upm55\makebox[0pt][l]{.}\ups89$ & $-30\arcd17\arcm18\makebox[0pt][l]{.}\arcs9$ & 14.66 & 15.79 & 0.0255401151(79) & 0.027133418(24) & 0.027149755(35) & 0.028792153(40) & - \\
16843 & $17\uph52\upm17\makebox[0pt][l]{.}\ups68$ & $-32\arcd42\arcm06\makebox[0pt][l]{.}\arcs3$ & 13.82 & 14.84 & 0.028433525(10) & 0.030864635(36) & - & - & - \\
16844 & $17\uph54\upm35\makebox[0pt][l]{.}\ups55$ & $-29\arcd48\arcm24\makebox[0pt][l]{.}\arcs4$ & 15.21 & 16.17 & 0.033516366(13) & 0.036939131(45) & 0.04307991(13) & 0.04435281(15) & 0.04541343(15) \\
16845 & $17\uph58\upm16\makebox[0pt][l]{.}\ups37$ & $-27\arcd54\arcm43\makebox[0pt][l]{.}\arcs8$ & 15.13 & 16.67 & 0.029935884(10) & 0.034300425(12) & 0.029934575(21) & 0.039805145(49) & 0.037234447(61) \\
16846 & $17\uph58\upm30\makebox[0pt][l]{.}\ups84$ & $-28\arcd25\arcm46\makebox[0pt][l]{.}\arcs4$ & 13.86 & 14.62 & 0.0230617618(57) & 0.0198797600(60) & 0.053054509(54) & 0.027531088(26) & 0.024992040(24) \\
16847 & $17\uph58\upm58\makebox[0pt][l]{.}\ups94$ & $-28\arcd03\arcm42\makebox[0pt][l]{.}\arcs6$ & 14.76 & 15.85 & 0.036425934(12) & 0.042694792(45) & 0.048473768(69) & 0.09609703(28) & 0.07844828(21) \\
16848 & $17\uph58\upm59\makebox[0pt][l]{.}\ups19$ & $-28\arcd45\arcm44\makebox[0pt][l]{.}\arcs2$ & 14.61 & 15.48 & 0.030000956(11) & 0.031942458(16) & - & - & - \\
16849 & $18\uph00\upm49\makebox[0pt][l]{.}\ups06$ & $-27\arcd35\arcm26\makebox[0pt][l]{.}\arcs0$ & 13.82 & 14.59 & 0.0311231198(67) & 0.041716525(14) & 0.035578305(13) & 0.033469123(20) & 0.036362040(27) \\
16850 & $18\uph02\upm25\makebox[0pt][l]{.}\ups70$ & $-27\arcd02\arcm23\makebox[0pt][l]{.}\arcs5$ & 14.48 & 15.41 & 0.0158377216(23) & 0.0166989591(23) & 0.0150967282(37) & 0.0167467952(47) & 0.0158803976(48) \\
16851 & $18\uph03\upm25\makebox[0pt][l]{.}\ups04$ & $-27\arcd01\arcm41\makebox[0pt][l]{.}\arcs7$ & 14.78 & 15.61 & 0.0168271841(31) & 0.0189019685(57) & 0.017930385(11) & 0.018901708(12) & 0.019266537(18) \\
16852 & $18\uph05\upm30\makebox[0pt][l]{.}\ups36$ & $-28\arcd28\arcm49\makebox[0pt][l]{.}\arcs4$ & 15.57 & 16.68 & 0.036513655(16) & 0.042300103(34) & 0.046980198(57) & 0.036513303(41) & - \\
16853 & $18\uph15\upm03\makebox[0pt][l]{.}\ups77$ & $-27\arcd33\arcm06\makebox[0pt][l]{.}\arcs7$ & 14.37 & 14.92 & 0.034369987(13) & 0.022029517(14) & - & - & - \\
16854 & $18\uph16\upm13\makebox[0pt][l]{.}\ups66$ & $-26\arcd53\arcm15\makebox[0pt][l]{.}\arcs3$ & 13.67 & 14.40 & 0.036765128(19) & 0.048780768(51) & 0.039913311(33) & 0.047063512(55) & 0.020849953(31) \\
16855 & $18\uph20\upm37\makebox[0pt][l]{.}\ups13$ & $-23\arcd00\arcm08\makebox[0pt][l]{.}\arcs0$ & 14.62 & 15.64 & 0.034885725(23) & 0.037769423(12) & 0.0361497383(56) & - & - \\
\hline
\end{tabular}}
\end{sidewaystable}
\Section{Results from the spectroscopic observations}
\begin{figure}
\centerline{\includegraphics[angle=90,width=105mm]{fig1.ps}}
\FigCap{Phased OGLE $I$-band light curves (left column), $V-I$ color {\it vs.\ } $I$
magnitude diagrams (middle column), and EFOSC2 low-resolution spectra (right column)
of 12 (out of 24) discovered $\delta$~Sct stars ordered with increasing
pulsation period (provided in the left column).}
\end{figure}
\begin{figure}
\centerline{\includegraphics[angle=90,width=105mm]{fig2.ps}}
\FigCap{Same as Figs.~1 for the remaining $\delta$~Sct stars.}
\end{figure}
In Figs.~1--2, for each star, we present the OGLE-IV $I$-band light curve
folded with the dominant period, a color-magnitude diagram with the position
of the star, and the obtained EFOSC2 spectrum. The spectra of all stars are
dominated by the Balmer series, which indicates type A. In all the stars,
one can also spot the blended sodium doublet Na~I (D1: 5896~$\rm \AA$,
D2: 5890~$\rm \AA$), which is primarily of interstellar origin for this
spectral type.
A detailed analysis of the spectra was carried out with the iSpec package
(Blanco-Cuaresma {\it et al.\ } 2014, Blanco-Cuaresma 2019). To determine atmospheric
parameters, the effective temperature $T_{\rm eff}$ and surface gravity $\log g$,
we used the MARCS model atmosphere (Castelli and Kurucz 2003), Synthe radiative
code (Kurucz 2005), solar abundances provided in Asplund {\it et al.\ } (2009), and VALD atomic
line list (Piskunov {\it et al.\ } 1995). Due to the low spectral resolution, our analysis was based
on the Balmer lines only (H$\alpha$, H$\beta$, and H$\gamma$). The parameters $T_{\rm eff}$
and $\log g$ were fitted with starting values of 7000 K and 4.0, respectively.
Microturbulence velocity was fixed at 2 km/s, while macroturbulence velocity at 4 km/s.
The rotational velocity $v \sin i$ was assumed to be of 5 km/s and the metallicity [M/H]
of 0.0 dex. Results of the analysis are collected in Table~2. In Fig.~3,
we present normalized spectra for three shortest-period $\delta$~Sct stars
in our sample with models obtained from the spectral synthesis.
\begin{table}[h]
\begin{center}
\caption{\small Atmospheric parameters of the observed stars}
\medskip
{\footnotesize
\begin{tabular}{ccc|ccc}
\hline
Name & $T_{\rm eff}$ & $\log g$ & Name & $T_{\rm eff}$ & $\log g$ \\
OGLE- & [K] & & OGLE- & [K] & \\
BLG-DSCT- & & & BLG-DSCT- & & \\
\hline
16832 & 7950 $\pm$ 260 & 3.69 $\pm$ 1.23 & 16844 & 7450 $\pm$ 220 & 4.31 $\pm$ 0.92 \\
16833 & 7610 $\pm$ 240 & 3.79 $\pm$ 1.03 & 16845 & 8170 $\pm$ 370 & 4.05 $\pm$ 1.05 \\
16834 & 7840 $\pm$ 190 & 4.68 $\pm$ 1.12 & 16846 & 8090 $\pm$ 310 & 4.73 $\pm$ 1.07 \\
16835 & 7570 $\pm$ 210 & 3.71 $\pm$ 0.64 & 16847 & 7770 $\pm$ 270 & 3.93 $\pm$ 0.83 \\
16836 & 8010 $\pm$ 220 & 4.35 $\pm$ 0.85 & 16848 & 8710 $\pm$ 360 & 4.08 $\pm$ 0.61 \\
16837 & 8230 $\pm$ 380 & 3.34 $\pm$ 0.58 & 16849 & 8130 $\pm$ 280 & 4.14 $\pm$ 0.91 \\
16838 & 7650 $\pm$ 300 & 4.90 $\pm$ 0.82 & 16850 & 8750 $\pm$ 500 & 4.47 $\pm$ 0.65 \\
16839 & 7730 $\pm$ 460 & 4.91 $\pm$ 0.80 & 16851 & 8930 $\pm$ 370 & 4.52 $\pm$ 0.39 \\
16840 & 8095 $\pm$ 183 & 4.97 $\pm$ 0.88 & 16852 & 7870 $\pm$ 300 & 4.08 $\pm$ 0.69 \\
16841 & 8410 $\pm$ 430 & 4.27 $\pm$ 1.06 & 16853 & 9480 $\pm$ 420 & 3.81 $\pm$ 0.62 \\
16842 & 8160 $\pm$ 290 & 4.59 $\pm$ 0.91 & 16854 & 7370 $\pm$ 230 & 4.34 $\pm$ 0.96 \\
16843 & 8390 $\pm$ 410 & 4.11 $\pm$ 0.83 & 16855 & 9180 $\pm$ 410 & 4.31 $\pm$ 0.68 \\
\hline
\end{tabular}}
\end{center}
\end{table}
\begin{figure}
\centerline{\includegraphics[angle=0,width=130mm]{fig3.ps}}
\FigCap{Fits to hydrogen lines in three $\delta$~Sct stars with periods below 30 min.}
\end{figure}
Fig.~4 shows the location of the 24 stars in the $T_{\rm eff}$ {\it vs.\ } $\log g$ plane.
All the objects have the atmospheric parameters typical for the main sequence A/F stars.
However, with the used spectral resolution, we cannot exclude that some objects
slightly evolved off the MS. Nevertheless, the spectra confirm that
all the stars are located in the area populated by $\delta$~Sct-type variables
in the Hertzsprung-Russell diagram.
We note that only one object, OGLE-BLG-DSCT-16883, was reported as
a variable source in Gaia DR3 (Gaia Collaboration 2021),
but it was incorrectly classified as a spotted star (of RS CVn type).
Measured parallaxes of 23 stars are between $0.20\pm0.06$ mas and $0.82\pm0.02$ mas,
which places them likely in the Galactic disk. The parallax of OGLE-BLG-DSCT-16864,
$0.08\pm0.06$ mas, indicates a more distant location, probably the Galactic bulge.
\begin{figure}[htb!]
\centerline{\includegraphics[angle=0,width=130mm]{fig4.ps}}
\FigCap{Positions of the 24 observed stars in the $T_{\rm eff}$ {\it vs.\ } $\log g$ plane. All
the objects are located in the area for MS stars of spectral type A and early F.}
\end{figure}
\Section{On-line Data}
The newly confirmed $\delta$~Sct stars supplement the OCVS.
The objects are arranged according to increasing right ascension and have
names from OGLE-BLG-DSCT-16860 to OGLE-BLG-DSCT-16883. The photometric
as well as spectroscopic data are available to the astronomical community
through the OGLE On-line Data Archive at {\it https://ogle.astrouw.edu.pl}.
\Acknow{
MG is supported by the EU Horizon 2020 research and innovation programme
under grant agreement No 101004719. This work is based on observations
collected at the European Southern Observatory under ESO programme 105.20EF.001.
We used data from the European Space Agency (ESA) mission Gaia,
processed by the Gaia Data Processing and Analysis Consortium (DPAC).
Funding for the DPAC has been provided by national institutions, in particular
the institutions participating in the Gaia Multilateral Agreement.
}
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.